rucc_driver/lib.rs
1//! The driver: command line parsing, the phase graph, job scheduling and the linker
2//! invocation.
3//!
4//! Design: `spec/04-driver-and-cli.md`. Layer rank 13, see `spec/18-package-layout.md`.
5//!
6//! This is the only crate that is allowed to know the process exists. It reads the command
7//! line, touches the file system, spawns the linker and writes to the terminal, and it hands
8//! everything below it a [`Session`]. The binary crate is a `main` that calls
9//! [`run`] and nothing else, so that the whole driver is reachable from a test.
10//!
11//! # Status
12//!
13//! `--help`, `--version` and `--print-config` are real, which is the `M0` exit criterion in
14//! `spec/17-milestones.md`. The phase graph is real and `-###` prints it, and the scheduler
15//! that will run it is real and tested.
16//!
17//! Two phases run. `-E` reads the file, runs phase 4 over it and writes the result, to `-o` or
18//! to standard output. `--emit=tast` carries on through phase 7, the parse and the checking,
19//! and writes the typed tree. The flags those two read are real with them, which is `-D`, `-U`,
20//! `-I`, `-I-`, `-iquote`, `-isystem`, `-idirafter`, `-iprefix`, `-iwithprefix`,
21//! `-iwithprefixbefore`, `-include`, `-imacros`, `--sysroot=`, `-isysroot`, `-P`, `-std=`,
22//! `-fgnuc-version=`, `-ansi`, `-ffreestanding`, `-fno-builtin`, `-fno-builtin-<name>`,
23//! `-fgnu89-inline`, `-pedantic` and `-Werror`.
24//! The phases after them still say they are not implemented.
25//!
26//! This crate is tier 3 in `spec/18-package-layout.md` section 18.5: its Rust API is
27//! explicitly unstable and will change without a major version bump.
28
29#![doc(html_root_url = "https://docs.rs/rucc-driver/0.15.3")]
30
31pub mod assemble;
32pub mod cache;
33pub mod compile;
34pub mod deps;
35pub mod fetch;
36mod glibc;
37pub mod install;
38pub mod library;
39pub mod link;
40mod map;
41pub mod msvc;
42pub mod phase;
43pub mod preprocess;
44pub mod schedule;
45mod shapes;
46pub mod trace;
47mod warnings;
48
49use std::fmt::Write as _;
50use std::io::Write as _;
51use std::path::PathBuf;
52
53use rucc_codegen::coverage::{self, Fired};
54use rucc_codegen::lowering::Lowerings;
55use rucc_codegen::pressure::Pressure;
56use rucc_pp::Dependency;
57use rucc_session::{
58 Compress, Control, Dumps, EmitKind, Hook, Math, Options, Pic, PrefixMap, Preinclude, Protector,
59 SaveTemps, Session, Std, Wrapping, runtime,
60};
61use rucc_sysroot::{Manifest, Sysroot};
62use rucc_target::{ObjectFormat, Triple};
63use rucc_tuple::TargetTuple;
64
65use crate::link::LinkOptions;
66
67pub use crate::assemble::assemble;
68pub use crate::compile::{Artifact, Compiled, Temps, compile, compile_ir};
69pub use crate::phase::{ArchiveJob, Input, InputKind, Job, LinkJob, Output, Phase, Plan, Role};
70pub use crate::preprocess::{OsFileSystem, Preprocessed, preprocess};
71pub use crate::schedule::Jobs;
72
73/// The compiler's version, taken from the workspace manifest.
74pub const VERSION: &str = env!("CARGO_PKG_VERSION");
75
76/// What the command line asked for.
77#[derive(Debug, Clone, PartialEq, Eq)]
78pub enum Action {
79 /// Print usage and exit successfully.
80 Help,
81 /// Print the version and exit successfully.
82 Version,
83 /// Print one line and exit successfully, which is what the `-dump` and `-print` family do.
84 ///
85 /// A build system asks these before it compiles anything, and what it does with the answer
86 /// is paste it into a path or into another command line, so each one is a single line with
87 /// no decoration around it.
88 Print(String),
89 /// Print the resolved configuration and exit successfully.
90 PrintConfig(Box<Options>),
91 /// Print the passes the level will run and exit successfully.
92 PrintPipeline(Box<Options>),
93 /// Print the phase plan and the link line and exit successfully, which is `-###`.
94 PrintPlan {
95 /// The resolved options, which is what says what the link line is for.
96 opts: Box<Options>,
97 /// What to do to each input, and in what order.
98 plan: Box<Plan>,
99 /// What the command line said about linking.
100 link: Box<LinkOptions>,
101 },
102 /// `--fetch <tuple>`, which gets the sysroot this release pins for a target and installs it.
103 ///
104 /// The only action in this compiler that may run another program to move bytes onto the
105 /// machine, which is `spec/cross-compile/13-distribution.md` section 13.8's rule rather than a
106 /// property of how this happens to be written: a compilation has no branch that reaches it.
107 Fetch {
108 /// The artifact, from the table in [`rucc_sysroot::artifact`]. Resolved here rather than where the
109 /// work happens, so that a target nothing is pinned for is a refusal from the parser like
110 /// every other thing a command line can ask for and not have.
111 what: &'static rucc_sysroot::Pinned,
112 /// The target, which names the directory under the cache the tree is installed at and is
113 /// checked against the record inside the artifact.
114 target: TargetTuple,
115 /// Where the cache is, read where everything else that needs it reads it.
116 cache: PathBuf,
117 },
118 /// `--fetch-msvc-sdk <tuple>`, which gets what is behind Microsoft's licence wall.
119 ///
120 /// The other action that may run another program to move bytes onto the machine, and the only
121 /// one that asks a person to accept somebody else's licence first.
122 /// `spec/cross-compile/13-distribution.md` section 13.4 is why no release pins an artifact
123 /// for these, and nothing about this may ever happen because a compile wanted it to. `--fetch`
124 /// of an MSVC target is this action too, starting from the build this release pins rather than
125 /// from the one Microsoft's channel names today.
126 FetchMsvcSdk {
127 /// The target, which says which architecture's CRT library package is wanted.
128 target: TargetTuple,
129 /// Whether `--accept-licence` was on the command line. Without it the licence and the list
130 /// are printed and nothing is downloaded, which is the whole of what the flag is for.
131 accepted: bool,
132 /// Where the cache is, read where everything else that needs it reads it.
133 cache: PathBuf,
134 /// Whether the documents at the top of the chain are [`rucc_sysroot::PINNED_BUILD`]'s,
135 /// which is `--fetch`, rather than the current channel's, which is `--fetch-msvc-sdk`.
136 pinned: bool,
137 },
138 /// Compile the given inputs.
139 Compile {
140 /// The resolved options.
141 opts: Box<Options>,
142 /// What to do to each input, and in what order.
143 plan: Box<Plan>,
144 /// What the command line said about linking.
145 link: Box<LinkOptions>,
146 /// How many translation units to compile at once.
147 jobs: Jobs,
148 /// Whether `-v` asked for the plan to be printed while it runs.
149 verbose: bool,
150 /// What is worth saying about the command line before anything is compiled, printed as
151 /// warnings and once for the whole run rather than once per file.
152 ///
153 /// These are not diagnostics. A diagnostic is about a piece of source and has a span to
154 /// point at, and these are about the way two flags were combined, so there is nothing to
155 /// point at and nowhere below the driver that knows both halves. `-w` does not reach them
156 /// for the same reason it does not reach a refusal from the parser.
157 notes: Vec<String>,
158 },
159}
160
161/// Why a command line was rejected.
162#[derive(Debug, Clone, PartialEq, Eq)]
163pub struct CliError {
164 /// The message, lowercase and without a trailing period, in the same shape as any other
165 /// diagnostic.
166 pub message: String,
167}
168
169impl std::fmt::Display for CliError {
170 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
171 f.write_str(&self.message)
172 }
173}
174
175impl std::error::Error for CliError {}
176
177fn err(message: impl Into<String>) -> CliError {
178 CliError { message: message.into() }
179}
180
181/// The two halves of one prefix mapping flag's argument, where `flag` includes its trailing `=`.
182///
183/// The split is at the last `=` in what follows the flag, not the first, which is gcc's rule and
184/// the only one that lets a directory whose name contains an `=` be the old half. It also means
185/// `-fmacro-prefix-map=a=b=c` rewrites `a=b` to `c` rather than `a` to `b=c`, which looks like a
186/// trap until you notice the alternative traps the far more common case.
187fn rewrite<'a>(arg: &'a str, flag: &str) -> Result<(&'a str, &'a str), CliError> {
188 let rest = &arg[flag.len()..];
189 PrefixMap::split(rest).ok_or_else(|| {
190 let flag = flag.trim_end_matches('=');
191 err(format!(
192 "`{rest}` is not a rewrite for `{flag}`, which is an old prefix, an `=` and a new one"
193 ))
194 })
195}
196
197/// A question the command line asked instead of asking for a compilation.
198///
199/// These are answered after the loop rather than where they are read, because every one of them
200/// is about the target or about the library search and the last word on both is the end of the
201/// command line.
202enum Query {
203 /// `-dumpmachine`, the triple.
204 Machine,
205 /// `-dumpversion`, the major number of the GCC release this compiler claims to be.
206 Version,
207 /// `-dumpfullversion`, the same release in all three numbers.
208 FullVersion,
209 /// `-print-multiarch`, the directory name a distribution files this target under.
210 Multiarch,
211 /// `-print-search-dirs`, in the three lines GCC prints.
212 SearchDirs,
213 /// `-print-sysroot`, the root the headers and the libraries are read under.
214 Sysroot,
215 /// `-print-sysroot-provenance`, what is in that root and where each of it came from.
216 SysrootProvenance,
217 /// `-print-sysroot-digest`, the one number that names all of it.
218 SysrootDigest,
219 /// `-print-file-name=<name>`, the full path of a library file.
220 FileName(String),
221 /// `-print-prog-name=<name>`, the full path of a program.
222 ProgName(String),
223 /// `-print-libgcc-file-name`, which is `-print-file-name=libgcc.a` under another spelling.
224 Libgcc,
225}
226
227/// Usage text.
228///
229/// Deliberately short. `spec/04-driver-and-cli.md` puts the full flag reference in the
230/// manual page, because a `--help` nobody can read in one screen is a `--help` nobody reads.
231pub const USAGE: &str = "\
232rucc, an optimizing C compiler
233
234usage: rucc [options] file...
235
236options:
237 -c compile and assemble, do not link
238 -S compile only, emit assembly
239 -E preprocess only
240 -o <file> write output to <file>, or to standard output for -
241 -D <name>[=<value>], -U <name> define a macro, or undefine one after every -D
242 -I <dir> add <dir> to the include search path
243 -iquote -isystem -idirafter <dir> the other chains, -nostdinc drops ours
244 -I-, -iprefix <p>, -iwithprefix[before] <dir> the older spellings of those
245 -include <file>, -imacros <file> read <file> first, the second for its macros only
246 --sysroot=<dir> look for the library's headers under <dir>, -isysroot too
247 -P, -dM with -E: leave out the markers, or dump the macros
248 -M -MM -MD -MMD write a make rule for the source, the last two compile as well
249 -MF <file> -MT <t> -MQ <t> -MP where the rule goes, what it builds, targets with no recipe
250 -std=<dialect> c89 through c2y, and the gnu spellings
251 -fgnuc-version=<v> -fms-compatibility-version=<v> the GCC (16.0.0) or MSVC (19.40) to claim
252 -x <lang> treat later inputs as <lang>, or none to stop
253 -O<level> optimize: 0, 1, 2, 3, s, z, fast
254 -fsafety=<tier> check memory safety: off, detect, enforce, kernel
255 -f[no-]sanitize=<what> the negative is taken, the positive is refused by name
256 -f[no-]safety-subobject a write has to stay inside the member it names
257 -f[no-]safety-restrict two restrict pointers of one block may not meet
258 -f<pass> -fno-<pass> -fdump-ir=<what> -fopt-info[-<kind>][=FILE]
259 -fpass-fuel=<pass>=<n>, -fpass-fuel-global=<n> stop a pass, or all of them, after n
260 -fdisable-<pass>[=<funcs>], -fenable-<pass>[=<funcs>] run a pass on some functions only
261 -g -g0 -gdwarf-5, -fno-omit-frame-pointer, -mno-red-zone debug info, frame pointer, red zone
262 -gz[=none|zlib|zlib-gnu|zstd] -gno-split-dwarf compress debug sections, one file not two
263 -flto[=auto|jobserver|<n>] -fno-lto -ffat-lto-objects read, and not done yet
264 -fprofile-use[=<path>] -fprofile-dir=<dir> read too, where -fprofile-generate is refused
265 -f[no-]stack-protector[-strong|-all], -f[no-]stack-clash-protection, -fcf-protection=<edges>
266 -ffunction-sections -fdata-sections a section per function or variable, for --gc-sections
267 -fvisibility=<what> default, hidden, internal or protected, when nothing in the source said
268 -l<name>, -L <dir>, -B <dir> link a library, where to look for one, where our own tools are
269 -fPIC -fpic -fPIE -fpie, -pipe what it does anyway, and -f[no-]common as the target's cc
270 -f[no-]strict-aliasing, -f[no-]delete-null-pointer-checks what it assumes anyway
271 -static -shared -pie -no-pie -nostdlib -nostartfiles -nodefaultlibs -rdynamic -s how to link
272 -Wl,<arg>, -Xlinker <arg>, -fuse-ld=<name> hand an argument to the linker, or pick one
273 -Werror -pedantic -pedantic-errors -w -W[no-]system-headers how much to say, and how fatal
274 -m64 -march= -mtune= -mcpu= -mabi= -mcmodel= what machine to generate for
275 -pg -p, -mfentry -mno-fentry call a profiler on the way in, and where that call goes
276 -fpatchable-function-entry=<n>[,<m>] room at the top of every function to patch later
277 -fwrapv, -fwrapv-pointer, -fno-strict-overflow, -ftrapv overflow wraps, or stops the program
278 -f[no-]exceptions, -f[no-]non-call-exceptions let an exception unwind through the code
279 -f[no-]signed-char, -f[no-]unsigned-char, -f[no-]short-enums change the ABI
280 -ffp-contract=<how> fuse a multiply and an addition: fast, on or off
281 -f[no-]fast-math and each of its members, -f[no-]rounding-math, -fexcess-precision=<how>
282 -ffile-prefix-map=<old>=<new> rewrite that front of every path we put in the output
283 -fmacro-prefix-map= -fdebug-prefix-map= -fprofile-prefix-map= the same, one output each
284 -pthread build for more than one thread, and link the library for it
285 -dumpmachine -dumpversion -print-multiarch -print-search-dirs what this compiler is
286 -print-file-name=<name> -print-prog-name=<name> where a file or a program is
287 -print-sysroot the root the headers and the libraries are read under
288 -print-sysroot-provenance every input under it, where it came from and its licence
289 -print-sysroot-digest the sha256 of that record, which names the whole sysroot in one line
290 --fetch <tuple> get the sysroot this release pins for <tuple> and install it in the cache
291 --fetch-msvc-sdk <tuple> the same for *-windows-msvc, from Microsoft's current build
292 --offline never download anything, which a compilation never does anyway
293 -j[n] compile n translation units at once, default all
294 -v, -### print each phase as it runs, or without running any
295 -save-temps[=cwd|obj], -fstack-usage, -time keep the .i and .s, write a .su, time each step
296 --target=<triple> generate code for <triple>, which a name like <triple>-rucc also does
297 --emit=<kind> exe, obj, archive, asm, preprocessed, tast, ir, mir-final,
298 safety-summary, type-granules
299 --print-config, --print-pipeline print the configuration or the pipeline, and exit
300 --version print the version and exit
301 -h, --help print this message and exit
302
303See spec/04-driver-and-cli.md for the full flag reference.
304";
305
306/// The argument of a flag that may be joined to it or may be the next word.
307///
308/// `-DFOO` and `-D FOO` are the same thing, and `at` is where the flag's own letters end.
309fn joined_or_next(
310 arg: &str,
311 at: usize,
312 args: &[String],
313 i: &mut usize,
314) -> Result<String, CliError> {
315 if arg.len() > at {
316 return Ok(arg[at..].to_owned());
317 }
318 let next = args.get(*i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
319 *i += 1;
320 Ok(next.clone())
321}
322
323/// The smallest boundary a function is put on when the command line asked for no alignment at all.
324///
325/// Eight bytes, which is what gcc 16 gives `-fno-align-functions` on x86-64 and is a boundary every
326/// target this compiler has is happy with. It is not zero: a function still has to start somewhere
327/// an instruction may start, and the flag asks for the target's minimum rather than for none.
328const MIN_FUNC_ALIGN: u32 = 8;
329
330/// What `-falign-functions=N` asks for, as a power of two, or `None` for the target's own answer.
331///
332/// Zero and one both mean the default, which is gcc's reading of them, and everything else is
333/// rounded up to the next power of two, which is also gcc's: `-falign-functions=3` puts a function
334/// on a four byte boundary rather than being refused. Gives back `Err` shaped as an outer `None`
335/// only when the text is not a number, since that is the one thing gcc will not read either. A
336/// number larger than any alignment makes sense at is clamped rather than refused, for the same
337/// reason: this is a preference about speed and a build that wrote a silly one still deserves to
338/// compile.
339fn function_alignment(text: &str) -> Option<Option<u32>> {
340 // gcc takes `N:M:N2:M2`, where everything after the first number is about how far it is willing
341 // to go to reach the boundary. Only the boundary is answerable here, so the rest is read to
342 // check that it is numbers and then dropped.
343 let mut parts = text.split(':');
344 let first = parts.next()?;
345 if parts.any(|part| part.parse::<u64>().is_err()) {
346 return None;
347 }
348 let want: u64 = first.parse().ok()?;
349 if want <= 1 {
350 return Some(None);
351 }
352 let bytes = want.min(1 << 16).next_power_of_two();
353 Some(Some(u32::try_from(bytes).ok()?))
354}
355
356/// Every name that may follow `-fsanitize=`, which is gcc 16's list and three of this compiler's
357/// own.
358///
359/// The three are on it because `spec/07-types-and-semantics.md` section 7.7 already promises them:
360/// each undefined behaviour this compiler exploits is listed there with the check that detects it,
361/// and `alias`, `restrict` and `memory` are checks gcc has no spelling for. gcc refuses `memory`
362/// outright, since the sanitizer of that name is clang's. A name being here means it is a name
363/// rather than a typo, and nothing more than that: every one of them is refused after the loop,
364/// because none of them is implemented.
365///
366/// `all` is deliberately absent. gcc takes it only in the negative, so it is handled where each of
367/// those two spellings is read rather than by being on this list.
368const SANITIZERS: [&str; 34] = [
369 "address",
370 "kernel-address",
371 "hwaddress",
372 "kernel-hwaddress",
373 "pointer-compare",
374 "pointer-subtract",
375 "thread",
376 "leak",
377 "undefined",
378 "shift",
379 "shift-base",
380 "shift-exponent",
381 "integer-divide-by-zero",
382 "unreachable",
383 "vla-bound",
384 "null",
385 "return",
386 "signed-integer-overflow",
387 "bounds",
388 "bounds-strict",
389 "alignment",
390 "object-size",
391 "float-divide-by-zero",
392 "float-cast-overflow",
393 "nonnull-attribute",
394 "returns-nonnull-attribute",
395 "bool",
396 "enum",
397 "vptr",
398 "pointer-overflow",
399 "builtin",
400 "alias",
401 "restrict",
402 "memory",
403];
404
405/// The command line with every `@file` replaced by the words in the file, the way gcc does it.
406///
407/// Meson writes the link of a large target this way, so that a command line holding a thousand
408/// objects stays under the limit the system puts on one. Postgres's `postgres` executable is the
409/// one link in its tree that meson writes as `@postgres.rsp`, and before this the name went to
410/// the linker as it was. GNU ld reads response files itself, so it opened the file and found
411/// `-Wl,--as-needed` in it, which is a driver flag it has never heard of.
412///
413/// The rules are libiberty's `expandargv`, since that is what gcc and every other GNU tool read
414/// these files with. Words are split on white space, a single or a double quote keeps white
415/// space in a word until the matching quote, and a backslash makes the character after it an
416/// ordinary one, inside quotes as well as outside. A word the file gives that starts with `@` is
417/// read as a response file in turn. A name that cannot be opened is left on the command line as
418/// it was, which is what gcc does and which is how a file really called `@x.c` still reaches the
419/// loop, where it is refused as an unknown input rather than swallowed. The depth is capped so a
420/// file that names itself is an error and not a hang.
421fn response_files(args: &[String]) -> Result<Vec<String>, CliError> {
422 const DEEPEST: usize = 64;
423 fn expand(args: &[String], depth: usize, out: &mut Vec<String>) -> Result<(), CliError> {
424 for arg in args {
425 let Some(name) = arg.strip_prefix('@') else {
426 out.push(arg.clone());
427 continue;
428 };
429 let Ok(text) = std::fs::read_to_string(name) else {
430 out.push(arg.clone());
431 continue;
432 };
433 if depth == DEEPEST {
434 return Err(err(format!("response file '{name}' is nested too deeply")));
435 }
436 expand(&response_words(&text), depth + 1, out)?;
437 }
438 Ok(())
439 }
440 if !args.iter().any(|arg| arg.starts_with('@')) {
441 return Ok(args.to_vec());
442 }
443 let mut out = Vec::with_capacity(args.len());
444 expand(args, 0, &mut out)?;
445 Ok(out)
446}
447
448/// The words of one response file, split the way libiberty's `buildargv` splits them.
449fn response_words(text: &str) -> Vec<String> {
450 let mut words = Vec::new();
451 let mut word = String::new();
452 // Whether a word has begun, which is not the same as `word` having something in it: `''` is
453 // an empty word of its own and has to reach the command line as one.
454 let mut begun = false;
455 let mut quote: Option<char> = None;
456 let mut chars = text.chars();
457 while let Some(c) = chars.next() {
458 match c {
459 '\\' => {
460 if let Some(next) = chars.next() {
461 word.push(next);
462 }
463 begun = true;
464 }
465 _ if quote == Some(c) => quote = None,
466 _ if quote.is_some() => word.push(c),
467 '\'' | '"' => {
468 quote = Some(c);
469 begun = true;
470 }
471 _ if c.is_whitespace() => {
472 if begun {
473 words.push(std::mem::take(&mut word));
474 begun = false;
475 }
476 }
477 _ => {
478 word.push(c);
479 begun = true;
480 }
481 }
482 }
483 if begun {
484 words.push(word);
485 }
486 words
487}
488
489/// The command line with every `-Wp,` this compiler understands spelled as its own flags.
490///
491/// The preprocessor is inside this compiler, so what a build hands it through `-Wp,` has to be
492/// read here. Kbuild is the reason: every object in the Linux kernel and in busybox is compiled
493/// with `-Wp,-MD,dir/.name.o.d`, which is cpp's spelling of `-MD -MF dir/.name.o.d`. cpp's `-MD`
494/// and `-MMD` take the file as their next word where the driver's do not, and the rest are the
495/// same flags in both. A `-Wp,` holding anything else is left as it was so the loop refuses it,
496/// because dropping part of what a build asked the preprocessor for would be the silent kind of
497/// wrong.
498fn preprocessor_args(args: &[String]) -> Vec<String> {
499 let mut out = Vec::with_capacity(args.len());
500 for arg in args {
501 let Some(list) = arg.strip_prefix("-Wp,") else {
502 out.push(arg.clone());
503 continue;
504 };
505 let words: Vec<&str> = list.split(',').collect();
506 let mut spelled = Vec::new();
507 let mut i = 0;
508 let understood = loop {
509 let Some(&word) = words.get(i) else {
510 break true;
511 };
512 i += 1;
513 match word {
514 "-MD" | "-MMD" | "-MF" | "-MT" | "-MQ" => {
515 let Some(&value) = words.get(i) else {
516 break false;
517 };
518 i += 1;
519 if word == "-MD" || word == "-MMD" {
520 spelled.extend([word.to_owned(), "-MF".to_owned()]);
521 } else {
522 spelled.push(word.to_owned());
523 }
524 spelled.push(value.to_owned());
525 }
526 "-MP" => spelled.push(word.to_owned()),
527 _ if word.len() > 2
528 && (word.starts_with("-D")
529 || word.starts_with("-U")
530 || word.starts_with("-I")) =>
531 {
532 spelled.push(word.to_owned());
533 }
534 _ => break false,
535 }
536 };
537 if understood {
538 out.extend(spelled);
539 } else {
540 out.push(arg.clone());
541 }
542 }
543 out
544}
545
546/// The extension a `-m` flag names and whether it turns it on, when it names one.
547///
548/// `-mno-` is the off form of every one of them, which is also how gcc spells it. A flag that is
549/// not an extension, `-mno-red-zone` say, is `None` and is left to the rest of the parser.
550fn isa_name(arg: &str) -> Option<(&str, rucc_target::Feature, bool)> {
551 let rest = arg.strip_prefix("-m")?;
552 let (name, on) = match rest.strip_prefix("no-") {
553 Some(name) => (name, false),
554 None => (rest, true),
555 };
556 let known =
557 if on { rucc_target::Feature::named(name) } else { rucc_target::Feature::named_off(name) };
558 known.map(|feature| (name, feature, on))
559}
560
561/// The extensions the machine running the compiler has, which is what `-march=native` means.
562///
563/// Asked of the processor with `cpuid`, through the standard library, and only when the compiler
564/// is running on an x86-64 at all. Anywhere else there is no processor to ask about an x86-64 one,
565/// and gcc on such a machine builds for the baseline, which is what this does. The list is the
566/// extensions whose names are stable in the standard library at this workspace's minimum Rust
567/// version, which covers everything [`rucc_target::Feature::honoured`] says yes to and a good deal
568/// that it does not.
569fn native_isa() -> rucc_target::Isa {
570 let base = rucc_target::Isa::baseline();
571 #[cfg(target_arch = "x86_64")]
572 {
573 let mut isa = rucc_target::Choices::new();
574 macro_rules! asked {
575 ($($detected:tt => $name:literal),* $(,)?) => {
576 $(if std::arch::is_x86_feature_detected!($detected) {
577 isa.read($name).expect("a name gcc knows");
578 })*
579 };
580 }
581 asked! {
582 "sse3" => "sse3",
583 "ssse3" => "ssse3",
584 "sse4.1" => "sse4.1",
585 "sse4.2" => "sse4.2",
586 "sse4a" => "sse4a",
587 "popcnt" => "popcnt",
588 "avx" => "avx",
589 "avx2" => "avx2",
590 "fma" => "fma",
591 "f16c" => "f16c",
592 "bmi1" => "bmi",
593 "bmi2" => "bmi2",
594 "lzcnt" => "lzcnt",
595 "xsave" => "xsave",
596 "aes" => "aes",
597 "pclmulqdq" => "pclmul",
598 "sha" => "sha",
599 "cmpxchg16b" => "cx16",
600 "adx" => "adx",
601 "rdrand" => "rdrnd",
602 "rdseed" => "rdseed",
603 }
604 isa.over(base)
605 }
606 #[cfg(not(target_arch = "x86_64"))]
607 base
608}
609
610/// Parses a command line, without the program name.
611///
612/// # Errors
613///
614/// Returns the message to print when the arguments do not name a compilation this compiler
615/// can attempt.
616pub fn parse_args(args: &[String]) -> Result<Action, CliError> {
617 let expanded = preprocessor_args(&response_files(args)?);
618 let args = expanded.as_slice();
619 let host = Triple::host()
620 .ok_or_else(|| err("this host is not a supported target and no --target was given"))?;
621 let mut opts = Options::new(host);
622 // Where the compiler is running, which is what `DW_AT_comp_dir` is and what a debugger joins a
623 // relative file name onto. Asked here rather than where the debug sections are written, because
624 // this is the one layer that is allowed to look at the process it is in, and because a command
625 // line that compiles four files should give the same answer for all four.
626 opts.working_dir = std::env::current_dir().ok().map(|dir| dir.to_string_lossy().into_owned());
627 let mut inputs: Vec<Input> = Vec::new();
628 let mut print_config = false;
629 let mut print_pipeline = false;
630 let mut print_plan = false;
631 let mut verbose = false;
632 let mut jobs = Jobs::default();
633 let mut nostdinc = false;
634 let mut sysroot: Option<PathBuf> = None;
635 // What the command line is worth warning about, filled in after the loop rather than during it,
636 // because every question of this kind is about two flags and the last word on both of them is
637 // the end of the loop.
638 let mut notes: Vec<String> = Vec::new();
639 // The whole ten field target, kept beside the three field one because `--target=` can pin a
640 // libc version and `Triple` has nowhere to put it. It decides `__GLIBC_MINOR__` and nothing
641 // else today, and `None` is a command line that named no target, which is this machine.
642 let mut pinned: Option<TargetTuple> = None;
643 let mut min_version: Option<rucc_tuple::Version> = None;
644 let mut output = None;
645 let mut link = LinkOptions::default();
646 let mut query: Option<Query> = None;
647 // What `--fetch` named, and whether `--offline` forbade it. Both are weighed after the loop
648 // because either can be written after the other.
649 let mut fetch: Option<String> = None;
650 // The other fetch, kept apart from the one above because they are different commands with
651 // different rules, and weighed after the loop for the same reason that one is.
652 let mut fetch_msvc: Option<String> = None;
653 let mut accepted = false;
654 let mut offline = false;
655 let mut threads = false;
656 // Which sanitizers are still asked for by the end of the command line. Accumulated across the
657 // loop rather than answered where it was read, because `-fno-sanitize=` turns one off and a
658 // build that asks for a check and then takes it back has asked for nothing. What happens to a
659 // set that is not empty is decided after the loop.
660 let mut sanitizers: Vec<&str> = Vec::new();
661 // The `-ffast-math` family in the order it was written, replayed after the loop on top of
662 // what `-Ofast` implies. gcc applies a level's defaults before any flag and the flags in order
663 // after that, so `-fno-fast-math -Ofast` is not fast math, and only a replay can say so.
664 let mut math_flags: Vec<&str> = Vec::new();
665 let mut ofast = false;
666 // `-mdaz-ftz` and `-mno-daz-ftz`, which decide the startup file directly and outrank the
667 // family on that one question.
668 let mut daz_ftz: Option<bool> = None;
669 // The instruction set extensions the `-m` flags named, in order, and the processor `-march`
670 // named last. Both are weighed after the loop, because a processor supplies only what no flag
671 // spoke for whichever order they came in, and because `--target=` may come after either and
672 // decide that neither means anything. See `rucc_target::isa`.
673 let mut isa = rucc_target::Choices::new();
674 let mut isa_flag: Option<&str> = None;
675 let mut march: Option<&str> = None;
676 // What `-fexceptions` and `-fno-exceptions` last said, if either was written. It is kept apart
677 // from the field because `-fnon-call-exceptions` turns exceptions on only when neither was,
678 // which is gcc's rule and is why `-fno-exceptions -fnon-call-exceptions` defines no
679 // `__EXCEPTIONS` whichever order the two come in.
680 let mut exceptions: Option<bool> = None;
681 // `-x` applies to inputs that come after it and stays in effect until the next one, which
682 // is why it is tracked across the loop rather than attached to a single argument.
683 let mut forced: Option<InputKind> = None;
684 // What `-iprefix` last said, stuck on the front of every later `-iwithprefix`. It applies to
685 // the flags after it and not the ones before, so a command line may set it more than once.
686 // GCC's default is its own installed header directory with the last component taken off,
687 // which is a path a cross compiler's build system knows and passes; there is no equivalent
688 // here, so with no `-iprefix` the prefix is nothing and `-iwithprefix` names a directory
689 // outright.
690 let mut iprefix = String::new();
691
692 let mut i = 0;
693 while i < args.len() {
694 let arg = args[i].as_str();
695 i += 1;
696 match arg {
697 "-h" | "--help" => return Ok(Action::Help),
698 "--version" => return Ok(Action::Version),
699 // The sysroot fetch, which is weighed after the loop rather than acted on here, because
700 // `--offline` written after it has to be able to forbid it. Both spellings, since a
701 // flag that takes a tuple gets written both ways and neither is a guess at what the
702 // other meant.
703 "--fetch" => {
704 let value = args
705 .get(i)
706 .ok_or_else(|| err("--fetch requires the target to get a sysroot for"))?;
707 i += 1;
708 fetch = Some(value.clone());
709 }
710 _ if arg.starts_with("--fetch=") => {
711 fetch = Some(arg["--fetch=".len()..].to_owned());
712 }
713 // The other fetch, which is section 13.4's. Same two spellings for the same reason,
714 // and weighed after the loop so that `--offline` and `--accept-licence` written after
715 // it are read whichever order somebody put them in.
716 "--fetch-msvc-sdk" => {
717 let value = args.get(i).ok_or_else(|| {
718 err("--fetch-msvc-sdk requires the target to get the SDK for")
719 })?;
720 i += 1;
721 fetch_msvc = Some(value.clone());
722 }
723 _ if arg.starts_with("--fetch-msvc-sdk=") => {
724 fetch_msvc = Some(arg["--fetch-msvc-sdk=".len()..].to_owned());
725 }
726 // Both spellings of the word, because the compiler's own prose uses one of them and
727 // most of the people typing this will reach for the other, and being told that a flag
728 // is not a flag over the letter in the middle of it is a puzzle rather than a message.
729 "--accept-licence" | "--accept-license" => accepted = true,
730 // Accepted on any command line and only ever read by the fetch, because an ordinary
731 // compile downloads nothing with or without it. So this flag takes nothing away today,
732 // which is the property section 13.2 asks for rather than an omission: a build that
733 // passes it is saying what it expects of this compiler, and what it expects is already
734 // true.
735 "--offline" => offline = true,
736 "--print-config" => print_config = true,
737 "--print-pipeline" => print_pipeline = true,
738 "-###" => print_plan = true,
739 "-v" => verbose = true,
740 // The files a compilation goes through, kept rather than thrown away. The bare
741 // spelling means `=obj` and not `=cwd`, which is not what the manual says and is what
742 // gcc 16 does; `SaveTemps::Object` carries the measurement.
743 "-save-temps" => opts.save_temps = SaveTemps::Object,
744 _ if arg.starts_with("-save-temps=") => {
745 opts.save_temps = arg["-save-temps=".len()..].parse().map_err(err)?;
746 }
747 // A `.su` beside every file compiled, one line per function saying how much stack it
748 // takes. Where the file goes is the plan's business, see `Job::stack_usage`.
749 "-fstack-usage" => opts.stack_usage = true,
750 "-fno-stack-usage" => opts.stack_usage = false,
751 // How long each step took. A misspelling of this is worth rejecting rather than
752 // ignoring, since a run that says nothing looks like a compilation that took no time.
753 "-time" => opts.time = true,
754 "-c" => opts.emit = EmitKind::Object,
755 "-S" => opts.emit = EmitKind::Asm,
756 "-E" => opts.emit = EmitKind::Preprocessed,
757 "-fsyntax-only" => opts.emit = EmitKind::SyntaxOnly,
758 "-g" => opts.debug_info = true,
759 // GCC's own levels of how much debug information to write. Zero is none and every
760 // other number is some, and this compiler has one amount, so the numbers above zero
761 // all mean the same thing here. `-ggdb` is the same flag asking for whatever the
762 // debugger on the machine prefers, which is what we emit anyway.
763 "-g0" => opts.debug_info = false,
764 "-g1" | "-g2" | "-g3" | "-ggdb" | "-ggdb1" | "-ggdb2" | "-ggdb3" => {
765 opts.debug_info = true;
766 }
767 // The version of DWARF to write. We write DWARF 5 and nothing else, so a build that
768 // asks for another version is told rather than handed a file it cannot read.
769 "-gdwarf" | "-gdwarf-5" => opts.debug_info = true,
770 _ if arg.starts_with("-gdwarf-") => {
771 return Err(err(format!(
772 "{arg}: this compiler writes DWARF 5 and no other version, see \
773 spec/11-debug-info.md"
774 )));
775 }
776 // Whether the debug information goes in a file of its own beside the object. gcc
777 // writes that `.dwo` whether or not it found anything to put in it, which means a
778 // build system that declares the file as an output gets one and a make rule that
779 // depends on it fires. Refused for that reason rather than taken: section 4.1 takes a
780 // flag that changes nothing and refuses one that changes what is produced, and a file
781 // that does not appear is the plainest change of that kind there is. The negative
782 // spelling is taken, because putting it all in the object is what happens anyway.
783 "-gno-split-dwarf" => {}
784 "-gsplit-dwarf" => {
785 return Err(err(format!(
786 "{arg}: this compiler writes no separate `.dwo` file, and a build that \
787 expects one beside each object would wait for a file that never arrives, \
788 see spec/11-debug-info.md"
789 )));
790 }
791 // How the debug sections are compressed. There are none yet, so every answer produces
792 // the same bytes and taking the flag promises nothing that is not kept. The value is
793 // still checked, because a typo in a distribution's flags is worth finding when the
794 // compiler reads it rather than when somebody later wonders why nothing got smaller.
795 // Bare `-gz` means `zlib`, which the manual leaves for the reader to discover.
796 "-gz" => opts.compress = Compress::Zlib,
797 _ if arg.starts_with("-gz=") => {
798 let how = &arg["-gz=".len()..];
799 opts.compress = how.parse().map_err(|()| {
800 err(format!(
801 "`{how}` is not a way to compress debug sections, which is none, zlib, \
802 zlib-gnu or zstd"
803 ))
804 })?;
805 }
806 "-Werror" => opts.warnings_are_errors = true,
807 // Nothing that is not fatal is said at all. Read at the one place a diagnostic goes
808 // through rather than here, so that a warning `-w` dropped is not counted either.
809 "-w" => opts.warnings = false,
810 // Off by default, the way gcc has it off. A header that came with the machine is not
811 // one the person compiling can change, so a warning about it is noise, and under
812 // `-Werror` it is a build that stops on a line nobody in the project wrote. Somebody
813 // porting a header does want to hear all of it, which is what the flag is for.
814 "-Wsystem-headers" => opts.system_header_warnings = true,
815 "-Wno-system-headers" => opts.system_header_warnings = false,
816 "-pedantic-errors" => {
817 opts.pedantic = true;
818 opts.warnings_are_errors = true;
819 }
820 "-P" => opts.line_markers = false,
821 // The dependency family, which section 4.4 calls required because every build system
822 // that generates its own makefiles asks for it. The two that end in `D` write a file
823 // beside the object and let the compilation happen, and the two that do not write to
824 // standard output and stop after it. Nothing here turns the system headers back on
825 // once a flag has turned them off, which is GCC's behaviour and is why `-MM -M` is
826 // `-MM`: the flag asking for fewer of them is the one with something to say.
827 "-M" => {
828 opts.deps.emit = true;
829 opts.deps.instead_of_compiling = true;
830 }
831 "-MM" => {
832 opts.deps.emit = true;
833 opts.deps.instead_of_compiling = true;
834 opts.deps.system_headers = false;
835 }
836 "-MD" => opts.deps.emit = true,
837 "-MMD" => {
838 opts.deps.emit = true;
839 opts.deps.system_headers = false;
840 }
841 "-MP" => opts.deps.phony = true,
842 // These three take a word and only in the separated form, which is how GCC spells
843 // them and how every build system writes them.
844 "-MF" | "-MT" | "-MQ" => {
845 let value =
846 args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
847 i += 1;
848 match arg {
849 "-MF" => opts.deps.file = Some(value.clone()),
850 // The whole of the difference between the two. `-MT` is for a build that has
851 // already escaped what it is passing, and `-MQ` is for one that has a name
852 // and wants it to arrive as that name.
853 "-MT" => opts.deps.targets.push(value.clone()),
854 _ => opts.deps.targets.push(deps::escaped(value)),
855 }
856 }
857 // The questions a build system asks before it compiles anything. Answered after the
858 // loop, because each one is about the target or the library search and the command
859 // line has not finished saying what those are.
860 "-dumpmachine" => query = Some(Query::Machine),
861 // Both answer with the GCC release in `__GNUC__` rather than our own version, because
862 // what asks is a build script deciding which GCC it is talking to, and `0.11` reads as
863 // a GCC too old to have anything. GCC 7 and later print only the major number for the
864 // first one, and that is the shape the scripts were written against.
865 "-dumpversion" => query = Some(Query::Version),
866 "-dumpfullversion" => query = Some(Query::FullVersion),
867 "-print-multiarch" => query = Some(Query::Multiarch),
868 "-print-search-dirs" => query = Some(Query::SearchDirs),
869 "-print-sysroot" => query = Some(Query::Sysroot),
870 // Both spellings, because this one is ours rather than GCC's and our own documents
871 // write it both ways: section 13.5 of `spec/cross-compile/13-distribution.md` gives it
872 // two dashes like the other flags we invented, and document 12's table gives it one
873 // like the `-print-` family it sits in. A person who reads either and types what it
874 // says is right, so neither is refused.
875 "-print-sysroot-provenance" | "--print-sysroot-provenance" => {
876 query = Some(Query::SysrootProvenance);
877 }
878 "-print-sysroot-digest" | "--print-sysroot-digest" => {
879 query = Some(Query::SysrootDigest);
880 }
881 "-print-libgcc-file-name" => query = Some(Query::Libgcc),
882 _ if arg.starts_with("-print-file-name=") => {
883 query = Some(Query::FileName(arg["-print-file-name=".len()..].to_owned()));
884 }
885 _ if arg.starts_with("-print-prog-name=") => {
886 query = Some(Query::ProgName(arg["-print-prog-name=".len()..].to_owned()));
887 }
888 // A program built to run in more than one thread. On every platform this compiler
889 // targets that is a macro the library's headers read and one more library on the
890 // link line, and the library is added after the loop so that it lands after the
891 // objects that refer to it.
892 "-pthread" | "-pthreads" => {
893 opts.defines.push("_REENTRANT".to_owned());
894 threads = true;
895 }
896 "-ansi" => {
897 opts.std = Std::C89;
898 opts.gnu_extensions = false;
899 }
900 // `-Wpedantic` is the same flag under the name the `-W` family gives it, which is
901 // the spelling a build system that groups its warning flags tends to write.
902 "-pedantic" | "-Wpedantic" => opts.pedantic = true,
903 // Both directions, because a build that needs this for one directory turns it back
904 // off for the next one rather than leaving it on for the whole tree.
905 "-fpermissive" => opts.permissive = true,
906 "-fno-permissive" => opts.permissive = false,
907 "-ffreestanding" => opts.hosted = false,
908 "-fhosted" => opts.hosted = true,
909 "-fno-builtin" => opts.builtins = false,
910 "-fbuiltin" => opts.builtins = true,
911 // The C89 dialects are under GNU's reading whatever this says, so turning it off
912 // there is turning off something the dialect asked for, which is accepted and does
913 // nothing. gcc refuses that command line, and there is nothing it could have meant.
914 "-fgnu89-inline" => opts.gnu89_inline = true,
915 "-fno-gnu89-inline" => opts.gnu89_inline = false,
916 // Both directions of each, because a build system that wants one of these usually
917 // writes it beside the flag that turns it back off for one directory.
918 "-fno-omit-frame-pointer" => opts.frame_pointer = Some(true),
919 "-fomit-frame-pointer" => opts.frame_pointer = Some(false),
920 // Both directions again, for the same reason, and a third answer for a command line
921 // that wrote neither: see `reorder_blocks` in `rucc_session`.
922 "-freorder-blocks" => opts.reorder_blocks = Some(true),
923 "-fno-reorder-blocks" => opts.reorder_blocks = Some(false),
924 // gcc's name for the scheduler that runs after the registers are handed out, which is
925 // the only one rucc has: see `schedule_insns` in `rucc_session`. gcc also takes
926 // `-fschedule-insns` for the pass before allocation, and taking that one here would be
927 // a flag that says a pass ran when none did.
928 "-fschedule-insns2" => opts.schedule_insns = Some(true),
929 "-fno-schedule-insns2" => opts.schedule_insns = Some(false),
930 // A call in tail position as a jump: see `sibling_calls` in `rucc_session`.
931 "-foptimize-sibling-calls" => opts.sibling_calls = Some(true),
932 "-fno-optimize-sibling-calls" => opts.sibling_calls = Some(false),
933 "-mno-red-zone" => opts.red_zone = false,
934 "-mred-zone" => opts.red_zone = true,
935 // Four flags rather than one with an argument, which is how gcc spells them and how
936 // every build line writes them. Last one wins, because a package build puts
937 // `-fstack-protector-strong` in its global flags and a directory that cannot have one
938 // turns it back off on the line after.
939 "-fno-stack-protector" | "-fno-stack-protector-all" | "-fno-stack-protector-strong" => {
940 opts.protector = Protector::None;
941 }
942 "-fstack-protector" => opts.protector = Protector::Buffers,
943 "-fstack-protector-strong" => opts.protector = Protector::Strong,
944 "-fstack-protector-all" => opts.protector = Protector::All,
945 // The other half of what a hardened build asks for, and it is a question about the
946 // frame rather than about the function, so it is a switch rather than a level.
947 "-fstack-clash-protection" => opts.stack_clash = true,
948 "-fno-stack-clash-protection" => opts.stack_clash = false,
949 // The third of them, and the one that is a question with an argument rather than a
950 // family of spellings, because what it asks about is which of the two edges of a
951 // control flow transfer is checked. Bare is both of them, which is what gcc does.
952 "-fcf-protection" => opts.control = Control::Full,
953 "-fno-cf-protection" => opts.control = Control::None,
954 // Two spellings of the same request, which is what gcc has as well. `-p` was the older
955 // profiler and `-pg` the one that also recorded who called whom, and on every platform
956 // this compiler targets there is now one hook and both ask for it.
957 "-pg" | "-p" => {
958 opts.profile = true;
959 link.profile = true;
960 }
961 // Accepted on their own and doing nothing on their own, which is gcc's behaviour: they
962 // say where the call goes and a command line that asked for no call has nowhere to put
963 // one. That matters because a build system that sets `-mfentry` globally and `-pg` per
964 // directory is a build system that would otherwise fail on every other directory.
965 "-mfentry" => opts.hook = Hook::Early,
966 "-mno-fentry" => opts.hook = Hook::Late,
967 // GCC drops its own include directory along with the system ones, because its
968 // headers are half of a pair with the library's and half a pair is worse than
969 // none. A build that passes this is supplying the whole set itself.
970 "-nostdinc" => nostdinc = true,
971 "-o" => {
972 output = Some(args.get(i).ok_or_else(|| err("-o requires an argument"))?.clone());
973 i += 1;
974 }
975 // What the files kept beside an output are named after, which is `-save-temps` and
976 // `-fstack-usage` so far. gcc takes each of the three in the separated form only, and
977 // its driver passes them to every compilation it runs, so a build that copied a
978 // command line out of gcc's `-v` has them. See `phase::aux_base` for what they do.
979 "-dumpbase" | "-dumpbase-ext" | "-dumpdir" => {
980 let value =
981 args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?.clone();
982 i += 1;
983 match arg {
984 "-dumpbase" => opts.dump_base = Some(value),
985 "-dumpbase-ext" => opts.dump_base_ext = Some(value),
986 _ => opts.dump_dir = Some(value),
987 }
988 }
989 // The flags that take a directory only in the separated form. GCC spells them
990 // this way and nothing writes `-iquotedir`, so accepting the joined form would
991 // mean guessing at a path that starts with the flag's own letters.
992 // Apple's spelling of `--sysroot`, and the one its own build systems pass. The
993 // two mean the same thing here: the configured directories are under there rather
994 // than under the root.
995 "-isysroot" => {
996 let dir = args.get(i).ok_or_else(|| err("-isysroot requires an argument"))?;
997 i += 1;
998 sysroot = Some(PathBuf::from(dir));
999 }
1000 "-iquote" | "-isystem" | "-idirafter" => {
1001 let dir = args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
1002 i += 1;
1003 match arg {
1004 "-iquote" => opts.search.push_quote(dir.clone()),
1005 "-isystem" => opts.search.push_system(dir.clone()),
1006 _ => opts.search.push_after(dir.clone()),
1007 }
1008 }
1009 "-iprefix" => {
1010 iprefix = args.get(i).ok_or_else(|| err("-iprefix requires an argument"))?.clone();
1011 i += 1;
1012 }
1013 // Where GCC puts these is not where its manual says it puts them, and this is the
1014 // measured answer rather than the documented one: `-iwithprefix` lands in the
1015 // `-isystem` slot and not the `-idirafter` slot, and `-iwithprefixbefore` lands in
1016 // the `-I` slot. A cross build that uses them is relying on the behaviour, since
1017 // that is the compiler it was developed against.
1018 "-iwithprefix" | "-iwithprefixbefore" => {
1019 let dir = args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
1020 i += 1;
1021 let dir = format!("{iprefix}{dir}");
1022 if arg == "-iwithprefix" {
1023 opts.search.push_system(dir);
1024 } else {
1025 opts.search.push_bracket(dir);
1026 }
1027 }
1028 "-include" | "-imacros" => {
1029 let name = args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
1030 i += 1;
1031 opts.preincludes
1032 .push(Preinclude { name: name.clone(), macros_only: arg == "-imacros" });
1033 }
1034 // The flag `-iquote` was introduced to replace, still passed by build systems old
1035 // enough to predate the replacement. It is not a directory: it says that every `-I`
1036 // so far is for quoted includes only, and that a quoted include stops looking next
1037 // to the file that wrote it.
1038 "-I-" => opts.search.split_quote_chain(),
1039 // `-x c` and `-xc`, both of which gcc takes. busybox and toybox probe the compiler
1040 // with the joined one.
1041 _ if arg.starts_with("-x") => {
1042 let lang = joined_or_next(arg, 2, args, &mut i)?;
1043 forced = if lang == "none" {
1044 None
1045 } else {
1046 Some(InputKind::from_x_arg(&lang).map_err(|e| err(format!("{e}")))?)
1047 };
1048 }
1049 // Not a GCC flag. spec/03-architecture.md section 3.5 compiles several
1050 // translation units in one process rather than making the build system fork, and
1051 // section 3.8's determinism check compares `-j1` against `-j16`, so the knob has
1052 // to exist and has to be spelled the way `make` spells it.
1053 // `-DFOO`, `-D FOO` and the same for `-U` and `-I`. Both forms are in wide use
1054 // and a build system may produce either, so both are read here rather than
1055 // being normalised by whatever generated the command line.
1056 _ if arg.starts_with("-D") => {
1057 let value = joined_or_next(arg, 2, args, &mut i)?;
1058 opts.defines.push(value);
1059 }
1060 _ if arg.starts_with("-U") => {
1061 let value = joined_or_next(arg, 2, args, &mut i)?;
1062 opts.undefines.push(value);
1063 }
1064 _ if arg.starts_with("-I") => {
1065 let dir = joined_or_next(arg, 2, args, &mut i)?;
1066 opts.search.push_bracket(dir);
1067 }
1068 _ if arg.starts_with("-std=") => {
1069 let name = &arg["-std=".len()..];
1070 let (std, gnu) = Std::from_flag(name)
1071 .ok_or_else(|| err(format!("unknown dialect `{name}`, see --help")))?;
1072 opts.std = std;
1073 opts.gnu_extensions = gnu;
1074 }
1075 // Section 4.5. The claim decides which half of glibc's `sys/cdefs.h` we are
1076 // handed, so a differential run that does not set it is comparing two compilers
1077 // that believe they are different compilers.
1078 // GCC packs these into one flag, so `-dDI` is two of them. Letters in the family
1079 // that we have not written yet are accepted and ignored, because a dump is a
1080 // debugging aid and a build that asks for one should still compile. A letter
1081 // outside the family falls through to the unknown option error, which is what
1082 // keeps `-dumpversion` from being read as a dump of nothing.
1083 _ if Dumps::is_family(arg) => {
1084 opts.dumps.add(&arg[2..]);
1085 }
1086 // One name at a time, which is what a build that means its own `memcpy` and the
1087 // library's everything else writes. The name is not checked against a list, because
1088 // the flag is about what the program means by a name and a program is allowed to mean
1089 // something by a name this compiler has never heard of.
1090 _ if arg.starts_with("-fno-builtin-") => {
1091 opts.no_builtin.push(arg["-fno-builtin-".len()..].to_owned());
1092 }
1093 _ if arg.starts_with("-fgnuc-version=") => {
1094 let v = &arg["-fgnuc-version=".len()..];
1095 opts.gnuc = v.parse().map_err(err)?;
1096 opts.gnuc_given = true;
1097 }
1098 // The MSVC release an MSVC row claims, which is `_MSC_VER` and nothing else here.
1099 _ if arg.starts_with("-fms-compatibility-version=") => {
1100 let v = &arg["-fms-compatibility-version=".len()..];
1101 opts.msc = v.parse().map_err(err)?;
1102 }
1103 // Which of Microsoft's C runtimes an MSVC row links against, in clang's spelling of
1104 // `cl.exe`'s `/MT` and `/MD`. The headers are told through `_DLL` and the link through
1105 // the libraries it names, so it is both a compile flag and a link one. The two debug
1106 // runtimes are refused by name rather than taken as the release ones, since what they
1107 // link is a different set of libraries and a program that asked for one wants its
1108 // checks.
1109 _ if arg.starts_with("-fms-runtime-lib=") => {
1110 let dll = match &arg["-fms-runtime-lib=".len()..] {
1111 "static" => false,
1112 "dll" => true,
1113 debug @ ("static_dbg" | "dll_dbg") => {
1114 return Err(err(format!(
1115 "-fms-runtime-lib={debug} asks for Microsoft's debug C runtime, which \
1116 this compiler does not link against yet. static and dll are the two \
1117 it has"
1118 )));
1119 }
1120 other => {
1121 return Err(err(format!(
1122 "-fms-runtime-lib= takes static or dll, and `{other}` is neither"
1123 )));
1124 }
1125 };
1126 opts.ms_dll_runtime = dll;
1127 link.crt = if dll { rucc_sysroot::Crt::Dll } else { rucc_sysroot::Crt::Static };
1128 }
1129 // spec/13-gnu-compat.md section 13.3 promises this flag an error that says why rather
1130 // than the unknown option one, because a build reaching for it is asking for a feature
1131 // and deserves to be told it is not coming rather than told the spelling is wrong.
1132 // The negative form is what this compiler does anyway, so it is taken and dropped.
1133 "-fnested-functions" => {
1134 return Err(err(
1135 "nested functions are not supported: a call to one goes through a trampoline \
1136 written on the stack, which no target that enforces an unexecutable stack \
1137 allows",
1138 ));
1139 }
1140 "-fno-nested-functions" => {}
1141 // Which of the two links the output is for, which is a real difference and not a
1142 // description of what happens anyway. Everything here is position independent either
1143 // way, and what these decide is whether a name may be one another object defines or
1144 // replaces, because a link that produces an executable puts every name in the same
1145 // program and a link that produces a shared library does not.
1146 //
1147 // It matters that they are accepted at all, whatever they then do. Every autoconf and
1148 // cmake build puts `-fPIC` on the compile line, so a compiler that rejects it cannot
1149 // be the `CC` of a project that has a configure script, whatever else it can do. That
1150 // is how this was found: building SQLite's test fixture stopped on it.
1151 "-fPIC" | "-fpic" => opts.pic = Pic::Library,
1152 // Not a synonym of the pair above, which is what they were treated as until #756. The
1153 // library is the expensive answer and gcc makes it the one that has to be asked for,
1154 // so this is also what nothing at all means.
1155 "-fPIE" | "-fpie" => opts.pic = Pic::Executable,
1156 // A different question from the pair above, and the one every distribution build of a
1157 // shared library answers. `-fPIC` decides how an address is reached, and this decides
1158 // whether the optimizer may believe a body it can see, because an exported name is one
1159 // the dynamic linker may find another definition of first. On by default, which is
1160 // gcc's arrangement and is the honest answer, and off is a promise the build makes and
1161 // nothing checks.
1162 "-fsemantic-interposition" => opts.interposition = true,
1163 "-fno-semantic-interposition" => opts.interposition = false,
1164 // Two requests rather than one, and the same table answers both, so what decides is
1165 // whether either of them is standing. gcc arranges it the same way: the asynchronous
1166 // one is the default here and it implies the other, and a line that asks for a table
1167 // and against an asynchronous one gets a table.
1168 "-fasynchronous-unwind-tables" => opts.async_unwind_tables = true,
1169 "-fno-asynchronous-unwind-tables" => opts.async_unwind_tables = false,
1170 "-funwind-tables" => opts.unwind_tables = true,
1171 "-fno-unwind-tables" => opts.unwind_tables = false,
1172 // The other direction is a request, not a description, and it is one this compiler
1173 // cannot grant, so it gets the treatment section 13.3 asks for rather than the unknown
1174 // option error. Answering it by carrying on would be answering a different question:
1175 // the code would still be position independent, which is correct everywhere an
1176 // ordinary program runs and is wrong in a kernel, where the flag is written precisely
1177 // because there is no loader to fill a global offset table in.
1178 "-fno-pic" | "-fno-pie" => {
1179 return Err(err(
1180 "position dependent code is not supported: an address that may be in another \
1181 object is loaded out of the global offset table, and nothing here emits the \
1182 absolute form this asks for. Use -no-pie if what you meant was how to link",
1183 ));
1184 }
1185 // A section per function and a section per variable, which is what makes
1186 // `--gc-sections` able to drop anything: a linker can leave out a section nothing
1187 // reaches and cannot leave out half of one. Both directions are taken, and the off
1188 // one is the default rather than a refusal, since a build that writes it is asking
1189 // for what happens anyway.
1190 "-ffunction-sections" => opts.function_sections = true,
1191 "-fno-function-sections" => opts.function_sections = false,
1192 "-fdata-sections" => opts.data_sections = true,
1193 "-fno-data-sections" => opts.data_sections = false,
1194 // Whether a file scope declaration with no initializer is offered to the linker as a
1195 // common symbol for it to merge, or written into `.bss` as an ordinary defined one.
1196 // Unwritten, the target answers, which is on for Darwin and off everywhere else.
1197 "-fcommon" => opts.common = Some(true),
1198 "-fno-common" => opts.common = Some(false),
1199 // What overflows rather than being undefined. Every one of these takes something away
1200 // from the optimizer rather than asking it to do anything, which is why the negative
1201 // spellings are the interesting ones and the positive spellings are the default.
1202 //
1203 // `-fno-strict-overflow` is both of the others, which is gcc's own reading of it: its
1204 // help text for `-fstrict-overflow` says "negated as -fwrapv -fwrapv-pointer". So it is
1205 // written here as the pair rather than kept as a third thing to test everywhere.
1206 //
1207 // `-ftrapv` is the exception and is the one that asks for something. It is the other
1208 // answer to the question `-fwrapv` answers, so the two cannot both hold and each clears
1209 // the other, which makes the last one on the command line the one that counts. That is
1210 // gcc 16's behaviour and was measured rather than read: `-ftrapv -fwrapv` emits no
1211 // checked calls and `-fwrapv -ftrapv` emits them. The positive spelling of the pointer
1212 // question is left alone by both, because neither has anything to say about it.
1213 "-fwrapv" => {
1214 opts.wrapping.signed = true;
1215 opts.wrapping.trap = false;
1216 }
1217 "-fno-wrapv" => opts.wrapping.signed = false,
1218 "-fwrapv-pointer" => opts.wrapping.pointer = true,
1219 "-fno-wrapv-pointer" => opts.wrapping.pointer = false,
1220 "-fno-strict-overflow" => opts.wrapping = Wrapping::ALL,
1221 // Which does not clear the checked one, because gcc does not: `-ftrapv
1222 // -fstrict-overflow` still emits the calls. It says what is assumed and not what
1223 // happens.
1224 "-fstrict-overflow" => {
1225 opts.wrapping.signed = false;
1226 opts.wrapping.pointer = false;
1227 }
1228 "-ftrapv" => {
1229 opts.wrapping.trap = true;
1230 opts.wrapping.signed = false;
1231 }
1232 "-fno-trapv" => opts.wrapping.trap = false,
1233 // The two flags that say what a plain `char` is, which is one question with two
1234 // spellings each: gcc reads `-fno-signed-char` as `-funsigned-char` and
1235 // `-fno-unsigned-char` as `-fsigned-char`, so there are four ways to write two
1236 // answers and the last one written wins. Nothing is set until one of them is given,
1237 // because the target's own ABI is the answer otherwise and it is not the same answer
1238 // everywhere: x86-64 and Apple's arm64 are signed, Linux's arm64 is not.
1239 "-fsigned-char" | "-fno-unsigned-char" => opts.char_signed = Some(true),
1240 "-funsigned-char" | "-fno-signed-char" => opts.char_signed = Some(false),
1241 // And the size of an enumeration, which is the other thing in this group that changes
1242 // the ABI rather than the code.
1243 "-fshort-enums" => opts.short_enums = true,
1244 "-fno-short-enums" => opts.short_enums = false,
1245 // And Microsoft's reading of an anonymous member, which changes the layout of every
1246 // record that writes a tag on one. Nothing is set until one of them is given, because
1247 // the target is the answer otherwise: gcc's mingw build has this on and its Linux
1248 // build has it off.
1249 "-fms-extensions" => opts.ms_extensions = Some(true),
1250 "-fno-ms-extensions" => opts.ms_extensions = Some(false),
1251 // Both directions of this one are recorded, and what they decide is whether lowering
1252 // names the type each access goes through. Turning it off is the front end leaving the
1253 // name off rather than a pass being told to ignore one it can see, which is one
1254 // condition in one place, and it is the reading that survives link time optimization:
1255 // a unit built with the flag off keeps its own answer when its bodies end up in a
1256 // module beside bodies that were not.
1257 //
1258 // Nothing in the pipeline reads those names yet. Layer 3 of the alias analysis does
1259 // and is tested, and no pass at any level asks the alias analysis anything today, so
1260 // no program compiles differently for having passed this. The flag is wired anyway,
1261 // because the change that makes a pass ask is not the change anybody will remember to
1262 // wire it in, and a flag that is taken and dropped once the names mean something is
1263 // the miscompilation `spec/04-driver-and-cli.md` section 4.1 warns about in as many
1264 // words.
1265 "-fstrict-aliasing" => opts.strict_aliasing = true,
1266 "-fno-strict-aliasing" => opts.strict_aliasing = false,
1267 // The same shape of answer for the same reason, and the flag the kernel writes beside
1268 // the one above it.
1269 //
1270 // Nothing here concludes that a pointer is not null from the fact that it was
1271 // dereferenced. There is no such conclusion to draw from, because no pass records one:
1272 // a load says where it read and nothing else, and a comparison against null is an
1273 // ordinary comparison of two values the optimizer has no fact about. So a function
1274 // that reads through a pointer and then tests it keeps the test, which is what the
1275 // kernel wants and what `-fno-delete-null-pointer-checks` asks for, and what gcc has
1276 // to be asked for because it draws the conclusion by default.
1277 //
1278 // `-fdelete-null-pointer-checks` is the request to draw it, and it goes the way
1279 // `-fstrict-aliasing` does: assuming less than was asked for costs speed and not
1280 // correctness, and `-O2` implies it, so refusing it would stop builds for nothing.
1281 "-fdelete-null-pointer-checks" | "-fno-delete-null-pointer-checks" => {}
1282 // The floating point group, which goes the same way and for the same reason, and which
1283 // is worth writing out because the reason is easy to get backwards.
1284 //
1285 // Each of these has a restrictive spelling and a permissive one. The restrictive ones,
1286 // `-frounding-math` and `-ftrapping-math`, say that the rounding mode may have been
1287 // changed and that an exception raised by an operation may be looked at, so an
1288 // arithmetic the compiler folds at compile time is an arithmetic whose rounding and
1289 // whose exception the program does not get. Nothing here folds any floating point
1290 // arithmetic in a function body: `0.1 + 0.2` is an `fadd` and `1.0 / 0.0` is a divide
1291 // that runs, at every level. So both of those describe what already happens.
1292 //
1293 // The permissive ones are the other half, and they are licences rather than requests
1294 // for an answer. `-fno-rounding-math` says the rounding mode is the default one and
1295 // `-fno-trapping-math` says nothing looks at the exceptions, which together are
1296 // permission to fold. Not folding is the conservative side of that permission and is
1297 // what a program is entitled to whichever was written, so `-fno-rounding-math` costs
1298 // speed and not correctness, which is the test section 4.1 puts a licence through.
1299 "-frounding-math" | "-fno-rounding-math" => {}
1300 // `-fno-trapping-math` is the one of the four that is kept, because there is one
1301 // conversion this compiler does not fold and gcc folds under it, and the two answers
1302 // differ. Converting a constant floating value to an integer type it does not fit in
1303 // is undefined behaviour rather than a value: left to the hardware it is one
1304 // instruction and the answer is the integer indefinite value, and folded it is the
1305 // nearest end of the integer's range. Both compilers leave it to the instruction by
1306 // default and gcc folds it under this flag, so a program built with it and compiled
1307 // without it gets a different number rather than a slower one. `-ftrapping-math` is
1308 // gcc's default, so a build spelling it out is asking for what it already has.
1309 //
1310 // The rest of the family goes with it, `-ffast-math` included, and all of them are
1311 // taken now. Each is a licence rather than a request and nothing here folds floating
1312 // point arithmetic, so the code does not change. What does change is the macros gcc
1313 // defines for each licence, which a header reads, and the startup file `-ffast-math`
1314 // links, which puts the hardware in flush to zero mode. Both are done after the loop,
1315 // because the family is a set of switches over the same fields and the last word on
1316 // each of them is the end of the command line.
1317 "-ftrapping-math"
1318 | "-fno-trapping-math"
1319 | "-ffast-math"
1320 | "-fno-fast-math"
1321 | "-funsafe-math-optimizations"
1322 | "-fno-unsafe-math-optimizations"
1323 | "-fmath-errno"
1324 | "-fno-math-errno"
1325 | "-ffinite-math-only"
1326 | "-fno-finite-math-only"
1327 | "-fsigned-zeros"
1328 | "-fno-signed-zeros"
1329 | "-freciprocal-math"
1330 | "-fno-reciprocal-math"
1331 | "-fassociative-math"
1332 | "-fno-associative-math" => math_flags.push(arg),
1333 // Whether the startup file that sets flush to zero is linked, asked directly. gcc
1334 // links it for a shared object too when this is written, which the family does not.
1335 "-mdaz-ftz" => daz_ftz = Some(true),
1336 "-mno-daz-ftz" => daz_ftz = Some(false),
1337 // About temporary files rather than about code. There is nothing between the phases of
1338 // one compilation here to write to a file in the first place.
1339 "-pipe" => {}
1340 // Preprocess the input, which a C compile always does. GCC has it for Fortran, and
1341 // meson writes it when it asks a compiler for its predefined macros.
1342 "-cpp" => {}
1343 // Nothing here writes colour, so all of these are the same answer, and it is the answer
1344 // that costs nothing: the diagnostics come out plain either way and no build depends on
1345 // an escape sequence being there. Taken rather than refused because cmake writes
1346 // `-fdiagnostics-color=always` on every compile line when the generator is ninja, which
1347 // makes this the second most common flag after `-fPIC` to stop a build over a question
1348 // about how the text looks.
1349 "-fdiagnostics-color" | "-fno-diagnostics-color" => {}
1350 _ if arg.starts_with("-fdiagnostics-color=") => {}
1351 // The link flags. None of them changes the compilation, which is why they are
1352 // collected apart from `opts` and why `-lm` on a `-c` line is a note rather than an
1353 // error: it is a thing said to a linker that is not going to run.
1354 "-static" => link.is_static = true,
1355 "-shared" => link.shared = true,
1356 "-r" => link.relocatable = true,
1357 "-pie" => link.pie = Some(true),
1358 "-no-pie" | "-nopie" => link.pie = Some(false),
1359 "-nostdlib" => link.no_stdlib = true,
1360 "-nostartfiles" => link.no_startfiles = true,
1361 "-nodefaultlibs" => link.no_defaultlibs = true,
1362 "-fno-builtins-lib" => link.no_builtins_lib = true,
1363 "-fbuiltins-lib" => link.no_builtins_lib = false,
1364 "-rdynamic" | "-export-dynamic" => link.export_dynamic = true,
1365 "-s" => link.strip = true,
1366 // mingw-w64's three. `-mwindows` and `-mconsole` pick the subsystem, last one wins,
1367 // and `-municode` picks the start file and tells the headers through `UNICODE`, which is
1368 // what gcc's spec does with it. All three are taken and ignored for other targets, as gcc
1369 // built for mingw is the only gcc that knows them and a Makefile written for it is what
1370 // passes them.
1371 "-mwindows" => link.gui = true,
1372 "-mconsole" => link.gui = false,
1373 "-municode" => {
1374 link.unicode = true;
1375 opts.defines.push("UNICODE".to_owned());
1376 }
1377 // Into the ordered input list rather than a list of its own, because a great many of
1378 // the linker's options are a bracket around the files after them and an option that
1379 // lost its place among them says nothing. `--whole-archive` is the one that found this.
1380 "-Xlinker" => {
1381 let next = args.get(i).ok_or_else(|| err("-Xlinker requires an argument"))?;
1382 i += 1;
1383 inputs.push(Input::linker(next));
1384 }
1385 _ if arg.starts_with("-Wl,") => {
1386 // Commas separate arguments rather than being part of one, which is what makes
1387 // `-Wl,-rpath,/opt/lib` two words to the linker and one word here.
1388 inputs.extend(arg["-Wl,".len()..].split(',').map(Input::linker));
1389 }
1390 _ if arg.starts_with("-fuse-ld=") => {
1391 link.use_ld = Some(arg["-fuse-ld=".len()..].to_owned());
1392 }
1393 _ if arg.starts_with("-l") && arg.len() > 2 => {
1394 inputs.push(Input::library(&arg[2..]));
1395 }
1396 "-l" => {
1397 let next = args.get(i).ok_or_else(|| err("-l requires an argument"))?;
1398 i += 1;
1399 inputs.push(Input::library(next));
1400 }
1401 _ if arg.starts_with("-L") => {
1402 link.search.push(PathBuf::from(joined_or_next(arg, 2, args, &mut i)?));
1403 }
1404 _ if arg.starts_with("-B") => {
1405 link.prefixes.push(PathBuf::from(joined_or_next(arg, 2, args, &mut i)?));
1406 }
1407 _ if arg.starts_with("-j") => {
1408 jobs = Jobs::parse(&arg[2..]).map_err(err)?;
1409 }
1410 _ if arg.starts_with("--sysroot=") => {
1411 sysroot = Some(PathBuf::from(&arg["--sysroot=".len()..]));
1412 }
1413 _ if arg.starts_with("--target=") => {
1414 let t = &arg["--target=".len()..];
1415 // The same string again, as the model that has room for a libc version. A spelling
1416 // the three field parser took and this one does not is not an error, because the
1417 // one that decides what is compiled has already accepted it and the only thing
1418 // lost is a version nobody asked for.
1419 pinned = t.parse().ok();
1420 // The other way round is a deployment target the three field parser has no room
1421 // for, `aarch64-macos.13`, and the triple is the one the tuple narrows to.
1422 opts.target = match t.parse() {
1423 Ok(triple) => triple,
1424 Err(e) => {
1425 pinned.and_then(Triple::from_tuple).ok_or_else(|| err(format!("{e}")))?
1426 }
1427 };
1428 }
1429 _ if arg.starts_with("--emit=") => {
1430 let k = &arg["--emit=".len()..];
1431 opts.emit = k
1432 .parse()
1433 .map_err(|()| err(format!("unknown --emit kind `{k}`, see --help")))?;
1434 }
1435 // A bare `-O` is `-O1`, which is what GCC has and what a hand written makefile tends
1436 // to write. `-Og` is GCC's level for a build somebody is going to step through, and
1437 // it is `-O1` with the transformations that move code around left out; this compiler
1438 // has no such level yet, so it is the nearest one and `--print-pipeline` says what
1439 // that came to rather than the flag pretending otherwise.
1440 "-O" | "-Og" => {
1441 opts.opt_level = rucc_session::OptLevel::O1;
1442 ofast = false;
1443 }
1444 // The union of `-O3` and `-ffast-math`. The second half is a default rather than a
1445 // flag, which is why it is remembered here and applied after the loop: a later level
1446 // takes it back, and so does a `-fno-fast-math` written on either side of it.
1447 "-Ofast" => {
1448 opts.opt_level = rucc_session::OptLevel::O3;
1449 ofast = true;
1450 }
1451 _ if arg.starts_with("-O") => {
1452 ofast = false;
1453 opts.opt_level = arg[2..]
1454 .parse()
1455 .map_err(|()| err(format!("unknown optimization level `{arg}`")))?;
1456 }
1457 // How far a multiply and an addition may be fused into one rounding. Before the
1458 // optimizer's `-f` family below for the reason the ones under it are, and kept rather
1459 // than dropped because it is the one flag in its group this compiler could act on: it
1460 // rides into the IR as an attribute on each function with a body, so the day the code
1461 // generator forms an `fma` it already knows which functions were given permission.
1462 // Nothing forms one today, under any value of this and under any `-march=`.
1463 _ if arg.starts_with("-ffp-contract=") => {
1464 let how = &arg["-ffp-contract=".len()..];
1465 opts.fp_contract = how.parse().map_err(|()| {
1466 err(format!("`{how}` is not a contraction, which is fast, on or off"))
1467 })?;
1468 }
1469 // How much of an expression may be computed wider than it was written. The values are
1470 // gcc's and so is the refusal of anything else, and none of the three changes anything
1471 // here: an operation is computed in the type C says it is on every target this compiler
1472 // has a back end for, so `__FLT_EVAL_METHOD__` is 0 and `standard` is already what
1473 // happens. `fast` and `16` are permission to be wider, which is a licence this takes
1474 // and does not use, the same way the two above are. The flag is worth taking because
1475 // glibc's headers and a good deal of configure output write it, and because the answer
1476 // it asks about is one this compiler can state rather than guess at: there is no x87
1477 // target here, which is the machine the whole question was invented for.
1478 // Whether a local and a spilled value that are never both wanted may be the same bytes
1479 // of the frame. gcc's three values, and two of them mean the same thing here: what rucc
1480 // shares is a local whose address provably never leaves the function, which is narrower
1481 // than `named_vars` and narrower still than `all`, so both of them get it. `none` is
1482 // the one that changes anything, and it is the flag a program that reads a local
1483 // through a pointer it kept past the end of the block writes.
1484 _ if arg.starts_with("-fstack-reuse=") => {
1485 let how = &arg["-fstack-reuse=".len()..];
1486 opts.stack_reuse = match how {
1487 "all" | "named_vars" => Some(true),
1488 "none" => Some(false),
1489 _ => {
1490 return Err(err(format!(
1491 "`{how}` is not a stack reuse, which is all, named_vars or none"
1492 )));
1493 }
1494 };
1495 }
1496 _ if arg.starts_with("-fexcess-precision=") => {
1497 let how = &arg["-fexcess-precision=".len()..];
1498 if !matches!(how, "16" | "fast" | "standard") {
1499 return Err(err(format!(
1500 "`{how}` is not an excess precision, which is 16, fast or standard"
1501 )));
1502 }
1503 }
1504 // Which front of a path is rewritten before it reaches the output, which is how a
1505 // build gets the same bytes out of two different directories. The four spellings are
1506 // one flag each into three lists, and `-ffile-prefix-map=` is the three of them at
1507 // once. Only the macro list does anything today, because `__FILE__` is the only place
1508 // a path reaches the output: there is no DWARF and no profile data yet, so the other
1509 // two are recorded for the work that will read them. The argument splits at the last
1510 // `=` rather than the first, which is gcc's rule and is what lets a directory with an
1511 // `=` in its name be the old half.
1512 _ if arg.starts_with("-fmacro-prefix-map=") => {
1513 let (old, new) = rewrite(arg, "-fmacro-prefix-map=")?;
1514 opts.prefix_map.macros.push(old, new);
1515 }
1516 _ if arg.starts_with("-fdebug-prefix-map=") => {
1517 let (old, new) = rewrite(arg, "-fdebug-prefix-map=")?;
1518 opts.prefix_map.debug.push(old, new);
1519 }
1520 _ if arg.starts_with("-fprofile-prefix-map=") => {
1521 let (old, new) = rewrite(arg, "-fprofile-prefix-map=")?;
1522 opts.prefix_map.profile.push(old, new);
1523 }
1524 _ if arg.starts_with("-ffile-prefix-map=") => {
1525 let (old, new) = rewrite(arg, "-ffile-prefix-map=")?;
1526 opts.prefix_map.macros.push(old, new);
1527 opts.prefix_map.debug.push(old, new);
1528 opts.prefix_map.profile.push(old, new);
1529 }
1530 // A whole optimization rather than a flag, and the family is taken rather than
1531 // refused because of what ignoring it does. There is none of it here yet, so a build
1532 // that asks for it gets a program that is correct and slower than it could have been,
1533 // which is what section 4.1 means by a hint about speed and what every compilation at
1534 // `-O0` already is. The objects settle the rest of the argument: gcc's `-flto` object
1535 // holds the bytecode and no machine code at all, and every object here holds the code,
1536 // which is exactly what `-ffat-lto-objects` asks gcc for. So a build passing `-flto`
1537 // to this compiler gets objects that are more usable than the ones it asked for rather
1538 // than different ones. Every value is still checked against gcc's, because somebody
1539 // who wrote `-flto=thin` meant clang and had better hear about it here.
1540 "-flto" => opts.lto.requested = true,
1541 "-fno-lto" => opts.lto.requested = false,
1542 _ if arg.starts_with("-flto=") => {
1543 let how = &arg["-flto=".len()..];
1544 opts.lto.jobs = how.parse().map_err(|()| {
1545 err(format!(
1546 "`{how}` is not a number of link time jobs, which is auto, jobserver or a \
1547 count above zero"
1548 ))
1549 })?;
1550 opts.lto.requested = true;
1551 }
1552 _ if arg.starts_with("-flto-partition=") => {
1553 let how = &arg["-flto-partition=".len()..];
1554 opts.lto.partition = how.parse().map_err(|()| {
1555 err(format!(
1556 "`{how}` is not a partitioning model, which is balanced, 1to1, one, max \
1557 or none"
1558 ))
1559 })?;
1560 }
1561 _ if arg.starts_with("-flto-compression-level=") => {
1562 let how = &arg["-flto-compression-level=".len()..];
1563 let level =
1564 how.parse::<u8>().ok().filter(|level| *level <= 19).ok_or_else(|| {
1565 err(format!("`{how}` is not a compression level, 0 to 19"))
1566 })?;
1567 opts.lto.compression = Some(level);
1568 }
1569 // Whether the object keeps its machine code as well as the bytecode. It always does
1570 // here, so the first of these describes what happens and the second asks for an object
1571 // with less in it, which is a smaller file and not a different program, so both are
1572 // taken.
1573 "-ffat-lto-objects" | "-fno-fat-lto-objects" => {}
1574 // Whether the linker is handed a plugin that does the link time work. The design in
1575 // `spec/09-optimizer.md` has this driver doing that work itself and never loading a
1576 // plugin into anybody, so neither answer is a question it has to hold.
1577 "-fuse-linker-plugin" | "-fno-use-linker-plugin" => {}
1578 // Reading a profile back. Taken for the reason the family above it is: nothing here
1579 // reads one, so a build that asks gets the program it would have got anyway, and gcc
1580 // itself produces a byte for byte identical object from `-fprofile-use` when there are
1581 // no counts beside the file. The path is recorded for the pass that will read it. The
1582 // warning gcc prints when it looked and found nothing is deliberately not copied,
1583 // because nothing here looks, and a warning about a file that was never opened would
1584 // fire on the builds that have a perfectly good profile as well as on the ones that
1585 // do not.
1586 "-fprofile-use" => opts.profile_data.requested = true,
1587 "-fno-profile-use" => opts.profile_data.requested = false,
1588 _ if arg.starts_with("-fprofile-use=") => {
1589 opts.profile_data.path = Some(arg["-fprofile-use=".len()..].to_string());
1590 opts.profile_data.requested = true;
1591 }
1592 _ if arg.starts_with("-fprofile-dir=") => {
1593 opts.profile_data.dir = Some(arg["-fprofile-dir=".len()..].to_string());
1594 }
1595 "-fprofile-abs-path" => opts.profile_data.absolute = true,
1596 "-fno-profile-abs-path" => opts.profile_data.absolute = false,
1597 "-fprofile-correction" => opts.profile_data.correction = true,
1598 "-fno-profile-correction" => opts.profile_data.correction = false,
1599 "-fprofile-partial-training" => opts.profile_data.partial_training = true,
1600 "-fno-profile-partial-training" => opts.profile_data.partial_training = false,
1601 // Writing the counts rather than reading them, which is refused rather than taken and
1602 // is the same line `-gsplit-dwarf` falls on the far side of. Ignoring these means a
1603 // file a build declared as an output never appears: the instrumented program writes a
1604 // `.gcda` as it exits and `-ftest-coverage` writes a `.gcno` beside the object, and a
1605 // two stage build that got neither would go on to optimize against no counts at all
1606 // and report coverage of nothing, with nothing along the way saying so. The objects
1607 // say the rest: gcc's `-fprofile-generate` object holds 375 bytes of code where a
1608 // plain one holds 71, and 296 bytes of counters that a plain one does not have, so
1609 // this is a flag that changes the output rather than a hint about speed.
1610 "-fprofile-arcs"
1611 | "--coverage"
1612 | "-fcondition-coverage"
1613 | "-fpath-coverage"
1614 | "-fprofile-generate" => {
1615 return Err(err(format!(
1616 "{arg}: this compiler does not instrument for profiling, and a build that \
1617 expects the counts a run of the instrumented program writes would optimize \
1618 against nothing on its second pass, see spec/04-driver-and-cli.md"
1619 )));
1620 }
1621 _ if arg.starts_with("-fprofile-generate=") => {
1622 return Err(err(format!(
1623 "{arg}: this compiler does not instrument for profiling, and a build that \
1624 expects the counts a run of the instrumented program writes would optimize \
1625 against nothing on its second pass, see spec/04-driver-and-cli.md"
1626 )));
1627 }
1628 "-ftest-coverage" => {
1629 return Err(err(format!(
1630 "{arg}: this compiler writes no `.gcno` file beside the object, and a build \
1631 that expects one would wait for a file that never arrives, see \
1632 spec/04-driver-and-cli.md"
1633 )));
1634 }
1635 // The rest of the family describes instrumentation that is refused above, so what is
1636 // left to do with them is check them and drop them. They are checked because a
1637 // misspelling in a distribution's flags is worth finding here rather than on the day
1638 // the instrumentation lands, and dropped because there is nothing for an answer about
1639 // how a counter is written to be an answer about.
1640 _ if arg.starts_with("-fprofile-update=") => {
1641 let how = &arg["-fprofile-update=".len()..];
1642 if !matches!(how, "single" | "atomic" | "prefer-atomic") {
1643 return Err(err(format!(
1644 "`{how}` is not a profile update method, which is single, atomic or \
1645 prefer-atomic"
1646 )));
1647 }
1648 }
1649 _ if arg.starts_with("-fprofile-reproducible=") => {
1650 let how = &arg["-fprofile-reproducible=".len()..];
1651 if !matches!(how, "serial" | "parallel-runs" | "multithreaded") {
1652 return Err(err(format!(
1653 "`{how}` is not a profile reproducibility method, which is serial, \
1654 parallel-runs or multithreaded"
1655 )));
1656 }
1657 }
1658 "-fprofile-values" | "-fno-profile-values" | "-fprofile-info-section" => {}
1659 "-fno-test-coverage" | "-fno-profile-arcs" | "-fno-profile-generate" => {}
1660 _ if arg.starts_with("-fprofile-filter-files=")
1661 || arg.starts_with("-fprofile-exclude-files=")
1662 || arg.starts_with("-fprofile-note=") => {}
1663 // What every name gets when nothing in the source said, which the attribute in the
1664 // source overrides rather than the other way round. Before the optimizer's `-f`
1665 // family below for the reason the tier below it is.
1666 _ if arg.starts_with("-fvisibility=") => {
1667 let seen = &arg["-fvisibility=".len()..];
1668 opts.visibility = seen.parse().map_err(|()| {
1669 err(format!(
1670 "`{seen}` is not a visibility, which is default, hidden, internal or \
1671 protected"
1672 ))
1673 })?;
1674 }
1675 // Which edges of a control flow transfer are checked. Before the optimizer's `-f`
1676 // family below for the reason the two above it are, and last of the three so that the
1677 // bare spelling and the negative one are matched exactly rather than by this.
1678 _ if arg.starts_with("-fcf-protection=") => {
1679 let edges = &arg["-fcf-protection=".len()..];
1680 opts.control = edges.parse().map_err(|()| {
1681 err(format!(
1682 "`{edges}` is not a control flow protection, which is full, branch, \
1683 return, none or check"
1684 ))
1685 })?;
1686 }
1687 // How much room every function opens with for something to be written over later.
1688 // Before the optimizer's `-f` family below for the reason the ones above it are.
1689 _ if arg.starts_with("-fpatchable-function-entry=") => {
1690 let room = &arg["-fpatchable-function-entry=".len()..];
1691 opts.patchable = room.parse().map_err(|()| {
1692 err(format!(
1693 "`{room}` is not an amount of room to reserve, which is a number of bytes and then, after a comma, how many of them go in front of the function's own label"
1694 ))
1695 })?;
1696 }
1697 // The memory safety monitor, from section 15.4 of
1698 // `spec/safe-memory/15-integration.md`. Before the optimizer's `-f` family below,
1699 // because a pass that took the name `safety=detect` would otherwise be handed the
1700 // flag, and the tier is not a pass.
1701 _ if arg.starts_with("-fsafety=") => {
1702 let tier = &arg["-fsafety=".len()..];
1703 opts.safety = tier.parse().map_err(|()| {
1704 err(format!(
1705 "`{tier}` is not a safety tier, which is off, detect, enforce or kernel"
1706 ))
1707 })?;
1708 }
1709 // Whether padding participates, from section 9.3 of document 09. Spelled out rather
1710 // than folded into the tier because it is a departure somebody who has read that
1711 // section makes, and the two defaults it describes are a property of what is being
1712 // built rather than of how much checking is wanted.
1713 _ if arg.starts_with("-fsafety-init=") => {
1714 let mode = &arg["-fsafety-init=".len()..];
1715 opts.padding = mode.parse().map_err(|()| {
1716 err(format!("`{mode}` is not a padding mode, which is padding or nopadding"))
1717 })?;
1718 }
1719 // Row S4, from section 9.4 of document 09. A bare flag with no value, because the
1720 // strict form of that section needs a member id the front end does not name yet and
1721 // accepting the spelling for it would be accepting a promise this build cannot keep.
1722 // Before `-fno-` is looked at below, for the reason the tier is.
1723 "-fsafety-subobject" => opts.subobject = rucc_session::Subobject::Members,
1724 "-fno-safety-subobject" => opts.subobject = rucc_session::Subobject::Off,
1725 _ if arg.starts_with("-fsafety-subobject=") => {
1726 let form = &arg["-fsafety-subobject=".len()..];
1727 return Err(err(format!(
1728 "`{form}` is not a form of -fsafety-subobject. The flag takes no value, and \
1729 the strict form of section 9.4 is tamnd/rucc#967"
1730 )));
1731 }
1732 // Row Y8, from section 9.6 of document 09. A bare flag with no value, for the reason
1733 // the one above has none: there is one form of this check and a spelling that suggested
1734 // otherwise would be promising something. Before `-fno-` is looked at below, the same
1735 // way.
1736 "-fsafety-restrict" => opts.promise = rucc_session::Promise::Blocks,
1737 "-fno-safety-restrict" => opts.promise = rucc_session::Promise::Off,
1738 _ if arg.starts_with("-fsafety-restrict=") => {
1739 let form = &arg["-fsafety-restrict=".len()..];
1740 return Err(err(format!(
1741 "`{form}` is not a form of -fsafety-restrict. The flag takes no value."
1742 )));
1743 }
1744 // Section 9.5's races, which take a value because the section gives them three modes
1745 // and the difference between two of them is which classes get reported rather than how
1746 // much is recorded. `-fno-` is the same as `=off` and is spelled out here for the same
1747 // reason the two above spell theirs out.
1748 _ if arg.starts_with("-fsafety-races=") => {
1749 let mode = &arg["-fsafety-races=".len()..];
1750 opts.races = mode.parse().map_err(|()| {
1751 err(format!("`{mode}` is not a race mode, which is off, metadata or pointer"))
1752 })?;
1753 }
1754 "-fno-safety-races" => opts.races = rucc_session::Races::Off,
1755 // The sanitizers of document 12, which are checks at run time rather than a way of
1756 // generating the same program. Each name is held to gcc 16's list, and what is still
1757 // asked for by the end of the line is answered after the loop, so that a command line
1758 // which turns one on and then off again is a command line that asked for nothing.
1759 //
1760 // Before the optimizer's `-f` family below, for the reason the tier above it is.
1761 _ if arg.starts_with("-fsanitize=") => {
1762 for one in arg["-fsanitize=".len()..].split(',') {
1763 if one == "all" {
1764 // gcc takes `all` only in the negative, because turning every check on at
1765 // once includes checks that contradict each other.
1766 return Err(err(
1767 "`-fsanitize=all` is not a gcc option, only `-fno-sanitize=all` is",
1768 ));
1769 }
1770 if !SANITIZERS.contains(&one) {
1771 return Err(err(format!(
1772 "`{one}` is not a sanitizer, see spec/04-driver-and-cli.md section 4.7"
1773 )));
1774 }
1775 if !sanitizers.contains(&one) {
1776 sanitizers.push(one);
1777 }
1778 }
1779 }
1780 _ if arg.starts_with("-fno-sanitize=") => {
1781 for one in arg["-fno-sanitize=".len()..].split(',') {
1782 if one == "all" {
1783 sanitizers.clear();
1784 continue;
1785 }
1786 if !SANITIZERS.contains(&one) {
1787 return Err(err(format!(
1788 "`{one}` is not a sanitizer, see spec/04-driver-and-cli.md section 4.7"
1789 )));
1790 }
1791 sanitizers.retain(|asked| *asked != one);
1792 }
1793 }
1794 // What a check does when it fires, and where the records about the checked objects go.
1795 // Each of them is an answer about the sanitizers refused after the loop, so there is
1796 // nothing left for them to change here. The names are still held to the list, because
1797 // a misspelling in a build's flags is worth finding when the compiler reads it.
1798 _ if arg.starts_with("-fsanitize-recover=")
1799 || arg.starts_with("-fno-sanitize-recover=")
1800 || arg.starts_with("-fsanitize-trap=")
1801 || arg.starts_with("-fno-sanitize-trap=") =>
1802 {
1803 // The guard above matched on a spelling that has an `=` in it, so the tail is
1804 // whatever follows the first one.
1805 let how = arg.split_once('=').map_or("", |(_, rest)| rest);
1806 for one in how.split(',') {
1807 if one != "all" && !SANITIZERS.contains(&one) {
1808 return Err(err(format!(
1809 "`{one}` is not a sanitizer, see spec/04-driver-and-cli.md section 4.7"
1810 )));
1811 }
1812 }
1813 }
1814 "-fsanitize-undefined-trap-on-error"
1815 | "-fsanitize-address-use-after-scope"
1816 | "-fno-sanitize-address-use-after-scope" => {}
1817 _ if arg.starts_with("-fsanitize-sections=") => {}
1818 // Counting which edges a run reached, which is how a fuzzer knows an input was worth
1819 // keeping. Refused rather than dropped, because a fuzzer whose calls into
1820 // `__sanitizer_cov_*` were never generated runs blind and reports coverage of nothing,
1821 // and there is no point in the campaign where that announces itself.
1822 _ if arg.starts_with("-fsanitize-coverage=") => {
1823 let how = &arg["-fsanitize-coverage=".len()..];
1824 for one in how.split(',') {
1825 if !matches!(one, "trace-pc" | "trace-cmp") {
1826 return Err(err(format!(
1827 "`{one}` is not a coverage instrumentation, which is trace-pc or \
1828 trace-cmp"
1829 )));
1830 }
1831 }
1832 return Err(err(format!(
1833 "{arg}: this compiler generates no coverage callbacks, and a fuzzer built \
1834 with it would run without any feedback at all, see \
1835 spec/04-driver-and-cli.md section 4.7"
1836 )));
1837 }
1838 // The optimizer's own flags, from section 9.10 of `spec/09-optimizer.md`. These come
1839 // after every `-f` the rest of the compiler answers to, so a pass can never take a
1840 // name that already means something else on the command line.
1841 _ if arg.starts_with("-fpass-fuel=") => {
1842 let (name, count) = arg["-fpass-fuel=".len()..]
1843 .split_once('=')
1844 .ok_or_else(|| err("-fpass-fuel= is spelled <pass>=<count>"))?;
1845 if rucc_opt::pass::find(name).is_none() {
1846 return Err(err(format!(
1847 "`{name}` is not a pass this compiler has, see --print-pipeline"
1848 )));
1849 }
1850 let count: u32 = count
1851 .parse()
1852 .map_err(|_| err(format!("`{count}` is not a number of transformations")))?;
1853 opts.pass_fuel.push((name.to_owned(), count));
1854 }
1855 _ if arg.starts_with("-fpass-fuel-global=") => {
1856 let count = &arg["-fpass-fuel-global=".len()..];
1857 let count: u32 = count
1858 .parse()
1859 .map_err(|_| err(format!("`{count}` is not a number of transformations")))?;
1860 opts.pass_fuel_global = Some(count);
1861 }
1862 _ if arg.starts_with("-frucc-trace=") => {
1863 let path = &arg["-frucc-trace=".len()..];
1864 if path.is_empty() {
1865 return Err(err("-frucc-trace= needs a file to write to"));
1866 }
1867 opts.trace = Some(path.to_owned());
1868 }
1869 // Everything from `-fopt-info` to the end of the argument, which is optional
1870 // keywords joined by hyphens and an optional `=<file>`. Checked here rather than
1871 // where the remarks are printed, because by then the compilation somebody wanted
1872 // to hear about is over.
1873 _ if arg == "-fopt-info"
1874 || arg.starts_with("-fopt-info=")
1875 || arg.starts_with("-fopt-info-") =>
1876 {
1877 let rest = &arg["-fopt-info".len()..];
1878 let (kinds, file) = match rest.split_once('=') {
1879 Some((kinds, file)) => (kinds, Some(file)),
1880 None => (rest, None),
1881 };
1882 let kinds = kinds.strip_prefix('-').unwrap_or(kinds);
1883 rucc_opt::Wants::none().add(kinds).map_err(err)?;
1884 opts.opt_info.push(kinds.to_owned());
1885 if let Some(file) = file {
1886 if file.is_empty() {
1887 return Err(err("-fopt-info= was given no file to write to"));
1888 }
1889 opts.opt_info_file = Some(file.to_owned());
1890 }
1891 }
1892 _ if arg.starts_with("-fdump-ir=") => {
1893 // Checked here rather than where the dumps are taken, because the compilation
1894 // that would have been dumped is over by then.
1895 let spec = &arg["-fdump-ir=".len()..];
1896 rucc_opt::Dumps::default().add(spec).map_err(err)?;
1897 opts.dump_ir.push(spec.to_owned());
1898 }
1899 // Before the bare `-f<pass>` below, because a pass called `enable-something` would
1900 // otherwise take the flag away from the gate. Checked here rather than where the
1901 // pipeline reads it, for the reason that applies to all of these: a misspelled pass
1902 // name that quietly gated nothing looks exactly like a pass that is not the guilty
1903 // one, and a bisection would carry on past the thing it was looking for.
1904 _ if arg.starts_with("-fdisable-") || arg.starts_with("-fenable-") => {
1905 let on = arg.starts_with("-fenable-");
1906 let spec = &arg[if on { "-fenable-".len() } else { "-fdisable-".len() }..];
1907 rucc_opt::Gates::default().add(on, spec).map_err(err)?;
1908 opts.pass_gates.push((on, spec.to_owned()));
1909 }
1910 // gcc's spelling for a pass this compiler has under a shorter name. It goes above the
1911 // two arms below rather than into the pile of gcc pass names further down, because the
1912 // pass is here: dropping the flag would leave a build that asked for unrolling without
1913 // it, and refusing it stops the build outright, which is what libtommath's makefile
1914 // ran into. `-funroll-all-loops` is deliberately not in here: gcc's is the one that
1915 // unrolls without a trip count, which is a different and usually worse thing.
1916 "-funroll-loops" => opts.passes.push(("unroll".to_owned(), true)),
1917 "-fno-unroll-loops" => opts.passes.push(("unroll".to_owned(), false)),
1918 // Here rather than through the two arms below, because what this names is not a
1919 // `rucc_opt::Pass`. Section 34.6's propagation is a module at a time and everything in
1920 // the pass list is one function at a time. `-fipa-cp-clone` is deliberately not here:
1921 // gcc turns that one on at `-O3` and it is in the list of what M4 does not build.
1922 "-fipa-cp" => opts.passes.push((rucc_opt::ipcp::NAME.to_owned(), true)),
1923 "-fno-ipa-cp" => opts.passes.push((rucc_opt::ipcp::NAME.to_owned(), false)),
1924 // The other half of the same section, here for the same reason, and `-fipa-sra` in gcc
1925 // is the aggregate splitting as well as the parameter removal. Asking for it gets the
1926 // half that is built.
1927 "-fipa-sra" => opts.passes.push((rucc_opt::ipasra::NAME.to_owned(), true)),
1928 "-fno-ipa-sra" => opts.passes.push((rucc_opt::ipasra::NAME.to_owned(), false)),
1929 // And the printf family fold, which is a module at a time for the same reason and so is
1930 // not a `rucc_opt::Pass` either. gcc has no flag of its own for this one, since
1931 // `-fno-builtin` already turns it off along with everything else the standard names
1932 // mean. This spelling is for taking one thing away during a bisection without taking
1933 // the rest of section 20.1 away with it.
1934 "-flibcall" => opts.passes.push((rucc_opt::libcall::NAME.to_owned(), true)),
1935 "-fno-libcall" => opts.passes.push((rucc_opt::libcall::NAME.to_owned(), false)),
1936 _ if arg.strip_prefix("-fno-").is_some_and(|n| rucc_opt::pass::find(n).is_some()) => {
1937 opts.passes.push((arg["-fno-".len()..].to_owned(), false));
1938 }
1939 _ if arg.strip_prefix("-f").is_some_and(|n| rucc_opt::pass::find(n).is_some()) => {
1940 opts.passes.push((arg["-f".len()..].to_owned(), true));
1941 }
1942 // The flags that name a pass of gcc's own. They arrive from the torture suite, where a
1943 // program reduced from a miscompilation usually names the pass that miscompiled it on
1944 // its `dg-options` line, and they arrive from hand written build files for the same
1945 // reason. Section 4.1 sorts a flag by what the output would be without it, and by that
1946 // rule these are one pile: a flag that turns one of gcc's passes on or off is asking
1947 // for a compiler that does not exist here, and the program it is attached to is a
1948 // correctness test that passes either way. Turning on a pass we do not have costs
1949 // speed, turning off a pass we do not have costs nothing, and neither changes what the
1950 // program computes.
1951 //
1952 // rucc's own pass names are matched above this, so `-fno-dce` turns off the dce this
1953 // compiler has rather than landing here, and the day one of these names becomes a pass
1954 // here it stops being taken and dropped without anybody editing this list.
1955 //
1956 // Two of them are prefixes rather than names, which is the one place this file takes a
1957 // family instead of a flag. gcc files its gimple passes under `-ftree-` and its
1958 // interprocedural passes under `-fipa-`, both namespaces are pass selection and
1959 // nothing else, and there is no member of either that changes the meaning of a program
1960 // that was already correct. The rest are written out one at a time, because they live
1961 // in the flat `-f` namespace where the neighbours do change meanings.
1962 _ if arg.starts_with("-ftree-") || arg.starts_with("-fno-tree-") => {}
1963 _ if arg.starts_with("-fipa-") || arg.starts_with("-fno-ipa-") => {}
1964 "-fexpensive-optimizations" | "-fno-expensive-optimizations" => {}
1965 "-fmodulo-sched" | "-fno-modulo-sched" => {}
1966 "-fvect-cost-model" | "-fno-vect-cost-model" => {}
1967 _ if arg.starts_with("-fvect-cost-model=") || arg.starts_with("-fsimd-cost-model=") => {
1968 }
1969 "-fearly-inlining" | "-fno-early-inlining" => {}
1970 // The one of the family that does reach the optimizer, since the step it names is built:
1971 // `-fno-inline` stops a function declared `inline` from being inlined and leaves
1972 // `always_inline` alone, which is what it does in gcc.
1973 "-finline" => opts.passes.push((rucc_opt::inline::NAME.to_owned(), true)),
1974 "-fno-inline" => opts.passes.push((rucc_opt::inline::NAME.to_owned(), false)),
1975 // The called once half of the same step, on its own, which leaves the `inline` hint and
1976 // `always_inline` as they are. tamnd/rucc#1966.
1977 "-finline-functions-called-once" => {
1978 opts.passes.push((rucc_opt::inline::ONCE.to_owned(), true));
1979 }
1980 "-fno-inline-functions-called-once" => {
1981 opts.passes.push((rucc_opt::inline::ONCE.to_owned(), false));
1982 }
1983 "-finline-functions"
1984 | "-fno-inline-functions"
1985 | "-finline-small-functions"
1986 | "-fno-inline-small-functions" => {}
1987 "-foptimize-strlen" | "-fno-optimize-strlen" => {}
1988 "-fira-share-spill-slots" | "-fno-ira-share-spill-slots" => {}
1989 // Where a function starts, which is a thing this compiler already decides and so is a
1990 // request it can answer rather than one it has to drop. The bare form asks for the
1991 // target's default and the default here is the sixteen bytes gcc also gives, so it
1992 // says nothing; a number is a floor under every function that did not ask for more
1993 // itself; and the negative form asks for the smallest boundary the target has. gcc 16
1994 // rounds a number that is not a power of two up rather than refusing it, which is what
1995 // `=3` giving `.p2align 2` on x86-64 means, so this rounds too.
1996 "-falign-functions" => opts.align_functions = None,
1997 "-fno-align-functions" => opts.align_functions = Some(MIN_FUNC_ALIGN),
1998 _ if arg.starts_with("-falign-functions=") => {
1999 opts.align_functions = function_alignment(&arg["-falign-functions=".len()..])
2000 .ok_or_else(|| {
2001 err(format!("{arg}: the alignment has to be a number of bytes"))
2002 })?;
2003 }
2004 // The head of every hot loop, which is padded when this is asked for so that a loop that
2005 // fits in a 64 byte line does not cross one. Both directions of the plain form are
2006 // answered. A number is taken and says nothing, because the boundary here is the
2007 // line's and a build that names another is asking for speed rather than for a
2008 // different program.
2009 "-falign-loops" => opts.align_loops = Some(true),
2010 "-fno-align-loops" => opts.align_loops = Some(false),
2011 // The other two of the family, which are about padding in front of any label and in
2012 // front of a label only a jump reaches. This compiler writes neither, and what they
2013 // ask for is speed: a label on a boundary computes what a label off one computes. So
2014 // they are taken and dropped for the reason `-march=` is, and the numbered form of
2015 // the loop flag with them.
2016 _ if arg.starts_with("-falign-labels")
2017 || arg.starts_with("-falign-loops=")
2018 || arg.starts_with("-falign-jumps")
2019 || arg.starts_with("-fno-align-labels")
2020 || arg.starts_with("-fno-align-jumps") => {}
2021 // The charset flags are not in that pile, because an encoding is a statement about
2022 // what the bytes of the source mean rather than about how fast the output is. The
2023 // preprocessor reads UTF-8 and has no converter, so the one name that describes what
2024 // already happens is taken and every other name is refused. Spelled without regard to
2025 // case and with both of the spellings iconv answers to, since a build writes whichever
2026 // one its author typed.
2027 _ if arg.starts_with("-finput-charset=") => {
2028 let name = &arg["-finput-charset=".len()..];
2029 if !name.eq_ignore_ascii_case("utf-8") && !name.eq_ignore_ascii_case("utf8") {
2030 return Err(err(format!(
2031 "-finput-charset={name}: the preprocessor reads UTF-8 and has no \
2032 converter, so a file in another encoding would be read as though it were \
2033 UTF-8 rather than converted",
2034 )));
2035 }
2036 }
2037 // What C has of exceptions, which is a `cleanup` handler an unwind has to run and the
2038 // `__EXCEPTIONS` that tells a header so. The walk is what turns down the handler it has
2039 // no landing pad for, so a unit with none of them is taken whole.
2040 "-fexceptions" => exceptions = Some(true),
2041 "-fno-exceptions" => exceptions = Some(false),
2042 "-fnon-call-exceptions" => opts.non_call_exceptions = true,
2043 "-fno-non-call-exceptions" => opts.non_call_exceptions = false,
2044 // Whether an instruction that could raise one may still be deleted when nothing uses
2045 // what it computes. Nothing here keeps a dead one, and neither does gcc in a C unit
2046 // with no handler around it, so both spellings describe the code as it is.
2047 "-fdelete-dead-exceptions" | "-fno-delete-dead-exceptions" => {}
2048 "-finstrument-functions" => opts.instrument_functions = true,
2049 "-fno-instrument-functions" => opts.instrument_functions = false,
2050 // The unstable options, spelled the way rustc spells them and carrying the same
2051 // promise, which is none: one of these may change or go away in any release. They are
2052 // measurements and debugging aids rather than things a build asks for, which is why
2053 // none of them is in the usage text and all of them are in section 4.11 of
2054 // `spec/04-driver-and-cli.md`.
2055 "-Zverify-each" => opts.verify_each = true,
2056 _ if arg.starts_with("-Zrule-coverage=") => {
2057 let file = &arg["-Zrule-coverage=".len()..];
2058 if file.is_empty() {
2059 return Err(err("-Zrule-coverage= needs a file to write to"));
2060 }
2061 opts.rule_coverage = Some(file.to_owned());
2062 }
2063 _ if arg.starts_with("-Zcycle-accurate-model=") => {
2064 let value = &arg["-Zcycle-accurate-model=".len()..];
2065 opts.cycle_accurate_model = match value {
2066 "yes" | "1" => Some(true),
2067 "no" | "0" => Some(false),
2068 _ => {
2069 return Err(err("-Zcycle-accurate-model= takes yes or no"));
2070 }
2071 };
2072 }
2073 _ if arg.starts_with("-Zregalloc=") => {
2074 opts.backtracking = match &arg["-Zregalloc=".len()..] {
2075 "backtracking" => Some(true),
2076 "single" => Some(false),
2077 _ => return Err(err("-Zregalloc= takes backtracking or single")),
2078 };
2079 }
2080 _ if arg.starts_with("-Zswitch=") => {
2081 let shape = &arg["-Zswitch=".len()..];
2082 if rucc_codegen::switch::Force::named(shape).is_none() {
2083 return Err(err("-Zswitch= takes table, tree or walk"));
2084 }
2085 opts.switch_shape = Some(shape.to_owned());
2086 }
2087 _ if arg.starts_with("-Zlowering=") => {
2088 let file = &arg["-Zlowering=".len()..];
2089 if file.is_empty() {
2090 return Err(err("-Zlowering= needs a file to write to"));
2091 }
2092 opts.lowering_dump = Some(file.to_owned());
2093 }
2094 _ if arg.starts_with("-Zregister-pressure=") => {
2095 let file = &arg["-Zregister-pressure=".len()..];
2096 if file.is_empty() {
2097 return Err(err("-Zregister-pressure= needs a file to write to"));
2098 }
2099 opts.register_pressure = Some(file.to_owned());
2100 }
2101 _ if arg.starts_with("-Z") => {
2102 return Err(err(format!(
2103 "`{arg}` is not an unstable option this compiler has, see \
2104 spec/04-driver-and-cli.md section 4.11 for the ones it does"
2105 )));
2106 }
2107 // The word size, which is a statement about the target and is taken as one. A build
2108 // that says the size the target already has is saying nothing, and one that says the
2109 // other size is asking for a target this compiler does not have, which it is told
2110 // rather than being given the wrong one.
2111 "-m64" | "-m32" | "-mx32" => {
2112 let want: u32 = match arg {
2113 "-m64" => 64,
2114 _ => 32,
2115 };
2116 let have = rucc_target::TargetInfo::new(opts.target).pointer_width;
2117 if have != want {
2118 return Err(err(format!(
2119 "{arg} asks for a {want} bit target and {} is {have} bit, use \
2120 --target= to name the one you mean",
2121 opts.target
2122 )));
2123 }
2124 }
2125 // One extension of the x86-64 instruction set, on or off, which is `-msse4.2` and its
2126 // relatives. Only the ones this compiler has the intrinsics for may be turned on for a
2127 // whole unit, because what turning one on does here is define the macro, and a macro
2128 // is a promise to a header that the names behind it exist. Turning one off is taken
2129 // for any name gcc knows, since nothing is promised by it, except for the baseline:
2130 // SSE2 is where the psABI passes a `double`, so a unit without it is a different
2131 // calling convention and not a smaller instruction set.
2132 _ if isa_name(arg).is_some() => {
2133 let Some((_, feature, on)) = isa_name(arg) else { continue };
2134 if on && !feature.honoured() {
2135 return Err(err(format!(
2136 "{arg}: this compiler has no intrinsics for {} yet, so it cannot build a \
2137 whole unit for it",
2138 feature.name()
2139 )));
2140 }
2141 if !on && rucc_target::Isa::baseline().has(feature) {
2142 return Err(err(format!(
2143 "{arg}: {} is part of the x86-64 baseline and the psABI passes values in \
2144 it, so a unit built without it would call and be called differently",
2145 feature.name()
2146 )));
2147 }
2148 isa.read(&arg["-m".len()..]).map_err(|_| err(format!("unknown option `{arg}`")))?;
2149 isa_flag.get_or_insert(arg);
2150 }
2151 // Which processor in the family to build for. What it decides is the extensions of
2152 // the instruction set the unit may assume, which is the macros, and only on x86-64;
2153 // see `rucc_target::isa`. A processor it has no list for is built for as the
2154 // baseline, which is a program that could have been faster rather than a program
2155 // that is wrong, and the same goes for every other target's processors. `-mtune=`
2156 // says what to schedule for and changes nothing a program can see.
2157 _ if arg.starts_with("-march=") => march = Some(&arg["-march=".len()..]),
2158 _ if arg.starts_with("-mtune=") || arg.starts_with("-mcpu=") => {}
2159 // The calling convention, which is not safe to ignore. Taken when it names the one
2160 // the target already uses and refused otherwise.
2161 _ if arg.starts_with("-mabi=") => {
2162 let want = &arg["-mabi=".len()..];
2163 let have = match opts.target.arch {
2164 rucc_target::Arch::X86_64 => "sysv",
2165 rucc_target::Arch::Aarch64 => "lp64",
2166 rucc_target::Arch::Riscv64 => "lp64d",
2167 };
2168 if want != have {
2169 return Err(err(format!(
2170 "{arg}: {} uses the {have} convention and this compiler has no other",
2171 opts.target
2172 )));
2173 }
2174 }
2175 // How far apart the pieces of the program may be. The small model is what we emit and
2176 // it is every hosted program's default; the kernel model is a different one and a
2177 // build that asks for it and does not get it links and then does not run.
2178 "-mcmodel=small" => {}
2179 // clang's spellings of the deployment target, which it takes over a version in the
2180 // tuple. gcc on a Mac takes the first. A target that is not Apple ignores it, as
2181 // clang does, so a makefile that always passes it still builds for Linux.
2182 _ if arg.starts_with("-mmacosx-version-min=")
2183 || arg.starts_with("-mmacos-version-min=") =>
2184 {
2185 let text = &arg[arg.find('=').map_or(arg.len(), |i| i + 1)..];
2186 let version = rucc_tuple::Version::parse(text)
2187 .ok_or_else(|| err(format!("`{text}` in `{arg}` is not a version")))?;
2188 min_version = Some(version);
2189 }
2190 _ if arg.starts_with("-mcmodel=") => {
2191 return Err(err(format!(
2192 "{arg}: this compiler emits the small code model and no other, see \
2193 spec/12-targets.md"
2194 )));
2195 }
2196 // GCC's own scripting language for how the driver builds a command line.
2197 // `spec/04-driver-and-cli.md` section 4.4 settles that we will not have it, so a
2198 // build reaching for it is told which flags do the same job.
2199 _ if arg.starts_with("-specs=") => {
2200 return Err(err(
2201 "-specs= is not supported: the parts of it builds rely on are -B, -L, \
2202 -nostdlib, -nostartfiles and -Wl,, see spec/04-driver-and-cli.md \
2203 section 4.4",
2204 ));
2205 }
2206 // Arguments meant for a separate assembler, which this compiler does not have: it is
2207 // inside it and does not read a command line. Refused rather than dropped, because
2208 // every one of these says something about the output and a build that asked for
2209 // `-Wa,--noexecstack` and was silently given an executable stack got the opposite of
2210 // what it asked for. The `-Wp,` ones this compiler understands were turned into its
2211 // own flags before the loop, so one that reaches here is one it does not.
2212 _ if arg.starts_with("-Wa,") || arg.starts_with("-Wp,") => {
2213 return Err(err(format!(
2214 "`{arg}` is an argument for a separate assembler or preprocessor, and both \
2215 are inside this compiler rather than programs it runs"
2216 )));
2217 }
2218 "-Xassembler" | "-Xpreprocessor" => {
2219 return Err(err(format!(
2220 "{arg} hands an argument to a separate assembler or preprocessor, and both \
2221 are inside this compiler rather than programs it runs"
2222 )));
2223 }
2224 // Everything else in the `-W` family. `spec/04-driver-and-cli.md` section 4.1 has
2225 // this one as a rule about build systems rather than about warnings: autoconf and
2226 // meson find out whether a warning flag exists by passing it and looking at the exit
2227 // status, so the answer has to be gcc's. A name gcc knows is accepted, and one it does
2228 // not is refused, the way gcc refuses clang's names. `-Wno-` of a name nobody knows is
2229 // accepted, because gcc accepts it too, but `-Werror=` and `-Wno-error=` of one are
2230 // not. None of them turns anything on yet, which #485 is about.
2231 _ if arg.starts_with("-W") => {
2232 let name = &arg["-W".len()..];
2233 let named = name.strip_prefix("error=").or_else(|| name.strip_prefix("no-error="));
2234 if let Some(named) = named {
2235 if !warnings::known(named) {
2236 return Err(err(format!("`{arg}`: no option `-W{named}`")));
2237 }
2238 } else if !name.is_empty() && !name.starts_with("no-") && !warnings::known(name) {
2239 return Err(err(format!("unknown option `{arg}`")));
2240 }
2241 }
2242 // Flags that name something this compiler does not do and would not do differently
2243 // if it did. `-fno-ident` is about a comment in the output that we do not write
2244 // either way, and the others are about a way of ordering the compilation that has
2245 // been GCC's only way for twenty years. `-mthreads` is mingw's, and what it links is
2246 // `libmingwthrd.a`, which mingw-w64 keeps as an empty archive because its CRT does the
2247 // thread cleanup for every program. Section 4.1 asks for the list to be short and for
2248 // adding to it to be deliberate, which is why it is written out here.
2249 "-fno-ident"
2250 | "-fident"
2251 | "-funit-at-a-time"
2252 | "-fno-unit-at-a-time"
2253 | "-shared-libgcc"
2254 | "-static-libgcc"
2255 | "-mthreads"
2256 | "-fpch-deps"
2257 | "-fno-pch-deps" => {}
2258 _ if arg.starts_with('-') && arg.len() > 1 => {
2259 // Silently ignoring an unknown flag is how a build ends up not doing what
2260 // its author asked. spec/13-gnu-compat.md section 13.4 makes this an error
2261 // for the flags that change code generation, and the safe default until the
2262 // flag table is populated is to reject everything we do not know.
2263 return Err(err(format!("unknown option `{arg}`")));
2264 }
2265 _ => inputs.push(Input { path: arg.to_owned(), forced, role: Role::File }),
2266 }
2267 }
2268
2269 // The fetch, before anything that resolves a compilation, because `--fetch` does not describe
2270 // one. It is here rather than in the loop so that `--offline` can forbid it whichever order the
2271 // two were written in, and it is before the refusals below so that a command line asking for a
2272 // sysroot is not told about a sanitizer.
2273 if let Some(named) = fetch {
2274 if fetch_msvc.is_some() {
2275 return Err(err(
2276 "--fetch and --fetch-msvc-sdk are two different commands and this command line \
2277 asked for both. --fetch gets what this release pins by URL and by hash, which for \
2278 a *-windows-msvc target is one Visual Studio build, and --fetch-msvc-sdk gets the \
2279 build Microsoft's channel names today. Run whichever one you meant",
2280 ));
2281 }
2282 return fetch_action(&named, offline, accepted, &inputs);
2283 }
2284 if let Some(named) = fetch_msvc {
2285 return fetch_msvc_action(&named, offline, accepted, &inputs);
2286 }
2287 if accepted {
2288 return Err(err(
2289 "--accept-licence says that Microsoft's Visual Studio Build Tools licence is accepted, \
2290 and nothing on this command line asked for anything that licence covers. \
2291 --fetch <tuple> or --fetch-msvc-sdk <tuple> for a *-windows-msvc target is the \
2292 command it belongs to, and an ordinary compile downloads nothing with it or \
2293 without it",
2294 ));
2295 }
2296
2297 // Last, so that it lands after every `-isystem` the command line gave. That is GCC's
2298 // order: a directory the user names outranks the compiler's own, and the compiler's own
2299 // outranks the library's. It is pushed after the loop rather than before it because
2300 // `SearchPath` appends within a group and the position is what the order is.
2301 // The same directory the headers were looked for under, because a sysroot is a statement
2302 // about a whole installation and not about half of one.
2303 // After the loop, because `-fno-sanitize=` can take back what an earlier flag asked for and a
2304 // command line that turns a check on and off again has asked for nothing. What is left is
2305 // refused rather than dropped, and it is the one place in this parser where the reason is not
2306 // that the output would differ. A sanitizer is a promise that the program is watched while it
2307 // runs, so a build that asks for one and is quietly given a program with no checks in it does
2308 // not get a slower program or a bigger file, it gets a test suite that passes for the wrong
2309 // reason. `-fsafety=` is the checking this compiler does have, and the message says so, because
2310 // somebody reaching for `-fsanitize=address` wants the nearest thing rather than a list of
2311 // options.
2312 if let Some(first) = sanitizers.first() {
2313 return Err(err(format!(
2314 "-fsanitize={first}: this compiler has no sanitizer instrumentation, and a build that \
2315 asked for one and got none would run its tests unchecked, see \
2316 spec/04-driver-and-cli.md section 4.7. `-fsafety=detect` is the memory checking this \
2317 compiler does have"
2318 )));
2319 }
2320 // The fast math family, replayed in order on top of what `-Ofast` implies. The startup file is
2321 // gcc's spec rather than the fields: it is linked when `-Ofast`, `-ffast-math` or
2322 // `-funsafe-math-optimizations` is still in force at the end of the line, whatever a later
2323 // member took back, and `-mdaz-ftz` decides it outright.
2324 let mut math = Math::default();
2325 let mut trapping = if ofast { math.set_fast(true) } else { true };
2326 for flag in &math_flags {
2327 match *flag {
2328 "-ftrapping-math" => trapping = true,
2329 "-fno-trapping-math" => trapping = false,
2330 "-ffast-math" => trapping = math.set_fast(true),
2331 "-fno-fast-math" => trapping = math.set_fast(false),
2332 "-funsafe-math-optimizations" => trapping = math.set_unsafe(true),
2333 "-fno-unsafe-math-optimizations" => trapping = math.set_unsafe(false),
2334 "-fmath-errno" => math.errno = true,
2335 "-fno-math-errno" => math.errno = false,
2336 "-ffinite-math-only" => math.finite_only = true,
2337 "-fno-finite-math-only" => math.finite_only = false,
2338 "-fsigned-zeros" => math.signed_zeros = true,
2339 "-fno-signed-zeros" => math.signed_zeros = false,
2340 "-freciprocal-math" => math.reciprocal = true,
2341 "-fno-reciprocal-math" => math.reciprocal = false,
2342 "-fassociative-math" => math.associative = true,
2343 "-fno-associative-math" => math.associative = false,
2344 _ => unreachable!("{flag} is not in the family"),
2345 }
2346 }
2347 opts.trapping_math = trapping;
2348 opts.math = math;
2349 let last = |on: &str, off: &str| {
2350 math_flags.iter().rev().find(|f| **f == on || **f == off).is_some_and(|f| *f == on)
2351 };
2352 link.fast_math = ofast
2353 || last("-ffast-math", "-fno-fast-math")
2354 || last("-funsafe-math-optimizations", "-fno-unsafe-math-optimizations");
2355 link.daz_ftz = daz_ftz;
2356 // The extensions, now that the target is known. On x86-64 the processor supplies whatever no
2357 // flag said. Anywhere else there are none to have, and a flag naming one is gcc's unknown
2358 // option there too, so it is refused the same way it would have been had it not looked like
2359 // an x86 flag.
2360 match opts.target.arch {
2361 rucc_target::Arch::X86_64 => {
2362 let base = match march {
2363 Some("native") => native_isa(),
2364 Some(name) => {
2365 rucc_target::Isa::level(name).unwrap_or_else(rucc_target::Isa::baseline)
2366 }
2367 None => rucc_target::Isa::baseline(),
2368 };
2369 opts.isa = isa.over(base);
2370 }
2371 rucc_target::Arch::Aarch64 | rucc_target::Arch::Riscv64 => {
2372 if let Some(flag) = isa_flag {
2373 return Err(err(format!("unknown option `{flag}`")));
2374 }
2375 opts.isa = rucc_target::Isa::NONE;
2376 }
2377 }
2378 opts.exceptions = exceptions.unwrap_or(opts.non_call_exceptions);
2379 link.sysroot = sysroot.clone();
2380 // Where a sysroot for a target that is not this machine would be. Read once, here, rather than
2381 // inside the link line, because a link line that read the environment could only be tested on a
2382 // machine whose environment said the right thing, and the link line is the last thing that
2383 // touches a binary. `spec/cross-compile/13-distribution.md` section 13.2 owns the answer.
2384 link.cache = Some(cache::dir());
2385 // And where a distribution's cross packages would have put a tree for the target, which is only
2386 // read when the target is not this machine and there is no sysroot of ours for it.
2387 link.usr = Some(PathBuf::from("/usr"));
2388 // And the ten field spelling of the target, because the release on it decides two things the
2389 // three field one cannot say: whether a target that is this architecture is still a cross
2390 // compile, and which directory under the cache it is against. After the loop because the last
2391 // `--target=` on the command line is the one that counts.
2392 link.pinned = pinned;
2393 // The deployment target, from the flag if there was one and from the tuple otherwise. Only an
2394 // Apple platform has one: anywhere else a version on the tuple is a libc or a preview number.
2395 if opts.target.os == rucc_target::Os::Darwin {
2396 opts.os_version = min_version.or_else(|| pinned.and_then(TargetTuple::os_version));
2397 link.os_version = opts.os_version;
2398 }
2399 // After the loop rather than where `-pthread` was read, so that it lands after the objects
2400 // that refer to it. A static link takes the definitions it needs from a library when it
2401 // reaches it and not afterwards, so a library before the objects is a library that answers
2402 // nothing.
2403 if threads {
2404 inputs.push(Input::library("pthread"));
2405 }
2406 if let Some(query) = query {
2407 return Ok(Action::Print(answer(&query, &opts, &link)?));
2408 }
2409 // `-M` and `-MM` produce the rule and nothing else, so the run stops after phase 4 whatever
2410 // else the command line asked for. Read here rather than where the flag was, because a `-c`
2411 // written after it has to lose and the loop cannot know that until it has ended. The output
2412 // file is where the rule goes rather than where an object would have gone, and the last
2413 // phase being the preprocessor is what makes that true without a second rule for it.
2414 if opts.deps.instead_of_compiling {
2415 opts.emit = EmitKind::Preprocessed;
2416 }
2417 if !nostdinc {
2418 opts.search.push_system(runtime::DIR);
2419 // And the library's after ours, which is the other half of the same order. They go on
2420 // here rather than at the point `--target=` or `--sysroot=` was read because either
2421 // one changes the answer and the last word on both is the end of the loop.
2422 //
2423 // Which library's is the question `link::cross_sysroot` answers, and it is asked here so
2424 // that the headers and the libraries come from the same place. A target that is this
2425 // machine reads this machine's headers, and a target that is not reads the ones in the
2426 // sysroot for it rather than the ones next door.
2427 let cross = link::cross_sysroot(opts.target, &link);
2428 let kernel = link::cross_kernel(opts.target, &link);
2429 let distro = link::distro_cross(opts.target, &link);
2430 // And the version of those headers, which only the bundled tree has an answer for. A host
2431 // glibc and a tree the user named both define `__GLIBC_MINOR__` in their own `features.h`,
2432 // and a second definition with a different value is a warning on every file, so the
2433 // condition is the same one that chose the directories.
2434 if cross.is_some() {
2435 let target = pinned.unwrap_or_else(|| opts.target.tuple());
2436 opts.glibc_minor = rucc_sysroot::bundled_glibc_minor(target).map_err(|skew| {
2437 err(format!(
2438 "{skew}; pin a release the tree has, or name a tree that has that one \
2439 with --sysroot"
2440 ))
2441 })?;
2442 }
2443 let system = library::header_dirs(
2444 opts.target,
2445 sysroot.as_deref(),
2446 cross.as_ref(),
2447 kernel.as_ref(),
2448 distro.as_ref(),
2449 );
2450 // The two licence walls of `spec/cross-compile/13-distribution.md` section 13.4, which are
2451 // the only way step 3 comes back with nothing on a hosted target. Section 8.6 asks for the
2452 // answer to name the licence and the lawful ways to get what is behind it, rather than
2453 // leaving a person with an `#include` that failed as though a directory had gone missing.
2454 //
2455 // It is left on the search path instead of refused here, because a program that includes
2456 // none of the library needs none of the SDK and section 8.6 is explicit that targeting the
2457 // platform has to keep working. So the reason waits until an include has actually failed,
2458 // which is the only moment it helps and the only moment it is true.
2459 //
2460 // The condition is that step 3 found nothing at all, so an `SDKROOT`, an `INCLUDE` or a mac
2461 // with Xcode on it all pass through untouched, and `-nostdinc` never reaches this block. A
2462 // `--sysroot` or `-isysroot` passes through as well, even when the tree it names turns out to
2463 // be empty or absent: somebody who wrote a path has already answered the question this
2464 // message asks, and answering it again over the top of a mistyped directory would hide the
2465 // mistake behind a licence notice.
2466 if system.is_empty() && sysroot.is_none() {
2467 let tuple = pinned.unwrap_or_else(|| opts.target.tuple());
2468 if let Some(wall) = rucc_sysroot::Wall::of(tuple) {
2469 opts.search.explain_missing_system(wall.no_headers(&tuple.to_canonical_string()));
2470 }
2471 }
2472 // And whether the tree somebody named is the release they asked for, which is the one
2473 // question left once the directories are settled and the only place both halves of it are
2474 // known. Only for a named tree, because that is the case where the release in the target
2475 // stops deciding anything, and `crate::glibc` is where the rest of the reasoning is.
2476 if sysroot.is_some() {
2477 notes.extend(glibc::skew(opts.target, pinned, &system));
2478 }
2479 for dir in system {
2480 opts.search.push_system(dir);
2481 }
2482 }
2483 // Once, here, rather than as each directory is pushed. A `-I` that names a system
2484 // directory has to lose to the system entry and the system entry is added last, so the
2485 // question cannot be answered until the whole path is known.
2486 opts.search.remove_duplicates();
2487
2488 // The target has to be resolved before the configuration is printed, so this check comes
2489 // after the loop rather than at the point `--print-config` was seen.
2490 if print_config {
2491 return Ok(Action::PrintConfig(Box::new(opts)));
2492 }
2493 if print_pipeline {
2494 return Ok(Action::PrintPipeline(Box::new(opts)));
2495 }
2496 let plan = Plan::new(&opts, &inputs, output.as_deref()).map_err(|e| err(e.message))?;
2497 if print_plan {
2498 return Ok(Action::PrintPlan {
2499 opts: Box::new(opts),
2500 plan: Box::new(plan),
2501 link: Box::new(link),
2502 });
2503 }
2504 Ok(Action::Compile {
2505 opts: Box::new(opts),
2506 plan: Box::new(plan),
2507 link: Box::new(link),
2508 jobs,
2509 verbose,
2510 notes,
2511 })
2512}
2513
2514/// What `--fetch <tuple>` asked for, or why it is not a thing that can be done.
2515///
2516/// The lookup happens here rather than at the point the bytes would move, so that a target this
2517/// release pins nothing for is a refusal from the parser and the only code that runs a downloader is
2518/// code that already knows what it is getting.
2519///
2520/// # Errors
2521///
2522/// [`CliError`] when `--offline` forbade it, when there are input files as well, when the tuple is
2523/// not a target this compiler knows, when its sysroot is behind Apple's licence wall, and when this
2524/// release pins no artifact for it.
2525fn fetch_action(
2526 named: &str,
2527 offline: bool,
2528 accepted: bool,
2529 inputs: &[Input],
2530) -> Result<Action, CliError> {
2531 // Not a precedence question. Section 13.2 says `--offline` forbids a fetch entirely, so a
2532 // command line that writes both has asked for two opposite things and the answer is to say so
2533 // rather than to pick one of them.
2534 if offline {
2535 return Err(err(
2536 "--fetch asks for a download and --offline forbids every download, so this command \
2537 line asks for two opposite things. Drop one of them: --offline is how a build says it \
2538 will not reach the network, and --fetch is one of the two things in this compiler \
2539 that reaches it",
2540 ));
2541 }
2542 if let Some(first) = inputs.first() {
2543 return Err(err(format!(
2544 "--fetch gets a sysroot and compiles nothing, so `{}` on the same command line is an \
2545 input that nothing would read",
2546 first.path
2547 )));
2548 }
2549 let target: TargetTuple = named
2550 .parse()
2551 .map_err(|why| err(format!("--fetch {named}: {why}, so there is no sysroot to get")))?;
2552 // The canonical spelling, because that is what a row is named by and what the directory under
2553 // the cache is called, and a person is free to write a tuple the long way round.
2554 let tuple = target.to_canonical_string();
2555 // Microsoft's side of the wall has something to fetch after all, which is the files its own
2556 // installer would fetch, from the build this release pins and only once the licence has been
2557 // accepted. Nothing of it is ours and nothing of it comes from us, which is why it is the other
2558 // action with a flag on it rather than a row in the table.
2559 if rucc_sysroot::Wall::of(target) == Some(rucc_sysroot::Wall::Microsoft) {
2560 return Ok(Action::FetchMsvcSdk { target, accepted, cache: cache::dir(), pinned: true });
2561 }
2562 // Before the table is consulted, because a target behind a licence wall is not a row that has not
2563 // been written yet. Section 13.4 is that no release pins one of these ever, so the message says
2564 // the licence and the two lawful ways rather than naming the producer that will publish the rest.
2565 if let Some(wall) = rucc_sysroot::Wall::of(target) {
2566 return Err(err(format!("--fetch {tuple}: {}", wall.no_fetch(&tuple))));
2567 }
2568 let Some(what) = rucc_sysroot::pinned_for_target(target) else {
2569 return Err(err(unpinned(&tuple)));
2570 };
2571 Ok(Action::Fetch { what, target, cache: cache::dir() })
2572}
2573
2574/// What `--fetch-msvc-sdk <tuple>` asks for, weighed the same way the fetch above is.
2575///
2576/// The target is resolved here rather than where the work happens, so that a tuple this compiler
2577/// does not know and a target that is not behind Microsoft's wall are refusals from the parser like
2578/// every other thing a command line can ask for and not have. Whether the licence was accepted is
2579/// carried rather than acted on, because what it changes is what the command does and not whether
2580/// the command line made sense.
2581///
2582/// # Errors
2583///
2584/// [`CliError`] when `--offline` forbade it, when there are input files as well, and when the tuple
2585/// is not a target this compiler knows.
2586fn fetch_msvc_action(
2587 named: &str,
2588 offline: bool,
2589 accepted: bool,
2590 inputs: &[Input],
2591) -> Result<Action, CliError> {
2592 if offline {
2593 return Err(err(
2594 "--fetch-msvc-sdk asks for a download and --offline forbids every download, so this \
2595 command line asks for two opposite things. Drop one of them: --offline is how a build \
2596 says it will not reach the network",
2597 ));
2598 }
2599 if let Some(first) = inputs.first() {
2600 return Err(err(format!(
2601 "--fetch-msvc-sdk gets an SDK and compiles nothing, so `{}` on the same command line \
2602 is an input that nothing would read",
2603 first.path
2604 )));
2605 }
2606 let target: TargetTuple = named.parse().map_err(|why| {
2607 err(format!("--fetch-msvc-sdk {named}: {why}, so there is no SDK to get"))
2608 })?;
2609 Ok(Action::FetchMsvcSdk { target, accepted, cache: cache::dir(), pinned: false })
2610}
2611
2612/// Why there is nothing to fetch for a target, which is a different sentence when the table is
2613/// empty.
2614///
2615/// A release that pins nothing and a release that pins eleven targets and not this one are two
2616/// situations, and a message that did not tell them apart would send somebody looking for a typo in
2617/// their tuple when the answer is that this work is not finished.
2618fn unpinned(tuple: &str) -> String {
2619 let pinned = rucc_sysroot::pinned_targets();
2620 if pinned.is_empty() {
2621 return format!(
2622 "this release pins no sysroot for {tuple}, and it pins none for any target yet. A \
2623 sysroot is built and published by the producer in tamnd/rucc-cross, per \
2624 spec/cross-compile/13-distribution.md section 13.8, and a release of this compiler \
2625 names one by URL and by hash afterwards. Until then, pass --sysroot=<dir> to compile \
2626 against a tree you have already"
2627 );
2628 }
2629 format!(
2630 "this release pins no sysroot for {tuple}. What it pins is {}. Pass --sysroot=<dir> to \
2631 compile against a tree you have already",
2632 pinned.join(", ")
2633 )
2634}
2635
2636/// Gets the artifact and installs it, saying what each step did.
2637///
2638/// The steps are section 13.8's and so are the messages: the transport is somebody else's program
2639/// and the check is ours, so a person reading this wants to know which downloader ran, that the
2640/// bytes matched, how many files the record named and where the tree ended up. A fetch of something
2641/// that is already there says that instead and moves nothing.
2642///
2643/// A Linux target is two artifacts, its own sysroot and the kernel header tree every Linux target
2644/// shares, and `kernel` is the second one when the target reads it. It is fetched after the sysroot
2645/// and by the same two steps, so a machine that has fetched one Linux target already has it and a
2646/// second target's fetch says so and moves nothing.
2647fn fetch_sysroot(
2648 what: &rucc_sysroot::Pinned,
2649 kernel: Option<&rucc_sysroot::Pinned>,
2650 target: TargetTuple,
2651 cache: &std::path::Path,
2652) -> i32 {
2653 let tuple = target.to_canonical_string();
2654 let say = |line: &str| println!("rucc: {tuple}: {line}");
2655 if let Err(why) = bring(what, cache, &say) {
2656 return complain(why);
2657 }
2658 let archive = what.archive_in(cache);
2659 match install::install(&archive, what.sha256, target, cache) {
2660 Ok(done) => report(&done, "sysroot", &say),
2661 Err(why) => return complain(why),
2662 }
2663 let Some(kernel) = kernel else { return 0 };
2664 if let Err(why) = bring(kernel, cache, &say) {
2665 return complain(why);
2666 }
2667 match install::install_kernel(&kernel.archive_in(cache), kernel.sha256, cache) {
2668 Ok(done) => {
2669 report(&done, "kernel header tree", &say);
2670 0
2671 }
2672 Err(why) => complain(why),
2673 }
2674}
2675
2676/// The download half of a fetch, for one artifact.
2677fn bring(
2678 what: &rucc_sysroot::Pinned,
2679 cache: &std::path::Path,
2680 say: &impl Fn(&str),
2681) -> Result<(), CliError> {
2682 let archive = what.archive_in(cache);
2683 match fetch::fetch(what.url, what.sha256, &archive)? {
2684 fetch::Fetched::AlreadyThere => {
2685 say(&format!("{} is already here and matches the hash", archive.display()));
2686 }
2687 fetch::Fetched::Downloaded(by) => {
2688 say(&format!("downloaded {} with {}", what.url, by.program()));
2689 }
2690 }
2691 Ok(())
2692}
2693
2694/// What an install did, in the words a person reading a fetch wants.
2695fn report(done: &install::Installed, what: &str, say: &impl Fn(&str)) {
2696 match &done.before {
2697 install::Before::Nothing => {
2698 say(&format!("{} files installed at {}", done.files, done.root.display()));
2699 }
2700 install::Before::TheSame => {
2701 say(&format!(
2702 "the same {what} is already at {}, so nothing moved",
2703 done.root.display()
2704 ));
2705 }
2706 install::Before::Different(was) => {
2707 say(&format!(
2708 "{} files installed at {}, over a tree whose record digested to {was}",
2709 done.files,
2710 done.root.display()
2711 ));
2712 }
2713 }
2714 say(&format!("the {what}'s record digests to {}", done.digest));
2715}
2716
2717/// What one of the `-dump` and `-print` flags prints.
2718///
2719/// GCC prints the name back unchanged when it cannot find the file a `-print` flag asked about,
2720/// which is what makes the answer safe to paste into a link line whether or not the file is
2721/// there, and this does the same.
2722fn answer(query: &Query, opts: &Options, link: &LinkOptions) -> Result<String, CliError> {
2723 let found = |name: &str| {
2724 link::find_in_search(link, opts.target, name)
2725 .map_or_else(|| name.to_owned(), |path| path.display().to_string())
2726 };
2727 Ok(match query {
2728 Query::Machine => opts.target.to_string(),
2729 Query::Version => opts.gnuc.major.to_string(),
2730 Query::FullVersion => {
2731 format!("{}.{}.{}", opts.gnuc.major, opts.gnuc.minor, opts.gnuc.patch)
2732 }
2733 Query::Multiarch => link::multiarch(opts.target),
2734 // The three lines GCC prints, in its order and with its punctuation, because what reads
2735 // them is a script written against that shape. There is no installation directory to
2736 // report: this compiler is one binary that works wherever it is copied, and the headers
2737 // it ships are inside it, so `install` is where the binary is and nothing is under it.
2738 Query::SearchDirs => {
2739 let here = std::env::current_exe()
2740 .ok()
2741 .and_then(|p| p.parent().map(std::path::Path::to_path_buf))
2742 .unwrap_or_default();
2743 let list = |dirs: &[PathBuf]| {
2744 dirs.iter().map(|d| d.display().to_string()).collect::<Vec<_>>().join(":")
2745 };
2746 let libraries = link::search_dirs(link, opts.target);
2747 format!(
2748 "install: {}\nprograms: ={}\nlibraries: ={}",
2749 here.display(),
2750 list(&link.prefixes),
2751 list(&libraries)
2752 )
2753 }
2754 // The root the rest of the answers are under, which a build system asks for when it wants
2755 // to find a file itself rather than ask for one by name, and which is the first thing to
2756 // look at when a cross build read a header nobody expected. A native compile has no
2757 // sysroot and the answer is the empty line, which is what GCC prints when it was
2758 // configured without one. `--sysroot` wins over ours because it wins everywhere else.
2759 Query::Sysroot => {
2760 sysroot_root(opts, link).map(|root| root.display().to_string()).unwrap_or_default()
2761 }
2762 // Section 13.5 of `spec/cross-compile/13-distribution.md`: for every input that is not this
2763 // compiler's own code, what it is, where it was got, its hash, its licence and whether it
2764 // was bundled, generated or fetched. What is printed is the manifest the sysroot already
2765 // carries rather than a second format saying the same things, because the three uses 13.5
2766 // gives for this are a licence notice, a reproducibility check and a security audit, and all
2767 // three are somebody else parsing it. One format is one parser to write.
2768 // Read and rendered rather than copied out, so that what comes back is the format this
2769 // build understands. The last newline comes off because whatever prints an answer adds
2770 // one, the way it does for every other query here. Keeping it would put a blank line at
2771 // the end of the one answer that is a file somebody diffs against the file it came from.
2772 Query::SysrootProvenance => match sysroot_manifest(opts, link)? {
2773 Some(manifest) => manifest.render().trim_end_matches('\n').to_string(),
2774 None => String::new(),
2775 },
2776 // Section 13.2 of the same document, which asks for the hash of a cache directory's
2777 // contents in the directory's name. A name cannot carry one, because the path has to be
2778 // computable before anything has been read, by the producer about to write the files and by
2779 // the compiler about to read them, and neither has the contents when it asks. So the number
2780 // is here instead, and it is the sha256 of the record rather than of a walk of the tree,
2781 // which means `sha256sum` over the manifest answers the same thing.
2782 Query::SysrootDigest => match sysroot_manifest(opts, link)? {
2783 Some(manifest) => manifest.digest(),
2784 None => String::new(),
2785 },
2786 Query::FileName(name) => found(name),
2787 // The name GCC gives the library of routines a compiler's output calls that the C
2788 // library does not have. Ours is built in and there is no file, so the answer is the
2789 // name itself, which is what GCC prints when it cannot find one either.
2790 Query::Libgcc => found("libgcc.a"),
2791 // A program rather than a library: the linker and the archiver are the ones a build asks
2792 // about, and this compiler finds them on the path or under `-B` rather than shipping
2793 // them, so the name back is the honest answer unless a `-B` prefix holds one.
2794 Query::ProgName(name) => link
2795 .prefixes
2796 .iter()
2797 .map(|dir| dir.join(name))
2798 .find(|path| path.is_file())
2799 .map_or_else(|| name.clone(), |path| path.display().to_string()),
2800 })
2801}
2802
2803/// The root every sysroot answer is about.
2804///
2805/// One function rather than a copy in each, because the other flags exist to say what is inside the
2806/// tree this one names, and two answers that disagreed about which tree that is would be a
2807/// difference nobody would think to look for. `--sysroot` wins over ours because it wins everywhere
2808/// else.
2809fn sysroot_root(opts: &Options, link: &LinkOptions) -> Option<PathBuf> {
2810 link.sysroot
2811 .clone()
2812 .or_else(|| link::cross_sysroot(opts.target, link).map(|at| at.root().to_path_buf()))
2813}
2814
2815/// The record of the sysroot this command line reads, when there is one to read.
2816///
2817/// [`None`] covers two cases that both print nothing, and they are different things. A compile for
2818/// this machine has no sysroot at all, and a tree somebody laid out themselves and pointed
2819/// `--sysroot` at carries no manifest, so nothing here knows where any of it came from. Saying
2820/// nothing is the only honest answer to either, and a reader can tell it from a manifest with no
2821/// inputs in it because that one still has its header lines.
2822///
2823/// # Errors
2824///
2825/// A manifest this build cannot parse, and anything else that went wrong reading the file. Passing a
2826/// record we could not read on to whoever asked would make their parser the one that finds the
2827/// problem, and every use section 13.5 gives for these two flags is somebody else reading the
2828/// output.
2829fn sysroot_manifest(opts: &Options, link: &LinkOptions) -> Result<Option<Manifest>, CliError> {
2830 let Some(root) = sysroot_root(opts, link) else {
2831 return Ok(None);
2832 };
2833 let path = Sysroot::at(root, opts.target.tuple()).manifest_path();
2834 match std::fs::read_to_string(&path) {
2835 Ok(text) => Manifest::parse(&text)
2836 .map(Some)
2837 .map_err(|why| err(format!("{}: {why}", path.display()))),
2838 Err(why) if why.kind() == std::io::ErrorKind::NotFound => Ok(None),
2839 Err(why) => Err(err(format!("{}: {why}", path.display()))),
2840 }
2841}
2842
2843/// Renders the passes this level will run, in order, with what each one does.
2844///
2845/// The level is the whole of the answer unless a `-f` flag edited it, which is section 9.1 of
2846/// `spec/09-optimizer.md`: a level is a list somebody wrote down rather than something that
2847/// emerges from which flags happen to be set, and this is how that list is read.
2848#[must_use]
2849pub fn print_pipeline(opts: &Options) -> String {
2850 let mut settings = rucc_opt::Options::for_level(opts.opt_level);
2851 settings.toggles.clone_from(&opts.passes);
2852 settings.global_fuel = opts.pass_fuel_global;
2853 for (on, spec) in &opts.pass_gates {
2854 // Every spelling was checked while the arguments were parsed, so there is nothing here
2855 // this can refuse, and a listing is not the place to report it if there were.
2856 let _ = settings.gates.add(*on, spec);
2857 }
2858 rucc_opt::pipeline::print(&settings)
2859}
2860
2861/// Renders the resolved configuration.
2862///
2863/// One `key: value` per line, sorted by nothing in particular but fixed in order, because
2864/// this output is diffed across hosts in CI and a reordering would read as a change.
2865#[must_use]
2866pub fn print_config(opts: &Options) -> String {
2867 let sess = Session::new(opts.clone());
2868 let t = &sess.target;
2869 let mut out = String::new();
2870 let _ = writeln!(out, "version: {VERSION}");
2871 // The three field triple the driver was given rather than the ten field tuple it widens to,
2872 // because this output is what a build system reads to find out what it asked for. The tuple is
2873 // the compiler's model of the machine and this line is a receipt for a command line.
2874 let _ = writeln!(out, "target: {}", opts.target);
2875 let _ = writeln!(out, "arch: {}", opts.target.arch.as_str());
2876 let _ = writeln!(out, "os: {}", opts.target.os.as_str());
2877 let _ = writeln!(out, "env: {}", opts.target.env.as_str());
2878 let _ = writeln!(out, "object-format: {}", t.object_format.as_str());
2879 let _ = writeln!(out, "pointer-width: {}", t.pointer_width);
2880 let _ = writeln!(out, "long-width: {}", t.long_width);
2881 let _ = writeln!(out, "long-double-width: {}", t.long_double_width);
2882 let _ = writeln!(out, "endian: {}", if t.little_endian { "little" } else { "big" });
2883 let _ = writeln!(out, "char-signed: {}", t.char_is_signed);
2884 let _ = writeln!(out, "va-list: {}", t.va_list.map_or("none", |list| list.as_str()));
2885 // The register file as a count per class, which is enough to tell a target whose registers
2886 // are described from one whose are not without printing sixteen names nobody asked for.
2887 let regs: Vec<String> = t
2888 .regs
2889 .classes()
2890 .map(|(class, info)| format!("{} {}", info.name, t.regs.len(class)))
2891 .collect();
2892 let _ = writeln!(
2893 out,
2894 "registers: {}",
2895 if regs.is_empty() { "none".to_string() } else { regs.join(", ") }
2896 );
2897 // What the schedule was chosen with, which is a sentence rather than a name on purpose: two
2898 // runs of a benchmark that disagree are usually two models and not two compilers.
2899 let _ = writeln!(out, "timing-model: {}", t.timing.map_or("none", |timing| timing.model));
2900 let _ = writeln!(out, "opt-level: {}", sess.opts.opt_level);
2901 let _ = writeln!(out, "safety: {}", sess.opts.safety);
2902 let _ = writeln!(out, "emit: {}", sess.opts.emit.as_str());
2903 let _ = writeln!(out, "debug-info: {}", sess.opts.debug_info);
2904 let _ = writeln!(out, "frame-pointer: {}", sess.opts.keeps_frame_pointer());
2905 let _ = writeln!(out, "red-zone: {}", sess.opts.red_zone);
2906 let _ = writeln!(out, "stack-protector: {}", sess.opts.protector);
2907 let _ = writeln!(out, "stack-clash-protection: {}", sess.opts.stack_clash);
2908 let _ = writeln!(out, "cf-protection: {}", sess.opts.control);
2909 let _ = writeln!(out, "patchable-function-entry: {}", sess.opts.patchable);
2910 let _ = writeln!(out, "profile: {}", sess.opts.profile);
2911 let _ = writeln!(out, "profile-hook: {}", sess.opts.hook);
2912 // Last because it is the one key with more than one line under it, and the only one
2913 // whose value is a property of the machine rather than of the command line.
2914 for dir in sess.opts.search.dirs() {
2915 let system = if dir.is_system { " (system)" } else { "" };
2916 let _ = writeln!(out, "include: {}{system}", dir.path.display());
2917 }
2918 out
2919}
2920
2921/// The output name the make target is taken from, which is the `-o` argument or nothing.
2922///
2923/// A run that stops at the preprocessor has not named an object, whatever its `-o` says: under
2924/// `-E` that argument is the preprocessed text and under `-M` it is the rule itself, and neither
2925/// is a file `make` would rebuild by running this rule. GCC agrees and falls back to the source
2926/// name in both, which is why a `-MD -E -o out.i` writes `out.d` holding a rule for `a.o`. From
2927/// `-S` on the argument does name what the rule builds, and it is used as written.
2928fn deps_target_output<'a>(opts: &Options, plan: &'a Plan) -> Option<&'a str> {
2929 if opts.emit == EmitKind::Preprocessed { None } else { plan.output.as_deref() }
2930}
2931
2932/// Writes to a path the command line named rather than one the plan derived, where `-` is
2933/// standard output.
2934fn write_named(path: &str, bytes: &[u8]) -> Result<(), String> {
2935 if path == "-" {
2936 return write_out(&Output::Stdout, bytes);
2937 }
2938 write_out(&Output::File(path.to_owned()), bytes)
2939}
2940
2941/// Writes the make rule for one input, and reports whether it got there.
2942///
2943/// A rule with no file of its own goes where the compilation it replaced would have written,
2944/// which is what makes the usual makefile recipe work: `rucc -M $< -o $@` leaves the rule in
2945/// `$@`, and the same line with the `-o` left off puts it on standard output.
2946fn write_deps(
2947 opts: &Options,
2948 plan: &Plan,
2949 job: &Job,
2950 found: &[Dependency],
2951 stderr: &mut impl std::io::Write,
2952) -> bool {
2953 let targets = if opts.deps.targets.is_empty() {
2954 vec![deps::default_target(&job.input, deps_target_output(opts, plan))]
2955 } else {
2956 opts.deps.targets.clone()
2957 };
2958 let rule = deps::rule(&opts.deps, &targets, &job.input, found);
2959 // The file, on the other hand, is named after the `-o` in every mode that still has one to
2960 // spend, which is every mode except the two that spend it on the rule.
2961 let wrote = match deps::default_file(&opts.deps, &job.input, plan.output.as_deref()) {
2962 // A `-MF` on a run that had nowhere else to put the rule leaves the file the `-o`
2963 // named empty rather than absent, because a makefile that named it as a target of its
2964 // own is a makefile that will look for it.
2965 Some(path) => write_named(&path, rule.as_bytes()).and_then(|()| {
2966 if opts.deps.instead_of_compiling { write_out(&job.output, b"") } else { Ok(()) }
2967 }),
2968 None => write_out(&job.output, rule.as_bytes()),
2969 };
2970 if let Err(e) = wrote {
2971 let _ = writeln!(stderr, "rucc: error: {e}");
2972 return false;
2973 }
2974 true
2975}
2976
2977/// Runs phase 4 over every input that has one, and writes what came out.
2978///
2979/// One input that fails does not stop the others. A build that reports every file it could
2980/// not preprocess in one run is worth more than one that stops at the first, and the exit
2981/// status is still a failure either way.
2982fn preprocess_all(opts: &Options, plan: &Plan) -> i32 {
2983 let fs = OsFileSystem::new();
2984 let mut stderr = std::io::stderr().lock();
2985 let mut failed = false;
2986 for job in &plan.jobs {
2987 if !job.phases.first().is_some_and(|p| *p == Phase::Preprocess) {
2988 // An input that is already preprocessed, or an object file. GCC passes these
2989 // through untouched, and the plan has already said so in its notes.
2990 continue;
2991 }
2992 let started = std::time::Instant::now();
2993 let result = preprocess(opts, &job.input, &fs);
2994 if opts.time {
2995 say_time(&job.input, started.elapsed(), &mut stderr);
2996 }
2997 for message in &result.messages {
2998 let _ = writeln!(stderr, "{message}");
2999 }
3000 if result.failed() {
3001 failed = true;
3002 continue;
3003 }
3004 if opts.deps.emit {
3005 failed |= !write_deps(opts, plan, job, &result.deps, &mut stderr);
3006 // `-M` and `-MM` asked for the rule instead of the text, so there is nothing else
3007 // to write. The other two asked for both and fall through to the text below.
3008 if opts.deps.instead_of_compiling {
3009 continue;
3010 }
3011 }
3012 if let Err(e) = write_out(&job.output, result.text.as_bytes()) {
3013 let _ = writeln!(stderr, "rucc: error: {e}");
3014 failed = true;
3015 }
3016 }
3017 i32::from(failed)
3018}
3019
3020/// Whether this job is a file of assembly that has to be assembled and that nothing here assembles.
3021///
3022/// The phases rather than the kind, because there are two kinds of assembly input and one of them
3023/// is preprocessed first, and because an object file also has no compile phase and is not this: it
3024/// has no phases at all and goes to the linker as it is. A `.s` on a `-c` line has exactly
3025/// [`Phase::Assemble`] left, and a `.S` has the preprocessor in front of it, and neither has
3026/// anything the front end can do.
3027fn needs_an_assembler(job: &Job) -> bool {
3028 job.phases.contains(&Phase::Assemble) && !job.phases.contains(&Phase::Compile)
3029}
3030
3031/// Whether the preprocessor runs over it on the way in, which is the whole difference between the
3032/// two kinds of assembly input.
3033fn assembly_wants_cpp(job: &Job) -> bool {
3034 job.phases.contains(&Phase::Preprocess)
3035}
3036
3037/// Runs the front end over every input that has a compile phase, and writes what came out.
3038///
3039/// The same rule as [`preprocess_all`]: one input that fails does not stop the others, and the
3040/// exit status is a failure either way. An input that is already assembly or an object has no
3041/// compile phase and is passed over here, which the plan has already said in its notes.
3042fn compile_all(opts: &Options, plan: &Plan) -> i32 {
3043 let fs = OsFileSystem::new();
3044 let mut stderr = std::io::stderr().lock();
3045 let mut failed = false;
3046 let (mut remarks, ok) = Remarks::new(opts.opt_info_file.as_ref(), &mut stderr);
3047 failed |= !ok;
3048 let mut fired = Fired::new();
3049 let mut pressure = Pressure::new();
3050 let mut lowerings = Lowerings::new();
3051 for job in &plan.jobs {
3052 if !job.phases.contains(&Phase::Compile) && !needs_an_assembler(job) {
3053 continue;
3054 }
3055 // An input of IR is read back rather than compiled, since the C it came from is not
3056 // here any more. A file of assembly does not go through the front end at all and is
3057 // read by the assembler instead. Everything after this is the same for all three, so
3058 // the paths meet again at the messages and the file the result is written to.
3059 let started = std::time::Instant::now();
3060 let result = if needs_an_assembler(job) {
3061 assemble(opts, &job.input, assembly_wants_cpp(job), &fs)
3062 } else if job.kind == InputKind::Ir {
3063 compile_ir(opts, &job.input, &fs)
3064 } else {
3065 compile(opts, &job.input, &fs)
3066 };
3067 if opts.time {
3068 say_time(&job.input, started.elapsed(), &mut stderr);
3069 }
3070 failed |= !write_trace(opts, job, started, &result, &mut stderr);
3071 fired.merge(&result.fired);
3072 pressure.merge(&result.pressure);
3073 lowerings.merge(&result.lowerings);
3074 failed |= !write_dumps(&job.input, &result.dumps, &mut stderr);
3075 failed |= !remarks.write(&result.remarks, &mut stderr);
3076 for message in &result.messages {
3077 let _ = writeln!(stderr, "{message}");
3078 }
3079 // Before the failure below, because a compilation that stopped in the back end is exactly
3080 // the one whose preprocessed source somebody wants to look at.
3081 failed |= !write_temps(job, &result.temps, &mut stderr);
3082 // Before it as well, because gcc leaves an empty report for a file that did not compile
3083 // and a build that looks for one beside every object should find one.
3084 failed |= !write_stack_usage(job, &result.stack_usage, &mut stderr);
3085 if result.failed() {
3086 failed = true;
3087 continue;
3088 }
3089 // `-MD` and `-MMD` write the rule beside the object and let the compilation happen, so
3090 // this is the one path where both files come out of the same run. An input of IR has no
3091 // dependencies to report and produces an empty list, which produces a rule naming only
3092 // itself, and that is the honest answer rather than a missing file.
3093 if opts.deps.emit {
3094 failed |= !write_deps(opts, plan, job, &result.deps, &mut stderr);
3095 }
3096 if let Err(e) = write_out(&job.output, result.artifact.bytes()) {
3097 let _ = writeln!(stderr, "rucc: error: {e}");
3098 failed = true;
3099 }
3100 }
3101 failed |= !write_coverage(opts, &fired, &mut stderr);
3102 failed |= !write_pressure(opts, &pressure, &mut stderr);
3103 failed |= !write_lowering(opts, &lowerings, &mut stderr);
3104 i32::from(failed)
3105}
3106
3107/// A directory for the object files only the link step ever sees, removed when it goes away.
3108///
3109/// `-c` writes its object where the user can see it and linking does not, which is the whole of
3110/// the difference: a `rucc a.c b.c` leaves an executable behind and nothing else, the same as
3111/// every other compiler. Removing them on drop rather than at the end of a function is so that a
3112/// link that failed leaves nothing behind either.
3113struct Scratch {
3114 /// Where the objects go.
3115 dir: PathBuf,
3116}
3117
3118impl Scratch {
3119 /// Makes one, under whatever the platform calls its temporary directory.
3120 ///
3121 /// The name carries the process id so that two compilers running at once do not share a
3122 /// directory, which they would otherwise do the moment two of them compiled a file of the
3123 /// same name.
3124 fn new() -> Result<Scratch, String> {
3125 let dir = std::env::temp_dir().join(format!("rucc-{}", std::process::id()));
3126 std::fs::create_dir_all(&dir).map_err(|e| format!("{}: {e}", dir.display()))?;
3127 Ok(Scratch { dir })
3128 }
3129}
3130
3131impl Drop for Scratch {
3132 fn drop(&mut self) {
3133 let _ = std::fs::remove_dir_all(&self.dir);
3134 }
3135}
3136
3137/// The link line the plan describes, for `-###`.
3138///
3139/// The names in it are the hints the plan carries rather than the temporaries a real compilation
3140/// would choose, because `-###` prints the line without having compiled anything and so has
3141/// nothing to point at. That also makes the printed line readable rather than naming a directory
3142/// that only exists while a compilation is running.
3143fn link_line(opts: &Options, link: &LinkOptions, job: &LinkJob) -> Result<String, link::Error> {
3144 let linker = link::find(opts.target, link)?;
3145 let args = link::line(opts.target, link, &job.inputs, &job.output)?;
3146 Ok(link::render(&linker, &args))
3147}
3148
3149/// Compiles everything, then links it.
3150///
3151/// The objects go in a directory that is removed afterwards, which is why this is not
3152/// [`compile_all`] followed by a link: the plan says an object feeding the linker is temporary
3153/// and does not say where, because where is a question that only has an answer once something is
3154/// running.
3155fn link_all(opts: &Options, plan: &Plan, link: &LinkOptions, verbose: bool) -> i32 {
3156 let Some(job) = &plan.link else {
3157 // Every path into here comes from a plan whose last phase is the link, and such a plan
3158 // has a link job. Saying so is cheaper than an unwrap that would have to be explained.
3159 let mut stderr = std::io::stderr().lock();
3160 let _ = writeln!(stderr, "rucc: error: there is nothing to link");
3161 return 1;
3162 };
3163 // Before anything is compiled, because a linker that is not on the machine is worth knowing
3164 // about in the second it takes to look rather than after the compilation.
3165 // And before that, whether this link has a line at all and whether what it reads is on the
3166 // machine. Both are answerable now, and a target whose sysroot has not been built is worth
3167 // saying so about before the compilation rather than after it.
3168 if let Err(why) = link::preflight(opts.target, link) {
3169 return complain(why);
3170 }
3171 let linker = match link::find(opts.target, link) {
3172 Ok(linker) => linker,
3173 Err(why) => return complain(why),
3174 };
3175 // Whether the one that was found can do this link is asked inside the search, which moves on
3176 // past an lld that is too old to a newer one somewhere else and refuses only when there is none.
3177 // The glibc stubs, which are the one part of a cross sysroot written here rather than fetched.
3178 // Before compiling for the same reason as the rest, and never for `-###`, which writes nothing.
3179 if let Err(why) = link::write_stubs(opts.target, link) {
3180 return complain(why);
3181 }
3182
3183 let scratch = match Scratch::new() {
3184 Ok(scratch) => scratch,
3185 Err(why) => return complain(format!("could not make a place for the object files: {why}")),
3186 };
3187
3188 let fs = OsFileSystem::new();
3189 let mut failed = false;
3190 // One per job, in job order, which is what lets the link line below be rebuilt with the real
3191 // paths in it: every job contributes exactly one file to the line and does so in this order.
3192 let mut produced: Vec<String> = Vec::with_capacity(plan.jobs.len());
3193 let mut fired = Fired::new();
3194 let mut pressure = Pressure::new();
3195 let mut lowerings = Lowerings::new();
3196 {
3197 let mut stderr = std::io::stderr().lock();
3198 let (mut remarks, ok) = Remarks::new(opts.opt_info_file.as_ref(), &mut stderr);
3199 failed |= !ok;
3200 for (at, job) in plan.jobs.iter().enumerate() {
3201 let out = match &job.output {
3202 Output::Temporary(hint) => {
3203 // The index because two inputs in different directories can have the same
3204 // name, and the two objects of `rucc a/x.c b/x.c` must not be one file.
3205 scratch.dir.join(format!("{at}-{hint}")).display().to_string()
3206 }
3207 Output::File(path) => path.clone(),
3208 // A job feeding the linker never writes to standard output, since the plan gives
3209 // it a temporary. This is here so that the match is total rather than a panic.
3210 Output::Stdout => continue,
3211 };
3212 produced.push(out.clone());
3213 if !job.phases.contains(&Phase::Compile) && !needs_an_assembler(job) {
3214 continue;
3215 }
3216 let started = std::time::Instant::now();
3217 let result = if needs_an_assembler(job) {
3218 assemble(opts, &job.input, assembly_wants_cpp(job), &fs)
3219 } else if job.kind == InputKind::Ir {
3220 compile_ir(opts, &job.input, &fs)
3221 } else {
3222 compile(opts, &job.input, &fs)
3223 };
3224 if opts.time {
3225 say_time(&job.input, started.elapsed(), &mut stderr);
3226 }
3227 failed |= !write_trace(opts, job, started, &result, &mut stderr);
3228 fired.merge(&result.fired);
3229 pressure.merge(&result.pressure);
3230 lowerings.merge(&result.lowerings);
3231 failed |= !write_dumps(&job.input, &result.dumps, &mut stderr);
3232 failed |= !remarks.write(&result.remarks, &mut stderr);
3233 for message in &result.messages {
3234 let _ = writeln!(stderr, "{message}");
3235 }
3236 failed |= !write_temps(job, &result.temps, &mut stderr);
3237 failed |= !write_stack_usage(job, &result.stack_usage, &mut stderr);
3238 if result.failed() {
3239 failed = true;
3240 continue;
3241 }
3242 // A `-MD` on a command line that links writes the rule next to the executable and
3243 // names the executable as its target, since that is the file this source builds
3244 // here. The object it went through is in a temporary directory and is gone by the
3245 // time `make` reads any of this.
3246 if opts.deps.emit {
3247 failed |= !write_deps(opts, plan, job, &result.deps, &mut stderr);
3248 }
3249 if !matches!(result.artifact, Artifact::Object { .. }) {
3250 // Worth saying rather than writing whatever it is and letting the linker read it.
3251 // An empty file is a valid empty linker script, so a link handed one gets as far
3252 // as reporting every symbol of this file undefined, which is a page of messages
3253 // about something that went wrong here.
3254 let _ = writeln!(
3255 stderr,
3256 "rucc: internal error: {}: no object file was produced for the link",
3257 job.input
3258 );
3259 failed = true;
3260 continue;
3261 }
3262 if let Err(e) = std::fs::write(&out, result.artifact.bytes()) {
3263 let _ = writeln!(stderr, "rucc: error: {out}: {e}");
3264 failed = true;
3265 }
3266 }
3267 failed |= !write_coverage(opts, &fired, &mut stderr);
3268 failed |= !write_pressure(opts, &pressure, &mut stderr);
3269 failed |= !write_lowering(opts, &lowerings, &mut stderr);
3270 failed |= !write_lowering(opts, &lowerings, &mut stderr);
3271 }
3272 if failed {
3273 // Nothing is linked from a compilation that did not finish. A linker run over the objects
3274 // that did compile would report every function of the file that did not as undefined,
3275 // which is a page of messages about a mistake already reported once.
3276 return 1;
3277 }
3278
3279 // The items in command line order with the temporaries filled in. A library and a word for the
3280 // linker contribute no job and pass through, and every file item takes the next job's real
3281 // output, which is what keeps whatever was written between two objects between them here.
3282 let mut outputs = produced.into_iter();
3283 let mut items = Vec::with_capacity(job.inputs.len());
3284 for item in &job.inputs {
3285 match item {
3286 link::Item::Library(name) => items.push(link::Item::Library(name.clone())),
3287 link::Item::Linker(arg) => items.push(link::Item::Linker(arg.clone())),
3288 link::Item::File(_) => match outputs.next() {
3289 Some(path) => items.push(link::Item::File(path)),
3290 None => return complain("the plan asks the linker for a file nothing produced"),
3291 },
3292 }
3293 }
3294
3295 let args = match link::line(opts.target, link, &items, &job.output) {
3296 Ok(args) => args,
3297 Err(why) => return complain(why),
3298 };
3299 if verbose {
3300 let mut stderr = std::io::stderr().lock();
3301 let _ = writeln!(stderr, "{}", link::render(&linker, &args));
3302 }
3303 let started = std::time::Instant::now();
3304 let ran = link::run(&linker, &args);
3305 if opts.time {
3306 // The one step of a compilation that really is another program, so this line is the same
3307 // measurement gcc's is and names the linker the way gcc names `collect2`.
3308 let mut stderr = std::io::stderr().lock();
3309 say_time(&linker.name, started.elapsed(), &mut stderr);
3310 }
3311 match ran {
3312 Ok(()) => 0,
3313 // The linker has already said what was wrong on its own error output, and repeating that
3314 // linking failed would only push its message further up the screen.
3315 Err(link::Error::Refused { .. }) => 1,
3316 Err(why) => complain(why),
3317 }
3318}
3319
3320/// Compiles everything and writes the objects into one static library.
3321///
3322/// No temporary directory and no second program. The objects never reach the file system at all:
3323/// they go from the compiler into the archive writer, which is both faster than writing a directory
3324/// of files for an `ar` to read back and the reason the symbol index can be written at all. A
3325/// member's index entries are the names the object writer says it wrote, and the only thing that
3326/// knows those is the run that wrote it.
3327///
3328/// `-save-temps` is the exception. It asked for the objects to be kept, the plan gave them names a
3329/// person can find, and they are written there as well as put in the archive.
3330fn archive_all(opts: &Options, plan: &Plan) -> i32 {
3331 let Some(job) = &plan.archive else {
3332 // Every path into here comes from a plan whose last phase is the archive, and such a plan
3333 // has an archive job. Saying so is cheaper than an unwrap that would have to be explained.
3334 return complain("there is nothing to put in an archive");
3335 };
3336 // Before anything is compiled, because a format this has no container for is worth knowing
3337 // about in the second it takes to look rather than after the whole compilation.
3338 let flavour = match opts.target.os.object_format() {
3339 ObjectFormat::Elf => rucc_archive::Flavour::Gnu,
3340 ObjectFormat::Coff => rucc_archive::Flavour::Coff,
3341 ObjectFormat::MachO => rucc_archive::Flavour::Bsd,
3342 // Wasm has no archives of its own at all.
3343 format @ ObjectFormat::Wasm => {
3344 return complain(format!(
3345 "there is no archive format for {} objects in this compiler yet",
3346 format.as_str()
3347 ));
3348 }
3349 };
3350
3351 let fs = OsFileSystem::new();
3352 let mut failed = false;
3353 let mut members: Vec<rucc_archive::Member> = Vec::with_capacity(plan.jobs.len());
3354 let mut names = job.members.iter();
3355 let mut fired = Fired::new();
3356 let mut pressure = Pressure::new();
3357 let mut lowerings = Lowerings::new();
3358 {
3359 let mut stderr = std::io::stderr().lock();
3360 let (mut remarks, ok) = Remarks::new(opts.opt_info_file.as_ref(), &mut stderr);
3361 failed |= !ok;
3362 for plan_job in &plan.jobs {
3363 // What the plan called this member. The two lists are walked together rather than the
3364 // name being worked out again here, so that what `-###` printed and what goes in the
3365 // file cannot come apart.
3366 let Some(member) = names.next() else {
3367 return complain("the plan asks the archive for a member nothing produced");
3368 };
3369 if !plan_job.phases.contains(&Phase::Compile) && !needs_an_assembler(plan_job) {
3370 // Neither something to compile nor something to assemble, so there is nothing to
3371 // put in, and an archive quietly missing a member is worse than a message.
3372 let _ = writeln!(
3373 &mut stderr,
3374 "rucc: error: {}: this compiler makes an archive out of what it compiles, and \
3375 there is nothing here for it to do",
3376 plan_job.input
3377 );
3378 failed = true;
3379 continue;
3380 }
3381 let started = std::time::Instant::now();
3382 let result = if needs_an_assembler(plan_job) {
3383 assemble(opts, &plan_job.input, assembly_wants_cpp(plan_job), &fs)
3384 } else if plan_job.kind == InputKind::Ir {
3385 compile_ir(opts, &plan_job.input, &fs)
3386 } else {
3387 compile(opts, &plan_job.input, &fs)
3388 };
3389 if opts.time {
3390 say_time(&plan_job.input, started.elapsed(), &mut stderr);
3391 }
3392 failed |= !write_trace(opts, plan_job, started, &result, &mut stderr);
3393 fired.merge(&result.fired);
3394 pressure.merge(&result.pressure);
3395 lowerings.merge(&result.lowerings);
3396 failed |= !write_dumps(&plan_job.input, &result.dumps, &mut stderr);
3397 failed |= !remarks.write(&result.remarks, &mut stderr);
3398 for message in &result.messages {
3399 let _ = writeln!(stderr, "{message}");
3400 }
3401 failed |= !write_temps(plan_job, &result.temps, &mut stderr);
3402 failed |= !write_stack_usage(plan_job, &result.stack_usage, &mut stderr);
3403 if result.failed() {
3404 failed = true;
3405 continue;
3406 }
3407 if opts.deps.emit {
3408 failed |= !write_deps(opts, plan, plan_job, &result.deps, &mut stderr);
3409 }
3410 let Artifact::Object { bytes, defines } = result.artifact else {
3411 let _ = writeln!(
3412 stderr,
3413 "rucc: internal error: {}: no object file was produced for the archive",
3414 plan_job.input
3415 );
3416 failed = true;
3417 continue;
3418 };
3419 // Under `-save-temps` the plan gave the object a name a person can find, so it is
3420 // written there too. Otherwise it is only ever a member and never a file.
3421 if let Output::File(path) = &plan_job.output {
3422 if let Err(e) = std::fs::write(path, &bytes) {
3423 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
3424 failed = true;
3425 }
3426 }
3427 members.push(rucc_archive::Member { name: member.clone(), body: bytes, defines });
3428 }
3429 failed |= !write_coverage(opts, &fired, &mut stderr);
3430 failed |= !write_pressure(opts, &pressure, &mut stderr);
3431 failed |= !write_lowering(opts, &lowerings, &mut stderr);
3432 failed |= !write_lowering(opts, &lowerings, &mut stderr);
3433 }
3434 if failed {
3435 // Nothing is written from a compilation that did not finish, for the reason the link gives:
3436 // an archive missing the file that failed is one a link reports every name of as undefined,
3437 // which is a page of messages about a mistake already reported once.
3438 return 1;
3439 }
3440
3441 let bytes = match rucc_archive::write(flavour, &members) {
3442 Ok(bytes) => bytes,
3443 // Every one of these is a bug here rather than a program's mistake: the names came from the
3444 // object writer and the bodies came from this process.
3445 Err(why) => return complain(format!("the archive could not be written: {why}")),
3446 };
3447 match std::fs::write(&job.output, &bytes) {
3448 Ok(()) => 0,
3449 Err(e) => complain(format!("{}: {e}", job.output)),
3450 }
3451}
3452
3453/// Prints one driver level message and gives back the exit status that goes with it.
3454fn complain(why: impl std::fmt::Display) -> i32 {
3455 let mut stderr = std::io::stderr().lock();
3456 let _ = writeln!(stderr, "rucc: error: {why}");
3457 1
3458}
3459
3460/// Writes what `-Zrule-coverage=FILE` asked for, and says whether it could.
3461///
3462/// Once for the whole command line rather than once per input, because the question is which
3463/// lowering rules this run of the compiler reached and a file per input would leave the reader
3464/// unioning files to find out something one process already knew.
3465///
3466/// A file that could not be written is a failure and not a warning. What asks for this is a
3467/// measurement run, and a measurement that quietly did not happen is worse than one that stopped.
3468fn write_coverage(opts: &Options, fired: &Fired, stderr: &mut impl std::io::Write) -> bool {
3469 let Some(path) = &opts.rule_coverage else { return true };
3470 let Some(table) = coverage::table(opts.target.arch) else {
3471 let _ = writeln!(
3472 stderr,
3473 "rucc: error: there are no lowering rules for {} yet, so there is no coverage of them \
3474 to report",
3475 opts.target
3476 );
3477 return false;
3478 };
3479 match std::fs::write(path, fired.listing(table)) {
3480 Ok(()) => true,
3481 Err(e) => {
3482 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
3483 false
3484 }
3485 }
3486}
3487
3488/// Writes what `-Zregister-pressure=FILE` asked for, and says whether it could.
3489///
3490/// Once for the whole command line, for the reason [`write_coverage`] gives, and a file that could
3491/// not be written is a failure for the reason it gives too. There is no equivalent of the missing
3492/// rule table here, since every target this compiles for has an allocator, and a run that reached
3493/// no back end at all writes an empty listing rather than nothing: a measurement of a build that
3494/// produced no code is still an answer and it is the honest one.
3495fn write_pressure(opts: &Options, pressure: &Pressure, stderr: &mut impl std::io::Write) -> bool {
3496 let Some(path) = &opts.register_pressure else { return true };
3497 match std::fs::write(path, pressure.listing()) {
3498 Ok(()) => true,
3499 Err(e) => {
3500 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
3501 false
3502 }
3503 }
3504}
3505
3506/// Writes what `-Zlowering=FILE` asked for, and says whether it could.
3507///
3508/// Once for the whole command line, for the reason [`write_coverage`] gives, and a file that could
3509/// not be written is a failure for the reason it gives too. A run that reached no back end writes
3510/// an empty listing rather than nothing, the way [`write_pressure`] does and for the same reason.
3511fn write_lowering(opts: &Options, lowerings: &Lowerings, stderr: &mut impl std::io::Write) -> bool {
3512 let Some(path) = &opts.lowering_dump else { return true };
3513 match std::fs::write(path, lowerings.listing()) {
3514 Ok(()) => true,
3515 Err(e) => {
3516 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
3517 false
3518 }
3519 }
3520}
3521
3522/// Where the `-fopt-info` remarks go, and how much of the run has already gone there.
3523///
3524/// Standard error by default, and one file for the whole run when `-fopt-info=<file>` named one.
3525/// A file rather than the diagnostic stream is what a harness wants: the corpus in
3526/// `tamnd/rucc-corpus` matches a rejection against what the compiler said on standard error, and
3527/// a few thousand remarks mixed into that would bury it.
3528struct Remarks {
3529 /// The file, if there is one.
3530 file: Option<String>,
3531 /// Whether anything has been written to it yet, which decides between truncating and
3532 /// appending. One file holds the whole run rather than the last input in it.
3533 started: bool,
3534}
3535
3536impl Remarks {
3537 /// Prepares the destination, emptying the file if there is one.
3538 ///
3539 /// Emptied here rather than at the first remark, because a run where no pass had anything to
3540 /// say should leave an empty file and not yesterday's. An absent file and an empty one are
3541 /// different facts and something reading this will act on the difference.
3542 fn new(file: Option<&String>, stderr: &mut impl std::io::Write) -> (Self, bool) {
3543 let mut ok = true;
3544 if let Some(path) = file {
3545 if let Err(e) = std::fs::write(path, "") {
3546 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
3547 ok = false;
3548 }
3549 }
3550 (Self { file: file.cloned(), started: false }, ok)
3551 }
3552
3553 /// Writes one input's remarks, and says whether that worked.
3554 ///
3555 /// A file that cannot be written is a failure and not a warning, for the reason
3556 /// [`write_dumps`] gives: remarks that quietly did not arrive look exactly like a compilation
3557 /// where nothing happened.
3558 fn write(&mut self, text: &str, stderr: &mut impl std::io::Write) -> bool {
3559 if text.is_empty() {
3560 return true;
3561 }
3562 let Some(path) = &self.file else {
3563 let _ = write!(stderr, "{text}");
3564 return true;
3565 };
3566 let opened = std::fs::OpenOptions::new()
3567 .write(true)
3568 .append(self.started)
3569 .truncate(!self.started)
3570 .create(true)
3571 .open(path);
3572 self.started = true;
3573 let result =
3574 opened.and_then(|mut file| std::io::Write::write_all(&mut file, text.as_bytes()));
3575 if let Err(e) = result {
3576 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
3577 return false;
3578 }
3579 true
3580 }
3581}
3582
3583/// Writes what `-fdump-ir=` asked to see, one file per dump.
3584///
3585/// The name is the input file with the dump's own name and `.ir` after it, so a directory listing
3586/// after a run is the passes in the order they ran, per input. They go in the working directory
3587/// rather than beside the output, because a dump is something a person asked for at a prompt and
3588/// the working directory is where that person is.
3589///
3590/// A file that could not be written is a failure and not a warning, for the reason
3591/// [`write_coverage`] gives: what asked for this is somebody debugging a pass, and a dump that
3592/// quietly did not happen looks exactly like a pass that did not run.
3593fn write_dumps(input: &str, dumps: &[rucc_opt::Dump], stderr: &mut impl std::io::Write) -> bool {
3594 let stem = std::path::Path::new(input)
3595 .file_name()
3596 .map_or_else(|| input.to_owned(), |name| name.to_string_lossy().into_owned());
3597 let mut ok = true;
3598 for dump in dumps {
3599 let path = format!("{stem}.{}.ir", dump.name);
3600 if let Err(e) = std::fs::write(&path, &dump.text) {
3601 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
3602 ok = false;
3603 }
3604 }
3605 ok
3606}
3607
3608/// Writes the files `-save-temps` kept, which is nothing at all unless it was given.
3609///
3610/// A file that could not be written is a failure rather than a warning, for the reason
3611/// [`write_dumps`] gives: somebody asked for these by name, and one that quietly did not happen
3612/// looks like a compilation that never went through that step.
3613fn write_temps(job: &Job, temps: &Temps, stderr: &mut impl std::io::Write) -> bool {
3614 let mut ok = true;
3615 let kept = [(job.saved_text(), &temps.preprocessed), (job.saved_asm(), &temps.assembly)];
3616 for (path, text) in kept {
3617 // A step the compilation did not reach has nothing to keep, and a job that is not keeping
3618 // that step has nowhere to put it. Either way there is no file here.
3619 let (Some(path), Some(text)) = (path, text) else { continue };
3620 if let Err(e) = std::fs::write(&path, text) {
3621 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
3622 ok = false;
3623 }
3624 }
3625 ok
3626}
3627
3628/// Writes the `.su` file `-fstack-usage` asked for, where the plan said it goes.
3629///
3630/// Written even when it is empty, because gcc writes an empty `.su` for a file with no functions,
3631/// for `-fsyntax-only` and for a file that did not compile, and a tool that looks for one beside
3632/// every object should find one.
3633fn write_stack_usage(job: &Job, text: &str, stderr: &mut impl std::io::Write) -> bool {
3634 let Some(path) = &job.stack_usage else { return true };
3635 if let Err(e) = std::fs::write(path, text) {
3636 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
3637 return false;
3638 }
3639 true
3640}
3641
3642/// Appends the file's line to the `-frucc-trace` file, when there is one.
3643///
3644/// Returns whether that went well, and says why on standard error when it did not.
3645fn write_trace(
3646 opts: &Options,
3647 job: &Job,
3648 started: std::time::Instant,
3649 result: &Compiled,
3650 stderr: &mut impl std::io::Write,
3651) -> bool {
3652 let Some(path) = &opts.trace else {
3653 return true;
3654 };
3655 let output = match &job.output {
3656 Output::Stdout => "-",
3657 Output::File(path) | Output::Temporary(path) => path,
3658 };
3659 let record = trace::Record {
3660 input: &job.input,
3661 output,
3662 ok: !result.failed(),
3663 total: started.elapsed(),
3664 timing: &result.timing,
3665 };
3666 match trace::append(path, &record) {
3667 Ok(()) => true,
3668 Err(e) => {
3669 let _ = writeln!(stderr, "rucc: error: {e}");
3670 false
3671 }
3672 }
3673}
3674
3675/// One line of `-time`, which is what a step was called and how long it took.
3676///
3677/// GCC's two numbers are the user and the system time of a subprocess it ran. This compiler runs
3678/// no subprocess for anything but the link, so what is measured here is the wall clock of the
3679/// step and the second column is always zero. The shape of the line is kept because a person
3680/// reading it next to gcc's should not have to work out which column is which.
3681fn say_time(name: &str, took: std::time::Duration, stderr: &mut impl std::io::Write) {
3682 let _ = writeln!(stderr, "# {name} {:.2} {:.2}", took.as_secs_f64(), 0.0);
3683}
3684
3685/// Writes one job's result where the plan said it goes.
3686///
3687/// # Errors
3688///
3689/// Returns the message to print, which names the file when there is one, because "permission
3690/// denied" on its own does not say which file was refused.
3691fn write_out(output: &Output, bytes: &[u8]) -> Result<(), String> {
3692 match output {
3693 Output::Stdout => {
3694 let mut stdout = std::io::stdout().lock();
3695 stdout.write_all(bytes).map_err(|e| format!("writing to standard output: {e}"))
3696 }
3697 Output::File(path) | Output::Temporary(path) => {
3698 std::fs::write(path, bytes).map_err(|e| format!("{path}: {e}"))
3699 }
3700 }
3701}
3702
3703/// The target a program name asks for, the way `aarch64-linux-gnu-gcc` is gcc for that target.
3704///
3705/// `program` is the path the compiler was started as. The name without its directory and without a
3706/// trailing `.exe` has to end in `-rucc`, and what comes before that has to be a target this
3707/// compiler knows, or there is no answer and the name means nothing. A link named `my-rucc` is
3708/// therefore just rucc and not an error.
3709pub fn target_from_program(program: &str) -> Option<String> {
3710 let name = program.rsplit(['/', '\\']).next()?;
3711 let name = name.strip_suffix(".exe").or_else(|| name.strip_suffix(".EXE")).unwrap_or(name);
3712 let triple = name.strip_suffix("-rucc")?;
3713 triple.parse::<Triple>().ok()?;
3714 Some(triple.to_owned())
3715}
3716
3717/// [`run`] for a compiler started as `program`, which is `argv[0]`.
3718///
3719/// A target taken from the name goes in front of `args`, so a `--target=` written on the command
3720/// line comes later and wins, which is what gcc and clang do with a prefixed name.
3721pub fn run_as(program: &str, args: &[String]) -> i32 {
3722 match target_from_program(program) {
3723 Some(triple) => {
3724 let mut all = Vec::with_capacity(args.len() + 1);
3725 all.push(format!("--target={triple}"));
3726 all.extend_from_slice(args);
3727 run(&all)
3728 }
3729 None => run(args),
3730 }
3731}
3732
3733/// What `--version` prints.
3734///
3735/// The first line is ours and is the one every harness we have reads. The second is for build
3736/// systems that decide what kind of compiler they have by reading this text. Meson takes the GNU
3737/// path only when it finds "Free Software Foundation" here, and otherwise stops with "Unknown
3738/// compiler" before it has asked a single question, which is how the whole of a meson build is
3739/// lost to one sentence. Past that point meson reads the version from `__GNUC__` and asks the
3740/// preprocessor everything else, so the line decides the path and nothing more. It says what is
3741/// true, that rucc speaks the dialect of GCC 16, and it does not claim to be GCC.
3742fn banner() -> String {
3743 format!(
3744 "rucc {VERSION}\nA C compiler for the GNU C dialect of GCC 16 from the Free Software Foundation.\nThis is free software under the Apache License 2.0. There is NO warranty.\n"
3745 )
3746}
3747
3748/// Runs the driver and returns the process exit code.
3749///
3750/// `args` excludes the program name. Output goes to `stdout` and errors to `stderr`, which
3751/// is the one place in the compiler that is true.
3752pub fn run(args: &[String]) -> i32 {
3753 match parse_args(args) {
3754 Ok(Action::Help) => {
3755 print!("{USAGE}");
3756 0
3757 }
3758 Ok(Action::Version) => {
3759 print!("{}", banner());
3760 0
3761 }
3762 Ok(Action::Print(line)) => {
3763 println!("{line}");
3764 0
3765 }
3766 Ok(Action::PrintConfig(opts)) => {
3767 print!("{}", print_config(&opts));
3768 0
3769 }
3770 Ok(Action::PrintPipeline(opts)) => {
3771 print!("{}", print_pipeline(&opts));
3772 0
3773 }
3774 Ok(Action::PrintPlan { opts, plan, link }) => {
3775 print!("{}", plan.render());
3776 // The line as it would be typed, which is the half of `-###` that section 4.3 says
3777 // arrives with the link. It is printed even when the linker is not on this machine,
3778 // because what a build wants from `-###` is what the compiler would do.
3779 if let Some(job) = &plan.link {
3780 match link_line(&opts, &link, job) {
3781 Ok(line) => println!("{line}"),
3782 Err(why) => {
3783 let mut stderr = std::io::stderr().lock();
3784 let _ = writeln!(stderr, "rucc: error: {why}");
3785 return 1;
3786 }
3787 }
3788 }
3789 0
3790 }
3791 Ok(Action::Fetch { what, target, cache }) => {
3792 let kernel = rucc_sysroot::Kernel::for_target(&cache, target)
3793 .map(|_| &rucc_sysroot::KERNEL_HEADERS);
3794 fetch_sysroot(what, kernel, target, &cache)
3795 }
3796 Ok(Action::FetchMsvcSdk { target, accepted, cache, pinned }) => msvc::fetch_msvc_sdk(
3797 target,
3798 accepted,
3799 &cache,
3800 pinned.then_some(&rucc_sysroot::PINNED_BUILD),
3801 ),
3802 Ok(Action::Compile { opts, plan, link, jobs, verbose, notes }) => {
3803 {
3804 let mut stderr = std::io::stderr().lock();
3805 // Before the plan rather than after it, because a note is about the command line
3806 // and the plan is what the command line was read as, so the reader wants the two
3807 // in that order.
3808 for note in ¬es {
3809 let _ = writeln!(stderr, "rucc: warning: {note}");
3810 }
3811 if verbose {
3812 let _ = write!(stderr, "{}", plan.render());
3813 let _ = writeln!(stderr, "workers: {}", jobs.count());
3814 // What `gcc -v` says about headers, because meson and cmake read it to find the
3815 // system directories.
3816 let _ = write!(stderr, "{}", opts.search.render_gcc());
3817 }
3818 }
3819 if opts.emit == EmitKind::Preprocessed {
3820 return preprocess_all(&opts, &plan);
3821 }
3822 if opts.emit == EmitKind::Archive {
3823 return archive_all(&opts, &plan);
3824 }
3825 if opts.emit != EmitKind::Executable {
3826 return compile_all(&opts, &plan);
3827 }
3828 if let Some(why) = unlinkable(&opts) {
3829 let _ = writeln!(std::io::stderr().lock(), "rucc: error: {why}");
3830 return 1;
3831 }
3832 link_all(&opts, &plan, &link, verbose)
3833 }
3834 Err(e) => {
3835 let mut stderr = std::io::stderr().lock();
3836 let _ = writeln!(stderr, "rucc: error: {e}");
3837 let _ = writeln!(stderr, "rucc: note: run `rucc --help` for usage");
3838 1
3839 }
3840 }
3841}
3842
3843/// Why a link that was asked for cannot be made, when that is known before anything is compiled.
3844///
3845/// The checked modes need `runtime/rucc-safe-rt` in the program, and that runtime is written
3846/// against Unix: shadow memory through `mmap`, reports through a signal handler, and the maps read
3847/// out of `/proc`. There is no Windows build of it, so a Windows program compiled with one used to
3848/// fail at the link with a page of undefined `__rucc_check_` names. Saying so before the link
3849/// is kinder until the port is done. Only the link is refused: an object or a listing built
3850/// with the checks in is still what was asked for, and is what a test of the instrumentation reads.
3851fn unlinkable(opts: &Options) -> Option<String> {
3852 (opts.safety.instruments() && opts.target.os == rucc_target::Os::Windows).then(|| {
3853 format!(
3854 "-fsafety={}: the checked modes are not available on a Windows target yet, because \
3855 the runtime they need has not been ported to Windows. -c still builds the object",
3856 opts.safety
3857 )
3858 })
3859}
3860
3861#[cfg(test)]
3862mod tests {
3863 use rucc_session::{
3864 Contract, GnucVersion, IncludeForm, LtoJobs, OptLevel, Partition, Patchable, Visibility,
3865 };
3866
3867 use super::*;
3868
3869 fn args(s: &[&str]) -> Vec<String> {
3870 s.iter().map(|x| (*x).to_owned()).collect()
3871 }
3872
3873 /// A target to write down where the host would otherwise decide, for the tests whose answer
3874 /// would be a different one on a different machine.
3875 ///
3876 /// Most of the tests here never name a target, which is right, because most of what the driver
3877 /// does with a command line is the same wherever it runs and a test that pinned one would be
3878 /// saying so in every case for the sake of the two that need it. The two that need it are the
3879 /// ones whose answer comes off the target rather than off the command line: the name an object
3880 /// gets, which is `a.o` here and `a.obj` on Windows, and whether Microsoft's reading of a
3881 /// nameless member is on, which is off here and on there. Both are the compiler being right, and
3882 /// a test that leaves the target to the host is asking a question with two correct answers.
3883 const LINUX: &str = "--target=x86_64-unknown-linux-gnu";
3884
3885 #[test]
3886 fn a_response_file_is_split_the_way_libiberty_splits_one() {
3887 let words = response_words("-Wl,--as-needed 'a b' \"c d\"\ne\\ f '' \"it's\" g\\\\h\n");
3888 assert_eq!(words, ["-Wl,--as-needed", "a b", "c d", "e f", "", "it's", "g\\h"]);
3889 assert!(response_words(" \n\t").is_empty());
3890 }
3891
3892 #[test]
3893 fn a_response_file_on_the_command_line_is_read_in_its_place() {
3894 let dir = std::env::temp_dir().join(format!("rucc-rsp-{}", std::process::id()));
3895 std::fs::create_dir_all(&dir).unwrap();
3896 let inner = dir.join("inner.rsp");
3897 std::fs::write(&inner, "-lm\n").unwrap();
3898 let outer = dir.join("outer.rsp");
3899 // Backslashes doubled, because the file is split the way libiberty splits one and a
3900 // Windows path is full of them.
3901 let named = inner.display().to_string().replace('\\', "\\\\");
3902 std::fs::write(&outer, format!("-o 'my prog' -Wl,--as-needed @{named}\n")).unwrap();
3903 let line = args(&["x.o", &format!("@{}", outer.display()), "@no-such-file"]);
3904 assert_eq!(
3905 response_files(&line).unwrap(),
3906 args(&["x.o", "-o", "my prog", "-Wl,--as-needed", "-lm", "@no-such-file"])
3907 );
3908 let itself = dir.join("itself.rsp");
3909 let named = itself.display().to_string().replace('\\', "\\\\");
3910 std::fs::write(&itself, format!("@{named}")).unwrap();
3911 let looped = response_files(&args(&[&format!("@{}", itself.display())]));
3912 assert!(looped.is_err(), "a file that names itself should be refused");
3913 std::fs::remove_dir_all(&dir).unwrap();
3914 }
3915
3916 #[test]
3917 fn help_and_version_win_over_everything_else() {
3918 assert_eq!(parse_args(&args(&["-c", "--help", "x.c"])).unwrap(), Action::Help);
3919 assert_eq!(parse_args(&args(&["--version"])).unwrap(), Action::Version);
3920 }
3921
3922 fn compile(s: &[&str]) -> (Box<Options>, Box<Plan>) {
3923 match parse_args(&args(s)).expect("expected a compilation") {
3924 Action::Compile { opts, plan, .. } => (opts, plan),
3925 other => panic!("expected a compilation, got {other:?}"),
3926 }
3927 }
3928
3929 fn linking(s: &[&str]) -> (Box<LinkOptions>, Box<Plan>) {
3930 match parse_args(&args(s)).expect("expected a compilation") {
3931 Action::Compile { link, plan, .. } => (link, plan),
3932 other => panic!("expected a compilation, got {other:?}"),
3933 }
3934 }
3935
3936 fn notes(s: &[&str]) -> Vec<String> {
3937 match parse_args(&args(s)).expect("expected a compilation") {
3938 Action::Compile { notes, .. } => notes,
3939 other => panic!("expected a compilation, got {other:?}"),
3940 }
3941 }
3942
3943 /// The ordinary command line has nothing to say about itself, which is the property that makes
3944 /// a note worth reading when there is one.
3945 #[test]
3946 fn a_command_line_with_nothing_wrong_with_it_carries_no_notes() {
3947 assert_eq!(notes(&["-c", "a.c"]), Vec::<String>::new());
3948 }
3949
3950 /// `-g` for Windows is kept, and says nothing, now that the COFF writer has DWARF sections.
3951 #[test]
3952 fn debug_information_for_coff_is_kept() {
3953 assert_eq!(
3954 notes(&["--target=x86_64-windows-gnu", "-g", "-c", "a.c"]),
3955 Vec::<String>::new()
3956 );
3957 let (opts, _) = compile(&["--target=x86_64-windows-gnu", "-g", "-c", "a.c"]);
3958 assert!(opts.debug_info);
3959 }
3960
3961 /// A directory that is not there contributes nothing to the search path, so there is no tree to
3962 /// read a release out of and nothing to compare the pin against. Said as a test because this is
3963 /// the shape a hermetic machine takes: the probe reads the disk and every other machine has a
3964 /// different disk, so what can be asserted here is the silence.
3965 #[test]
3966 fn a_named_tree_that_is_not_on_the_machine_is_not_a_release_mismatch() {
3967 let said =
3968 notes(&["--target=x86_64-linux-gnu.2.28", "--sysroot=/nowhere-at-all", "-c", "a.c"]);
3969 assert_eq!(said, Vec::<String>::new());
3970 }
3971
3972 #[test]
3973 fn collects_inputs_and_flags() {
3974 let (opts, plan) = compile(&["-c", "-O2", "-g", "a.c", "b.c"]);
3975 let paths: Vec<&str> = plan.jobs.iter().map(|j| j.input.as_str()).collect();
3976 assert_eq!(paths, vec!["a.c", "b.c"]);
3977 assert_eq!(opts.opt_level, OptLevel::O2);
3978 assert_eq!(opts.emit, EmitKind::Object);
3979 assert!(opts.debug_info);
3980 }
3981
3982 /// The unstable options, which are spelled apart from everything else on purpose: what is
3983 /// under `-Z` promises nothing, and a build that reaches for one should have had to say so.
3984 #[test]
3985 fn an_unstable_option_is_taken_and_one_that_does_not_exist_is_refused() {
3986 let (opts, _) = compile(&["-c", "-Zrule-coverage=/tmp/rules.cov", "a.c"]);
3987 assert_eq!(opts.rule_coverage.as_deref(), Some("/tmp/rules.cov"));
3988
3989 let (plain, _) = compile(&["-c", "a.c"]);
3990 assert_eq!(plain.rule_coverage, None, "nothing is measured unless it was asked for");
3991
3992 assert!(parse_args(&args(&["-Zrule-coverage=", "a.c"])).is_err(), "a file with no name");
3993 let unknown = parse_args(&args(&["-Zwhat", "a.c"])).expect_err("there is no such option");
3994 assert!(unknown.message.contains("4.11"), "{}", unknown.message);
3995 }
3996
3997 /// The other measurement written to a file, which reads the same way and fails the same way.
3998 #[test]
3999 fn where_the_register_pressure_goes_is_asked_for_the_same_way() {
4000 let (opts, _) = compile(&["-c", "-O2", "-Zregister-pressure=/tmp/spills.txt", "a.c"]);
4001 assert_eq!(opts.register_pressure.as_deref(), Some("/tmp/spills.txt"));
4002
4003 let (plain, _) = compile(&["-c", "a.c"]);
4004 assert_eq!(plain.register_pressure, None, "nothing is measured unless it was asked for");
4005
4006 assert!(parse_args(&args(&["-Zregister-pressure=", "a.c"])).is_err(), "no file named");
4007 }
4008
4009 /// Which register allocator runs, asked for by name, and left to the level when it is not.
4010 #[test]
4011 fn the_register_allocator_is_asked_for_by_name() {
4012 let (opts, _) = compile(&["-c", "-O2", "-Zregalloc=backtracking", "a.c"]);
4013 assert_eq!(opts.backtracking, Some(true));
4014 let (opts, _) = compile(&["-c", "-O2", "-Zregalloc=single", "a.c"]);
4015 assert_eq!(opts.backtracking, Some(false));
4016 let (plain, _) = compile(&["-c", "-O2", "a.c"]);
4017 assert_eq!(plain.backtracking, None, "the level decides unless it was asked for");
4018 assert!(parse_args(&args(&["-Zregalloc=graph", "a.c"])).is_err(), "not an allocator");
4019 }
4020
4021 /// The third one, which says what the pre-selection lowering group did.
4022 #[test]
4023 fn a_switch_shape_is_forced_by_name_and_only_by_one_it_has() {
4024 let (opts, _) = compile(&["-c", "-O2", "-Zswitch=walk", "a.c"]);
4025 assert_eq!(opts.switch_shape.as_deref(), Some("walk"));
4026 let (plain, _) = compile(&["-c", "-O2", "a.c"]);
4027 assert_eq!(plain.switch_shape, None, "nothing is forced unless it was asked for");
4028 assert!(parse_args(&args(&["-Zswitch=bit-test", "a.c"])).is_err(), "not a shape it forces");
4029 }
4030
4031 #[test]
4032 fn where_the_lowering_dump_goes_is_asked_for_the_same_way() {
4033 let (opts, _) = compile(&["-c", "-O2", "-Zlowering=/tmp/lowering.txt", "a.c"]);
4034 assert_eq!(opts.lowering_dump.as_deref(), Some("/tmp/lowering.txt"));
4035
4036 let (plain, _) = compile(&["-c", "a.c"]);
4037 assert_eq!(plain.lowering_dump, None, "nothing is dumped unless it was asked for");
4038
4039 assert!(parse_args(&args(&["-Zlowering=", "a.c"])).is_err(), "no file named");
4040 }
4041
4042 /// Scheduling, which has the three way answer every optimization flag has: on, off, and
4043 /// nothing said, which is whatever the optimization level asks for. The name is gcc's, and
4044 /// gcc's has a two in it because gcc has a scheduler before allocation and one after and this
4045 /// is the one after.
4046 #[test]
4047 fn scheduling_can_be_turned_on_and_off_and_left_to_the_optimization_level() {
4048 let (on, _) = compile(&["-c", "-O0", "-fschedule-insns2", "a.c"]);
4049 assert_eq!(on.schedule_insns, Some(true));
4050
4051 let (off, _) = compile(&["-c", "-O2", "-fno-schedule-insns2", "a.c"]);
4052 assert_eq!(off.schedule_insns, Some(false));
4053
4054 let (quiet, _) = compile(&["-c", "-O2", "a.c"]);
4055 assert_eq!(quiet.schedule_insns, None, "nothing said, so the level decides");
4056 assert!(quiet.opt_level.schedules(), "and at this level the level says yes");
4057
4058 let (none, _) = compile(&["-c", "a.c"]);
4059 assert!(!none.opt_level.schedules(), "at no optimization it says no");
4060 }
4061
4062 /// Tail calls, which gcc spells as sibling calls and turns on at `-O2` and `-Os`.
4063 #[test]
4064 fn sibling_calls_can_be_turned_on_and_off_and_left_to_the_optimization_level() {
4065 let (on, _) = compile(&["-c", "-O1", "-foptimize-sibling-calls", "a.c"]);
4066 assert_eq!(on.sibling_calls, Some(true));
4067
4068 let (off, _) = compile(&["-c", "-O2", "-fno-optimize-sibling-calls", "a.c"]);
4069 assert_eq!(off.sibling_calls, Some(false));
4070
4071 let (quiet, _) = compile(&["-c", "-Os", "a.c"]);
4072 assert_eq!(quiet.sibling_calls, None, "nothing said, so the level decides");
4073 assert!(quiet.opt_level.sibling_calls(), "and at this level the level says yes");
4074
4075 let (one, _) = compile(&["-c", "-O1", "a.c"]);
4076 assert!(!one.opt_level.sibling_calls(), "gcc leaves them off at -O1");
4077 }
4078
4079 /// Whether the timing model is worth holding an instruction back over, which is a `-Z` because
4080 /// it is a question about a target's description rather than about the program being compiled.
4081 #[test]
4082 fn whether_the_timing_model_is_cycle_accurate_can_be_overridden() {
4083 let (yes, _) = compile(&["-c", "-O2", "-Zcycle-accurate-model=yes", "a.c"]);
4084 assert_eq!(yes.cycle_accurate_model, Some(true));
4085
4086 let (no, _) = compile(&["-c", "-O2", "-Zcycle-accurate-model=no", "a.c"]);
4087 assert_eq!(no.cycle_accurate_model, Some(false));
4088
4089 let (plain, _) = compile(&["-c", "-O2", "a.c"]);
4090 assert_eq!(plain.cycle_accurate_model, None, "the target's own answer stands");
4091
4092 let bad = parse_args(&args(&["-Zcycle-accurate-model=maybe", "a.c"]))
4093 .expect_err("it takes yes or no");
4094 assert!(bad.message.contains("yes or no"), "{}", bad.message);
4095 }
4096
4097 #[test]
4098 fn a_bare_dash_o_means_o1_the_way_gcc_reads_it() {
4099 let (opts, _) = compile(&["-O", "a.c"]);
4100 assert_eq!(opts.opt_level, OptLevel::O1);
4101 }
4102
4103 #[test]
4104 fn dash_x_applies_to_later_inputs_only_and_none_stops_it() {
4105 let (_, plan) = compile(&["a.o", "-x", "c", "b.txt", "-x", "none", "c.o"]);
4106 assert_eq!(plan.jobs[0].kind, InputKind::LinkerInput);
4107 assert_eq!(plan.jobs[1].kind, InputKind::C);
4108 assert_eq!(plan.jobs[2].kind, InputKind::LinkerInput);
4109 }
4110
4111 #[test]
4112 fn dash_x_can_be_joined_to_its_language() {
4113 let (_, plan) = compile(&["a.o", "-xc", "b.txt", "-xnone", "c.o"]);
4114 assert_eq!(plan.jobs[0].kind, InputKind::LinkerInput);
4115 assert_eq!(plan.jobs[1].kind, InputKind::C);
4116 assert_eq!(plan.jobs[2].kind, InputKind::LinkerInput);
4117 }
4118
4119 #[test]
4120 fn dash_j_reaches_the_scheduler_and_defaults_to_the_machine() {
4121 let (_, _, jobs) = match parse_args(&args(&["-j4", "a.c"])).unwrap() {
4122 Action::Compile { opts, plan, jobs, .. } => (opts, plan, jobs),
4123 other => panic!("expected a compilation, got {other:?}"),
4124 };
4125 assert_eq!(jobs.count(), 4);
4126
4127 let default = match parse_args(&args(&["a.c"])).unwrap() {
4128 Action::Compile { jobs, .. } => jobs,
4129 other => panic!("expected a compilation, got {other:?}"),
4130 };
4131 assert_eq!(default, Jobs::available());
4132 assert!(parse_args(&args(&["-j0", "a.c"])).is_err());
4133 }
4134
4135 #[test]
4136 fn triple_hash_prints_the_plan_and_runs_nothing() {
4137 let a = parse_args(&args(&["-###", "-c", "a.c"])).unwrap();
4138 let Action::PrintPlan { plan, .. } = a else { panic!("expected a plan dump") };
4139 assert!(plan.render().contains("a.c: preprocess, compile, assemble -> a.o"));
4140 }
4141
4142 #[test]
4143 fn the_flag_that_keeps_the_intermediate_files_has_three_spellings_and_two_meanings() {
4144 // The bare one is `=obj` and not `=cwd`. gcc's manual says the opposite and gcc 16 does
4145 // this, and following the compiler is what makes a build that reads either of them find
4146 // the files where they are.
4147 assert_eq!(compile(&["-c", "-save-temps", "a.c"]).0.save_temps, SaveTemps::Object);
4148 assert_eq!(compile(&["-c", "-save-temps=obj", "a.c"]).0.save_temps, SaveTemps::Object);
4149 assert_eq!(compile(&["-c", "-save-temps=cwd", "a.c"]).0.save_temps, SaveTemps::Cwd);
4150 assert_eq!(compile(&["-c", "a.c"]).0.save_temps, SaveTemps::No);
4151 // The last one on the line decides, the way it does for every other flag with an
4152 // argument, and a keyword that is neither is fatal rather than ignored: a run that kept
4153 // nothing and said nothing looks exactly like one where the files were not produced.
4154 let (opts, _) = compile(&["-c", "-save-temps", "-save-temps=cwd", "a.c"]);
4155 assert_eq!(opts.save_temps, SaveTemps::Cwd);
4156 let e = parse_args(&args(&["-c", "-save-temps=nowhere", "a.c"])).unwrap_err();
4157 assert!(e.message.contains("accepted: cwd, obj"), "{}", e.message);
4158 }
4159
4160 #[test]
4161 fn the_flag_that_times_each_step_reaches_the_options_and_changes_nothing_else() {
4162 let (opts, plan) = compile(&["-c", "-time", "a.c"]);
4163 let (plain, without) = compile(&["-c", "a.c"]);
4164 assert!(opts.time);
4165 assert!(!plain.time);
4166 // Against the same line without the flag rather than against a spelling of the object's
4167 // name, since what the object is called is the host's business and this is not about that.
4168 assert_eq!(plan.jobs[0].output, without.jobs[0].output);
4169 }
4170
4171 #[test]
4172 fn dash_x_names_what_it_accepts_when_it_does_not_know_a_language() {
4173 let e = parse_args(&args(&["-x", "fortran", "a.c"])).unwrap_err();
4174 assert!(e.message.contains("assembler-with-cpp"), "{}", e.message);
4175 }
4176
4177 /// What `--fetch` says for a target this release pins nothing for, which today is every target
4178 /// but the three windows-gnu ones, the four musl ones and the eight glibc ones.
4179 #[test]
4180 fn a_fetch_of_a_target_nothing_is_pinned_for_says_so_rather_than_reaching_the_network() {
4181 let e = parse_args(&args(&["--fetch", "x86_64-linux-gnux32"])).unwrap_err();
4182 assert!(e.message.contains("pins no sysroot for x86_64-linux-gnux32"), "{}", e.message);
4183 // And what it does pin, because a release with some rows in the table and a release with
4184 // none are two situations and the second sentence is what tells them apart.
4185 assert!(e.message.contains("x86_64-windows-gnu"), "{}", e.message);
4186 // The joined spelling is the same flag.
4187 let joined = parse_args(&args(&["--fetch=x86_64-linux-gnux32"])).unwrap_err();
4188 assert_eq!(joined, e);
4189 }
4190
4191 /// The two targets a release will never pin, which is a different answer from the one above.
4192 ///
4193 /// Section 13.4. A person who reads "this release pins no sysroot yet" waits for a release that
4194 /// does, and no release of this compiler can ship either of these. An Apple target gets the
4195 /// licence and what to do instead, and a Microsoft one gets Microsoft's own files from
4196 /// Microsoft, which is the one lawful download either wall has behind it.
4197 #[test]
4198 fn a_fetch_of_a_target_behind_a_licence_wall_says_so_rather_than_saying_not_yet() {
4199 let e = parse_args(&args(&["--fetch", "aarch64-macos"])).unwrap_err();
4200 assert!(e.message.contains("Xcode licence"), "{}", e.message);
4201 assert!(e.message.contains("there never will be"), "{}", e.message);
4202 assert!(!e.message.contains("tamnd/rucc-cross"), "{}", e.message);
4203
4204 // Microsoft's side has a download behind it, which is Microsoft's own files from the build
4205 // this release pins, and the licence still has to be accepted for anything to move.
4206 let action = parse_args(&args(&["--fetch", "x86_64-windows-msvc"])).expect("pinned build");
4207 let Action::FetchMsvcSdk { target, accepted, pinned, .. } = action else {
4208 panic!("{action:?}")
4209 };
4210 assert_eq!(target.to_canonical_string(), "x86_64-windows-msvc");
4211 assert!(pinned);
4212 assert!(!accepted);
4213 let action = parse_args(&args(&["--fetch=aarch64-windows-msvc", "--accept-licence"]))
4214 .expect("pinned build");
4215 let Action::FetchMsvcSdk { accepted, pinned, .. } = action else { panic!("{action:?}") };
4216 assert!(accepted && pinned);
4217 // And the mingw-w64 target next to it is ours to ship and published, so the same flag has
4218 // something to get rather than a licence to explain.
4219 let action = parse_args(&args(&["--fetch", "x86_64-windows-gnu"])).expect("it is pinned");
4220 let Action::Fetch { what, .. } = action else { panic!("{action:?}") };
4221 assert_eq!(what.tuple, "x86_64-windows-gnu");
4222 }
4223
4224 #[test]
4225 fn the_other_fetch_takes_a_target_behind_microsofts_wall_and_carries_the_acceptance() {
4226 // Both spellings of the flag, because a flag that takes a tuple gets written both ways.
4227 for line in [
4228 vec!["--fetch-msvc-sdk", "x86_64-windows-msvc"],
4229 vec!["--fetch-msvc-sdk=x86_64-windows-msvc"],
4230 ] {
4231 let action = parse_args(&args(&line)).expect("that is a target behind the wall");
4232 let Action::FetchMsvcSdk { target, accepted, pinned, .. } = action else {
4233 panic!("{action:?}")
4234 };
4235 assert_eq!(target.to_canonical_string(), "x86_64-windows-msvc");
4236 // Nothing on the line accepted anything, so nothing did.
4237 assert!(!accepted);
4238 // This one follows Microsoft's channel to whatever it names today.
4239 assert!(!pinned);
4240 }
4241
4242 // And both spellings of the word, because the prose here uses one and most of the people
4243 // typing this will reach for the other.
4244 for word in ["--accept-licence", "--accept-license"] {
4245 let action = parse_args(&args(&["--fetch-msvc-sdk", "aarch64-windows-msvc", word]))
4246 .expect("that is a target behind the wall");
4247 let Action::FetchMsvcSdk { target, accepted, .. } = action else {
4248 panic!("{action:?}")
4249 };
4250 assert_eq!(target.to_canonical_string(), "aarch64-windows-msvc");
4251 assert!(accepted, "{word} should have been read");
4252 }
4253 }
4254
4255 #[test]
4256 fn the_other_fetch_refuses_the_command_lines_that_do_not_mean_anything() {
4257 // A tuple is what it gets, so a flag with nothing after it is not a command.
4258 let e = parse_args(&args(&["--fetch-msvc-sdk"])).unwrap_err();
4259 assert!(e.message.contains("requires the target"), "{}", e.message);
4260 let e = parse_args(&args(&["--fetch-msvc-sdk", "not-a-target"])).unwrap_err();
4261 assert!(e.message.contains("there is no SDK to get"), "{}", e.message);
4262
4263 // `--offline` forbids every download and this one asks for one, whichever order they came
4264 // in, which is the same answer `--fetch` gives.
4265 for line in [
4266 vec!["--offline", "--fetch-msvc-sdk", "x86_64-windows-msvc"],
4267 vec!["--fetch-msvc-sdk", "x86_64-windows-msvc", "--offline"],
4268 ] {
4269 let e = parse_args(&args(&line)).unwrap_err();
4270 assert!(e.message.contains("two opposite things"), "{}", e.message);
4271 }
4272
4273 // It gets an SDK and compiles nothing, so a file on the same line would be read by nothing.
4274 let e = parse_args(&args(&["--fetch-msvc-sdk", "x86_64-windows-msvc", "a.c"])).unwrap_err();
4275 assert!(e.message.contains("compiles nothing"), "{}", e.message);
4276
4277 // The two fetches are two commands and a line that asked for both asked for neither.
4278 let e = parse_args(&args(&[
4279 "--fetch",
4280 "x86_64-windows-gnu",
4281 "--fetch-msvc-sdk",
4282 "x86_64-windows-msvc",
4283 ]))
4284 .unwrap_err();
4285 assert!(e.message.contains("two different commands"), "{}", e.message);
4286
4287 // And an acceptance with nothing to accept for is a command line that says something about
4288 // a licence no part of it goes near.
4289 let e = parse_args(&args(&["--accept-licence", "-c", "a.c"])).unwrap_err();
4290 assert!(e.message.contains("--fetch-msvc-sdk <tuple> for a"), "{}", e.message);
4291 }
4292
4293 /// An Apple target on a machine with no SDK, which is section 8.6's other host.
4294 ///
4295 /// Not run on a mac, where the SDK this is about is installed and the compile is the ordinary one
4296 /// that uses it. What the reason says is asserted in `rucc_sysroot::wall` and where it is printed
4297 /// is asserted in `rucc-pp`, so what is left here is that the driver works it out and leaves it
4298 /// where the preprocessor will find it, and that neither way past the wall leaves one behind.
4299 #[test]
4300 fn an_apple_target_with_no_sdk_anywhere_carries_the_licence_rather_than_a_missing_directory() {
4301 if cfg!(target_os = "macos") || std::env::var_os("SDKROOT").is_some() {
4302 return;
4303 }
4304 let (opts, _) = compile(&["--target=aarch64-macos", "-c", "a.c"]);
4305 let why = opts.search.missing_system().expect("the wall is the reason there are none");
4306 assert!(why.contains("aarch64-macos needs a macOS SDK"), "{why}");
4307 assert!(why.contains("Xcode licence"), "{why}");
4308 assert!(why.contains("-isysroot"), "{why}");
4309
4310 // A program that includes none of the library needs none of the SDK, which is what section
4311 // 8.6 means by being able to target the platform without one, so there is nothing to explain.
4312 let (opts, _) = compile(&["--target=aarch64-macos", "-nostdinc", "-c", "a.c"]);
4313 assert_eq!(opts.search.missing_system(), None);
4314 // And naming a path is the other way through, whether or not the path is there: a mistyped
4315 // directory is a mistake to report on its own terms rather than a licence to explain.
4316 let (opts, _) = compile(&["--target=aarch64-macos", "-isysroot", "/opt/sdk", "-c", "a.c"]);
4317 assert_eq!(opts.search.missing_system(), None);
4318 }
4319
4320 /// The same wall on the compile side of an MSVC target, where the way past it is a tuple.
4321 ///
4322 /// Not run on Windows, for the same reason the one above is not run on a mac: the wall stands in
4323 /// front of an SDK this machine does not have, and a Windows machine is the kind that does. The
4324 /// driver asks `vswhere` where Visual Studio is and takes the newest kit under it, so on a box
4325 /// with the build tools installed there are headers, no wall and nothing here to be about.
4326 /// `INCLUDE` is the other way a machine has one and is the other half of the guard, since a
4327 /// person can set that anywhere while Visual Studio is only found on the platform it runs on.
4328 #[test]
4329 fn an_msvc_target_with_no_sdk_named_says_which_environment_needs_nothing_installed() {
4330 if cfg!(target_os = "windows") || std::env::var_os("INCLUDE").is_some() {
4331 return;
4332 }
4333 let (opts, _) = compile(&["--target=x86_64-windows-msvc", "-c", "a.c"]);
4334 let why = opts.search.missing_system().expect("the wall is the reason there are none");
4335 assert!(why.contains("the Windows SDK and its universal CRT"), "{why}");
4336 assert!(why.contains("mingw-w64"), "{why}");
4337 // And the mingw-w64 target has its headers from us, so nothing is missing to explain.
4338 let (opts, _) = compile(&["--target=x86_64-windows-gnu", "-c", "a.c"]);
4339 assert_eq!(opts.search.missing_system(), None);
4340 }
4341
4342 #[test]
4343 fn a_fetch_with_no_target_and_a_fetch_of_a_tuple_that_is_not_one_both_say_which() {
4344 let e = parse_args(&args(&["--fetch"])).unwrap_err();
4345 assert!(e.message.contains("--fetch requires"), "{}", e.message);
4346 let e = parse_args(&args(&["--fetch", "sparc64-solaris-gnu"])).unwrap_err();
4347 assert!(e.message.contains("--fetch sparc64-solaris-gnu"), "{}", e.message);
4348 assert!(e.message.contains("no sysroot to get"), "{}", e.message);
4349 }
4350
4351 /// Both flags on one line ask for opposite things, in either order.
4352 #[test]
4353 fn a_fetch_and_offline_together_is_a_refusal_whichever_way_round_they_are_written() {
4354 for line in [
4355 vec!["--offline", "--fetch", "x86_64-linux-musl"],
4356 vec!["--fetch", "x86_64-linux-musl", "--offline"],
4357 ] {
4358 let e = parse_args(&args(&line)).unwrap_err();
4359 assert!(e.message.contains("two opposite things"), "{}", e.message);
4360 }
4361 }
4362
4363 #[test]
4364 fn a_fetch_does_not_compile_anything_and_says_so_when_it_is_handed_a_file() {
4365 let e = parse_args(&args(&["--fetch", "x86_64-linux-musl", "a.c"])).unwrap_err();
4366 assert!(e.message.contains("compiles nothing"), "{}", e.message);
4367 assert!(e.message.contains("a.c"), "{}", e.message);
4368 }
4369
4370 /// `--offline` on its own is accepted and changes nothing, because an ordinary compile
4371 /// downloads nothing with or without it. A build that passes it everywhere is the case this is
4372 /// for, and it must not lose the compilation it was passed beside.
4373 #[test]
4374 fn offline_on_a_compilation_is_the_same_compilation() {
4375 let (opts, plan) = compile(&["-c", "--offline", "a.c"]);
4376 let (plain, without) = compile(&["-c", "a.c"]);
4377 assert_eq!(opts.target, plain.target);
4378 assert_eq!(plan.jobs.len(), without.jobs.len());
4379 assert_eq!(plan.jobs[0].output, without.jobs[0].output);
4380 }
4381
4382 #[test]
4383 fn a_deployment_target_comes_from_the_tuple_or_from_the_flag() {
4384 let version = |v: &str| rucc_tuple::Version::parse(v);
4385 let (opts, _) = compile(&["--target=aarch64-macos.13", "-c", "a.c"]);
4386 assert_eq!(opts.target, "aarch64-apple-darwin".parse().unwrap());
4387 assert_eq!(opts.os_version, version("13"));
4388 // The flag wins over the tuple, as it does under clang, and either spelling of it works.
4389 let (opts, _) =
4390 compile(&["--target=aarch64-macos.13", "-mmacosx-version-min=14.2", "-c", "a.c"]);
4391 assert_eq!(opts.os_version, version("14.2"));
4392 let (opts, _) = compile(&["--target=x86_64-macos", "-mmacos-version-min=12", "-c", "a.c"]);
4393 assert_eq!(opts.os_version, version("12"));
4394 // Nothing said leaves the platform's default to the target description.
4395 let (opts, _) = compile(&["--target=aarch64-macos", "-c", "a.c"]);
4396 assert_eq!(opts.os_version, None);
4397 // A Linux build that always passes the flag is not an Apple build because of it.
4398 let (opts, _) =
4399 compile(&["--target=aarch64-linux-gnu", "-mmacosx-version-min=13", "-c", "a.c"]);
4400 assert_eq!(opts.os_version, None);
4401 let e = parse_args(&args(&["-mmacosx-version-min=thirteen", "a.c"])).unwrap_err();
4402 assert!(e.message.contains("is not a version"), "{}", e.message);
4403 }
4404
4405 #[test]
4406 fn an_unknown_flag_is_an_error_rather_than_a_shrug() {
4407 let e = parse_args(&args(&["-fno-such-thing", "a.c"])).unwrap_err();
4408 assert!(e.message.contains("unknown option"), "{}", e.message);
4409 }
4410
4411 /// `-fpermissive` and the flag that turns it back off, which a build writes beside it when
4412 /// one directory needs the older rules and the rest of the tree does not.
4413 #[test]
4414 fn permissive_reads_in_both_directions_and_the_last_one_wins() {
4415 let (opts, _) = compile(&["-c", "a.c"]);
4416 assert!(!opts.permissive, "off unless it is asked for");
4417
4418 let (opts, _) = compile(&["-c", "-fpermissive", "a.c"]);
4419 assert!(opts.permissive);
4420
4421 let (opts, _) = compile(&["-c", "-fpermissive", "-fno-permissive", "a.c"]);
4422 assert!(!opts.permissive);
4423 }
4424
4425 #[test]
4426 fn asking_for_nested_functions_is_told_why_it_is_not_coming() {
4427 let e = parse_args(&args(&["-fnested-functions", "a.c"])).unwrap_err();
4428 assert!(e.message.contains("trampoline"), "{}", e.message);
4429 assert!(parse_args(&args(&["-fno-nested-functions", "a.c"])).is_ok());
4430 }
4431
4432 #[test]
4433 fn the_flag_every_configure_script_writes_is_taken() {
4434 // All four spellings, because a build writes whichever one its macros picked and a
4435 // compiler that takes three of them is a compiler that fails on the fourth.
4436 for flag in ["-fPIC", "-fpic", "-fPIE", "-fpie"] {
4437 let (opts, _) = compile(&["-c", flag, "a.c"]);
4438 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
4439 }
4440 }
4441
4442 #[test]
4443 fn a_table_is_written_unless_the_build_says_nothing_will_walk_it() {
4444 let (opts, _) = compile(&["-c", "a.c"]);
4445 assert!(opts.unwinds(), "the default is off");
4446 let (opts, _) = compile(&["-c", "-fno-asynchronous-unwind-tables", "a.c"]);
4447 assert!(!opts.unwinds(), "the build was not taken at its word");
4448 let (opts, _) = compile(&[
4449 "-c",
4450 "-fno-asynchronous-unwind-tables",
4451 "-fasynchronous-unwind-tables",
4452 "a.c",
4453 ]);
4454 assert!(opts.unwinds(), "the last flag did not win");
4455 // The weaker request, which the same table answers, so a line that asks for a table and
4456 // against an asynchronous one gets one. That is gcc's arrangement and it turns up when a
4457 // build turns the asynchronous one off globally and a directory asks for a table back.
4458 let (opts, _) =
4459 compile(&["-c", "-fno-asynchronous-unwind-tables", "-funwind-tables", "a.c"]);
4460 assert!(opts.unwinds(), "the weaker request was dropped");
4461 let (opts, _) = compile(&["-c", "-fno-unwind-tables", "a.c"]);
4462 assert!(opts.unwinds(), "the weaker negative turned off the stronger request");
4463 let (opts, _) =
4464 compile(&["-c", "-fno-unwind-tables", "-fno-asynchronous-unwind-tables", "a.c"]);
4465 assert!(!opts.unwinds(), "both were turned off and one stayed on");
4466 }
4467
4468 #[test]
4469 fn the_flags_that_describe_what_this_compiler_already_does_are_taken() {
4470 // Every one of these is on a real build line somewhere and every one of them was an
4471 // unknown option. What they have in common is that the answer rucc gives is the answer
4472 // they ask for, so there is nothing to implement and nothing to refuse.
4473 for flag in [
4474 "-fstrict-aliasing",
4475 "-fno-strict-aliasing",
4476 "-fdelete-null-pointer-checks",
4477 "-fno-delete-null-pointer-checks",
4478 "-frounding-math",
4479 "-fno-rounding-math",
4480 "-fexcess-precision=standard",
4481 "-fexcess-precision=fast",
4482 "-fexcess-precision=16",
4483 "-pipe",
4484 "-cpp",
4485 "-fdiagnostics-color",
4486 "-fno-diagnostics-color",
4487 "-fdiagnostics-color=always",
4488 "-fdiagnostics-color=never",
4489 "-fdiagnostics-color=auto",
4490 ] {
4491 let (opts, _) = compile(&["-c", flag, "a.c"]);
4492 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
4493 }
4494 }
4495
4496 #[test]
4497 fn whether_an_exception_is_looked_at_is_kept_and_defaults_to_gccs_answer() {
4498 let (opts, _) = compile(&["-c", "a.c"]);
4499 assert!(opts.trapping_math, "the default was not gcc's");
4500 let (opts, _) = compile(&["-c", "-fno-trapping-math", "a.c"]);
4501 assert!(!opts.trapping_math);
4502 let (opts, _) = compile(&["-c", "-ftrapping-math", "a.c"]);
4503 assert!(opts.trapping_math, "spelling out the default turned it off");
4504 // The last one written wins, which is how a build line that inherits a flag from one
4505 // place and overrides it in another is read.
4506 let (opts, _) = compile(&["-c", "-fno-trapping-math", "-ftrapping-math", "a.c"]);
4507 assert!(opts.trapping_math);
4508 }
4509
4510 /// The flags a torture program writes on its own `dg-options` line, which is where most of
4511 /// these come from: a program reduced from a miscompilation names the pass that miscompiled
4512 /// it. Eighteen programs in the suite stopped on the driver before anything read them, and
4513 /// tamnd/rucc#1019 is the list.
4514 #[test]
4515 fn no_inline_turns_off_the_inlining_of_a_function_declared_inline() {
4516 let (opts, _) = compile(&["-c", "-O2", "-fno-inline", "a.c"]);
4517 assert_eq!(opts.passes, [(rucc_opt::inline::NAME.to_owned(), false)]);
4518 }
4519
4520 #[test]
4521 fn inlining_a_function_called_once_is_turned_off_and_on_by_its_own_flag() {
4522 for level in ["-O0", "-O1", "-O2", "-O3", "-Os", "-Oz", "-Og"] {
4523 let (opts, _) = compile(&["-c", level, "-fno-inline-functions-called-once", "a.c"]);
4524 assert_eq!(opts.passes, [(rucc_opt::inline::ONCE.to_owned(), false)], "{level}");
4525 let (opts, _) = compile(&["-c", level, "-finline-functions-called-once", "a.c"]);
4526 assert_eq!(opts.passes, [(rucc_opt::inline::ONCE.to_owned(), true)], "{level}");
4527 }
4528 }
4529
4530 #[test]
4531 fn the_flags_that_name_a_pass_of_gccs_own_are_taken_and_dropped() {
4532 for flag in [
4533 "-fno-tree-ccp",
4534 "-fno-tree-dominator-opts",
4535 "-fno-tree-vrp",
4536 "-fno-tree-bit-ccp",
4537 "-fno-tree-coalesce-vars",
4538 "-ftree-vectorize",
4539 "-ftree-loop-distribution",
4540 "-fipa-pta",
4541 "-fmodulo-sched",
4542 "-fno-vect-cost-model",
4543 "-fvect-cost-model=unlimited",
4544 "-fsimd-cost-model=cheap",
4545 "-fexpensive-optimizations",
4546 "-fno-early-inlining",
4547 "-finline-functions",
4548 "-foptimize-strlen",
4549 "-fno-ira-share-spill-slots",
4550 ] {
4551 let (opts, _) = compile(&["-c", flag, "a.c"]);
4552 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
4553 assert!(opts.passes.is_empty(), "{flag} named a pass of gcc's and not one of ours");
4554 }
4555 }
4556
4557 /// The two namespaces are taken whole, so a name neither this test nor gcc 16 has heard of
4558 /// goes the same way as the ones above rather than stopping a build on the day gcc adds it.
4559 #[test]
4560 fn a_pass_name_in_either_family_is_taken_whether_or_not_it_is_one_gcc_has() {
4561 for flag in ["-ftree-no-such-pass", "-fno-ipa-no-such-pass"] {
4562 let (opts, _) = compile(&["-c", flag, "a.c"]);
4563 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
4564 }
4565 }
4566
4567 /// A pass this compiler has keeps its flag, since the arms that read the registry are above
4568 /// the family arms. `dce` is the one both compilers have a name for, and `execute/pr97421-2.c`
4569 /// is the program that writes it.
4570 #[test]
4571 fn a_pass_name_this_compiler_has_is_still_read_as_a_pass() {
4572 let (opts, _) = compile(&["-c", "-fno-dce", "a.c"]);
4573 assert_eq!(opts.passes, vec![("dce".to_owned(), false)]);
4574 }
4575
4576 /// gcc's name for the unroller reaches the unroller, in both directions. libtommath puts
4577 /// `-funroll-loops` in `CFLAGS` unconditionally, and before this it was an unknown option and
4578 /// the build stopped on its first file.
4579 #[test]
4580 fn the_gcc_spelling_of_the_unroller_turns_the_unroller_on_and_off() {
4581 let (opts, _) = compile(&["-c", "-funroll-loops", "a.c"]);
4582 assert_eq!(opts.passes, vec![("unroll".to_owned(), true)]);
4583 let (opts, _) = compile(&["-c", "-fno-unroll-loops", "a.c"]);
4584 assert_eq!(opts.passes, vec![("unroll".to_owned(), false)]);
4585 }
4586
4587 /// The three transformations that are a module at a time are named by a flag as well, even
4588 /// though none of them is a `rucc_opt::Pass` and so none is reached by the generic arms.
4589 ///
4590 /// A bisection over a miscompilation turns one thing off at a time, and a transformation with
4591 /// no spelling of its own cannot be the one turned off.
4592 #[test]
4593 fn the_transformations_that_are_not_passes_are_still_named_by_a_flag() {
4594 let (opts, _) = compile(&["-c", "-fno-ipa-cp", "-fipa-sra", "-fno-libcall", "a.c"]);
4595 assert_eq!(
4596 opts.passes,
4597 vec![
4598 (rucc_opt::ipcp::NAME.to_owned(), false),
4599 (rucc_opt::ipasra::NAME.to_owned(), true),
4600 (rucc_opt::libcall::NAME.to_owned(), false),
4601 ]
4602 );
4603 let (opts, _) = compile(&["-c", "-flibcall", "a.c"]);
4604 assert_eq!(opts.passes, vec![(rucc_opt::libcall::NAME.to_owned(), true)]);
4605 }
4606
4607 /// Where a function starts is a question this compiler answers, so the flag that asks about it
4608 /// is answered rather than dropped. femtolisp's Makefile writes the bare form on every compile
4609 /// of the project, and before this it was an unknown option and the build stopped on its first
4610 /// file. The numbers are gcc 16's, read off `-S` on x86-64: nothing and the bare form both
4611 /// give `.p2align 4`, `=32` gives 5, `=3` gives 2, and the negative form gives `.align 8`.
4612 #[test]
4613 fn the_alignment_of_a_function_is_a_request_this_compiler_can_answer() {
4614 let (opts, _) = compile(&["-c", "-falign-functions", "a.c"]);
4615 assert_eq!(opts.align_functions, None, "the bare form asks for the default");
4616
4617 let (opts, _) = compile(&["-c", "-falign-functions=32", "a.c"]);
4618 assert_eq!(opts.align_functions, Some(32));
4619
4620 let (opts, _) = compile(&["-c", "-falign-functions=3", "a.c"]);
4621 assert_eq!(opts.align_functions, Some(4), "rounded up rather than refused");
4622
4623 let (opts, _) = compile(&["-c", "-falign-functions=32:8", "a.c"]);
4624 assert_eq!(opts.align_functions, Some(32), "the boundary is the answerable half");
4625
4626 for flag in ["-falign-functions=0", "-falign-functions=1"] {
4627 let (opts, _) = compile(&["-c", flag, "a.c"]);
4628 assert_eq!(opts.align_functions, None, "{flag} means the default");
4629 }
4630
4631 let (opts, _) = compile(&["-c", "-fno-align-functions", "a.c"]);
4632 assert_eq!(opts.align_functions, Some(8), "the smallest boundary the target has");
4633
4634 // The last one on the line wins, which is how gcc reads a repeated flag.
4635 let (opts, _) = compile(&["-c", "-falign-functions=32", "-falign-functions", "a.c"]);
4636 assert_eq!(opts.align_functions, None);
4637
4638 let e = parse_args(&args(&["-c", "-falign-functions=big", "a.c"])).unwrap_err();
4639 assert!(e.message.contains("number of bytes"), "{}", e.message);
4640 }
4641
4642 /// The other three of the family are about padding inside a body, so none of them is about
4643 /// where a function starts. Every spelling of each, since a build writes whichever one its
4644 /// author typed.
4645 #[test]
4646 fn the_alignment_flags_about_the_inside_of_a_body_are_taken_and_say_nothing() {
4647 for flag in [
4648 "-falign-labels",
4649 "-falign-loops",
4650 "-falign-jumps",
4651 "-falign-loops=16",
4652 "-falign-labels=32",
4653 "-fno-align-loops",
4654 "-fno-align-labels",
4655 "-fno-align-jumps",
4656 ] {
4657 let (opts, _) = compile(&["-c", flag, "a.c"]);
4658 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
4659 assert_eq!(opts.align_functions, None, "{flag} is not about where a function starts");
4660 }
4661 }
4662
4663 /// The loop flag in either direction is an answer, and a command line that wrote neither
4664 /// leaves the level to decide.
4665 #[test]
4666 fn the_loop_alignment_flag_is_answered_both_ways() {
4667 assert_eq!(compile(&["-c", "-O2", "a.c"]).0.align_loops, None);
4668 assert_eq!(compile(&["-c", "-O0", "-falign-loops", "a.c"]).0.align_loops, Some(true));
4669 assert_eq!(compile(&["-c", "-O2", "-fno-align-loops", "a.c"]).0.align_loops, Some(false));
4670 assert_eq!(compile(&["-c", "-falign-loops=32", "a.c"]).0.align_loops, None, "a number");
4671 }
4672
4673 /// The encoding of the source is not a question about speed, so the one name that describes
4674 /// what the preprocessor does is taken and every other name is refused.
4675 #[test]
4676 fn the_input_charset_is_taken_when_it_names_the_one_that_is_read() {
4677 for flag in ["-finput-charset=utf-8", "-finput-charset=UTF-8", "-finput-charset=utf8"] {
4678 let (opts, _) = compile(&["-c", flag, "a.c"]);
4679 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
4680 }
4681
4682 let e = parse_args(&args(&["-c", "-finput-charset=latin1", "a.c"])).unwrap_err();
4683 assert!(e.message.contains("latin1"), "{}", e.message);
4684 assert!(e.message.contains("UTF-8"), "what is read is worth saying: {}", e.message);
4685 }
4686
4687 /// `-fnon-call-exceptions` turns exceptions on unless `-fexceptions` or `-fno-exceptions` was
4688 /// written, and the one written wins whichever side of it it is on, which is gcc 16's reading.
4689 #[test]
4690 fn exceptions_are_on_when_asked_for_and_non_call_ones_ask_unless_told_not_to() {
4691 let (opts, _) = compile(&["-c", "a.c"]);
4692 assert!(!opts.exceptions && !opts.non_call_exceptions, "gcc's default for C is off");
4693 let (opts, _) = compile(&["-c", "-fexceptions", "a.c"]);
4694 assert!(opts.exceptions && !opts.non_call_exceptions);
4695 let (opts, _) = compile(&["-c", "-fexceptions", "-fno-exceptions", "a.c"]);
4696 assert!(!opts.exceptions);
4697 let (opts, _) = compile(&["-c", "-fnon-call-exceptions", "a.c"]);
4698 assert!(opts.exceptions && opts.non_call_exceptions);
4699 for line in [
4700 ["-fno-exceptions", "-fnon-call-exceptions"],
4701 ["-fnon-call-exceptions", "-fno-exceptions"],
4702 ] {
4703 let (opts, _) = compile(&["-c", line[0], line[1], "a.c"]);
4704 assert!(!opts.exceptions && opts.non_call_exceptions, "{line:?}");
4705 }
4706 let (opts, _) =
4707 compile(&["-c", "-fnon-call-exceptions", "-fno-non-call-exceptions", "a.c"]);
4708 assert!(!opts.exceptions && !opts.non_call_exceptions);
4709 let (opts, _) = compile(&["-c", "-fno-delete-dead-exceptions", "a.c"]);
4710 assert_eq!(opts.emit, EmitKind::Object);
4711 }
4712
4713 /// `-ffast-math` used to be refused beside it and is the family it names now, with each
4714 /// member settable on its own and the last word on each winning, which is gcc's reading.
4715 #[test]
4716 fn fast_math_is_the_family_it_names_and_the_last_word_on_each_member_wins() {
4717 let both = |line: &[&str]| {
4718 let (opts, _) = compile(&[&["-c"], line, &["a.c"]].concat());
4719 let (link, _) = linking(&[line, &["a.c"]].concat());
4720 (opts, link)
4721 };
4722 let (opts, link) = both(&[]);
4723 assert_eq!(opts.math, Math::default());
4724 assert!(opts.trapping_math);
4725 assert!(!link.fast_math);
4726
4727 let (opts, link) = both(&["-ffast-math"]);
4728 assert!(opts.math.fast(opts.trapping_math), "{:?}", opts.math);
4729 assert!(!opts.trapping_math, "fast math turns trapping off");
4730 assert!(link.fast_math, "and it links the startup file");
4731
4732 // Taking one member back leaves the rest, and the whole is not fast math any more.
4733 let (opts, link) = both(&["-ffast-math", "-fno-finite-math-only"]);
4734 assert!(!opts.math.finite_only);
4735 assert!(!opts.math.errno && !opts.math.signed_zeros && opts.math.reciprocal);
4736 assert!(!opts.math.fast(opts.trapping_math));
4737 assert!(link.fast_math, "gcc's spec reads the flag and not the fields");
4738
4739 let (opts, _) = both(&["-ffast-math", "-ftrapping-math"]);
4740 assert!(opts.trapping_math);
4741 assert!(!opts.math.fast(opts.trapping_math));
4742 assert!(!opts.math.associative(opts.trapping_math));
4743
4744 let (opts, link) = both(&["-ffast-math", "-fno-fast-math"]);
4745 assert_eq!(opts.math, Math::default());
4746 assert!(opts.trapping_math);
4747 assert!(!link.fast_math);
4748
4749 // A member written alone is only that member.
4750 let (opts, link) = both(&["-fno-math-errno"]);
4751 assert_eq!(opts.math, Math { errno: false, ..Math::default() });
4752 assert!(opts.math.iec_559(opts.trapping_math), "errno is not an IEC 60559 question");
4753 assert!(!link.fast_math);
4754
4755 let (opts, link) = both(&["-funsafe-math-optimizations"]);
4756 assert!(opts.math.unsafe_math && opts.math.associative(opts.trapping_math));
4757 assert!(opts.math.errno && !opts.math.finite_only);
4758 assert!(link.fast_math);
4759 }
4760
4761 /// `-Ofast` is `-O3` with fast math as a default, which a later level and a
4762 /// `-fno-fast-math` on either side of it both take back.
4763 #[test]
4764 fn ofast_is_o3_with_fast_math_as_a_default_a_flag_can_take_back() {
4765 let both = |line: &[&str]| {
4766 let (opts, _) = compile(&[&["-c"], line, &["a.c"]].concat());
4767 let (link, _) = linking(&[line, &["a.c"]].concat());
4768 (opts, link)
4769 };
4770 let (opts, link) = both(&["-Ofast"]);
4771 assert_eq!(opts.opt_level, OptLevel::O3);
4772 assert!(opts.math.fast(opts.trapping_math));
4773 assert!(link.fast_math);
4774
4775 for line in [&["-Ofast", "-O2"][..], &["-fno-fast-math", "-Ofast"]] {
4776 let (opts, _) = both(line);
4777 assert!(!opts.math.fast(opts.trapping_math), "{line:?}");
4778 }
4779
4780 let (_, link) = both(&["-Ofast", "-mno-daz-ftz"]);
4781 assert_eq!(link.daz_ftz, Some(false));
4782 }
4783
4784 /// `-finstrument-functions` used to be refused beside those two, and it is taken now that the
4785 /// hooks are called. The last of it and its negative is the one that counts, as with any pair.
4786 #[test]
4787 fn instrument_functions_is_taken_and_the_last_of_the_pair_wins() {
4788 let (opts, _) = compile(&["-c", "-finstrument-functions", "a.c"]);
4789 assert!(opts.instrument_functions);
4790 let (opts, _) =
4791 compile(&["-c", "-finstrument-functions", "-fno-instrument-functions", "a.c"]);
4792 assert!(!opts.instrument_functions);
4793 }
4794
4795 #[test]
4796 fn a_tentative_definition_is_common_on_darwin_unless_told_otherwise() {
4797 // Two files each writing `int g;` link under `-fcommon` and do not without it, and Apple's
4798 // clang has it on where every other compiler rucc stands in for has it off.
4799 let (opts, _) = compile(&[LINUX, "-c", "a.c"]);
4800 assert!(!Session::new(*opts).common());
4801
4802 let (opts, _) = compile(&["-c", "--target=aarch64-apple-darwin", "a.c"]);
4803 assert!(Session::new(*opts).common());
4804
4805 let (opts, _) = compile(&[LINUX, "-c", "-fcommon", "a.c"]);
4806 assert!(Session::new(*opts).common());
4807
4808 let (opts, _) =
4809 compile(&["-c", "--target=aarch64-apple-darwin", "-fcommon", "-fno-common", "a.c"]);
4810 assert!(!Session::new(*opts).common());
4811 }
4812
4813 #[test]
4814 fn asking_for_position_dependent_code_is_told_why_it_is_not_coming() {
4815 for flag in ["-fno-pic", "-fno-pie"] {
4816 let e = parse_args(&args(&[flag, "a.c"])).unwrap_err();
4817 assert!(e.message.contains("global offset table"), "{flag}: {}", e.message);
4818 // The one it may have meant, since the two are a letter apart and one of them is
4819 // about linking and is taken.
4820 assert!(e.message.contains("-no-pie"), "{flag}: {}", e.message);
4821 }
4822 }
4823
4824 #[test]
4825 fn a_program_name_with_a_known_target_in_front_of_rucc_picks_that_target() {
4826 let t = |p: &str| target_from_program(p);
4827 assert_eq!(t("aarch64-linux-gnu-rucc").as_deref(), Some("aarch64-linux-gnu"));
4828 assert_eq!(t("/usr/bin/riscv64-linux-musl-rucc").as_deref(), Some("riscv64-linux-musl"));
4829 assert_eq!(t(r"C:\bin\x86_64-windows-gnu-rucc.exe").as_deref(), Some("x86_64-windows-gnu"));
4830 assert_eq!(t("rucc"), None);
4831 assert_eq!(t("/usr/local/bin/rucc"), None);
4832 assert_eq!(t("my-rucc"), None);
4833 assert_eq!(t("sparc64-linux-gnu-rucc"), None);
4834 assert_eq!(t("aarch64-linux-gnu-gcc"), None);
4835 }
4836
4837 #[test]
4838 fn an_unsupported_target_names_itself() {
4839 let e = parse_args(&args(&["--target=sparc64-linux-gnu", "a.c"])).unwrap_err();
4840 assert!(e.message.contains("sparc64"), "{}", e.message);
4841 }
4842
4843 #[test]
4844 fn no_inputs_is_an_error_but_print_config_needs_none() {
4845 assert!(parse_args(&args(&[])).is_err());
4846 assert!(matches!(parse_args(&args(&["--print-config"])), Ok(Action::PrintConfig(_))));
4847 }
4848
4849 #[test]
4850 fn print_config_reports_the_target_it_was_given_not_the_host() {
4851 let a = parse_args(&args(&["--print-config", "--target=riscv64-linux-musl"])).unwrap();
4852 let Action::PrintConfig(opts) = a else { panic!("expected a configuration dump") };
4853 let text = print_config(&opts);
4854 assert!(text.contains("target: riscv64-unknown-linux-musl"), "{text}");
4855 assert!(text.contains("char-signed: false"), "{text}");
4856 assert!(text.contains("object-format: elf"), "{text}");
4857 assert!(text.contains("va-list: void-pointer"), "{text}");
4858 // RISC-V has a register file and this compiler has not written it down yet, and the
4859 // dump says which of those two it is rather than leaving the line out.
4860 assert!(text.contains("registers: none"), "{text}");
4861 assert!(text.contains("timing-model: none"), "{text}");
4862 }
4863
4864 /// The model the schedule was chosen with, which is a receipt anybody comparing two runs of a
4865 /// benchmark needs: two numbers that disagree are usually two models and not two compilers.
4866 #[test]
4867 fn print_config_names_the_model_the_schedule_was_chosen_with() {
4868 let opts = Options::new("x86_64-unknown-linux-gnu".parse().unwrap());
4869 let text = print_config(&opts);
4870 let line = text.lines().find(|l| l.starts_with("timing-model:")).expect("the model");
4871 assert!(line.contains("Skylake"), "{line}");
4872 assert!(line.contains("published"), "a sentence saying where it came from: {line}");
4873 }
4874
4875 #[test]
4876 fn print_config_has_one_key_per_line_and_a_fixed_order() {
4877 let opts = Options::new("x86_64-unknown-linux-gnu".parse().unwrap());
4878 let text = print_config(&opts);
4879 let keys: Vec<&str> =
4880 text.lines().map(|l| l.split(':').next().unwrap_or_default()).collect();
4881 assert_eq!(keys[0], "version");
4882 assert_eq!(keys[1], "target");
4883 assert_eq!(keys.len(), 26);
4884 assert!(text.ends_with('\n'));
4885 }
4886
4887 #[test]
4888 fn the_safety_tier_is_read_off_the_command_line_and_a_wrong_one_is_refused() {
4889 let (opts, _) = compile(&["a.c"]);
4890 assert_eq!(opts.safety, rucc_session::Safety::Off);
4891
4892 for (flag, tier) in [
4893 ("-fsafety=detect", rucc_session::Safety::Detect),
4894 ("-fsafety=enforce", rucc_session::Safety::Enforce),
4895 ("-fsafety=kernel", rucc_session::Safety::Kernel),
4896 ("-fsafety=off", rucc_session::Safety::Off),
4897 ] {
4898 let (opts, _) = compile(&[flag, "a.c"]);
4899 assert_eq!(opts.safety, tier, "{flag}");
4900 }
4901
4902 // The last one wins, the way every other repeated flag on this command line does.
4903 let (opts, _) = compile(&["-fsafety=enforce", "-fsafety=off", "a.c"]);
4904 assert_eq!(opts.safety, rucc_session::Safety::Off);
4905
4906 // A misspelled tier is refused rather than ignored. Silently compiling without the
4907 // monitor a build asked for is the one failure mode this feature cannot have.
4908 let e = parse_args(&args(&["-fsafety=on", "a.c"])).unwrap_err();
4909 assert!(e.message.contains("is not a safety tier"), "{}", e.message);
4910 assert!(parse_args(&args(&["-fsafety", "a.c"])).is_err());
4911 }
4912
4913 /// A checked mode on a Windows target is refused at the link, with a message that says why,
4914 /// rather than left to fail there on names the runtime would have defined. The object is
4915 /// still built, and every other target links as before.
4916 #[test]
4917 fn a_checked_mode_on_windows_is_refused_at_the_link_and_nowhere_else() {
4918 let (opts, _) = compile(&["--target=x86_64-windows-gnu", "-fsafety=detect", "a.c"]);
4919 let why = unlinkable(&opts).expect("a refusal");
4920 assert!(why.contains("not available on a Windows target"), "{why}");
4921 let (opts, _) = compile(&["--target=x86_64-windows-gnu", "-fsafety=off", "a.c"]);
4922 assert_eq!(unlinkable(&opts), None);
4923 let (opts, _) = compile(&["--target=x86_64-linux-gnu", "-fsafety=detect", "a.c"]);
4924 assert_eq!(unlinkable(&opts), None);
4925 }
4926
4927 #[test]
4928 fn the_padding_mode_is_read_off_the_command_line_and_a_wrong_one_is_refused() {
4929 // The default is the one section 9.3 of document 09 gives library code, which is that
4930 // padding does not participate, so a record filled a member at a time is not reported.
4931 let (opts, _) = compile(&["a.c"]);
4932 assert_eq!(opts.padding, rucc_session::Padding::Ignored);
4933
4934 let (opts, _) = compile(&["-fsafety=detect", "-fsafety-init=padding", "a.c"]);
4935 assert_eq!(opts.padding, rucc_session::Padding::Tracked);
4936
4937 let (opts, _) = compile(&["-fsafety-init=padding", "-fsafety-init=nopadding", "a.c"]);
4938 assert_eq!(opts.padding, rucc_session::Padding::Ignored);
4939
4940 // The tier is still a tier. A flag whose name starts the same way must not be eaten by
4941 // the one above it, which is the thing worth pinning about a pair of names like these.
4942 let (opts, _) = compile(&["-fsafety-init=padding", "a.c"]);
4943 assert_eq!(opts.safety, rucc_session::Safety::Off);
4944
4945 let e = parse_args(&args(&["-fsafety-init=some", "a.c"])).unwrap_err();
4946 assert!(e.message.contains("is not a padding mode"), "{}", e.message);
4947 }
4948
4949 #[test]
4950 fn whether_a_write_has_to_stay_inside_its_member_is_read_off_the_command_line() {
4951 // Off by default, because a store to allocated storage sets its effective type and C 6.5
4952 // lets a program reuse a buffer as something else. Row S4 is a build opting out of that.
4953 let (opts, _) = compile(&["a.c"]);
4954 assert_eq!(opts.subobject, rucc_session::Subobject::Off);
4955
4956 let (opts, _) = compile(&["-fsafety=detect", "-fsafety-subobject", "a.c"]);
4957 assert_eq!(opts.subobject, rucc_session::Subobject::Members);
4958
4959 let (opts, _) = compile(&["-fsafety-subobject", "-fno-safety-subobject", "a.c"]);
4960 assert_eq!(opts.subobject, rucc_session::Subobject::Off);
4961
4962 // It takes no value. The form that would take one is the strict reading of section 9.4,
4963 // which is not written yet, so say so rather than accept a spelling that does nothing.
4964 let e = parse_args(&args(&["-fsafety-subobject=strict", "a.c"])).unwrap_err();
4965 assert!(e.message.contains("tamnd/rucc#967"), "{}", e.message);
4966 }
4967
4968 #[test]
4969 fn whether_two_restrict_pointers_may_meet_is_read_off_the_command_line() {
4970 // Off by default, because the record a block keeps is the union of what each pointer
4971 // reached, so two pointers striding through one array without landing on the same byte are
4972 // reported and by the letter of the standard those are different objects. Row Y8 is a build
4973 // deciding it would rather know.
4974 let (opts, _) = compile(&["a.c"]);
4975 assert_eq!(opts.promise, rucc_session::Promise::Off);
4976
4977 let (opts, _) = compile(&["-fsafety=detect", "-fsafety-restrict", "a.c"]);
4978 assert_eq!(opts.promise, rucc_session::Promise::Blocks);
4979
4980 let (opts, _) = compile(&["-fsafety-restrict", "-fno-safety-restrict", "a.c"]);
4981 assert_eq!(opts.promise, rucc_session::Promise::Off);
4982
4983 // The tier is still a tier, which is the thing worth pinning about a pair of names where
4984 // one is the front of the other.
4985 let (opts, _) = compile(&["-fsafety-restrict", "a.c"]);
4986 assert_eq!(opts.safety, rucc_session::Safety::Off);
4987
4988 let e = parse_args(&args(&["-fsafety-restrict=blocks", "a.c"])).unwrap_err();
4989 assert!(e.message.contains("takes no value"), "{}", e.message);
4990 }
4991
4992 #[test]
4993 fn safety_races_takes_a_mode_and_defaults_to_watching_nothing() {
4994 // Three modes rather than a bare flag, because section 9.5 gives two answers that record
4995 // the same thing and report different classes, so a flag with no value could not say which
4996 // was wanted. Off by default for the reason on `rucc_session::Races`, which is not a cost
4997 // argument: this is the one plane where an edge nobody interposed costs a false report.
4998 let (opts, _) = compile(&["a.c"]);
4999 assert_eq!(opts.races, rucc_session::Races::Off);
5000
5001 let (opts, _) = compile(&["-fsafety-races=metadata", "a.c"]);
5002 assert_eq!(opts.races, rucc_session::Races::Metadata);
5003
5004 let (opts, _) = compile(&["-fsafety-races=pointer", "a.c"]);
5005 assert_eq!(opts.races, rucc_session::Races::Pointer);
5006
5007 // Last one wins, as it does for every other mode flag here.
5008 let (opts, _) = compile(&["-fsafety-races=pointer", "-fno-safety-races", "a.c"]);
5009 assert_eq!(opts.races, rucc_session::Races::Off);
5010
5011 let e = parse_args(&args(&["-fsafety-races=all", "a.c"])).unwrap_err();
5012 assert!(e.message.contains("off, metadata or pointer"), "{}", e.message);
5013 }
5014
5015 #[test]
5016 fn print_pipeline_answers_with_the_passes_the_level_asked_for() {
5017 let a = parse_args(&args(&["--print-pipeline", "-O2"])).unwrap();
5018 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
5019 let text = print_pipeline(&opts);
5020 assert!(text.starts_with("level: -O2\n"), "{text}");
5021 assert!(text.contains("fold"), "{text}");
5022
5023 let a = parse_args(&args(&["--print-pipeline"])).unwrap();
5024 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
5025 // Two passes run at `-O0` and neither is an optimization. The first moves what
5026 // `__builtin_expect` said onto the branch and takes the instruction away, so that nothing
5027 // past the optimizer has to know the instruction exists. The second removes code nothing
5028 // reaches. See issue 359.
5029 assert!(print_pipeline(&opts).contains("1: expect,"), "{}", print_pipeline(&opts));
5030 assert!(print_pipeline(&opts).contains("2: simplify-cfg,"), "{}", print_pipeline(&opts));
5031
5032 let a = parse_args(&args(&["--print-pipeline", "-fno-simplify-cfg"])).unwrap();
5033 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
5034 // The second turns off and the first does not, because nothing below the optimizer lowers
5035 // what it removes, so `-fno-expect` is a compile that stops rather than one that runs.
5036 let text = print_pipeline(&opts);
5037 assert!(text.contains("1: expect,"), "{text}");
5038 assert!(!text.contains("simplify-cfg"), "{text}");
5039 }
5040
5041 #[test]
5042 fn print_pipeline_takes_the_toggles_into_account() {
5043 let a = parse_args(&args(&["--print-pipeline", "-O2", "-fno-fold"])).unwrap();
5044 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
5045 let text = print_pipeline(&opts);
5046 // The one that was named is gone and the rest of the level is not, which is the whole
5047 // of what a toggle promises.
5048 assert!(!text.contains("fold"), "{text}");
5049 assert!(text.contains("dce"), "{text}");
5050
5051 // Every pass the compiler has, named off. Built from the registry rather than written
5052 // out, so a pass added later is turned off here too and this keeps testing the thing it
5053 // is about, which is that the toggles can empty a level down to the passes that are not
5054 // optional. Those are named, because a listing that is all of them is a level nobody
5055 // emptied and the assertion would pass while saying nothing.
5056 let mut off = vec!["--print-pipeline".to_owned(), "-O2".to_owned()];
5057 off.extend(rucc_opt::PASSES.iter().map(|p| format!("-fno-{}", p.name())));
5058 let spelled: Vec<&str> = off.iter().map(String::as_str).collect();
5059 let a = parse_args(&args(&spelled)).unwrap();
5060 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
5061 let text = print_pipeline(&opts);
5062 let left: Vec<&str> =
5063 rucc_opt::PASSES.iter().filter(|p| p.required()).map(|p| p.name()).collect();
5064 assert_eq!(left, vec!["expect", "constant-p"], "{text}");
5065 for (at, name) in left.iter().enumerate() {
5066 assert!(text.contains(&format!("{}: {name},", at + 1)), "{text}");
5067 }
5068 assert!(!text.contains("dce"), "{text}");
5069 }
5070
5071 #[test]
5072 fn print_pipeline_says_when_a_budget_will_stop_the_run_short() {
5073 let a = parse_args(&args(&["--print-pipeline", "-O2"])).unwrap();
5074 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
5075 assert!(!print_pipeline(&opts).contains("global fuel"));
5076
5077 let a = parse_args(&args(&["--print-pipeline", "-O2", "-fpass-fuel-global=4"])).unwrap();
5078 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
5079 let text = print_pipeline(&opts);
5080 // Because the listing is the answer to what this compilation will do, and a run that
5081 // stops after four rewrites is not doing what the level says it does.
5082 assert!(text.contains("global fuel: 4"), "{text}");
5083 }
5084
5085 /// A pass is turned on and off by its own name, and the order the flags were given in is
5086 /// kept, because the last spelling of a name is the one that decides.
5087 #[test]
5088 fn a_pass_is_named_by_dash_f_and_unnamed_by_dash_f_no() {
5089 let (opts, _) = compile(&["-c", "-O0", "-ffold", "-fno-fold", "-ffold", "a.c"]);
5090 assert_eq!(
5091 opts.passes,
5092 [("fold".to_owned(), true), ("fold".to_owned(), false), ("fold".to_owned(), true)]
5093 );
5094
5095 let e = parse_args(&args(&["-fno-such-pass", "a.c"])).unwrap_err();
5096 assert!(e.message.contains("unknown option"), "{}", e.message);
5097 }
5098
5099 #[test]
5100 fn pass_fuel_names_a_pass_and_a_count_and_refuses_anything_else() {
5101 let (opts, _) = compile(&["-c", "-O2", "-fpass-fuel=fold=3", "a.c"]);
5102 assert_eq!(opts.pass_fuel, [("fold".to_owned(), 3)]);
5103
5104 let e = parse_args(&args(&["-fpass-fuel=fold", "a.c"])).unwrap_err();
5105 assert!(e.message.contains("<pass>=<count>"), "{}", e.message);
5106 let e = parse_args(&args(&["-fpass-fuel=nosuch=3", "a.c"])).unwrap_err();
5107 assert!(e.message.contains("--print-pipeline"), "{}", e.message);
5108 let e = parse_args(&args(&["-fpass-fuel=fold=lots", "a.c"])).unwrap_err();
5109 assert!(e.message.contains("not a number"), "{}", e.message);
5110 }
5111
5112 #[test]
5113 fn global_pass_fuel_is_a_count_on_its_own_and_defaults_to_no_limit() {
5114 let (opts, _) = compile(&["-c", "-O2", "a.c"]);
5115 assert_eq!(opts.pass_fuel_global, None);
5116
5117 let (opts, _) = compile(&["-c", "-O2", "-fpass-fuel-global=12", "a.c"]);
5118 assert_eq!(opts.pass_fuel_global, Some(12));
5119 // And it is not the per pass flag with a longer name, so neither spelling swallows the
5120 // other.
5121 assert!(opts.pass_fuel.is_empty());
5122
5123 let e = parse_args(&args(&["-fpass-fuel-global=lots", "a.c"])).unwrap_err();
5124 assert!(e.message.contains("not a number"), "{}", e.message);
5125 }
5126
5127 #[test]
5128 fn the_trace_file_is_taken_from_the_flag_and_an_empty_one_is_refused() {
5129 let (opts, _) = compile(&["-c", "a.c"]);
5130 assert_eq!(opts.trace, None);
5131 let (opts, _) = compile(&["-c", "-frucc-trace=/tmp/compile.jsonl", "a.c"]);
5132 assert_eq!(opts.trace.as_deref(), Some("/tmp/compile.jsonl"));
5133 let e = parse_args(&args(&["-frucc-trace=", "a.c"])).unwrap_err();
5134 assert!(e.message.contains("needs a file"), "{}", e.message);
5135 }
5136
5137 #[test]
5138 fn a_gate_names_a_pass_and_optionally_the_functions_it_covers() {
5139 let (opts, _) = compile(&["-c", "-O2", "-fdisable-fold", "-fenable-fold=2-4,main", "a.c"]);
5140 assert_eq!(
5141 opts.pass_gates,
5142 [(false, "fold".to_owned()), (true, "fold=2-4,main".to_owned())],
5143 "the order is what decides, so it has to survive the parse"
5144 );
5145
5146 let e = parse_args(&args(&["-fdisable-nosuch", "a.c"])).unwrap_err();
5147 assert!(e.message.contains("--print-pipeline"), "{}", e.message);
5148 let e = parse_args(&args(&["-fenable-fold=9-2", "a.c"])).unwrap_err();
5149 assert!(e.message.contains("ends before it starts"), "{}", e.message);
5150 let e = parse_args(&args(&["-fdisable-fold=", "a.c"])).unwrap_err();
5151 assert!(e.message.contains("is empty"), "{}", e.message);
5152 }
5153
5154 #[test]
5155 fn the_pipeline_listing_says_which_passes_a_gate_touched() {
5156 let (opts, _) = compile(&["-c", "-O2", "-fdisable-fold=main", "a.c"]);
5157 let text = print_pipeline(&opts);
5158 assert!(text.contains("fold, "), "{text}");
5159 assert!(text.contains("[off for main]"), "{text}");
5160 }
5161
5162 /// The spelling is checked while the arguments are read, because a dump that names a pass
5163 /// this compiler does not have is a typo, and a typo found after the compilation has run is
5164 /// found too late to be any use.
5165 #[test]
5166 fn a_dump_is_checked_when_it_is_asked_for_rather_than_when_it_is_taken() {
5167 let (opts, _) = compile(&["-c", "-O2", "-fdump-ir=all", "-fdump-ir=after-fold", "a.c"]);
5168 assert_eq!(opts.dump_ir, ["all", "after-fold"]);
5169
5170 let e = parse_args(&args(&["-fdump-ir=after-nosuch", "a.c"])).unwrap_err();
5171 assert!(e.message.contains("nosuch"), "{}", e.message);
5172 assert!(parse_args(&args(&["-fdump-ir=sideways-fold", "a.c"])).is_err());
5173 }
5174
5175 /// Every spelling `-fopt-info` takes, and the one it does not.
5176 ///
5177 /// The keywords are checked here for the same reason a dump's pass name is: a person who
5178 /// misspelled one gets no output, and no output is also what a compilation where nothing
5179 /// happened looks like. Telling those two apart is the entire reason to reach for this flag.
5180 #[test]
5181 fn opt_info_takes_kinds_and_a_file_and_refuses_a_kind_it_does_not_have() {
5182 let (opts, _) = compile(&["-c", "-O2", "-fopt-info", "a.c"]);
5183 assert_eq!(opts.opt_info, [""], "a bare flag asks for the rewrites");
5184 assert_eq!(opts.opt_info_file, None, "and goes to standard error");
5185
5186 let (opts, _) = compile(&["-c", "-O2", "-fopt-info-missed-note", "a.c"]);
5187 assert_eq!(opts.opt_info, ["missed-note"]);
5188
5189 // Two flags add up rather than the second replacing the first, and the file is the last
5190 // one that named a file, which is how GCC treats both.
5191 let (opts, _) =
5192 compile(&["-c", "-O2", "-fopt-info-missed=one.txt", "-fopt-info-all=two.txt", "a.c"]);
5193 assert_eq!(opts.opt_info, ["missed", "all"]);
5194 assert_eq!(opts.opt_info_file.as_deref(), Some("two.txt"));
5195
5196 let e = parse_args(&args(&["-fopt-info-vectorized", "a.c"])).unwrap_err();
5197 assert!(e.message.contains("vectorized"), "{}", e.message);
5198 assert!(e.message.contains("`missed`"), "{}", e.message);
5199 let e = parse_args(&args(&["-fopt-info-missed=", "a.c"])).unwrap_err();
5200 assert!(e.message.contains("no file"), "{}", e.message);
5201 }
5202
5203 #[test]
5204 fn verify_each_is_unstable_and_off_unless_it_was_asked_for() {
5205 let (opts, _) = compile(&["-c", "-Zverify-each", "a.c"]);
5206 assert!(opts.verify_each);
5207 assert!(!USAGE.contains("verify-each"), "an unstable option stays out of the usage text");
5208 }
5209
5210 #[test]
5211 fn dash_o_needs_an_argument() {
5212 let e = parse_args(&args(&["a.c", "-o"])).unwrap_err();
5213 assert_eq!(e.message, "-o requires an argument");
5214 }
5215
5216 #[test]
5217 fn dash_d_and_dash_u_are_read_joined_or_separated_and_keep_their_order() {
5218 let (opts, _) = compile(&["-DFOO=1", "-D", "BAR", "-UBAZ", "-U", "QUX", "a.c"]);
5219 assert_eq!(opts.defines, ["FOO=1", "BAR"]);
5220 assert_eq!(opts.undefines, ["BAZ", "QUX"]);
5221 }
5222
5223 #[test]
5224 fn the_include_flags_land_on_the_chain_each_one_names() {
5225 // A sysroot with nothing under it, so that the library's own directories are the
5226 // same on every machine this test runs on, which is none of them.
5227 let (opts, _) = compile(&[
5228 "-Ii",
5229 "-iquote",
5230 "q",
5231 "-isystem",
5232 "sys",
5233 "-idirafter",
5234 "after",
5235 "--sysroot=/nowhere-at-all",
5236 "a.c",
5237 ]);
5238 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
5239 // The compiler's own headers sit after every `-isystem` and before `-idirafter`,
5240 // which is where GCC puts its own: a directory the user named outranks ours.
5241 assert_eq!(dirs, ["q", "i", "sys", runtime::DIR, "after"]);
5242 assert!(!opts.search.dirs()[1].is_system);
5243 assert!(opts.search.dirs()[2].is_system);
5244 }
5245
5246 #[test]
5247 fn the_librarys_headers_come_after_the_compilers_own_and_go_away_with_them() {
5248 // Which machine this runs on decides what is on the path, so the test is about the
5249 // order rather than about the names: ours is on it, the library's follow it, and
5250 // `-nostdinc` is the one flag that takes both halves of the pair off at once.
5251 let (opts, _) = compile(&["a.c"]);
5252 let dirs = opts.search.dirs();
5253 let ours = dirs.iter().position(|d| d.path.to_str() == Some(runtime::DIR));
5254 assert_eq!(ours, Some(0), "{dirs:?}");
5255 assert!(dirs[1..].iter().all(|d| d.is_system), "{dirs:?}");
5256 let (bare, _) = compile(&["-nostdinc", "a.c"]);
5257 assert!(bare.search.dirs().is_empty(), "{:?}", bare.search.dirs());
5258 }
5259
5260 #[test]
5261 fn a_sysroot_moves_the_librarys_directories_and_nothing_else() {
5262 let (opts, _) = compile(&["-isystem", "sys", "--sysroot=/nowhere-at-all", "a.c"]);
5263 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
5264 assert_eq!(dirs, ["sys", runtime::DIR]);
5265 }
5266
5267 #[test]
5268 fn a_cross_compile_reads_the_targets_own_headers_rather_than_the_ones_next_door() {
5269 // The target is not the machine this test runs on wherever it runs, so the answer is the
5270 // same on all of them: the libc's two include directories for that target, the kernel's
5271 // two, and nothing from here. A header read from here is the quiet failure of section 8.5, a
5272 // program that builds on the build machine and is wrong everywhere else.
5273 let (opts, _) = compile(&["--target=riscv64-linux-musl", "-c", "a.c"]);
5274 let dirs: Vec<&std::path::Path> =
5275 opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
5276 let root = cache::dir().join("sysroots").join("riscv64-linux-musl");
5277 let kernel = cache::dir().join("kernel-headers");
5278 assert_eq!(dirs.len(), 5, "{dirs:?}");
5279 assert_eq!(dirs[0], std::path::Path::new(runtime::DIR));
5280 assert_eq!(dirs[1], root.join("include").join("riscv64"));
5281 assert_eq!(dirs[2], root.join("include").join("generic"));
5282 // The kernel's, which are beside the sysroots rather than inside one, because every target
5283 // that shares an architecture reads the same files.
5284 assert_eq!(dirs[3], kernel.join("riscv"));
5285 assert_eq!(dirs[4], kernel.join("generic"));
5286 }
5287
5288 #[test]
5289 fn a_cross_compile_to_something_that_is_not_linux_reads_no_kernel_headers() {
5290 // The other side of the same answer. Windows has its own system headers and no `linux/` at
5291 // all, so the list is the libc's own and the question never arises, which is the `None` that
5292 // `link::cross_kernel` returns rather than a directory nothing would be found in.
5293 //
5294 // The libc's own is one directory rather than two here, because mingw-w64 publishes a single
5295 // header tree for every architecture and `Sysroot::splits_by_arch` says so.
5296 let (opts, _) = compile(&["--target=x86_64-pc-windows-gnu", "-c", "a.c"]);
5297 let dirs: Vec<&std::path::Path> =
5298 opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
5299 assert_eq!(dirs.len(), 2, "{dirs:?}");
5300 assert!(!dirs.iter().any(|dir| dir.ends_with("kernel-headers")), "{dirs:?}");
5301 }
5302
5303 #[test]
5304 fn the_glibc_version_macro_goes_with_the_bundled_tree_and_with_nothing_else() {
5305 // One tree serves every glibc release, so the release is what the target supplies, and the
5306 // condition is the same one that chose the directories. A host glibc and a tree somebody
5307 // named both define `__GLIBC_MINOR__` in their own `features.h`, and two definitions with
5308 // different values is a warning on every compilation of every file.
5309 //
5310 // The architecture is chosen against this machine's rather than written down, because the
5311 // bundled tree is only in effect for a target that is not this machine. The first version of
5312 // this test said x86_64-linux-gnu, which is a cross compile on a mac and this machine on a
5313 // Linux runner, so it passed here and failed there.
5314 //
5315 // Unless this machine has the distribution's cross packages for it and nothing fetched, and
5316 // then those are the headers and their own `features.h` says the release, as it does for a
5317 // tree somebody named.
5318 let gnu = format!("--target={}-linux-gnu", cross_arch());
5319 let (bundled, _) = compile(&[&gnu, "-c", "a.c"]);
5320 let (link, _) = linking(&[&gnu, "-c", "a.c"]);
5321 let distro = link::distro_cross(bundled.target, &link).is_some();
5322 assert_eq!(bundled.glibc_minor, if distro { None } else { Some(44) });
5323 let pin = format!("{gnu}.2.28");
5324 let (pinned, _) = compile(&[&pin, "-c", "a.c"]);
5325 assert_eq!(pinned.glibc_minor, Some(28));
5326
5327 let (named, _) = compile(&[&gnu, "--sysroot=/nowhere-at-all", "-c", "a.c"]);
5328 assert_eq!(named.glibc_minor, None);
5329 let (none, _) = compile(&[&gnu, "-nostdinc", "-c", "a.c"]);
5330 assert_eq!(none.glibc_minor, None);
5331 let musl = format!("--target={}-linux-musl", cross_arch());
5332 let (musl, _) = compile(&[&musl, "-c", "a.c"]);
5333 assert_eq!(musl.glibc_minor, None);
5334
5335 // And this machine's own target gets nothing, whatever this machine is, because its headers
5336 // come from the machine and its own `features.h` defines the macro. On a glibc Linux box
5337 // that is the case this test had backwards; on a mac it is true for the other reason, which
5338 // is that Darwin is not a glibc target at all.
5339 if let Some(host) = Triple::host() {
5340 let native = format!("--target={}", host.tuple());
5341 let (native, _) = compile(&[&native, "-c", "a.c"]);
5342 assert_eq!(native.glibc_minor, None);
5343 }
5344 }
5345
5346 #[test]
5347 fn a_pinned_release_on_this_machines_own_target_reads_the_bundled_tree() {
5348 // The end to end half of the answer in `link::cross_for`. A release named for this machine's
5349 // own target is a cross compile, so the headers are the bundled tree's and the macro says
5350 // what was asked for rather than what this machine has.
5351 //
5352 // Only on a glibc box, because a release is a glibc release: a mac has no `__GLIBC_MINOR__`
5353 // to get wrong and nothing to pin. That makes this a test the Linux runners carry, which is
5354 // where the case lives.
5355 let Some(host) = Triple::host() else { return };
5356 if host.os != rucc_target::Os::Linux || host.env != rucc_target::Env::Gnu {
5357 return;
5358 }
5359 let pin = format!("--target={}.2.28", host.tuple());
5360 let (opts, _) = compile(&[&pin, "-c", "a.c"]);
5361 assert_eq!(opts.glibc_minor, Some(28));
5362 let root = cache::dir().join("sysroots").join(format!("{}.2.28", host.tuple()));
5363 let dirs: Vec<&std::path::Path> =
5364 opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
5365 assert!(dirs.iter().any(|dir| dir.starts_with(&root)), "{dirs:?}");
5366 // And nothing of this machine's, which is the failure this was: a program compiled against
5367 // 2.44 declarations and told it was 2.28.
5368 assert!(!dirs.iter().any(|dir| *dir == std::path::Path::new("/usr/include")), "{dirs:?}");
5369 }
5370
5371 /// An architecture that is not this machine's, out of the three the driver has targets for.
5372 ///
5373 /// A test about the bundled sysroot has to name a target that is not the host, because a target
5374 /// that is the host reads the host's own headers and libraries. Asking which machine this is
5375 /// beats picking a row and hoping, and it is two lines.
5376 fn cross_arch() -> &'static str {
5377 match Triple::host().map(|host| host.arch) {
5378 Some(rucc_target::Arch::X86_64) => "aarch64",
5379 _ => "x86_64",
5380 }
5381 }
5382
5383 #[test]
5384 fn a_glibc_newer_than_the_bundled_tree_is_refused_by_name() {
5385 // Both versions in the message, because the two things a person can do about it are pin a
5386 // release the tree has and name a sysroot that has the one they asked for, and neither is a
5387 // choice they can make without knowing which release the tree is.
5388 //
5389 // Not this machine's architecture, for the reason the test above gives: the refusal is about
5390 // the bundled tree, and the bundled tree is not what a target that is this machine reads.
5391 let target = format!("--target={}-linux-gnu.2.99", cross_arch());
5392 let message = refused(&[&target, "-c", "a.c"]);
5393 assert!(message.contains("asked for glibc 2.99"), "{message}");
5394 assert!(message.contains("bundled headers are glibc 2.44"), "{message}");
5395 assert!(message.contains("--sysroot"), "{message}");
5396 }
5397
5398 #[test]
5399 fn a_sysroot_the_user_named_is_still_what_a_cross_compile_reads() {
5400 // The tree somebody assembled beats the one we would build, on the headers as on the
5401 // libraries. It is empty here, which is why the list comes out short: the directories under
5402 // it are checked for rather than assumed, and a tree that is not there offers nothing.
5403 let (opts, _) =
5404 compile(&["--target=riscv64-linux-musl", "--sysroot=/nowhere-at-all", "-c", "a.c"]);
5405 let dirs: Vec<&std::path::Path> =
5406 opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
5407 assert_eq!(dirs, [std::path::Path::new(runtime::DIR)]);
5408 }
5409
5410 #[test]
5411 fn dash_i_dash_moves_the_bracket_directories_into_the_quoted_chain() {
5412 let (opts, _) =
5413 compile(&["-Iinc1", "-iquote", "inc2", "-I-", "-Iinc3", "-nostdinc", "a.c"]);
5414 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
5415 assert_eq!(dirs, ["inc1", "inc2", "inc3"]);
5416 // An angled include sees only what came after the flag.
5417 assert_eq!(opts.search.start(IncludeForm::Angled), 2);
5418 assert!(!opts.search.searches_current_dir());
5419 }
5420
5421 #[test]
5422 fn the_prefix_flags_stick_what_iprefix_said_on_the_front_of_what_follows_it() {
5423 let (opts, _) = compile(&[
5424 "-iprefix",
5425 "/tools/",
5426 "-iwithprefix",
5427 "late",
5428 "-iwithprefixbefore",
5429 "early",
5430 "-iprefix",
5431 "/other/",
5432 "-iwithprefix",
5433 "last",
5434 "-nostdinc",
5435 "a.c",
5436 ]);
5437 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
5438 // `-iwithprefixbefore` is an `-I` and the other two are `-isystem`, which is where GCC
5439 // puts them rather than where its manual says it does.
5440 assert_eq!(dirs, ["/tools/early", "/tools/late", "/other/last"]);
5441 assert!(!opts.search.dirs()[0].is_system);
5442 assert!(opts.search.dirs()[1].is_system);
5443 }
5444
5445 #[test]
5446 fn the_files_named_on_the_command_line_keep_their_order_and_which_flag_named_them() {
5447 let (opts, _) =
5448 compile(&["-include", "one.h", "-imacros", "two.h", "-include", "3.h", "a.c"]);
5449 let names: Vec<&str> = opts.preincludes.iter().map(|p| p.name.as_str()).collect();
5450 assert_eq!(names, ["one.h", "two.h", "3.h"]);
5451 assert_eq!(opts.preincludes.iter().filter(|p| p.macros_only).count(), 1);
5452 }
5453
5454 #[test]
5455 fn nostdinc_takes_the_compilers_own_headers_off_the_path() {
5456 let (opts, _) = compile(&["-Ii", "-nostdinc", "a.c"]);
5457 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
5458 assert_eq!(dirs, ["i"]);
5459 }
5460
5461 #[test]
5462 fn the_dialect_flags_set_the_language_and_the_extensions_separately() {
5463 let (opts, _) = compile(&["-std=gnu11", "a.c"]);
5464 assert_eq!(opts.std, Std::C11);
5465 assert!(opts.gnu_extensions);
5466
5467 let (opts, _) = compile(&["-std=iso9899:1999", "a.c"]);
5468 assert_eq!(opts.std, Std::C99);
5469 assert!(!opts.gnu_extensions);
5470
5471 let (opts, _) = compile(&["-ansi", "a.c"]);
5472 assert_eq!(opts.std, Std::C89);
5473 assert!(!opts.gnu_extensions);
5474
5475 let (opts, _) = compile(&["-std=gnu2y", "a.c"]);
5476 assert_eq!(opts.std, Std::C2y);
5477 assert!(opts.gnu_extensions);
5478
5479 let e = parse_args(&args(&["-std=c94jr", "a.c"])).unwrap_err();
5480 assert!(e.message.contains("unknown dialect"), "{}", e.message);
5481 }
5482
5483 #[test]
5484 fn the_dump_letters_are_a_family_and_everything_else_beginning_with_d_is_not() {
5485 let (opts, _) = compile(&["-dM", "a.c"]);
5486 assert!(opts.dumps.macros);
5487
5488 // Packed, the way GCC takes them, and a letter in the family we have not written yet
5489 // is accepted and does nothing rather than failing a build.
5490 let (opts, _) = compile(&["-dDM", "a.c"]);
5491 assert!(opts.dumps.macros);
5492 let (opts, _) = compile(&["-dD", "a.c"]);
5493 assert!(!opts.dumps.macros);
5494
5495 let (opts, _) = compile(&["a.c"]);
5496 assert!(!opts.dumps.any());
5497
5498 // `-dumpversion` is a different flag that happens to start the same way, and it is read
5499 // as itself rather than as a dump of nothing.
5500 assert_eq!(printed(&["-dumpversion", "a.c"]), "16");
5501 }
5502
5503 #[test]
5504 fn the_msvc_runtime_is_a_compile_flag_and_a_link_one() {
5505 let (link, _) = linking(&["a.c"]);
5506 assert_eq!(link.crt, rucc_sysroot::Crt::Static);
5507 let (link, _) = linking(&["-fms-runtime-lib=dll", "a.c"]);
5508 assert_eq!(link.crt, rucc_sysroot::Crt::Dll);
5509 let (opts, _) = compile(&["-fms-runtime-lib=dll", "-c", "a.c"]);
5510 assert!(opts.ms_dll_runtime);
5511 let (opts, _) = compile(&["-fms-runtime-lib=dll", "-fms-runtime-lib=static", "-c", "a.c"]);
5512 assert!(!opts.ms_dll_runtime, "the last one wins");
5513 let e = parse_args(&args(&["-fms-runtime-lib=dll_dbg", "a.c"])).unwrap_err();
5514 assert!(e.message.contains("debug"), "{}", e.message);
5515 let e = parse_args(&args(&["-fms-runtime-lib=shared", "a.c"])).unwrap_err();
5516 assert!(e.message.contains("static or dll"), "{}", e.message);
5517 }
5518
5519 #[test]
5520 fn the_gcc_version_claimed_is_a_flag_and_the_short_spellings_are_the_ones_people_write() {
5521 let (opts, _) = compile(&["a.c"]);
5522 assert_eq!(
5523 opts.gnuc,
5524 GnucVersion { major: 16, minor: 0, patch: 0 },
5525 "the release this compiler is written against, and the earliest one of that series"
5526 );
5527
5528 let (opts, _) = compile(&["-fgnuc-version=15.1.0", "a.c"]);
5529 assert_eq!(opts.gnuc, GnucVersion { major: 15, minor: 1, patch: 0 });
5530
5531 // A missing component is zero. `gcc -dumpversion` says `15` on a release with no
5532 // patchlevel and a harness that pastes that back has to be understood.
5533 let (opts, _) = compile(&["-fgnuc-version=15", "a.c"]);
5534 assert_eq!(opts.gnuc, GnucVersion { major: 15, minor: 0, patch: 0 });
5535
5536 assert!(opts.gnuc_given, "a version that was written down is one that was given");
5537 assert!(!compile(&["a.c"]).0.gnuc_given);
5538
5539 let (opts, _) = compile(&["-fms-compatibility-version=19.29.30133", "a.c"]);
5540 assert_eq!(opts.msc.msc_ver(), 1929);
5541 assert_eq!(opts.msc.msc_full_ver(), 192_930_133);
5542
5543 let (opts, _) = compile(&["-fgnuc-version=13.2", "a.c"]);
5544 assert_eq!(opts.gnuc, GnucVersion { major: 13, minor: 2, patch: 0 });
5545
5546 let e = parse_args(&args(&["-fgnuc-version=15.x", "a.c"])).unwrap_err();
5547 assert!(e.message.contains("minor that is not a number"), "{}", e.message);
5548
5549 let e = parse_args(&args(&["-fgnuc-version=1.2.3.4", "a.c"])).unwrap_err();
5550 assert!(e.message.contains("more than three"), "{}", e.message);
5551 }
5552
5553 #[test]
5554 fn pedantic_has_two_spellings_and_is_not_the_same_knob_as_the_dialect() {
5555 let (opts, _) = compile(&["-std=c17", "-pedantic", "a.c"]);
5556 assert!(opts.pedantic);
5557 assert_eq!(opts.std, Std::C17);
5558
5559 // The `-W` family's name for it, which is what a build that groups its warning flags
5560 // tends to write.
5561 let (opts, _) = compile(&["-Wpedantic", "a.c"]);
5562 assert!(opts.pedantic);
5563
5564 let (opts, _) = compile(&["-std=c17", "a.c"]);
5565 assert!(!opts.pedantic, "a dialect on its own does not diagnose an extension");
5566 }
5567
5568 #[test]
5569 fn dash_p_and_dash_ffreestanding_reach_the_options() {
5570 let (opts, _) = compile(&["-E", "-P", "-ffreestanding", "a.c"]);
5571 assert!(!opts.line_markers);
5572 assert!(!opts.hosted);
5573 assert_eq!(opts.emit, EmitKind::Preprocessed);
5574 }
5575
5576 /// The two ways a build says it means its own function by a name the C library also has.
5577 ///
5578 /// `-fno-builtin` is all of them and `-fno-builtin-<name>` is one, and the second is what a
5579 /// build writes when it means its own `memcpy` and the library's everything else. The name is
5580 /// kept as it was written and not checked against anything, because a program is allowed to
5581 /// mean something by a name this compiler has never heard of.
5582 #[test]
5583 fn the_builtin_flags_are_read_in_both_directions_and_one_name_at_a_time() {
5584 let (opts, _) = compile(&["-c", "a.c"]);
5585 assert!(opts.builtins, "a library name means the library function by default");
5586 assert!(opts.no_builtin.is_empty());
5587
5588 let (opts, _) = compile(&["-c", "-fno-builtin", "a.c"]);
5589 assert!(!opts.builtins);
5590
5591 let (opts, _) = compile(&["-c", "-fno-builtin", "-fbuiltin", "a.c"]);
5592 assert!(opts.builtins, "the last mention decides");
5593
5594 let (opts, _) = compile(&["-c", "-fno-builtin-memcpy", "-fno-builtin-nonesuch", "a.c"]);
5595 assert!(opts.builtins, "one name is not the family");
5596 assert_eq!(opts.no_builtin, vec!["memcpy".to_owned(), "nonesuch".to_owned()]);
5597 }
5598
5599 /// `-fvisibility=`, which is on every cmake project that cares about which names it exports
5600 /// and which was refused as an unknown option until now.
5601 ///
5602 /// Four spellings and three answers. `internal` is hidden plus a promise about never taking
5603 /// the address across a component boundary, and nothing derives anything from that promise
5604 /// here, so it comes out as the weaker of the two rather than as a refusal that stops a build
5605 /// over a distinction this compiler does not make.
5606 #[test]
5607 fn visibility_takes_the_four_spellings_gcc_takes_and_refuses_the_rest() {
5608 let (opts, _) = compile(&["-c", "a.c"]);
5609 assert_eq!(opts.visibility, Visibility::Default, "exported unless something says not");
5610
5611 for (written, wanted) in [
5612 ("default", Visibility::Default),
5613 ("hidden", Visibility::Hidden),
5614 ("internal", Visibility::Hidden),
5615 ("protected", Visibility::Protected),
5616 ] {
5617 let (opts, _) = compile(&["-c", &format!("-fvisibility={written}"), "a.c"]);
5618 assert_eq!(opts.visibility, wanted, "{written}");
5619 }
5620
5621 // The last mention decides, which is what every other flag of this shape does and what a
5622 // build that turns something off for one directory relies on.
5623 let (opts, _) = compile(&["-c", "-fvisibility=hidden", "-fvisibility=default", "a.c"]);
5624 assert_eq!(opts.visibility, Visibility::Default, "the last mention decides");
5625
5626 // A spelling gcc does not take is refused rather than read as the default, because a
5627 // build that meant hidden and got exported is a library with the wrong interface and
5628 // nothing said about it anywhere.
5629 let failed = parse_args(&args(&["-fvisibility=none", "a.c"])).expect_err("refused");
5630 assert!(failed.to_string().contains("is not a visibility"), "{failed}");
5631 }
5632
5633 /// `-ffp-contract=`, which is the one flag in the floating point group that is kept rather than
5634 /// described, and the values are gcc 16's three.
5635 #[test]
5636 fn how_far_a_multiply_and_an_addition_may_be_fused_is_asked_for() {
5637 let (opts, _) = compile(&["-c", "a.c"]);
5638 assert_eq!(opts.fp_contract, Contract::Off, "a licence nobody granted is not assumed");
5639
5640 for (written, wanted) in
5641 [("off", Contract::Off), ("on", Contract::On), ("fast", Contract::Fast)]
5642 {
5643 let (opts, _) = compile(&["-c", &format!("-ffp-contract={written}"), "a.c"]);
5644 assert_eq!(opts.fp_contract, wanted, "{written}");
5645 }
5646
5647 let (opts, _) = compile(&["-c", "-ffp-contract=fast", "-ffp-contract=off", "a.c"]);
5648 assert_eq!(opts.fp_contract, Contract::Off, "the last mention decides");
5649
5650 // Refused rather than read as one of the three, because a build that asked for no fusing
5651 // and was given the default would be one whose numbers change and whose command line says
5652 // they should not. gcc refuses the same spellings and names the same three in its message.
5653 for bad in ["-ffp-contract=none", "-ffp-contract=", "-ffp-contract=Fast"] {
5654 let failed = parse_args(&args(&[bad, "a.c"])).expect_err("refused");
5655 assert!(failed.to_string().contains("is not a contraction"), "{bad}: {failed}");
5656 }
5657
5658 // And the other one that takes a value, which is taken and kept nowhere: every operation
5659 // here is computed in the type it was written in, so `standard` is what happens and the
5660 // other two are permission to do something this does not do.
5661 let failed = parse_args(&args(&["-fexcess-precision=long", "a.c"])).expect_err("refused");
5662 assert!(failed.to_string().contains("is not an excess precision"), "{failed}");
5663 }
5664
5665 /// The four prefix mapping flags, which are what a distribution passes to get the same bytes
5666 /// out of `/build/pkg-1.2` and out of `/home/someone/pkg-1.2`. Three lists rather than one
5667 /// because gcc has three, and `-ffile-prefix-map=` is the three of them at once.
5668 #[test]
5669 fn a_prefix_mapping_flag_goes_on_the_list_its_spelling_names() {
5670 let (opts, _) = compile(&["-c", "a.c"]);
5671 assert!(opts.prefix_map.macros.is_empty(), "nothing is rewritten unless it is asked for");
5672 assert!(opts.prefix_map.debug.is_empty(), "nor here");
5673 assert!(opts.prefix_map.profile.is_empty(), "nor here");
5674
5675 let (opts, _) = compile(&["-c", "-fmacro-prefix-map=/build=.", "a.c"]);
5676 assert_eq!(opts.prefix_map.macros.apply("/build/a.c"), "./a.c", "the one it names");
5677 assert!(opts.prefix_map.debug.is_empty(), "and not the two it does not");
5678
5679 let (opts, _) = compile(&["-c", "-fdebug-prefix-map=/build=.", "a.c"]);
5680 assert_eq!(opts.prefix_map.debug.apply("/build/a.c"), "./a.c", "the one it names");
5681 assert!(opts.prefix_map.macros.is_empty(), "and not the two it does not");
5682
5683 let (opts, _) = compile(&["-c", "-fprofile-prefix-map=/build=.", "a.c"]);
5684 assert_eq!(opts.prefix_map.profile.apply("/build/a.c"), "./a.c", "the one it names");
5685 assert!(opts.prefix_map.macros.is_empty(), "and not the two it does not");
5686
5687 let (opts, _) = compile(&["-c", "-ffile-prefix-map=/build=.", "a.c"]);
5688 for list in [&opts.prefix_map.macros, &opts.prefix_map.debug, &opts.prefix_map.profile] {
5689 assert_eq!(list.apply("/build/a.c"), "./a.c", "all three at once");
5690 }
5691
5692 // Every mention is kept and the last one that matches wins, unlike the flags above whose
5693 // last mention replaces the earlier ones. A build writes one of these per source root and
5694 // expects all of them to be in force, which is the whole point of a list.
5695 let (opts, _) =
5696 compile(&["-c", "-ffile-prefix-map=/a=one", "-ffile-prefix-map=/b=two", "a.c"]);
5697 assert_eq!(opts.prefix_map.macros.apply("/a/x.c"), "one/x.c", "the earlier one still acts");
5698 assert_eq!(opts.prefix_map.macros.apply("/b/x.c"), "two/x.c", "and so does the later one");
5699
5700 // An argument with no `=` is refused rather than ignored, because a build whose paths were
5701 // meant to be rewritten and were not is one that ships the build directory's name and says
5702 // nothing about it. gcc refuses the same thing.
5703 for bad in ["-fmacro-prefix-map=nope", "-ffile-prefix-map=", "-fdebug-prefix-map=/build"] {
5704 let failed = parse_args(&args(&[bad, "a.c"])).expect_err("refused");
5705 assert!(failed.to_string().contains("is not a rewrite for"), "{bad}: {failed}");
5706 }
5707 }
5708
5709 /// `-ffunction-sections` and `-fdata-sections`, which are what make `--gc-sections` able to
5710 /// drop anything: a linker can leave out a section nothing reaches and cannot leave out half of
5711 /// one. A kernel and an embedded image are both linked that way.
5712 ///
5713 /// Two flags rather than one because gcc has two, and a build that asks for one of them and not
5714 /// the other is a build that measured something: splitting the code is nearly free at link time
5715 /// and splitting the data can defeat the linker's ordering of what is next to what.
5716 #[test]
5717 fn a_section_per_function_and_a_section_per_variable_are_asked_for_one_at_a_time() {
5718 let (opts, _) = compile(&["-c", "a.c"]);
5719 assert!(!opts.function_sections, "one text section unless something says otherwise");
5720 assert!(!opts.data_sections);
5721
5722 let (opts, _) = compile(&["-c", "-ffunction-sections", "a.c"]);
5723 assert!(opts.function_sections);
5724 assert!(!opts.data_sections, "one flag is not the other");
5725
5726 let (opts, _) = compile(&["-c", "-fdata-sections", "a.c"]);
5727 assert!(opts.data_sections);
5728 assert!(!opts.function_sections);
5729
5730 // Both directions taken, and the off one is what happens anyway rather than a refusal,
5731 // since a build that writes it is asking for the default.
5732 let (opts, _) = compile(&[
5733 "-c",
5734 "-ffunction-sections",
5735 "-fno-function-sections",
5736 "-fdata-sections",
5737 "-fno-data-sections",
5738 "a.c",
5739 ]);
5740 assert!(!opts.function_sections, "the last mention decides");
5741 assert!(!opts.data_sections, "the last mention decides");
5742 }
5743
5744 /// `-fgnu89-inline`, which is off by default and is not implied by anything on the command
5745 /// line, since the dialect asks for GNU's reading further in rather than through this.
5746 #[test]
5747 fn gnu89_inline_is_off_until_it_is_asked_for_and_the_last_mention_decides() {
5748 let (opts, _) = compile(&["-c", "a.c"]);
5749 assert!(!opts.gnu89_inline, "C's reading of inline by default");
5750
5751 let (opts, _) = compile(&["-c", "-fgnu89-inline", "a.c"]);
5752 assert!(opts.gnu89_inline);
5753
5754 let (opts, _) = compile(&["-c", "-fgnu89-inline", "-fno-gnu89-inline", "a.c"]);
5755 assert!(!opts.gnu89_inline, "the last mention decides");
5756
5757 // The C89 dialects are under GNU's reading whether this was written or not, so the flag
5758 // stays off there and the dialect is what the checker and the macro set both ask. That is
5759 // also why `-std=c89 -fno-gnu89-inline` needs no diagnostic: it asks for the reading the
5760 // dialect already has. gcc refuses that command line, which is measured in the issue.
5761 let (opts, _) = compile(&["-c", "-std=c89", "a.c"]);
5762 assert!(!opts.gnu89_inline);
5763 }
5764
5765 /// Both spellings of both frame flags, since a build that wants one usually writes the
5766 /// other beside it for the one file that has to be compiled the ordinary way.
5767 #[test]
5768 fn the_two_frame_flags_are_read_in_both_directions() {
5769 let (opts, _) = compile(&["-c", "a.c"]);
5770 assert_eq!(opts.frame_pointer, None, "nothing said, so the level decides");
5771 assert!(opts.keeps_frame_pointer(), "and at -O0 gcc keeps one, so this does too");
5772 let (opts, _) = compile(&["-c", "-O1", "a.c"]);
5773 assert!(!opts.keeps_frame_pointer(), "gcc omits it above -O0 and so does this");
5774 assert!(opts.red_zone, "the psABI has one and nothing said not to use it");
5775
5776 let (opts, _) = compile(&["-c", "-fno-omit-frame-pointer", "-mno-red-zone", "a.c"]);
5777 assert_eq!(opts.frame_pointer, Some(true));
5778 assert!(!opts.red_zone);
5779
5780 let (opts, _) = compile(&[
5781 "-c",
5782 "-fno-omit-frame-pointer",
5783 "-fomit-frame-pointer",
5784 "-mno-red-zone",
5785 "-mred-zone",
5786 "a.c",
5787 ]);
5788 assert_eq!(opts.frame_pointer, Some(false), "the last one wins, as it does in gcc");
5789 assert!(!opts.keeps_frame_pointer(), "and it wins over the level too");
5790 assert!(opts.red_zone);
5791 }
5792
5793 /// Four flags rather than one with an argument, which is how gcc spells them, and the negative
5794 /// spelled three ways because a build that turns one off writes whichever it turned on.
5795 #[test]
5796 fn the_stack_protector_is_four_flags_and_the_last_one_wins() {
5797 let (opts, _) = compile(&["-c", "a.c"]);
5798 assert_eq!(opts.protector, Protector::None, "gcc protects nothing unless it was asked");
5799
5800 for (flag, want) in [
5801 ("-fstack-protector", Protector::Buffers),
5802 ("-fstack-protector-strong", Protector::Strong),
5803 ("-fstack-protector-all", Protector::All),
5804 ] {
5805 let (opts, _) = compile(&["-c", flag, "a.c"]);
5806 assert_eq!(opts.protector, want, "{flag}");
5807 }
5808
5809 // What a package build does: the strong one in the global flags and one directory that
5810 // cannot have a protector turning it off on the line after.
5811 for off in ["-fno-stack-protector", "-fno-stack-protector-strong"] {
5812 let (opts, _) = compile(&["-c", "-fstack-protector-strong", off, "a.c"]);
5813 assert_eq!(opts.protector, Protector::None, "{off}");
5814 }
5815 let (opts, _) = compile(&["-c", "-fno-stack-protector", "-fstack-protector-all", "a.c"]);
5816 assert_eq!(opts.protector, Protector::All, "the last one wins either way round");
5817 }
5818
5819 /// A switch rather than a level, because how a frame is taken is one question and which
5820 /// functions get a canary is another, and gcc spells it that way for the same reason.
5821 #[test]
5822 fn taking_a_frame_a_page_at_a_time_is_off_until_it_is_asked_for() {
5823 let (opts, _) = compile(&["-c", "a.c"]);
5824 assert!(!opts.stack_clash, "gcc takes a frame in one subtraction unless it was asked");
5825
5826 let (opts, _) = compile(&["-c", "-fstack-clash-protection", "a.c"]);
5827 assert!(opts.stack_clash);
5828
5829 // The same shape a package build uses for the protector: on in the global flags and off
5830 // for the one directory that cannot have it.
5831 let (opts, _) =
5832 compile(&["-c", "-fstack-clash-protection", "-fno-stack-clash-protection", "a.c"]);
5833 assert!(!opts.stack_clash);
5834 let (opts, _) =
5835 compile(&["-c", "-fno-stack-clash-protection", "-fstack-clash-protection", "a.c"]);
5836 assert!(opts.stack_clash, "the last one wins either way round");
5837
5838 // The two are independent, since one is about the frame and the other about the function.
5839 let (opts, _) =
5840 compile(&["-c", "-fstack-clash-protection", "-fstack-protector-strong", "a.c"]);
5841 assert!(opts.stack_clash);
5842 assert_eq!(opts.protector, Protector::Strong);
5843 }
5844
5845 /// One flag with an argument rather than a family of spellings, because what it asks about is
5846 /// which of the two edges of a control flow transfer is checked and the two are not separate
5847 /// questions to the hardware.
5848 #[test]
5849 fn which_control_flow_edges_are_checked_is_asked_for_by_name() {
5850 let (opts, _) = compile(&["-c", "a.c"]);
5851 assert_eq!(opts.control, Control::None, "gcc's default on the targets this compiler has");
5852
5853 for (arg, want) in [
5854 ("-fcf-protection", Control::Full),
5855 ("-fcf-protection=full", Control::Full),
5856 ("-fcf-protection=branch", Control::Branch),
5857 ("-fcf-protection=return", Control::Return),
5858 ("-fcf-protection=none", Control::None),
5859 ("-fcf-protection=check", Control::Check),
5860 ] {
5861 let (opts, _) = compile(&["-c", arg, "a.c"]);
5862 assert_eq!(opts.control, want, "{arg}");
5863 }
5864
5865 // The shape a package build uses: on in the global flags and off for the one directory
5866 // that cannot have it, whichever of the two spellings of off it reaches for.
5867 let (opts, _) = compile(&["-c", "-fcf-protection=full", "-fno-cf-protection", "a.c"]);
5868 assert_eq!(opts.control, Control::None);
5869 let (opts, _) = compile(&["-c", "-fno-cf-protection", "-fcf-protection=branch", "a.c"]);
5870 assert_eq!(opts.control, Control::Branch, "the last one wins either way round");
5871 }
5872
5873 /// The profiler is asked for by two spellings, and where its hook goes by two more.
5874 ///
5875 /// The two halves are separate on purpose. `-mfentry` on its own says where a call would go and
5876 /// asks for no call, which is what gcc does with it, and a build system that sets it globally
5877 /// and asks for the profile per directory needs that to be true rather than an error.
5878 ///
5879 /// The link is asserted alongside, because the flag changes it too and a build that compiled
5880 /// with it and linked without it is a program that calls the hook everywhere and never writes a
5881 /// profile.
5882 #[test]
5883 fn the_profiler_and_where_its_hook_goes_are_two_separate_questions() {
5884 let (opts, _) = compile(&["-c", "a.c"]);
5885 assert!(!opts.profile);
5886 assert_eq!(opts.hook, Hook::Platform, "neither was named, so the target decides");
5887
5888 for arg in ["-pg", "-p"] {
5889 let (opts, _) = compile(&["-c", arg, "a.c"]);
5890 assert!(opts.profile, "{arg}");
5891 let (link, _) = linking(&[arg, "a.c"]);
5892 assert!(link.profile, "{arg} changes the link as well");
5893 }
5894
5895 for (arg, want) in [("-mfentry", Hook::Early), ("-mno-fentry", Hook::Late)] {
5896 let (opts, _) = compile(&["-c", arg, "a.c"]);
5897 assert_eq!(opts.hook, want, "{arg}");
5898 assert!(!opts.profile, "{arg} asks for no call of its own");
5899 }
5900
5901 let (opts, _) = compile(&["-c", "-mfentry", "-mno-fentry", "-pg", "a.c"]);
5902 assert_eq!(opts.hook, Hook::Late, "the last one wins");
5903 assert!(opts.profile);
5904 }
5905
5906 /// How much room a patcher is promised, which is one number or two.
5907 ///
5908 /// A command line that did not ask is asserted alongside, because the flag has to be written to
5909 /// mean anything and a build that reserved room nobody asked for would grow every function in
5910 /// it for nothing.
5911 #[test]
5912 fn the_room_a_patcher_is_promised_is_a_number_of_bytes_and_where_they_go() {
5913 let (opts, _) = compile(&["-c", "a.c"]);
5914 assert_eq!(opts.patchable, Patchable::default());
5915 assert!(!opts.patchable.any(), "nothing is reserved unless it was asked for");
5916
5917 let (opts, _) = compile(&["-c", "-fpatchable-function-entry=16", "a.c"]);
5918 assert_eq!(opts.patchable, Patchable { total: 16, before: 0 });
5919
5920 let (opts, _) = compile(&["-c", "-fpatchable-function-entry=5,3", "a.c"]);
5921 assert_eq!(opts.patchable, Patchable { total: 5, before: 3 });
5922 assert_eq!(opts.patchable.after(), 2);
5923
5924 // The last one wins, which is what every other flag of this shape does and what a build
5925 // that adds one to a command line it did not write is relying on.
5926 let (opts, _) = compile(&[
5927 "-c",
5928 "-fpatchable-function-entry=5,3",
5929 "-fpatchable-function-entry=2",
5930 "a.c",
5931 ]);
5932 assert_eq!(opts.patchable, Patchable { total: 2, before: 0 });
5933 }
5934
5935 /// And a request nothing could satisfy is refused rather than rounded into one that can be.
5936 #[test]
5937 fn room_in_front_of_the_label_that_is_more_than_the_room_asked_for_is_refused() {
5938 for arg in ["-fpatchable-function-entry=1,2", "-fpatchable-function-entry=x"] {
5939 let e = parse_args(&args(&["-c", arg, "a.c"])).unwrap_err();
5940 assert!(e.message.contains("is not an amount of room to reserve"), "{}", e.message);
5941 }
5942 }
5943
5944 /// What wraps rather than being undefined, which is two questions and three flags.
5945 ///
5946 /// The older flag is the pair of the newer two, which is gcc's own reading of it, so a build
5947 /// that writes `-fno-strict-overflow` gets both and a build that writes one of the others gets
5948 /// only what it asked for.
5949 #[test]
5950 fn what_overflows_rather_than_being_undefined_is_asked_for_two_ways() {
5951 let (opts, _) = compile(&["-c", "a.c"]);
5952 assert_eq!(opts.wrapping, Wrapping::NONE, "nothing wraps unless it was asked for");
5953
5954 let (opts, _) = compile(&["-c", "-fwrapv", "a.c"]);
5955 assert_eq!(opts.wrapping, Wrapping { signed: true, pointer: false, trap: false });
5956
5957 let (opts, _) = compile(&["-c", "-fwrapv-pointer", "a.c"]);
5958 assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: true, trap: false });
5959
5960 let (opts, _) = compile(&["-c", "-fno-strict-overflow", "a.c"]);
5961 assert_eq!(opts.wrapping, Wrapping::ALL);
5962
5963 // And the last one wins, in both directions. A build that turns one of these on globally
5964 // and off for one directory is relying on that, and so is one that writes the pair and
5965 // then takes half of it back.
5966 let (opts, _) = compile(&["-c", "-fwrapv", "-fno-wrapv", "a.c"]);
5967 assert_eq!(opts.wrapping, Wrapping::NONE);
5968
5969 let (opts, _) = compile(&["-c", "-fno-strict-overflow", "-fstrict-overflow", "a.c"]);
5970 assert_eq!(opts.wrapping, Wrapping::NONE);
5971
5972 let (opts, _) = compile(&["-c", "-fno-strict-overflow", "-fno-wrapv-pointer", "a.c"]);
5973 assert_eq!(opts.wrapping, Wrapping { signed: true, pointer: false, trap: false });
5974 }
5975
5976 /// And the other answer to the signed question cannot be held at the same time as the first.
5977 ///
5978 /// A program cannot both wrap and stop, so writing both is writing a contradiction, and gcc
5979 /// resolves it by letting the last one win rather than by reporting anything. That was measured
5980 /// against gcc 16 rather than read out of the manual, which says nothing about it: `-ftrapv
5981 /// -fwrapv` emits no checked calls and `-fwrapv -ftrapv` emits them.
5982 #[test]
5983 fn a_signed_overflow_that_stops_is_the_other_answer_and_not_a_third_one() {
5984 let (opts, _) = compile(&["-c", "-ftrapv", "a.c"]);
5985 assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: false, trap: true });
5986
5987 let (opts, _) = compile(&["-c", "-fwrapv", "-ftrapv", "a.c"]);
5988 assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: false, trap: true });
5989
5990 let (opts, _) = compile(&["-c", "-ftrapv", "-fwrapv", "a.c"]);
5991 assert_eq!(opts.wrapping, Wrapping { signed: true, pointer: false, trap: false });
5992
5993 let (opts, _) = compile(&["-c", "-ftrapv", "-fno-strict-overflow", "a.c"]);
5994 assert_eq!(opts.wrapping, Wrapping::ALL);
5995
5996 let (opts, _) = compile(&["-c", "-ftrapv", "-fno-trapv", "a.c"]);
5997 assert_eq!(opts.wrapping, Wrapping::NONE);
5998
5999 // And the flag that says what may be assumed says nothing about what happens, so it leaves
6000 // this alone where it takes the wrapping away. gcc does the same.
6001 let (opts, _) = compile(&["-c", "-ftrapv", "-fstrict-overflow", "a.c"]);
6002 assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: false, trap: true });
6003 }
6004
6005 /// What a plain `char` is, which is four spellings of two answers and nothing by default.
6006 ///
6007 /// Nothing is the target's own answer and has to stay distinct from both of the others, since
6008 /// the same command line means a signed `char` on x86-64 and an unsigned one on Linux's arm64.
6009 /// The negative spellings are the other flag rather than a way of asking for the default, which
6010 /// was measured against gcc 16: `-fno-signed-char` defines `__CHAR_UNSIGNED__` and
6011 /// `-fno-unsigned-char` does not.
6012 #[test]
6013 fn the_signedness_of_a_plain_char_is_asked_for_in_four_ways() {
6014 let (opts, _) = compile(&["-c", "a.c"]);
6015 assert_eq!(opts.char_signed, None);
6016
6017 for flag in ["-fsigned-char", "-fno-unsigned-char"] {
6018 let (opts, _) = compile(&["-c", flag, "a.c"]);
6019 assert_eq!(opts.char_signed, Some(true), "{flag}");
6020 }
6021
6022 for flag in ["-funsigned-char", "-fno-signed-char"] {
6023 let (opts, _) = compile(&["-c", flag, "a.c"]);
6024 assert_eq!(opts.char_signed, Some(false), "{flag}");
6025 }
6026
6027 // And the last one wins, which is what a build that sets one globally and the other for a
6028 // directory relies on.
6029 let (opts, _) = compile(&["-c", "-funsigned-char", "-fsigned-char", "a.c"]);
6030 assert_eq!(opts.char_signed, Some(true));
6031
6032 // And what is asked for reaches the target, because that is what every other part of the
6033 // compiler asks. The triple is one whose own answer is the opposite, so a session that
6034 // ignored the flag would still read as signed here.
6035 let (opts, _) =
6036 compile(&["-c", "--target=aarch64-unknown-linux-gnu", "-fsigned-char", "a.c"]);
6037 assert!(Session::new(*opts).target.char_is_signed);
6038 let (opts, _) = compile(&["-c", "--target=aarch64-unknown-linux-gnu", "a.c"]);
6039 assert!(!Session::new(*opts).target.char_is_signed);
6040 }
6041
6042 /// And the size of an enumeration, which is one question with two spellings.
6043 #[test]
6044 fn the_smallest_enumeration_is_asked_for_and_taken_back() {
6045 let (opts, _) = compile(&["-c", "a.c"]);
6046 assert!(!opts.short_enums);
6047
6048 let (opts, _) = compile(&["-c", "-fshort-enums", "a.c"]);
6049 assert!(opts.short_enums);
6050
6051 let (opts, _) = compile(&["-c", "-fshort-enums", "-fno-short-enums", "a.c"]);
6052 assert!(!opts.short_enums);
6053
6054 let (opts, _) = compile(&["-c", "-fno-short-enums", "-fshort-enums", "a.c"]);
6055 assert!(opts.short_enums);
6056 }
6057
6058 /// And Microsoft's reading of an anonymous member, which the target answers where the command
6059 /// line said nothing. gcc's mingw build has it on and its Linux build has it off, so a header
6060 /// that closes a nameless union with a macro that expands to nothing is read the way the
6061 /// compiler that platform ships would read it.
6062 #[test]
6063 fn the_microsoft_reading_of_a_member_follows_the_target_until_it_is_asked_for() {
6064 // Named rather than left to the host, since the answer this asks for is the one a target
6065 // that is not Windows gives and on a Windows machine the host is not one of those.
6066 let (opts, _) = compile(&[LINUX, "-c", "a.c"]);
6067 assert!(!Session::new(*opts).ms_extensions());
6068
6069 let (opts, _) = compile(&["-c", "--target=x86_64-pc-windows-gnu", "a.c"]);
6070 assert!(Session::new(*opts).ms_extensions());
6071
6072 let (opts, _) = compile(&["-c", "-fms-extensions", "a.c"]);
6073 assert!(Session::new(*opts).ms_extensions());
6074
6075 let (opts, _) =
6076 compile(&["-c", "--target=x86_64-pc-windows-gnu", "-fno-ms-extensions", "a.c"]);
6077 assert!(!Session::new(*opts).ms_extensions());
6078 }
6079
6080 /// And a value nothing means is refused rather than taken for the nearest thing it looks like.
6081 ///
6082 /// `-fcf-protection=all` is the spelling somebody writes from memory, and a compiler that read
6083 /// it as `full` would be guessing, while one that let it fall through to the optimizer's `-f`
6084 /// family would report it as an unknown pass. Neither is the news the build wants.
6085 #[test]
6086 fn a_control_flow_protection_nothing_means_is_refused() {
6087 let e = parse_args(&args(&["-c", "-fcf-protection=all", "a.c"])).unwrap_err();
6088 assert!(e.message.contains("is not a control flow protection"), "{}", e.message);
6089 assert!(e.message.contains("full, branch, return, none or check"), "{}", e.message);
6090 }
6091
6092 #[test]
6093 fn mingw_subsystem_and_unicode_flags_are_taken_last_one_winning() {
6094 let (link, _) = linking(&["-mwindows", "-municode", "-mthreads", "-static-libgcc", "a.c"]);
6095 assert!(link.gui && link.unicode);
6096 let (link, _) = linking(&["-mwindows", "-mconsole", "a.c"]);
6097 assert!(!link.gui);
6098 let (opts, _) = compile(&["-municode", "-c", "a.c"]);
6099 assert!(opts.defines.iter().any(|define| define == "UNICODE"), "{:?}", opts.defines);
6100 }
6101
6102 #[test]
6103 fn the_link_flags_are_collected_apart_from_the_compilation() {
6104 let (link, _) = linking(&[
6105 "-static",
6106 "-nostartfiles",
6107 "-rdynamic",
6108 "-s",
6109 "-fuse-ld=mold",
6110 "-L/opt/lib",
6111 "-B",
6112 "/opt/tools",
6113 "a.c",
6114 ]);
6115 assert!(link.is_static);
6116 assert!(link.no_startfiles);
6117 assert!(link.export_dynamic);
6118 assert!(link.strip);
6119 assert_eq!(link.use_ld.as_deref(), Some("mold"));
6120 assert_eq!(link.search, vec![PathBuf::from("/opt/lib")]);
6121 assert_eq!(link.prefixes, vec![PathBuf::from("/opt/tools")]);
6122 }
6123
6124 #[test]
6125 fn a_mingw_link_names_its_output_the_way_mingw_gcc_does() {
6126 // gcc puts `.exe` on a DLL's name as well when it has no extension, so `-shared` is not an
6127 // exception, and `-c` links nothing and is. tamnd/rucc#2152.
6128 let mingw = "--target=x86_64-windows-gnu";
6129 let (_, plan) = linking(&[mingw, "a.c", "-o", "foo"]);
6130 assert_eq!(plan.link.expect("expected a link step").output, "foo.exe");
6131 let (_, plan) = linking(&[mingw, "-shared", "a.c", "-o", "x"]);
6132 assert_eq!(plan.link.expect("expected a link step").output, "x.exe");
6133 let (_, plan) = linking(&[mingw, "-shared", "a.c", "-o", "x.dll"]);
6134 assert_eq!(plan.link.expect("expected a link step").output, "x.dll");
6135 let (_, plan) = compile(&[mingw, "-c", "a.c", "-o", "x"]);
6136 assert!(plan.link.is_none());
6137 assert_eq!(plan.jobs[0].output, Output::File("x".into()));
6138 let (_, plan) = linking(&[LINUX, "a.c", "-o", "foo"]);
6139 assert_eq!(plan.link.expect("expected a link step").output, "foo");
6140 }
6141
6142 #[test]
6143 fn a_comma_in_dash_wl_separates_two_arguments() {
6144 // The target is written down because the name of the object is derived from it, and `a.o`
6145 // on a Linux host is `a.obj` on a Windows one. What is under test is the splitting of the
6146 // argument, which has nothing to do with either.
6147 let (_, plan) = linking(&[LINUX, "-Wl,-rpath,/opt/lib", "-Xlinker", "--as-needed", "a.c"]);
6148 let link = plan.link.expect("expected a link step");
6149 assert_eq!(
6150 link.inputs,
6151 vec![
6152 link::Item::Linker("-rpath".into()),
6153 link::Item::Linker("/opt/lib".into()),
6154 link::Item::Linker("--as-needed".into()),
6155 link::Item::File("a.o".into()),
6156 ]
6157 );
6158 }
6159
6160 #[test]
6161 fn a_word_for_the_linker_keeps_its_place_among_the_files_too() {
6162 // What libtool writes around a set of convenience archives, and what #1279 was. Both words
6163 // are about the files between them, so the pair collected out of the line and appended to
6164 // the end is two options that bracket nothing and an archive that went in empty.
6165 let (_, plan) = linking(&[
6166 "--target=x86_64-unknown-linux-gnu",
6167 "a.c",
6168 "-Wl,--whole-archive",
6169 "libaesni.a",
6170 "-Wl,--no-whole-archive",
6171 "-lm",
6172 ]);
6173 let link = plan.link.expect("expected a link step");
6174 assert_eq!(
6175 link.inputs,
6176 vec![
6177 link::Item::File("a.o".into()),
6178 link::Item::Linker("--whole-archive".into()),
6179 link::Item::File("libaesni.a".into()),
6180 link::Item::Linker("--no-whole-archive".into()),
6181 link::Item::Library("m".into()),
6182 ]
6183 );
6184 // And it is not a job, because there is nothing to compile in a word for the linker.
6185 assert_eq!(plan.jobs.len(), 2);
6186 }
6187
6188 #[test]
6189 fn a_word_for_the_linker_on_a_dash_c_line_is_dropped_without_a_word() {
6190 // GCC says nothing about one either. `-Wl,` on a compile line is what a build system
6191 // writes when one variable holds the flags for both, and a note here would be a note on
6192 // every compile of every autotools project.
6193 let (_, plan) = linking(&["-c", "-Wl,--as-needed", "a.c"]);
6194 assert!(plan.link.is_none());
6195 assert!(plan.notes.is_empty(), "{:?}", plan.notes);
6196 assert_eq!(plan.jobs.len(), 1);
6197 }
6198
6199 #[test]
6200 fn a_library_keeps_its_place_between_the_objects() {
6201 // Link order is semantic: `-lm` written between two files resolves for the one before
6202 // it and not for the one after, so a library cannot be collected into a list of its own.
6203 // The target is named because the suffix of an object is the target's and this asserts
6204 // on the names: the same command line on a Windows host plans two `.obj` files.
6205 let (_, plan) = linking(&["--target=x86_64-unknown-linux-gnu", "a.c", "-lm", "b.c"]);
6206 let link = plan.link.expect("expected a link step");
6207 assert_eq!(
6208 link.inputs,
6209 vec![
6210 link::Item::File("a.o".into()),
6211 link::Item::Library("m".into()),
6212 link::Item::File("b.o".into()),
6213 ]
6214 );
6215 // And it is not a job, because there is nothing to compile in a library.
6216 assert_eq!(plan.jobs.len(), 2);
6217 }
6218
6219 #[test]
6220 fn a_library_on_a_dash_c_line_is_a_note_rather_than_an_error() {
6221 let (_, plan) = linking(&["-c", "-lm", "a.c"]);
6222 assert!(plan.link.is_none());
6223 assert!(plan.notes.iter().any(|n| n.contains("-lm")), "{:?}", plan.notes);
6224 }
6225
6226 #[test]
6227 fn the_sysroot_reaches_the_linker_as_well_as_the_headers() {
6228 let (link, _) = linking(&["--sysroot=/opt/root", "a.c"]);
6229 assert_eq!(link.sysroot, Some(PathBuf::from("/opt/root")));
6230 }
6231
6232 fn printed(s: &[&str]) -> String {
6233 match parse_args(&args(s)).expect("expected an answer") {
6234 Action::Print(line) => line,
6235 other => panic!("expected an answer, got {other:?}"),
6236 }
6237 }
6238
6239 fn refused(s: &[&str]) -> String {
6240 parse_args(&args(s)).expect_err("expected a refusal").message
6241 }
6242
6243 #[test]
6244 fn a_warning_flag_gcc_knows_is_taken_even_though_nothing_reads_it() {
6245 // The rule in section 4.1, and the reason for it is autoconf: a configure script finds
6246 // out whether a warning flag exists by passing it and looking at the exit status, so a
6247 // compiler that refuses one gcc knows fails a script written for gcc.
6248 let (opts, _) = compile(&["-Wall", "-Wextra", "-Wno-format-truncation", "-c", "a.c"]);
6249 assert!(!opts.warnings_are_errors);
6250 assert!(opts.warnings);
6251 // The two spellings that do mean something are still read.
6252 let (opts, _) = compile(&["-Werror", "-c", "a.c"]);
6253 assert!(opts.warnings_are_errors);
6254 let (opts, _) = compile(&["-w", "-c", "a.c"]);
6255 assert!(!opts.warnings);
6256 // Off without being asked, the way gcc has it off, and both spellings are read.
6257 let (opts, _) = compile(&["-c", "a.c"]);
6258 assert!(!opts.system_header_warnings);
6259 let (opts, _) = compile(&["-Wsystem-headers", "-c", "a.c"]);
6260 assert!(opts.system_header_warnings);
6261 let (opts, _) = compile(&["-Wsystem-headers", "-Wno-system-headers", "-c", "a.c"]);
6262 assert!(!opts.system_header_warnings);
6263 let (opts, _) = compile(&["-pedantic-errors", "-c", "a.c"]);
6264 assert!(opts.pedantic && opts.warnings_are_errors);
6265 }
6266
6267 #[test]
6268 fn a_warning_flag_gcc_refuses_is_refused_here_too() {
6269 // Postgres's meson build probes these, and with rucc taking them it ended up passing four
6270 // clang warnings that the gcc build had dropped.
6271 for flag in ["-Wcast-function-type-strict", "-Wunused-command-line-argument"] {
6272 assert_eq!(refused(&[flag, "-c", "a.c"]), format!("unknown option `{flag}`"));
6273 }
6274 assert_eq!(
6275 refused(&["-Werror=unguarded-availability-new", "-c", "a.c"]),
6276 "`-Werror=unguarded-availability-new`: no option `-Wunguarded-availability-new`"
6277 );
6278 assert!(refused(&["-Wno-error=nonsense", "-c", "a.c"]).contains("no option `-Wnonsense`"));
6279 // gcc takes `-Wno-` of a name it does not know, and says nothing unless something else
6280 // is said, and it takes C++ and Fortran names on a C compile.
6281 for flag in
6282 ["-Wno-cast-function-type-strict", "-Werror=format", "-Wformat=2", "-Wabi-tag", "-W"]
6283 {
6284 compile(&[flag, "-c", "a.c"]);
6285 }
6286 }
6287
6288 #[test]
6289 fn an_argument_for_a_separate_tool_is_refused_rather_than_dropped() {
6290 // Every one of these says something about the output, so the wrong answer is silence.
6291 assert!(refused(&["-Wa,--noexecstack", "-c", "a.c"]).contains("separate assembler"));
6292 assert!(refused(&["-Wp,-C", "-c", "a.c"]).contains("separate assembler"));
6293 assert!(refused(&["-specs=/x", "a.c"]).contains("-specs= is not supported"));
6294 assert!(refused(&["-mcmodel=kernel", "-c", "a.c"]).contains("small code model"));
6295 assert!(refused(&["-gdwarf-4", "-c", "a.c"]).contains("DWARF 5"));
6296 // The word size the target does not have, which is a target this compiler was not asked
6297 // for rather than a flag it does not know.
6298 let no32 = refused(&["--target=x86_64-unknown-linux-gnu", "-m32", "-c", "a.c"]);
6299 assert!(no32.contains("32 bit target"), "{no32}");
6300 }
6301
6302 /// `-gz` and the two spellings of the split, which are the two questions about the shape of
6303 /// the debug output rather than about how much of it there is.
6304 ///
6305 /// Both answers here are about what happens when there is debug information to shape, and
6306 /// there is none yet, so what is being asserted is that the flags are read and remembered
6307 /// rather than that anything changed in the output. That is the whole of what taking them
6308 /// claims, and it is worth a test because the day `rucc-debug` writes a section this is where
6309 /// it comes to find out what the command line said.
6310 #[test]
6311 fn the_shape_of_the_debug_output_is_recorded_even_where_there_is_none_of_it() {
6312 let (opts, _) = compile(&["-c", "a.c"]);
6313 assert_eq!(opts.compress, Compress::None, "uncompressed unless somebody asks");
6314
6315 // Bare `-gz` is `-gz=zlib`, measured against gcc 16 rather than read out of the manual,
6316 // which describes the flag without ever saying which algorithm it picks.
6317 assert_eq!(compile(&["-gz", "-c", "a.c"]).0.compress, Compress::Zlib);
6318 for (spelling, want) in [
6319 ("none", Compress::None),
6320 ("zlib", Compress::Zlib),
6321 ("zlib-gnu", Compress::ZlibGnu),
6322 ("zstd", Compress::Zstd),
6323 ] {
6324 let (opts, _) = compile(&[&format!("-gz={spelling}"), "-c", "a.c"]);
6325 assert_eq!(opts.compress, want, "{spelling}");
6326 }
6327
6328 // A value nothing here has heard of is refused rather than rounded to the nearest one,
6329 // because a build that asked for `zstd` and quietly got `zlib` would ship a file its
6330 // reader may not understand and would have no way of finding out.
6331 for bad in ["-gz=gzip", "-gz="] {
6332 let failed = refused(&[bad, "-c", "a.c"]);
6333 assert!(failed.contains("is not a way to compress"), "{bad}: {failed}");
6334 }
6335
6336 // The split is refused in the direction that would have written a file and taken in the
6337 // direction that describes what happens. A build system that names the `.dwo` as an
6338 // output has to hear about it now rather than at the point the file is missing.
6339 let (opts, _) = compile(&["-gno-split-dwarf", "-g", "-c", "a.c"]);
6340 assert!(opts.debug_info, "the negative spelling says nothing about how much");
6341 let failed = refused(&["-gsplit-dwarf", "-c", "a.c"]);
6342 assert!(failed.contains(".dwo"), "the refusal names the file it would have written");
6343 }
6344
6345 /// The `-flto` family, which is the whole of an optimization this compiler does not do.
6346 ///
6347 /// Taken rather than refused because ignoring it gives a correct program that is slower than
6348 /// it could have been, which is section 4.1's hint about speed. The values are still held to
6349 /// gcc's, so a command line written for clang is told rather than quietly taken.
6350 #[test]
6351 fn the_link_time_family_is_read_and_checked_and_nothing_is_done_about_it() {
6352 let (opts, _) = compile(&["-c", "a.c"]);
6353 assert!(!opts.lto.requested, "nothing asks unless the command line does");
6354
6355 let (opts, _) = compile(&["-flto", "-c", "a.c"]);
6356 assert!(opts.lto.requested);
6357 assert_eq!(opts.lto.jobs, LtoJobs::One, "bare -flto is one process, the way gcc reads it");
6358
6359 // The last of the two directions wins, the same as every other pair of `-f` spellings.
6360 assert!(!compile(&["-flto", "-fno-lto", "-c", "a.c"]).0.lto.requested);
6361 assert!(compile(&["-fno-lto", "-flto", "-c", "a.c"]).0.lto.requested);
6362
6363 // A count is a count, and asking for one implies asking for the optimization.
6364 for (spelling, want) in [
6365 ("auto", LtoJobs::Auto),
6366 ("jobserver", LtoJobs::Jobserver),
6367 ("1", LtoJobs::One),
6368 ("8", LtoJobs::Count(8)),
6369 ] {
6370 let (opts, _) = compile(&[&format!("-flto={spelling}"), "-c", "a.c"]);
6371 assert_eq!(opts.lto.jobs, want, "{spelling}");
6372 assert!(opts.lto.requested, "{spelling} asks for it too");
6373 }
6374
6375 // gcc refuses a zero rather than reading it as `-fno-lto`, and `thin` is clang's spelling
6376 // of a question gcc answers with `-flto-partition=`, so somebody who wrote it meant a
6377 // different compiler and gets told so here rather than getting a serial link.
6378 for bad in ["-flto=0", "-flto=thin", "-flto=full", "-flto=-1"] {
6379 let failed = refused(&[bad, "-c", "a.c"]);
6380 assert!(failed.contains("link time jobs"), "{bad}: {failed}");
6381 }
6382
6383 // How the program is cut up before the work is spread over it.
6384 assert_eq!(compile(&["-c", "a.c"]).0.lto.partition, Partition::Balanced, "gcc's default");
6385 for (spelling, want) in [
6386 ("balanced", Partition::Balanced),
6387 ("1to1", Partition::OneToOne),
6388 ("one", Partition::One),
6389 ("max", Partition::Max),
6390 ("none", Partition::None),
6391 ] {
6392 let (opts, _) = compile(&[&format!("-flto-partition={spelling}"), "-c", "a.c"]);
6393 assert_eq!(opts.lto.partition, want, "{spelling}");
6394 }
6395 assert!(refused(&["-flto-partition=big", "-c", "a.c"]).contains("partitioning model"));
6396
6397 // And how hard the bytecode is compressed on its way into the object, which is zstd's
6398 // range of levels and is the range gcc checks an argument against.
6399 assert_eq!(compile(&["-c", "a.c"]).0.lto.compression, None, "whatever it does by default");
6400 assert_eq!(compile(&["-flto-compression-level=0", "-c", "a.c"]).0.lto.compression, Some(0));
6401 let (opts, _) = compile(&["-flto-compression-level=19", "-c", "a.c"]);
6402 assert_eq!(opts.lto.compression, Some(19));
6403 for bad in ["-flto-compression-level=20", "-flto-compression-level=-1"] {
6404 let failed = refused(&[bad, "-c", "a.c"]);
6405 assert!(failed.contains("compression level"), "{bad}: {failed}");
6406 }
6407
6408 // The two pairs that describe an arrangement rather than ask for one. Every object here
6409 // holds its machine code, so the fat spelling is what already happens and the other is a
6410 // smaller file rather than a different program, and the plugin pair is about a tool the
6411 // design in `spec/09-optimizer.md` never loads.
6412 for taken in [
6413 "-ffat-lto-objects",
6414 "-fno-fat-lto-objects",
6415 "-fuse-linker-plugin",
6416 "-fno-use-linker-plugin",
6417 ] {
6418 let (opts, _) = compile(&[taken, "-c", "a.c"]);
6419 assert!(!opts.lto.requested, "{taken} says nothing about whether to do it");
6420 }
6421 }
6422
6423 /// The profile family, which is the only one here that splits down the middle.
6424 ///
6425 /// Reading a profile is taken and writing one is refused, and the line between them is the one
6426 /// section 4.1 draws: ignoring a request to read the counts gives a correct program that is
6427 /// slower than it could have been, and ignoring a request to write them means a file the build
6428 /// declared as an output never appears.
6429 #[test]
6430 fn reading_a_profile_is_taken_and_writing_one_is_refused() {
6431 let (opts, _) = compile(&["-c", "a.c"]);
6432 assert!(!opts.profile_data.requested, "nothing asks unless the command line does");
6433 assert_eq!(opts.profile_data.path, None);
6434
6435 let (opts, _) = compile(&["-fprofile-use", "-c", "a.c"]);
6436 assert!(opts.profile_data.requested);
6437 assert_eq!(opts.profile_data.path, None, "beside the object, the way gcc looks");
6438
6439 let (opts, _) = compile(&["-fprofile-use=/counts", "-c", "a.c"]);
6440 assert!(opts.profile_data.requested, "naming a path asks for it too");
6441 assert_eq!(opts.profile_data.path.as_deref(), Some("/counts"));
6442
6443 // The last of the two directions wins, the same as every other pair of `-f` spellings.
6444 assert!(
6445 !compile(&["-fprofile-use", "-fno-profile-use", "-c", "a.c"]).0.profile_data.requested
6446 );
6447 assert!(
6448 compile(&["-fno-profile-use", "-fprofile-use", "-c", "a.c"]).0.profile_data.requested
6449 );
6450
6451 // The rest of the reading half, which is where the files are and three answers about what
6452 // to make of what is in them.
6453 let (opts, _) = compile(&[
6454 "-fprofile-dir=/build/profiles",
6455 "-fprofile-abs-path",
6456 "-fprofile-correction",
6457 "-fprofile-partial-training",
6458 "-c",
6459 "a.c",
6460 ]);
6461 assert_eq!(opts.profile_data.dir.as_deref(), Some("/build/profiles"));
6462 assert!(opts.profile_data.absolute);
6463 assert!(opts.profile_data.correction);
6464 assert!(opts.profile_data.partial_training);
6465
6466 // Writing one, which is refused by name. The first four instrument the program and the
6467 // last writes a file beside the object, and a build that got neither and no message would
6468 // go on to optimize against counts that were never gathered.
6469 for writing in [
6470 "-fprofile-generate",
6471 "-fprofile-generate=/build/profiles",
6472 "-fprofile-arcs",
6473 "--coverage",
6474 "-fcondition-coverage",
6475 "-fpath-coverage",
6476 ] {
6477 let failed = refused(&[writing, "-c", "a.c"]);
6478 assert!(failed.contains("instrument"), "{writing}: {failed}");
6479 }
6480 assert!(refused(&["-ftest-coverage", "-c", "a.c"]).contains(".gcno"), "it names the file");
6481
6482 // The negative spellings of the refused half are what already happens, so they are taken.
6483 for taken in ["-fno-profile-generate", "-fno-profile-arcs", "-fno-test-coverage"] {
6484 let (opts, _) = compile(&[taken, "-c", "a.c"]);
6485 assert!(!opts.profile_data.requested, "{taken} asks for nothing");
6486 }
6487
6488 // And the flags that describe the instrumentation that is refused above, which are checked
6489 // and dropped. Checked because a typo is worth finding here rather than on the day the
6490 // instrumentation lands.
6491 for taken in [
6492 "-fprofile-update=single",
6493 "-fprofile-update=atomic",
6494 "-fprofile-update=prefer-atomic",
6495 "-fprofile-reproducible=serial",
6496 "-fprofile-reproducible=parallel-runs",
6497 "-fprofile-reproducible=multithreaded",
6498 "-fprofile-values",
6499 "-fno-profile-values",
6500 "-fprofile-info-section",
6501 "-fprofile-filter-files=a.c",
6502 "-fprofile-exclude-files=b.c",
6503 "-fprofile-note=a.gcno",
6504 ] {
6505 let (opts, _) = compile(&[taken, "-c", "a.c"]);
6506 assert!(!opts.profile_data.requested, "{taken} says nothing about reading one");
6507 }
6508 assert!(refused(&["-fprofile-update=none", "-c", "a.c"]).contains("update method"));
6509 assert!(refused(&["-fprofile-reproducible=any", "-c", "a.c"]).contains("reproducibility"));
6510 }
6511
6512 /// The sanitizers, which are refused by name and are the one family refused for a reason that
6513 /// is not about the bytes.
6514 ///
6515 /// A sanitizer is a promise that the program is watched while it runs, so a build that asked
6516 /// for one and was quietly given a program with no checks in it gets a test suite that passes
6517 /// for the wrong reason rather than a slower program.
6518 #[test]
6519 fn a_sanitizer_that_is_still_asked_for_at_the_end_of_the_line_is_refused_by_name() {
6520 for asked in ["address", "undefined", "thread", "kernel-address", "leak", "memory"] {
6521 let failed = refused(&[&format!("-fsanitize={asked}"), "-c", "a.c"]);
6522 assert!(failed.contains(asked), "the refusal names what was asked for: {failed}");
6523 assert!(failed.contains("-fsafety=detect"), "and the nearest thing: {failed}");
6524 }
6525
6526 // A list is every name in it, and the first one still standing is the one named.
6527 let failed = refused(&["-fsanitize=address,undefined", "-c", "a.c"]);
6528 assert!(failed.contains("address"), "{failed}");
6529
6530 // A name that is not one, which is worth its own message: somebody who wrote `-fsanitize`
6531 // with a typo in it has a different problem from somebody who wrote a real one.
6532 for bad in ["-fsanitize=bogus", "-fsanitize=address,bogus", "-fno-sanitize=bogus"] {
6533 let failed = refused(&[bad, "-c", "a.c"]);
6534 assert!(failed.contains("is not a sanitizer"), "{bad}: {failed}");
6535 }
6536
6537 // gcc takes `all` only in the negative, and so does this.
6538 assert!(refused(&["-fsanitize=all", "-c", "a.c"]).contains("only `-fno-sanitize=all`"));
6539
6540 // Asking and then taking it back is asking for nothing, which is why the answer waits for
6541 // the end of the line. A build whose shared flags turn a check on and whose rule for one
6542 // file turns it off again compiles that file here.
6543 for pair in [
6544 ["-fsanitize=address", "-fno-sanitize=address"],
6545 ["-fsanitize=address,undefined", "-fno-sanitize=all"],
6546 ["-fsanitize=undefined", "-fno-sanitize=undefined"],
6547 ] {
6548 let (opts, _) = compile(&[pair[0], pair[1], "-c", "a.c"]);
6549 assert_eq!(opts.safety, rucc_session::Safety::Off, "{pair:?} asked for nothing");
6550 }
6551 // And the other order still asks, because the last word is the one that counts.
6552 assert!(!refused(&["-fno-sanitize=address", "-fsanitize=address", "-c", "a.c"]).is_empty());
6553
6554 // What a check does when it fires is an answer about checks that are refused, so there is
6555 // nothing left for it to change and it is taken.
6556 for taken in [
6557 "-fsanitize-recover=undefined",
6558 "-fno-sanitize-recover=all",
6559 "-fsanitize-trap=undefined",
6560 "-fno-sanitize-trap=all",
6561 "-fsanitize-undefined-trap-on-error",
6562 "-fsanitize-address-use-after-scope",
6563 "-fno-sanitize-address-use-after-scope",
6564 "-fsanitize-sections=.data",
6565 ] {
6566 let (opts, _) = compile(&[taken, "-c", "a.c"]);
6567 assert_eq!(opts.safety, rucc_session::Safety::Off, "{taken} asks for no checking");
6568 }
6569 assert!(refused(&["-fsanitize-recover=bogus", "-c", "a.c"]).contains("is not a sanitizer"));
6570
6571 // Coverage instrumentation is refused rather than dropped, because a fuzzer with no
6572 // feedback runs blind and never says so.
6573 let failed = refused(&["-fsanitize-coverage=trace-pc", "-c", "a.c"]);
6574 assert!(failed.contains("feedback"), "{failed}");
6575 let failed = refused(&["-fsanitize-coverage=trace-pc-guard", "-c", "a.c"]);
6576 assert!(failed.contains("trace-pc or trace-cmp"), "gcc takes two of them: {failed}");
6577 }
6578
6579 #[test]
6580 fn the_levels_gcc_spells_differently_are_the_levels_they_mean() {
6581 assert_eq!(compile(&["-O", "-c", "a.c"]).0.opt_level, OptLevel::O1);
6582 assert_eq!(compile(&["-Og", "-c", "a.c"]).0.opt_level, OptLevel::O1);
6583 assert_eq!(compile(&["-O2", "-c", "a.c"]).0.opt_level, OptLevel::O2);
6584 }
6585
6586 #[test]
6587 fn the_machine_flags_that_name_what_we_already_do_are_taken_and_the_rest_are_not() {
6588 let line = ["--target=x86_64-unknown-linux-gnu", "-m64", "-march=x86-64-v3"];
6589 let (opts, _) =
6590 compile(&[&line[..], &["-mtune=native", "-mabi=sysv", "-c", "a.c"]].concat());
6591 assert_eq!(opts.target.to_string(), "x86_64-unknown-linux-gnu");
6592 let wrong = refused(&["--target=x86_64-unknown-linux-gnu", "-mabi=ms", "-c", "a.c"]);
6593 assert!(wrong.contains("sysv convention"), "{wrong}");
6594 }
6595
6596 /// Whether a unit built with that command line has the extension called `name`.
6597 fn has(line: &[&str], name: &str) -> bool {
6598 let x86 = ["--target=x86_64-unknown-linux-gnu", "-c", "a.c"];
6599 let (opts, _) = compile(&[&x86[..], line].concat());
6600 opts.isa.has(rucc_target::Feature::named(name).expect("a feature"))
6601 }
6602
6603 #[test]
6604 fn the_sse_flags_and_the_processor_levels_name_extensions() {
6605 // tamnd/rucc#2003. Every one of these was an unknown option before, and Postgres's
6606 // configure probe for the CRC-32C intrinsics is compiled with the first.
6607 assert!(has(&["-msse4.2"], "sse4.2") && has(&["-msse4.2"], "crc32"));
6608 assert!(has(&["-msse4.2"], "ssse3") && has(&["-msse4.2"], "popcnt"));
6609 assert!(!has(&[], "sse3") && !has(&[], "popcnt"));
6610 assert!(has(&["-mssse3"], "sse3") && !has(&["-mssse3"], "sse4.1"));
6611 assert!(has(&["-msse4"], "sse4.2") && !has(&["-msse4", "-mno-sse4"], "sse4.1"));
6612 assert!(has(&["-mpopcnt"], "popcnt") && !has(&["-mpopcnt"], "sse3"));
6613 assert!(has(&["-mcrc32"], "crc32"));
6614 assert!(has(&["-mxsave"], "xsave") && !has(&["-mxsave", "-mno-xsave"], "xsave"));
6615 assert!(!has(&["-msse4.2", "-mno-popcnt"], "popcnt"));
6616 // A processor supplies what no flag spoke for, whichever order they came in.
6617 assert!(has(&["-march=x86-64-v2"], "sse4.2"));
6618 assert!(!has(&["-march=x86-64-v2", "-mno-sse4.2"], "sse4.2"));
6619 assert!(!has(&["-mno-sse4.2", "-march=x86-64-v2"], "sse4.2"));
6620 assert!(has(&["-mno-sse4.2", "-march=x86-64-v2"], "sse4.1"));
6621 assert!(!has(&["-march=x86-64-v2", "-march=x86-64"], "sse3"));
6622 // One it has no list for is the baseline, as it was when all of them were.
6623 assert!(!has(&["-march=pentium-m"], "sse3"));
6624 assert!(has(&["-march=x86-64-v3"], "avx2"));
6625 // Turning off what is never on is nothing, and the flag is still gcc's.
6626 assert!(!has(&["-mno-avx512f"], "avx512f"));
6627 }
6628
6629 #[test]
6630 fn an_extension_this_compiler_cannot_provide_for_a_whole_unit_is_refused() {
6631 let x86 = ["--target=x86_64-unknown-linux-gnu", "-c", "a.c"];
6632 let said = refused(&[&x86[..], &["-mavx2"]].concat());
6633 assert!(said.contains("no intrinsics for avx2"), "{said}");
6634 let said = refused(&[&x86[..], &["-mno-sse2"]].concat());
6635 assert!(said.contains("baseline"), "{said}");
6636 assert!(refused(&[&x86[..], &["-msse5"]].concat()).contains("unknown option"));
6637 // No other target has these, whichever side of the target the flag was written on.
6638 let said = refused(&["-msse4.2", "--target=aarch64-linux-gnu", "-c", "a.c"]);
6639 assert!(said.contains("unknown option `-msse4.2`"), "{said}");
6640 let (opts, _) = compile(&["--target=aarch64-linux-gnu", "-march=armv8-a+crc", "-c", "a.c"]);
6641 assert_eq!(opts.isa, rucc_target::Isa::NONE);
6642 }
6643
6644 #[test]
6645 fn the_thread_flag_is_a_macro_and_a_library_and_the_library_goes_last() {
6646 let (opts, plan) = compile(&["-pthread", "-c", "a.c"]);
6647 assert!(opts.defines.iter().any(|d| d == "_REENTRANT"));
6648 // After the input, because a static link takes what it needs from a library when it
6649 // reaches it and not afterwards.
6650 let names: Vec<&str> = plan.jobs.iter().map(|j| j.input.as_str()).collect();
6651 assert_eq!(names, vec!["a.c"]);
6652 }
6653
6654 #[test]
6655 fn the_version_banner_keeps_our_first_line_and_takes_meson_down_the_gnu_path() {
6656 let text = banner();
6657 let mut lines = text.lines();
6658 // Every harness we have reads the first line and nothing else.
6659 assert_eq!(lines.next(), Some(format!("rucc {VERSION}").as_str()));
6660 // The words meson looks for, in `mesonbuild/compilers/detect.py`.
6661 assert!(text.contains("Free Software Foundation"), "{text}");
6662 // GCC's own banner has three lines and so does this one, and the claim is the dialect.
6663 assert!(lines.next().is_some_and(|l| l.contains("GCC 16")), "{text}");
6664 assert!(lines.next().is_some() && lines.next().is_none(), "{text}");
6665 }
6666
6667 #[test]
6668 fn the_questions_a_build_system_asks_before_it_compiles_anything() {
6669 let target = "--target=x86_64-unknown-linux-gnu";
6670 assert_eq!(printed(&[target, "-dumpmachine"]), "x86_64-unknown-linux-gnu");
6671 assert_eq!(printed(&[target, "-dumpversion"]), "16");
6672 assert_eq!(printed(&[target, "-dumpfullversion"]), "16.0.0");
6673 // They follow the release claimed, since that is the one `__GNUC__` says.
6674 assert_eq!(printed(&[target, "-fgnuc-version=15.2", "-dumpversion"]), "15");
6675 assert_eq!(printed(&[target, "-fgnuc-version=15.2", "-dumpfullversion"]), "15.2.0");
6676 assert_eq!(printed(&[target, "-print-multiarch"]), "x86_64-linux-gnu");
6677 // A name nothing holds comes back unchanged, which is GCC's rule and is what makes the
6678 // answer safe to paste into a link line whether or not the file is there.
6679 assert_eq!(printed(&[target, "-print-file-name=no-such-library.a"]), "no-such-library.a");
6680 assert_eq!(printed(&[target, "-print-prog-name=ld"]), "ld");
6681 let dirs = printed(&[target, "-print-search-dirs"]);
6682 assert!(dirs.starts_with("install: "), "{dirs}");
6683 assert!(dirs.contains("\nlibraries: ="), "{dirs}");
6684 }
6685
6686 #[test]
6687 fn the_sysroot_in_effect_is_the_one_the_command_line_named_or_the_one_for_the_target() {
6688 // A tree the user named is the answer whatever the target is, because it is the answer to
6689 // every other question too.
6690 assert_eq!(printed(&["--sysroot=/opt/cross", "-print-sysroot"]), "/opt/cross");
6691
6692 // A target that is no machine this suite runs on is read under the cache, and the answer is
6693 // the root rather than one of the directories under it, since what asks is looking for a
6694 // file of its own.
6695 let root = cache::dir().join("sysroots").join("riscv64-linux-musl");
6696 assert_eq!(
6697 printed(&["--target=riscv64-linux-musl", "-print-sysroot"]),
6698 root.display().to_string()
6699 );
6700
6701 // And a compile for this machine has no sysroot, which is the empty line GCC prints when it
6702 // was configured without one rather than a `/` that would be a claim about the filesystem.
6703 // Except on Windows, which has no C library of its own and reads the fetched tree for its
6704 // own target as it would for any other.
6705 let host = Triple::host().expect("a host this compiler knows");
6706 let own = printed(&[&format!("--target={host}"), "-print-sysroot"]);
6707 if host.os == rucc_target::Os::Windows {
6708 let root = cache::dir().join("sysroots").join(host.tuple().to_string());
6709 assert_eq!(own, root.display().to_string());
6710 } else {
6711 assert_eq!(own, "");
6712 }
6713 }
6714
6715 #[test]
6716 fn the_provenance_of_a_sysroot_is_the_manifest_it_carries() {
6717 // Section 13.5 wants seven things per input and wants them machine readable, and the manifest
6718 // is the record that already has them, so the flag prints that rather than a second format.
6719 let manifest = "rucc sysroot manifest 3\n\
6720 target\tx86_64-linux-musl\n\
6721 kernel\t6.12\n\
6722 include/generic/stdio.h\tmusl-1.2.5\t\
6723 https://musl.libc.org/releases/musl-1.2.5.tar.gz\t\
6724 0000000000000000000000000000000000000000000000000000000000000000\tmit\t\
6725 bundled\n\
6726 lib/libc.so\tmusl-1.2.5\t\
6727 https://musl.libc.org/releases/musl-1.2.5.tar.gz\t\
6728 1111111111111111111111111111111111111111111111111111111111111111\tmit\t\
6729 generated\n";
6730 let tree = TempTree::new("provenance", &[("manifest", manifest)]);
6731 let sysroot = format!("--sysroot={}", tree.0.display());
6732 // The kernel line of tamnd/rucc#934 is in the answer without anything here naming it, because
6733 // the flag parses the record and renders it again rather than picking fields out of it. That
6734 // is the reason it prints a manifest and not a format of its own.
6735 //
6736 // The answer is the file without its last newline, because whatever prints it adds one. The
6737 // file is what somebody diffs the output against, so the two have to be the same bytes.
6738 assert_eq!(printed(&[&sysroot, "-print-sysroot-provenance"]) + "\n", manifest);
6739
6740 // A tree with no manifest in it is a tree somebody assembled themselves, and nothing here
6741 // knows where any of it came from. Saying nothing is the only honest answer, and a reader can
6742 // tell it from a manifest with no inputs because that one still has its two header lines.
6743 let bare = TempTree::new("provenance-bare", &[]);
6744 assert_eq!(
6745 printed(&[&format!("--sysroot={}", bare.0.display()), "-print-sysroot-provenance"]),
6746 ""
6747 );
6748
6749 // And a compile for this machine has no sysroot at all, which is the same empty answer
6750 // `-print-sysroot` gives for it.
6751 let host = Triple::host().expect("a host this compiler knows");
6752 assert_eq!(printed(&[&format!("--target={host}"), "-print-sysroot-provenance"]), "");
6753
6754 // And the other spelling, which section 13.5 is the document that writes.
6755 assert_eq!(printed(&[&sysroot, "--print-sysroot-provenance"]) + "\n", manifest);
6756
6757 // tamnd/rucc#1021. The digest of the same tree is the sha256 of that record, so it is one
6758 // line where the provenance is a few hundred, and it is checkable with `sha256sum` because
6759 // the bytes it is over are the bytes of the file. The number here is that hash of the
6760 // fixture above, computed by `sha256sum` rather than by this compiler.
6761 assert_eq!(
6762 printed(&[&sysroot, "-print-sysroot-digest"]),
6763 "d705ae6ebeafeb7fda4bd57cecc7882bf49784b17015664a09cfae25a1b2000a"
6764 );
6765 assert_eq!(
6766 printed(&[&sysroot, "--print-sysroot-digest"]),
6767 printed(&[&sysroot, "-print-sysroot-digest"])
6768 );
6769
6770 // And the two empty answers are empty here too, because a digest of nothing would read as a
6771 // claim about a sysroot rather than as the absence of one.
6772 assert_eq!(
6773 printed(&[&format!("--sysroot={}", bare.0.display()), "-print-sysroot-digest"]),
6774 ""
6775 );
6776 assert_eq!(printed(&[&format!("--target={host}"), "-print-sysroot-digest"]), "");
6777 }
6778
6779 #[test]
6780 fn a_manifest_this_build_cannot_read_is_refused_rather_than_printed() {
6781 // Passing a file we could not parse to whoever asked would make their parser the one that
6782 // finds the problem, and the three uses section 13.5 gives for this are all somebody else
6783 // parsing it.
6784 let tree = TempTree::new(
6785 "provenance-bad",
6786 &[("manifest", "rucc sysroot manifest 3\ntarget\tx86_64-linux-musl\nlib/libc.a\n")],
6787 );
6788 let message =
6789 refused(&[&format!("--sysroot={}", tree.0.display()), "-print-sysroot-provenance"]);
6790 assert!(message.contains("manifest"), "{message}");
6791 assert!(message.contains("1 fields where an input has six"), "{message}");
6792
6793 // The digest is refused for the same file and for a stronger reason: a hash of bytes this
6794 // build cannot read would be a number that names a record nobody can act on.
6795 let digest =
6796 refused(&[&format!("--sysroot={}", tree.0.display()), "-print-sysroot-digest"]);
6797 assert_eq!(digest, message);
6798 }
6799
6800 #[test]
6801 fn the_two_dependency_flags_that_stop_after_the_rule_stop_after_the_rule() {
6802 let (opts, _) = compile(&["-M", "a.c"]);
6803 assert!(opts.deps.emit && opts.deps.instead_of_compiling);
6804 assert!(opts.deps.system_headers, "plain -M lists them");
6805 assert_eq!(opts.emit, EmitKind::Preprocessed);
6806
6807 // Even where a later flag asked for something else, because the family is a mode and
6808 // the mode is what the run is for.
6809 let (opts, _) = compile(&["-M", "-c", "a.c"]);
6810 assert_eq!(opts.emit, EmitKind::Preprocessed);
6811
6812 let (opts, _) = compile(&["-MM", "a.c"]);
6813 assert!(!opts.deps.system_headers);
6814 }
6815
6816 #[test]
6817 fn the_two_that_end_in_d_leave_the_compilation_alone() {
6818 let (opts, _) = compile(&["-MD", "-c", "a.c"]);
6819 assert!(opts.deps.emit && !opts.deps.instead_of_compiling);
6820 assert!(opts.deps.system_headers);
6821 assert_eq!(opts.emit, EmitKind::Object);
6822
6823 let (opts, _) = compile(&["-MMD", "-c", "a.c"]);
6824 assert!(opts.deps.emit && !opts.deps.instead_of_compiling);
6825 assert!(!opts.deps.system_headers);
6826 }
6827
6828 #[test]
6829 fn nothing_puts_the_system_headers_back_once_a_flag_has_taken_them_out() {
6830 // GCC's rule, and not an oversight in it. The flag asking for fewer of them is read as
6831 // the answer, because the other one never asked the question.
6832 let (opts, _) = compile(&["-MM", "-M", "a.c"]);
6833 assert!(!opts.deps.system_headers);
6834 let (opts, _) = compile(&["-MD", "-MMD", "-c", "a.c"]);
6835 assert!(!opts.deps.system_headers);
6836 let (opts, _) = compile(&["-MMD", "-MD", "-c", "a.c"]);
6837 assert!(!opts.deps.system_headers);
6838 }
6839
6840 #[test]
6841 fn a_target_arrives_escaped_from_one_flag_and_untouched_from_the_other() {
6842 let (opts, _) = compile(&["-MM", "-MT", "a b.o", "-MQ", "a b.o", "a.c"]);
6843 assert_eq!(opts.deps.targets, vec!["a b.o".to_owned(), "a\\ b.o".to_owned()]);
6844 }
6845
6846 #[test]
6847 fn the_rest_of_the_family_is_a_file_and_a_switch() {
6848 let (opts, _) = compile(&["-MM", "-MF", "dep.d", "-MP", "a.c"]);
6849 assert_eq!(opts.deps.file.as_deref(), Some("dep.d"));
6850 assert!(opts.deps.phony);
6851
6852 for flag in ["-MF", "-MT", "-MQ"] {
6853 let e = parse_args(&args(&[flag])).unwrap_err();
6854 assert!(e.message.contains("requires an argument"), "{}", e.message);
6855 }
6856 }
6857
6858 /// Kbuild's spelling, which is how busybox and the kernel ask for every dependency file.
6859 #[test]
6860 fn a_dependency_file_asked_for_through_the_preprocessor_is_written_where_it_said() {
6861 let (opts, _) = compile(&["-Wp,-MD,applets/.applets.o.d", "-c", "a.c"]);
6862 assert!(opts.deps.emit);
6863 assert!(opts.deps.system_headers);
6864 assert_eq!(opts.deps.file.as_deref(), Some("applets/.applets.o.d"));
6865
6866 let (opts, _) = compile(&["-Wp,-MMD,x.d,-MP,-MT,x.o", "-c", "a.c"]);
6867 assert!(!opts.deps.system_headers);
6868 assert!(opts.deps.phony);
6869 assert_eq!(opts.deps.file.as_deref(), Some("x.d"));
6870 assert_eq!(opts.deps.targets, vec!["x.o".to_owned()]);
6871 }
6872
6873 #[test]
6874 fn a_preprocessor_flag_this_compiler_does_not_read_is_still_refused_whole() {
6875 assert!(refused(&["-Wp,-MD", "-c", "a.c"]).contains("separate assembler"));
6876 assert!(refused(&["-Wp,-MD,x.d,-C", "-c", "a.c"]).contains("-Wp,-MD,x.d,-C"));
6877 }
6878
6879 /// A directory of sources for one test, removed when the test is done with it.
6880 struct TempTree(PathBuf);
6881
6882 impl Drop for TempTree {
6883 fn drop(&mut self) {
6884 let _ = std::fs::remove_dir_all(&self.0);
6885 }
6886 }
6887
6888 impl TempTree {
6889 fn new(name: &str, files: &[(&str, &str)]) -> TempTree {
6890 let dir = std::env::temp_dir().join(format!("rucc-deps-{}-{name}", std::process::id()));
6891 let _ = std::fs::remove_dir_all(&dir);
6892 std::fs::create_dir_all(&dir).expect("temporary directory should be writable");
6893 for (path, text) in files {
6894 let at = dir.join(path);
6895 if let Some(parent) = at.parent() {
6896 std::fs::create_dir_all(parent).expect("creating a subdirectory should work");
6897 }
6898 std::fs::write(&at, text).expect("writing a temporary file should work");
6899 }
6900 TempTree(dir)
6901 }
6902
6903 fn path(&self, name: &str) -> String {
6904 self.0.join(name).to_string_lossy().into_owned()
6905 }
6906 }
6907
6908 #[test]
6909 fn the_rule_names_what_the_includes_found_and_names_each_of_them_once() {
6910 // End to end, because the list comes from the preprocessor and the format comes from
6911 // somewhere else, and a test of either half on its own would pass with the two of them
6912 // wired up backwards.
6913 let tree = TempTree::new(
6914 "found",
6915 &[
6916 ("a.c", "#include \"one.h\"\n#include \"two.h\"\nint main(void) { return X; }\n"),
6917 ("one.h", "#define X 0\n"),
6918 ("two.h", "#include \"one.h\"\n"),
6919 ],
6920 );
6921 let out = tree.path("dep.d");
6922 let code = run(&args(&["-MM", "-MF", &out, "-o", &tree.path("a.i"), &tree.path("a.c")]));
6923 assert_eq!(code, 0);
6924
6925 let text = std::fs::read_to_string(&out).expect("the rule should have been written");
6926 let names: Vec<&str> = text.split_whitespace().collect();
6927 // The target, the source, and each header once however many times it was reached.
6928 assert_eq!(names.first(), Some(&"a.o:"), "{text}");
6929 assert_eq!(names.iter().filter(|n| n.ends_with("one.h")).count(), 1, "{text}");
6930 assert_eq!(names.iter().filter(|n| n.ends_with("two.h")).count(), 1, "{text}");
6931 // And the `-o` went to the file the rule replaced, which is left empty rather than
6932 // absent because a makefile that named it as a target will look for it.
6933 assert_eq!(std::fs::read(tree.path("a.i")).expect("the output should exist"), b"");
6934 }
6935
6936 #[test]
6937 fn syntax_only_checks_the_file_and_writes_nothing() {
6938 // What meson's `has_header_symbol` probe does: compile with `-fsyntax-only` and read the
6939 // exit status. A good file passes and leaves no output behind, a bad one fails.
6940 let tree = TempTree::new(
6941 "syntax-only",
6942 &[
6943 ("good.c", "int f(int x) { return x + 1; }\n"),
6944 ("bad.c", "int f(void) { return y; }\n"),
6945 ],
6946 );
6947 let (opts, _) = compile(&["-fsyntax-only", "a.c"]);
6948 assert_eq!(opts.emit, EmitKind::SyntaxOnly);
6949
6950 let out = tree.path("good.o");
6951 assert_eq!(run(&args(&["-fsyntax-only", "-o", &out, &tree.path("good.c")])), 0);
6952 assert!(!std::path::Path::new(&out).exists(), "-fsyntax-only wrote {out}");
6953 assert!(!std::path::Path::new(&tree.path("good.s")).exists());
6954 assert_ne!(run(&args(&["-fsyntax-only", &tree.path("bad.c")])), 0);
6955 }
6956
6957 /// Where `-fstack-usage` puts each job's report, one entry per job, for a command line.
6958 fn stack_usage_files(line: &[&str]) -> Vec<Option<String>> {
6959 let mut words = vec![LINUX, "-fstack-usage"];
6960 words.extend_from_slice(line);
6961 let (_, plan) = compile(&words);
6962 plan.jobs.iter().map(|job| job.stack_usage.clone()).collect()
6963 }
6964
6965 #[test]
6966 fn a_stack_usage_file_is_named_the_way_gcc_names_it() {
6967 // Every row was run through gcc 16 with the same command line, and the name is the one it
6968 // wrote. `rpg frames` finds gcc's file and this compiler's by the same rule, so a name that
6969 // differs is a function that goes missing from the comparison.
6970 let cases: &[(&[&str], &[Option<&str>])] = &[
6971 (&["-c", "sub/a.c"], &[Some("a.su")]),
6972 (&["-c", "sub/a.c", "-o", "out/x.o"], &[Some("out/x.su")]),
6973 (&["-c", "sub/a.c", "b.c"], &[Some("a.su"), Some("b.su")]),
6974 (&["-S", "sub/a.c", "-o", "out/y.s"], &[Some("out/y.su")]),
6975 (&["-S", "sub/a.c", "-o", "-"], &[Some("a.su")]),
6976 (&["-E", "sub/a.c", "-o", "out/z.i"], &[None]),
6977 (&["-fsyntax-only", "sub/a.c"], &[Some("a.su")]),
6978 (&["-fsyntax-only", "sub/a.c", "-o", "out/x.o"], &[Some("out/x.o-a.su")]),
6979 (&["sub/a.c"], &[Some("a.su")]),
6980 (&["sub/a.c", "-lm"], &[Some("a.su")]),
6981 (&["sub/a.c", "b.c"], &[Some("a-a.su"), Some("a-b.su")]),
6982 (&["sub/a.c", "b.o"], &[Some("a-a.su"), None]),
6983 (&["sub/a.c", "-o", "out/prog"], &[Some("out/prog-a.su")]),
6984 (&["sub/a.c", "-o", "out/lib.so"], &[Some("out/lib.so-a.su")]),
6985 (&["sub/a.c", "-o", "out/prog.exe"], &[Some("out/prog-a.su")]),
6986 (&["sub/a.c", "-o", "out/prog", "-dumpbase", "zz"], &[Some("out/zz-a.su")]),
6987 (&["sub/a.c", "-o", "out/prog", "-dumpdir", "dd-"], &[Some("dd-a.su")]),
6988 (
6989 &["sub/a.c", "b.c", "-dumpdir", "dd/", "-dumpbase", "zz"],
6990 &[Some("dd/zz-a.su"), Some("dd/zz-b.su")],
6991 ),
6992 (&["-c", "sub/a.c", "-dumpbase", "foo", "-o", "out/w.o"], &[Some("out/foo.su")]),
6993 (&["-c", "sub/a.c", "-dumpdir", "dd/", "-dumpbase", "sub/zz"], &[Some("sub/zz.su")]),
6994 (&["-c", "sub/a.c", "-dumpbase", "zz.c", "-dumpbase-ext", ".c"], &[Some("zz.su")]),
6995 (&["-c", "sub/a.c", "-dumpdir", "pre", "-o", "out/x.o"], &[Some("prex.su")]),
6996 (&["-c", "sub/a.c", "-save-temps=cwd", "-o", "out/x.o"], &[Some("x.su")]),
6997 ];
6998 for (line, want) in cases {
6999 let want: Vec<Option<String>> = want.iter().map(|w| w.map(str::to_owned)).collect();
7000 assert_eq!(stack_usage_files(line), want, "{line:?}");
7001 }
7002 // Nothing at all without the flag.
7003 let (_, plan) = compile(&[LINUX, "-c", "sub/a.c"]);
7004 assert_eq!(plan.jobs[0].stack_usage, None);
7005 }
7006
7007 #[test]
7008 fn a_stack_usage_file_has_a_line_per_function_where_gcc_would_put_it() {
7009 let tree = TempTree::new(
7010 "stack-usage",
7011 &[
7012 ("inc/h.h", "static inline int twice(int x) { return x * 2; }\n"),
7013 (
7014 "a.c",
7015 "#include \"inc/h.h\"\n\
7016 static int helper(int);\n\
7017 int grows(int n) { char v[n]; v[0] = (char)n; return v[n - 1] + twice(n); }\n\
7018 static int\n\
7019 helper(int x)\n\
7020 {\n\
7021 return x + 1;\n\
7022 }\n\
7023 int calls(int x) { return helper(x) + grows(x); }\n",
7024 ),
7025 ],
7026 );
7027 let (source, object) = (tree.path("a.c"), tree.path("a.o"));
7028 assert_eq!(run(&args(&["-O0", "-fstack-usage", "-c", &source, "-o", &object])), 0);
7029 let text = std::fs::read_to_string(tree.path("a.su")).expect("a.su should be written");
7030
7031 let line = |function: &str| {
7032 let suffix = format!(":{function}");
7033 let line =
7034 text.lines().find(|line| line.split('\t').next().unwrap().ends_with(&suffix));
7035 line.unwrap_or_else(|| panic!("no line for {function} in\n{text}"))
7036 };
7037 let expect = |function: &str, at: String, qualifier: &str| {
7038 let fields: Vec<&str> = line(function).split('\t').collect();
7039 assert_eq!(fields.len(), 3, "{text}");
7040 assert_eq!(fields[0], format!("{at}:{function}"), "{text}");
7041 let bytes: u32 = fields[1].parse().expect("the bytes should be a number");
7042 assert!(bytes >= 8 && bytes % 8 == 0, "{function} takes {bytes} bytes");
7043 assert_eq!(fields[2], qualifier, "{text}");
7044 };
7045 // A variable length array makes the frame grow while the function runs.
7046 expect("grows", format!("{source}:3:5"), "dynamic");
7047 // The definition rather than the declaration above it, and the line the name is on
7048 // rather than the one the type is on.
7049 expect("helper", format!("{source}:5:1"), "static");
7050 expect("calls", format!("{source}:9:5"), "static");
7051 // A function from a header is reported against the header.
7052 expect("twice", format!("{}:1:19", tree.path("inc/h.h")), "static");
7053 assert_eq!(text.lines().count(), 4, "{text}");
7054 }
7055
7056 #[test]
7057 fn a_stack_usage_file_is_empty_when_there_is_nothing_to_report_and_absent_under_dash_e() {
7058 let tree = TempTree::new(
7059 "stack-usage-empty",
7060 &[
7061 ("good.c", "int f(int x) { return x + 1; }\n"),
7062 ("bad.c", "int f(void) { return y; }\n"),
7063 ],
7064 );
7065 let good = tree.path("good.c");
7066 // gcc writes an empty file for a check that compiles nothing and for a file that failed,
7067 // and a build that looks for one beside every object finds one.
7068 assert_eq!(
7069 run(&args(&["-fstack-usage", "-fsyntax-only", &good, "-o", &tree.path("x")])),
7070 0
7071 );
7072 assert_eq!(std::fs::read_to_string(tree.path("x-good.su")).unwrap(), "");
7073 let bad = tree.path("bad.c");
7074 assert_ne!(run(&args(&["-fstack-usage", "-c", &bad, "-o", &tree.path("bad.o")])), 0);
7075 assert_eq!(std::fs::read_to_string(tree.path("bad.su")).unwrap(), "");
7076 // And none under `-E`, which never reaches a function.
7077 assert_eq!(run(&args(&["-fstack-usage", "-E", &good, "-o", &tree.path("e.i")])), 0);
7078 assert!(!std::path::Path::new(&tree.path("e.su")).exists());
7079 }
7080
7081 #[test]
7082 fn a_header_that_is_only_reached_under_a_guard_is_still_a_dependency() {
7083 // The multiple-include optimization means the second reach never opens the file. It is
7084 // still a file this translation unit was built from, so it is still in the rule.
7085 let tree = TempTree::new(
7086 "guarded",
7087 &[
7088 ("a.c", "#include \"g.h\"\n#include \"g.h\"\nint main(void) { return 0; }\n"),
7089 ("g.h", "#ifndef G\n#define G\n#endif\n"),
7090 ],
7091 );
7092 let out = tree.path("dep.d");
7093 let code = run(&args(&["-MM", "-MF", &out, "-o", &tree.path("a.i"), &tree.path("a.c")]));
7094 assert_eq!(code, 0);
7095 let text = std::fs::read_to_string(&out).expect("the rule should have been written");
7096 assert_eq!(text.split_whitespace().filter(|n| n.ends_with("g.h")).count(), 1, "{text}");
7097 }
7098
7099 #[test]
7100 fn every_imacros_file_is_read_before_every_include_file_whatever_order_they_were_written() {
7101 // Measured against GCC rather than read: the two flags the other way round produce the
7102 // same output byte for byte, so the command line order between the two families does not
7103 // decide anything and the order within one does. The `-include` file here can only see
7104 // the definition if the `-imacros` file that was written after it ran first.
7105 let tree = TempTree::new(
7106 "preinclude",
7107 &[
7108 ("a.c", "int main(void) { return 0; }\n"),
7109 ("i.h", "#ifdef FROM_MACROS\nint saw_it;\n#else\nint missed_it;\n#endif\n"),
7110 ("m.h", "#define FROM_MACROS 1\nint macros_text;\n"),
7111 ],
7112 );
7113 let out = tree.path("a.i");
7114 let code = run(&args(&[
7115 "-E",
7116 "-include",
7117 &tree.path("i.h"),
7118 "-imacros",
7119 &tree.path("m.h"),
7120 "-o",
7121 &out,
7122 &tree.path("a.c"),
7123 ]));
7124 assert_eq!(code, 0);
7125 let text = std::fs::read_to_string(&out).expect("the output should have been written");
7126 assert!(text.contains("saw_it"), "{text}");
7127 // And the text of the `-imacros` file is thrown away, which is the whole difference
7128 // between the two flags.
7129 assert!(!text.contains("macros_text"), "{text}");
7130 }
7131
7132 #[test]
7133 fn a_file_the_command_line_named_is_a_prerequisite_the_same_as_one_a_directive_named() {
7134 let tree = TempTree::new(
7135 "preinclude-deps",
7136 &[
7137 ("a.c", "int main(void) { return 0; }\n"),
7138 ("i.h", "int from_include;\n"),
7139 ("m.h", "#define M 1\n"),
7140 ],
7141 );
7142 let out = tree.path("dep.d");
7143 let code = run(&args(&[
7144 "-MM",
7145 "-MF",
7146 &out,
7147 "-include",
7148 &tree.path("i.h"),
7149 "-imacros",
7150 &tree.path("m.h"),
7151 "-o",
7152 &tree.path("a.i"),
7153 &tree.path("a.c"),
7154 ]));
7155 assert_eq!(code, 0);
7156 let text = std::fs::read_to_string(&out).expect("the rule should have been written");
7157 assert!(text.contains("i.h"), "{text}");
7158 assert!(text.contains("m.h"), "{text}");
7159 }
7160
7161 #[test]
7162 fn a_command_line_include_that_is_nowhere_on_the_path_is_an_error_and_not_a_warning() {
7163 // Including the directory of the source file, which is not on the path for these: the
7164 // command line was not written there, so a name in it is relative to where the compiler
7165 // was run rather than to where the source sits.
7166 let tree = TempTree::new(
7167 "preinclude-missing",
7168 &[("sub/a.c", "int main(void) { return 0; }\n"), ("sub/beside.h", "int x;\n")],
7169 );
7170 let code = run(&args(&["-E", "-include", "beside.h", "-o", "-", &tree.path("sub/a.c")]));
7171 assert_eq!(code, 1);
7172 }
7173
7174 #[test]
7175 fn a_command_line_that_links_names_the_executable_and_not_the_object_it_went_through() {
7176 // The object a link goes through is in a temporary directory and is gone before `make`
7177 // reads any of this, so the rule that named it would be a rule for a file that is never
7178 // there. The target and the file are both the `-o`, which is the executable.
7179 let (opts, plan) = compile(&["-MD", "sub/a.c", "-o", "prog"]);
7180 assert_eq!(plan.output.as_deref(), Some("prog"));
7181 assert_eq!(deps::default_target("sub/a.c", deps_target_output(&opts, &plan)), "prog");
7182 assert_eq!(
7183 deps::default_file(&opts.deps, "sub/a.c", plan.output.as_deref()).as_deref(),
7184 Some("prog.d")
7185 );
7186 }
7187
7188 #[test]
7189 fn the_plan_keeps_the_output_name_because_the_rule_is_written_from_it() {
7190 let (_, plan) = compile(&["-MMD", "-c", "sub/a.c", "-o", "obj/x.o"]);
7191 assert_eq!(plan.output.as_deref(), Some("obj/x.o"));
7192 let (_, plan) = compile(&["-MMD", "-c", "sub/a.c"]);
7193 assert_eq!(plan.output, None);
7194 }
7195
7196 #[test]
7197 fn usage_fits_on_a_screen() {
7198 // Not a style preference. A help text that scrolls is one nobody reads, and this is
7199 // the cheapest way to keep it honest as flags accumulate. The number goes up only when
7200 // a family of flags arrives that has nowhere to share a line, which the two pass gates
7201 // were and which the two fuel flags and `-fsafety=` now are, and it goes up by exactly
7202 // the lines that family took. The four it went up by last are the flags a build system
7203 // passes without being asked to: how much to say, what machine to generate for, threads,
7204 // and the questions `configure` asks before it compiles anything. The one it went up by
7205 // last is the second line of `--emit`, whose kinds are a family that has now outgrown
7206 // one line and has nowhere else to go. The two it went up by last are the dependency
7207 // family, which is eight flags that share nothing with anything above them. The one it
7208 // went up by last is the four spellings of position independent code, which every
7209 // configure script writes and which could only have shared the link line, and that line
7210 // is already four characters short of the limit. The two it went up by last are the rest
7211 // of the include family, which is six more flags that change where a header is looked for
7212 // and two that name a header outright. The one it went up by last is the pair that keeps
7213 // the intermediate files and times the steps, which belong next to the two flags above
7214 // them that are also about watching a compilation rather than changing one. The two it
7215 // went up by last are the section flags and the visibility flag, which are what a build
7216 // that cares about the size of what it ships and about which names it exports writes, and
7217 // the second of them was already taken and only missing from here. The one it went up by
7218 // last is the stack protector, which is four spellings of one question and which every
7219 // distribution puts on every command line it issues, so a build that reads this list
7220 // looking for it and does not find it has to go and read the specification instead. The one
7221 // it went up by last is the profiler, which is two spellings of the request and two of
7222 // where the call goes, and which is about watching a program run rather than about what is
7223 // generated, so it shares its subject with nothing above it. The one it went up by last is
7224 // the room a function opens with for something to be written over it later, which takes an
7225 // argument of its own shape and is what a kernel build asks for, so it fits beside the
7226 // profiler and nothing else. The one it went up by last is what overflows rather than being
7227 // undefined, which is three spellings of two questions and which a kernel build and a great
7228 // deal of code written before the standard settled both pass. The one it went up by last is
7229 // the other answer to the first of those questions, which could not share the line because
7230 // what it asks for is the opposite of what the flags on that line ask for. The one it went
7231 // up by last is the split of the line that lists what this compiler does anyway into that
7232 // and what it assumes anyway, which are two different claims that were sharing a line until
7233 // the second of them got a second flag and the line stopped fitting. The one it went up by
7234 // last is the three flags that change the ABI rather than the code, which have to be given
7235 // to every file in a program or none of them and which therefore belong somewhere a person
7236 // reading this list will see them. The one it went up by last is the floating point group,
7237 // which is two lines rather than one because the first of them is a choice this compiler
7238 // records and the rest are claims about what it does anyway, and putting a real setting on
7239 // the same line as three flags that change nothing would be misleading about both. The one
7240 // it went up by last is the flag that says a write has to stay inside the member it names,
7241 // which is a setting rather than a claim and so cannot share the line above it, that being
7242 // the one that picks a tier. The two it went up by last are the prefix mapping family,
7243 // which is four flags whose whole job is to keep a build's output the same from two
7244 // different directories, and which a person chasing a reproducible build comes here
7245 // looking for by name. The one it went up by last is how the debug sections are compressed
7246 // and whether they go in a file of their own, which are two questions about the shape of
7247 // the debug output, where the line above them is about how much of it there is. The one it
7248 // went up by last is the `restrict` contract, which is a setting for the same reason the
7249 // flag that keeps a write inside its member is and which is the check a person who has been
7250 // bitten by a vectorizer comes here looking for. The one it went up by last is link time
7251 // optimization, which is a whole optimization rather than a flag and which says so on its
7252 // own line, because a build that passes it and reads this looking for what it got is
7253 // asking a question no other line here answers. The one it went up by last is the sysroot,
7254 // which is the question somebody asks when a cross build read a file nobody expected, and
7255 // which has no room on the line above it because the answers there are a path each and this
7256 // one is the root all of them are under. The one it went up by last is what is inside that
7257 // root and where each of it came from, which is a question about a whole tree rather than
7258 // about a path and which is long enough on its own that it could not have shared a line with
7259 // anything. The one it went up by last is the profile family, which splits down the middle
7260 // where no other family here does, so the line has to name the half that is taken and the
7261 // half that is refused or it would be read as taking both. The one it went up by last is
7262 // the sanitizers, which are what somebody reaching for a checked build writes first and
7263 // which belong beside the tier that is the nearest thing here to what they asked for. The
7264 // one it went up by last is the digest of that record, which is the same tree as one number
7265 // and could not share the line above it because that line prints a few hundred lines and
7266 // this one prints sixty four characters, and a reader who wants the short answer is looking
7267 // for it by name rather than reading the long one. The one it went up by last is the
7268 // sysroot fetch, which is the only command here that gets something from somewhere else and
7269 // is therefore the one a person wants to have read before they run it rather than after.
7270 // And the flag beside it that forbids every download, which earns its line by being what a
7271 // build in a sealed environment passes and by meaning something even though an ordinary
7272 // compile downloads nothing either way. The one it went up by last is the other fetch, the
7273 // one behind Microsoft's licence wall, which is a line rather than a paragraph because what
7274 // a person needs from here is that the command exists and that it will not do anything
7275 // until they have read a licence it prints for them.
7276 assert!(USAGE.lines().count() < 73, "usage text has grown past one screen");
7277 }
7278}