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/0.10.78")]
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;
46
47use std::fmt::Write as _;
48use std::io::Write as _;
49use std::path::PathBuf;
50
51use rucc_codegen::coverage::{self, Fired};
52use rucc_codegen::lowering::Lowerings;
53use rucc_codegen::pressure::Pressure;
54use rucc_pp::Dependency;
55use rucc_session::{
56 Compress, Control, Dumps, EmitKind, Hook, Options, Pic, PrefixMap, Preinclude, Protector,
57 SaveTemps, Session, Std, Wrapping, runtime,
58};
59use rucc_sysroot::{Manifest, Sysroot};
60use rucc_target::{ObjectFormat, Triple};
61use rucc_tuple::TargetTuple;
62
63use crate::link::LinkOptions;
64
65pub use crate::assemble::assemble;
66pub use crate::compile::{Artifact, Compiled, Temps, compile, compile_ir};
67pub use crate::phase::{ArchiveJob, Input, InputKind, Job, LinkJob, Output, Phase, Plan, Role};
68pub use crate::preprocess::{OsFileSystem, Preprocessed, preprocess};
69pub use crate::schedule::Jobs;
70
71/// The compiler's version, taken from the workspace manifest.
72pub const VERSION: &str = env!("CARGO_PKG_VERSION");
73
74/// What the command line asked for.
75#[derive(Debug, Clone, PartialEq, Eq)]
76pub enum Action {
77 /// Print usage and exit successfully.
78 Help,
79 /// Print the version and exit successfully.
80 Version,
81 /// Print one line and exit successfully, which is what the `-dump` and `-print` family do.
82 ///
83 /// A build system asks these before it compiles anything, and what it does with the answer
84 /// is paste it into a path or into another command line, so each one is a single line with
85 /// no decoration around it.
86 Print(String),
87 /// Print the resolved configuration and exit successfully.
88 PrintConfig(Box<Options>),
89 /// Print the passes the level will run and exit successfully.
90 PrintPipeline(Box<Options>),
91 /// Print the phase plan and the link line and exit successfully, which is `-###`.
92 PrintPlan {
93 /// The resolved options, which is what says what the link line is for.
94 opts: Box<Options>,
95 /// What to do to each input, and in what order.
96 plan: Box<Plan>,
97 /// What the command line said about linking.
98 link: Box<LinkOptions>,
99 },
100 /// `--fetch <tuple>`, which gets the sysroot this release pins for a target and installs it.
101 ///
102 /// The only action in this compiler that may run another program to move bytes onto the
103 /// machine, which is `spec/cross-compile/13-distribution.md` section 13.8's rule rather than a
104 /// property of how this happens to be written: a compilation has no branch that reaches it.
105 Fetch {
106 /// The artifact, from the table in [`rucc_sysroot::artifact`]. Resolved here rather than where the
107 /// work happens, so that a target nothing is pinned for is a refusal from the parser like
108 /// every other thing a command line can ask for and not have.
109 what: &'static rucc_sysroot::Pinned,
110 /// The target, which names the directory under the cache the tree is installed at and is
111 /// checked against the record inside the artifact.
112 target: TargetTuple,
113 /// Where the cache is, read where everything else that needs it reads it.
114 cache: PathBuf,
115 },
116 /// `--fetch-msvc-sdk <tuple>`, which gets what is behind Microsoft's licence wall.
117 ///
118 /// The other action that may run another program to move bytes onto the machine, and the only
119 /// one that asks a person to accept somebody else's licence first.
120 /// `spec/cross-compile/13-distribution.md` section 13.4 is why it is a command of its own
121 /// rather than something `--fetch` does when it recognises the target: no release pins an
122 /// artifact for these, and nothing about this may ever happen because a compile wanted it to.
123 FetchMsvcSdk {
124 /// The target, which says which architecture's CRT library package is wanted.
125 target: TargetTuple,
126 /// Whether `--accept-licence` was on the command line. Without it the licence and the list
127 /// are printed and nothing is downloaded, which is the whole of what the flag is for.
128 accepted: bool,
129 /// Where the cache is, read where everything else that needs it reads it.
130 cache: PathBuf,
131 },
132 /// Compile the given inputs.
133 Compile {
134 /// The resolved options.
135 opts: Box<Options>,
136 /// What to do to each input, and in what order.
137 plan: Box<Plan>,
138 /// What the command line said about linking.
139 link: Box<LinkOptions>,
140 /// How many translation units to compile at once.
141 jobs: Jobs,
142 /// Whether `-v` asked for the plan to be printed while it runs.
143 verbose: bool,
144 /// What is worth saying about the command line before anything is compiled, printed as
145 /// warnings and once for the whole run rather than once per file.
146 ///
147 /// These are not diagnostics. A diagnostic is about a piece of source and has a span to
148 /// point at, and these are about the way two flags were combined, so there is nothing to
149 /// point at and nowhere below the driver that knows both halves. `-w` does not reach them
150 /// for the same reason it does not reach a refusal from the parser.
151 notes: Vec<String>,
152 },
153}
154
155/// Why a command line was rejected.
156#[derive(Debug, Clone, PartialEq, Eq)]
157pub struct CliError {
158 /// The message, lowercase and without a trailing period, in the same shape as any other
159 /// diagnostic.
160 pub message: String,
161}
162
163impl std::fmt::Display for CliError {
164 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
165 f.write_str(&self.message)
166 }
167}
168
169impl std::error::Error for CliError {}
170
171fn err(message: impl Into<String>) -> CliError {
172 CliError { message: message.into() }
173}
174
175/// The two halves of one prefix mapping flag's argument, where `flag` includes its trailing `=`.
176///
177/// The split is at the last `=` in what follows the flag, not the first, which is gcc's rule and
178/// the only one that lets a directory whose name contains an `=` be the old half. It also means
179/// `-fmacro-prefix-map=a=b=c` rewrites `a=b` to `c` rather than `a` to `b=c`, which looks like a
180/// trap until you notice the alternative traps the far more common case.
181fn rewrite<'a>(arg: &'a str, flag: &str) -> Result<(&'a str, &'a str), CliError> {
182 let rest = &arg[flag.len()..];
183 PrefixMap::split(rest).ok_or_else(|| {
184 let flag = flag.trim_end_matches('=');
185 err(format!(
186 "`{rest}` is not a rewrite for `{flag}`, which is an old prefix, an `=` and a new one"
187 ))
188 })
189}
190
191/// A question the command line asked instead of asking for a compilation.
192///
193/// These are answered after the loop rather than where they are read, because every one of them
194/// is about the target or about the library search and the last word on both is the end of the
195/// command line.
196enum Query {
197 /// `-dumpmachine`, the triple.
198 Machine,
199 /// `-dumpversion` and `-dumpfullversion`, which are the same three numbers here.
200 Version,
201 /// `-print-multiarch`, the directory name a distribution files this target under.
202 Multiarch,
203 /// `-print-search-dirs`, in the three lines GCC prints.
204 SearchDirs,
205 /// `-print-sysroot`, the root the headers and the libraries are read under.
206 Sysroot,
207 /// `-print-sysroot-provenance`, what is in that root and where each of it came from.
208 SysrootProvenance,
209 /// `-print-sysroot-digest`, the one number that names all of it.
210 SysrootDigest,
211 /// `-print-file-name=<name>`, the full path of a library file.
212 FileName(String),
213 /// `-print-prog-name=<name>`, the full path of a program.
214 ProgName(String),
215 /// `-print-libgcc-file-name`, which is `-print-file-name=libgcc.a` under another spelling.
216 Libgcc,
217}
218
219/// Usage text.
220///
221/// Deliberately short. `spec/04-driver-and-cli.md` puts the full flag reference in the
222/// manual page, because a `--help` nobody can read in one screen is a `--help` nobody reads.
223pub const USAGE: &str = "\
224rucc, an optimizing C compiler
225
226usage: rucc [options] file...
227
228options:
229 -c compile and assemble, do not link
230 -S compile only, emit assembly
231 -E preprocess only
232 -o <file> write output to <file>, or to standard output for -
233 -D <name>[=<value>], -U <name> define a macro, or undefine one after every -D
234 -I <dir> add <dir> to the include search path
235 -iquote -isystem -idirafter <dir> the other chains, -nostdinc drops ours
236 -I-, -iprefix <p>, -iwithprefix[before] <dir> the older spellings of those
237 -include <file>, -imacros <file> read <file> first, the second for its macros only
238 --sysroot=<dir> look for the library's headers under <dir>, -isysroot too
239 -P, -dM with -E: leave out the markers, or dump the macros
240 -M -MM -MD -MMD write a make rule for the source, the last two compile as well
241 -MF <file> -MT <t> -MQ <t> -MP where the rule goes, what it builds, targets with no recipe
242 -std=<dialect> c89 through c2y, and the gnu spellings
243 -fgnuc-version=<v> the GCC release to claim, default 16.0.0
244 -x <lang> treat later inputs as <lang>, or none to stop
245 -O<level> optimize: 0, 1, 2, 3, s, z
246 -fsafety=<tier> check memory safety: off, detect, enforce, kernel
247 -f[no-]sanitize=<what> the negative is taken, the positive is refused by name
248 -f[no-]safety-subobject a write has to stay inside the member it names
249 -f[no-]safety-restrict two restrict pointers of one block may not meet
250 -f<pass> -fno-<pass> -fdump-ir=<what> -fopt-info[-<kind>][=FILE]
251 -fpass-fuel=<pass>=<n>, -fpass-fuel-global=<n> stop a pass, or all of them, after n
252 -fdisable-<pass>[=<funcs>], -fenable-<pass>[=<funcs>] run a pass on some functions only
253 -g -g0 -gdwarf-5, -fno-omit-frame-pointer, -mno-red-zone debug info, frame pointer, red zone
254 -gz[=none|zlib|zlib-gnu|zstd] -gno-split-dwarf compress debug sections, one file not two
255 -flto[=auto|jobserver|<n>] -fno-lto -ffat-lto-objects read, and not done yet
256 -fprofile-use[=<path>] -fprofile-dir=<dir> read too, where -fprofile-generate is refused
257 -f[no-]stack-protector[-strong|-all], -f[no-]stack-clash-protection, -fcf-protection=<edges>
258 -ffunction-sections -fdata-sections a section per function or variable, for --gc-sections
259 -fvisibility=<what> default, hidden, internal or protected, when nothing in the source said
260 -l<name>, -L <dir>, -B <dir> link a library, where to look for one, where our own tools are
261 -fPIC -fpic -fPIE -fpie, -fno-common, -pipe what it does anyway
262 -f[no-]strict-aliasing, -f[no-]delete-null-pointer-checks what it assumes anyway
263 -static -shared -pie -no-pie -nostdlib -nostartfiles -nodefaultlibs -rdynamic -s how to link
264 -Wl,<arg>, -Xlinker <arg>, -fuse-ld=<name> hand an argument to the linker, or pick one
265 -Werror -pedantic -pedantic-errors -w -W[no-]system-headers how much to say, and how fatal
266 -m64 -march= -mtune= -mcpu= -mabi= -mcmodel= what machine to generate for
267 -pg -p, -mfentry -mno-fentry call a profiler on the way in, and where that call goes
268 -fpatchable-function-entry=<n>[,<m>] room at the top of every function to patch later
269 -fwrapv, -fwrapv-pointer, -fno-strict-overflow signed or pointer overflow wraps
270 -ftrapv signed overflow stops the program instead
271 -f[no-]signed-char, -f[no-]unsigned-char, -f[no-]short-enums change the ABI
272 -ffp-contract=<how> fuse a multiply and an addition: fast, on or off
273 -fexcess-precision=<how>, -f[no-]rounding-math, -f[no-]trapping-math what may be folded
274 -ffile-prefix-map=<old>=<new> rewrite that front of every path we put in the output
275 -fmacro-prefix-map= -fdebug-prefix-map= -fprofile-prefix-map= the same, one output each
276 -pthread build for more than one thread, and link the library for it
277 -dumpmachine -dumpversion -print-multiarch -print-search-dirs what this compiler is
278 -print-file-name=<name> -print-prog-name=<name> where a file or a program is
279 -print-sysroot the root the headers and the libraries are read under
280 -print-sysroot-provenance every input under it, where it came from and its licence
281 -print-sysroot-digest the sha256 of that record, which names the whole sysroot in one line
282 --fetch <tuple> get the sysroot this release pins for <tuple> and install it in the cache
283 --fetch-msvc-sdk <tuple> Microsoft's licence, then the SDK behind it with --accept-licence
284 --offline never download anything, which a compilation never does anyway
285 -j[n] compile n translation units at once, default all
286 -v, -### print each phase as it runs, or without running any
287 -save-temps[=cwd|obj], -time keep the .i and the .s, say how long each step took
288 --target=<triple> generate code for <triple>
289 --emit=<kind> exe, obj, archive, asm, preprocessed, tast, ir, mir-final,
290 safety-summary, type-granules
291 --print-config, --print-pipeline print the configuration or the pipeline, and exit
292 --version print the version and exit
293 -h, --help print this message and exit
294
295See spec/04-driver-and-cli.md for the full flag reference.
296";
297
298/// The argument of a flag that may be joined to it or may be the next word.
299///
300/// `-DFOO` and `-D FOO` are the same thing, and `at` is where the flag's own letters end.
301fn joined_or_next(
302 arg: &str,
303 at: usize,
304 args: &[String],
305 i: &mut usize,
306) -> Result<String, CliError> {
307 if arg.len() > at {
308 return Ok(arg[at..].to_owned());
309 }
310 let next = args.get(*i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
311 *i += 1;
312 Ok(next.clone())
313}
314
315/// The smallest boundary a function is put on when the command line asked for no alignment at all.
316///
317/// Eight bytes, which is what gcc 16 gives `-fno-align-functions` on x86-64 and is a boundary every
318/// target this compiler has is happy with. It is not zero: a function still has to start somewhere
319/// an instruction may start, and the flag asks for the target's minimum rather than for none.
320const MIN_FUNC_ALIGN: u32 = 8;
321
322/// What `-falign-functions=N` asks for, as a power of two, or `None` for the target's own answer.
323///
324/// Zero and one both mean the default, which is gcc's reading of them, and everything else is
325/// rounded up to the next power of two, which is also gcc's: `-falign-functions=3` puts a function
326/// on a four byte boundary rather than being refused. Gives back `Err` shaped as an outer `None`
327/// only when the text is not a number, since that is the one thing gcc will not read either. A
328/// number larger than any alignment makes sense at is clamped rather than refused, for the same
329/// reason: this is a preference about speed and a build that wrote a silly one still deserves to
330/// compile.
331fn function_alignment(text: &str) -> Option<Option<u32>> {
332 // gcc takes `N:M:N2:M2`, where everything after the first number is about how far it is willing
333 // to go to reach the boundary. Only the boundary is answerable here, so the rest is read to
334 // check that it is numbers and then dropped.
335 let mut parts = text.split(':');
336 let first = parts.next()?;
337 if parts.any(|part| part.parse::<u64>().is_err()) {
338 return None;
339 }
340 let want: u64 = first.parse().ok()?;
341 if want <= 1 {
342 return Some(None);
343 }
344 let bytes = want.min(1 << 16).next_power_of_two();
345 Some(Some(u32::try_from(bytes).ok()?))
346}
347
348/// Every name that may follow `-fsanitize=`, which is gcc 16's list and three of this compiler's
349/// own.
350///
351/// The three are on it because `spec/07-types-and-semantics.md` section 7.7 already promises them:
352/// each undefined behaviour this compiler exploits is listed there with the check that detects it,
353/// and `alias`, `restrict` and `memory` are checks gcc has no spelling for. gcc refuses `memory`
354/// outright, since the sanitizer of that name is clang's. A name being here means it is a name
355/// rather than a typo, and nothing more than that: every one of them is refused after the loop,
356/// because none of them is implemented.
357///
358/// `all` is deliberately absent. gcc takes it only in the negative, so it is handled where each of
359/// those two spellings is read rather than by being on this list.
360const SANITIZERS: [&str; 34] = [
361 "address",
362 "kernel-address",
363 "hwaddress",
364 "kernel-hwaddress",
365 "pointer-compare",
366 "pointer-subtract",
367 "thread",
368 "leak",
369 "undefined",
370 "shift",
371 "shift-base",
372 "shift-exponent",
373 "integer-divide-by-zero",
374 "unreachable",
375 "vla-bound",
376 "null",
377 "return",
378 "signed-integer-overflow",
379 "bounds",
380 "bounds-strict",
381 "alignment",
382 "object-size",
383 "float-divide-by-zero",
384 "float-cast-overflow",
385 "nonnull-attribute",
386 "returns-nonnull-attribute",
387 "bool",
388 "enum",
389 "vptr",
390 "pointer-overflow",
391 "builtin",
392 "alias",
393 "restrict",
394 "memory",
395];
396
397/// Parses a command line, without the program name.
398///
399/// # Errors
400///
401/// Returns the message to print when the arguments do not name a compilation this compiler
402/// can attempt.
403pub fn parse_args(args: &[String]) -> Result<Action, CliError> {
404 let host = Triple::host()
405 .ok_or_else(|| err("this host is not a supported target and no --target was given"))?;
406 let mut opts = Options::new(host);
407 // Where the compiler is running, which is what `DW_AT_comp_dir` is and what a debugger joins a
408 // relative file name onto. Asked here rather than where the debug sections are written, because
409 // this is the one layer that is allowed to look at the process it is in, and because a command
410 // line that compiles four files should give the same answer for all four.
411 opts.working_dir = std::env::current_dir().ok().map(|dir| dir.to_string_lossy().into_owned());
412 let mut inputs: Vec<Input> = Vec::new();
413 let mut print_config = false;
414 let mut print_pipeline = false;
415 let mut print_plan = false;
416 let mut verbose = false;
417 let mut jobs = Jobs::default();
418 let mut nostdinc = false;
419 let mut sysroot: Option<PathBuf> = None;
420 // What the command line is worth warning about, filled in after the loop rather than during it,
421 // because every question of this kind is about two flags and the last word on both of them is
422 // the end of the loop.
423 let mut notes: Vec<String> = Vec::new();
424 // The whole ten field target, kept beside the three field one because `--target=` can pin a
425 // libc version and `Triple` has nowhere to put it. It decides `__GLIBC_MINOR__` and nothing
426 // else today, and `None` is a command line that named no target, which is this machine.
427 let mut pinned: Option<TargetTuple> = None;
428 let mut output = None;
429 let mut link = LinkOptions::default();
430 let mut query: Option<Query> = None;
431 // What `--fetch` named, and whether `--offline` forbade it. Both are weighed after the loop
432 // because either can be written after the other.
433 let mut fetch: Option<String> = None;
434 // The other fetch, kept apart from the one above because they are different commands with
435 // different rules, and weighed after the loop for the same reason that one is.
436 let mut fetch_msvc: Option<String> = None;
437 let mut accepted = false;
438 let mut offline = false;
439 let mut threads = false;
440 // Which sanitizers are still asked for by the end of the command line. Accumulated across the
441 // loop rather than answered where it was read, because `-fno-sanitize=` turns one off and a
442 // build that asks for a check and then takes it back has asked for nothing. What happens to a
443 // set that is not empty is decided after the loop.
444 let mut sanitizers: Vec<&str> = Vec::new();
445 // `-x` applies to inputs that come after it and stays in effect until the next one, which
446 // is why it is tracked across the loop rather than attached to a single argument.
447 let mut forced: Option<InputKind> = None;
448 // What `-iprefix` last said, stuck on the front of every later `-iwithprefix`. It applies to
449 // the flags after it and not the ones before, so a command line may set it more than once.
450 // GCC's default is its own installed header directory with the last component taken off,
451 // which is a path a cross compiler's build system knows and passes; there is no equivalent
452 // here, so with no `-iprefix` the prefix is nothing and `-iwithprefix` names a directory
453 // outright.
454 let mut iprefix = String::new();
455
456 let mut i = 0;
457 while i < args.len() {
458 let arg = args[i].as_str();
459 i += 1;
460 match arg {
461 "-h" | "--help" => return Ok(Action::Help),
462 "--version" => return Ok(Action::Version),
463 // The sysroot fetch, which is weighed after the loop rather than acted on here, because
464 // `--offline` written after it has to be able to forbid it. Both spellings, since a
465 // flag that takes a tuple gets written both ways and neither is a guess at what the
466 // other meant.
467 "--fetch" => {
468 let value = args
469 .get(i)
470 .ok_or_else(|| err("--fetch requires the target to get a sysroot for"))?;
471 i += 1;
472 fetch = Some(value.clone());
473 }
474 _ if arg.starts_with("--fetch=") => {
475 fetch = Some(arg["--fetch=".len()..].to_owned());
476 }
477 // The other fetch, which is section 13.4's. Same two spellings for the same reason,
478 // and weighed after the loop so that `--offline` and `--accept-licence` written after
479 // it are read whichever order somebody put them in.
480 "--fetch-msvc-sdk" => {
481 let value = args.get(i).ok_or_else(|| {
482 err("--fetch-msvc-sdk requires the target to get the SDK for")
483 })?;
484 i += 1;
485 fetch_msvc = Some(value.clone());
486 }
487 _ if arg.starts_with("--fetch-msvc-sdk=") => {
488 fetch_msvc = Some(arg["--fetch-msvc-sdk=".len()..].to_owned());
489 }
490 // Both spellings of the word, because the compiler's own prose uses one of them and
491 // most of the people typing this will reach for the other, and being told that a flag
492 // is not a flag over the letter in the middle of it is a puzzle rather than a message.
493 "--accept-licence" | "--accept-license" => accepted = true,
494 // Accepted on any command line and only ever read by the fetch, because an ordinary
495 // compile downloads nothing with or without it. So this flag takes nothing away today,
496 // which is the property section 13.2 asks for rather than an omission: a build that
497 // passes it is saying what it expects of this compiler, and what it expects is already
498 // true.
499 "--offline" => offline = true,
500 "--print-config" => print_config = true,
501 "--print-pipeline" => print_pipeline = true,
502 "-###" => print_plan = true,
503 "-v" => verbose = true,
504 // The files a compilation goes through, kept rather than thrown away. The bare
505 // spelling means `=obj` and not `=cwd`, which is not what the manual says and is what
506 // gcc 16 does; `SaveTemps::Object` carries the measurement.
507 "-save-temps" => opts.save_temps = SaveTemps::Object,
508 _ if arg.starts_with("-save-temps=") => {
509 opts.save_temps = arg["-save-temps=".len()..].parse().map_err(err)?;
510 }
511 // How long each step took. A misspelling of this is worth rejecting rather than
512 // ignoring, since a run that says nothing looks like a compilation that took no time.
513 "-time" => opts.time = true,
514 "-c" => opts.emit = EmitKind::Object,
515 "-S" => opts.emit = EmitKind::Asm,
516 "-E" => opts.emit = EmitKind::Preprocessed,
517 "-g" => opts.debug_info = true,
518 // GCC's own levels of how much debug information to write. Zero is none and every
519 // other number is some, and this compiler has one amount, so the numbers above zero
520 // all mean the same thing here. `-ggdb` is the same flag asking for whatever the
521 // debugger on the machine prefers, which is what we emit anyway.
522 "-g0" => opts.debug_info = false,
523 "-g1" | "-g2" | "-g3" | "-ggdb" | "-ggdb1" | "-ggdb2" | "-ggdb3" => {
524 opts.debug_info = true;
525 }
526 // The version of DWARF to write. We write DWARF 5 and nothing else, so a build that
527 // asks for another version is told rather than handed a file it cannot read.
528 "-gdwarf" | "-gdwarf-5" => opts.debug_info = true,
529 _ if arg.starts_with("-gdwarf-") => {
530 return Err(err(format!(
531 "{arg}: this compiler writes DWARF 5 and no other version, see \
532 spec/11-debug-info.md"
533 )));
534 }
535 // Whether the debug information goes in a file of its own beside the object. gcc
536 // writes that `.dwo` whether or not it found anything to put in it, which means a
537 // build system that declares the file as an output gets one and a make rule that
538 // depends on it fires. Refused for that reason rather than taken: section 4.1 takes a
539 // flag that changes nothing and refuses one that changes what is produced, and a file
540 // that does not appear is the plainest change of that kind there is. The negative
541 // spelling is taken, because putting it all in the object is what happens anyway.
542 "-gno-split-dwarf" => {}
543 "-gsplit-dwarf" => {
544 return Err(err(format!(
545 "{arg}: this compiler writes no separate `.dwo` file, and a build that \
546 expects one beside each object would wait for a file that never arrives, \
547 see spec/11-debug-info.md"
548 )));
549 }
550 // How the debug sections are compressed. There are none yet, so every answer produces
551 // the same bytes and taking the flag promises nothing that is not kept. The value is
552 // still checked, because a typo in a distribution's flags is worth finding when the
553 // compiler reads it rather than when somebody later wonders why nothing got smaller.
554 // Bare `-gz` means `zlib`, which the manual leaves for the reader to discover.
555 "-gz" => opts.compress = Compress::Zlib,
556 _ if arg.starts_with("-gz=") => {
557 let how = &arg["-gz=".len()..];
558 opts.compress = how.parse().map_err(|()| {
559 err(format!(
560 "`{how}` is not a way to compress debug sections, which is none, zlib, \
561 zlib-gnu or zstd"
562 ))
563 })?;
564 }
565 "-Werror" => opts.warnings_are_errors = true,
566 // Nothing that is not fatal is said at all. Read at the one place a diagnostic goes
567 // through rather than here, so that a warning `-w` dropped is not counted either.
568 "-w" => opts.warnings = false,
569 // Off by default, the way gcc has it off. A header that came with the machine is not
570 // one the person compiling can change, so a warning about it is noise, and under
571 // `-Werror` it is a build that stops on a line nobody in the project wrote. Somebody
572 // porting a header does want to hear all of it, which is what the flag is for.
573 "-Wsystem-headers" => opts.system_header_warnings = true,
574 "-Wno-system-headers" => opts.system_header_warnings = false,
575 "-pedantic-errors" => {
576 opts.pedantic = true;
577 opts.warnings_are_errors = true;
578 }
579 "-P" => opts.line_markers = false,
580 // The dependency family, which section 4.4 calls required because every build system
581 // that generates its own makefiles asks for it. The two that end in `D` write a file
582 // beside the object and let the compilation happen, and the two that do not write to
583 // standard output and stop after it. Nothing here turns the system headers back on
584 // once a flag has turned them off, which is GCC's behaviour and is why `-MM -M` is
585 // `-MM`: the flag asking for fewer of them is the one with something to say.
586 "-M" => {
587 opts.deps.emit = true;
588 opts.deps.instead_of_compiling = true;
589 }
590 "-MM" => {
591 opts.deps.emit = true;
592 opts.deps.instead_of_compiling = true;
593 opts.deps.system_headers = false;
594 }
595 "-MD" => opts.deps.emit = true,
596 "-MMD" => {
597 opts.deps.emit = true;
598 opts.deps.system_headers = false;
599 }
600 "-MP" => opts.deps.phony = true,
601 // These three take a word and only in the separated form, which is how GCC spells
602 // them and how every build system writes them.
603 "-MF" | "-MT" | "-MQ" => {
604 let value =
605 args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
606 i += 1;
607 match arg {
608 "-MF" => opts.deps.file = Some(value.clone()),
609 // The whole of the difference between the two. `-MT` is for a build that has
610 // already escaped what it is passing, and `-MQ` is for one that has a name
611 // and wants it to arrive as that name.
612 "-MT" => opts.deps.targets.push(value.clone()),
613 _ => opts.deps.targets.push(deps::escaped(value)),
614 }
615 }
616 // The questions a build system asks before it compiles anything. Answered after the
617 // loop, because each one is about the target or the library search and the command
618 // line has not finished saying what those are.
619 "-dumpmachine" => query = Some(Query::Machine),
620 "-dumpversion" | "-dumpfullversion" => query = Some(Query::Version),
621 "-print-multiarch" => query = Some(Query::Multiarch),
622 "-print-search-dirs" => query = Some(Query::SearchDirs),
623 "-print-sysroot" => query = Some(Query::Sysroot),
624 // Both spellings, because this one is ours rather than GCC's and our own documents
625 // write it both ways: section 13.5 of `spec/cross-compile/13-distribution.md` gives it
626 // two dashes like the other flags we invented, and document 12's table gives it one
627 // like the `-print-` family it sits in. A person who reads either and types what it
628 // says is right, so neither is refused.
629 "-print-sysroot-provenance" | "--print-sysroot-provenance" => {
630 query = Some(Query::SysrootProvenance);
631 }
632 "-print-sysroot-digest" | "--print-sysroot-digest" => {
633 query = Some(Query::SysrootDigest);
634 }
635 "-print-libgcc-file-name" => query = Some(Query::Libgcc),
636 _ if arg.starts_with("-print-file-name=") => {
637 query = Some(Query::FileName(arg["-print-file-name=".len()..].to_owned()));
638 }
639 _ if arg.starts_with("-print-prog-name=") => {
640 query = Some(Query::ProgName(arg["-print-prog-name=".len()..].to_owned()));
641 }
642 // A program built to run in more than one thread. On every platform this compiler
643 // targets that is a macro the library's headers read and one more library on the
644 // link line, and the library is added after the loop so that it lands after the
645 // objects that refer to it.
646 "-pthread" | "-pthreads" => {
647 opts.defines.push("_REENTRANT".to_owned());
648 threads = true;
649 }
650 "-ansi" => {
651 opts.std = Std::C89;
652 opts.gnu_extensions = false;
653 }
654 // `-Wpedantic` is the same flag under the name the `-W` family gives it, which is
655 // the spelling a build system that groups its warning flags tends to write.
656 "-pedantic" | "-Wpedantic" => opts.pedantic = true,
657 // Both directions, because a build that needs this for one directory turns it back
658 // off for the next one rather than leaving it on for the whole tree.
659 "-fpermissive" => opts.permissive = true,
660 "-fno-permissive" => opts.permissive = false,
661 "-ffreestanding" => opts.hosted = false,
662 "-fhosted" => opts.hosted = true,
663 "-fno-builtin" => opts.builtins = false,
664 "-fbuiltin" => opts.builtins = true,
665 // The C89 dialects are under GNU's reading whatever this says, so turning it off
666 // there is turning off something the dialect asked for, which is accepted and does
667 // nothing. gcc refuses that command line, and there is nothing it could have meant.
668 "-fgnu89-inline" => opts.gnu89_inline = true,
669 "-fno-gnu89-inline" => opts.gnu89_inline = false,
670 // Both directions of each, because a build system that wants one of these usually
671 // writes it beside the flag that turns it back off for one directory.
672 "-fno-omit-frame-pointer" => opts.frame_pointer = true,
673 "-fomit-frame-pointer" => opts.frame_pointer = false,
674 // Both directions again, for the same reason, and a third answer for a command line
675 // that wrote neither: see `reorder_blocks` in `rucc_session`.
676 "-freorder-blocks" => opts.reorder_blocks = Some(true),
677 "-fno-reorder-blocks" => opts.reorder_blocks = Some(false),
678 // gcc's name for the scheduler that runs after the registers are handed out, which is
679 // the only one rucc has: see `schedule_insns` in `rucc_session`. gcc also takes
680 // `-fschedule-insns` for the pass before allocation, and taking that one here would be
681 // a flag that says a pass ran when none did.
682 "-fschedule-insns2" => opts.schedule_insns = Some(true),
683 "-fno-schedule-insns2" => opts.schedule_insns = Some(false),
684 "-mno-red-zone" => opts.red_zone = false,
685 "-mred-zone" => opts.red_zone = true,
686 // Four flags rather than one with an argument, which is how gcc spells them and how
687 // every build line writes them. Last one wins, because a package build puts
688 // `-fstack-protector-strong` in its global flags and a directory that cannot have one
689 // turns it back off on the line after.
690 "-fno-stack-protector" | "-fno-stack-protector-all" | "-fno-stack-protector-strong" => {
691 opts.protector = Protector::None;
692 }
693 "-fstack-protector" => opts.protector = Protector::Buffers,
694 "-fstack-protector-strong" => opts.protector = Protector::Strong,
695 "-fstack-protector-all" => opts.protector = Protector::All,
696 // The other half of what a hardened build asks for, and it is a question about the
697 // frame rather than about the function, so it is a switch rather than a level.
698 "-fstack-clash-protection" => opts.stack_clash = true,
699 "-fno-stack-clash-protection" => opts.stack_clash = false,
700 // The third of them, and the one that is a question with an argument rather than a
701 // family of spellings, because what it asks about is which of the two edges of a
702 // control flow transfer is checked. Bare is both of them, which is what gcc does.
703 "-fcf-protection" => opts.control = Control::Full,
704 "-fno-cf-protection" => opts.control = Control::None,
705 // Two spellings of the same request, which is what gcc has as well. `-p` was the older
706 // profiler and `-pg` the one that also recorded who called whom, and on every platform
707 // this compiler targets there is now one hook and both ask for it.
708 "-pg" | "-p" => {
709 opts.profile = true;
710 link.profile = true;
711 }
712 // Accepted on their own and doing nothing on their own, which is gcc's behaviour: they
713 // say where the call goes and a command line that asked for no call has nowhere to put
714 // one. That matters because a build system that sets `-mfentry` globally and `-pg` per
715 // directory is a build system that would otherwise fail on every other directory.
716 "-mfentry" => opts.hook = Hook::Early,
717 "-mno-fentry" => opts.hook = Hook::Late,
718 // GCC drops its own include directory along with the system ones, because its
719 // headers are half of a pair with the library's and half a pair is worse than
720 // none. A build that passes this is supplying the whole set itself.
721 "-nostdinc" => nostdinc = true,
722 "-o" => {
723 output = Some(args.get(i).ok_or_else(|| err("-o requires an argument"))?.clone());
724 i += 1;
725 }
726 // The flags that take a directory only in the separated form. GCC spells them
727 // this way and nothing writes `-iquotedir`, so accepting the joined form would
728 // mean guessing at a path that starts with the flag's own letters.
729 // Apple's spelling of `--sysroot`, and the one its own build systems pass. The
730 // two mean the same thing here: the configured directories are under there rather
731 // than under the root.
732 "-isysroot" => {
733 let dir = args.get(i).ok_or_else(|| err("-isysroot requires an argument"))?;
734 i += 1;
735 sysroot = Some(PathBuf::from(dir));
736 }
737 "-iquote" | "-isystem" | "-idirafter" => {
738 let dir = args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
739 i += 1;
740 match arg {
741 "-iquote" => opts.search.push_quote(dir.clone()),
742 "-isystem" => opts.search.push_system(dir.clone()),
743 _ => opts.search.push_after(dir.clone()),
744 }
745 }
746 "-iprefix" => {
747 iprefix = args.get(i).ok_or_else(|| err("-iprefix requires an argument"))?.clone();
748 i += 1;
749 }
750 // Where GCC puts these is not where its manual says it puts them, and this is the
751 // measured answer rather than the documented one: `-iwithprefix` lands in the
752 // `-isystem` slot and not the `-idirafter` slot, and `-iwithprefixbefore` lands in
753 // the `-I` slot. A cross build that uses them is relying on the behaviour, since
754 // that is the compiler it was developed against.
755 "-iwithprefix" | "-iwithprefixbefore" => {
756 let dir = args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
757 i += 1;
758 let dir = format!("{iprefix}{dir}");
759 if arg == "-iwithprefix" {
760 opts.search.push_system(dir);
761 } else {
762 opts.search.push_bracket(dir);
763 }
764 }
765 "-include" | "-imacros" => {
766 let name = args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
767 i += 1;
768 opts.preincludes
769 .push(Preinclude { name: name.clone(), macros_only: arg == "-imacros" });
770 }
771 // The flag `-iquote` was introduced to replace, still passed by build systems old
772 // enough to predate the replacement. It is not a directory: it says that every `-I`
773 // so far is for quoted includes only, and that a quoted include stops looking next
774 // to the file that wrote it.
775 "-I-" => opts.search.split_quote_chain(),
776 "-x" => {
777 let lang = args.get(i).ok_or_else(|| err("-x requires an argument"))?;
778 i += 1;
779 forced = if lang == "none" {
780 None
781 } else {
782 Some(InputKind::from_x_arg(lang).map_err(|e| err(format!("{e}")))?)
783 };
784 }
785 // Not a GCC flag. spec/03-architecture.md section 3.5 compiles several
786 // translation units in one process rather than making the build system fork, and
787 // section 3.8's determinism check compares `-j1` against `-j16`, so the knob has
788 // to exist and has to be spelled the way `make` spells it.
789 // `-DFOO`, `-D FOO` and the same for `-U` and `-I`. Both forms are in wide use
790 // and a build system may produce either, so both are read here rather than
791 // being normalised by whatever generated the command line.
792 _ if arg.starts_with("-D") => {
793 let value = joined_or_next(arg, 2, args, &mut i)?;
794 opts.defines.push(value);
795 }
796 _ if arg.starts_with("-U") => {
797 let value = joined_or_next(arg, 2, args, &mut i)?;
798 opts.undefines.push(value);
799 }
800 _ if arg.starts_with("-I") => {
801 let dir = joined_or_next(arg, 2, args, &mut i)?;
802 opts.search.push_bracket(dir);
803 }
804 _ if arg.starts_with("-std=") => {
805 let name = &arg["-std=".len()..];
806 let (std, gnu) = Std::from_flag(name)
807 .ok_or_else(|| err(format!("unknown dialect `{name}`, see --help")))?;
808 opts.std = std;
809 opts.gnu_extensions = gnu;
810 }
811 // Section 4.5. The claim decides which half of glibc's `sys/cdefs.h` we are
812 // handed, so a differential run that does not set it is comparing two compilers
813 // that believe they are different compilers.
814 // GCC packs these into one flag, so `-dDI` is two of them. Letters in the family
815 // that we have not written yet are accepted and ignored, because a dump is a
816 // debugging aid and a build that asks for one should still compile. A letter
817 // outside the family falls through to the unknown option error, which is what
818 // keeps `-dumpversion` from being read as a dump of nothing.
819 _ if Dumps::is_family(arg) => {
820 opts.dumps.add(&arg[2..]);
821 }
822 // One name at a time, which is what a build that means its own `memcpy` and the
823 // library's everything else writes. The name is not checked against a list, because
824 // the flag is about what the program means by a name and a program is allowed to mean
825 // something by a name this compiler has never heard of.
826 _ if arg.starts_with("-fno-builtin-") => {
827 opts.no_builtin.push(arg["-fno-builtin-".len()..].to_owned());
828 }
829 _ if arg.starts_with("-fgnuc-version=") => {
830 let v = &arg["-fgnuc-version=".len()..];
831 opts.gnuc = v.parse().map_err(err)?;
832 }
833 // spec/13-gnu-compat.md section 13.3 promises this flag an error that says why rather
834 // than the unknown option one, because a build reaching for it is asking for a feature
835 // and deserves to be told it is not coming rather than told the spelling is wrong.
836 // The negative form is what this compiler does anyway, so it is taken and dropped.
837 "-fnested-functions" => {
838 return Err(err(
839 "nested functions are not supported: a call to one goes through a trampoline \
840 written on the stack, which no target that enforces an unexecutable stack \
841 allows",
842 ));
843 }
844 "-fno-nested-functions" => {}
845 // Which of the two links the output is for, which is a real difference and not a
846 // description of what happens anyway. Everything here is position independent either
847 // way, and what these decide is whether a name may be one another object defines or
848 // replaces, because a link that produces an executable puts every name in the same
849 // program and a link that produces a shared library does not.
850 //
851 // It matters that they are accepted at all, whatever they then do. Every autoconf and
852 // cmake build puts `-fPIC` on the compile line, so a compiler that rejects it cannot
853 // be the `CC` of a project that has a configure script, whatever else it can do. That
854 // is how this was found: building SQLite's test fixture stopped on it.
855 "-fPIC" | "-fpic" => opts.pic = Pic::Library,
856 // Not a synonym of the pair above, which is what they were treated as until #756. The
857 // library is the expensive answer and gcc makes it the one that has to be asked for,
858 // so this is also what nothing at all means.
859 "-fPIE" | "-fpie" => opts.pic = Pic::Executable,
860 // A different question from the pair above, and the one every distribution build of a
861 // shared library answers. `-fPIC` decides how an address is reached, and this decides
862 // whether the optimizer may believe a body it can see, because an exported name is one
863 // the dynamic linker may find another definition of first. On by default, which is
864 // gcc's arrangement and is the honest answer, and off is a promise the build makes and
865 // nothing checks.
866 "-fsemantic-interposition" => opts.interposition = true,
867 "-fno-semantic-interposition" => opts.interposition = false,
868 // Two requests rather than one, and the same table answers both, so what decides is
869 // whether either of them is standing. gcc arranges it the same way: the asynchronous
870 // one is the default here and it implies the other, and a line that asks for a table
871 // and against an asynchronous one gets a table.
872 "-fasynchronous-unwind-tables" => opts.async_unwind_tables = true,
873 "-fno-asynchronous-unwind-tables" => opts.async_unwind_tables = false,
874 "-funwind-tables" => opts.unwind_tables = true,
875 "-fno-unwind-tables" => opts.unwind_tables = false,
876 // The other direction is a request, not a description, and it is one this compiler
877 // cannot grant, so it gets the treatment section 13.3 asks for rather than the unknown
878 // option error. Answering it by carrying on would be answering a different question:
879 // the code would still be position independent, which is correct everywhere an
880 // ordinary program runs and is wrong in a kernel, where the flag is written precisely
881 // because there is no loader to fill a global offset table in.
882 "-fno-pic" | "-fno-pie" => {
883 return Err(err(
884 "position dependent code is not supported: an address that may be in another \
885 object is loaded out of the global offset table, and nothing here emits the \
886 absolute form this asks for. Use -no-pie if what you meant was how to link",
887 ));
888 }
889 // A section per function and a section per variable, which is what makes
890 // `--gc-sections` able to drop anything: a linker can leave out a section nothing
891 // reaches and cannot leave out half of one. Both directions are taken, and the off
892 // one is the default rather than a refusal, since a build that writes it is asking
893 // for what happens anyway.
894 "-ffunction-sections" => opts.function_sections = true,
895 "-fno-function-sections" => opts.function_sections = false,
896 "-fdata-sections" => opts.data_sections = true,
897 "-fno-data-sections" => opts.data_sections = false,
898 // Another description of what this compiler does. A file scope declaration with no
899 // initializer is written into `.bss` as an ordinary defined symbol, not offered to the
900 // linker as a common one for it to merge, which is what `-fno-common` asks for and what
901 // gcc has done by default since 10. Nothing in the front end produces `Linkage::Common`
902 // at all.
903 "-fno-common" => {}
904 // What overflows rather than being undefined. Every one of these takes something away
905 // from the optimizer rather than asking it to do anything, which is why the negative
906 // spellings are the interesting ones and the positive spellings are the default.
907 //
908 // `-fno-strict-overflow` is both of the others, which is gcc's own reading of it: its
909 // help text for `-fstrict-overflow` says "negated as -fwrapv -fwrapv-pointer". So it is
910 // written here as the pair rather than kept as a third thing to test everywhere.
911 //
912 // `-ftrapv` is the exception and is the one that asks for something. It is the other
913 // answer to the question `-fwrapv` answers, so the two cannot both hold and each clears
914 // the other, which makes the last one on the command line the one that counts. That is
915 // gcc 16's behaviour and was measured rather than read: `-ftrapv -fwrapv` emits no
916 // checked calls and `-fwrapv -ftrapv` emits them. The positive spelling of the pointer
917 // question is left alone by both, because neither has anything to say about it.
918 "-fwrapv" => {
919 opts.wrapping.signed = true;
920 opts.wrapping.trap = false;
921 }
922 "-fno-wrapv" => opts.wrapping.signed = false,
923 "-fwrapv-pointer" => opts.wrapping.pointer = true,
924 "-fno-wrapv-pointer" => opts.wrapping.pointer = false,
925 "-fno-strict-overflow" => opts.wrapping = Wrapping::ALL,
926 // Which does not clear the checked one, because gcc does not: `-ftrapv
927 // -fstrict-overflow` still emits the calls. It says what is assumed and not what
928 // happens.
929 "-fstrict-overflow" => {
930 opts.wrapping.signed = false;
931 opts.wrapping.pointer = false;
932 }
933 "-ftrapv" => {
934 opts.wrapping.trap = true;
935 opts.wrapping.signed = false;
936 }
937 "-fno-trapv" => opts.wrapping.trap = false,
938 // The two flags that say what a plain `char` is, which is one question with two
939 // spellings each: gcc reads `-fno-signed-char` as `-funsigned-char` and
940 // `-fno-unsigned-char` as `-fsigned-char`, so there are four ways to write two
941 // answers and the last one written wins. Nothing is set until one of them is given,
942 // because the target's own ABI is the answer otherwise and it is not the same answer
943 // everywhere: x86-64 and Apple's arm64 are signed, Linux's arm64 is not.
944 "-fsigned-char" | "-fno-unsigned-char" => opts.char_signed = Some(true),
945 "-funsigned-char" | "-fno-signed-char" => opts.char_signed = Some(false),
946 // And the size of an enumeration, which is the other thing in this group that changes
947 // the ABI rather than the code.
948 "-fshort-enums" => opts.short_enums = true,
949 "-fno-short-enums" => opts.short_enums = false,
950 // And Microsoft's reading of an anonymous member, which changes the layout of every
951 // record that writes a tag on one. Nothing is set until one of them is given, because
952 // the target is the answer otherwise: gcc's mingw build has this on and its Linux
953 // build has it off.
954 "-fms-extensions" => opts.ms_extensions = Some(true),
955 "-fno-ms-extensions" => opts.ms_extensions = Some(false),
956 // And the request, which is the one that cannot be granted. It is a real difference and
957 // not a preference: two files each writing `int g;` link under `-fcommon` and are a
958 // duplicate definition without it, which is the whole reason the flag survives.
959 "-fcommon" => {
960 return Err(err(
961 "a tentative definition is written into .bss as its own symbol here, and \
962 nothing emits the common symbol this asks the linker to merge. Give the \
963 variable a definition in one file and declare it extern in the others",
964 ));
965 }
966 // Both directions of this one are recorded, and what they decide is whether lowering
967 // names the type each access goes through. Turning it off is the front end leaving the
968 // name off rather than a pass being told to ignore one it can see, which is one
969 // condition in one place, and it is the reading that survives link time optimization:
970 // a unit built with the flag off keeps its own answer when its bodies end up in a
971 // module beside bodies that were not.
972 //
973 // Nothing in the pipeline reads those names yet. Layer 3 of the alias analysis does
974 // and is tested, and no pass at any level asks the alias analysis anything today, so
975 // no program compiles differently for having passed this. The flag is wired anyway,
976 // because the change that makes a pass ask is not the change anybody will remember to
977 // wire it in, and a flag that is taken and dropped once the names mean something is
978 // the miscompilation `spec/04-driver-and-cli.md` section 4.1 warns about in as many
979 // words.
980 "-fstrict-aliasing" => opts.strict_aliasing = true,
981 "-fno-strict-aliasing" => opts.strict_aliasing = false,
982 // The same shape of answer for the same reason, and the flag the kernel writes beside
983 // the one above it.
984 //
985 // Nothing here concludes that a pointer is not null from the fact that it was
986 // dereferenced. There is no such conclusion to draw from, because no pass records one:
987 // a load says where it read and nothing else, and a comparison against null is an
988 // ordinary comparison of two values the optimizer has no fact about. So a function
989 // that reads through a pointer and then tests it keeps the test, which is what the
990 // kernel wants and what `-fno-delete-null-pointer-checks` asks for, and what gcc has
991 // to be asked for because it draws the conclusion by default.
992 //
993 // `-fdelete-null-pointer-checks` is the request to draw it, and it goes the way
994 // `-fstrict-aliasing` does: assuming less than was asked for costs speed and not
995 // correctness, and `-O2` implies it, so refusing it would stop builds for nothing.
996 "-fdelete-null-pointer-checks" | "-fno-delete-null-pointer-checks" => {}
997 // The floating point group, which goes the same way and for the same reason, and which
998 // is worth writing out because the reason is easy to get backwards.
999 //
1000 // Each of these has a restrictive spelling and a permissive one. The restrictive ones,
1001 // `-frounding-math` and `-ftrapping-math`, say that the rounding mode may have been
1002 // changed and that an exception raised by an operation may be looked at, so an
1003 // arithmetic the compiler folds at compile time is an arithmetic whose rounding and
1004 // whose exception the program does not get. Nothing here folds any floating point
1005 // arithmetic in a function body: `0.1 + 0.2` is an `fadd` and `1.0 / 0.0` is a divide
1006 // that runs, at every level. So both of those describe what already happens.
1007 //
1008 // The permissive ones are the other half, and they are licences rather than requests
1009 // for an answer. `-fno-rounding-math` says the rounding mode is the default one and
1010 // `-fno-trapping-math` says nothing looks at the exceptions, which together are
1011 // permission to fold. Not folding is the conservative side of that permission and is
1012 // what a program is entitled to whichever was written, so `-fno-rounding-math` costs
1013 // speed and not correctness, which is the test section 4.1 puts a licence through.
1014 "-frounding-math" | "-fno-rounding-math" => {}
1015 // `-fno-trapping-math` is the one of the four that is kept, because there is one
1016 // conversion this compiler does not fold and gcc folds under it, and the two answers
1017 // differ. Converting a constant floating value to an integer type it does not fit in
1018 // is undefined behaviour rather than a value: left to the hardware it is one
1019 // instruction and the answer is the integer indefinite value, and folded it is the
1020 // nearest end of the integer's range. Both compilers leave it to the instruction by
1021 // default and gcc folds it under this flag, so a program built with it and compiled
1022 // without it gets a different number rather than a slower one. `-ftrapping-math` is
1023 // gcc's default, so a build spelling it out is asking for what it already has.
1024 "-ftrapping-math" => opts.trapping_math = true,
1025 "-fno-trapping-math" => opts.trapping_math = false,
1026 // About temporary files rather than about code. There is nothing between the phases of
1027 // one compilation here to write to a file in the first place.
1028 "-pipe" => {}
1029 // Nothing here writes colour, so all of these are the same answer, and it is the answer
1030 // that costs nothing: the diagnostics come out plain either way and no build depends on
1031 // an escape sequence being there. Taken rather than refused because cmake writes
1032 // `-fdiagnostics-color=always` on every compile line when the generator is ninja, which
1033 // makes this the second most common flag after `-fPIC` to stop a build over a question
1034 // about how the text looks.
1035 "-fdiagnostics-color" | "-fno-diagnostics-color" => {}
1036 _ if arg.starts_with("-fdiagnostics-color=") => {}
1037 // The link flags. None of them changes the compilation, which is why they are
1038 // collected apart from `opts` and why `-lm` on a `-c` line is a note rather than an
1039 // error: it is a thing said to a linker that is not going to run.
1040 "-static" => link.is_static = true,
1041 "-shared" => link.shared = true,
1042 "-pie" => link.pie = Some(true),
1043 "-no-pie" | "-nopie" => link.pie = Some(false),
1044 "-nostdlib" => link.no_stdlib = true,
1045 "-nostartfiles" => link.no_startfiles = true,
1046 "-nodefaultlibs" => link.no_defaultlibs = true,
1047 "-fno-builtins-lib" => link.no_builtins_lib = true,
1048 "-fbuiltins-lib" => link.no_builtins_lib = false,
1049 "-rdynamic" | "-export-dynamic" => link.export_dynamic = true,
1050 "-s" => link.strip = true,
1051 // Into the ordered input list rather than a list of its own, because a great many of
1052 // the linker's options are a bracket around the files after them and an option that
1053 // lost its place among them says nothing. `--whole-archive` is the one that found this.
1054 "-Xlinker" => {
1055 let next = args.get(i).ok_or_else(|| err("-Xlinker requires an argument"))?;
1056 i += 1;
1057 inputs.push(Input::linker(next));
1058 }
1059 _ if arg.starts_with("-Wl,") => {
1060 // Commas separate arguments rather than being part of one, which is what makes
1061 // `-Wl,-rpath,/opt/lib` two words to the linker and one word here.
1062 inputs.extend(arg["-Wl,".len()..].split(',').map(Input::linker));
1063 }
1064 _ if arg.starts_with("-fuse-ld=") => {
1065 link.use_ld = Some(arg["-fuse-ld=".len()..].to_owned());
1066 }
1067 _ if arg.starts_with("-l") && arg.len() > 2 => {
1068 inputs.push(Input::library(&arg[2..]));
1069 }
1070 "-l" => {
1071 let next = args.get(i).ok_or_else(|| err("-l requires an argument"))?;
1072 i += 1;
1073 inputs.push(Input::library(next));
1074 }
1075 _ if arg.starts_with("-L") => {
1076 link.search.push(PathBuf::from(joined_or_next(arg, 2, args, &mut i)?));
1077 }
1078 _ if arg.starts_with("-B") => {
1079 link.prefixes.push(PathBuf::from(joined_or_next(arg, 2, args, &mut i)?));
1080 }
1081 _ if arg.starts_with("-j") => {
1082 jobs = Jobs::parse(&arg[2..]).map_err(err)?;
1083 }
1084 _ if arg.starts_with("--sysroot=") => {
1085 sysroot = Some(PathBuf::from(&arg["--sysroot=".len()..]));
1086 }
1087 _ if arg.starts_with("--target=") => {
1088 let t = &arg["--target=".len()..];
1089 opts.target = t.parse().map_err(|e| err(format!("{e}")))?;
1090 // The same string again, as the model that has room for a libc version. A spelling
1091 // the three field parser took and this one does not is not an error, because the
1092 // one that decides what is compiled has already accepted it and the only thing
1093 // lost is a version nobody asked for.
1094 pinned = t.parse().ok();
1095 }
1096 _ if arg.starts_with("--emit=") => {
1097 let k = &arg["--emit=".len()..];
1098 opts.emit = k
1099 .parse()
1100 .map_err(|()| err(format!("unknown --emit kind `{k}`, see --help")))?;
1101 }
1102 // A bare `-O` is `-O1`, which is what GCC has and what a hand written makefile tends
1103 // to write. `-Og` is GCC's level for a build somebody is going to step through, and
1104 // it is `-O1` with the transformations that move code around left out; this compiler
1105 // has no such level yet, so it is the nearest one and `--print-pipeline` says what
1106 // that came to rather than the flag pretending otherwise.
1107 "-O" | "-Og" => opts.opt_level = rucc_session::OptLevel::O1,
1108 // The union of `-O3` and `-ffast-math`, and the second half of that changes what
1109 // floating point arithmetic means. Refused rather than taken as `-O3`, because a
1110 // build that asks for fast math and is quietly given ordinary arithmetic gets a
1111 // slower program than it asked for and a build that is given fast math it did not
1112 // ask for gets a wrong one.
1113 "-Ofast" => {
1114 return Err(err(
1115 "-Ofast is -O3 with fast math, and fast math is not implemented, see \
1116 spec/04-driver-and-cli.md section 4.6",
1117 ));
1118 }
1119 _ if arg.starts_with("-O") => {
1120 opts.opt_level = arg[2..]
1121 .parse()
1122 .map_err(|()| err(format!("unknown optimization level `{arg}`")))?;
1123 }
1124 // How far a multiply and an addition may be fused into one rounding. Before the
1125 // optimizer's `-f` family below for the reason the ones under it are, and kept rather
1126 // than dropped because it is the one flag in its group this compiler could act on: it
1127 // rides into the IR as an attribute on each function with a body, so the day the code
1128 // generator forms an `fma` it already knows which functions were given permission.
1129 // Nothing forms one today, under any value of this and under any `-march=`.
1130 _ if arg.starts_with("-ffp-contract=") => {
1131 let how = &arg["-ffp-contract=".len()..];
1132 opts.fp_contract = how.parse().map_err(|()| {
1133 err(format!("`{how}` is not a contraction, which is fast, on or off"))
1134 })?;
1135 }
1136 // How much of an expression may be computed wider than it was written. The values are
1137 // gcc's and so is the refusal of anything else, and none of the three changes anything
1138 // here: an operation is computed in the type C says it is on every target this compiler
1139 // has a back end for, so `__FLT_EVAL_METHOD__` is 0 and `standard` is already what
1140 // happens. `fast` and `16` are permission to be wider, which is a licence this takes
1141 // and does not use, the same way the two above are. The flag is worth taking because
1142 // glibc's headers and a good deal of configure output write it, and because the answer
1143 // it asks about is one this compiler can state rather than guess at: there is no x87
1144 // target here, which is the machine the whole question was invented for.
1145 // Whether a local and a spilled value that are never both wanted may be the same bytes
1146 // of the frame. gcc's three values, and two of them mean the same thing here: what rucc
1147 // shares is a local whose address provably never leaves the function, which is narrower
1148 // than `named_vars` and narrower still than `all`, so both of them get it. `none` is
1149 // the one that changes anything, and it is the flag a program that reads a local
1150 // through a pointer it kept past the end of the block writes.
1151 _ if arg.starts_with("-fstack-reuse=") => {
1152 let how = &arg["-fstack-reuse=".len()..];
1153 opts.stack_reuse = match how {
1154 "all" | "named_vars" => Some(true),
1155 "none" => Some(false),
1156 _ => {
1157 return Err(err(format!(
1158 "`{how}` is not a stack reuse, which is all, named_vars or none"
1159 )));
1160 }
1161 };
1162 }
1163 _ if arg.starts_with("-fexcess-precision=") => {
1164 let how = &arg["-fexcess-precision=".len()..];
1165 if !matches!(how, "16" | "fast" | "standard") {
1166 return Err(err(format!(
1167 "`{how}` is not an excess precision, which is 16, fast or standard"
1168 )));
1169 }
1170 }
1171 // Which front of a path is rewritten before it reaches the output, which is how a
1172 // build gets the same bytes out of two different directories. The four spellings are
1173 // one flag each into three lists, and `-ffile-prefix-map=` is the three of them at
1174 // once. Only the macro list does anything today, because `__FILE__` is the only place
1175 // a path reaches the output: there is no DWARF and no profile data yet, so the other
1176 // two are recorded for the work that will read them. The argument splits at the last
1177 // `=` rather than the first, which is gcc's rule and is what lets a directory with an
1178 // `=` in its name be the old half.
1179 _ if arg.starts_with("-fmacro-prefix-map=") => {
1180 let (old, new) = rewrite(arg, "-fmacro-prefix-map=")?;
1181 opts.prefix_map.macros.push(old, new);
1182 }
1183 _ if arg.starts_with("-fdebug-prefix-map=") => {
1184 let (old, new) = rewrite(arg, "-fdebug-prefix-map=")?;
1185 opts.prefix_map.debug.push(old, new);
1186 }
1187 _ if arg.starts_with("-fprofile-prefix-map=") => {
1188 let (old, new) = rewrite(arg, "-fprofile-prefix-map=")?;
1189 opts.prefix_map.profile.push(old, new);
1190 }
1191 _ if arg.starts_with("-ffile-prefix-map=") => {
1192 let (old, new) = rewrite(arg, "-ffile-prefix-map=")?;
1193 opts.prefix_map.macros.push(old, new);
1194 opts.prefix_map.debug.push(old, new);
1195 opts.prefix_map.profile.push(old, new);
1196 }
1197 // A whole optimization rather than a flag, and the family is taken rather than
1198 // refused because of what ignoring it does. There is none of it here yet, so a build
1199 // that asks for it gets a program that is correct and slower than it could have been,
1200 // which is what section 4.1 means by a hint about speed and what every compilation at
1201 // `-O0` already is. The objects settle the rest of the argument: gcc's `-flto` object
1202 // holds the bytecode and no machine code at all, and every object here holds the code,
1203 // which is exactly what `-ffat-lto-objects` asks gcc for. So a build passing `-flto`
1204 // to this compiler gets objects that are more usable than the ones it asked for rather
1205 // than different ones. Every value is still checked against gcc's, because somebody
1206 // who wrote `-flto=thin` meant clang and had better hear about it here.
1207 "-flto" => opts.lto.requested = true,
1208 "-fno-lto" => opts.lto.requested = false,
1209 _ if arg.starts_with("-flto=") => {
1210 let how = &arg["-flto=".len()..];
1211 opts.lto.jobs = how.parse().map_err(|()| {
1212 err(format!(
1213 "`{how}` is not a number of link time jobs, which is auto, jobserver or a \
1214 count above zero"
1215 ))
1216 })?;
1217 opts.lto.requested = true;
1218 }
1219 _ if arg.starts_with("-flto-partition=") => {
1220 let how = &arg["-flto-partition=".len()..];
1221 opts.lto.partition = how.parse().map_err(|()| {
1222 err(format!(
1223 "`{how}` is not a partitioning model, which is balanced, 1to1, one, max \
1224 or none"
1225 ))
1226 })?;
1227 }
1228 _ if arg.starts_with("-flto-compression-level=") => {
1229 let how = &arg["-flto-compression-level=".len()..];
1230 let level =
1231 how.parse::<u8>().ok().filter(|level| *level <= 19).ok_or_else(|| {
1232 err(format!("`{how}` is not a compression level, 0 to 19"))
1233 })?;
1234 opts.lto.compression = Some(level);
1235 }
1236 // Whether the object keeps its machine code as well as the bytecode. It always does
1237 // here, so the first of these describes what happens and the second asks for an object
1238 // with less in it, which is a smaller file and not a different program, so both are
1239 // taken.
1240 "-ffat-lto-objects" | "-fno-fat-lto-objects" => {}
1241 // Whether the linker is handed a plugin that does the link time work. The design in
1242 // `spec/09-optimizer.md` has this driver doing that work itself and never loading a
1243 // plugin into anybody, so neither answer is a question it has to hold.
1244 "-fuse-linker-plugin" | "-fno-use-linker-plugin" => {}
1245 // Reading a profile back. Taken for the reason the family above it is: nothing here
1246 // reads one, so a build that asks gets the program it would have got anyway, and gcc
1247 // itself produces a byte for byte identical object from `-fprofile-use` when there are
1248 // no counts beside the file. The path is recorded for the pass that will read it. The
1249 // warning gcc prints when it looked and found nothing is deliberately not copied,
1250 // because nothing here looks, and a warning about a file that was never opened would
1251 // fire on the builds that have a perfectly good profile as well as on the ones that
1252 // do not.
1253 "-fprofile-use" => opts.profile_data.requested = true,
1254 "-fno-profile-use" => opts.profile_data.requested = false,
1255 _ if arg.starts_with("-fprofile-use=") => {
1256 opts.profile_data.path = Some(arg["-fprofile-use=".len()..].to_string());
1257 opts.profile_data.requested = true;
1258 }
1259 _ if arg.starts_with("-fprofile-dir=") => {
1260 opts.profile_data.dir = Some(arg["-fprofile-dir=".len()..].to_string());
1261 }
1262 "-fprofile-abs-path" => opts.profile_data.absolute = true,
1263 "-fno-profile-abs-path" => opts.profile_data.absolute = false,
1264 "-fprofile-correction" => opts.profile_data.correction = true,
1265 "-fno-profile-correction" => opts.profile_data.correction = false,
1266 "-fprofile-partial-training" => opts.profile_data.partial_training = true,
1267 "-fno-profile-partial-training" => opts.profile_data.partial_training = false,
1268 // Writing the counts rather than reading them, which is refused rather than taken and
1269 // is the same line `-gsplit-dwarf` falls on the far side of. Ignoring these means a
1270 // file a build declared as an output never appears: the instrumented program writes a
1271 // `.gcda` as it exits and `-ftest-coverage` writes a `.gcno` beside the object, and a
1272 // two stage build that got neither would go on to optimize against no counts at all
1273 // and report coverage of nothing, with nothing along the way saying so. The objects
1274 // say the rest: gcc's `-fprofile-generate` object holds 375 bytes of code where a
1275 // plain one holds 71, and 296 bytes of counters that a plain one does not have, so
1276 // this is a flag that changes the output rather than a hint about speed.
1277 "-fprofile-arcs"
1278 | "--coverage"
1279 | "-fcondition-coverage"
1280 | "-fpath-coverage"
1281 | "-fprofile-generate" => {
1282 return Err(err(format!(
1283 "{arg}: this compiler does not instrument for profiling, and a build that \
1284 expects the counts a run of the instrumented program writes would optimize \
1285 against nothing on its second pass, see spec/04-driver-and-cli.md"
1286 )));
1287 }
1288 _ if arg.starts_with("-fprofile-generate=") => {
1289 return Err(err(format!(
1290 "{arg}: this compiler does not instrument for profiling, and a build that \
1291 expects the counts a run of the instrumented program writes would optimize \
1292 against nothing on its second pass, see spec/04-driver-and-cli.md"
1293 )));
1294 }
1295 "-ftest-coverage" => {
1296 return Err(err(format!(
1297 "{arg}: this compiler writes no `.gcno` file beside the object, and a build \
1298 that expects one would wait for a file that never arrives, see \
1299 spec/04-driver-and-cli.md"
1300 )));
1301 }
1302 // The rest of the family describes instrumentation that is refused above, so what is
1303 // left to do with them is check them and drop them. They are checked because a
1304 // misspelling in a distribution's flags is worth finding here rather than on the day
1305 // the instrumentation lands, and dropped because there is nothing for an answer about
1306 // how a counter is written to be an answer about.
1307 _ if arg.starts_with("-fprofile-update=") => {
1308 let how = &arg["-fprofile-update=".len()..];
1309 if !matches!(how, "single" | "atomic" | "prefer-atomic") {
1310 return Err(err(format!(
1311 "`{how}` is not a profile update method, which is single, atomic or \
1312 prefer-atomic"
1313 )));
1314 }
1315 }
1316 _ if arg.starts_with("-fprofile-reproducible=") => {
1317 let how = &arg["-fprofile-reproducible=".len()..];
1318 if !matches!(how, "serial" | "parallel-runs" | "multithreaded") {
1319 return Err(err(format!(
1320 "`{how}` is not a profile reproducibility method, which is serial, \
1321 parallel-runs or multithreaded"
1322 )));
1323 }
1324 }
1325 "-fprofile-values" | "-fno-profile-values" | "-fprofile-info-section" => {}
1326 "-fno-test-coverage" | "-fno-profile-arcs" | "-fno-profile-generate" => {}
1327 _ if arg.starts_with("-fprofile-filter-files=")
1328 || arg.starts_with("-fprofile-exclude-files=")
1329 || arg.starts_with("-fprofile-note=") => {}
1330 // What every name gets when nothing in the source said, which the attribute in the
1331 // source overrides rather than the other way round. Before the optimizer's `-f`
1332 // family below for the reason the tier below it is.
1333 _ if arg.starts_with("-fvisibility=") => {
1334 let seen = &arg["-fvisibility=".len()..];
1335 opts.visibility = seen.parse().map_err(|()| {
1336 err(format!(
1337 "`{seen}` is not a visibility, which is default, hidden, internal or \
1338 protected"
1339 ))
1340 })?;
1341 }
1342 // Which edges of a control flow transfer are checked. Before the optimizer's `-f`
1343 // family below for the reason the two above it are, and last of the three so that the
1344 // bare spelling and the negative one are matched exactly rather than by this.
1345 _ if arg.starts_with("-fcf-protection=") => {
1346 let edges = &arg["-fcf-protection=".len()..];
1347 opts.control = edges.parse().map_err(|()| {
1348 err(format!(
1349 "`{edges}` is not a control flow protection, which is full, branch, \
1350 return, none or check"
1351 ))
1352 })?;
1353 }
1354 // How much room every function opens with for something to be written over later.
1355 // Before the optimizer's `-f` family below for the reason the ones above it are.
1356 _ if arg.starts_with("-fpatchable-function-entry=") => {
1357 let room = &arg["-fpatchable-function-entry=".len()..];
1358 opts.patchable = room.parse().map_err(|()| {
1359 err(format!(
1360 "`{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"
1361 ))
1362 })?;
1363 }
1364 // The memory safety monitor, from section 15.4 of
1365 // `spec/safe-memory/15-integration.md`. Before the optimizer's `-f` family below,
1366 // because a pass that took the name `safety=detect` would otherwise be handed the
1367 // flag, and the tier is not a pass.
1368 _ if arg.starts_with("-fsafety=") => {
1369 let tier = &arg["-fsafety=".len()..];
1370 opts.safety = tier.parse().map_err(|()| {
1371 err(format!(
1372 "`{tier}` is not a safety tier, which is off, detect, enforce or kernel"
1373 ))
1374 })?;
1375 }
1376 // Whether padding participates, from section 9.3 of document 09. Spelled out rather
1377 // than folded into the tier because it is a departure somebody who has read that
1378 // section makes, and the two defaults it describes are a property of what is being
1379 // built rather than of how much checking is wanted.
1380 _ if arg.starts_with("-fsafety-init=") => {
1381 let mode = &arg["-fsafety-init=".len()..];
1382 opts.padding = mode.parse().map_err(|()| {
1383 err(format!("`{mode}` is not a padding mode, which is padding or nopadding"))
1384 })?;
1385 }
1386 // Row S4, from section 9.4 of document 09. A bare flag with no value, because the
1387 // strict form of that section needs a member id the front end does not name yet and
1388 // accepting the spelling for it would be accepting a promise this build cannot keep.
1389 // Before `-fno-` is looked at below, for the reason the tier is.
1390 "-fsafety-subobject" => opts.subobject = rucc_session::Subobject::Members,
1391 "-fno-safety-subobject" => opts.subobject = rucc_session::Subobject::Off,
1392 _ if arg.starts_with("-fsafety-subobject=") => {
1393 let form = &arg["-fsafety-subobject=".len()..];
1394 return Err(err(format!(
1395 "`{form}` is not a form of -fsafety-subobject. The flag takes no value, and \
1396 the strict form of section 9.4 is tamnd/rucc#967"
1397 )));
1398 }
1399 // Row Y8, from section 9.6 of document 09. A bare flag with no value, for the reason
1400 // the one above has none: there is one form of this check and a spelling that suggested
1401 // otherwise would be promising something. Before `-fno-` is looked at below, the same
1402 // way.
1403 "-fsafety-restrict" => opts.promise = rucc_session::Promise::Blocks,
1404 "-fno-safety-restrict" => opts.promise = rucc_session::Promise::Off,
1405 _ if arg.starts_with("-fsafety-restrict=") => {
1406 let form = &arg["-fsafety-restrict=".len()..];
1407 return Err(err(format!(
1408 "`{form}` is not a form of -fsafety-restrict. The flag takes no value."
1409 )));
1410 }
1411 // Section 9.5's races, which take a value because the section gives them three modes
1412 // and the difference between two of them is which classes get reported rather than how
1413 // much is recorded. `-fno-` is the same as `=off` and is spelled out here for the same
1414 // reason the two above spell theirs out.
1415 _ if arg.starts_with("-fsafety-races=") => {
1416 let mode = &arg["-fsafety-races=".len()..];
1417 opts.races = mode.parse().map_err(|()| {
1418 err(format!("`{mode}` is not a race mode, which is off, metadata or pointer"))
1419 })?;
1420 }
1421 "-fno-safety-races" => opts.races = rucc_session::Races::Off,
1422 // The sanitizers of document 12, which are checks at run time rather than a way of
1423 // generating the same program. Each name is held to gcc 16's list, and what is still
1424 // asked for by the end of the line is answered after the loop, so that a command line
1425 // which turns one on and then off again is a command line that asked for nothing.
1426 //
1427 // Before the optimizer's `-f` family below, for the reason the tier above it is.
1428 _ if arg.starts_with("-fsanitize=") => {
1429 for one in arg["-fsanitize=".len()..].split(',') {
1430 if one == "all" {
1431 // gcc takes `all` only in the negative, because turning every check on at
1432 // once includes checks that contradict each other.
1433 return Err(err(
1434 "`-fsanitize=all` is not a gcc option, only `-fno-sanitize=all` is",
1435 ));
1436 }
1437 if !SANITIZERS.contains(&one) {
1438 return Err(err(format!(
1439 "`{one}` is not a sanitizer, see spec/04-driver-and-cli.md section 4.7"
1440 )));
1441 }
1442 if !sanitizers.contains(&one) {
1443 sanitizers.push(one);
1444 }
1445 }
1446 }
1447 _ if arg.starts_with("-fno-sanitize=") => {
1448 for one in arg["-fno-sanitize=".len()..].split(',') {
1449 if one == "all" {
1450 sanitizers.clear();
1451 continue;
1452 }
1453 if !SANITIZERS.contains(&one) {
1454 return Err(err(format!(
1455 "`{one}` is not a sanitizer, see spec/04-driver-and-cli.md section 4.7"
1456 )));
1457 }
1458 sanitizers.retain(|asked| *asked != one);
1459 }
1460 }
1461 // What a check does when it fires, and where the records about the checked objects go.
1462 // Each of them is an answer about the sanitizers refused after the loop, so there is
1463 // nothing left for them to change here. The names are still held to the list, because
1464 // a misspelling in a build's flags is worth finding when the compiler reads it.
1465 _ if arg.starts_with("-fsanitize-recover=")
1466 || arg.starts_with("-fno-sanitize-recover=")
1467 || arg.starts_with("-fsanitize-trap=")
1468 || arg.starts_with("-fno-sanitize-trap=") =>
1469 {
1470 // The guard above matched on a spelling that has an `=` in it, so the tail is
1471 // whatever follows the first one.
1472 let how = arg.split_once('=').map_or("", |(_, rest)| rest);
1473 for one in how.split(',') {
1474 if one != "all" && !SANITIZERS.contains(&one) {
1475 return Err(err(format!(
1476 "`{one}` is not a sanitizer, see spec/04-driver-and-cli.md section 4.7"
1477 )));
1478 }
1479 }
1480 }
1481 "-fsanitize-undefined-trap-on-error"
1482 | "-fsanitize-address-use-after-scope"
1483 | "-fno-sanitize-address-use-after-scope" => {}
1484 _ if arg.starts_with("-fsanitize-sections=") => {}
1485 // Counting which edges a run reached, which is how a fuzzer knows an input was worth
1486 // keeping. Refused rather than dropped, because a fuzzer whose calls into
1487 // `__sanitizer_cov_*` were never generated runs blind and reports coverage of nothing,
1488 // and there is no point in the campaign where that announces itself.
1489 _ if arg.starts_with("-fsanitize-coverage=") => {
1490 let how = &arg["-fsanitize-coverage=".len()..];
1491 for one in how.split(',') {
1492 if !matches!(one, "trace-pc" | "trace-cmp") {
1493 return Err(err(format!(
1494 "`{one}` is not a coverage instrumentation, which is trace-pc or \
1495 trace-cmp"
1496 )));
1497 }
1498 }
1499 return Err(err(format!(
1500 "{arg}: this compiler generates no coverage callbacks, and a fuzzer built \
1501 with it would run without any feedback at all, see \
1502 spec/04-driver-and-cli.md section 4.7"
1503 )));
1504 }
1505 // The optimizer's own flags, from section 9.10 of `spec/09-optimizer.md`. These come
1506 // after every `-f` the rest of the compiler answers to, so a pass can never take a
1507 // name that already means something else on the command line.
1508 _ if arg.starts_with("-fpass-fuel=") => {
1509 let (name, count) = arg["-fpass-fuel=".len()..]
1510 .split_once('=')
1511 .ok_or_else(|| err("-fpass-fuel= is spelled <pass>=<count>"))?;
1512 if rucc_opt::pass::find(name).is_none() {
1513 return Err(err(format!(
1514 "`{name}` is not a pass this compiler has, see --print-pipeline"
1515 )));
1516 }
1517 let count: u32 = count
1518 .parse()
1519 .map_err(|_| err(format!("`{count}` is not a number of transformations")))?;
1520 opts.pass_fuel.push((name.to_owned(), count));
1521 }
1522 _ if arg.starts_with("-fpass-fuel-global=") => {
1523 let count = &arg["-fpass-fuel-global=".len()..];
1524 let count: u32 = count
1525 .parse()
1526 .map_err(|_| err(format!("`{count}` is not a number of transformations")))?;
1527 opts.pass_fuel_global = Some(count);
1528 }
1529 // Everything from `-fopt-info` to the end of the argument, which is optional
1530 // keywords joined by hyphens and an optional `=<file>`. Checked here rather than
1531 // where the remarks are printed, because by then the compilation somebody wanted
1532 // to hear about is over.
1533 _ if arg == "-fopt-info"
1534 || arg.starts_with("-fopt-info=")
1535 || arg.starts_with("-fopt-info-") =>
1536 {
1537 let rest = &arg["-fopt-info".len()..];
1538 let (kinds, file) = match rest.split_once('=') {
1539 Some((kinds, file)) => (kinds, Some(file)),
1540 None => (rest, None),
1541 };
1542 let kinds = kinds.strip_prefix('-').unwrap_or(kinds);
1543 rucc_opt::Wants::none().add(kinds).map_err(err)?;
1544 opts.opt_info.push(kinds.to_owned());
1545 if let Some(file) = file {
1546 if file.is_empty() {
1547 return Err(err("-fopt-info= was given no file to write to"));
1548 }
1549 opts.opt_info_file = Some(file.to_owned());
1550 }
1551 }
1552 _ if arg.starts_with("-fdump-ir=") => {
1553 // Checked here rather than where the dumps are taken, because the compilation
1554 // that would have been dumped is over by then.
1555 let spec = &arg["-fdump-ir=".len()..];
1556 rucc_opt::Dumps::default().add(spec).map_err(err)?;
1557 opts.dump_ir.push(spec.to_owned());
1558 }
1559 // Before the bare `-f<pass>` below, because a pass called `enable-something` would
1560 // otherwise take the flag away from the gate. Checked here rather than where the
1561 // pipeline reads it, for the reason that applies to all of these: a misspelled pass
1562 // name that quietly gated nothing looks exactly like a pass that is not the guilty
1563 // one, and a bisection would carry on past the thing it was looking for.
1564 _ if arg.starts_with("-fdisable-") || arg.starts_with("-fenable-") => {
1565 let on = arg.starts_with("-fenable-");
1566 let spec = &arg[if on { "-fenable-".len() } else { "-fdisable-".len() }..];
1567 rucc_opt::Gates::default().add(on, spec).map_err(err)?;
1568 opts.pass_gates.push((on, spec.to_owned()));
1569 }
1570 // gcc's spelling for a pass this compiler has under a shorter name. It goes above the
1571 // two arms below rather than into the pile of gcc pass names further down, because the
1572 // pass is here: dropping the flag would leave a build that asked for unrolling without
1573 // it, and refusing it stops the build outright, which is what libtommath's makefile
1574 // ran into. `-funroll-all-loops` is deliberately not in here: gcc's is the one that
1575 // unrolls without a trip count, which is a different and usually worse thing.
1576 "-funroll-loops" => opts.passes.push(("unroll".to_owned(), true)),
1577 "-fno-unroll-loops" => opts.passes.push(("unroll".to_owned(), false)),
1578 // Here rather than through the two arms below, because what this names is not a
1579 // `rucc_opt::Pass`. Section 34.6's propagation is a module at a time and everything in
1580 // the pass list is one function at a time. `-fipa-cp-clone` is deliberately not here:
1581 // gcc turns that one on at `-O3` and it is in the list of what M4 does not build.
1582 "-fipa-cp" => opts.passes.push((rucc_opt::ipcp::NAME.to_owned(), true)),
1583 "-fno-ipa-cp" => opts.passes.push((rucc_opt::ipcp::NAME.to_owned(), false)),
1584 // The other half of the same section, here for the same reason, and `-fipa-sra` in gcc
1585 // is the aggregate splitting as well as the parameter removal. Asking for it gets the
1586 // half that is built.
1587 "-fipa-sra" => opts.passes.push((rucc_opt::ipasra::NAME.to_owned(), true)),
1588 "-fno-ipa-sra" => opts.passes.push((rucc_opt::ipasra::NAME.to_owned(), false)),
1589 // And the printf family fold, which is a module at a time for the same reason and so is
1590 // not a `rucc_opt::Pass` either. gcc has no flag of its own for this one, since
1591 // `-fno-builtin` already turns it off along with everything else the standard names
1592 // mean. This spelling is for taking one thing away during a bisection without taking
1593 // the rest of section 20.1 away with it.
1594 "-flibcall" => opts.passes.push((rucc_opt::libcall::NAME.to_owned(), true)),
1595 "-fno-libcall" => opts.passes.push((rucc_opt::libcall::NAME.to_owned(), false)),
1596 _ if arg.strip_prefix("-fno-").is_some_and(|n| rucc_opt::pass::find(n).is_some()) => {
1597 opts.passes.push((arg["-fno-".len()..].to_owned(), false));
1598 }
1599 _ if arg.strip_prefix("-f").is_some_and(|n| rucc_opt::pass::find(n).is_some()) => {
1600 opts.passes.push((arg["-f".len()..].to_owned(), true));
1601 }
1602 // The flags that name a pass of gcc's own. They arrive from the torture suite, where a
1603 // program reduced from a miscompilation usually names the pass that miscompiled it on
1604 // its `dg-options` line, and they arrive from hand written build files for the same
1605 // reason. Section 4.1 sorts a flag by what the output would be without it, and by that
1606 // rule these are one pile: a flag that turns one of gcc's passes on or off is asking
1607 // for a compiler that does not exist here, and the program it is attached to is a
1608 // correctness test that passes either way. Turning on a pass we do not have costs
1609 // speed, turning off a pass we do not have costs nothing, and neither changes what the
1610 // program computes.
1611 //
1612 // rucc's own pass names are matched above this, so `-fno-dce` turns off the dce this
1613 // compiler has rather than landing here, and the day one of these names becomes a pass
1614 // here it stops being taken and dropped without anybody editing this list.
1615 //
1616 // Two of them are prefixes rather than names, which is the one place this file takes a
1617 // family instead of a flag. gcc files its gimple passes under `-ftree-` and its
1618 // interprocedural passes under `-fipa-`, both namespaces are pass selection and
1619 // nothing else, and there is no member of either that changes the meaning of a program
1620 // that was already correct. The rest are written out one at a time, because they live
1621 // in the flat `-f` namespace where the neighbours do change meanings.
1622 _ if arg.starts_with("-ftree-") || arg.starts_with("-fno-tree-") => {}
1623 _ if arg.starts_with("-fipa-") || arg.starts_with("-fno-ipa-") => {}
1624 "-fexpensive-optimizations" | "-fno-expensive-optimizations" => {}
1625 "-fmodulo-sched" | "-fno-modulo-sched" => {}
1626 "-fvect-cost-model" | "-fno-vect-cost-model" => {}
1627 _ if arg.starts_with("-fvect-cost-model=") || arg.starts_with("-fsimd-cost-model=") => {
1628 }
1629 "-fearly-inlining" | "-fno-early-inlining" => {}
1630 "-finline"
1631 | "-fno-inline"
1632 | "-finline-functions"
1633 | "-fno-inline-functions"
1634 | "-finline-small-functions"
1635 | "-fno-inline-small-functions"
1636 | "-finline-functions-called-once"
1637 | "-fno-inline-functions-called-once" => {}
1638 "-foptimize-strlen" | "-fno-optimize-strlen" => {}
1639 "-fira-share-spill-slots" | "-fno-ira-share-spill-slots" => {}
1640 // Where a function starts, which is a thing this compiler already decides and so is a
1641 // request it can answer rather than one it has to drop. The bare form asks for the
1642 // target's default and the default here is the sixteen bytes gcc also gives, so it
1643 // says nothing; a number is a floor under every function that did not ask for more
1644 // itself; and the negative form asks for the smallest boundary the target has. gcc 16
1645 // rounds a number that is not a power of two up rather than refusing it, which is what
1646 // `=3` giving `.p2align 2` on x86-64 means, so this rounds too.
1647 "-falign-functions" => opts.align_functions = None,
1648 "-fno-align-functions" => opts.align_functions = Some(MIN_FUNC_ALIGN),
1649 _ if arg.starts_with("-falign-functions=") => {
1650 opts.align_functions = function_alignment(&arg["-falign-functions=".len()..])
1651 .ok_or_else(|| {
1652 err(format!("{arg}: the alignment has to be a number of bytes"))
1653 })?;
1654 }
1655 // The other three of the family, which are about padding in front of a label inside a
1656 // body. This compiler writes none, and what they ask for is speed: a loop that starts
1657 // on a cache line boundary computes what a loop that does not computes. So they are
1658 // taken and dropped for the reason `-march=` is, and they are kept out of the arm
1659 // above because the question they ask is a different one and the day one of them is
1660 // answered it will be answered separately.
1661 _ if arg.starts_with("-falign-labels")
1662 || arg.starts_with("-falign-loops")
1663 || arg.starts_with("-falign-jumps")
1664 || arg.starts_with("-fno-align-labels")
1665 || arg.starts_with("-fno-align-loops")
1666 || arg.starts_with("-fno-align-jumps") => {}
1667 // The charset flags are not in that pile, because an encoding is a statement about
1668 // what the bytes of the source mean rather than about how fast the output is. The
1669 // preprocessor reads UTF-8 and has no converter, so the one name that describes what
1670 // already happens is taken and every other name is refused. Spelled without regard to
1671 // case and with both of the spellings iconv answers to, since a build writes whichever
1672 // one its author typed.
1673 _ if arg.starts_with("-finput-charset=") => {
1674 let name = &arg["-finput-charset=".len()..];
1675 if !name.eq_ignore_ascii_case("utf-8") && !name.eq_ignore_ascii_case("utf8") {
1676 return Err(err(format!(
1677 "-finput-charset={name}: the preprocessor reads UTF-8 and has no \
1678 converter, so a file in another encoding would be read as though it were \
1679 UTF-8 rather than converted",
1680 )));
1681 }
1682 }
1683 // The three that come in on the same `dg-options` lines and are the other half of
1684 // section 4.1's rule, because each of them changes what the program does and not how
1685 // fast it does it. The negative form of each is what this compiler does anyway, so it
1686 // is taken and dropped, which is the shape `-fnested-functions` has above.
1687 "-ffast-math" => {
1688 return Err(err(
1689 "-ffast-math is a licence to answer a floating point arithmetic differently \
1690 from the way the source wrote it, and it is not one flag: it defines \
1691 __FAST_MATH__, which a library header reads, and gcc links a startup file \
1692 that puts the hardware in flush to zero mode for the whole process. Taking it \
1693 and dropping it would change what other objects in the same program answer. \
1694 -ffp-contract= and -fexcess-precision= are the parts of it this compiler has",
1695 ));
1696 }
1697 "-fno-fast-math" => {}
1698 "-fnon-call-exceptions" => {
1699 return Err(err(
1700 "-fnon-call-exceptions is a promise that an instruction which is not a call \
1701 can raise an exception the unwinder finds a handler for, and nothing here \
1702 produces a landing pad for a trapping instruction. A program built without it \
1703 would unwind past the handler it wrote",
1704 ));
1705 }
1706 "-fno-non-call-exceptions" => {}
1707 "-finstrument-functions" => {
1708 return Err(err(
1709 "-finstrument-functions calls __cyg_profile_func_enter on entry to every \
1710 function and __cyg_profile_func_exit on the way out, and nothing here emits \
1711 either call. A program that asks for them usually counts them, so taking the \
1712 flag and dropping it would turn a program that fails loudly into one that \
1713 fails quietly",
1714 ));
1715 }
1716 "-fno-instrument-functions" => {}
1717 // The unstable options, spelled the way rustc spells them and carrying the same
1718 // promise, which is none: one of these may change or go away in any release. They are
1719 // measurements and debugging aids rather than things a build asks for, which is why
1720 // none of them is in the usage text and all of them are in section 4.11 of
1721 // `spec/04-driver-and-cli.md`.
1722 "-Zverify-each" => opts.verify_each = true,
1723 _ if arg.starts_with("-Zrule-coverage=") => {
1724 let file = &arg["-Zrule-coverage=".len()..];
1725 if file.is_empty() {
1726 return Err(err("-Zrule-coverage= needs a file to write to"));
1727 }
1728 opts.rule_coverage = Some(file.to_owned());
1729 }
1730 _ if arg.starts_with("-Zcycle-accurate-model=") => {
1731 let value = &arg["-Zcycle-accurate-model=".len()..];
1732 opts.cycle_accurate_model = match value {
1733 "yes" | "1" => Some(true),
1734 "no" | "0" => Some(false),
1735 _ => {
1736 return Err(err("-Zcycle-accurate-model= takes yes or no"));
1737 }
1738 };
1739 }
1740 _ if arg.starts_with("-Zlowering=") => {
1741 let file = &arg["-Zlowering=".len()..];
1742 if file.is_empty() {
1743 return Err(err("-Zlowering= needs a file to write to"));
1744 }
1745 opts.lowering_dump = Some(file.to_owned());
1746 }
1747 _ if arg.starts_with("-Zregister-pressure=") => {
1748 let file = &arg["-Zregister-pressure=".len()..];
1749 if file.is_empty() {
1750 return Err(err("-Zregister-pressure= needs a file to write to"));
1751 }
1752 opts.register_pressure = Some(file.to_owned());
1753 }
1754 _ if arg.starts_with("-Z") => {
1755 return Err(err(format!(
1756 "`{arg}` is not an unstable option this compiler has, see \
1757 spec/04-driver-and-cli.md section 4.11 for the ones it does"
1758 )));
1759 }
1760 // The word size, which is a statement about the target and is taken as one. A build
1761 // that says the size the target already has is saying nothing, and one that says the
1762 // other size is asking for a target this compiler does not have, which it is told
1763 // rather than being given the wrong one.
1764 "-m64" | "-m32" | "-mx32" => {
1765 let want: u32 = match arg {
1766 "-m64" => 64,
1767 _ => 32,
1768 };
1769 let have = rucc_target::TargetInfo::new(opts.target).pointer_width;
1770 if have != want {
1771 return Err(err(format!(
1772 "{arg} asks for a {want} bit target and {} is {have} bit, use \
1773 --target= to name the one you mean",
1774 opts.target
1775 )));
1776 }
1777 }
1778 // Which processor in the family to generate for. This compiler emits the base
1779 // instruction set of the architecture and nothing above it, so a program built with
1780 // any of these runs on the machine that was named; it is a program that could have
1781 // been faster rather than a program that is wrong, which is what makes these safe to
1782 // take and ignore where a flag that changed the meaning of the code would not be.
1783 _ if arg.starts_with("-march=")
1784 || arg.starts_with("-mtune=")
1785 || arg.starts_with("-mcpu=") => {}
1786 // The calling convention, which is not safe to ignore. Taken when it names the one
1787 // the target already uses and refused otherwise.
1788 _ if arg.starts_with("-mabi=") => {
1789 let want = &arg["-mabi=".len()..];
1790 let have = match opts.target.arch {
1791 rucc_target::Arch::X86_64 => "sysv",
1792 rucc_target::Arch::Aarch64 => "lp64",
1793 rucc_target::Arch::Riscv64 => "lp64d",
1794 };
1795 if want != have {
1796 return Err(err(format!(
1797 "{arg}: {} uses the {have} convention and this compiler has no other",
1798 opts.target
1799 )));
1800 }
1801 }
1802 // How far apart the pieces of the program may be. The small model is what we emit and
1803 // it is every hosted program's default; the kernel model is a different one and a
1804 // build that asks for it and does not get it links and then does not run.
1805 "-mcmodel=small" => {}
1806 _ if arg.starts_with("-mcmodel=") => {
1807 return Err(err(format!(
1808 "{arg}: this compiler emits the small code model and no other, see \
1809 spec/12-targets.md"
1810 )));
1811 }
1812 // GCC's own scripting language for how the driver builds a command line.
1813 // `spec/04-driver-and-cli.md` section 4.4 settles that we will not have it, so a
1814 // build reaching for it is told which flags do the same job.
1815 _ if arg.starts_with("-specs=") => {
1816 return Err(err(
1817 "-specs= is not supported: the parts of it builds rely on are -B, -L, \
1818 -nostdlib, -nostartfiles and -Wl,, see spec/04-driver-and-cli.md \
1819 section 4.4",
1820 ));
1821 }
1822 // Arguments meant for a separate assembler or preprocessor, which this compiler does
1823 // not have: both are inside it and neither reads a command line. Refused rather than
1824 // dropped, because every one of these says something about the output and a build
1825 // that asked for `-Wa,--noexecstack` and was silently given an executable stack got
1826 // the opposite of what it asked for.
1827 _ if arg.starts_with("-Wa,") || arg.starts_with("-Wp,") => {
1828 return Err(err(format!(
1829 "`{arg}` is an argument for a separate assembler or preprocessor, and both \
1830 are inside this compiler rather than programs it runs"
1831 )));
1832 }
1833 "-Xassembler" | "-Xpreprocessor" => {
1834 return Err(err(format!(
1835 "{arg} hands an argument to a separate assembler or preprocessor, and both \
1836 are inside this compiler rather than programs it runs"
1837 )));
1838 }
1839 // Everything else in the `-W` family. `spec/04-driver-and-cli.md` section 4.1 has
1840 // this one as a rule about build systems rather than about warnings: autoconf finds
1841 // out whether a warning flag exists by passing it and looking at the exit status, so
1842 // a compiler that refuses one it has not heard of fails a configure script written
1843 // for a GCC newer than itself. The names are not checked against a list because this
1844 // compiler has no warning groups for a list to be of, which #485 is about.
1845 _ if arg.starts_with("-W") => {}
1846 // Flags that name something this compiler does not do and would not do differently
1847 // if it did. `-fno-ident` is about a comment in the output that we do not write
1848 // either way, and the others are about a way of ordering the compilation that has
1849 // been GCC's only way for twenty years. Section 4.1 asks for the list to be short
1850 // and for adding to it to be deliberate, which is why it is written out here.
1851 "-fno-ident"
1852 | "-fident"
1853 | "-funit-at-a-time"
1854 | "-fno-unit-at-a-time"
1855 | "-shared-libgcc"
1856 | "-static-libgcc" => {}
1857 _ if arg.starts_with('-') && arg.len() > 1 => {
1858 // Silently ignoring an unknown flag is how a build ends up not doing what
1859 // its author asked. spec/13-gnu-compat.md section 13.4 makes this an error
1860 // for the flags that change code generation, and the safe default until the
1861 // flag table is populated is to reject everything we do not know.
1862 return Err(err(format!("unknown option `{arg}`")));
1863 }
1864 _ => inputs.push(Input { path: arg.to_owned(), forced, role: Role::File }),
1865 }
1866 }
1867
1868 // The fetch, before anything that resolves a compilation, because `--fetch` does not describe
1869 // one. It is here rather than in the loop so that `--offline` can forbid it whichever order the
1870 // two were written in, and it is before the refusals below so that a command line asking for a
1871 // sysroot is not told about a sanitizer.
1872 if let Some(named) = fetch {
1873 if fetch_msvc.is_some() {
1874 return Err(err(
1875 "--fetch and --fetch-msvc-sdk are two different commands and this command line \
1876 asked for both. --fetch gets a sysroot this release pins by URL and by hash, and \
1877 --fetch-msvc-sdk gets what is behind Microsoft's licence wall, which no release \
1878 pins and which nobody may republish. Run whichever one you meant",
1879 ));
1880 }
1881 return fetch_action(&named, offline, &inputs);
1882 }
1883 if let Some(named) = fetch_msvc {
1884 return fetch_msvc_action(&named, offline, accepted, &inputs);
1885 }
1886 if accepted {
1887 return Err(err(
1888 "--accept-licence says that Microsoft's Visual Studio Build Tools licence is accepted, \
1889 and nothing on this command line asked for anything that licence covers. \
1890 --fetch-msvc-sdk <tuple> is the command it belongs to, and an ordinary compile \
1891 downloads nothing with it or without it",
1892 ));
1893 }
1894
1895 // Last, so that it lands after every `-isystem` the command line gave. That is GCC's
1896 // order: a directory the user names outranks the compiler's own, and the compiler's own
1897 // outranks the library's. It is pushed after the loop rather than before it because
1898 // `SearchPath` appends within a group and the position is what the order is.
1899 // The same directory the headers were looked for under, because a sysroot is a statement
1900 // about a whole installation and not about half of one.
1901 // After the loop, because `-fno-sanitize=` can take back what an earlier flag asked for and a
1902 // command line that turns a check on and off again has asked for nothing. What is left is
1903 // refused rather than dropped, and it is the one place in this parser where the reason is not
1904 // that the output would differ. A sanitizer is a promise that the program is watched while it
1905 // runs, so a build that asks for one and is quietly given a program with no checks in it does
1906 // not get a slower program or a bigger file, it gets a test suite that passes for the wrong
1907 // reason. `-fsafety=` is the checking this compiler does have, and the message says so, because
1908 // somebody reaching for `-fsanitize=address` wants the nearest thing rather than a list of
1909 // options.
1910 if let Some(first) = sanitizers.first() {
1911 return Err(err(format!(
1912 "-fsanitize={first}: this compiler has no sanitizer instrumentation, and a build that \
1913 asked for one and got none would run its tests unchecked, see \
1914 spec/04-driver-and-cli.md section 4.7. `-fsafety=detect` is the memory checking this \
1915 compiler does have"
1916 )));
1917 }
1918 link.sysroot = sysroot.clone();
1919 // Where a sysroot for a target that is not this machine would be. Read once, here, rather than
1920 // inside the link line, because a link line that read the environment could only be tested on a
1921 // machine whose environment said the right thing, and the link line is the last thing that
1922 // touches a binary. `spec/cross-compile/13-distribution.md` section 13.2 owns the answer.
1923 link.cache = Some(cache::dir());
1924 // And the ten field spelling of the target, because the release on it decides two things the
1925 // three field one cannot say: whether a target that is this architecture is still a cross
1926 // compile, and which directory under the cache it is against. After the loop because the last
1927 // `--target=` on the command line is the one that counts.
1928 link.pinned = pinned;
1929 // After the loop rather than where `-pthread` was read, so that it lands after the objects
1930 // that refer to it. A static link takes the definitions it needs from a library when it
1931 // reaches it and not afterwards, so a library before the objects is a library that answers
1932 // nothing.
1933 if threads {
1934 inputs.push(Input::library("pthread"));
1935 }
1936 if let Some(query) = query {
1937 return Ok(Action::Print(answer(&query, &opts, &link)?));
1938 }
1939 // `-M` and `-MM` produce the rule and nothing else, so the run stops after phase 4 whatever
1940 // else the command line asked for. Read here rather than where the flag was, because a `-c`
1941 // written after it has to lose and the loop cannot know that until it has ended. The output
1942 // file is where the rule goes rather than where an object would have gone, and the last
1943 // phase being the preprocessor is what makes that true without a second rule for it.
1944 if opts.deps.instead_of_compiling {
1945 opts.emit = EmitKind::Preprocessed;
1946 }
1947 if !nostdinc {
1948 opts.search.push_system(runtime::DIR);
1949 // And the library's after ours, which is the other half of the same order. They go on
1950 // here rather than at the point `--target=` or `--sysroot=` was read because either
1951 // one changes the answer and the last word on both is the end of the loop.
1952 //
1953 // Which library's is the question `link::cross_sysroot` answers, and it is asked here so
1954 // that the headers and the libraries come from the same place. A target that is this
1955 // machine reads this machine's headers, and a target that is not reads the ones in the
1956 // sysroot for it rather than the ones next door.
1957 let cross = link::cross_sysroot(opts.target, &link);
1958 let kernel = link::cross_kernel(opts.target, &link);
1959 // And the version of those headers, which only the bundled tree has an answer for. A host
1960 // glibc and a tree the user named both define `__GLIBC_MINOR__` in their own `features.h`,
1961 // and a second definition with a different value is a warning on every file, so the
1962 // condition is the same one that chose the directories.
1963 if cross.is_some() {
1964 let target = pinned.unwrap_or_else(|| opts.target.tuple());
1965 opts.glibc_minor = rucc_sysroot::bundled_glibc_minor(target).map_err(|skew| {
1966 err(format!(
1967 "{skew}; pin a release the tree has, or name a tree that has that one \
1968 with --sysroot"
1969 ))
1970 })?;
1971 }
1972 let system =
1973 library::header_dirs(opts.target, sysroot.as_deref(), cross.as_ref(), kernel.as_ref());
1974 // The two licence walls of `spec/cross-compile/13-distribution.md` section 13.4, which are
1975 // the only way step 3 comes back with nothing on a hosted target. Section 8.6 asks for the
1976 // answer to name the licence and the lawful ways to get what is behind it, rather than
1977 // leaving a person with an `#include` that failed as though a directory had gone missing.
1978 //
1979 // It is left on the search path instead of refused here, because a program that includes
1980 // none of the library needs none of the SDK and section 8.6 is explicit that targeting the
1981 // platform has to keep working. So the reason waits until an include has actually failed,
1982 // which is the only moment it helps and the only moment it is true.
1983 //
1984 // The condition is that step 3 found nothing at all, so an `SDKROOT`, an `INCLUDE` or a mac
1985 // with Xcode on it all pass through untouched, and `-nostdinc` never reaches this block. A
1986 // `--sysroot` or `-isysroot` passes through as well, even when the tree it names turns out to
1987 // be empty or absent: somebody who wrote a path has already answered the question this
1988 // message asks, and answering it again over the top of a mistyped directory would hide the
1989 // mistake behind a licence notice.
1990 if system.is_empty() && sysroot.is_none() {
1991 let tuple = pinned.unwrap_or_else(|| opts.target.tuple());
1992 if let Some(wall) = rucc_sysroot::Wall::of(tuple) {
1993 opts.search.explain_missing_system(wall.no_headers(&tuple.to_canonical_string()));
1994 }
1995 }
1996 // And whether the tree somebody named is the release they asked for, which is the one
1997 // question left once the directories are settled and the only place both halves of it are
1998 // known. Only for a named tree, because that is the case where the release in the target
1999 // stops deciding anything, and `crate::glibc` is where the rest of the reasoning is.
2000 if sysroot.is_some() {
2001 notes.extend(glibc::skew(opts.target, pinned, &system));
2002 }
2003 for dir in system {
2004 opts.search.push_system(dir);
2005 }
2006 }
2007 // Once, here, rather than as each directory is pushed. A `-I` that names a system
2008 // directory has to lose to the system entry and the system entry is added last, so the
2009 // question cannot be answered until the whole path is known.
2010 opts.search.remove_duplicates();
2011
2012 // The target has to be resolved before the configuration is printed, so this check comes
2013 // after the loop rather than at the point `--print-config` was seen.
2014 if print_config {
2015 return Ok(Action::PrintConfig(Box::new(opts)));
2016 }
2017 if print_pipeline {
2018 return Ok(Action::PrintPipeline(Box::new(opts)));
2019 }
2020 let plan = Plan::new(&opts, &inputs, output.as_deref()).map_err(|e| err(e.message))?;
2021 if print_plan {
2022 return Ok(Action::PrintPlan {
2023 opts: Box::new(opts),
2024 plan: Box::new(plan),
2025 link: Box::new(link),
2026 });
2027 }
2028 Ok(Action::Compile {
2029 opts: Box::new(opts),
2030 plan: Box::new(plan),
2031 link: Box::new(link),
2032 jobs,
2033 verbose,
2034 notes,
2035 })
2036}
2037
2038/// What `--fetch <tuple>` asked for, or why it is not a thing that can be done.
2039///
2040/// The lookup happens here rather than at the point the bytes would move, so that a target this
2041/// release pins nothing for is a refusal from the parser and the only code that runs a downloader is
2042/// code that already knows what it is getting.
2043///
2044/// # Errors
2045///
2046/// [`CliError`] when `--offline` forbade it, when there are input files as well, when the tuple is
2047/// not a target this compiler knows, when its sysroot is behind one of section 13.4's licence walls,
2048/// and when this release pins no artifact for it.
2049fn fetch_action(named: &str, offline: bool, inputs: &[Input]) -> Result<Action, CliError> {
2050 // Not a precedence question. Section 13.2 says `--offline` forbids a fetch entirely, so a
2051 // command line that writes both has asked for two opposite things and the answer is to say so
2052 // rather than to pick one of them.
2053 if offline {
2054 return Err(err(
2055 "--fetch asks for a download and --offline forbids every download, so this command \
2056 line asks for two opposite things. Drop one of them: --offline is how a build says it \
2057 will not reach the network, and --fetch is one of the two things in this compiler \
2058 that reaches it",
2059 ));
2060 }
2061 if let Some(first) = inputs.first() {
2062 return Err(err(format!(
2063 "--fetch gets a sysroot and compiles nothing, so `{}` on the same command line is an \
2064 input that nothing would read",
2065 first.path
2066 )));
2067 }
2068 let target: TargetTuple = named
2069 .parse()
2070 .map_err(|why| err(format!("--fetch {named}: {why}, so there is no sysroot to get")))?;
2071 // The canonical spelling, because that is what a row is named by and what the directory under
2072 // the cache is called, and a person is free to write a tuple the long way round.
2073 let tuple = target.to_canonical_string();
2074 // Before the table is consulted, because a target behind a licence wall is not a row that has not
2075 // been written yet. Section 13.4 is that no release pins one of these ever, so the message says
2076 // the licence and the two lawful ways rather than naming the producer that will publish the rest.
2077 if let Some(wall) = rucc_sysroot::Wall::of(target) {
2078 return Err(err(format!("--fetch {tuple}: {}", wall.no_fetch(&tuple))));
2079 }
2080 let Some(what) = rucc_sysroot::pinned_for(&tuple) else {
2081 return Err(err(unpinned(&tuple)));
2082 };
2083 Ok(Action::Fetch { what, target, cache: cache::dir() })
2084}
2085
2086/// What `--fetch-msvc-sdk <tuple>` asks for, weighed the same way the fetch above is.
2087///
2088/// The target is resolved here rather than where the work happens, so that a tuple this compiler
2089/// does not know and a target that is not behind Microsoft's wall are refusals from the parser like
2090/// every other thing a command line can ask for and not have. Whether the licence was accepted is
2091/// carried rather than acted on, because what it changes is what the command does and not whether
2092/// the command line made sense.
2093///
2094/// # Errors
2095///
2096/// [`CliError`] when `--offline` forbade it, when there are input files as well, and when the tuple
2097/// is not a target this compiler knows.
2098fn fetch_msvc_action(
2099 named: &str,
2100 offline: bool,
2101 accepted: bool,
2102 inputs: &[Input],
2103) -> Result<Action, CliError> {
2104 if offline {
2105 return Err(err(
2106 "--fetch-msvc-sdk asks for a download and --offline forbids every download, so this \
2107 command line asks for two opposite things. Drop one of them: --offline is how a build \
2108 says it will not reach the network",
2109 ));
2110 }
2111 if let Some(first) = inputs.first() {
2112 return Err(err(format!(
2113 "--fetch-msvc-sdk gets an SDK and compiles nothing, so `{}` on the same command line \
2114 is an input that nothing would read",
2115 first.path
2116 )));
2117 }
2118 let target: TargetTuple = named.parse().map_err(|why| {
2119 err(format!("--fetch-msvc-sdk {named}: {why}, so there is no SDK to get"))
2120 })?;
2121 Ok(Action::FetchMsvcSdk { target, accepted, cache: cache::dir() })
2122}
2123
2124/// Why there is nothing to fetch for a target, which is a different sentence when the table is
2125/// empty.
2126///
2127/// A release that pins nothing and a release that pins eleven targets and not this one are two
2128/// situations, and a message that did not tell them apart would send somebody looking for a typo in
2129/// their tuple when the answer is that this work is not finished.
2130fn unpinned(tuple: &str) -> String {
2131 let pinned = rucc_sysroot::pinned_targets();
2132 if pinned.is_empty() {
2133 return format!(
2134 "this release pins no sysroot for {tuple}, and it pins none for any target yet. A \
2135 sysroot is built and published by the producer in tamnd/rucc-cross, per \
2136 spec/cross-compile/13-distribution.md section 13.8, and a release of this compiler \
2137 names one by URL and by hash afterwards. Until then, pass --sysroot=<dir> to compile \
2138 against a tree you have already"
2139 );
2140 }
2141 format!(
2142 "this release pins no sysroot for {tuple}. What it pins is {}. Pass --sysroot=<dir> to \
2143 compile against a tree you have already",
2144 pinned.join(", ")
2145 )
2146}
2147
2148/// Gets the artifact and installs it, saying what each step did.
2149///
2150/// The steps are section 13.8's and so are the messages: the transport is somebody else's program
2151/// and the check is ours, so a person reading this wants to know which downloader ran, that the
2152/// bytes matched, how many files the record named and where the tree ended up. A fetch of something
2153/// that is already there says that instead and moves nothing.
2154fn fetch_sysroot(what: &rucc_sysroot::Pinned, target: TargetTuple, cache: &std::path::Path) -> i32 {
2155 let tuple = target.to_canonical_string();
2156 let archive = what.archive_in(cache);
2157 let say = |line: &str| println!("rucc: {tuple}: {line}");
2158 match fetch::fetch(what.url, what.sha256, &archive) {
2159 Ok(fetch::Fetched::AlreadyThere) => {
2160 say(&format!("{} is already here and matches the hash", archive.display()));
2161 }
2162 Ok(fetch::Fetched::Downloaded(by)) => {
2163 say(&format!("downloaded {} with {}", what.url, by.program()));
2164 }
2165 Err(why) => return complain(why),
2166 }
2167 match install::install(&archive, what.sha256, target, cache) {
2168 Ok(done) => {
2169 match &done.before {
2170 install::Before::Nothing => {
2171 say(&format!("{} files installed at {}", done.files, done.root.display()));
2172 }
2173 install::Before::TheSame => {
2174 say(&format!(
2175 "the same sysroot is already at {}, so nothing moved",
2176 done.root.display()
2177 ));
2178 }
2179 install::Before::Different(was) => {
2180 say(&format!(
2181 "{} files installed at {}, over a tree whose record digested to {was}",
2182 done.files,
2183 done.root.display()
2184 ));
2185 }
2186 }
2187 say(&format!("the record digests to {}", done.digest));
2188 0
2189 }
2190 Err(why) => complain(why),
2191 }
2192}
2193
2194/// What one of the `-dump` and `-print` flags prints.
2195///
2196/// GCC prints the name back unchanged when it cannot find the file a `-print` flag asked about,
2197/// which is what makes the answer safe to paste into a link line whether or not the file is
2198/// there, and this does the same.
2199fn answer(query: &Query, opts: &Options, link: &LinkOptions) -> Result<String, CliError> {
2200 let found = |name: &str| {
2201 link::find_in_search(link, opts.target, name)
2202 .map_or_else(|| name.to_owned(), |path| path.display().to_string())
2203 };
2204 Ok(match query {
2205 Query::Machine => opts.target.to_string(),
2206 Query::Version => VERSION.to_owned(),
2207 Query::Multiarch => link::multiarch(opts.target),
2208 // The three lines GCC prints, in its order and with its punctuation, because what reads
2209 // them is a script written against that shape. There is no installation directory to
2210 // report: this compiler is one binary that works wherever it is copied, and the headers
2211 // it ships are inside it, so `install` is where the binary is and nothing is under it.
2212 Query::SearchDirs => {
2213 let here = std::env::current_exe()
2214 .ok()
2215 .and_then(|p| p.parent().map(std::path::Path::to_path_buf))
2216 .unwrap_or_default();
2217 let list = |dirs: &[PathBuf]| {
2218 dirs.iter().map(|d| d.display().to_string()).collect::<Vec<_>>().join(":")
2219 };
2220 let libraries = link::search_dirs(link, opts.target);
2221 format!(
2222 "install: {}\nprograms: ={}\nlibraries: ={}",
2223 here.display(),
2224 list(&link.prefixes),
2225 list(&libraries)
2226 )
2227 }
2228 // The root the rest of the answers are under, which a build system asks for when it wants
2229 // to find a file itself rather than ask for one by name, and which is the first thing to
2230 // look at when a cross build read a header nobody expected. A native compile has no
2231 // sysroot and the answer is the empty line, which is what GCC prints when it was
2232 // configured without one. `--sysroot` wins over ours because it wins everywhere else.
2233 Query::Sysroot => {
2234 sysroot_root(opts, link).map(|root| root.display().to_string()).unwrap_or_default()
2235 }
2236 // Section 13.5 of `spec/cross-compile/13-distribution.md`: for every input that is not this
2237 // compiler's own code, what it is, where it was got, its hash, its licence and whether it
2238 // was bundled, generated or fetched. What is printed is the manifest the sysroot already
2239 // carries rather than a second format saying the same things, because the three uses 13.5
2240 // gives for this are a licence notice, a reproducibility check and a security audit, and all
2241 // three are somebody else parsing it. One format is one parser to write.
2242 // Read and rendered rather than copied out, so that what comes back is the format this
2243 // build understands. The last newline comes off because whatever prints an answer adds
2244 // one, the way it does for every other query here. Keeping it would put a blank line at
2245 // the end of the one answer that is a file somebody diffs against the file it came from.
2246 Query::SysrootProvenance => match sysroot_manifest(opts, link)? {
2247 Some(manifest) => manifest.render().trim_end_matches('\n').to_string(),
2248 None => String::new(),
2249 },
2250 // Section 13.2 of the same document, which asks for the hash of a cache directory's
2251 // contents in the directory's name. A name cannot carry one, because the path has to be
2252 // computable before anything has been read, by the producer about to write the files and by
2253 // the compiler about to read them, and neither has the contents when it asks. So the number
2254 // is here instead, and it is the sha256 of the record rather than of a walk of the tree,
2255 // which means `sha256sum` over the manifest answers the same thing.
2256 Query::SysrootDigest => match sysroot_manifest(opts, link)? {
2257 Some(manifest) => manifest.digest(),
2258 None => String::new(),
2259 },
2260 Query::FileName(name) => found(name),
2261 // The name GCC gives the library of routines a compiler's output calls that the C
2262 // library does not have. Ours is built in and there is no file, so the answer is the
2263 // name itself, which is what GCC prints when it cannot find one either.
2264 Query::Libgcc => found("libgcc.a"),
2265 // A program rather than a library: the linker and the archiver are the ones a build asks
2266 // about, and this compiler finds them on the path or under `-B` rather than shipping
2267 // them, so the name back is the honest answer unless a `-B` prefix holds one.
2268 Query::ProgName(name) => link
2269 .prefixes
2270 .iter()
2271 .map(|dir| dir.join(name))
2272 .find(|path| path.is_file())
2273 .map_or_else(|| name.clone(), |path| path.display().to_string()),
2274 })
2275}
2276
2277/// The root every sysroot answer is about.
2278///
2279/// One function rather than a copy in each, because the other flags exist to say what is inside the
2280/// tree this one names, and two answers that disagreed about which tree that is would be a
2281/// difference nobody would think to look for. `--sysroot` wins over ours because it wins everywhere
2282/// else.
2283fn sysroot_root(opts: &Options, link: &LinkOptions) -> Option<PathBuf> {
2284 link.sysroot
2285 .clone()
2286 .or_else(|| link::cross_sysroot(opts.target, link).map(|at| at.root().to_path_buf()))
2287}
2288
2289/// The record of the sysroot this command line reads, when there is one to read.
2290///
2291/// [`None`] covers two cases that both print nothing, and they are different things. A compile for
2292/// this machine has no sysroot at all, and a tree somebody laid out themselves and pointed
2293/// `--sysroot` at carries no manifest, so nothing here knows where any of it came from. Saying
2294/// nothing is the only honest answer to either, and a reader can tell it from a manifest with no
2295/// inputs in it because that one still has its header lines.
2296///
2297/// # Errors
2298///
2299/// A manifest this build cannot parse, and anything else that went wrong reading the file. Passing a
2300/// record we could not read on to whoever asked would make their parser the one that finds the
2301/// problem, and every use section 13.5 gives for these two flags is somebody else reading the
2302/// output.
2303fn sysroot_manifest(opts: &Options, link: &LinkOptions) -> Result<Option<Manifest>, CliError> {
2304 let Some(root) = sysroot_root(opts, link) else {
2305 return Ok(None);
2306 };
2307 let path = Sysroot::at(root, opts.target.tuple()).manifest_path();
2308 match std::fs::read_to_string(&path) {
2309 Ok(text) => Manifest::parse(&text)
2310 .map(Some)
2311 .map_err(|why| err(format!("{}: {why}", path.display()))),
2312 Err(why) if why.kind() == std::io::ErrorKind::NotFound => Ok(None),
2313 Err(why) => Err(err(format!("{}: {why}", path.display()))),
2314 }
2315}
2316
2317/// Renders the passes this level will run, in order, with what each one does.
2318///
2319/// The level is the whole of the answer unless a `-f` flag edited it, which is section 9.1 of
2320/// `spec/09-optimizer.md`: a level is a list somebody wrote down rather than something that
2321/// emerges from which flags happen to be set, and this is how that list is read.
2322#[must_use]
2323pub fn print_pipeline(opts: &Options) -> String {
2324 let mut settings = rucc_opt::Options::for_level(opts.opt_level);
2325 settings.toggles.clone_from(&opts.passes);
2326 settings.global_fuel = opts.pass_fuel_global;
2327 for (on, spec) in &opts.pass_gates {
2328 // Every spelling was checked while the arguments were parsed, so there is nothing here
2329 // this can refuse, and a listing is not the place to report it if there were.
2330 let _ = settings.gates.add(*on, spec);
2331 }
2332 rucc_opt::pipeline::print(&settings)
2333}
2334
2335/// Renders the resolved configuration.
2336///
2337/// One `key: value` per line, sorted by nothing in particular but fixed in order, because
2338/// this output is diffed across hosts in CI and a reordering would read as a change.
2339#[must_use]
2340pub fn print_config(opts: &Options) -> String {
2341 let sess = Session::new(opts.clone());
2342 let t = &sess.target;
2343 let mut out = String::new();
2344 let _ = writeln!(out, "version: {VERSION}");
2345 // The three field triple the driver was given rather than the ten field tuple it widens to,
2346 // because this output is what a build system reads to find out what it asked for. The tuple is
2347 // the compiler's model of the machine and this line is a receipt for a command line.
2348 let _ = writeln!(out, "target: {}", opts.target);
2349 let _ = writeln!(out, "arch: {}", opts.target.arch.as_str());
2350 let _ = writeln!(out, "os: {}", opts.target.os.as_str());
2351 let _ = writeln!(out, "env: {}", opts.target.env.as_str());
2352 let _ = writeln!(out, "object-format: {}", t.object_format.as_str());
2353 let _ = writeln!(out, "pointer-width: {}", t.pointer_width);
2354 let _ = writeln!(out, "long-width: {}", t.long_width);
2355 let _ = writeln!(out, "long-double-width: {}", t.long_double_width);
2356 let _ = writeln!(out, "endian: {}", if t.little_endian { "little" } else { "big" });
2357 let _ = writeln!(out, "char-signed: {}", t.char_is_signed);
2358 let _ = writeln!(out, "va-list: {}", t.va_list.map_or("none", |list| list.as_str()));
2359 // The register file as a count per class, which is enough to tell a target whose registers
2360 // are described from one whose are not without printing sixteen names nobody asked for.
2361 let regs: Vec<String> = t
2362 .regs
2363 .classes()
2364 .map(|(class, info)| format!("{} {}", info.name, t.regs.len(class)))
2365 .collect();
2366 let _ = writeln!(
2367 out,
2368 "registers: {}",
2369 if regs.is_empty() { "none".to_string() } else { regs.join(", ") }
2370 );
2371 // What the schedule was chosen with, which is a sentence rather than a name on purpose: two
2372 // runs of a benchmark that disagree are usually two models and not two compilers.
2373 let _ = writeln!(out, "timing-model: {}", t.timing.map_or("none", |timing| timing.model));
2374 let _ = writeln!(out, "opt-level: {}", sess.opts.opt_level);
2375 let _ = writeln!(out, "safety: {}", sess.opts.safety);
2376 let _ = writeln!(out, "emit: {}", sess.opts.emit.as_str());
2377 let _ = writeln!(out, "debug-info: {}", sess.opts.debug_info);
2378 let _ = writeln!(out, "frame-pointer: {}", sess.opts.frame_pointer);
2379 let _ = writeln!(out, "red-zone: {}", sess.opts.red_zone);
2380 let _ = writeln!(out, "stack-protector: {}", sess.opts.protector);
2381 let _ = writeln!(out, "stack-clash-protection: {}", sess.opts.stack_clash);
2382 let _ = writeln!(out, "cf-protection: {}", sess.opts.control);
2383 let _ = writeln!(out, "patchable-function-entry: {}", sess.opts.patchable);
2384 let _ = writeln!(out, "profile: {}", sess.opts.profile);
2385 let _ = writeln!(out, "profile-hook: {}", sess.opts.hook);
2386 // Last because it is the one key with more than one line under it, and the only one
2387 // whose value is a property of the machine rather than of the command line.
2388 for dir in sess.opts.search.dirs() {
2389 let system = if dir.is_system { " (system)" } else { "" };
2390 let _ = writeln!(out, "include: {}{system}", dir.path.display());
2391 }
2392 out
2393}
2394
2395/// The output name the make target is taken from, which is the `-o` argument or nothing.
2396///
2397/// A run that stops at the preprocessor has not named an object, whatever its `-o` says: under
2398/// `-E` that argument is the preprocessed text and under `-M` it is the rule itself, and neither
2399/// is a file `make` would rebuild by running this rule. GCC agrees and falls back to the source
2400/// name in both, which is why a `-MD -E -o out.i` writes `out.d` holding a rule for `a.o`. From
2401/// `-S` on the argument does name what the rule builds, and it is used as written.
2402fn deps_target_output<'a>(opts: &Options, plan: &'a Plan) -> Option<&'a str> {
2403 if opts.emit == EmitKind::Preprocessed { None } else { plan.output.as_deref() }
2404}
2405
2406/// Writes to a path the command line named rather than one the plan derived, where `-` is
2407/// standard output.
2408fn write_named(path: &str, bytes: &[u8]) -> Result<(), String> {
2409 if path == "-" {
2410 return write_out(&Output::Stdout, bytes);
2411 }
2412 write_out(&Output::File(path.to_owned()), bytes)
2413}
2414
2415/// Writes the make rule for one input, and reports whether it got there.
2416///
2417/// A rule with no file of its own goes where the compilation it replaced would have written,
2418/// which is what makes the usual makefile recipe work: `rucc -M $< -o $@` leaves the rule in
2419/// `$@`, and the same line with the `-o` left off puts it on standard output.
2420fn write_deps(
2421 opts: &Options,
2422 plan: &Plan,
2423 job: &Job,
2424 found: &[Dependency],
2425 stderr: &mut impl std::io::Write,
2426) -> bool {
2427 let targets = if opts.deps.targets.is_empty() {
2428 vec![deps::default_target(&job.input, deps_target_output(opts, plan))]
2429 } else {
2430 opts.deps.targets.clone()
2431 };
2432 let rule = deps::rule(&opts.deps, &targets, &job.input, found);
2433 // The file, on the other hand, is named after the `-o` in every mode that still has one to
2434 // spend, which is every mode except the two that spend it on the rule.
2435 let wrote = match deps::default_file(&opts.deps, &job.input, plan.output.as_deref()) {
2436 // A `-MF` on a run that had nowhere else to put the rule leaves the file the `-o`
2437 // named empty rather than absent, because a makefile that named it as a target of its
2438 // own is a makefile that will look for it.
2439 Some(path) => write_named(&path, rule.as_bytes()).and_then(|()| {
2440 if opts.deps.instead_of_compiling { write_out(&job.output, b"") } else { Ok(()) }
2441 }),
2442 None => write_out(&job.output, rule.as_bytes()),
2443 };
2444 if let Err(e) = wrote {
2445 let _ = writeln!(stderr, "rucc: error: {e}");
2446 return false;
2447 }
2448 true
2449}
2450
2451/// Runs phase 4 over every input that has one, and writes what came out.
2452///
2453/// One input that fails does not stop the others. A build that reports every file it could
2454/// not preprocess in one run is worth more than one that stops at the first, and the exit
2455/// status is still a failure either way.
2456fn preprocess_all(opts: &Options, plan: &Plan) -> i32 {
2457 let fs = OsFileSystem::new();
2458 let mut stderr = std::io::stderr().lock();
2459 let mut failed = false;
2460 for job in &plan.jobs {
2461 if !job.phases.first().is_some_and(|p| *p == Phase::Preprocess) {
2462 // An input that is already preprocessed, or an object file. GCC passes these
2463 // through untouched, and the plan has already said so in its notes.
2464 continue;
2465 }
2466 let started = std::time::Instant::now();
2467 let result = preprocess(opts, &job.input, &fs);
2468 if opts.time {
2469 say_time(&job.input, started.elapsed(), &mut stderr);
2470 }
2471 for message in &result.messages {
2472 let _ = writeln!(stderr, "{message}");
2473 }
2474 if result.failed() {
2475 failed = true;
2476 continue;
2477 }
2478 if opts.deps.emit {
2479 failed |= !write_deps(opts, plan, job, &result.deps, &mut stderr);
2480 // `-M` and `-MM` asked for the rule instead of the text, so there is nothing else
2481 // to write. The other two asked for both and fall through to the text below.
2482 if opts.deps.instead_of_compiling {
2483 continue;
2484 }
2485 }
2486 if let Err(e) = write_out(&job.output, result.text.as_bytes()) {
2487 let _ = writeln!(stderr, "rucc: error: {e}");
2488 failed = true;
2489 }
2490 }
2491 i32::from(failed)
2492}
2493
2494/// Whether this job is a file of assembly that has to be assembled and that nothing here assembles.
2495///
2496/// The phases rather than the kind, because there are two kinds of assembly input and one of them
2497/// is preprocessed first, and because an object file also has no compile phase and is not this: it
2498/// has no phases at all and goes to the linker as it is. A `.s` on a `-c` line has exactly
2499/// [`Phase::Assemble`] left, and a `.S` has the preprocessor in front of it, and neither has
2500/// anything the front end can do.
2501fn needs_an_assembler(job: &Job) -> bool {
2502 job.phases.contains(&Phase::Assemble) && !job.phases.contains(&Phase::Compile)
2503}
2504
2505/// Whether the preprocessor runs over it on the way in, which is the whole difference between the
2506/// two kinds of assembly input.
2507fn assembly_wants_cpp(job: &Job) -> bool {
2508 job.phases.contains(&Phase::Preprocess)
2509}
2510
2511/// Runs the front end over every input that has a compile phase, and writes what came out.
2512///
2513/// The same rule as [`preprocess_all`]: one input that fails does not stop the others, and the
2514/// exit status is a failure either way. An input that is already assembly or an object has no
2515/// compile phase and is passed over here, which the plan has already said in its notes.
2516fn compile_all(opts: &Options, plan: &Plan) -> i32 {
2517 let fs = OsFileSystem::new();
2518 let mut stderr = std::io::stderr().lock();
2519 let mut failed = false;
2520 let (mut remarks, ok) = Remarks::new(opts.opt_info_file.as_ref(), &mut stderr);
2521 failed |= !ok;
2522 let mut fired = Fired::new();
2523 let mut pressure = Pressure::new();
2524 let mut lowerings = Lowerings::new();
2525 for job in &plan.jobs {
2526 if !job.phases.contains(&Phase::Compile) && !needs_an_assembler(job) {
2527 continue;
2528 }
2529 // An input of IR is read back rather than compiled, since the C it came from is not
2530 // here any more. A file of assembly does not go through the front end at all and is
2531 // read by the assembler instead. Everything after this is the same for all three, so
2532 // the paths meet again at the messages and the file the result is written to.
2533 let started = std::time::Instant::now();
2534 let result = if needs_an_assembler(job) {
2535 assemble(opts, &job.input, assembly_wants_cpp(job), &fs)
2536 } else if job.kind == InputKind::Ir {
2537 compile_ir(opts, &job.input, &fs)
2538 } else {
2539 compile(opts, &job.input, &fs)
2540 };
2541 if opts.time {
2542 say_time(&job.input, started.elapsed(), &mut stderr);
2543 }
2544 fired.merge(&result.fired);
2545 pressure.merge(&result.pressure);
2546 lowerings.merge(&result.lowerings);
2547 failed |= !write_dumps(&job.input, &result.dumps, &mut stderr);
2548 failed |= !remarks.write(&result.remarks, &mut stderr);
2549 for message in &result.messages {
2550 let _ = writeln!(stderr, "{message}");
2551 }
2552 // Before the failure below, because a compilation that stopped in the back end is exactly
2553 // the one whose preprocessed source somebody wants to look at.
2554 failed |= !write_temps(job, &result.temps, &mut stderr);
2555 if result.failed() {
2556 failed = true;
2557 continue;
2558 }
2559 // `-MD` and `-MMD` write the rule beside the object and let the compilation happen, so
2560 // this is the one path where both files come out of the same run. An input of IR has no
2561 // dependencies to report and produces an empty list, which produces a rule naming only
2562 // itself, and that is the honest answer rather than a missing file.
2563 if opts.deps.emit {
2564 failed |= !write_deps(opts, plan, job, &result.deps, &mut stderr);
2565 }
2566 if let Err(e) = write_out(&job.output, result.artifact.bytes()) {
2567 let _ = writeln!(stderr, "rucc: error: {e}");
2568 failed = true;
2569 }
2570 }
2571 failed |= !write_coverage(opts, &fired, &mut stderr);
2572 failed |= !write_pressure(opts, &pressure, &mut stderr);
2573 failed |= !write_lowering(opts, &lowerings, &mut stderr);
2574 i32::from(failed)
2575}
2576
2577/// A directory for the object files only the link step ever sees, removed when it goes away.
2578///
2579/// `-c` writes its object where the user can see it and linking does not, which is the whole of
2580/// the difference: a `rucc a.c b.c` leaves an executable behind and nothing else, the same as
2581/// every other compiler. Removing them on drop rather than at the end of a function is so that a
2582/// link that failed leaves nothing behind either.
2583struct Scratch {
2584 /// Where the objects go.
2585 dir: PathBuf,
2586}
2587
2588impl Scratch {
2589 /// Makes one, under whatever the platform calls its temporary directory.
2590 ///
2591 /// The name carries the process id so that two compilers running at once do not share a
2592 /// directory, which they would otherwise do the moment two of them compiled a file of the
2593 /// same name.
2594 fn new() -> Result<Scratch, String> {
2595 let dir = std::env::temp_dir().join(format!("rucc-{}", std::process::id()));
2596 std::fs::create_dir_all(&dir).map_err(|e| format!("{}: {e}", dir.display()))?;
2597 Ok(Scratch { dir })
2598 }
2599}
2600
2601impl Drop for Scratch {
2602 fn drop(&mut self) {
2603 let _ = std::fs::remove_dir_all(&self.dir);
2604 }
2605}
2606
2607/// The link line the plan describes, for `-###`.
2608///
2609/// The names in it are the hints the plan carries rather than the temporaries a real compilation
2610/// would choose, because `-###` prints the line without having compiled anything and so has
2611/// nothing to point at. That also makes the printed line readable rather than naming a directory
2612/// that only exists while a compilation is running.
2613fn link_line(opts: &Options, link: &LinkOptions, job: &LinkJob) -> Result<String, link::Error> {
2614 let linker = link::find(opts.target, link)?;
2615 let args = link::line(opts.target, link, &job.inputs, &job.output)?;
2616 Ok(link::render(&linker, &args))
2617}
2618
2619/// Compiles everything, then links it.
2620///
2621/// The objects go in a directory that is removed afterwards, which is why this is not
2622/// [`compile_all`] followed by a link: the plan says an object feeding the linker is temporary
2623/// and does not say where, because where is a question that only has an answer once something is
2624/// running.
2625fn link_all(opts: &Options, plan: &Plan, link: &LinkOptions, verbose: bool) -> i32 {
2626 let Some(job) = &plan.link else {
2627 // Every path into here comes from a plan whose last phase is the link, and such a plan
2628 // has a link job. Saying so is cheaper than an unwrap that would have to be explained.
2629 let mut stderr = std::io::stderr().lock();
2630 let _ = writeln!(stderr, "rucc: error: there is nothing to link");
2631 return 1;
2632 };
2633 // Before anything is compiled, because a linker that is not on the machine is worth knowing
2634 // about in the second it takes to look rather than after the compilation.
2635 // And before that, whether this link has a line at all and whether what it reads is on the
2636 // machine. Both are answerable now, and a target whose sysroot has not been built is worth
2637 // saying so about before the compilation rather than after it.
2638 if let Err(why) = link::preflight(opts.target, link) {
2639 return complain(why);
2640 }
2641 let linker = match link::find(opts.target, link) {
2642 Ok(linker) => linker,
2643 Err(why) => return complain(why),
2644 };
2645 // And whether the one that was found can do this link, which for one linker and one target is
2646 // a question only the linker itself can answer. Here rather than inside the search, because
2647 // what it does is refuse rather than move on to the next candidate: nothing else in the list
2648 // links a produced Windows sysroot either.
2649 if let Err(why) = link::suitable(opts.target, &linker) {
2650 return complain(why);
2651 }
2652
2653 let scratch = match Scratch::new() {
2654 Ok(scratch) => scratch,
2655 Err(why) => return complain(format!("could not make a place for the object files: {why}")),
2656 };
2657
2658 let fs = OsFileSystem::new();
2659 let mut failed = false;
2660 // One per job, in job order, which is what lets the link line below be rebuilt with the real
2661 // paths in it: every job contributes exactly one file to the line and does so in this order.
2662 let mut produced: Vec<String> = Vec::with_capacity(plan.jobs.len());
2663 let mut fired = Fired::new();
2664 let mut pressure = Pressure::new();
2665 let mut lowerings = Lowerings::new();
2666 {
2667 let mut stderr = std::io::stderr().lock();
2668 let (mut remarks, ok) = Remarks::new(opts.opt_info_file.as_ref(), &mut stderr);
2669 failed |= !ok;
2670 for (at, job) in plan.jobs.iter().enumerate() {
2671 let out = match &job.output {
2672 Output::Temporary(hint) => {
2673 // The index because two inputs in different directories can have the same
2674 // name, and the two objects of `rucc a/x.c b/x.c` must not be one file.
2675 scratch.dir.join(format!("{at}-{hint}")).display().to_string()
2676 }
2677 Output::File(path) => path.clone(),
2678 // A job feeding the linker never writes to standard output, since the plan gives
2679 // it a temporary. This is here so that the match is total rather than a panic.
2680 Output::Stdout => continue,
2681 };
2682 produced.push(out.clone());
2683 if !job.phases.contains(&Phase::Compile) && !needs_an_assembler(job) {
2684 continue;
2685 }
2686 let started = std::time::Instant::now();
2687 let result = if needs_an_assembler(job) {
2688 assemble(opts, &job.input, assembly_wants_cpp(job), &fs)
2689 } else if job.kind == InputKind::Ir {
2690 compile_ir(opts, &job.input, &fs)
2691 } else {
2692 compile(opts, &job.input, &fs)
2693 };
2694 if opts.time {
2695 say_time(&job.input, started.elapsed(), &mut stderr);
2696 }
2697 fired.merge(&result.fired);
2698 pressure.merge(&result.pressure);
2699 lowerings.merge(&result.lowerings);
2700 failed |= !write_dumps(&job.input, &result.dumps, &mut stderr);
2701 failed |= !remarks.write(&result.remarks, &mut stderr);
2702 for message in &result.messages {
2703 let _ = writeln!(stderr, "{message}");
2704 }
2705 failed |= !write_temps(job, &result.temps, &mut stderr);
2706 if result.failed() {
2707 failed = true;
2708 continue;
2709 }
2710 // A `-MD` on a command line that links writes the rule next to the executable and
2711 // names the executable as its target, since that is the file this source builds
2712 // here. The object it went through is in a temporary directory and is gone by the
2713 // time `make` reads any of this.
2714 if opts.deps.emit {
2715 failed |= !write_deps(opts, plan, job, &result.deps, &mut stderr);
2716 }
2717 if !matches!(result.artifact, Artifact::Object { .. }) {
2718 // Worth saying rather than writing whatever it is and letting the linker read it.
2719 // An empty file is a valid empty linker script, so a link handed one gets as far
2720 // as reporting every symbol of this file undefined, which is a page of messages
2721 // about something that went wrong here.
2722 let _ = writeln!(
2723 stderr,
2724 "rucc: internal error: {}: no object file was produced for the link",
2725 job.input
2726 );
2727 failed = true;
2728 continue;
2729 }
2730 if let Err(e) = std::fs::write(&out, result.artifact.bytes()) {
2731 let _ = writeln!(stderr, "rucc: error: {out}: {e}");
2732 failed = true;
2733 }
2734 }
2735 failed |= !write_coverage(opts, &fired, &mut stderr);
2736 failed |= !write_pressure(opts, &pressure, &mut stderr);
2737 failed |= !write_lowering(opts, &lowerings, &mut stderr);
2738 failed |= !write_lowering(opts, &lowerings, &mut stderr);
2739 }
2740 if failed {
2741 // Nothing is linked from a compilation that did not finish. A linker run over the objects
2742 // that did compile would report every function of the file that did not as undefined,
2743 // which is a page of messages about a mistake already reported once.
2744 return 1;
2745 }
2746
2747 // The items in command line order with the temporaries filled in. A library and a word for the
2748 // linker contribute no job and pass through, and every file item takes the next job's real
2749 // output, which is what keeps whatever was written between two objects between them here.
2750 let mut outputs = produced.into_iter();
2751 let mut items = Vec::with_capacity(job.inputs.len());
2752 for item in &job.inputs {
2753 match item {
2754 link::Item::Library(name) => items.push(link::Item::Library(name.clone())),
2755 link::Item::Linker(arg) => items.push(link::Item::Linker(arg.clone())),
2756 link::Item::File(_) => match outputs.next() {
2757 Some(path) => items.push(link::Item::File(path)),
2758 None => return complain("the plan asks the linker for a file nothing produced"),
2759 },
2760 }
2761 }
2762
2763 let args = match link::line(opts.target, link, &items, &job.output) {
2764 Ok(args) => args,
2765 Err(why) => return complain(why),
2766 };
2767 if verbose {
2768 let mut stderr = std::io::stderr().lock();
2769 let _ = writeln!(stderr, "{}", link::render(&linker, &args));
2770 }
2771 let started = std::time::Instant::now();
2772 let ran = link::run(&linker, &args);
2773 if opts.time {
2774 // The one step of a compilation that really is another program, so this line is the same
2775 // measurement gcc's is and names the linker the way gcc names `collect2`.
2776 let mut stderr = std::io::stderr().lock();
2777 say_time(&linker.name, started.elapsed(), &mut stderr);
2778 }
2779 match ran {
2780 Ok(()) => 0,
2781 // The linker has already said what was wrong on its own error output, and repeating that
2782 // linking failed would only push its message further up the screen.
2783 Err(link::Error::Refused { .. }) => 1,
2784 Err(why) => complain(why),
2785 }
2786}
2787
2788/// Compiles everything and writes the objects into one static library.
2789///
2790/// No temporary directory and no second program. The objects never reach the file system at all:
2791/// they go from the compiler into the archive writer, which is both faster than writing a directory
2792/// of files for an `ar` to read back and the reason the symbol index can be written at all. A
2793/// member's index entries are the names the object writer says it wrote, and the only thing that
2794/// knows those is the run that wrote it.
2795///
2796/// `-save-temps` is the exception. It asked for the objects to be kept, the plan gave them names a
2797/// person can find, and they are written there as well as put in the archive.
2798fn archive_all(opts: &Options, plan: &Plan) -> i32 {
2799 let Some(job) = &plan.archive else {
2800 // Every path into here comes from a plan whose last phase is the archive, and such a plan
2801 // has an archive job. Saying so is cheaper than an unwrap that would have to be explained.
2802 return complain("there is nothing to put in an archive");
2803 };
2804 // Before anything is compiled, because a format this has no container for is worth knowing
2805 // about in the second it takes to look rather than after the whole compilation.
2806 let flavour = match opts.target.os.object_format() {
2807 ObjectFormat::Elf => rucc_archive::Flavour::Gnu,
2808 ObjectFormat::Coff => rucc_archive::Flavour::Coff,
2809 // Mach-O wants the BSD flavour, whose index is a different member under a different name,
2810 // and wasm has no archives of its own at all. Neither has an object writer either, so a
2811 // command line reaching this would have failed in the next step regardless.
2812 format @ (ObjectFormat::MachO | ObjectFormat::Wasm) => {
2813 return complain(format!(
2814 "there is no archive format for {} objects in this compiler yet",
2815 format.as_str()
2816 ));
2817 }
2818 };
2819
2820 let fs = OsFileSystem::new();
2821 let mut failed = false;
2822 let mut members: Vec<rucc_archive::Member> = Vec::with_capacity(plan.jobs.len());
2823 let mut names = job.members.iter();
2824 let mut fired = Fired::new();
2825 let mut pressure = Pressure::new();
2826 let mut lowerings = Lowerings::new();
2827 {
2828 let mut stderr = std::io::stderr().lock();
2829 let (mut remarks, ok) = Remarks::new(opts.opt_info_file.as_ref(), &mut stderr);
2830 failed |= !ok;
2831 for plan_job in &plan.jobs {
2832 // What the plan called this member. The two lists are walked together rather than the
2833 // name being worked out again here, so that what `-###` printed and what goes in the
2834 // file cannot come apart.
2835 let Some(member) = names.next() else {
2836 return complain("the plan asks the archive for a member nothing produced");
2837 };
2838 if !plan_job.phases.contains(&Phase::Compile) && !needs_an_assembler(plan_job) {
2839 // Neither something to compile nor something to assemble, so there is nothing to
2840 // put in, and an archive quietly missing a member is worse than a message.
2841 let _ = writeln!(
2842 &mut stderr,
2843 "rucc: error: {}: this compiler makes an archive out of what it compiles, and \
2844 there is nothing here for it to do",
2845 plan_job.input
2846 );
2847 failed = true;
2848 continue;
2849 }
2850 let started = std::time::Instant::now();
2851 let result = if needs_an_assembler(plan_job) {
2852 assemble(opts, &plan_job.input, assembly_wants_cpp(plan_job), &fs)
2853 } else if plan_job.kind == InputKind::Ir {
2854 compile_ir(opts, &plan_job.input, &fs)
2855 } else {
2856 compile(opts, &plan_job.input, &fs)
2857 };
2858 if opts.time {
2859 say_time(&plan_job.input, started.elapsed(), &mut stderr);
2860 }
2861 fired.merge(&result.fired);
2862 pressure.merge(&result.pressure);
2863 lowerings.merge(&result.lowerings);
2864 failed |= !write_dumps(&plan_job.input, &result.dumps, &mut stderr);
2865 failed |= !remarks.write(&result.remarks, &mut stderr);
2866 for message in &result.messages {
2867 let _ = writeln!(stderr, "{message}");
2868 }
2869 failed |= !write_temps(plan_job, &result.temps, &mut stderr);
2870 if result.failed() {
2871 failed = true;
2872 continue;
2873 }
2874 if opts.deps.emit {
2875 failed |= !write_deps(opts, plan, plan_job, &result.deps, &mut stderr);
2876 }
2877 let Artifact::Object { bytes, defines } = result.artifact else {
2878 let _ = writeln!(
2879 stderr,
2880 "rucc: internal error: {}: no object file was produced for the archive",
2881 plan_job.input
2882 );
2883 failed = true;
2884 continue;
2885 };
2886 // Under `-save-temps` the plan gave the object a name a person can find, so it is
2887 // written there too. Otherwise it is only ever a member and never a file.
2888 if let Output::File(path) = &plan_job.output {
2889 if let Err(e) = std::fs::write(path, &bytes) {
2890 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2891 failed = true;
2892 }
2893 }
2894 members.push(rucc_archive::Member { name: member.clone(), body: bytes, defines });
2895 }
2896 failed |= !write_coverage(opts, &fired, &mut stderr);
2897 failed |= !write_pressure(opts, &pressure, &mut stderr);
2898 failed |= !write_lowering(opts, &lowerings, &mut stderr);
2899 failed |= !write_lowering(opts, &lowerings, &mut stderr);
2900 }
2901 if failed {
2902 // Nothing is written from a compilation that did not finish, for the reason the link gives:
2903 // an archive missing the file that failed is one a link reports every name of as undefined,
2904 // which is a page of messages about a mistake already reported once.
2905 return 1;
2906 }
2907
2908 let bytes = match rucc_archive::write(flavour, &members) {
2909 Ok(bytes) => bytes,
2910 // Every one of these is a bug here rather than a program's mistake: the names came from the
2911 // object writer and the bodies came from this process.
2912 Err(why) => return complain(format!("the archive could not be written: {why}")),
2913 };
2914 match std::fs::write(&job.output, &bytes) {
2915 Ok(()) => 0,
2916 Err(e) => complain(format!("{}: {e}", job.output)),
2917 }
2918}
2919
2920/// Prints one driver level message and gives back the exit status that goes with it.
2921fn complain(why: impl std::fmt::Display) -> i32 {
2922 let mut stderr = std::io::stderr().lock();
2923 let _ = writeln!(stderr, "rucc: error: {why}");
2924 1
2925}
2926
2927/// Writes what `-Zrule-coverage=FILE` asked for, and says whether it could.
2928///
2929/// Once for the whole command line rather than once per input, because the question is which
2930/// lowering rules this run of the compiler reached and a file per input would leave the reader
2931/// unioning files to find out something one process already knew.
2932///
2933/// A file that could not be written is a failure and not a warning. What asks for this is a
2934/// measurement run, and a measurement that quietly did not happen is worse than one that stopped.
2935fn write_coverage(opts: &Options, fired: &Fired, stderr: &mut impl std::io::Write) -> bool {
2936 let Some(path) = &opts.rule_coverage else { return true };
2937 let Some(table) = coverage::table(opts.target.arch) else {
2938 let _ = writeln!(
2939 stderr,
2940 "rucc: error: there are no lowering rules for {} yet, so there is no coverage of them \
2941 to report",
2942 opts.target
2943 );
2944 return false;
2945 };
2946 match std::fs::write(path, fired.listing(table)) {
2947 Ok(()) => true,
2948 Err(e) => {
2949 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2950 false
2951 }
2952 }
2953}
2954
2955/// Writes what `-Zregister-pressure=FILE` asked for, and says whether it could.
2956///
2957/// Once for the whole command line, for the reason [`write_coverage`] gives, and a file that could
2958/// not be written is a failure for the reason it gives too. There is no equivalent of the missing
2959/// rule table here, since every target this compiles for has an allocator, and a run that reached
2960/// no back end at all writes an empty listing rather than nothing: a measurement of a build that
2961/// produced no code is still an answer and it is the honest one.
2962fn write_pressure(opts: &Options, pressure: &Pressure, stderr: &mut impl std::io::Write) -> bool {
2963 let Some(path) = &opts.register_pressure else { return true };
2964 match std::fs::write(path, pressure.listing()) {
2965 Ok(()) => true,
2966 Err(e) => {
2967 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2968 false
2969 }
2970 }
2971}
2972
2973/// Writes what `-Zlowering=FILE` asked for, and says whether it could.
2974///
2975/// Once for the whole command line, for the reason [`write_coverage`] gives, and a file that could
2976/// not be written is a failure for the reason it gives too. A run that reached no back end writes
2977/// an empty listing rather than nothing, the way [`write_pressure`] does and for the same reason.
2978fn write_lowering(opts: &Options, lowerings: &Lowerings, stderr: &mut impl std::io::Write) -> bool {
2979 let Some(path) = &opts.lowering_dump else { return true };
2980 match std::fs::write(path, lowerings.listing()) {
2981 Ok(()) => true,
2982 Err(e) => {
2983 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2984 false
2985 }
2986 }
2987}
2988
2989/// Where the `-fopt-info` remarks go, and how much of the run has already gone there.
2990///
2991/// Standard error by default, and one file for the whole run when `-fopt-info=<file>` named one.
2992/// A file rather than the diagnostic stream is what a harness wants: the corpus in
2993/// `tamnd/rucc-corpus` matches a rejection against what the compiler said on standard error, and
2994/// a few thousand remarks mixed into that would bury it.
2995struct Remarks {
2996 /// The file, if there is one.
2997 file: Option<String>,
2998 /// Whether anything has been written to it yet, which decides between truncating and
2999 /// appending. One file holds the whole run rather than the last input in it.
3000 started: bool,
3001}
3002
3003impl Remarks {
3004 /// Prepares the destination, emptying the file if there is one.
3005 ///
3006 /// Emptied here rather than at the first remark, because a run where no pass had anything to
3007 /// say should leave an empty file and not yesterday's. An absent file and an empty one are
3008 /// different facts and something reading this will act on the difference.
3009 fn new(file: Option<&String>, stderr: &mut impl std::io::Write) -> (Self, bool) {
3010 let mut ok = true;
3011 if let Some(path) = file {
3012 if let Err(e) = std::fs::write(path, "") {
3013 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
3014 ok = false;
3015 }
3016 }
3017 (Self { file: file.cloned(), started: false }, ok)
3018 }
3019
3020 /// Writes one input's remarks, and says whether that worked.
3021 ///
3022 /// A file that cannot be written is a failure and not a warning, for the reason
3023 /// [`write_dumps`] gives: remarks that quietly did not arrive look exactly like a compilation
3024 /// where nothing happened.
3025 fn write(&mut self, text: &str, stderr: &mut impl std::io::Write) -> bool {
3026 if text.is_empty() {
3027 return true;
3028 }
3029 let Some(path) = &self.file else {
3030 let _ = write!(stderr, "{text}");
3031 return true;
3032 };
3033 let opened = std::fs::OpenOptions::new()
3034 .write(true)
3035 .append(self.started)
3036 .truncate(!self.started)
3037 .create(true)
3038 .open(path);
3039 self.started = true;
3040 let result =
3041 opened.and_then(|mut file| std::io::Write::write_all(&mut file, text.as_bytes()));
3042 if let Err(e) = result {
3043 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
3044 return false;
3045 }
3046 true
3047 }
3048}
3049
3050/// Writes what `-fdump-ir=` asked to see, one file per dump.
3051///
3052/// The name is the input file with the dump's own name and `.ir` after it, so a directory listing
3053/// after a run is the passes in the order they ran, per input. They go in the working directory
3054/// rather than beside the output, because a dump is something a person asked for at a prompt and
3055/// the working directory is where that person is.
3056///
3057/// A file that could not be written is a failure and not a warning, for the reason
3058/// [`write_coverage`] gives: what asked for this is somebody debugging a pass, and a dump that
3059/// quietly did not happen looks exactly like a pass that did not run.
3060fn write_dumps(input: &str, dumps: &[rucc_opt::Dump], stderr: &mut impl std::io::Write) -> bool {
3061 let stem = std::path::Path::new(input)
3062 .file_name()
3063 .map_or_else(|| input.to_owned(), |name| name.to_string_lossy().into_owned());
3064 let mut ok = true;
3065 for dump in dumps {
3066 let path = format!("{stem}.{}.ir", dump.name);
3067 if let Err(e) = std::fs::write(&path, &dump.text) {
3068 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
3069 ok = false;
3070 }
3071 }
3072 ok
3073}
3074
3075/// Writes the files `-save-temps` kept, which is nothing at all unless it was given.
3076///
3077/// A file that could not be written is a failure rather than a warning, for the reason
3078/// [`write_dumps`] gives: somebody asked for these by name, and one that quietly did not happen
3079/// looks like a compilation that never went through that step.
3080fn write_temps(job: &Job, temps: &Temps, stderr: &mut impl std::io::Write) -> bool {
3081 let mut ok = true;
3082 let kept = [(job.saved_text(), &temps.preprocessed), (job.saved_asm(), &temps.assembly)];
3083 for (path, text) in kept {
3084 // A step the compilation did not reach has nothing to keep, and a job that is not keeping
3085 // that step has nowhere to put it. Either way there is no file here.
3086 let (Some(path), Some(text)) = (path, text) else { continue };
3087 if let Err(e) = std::fs::write(&path, text) {
3088 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
3089 ok = false;
3090 }
3091 }
3092 ok
3093}
3094
3095/// One line of `-time`, which is what a step was called and how long it took.
3096///
3097/// GCC's two numbers are the user and the system time of a subprocess it ran. This compiler runs
3098/// no subprocess for anything but the link, so what is measured here is the wall clock of the
3099/// step and the second column is always zero. The shape of the line is kept because a person
3100/// reading it next to gcc's should not have to work out which column is which.
3101fn say_time(name: &str, took: std::time::Duration, stderr: &mut impl std::io::Write) {
3102 let _ = writeln!(stderr, "# {name} {:.2} {:.2}", took.as_secs_f64(), 0.0);
3103}
3104
3105/// Writes one job's result where the plan said it goes.
3106///
3107/// # Errors
3108///
3109/// Returns the message to print, which names the file when there is one, because "permission
3110/// denied" on its own does not say which file was refused.
3111fn write_out(output: &Output, bytes: &[u8]) -> Result<(), String> {
3112 match output {
3113 Output::Stdout => {
3114 let mut stdout = std::io::stdout().lock();
3115 stdout.write_all(bytes).map_err(|e| format!("writing to standard output: {e}"))
3116 }
3117 Output::File(path) | Output::Temporary(path) => {
3118 std::fs::write(path, bytes).map_err(|e| format!("{path}: {e}"))
3119 }
3120 }
3121}
3122
3123/// Runs the driver and returns the process exit code.
3124///
3125/// `args` excludes the program name. Output goes to `stdout` and errors to `stderr`, which
3126/// is the one place in the compiler that is true.
3127pub fn run(args: &[String]) -> i32 {
3128 match parse_args(args) {
3129 Ok(Action::Help) => {
3130 print!("{USAGE}");
3131 0
3132 }
3133 Ok(Action::Version) => {
3134 println!("rucc {VERSION}");
3135 0
3136 }
3137 Ok(Action::Print(line)) => {
3138 println!("{line}");
3139 0
3140 }
3141 Ok(Action::PrintConfig(opts)) => {
3142 print!("{}", print_config(&opts));
3143 0
3144 }
3145 Ok(Action::PrintPipeline(opts)) => {
3146 print!("{}", print_pipeline(&opts));
3147 0
3148 }
3149 Ok(Action::PrintPlan { opts, plan, link }) => {
3150 print!("{}", plan.render());
3151 // The line as it would be typed, which is the half of `-###` that section 4.3 says
3152 // arrives with the link. It is printed even when the linker is not on this machine,
3153 // because what a build wants from `-###` is what the compiler would do.
3154 if let Some(job) = &plan.link {
3155 match link_line(&opts, &link, job) {
3156 Ok(line) => println!("{line}"),
3157 Err(why) => {
3158 let mut stderr = std::io::stderr().lock();
3159 let _ = writeln!(stderr, "rucc: error: {why}");
3160 return 1;
3161 }
3162 }
3163 }
3164 0
3165 }
3166 Ok(Action::Fetch { what, target, cache }) => fetch_sysroot(what, target, &cache),
3167 Ok(Action::FetchMsvcSdk { target, accepted, cache }) => {
3168 msvc::fetch_msvc_sdk(target, accepted, &cache)
3169 }
3170 Ok(Action::Compile { opts, plan, link, jobs, verbose, notes }) => {
3171 {
3172 let mut stderr = std::io::stderr().lock();
3173 // Before the plan rather than after it, because a note is about the command line
3174 // and the plan is what the command line was read as, so the reader wants the two
3175 // in that order.
3176 for note in ¬es {
3177 let _ = writeln!(stderr, "rucc: warning: {note}");
3178 }
3179 if verbose {
3180 let _ = write!(stderr, "{}", plan.render());
3181 let _ = writeln!(stderr, "workers: {}", jobs.count());
3182 }
3183 }
3184 if opts.emit == EmitKind::Preprocessed {
3185 return preprocess_all(&opts, &plan);
3186 }
3187 if opts.emit == EmitKind::Archive {
3188 return archive_all(&opts, &plan);
3189 }
3190 if opts.emit != EmitKind::Executable {
3191 return compile_all(&opts, &plan);
3192 }
3193 link_all(&opts, &plan, &link, verbose)
3194 }
3195 Err(e) => {
3196 let mut stderr = std::io::stderr().lock();
3197 let _ = writeln!(stderr, "rucc: error: {e}");
3198 let _ = writeln!(stderr, "rucc: note: run `rucc --help` for usage");
3199 1
3200 }
3201 }
3202}
3203
3204#[cfg(test)]
3205mod tests {
3206 use rucc_session::{
3207 Contract, GnucVersion, IncludeForm, LtoJobs, OptLevel, Partition, Patchable, Visibility,
3208 };
3209
3210 use super::*;
3211
3212 fn args(s: &[&str]) -> Vec<String> {
3213 s.iter().map(|x| (*x).to_owned()).collect()
3214 }
3215
3216 /// A target to write down where the host would otherwise decide, for the tests whose answer
3217 /// would be a different one on a different machine.
3218 ///
3219 /// Most of the tests here never name a target, which is right, because most of what the driver
3220 /// does with a command line is the same wherever it runs and a test that pinned one would be
3221 /// saying so in every case for the sake of the two that need it. The two that need it are the
3222 /// ones whose answer comes off the target rather than off the command line: the name an object
3223 /// gets, which is `a.o` here and `a.obj` on Windows, and whether Microsoft's reading of a
3224 /// nameless member is on, which is off here and on there. Both are the compiler being right, and
3225 /// a test that leaves the target to the host is asking a question with two correct answers.
3226 const LINUX: &str = "--target=x86_64-unknown-linux-gnu";
3227
3228 #[test]
3229 fn help_and_version_win_over_everything_else() {
3230 assert_eq!(parse_args(&args(&["-c", "--help", "x.c"])).unwrap(), Action::Help);
3231 assert_eq!(parse_args(&args(&["--version"])).unwrap(), Action::Version);
3232 }
3233
3234 fn compile(s: &[&str]) -> (Box<Options>, Box<Plan>) {
3235 match parse_args(&args(s)).expect("expected a compilation") {
3236 Action::Compile { opts, plan, .. } => (opts, plan),
3237 other => panic!("expected a compilation, got {other:?}"),
3238 }
3239 }
3240
3241 fn linking(s: &[&str]) -> (Box<LinkOptions>, Box<Plan>) {
3242 match parse_args(&args(s)).expect("expected a compilation") {
3243 Action::Compile { link, plan, .. } => (link, plan),
3244 other => panic!("expected a compilation, got {other:?}"),
3245 }
3246 }
3247
3248 fn notes(s: &[&str]) -> Vec<String> {
3249 match parse_args(&args(s)).expect("expected a compilation") {
3250 Action::Compile { notes, .. } => notes,
3251 other => panic!("expected a compilation, got {other:?}"),
3252 }
3253 }
3254
3255 /// The ordinary command line has nothing to say about itself, which is the property that makes
3256 /// a note worth reading when there is one.
3257 #[test]
3258 fn a_command_line_with_nothing_wrong_with_it_carries_no_notes() {
3259 assert_eq!(notes(&["-c", "a.c"]), Vec::<String>::new());
3260 }
3261
3262 /// A directory that is not there contributes nothing to the search path, so there is no tree to
3263 /// read a release out of and nothing to compare the pin against. Said as a test because this is
3264 /// the shape a hermetic machine takes: the probe reads the disk and every other machine has a
3265 /// different disk, so what can be asserted here is the silence.
3266 #[test]
3267 fn a_named_tree_that_is_not_on_the_machine_is_not_a_release_mismatch() {
3268 let said =
3269 notes(&["--target=x86_64-linux-gnu.2.28", "--sysroot=/nowhere-at-all", "-c", "a.c"]);
3270 assert_eq!(said, Vec::<String>::new());
3271 }
3272
3273 #[test]
3274 fn collects_inputs_and_flags() {
3275 let (opts, plan) = compile(&["-c", "-O2", "-g", "a.c", "b.c"]);
3276 let paths: Vec<&str> = plan.jobs.iter().map(|j| j.input.as_str()).collect();
3277 assert_eq!(paths, vec!["a.c", "b.c"]);
3278 assert_eq!(opts.opt_level, OptLevel::O2);
3279 assert_eq!(opts.emit, EmitKind::Object);
3280 assert!(opts.debug_info);
3281 }
3282
3283 /// The unstable options, which are spelled apart from everything else on purpose: what is
3284 /// under `-Z` promises nothing, and a build that reaches for one should have had to say so.
3285 #[test]
3286 fn an_unstable_option_is_taken_and_one_that_does_not_exist_is_refused() {
3287 let (opts, _) = compile(&["-c", "-Zrule-coverage=/tmp/rules.cov", "a.c"]);
3288 assert_eq!(opts.rule_coverage.as_deref(), Some("/tmp/rules.cov"));
3289
3290 let (plain, _) = compile(&["-c", "a.c"]);
3291 assert_eq!(plain.rule_coverage, None, "nothing is measured unless it was asked for");
3292
3293 assert!(parse_args(&args(&["-Zrule-coverage=", "a.c"])).is_err(), "a file with no name");
3294 let unknown = parse_args(&args(&["-Zwhat", "a.c"])).expect_err("there is no such option");
3295 assert!(unknown.message.contains("4.11"), "{}", unknown.message);
3296 }
3297
3298 /// The other measurement written to a file, which reads the same way and fails the same way.
3299 #[test]
3300 fn where_the_register_pressure_goes_is_asked_for_the_same_way() {
3301 let (opts, _) = compile(&["-c", "-O2", "-Zregister-pressure=/tmp/spills.txt", "a.c"]);
3302 assert_eq!(opts.register_pressure.as_deref(), Some("/tmp/spills.txt"));
3303
3304 let (plain, _) = compile(&["-c", "a.c"]);
3305 assert_eq!(plain.register_pressure, None, "nothing is measured unless it was asked for");
3306
3307 assert!(parse_args(&args(&["-Zregister-pressure=", "a.c"])).is_err(), "no file named");
3308 }
3309
3310 /// The third one, which says what the pre-selection lowering group did.
3311 #[test]
3312 fn where_the_lowering_dump_goes_is_asked_for_the_same_way() {
3313 let (opts, _) = compile(&["-c", "-O2", "-Zlowering=/tmp/lowering.txt", "a.c"]);
3314 assert_eq!(opts.lowering_dump.as_deref(), Some("/tmp/lowering.txt"));
3315
3316 let (plain, _) = compile(&["-c", "a.c"]);
3317 assert_eq!(plain.lowering_dump, None, "nothing is dumped unless it was asked for");
3318
3319 assert!(parse_args(&args(&["-Zlowering=", "a.c"])).is_err(), "no file named");
3320 }
3321
3322 /// Scheduling, which has the three way answer every optimization flag has: on, off, and
3323 /// nothing said, which is whatever the optimization level asks for. The name is gcc's, and
3324 /// gcc's has a two in it because gcc has a scheduler before allocation and one after and this
3325 /// is the one after.
3326 #[test]
3327 fn scheduling_can_be_turned_on_and_off_and_left_to_the_optimization_level() {
3328 let (on, _) = compile(&["-c", "-O0", "-fschedule-insns2", "a.c"]);
3329 assert_eq!(on.schedule_insns, Some(true));
3330
3331 let (off, _) = compile(&["-c", "-O2", "-fno-schedule-insns2", "a.c"]);
3332 assert_eq!(off.schedule_insns, Some(false));
3333
3334 let (quiet, _) = compile(&["-c", "-O2", "a.c"]);
3335 assert_eq!(quiet.schedule_insns, None, "nothing said, so the level decides");
3336 assert!(quiet.opt_level.schedules(), "and at this level the level says yes");
3337
3338 let (none, _) = compile(&["-c", "a.c"]);
3339 assert!(!none.opt_level.schedules(), "at no optimization it says no");
3340 }
3341
3342 /// Whether the timing model is worth holding an instruction back over, which is a `-Z` because
3343 /// it is a question about a target's description rather than about the program being compiled.
3344 #[test]
3345 fn whether_the_timing_model_is_cycle_accurate_can_be_overridden() {
3346 let (yes, _) = compile(&["-c", "-O2", "-Zcycle-accurate-model=yes", "a.c"]);
3347 assert_eq!(yes.cycle_accurate_model, Some(true));
3348
3349 let (no, _) = compile(&["-c", "-O2", "-Zcycle-accurate-model=no", "a.c"]);
3350 assert_eq!(no.cycle_accurate_model, Some(false));
3351
3352 let (plain, _) = compile(&["-c", "-O2", "a.c"]);
3353 assert_eq!(plain.cycle_accurate_model, None, "the target's own answer stands");
3354
3355 let bad = parse_args(&args(&["-Zcycle-accurate-model=maybe", "a.c"]))
3356 .expect_err("it takes yes or no");
3357 assert!(bad.message.contains("yes or no"), "{}", bad.message);
3358 }
3359
3360 #[test]
3361 fn a_bare_dash_o_means_o1_the_way_gcc_reads_it() {
3362 let (opts, _) = compile(&["-O", "a.c"]);
3363 assert_eq!(opts.opt_level, OptLevel::O1);
3364 }
3365
3366 #[test]
3367 fn dash_x_applies_to_later_inputs_only_and_none_stops_it() {
3368 let (_, plan) = compile(&["a.o", "-x", "c", "b.txt", "-x", "none", "c.o"]);
3369 assert_eq!(plan.jobs[0].kind, InputKind::LinkerInput);
3370 assert_eq!(plan.jobs[1].kind, InputKind::C);
3371 assert_eq!(plan.jobs[2].kind, InputKind::LinkerInput);
3372 }
3373
3374 #[test]
3375 fn dash_j_reaches_the_scheduler_and_defaults_to_the_machine() {
3376 let (_, _, jobs) = match parse_args(&args(&["-j4", "a.c"])).unwrap() {
3377 Action::Compile { opts, plan, jobs, .. } => (opts, plan, jobs),
3378 other => panic!("expected a compilation, got {other:?}"),
3379 };
3380 assert_eq!(jobs.count(), 4);
3381
3382 let default = match parse_args(&args(&["a.c"])).unwrap() {
3383 Action::Compile { jobs, .. } => jobs,
3384 other => panic!("expected a compilation, got {other:?}"),
3385 };
3386 assert_eq!(default, Jobs::available());
3387 assert!(parse_args(&args(&["-j0", "a.c"])).is_err());
3388 }
3389
3390 #[test]
3391 fn triple_hash_prints_the_plan_and_runs_nothing() {
3392 let a = parse_args(&args(&["-###", "-c", "a.c"])).unwrap();
3393 let Action::PrintPlan { plan, .. } = a else { panic!("expected a plan dump") };
3394 assert!(plan.render().contains("a.c: preprocess, compile, assemble -> a.o"));
3395 }
3396
3397 #[test]
3398 fn the_flag_that_keeps_the_intermediate_files_has_three_spellings_and_two_meanings() {
3399 // The bare one is `=obj` and not `=cwd`. gcc's manual says the opposite and gcc 16 does
3400 // this, and following the compiler is what makes a build that reads either of them find
3401 // the files where they are.
3402 assert_eq!(compile(&["-c", "-save-temps", "a.c"]).0.save_temps, SaveTemps::Object);
3403 assert_eq!(compile(&["-c", "-save-temps=obj", "a.c"]).0.save_temps, SaveTemps::Object);
3404 assert_eq!(compile(&["-c", "-save-temps=cwd", "a.c"]).0.save_temps, SaveTemps::Cwd);
3405 assert_eq!(compile(&["-c", "a.c"]).0.save_temps, SaveTemps::No);
3406 // The last one on the line decides, the way it does for every other flag with an
3407 // argument, and a keyword that is neither is fatal rather than ignored: a run that kept
3408 // nothing and said nothing looks exactly like one where the files were not produced.
3409 let (opts, _) = compile(&["-c", "-save-temps", "-save-temps=cwd", "a.c"]);
3410 assert_eq!(opts.save_temps, SaveTemps::Cwd);
3411 let e = parse_args(&args(&["-c", "-save-temps=nowhere", "a.c"])).unwrap_err();
3412 assert!(e.message.contains("accepted: cwd, obj"), "{}", e.message);
3413 }
3414
3415 #[test]
3416 fn the_flag_that_times_each_step_reaches_the_options_and_changes_nothing_else() {
3417 let (opts, plan) = compile(&["-c", "-time", "a.c"]);
3418 let (plain, without) = compile(&["-c", "a.c"]);
3419 assert!(opts.time);
3420 assert!(!plain.time);
3421 // Against the same line without the flag rather than against a spelling of the object's
3422 // name, since what the object is called is the host's business and this is not about that.
3423 assert_eq!(plan.jobs[0].output, without.jobs[0].output);
3424 }
3425
3426 #[test]
3427 fn dash_x_names_what_it_accepts_when_it_does_not_know_a_language() {
3428 let e = parse_args(&args(&["-x", "fortran", "a.c"])).unwrap_err();
3429 assert!(e.message.contains("assembler-with-cpp"), "{}", e.message);
3430 }
3431
3432 /// What `--fetch` says for a target this release pins nothing for, which is every target except
3433 /// the three windows-gnu ones today.
3434 #[test]
3435 fn a_fetch_of_a_target_nothing_is_pinned_for_says_so_rather_than_reaching_the_network() {
3436 let e = parse_args(&args(&["--fetch", "x86_64-linux-musl"])).unwrap_err();
3437 assert!(e.message.contains("pins no sysroot for x86_64-linux-musl"), "{}", e.message);
3438 // And what it does pin, because a release with some rows in the table and a release with
3439 // none are two situations and the second sentence is what tells them apart.
3440 assert!(e.message.contains("x86_64-windows-gnu"), "{}", e.message);
3441 // The joined spelling is the same flag.
3442 let joined = parse_args(&args(&["--fetch=x86_64-linux-musl"])).unwrap_err();
3443 assert_eq!(joined, e);
3444 }
3445
3446 /// The two targets a release will never pin, which is a different answer from the one above.
3447 ///
3448 /// Section 13.4. A person who reads "this release pins no sysroot yet" waits for a release that
3449 /// does, and no release of this compiler can ship either of these, so the message names the
3450 /// licence that decides it and what to do instead.
3451 #[test]
3452 fn a_fetch_of_a_target_behind_a_licence_wall_says_so_rather_than_saying_not_yet() {
3453 let e = parse_args(&args(&["--fetch", "aarch64-macos"])).unwrap_err();
3454 assert!(e.message.contains("Xcode licence"), "{}", e.message);
3455 assert!(e.message.contains("there never will be"), "{}", e.message);
3456 assert!(!e.message.contains("tamnd/rucc-cross"), "{}", e.message);
3457
3458 let e = parse_args(&args(&["--fetch", "x86_64-windows-msvc"])).unwrap_err();
3459 assert!(e.message.contains("redistributed"), "{}", e.message);
3460 // The way out of this one is a target rather than a download, and it is the default already.
3461 assert!(e.message.contains("mingw-w64"), "{}", e.message);
3462 // And the mingw-w64 target next to it is ours to ship and published, so the same flag has
3463 // something to get rather than a licence to explain.
3464 let action = parse_args(&args(&["--fetch", "x86_64-windows-gnu"])).expect("it is pinned");
3465 let Action::Fetch { what, .. } = action else { panic!("{action:?}") };
3466 assert_eq!(what.tuple, "x86_64-windows-gnu");
3467 }
3468
3469 #[test]
3470 fn the_other_fetch_takes_a_target_behind_microsofts_wall_and_carries_the_acceptance() {
3471 // Both spellings of the flag, because a flag that takes a tuple gets written both ways.
3472 for line in [
3473 vec!["--fetch-msvc-sdk", "x86_64-windows-msvc"],
3474 vec!["--fetch-msvc-sdk=x86_64-windows-msvc"],
3475 ] {
3476 let action = parse_args(&args(&line)).expect("that is a target behind the wall");
3477 let Action::FetchMsvcSdk { target, accepted, .. } = action else {
3478 panic!("{action:?}")
3479 };
3480 assert_eq!(target.to_canonical_string(), "x86_64-windows-msvc");
3481 // Nothing on the line accepted anything, so nothing did.
3482 assert!(!accepted);
3483 }
3484
3485 // And both spellings of the word, because the prose here uses one and most of the people
3486 // typing this will reach for the other.
3487 for word in ["--accept-licence", "--accept-license"] {
3488 let action = parse_args(&args(&["--fetch-msvc-sdk", "aarch64-windows-msvc", word]))
3489 .expect("that is a target behind the wall");
3490 let Action::FetchMsvcSdk { target, accepted, .. } = action else {
3491 panic!("{action:?}")
3492 };
3493 assert_eq!(target.to_canonical_string(), "aarch64-windows-msvc");
3494 assert!(accepted, "{word} should have been read");
3495 }
3496 }
3497
3498 #[test]
3499 fn the_other_fetch_refuses_the_command_lines_that_do_not_mean_anything() {
3500 // A tuple is what it gets, so a flag with nothing after it is not a command.
3501 let e = parse_args(&args(&["--fetch-msvc-sdk"])).unwrap_err();
3502 assert!(e.message.contains("requires the target"), "{}", e.message);
3503 let e = parse_args(&args(&["--fetch-msvc-sdk", "not-a-target"])).unwrap_err();
3504 assert!(e.message.contains("there is no SDK to get"), "{}", e.message);
3505
3506 // `--offline` forbids every download and this one asks for one, whichever order they came
3507 // in, which is the same answer `--fetch` gives.
3508 for line in [
3509 vec!["--offline", "--fetch-msvc-sdk", "x86_64-windows-msvc"],
3510 vec!["--fetch-msvc-sdk", "x86_64-windows-msvc", "--offline"],
3511 ] {
3512 let e = parse_args(&args(&line)).unwrap_err();
3513 assert!(e.message.contains("two opposite things"), "{}", e.message);
3514 }
3515
3516 // It gets an SDK and compiles nothing, so a file on the same line would be read by nothing.
3517 let e = parse_args(&args(&["--fetch-msvc-sdk", "x86_64-windows-msvc", "a.c"])).unwrap_err();
3518 assert!(e.message.contains("compiles nothing"), "{}", e.message);
3519
3520 // The two fetches are two commands and a line that asked for both asked for neither.
3521 let e = parse_args(&args(&[
3522 "--fetch",
3523 "x86_64-windows-gnu",
3524 "--fetch-msvc-sdk",
3525 "x86_64-windows-msvc",
3526 ]))
3527 .unwrap_err();
3528 assert!(e.message.contains("two different commands"), "{}", e.message);
3529
3530 // And an acceptance with nothing to accept for is a command line that says something about
3531 // a licence no part of it goes near.
3532 let e = parse_args(&args(&["--accept-licence", "-c", "a.c"])).unwrap_err();
3533 assert!(e.message.contains("--fetch-msvc-sdk <tuple> is the command"), "{}", e.message);
3534 }
3535
3536 /// An Apple target on a machine with no SDK, which is section 8.6's other host.
3537 ///
3538 /// Not run on a mac, where the SDK this is about is installed and the compile is the ordinary one
3539 /// that uses it. What the reason says is asserted in `rucc_sysroot::wall` and where it is printed
3540 /// is asserted in `rucc-pp`, so what is left here is that the driver works it out and leaves it
3541 /// where the preprocessor will find it, and that neither way past the wall leaves one behind.
3542 #[test]
3543 fn an_apple_target_with_no_sdk_anywhere_carries_the_licence_rather_than_a_missing_directory() {
3544 if cfg!(target_os = "macos") || std::env::var_os("SDKROOT").is_some() {
3545 return;
3546 }
3547 let (opts, _) = compile(&["--target=aarch64-macos", "-c", "a.c"]);
3548 let why = opts.search.missing_system().expect("the wall is the reason there are none");
3549 assert!(why.contains("aarch64-macos needs a macOS SDK"), "{why}");
3550 assert!(why.contains("Xcode licence"), "{why}");
3551 assert!(why.contains("-isysroot"), "{why}");
3552
3553 // A program that includes none of the library needs none of the SDK, which is what section
3554 // 8.6 means by being able to target the platform without one, so there is nothing to explain.
3555 let (opts, _) = compile(&["--target=aarch64-macos", "-nostdinc", "-c", "a.c"]);
3556 assert_eq!(opts.search.missing_system(), None);
3557 // And naming a path is the other way through, whether or not the path is there: a mistyped
3558 // directory is a mistake to report on its own terms rather than a licence to explain.
3559 let (opts, _) = compile(&["--target=aarch64-macos", "-isysroot", "/opt/sdk", "-c", "a.c"]);
3560 assert_eq!(opts.search.missing_system(), None);
3561 }
3562
3563 /// The same wall on the compile side of an MSVC target, where the way past it is a tuple.
3564 ///
3565 /// Not run on Windows, for the same reason the one above is not run on a mac: the wall stands in
3566 /// front of an SDK this machine does not have, and a Windows machine is the kind that does. The
3567 /// driver asks `vswhere` where Visual Studio is and takes the newest kit under it, so on a box
3568 /// with the build tools installed there are headers, no wall and nothing here to be about.
3569 /// `INCLUDE` is the other way a machine has one and is the other half of the guard, since a
3570 /// person can set that anywhere while Visual Studio is only found on the platform it runs on.
3571 #[test]
3572 fn an_msvc_target_with_no_sdk_named_says_which_environment_needs_nothing_installed() {
3573 if cfg!(target_os = "windows") || std::env::var_os("INCLUDE").is_some() {
3574 return;
3575 }
3576 let (opts, _) = compile(&["--target=x86_64-windows-msvc", "-c", "a.c"]);
3577 let why = opts.search.missing_system().expect("the wall is the reason there are none");
3578 assert!(why.contains("the Windows SDK and its universal CRT"), "{why}");
3579 assert!(why.contains("mingw-w64"), "{why}");
3580 // And the mingw-w64 target has its headers from us, so nothing is missing to explain.
3581 let (opts, _) = compile(&["--target=x86_64-windows-gnu", "-c", "a.c"]);
3582 assert_eq!(opts.search.missing_system(), None);
3583 }
3584
3585 #[test]
3586 fn a_fetch_with_no_target_and_a_fetch_of_a_tuple_that_is_not_one_both_say_which() {
3587 let e = parse_args(&args(&["--fetch"])).unwrap_err();
3588 assert!(e.message.contains("--fetch requires"), "{}", e.message);
3589 let e = parse_args(&args(&["--fetch", "sparc64-solaris-gnu"])).unwrap_err();
3590 assert!(e.message.contains("--fetch sparc64-solaris-gnu"), "{}", e.message);
3591 assert!(e.message.contains("no sysroot to get"), "{}", e.message);
3592 }
3593
3594 /// Both flags on one line ask for opposite things, in either order.
3595 #[test]
3596 fn a_fetch_and_offline_together_is_a_refusal_whichever_way_round_they_are_written() {
3597 for line in [
3598 vec!["--offline", "--fetch", "x86_64-linux-musl"],
3599 vec!["--fetch", "x86_64-linux-musl", "--offline"],
3600 ] {
3601 let e = parse_args(&args(&line)).unwrap_err();
3602 assert!(e.message.contains("two opposite things"), "{}", e.message);
3603 }
3604 }
3605
3606 #[test]
3607 fn a_fetch_does_not_compile_anything_and_says_so_when_it_is_handed_a_file() {
3608 let e = parse_args(&args(&["--fetch", "x86_64-linux-musl", "a.c"])).unwrap_err();
3609 assert!(e.message.contains("compiles nothing"), "{}", e.message);
3610 assert!(e.message.contains("a.c"), "{}", e.message);
3611 }
3612
3613 /// `--offline` on its own is accepted and changes nothing, because an ordinary compile
3614 /// downloads nothing with or without it. A build that passes it everywhere is the case this is
3615 /// for, and it must not lose the compilation it was passed beside.
3616 #[test]
3617 fn offline_on_a_compilation_is_the_same_compilation() {
3618 let (opts, plan) = compile(&["-c", "--offline", "a.c"]);
3619 let (plain, without) = compile(&["-c", "a.c"]);
3620 assert_eq!(opts.target, plain.target);
3621 assert_eq!(plan.jobs.len(), without.jobs.len());
3622 assert_eq!(plan.jobs[0].output, without.jobs[0].output);
3623 }
3624
3625 #[test]
3626 fn an_unknown_flag_is_an_error_rather_than_a_shrug() {
3627 let e = parse_args(&args(&["-fno-such-thing", "a.c"])).unwrap_err();
3628 assert!(e.message.contains("unknown option"), "{}", e.message);
3629 }
3630
3631 /// `-fpermissive` and the flag that turns it back off, which a build writes beside it when
3632 /// one directory needs the older rules and the rest of the tree does not.
3633 #[test]
3634 fn permissive_reads_in_both_directions_and_the_last_one_wins() {
3635 let (opts, _) = compile(&["-c", "a.c"]);
3636 assert!(!opts.permissive, "off unless it is asked for");
3637
3638 let (opts, _) = compile(&["-c", "-fpermissive", "a.c"]);
3639 assert!(opts.permissive);
3640
3641 let (opts, _) = compile(&["-c", "-fpermissive", "-fno-permissive", "a.c"]);
3642 assert!(!opts.permissive);
3643 }
3644
3645 #[test]
3646 fn asking_for_nested_functions_is_told_why_it_is_not_coming() {
3647 let e = parse_args(&args(&["-fnested-functions", "a.c"])).unwrap_err();
3648 assert!(e.message.contains("trampoline"), "{}", e.message);
3649 assert!(parse_args(&args(&["-fno-nested-functions", "a.c"])).is_ok());
3650 }
3651
3652 #[test]
3653 fn the_flag_every_configure_script_writes_is_taken() {
3654 // All four spellings, because a build writes whichever one its macros picked and a
3655 // compiler that takes three of them is a compiler that fails on the fourth.
3656 for flag in ["-fPIC", "-fpic", "-fPIE", "-fpie"] {
3657 let (opts, _) = compile(&["-c", flag, "a.c"]);
3658 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3659 }
3660 }
3661
3662 #[test]
3663 fn a_table_is_written_unless_the_build_says_nothing_will_walk_it() {
3664 let (opts, _) = compile(&["-c", "a.c"]);
3665 assert!(opts.unwinds(), "the default is off");
3666 let (opts, _) = compile(&["-c", "-fno-asynchronous-unwind-tables", "a.c"]);
3667 assert!(!opts.unwinds(), "the build was not taken at its word");
3668 let (opts, _) = compile(&[
3669 "-c",
3670 "-fno-asynchronous-unwind-tables",
3671 "-fasynchronous-unwind-tables",
3672 "a.c",
3673 ]);
3674 assert!(opts.unwinds(), "the last flag did not win");
3675 // The weaker request, which the same table answers, so a line that asks for a table and
3676 // against an asynchronous one gets one. That is gcc's arrangement and it turns up when a
3677 // build turns the asynchronous one off globally and a directory asks for a table back.
3678 let (opts, _) =
3679 compile(&["-c", "-fno-asynchronous-unwind-tables", "-funwind-tables", "a.c"]);
3680 assert!(opts.unwinds(), "the weaker request was dropped");
3681 let (opts, _) = compile(&["-c", "-fno-unwind-tables", "a.c"]);
3682 assert!(opts.unwinds(), "the weaker negative turned off the stronger request");
3683 let (opts, _) =
3684 compile(&["-c", "-fno-unwind-tables", "-fno-asynchronous-unwind-tables", "a.c"]);
3685 assert!(!opts.unwinds(), "both were turned off and one stayed on");
3686 }
3687
3688 #[test]
3689 fn the_flags_that_describe_what_this_compiler_already_does_are_taken() {
3690 // Every one of these is on a real build line somewhere and every one of them was an
3691 // unknown option. What they have in common is that the answer rucc gives is the answer
3692 // they ask for, so there is nothing to implement and nothing to refuse.
3693 for flag in [
3694 "-fno-common",
3695 "-fstrict-aliasing",
3696 "-fno-strict-aliasing",
3697 "-fdelete-null-pointer-checks",
3698 "-fno-delete-null-pointer-checks",
3699 "-frounding-math",
3700 "-fno-rounding-math",
3701 "-fexcess-precision=standard",
3702 "-fexcess-precision=fast",
3703 "-fexcess-precision=16",
3704 "-pipe",
3705 "-fdiagnostics-color",
3706 "-fno-diagnostics-color",
3707 "-fdiagnostics-color=always",
3708 "-fdiagnostics-color=never",
3709 "-fdiagnostics-color=auto",
3710 ] {
3711 let (opts, _) = compile(&["-c", flag, "a.c"]);
3712 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3713 }
3714 }
3715
3716 #[test]
3717 fn whether_an_exception_is_looked_at_is_kept_and_defaults_to_gccs_answer() {
3718 let (opts, _) = compile(&["-c", "a.c"]);
3719 assert!(opts.trapping_math, "the default was not gcc's");
3720 let (opts, _) = compile(&["-c", "-fno-trapping-math", "a.c"]);
3721 assert!(!opts.trapping_math);
3722 let (opts, _) = compile(&["-c", "-ftrapping-math", "a.c"]);
3723 assert!(opts.trapping_math, "spelling out the default turned it off");
3724 // The last one written wins, which is how a build line that inherits a flag from one
3725 // place and overrides it in another is read.
3726 let (opts, _) = compile(&["-c", "-fno-trapping-math", "-ftrapping-math", "a.c"]);
3727 assert!(opts.trapping_math);
3728 }
3729
3730 /// The flags a torture program writes on its own `dg-options` line, which is where most of
3731 /// these come from: a program reduced from a miscompilation names the pass that miscompiled
3732 /// it. Eighteen programs in the suite stopped on the driver before anything read them, and
3733 /// tamnd/rucc#1019 is the list.
3734 #[test]
3735 fn the_flags_that_name_a_pass_of_gccs_own_are_taken_and_dropped() {
3736 for flag in [
3737 "-fno-tree-ccp",
3738 "-fno-tree-dominator-opts",
3739 "-fno-tree-vrp",
3740 "-fno-tree-bit-ccp",
3741 "-fno-tree-coalesce-vars",
3742 "-ftree-vectorize",
3743 "-ftree-loop-distribution",
3744 "-fipa-pta",
3745 "-fmodulo-sched",
3746 "-fno-vect-cost-model",
3747 "-fvect-cost-model=unlimited",
3748 "-fsimd-cost-model=cheap",
3749 "-fexpensive-optimizations",
3750 "-fno-early-inlining",
3751 "-fno-inline",
3752 "-finline-functions",
3753 "-foptimize-strlen",
3754 "-fno-ira-share-spill-slots",
3755 ] {
3756 let (opts, _) = compile(&["-c", flag, "a.c"]);
3757 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3758 assert!(opts.passes.is_empty(), "{flag} named a pass of gcc's and not one of ours");
3759 }
3760 }
3761
3762 /// The two namespaces are taken whole, so a name neither this test nor gcc 16 has heard of
3763 /// goes the same way as the ones above rather than stopping a build on the day gcc adds it.
3764 #[test]
3765 fn a_pass_name_in_either_family_is_taken_whether_or_not_it_is_one_gcc_has() {
3766 for flag in ["-ftree-no-such-pass", "-fno-ipa-no-such-pass"] {
3767 let (opts, _) = compile(&["-c", flag, "a.c"]);
3768 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3769 }
3770 }
3771
3772 /// A pass this compiler has keeps its flag, since the arms that read the registry are above
3773 /// the family arms. `dce` is the one both compilers have a name for, and `execute/pr97421-2.c`
3774 /// is the program that writes it.
3775 #[test]
3776 fn a_pass_name_this_compiler_has_is_still_read_as_a_pass() {
3777 let (opts, _) = compile(&["-c", "-fno-dce", "a.c"]);
3778 assert_eq!(opts.passes, vec![("dce".to_owned(), false)]);
3779 }
3780
3781 /// gcc's name for the unroller reaches the unroller, in both directions. libtommath puts
3782 /// `-funroll-loops` in `CFLAGS` unconditionally, and before this it was an unknown option and
3783 /// the build stopped on its first file.
3784 #[test]
3785 fn the_gcc_spelling_of_the_unroller_turns_the_unroller_on_and_off() {
3786 let (opts, _) = compile(&["-c", "-funroll-loops", "a.c"]);
3787 assert_eq!(opts.passes, vec![("unroll".to_owned(), true)]);
3788 let (opts, _) = compile(&["-c", "-fno-unroll-loops", "a.c"]);
3789 assert_eq!(opts.passes, vec![("unroll".to_owned(), false)]);
3790 }
3791
3792 /// The three transformations that are a module at a time are named by a flag as well, even
3793 /// though none of them is a `rucc_opt::Pass` and so none is reached by the generic arms.
3794 ///
3795 /// A bisection over a miscompilation turns one thing off at a time, and a transformation with
3796 /// no spelling of its own cannot be the one turned off.
3797 #[test]
3798 fn the_transformations_that_are_not_passes_are_still_named_by_a_flag() {
3799 let (opts, _) = compile(&["-c", "-fno-ipa-cp", "-fipa-sra", "-fno-libcall", "a.c"]);
3800 assert_eq!(
3801 opts.passes,
3802 vec![
3803 (rucc_opt::ipcp::NAME.to_owned(), false),
3804 (rucc_opt::ipasra::NAME.to_owned(), true),
3805 (rucc_opt::libcall::NAME.to_owned(), false),
3806 ]
3807 );
3808 let (opts, _) = compile(&["-c", "-flibcall", "a.c"]);
3809 assert_eq!(opts.passes, vec![(rucc_opt::libcall::NAME.to_owned(), true)]);
3810 }
3811
3812 /// Where a function starts is a question this compiler answers, so the flag that asks about it
3813 /// is answered rather than dropped. femtolisp's Makefile writes the bare form on every compile
3814 /// of the project, and before this it was an unknown option and the build stopped on its first
3815 /// file. The numbers are gcc 16's, read off `-S` on x86-64: nothing and the bare form both
3816 /// give `.p2align 4`, `=32` gives 5, `=3` gives 2, and the negative form gives `.align 8`.
3817 #[test]
3818 fn the_alignment_of_a_function_is_a_request_this_compiler_can_answer() {
3819 let (opts, _) = compile(&["-c", "-falign-functions", "a.c"]);
3820 assert_eq!(opts.align_functions, None, "the bare form asks for the default");
3821
3822 let (opts, _) = compile(&["-c", "-falign-functions=32", "a.c"]);
3823 assert_eq!(opts.align_functions, Some(32));
3824
3825 let (opts, _) = compile(&["-c", "-falign-functions=3", "a.c"]);
3826 assert_eq!(opts.align_functions, Some(4), "rounded up rather than refused");
3827
3828 let (opts, _) = compile(&["-c", "-falign-functions=32:8", "a.c"]);
3829 assert_eq!(opts.align_functions, Some(32), "the boundary is the answerable half");
3830
3831 for flag in ["-falign-functions=0", "-falign-functions=1"] {
3832 let (opts, _) = compile(&["-c", flag, "a.c"]);
3833 assert_eq!(opts.align_functions, None, "{flag} means the default");
3834 }
3835
3836 let (opts, _) = compile(&["-c", "-fno-align-functions", "a.c"]);
3837 assert_eq!(opts.align_functions, Some(8), "the smallest boundary the target has");
3838
3839 // The last one on the line wins, which is how gcc reads a repeated flag.
3840 let (opts, _) = compile(&["-c", "-falign-functions=32", "-falign-functions", "a.c"]);
3841 assert_eq!(opts.align_functions, None);
3842
3843 let e = parse_args(&args(&["-c", "-falign-functions=big", "a.c"])).unwrap_err();
3844 assert!(e.message.contains("number of bytes"), "{}", e.message);
3845 }
3846
3847 /// The other three of the family are about padding inside a body, which nothing here writes,
3848 /// so they are taken and say nothing. Every spelling of each, since a build writes whichever
3849 /// one its author typed.
3850 #[test]
3851 fn the_alignment_flags_about_the_inside_of_a_body_are_taken_and_say_nothing() {
3852 for flag in [
3853 "-falign-labels",
3854 "-falign-loops",
3855 "-falign-jumps",
3856 "-falign-loops=16",
3857 "-falign-labels=32",
3858 "-fno-align-loops",
3859 "-fno-align-labels",
3860 "-fno-align-jumps",
3861 ] {
3862 let (opts, _) = compile(&["-c", flag, "a.c"]);
3863 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3864 assert_eq!(opts.align_functions, None, "{flag} is not about where a function starts");
3865 }
3866 }
3867
3868 /// The encoding of the source is not a question about speed, so the one name that describes
3869 /// what the preprocessor does is taken and every other name is refused.
3870 #[test]
3871 fn the_input_charset_is_taken_when_it_names_the_one_that_is_read() {
3872 for flag in ["-finput-charset=utf-8", "-finput-charset=UTF-8", "-finput-charset=utf8"] {
3873 let (opts, _) = compile(&["-c", flag, "a.c"]);
3874 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3875 }
3876
3877 let e = parse_args(&args(&["-c", "-finput-charset=latin1", "a.c"])).unwrap_err();
3878 assert!(e.message.contains("latin1"), "{}", e.message);
3879 assert!(e.message.contains("UTF-8"), "what is read is worth saying: {}", e.message);
3880 }
3881
3882 /// The other half of the same rule. Each of these changes what the program does rather than
3883 /// how fast it does it, so each is refused with the reason, and the negative of each is what
3884 /// happens anyway and is taken.
3885 #[test]
3886 fn the_three_that_change_the_answer_are_refused_and_their_negatives_are_taken() {
3887 for (flag, word) in [
3888 ("-ffast-math", "__FAST_MATH__"),
3889 ("-fnon-call-exceptions", "landing pad"),
3890 ("-finstrument-functions", "__cyg_profile_func_enter"),
3891 ] {
3892 let e = parse_args(&args(&["-c", flag, "a.c"])).unwrap_err();
3893 assert!(e.message.contains(word), "{flag}: {}", e.message);
3894 assert!(!e.message.contains("unknown option"), "{flag} deserves a reason");
3895
3896 let off = format!("-fno-{}", flag.trim_start_matches("-f"));
3897 let (opts, _) = compile(&["-c", &off, "a.c"]);
3898 assert_eq!(opts.emit, EmitKind::Object, "{off}");
3899 }
3900 }
3901
3902 #[test]
3903 fn asking_the_linker_to_merge_tentative_definitions_is_told_why_it_is_not_coming() {
3904 // The one of that family that is a request rather than a description, and it is a real
3905 // difference: two files each writing `int g;` link under it and do not without it.
3906 let e = parse_args(&args(&["-fcommon", "a.c"])).unwrap_err();
3907 assert!(e.message.contains(".bss"), "{}", e.message);
3908 assert!(e.message.contains("extern"), "the way out is worth saying: {}", e.message);
3909 }
3910
3911 #[test]
3912 fn asking_for_position_dependent_code_is_told_why_it_is_not_coming() {
3913 for flag in ["-fno-pic", "-fno-pie"] {
3914 let e = parse_args(&args(&[flag, "a.c"])).unwrap_err();
3915 assert!(e.message.contains("global offset table"), "{flag}: {}", e.message);
3916 // The one it may have meant, since the two are a letter apart and one of them is
3917 // about linking and is taken.
3918 assert!(e.message.contains("-no-pie"), "{flag}: {}", e.message);
3919 }
3920 }
3921
3922 #[test]
3923 fn an_unsupported_target_names_itself() {
3924 let e = parse_args(&args(&["--target=sparc64-linux-gnu", "a.c"])).unwrap_err();
3925 assert!(e.message.contains("sparc64"), "{}", e.message);
3926 }
3927
3928 #[test]
3929 fn no_inputs_is_an_error_but_print_config_needs_none() {
3930 assert!(parse_args(&args(&[])).is_err());
3931 assert!(matches!(parse_args(&args(&["--print-config"])), Ok(Action::PrintConfig(_))));
3932 }
3933
3934 #[test]
3935 fn print_config_reports_the_target_it_was_given_not_the_host() {
3936 let a = parse_args(&args(&["--print-config", "--target=riscv64-linux-musl"])).unwrap();
3937 let Action::PrintConfig(opts) = a else { panic!("expected a configuration dump") };
3938 let text = print_config(&opts);
3939 assert!(text.contains("target: riscv64-unknown-linux-musl"), "{text}");
3940 assert!(text.contains("char-signed: false"), "{text}");
3941 assert!(text.contains("object-format: elf"), "{text}");
3942 assert!(text.contains("va-list: void-pointer"), "{text}");
3943 // RISC-V has a register file and this compiler has not written it down yet, and the
3944 // dump says which of those two it is rather than leaving the line out.
3945 assert!(text.contains("registers: none"), "{text}");
3946 assert!(text.contains("timing-model: none"), "{text}");
3947 }
3948
3949 /// The model the schedule was chosen with, which is a receipt anybody comparing two runs of a
3950 /// benchmark needs: two numbers that disagree are usually two models and not two compilers.
3951 #[test]
3952 fn print_config_names_the_model_the_schedule_was_chosen_with() {
3953 let opts = Options::new("x86_64-unknown-linux-gnu".parse().unwrap());
3954 let text = print_config(&opts);
3955 let line = text.lines().find(|l| l.starts_with("timing-model:")).expect("the model");
3956 assert!(line.contains("Skylake"), "{line}");
3957 assert!(line.contains("published"), "a sentence saying where it came from: {line}");
3958 }
3959
3960 #[test]
3961 fn print_config_has_one_key_per_line_and_a_fixed_order() {
3962 let opts = Options::new("x86_64-unknown-linux-gnu".parse().unwrap());
3963 let text = print_config(&opts);
3964 let keys: Vec<&str> =
3965 text.lines().map(|l| l.split(':').next().unwrap_or_default()).collect();
3966 assert_eq!(keys[0], "version");
3967 assert_eq!(keys[1], "target");
3968 assert_eq!(keys.len(), 26);
3969 assert!(text.ends_with('\n'));
3970 }
3971
3972 #[test]
3973 fn the_safety_tier_is_read_off_the_command_line_and_a_wrong_one_is_refused() {
3974 let (opts, _) = compile(&["a.c"]);
3975 assert_eq!(opts.safety, rucc_session::Safety::Off);
3976
3977 for (flag, tier) in [
3978 ("-fsafety=detect", rucc_session::Safety::Detect),
3979 ("-fsafety=enforce", rucc_session::Safety::Enforce),
3980 ("-fsafety=kernel", rucc_session::Safety::Kernel),
3981 ("-fsafety=off", rucc_session::Safety::Off),
3982 ] {
3983 let (opts, _) = compile(&[flag, "a.c"]);
3984 assert_eq!(opts.safety, tier, "{flag}");
3985 }
3986
3987 // The last one wins, the way every other repeated flag on this command line does.
3988 let (opts, _) = compile(&["-fsafety=enforce", "-fsafety=off", "a.c"]);
3989 assert_eq!(opts.safety, rucc_session::Safety::Off);
3990
3991 // A misspelled tier is refused rather than ignored. Silently compiling without the
3992 // monitor a build asked for is the one failure mode this feature cannot have.
3993 let e = parse_args(&args(&["-fsafety=on", "a.c"])).unwrap_err();
3994 assert!(e.message.contains("is not a safety tier"), "{}", e.message);
3995 assert!(parse_args(&args(&["-fsafety", "a.c"])).is_err());
3996 }
3997
3998 #[test]
3999 fn the_padding_mode_is_read_off_the_command_line_and_a_wrong_one_is_refused() {
4000 // The default is the one section 9.3 of document 09 gives library code, which is that
4001 // padding does not participate, so a record filled a member at a time is not reported.
4002 let (opts, _) = compile(&["a.c"]);
4003 assert_eq!(opts.padding, rucc_session::Padding::Ignored);
4004
4005 let (opts, _) = compile(&["-fsafety=detect", "-fsafety-init=padding", "a.c"]);
4006 assert_eq!(opts.padding, rucc_session::Padding::Tracked);
4007
4008 let (opts, _) = compile(&["-fsafety-init=padding", "-fsafety-init=nopadding", "a.c"]);
4009 assert_eq!(opts.padding, rucc_session::Padding::Ignored);
4010
4011 // The tier is still a tier. A flag whose name starts the same way must not be eaten by
4012 // the one above it, which is the thing worth pinning about a pair of names like these.
4013 let (opts, _) = compile(&["-fsafety-init=padding", "a.c"]);
4014 assert_eq!(opts.safety, rucc_session::Safety::Off);
4015
4016 let e = parse_args(&args(&["-fsafety-init=some", "a.c"])).unwrap_err();
4017 assert!(e.message.contains("is not a padding mode"), "{}", e.message);
4018 }
4019
4020 #[test]
4021 fn whether_a_write_has_to_stay_inside_its_member_is_read_off_the_command_line() {
4022 // Off by default, because a store to allocated storage sets its effective type and C 6.5
4023 // lets a program reuse a buffer as something else. Row S4 is a build opting out of that.
4024 let (opts, _) = compile(&["a.c"]);
4025 assert_eq!(opts.subobject, rucc_session::Subobject::Off);
4026
4027 let (opts, _) = compile(&["-fsafety=detect", "-fsafety-subobject", "a.c"]);
4028 assert_eq!(opts.subobject, rucc_session::Subobject::Members);
4029
4030 let (opts, _) = compile(&["-fsafety-subobject", "-fno-safety-subobject", "a.c"]);
4031 assert_eq!(opts.subobject, rucc_session::Subobject::Off);
4032
4033 // It takes no value. The form that would take one is the strict reading of section 9.4,
4034 // which is not written yet, so say so rather than accept a spelling that does nothing.
4035 let e = parse_args(&args(&["-fsafety-subobject=strict", "a.c"])).unwrap_err();
4036 assert!(e.message.contains("tamnd/rucc#967"), "{}", e.message);
4037 }
4038
4039 #[test]
4040 fn whether_two_restrict_pointers_may_meet_is_read_off_the_command_line() {
4041 // Off by default, because the record a block keeps is the union of what each pointer
4042 // reached, so two pointers striding through one array without landing on the same byte are
4043 // reported and by the letter of the standard those are different objects. Row Y8 is a build
4044 // deciding it would rather know.
4045 let (opts, _) = compile(&["a.c"]);
4046 assert_eq!(opts.promise, rucc_session::Promise::Off);
4047
4048 let (opts, _) = compile(&["-fsafety=detect", "-fsafety-restrict", "a.c"]);
4049 assert_eq!(opts.promise, rucc_session::Promise::Blocks);
4050
4051 let (opts, _) = compile(&["-fsafety-restrict", "-fno-safety-restrict", "a.c"]);
4052 assert_eq!(opts.promise, rucc_session::Promise::Off);
4053
4054 // The tier is still a tier, which is the thing worth pinning about a pair of names where
4055 // one is the front of the other.
4056 let (opts, _) = compile(&["-fsafety-restrict", "a.c"]);
4057 assert_eq!(opts.safety, rucc_session::Safety::Off);
4058
4059 let e = parse_args(&args(&["-fsafety-restrict=blocks", "a.c"])).unwrap_err();
4060 assert!(e.message.contains("takes no value"), "{}", e.message);
4061 }
4062
4063 #[test]
4064 fn safety_races_takes_a_mode_and_defaults_to_watching_nothing() {
4065 // Three modes rather than a bare flag, because section 9.5 gives two answers that record
4066 // the same thing and report different classes, so a flag with no value could not say which
4067 // was wanted. Off by default for the reason on `rucc_session::Races`, which is not a cost
4068 // argument: this is the one plane where an edge nobody interposed costs a false report.
4069 let (opts, _) = compile(&["a.c"]);
4070 assert_eq!(opts.races, rucc_session::Races::Off);
4071
4072 let (opts, _) = compile(&["-fsafety-races=metadata", "a.c"]);
4073 assert_eq!(opts.races, rucc_session::Races::Metadata);
4074
4075 let (opts, _) = compile(&["-fsafety-races=pointer", "a.c"]);
4076 assert_eq!(opts.races, rucc_session::Races::Pointer);
4077
4078 // Last one wins, as it does for every other mode flag here.
4079 let (opts, _) = compile(&["-fsafety-races=pointer", "-fno-safety-races", "a.c"]);
4080 assert_eq!(opts.races, rucc_session::Races::Off);
4081
4082 let e = parse_args(&args(&["-fsafety-races=all", "a.c"])).unwrap_err();
4083 assert!(e.message.contains("off, metadata or pointer"), "{}", e.message);
4084 }
4085
4086 #[test]
4087 fn print_pipeline_answers_with_the_passes_the_level_asked_for() {
4088 let a = parse_args(&args(&["--print-pipeline", "-O2"])).unwrap();
4089 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
4090 let text = print_pipeline(&opts);
4091 assert!(text.starts_with("level: -O2\n"), "{text}");
4092 assert!(text.contains("fold"), "{text}");
4093
4094 let a = parse_args(&args(&["--print-pipeline"])).unwrap();
4095 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
4096 // Two passes run at `-O0` and neither is an optimization. The first moves what
4097 // `__builtin_expect` said onto the branch and takes the instruction away, so that nothing
4098 // past the optimizer has to know the instruction exists. The second removes code nothing
4099 // reaches. See issue 359.
4100 assert!(print_pipeline(&opts).contains("1: expect,"), "{}", print_pipeline(&opts));
4101 assert!(print_pipeline(&opts).contains("2: simplify-cfg,"), "{}", print_pipeline(&opts));
4102
4103 let a = parse_args(&args(&["--print-pipeline", "-fno-simplify-cfg"])).unwrap();
4104 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
4105 // The second turns off and the first does not, because nothing below the optimizer lowers
4106 // what it removes, so `-fno-expect` is a compile that stops rather than one that runs.
4107 let text = print_pipeline(&opts);
4108 assert!(text.contains("1: expect,"), "{text}");
4109 assert!(!text.contains("simplify-cfg"), "{text}");
4110 }
4111
4112 #[test]
4113 fn print_pipeline_takes_the_toggles_into_account() {
4114 let a = parse_args(&args(&["--print-pipeline", "-O2", "-fno-fold"])).unwrap();
4115 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
4116 let text = print_pipeline(&opts);
4117 // The one that was named is gone and the rest of the level is not, which is the whole
4118 // of what a toggle promises.
4119 assert!(!text.contains("fold"), "{text}");
4120 assert!(text.contains("dce"), "{text}");
4121
4122 // Every pass the compiler has, named off. Built from the registry rather than written
4123 // out, so a pass added later is turned off here too and this keeps testing the thing it
4124 // is about, which is that the toggles can empty a level down to the passes that are not
4125 // optional. Those are named, because a listing that is all of them is a level nobody
4126 // emptied and the assertion would pass while saying nothing.
4127 let mut off = vec!["--print-pipeline".to_owned(), "-O2".to_owned()];
4128 off.extend(rucc_opt::PASSES.iter().map(|p| format!("-fno-{}", p.name())));
4129 let spelled: Vec<&str> = off.iter().map(String::as_str).collect();
4130 let a = parse_args(&args(&spelled)).unwrap();
4131 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
4132 let text = print_pipeline(&opts);
4133 let left: Vec<&str> =
4134 rucc_opt::PASSES.iter().filter(|p| p.required()).map(|p| p.name()).collect();
4135 assert_eq!(left, vec!["expect"], "{text}");
4136 for (at, name) in left.iter().enumerate() {
4137 assert!(text.contains(&format!("{}: {name},", at + 1)), "{text}");
4138 }
4139 assert!(!text.contains("dce"), "{text}");
4140 }
4141
4142 #[test]
4143 fn print_pipeline_says_when_a_budget_will_stop_the_run_short() {
4144 let a = parse_args(&args(&["--print-pipeline", "-O2"])).unwrap();
4145 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
4146 assert!(!print_pipeline(&opts).contains("global fuel"));
4147
4148 let a = parse_args(&args(&["--print-pipeline", "-O2", "-fpass-fuel-global=4"])).unwrap();
4149 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
4150 let text = print_pipeline(&opts);
4151 // Because the listing is the answer to what this compilation will do, and a run that
4152 // stops after four rewrites is not doing what the level says it does.
4153 assert!(text.contains("global fuel: 4"), "{text}");
4154 }
4155
4156 /// A pass is turned on and off by its own name, and the order the flags were given in is
4157 /// kept, because the last spelling of a name is the one that decides.
4158 #[test]
4159 fn a_pass_is_named_by_dash_f_and_unnamed_by_dash_f_no() {
4160 let (opts, _) = compile(&["-c", "-O0", "-ffold", "-fno-fold", "-ffold", "a.c"]);
4161 assert_eq!(
4162 opts.passes,
4163 [("fold".to_owned(), true), ("fold".to_owned(), false), ("fold".to_owned(), true)]
4164 );
4165
4166 let e = parse_args(&args(&["-fno-such-pass", "a.c"])).unwrap_err();
4167 assert!(e.message.contains("unknown option"), "{}", e.message);
4168 }
4169
4170 #[test]
4171 fn pass_fuel_names_a_pass_and_a_count_and_refuses_anything_else() {
4172 let (opts, _) = compile(&["-c", "-O2", "-fpass-fuel=fold=3", "a.c"]);
4173 assert_eq!(opts.pass_fuel, [("fold".to_owned(), 3)]);
4174
4175 let e = parse_args(&args(&["-fpass-fuel=fold", "a.c"])).unwrap_err();
4176 assert!(e.message.contains("<pass>=<count>"), "{}", e.message);
4177 let e = parse_args(&args(&["-fpass-fuel=nosuch=3", "a.c"])).unwrap_err();
4178 assert!(e.message.contains("--print-pipeline"), "{}", e.message);
4179 let e = parse_args(&args(&["-fpass-fuel=fold=lots", "a.c"])).unwrap_err();
4180 assert!(e.message.contains("not a number"), "{}", e.message);
4181 }
4182
4183 #[test]
4184 fn global_pass_fuel_is_a_count_on_its_own_and_defaults_to_no_limit() {
4185 let (opts, _) = compile(&["-c", "-O2", "a.c"]);
4186 assert_eq!(opts.pass_fuel_global, None);
4187
4188 let (opts, _) = compile(&["-c", "-O2", "-fpass-fuel-global=12", "a.c"]);
4189 assert_eq!(opts.pass_fuel_global, Some(12));
4190 // And it is not the per pass flag with a longer name, so neither spelling swallows the
4191 // other.
4192 assert!(opts.pass_fuel.is_empty());
4193
4194 let e = parse_args(&args(&["-fpass-fuel-global=lots", "a.c"])).unwrap_err();
4195 assert!(e.message.contains("not a number"), "{}", e.message);
4196 }
4197
4198 #[test]
4199 fn a_gate_names_a_pass_and_optionally_the_functions_it_covers() {
4200 let (opts, _) = compile(&["-c", "-O2", "-fdisable-fold", "-fenable-fold=2-4,main", "a.c"]);
4201 assert_eq!(
4202 opts.pass_gates,
4203 [(false, "fold".to_owned()), (true, "fold=2-4,main".to_owned())],
4204 "the order is what decides, so it has to survive the parse"
4205 );
4206
4207 let e = parse_args(&args(&["-fdisable-nosuch", "a.c"])).unwrap_err();
4208 assert!(e.message.contains("--print-pipeline"), "{}", e.message);
4209 let e = parse_args(&args(&["-fenable-fold=9-2", "a.c"])).unwrap_err();
4210 assert!(e.message.contains("ends before it starts"), "{}", e.message);
4211 let e = parse_args(&args(&["-fdisable-fold=", "a.c"])).unwrap_err();
4212 assert!(e.message.contains("is empty"), "{}", e.message);
4213 }
4214
4215 #[test]
4216 fn the_pipeline_listing_says_which_passes_a_gate_touched() {
4217 let (opts, _) = compile(&["-c", "-O2", "-fdisable-fold=main", "a.c"]);
4218 let text = print_pipeline(&opts);
4219 assert!(text.contains("fold, "), "{text}");
4220 assert!(text.contains("[off for main]"), "{text}");
4221 }
4222
4223 /// The spelling is checked while the arguments are read, because a dump that names a pass
4224 /// this compiler does not have is a typo, and a typo found after the compilation has run is
4225 /// found too late to be any use.
4226 #[test]
4227 fn a_dump_is_checked_when_it_is_asked_for_rather_than_when_it_is_taken() {
4228 let (opts, _) = compile(&["-c", "-O2", "-fdump-ir=all", "-fdump-ir=after-fold", "a.c"]);
4229 assert_eq!(opts.dump_ir, ["all", "after-fold"]);
4230
4231 let e = parse_args(&args(&["-fdump-ir=after-nosuch", "a.c"])).unwrap_err();
4232 assert!(e.message.contains("nosuch"), "{}", e.message);
4233 assert!(parse_args(&args(&["-fdump-ir=sideways-fold", "a.c"])).is_err());
4234 }
4235
4236 /// Every spelling `-fopt-info` takes, and the one it does not.
4237 ///
4238 /// The keywords are checked here for the same reason a dump's pass name is: a person who
4239 /// misspelled one gets no output, and no output is also what a compilation where nothing
4240 /// happened looks like. Telling those two apart is the entire reason to reach for this flag.
4241 #[test]
4242 fn opt_info_takes_kinds_and_a_file_and_refuses_a_kind_it_does_not_have() {
4243 let (opts, _) = compile(&["-c", "-O2", "-fopt-info", "a.c"]);
4244 assert_eq!(opts.opt_info, [""], "a bare flag asks for the rewrites");
4245 assert_eq!(opts.opt_info_file, None, "and goes to standard error");
4246
4247 let (opts, _) = compile(&["-c", "-O2", "-fopt-info-missed-note", "a.c"]);
4248 assert_eq!(opts.opt_info, ["missed-note"]);
4249
4250 // Two flags add up rather than the second replacing the first, and the file is the last
4251 // one that named a file, which is how GCC treats both.
4252 let (opts, _) =
4253 compile(&["-c", "-O2", "-fopt-info-missed=one.txt", "-fopt-info-all=two.txt", "a.c"]);
4254 assert_eq!(opts.opt_info, ["missed", "all"]);
4255 assert_eq!(opts.opt_info_file.as_deref(), Some("two.txt"));
4256
4257 let e = parse_args(&args(&["-fopt-info-vectorized", "a.c"])).unwrap_err();
4258 assert!(e.message.contains("vectorized"), "{}", e.message);
4259 assert!(e.message.contains("`missed`"), "{}", e.message);
4260 let e = parse_args(&args(&["-fopt-info-missed=", "a.c"])).unwrap_err();
4261 assert!(e.message.contains("no file"), "{}", e.message);
4262 }
4263
4264 #[test]
4265 fn verify_each_is_unstable_and_off_unless_it_was_asked_for() {
4266 let (opts, _) = compile(&["-c", "-Zverify-each", "a.c"]);
4267 assert!(opts.verify_each);
4268 assert!(!USAGE.contains("verify-each"), "an unstable option stays out of the usage text");
4269 }
4270
4271 #[test]
4272 fn dash_o_needs_an_argument() {
4273 let e = parse_args(&args(&["a.c", "-o"])).unwrap_err();
4274 assert_eq!(e.message, "-o requires an argument");
4275 }
4276
4277 #[test]
4278 fn dash_d_and_dash_u_are_read_joined_or_separated_and_keep_their_order() {
4279 let (opts, _) = compile(&["-DFOO=1", "-D", "BAR", "-UBAZ", "-U", "QUX", "a.c"]);
4280 assert_eq!(opts.defines, ["FOO=1", "BAR"]);
4281 assert_eq!(opts.undefines, ["BAZ", "QUX"]);
4282 }
4283
4284 #[test]
4285 fn the_include_flags_land_on_the_chain_each_one_names() {
4286 // A sysroot with nothing under it, so that the library's own directories are the
4287 // same on every machine this test runs on, which is none of them.
4288 let (opts, _) = compile(&[
4289 "-Ii",
4290 "-iquote",
4291 "q",
4292 "-isystem",
4293 "sys",
4294 "-idirafter",
4295 "after",
4296 "--sysroot=/nowhere-at-all",
4297 "a.c",
4298 ]);
4299 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
4300 // The compiler's own headers sit after every `-isystem` and before `-idirafter`,
4301 // which is where GCC puts its own: a directory the user named outranks ours.
4302 assert_eq!(dirs, ["q", "i", "sys", runtime::DIR, "after"]);
4303 assert!(!opts.search.dirs()[1].is_system);
4304 assert!(opts.search.dirs()[2].is_system);
4305 }
4306
4307 #[test]
4308 fn the_librarys_headers_come_after_the_compilers_own_and_go_away_with_them() {
4309 // Which machine this runs on decides what is on the path, so the test is about the
4310 // order rather than about the names: ours is on it, the library's follow it, and
4311 // `-nostdinc` is the one flag that takes both halves of the pair off at once.
4312 let (opts, _) = compile(&["a.c"]);
4313 let dirs = opts.search.dirs();
4314 let ours = dirs.iter().position(|d| d.path.to_str() == Some(runtime::DIR));
4315 assert_eq!(ours, Some(0), "{dirs:?}");
4316 assert!(dirs[1..].iter().all(|d| d.is_system), "{dirs:?}");
4317 let (bare, _) = compile(&["-nostdinc", "a.c"]);
4318 assert!(bare.search.dirs().is_empty(), "{:?}", bare.search.dirs());
4319 }
4320
4321 #[test]
4322 fn a_sysroot_moves_the_librarys_directories_and_nothing_else() {
4323 let (opts, _) = compile(&["-isystem", "sys", "--sysroot=/nowhere-at-all", "a.c"]);
4324 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
4325 assert_eq!(dirs, ["sys", runtime::DIR]);
4326 }
4327
4328 #[test]
4329 fn a_cross_compile_reads_the_targets_own_headers_rather_than_the_ones_next_door() {
4330 // The target is not the machine this test runs on wherever it runs, so the answer is the
4331 // same on all of them: the libc's two include directories for that target, the kernel's
4332 // two, and nothing from here. A header read from here is the quiet failure of section 8.5, a
4333 // program that builds on the build machine and is wrong everywhere else.
4334 let (opts, _) = compile(&["--target=riscv64-linux-musl", "-c", "a.c"]);
4335 let dirs: Vec<&std::path::Path> =
4336 opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
4337 let root = cache::dir().join("sysroots").join("riscv64-linux-musl");
4338 let kernel = cache::dir().join("kernel-headers");
4339 assert_eq!(dirs.len(), 5, "{dirs:?}");
4340 assert_eq!(dirs[0], std::path::Path::new(runtime::DIR));
4341 assert_eq!(dirs[1], root.join("include").join("riscv64"));
4342 assert_eq!(dirs[2], root.join("include").join("generic"));
4343 // The kernel's, which are beside the sysroots rather than inside one, because every target
4344 // that shares an architecture reads the same files.
4345 assert_eq!(dirs[3], kernel.join("riscv"));
4346 assert_eq!(dirs[4], kernel.join("generic"));
4347 }
4348
4349 #[test]
4350 fn a_cross_compile_to_something_that_is_not_linux_reads_no_kernel_headers() {
4351 // The other side of the same answer. Windows has its own system headers and no `linux/` at
4352 // all, so the list is the libc's own and the question never arises, which is the `None` that
4353 // `link::cross_kernel` returns rather than a directory nothing would be found in.
4354 //
4355 // The libc's own is one directory rather than two here, because mingw-w64 publishes a single
4356 // header tree for every architecture and `Sysroot::splits_by_arch` says so.
4357 let (opts, _) = compile(&["--target=x86_64-pc-windows-gnu", "-c", "a.c"]);
4358 let dirs: Vec<&std::path::Path> =
4359 opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
4360 assert_eq!(dirs.len(), 2, "{dirs:?}");
4361 assert!(!dirs.iter().any(|dir| dir.ends_with("kernel-headers")), "{dirs:?}");
4362 }
4363
4364 #[test]
4365 fn the_glibc_version_macro_goes_with_the_bundled_tree_and_with_nothing_else() {
4366 // One tree serves every glibc release, so the release is what the target supplies, and the
4367 // condition is the same one that chose the directories. A host glibc and a tree somebody
4368 // named both define `__GLIBC_MINOR__` in their own `features.h`, and two definitions with
4369 // different values is a warning on every compilation of every file.
4370 //
4371 // The architecture is chosen against this machine's rather than written down, because the
4372 // bundled tree is only in effect for a target that is not this machine. The first version of
4373 // this test said x86_64-linux-gnu, which is a cross compile on a mac and this machine on a
4374 // Linux runner, so it passed here and failed there.
4375 let gnu = format!("--target={}-linux-gnu", cross_arch());
4376 let (bundled, _) = compile(&[&gnu, "-c", "a.c"]);
4377 assert_eq!(bundled.glibc_minor, Some(44));
4378 let pin = format!("{gnu}.2.28");
4379 let (pinned, _) = compile(&[&pin, "-c", "a.c"]);
4380 assert_eq!(pinned.glibc_minor, Some(28));
4381
4382 let (named, _) = compile(&[&gnu, "--sysroot=/nowhere-at-all", "-c", "a.c"]);
4383 assert_eq!(named.glibc_minor, None);
4384 let (none, _) = compile(&[&gnu, "-nostdinc", "-c", "a.c"]);
4385 assert_eq!(none.glibc_minor, None);
4386 let musl = format!("--target={}-linux-musl", cross_arch());
4387 let (musl, _) = compile(&[&musl, "-c", "a.c"]);
4388 assert_eq!(musl.glibc_minor, None);
4389
4390 // And this machine's own target gets nothing, whatever this machine is, because its headers
4391 // come from the machine and its own `features.h` defines the macro. On a glibc Linux box
4392 // that is the case this test had backwards; on a mac it is true for the other reason, which
4393 // is that Darwin is not a glibc target at all.
4394 if let Some(host) = Triple::host() {
4395 let native = format!("--target={}", host.tuple());
4396 let (native, _) = compile(&[&native, "-c", "a.c"]);
4397 assert_eq!(native.glibc_minor, None);
4398 }
4399 }
4400
4401 #[test]
4402 fn a_pinned_release_on_this_machines_own_target_reads_the_bundled_tree() {
4403 // The end to end half of the answer in `link::cross_for`. A release named for this machine's
4404 // own target is a cross compile, so the headers are the bundled tree's and the macro says
4405 // what was asked for rather than what this machine has.
4406 //
4407 // Only on a glibc box, because a release is a glibc release: a mac has no `__GLIBC_MINOR__`
4408 // to get wrong and nothing to pin. That makes this a test the Linux runners carry, which is
4409 // where the case lives.
4410 let Some(host) = Triple::host() else { return };
4411 if host.env != rucc_target::Env::Gnu {
4412 return;
4413 }
4414 let pin = format!("--target={}.2.28", host.tuple());
4415 let (opts, _) = compile(&[&pin, "-c", "a.c"]);
4416 assert_eq!(opts.glibc_minor, Some(28));
4417 let root = cache::dir().join("sysroots").join(format!("{}.2.28", host.tuple()));
4418 let dirs: Vec<&std::path::Path> =
4419 opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
4420 assert!(dirs.iter().any(|dir| dir.starts_with(&root)), "{dirs:?}");
4421 // And nothing of this machine's, which is the failure this was: a program compiled against
4422 // 2.44 declarations and told it was 2.28.
4423 assert!(!dirs.iter().any(|dir| *dir == std::path::Path::new("/usr/include")), "{dirs:?}");
4424 }
4425
4426 /// An architecture that is not this machine's, out of the three the driver has targets for.
4427 ///
4428 /// A test about the bundled sysroot has to name a target that is not the host, because a target
4429 /// that is the host reads the host's own headers and libraries. Asking which machine this is
4430 /// beats picking a row and hoping, and it is two lines.
4431 fn cross_arch() -> &'static str {
4432 match Triple::host().map(|host| host.arch) {
4433 Some(rucc_target::Arch::X86_64) => "aarch64",
4434 _ => "x86_64",
4435 }
4436 }
4437
4438 #[test]
4439 fn a_glibc_newer_than_the_bundled_tree_is_refused_by_name() {
4440 // Both versions in the message, because the two things a person can do about it are pin a
4441 // release the tree has and name a sysroot that has the one they asked for, and neither is a
4442 // choice they can make without knowing which release the tree is.
4443 //
4444 // Not this machine's architecture, for the reason the test above gives: the refusal is about
4445 // the bundled tree, and the bundled tree is not what a target that is this machine reads.
4446 let target = format!("--target={}-linux-gnu.2.99", cross_arch());
4447 let message = refused(&[&target, "-c", "a.c"]);
4448 assert!(message.contains("asked for glibc 2.99"), "{message}");
4449 assert!(message.contains("bundled headers are glibc 2.44"), "{message}");
4450 assert!(message.contains("--sysroot"), "{message}");
4451 }
4452
4453 #[test]
4454 fn a_sysroot_the_user_named_is_still_what_a_cross_compile_reads() {
4455 // The tree somebody assembled beats the one we would build, on the headers as on the
4456 // libraries. It is empty here, which is why the list comes out short: the directories under
4457 // it are checked for rather than assumed, and a tree that is not there offers nothing.
4458 let (opts, _) =
4459 compile(&["--target=riscv64-linux-musl", "--sysroot=/nowhere-at-all", "-c", "a.c"]);
4460 let dirs: Vec<&std::path::Path> =
4461 opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
4462 assert_eq!(dirs, [std::path::Path::new(runtime::DIR)]);
4463 }
4464
4465 #[test]
4466 fn dash_i_dash_moves_the_bracket_directories_into_the_quoted_chain() {
4467 let (opts, _) =
4468 compile(&["-Iinc1", "-iquote", "inc2", "-I-", "-Iinc3", "-nostdinc", "a.c"]);
4469 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
4470 assert_eq!(dirs, ["inc1", "inc2", "inc3"]);
4471 // An angled include sees only what came after the flag.
4472 assert_eq!(opts.search.start(IncludeForm::Angled), 2);
4473 assert!(!opts.search.searches_current_dir());
4474 }
4475
4476 #[test]
4477 fn the_prefix_flags_stick_what_iprefix_said_on_the_front_of_what_follows_it() {
4478 let (opts, _) = compile(&[
4479 "-iprefix",
4480 "/tools/",
4481 "-iwithprefix",
4482 "late",
4483 "-iwithprefixbefore",
4484 "early",
4485 "-iprefix",
4486 "/other/",
4487 "-iwithprefix",
4488 "last",
4489 "-nostdinc",
4490 "a.c",
4491 ]);
4492 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
4493 // `-iwithprefixbefore` is an `-I` and the other two are `-isystem`, which is where GCC
4494 // puts them rather than where its manual says it does.
4495 assert_eq!(dirs, ["/tools/early", "/tools/late", "/other/last"]);
4496 assert!(!opts.search.dirs()[0].is_system);
4497 assert!(opts.search.dirs()[1].is_system);
4498 }
4499
4500 #[test]
4501 fn the_files_named_on_the_command_line_keep_their_order_and_which_flag_named_them() {
4502 let (opts, _) =
4503 compile(&["-include", "one.h", "-imacros", "two.h", "-include", "3.h", "a.c"]);
4504 let names: Vec<&str> = opts.preincludes.iter().map(|p| p.name.as_str()).collect();
4505 assert_eq!(names, ["one.h", "two.h", "3.h"]);
4506 assert_eq!(opts.preincludes.iter().filter(|p| p.macros_only).count(), 1);
4507 }
4508
4509 #[test]
4510 fn nostdinc_takes_the_compilers_own_headers_off_the_path() {
4511 let (opts, _) = compile(&["-Ii", "-nostdinc", "a.c"]);
4512 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
4513 assert_eq!(dirs, ["i"]);
4514 }
4515
4516 #[test]
4517 fn the_dialect_flags_set_the_language_and_the_extensions_separately() {
4518 let (opts, _) = compile(&["-std=gnu11", "a.c"]);
4519 assert_eq!(opts.std, Std::C11);
4520 assert!(opts.gnu_extensions);
4521
4522 let (opts, _) = compile(&["-std=iso9899:1999", "a.c"]);
4523 assert_eq!(opts.std, Std::C99);
4524 assert!(!opts.gnu_extensions);
4525
4526 let (opts, _) = compile(&["-ansi", "a.c"]);
4527 assert_eq!(opts.std, Std::C89);
4528 assert!(!opts.gnu_extensions);
4529
4530 let (opts, _) = compile(&["-std=gnu2y", "a.c"]);
4531 assert_eq!(opts.std, Std::C2y);
4532 assert!(opts.gnu_extensions);
4533
4534 let e = parse_args(&args(&["-std=c94jr", "a.c"])).unwrap_err();
4535 assert!(e.message.contains("unknown dialect"), "{}", e.message);
4536 }
4537
4538 #[test]
4539 fn the_dump_letters_are_a_family_and_everything_else_beginning_with_d_is_not() {
4540 let (opts, _) = compile(&["-dM", "a.c"]);
4541 assert!(opts.dumps.macros);
4542
4543 // Packed, the way GCC takes them, and a letter in the family we have not written yet
4544 // is accepted and does nothing rather than failing a build.
4545 let (opts, _) = compile(&["-dDM", "a.c"]);
4546 assert!(opts.dumps.macros);
4547 let (opts, _) = compile(&["-dD", "a.c"]);
4548 assert!(!opts.dumps.macros);
4549
4550 let (opts, _) = compile(&["a.c"]);
4551 assert!(!opts.dumps.any());
4552
4553 // `-dumpversion` is a different flag that happens to start the same way, and it is read
4554 // as itself rather than as a dump of nothing.
4555 assert_eq!(printed(&["-dumpversion", "a.c"]), VERSION);
4556 }
4557
4558 #[test]
4559 fn the_gcc_version_claimed_is_a_flag_and_the_short_spellings_are_the_ones_people_write() {
4560 let (opts, _) = compile(&["a.c"]);
4561 assert_eq!(
4562 opts.gnuc,
4563 GnucVersion { major: 16, minor: 0, patch: 0 },
4564 "the release this compiler is written against, and the earliest one of that series"
4565 );
4566
4567 let (opts, _) = compile(&["-fgnuc-version=15.1.0", "a.c"]);
4568 assert_eq!(opts.gnuc, GnucVersion { major: 15, minor: 1, patch: 0 });
4569
4570 // A missing component is zero. `gcc -dumpversion` says `15` on a release with no
4571 // patchlevel and a harness that pastes that back has to be understood.
4572 let (opts, _) = compile(&["-fgnuc-version=15", "a.c"]);
4573 assert_eq!(opts.gnuc, GnucVersion { major: 15, minor: 0, patch: 0 });
4574
4575 let (opts, _) = compile(&["-fgnuc-version=13.2", "a.c"]);
4576 assert_eq!(opts.gnuc, GnucVersion { major: 13, minor: 2, patch: 0 });
4577
4578 let e = parse_args(&args(&["-fgnuc-version=15.x", "a.c"])).unwrap_err();
4579 assert!(e.message.contains("minor that is not a number"), "{}", e.message);
4580
4581 let e = parse_args(&args(&["-fgnuc-version=1.2.3.4", "a.c"])).unwrap_err();
4582 assert!(e.message.contains("more than three"), "{}", e.message);
4583 }
4584
4585 #[test]
4586 fn pedantic_has_two_spellings_and_is_not_the_same_knob_as_the_dialect() {
4587 let (opts, _) = compile(&["-std=c17", "-pedantic", "a.c"]);
4588 assert!(opts.pedantic);
4589 assert_eq!(opts.std, Std::C17);
4590
4591 // The `-W` family's name for it, which is what a build that groups its warning flags
4592 // tends to write.
4593 let (opts, _) = compile(&["-Wpedantic", "a.c"]);
4594 assert!(opts.pedantic);
4595
4596 let (opts, _) = compile(&["-std=c17", "a.c"]);
4597 assert!(!opts.pedantic, "a dialect on its own does not diagnose an extension");
4598 }
4599
4600 #[test]
4601 fn dash_p_and_dash_ffreestanding_reach_the_options() {
4602 let (opts, _) = compile(&["-E", "-P", "-ffreestanding", "a.c"]);
4603 assert!(!opts.line_markers);
4604 assert!(!opts.hosted);
4605 assert_eq!(opts.emit, EmitKind::Preprocessed);
4606 }
4607
4608 /// The two ways a build says it means its own function by a name the C library also has.
4609 ///
4610 /// `-fno-builtin` is all of them and `-fno-builtin-<name>` is one, and the second is what a
4611 /// build writes when it means its own `memcpy` and the library's everything else. The name is
4612 /// kept as it was written and not checked against anything, because a program is allowed to
4613 /// mean something by a name this compiler has never heard of.
4614 #[test]
4615 fn the_builtin_flags_are_read_in_both_directions_and_one_name_at_a_time() {
4616 let (opts, _) = compile(&["-c", "a.c"]);
4617 assert!(opts.builtins, "a library name means the library function by default");
4618 assert!(opts.no_builtin.is_empty());
4619
4620 let (opts, _) = compile(&["-c", "-fno-builtin", "a.c"]);
4621 assert!(!opts.builtins);
4622
4623 let (opts, _) = compile(&["-c", "-fno-builtin", "-fbuiltin", "a.c"]);
4624 assert!(opts.builtins, "the last mention decides");
4625
4626 let (opts, _) = compile(&["-c", "-fno-builtin-memcpy", "-fno-builtin-nonesuch", "a.c"]);
4627 assert!(opts.builtins, "one name is not the family");
4628 assert_eq!(opts.no_builtin, vec!["memcpy".to_owned(), "nonesuch".to_owned()]);
4629 }
4630
4631 /// `-fvisibility=`, which is on every cmake project that cares about which names it exports
4632 /// and which was refused as an unknown option until now.
4633 ///
4634 /// Four spellings and three answers. `internal` is hidden plus a promise about never taking
4635 /// the address across a component boundary, and nothing derives anything from that promise
4636 /// here, so it comes out as the weaker of the two rather than as a refusal that stops a build
4637 /// over a distinction this compiler does not make.
4638 #[test]
4639 fn visibility_takes_the_four_spellings_gcc_takes_and_refuses_the_rest() {
4640 let (opts, _) = compile(&["-c", "a.c"]);
4641 assert_eq!(opts.visibility, Visibility::Default, "exported unless something says not");
4642
4643 for (written, wanted) in [
4644 ("default", Visibility::Default),
4645 ("hidden", Visibility::Hidden),
4646 ("internal", Visibility::Hidden),
4647 ("protected", Visibility::Protected),
4648 ] {
4649 let (opts, _) = compile(&["-c", &format!("-fvisibility={written}"), "a.c"]);
4650 assert_eq!(opts.visibility, wanted, "{written}");
4651 }
4652
4653 // The last mention decides, which is what every other flag of this shape does and what a
4654 // build that turns something off for one directory relies on.
4655 let (opts, _) = compile(&["-c", "-fvisibility=hidden", "-fvisibility=default", "a.c"]);
4656 assert_eq!(opts.visibility, Visibility::Default, "the last mention decides");
4657
4658 // A spelling gcc does not take is refused rather than read as the default, because a
4659 // build that meant hidden and got exported is a library with the wrong interface and
4660 // nothing said about it anywhere.
4661 let failed = parse_args(&args(&["-fvisibility=none", "a.c"])).expect_err("refused");
4662 assert!(failed.to_string().contains("is not a visibility"), "{failed}");
4663 }
4664
4665 /// `-ffp-contract=`, which is the one flag in the floating point group that is kept rather than
4666 /// described, and the values are gcc 16's three.
4667 #[test]
4668 fn how_far_a_multiply_and_an_addition_may_be_fused_is_asked_for() {
4669 let (opts, _) = compile(&["-c", "a.c"]);
4670 assert_eq!(opts.fp_contract, Contract::Off, "a licence nobody granted is not assumed");
4671
4672 for (written, wanted) in
4673 [("off", Contract::Off), ("on", Contract::On), ("fast", Contract::Fast)]
4674 {
4675 let (opts, _) = compile(&["-c", &format!("-ffp-contract={written}"), "a.c"]);
4676 assert_eq!(opts.fp_contract, wanted, "{written}");
4677 }
4678
4679 let (opts, _) = compile(&["-c", "-ffp-contract=fast", "-ffp-contract=off", "a.c"]);
4680 assert_eq!(opts.fp_contract, Contract::Off, "the last mention decides");
4681
4682 // Refused rather than read as one of the three, because a build that asked for no fusing
4683 // and was given the default would be one whose numbers change and whose command line says
4684 // they should not. gcc refuses the same spellings and names the same three in its message.
4685 for bad in ["-ffp-contract=none", "-ffp-contract=", "-ffp-contract=Fast"] {
4686 let failed = parse_args(&args(&[bad, "a.c"])).expect_err("refused");
4687 assert!(failed.to_string().contains("is not a contraction"), "{bad}: {failed}");
4688 }
4689
4690 // And the other one that takes a value, which is taken and kept nowhere: every operation
4691 // here is computed in the type it was written in, so `standard` is what happens and the
4692 // other two are permission to do something this does not do.
4693 let failed = parse_args(&args(&["-fexcess-precision=long", "a.c"])).expect_err("refused");
4694 assert!(failed.to_string().contains("is not an excess precision"), "{failed}");
4695 }
4696
4697 /// The four prefix mapping flags, which are what a distribution passes to get the same bytes
4698 /// out of `/build/pkg-1.2` and out of `/home/someone/pkg-1.2`. Three lists rather than one
4699 /// because gcc has three, and `-ffile-prefix-map=` is the three of them at once.
4700 #[test]
4701 fn a_prefix_mapping_flag_goes_on_the_list_its_spelling_names() {
4702 let (opts, _) = compile(&["-c", "a.c"]);
4703 assert!(opts.prefix_map.macros.is_empty(), "nothing is rewritten unless it is asked for");
4704 assert!(opts.prefix_map.debug.is_empty(), "nor here");
4705 assert!(opts.prefix_map.profile.is_empty(), "nor here");
4706
4707 let (opts, _) = compile(&["-c", "-fmacro-prefix-map=/build=.", "a.c"]);
4708 assert_eq!(opts.prefix_map.macros.apply("/build/a.c"), "./a.c", "the one it names");
4709 assert!(opts.prefix_map.debug.is_empty(), "and not the two it does not");
4710
4711 let (opts, _) = compile(&["-c", "-fdebug-prefix-map=/build=.", "a.c"]);
4712 assert_eq!(opts.prefix_map.debug.apply("/build/a.c"), "./a.c", "the one it names");
4713 assert!(opts.prefix_map.macros.is_empty(), "and not the two it does not");
4714
4715 let (opts, _) = compile(&["-c", "-fprofile-prefix-map=/build=.", "a.c"]);
4716 assert_eq!(opts.prefix_map.profile.apply("/build/a.c"), "./a.c", "the one it names");
4717 assert!(opts.prefix_map.macros.is_empty(), "and not the two it does not");
4718
4719 let (opts, _) = compile(&["-c", "-ffile-prefix-map=/build=.", "a.c"]);
4720 for list in [&opts.prefix_map.macros, &opts.prefix_map.debug, &opts.prefix_map.profile] {
4721 assert_eq!(list.apply("/build/a.c"), "./a.c", "all three at once");
4722 }
4723
4724 // Every mention is kept and the last one that matches wins, unlike the flags above whose
4725 // last mention replaces the earlier ones. A build writes one of these per source root and
4726 // expects all of them to be in force, which is the whole point of a list.
4727 let (opts, _) =
4728 compile(&["-c", "-ffile-prefix-map=/a=one", "-ffile-prefix-map=/b=two", "a.c"]);
4729 assert_eq!(opts.prefix_map.macros.apply("/a/x.c"), "one/x.c", "the earlier one still acts");
4730 assert_eq!(opts.prefix_map.macros.apply("/b/x.c"), "two/x.c", "and so does the later one");
4731
4732 // An argument with no `=` is refused rather than ignored, because a build whose paths were
4733 // meant to be rewritten and were not is one that ships the build directory's name and says
4734 // nothing about it. gcc refuses the same thing.
4735 for bad in ["-fmacro-prefix-map=nope", "-ffile-prefix-map=", "-fdebug-prefix-map=/build"] {
4736 let failed = parse_args(&args(&[bad, "a.c"])).expect_err("refused");
4737 assert!(failed.to_string().contains("is not a rewrite for"), "{bad}: {failed}");
4738 }
4739 }
4740
4741 /// `-ffunction-sections` and `-fdata-sections`, which are what make `--gc-sections` able to
4742 /// drop anything: a linker can leave out a section nothing reaches and cannot leave out half of
4743 /// one. A kernel and an embedded image are both linked that way.
4744 ///
4745 /// Two flags rather than one because gcc has two, and a build that asks for one of them and not
4746 /// the other is a build that measured something: splitting the code is nearly free at link time
4747 /// and splitting the data can defeat the linker's ordering of what is next to what.
4748 #[test]
4749 fn a_section_per_function_and_a_section_per_variable_are_asked_for_one_at_a_time() {
4750 let (opts, _) = compile(&["-c", "a.c"]);
4751 assert!(!opts.function_sections, "one text section unless something says otherwise");
4752 assert!(!opts.data_sections);
4753
4754 let (opts, _) = compile(&["-c", "-ffunction-sections", "a.c"]);
4755 assert!(opts.function_sections);
4756 assert!(!opts.data_sections, "one flag is not the other");
4757
4758 let (opts, _) = compile(&["-c", "-fdata-sections", "a.c"]);
4759 assert!(opts.data_sections);
4760 assert!(!opts.function_sections);
4761
4762 // Both directions taken, and the off one is what happens anyway rather than a refusal,
4763 // since a build that writes it is asking for the default.
4764 let (opts, _) = compile(&[
4765 "-c",
4766 "-ffunction-sections",
4767 "-fno-function-sections",
4768 "-fdata-sections",
4769 "-fno-data-sections",
4770 "a.c",
4771 ]);
4772 assert!(!opts.function_sections, "the last mention decides");
4773 assert!(!opts.data_sections, "the last mention decides");
4774 }
4775
4776 /// `-fgnu89-inline`, which is off by default and is not implied by anything on the command
4777 /// line, since the dialect asks for GNU's reading further in rather than through this.
4778 #[test]
4779 fn gnu89_inline_is_off_until_it_is_asked_for_and_the_last_mention_decides() {
4780 let (opts, _) = compile(&["-c", "a.c"]);
4781 assert!(!opts.gnu89_inline, "C's reading of inline by default");
4782
4783 let (opts, _) = compile(&["-c", "-fgnu89-inline", "a.c"]);
4784 assert!(opts.gnu89_inline);
4785
4786 let (opts, _) = compile(&["-c", "-fgnu89-inline", "-fno-gnu89-inline", "a.c"]);
4787 assert!(!opts.gnu89_inline, "the last mention decides");
4788
4789 // The C89 dialects are under GNU's reading whether this was written or not, so the flag
4790 // stays off there and the dialect is what the checker and the macro set both ask. That is
4791 // also why `-std=c89 -fno-gnu89-inline` needs no diagnostic: it asks for the reading the
4792 // dialect already has. gcc refuses that command line, which is measured in the issue.
4793 let (opts, _) = compile(&["-c", "-std=c89", "a.c"]);
4794 assert!(!opts.gnu89_inline);
4795 }
4796
4797 /// Both spellings of both frame flags, since a build that wants one usually writes the
4798 /// other beside it for the one file that has to be compiled the ordinary way.
4799 #[test]
4800 fn the_two_frame_flags_are_read_in_both_directions() {
4801 let (opts, _) = compile(&["-c", "a.c"]);
4802 assert!(!opts.frame_pointer, "gcc omits it above -O0 and so does this");
4803 assert!(opts.red_zone, "the psABI has one and nothing said not to use it");
4804
4805 let (opts, _) = compile(&["-c", "-fno-omit-frame-pointer", "-mno-red-zone", "a.c"]);
4806 assert!(opts.frame_pointer);
4807 assert!(!opts.red_zone);
4808
4809 let (opts, _) = compile(&[
4810 "-c",
4811 "-fno-omit-frame-pointer",
4812 "-fomit-frame-pointer",
4813 "-mno-red-zone",
4814 "-mred-zone",
4815 "a.c",
4816 ]);
4817 assert!(!opts.frame_pointer, "the last one wins, as it does in gcc");
4818 assert!(opts.red_zone);
4819 }
4820
4821 /// Four flags rather than one with an argument, which is how gcc spells them, and the negative
4822 /// spelled three ways because a build that turns one off writes whichever it turned on.
4823 #[test]
4824 fn the_stack_protector_is_four_flags_and_the_last_one_wins() {
4825 let (opts, _) = compile(&["-c", "a.c"]);
4826 assert_eq!(opts.protector, Protector::None, "gcc protects nothing unless it was asked");
4827
4828 for (flag, want) in [
4829 ("-fstack-protector", Protector::Buffers),
4830 ("-fstack-protector-strong", Protector::Strong),
4831 ("-fstack-protector-all", Protector::All),
4832 ] {
4833 let (opts, _) = compile(&["-c", flag, "a.c"]);
4834 assert_eq!(opts.protector, want, "{flag}");
4835 }
4836
4837 // What a package build does: the strong one in the global flags and one directory that
4838 // cannot have a protector turning it off on the line after.
4839 for off in ["-fno-stack-protector", "-fno-stack-protector-strong"] {
4840 let (opts, _) = compile(&["-c", "-fstack-protector-strong", off, "a.c"]);
4841 assert_eq!(opts.protector, Protector::None, "{off}");
4842 }
4843 let (opts, _) = compile(&["-c", "-fno-stack-protector", "-fstack-protector-all", "a.c"]);
4844 assert_eq!(opts.protector, Protector::All, "the last one wins either way round");
4845 }
4846
4847 /// A switch rather than a level, because how a frame is taken is one question and which
4848 /// functions get a canary is another, and gcc spells it that way for the same reason.
4849 #[test]
4850 fn taking_a_frame_a_page_at_a_time_is_off_until_it_is_asked_for() {
4851 let (opts, _) = compile(&["-c", "a.c"]);
4852 assert!(!opts.stack_clash, "gcc takes a frame in one subtraction unless it was asked");
4853
4854 let (opts, _) = compile(&["-c", "-fstack-clash-protection", "a.c"]);
4855 assert!(opts.stack_clash);
4856
4857 // The same shape a package build uses for the protector: on in the global flags and off
4858 // for the one directory that cannot have it.
4859 let (opts, _) =
4860 compile(&["-c", "-fstack-clash-protection", "-fno-stack-clash-protection", "a.c"]);
4861 assert!(!opts.stack_clash);
4862 let (opts, _) =
4863 compile(&["-c", "-fno-stack-clash-protection", "-fstack-clash-protection", "a.c"]);
4864 assert!(opts.stack_clash, "the last one wins either way round");
4865
4866 // The two are independent, since one is about the frame and the other about the function.
4867 let (opts, _) =
4868 compile(&["-c", "-fstack-clash-protection", "-fstack-protector-strong", "a.c"]);
4869 assert!(opts.stack_clash);
4870 assert_eq!(opts.protector, Protector::Strong);
4871 }
4872
4873 /// One flag with an argument rather than a family of spellings, because what it asks about is
4874 /// which of the two edges of a control flow transfer is checked and the two are not separate
4875 /// questions to the hardware.
4876 #[test]
4877 fn which_control_flow_edges_are_checked_is_asked_for_by_name() {
4878 let (opts, _) = compile(&["-c", "a.c"]);
4879 assert_eq!(opts.control, Control::None, "gcc's default on the targets this compiler has");
4880
4881 for (arg, want) in [
4882 ("-fcf-protection", Control::Full),
4883 ("-fcf-protection=full", Control::Full),
4884 ("-fcf-protection=branch", Control::Branch),
4885 ("-fcf-protection=return", Control::Return),
4886 ("-fcf-protection=none", Control::None),
4887 ("-fcf-protection=check", Control::Check),
4888 ] {
4889 let (opts, _) = compile(&["-c", arg, "a.c"]);
4890 assert_eq!(opts.control, want, "{arg}");
4891 }
4892
4893 // The shape a package build uses: on in the global flags and off for the one directory
4894 // that cannot have it, whichever of the two spellings of off it reaches for.
4895 let (opts, _) = compile(&["-c", "-fcf-protection=full", "-fno-cf-protection", "a.c"]);
4896 assert_eq!(opts.control, Control::None);
4897 let (opts, _) = compile(&["-c", "-fno-cf-protection", "-fcf-protection=branch", "a.c"]);
4898 assert_eq!(opts.control, Control::Branch, "the last one wins either way round");
4899 }
4900
4901 /// The profiler is asked for by two spellings, and where its hook goes by two more.
4902 ///
4903 /// The two halves are separate on purpose. `-mfentry` on its own says where a call would go and
4904 /// asks for no call, which is what gcc does with it, and a build system that sets it globally
4905 /// and asks for the profile per directory needs that to be true rather than an error.
4906 ///
4907 /// The link is asserted alongside, because the flag changes it too and a build that compiled
4908 /// with it and linked without it is a program that calls the hook everywhere and never writes a
4909 /// profile.
4910 #[test]
4911 fn the_profiler_and_where_its_hook_goes_are_two_separate_questions() {
4912 let (opts, _) = compile(&["-c", "a.c"]);
4913 assert!(!opts.profile);
4914 assert_eq!(opts.hook, Hook::Platform, "neither was named, so the target decides");
4915
4916 for arg in ["-pg", "-p"] {
4917 let (opts, _) = compile(&["-c", arg, "a.c"]);
4918 assert!(opts.profile, "{arg}");
4919 let (link, _) = linking(&[arg, "a.c"]);
4920 assert!(link.profile, "{arg} changes the link as well");
4921 }
4922
4923 for (arg, want) in [("-mfentry", Hook::Early), ("-mno-fentry", Hook::Late)] {
4924 let (opts, _) = compile(&["-c", arg, "a.c"]);
4925 assert_eq!(opts.hook, want, "{arg}");
4926 assert!(!opts.profile, "{arg} asks for no call of its own");
4927 }
4928
4929 let (opts, _) = compile(&["-c", "-mfentry", "-mno-fentry", "-pg", "a.c"]);
4930 assert_eq!(opts.hook, Hook::Late, "the last one wins");
4931 assert!(opts.profile);
4932 }
4933
4934 /// How much room a patcher is promised, which is one number or two.
4935 ///
4936 /// A command line that did not ask is asserted alongside, because the flag has to be written to
4937 /// mean anything and a build that reserved room nobody asked for would grow every function in
4938 /// it for nothing.
4939 #[test]
4940 fn the_room_a_patcher_is_promised_is_a_number_of_bytes_and_where_they_go() {
4941 let (opts, _) = compile(&["-c", "a.c"]);
4942 assert_eq!(opts.patchable, Patchable::default());
4943 assert!(!opts.patchable.any(), "nothing is reserved unless it was asked for");
4944
4945 let (opts, _) = compile(&["-c", "-fpatchable-function-entry=16", "a.c"]);
4946 assert_eq!(opts.patchable, Patchable { total: 16, before: 0 });
4947
4948 let (opts, _) = compile(&["-c", "-fpatchable-function-entry=5,3", "a.c"]);
4949 assert_eq!(opts.patchable, Patchable { total: 5, before: 3 });
4950 assert_eq!(opts.patchable.after(), 2);
4951
4952 // The last one wins, which is what every other flag of this shape does and what a build
4953 // that adds one to a command line it did not write is relying on.
4954 let (opts, _) = compile(&[
4955 "-c",
4956 "-fpatchable-function-entry=5,3",
4957 "-fpatchable-function-entry=2",
4958 "a.c",
4959 ]);
4960 assert_eq!(opts.patchable, Patchable { total: 2, before: 0 });
4961 }
4962
4963 /// And a request nothing could satisfy is refused rather than rounded into one that can be.
4964 #[test]
4965 fn room_in_front_of_the_label_that_is_more_than_the_room_asked_for_is_refused() {
4966 for arg in ["-fpatchable-function-entry=1,2", "-fpatchable-function-entry=x"] {
4967 let e = parse_args(&args(&["-c", arg, "a.c"])).unwrap_err();
4968 assert!(e.message.contains("is not an amount of room to reserve"), "{}", e.message);
4969 }
4970 }
4971
4972 /// What wraps rather than being undefined, which is two questions and three flags.
4973 ///
4974 /// The older flag is the pair of the newer two, which is gcc's own reading of it, so a build
4975 /// that writes `-fno-strict-overflow` gets both and a build that writes one of the others gets
4976 /// only what it asked for.
4977 #[test]
4978 fn what_overflows_rather_than_being_undefined_is_asked_for_two_ways() {
4979 let (opts, _) = compile(&["-c", "a.c"]);
4980 assert_eq!(opts.wrapping, Wrapping::NONE, "nothing wraps unless it was asked for");
4981
4982 let (opts, _) = compile(&["-c", "-fwrapv", "a.c"]);
4983 assert_eq!(opts.wrapping, Wrapping { signed: true, pointer: false, trap: false });
4984
4985 let (opts, _) = compile(&["-c", "-fwrapv-pointer", "a.c"]);
4986 assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: true, trap: false });
4987
4988 let (opts, _) = compile(&["-c", "-fno-strict-overflow", "a.c"]);
4989 assert_eq!(opts.wrapping, Wrapping::ALL);
4990
4991 // And the last one wins, in both directions. A build that turns one of these on globally
4992 // and off for one directory is relying on that, and so is one that writes the pair and
4993 // then takes half of it back.
4994 let (opts, _) = compile(&["-c", "-fwrapv", "-fno-wrapv", "a.c"]);
4995 assert_eq!(opts.wrapping, Wrapping::NONE);
4996
4997 let (opts, _) = compile(&["-c", "-fno-strict-overflow", "-fstrict-overflow", "a.c"]);
4998 assert_eq!(opts.wrapping, Wrapping::NONE);
4999
5000 let (opts, _) = compile(&["-c", "-fno-strict-overflow", "-fno-wrapv-pointer", "a.c"]);
5001 assert_eq!(opts.wrapping, Wrapping { signed: true, pointer: false, trap: false });
5002 }
5003
5004 /// And the other answer to the signed question cannot be held at the same time as the first.
5005 ///
5006 /// A program cannot both wrap and stop, so writing both is writing a contradiction, and gcc
5007 /// resolves it by letting the last one win rather than by reporting anything. That was measured
5008 /// against gcc 16 rather than read out of the manual, which says nothing about it: `-ftrapv
5009 /// -fwrapv` emits no checked calls and `-fwrapv -ftrapv` emits them.
5010 #[test]
5011 fn a_signed_overflow_that_stops_is_the_other_answer_and_not_a_third_one() {
5012 let (opts, _) = compile(&["-c", "-ftrapv", "a.c"]);
5013 assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: false, trap: true });
5014
5015 let (opts, _) = compile(&["-c", "-fwrapv", "-ftrapv", "a.c"]);
5016 assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: false, trap: true });
5017
5018 let (opts, _) = compile(&["-c", "-ftrapv", "-fwrapv", "a.c"]);
5019 assert_eq!(opts.wrapping, Wrapping { signed: true, pointer: false, trap: false });
5020
5021 let (opts, _) = compile(&["-c", "-ftrapv", "-fno-strict-overflow", "a.c"]);
5022 assert_eq!(opts.wrapping, Wrapping::ALL);
5023
5024 let (opts, _) = compile(&["-c", "-ftrapv", "-fno-trapv", "a.c"]);
5025 assert_eq!(opts.wrapping, Wrapping::NONE);
5026
5027 // And the flag that says what may be assumed says nothing about what happens, so it leaves
5028 // this alone where it takes the wrapping away. gcc does the same.
5029 let (opts, _) = compile(&["-c", "-ftrapv", "-fstrict-overflow", "a.c"]);
5030 assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: false, trap: true });
5031 }
5032
5033 /// What a plain `char` is, which is four spellings of two answers and nothing by default.
5034 ///
5035 /// Nothing is the target's own answer and has to stay distinct from both of the others, since
5036 /// the same command line means a signed `char` on x86-64 and an unsigned one on Linux's arm64.
5037 /// The negative spellings are the other flag rather than a way of asking for the default, which
5038 /// was measured against gcc 16: `-fno-signed-char` defines `__CHAR_UNSIGNED__` and
5039 /// `-fno-unsigned-char` does not.
5040 #[test]
5041 fn the_signedness_of_a_plain_char_is_asked_for_in_four_ways() {
5042 let (opts, _) = compile(&["-c", "a.c"]);
5043 assert_eq!(opts.char_signed, None);
5044
5045 for flag in ["-fsigned-char", "-fno-unsigned-char"] {
5046 let (opts, _) = compile(&["-c", flag, "a.c"]);
5047 assert_eq!(opts.char_signed, Some(true), "{flag}");
5048 }
5049
5050 for flag in ["-funsigned-char", "-fno-signed-char"] {
5051 let (opts, _) = compile(&["-c", flag, "a.c"]);
5052 assert_eq!(opts.char_signed, Some(false), "{flag}");
5053 }
5054
5055 // And the last one wins, which is what a build that sets one globally and the other for a
5056 // directory relies on.
5057 let (opts, _) = compile(&["-c", "-funsigned-char", "-fsigned-char", "a.c"]);
5058 assert_eq!(opts.char_signed, Some(true));
5059
5060 // And what is asked for reaches the target, because that is what every other part of the
5061 // compiler asks. The triple is one whose own answer is the opposite, so a session that
5062 // ignored the flag would still read as signed here.
5063 let (opts, _) =
5064 compile(&["-c", "--target=aarch64-unknown-linux-gnu", "-fsigned-char", "a.c"]);
5065 assert!(Session::new(*opts).target.char_is_signed);
5066 let (opts, _) = compile(&["-c", "--target=aarch64-unknown-linux-gnu", "a.c"]);
5067 assert!(!Session::new(*opts).target.char_is_signed);
5068 }
5069
5070 /// And the size of an enumeration, which is one question with two spellings.
5071 #[test]
5072 fn the_smallest_enumeration_is_asked_for_and_taken_back() {
5073 let (opts, _) = compile(&["-c", "a.c"]);
5074 assert!(!opts.short_enums);
5075
5076 let (opts, _) = compile(&["-c", "-fshort-enums", "a.c"]);
5077 assert!(opts.short_enums);
5078
5079 let (opts, _) = compile(&["-c", "-fshort-enums", "-fno-short-enums", "a.c"]);
5080 assert!(!opts.short_enums);
5081
5082 let (opts, _) = compile(&["-c", "-fno-short-enums", "-fshort-enums", "a.c"]);
5083 assert!(opts.short_enums);
5084 }
5085
5086 /// And Microsoft's reading of an anonymous member, which the target answers where the command
5087 /// line said nothing. gcc's mingw build has it on and its Linux build has it off, so a header
5088 /// that closes a nameless union with a macro that expands to nothing is read the way the
5089 /// compiler that platform ships would read it.
5090 #[test]
5091 fn the_microsoft_reading_of_a_member_follows_the_target_until_it_is_asked_for() {
5092 // Named rather than left to the host, since the answer this asks for is the one a target
5093 // that is not Windows gives and on a Windows machine the host is not one of those.
5094 let (opts, _) = compile(&[LINUX, "-c", "a.c"]);
5095 assert!(!Session::new(*opts).ms_extensions());
5096
5097 let (opts, _) = compile(&["-c", "--target=x86_64-pc-windows-gnu", "a.c"]);
5098 assert!(Session::new(*opts).ms_extensions());
5099
5100 let (opts, _) = compile(&["-c", "-fms-extensions", "a.c"]);
5101 assert!(Session::new(*opts).ms_extensions());
5102
5103 let (opts, _) =
5104 compile(&["-c", "--target=x86_64-pc-windows-gnu", "-fno-ms-extensions", "a.c"]);
5105 assert!(!Session::new(*opts).ms_extensions());
5106 }
5107
5108 /// And a value nothing means is refused rather than taken for the nearest thing it looks like.
5109 ///
5110 /// `-fcf-protection=all` is the spelling somebody writes from memory, and a compiler that read
5111 /// it as `full` would be guessing, while one that let it fall through to the optimizer's `-f`
5112 /// family would report it as an unknown pass. Neither is the news the build wants.
5113 #[test]
5114 fn a_control_flow_protection_nothing_means_is_refused() {
5115 let e = parse_args(&args(&["-c", "-fcf-protection=all", "a.c"])).unwrap_err();
5116 assert!(e.message.contains("is not a control flow protection"), "{}", e.message);
5117 assert!(e.message.contains("full, branch, return, none or check"), "{}", e.message);
5118 }
5119
5120 #[test]
5121 fn the_link_flags_are_collected_apart_from_the_compilation() {
5122 let (link, _) = linking(&[
5123 "-static",
5124 "-nostartfiles",
5125 "-rdynamic",
5126 "-s",
5127 "-fuse-ld=mold",
5128 "-L/opt/lib",
5129 "-B",
5130 "/opt/tools",
5131 "a.c",
5132 ]);
5133 assert!(link.is_static);
5134 assert!(link.no_startfiles);
5135 assert!(link.export_dynamic);
5136 assert!(link.strip);
5137 assert_eq!(link.use_ld.as_deref(), Some("mold"));
5138 assert_eq!(link.search, vec![PathBuf::from("/opt/lib")]);
5139 assert_eq!(link.prefixes, vec![PathBuf::from("/opt/tools")]);
5140 }
5141
5142 #[test]
5143 fn a_comma_in_dash_wl_separates_two_arguments() {
5144 // The target is written down because the name of the object is derived from it, and `a.o`
5145 // on a Linux host is `a.obj` on a Windows one. What is under test is the splitting of the
5146 // argument, which has nothing to do with either.
5147 let (_, plan) = linking(&[LINUX, "-Wl,-rpath,/opt/lib", "-Xlinker", "--as-needed", "a.c"]);
5148 let link = plan.link.expect("expected a link step");
5149 assert_eq!(
5150 link.inputs,
5151 vec![
5152 link::Item::Linker("-rpath".into()),
5153 link::Item::Linker("/opt/lib".into()),
5154 link::Item::Linker("--as-needed".into()),
5155 link::Item::File("a.o".into()),
5156 ]
5157 );
5158 }
5159
5160 #[test]
5161 fn a_word_for_the_linker_keeps_its_place_among_the_files_too() {
5162 // What libtool writes around a set of convenience archives, and what #1279 was. Both words
5163 // are about the files between them, so the pair collected out of the line and appended to
5164 // the end is two options that bracket nothing and an archive that went in empty.
5165 let (_, plan) = linking(&[
5166 "--target=x86_64-unknown-linux-gnu",
5167 "a.c",
5168 "-Wl,--whole-archive",
5169 "libaesni.a",
5170 "-Wl,--no-whole-archive",
5171 "-lm",
5172 ]);
5173 let link = plan.link.expect("expected a link step");
5174 assert_eq!(
5175 link.inputs,
5176 vec![
5177 link::Item::File("a.o".into()),
5178 link::Item::Linker("--whole-archive".into()),
5179 link::Item::File("libaesni.a".into()),
5180 link::Item::Linker("--no-whole-archive".into()),
5181 link::Item::Library("m".into()),
5182 ]
5183 );
5184 // And it is not a job, because there is nothing to compile in a word for the linker.
5185 assert_eq!(plan.jobs.len(), 2);
5186 }
5187
5188 #[test]
5189 fn a_word_for_the_linker_on_a_dash_c_line_is_dropped_without_a_word() {
5190 // GCC says nothing about one either. `-Wl,` on a compile line is what a build system
5191 // writes when one variable holds the flags for both, and a note here would be a note on
5192 // every compile of every autotools project.
5193 let (_, plan) = linking(&["-c", "-Wl,--as-needed", "a.c"]);
5194 assert!(plan.link.is_none());
5195 assert!(plan.notes.is_empty(), "{:?}", plan.notes);
5196 assert_eq!(plan.jobs.len(), 1);
5197 }
5198
5199 #[test]
5200 fn a_library_keeps_its_place_between_the_objects() {
5201 // Link order is semantic: `-lm` written between two files resolves for the one before
5202 // it and not for the one after, so a library cannot be collected into a list of its own.
5203 // The target is named because the suffix of an object is the target's and this asserts
5204 // on the names: the same command line on a Windows host plans two `.obj` files.
5205 let (_, plan) = linking(&["--target=x86_64-unknown-linux-gnu", "a.c", "-lm", "b.c"]);
5206 let link = plan.link.expect("expected a link step");
5207 assert_eq!(
5208 link.inputs,
5209 vec![
5210 link::Item::File("a.o".into()),
5211 link::Item::Library("m".into()),
5212 link::Item::File("b.o".into()),
5213 ]
5214 );
5215 // And it is not a job, because there is nothing to compile in a library.
5216 assert_eq!(plan.jobs.len(), 2);
5217 }
5218
5219 #[test]
5220 fn a_library_on_a_dash_c_line_is_a_note_rather_than_an_error() {
5221 let (_, plan) = linking(&["-c", "-lm", "a.c"]);
5222 assert!(plan.link.is_none());
5223 assert!(plan.notes.iter().any(|n| n.contains("-lm")), "{:?}", plan.notes);
5224 }
5225
5226 #[test]
5227 fn the_sysroot_reaches_the_linker_as_well_as_the_headers() {
5228 let (link, _) = linking(&["--sysroot=/opt/root", "a.c"]);
5229 assert_eq!(link.sysroot, Some(PathBuf::from("/opt/root")));
5230 }
5231
5232 fn printed(s: &[&str]) -> String {
5233 match parse_args(&args(s)).expect("expected an answer") {
5234 Action::Print(line) => line,
5235 other => panic!("expected an answer, got {other:?}"),
5236 }
5237 }
5238
5239 fn refused(s: &[&str]) -> String {
5240 parse_args(&args(s)).expect_err("expected a refusal").message
5241 }
5242
5243 #[test]
5244 fn a_warning_flag_this_compiler_has_not_heard_of_is_taken_rather_than_refused() {
5245 // The rule in section 4.1, and the reason for it is autoconf: a configure script finds
5246 // out whether a warning flag exists by passing it and looking at the exit status, so a
5247 // compiler that refuses one it does not know fails a script written for a newer GCC.
5248 let (opts, _) = compile(&["-Wall", "-Wextra", "-Wno-format-truncation", "-c", "a.c"]);
5249 assert!(!opts.warnings_are_errors);
5250 assert!(opts.warnings);
5251 // The two spellings that do mean something are still read.
5252 let (opts, _) = compile(&["-Werror", "-c", "a.c"]);
5253 assert!(opts.warnings_are_errors);
5254 let (opts, _) = compile(&["-w", "-c", "a.c"]);
5255 assert!(!opts.warnings);
5256 // Off without being asked, the way gcc has it off, and both spellings are read.
5257 let (opts, _) = compile(&["-c", "a.c"]);
5258 assert!(!opts.system_header_warnings);
5259 let (opts, _) = compile(&["-Wsystem-headers", "-c", "a.c"]);
5260 assert!(opts.system_header_warnings);
5261 let (opts, _) = compile(&["-Wsystem-headers", "-Wno-system-headers", "-c", "a.c"]);
5262 assert!(!opts.system_header_warnings);
5263 let (opts, _) = compile(&["-pedantic-errors", "-c", "a.c"]);
5264 assert!(opts.pedantic && opts.warnings_are_errors);
5265 }
5266
5267 #[test]
5268 fn an_argument_for_a_separate_tool_is_refused_rather_than_dropped() {
5269 // Every one of these says something about the output, so the wrong answer is silence.
5270 assert!(refused(&["-Wa,--noexecstack", "-c", "a.c"]).contains("separate assembler"));
5271 assert!(refused(&["-Wp,-DX", "-c", "a.c"]).contains("separate assembler"));
5272 assert!(refused(&["-specs=/x", "a.c"]).contains("-specs= is not supported"));
5273 assert!(refused(&["-mcmodel=kernel", "-c", "a.c"]).contains("small code model"));
5274 assert!(refused(&["-gdwarf-4", "-c", "a.c"]).contains("DWARF 5"));
5275 assert!(refused(&["-Ofast", "-c", "a.c"]).contains("fast math"));
5276 // The word size the target does not have, which is a target this compiler was not asked
5277 // for rather than a flag it does not know.
5278 let no32 = refused(&["--target=x86_64-unknown-linux-gnu", "-m32", "-c", "a.c"]);
5279 assert!(no32.contains("32 bit target"), "{no32}");
5280 }
5281
5282 /// `-gz` and the two spellings of the split, which are the two questions about the shape of
5283 /// the debug output rather than about how much of it there is.
5284 ///
5285 /// Both answers here are about what happens when there is debug information to shape, and
5286 /// there is none yet, so what is being asserted is that the flags are read and remembered
5287 /// rather than that anything changed in the output. That is the whole of what taking them
5288 /// claims, and it is worth a test because the day `rucc-debug` writes a section this is where
5289 /// it comes to find out what the command line said.
5290 #[test]
5291 fn the_shape_of_the_debug_output_is_recorded_even_where_there_is_none_of_it() {
5292 let (opts, _) = compile(&["-c", "a.c"]);
5293 assert_eq!(opts.compress, Compress::None, "uncompressed unless somebody asks");
5294
5295 // Bare `-gz` is `-gz=zlib`, measured against gcc 16 rather than read out of the manual,
5296 // which describes the flag without ever saying which algorithm it picks.
5297 assert_eq!(compile(&["-gz", "-c", "a.c"]).0.compress, Compress::Zlib);
5298 for (spelling, want) in [
5299 ("none", Compress::None),
5300 ("zlib", Compress::Zlib),
5301 ("zlib-gnu", Compress::ZlibGnu),
5302 ("zstd", Compress::Zstd),
5303 ] {
5304 let (opts, _) = compile(&[&format!("-gz={spelling}"), "-c", "a.c"]);
5305 assert_eq!(opts.compress, want, "{spelling}");
5306 }
5307
5308 // A value nothing here has heard of is refused rather than rounded to the nearest one,
5309 // because a build that asked for `zstd` and quietly got `zlib` would ship a file its
5310 // reader may not understand and would have no way of finding out.
5311 for bad in ["-gz=gzip", "-gz="] {
5312 let failed = refused(&[bad, "-c", "a.c"]);
5313 assert!(failed.contains("is not a way to compress"), "{bad}: {failed}");
5314 }
5315
5316 // The split is refused in the direction that would have written a file and taken in the
5317 // direction that describes what happens. A build system that names the `.dwo` as an
5318 // output has to hear about it now rather than at the point the file is missing.
5319 let (opts, _) = compile(&["-gno-split-dwarf", "-g", "-c", "a.c"]);
5320 assert!(opts.debug_info, "the negative spelling says nothing about how much");
5321 let failed = refused(&["-gsplit-dwarf", "-c", "a.c"]);
5322 assert!(failed.contains(".dwo"), "the refusal names the file it would have written");
5323 }
5324
5325 /// The `-flto` family, which is the whole of an optimization this compiler does not do.
5326 ///
5327 /// Taken rather than refused because ignoring it gives a correct program that is slower than
5328 /// it could have been, which is section 4.1's hint about speed. The values are still held to
5329 /// gcc's, so a command line written for clang is told rather than quietly taken.
5330 #[test]
5331 fn the_link_time_family_is_read_and_checked_and_nothing_is_done_about_it() {
5332 let (opts, _) = compile(&["-c", "a.c"]);
5333 assert!(!opts.lto.requested, "nothing asks unless the command line does");
5334
5335 let (opts, _) = compile(&["-flto", "-c", "a.c"]);
5336 assert!(opts.lto.requested);
5337 assert_eq!(opts.lto.jobs, LtoJobs::One, "bare -flto is one process, the way gcc reads it");
5338
5339 // The last of the two directions wins, the same as every other pair of `-f` spellings.
5340 assert!(!compile(&["-flto", "-fno-lto", "-c", "a.c"]).0.lto.requested);
5341 assert!(compile(&["-fno-lto", "-flto", "-c", "a.c"]).0.lto.requested);
5342
5343 // A count is a count, and asking for one implies asking for the optimization.
5344 for (spelling, want) in [
5345 ("auto", LtoJobs::Auto),
5346 ("jobserver", LtoJobs::Jobserver),
5347 ("1", LtoJobs::One),
5348 ("8", LtoJobs::Count(8)),
5349 ] {
5350 let (opts, _) = compile(&[&format!("-flto={spelling}"), "-c", "a.c"]);
5351 assert_eq!(opts.lto.jobs, want, "{spelling}");
5352 assert!(opts.lto.requested, "{spelling} asks for it too");
5353 }
5354
5355 // gcc refuses a zero rather than reading it as `-fno-lto`, and `thin` is clang's spelling
5356 // of a question gcc answers with `-flto-partition=`, so somebody who wrote it meant a
5357 // different compiler and gets told so here rather than getting a serial link.
5358 for bad in ["-flto=0", "-flto=thin", "-flto=full", "-flto=-1"] {
5359 let failed = refused(&[bad, "-c", "a.c"]);
5360 assert!(failed.contains("link time jobs"), "{bad}: {failed}");
5361 }
5362
5363 // How the program is cut up before the work is spread over it.
5364 assert_eq!(compile(&["-c", "a.c"]).0.lto.partition, Partition::Balanced, "gcc's default");
5365 for (spelling, want) in [
5366 ("balanced", Partition::Balanced),
5367 ("1to1", Partition::OneToOne),
5368 ("one", Partition::One),
5369 ("max", Partition::Max),
5370 ("none", Partition::None),
5371 ] {
5372 let (opts, _) = compile(&[&format!("-flto-partition={spelling}"), "-c", "a.c"]);
5373 assert_eq!(opts.lto.partition, want, "{spelling}");
5374 }
5375 assert!(refused(&["-flto-partition=big", "-c", "a.c"]).contains("partitioning model"));
5376
5377 // And how hard the bytecode is compressed on its way into the object, which is zstd's
5378 // range of levels and is the range gcc checks an argument against.
5379 assert_eq!(compile(&["-c", "a.c"]).0.lto.compression, None, "whatever it does by default");
5380 assert_eq!(compile(&["-flto-compression-level=0", "-c", "a.c"]).0.lto.compression, Some(0));
5381 let (opts, _) = compile(&["-flto-compression-level=19", "-c", "a.c"]);
5382 assert_eq!(opts.lto.compression, Some(19));
5383 for bad in ["-flto-compression-level=20", "-flto-compression-level=-1"] {
5384 let failed = refused(&[bad, "-c", "a.c"]);
5385 assert!(failed.contains("compression level"), "{bad}: {failed}");
5386 }
5387
5388 // The two pairs that describe an arrangement rather than ask for one. Every object here
5389 // holds its machine code, so the fat spelling is what already happens and the other is a
5390 // smaller file rather than a different program, and the plugin pair is about a tool the
5391 // design in `spec/09-optimizer.md` never loads.
5392 for taken in [
5393 "-ffat-lto-objects",
5394 "-fno-fat-lto-objects",
5395 "-fuse-linker-plugin",
5396 "-fno-use-linker-plugin",
5397 ] {
5398 let (opts, _) = compile(&[taken, "-c", "a.c"]);
5399 assert!(!opts.lto.requested, "{taken} says nothing about whether to do it");
5400 }
5401 }
5402
5403 /// The profile family, which is the only one here that splits down the middle.
5404 ///
5405 /// Reading a profile is taken and writing one is refused, and the line between them is the one
5406 /// section 4.1 draws: ignoring a request to read the counts gives a correct program that is
5407 /// slower than it could have been, and ignoring a request to write them means a file the build
5408 /// declared as an output never appears.
5409 #[test]
5410 fn reading_a_profile_is_taken_and_writing_one_is_refused() {
5411 let (opts, _) = compile(&["-c", "a.c"]);
5412 assert!(!opts.profile_data.requested, "nothing asks unless the command line does");
5413 assert_eq!(opts.profile_data.path, None);
5414
5415 let (opts, _) = compile(&["-fprofile-use", "-c", "a.c"]);
5416 assert!(opts.profile_data.requested);
5417 assert_eq!(opts.profile_data.path, None, "beside the object, the way gcc looks");
5418
5419 let (opts, _) = compile(&["-fprofile-use=/counts", "-c", "a.c"]);
5420 assert!(opts.profile_data.requested, "naming a path asks for it too");
5421 assert_eq!(opts.profile_data.path.as_deref(), Some("/counts"));
5422
5423 // The last of the two directions wins, the same as every other pair of `-f` spellings.
5424 assert!(
5425 !compile(&["-fprofile-use", "-fno-profile-use", "-c", "a.c"]).0.profile_data.requested
5426 );
5427 assert!(
5428 compile(&["-fno-profile-use", "-fprofile-use", "-c", "a.c"]).0.profile_data.requested
5429 );
5430
5431 // The rest of the reading half, which is where the files are and three answers about what
5432 // to make of what is in them.
5433 let (opts, _) = compile(&[
5434 "-fprofile-dir=/build/profiles",
5435 "-fprofile-abs-path",
5436 "-fprofile-correction",
5437 "-fprofile-partial-training",
5438 "-c",
5439 "a.c",
5440 ]);
5441 assert_eq!(opts.profile_data.dir.as_deref(), Some("/build/profiles"));
5442 assert!(opts.profile_data.absolute);
5443 assert!(opts.profile_data.correction);
5444 assert!(opts.profile_data.partial_training);
5445
5446 // Writing one, which is refused by name. The first four instrument the program and the
5447 // last writes a file beside the object, and a build that got neither and no message would
5448 // go on to optimize against counts that were never gathered.
5449 for writing in [
5450 "-fprofile-generate",
5451 "-fprofile-generate=/build/profiles",
5452 "-fprofile-arcs",
5453 "--coverage",
5454 "-fcondition-coverage",
5455 "-fpath-coverage",
5456 ] {
5457 let failed = refused(&[writing, "-c", "a.c"]);
5458 assert!(failed.contains("instrument"), "{writing}: {failed}");
5459 }
5460 assert!(refused(&["-ftest-coverage", "-c", "a.c"]).contains(".gcno"), "it names the file");
5461
5462 // The negative spellings of the refused half are what already happens, so they are taken.
5463 for taken in ["-fno-profile-generate", "-fno-profile-arcs", "-fno-test-coverage"] {
5464 let (opts, _) = compile(&[taken, "-c", "a.c"]);
5465 assert!(!opts.profile_data.requested, "{taken} asks for nothing");
5466 }
5467
5468 // And the flags that describe the instrumentation that is refused above, which are checked
5469 // and dropped. Checked because a typo is worth finding here rather than on the day the
5470 // instrumentation lands.
5471 for taken in [
5472 "-fprofile-update=single",
5473 "-fprofile-update=atomic",
5474 "-fprofile-update=prefer-atomic",
5475 "-fprofile-reproducible=serial",
5476 "-fprofile-reproducible=parallel-runs",
5477 "-fprofile-reproducible=multithreaded",
5478 "-fprofile-values",
5479 "-fno-profile-values",
5480 "-fprofile-info-section",
5481 "-fprofile-filter-files=a.c",
5482 "-fprofile-exclude-files=b.c",
5483 "-fprofile-note=a.gcno",
5484 ] {
5485 let (opts, _) = compile(&[taken, "-c", "a.c"]);
5486 assert!(!opts.profile_data.requested, "{taken} says nothing about reading one");
5487 }
5488 assert!(refused(&["-fprofile-update=none", "-c", "a.c"]).contains("update method"));
5489 assert!(refused(&["-fprofile-reproducible=any", "-c", "a.c"]).contains("reproducibility"));
5490 }
5491
5492 /// The sanitizers, which are refused by name and are the one family refused for a reason that
5493 /// is not about the bytes.
5494 ///
5495 /// A sanitizer is a promise that the program is watched while it runs, so a build that asked
5496 /// for one and was quietly given a program with no checks in it gets a test suite that passes
5497 /// for the wrong reason rather than a slower program.
5498 #[test]
5499 fn a_sanitizer_that_is_still_asked_for_at_the_end_of_the_line_is_refused_by_name() {
5500 for asked in ["address", "undefined", "thread", "kernel-address", "leak", "memory"] {
5501 let failed = refused(&[&format!("-fsanitize={asked}"), "-c", "a.c"]);
5502 assert!(failed.contains(asked), "the refusal names what was asked for: {failed}");
5503 assert!(failed.contains("-fsafety=detect"), "and the nearest thing: {failed}");
5504 }
5505
5506 // A list is every name in it, and the first one still standing is the one named.
5507 let failed = refused(&["-fsanitize=address,undefined", "-c", "a.c"]);
5508 assert!(failed.contains("address"), "{failed}");
5509
5510 // A name that is not one, which is worth its own message: somebody who wrote `-fsanitize`
5511 // with a typo in it has a different problem from somebody who wrote a real one.
5512 for bad in ["-fsanitize=bogus", "-fsanitize=address,bogus", "-fno-sanitize=bogus"] {
5513 let failed = refused(&[bad, "-c", "a.c"]);
5514 assert!(failed.contains("is not a sanitizer"), "{bad}: {failed}");
5515 }
5516
5517 // gcc takes `all` only in the negative, and so does this.
5518 assert!(refused(&["-fsanitize=all", "-c", "a.c"]).contains("only `-fno-sanitize=all`"));
5519
5520 // Asking and then taking it back is asking for nothing, which is why the answer waits for
5521 // the end of the line. A build whose shared flags turn a check on and whose rule for one
5522 // file turns it off again compiles that file here.
5523 for pair in [
5524 ["-fsanitize=address", "-fno-sanitize=address"],
5525 ["-fsanitize=address,undefined", "-fno-sanitize=all"],
5526 ["-fsanitize=undefined", "-fno-sanitize=undefined"],
5527 ] {
5528 let (opts, _) = compile(&[pair[0], pair[1], "-c", "a.c"]);
5529 assert_eq!(opts.safety, rucc_session::Safety::Off, "{pair:?} asked for nothing");
5530 }
5531 // And the other order still asks, because the last word is the one that counts.
5532 assert!(!refused(&["-fno-sanitize=address", "-fsanitize=address", "-c", "a.c"]).is_empty());
5533
5534 // What a check does when it fires is an answer about checks that are refused, so there is
5535 // nothing left for it to change and it is taken.
5536 for taken in [
5537 "-fsanitize-recover=undefined",
5538 "-fno-sanitize-recover=all",
5539 "-fsanitize-trap=undefined",
5540 "-fno-sanitize-trap=all",
5541 "-fsanitize-undefined-trap-on-error",
5542 "-fsanitize-address-use-after-scope",
5543 "-fno-sanitize-address-use-after-scope",
5544 "-fsanitize-sections=.data",
5545 ] {
5546 let (opts, _) = compile(&[taken, "-c", "a.c"]);
5547 assert_eq!(opts.safety, rucc_session::Safety::Off, "{taken} asks for no checking");
5548 }
5549 assert!(refused(&["-fsanitize-recover=bogus", "-c", "a.c"]).contains("is not a sanitizer"));
5550
5551 // Coverage instrumentation is refused rather than dropped, because a fuzzer with no
5552 // feedback runs blind and never says so.
5553 let failed = refused(&["-fsanitize-coverage=trace-pc", "-c", "a.c"]);
5554 assert!(failed.contains("feedback"), "{failed}");
5555 let failed = refused(&["-fsanitize-coverage=trace-pc-guard", "-c", "a.c"]);
5556 assert!(failed.contains("trace-pc or trace-cmp"), "gcc takes two of them: {failed}");
5557 }
5558
5559 #[test]
5560 fn the_levels_gcc_spells_differently_are_the_levels_they_mean() {
5561 assert_eq!(compile(&["-O", "-c", "a.c"]).0.opt_level, OptLevel::O1);
5562 assert_eq!(compile(&["-Og", "-c", "a.c"]).0.opt_level, OptLevel::O1);
5563 assert_eq!(compile(&["-O2", "-c", "a.c"]).0.opt_level, OptLevel::O2);
5564 }
5565
5566 #[test]
5567 fn the_machine_flags_that_name_what_we_already_do_are_taken_and_the_rest_are_not() {
5568 let line = ["--target=x86_64-unknown-linux-gnu", "-m64", "-march=x86-64-v3"];
5569 let (opts, _) =
5570 compile(&[&line[..], &["-mtune=native", "-mabi=sysv", "-c", "a.c"]].concat());
5571 assert_eq!(opts.target.to_string(), "x86_64-unknown-linux-gnu");
5572 let wrong = refused(&["--target=x86_64-unknown-linux-gnu", "-mabi=ms", "-c", "a.c"]);
5573 assert!(wrong.contains("sysv convention"), "{wrong}");
5574 }
5575
5576 #[test]
5577 fn the_thread_flag_is_a_macro_and_a_library_and_the_library_goes_last() {
5578 let (opts, plan) = compile(&["-pthread", "-c", "a.c"]);
5579 assert!(opts.defines.iter().any(|d| d == "_REENTRANT"));
5580 // After the input, because a static link takes what it needs from a library when it
5581 // reaches it and not afterwards.
5582 let names: Vec<&str> = plan.jobs.iter().map(|j| j.input.as_str()).collect();
5583 assert_eq!(names, vec!["a.c"]);
5584 }
5585
5586 #[test]
5587 fn the_questions_a_build_system_asks_before_it_compiles_anything() {
5588 let target = "--target=x86_64-unknown-linux-gnu";
5589 assert_eq!(printed(&[target, "-dumpmachine"]), "x86_64-unknown-linux-gnu");
5590 assert_eq!(printed(&[target, "-dumpversion"]), VERSION);
5591 assert_eq!(printed(&[target, "-dumpfullversion"]), VERSION);
5592 assert_eq!(printed(&[target, "-print-multiarch"]), "x86_64-linux-gnu");
5593 // A name nothing holds comes back unchanged, which is GCC's rule and is what makes the
5594 // answer safe to paste into a link line whether or not the file is there.
5595 assert_eq!(printed(&[target, "-print-file-name=no-such-library.a"]), "no-such-library.a");
5596 assert_eq!(printed(&[target, "-print-prog-name=ld"]), "ld");
5597 let dirs = printed(&[target, "-print-search-dirs"]);
5598 assert!(dirs.starts_with("install: "), "{dirs}");
5599 assert!(dirs.contains("\nlibraries: ="), "{dirs}");
5600 }
5601
5602 #[test]
5603 fn the_sysroot_in_effect_is_the_one_the_command_line_named_or_the_one_for_the_target() {
5604 // A tree the user named is the answer whatever the target is, because it is the answer to
5605 // every other question too.
5606 assert_eq!(printed(&["--sysroot=/opt/cross", "-print-sysroot"]), "/opt/cross");
5607
5608 // A target that is no machine this suite runs on is read under the cache, and the answer is
5609 // the root rather than one of the directories under it, since what asks is looking for a
5610 // file of its own.
5611 let root = cache::dir().join("sysroots").join("riscv64-linux-musl");
5612 assert_eq!(
5613 printed(&["--target=riscv64-linux-musl", "-print-sysroot"]),
5614 root.display().to_string()
5615 );
5616
5617 // And a compile for this machine has no sysroot, which is the empty line GCC prints when it
5618 // was configured without one rather than a `/` that would be a claim about the filesystem.
5619 let host = Triple::host().expect("a host this compiler knows");
5620 assert_eq!(printed(&[&format!("--target={host}"), "-print-sysroot"]), "");
5621 }
5622
5623 #[test]
5624 fn the_provenance_of_a_sysroot_is_the_manifest_it_carries() {
5625 // Section 13.5 wants seven things per input and wants them machine readable, and the manifest
5626 // is the record that already has them, so the flag prints that rather than a second format.
5627 let manifest = "rucc sysroot manifest 3\n\
5628 target\tx86_64-linux-musl\n\
5629 kernel\t6.12\n\
5630 include/generic/stdio.h\tmusl-1.2.5\t\
5631 https://musl.libc.org/releases/musl-1.2.5.tar.gz\t\
5632 0000000000000000000000000000000000000000000000000000000000000000\tmit\t\
5633 bundled\n\
5634 lib/libc.so\tmusl-1.2.5\t\
5635 https://musl.libc.org/releases/musl-1.2.5.tar.gz\t\
5636 1111111111111111111111111111111111111111111111111111111111111111\tmit\t\
5637 generated\n";
5638 let tree = TempTree::new("provenance", &[("manifest", manifest)]);
5639 let sysroot = format!("--sysroot={}", tree.0.display());
5640 // The kernel line of tamnd/rucc#934 is in the answer without anything here naming it, because
5641 // the flag parses the record and renders it again rather than picking fields out of it. That
5642 // is the reason it prints a manifest and not a format of its own.
5643 //
5644 // The answer is the file without its last newline, because whatever prints it adds one. The
5645 // file is what somebody diffs the output against, so the two have to be the same bytes.
5646 assert_eq!(printed(&[&sysroot, "-print-sysroot-provenance"]) + "\n", manifest);
5647
5648 // A tree with no manifest in it is a tree somebody assembled themselves, and nothing here
5649 // knows where any of it came from. Saying nothing is the only honest answer, and a reader can
5650 // tell it from a manifest with no inputs because that one still has its two header lines.
5651 let bare = TempTree::new("provenance-bare", &[]);
5652 assert_eq!(
5653 printed(&[&format!("--sysroot={}", bare.0.display()), "-print-sysroot-provenance"]),
5654 ""
5655 );
5656
5657 // And a compile for this machine has no sysroot at all, which is the same empty answer
5658 // `-print-sysroot` gives for it.
5659 let host = Triple::host().expect("a host this compiler knows");
5660 assert_eq!(printed(&[&format!("--target={host}"), "-print-sysroot-provenance"]), "");
5661
5662 // And the other spelling, which section 13.5 is the document that writes.
5663 assert_eq!(printed(&[&sysroot, "--print-sysroot-provenance"]) + "\n", manifest);
5664
5665 // tamnd/rucc#1021. The digest of the same tree is the sha256 of that record, so it is one
5666 // line where the provenance is a few hundred, and it is checkable with `sha256sum` because
5667 // the bytes it is over are the bytes of the file. The number here is that hash of the
5668 // fixture above, computed by `sha256sum` rather than by this compiler.
5669 assert_eq!(
5670 printed(&[&sysroot, "-print-sysroot-digest"]),
5671 "d705ae6ebeafeb7fda4bd57cecc7882bf49784b17015664a09cfae25a1b2000a"
5672 );
5673 assert_eq!(
5674 printed(&[&sysroot, "--print-sysroot-digest"]),
5675 printed(&[&sysroot, "-print-sysroot-digest"])
5676 );
5677
5678 // And the two empty answers are empty here too, because a digest of nothing would read as a
5679 // claim about a sysroot rather than as the absence of one.
5680 assert_eq!(
5681 printed(&[&format!("--sysroot={}", bare.0.display()), "-print-sysroot-digest"]),
5682 ""
5683 );
5684 assert_eq!(printed(&[&format!("--target={host}"), "-print-sysroot-digest"]), "");
5685 }
5686
5687 #[test]
5688 fn a_manifest_this_build_cannot_read_is_refused_rather_than_printed() {
5689 // Passing a file we could not parse to whoever asked would make their parser the one that
5690 // finds the problem, and the three uses section 13.5 gives for this are all somebody else
5691 // parsing it.
5692 let tree = TempTree::new(
5693 "provenance-bad",
5694 &[("manifest", "rucc sysroot manifest 3\ntarget\tx86_64-linux-musl\nlib/libc.a\n")],
5695 );
5696 let message =
5697 refused(&[&format!("--sysroot={}", tree.0.display()), "-print-sysroot-provenance"]);
5698 assert!(message.contains("manifest"), "{message}");
5699 assert!(message.contains("1 fields where an input has six"), "{message}");
5700
5701 // The digest is refused for the same file and for a stronger reason: a hash of bytes this
5702 // build cannot read would be a number that names a record nobody can act on.
5703 let digest =
5704 refused(&[&format!("--sysroot={}", tree.0.display()), "-print-sysroot-digest"]);
5705 assert_eq!(digest, message);
5706 }
5707
5708 #[test]
5709 fn the_two_dependency_flags_that_stop_after_the_rule_stop_after_the_rule() {
5710 let (opts, _) = compile(&["-M", "a.c"]);
5711 assert!(opts.deps.emit && opts.deps.instead_of_compiling);
5712 assert!(opts.deps.system_headers, "plain -M lists them");
5713 assert_eq!(opts.emit, EmitKind::Preprocessed);
5714
5715 // Even where a later flag asked for something else, because the family is a mode and
5716 // the mode is what the run is for.
5717 let (opts, _) = compile(&["-M", "-c", "a.c"]);
5718 assert_eq!(opts.emit, EmitKind::Preprocessed);
5719
5720 let (opts, _) = compile(&["-MM", "a.c"]);
5721 assert!(!opts.deps.system_headers);
5722 }
5723
5724 #[test]
5725 fn the_two_that_end_in_d_leave_the_compilation_alone() {
5726 let (opts, _) = compile(&["-MD", "-c", "a.c"]);
5727 assert!(opts.deps.emit && !opts.deps.instead_of_compiling);
5728 assert!(opts.deps.system_headers);
5729 assert_eq!(opts.emit, EmitKind::Object);
5730
5731 let (opts, _) = compile(&["-MMD", "-c", "a.c"]);
5732 assert!(opts.deps.emit && !opts.deps.instead_of_compiling);
5733 assert!(!opts.deps.system_headers);
5734 }
5735
5736 #[test]
5737 fn nothing_puts_the_system_headers_back_once_a_flag_has_taken_them_out() {
5738 // GCC's rule, and not an oversight in it. The flag asking for fewer of them is read as
5739 // the answer, because the other one never asked the question.
5740 let (opts, _) = compile(&["-MM", "-M", "a.c"]);
5741 assert!(!opts.deps.system_headers);
5742 let (opts, _) = compile(&["-MD", "-MMD", "-c", "a.c"]);
5743 assert!(!opts.deps.system_headers);
5744 let (opts, _) = compile(&["-MMD", "-MD", "-c", "a.c"]);
5745 assert!(!opts.deps.system_headers);
5746 }
5747
5748 #[test]
5749 fn a_target_arrives_escaped_from_one_flag_and_untouched_from_the_other() {
5750 let (opts, _) = compile(&["-MM", "-MT", "a b.o", "-MQ", "a b.o", "a.c"]);
5751 assert_eq!(opts.deps.targets, vec!["a b.o".to_owned(), "a\\ b.o".to_owned()]);
5752 }
5753
5754 #[test]
5755 fn the_rest_of_the_family_is_a_file_and_a_switch() {
5756 let (opts, _) = compile(&["-MM", "-MF", "dep.d", "-MP", "a.c"]);
5757 assert_eq!(opts.deps.file.as_deref(), Some("dep.d"));
5758 assert!(opts.deps.phony);
5759
5760 for flag in ["-MF", "-MT", "-MQ"] {
5761 let e = parse_args(&args(&[flag])).unwrap_err();
5762 assert!(e.message.contains("requires an argument"), "{}", e.message);
5763 }
5764 }
5765
5766 /// A directory of sources for one test, removed when the test is done with it.
5767 struct TempTree(PathBuf);
5768
5769 impl Drop for TempTree {
5770 fn drop(&mut self) {
5771 let _ = std::fs::remove_dir_all(&self.0);
5772 }
5773 }
5774
5775 impl TempTree {
5776 fn new(name: &str, files: &[(&str, &str)]) -> TempTree {
5777 let dir = std::env::temp_dir().join(format!("rucc-deps-{}-{name}", std::process::id()));
5778 let _ = std::fs::remove_dir_all(&dir);
5779 std::fs::create_dir_all(&dir).expect("temporary directory should be writable");
5780 for (path, text) in files {
5781 let at = dir.join(path);
5782 if let Some(parent) = at.parent() {
5783 std::fs::create_dir_all(parent).expect("creating a subdirectory should work");
5784 }
5785 std::fs::write(&at, text).expect("writing a temporary file should work");
5786 }
5787 TempTree(dir)
5788 }
5789
5790 fn path(&self, name: &str) -> String {
5791 self.0.join(name).to_string_lossy().into_owned()
5792 }
5793 }
5794
5795 #[test]
5796 fn the_rule_names_what_the_includes_found_and_names_each_of_them_once() {
5797 // End to end, because the list comes from the preprocessor and the format comes from
5798 // somewhere else, and a test of either half on its own would pass with the two of them
5799 // wired up backwards.
5800 let tree = TempTree::new(
5801 "found",
5802 &[
5803 ("a.c", "#include \"one.h\"\n#include \"two.h\"\nint main(void) { return X; }\n"),
5804 ("one.h", "#define X 0\n"),
5805 ("two.h", "#include \"one.h\"\n"),
5806 ],
5807 );
5808 let out = tree.path("dep.d");
5809 let code = run(&args(&["-MM", "-MF", &out, "-o", &tree.path("a.i"), &tree.path("a.c")]));
5810 assert_eq!(code, 0);
5811
5812 let text = std::fs::read_to_string(&out).expect("the rule should have been written");
5813 let names: Vec<&str> = text.split_whitespace().collect();
5814 // The target, the source, and each header once however many times it was reached.
5815 assert_eq!(names.first(), Some(&"a.o:"), "{text}");
5816 assert_eq!(names.iter().filter(|n| n.ends_with("one.h")).count(), 1, "{text}");
5817 assert_eq!(names.iter().filter(|n| n.ends_with("two.h")).count(), 1, "{text}");
5818 // And the `-o` went to the file the rule replaced, which is left empty rather than
5819 // absent because a makefile that named it as a target will look for it.
5820 assert_eq!(std::fs::read(tree.path("a.i")).expect("the output should exist"), b"");
5821 }
5822
5823 #[test]
5824 fn a_header_that_is_only_reached_under_a_guard_is_still_a_dependency() {
5825 // The multiple-include optimization means the second reach never opens the file. It is
5826 // still a file this translation unit was built from, so it is still in the rule.
5827 let tree = TempTree::new(
5828 "guarded",
5829 &[
5830 ("a.c", "#include \"g.h\"\n#include \"g.h\"\nint main(void) { return 0; }\n"),
5831 ("g.h", "#ifndef G\n#define G\n#endif\n"),
5832 ],
5833 );
5834 let out = tree.path("dep.d");
5835 let code = run(&args(&["-MM", "-MF", &out, "-o", &tree.path("a.i"), &tree.path("a.c")]));
5836 assert_eq!(code, 0);
5837 let text = std::fs::read_to_string(&out).expect("the rule should have been written");
5838 assert_eq!(text.split_whitespace().filter(|n| n.ends_with("g.h")).count(), 1, "{text}");
5839 }
5840
5841 #[test]
5842 fn every_imacros_file_is_read_before_every_include_file_whatever_order_they_were_written() {
5843 // Measured against GCC rather than read: the two flags the other way round produce the
5844 // same output byte for byte, so the command line order between the two families does not
5845 // decide anything and the order within one does. The `-include` file here can only see
5846 // the definition if the `-imacros` file that was written after it ran first.
5847 let tree = TempTree::new(
5848 "preinclude",
5849 &[
5850 ("a.c", "int main(void) { return 0; }\n"),
5851 ("i.h", "#ifdef FROM_MACROS\nint saw_it;\n#else\nint missed_it;\n#endif\n"),
5852 ("m.h", "#define FROM_MACROS 1\nint macros_text;\n"),
5853 ],
5854 );
5855 let out = tree.path("a.i");
5856 let code = run(&args(&[
5857 "-E",
5858 "-include",
5859 &tree.path("i.h"),
5860 "-imacros",
5861 &tree.path("m.h"),
5862 "-o",
5863 &out,
5864 &tree.path("a.c"),
5865 ]));
5866 assert_eq!(code, 0);
5867 let text = std::fs::read_to_string(&out).expect("the output should have been written");
5868 assert!(text.contains("saw_it"), "{text}");
5869 // And the text of the `-imacros` file is thrown away, which is the whole difference
5870 // between the two flags.
5871 assert!(!text.contains("macros_text"), "{text}");
5872 }
5873
5874 #[test]
5875 fn a_file_the_command_line_named_is_a_prerequisite_the_same_as_one_a_directive_named() {
5876 let tree = TempTree::new(
5877 "preinclude-deps",
5878 &[
5879 ("a.c", "int main(void) { return 0; }\n"),
5880 ("i.h", "int from_include;\n"),
5881 ("m.h", "#define M 1\n"),
5882 ],
5883 );
5884 let out = tree.path("dep.d");
5885 let code = run(&args(&[
5886 "-MM",
5887 "-MF",
5888 &out,
5889 "-include",
5890 &tree.path("i.h"),
5891 "-imacros",
5892 &tree.path("m.h"),
5893 "-o",
5894 &tree.path("a.i"),
5895 &tree.path("a.c"),
5896 ]));
5897 assert_eq!(code, 0);
5898 let text = std::fs::read_to_string(&out).expect("the rule should have been written");
5899 assert!(text.contains("i.h"), "{text}");
5900 assert!(text.contains("m.h"), "{text}");
5901 }
5902
5903 #[test]
5904 fn a_command_line_include_that_is_nowhere_on_the_path_is_an_error_and_not_a_warning() {
5905 // Including the directory of the source file, which is not on the path for these: the
5906 // command line was not written there, so a name in it is relative to where the compiler
5907 // was run rather than to where the source sits.
5908 let tree = TempTree::new(
5909 "preinclude-missing",
5910 &[("sub/a.c", "int main(void) { return 0; }\n"), ("sub/beside.h", "int x;\n")],
5911 );
5912 let code = run(&args(&["-E", "-include", "beside.h", "-o", "-", &tree.path("sub/a.c")]));
5913 assert_eq!(code, 1);
5914 }
5915
5916 #[test]
5917 fn a_command_line_that_links_names_the_executable_and_not_the_object_it_went_through() {
5918 // The object a link goes through is in a temporary directory and is gone before `make`
5919 // reads any of this, so the rule that named it would be a rule for a file that is never
5920 // there. The target and the file are both the `-o`, which is the executable.
5921 let (opts, plan) = compile(&["-MD", "sub/a.c", "-o", "prog"]);
5922 assert_eq!(plan.output.as_deref(), Some("prog"));
5923 assert_eq!(deps::default_target("sub/a.c", deps_target_output(&opts, &plan)), "prog");
5924 assert_eq!(
5925 deps::default_file(&opts.deps, "sub/a.c", plan.output.as_deref()).as_deref(),
5926 Some("prog.d")
5927 );
5928 }
5929
5930 #[test]
5931 fn the_plan_keeps_the_output_name_because_the_rule_is_written_from_it() {
5932 let (_, plan) = compile(&["-MMD", "-c", "sub/a.c", "-o", "obj/x.o"]);
5933 assert_eq!(plan.output.as_deref(), Some("obj/x.o"));
5934 let (_, plan) = compile(&["-MMD", "-c", "sub/a.c"]);
5935 assert_eq!(plan.output, None);
5936 }
5937
5938 #[test]
5939 fn usage_fits_on_a_screen() {
5940 // Not a style preference. A help text that scrolls is one nobody reads, and this is
5941 // the cheapest way to keep it honest as flags accumulate. The number goes up only when
5942 // a family of flags arrives that has nowhere to share a line, which the two pass gates
5943 // were and which the two fuel flags and `-fsafety=` now are, and it goes up by exactly
5944 // the lines that family took. The four it went up by last are the flags a build system
5945 // passes without being asked to: how much to say, what machine to generate for, threads,
5946 // and the questions `configure` asks before it compiles anything. The one it went up by
5947 // last is the second line of `--emit`, whose kinds are a family that has now outgrown
5948 // one line and has nowhere else to go. The two it went up by last are the dependency
5949 // family, which is eight flags that share nothing with anything above them. The one it
5950 // went up by last is the four spellings of position independent code, which every
5951 // configure script writes and which could only have shared the link line, and that line
5952 // is already four characters short of the limit. The two it went up by last are the rest
5953 // of the include family, which is six more flags that change where a header is looked for
5954 // and two that name a header outright. The one it went up by last is the pair that keeps
5955 // the intermediate files and times the steps, which belong next to the two flags above
5956 // them that are also about watching a compilation rather than changing one. The two it
5957 // went up by last are the section flags and the visibility flag, which are what a build
5958 // that cares about the size of what it ships and about which names it exports writes, and
5959 // the second of them was already taken and only missing from here. The one it went up by
5960 // last is the stack protector, which is four spellings of one question and which every
5961 // distribution puts on every command line it issues, so a build that reads this list
5962 // looking for it and does not find it has to go and read the specification instead. The one
5963 // it went up by last is the profiler, which is two spellings of the request and two of
5964 // where the call goes, and which is about watching a program run rather than about what is
5965 // generated, so it shares its subject with nothing above it. The one it went up by last is
5966 // the room a function opens with for something to be written over it later, which takes an
5967 // argument of its own shape and is what a kernel build asks for, so it fits beside the
5968 // profiler and nothing else. The one it went up by last is what overflows rather than being
5969 // undefined, which is three spellings of two questions and which a kernel build and a great
5970 // deal of code written before the standard settled both pass. The one it went up by last is
5971 // the other answer to the first of those questions, which could not share the line because
5972 // what it asks for is the opposite of what the flags on that line ask for. The one it went
5973 // up by last is the split of the line that lists what this compiler does anyway into that
5974 // and what it assumes anyway, which are two different claims that were sharing a line until
5975 // the second of them got a second flag and the line stopped fitting. The one it went up by
5976 // last is the three flags that change the ABI rather than the code, which have to be given
5977 // to every file in a program or none of them and which therefore belong somewhere a person
5978 // reading this list will see them. The one it went up by last is the floating point group,
5979 // which is two lines rather than one because the first of them is a choice this compiler
5980 // records and the rest are claims about what it does anyway, and putting a real setting on
5981 // the same line as three flags that change nothing would be misleading about both. The one
5982 // it went up by last is the flag that says a write has to stay inside the member it names,
5983 // which is a setting rather than a claim and so cannot share the line above it, that being
5984 // the one that picks a tier. The two it went up by last are the prefix mapping family,
5985 // which is four flags whose whole job is to keep a build's output the same from two
5986 // different directories, and which a person chasing a reproducible build comes here
5987 // looking for by name. The one it went up by last is how the debug sections are compressed
5988 // and whether they go in a file of their own, which are two questions about the shape of
5989 // the debug output, where the line above them is about how much of it there is. The one it
5990 // went up by last is the `restrict` contract, which is a setting for the same reason the
5991 // flag that keeps a write inside its member is and which is the check a person who has been
5992 // bitten by a vectorizer comes here looking for. The one it went up by last is link time
5993 // optimization, which is a whole optimization rather than a flag and which says so on its
5994 // own line, because a build that passes it and reads this looking for what it got is
5995 // asking a question no other line here answers. The one it went up by last is the sysroot,
5996 // which is the question somebody asks when a cross build read a file nobody expected, and
5997 // which has no room on the line above it because the answers there are a path each and this
5998 // one is the root all of them are under. The one it went up by last is what is inside that
5999 // root and where each of it came from, which is a question about a whole tree rather than
6000 // about a path and which is long enough on its own that it could not have shared a line with
6001 // anything. The one it went up by last is the profile family, which splits down the middle
6002 // where no other family here does, so the line has to name the half that is taken and the
6003 // half that is refused or it would be read as taking both. The one it went up by last is
6004 // the sanitizers, which are what somebody reaching for a checked build writes first and
6005 // which belong beside the tier that is the nearest thing here to what they asked for. The
6006 // one it went up by last is the digest of that record, which is the same tree as one number
6007 // and could not share the line above it because that line prints a few hundred lines and
6008 // this one prints sixty four characters, and a reader who wants the short answer is looking
6009 // for it by name rather than reading the long one. The one it went up by last is the
6010 // sysroot fetch, which is the only command here that gets something from somewhere else and
6011 // is therefore the one a person wants to have read before they run it rather than after.
6012 // And the flag beside it that forbids every download, which earns its line by being what a
6013 // build in a sealed environment passes and by meaning something even though an ordinary
6014 // compile downloads nothing either way. The one it went up by last is the other fetch, the
6015 // one behind Microsoft's licence wall, which is a line rather than a paragraph because what
6016 // a person needs from here is that the command exists and that it will not do anything
6017 // until they have read a licence it prints for them.
6018 assert!(USAGE.lines().count() < 73, "usage text has grown past one screen");
6019 }
6020}