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/0.10.46")]
30
31pub mod cache;
32pub mod compile;
33pub mod deps;
34pub mod fetch;
35mod glibc;
36pub mod install;
37pub mod library;
38pub mod link;
39mod map;
40pub mod phase;
41pub mod preprocess;
42pub mod schedule;
43
44use std::fmt::Write as _;
45use std::io::Write as _;
46use std::path::PathBuf;
47
48use rucc_codegen::coverage::{self, Fired};
49use rucc_codegen::pressure::Pressure;
50use rucc_pp::Dependency;
51use rucc_session::{
52 Compress, Control, Dumps, EmitKind, Hook, Options, Pic, PrefixMap, Preinclude, Protector,
53 SaveTemps, Session, Std, Wrapping, runtime,
54};
55use rucc_sysroot::{Manifest, Sysroot};
56use rucc_target::{ObjectFormat, Triple};
57use rucc_tuple::TargetTuple;
58
59use crate::link::LinkOptions;
60
61pub use crate::compile::{Artifact, Compiled, Temps, compile, compile_ir};
62pub use crate::phase::{ArchiveJob, Input, InputKind, Job, LinkJob, Output, Phase, Plan};
63pub use crate::preprocess::{OsFileSystem, Preprocessed, preprocess};
64pub use crate::schedule::Jobs;
65
66/// The compiler's version, taken from the workspace manifest.
67pub const VERSION: &str = env!("CARGO_PKG_VERSION");
68
69/// What the command line asked for.
70#[derive(Debug, Clone, PartialEq, Eq)]
71pub enum Action {
72 /// Print usage and exit successfully.
73 Help,
74 /// Print the version and exit successfully.
75 Version,
76 /// Print one line and exit successfully, which is what the `-dump` and `-print` family do.
77 ///
78 /// A build system asks these before it compiles anything, and what it does with the answer
79 /// is paste it into a path or into another command line, so each one is a single line with
80 /// no decoration around it.
81 Print(String),
82 /// Print the resolved configuration and exit successfully.
83 PrintConfig(Box<Options>),
84 /// Print the passes the level will run and exit successfully.
85 PrintPipeline(Box<Options>),
86 /// Print the phase plan and the link line and exit successfully, which is `-###`.
87 PrintPlan {
88 /// The resolved options, which is what says what the link line is for.
89 opts: Box<Options>,
90 /// What to do to each input, and in what order.
91 plan: Box<Plan>,
92 /// What the command line said about linking.
93 link: Box<LinkOptions>,
94 },
95 /// `--fetch <tuple>`, which gets the sysroot this release pins for a target and installs it.
96 ///
97 /// The only action in this compiler that may run another program to move bytes onto the
98 /// machine, which is `spec/cross-compile/13-distribution.md` section 13.8's rule rather than a
99 /// property of how this happens to be written: a compilation has no branch that reaches it.
100 Fetch {
101 /// The artifact, from the table in [`rucc_sysroot::artifact`]. Resolved here rather than where the
102 /// work happens, so that a target nothing is pinned for is a refusal from the parser like
103 /// every other thing a command line can ask for and not have.
104 what: &'static rucc_sysroot::Pinned,
105 /// The target, which names the directory under the cache the tree is installed at and is
106 /// checked against the record inside the artifact.
107 target: TargetTuple,
108 /// Where the cache is, read where everything else that needs it reads it.
109 cache: PathBuf,
110 },
111 /// Compile the given inputs.
112 Compile {
113 /// The resolved options.
114 opts: Box<Options>,
115 /// What to do to each input, and in what order.
116 plan: Box<Plan>,
117 /// What the command line said about linking.
118 link: Box<LinkOptions>,
119 /// How many translation units to compile at once.
120 jobs: Jobs,
121 /// Whether `-v` asked for the plan to be printed while it runs.
122 verbose: bool,
123 /// What is worth saying about the command line before anything is compiled, printed as
124 /// warnings and once for the whole run rather than once per file.
125 ///
126 /// These are not diagnostics. A diagnostic is about a piece of source and has a span to
127 /// point at, and these are about the way two flags were combined, so there is nothing to
128 /// point at and nowhere below the driver that knows both halves. `-w` does not reach them
129 /// for the same reason it does not reach a refusal from the parser.
130 notes: Vec<String>,
131 },
132}
133
134/// Why a command line was rejected.
135#[derive(Debug, Clone, PartialEq, Eq)]
136pub struct CliError {
137 /// The message, lowercase and without a trailing period, in the same shape as any other
138 /// diagnostic.
139 pub message: String,
140}
141
142impl std::fmt::Display for CliError {
143 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
144 f.write_str(&self.message)
145 }
146}
147
148impl std::error::Error for CliError {}
149
150fn err(message: impl Into<String>) -> CliError {
151 CliError { message: message.into() }
152}
153
154/// The two halves of one prefix mapping flag's argument, where `flag` includes its trailing `=`.
155///
156/// The split is at the last `=` in what follows the flag, not the first, which is gcc's rule and
157/// the only one that lets a directory whose name contains an `=` be the old half. It also means
158/// `-fmacro-prefix-map=a=b=c` rewrites `a=b` to `c` rather than `a` to `b=c`, which looks like a
159/// trap until you notice the alternative traps the far more common case.
160fn rewrite<'a>(arg: &'a str, flag: &str) -> Result<(&'a str, &'a str), CliError> {
161 let rest = &arg[flag.len()..];
162 PrefixMap::split(rest).ok_or_else(|| {
163 let flag = flag.trim_end_matches('=');
164 err(format!(
165 "`{rest}` is not a rewrite for `{flag}`, which is an old prefix, an `=` and a new one"
166 ))
167 })
168}
169
170/// A question the command line asked instead of asking for a compilation.
171///
172/// These are answered after the loop rather than where they are read, because every one of them
173/// is about the target or about the library search and the last word on both is the end of the
174/// command line.
175enum Query {
176 /// `-dumpmachine`, the triple.
177 Machine,
178 /// `-dumpversion` and `-dumpfullversion`, which are the same three numbers here.
179 Version,
180 /// `-print-multiarch`, the directory name a distribution files this target under.
181 Multiarch,
182 /// `-print-search-dirs`, in the three lines GCC prints.
183 SearchDirs,
184 /// `-print-sysroot`, the root the headers and the libraries are read under.
185 Sysroot,
186 /// `-print-sysroot-provenance`, what is in that root and where each of it came from.
187 SysrootProvenance,
188 /// `-print-sysroot-digest`, the one number that names all of it.
189 SysrootDigest,
190 /// `-print-file-name=<name>`, the full path of a library file.
191 FileName(String),
192 /// `-print-prog-name=<name>`, the full path of a program.
193 ProgName(String),
194 /// `-print-libgcc-file-name`, which is `-print-file-name=libgcc.a` under another spelling.
195 Libgcc,
196}
197
198/// Usage text.
199///
200/// Deliberately short. `spec/04-driver-and-cli.md` puts the full flag reference in the
201/// manual page, because a `--help` nobody can read in one screen is a `--help` nobody reads.
202pub const USAGE: &str = "\
203rucc, an optimizing C compiler
204
205usage: rucc [options] file...
206
207options:
208 -c compile and assemble, do not link
209 -S compile only, emit assembly
210 -E preprocess only
211 -o <file> write output to <file>, or to standard output for -
212 -D <name>[=<value>], -U <name> define a macro, or undefine one after every -D
213 -I <dir> add <dir> to the include search path
214 -iquote -isystem -idirafter <dir> the other chains, -nostdinc drops ours
215 -I-, -iprefix <p>, -iwithprefix[before] <dir> the older spellings of those
216 -include <file>, -imacros <file> read <file> first, the second for its macros only
217 --sysroot=<dir> look for the library's headers under <dir>, -isysroot too
218 -P, -dM with -E: leave out the markers, or dump the macros
219 -M -MM -MD -MMD write a make rule for the source, the last two compile as well
220 -MF <file> -MT <t> -MQ <t> -MP where the rule goes, what it builds, targets with no recipe
221 -std=<dialect> c89 through c23, and the gnu spellings
222 -fgnuc-version=<v> the GCC release to claim, default 7.0.0
223 -x <lang> treat later inputs as <lang>, or none to stop
224 -O<level> optimize: 0, 1, 2, 3, s, z
225 -fsafety=<tier> check memory safety: off, detect, enforce, kernel
226 -f[no-]sanitize=<what> the negative is taken, the positive is refused by name
227 -f[no-]safety-subobject a write has to stay inside the member it names
228 -f[no-]safety-restrict two restrict pointers of one block may not meet
229 -f<pass> -fno-<pass> -fdump-ir=<what> -fopt-info[-<kind>][=FILE]
230 -fpass-fuel=<pass>=<n>, -fpass-fuel-global=<n> stop a pass, or all of them, after n
231 -fdisable-<pass>[=<funcs>], -fenable-<pass>[=<funcs>] run a pass on some functions only
232 -g -g0 -gdwarf-5, -fno-omit-frame-pointer, -mno-red-zone debug info, frame pointer, red zone
233 -gz[=none|zlib|zlib-gnu|zstd] -gno-split-dwarf compress debug sections, one file not two
234 -flto[=auto|jobserver|<n>] -fno-lto -ffat-lto-objects read, and not done yet
235 -fprofile-use[=<path>] -fprofile-dir=<dir> read too, where -fprofile-generate is refused
236 -f[no-]stack-protector[-strong|-all], -f[no-]stack-clash-protection, -fcf-protection=<edges>
237 -ffunction-sections -fdata-sections a section per function or variable, for --gc-sections
238 -fvisibility=<what> default, hidden, internal or protected, when nothing in the source said
239 -l<name>, -L <dir>, -B <dir> link a library, where to look for one, where our own tools are
240 -fPIC -fpic -fPIE -fpie, -fno-common, -pipe what it does anyway
241 -f[no-]strict-aliasing, -f[no-]delete-null-pointer-checks what it assumes anyway
242 -static -shared -pie -no-pie -nostdlib -nostartfiles -nodefaultlibs -rdynamic -s how to link
243 -Wl,<arg>, -Xlinker <arg>, -fuse-ld=<name> hand an argument to the linker, or pick one
244 -Werror -pedantic -pedantic-errors -w how much to say, and whether it is fatal
245 -m64 -march= -mtune= -mcpu= -mabi= -mcmodel= what machine to generate for
246 -pg -p, -mfentry -mno-fentry call a profiler on the way in, and where that call goes
247 -fpatchable-function-entry=<n>[,<m>] room at the top of every function to patch later
248 -fwrapv, -fwrapv-pointer, -fno-strict-overflow signed or pointer overflow wraps
249 -ftrapv signed overflow stops the program instead
250 -f[no-]signed-char, -f[no-]unsigned-char, -f[no-]short-enums change the ABI
251 -ffp-contract=<how> fuse a multiply and an addition: fast, on or off
252 -fexcess-precision=<how>, -f[no-]rounding-math, -f[no-]trapping-math what may be folded
253 -ffile-prefix-map=<old>=<new> rewrite that front of every path we put in the output
254 -fmacro-prefix-map= -fdebug-prefix-map= -fprofile-prefix-map= the same, one output each
255 -pthread build for more than one thread, and link the library for it
256 -dumpmachine -dumpversion -print-multiarch -print-search-dirs what this compiler is
257 -print-file-name=<name> -print-prog-name=<name> where a file or a program is
258 -print-sysroot the root the headers and the libraries are read under
259 -print-sysroot-provenance every input under it, where it came from and its licence
260 -print-sysroot-digest the sha256 of that record, which names the whole sysroot in one line
261 --fetch <tuple> get the sysroot this release pins for <tuple> and install it in the cache
262 --offline never download anything, which a compilation never does anyway
263 -j[n] compile n translation units at once, default all
264 -v, -### print each phase as it runs, or without running any
265 -save-temps[=cwd|obj], -time keep the .i and the .s, say how long each step took
266 --target=<triple> generate code for <triple>
267 --emit=<kind> exe, obj, archive, asm, preprocessed, tast, ir, mir-final,
268 safety-summary, type-granules
269 --print-config, --print-pipeline print the configuration or the pipeline, and exit
270 --version print the version and exit
271 -h, --help print this message and exit
272
273See spec/04-driver-and-cli.md for the full flag reference.
274";
275
276/// The argument of a flag that may be joined to it or may be the next word.
277///
278/// `-DFOO` and `-D FOO` are the same thing, and `at` is where the flag's own letters end.
279fn joined_or_next(
280 arg: &str,
281 at: usize,
282 args: &[String],
283 i: &mut usize,
284) -> Result<String, CliError> {
285 if arg.len() > at {
286 return Ok(arg[at..].to_owned());
287 }
288 let next = args.get(*i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
289 *i += 1;
290 Ok(next.clone())
291}
292
293/// Every name that may follow `-fsanitize=`, which is gcc 16's list and three of this compiler's
294/// own.
295///
296/// The three are on it because `spec/07-types-and-semantics.md` section 7.7 already promises them:
297/// each undefined behaviour this compiler exploits is listed there with the check that detects it,
298/// and `alias`, `restrict` and `memory` are checks gcc has no spelling for. gcc refuses `memory`
299/// outright, since the sanitizer of that name is clang's. A name being here means it is a name
300/// rather than a typo, and nothing more than that: every one of them is refused after the loop,
301/// because none of them is implemented.
302///
303/// `all` is deliberately absent. gcc takes it only in the negative, so it is handled where each of
304/// those two spellings is read rather than by being on this list.
305const SANITIZERS: [&str; 34] = [
306 "address",
307 "kernel-address",
308 "hwaddress",
309 "kernel-hwaddress",
310 "pointer-compare",
311 "pointer-subtract",
312 "thread",
313 "leak",
314 "undefined",
315 "shift",
316 "shift-base",
317 "shift-exponent",
318 "integer-divide-by-zero",
319 "unreachable",
320 "vla-bound",
321 "null",
322 "return",
323 "signed-integer-overflow",
324 "bounds",
325 "bounds-strict",
326 "alignment",
327 "object-size",
328 "float-divide-by-zero",
329 "float-cast-overflow",
330 "nonnull-attribute",
331 "returns-nonnull-attribute",
332 "bool",
333 "enum",
334 "vptr",
335 "pointer-overflow",
336 "builtin",
337 "alias",
338 "restrict",
339 "memory",
340];
341
342/// Parses a command line, without the program name.
343///
344/// # Errors
345///
346/// Returns the message to print when the arguments do not name a compilation this compiler
347/// can attempt.
348pub fn parse_args(args: &[String]) -> Result<Action, CliError> {
349 let host = Triple::host()
350 .ok_or_else(|| err("this host is not a supported target and no --target was given"))?;
351 let mut opts = Options::new(host);
352 let mut inputs: Vec<Input> = Vec::new();
353 let mut print_config = false;
354 let mut print_pipeline = false;
355 let mut print_plan = false;
356 let mut verbose = false;
357 let mut jobs = Jobs::default();
358 let mut nostdinc = false;
359 let mut sysroot: Option<PathBuf> = None;
360 // What the command line is worth warning about, filled in after the loop rather than during it,
361 // because every question of this kind is about two flags and the last word on both of them is
362 // the end of the loop.
363 let mut notes: Vec<String> = Vec::new();
364 // The whole ten field target, kept beside the three field one because `--target=` can pin a
365 // libc version and `Triple` has nowhere to put it. It decides `__GLIBC_MINOR__` and nothing
366 // else today, and `None` is a command line that named no target, which is this machine.
367 let mut pinned: Option<TargetTuple> = None;
368 let mut output = None;
369 let mut link = LinkOptions::default();
370 let mut query: Option<Query> = None;
371 // What `--fetch` named, and whether `--offline` forbade it. Both are weighed after the loop
372 // because either can be written after the other.
373 let mut fetch: Option<String> = None;
374 let mut offline = false;
375 let mut threads = false;
376 // Which sanitizers are still asked for by the end of the command line. Accumulated across the
377 // loop rather than answered where it was read, because `-fno-sanitize=` turns one off and a
378 // build that asks for a check and then takes it back has asked for nothing. What happens to a
379 // set that is not empty is decided after the loop.
380 let mut sanitizers: Vec<&str> = Vec::new();
381 // `-x` applies to inputs that come after it and stays in effect until the next one, which
382 // is why it is tracked across the loop rather than attached to a single argument.
383 let mut forced: Option<InputKind> = None;
384 // What `-iprefix` last said, stuck on the front of every later `-iwithprefix`. It applies to
385 // the flags after it and not the ones before, so a command line may set it more than once.
386 // GCC's default is its own installed header directory with the last component taken off,
387 // which is a path a cross compiler's build system knows and passes; there is no equivalent
388 // here, so with no `-iprefix` the prefix is nothing and `-iwithprefix` names a directory
389 // outright.
390 let mut iprefix = String::new();
391
392 let mut i = 0;
393 while i < args.len() {
394 let arg = args[i].as_str();
395 i += 1;
396 match arg {
397 "-h" | "--help" => return Ok(Action::Help),
398 "--version" => return Ok(Action::Version),
399 // The sysroot fetch, which is weighed after the loop rather than acted on here, because
400 // `--offline` written after it has to be able to forbid it. Both spellings, since a
401 // flag that takes a tuple gets written both ways and neither is a guess at what the
402 // other meant.
403 "--fetch" => {
404 let value = args
405 .get(i)
406 .ok_or_else(|| err("--fetch requires the target to get a sysroot for"))?;
407 i += 1;
408 fetch = Some(value.clone());
409 }
410 _ if arg.starts_with("--fetch=") => {
411 fetch = Some(arg["--fetch=".len()..].to_owned());
412 }
413 // Accepted on any command line and only ever read by the fetch, because an ordinary
414 // compile downloads nothing with or without it. So this flag takes nothing away today,
415 // which is the property section 13.2 asks for rather than an omission: a build that
416 // passes it is saying what it expects of this compiler, and what it expects is already
417 // true.
418 "--offline" => offline = true,
419 "--print-config" => print_config = true,
420 "--print-pipeline" => print_pipeline = true,
421 "-###" => print_plan = true,
422 "-v" => verbose = true,
423 // The files a compilation goes through, kept rather than thrown away. The bare
424 // spelling means `=obj` and not `=cwd`, which is not what the manual says and is what
425 // gcc 16 does; `SaveTemps::Object` carries the measurement.
426 "-save-temps" => opts.save_temps = SaveTemps::Object,
427 _ if arg.starts_with("-save-temps=") => {
428 opts.save_temps = arg["-save-temps=".len()..].parse().map_err(err)?;
429 }
430 // How long each step took. A misspelling of this is worth rejecting rather than
431 // ignoring, since a run that says nothing looks like a compilation that took no time.
432 "-time" => opts.time = true,
433 "-c" => opts.emit = EmitKind::Object,
434 "-S" => opts.emit = EmitKind::Asm,
435 "-E" => opts.emit = EmitKind::Preprocessed,
436 "-g" => opts.debug_info = true,
437 // GCC's own levels of how much debug information to write. Zero is none and every
438 // other number is some, and this compiler has one amount, so the numbers above zero
439 // all mean the same thing here. `-ggdb` is the same flag asking for whatever the
440 // debugger on the machine prefers, which is what we emit anyway.
441 "-g0" => opts.debug_info = false,
442 "-g1" | "-g2" | "-g3" | "-ggdb" | "-ggdb1" | "-ggdb2" | "-ggdb3" => {
443 opts.debug_info = true;
444 }
445 // The version of DWARF to write. We write DWARF 5 and nothing else, so a build that
446 // asks for another version is told rather than handed a file it cannot read.
447 "-gdwarf" | "-gdwarf-5" => opts.debug_info = true,
448 _ if arg.starts_with("-gdwarf-") => {
449 return Err(err(format!(
450 "{arg}: this compiler writes DWARF 5 and no other version, see \
451 spec/11-debug-info.md"
452 )));
453 }
454 // Whether the debug information goes in a file of its own beside the object. gcc
455 // writes that `.dwo` whether or not it found anything to put in it, which means a
456 // build system that declares the file as an output gets one and a make rule that
457 // depends on it fires. Refused for that reason rather than taken: section 4.1 takes a
458 // flag that changes nothing and refuses one that changes what is produced, and a file
459 // that does not appear is the plainest change of that kind there is. The negative
460 // spelling is taken, because putting it all in the object is what happens anyway.
461 "-gno-split-dwarf" => {}
462 "-gsplit-dwarf" => {
463 return Err(err(format!(
464 "{arg}: this compiler writes no separate `.dwo` file, and a build that \
465 expects one beside each object would wait for a file that never arrives, \
466 see spec/11-debug-info.md"
467 )));
468 }
469 // How the debug sections are compressed. There are none yet, so every answer produces
470 // the same bytes and taking the flag promises nothing that is not kept. The value is
471 // still checked, because a typo in a distribution's flags is worth finding when the
472 // compiler reads it rather than when somebody later wonders why nothing got smaller.
473 // Bare `-gz` means `zlib`, which the manual leaves for the reader to discover.
474 "-gz" => opts.compress = Compress::Zlib,
475 _ if arg.starts_with("-gz=") => {
476 let how = &arg["-gz=".len()..];
477 opts.compress = how.parse().map_err(|()| {
478 err(format!(
479 "`{how}` is not a way to compress debug sections, which is none, zlib, \
480 zlib-gnu or zstd"
481 ))
482 })?;
483 }
484 "-Werror" => opts.warnings_are_errors = true,
485 // Nothing that is not fatal is said at all. Read at the one place a diagnostic goes
486 // through rather than here, so that a warning `-w` dropped is not counted either.
487 "-w" => opts.warnings = false,
488 "-pedantic-errors" => {
489 opts.pedantic = true;
490 opts.warnings_are_errors = true;
491 }
492 "-P" => opts.line_markers = false,
493 // The dependency family, which section 4.4 calls required because every build system
494 // that generates its own makefiles asks for it. The two that end in `D` write a file
495 // beside the object and let the compilation happen, and the two that do not write to
496 // standard output and stop after it. Nothing here turns the system headers back on
497 // once a flag has turned them off, which is GCC's behaviour and is why `-MM -M` is
498 // `-MM`: the flag asking for fewer of them is the one with something to say.
499 "-M" => {
500 opts.deps.emit = true;
501 opts.deps.instead_of_compiling = true;
502 }
503 "-MM" => {
504 opts.deps.emit = true;
505 opts.deps.instead_of_compiling = true;
506 opts.deps.system_headers = false;
507 }
508 "-MD" => opts.deps.emit = true,
509 "-MMD" => {
510 opts.deps.emit = true;
511 opts.deps.system_headers = false;
512 }
513 "-MP" => opts.deps.phony = true,
514 // These three take a word and only in the separated form, which is how GCC spells
515 // them and how every build system writes them.
516 "-MF" | "-MT" | "-MQ" => {
517 let value =
518 args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
519 i += 1;
520 match arg {
521 "-MF" => opts.deps.file = Some(value.clone()),
522 // The whole of the difference between the two. `-MT` is for a build that has
523 // already escaped what it is passing, and `-MQ` is for one that has a name
524 // and wants it to arrive as that name.
525 "-MT" => opts.deps.targets.push(value.clone()),
526 _ => opts.deps.targets.push(deps::escaped(value)),
527 }
528 }
529 // The questions a build system asks before it compiles anything. Answered after the
530 // loop, because each one is about the target or the library search and the command
531 // line has not finished saying what those are.
532 "-dumpmachine" => query = Some(Query::Machine),
533 "-dumpversion" | "-dumpfullversion" => query = Some(Query::Version),
534 "-print-multiarch" => query = Some(Query::Multiarch),
535 "-print-search-dirs" => query = Some(Query::SearchDirs),
536 "-print-sysroot" => query = Some(Query::Sysroot),
537 // Both spellings, because this one is ours rather than GCC's and our own documents
538 // write it both ways: section 13.5 of `spec/cross-compile/13-distribution.md` gives it
539 // two dashes like the other flags we invented, and document 12's table gives it one
540 // like the `-print-` family it sits in. A person who reads either and types what it
541 // says is right, so neither is refused.
542 "-print-sysroot-provenance" | "--print-sysroot-provenance" => {
543 query = Some(Query::SysrootProvenance);
544 }
545 "-print-sysroot-digest" | "--print-sysroot-digest" => {
546 query = Some(Query::SysrootDigest);
547 }
548 "-print-libgcc-file-name" => query = Some(Query::Libgcc),
549 _ if arg.starts_with("-print-file-name=") => {
550 query = Some(Query::FileName(arg["-print-file-name=".len()..].to_owned()));
551 }
552 _ if arg.starts_with("-print-prog-name=") => {
553 query = Some(Query::ProgName(arg["-print-prog-name=".len()..].to_owned()));
554 }
555 // A program built to run in more than one thread. On every platform this compiler
556 // targets that is a macro the library's headers read and one more library on the
557 // link line, and the library is added after the loop so that it lands after the
558 // objects that refer to it.
559 "-pthread" | "-pthreads" => {
560 opts.defines.push("_REENTRANT".to_owned());
561 threads = true;
562 }
563 "-ansi" => {
564 opts.std = Std::C89;
565 opts.gnu_extensions = false;
566 }
567 // `-Wpedantic` is the same flag under the name the `-W` family gives it, which is
568 // the spelling a build system that groups its warning flags tends to write.
569 "-pedantic" | "-Wpedantic" => opts.pedantic = true,
570 // Both directions, because a build that needs this for one directory turns it back
571 // off for the next one rather than leaving it on for the whole tree.
572 "-fpermissive" => opts.permissive = true,
573 "-fno-permissive" => opts.permissive = false,
574 "-ffreestanding" => opts.hosted = false,
575 "-fhosted" => opts.hosted = true,
576 "-fno-builtin" => opts.builtins = false,
577 "-fbuiltin" => opts.builtins = true,
578 // The C89 dialects are under GNU's reading whatever this says, so turning it off
579 // there is turning off something the dialect asked for, which is accepted and does
580 // nothing. gcc refuses that command line, and there is nothing it could have meant.
581 "-fgnu89-inline" => opts.gnu89_inline = true,
582 "-fno-gnu89-inline" => opts.gnu89_inline = false,
583 // Both directions of each, because a build system that wants one of these usually
584 // writes it beside the flag that turns it back off for one directory.
585 "-fno-omit-frame-pointer" => opts.frame_pointer = true,
586 "-fomit-frame-pointer" => opts.frame_pointer = false,
587 // Both directions again, for the same reason, and a third answer for a command line
588 // that wrote neither: see `reorder_blocks` in `rucc_session`.
589 "-freorder-blocks" => opts.reorder_blocks = Some(true),
590 "-fno-reorder-blocks" => opts.reorder_blocks = Some(false),
591 // gcc's name for the scheduler that runs after the registers are handed out, which is
592 // the only one rucc has: see `schedule_insns` in `rucc_session`. gcc also takes
593 // `-fschedule-insns` for the pass before allocation, and taking that one here would be
594 // a flag that says a pass ran when none did.
595 "-fschedule-insns2" => opts.schedule_insns = Some(true),
596 "-fno-schedule-insns2" => opts.schedule_insns = Some(false),
597 "-mno-red-zone" => opts.red_zone = false,
598 "-mred-zone" => opts.red_zone = true,
599 // Four flags rather than one with an argument, which is how gcc spells them and how
600 // every build line writes them. Last one wins, because a package build puts
601 // `-fstack-protector-strong` in its global flags and a directory that cannot have one
602 // turns it back off on the line after.
603 "-fno-stack-protector" | "-fno-stack-protector-all" | "-fno-stack-protector-strong" => {
604 opts.protector = Protector::None;
605 }
606 "-fstack-protector" => opts.protector = Protector::Buffers,
607 "-fstack-protector-strong" => opts.protector = Protector::Strong,
608 "-fstack-protector-all" => opts.protector = Protector::All,
609 // The other half of what a hardened build asks for, and it is a question about the
610 // frame rather than about the function, so it is a switch rather than a level.
611 "-fstack-clash-protection" => opts.stack_clash = true,
612 "-fno-stack-clash-protection" => opts.stack_clash = false,
613 // The third of them, and the one that is a question with an argument rather than a
614 // family of spellings, because what it asks about is which of the two edges of a
615 // control flow transfer is checked. Bare is both of them, which is what gcc does.
616 "-fcf-protection" => opts.control = Control::Full,
617 "-fno-cf-protection" => opts.control = Control::None,
618 // Two spellings of the same request, which is what gcc has as well. `-p` was the older
619 // profiler and `-pg` the one that also recorded who called whom, and on every platform
620 // this compiler targets there is now one hook and both ask for it.
621 "-pg" | "-p" => {
622 opts.profile = true;
623 link.profile = true;
624 }
625 // Accepted on their own and doing nothing on their own, which is gcc's behaviour: they
626 // say where the call goes and a command line that asked for no call has nowhere to put
627 // one. That matters because a build system that sets `-mfentry` globally and `-pg` per
628 // directory is a build system that would otherwise fail on every other directory.
629 "-mfentry" => opts.hook = Hook::Early,
630 "-mno-fentry" => opts.hook = Hook::Late,
631 // GCC drops its own include directory along with the system ones, because its
632 // headers are half of a pair with the library's and half a pair is worse than
633 // none. A build that passes this is supplying the whole set itself.
634 "-nostdinc" => nostdinc = true,
635 "-o" => {
636 output = Some(args.get(i).ok_or_else(|| err("-o requires an argument"))?.clone());
637 i += 1;
638 }
639 // The flags that take a directory only in the separated form. GCC spells them
640 // this way and nothing writes `-iquotedir`, so accepting the joined form would
641 // mean guessing at a path that starts with the flag's own letters.
642 // Apple's spelling of `--sysroot`, and the one its own build systems pass. The
643 // two mean the same thing here: the configured directories are under there rather
644 // than under the root.
645 "-isysroot" => {
646 let dir = args.get(i).ok_or_else(|| err("-isysroot requires an argument"))?;
647 i += 1;
648 sysroot = Some(PathBuf::from(dir));
649 }
650 "-iquote" | "-isystem" | "-idirafter" => {
651 let dir = args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
652 i += 1;
653 match arg {
654 "-iquote" => opts.search.push_quote(dir.clone()),
655 "-isystem" => opts.search.push_system(dir.clone()),
656 _ => opts.search.push_after(dir.clone()),
657 }
658 }
659 "-iprefix" => {
660 iprefix = args.get(i).ok_or_else(|| err("-iprefix requires an argument"))?.clone();
661 i += 1;
662 }
663 // Where GCC puts these is not where its manual says it puts them, and this is the
664 // measured answer rather than the documented one: `-iwithprefix` lands in the
665 // `-isystem` slot and not the `-idirafter` slot, and `-iwithprefixbefore` lands in
666 // the `-I` slot. A cross build that uses them is relying on the behaviour, since
667 // that is the compiler it was developed against.
668 "-iwithprefix" | "-iwithprefixbefore" => {
669 let dir = args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
670 i += 1;
671 let dir = format!("{iprefix}{dir}");
672 if arg == "-iwithprefix" {
673 opts.search.push_system(dir);
674 } else {
675 opts.search.push_bracket(dir);
676 }
677 }
678 "-include" | "-imacros" => {
679 let name = args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
680 i += 1;
681 opts.preincludes
682 .push(Preinclude { name: name.clone(), macros_only: arg == "-imacros" });
683 }
684 // The flag `-iquote` was introduced to replace, still passed by build systems old
685 // enough to predate the replacement. It is not a directory: it says that every `-I`
686 // so far is for quoted includes only, and that a quoted include stops looking next
687 // to the file that wrote it.
688 "-I-" => opts.search.split_quote_chain(),
689 "-x" => {
690 let lang = args.get(i).ok_or_else(|| err("-x requires an argument"))?;
691 i += 1;
692 forced = if lang == "none" {
693 None
694 } else {
695 Some(InputKind::from_x_arg(lang).map_err(|e| err(format!("{e}")))?)
696 };
697 }
698 // Not a GCC flag. spec/03-architecture.md section 3.5 compiles several
699 // translation units in one process rather than making the build system fork, and
700 // section 3.8's determinism check compares `-j1` against `-j16`, so the knob has
701 // to exist and has to be spelled the way `make` spells it.
702 // `-DFOO`, `-D FOO` and the same for `-U` and `-I`. Both forms are in wide use
703 // and a build system may produce either, so both are read here rather than
704 // being normalised by whatever generated the command line.
705 _ if arg.starts_with("-D") => {
706 let value = joined_or_next(arg, 2, args, &mut i)?;
707 opts.defines.push(value);
708 }
709 _ if arg.starts_with("-U") => {
710 let value = joined_or_next(arg, 2, args, &mut i)?;
711 opts.undefines.push(value);
712 }
713 _ if arg.starts_with("-I") => {
714 let dir = joined_or_next(arg, 2, args, &mut i)?;
715 opts.search.push_bracket(dir);
716 }
717 _ if arg.starts_with("-std=") => {
718 let name = &arg["-std=".len()..];
719 let (std, gnu) = Std::from_flag(name)
720 .ok_or_else(|| err(format!("unknown dialect `{name}`, see --help")))?;
721 opts.std = std;
722 opts.gnu_extensions = gnu;
723 }
724 // Section 4.5. The claim decides which half of glibc's `sys/cdefs.h` we are
725 // handed, so a differential run that does not set it is comparing two compilers
726 // that believe they are different compilers.
727 // GCC packs these into one flag, so `-dDI` is two of them. Letters in the family
728 // that we have not written yet are accepted and ignored, because a dump is a
729 // debugging aid and a build that asks for one should still compile. A letter
730 // outside the family falls through to the unknown option error, which is what
731 // keeps `-dumpversion` from being read as a dump of nothing.
732 _ if Dumps::is_family(arg) => {
733 opts.dumps.add(&arg[2..]);
734 }
735 // One name at a time, which is what a build that means its own `memcpy` and the
736 // library's everything else writes. The name is not checked against a list, because
737 // the flag is about what the program means by a name and a program is allowed to mean
738 // something by a name this compiler has never heard of.
739 _ if arg.starts_with("-fno-builtin-") => {
740 opts.no_builtin.push(arg["-fno-builtin-".len()..].to_owned());
741 }
742 _ if arg.starts_with("-fgnuc-version=") => {
743 let v = &arg["-fgnuc-version=".len()..];
744 opts.gnuc = v.parse().map_err(err)?;
745 }
746 // spec/13-gnu-compat.md section 13.3 promises this flag an error that says why rather
747 // than the unknown option one, because a build reaching for it is asking for a feature
748 // and deserves to be told it is not coming rather than told the spelling is wrong.
749 // The negative form is what this compiler does anyway, so it is taken and dropped.
750 "-fnested-functions" => {
751 return Err(err(
752 "nested functions are not supported: a call to one goes through a trampoline \
753 written on the stack, which no target that enforces an unexecutable stack \
754 allows",
755 ));
756 }
757 "-fno-nested-functions" => {}
758 // Which of the two links the output is for, which is a real difference and not a
759 // description of what happens anyway. Everything here is position independent either
760 // way, and what these decide is whether a name may be one another object defines or
761 // replaces, because a link that produces an executable puts every name in the same
762 // program and a link that produces a shared library does not.
763 //
764 // It matters that they are accepted at all, whatever they then do. Every autoconf and
765 // cmake build puts `-fPIC` on the compile line, so a compiler that rejects it cannot
766 // be the `CC` of a project that has a configure script, whatever else it can do. That
767 // is how this was found: building SQLite's test fixture stopped on it.
768 "-fPIC" | "-fpic" => opts.pic = Pic::Library,
769 // Not a synonym of the pair above, which is what they were treated as until #756. The
770 // library is the expensive answer and gcc makes it the one that has to be asked for,
771 // so this is also what nothing at all means.
772 "-fPIE" | "-fpie" => opts.pic = Pic::Executable,
773 // A different question from the pair above, and the one every distribution build of a
774 // shared library answers. `-fPIC` decides how an address is reached, and this decides
775 // whether the optimizer may believe a body it can see, because an exported name is one
776 // the dynamic linker may find another definition of first. On by default, which is
777 // gcc's arrangement and is the honest answer, and off is a promise the build makes and
778 // nothing checks.
779 "-fsemantic-interposition" => opts.interposition = true,
780 "-fno-semantic-interposition" => opts.interposition = false,
781 // Two requests rather than one, and the same table answers both, so what decides is
782 // whether either of them is standing. gcc arranges it the same way: the asynchronous
783 // one is the default here and it implies the other, and a line that asks for a table
784 // and against an asynchronous one gets a table.
785 "-fasynchronous-unwind-tables" => opts.async_unwind_tables = true,
786 "-fno-asynchronous-unwind-tables" => opts.async_unwind_tables = false,
787 "-funwind-tables" => opts.unwind_tables = true,
788 "-fno-unwind-tables" => opts.unwind_tables = false,
789 // The other direction is a request, not a description, and it is one this compiler
790 // cannot grant, so it gets the treatment section 13.3 asks for rather than the unknown
791 // option error. Answering it by carrying on would be answering a different question:
792 // the code would still be position independent, which is correct everywhere an
793 // ordinary program runs and is wrong in a kernel, where the flag is written precisely
794 // because there is no loader to fill a global offset table in.
795 "-fno-pic" | "-fno-pie" => {
796 return Err(err(
797 "position dependent code is not supported: an address that may be in another \
798 object is loaded out of the global offset table, and nothing here emits the \
799 absolute form this asks for. Use -no-pie if what you meant was how to link",
800 ));
801 }
802 // A section per function and a section per variable, which is what makes
803 // `--gc-sections` able to drop anything: a linker can leave out a section nothing
804 // reaches and cannot leave out half of one. Both directions are taken, and the off
805 // one is the default rather than a refusal, since a build that writes it is asking
806 // for what happens anyway.
807 "-ffunction-sections" => opts.function_sections = true,
808 "-fno-function-sections" => opts.function_sections = false,
809 "-fdata-sections" => opts.data_sections = true,
810 "-fno-data-sections" => opts.data_sections = false,
811 // Another description of what this compiler does. A file scope declaration with no
812 // initializer is written into `.bss` as an ordinary defined symbol, not offered to the
813 // linker as a common one for it to merge, which is what `-fno-common` asks for and what
814 // gcc has done by default since 10. Nothing in the front end produces `Linkage::Common`
815 // at all.
816 "-fno-common" => {}
817 // What overflows rather than being undefined. Every one of these takes something away
818 // from the optimizer rather than asking it to do anything, which is why the negative
819 // spellings are the interesting ones and the positive spellings are the default.
820 //
821 // `-fno-strict-overflow` is both of the others, which is gcc's own reading of it: its
822 // help text for `-fstrict-overflow` says "negated as -fwrapv -fwrapv-pointer". So it is
823 // written here as the pair rather than kept as a third thing to test everywhere.
824 //
825 // `-ftrapv` is the exception and is the one that asks for something. It is the other
826 // answer to the question `-fwrapv` answers, so the two cannot both hold and each clears
827 // the other, which makes the last one on the command line the one that counts. That is
828 // gcc 16's behaviour and was measured rather than read: `-ftrapv -fwrapv` emits no
829 // checked calls and `-fwrapv -ftrapv` emits them. The positive spelling of the pointer
830 // question is left alone by both, because neither has anything to say about it.
831 "-fwrapv" => {
832 opts.wrapping.signed = true;
833 opts.wrapping.trap = false;
834 }
835 "-fno-wrapv" => opts.wrapping.signed = false,
836 "-fwrapv-pointer" => opts.wrapping.pointer = true,
837 "-fno-wrapv-pointer" => opts.wrapping.pointer = false,
838 "-fno-strict-overflow" => opts.wrapping = Wrapping::ALL,
839 // Which does not clear the checked one, because gcc does not: `-ftrapv
840 // -fstrict-overflow` still emits the calls. It says what is assumed and not what
841 // happens.
842 "-fstrict-overflow" => {
843 opts.wrapping.signed = false;
844 opts.wrapping.pointer = false;
845 }
846 "-ftrapv" => {
847 opts.wrapping.trap = true;
848 opts.wrapping.signed = false;
849 }
850 "-fno-trapv" => opts.wrapping.trap = false,
851 // The two flags that say what a plain `char` is, which is one question with two
852 // spellings each: gcc reads `-fno-signed-char` as `-funsigned-char` and
853 // `-fno-unsigned-char` as `-fsigned-char`, so there are four ways to write two
854 // answers and the last one written wins. Nothing is set until one of them is given,
855 // because the target's own ABI is the answer otherwise and it is not the same answer
856 // everywhere: x86-64 and Apple's arm64 are signed, Linux's arm64 is not.
857 "-fsigned-char" | "-fno-unsigned-char" => opts.char_signed = Some(true),
858 "-funsigned-char" | "-fno-signed-char" => opts.char_signed = Some(false),
859 // And the size of an enumeration, which is the other thing in this group that changes
860 // the ABI rather than the code.
861 "-fshort-enums" => opts.short_enums = true,
862 "-fno-short-enums" => opts.short_enums = false,
863 // And the request, which is the one that cannot be granted. It is a real difference and
864 // not a preference: two files each writing `int g;` link under `-fcommon` and are a
865 // duplicate definition without it, which is the whole reason the flag survives.
866 "-fcommon" => {
867 return Err(err(
868 "a tentative definition is written into .bss as its own symbol here, and \
869 nothing emits the common symbol this asks the linker to merge. Give the \
870 variable a definition in one file and declare it extern in the others",
871 ));
872 }
873 // Both directions of this one are recorded, and what they decide is whether lowering
874 // names the type each access goes through. Turning it off is the front end leaving the
875 // name off rather than a pass being told to ignore one it can see, which is one
876 // condition in one place, and it is the reading that survives link time optimization:
877 // a unit built with the flag off keeps its own answer when its bodies end up in a
878 // module beside bodies that were not.
879 //
880 // Nothing in the pipeline reads those names yet. Layer 3 of the alias analysis does
881 // and is tested, and no pass at any level asks the alias analysis anything today, so
882 // no program compiles differently for having passed this. The flag is wired anyway,
883 // because the change that makes a pass ask is not the change anybody will remember to
884 // wire it in, and a flag that is taken and dropped once the names mean something is
885 // the miscompilation `spec/04-driver-and-cli.md` section 4.1 warns about in as many
886 // words.
887 "-fstrict-aliasing" => opts.strict_aliasing = true,
888 "-fno-strict-aliasing" => opts.strict_aliasing = false,
889 // The same shape of answer for the same reason, and the flag the kernel writes beside
890 // the one above it.
891 //
892 // Nothing here concludes that a pointer is not null from the fact that it was
893 // dereferenced. There is no such conclusion to draw from, because no pass records one:
894 // a load says where it read and nothing else, and a comparison against null is an
895 // ordinary comparison of two values the optimizer has no fact about. So a function
896 // that reads through a pointer and then tests it keeps the test, which is what the
897 // kernel wants and what `-fno-delete-null-pointer-checks` asks for, and what gcc has
898 // to be asked for because it draws the conclusion by default.
899 //
900 // `-fdelete-null-pointer-checks` is the request to draw it, and it goes the way
901 // `-fstrict-aliasing` does: assuming less than was asked for costs speed and not
902 // correctness, and `-O2` implies it, so refusing it would stop builds for nothing.
903 "-fdelete-null-pointer-checks" | "-fno-delete-null-pointer-checks" => {}
904 // The floating point group, which goes the same way and for the same reason, and which
905 // is worth writing out because the reason is easy to get backwards.
906 //
907 // Each of these has a restrictive spelling and a permissive one. The restrictive ones,
908 // `-frounding-math` and `-ftrapping-math`, say that the rounding mode may have been
909 // changed and that an exception raised by an operation may be looked at, so an
910 // arithmetic the compiler folds at compile time is an arithmetic whose rounding and
911 // whose exception the program does not get. Nothing here folds any floating point
912 // arithmetic in a function body: `0.1 + 0.2` is an `fadd` and `1.0 / 0.0` is a divide
913 // that runs, at every level. So both of those describe what already happens.
914 //
915 // The permissive ones are the other half, and they are licences rather than requests
916 // for an answer. `-fno-rounding-math` says the rounding mode is the default one and
917 // `-fno-trapping-math` says nothing looks at the exceptions, which together are
918 // permission to fold. Not folding is the conservative side of that permission and is
919 // what a program is entitled to whichever was written, so `-fno-rounding-math` costs
920 // speed and not correctness, which is the test section 4.1 puts a licence through.
921 "-frounding-math" | "-fno-rounding-math" => {}
922 // `-fno-trapping-math` is the one of the four that is kept, because there is one
923 // conversion this compiler does not fold and gcc folds under it, and the two answers
924 // differ. Converting a constant floating value to an integer type it does not fit in
925 // is undefined behaviour rather than a value: left to the hardware it is one
926 // instruction and the answer is the integer indefinite value, and folded it is the
927 // nearest end of the integer's range. Both compilers leave it to the instruction by
928 // default and gcc folds it under this flag, so a program built with it and compiled
929 // without it gets a different number rather than a slower one. `-ftrapping-math` is
930 // gcc's default, so a build spelling it out is asking for what it already has.
931 "-ftrapping-math" => opts.trapping_math = true,
932 "-fno-trapping-math" => opts.trapping_math = false,
933 // About temporary files rather than about code. There is nothing between the phases of
934 // one compilation here to write to a file in the first place.
935 "-pipe" => {}
936 // Nothing here writes colour, so all of these are the same answer, and it is the answer
937 // that costs nothing: the diagnostics come out plain either way and no build depends on
938 // an escape sequence being there. Taken rather than refused because cmake writes
939 // `-fdiagnostics-color=always` on every compile line when the generator is ninja, which
940 // makes this the second most common flag after `-fPIC` to stop a build over a question
941 // about how the text looks.
942 "-fdiagnostics-color" | "-fno-diagnostics-color" => {}
943 _ if arg.starts_with("-fdiagnostics-color=") => {}
944 // The link flags. None of them changes the compilation, which is why they are
945 // collected apart from `opts` and why `-lm` on a `-c` line is a note rather than an
946 // error: it is a thing said to a linker that is not going to run.
947 "-static" => link.is_static = true,
948 "-shared" => link.shared = true,
949 "-pie" => link.pie = Some(true),
950 "-no-pie" | "-nopie" => link.pie = Some(false),
951 "-nostdlib" => link.no_stdlib = true,
952 "-nostartfiles" => link.no_startfiles = true,
953 "-nodefaultlibs" => link.no_defaultlibs = true,
954 "-fno-builtins-lib" => link.no_builtins_lib = true,
955 "-fbuiltins-lib" => link.no_builtins_lib = false,
956 "-rdynamic" | "-export-dynamic" => link.export_dynamic = true,
957 "-s" => link.strip = true,
958 "-Xlinker" => {
959 let next = args.get(i).ok_or_else(|| err("-Xlinker requires an argument"))?;
960 i += 1;
961 link.passthrough.push(next.clone());
962 }
963 _ if arg.starts_with("-Wl,") => {
964 // Commas separate arguments rather than being part of one, which is what makes
965 // `-Wl,-rpath,/opt/lib` two words to the linker and one word here.
966 link.passthrough.extend(arg["-Wl,".len()..].split(',').map(str::to_owned));
967 }
968 _ if arg.starts_with("-fuse-ld=") => {
969 link.use_ld = Some(arg["-fuse-ld=".len()..].to_owned());
970 }
971 _ if arg.starts_with("-l") && arg.len() > 2 => {
972 inputs.push(Input::library(&arg[2..]));
973 }
974 "-l" => {
975 let next = args.get(i).ok_or_else(|| err("-l requires an argument"))?;
976 i += 1;
977 inputs.push(Input::library(next));
978 }
979 _ if arg.starts_with("-L") => {
980 link.search.push(PathBuf::from(joined_or_next(arg, 2, args, &mut i)?));
981 }
982 _ if arg.starts_with("-B") => {
983 link.prefixes.push(PathBuf::from(joined_or_next(arg, 2, args, &mut i)?));
984 }
985 _ if arg.starts_with("-j") => {
986 jobs = Jobs::parse(&arg[2..]).map_err(err)?;
987 }
988 _ if arg.starts_with("--sysroot=") => {
989 sysroot = Some(PathBuf::from(&arg["--sysroot=".len()..]));
990 }
991 _ if arg.starts_with("--target=") => {
992 let t = &arg["--target=".len()..];
993 opts.target = t.parse().map_err(|e| err(format!("{e}")))?;
994 // The same string again, as the model that has room for a libc version. A spelling
995 // the three field parser took and this one does not is not an error, because the
996 // one that decides what is compiled has already accepted it and the only thing
997 // lost is a version nobody asked for.
998 pinned = t.parse().ok();
999 }
1000 _ if arg.starts_with("--emit=") => {
1001 let k = &arg["--emit=".len()..];
1002 opts.emit = k
1003 .parse()
1004 .map_err(|()| err(format!("unknown --emit kind `{k}`, see --help")))?;
1005 }
1006 // A bare `-O` is `-O1`, which is what GCC has and what a hand written makefile tends
1007 // to write. `-Og` is GCC's level for a build somebody is going to step through, and
1008 // it is `-O1` with the transformations that move code around left out; this compiler
1009 // has no such level yet, so it is the nearest one and `--print-pipeline` says what
1010 // that came to rather than the flag pretending otherwise.
1011 "-O" | "-Og" => opts.opt_level = rucc_session::OptLevel::O1,
1012 // The union of `-O3` and `-ffast-math`, and the second half of that changes what
1013 // floating point arithmetic means. Refused rather than taken as `-O3`, because a
1014 // build that asks for fast math and is quietly given ordinary arithmetic gets a
1015 // slower program than it asked for and a build that is given fast math it did not
1016 // ask for gets a wrong one.
1017 "-Ofast" => {
1018 return Err(err(
1019 "-Ofast is -O3 with fast math, and fast math is not implemented, see \
1020 spec/04-driver-and-cli.md section 4.6",
1021 ));
1022 }
1023 _ if arg.starts_with("-O") => {
1024 opts.opt_level = arg[2..]
1025 .parse()
1026 .map_err(|()| err(format!("unknown optimization level `{arg}`")))?;
1027 }
1028 // How far a multiply and an addition may be fused into one rounding. Before the
1029 // optimizer's `-f` family below for the reason the ones under it are, and kept rather
1030 // than dropped because it is the one flag in its group this compiler could act on: it
1031 // rides into the IR as an attribute on each function with a body, so the day the code
1032 // generator forms an `fma` it already knows which functions were given permission.
1033 // Nothing forms one today, under any value of this and under any `-march=`.
1034 _ if arg.starts_with("-ffp-contract=") => {
1035 let how = &arg["-ffp-contract=".len()..];
1036 opts.fp_contract = how.parse().map_err(|()| {
1037 err(format!("`{how}` is not a contraction, which is fast, on or off"))
1038 })?;
1039 }
1040 // How much of an expression may be computed wider than it was written. The values are
1041 // gcc's and so is the refusal of anything else, and none of the three changes anything
1042 // here: an operation is computed in the type C says it is on every target this compiler
1043 // has a back end for, so `__FLT_EVAL_METHOD__` is 0 and `standard` is already what
1044 // happens. `fast` and `16` are permission to be wider, which is a licence this takes
1045 // and does not use, the same way the two above are. The flag is worth taking because
1046 // glibc's headers and a good deal of configure output write it, and because the answer
1047 // it asks about is one this compiler can state rather than guess at: there is no x87
1048 // target here, which is the machine the whole question was invented for.
1049 // Whether a local and a spilled value that are never both wanted may be the same bytes
1050 // of the frame. gcc's three values, and two of them mean the same thing here: what rucc
1051 // shares is a local whose address provably never leaves the function, which is narrower
1052 // than `named_vars` and narrower still than `all`, so both of them get it. `none` is
1053 // the one that changes anything, and it is the flag a program that reads a local
1054 // through a pointer it kept past the end of the block writes.
1055 _ if arg.starts_with("-fstack-reuse=") => {
1056 let how = &arg["-fstack-reuse=".len()..];
1057 opts.stack_reuse = match how {
1058 "all" | "named_vars" => Some(true),
1059 "none" => Some(false),
1060 _ => {
1061 return Err(err(format!(
1062 "`{how}` is not a stack reuse, which is all, named_vars or none"
1063 )));
1064 }
1065 };
1066 }
1067 _ if arg.starts_with("-fexcess-precision=") => {
1068 let how = &arg["-fexcess-precision=".len()..];
1069 if !matches!(how, "16" | "fast" | "standard") {
1070 return Err(err(format!(
1071 "`{how}` is not an excess precision, which is 16, fast or standard"
1072 )));
1073 }
1074 }
1075 // Which front of a path is rewritten before it reaches the output, which is how a
1076 // build gets the same bytes out of two different directories. The four spellings are
1077 // one flag each into three lists, and `-ffile-prefix-map=` is the three of them at
1078 // once. Only the macro list does anything today, because `__FILE__` is the only place
1079 // a path reaches the output: there is no DWARF and no profile data yet, so the other
1080 // two are recorded for the work that will read them. The argument splits at the last
1081 // `=` rather than the first, which is gcc's rule and is what lets a directory with an
1082 // `=` in its name be the old half.
1083 _ if arg.starts_with("-fmacro-prefix-map=") => {
1084 let (old, new) = rewrite(arg, "-fmacro-prefix-map=")?;
1085 opts.prefix_map.macros.push(old, new);
1086 }
1087 _ if arg.starts_with("-fdebug-prefix-map=") => {
1088 let (old, new) = rewrite(arg, "-fdebug-prefix-map=")?;
1089 opts.prefix_map.debug.push(old, new);
1090 }
1091 _ if arg.starts_with("-fprofile-prefix-map=") => {
1092 let (old, new) = rewrite(arg, "-fprofile-prefix-map=")?;
1093 opts.prefix_map.profile.push(old, new);
1094 }
1095 _ if arg.starts_with("-ffile-prefix-map=") => {
1096 let (old, new) = rewrite(arg, "-ffile-prefix-map=")?;
1097 opts.prefix_map.macros.push(old, new);
1098 opts.prefix_map.debug.push(old, new);
1099 opts.prefix_map.profile.push(old, new);
1100 }
1101 // A whole optimization rather than a flag, and the family is taken rather than
1102 // refused because of what ignoring it does. There is none of it here yet, so a build
1103 // that asks for it gets a program that is correct and slower than it could have been,
1104 // which is what section 4.1 means by a hint about speed and what every compilation at
1105 // `-O0` already is. The objects settle the rest of the argument: gcc's `-flto` object
1106 // holds the bytecode and no machine code at all, and every object here holds the code,
1107 // which is exactly what `-ffat-lto-objects` asks gcc for. So a build passing `-flto`
1108 // to this compiler gets objects that are more usable than the ones it asked for rather
1109 // than different ones. Every value is still checked against gcc's, because somebody
1110 // who wrote `-flto=thin` meant clang and had better hear about it here.
1111 "-flto" => opts.lto.requested = true,
1112 "-fno-lto" => opts.lto.requested = false,
1113 _ if arg.starts_with("-flto=") => {
1114 let how = &arg["-flto=".len()..];
1115 opts.lto.jobs = how.parse().map_err(|()| {
1116 err(format!(
1117 "`{how}` is not a number of link time jobs, which is auto, jobserver or a \
1118 count above zero"
1119 ))
1120 })?;
1121 opts.lto.requested = true;
1122 }
1123 _ if arg.starts_with("-flto-partition=") => {
1124 let how = &arg["-flto-partition=".len()..];
1125 opts.lto.partition = how.parse().map_err(|()| {
1126 err(format!(
1127 "`{how}` is not a partitioning model, which is balanced, 1to1, one, max \
1128 or none"
1129 ))
1130 })?;
1131 }
1132 _ if arg.starts_with("-flto-compression-level=") => {
1133 let how = &arg["-flto-compression-level=".len()..];
1134 let level =
1135 how.parse::<u8>().ok().filter(|level| *level <= 19).ok_or_else(|| {
1136 err(format!("`{how}` is not a compression level, 0 to 19"))
1137 })?;
1138 opts.lto.compression = Some(level);
1139 }
1140 // Whether the object keeps its machine code as well as the bytecode. It always does
1141 // here, so the first of these describes what happens and the second asks for an object
1142 // with less in it, which is a smaller file and not a different program, so both are
1143 // taken.
1144 "-ffat-lto-objects" | "-fno-fat-lto-objects" => {}
1145 // Whether the linker is handed a plugin that does the link time work. The design in
1146 // `spec/09-optimizer.md` has this driver doing that work itself and never loading a
1147 // plugin into anybody, so neither answer is a question it has to hold.
1148 "-fuse-linker-plugin" | "-fno-use-linker-plugin" => {}
1149 // Reading a profile back. Taken for the reason the family above it is: nothing here
1150 // reads one, so a build that asks gets the program it would have got anyway, and gcc
1151 // itself produces a byte for byte identical object from `-fprofile-use` when there are
1152 // no counts beside the file. The path is recorded for the pass that will read it. The
1153 // warning gcc prints when it looked and found nothing is deliberately not copied,
1154 // because nothing here looks, and a warning about a file that was never opened would
1155 // fire on the builds that have a perfectly good profile as well as on the ones that
1156 // do not.
1157 "-fprofile-use" => opts.profile_data.requested = true,
1158 "-fno-profile-use" => opts.profile_data.requested = false,
1159 _ if arg.starts_with("-fprofile-use=") => {
1160 opts.profile_data.path = Some(arg["-fprofile-use=".len()..].to_string());
1161 opts.profile_data.requested = true;
1162 }
1163 _ if arg.starts_with("-fprofile-dir=") => {
1164 opts.profile_data.dir = Some(arg["-fprofile-dir=".len()..].to_string());
1165 }
1166 "-fprofile-abs-path" => opts.profile_data.absolute = true,
1167 "-fno-profile-abs-path" => opts.profile_data.absolute = false,
1168 "-fprofile-correction" => opts.profile_data.correction = true,
1169 "-fno-profile-correction" => opts.profile_data.correction = false,
1170 "-fprofile-partial-training" => opts.profile_data.partial_training = true,
1171 "-fno-profile-partial-training" => opts.profile_data.partial_training = false,
1172 // Writing the counts rather than reading them, which is refused rather than taken and
1173 // is the same line `-gsplit-dwarf` falls on the far side of. Ignoring these means a
1174 // file a build declared as an output never appears: the instrumented program writes a
1175 // `.gcda` as it exits and `-ftest-coverage` writes a `.gcno` beside the object, and a
1176 // two stage build that got neither would go on to optimize against no counts at all
1177 // and report coverage of nothing, with nothing along the way saying so. The objects
1178 // say the rest: gcc's `-fprofile-generate` object holds 375 bytes of code where a
1179 // plain one holds 71, and 296 bytes of counters that a plain one does not have, so
1180 // this is a flag that changes the output rather than a hint about speed.
1181 "-fprofile-arcs"
1182 | "--coverage"
1183 | "-fcondition-coverage"
1184 | "-fpath-coverage"
1185 | "-fprofile-generate" => {
1186 return Err(err(format!(
1187 "{arg}: this compiler does not instrument for profiling, and a build that \
1188 expects the counts a run of the instrumented program writes would optimize \
1189 against nothing on its second pass, see spec/04-driver-and-cli.md"
1190 )));
1191 }
1192 _ if arg.starts_with("-fprofile-generate=") => {
1193 return Err(err(format!(
1194 "{arg}: this compiler does not instrument for profiling, and a build that \
1195 expects the counts a run of the instrumented program writes would optimize \
1196 against nothing on its second pass, see spec/04-driver-and-cli.md"
1197 )));
1198 }
1199 "-ftest-coverage" => {
1200 return Err(err(format!(
1201 "{arg}: this compiler writes no `.gcno` file beside the object, and a build \
1202 that expects one would wait for a file that never arrives, see \
1203 spec/04-driver-and-cli.md"
1204 )));
1205 }
1206 // The rest of the family describes instrumentation that is refused above, so what is
1207 // left to do with them is check them and drop them. They are checked because a
1208 // misspelling in a distribution's flags is worth finding here rather than on the day
1209 // the instrumentation lands, and dropped because there is nothing for an answer about
1210 // how a counter is written to be an answer about.
1211 _ if arg.starts_with("-fprofile-update=") => {
1212 let how = &arg["-fprofile-update=".len()..];
1213 if !matches!(how, "single" | "atomic" | "prefer-atomic") {
1214 return Err(err(format!(
1215 "`{how}` is not a profile update method, which is single, atomic or \
1216 prefer-atomic"
1217 )));
1218 }
1219 }
1220 _ if arg.starts_with("-fprofile-reproducible=") => {
1221 let how = &arg["-fprofile-reproducible=".len()..];
1222 if !matches!(how, "serial" | "parallel-runs" | "multithreaded") {
1223 return Err(err(format!(
1224 "`{how}` is not a profile reproducibility method, which is serial, \
1225 parallel-runs or multithreaded"
1226 )));
1227 }
1228 }
1229 "-fprofile-values" | "-fno-profile-values" | "-fprofile-info-section" => {}
1230 "-fno-test-coverage" | "-fno-profile-arcs" | "-fno-profile-generate" => {}
1231 _ if arg.starts_with("-fprofile-filter-files=")
1232 || arg.starts_with("-fprofile-exclude-files=")
1233 || arg.starts_with("-fprofile-note=") => {}
1234 // What every name gets when nothing in the source said, which the attribute in the
1235 // source overrides rather than the other way round. Before the optimizer's `-f`
1236 // family below for the reason the tier below it is.
1237 _ if arg.starts_with("-fvisibility=") => {
1238 let seen = &arg["-fvisibility=".len()..];
1239 opts.visibility = seen.parse().map_err(|()| {
1240 err(format!(
1241 "`{seen}` is not a visibility, which is default, hidden, internal or \
1242 protected"
1243 ))
1244 })?;
1245 }
1246 // Which edges of a control flow transfer are checked. Before the optimizer's `-f`
1247 // family below for the reason the two above it are, and last of the three so that the
1248 // bare spelling and the negative one are matched exactly rather than by this.
1249 _ if arg.starts_with("-fcf-protection=") => {
1250 let edges = &arg["-fcf-protection=".len()..];
1251 opts.control = edges.parse().map_err(|()| {
1252 err(format!(
1253 "`{edges}` is not a control flow protection, which is full, branch, \
1254 return, none or check"
1255 ))
1256 })?;
1257 }
1258 // How much room every function opens with for something to be written over later.
1259 // Before the optimizer's `-f` family below for the reason the ones above it are.
1260 _ if arg.starts_with("-fpatchable-function-entry=") => {
1261 let room = &arg["-fpatchable-function-entry=".len()..];
1262 opts.patchable = room.parse().map_err(|()| {
1263 err(format!(
1264 "`{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"
1265 ))
1266 })?;
1267 }
1268 // The memory safety monitor, from section 15.4 of
1269 // `spec/safe-memory/15-integration.md`. Before the optimizer's `-f` family below,
1270 // because a pass that took the name `safety=detect` would otherwise be handed the
1271 // flag, and the tier is not a pass.
1272 _ if arg.starts_with("-fsafety=") => {
1273 let tier = &arg["-fsafety=".len()..];
1274 opts.safety = tier.parse().map_err(|()| {
1275 err(format!(
1276 "`{tier}` is not a safety tier, which is off, detect, enforce or kernel"
1277 ))
1278 })?;
1279 }
1280 // Whether padding participates, from section 9.3 of document 09. Spelled out rather
1281 // than folded into the tier because it is a departure somebody who has read that
1282 // section makes, and the two defaults it describes are a property of what is being
1283 // built rather than of how much checking is wanted.
1284 _ if arg.starts_with("-fsafety-init=") => {
1285 let mode = &arg["-fsafety-init=".len()..];
1286 opts.padding = mode.parse().map_err(|()| {
1287 err(format!("`{mode}` is not a padding mode, which is padding or nopadding"))
1288 })?;
1289 }
1290 // Row S4, from section 9.4 of document 09. A bare flag with no value, because the
1291 // strict form of that section needs a member id the front end does not name yet and
1292 // accepting the spelling for it would be accepting a promise this build cannot keep.
1293 // Before `-fno-` is looked at below, for the reason the tier is.
1294 "-fsafety-subobject" => opts.subobject = rucc_session::Subobject::Members,
1295 "-fno-safety-subobject" => opts.subobject = rucc_session::Subobject::Off,
1296 _ if arg.starts_with("-fsafety-subobject=") => {
1297 let form = &arg["-fsafety-subobject=".len()..];
1298 return Err(err(format!(
1299 "`{form}` is not a form of -fsafety-subobject. The flag takes no value, and \
1300 the strict form of section 9.4 is tamnd/rucc#967"
1301 )));
1302 }
1303 // Row Y8, from section 9.6 of document 09. A bare flag with no value, for the reason
1304 // the one above has none: there is one form of this check and a spelling that suggested
1305 // otherwise would be promising something. Before `-fno-` is looked at below, the same
1306 // way.
1307 "-fsafety-restrict" => opts.promise = rucc_session::Promise::Blocks,
1308 "-fno-safety-restrict" => opts.promise = rucc_session::Promise::Off,
1309 _ if arg.starts_with("-fsafety-restrict=") => {
1310 let form = &arg["-fsafety-restrict=".len()..];
1311 return Err(err(format!(
1312 "`{form}` is not a form of -fsafety-restrict. The flag takes no value."
1313 )));
1314 }
1315 // Section 9.5's races, which take a value because the section gives them three modes
1316 // and the difference between two of them is which classes get reported rather than how
1317 // much is recorded. `-fno-` is the same as `=off` and is spelled out here for the same
1318 // reason the two above spell theirs out.
1319 _ if arg.starts_with("-fsafety-races=") => {
1320 let mode = &arg["-fsafety-races=".len()..];
1321 opts.races = mode.parse().map_err(|()| {
1322 err(format!("`{mode}` is not a race mode, which is off, metadata or pointer"))
1323 })?;
1324 }
1325 "-fno-safety-races" => opts.races = rucc_session::Races::Off,
1326 // The sanitizers of document 12, which are checks at run time rather than a way of
1327 // generating the same program. Each name is held to gcc 16's list, and what is still
1328 // asked for by the end of the line is answered after the loop, so that a command line
1329 // which turns one on and then off again is a command line that asked for nothing.
1330 //
1331 // Before the optimizer's `-f` family below, for the reason the tier above it is.
1332 _ if arg.starts_with("-fsanitize=") => {
1333 for one in arg["-fsanitize=".len()..].split(',') {
1334 if one == "all" {
1335 // gcc takes `all` only in the negative, because turning every check on at
1336 // once includes checks that contradict each other.
1337 return Err(err(
1338 "`-fsanitize=all` is not a gcc option, only `-fno-sanitize=all` is",
1339 ));
1340 }
1341 if !SANITIZERS.contains(&one) {
1342 return Err(err(format!(
1343 "`{one}` is not a sanitizer, see spec/04-driver-and-cli.md section 4.7"
1344 )));
1345 }
1346 if !sanitizers.contains(&one) {
1347 sanitizers.push(one);
1348 }
1349 }
1350 }
1351 _ if arg.starts_with("-fno-sanitize=") => {
1352 for one in arg["-fno-sanitize=".len()..].split(',') {
1353 if one == "all" {
1354 sanitizers.clear();
1355 continue;
1356 }
1357 if !SANITIZERS.contains(&one) {
1358 return Err(err(format!(
1359 "`{one}` is not a sanitizer, see spec/04-driver-and-cli.md section 4.7"
1360 )));
1361 }
1362 sanitizers.retain(|asked| *asked != one);
1363 }
1364 }
1365 // What a check does when it fires, and where the records about the checked objects go.
1366 // Each of them is an answer about the sanitizers refused after the loop, so there is
1367 // nothing left for them to change here. The names are still held to the list, because
1368 // a misspelling in a build's flags is worth finding when the compiler reads it.
1369 _ if arg.starts_with("-fsanitize-recover=")
1370 || arg.starts_with("-fno-sanitize-recover=")
1371 || arg.starts_with("-fsanitize-trap=")
1372 || arg.starts_with("-fno-sanitize-trap=") =>
1373 {
1374 // The guard above matched on a spelling that has an `=` in it, so the tail is
1375 // whatever follows the first one.
1376 let how = arg.split_once('=').map_or("", |(_, rest)| rest);
1377 for one in how.split(',') {
1378 if one != "all" && !SANITIZERS.contains(&one) {
1379 return Err(err(format!(
1380 "`{one}` is not a sanitizer, see spec/04-driver-and-cli.md section 4.7"
1381 )));
1382 }
1383 }
1384 }
1385 "-fsanitize-undefined-trap-on-error"
1386 | "-fsanitize-address-use-after-scope"
1387 | "-fno-sanitize-address-use-after-scope" => {}
1388 _ if arg.starts_with("-fsanitize-sections=") => {}
1389 // Counting which edges a run reached, which is how a fuzzer knows an input was worth
1390 // keeping. Refused rather than dropped, because a fuzzer whose calls into
1391 // `__sanitizer_cov_*` were never generated runs blind and reports coverage of nothing,
1392 // and there is no point in the campaign where that announces itself.
1393 _ if arg.starts_with("-fsanitize-coverage=") => {
1394 let how = &arg["-fsanitize-coverage=".len()..];
1395 for one in how.split(',') {
1396 if !matches!(one, "trace-pc" | "trace-cmp") {
1397 return Err(err(format!(
1398 "`{one}` is not a coverage instrumentation, which is trace-pc or \
1399 trace-cmp"
1400 )));
1401 }
1402 }
1403 return Err(err(format!(
1404 "{arg}: this compiler generates no coverage callbacks, and a fuzzer built \
1405 with it would run without any feedback at all, see \
1406 spec/04-driver-and-cli.md section 4.7"
1407 )));
1408 }
1409 // The optimizer's own flags, from section 9.10 of `spec/09-optimizer.md`. These come
1410 // after every `-f` the rest of the compiler answers to, so a pass can never take a
1411 // name that already means something else on the command line.
1412 _ if arg.starts_with("-fpass-fuel=") => {
1413 let (name, count) = arg["-fpass-fuel=".len()..]
1414 .split_once('=')
1415 .ok_or_else(|| err("-fpass-fuel= is spelled <pass>=<count>"))?;
1416 if rucc_opt::pass::find(name).is_none() {
1417 return Err(err(format!(
1418 "`{name}` is not a pass this compiler has, see --print-pipeline"
1419 )));
1420 }
1421 let count: u32 = count
1422 .parse()
1423 .map_err(|_| err(format!("`{count}` is not a number of transformations")))?;
1424 opts.pass_fuel.push((name.to_owned(), count));
1425 }
1426 _ if arg.starts_with("-fpass-fuel-global=") => {
1427 let count = &arg["-fpass-fuel-global=".len()..];
1428 let count: u32 = count
1429 .parse()
1430 .map_err(|_| err(format!("`{count}` is not a number of transformations")))?;
1431 opts.pass_fuel_global = Some(count);
1432 }
1433 // Everything from `-fopt-info` to the end of the argument, which is optional
1434 // keywords joined by hyphens and an optional `=<file>`. Checked here rather than
1435 // where the remarks are printed, because by then the compilation somebody wanted
1436 // to hear about is over.
1437 _ if arg == "-fopt-info"
1438 || arg.starts_with("-fopt-info=")
1439 || arg.starts_with("-fopt-info-") =>
1440 {
1441 let rest = &arg["-fopt-info".len()..];
1442 let (kinds, file) = match rest.split_once('=') {
1443 Some((kinds, file)) => (kinds, Some(file)),
1444 None => (rest, None),
1445 };
1446 let kinds = kinds.strip_prefix('-').unwrap_or(kinds);
1447 rucc_opt::Wants::none().add(kinds).map_err(err)?;
1448 opts.opt_info.push(kinds.to_owned());
1449 if let Some(file) = file {
1450 if file.is_empty() {
1451 return Err(err("-fopt-info= was given no file to write to"));
1452 }
1453 opts.opt_info_file = Some(file.to_owned());
1454 }
1455 }
1456 _ if arg.starts_with("-fdump-ir=") => {
1457 // Checked here rather than where the dumps are taken, because the compilation
1458 // that would have been dumped is over by then.
1459 let spec = &arg["-fdump-ir=".len()..];
1460 rucc_opt::Dumps::default().add(spec).map_err(err)?;
1461 opts.dump_ir.push(spec.to_owned());
1462 }
1463 // Before the bare `-f<pass>` below, because a pass called `enable-something` would
1464 // otherwise take the flag away from the gate. Checked here rather than where the
1465 // pipeline reads it, for the reason that applies to all of these: a misspelled pass
1466 // name that quietly gated nothing looks exactly like a pass that is not the guilty
1467 // one, and a bisection would carry on past the thing it was looking for.
1468 _ if arg.starts_with("-fdisable-") || arg.starts_with("-fenable-") => {
1469 let on = arg.starts_with("-fenable-");
1470 let spec = &arg[if on { "-fenable-".len() } else { "-fdisable-".len() }..];
1471 rucc_opt::Gates::default().add(on, spec).map_err(err)?;
1472 opts.pass_gates.push((on, spec.to_owned()));
1473 }
1474 // gcc's spelling for a pass this compiler has under a shorter name. It goes above the
1475 // two arms below rather than into the pile of gcc pass names further down, because the
1476 // pass is here: dropping the flag would leave a build that asked for unrolling without
1477 // it, and refusing it stops the build outright, which is what libtommath's makefile
1478 // ran into. `-funroll-all-loops` is deliberately not in here: gcc's is the one that
1479 // unrolls without a trip count, which is a different and usually worse thing.
1480 "-funroll-loops" => opts.passes.push(("unroll".to_owned(), true)),
1481 "-fno-unroll-loops" => opts.passes.push(("unroll".to_owned(), false)),
1482 _ if arg.strip_prefix("-fno-").is_some_and(|n| rucc_opt::pass::find(n).is_some()) => {
1483 opts.passes.push((arg["-fno-".len()..].to_owned(), false));
1484 }
1485 _ if arg.strip_prefix("-f").is_some_and(|n| rucc_opt::pass::find(n).is_some()) => {
1486 opts.passes.push((arg["-f".len()..].to_owned(), true));
1487 }
1488 // The flags that name a pass of gcc's own. They arrive from the torture suite, where a
1489 // program reduced from a miscompilation usually names the pass that miscompiled it on
1490 // its `dg-options` line, and they arrive from hand written build files for the same
1491 // reason. Section 4.1 sorts a flag by what the output would be without it, and by that
1492 // rule these are one pile: a flag that turns one of gcc's passes on or off is asking
1493 // for a compiler that does not exist here, and the program it is attached to is a
1494 // correctness test that passes either way. Turning on a pass we do not have costs
1495 // speed, turning off a pass we do not have costs nothing, and neither changes what the
1496 // program computes.
1497 //
1498 // rucc's own pass names are matched above this, so `-fno-dce` turns off the dce this
1499 // compiler has rather than landing here, and the day one of these names becomes a pass
1500 // here it stops being taken and dropped without anybody editing this list.
1501 //
1502 // Two of them are prefixes rather than names, which is the one place this file takes a
1503 // family instead of a flag. gcc files its gimple passes under `-ftree-` and its
1504 // interprocedural passes under `-fipa-`, both namespaces are pass selection and
1505 // nothing else, and there is no member of either that changes the meaning of a program
1506 // that was already correct. The rest are written out one at a time, because they live
1507 // in the flat `-f` namespace where the neighbours do change meanings.
1508 _ if arg.starts_with("-ftree-") || arg.starts_with("-fno-tree-") => {}
1509 _ if arg.starts_with("-fipa-") || arg.starts_with("-fno-ipa-") => {}
1510 "-fexpensive-optimizations" | "-fno-expensive-optimizations" => {}
1511 "-fmodulo-sched" | "-fno-modulo-sched" => {}
1512 "-fvect-cost-model" | "-fno-vect-cost-model" => {}
1513 _ if arg.starts_with("-fvect-cost-model=") || arg.starts_with("-fsimd-cost-model=") => {
1514 }
1515 "-fearly-inlining" | "-fno-early-inlining" => {}
1516 "-finline"
1517 | "-fno-inline"
1518 | "-finline-functions"
1519 | "-fno-inline-functions"
1520 | "-finline-small-functions"
1521 | "-fno-inline-small-functions"
1522 | "-finline-functions-called-once"
1523 | "-fno-inline-functions-called-once" => {}
1524 "-foptimize-strlen" | "-fno-optimize-strlen" => {}
1525 "-fira-share-spill-slots" | "-fno-ira-share-spill-slots" => {}
1526 // The charset flags are not in that pile, because an encoding is a statement about
1527 // what the bytes of the source mean rather than about how fast the output is. The
1528 // preprocessor reads UTF-8 and has no converter, so the one name that describes what
1529 // already happens is taken and every other name is refused. Spelled without regard to
1530 // case and with both of the spellings iconv answers to, since a build writes whichever
1531 // one its author typed.
1532 _ if arg.starts_with("-finput-charset=") => {
1533 let name = &arg["-finput-charset=".len()..];
1534 if !name.eq_ignore_ascii_case("utf-8") && !name.eq_ignore_ascii_case("utf8") {
1535 return Err(err(format!(
1536 "-finput-charset={name}: the preprocessor reads UTF-8 and has no \
1537 converter, so a file in another encoding would be read as though it were \
1538 UTF-8 rather than converted",
1539 )));
1540 }
1541 }
1542 // The three that come in on the same `dg-options` lines and are the other half of
1543 // section 4.1's rule, because each of them changes what the program does and not how
1544 // fast it does it. The negative form of each is what this compiler does anyway, so it
1545 // is taken and dropped, which is the shape `-fnested-functions` has above.
1546 "-ffast-math" => {
1547 return Err(err(
1548 "-ffast-math is a licence to answer a floating point arithmetic differently \
1549 from the way the source wrote it, and it is not one flag: it defines \
1550 __FAST_MATH__, which a library header reads, and gcc links a startup file \
1551 that puts the hardware in flush to zero mode for the whole process. Taking it \
1552 and dropping it would change what other objects in the same program answer. \
1553 -ffp-contract= and -fexcess-precision= are the parts of it this compiler has",
1554 ));
1555 }
1556 "-fno-fast-math" => {}
1557 "-fnon-call-exceptions" => {
1558 return Err(err(
1559 "-fnon-call-exceptions is a promise that an instruction which is not a call \
1560 can raise an exception the unwinder finds a handler for, and nothing here \
1561 produces a landing pad for a trapping instruction. A program built without it \
1562 would unwind past the handler it wrote",
1563 ));
1564 }
1565 "-fno-non-call-exceptions" => {}
1566 "-finstrument-functions" => {
1567 return Err(err(
1568 "-finstrument-functions calls __cyg_profile_func_enter on entry to every \
1569 function and __cyg_profile_func_exit on the way out, and nothing here emits \
1570 either call. A program that asks for them usually counts them, so taking the \
1571 flag and dropping it would turn a program that fails loudly into one that \
1572 fails quietly",
1573 ));
1574 }
1575 "-fno-instrument-functions" => {}
1576 // The unstable options, spelled the way rustc spells them and carrying the same
1577 // promise, which is none: one of these may change or go away in any release. They are
1578 // measurements and debugging aids rather than things a build asks for, which is why
1579 // none of them is in the usage text and all of them are in section 4.11 of
1580 // `spec/04-driver-and-cli.md`.
1581 "-Zverify-each" => opts.verify_each = true,
1582 _ if arg.starts_with("-Zrule-coverage=") => {
1583 let file = &arg["-Zrule-coverage=".len()..];
1584 if file.is_empty() {
1585 return Err(err("-Zrule-coverage= needs a file to write to"));
1586 }
1587 opts.rule_coverage = Some(file.to_owned());
1588 }
1589 _ if arg.starts_with("-Zcycle-accurate-model=") => {
1590 let value = &arg["-Zcycle-accurate-model=".len()..];
1591 opts.cycle_accurate_model = match value {
1592 "yes" | "1" => Some(true),
1593 "no" | "0" => Some(false),
1594 _ => {
1595 return Err(err("-Zcycle-accurate-model= takes yes or no"));
1596 }
1597 };
1598 }
1599 _ if arg.starts_with("-Zregister-pressure=") => {
1600 let file = &arg["-Zregister-pressure=".len()..];
1601 if file.is_empty() {
1602 return Err(err("-Zregister-pressure= needs a file to write to"));
1603 }
1604 opts.register_pressure = Some(file.to_owned());
1605 }
1606 _ if arg.starts_with("-Z") => {
1607 return Err(err(format!(
1608 "`{arg}` is not an unstable option this compiler has, see \
1609 spec/04-driver-and-cli.md section 4.11 for the ones it does"
1610 )));
1611 }
1612 // The word size, which is a statement about the target and is taken as one. A build
1613 // that says the size the target already has is saying nothing, and one that says the
1614 // other size is asking for a target this compiler does not have, which it is told
1615 // rather than being given the wrong one.
1616 "-m64" | "-m32" | "-mx32" => {
1617 let want: u32 = match arg {
1618 "-m64" => 64,
1619 _ => 32,
1620 };
1621 let have = rucc_target::TargetInfo::new(opts.target).pointer_width;
1622 if have != want {
1623 return Err(err(format!(
1624 "{arg} asks for a {want} bit target and {} is {have} bit, use \
1625 --target= to name the one you mean",
1626 opts.target
1627 )));
1628 }
1629 }
1630 // Which processor in the family to generate for. This compiler emits the base
1631 // instruction set of the architecture and nothing above it, so a program built with
1632 // any of these runs on the machine that was named; it is a program that could have
1633 // been faster rather than a program that is wrong, which is what makes these safe to
1634 // take and ignore where a flag that changed the meaning of the code would not be.
1635 _ if arg.starts_with("-march=")
1636 || arg.starts_with("-mtune=")
1637 || arg.starts_with("-mcpu=") => {}
1638 // The calling convention, which is not safe to ignore. Taken when it names the one
1639 // the target already uses and refused otherwise.
1640 _ if arg.starts_with("-mabi=") => {
1641 let want = &arg["-mabi=".len()..];
1642 let have = match opts.target.arch {
1643 rucc_target::Arch::X86_64 => "sysv",
1644 rucc_target::Arch::Aarch64 => "lp64",
1645 rucc_target::Arch::Riscv64 => "lp64d",
1646 };
1647 if want != have {
1648 return Err(err(format!(
1649 "{arg}: {} uses the {have} convention and this compiler has no other",
1650 opts.target
1651 )));
1652 }
1653 }
1654 // How far apart the pieces of the program may be. The small model is what we emit and
1655 // it is every hosted program's default; the kernel model is a different one and a
1656 // build that asks for it and does not get it links and then does not run.
1657 "-mcmodel=small" => {}
1658 _ if arg.starts_with("-mcmodel=") => {
1659 return Err(err(format!(
1660 "{arg}: this compiler emits the small code model and no other, see \
1661 spec/12-targets.md"
1662 )));
1663 }
1664 // GCC's own scripting language for how the driver builds a command line.
1665 // `spec/04-driver-and-cli.md` section 4.4 settles that we will not have it, so a
1666 // build reaching for it is told which flags do the same job.
1667 _ if arg.starts_with("-specs=") => {
1668 return Err(err(
1669 "-specs= is not supported: the parts of it builds rely on are -B, -L, \
1670 -nostdlib, -nostartfiles and -Wl,, see spec/04-driver-and-cli.md \
1671 section 4.4",
1672 ));
1673 }
1674 // Arguments meant for a separate assembler or preprocessor, which this compiler does
1675 // not have: both are inside it and neither reads a command line. Refused rather than
1676 // dropped, because every one of these says something about the output and a build
1677 // that asked for `-Wa,--noexecstack` and was silently given an executable stack got
1678 // the opposite of what it asked for.
1679 _ if arg.starts_with("-Wa,") || arg.starts_with("-Wp,") => {
1680 return Err(err(format!(
1681 "`{arg}` is an argument for a separate assembler or preprocessor, and both \
1682 are inside this compiler rather than programs it runs"
1683 )));
1684 }
1685 "-Xassembler" | "-Xpreprocessor" => {
1686 return Err(err(format!(
1687 "{arg} hands an argument to a separate assembler or preprocessor, and both \
1688 are inside this compiler rather than programs it runs"
1689 )));
1690 }
1691 // Everything else in the `-W` family. `spec/04-driver-and-cli.md` section 4.1 has
1692 // this one as a rule about build systems rather than about warnings: autoconf finds
1693 // out whether a warning flag exists by passing it and looking at the exit status, so
1694 // a compiler that refuses one it has not heard of fails a configure script written
1695 // for a GCC newer than itself. The names are not checked against a list because this
1696 // compiler has no warning groups for a list to be of, which #485 is about.
1697 _ if arg.starts_with("-W") => {}
1698 // Flags that name something this compiler does not do and would not do differently
1699 // if it did. `-fno-ident` is about a comment in the output that we do not write
1700 // either way, and the others are about a way of ordering the compilation that has
1701 // been GCC's only way for twenty years. Section 4.1 asks for the list to be short
1702 // and for adding to it to be deliberate, which is why it is written out here.
1703 "-fno-ident"
1704 | "-fident"
1705 | "-funit-at-a-time"
1706 | "-fno-unit-at-a-time"
1707 | "-shared-libgcc"
1708 | "-static-libgcc" => {}
1709 _ if arg.starts_with('-') && arg.len() > 1 => {
1710 // Silently ignoring an unknown flag is how a build ends up not doing what
1711 // its author asked. spec/13-gnu-compat.md section 13.4 makes this an error
1712 // for the flags that change code generation, and the safe default until the
1713 // flag table is populated is to reject everything we do not know.
1714 return Err(err(format!("unknown option `{arg}`")));
1715 }
1716 _ => inputs.push(Input { path: arg.to_owned(), forced, library: false }),
1717 }
1718 }
1719
1720 // The fetch, before anything that resolves a compilation, because `--fetch` does not describe
1721 // one. It is here rather than in the loop so that `--offline` can forbid it whichever order the
1722 // two were written in, and it is before the refusals below so that a command line asking for a
1723 // sysroot is not told about a sanitizer.
1724 if let Some(named) = fetch {
1725 return fetch_action(&named, offline, &inputs);
1726 }
1727
1728 // Last, so that it lands after every `-isystem` the command line gave. That is GCC's
1729 // order: a directory the user names outranks the compiler's own, and the compiler's own
1730 // outranks the library's. It is pushed after the loop rather than before it because
1731 // `SearchPath` appends within a group and the position is what the order is.
1732 // The same directory the headers were looked for under, because a sysroot is a statement
1733 // about a whole installation and not about half of one.
1734 // After the loop, because `-fno-sanitize=` can take back what an earlier flag asked for and a
1735 // command line that turns a check on and off again has asked for nothing. What is left is
1736 // refused rather than dropped, and it is the one place in this parser where the reason is not
1737 // that the output would differ. A sanitizer is a promise that the program is watched while it
1738 // runs, so a build that asks for one and is quietly given a program with no checks in it does
1739 // not get a slower program or a bigger file, it gets a test suite that passes for the wrong
1740 // reason. `-fsafety=` is the checking this compiler does have, and the message says so, because
1741 // somebody reaching for `-fsanitize=address` wants the nearest thing rather than a list of
1742 // options.
1743 if let Some(first) = sanitizers.first() {
1744 return Err(err(format!(
1745 "-fsanitize={first}: this compiler has no sanitizer instrumentation, and a build that \
1746 asked for one and got none would run its tests unchecked, see \
1747 spec/04-driver-and-cli.md section 4.7. `-fsafety=detect` is the memory checking this \
1748 compiler does have"
1749 )));
1750 }
1751 link.sysroot = sysroot.clone();
1752 // Where a sysroot for a target that is not this machine would be. Read once, here, rather than
1753 // inside the link line, because a link line that read the environment could only be tested on a
1754 // machine whose environment said the right thing, and the link line is the last thing that
1755 // touches a binary. `spec/cross-compile/13-distribution.md` section 13.2 owns the answer.
1756 link.cache = Some(cache::dir());
1757 // And the ten field spelling of the target, because the release on it decides two things the
1758 // three field one cannot say: whether a target that is this architecture is still a cross
1759 // compile, and which directory under the cache it is against. After the loop because the last
1760 // `--target=` on the command line is the one that counts.
1761 link.pinned = pinned;
1762 // After the loop rather than where `-pthread` was read, so that it lands after the objects
1763 // that refer to it. A static link takes the definitions it needs from a library when it
1764 // reaches it and not afterwards, so a library before the objects is a library that answers
1765 // nothing.
1766 if threads {
1767 inputs.push(Input::library("pthread"));
1768 }
1769 if let Some(query) = query {
1770 return Ok(Action::Print(answer(&query, &opts, &link)?));
1771 }
1772 // `-M` and `-MM` produce the rule and nothing else, so the run stops after phase 4 whatever
1773 // else the command line asked for. Read here rather than where the flag was, because a `-c`
1774 // written after it has to lose and the loop cannot know that until it has ended. The output
1775 // file is where the rule goes rather than where an object would have gone, and the last
1776 // phase being the preprocessor is what makes that true without a second rule for it.
1777 if opts.deps.instead_of_compiling {
1778 opts.emit = EmitKind::Preprocessed;
1779 }
1780 if !nostdinc {
1781 opts.search.push_system(runtime::DIR);
1782 // And the library's after ours, which is the other half of the same order. They go on
1783 // here rather than at the point `--target=` or `--sysroot=` was read because either
1784 // one changes the answer and the last word on both is the end of the loop.
1785 //
1786 // Which library's is the question `link::cross_sysroot` answers, and it is asked here so
1787 // that the headers and the libraries come from the same place. A target that is this
1788 // machine reads this machine's headers, and a target that is not reads the ones in the
1789 // sysroot for it rather than the ones next door.
1790 let cross = link::cross_sysroot(opts.target, &link);
1791 let kernel = link::cross_kernel(opts.target, &link);
1792 // And the version of those headers, which only the bundled tree has an answer for. A host
1793 // glibc and a tree the user named both define `__GLIBC_MINOR__` in their own `features.h`,
1794 // and a second definition with a different value is a warning on every file, so the
1795 // condition is the same one that chose the directories.
1796 if cross.is_some() {
1797 let target = pinned.unwrap_or_else(|| opts.target.tuple());
1798 opts.glibc_minor = rucc_sysroot::bundled_glibc_minor(target).map_err(|skew| {
1799 err(format!(
1800 "{skew}; pin a release the tree has, or name a tree that has that one \
1801 with --sysroot"
1802 ))
1803 })?;
1804 }
1805 let system =
1806 library::header_dirs(opts.target, sysroot.as_deref(), cross.as_ref(), kernel.as_ref());
1807 // The two licence walls of `spec/cross-compile/13-distribution.md` section 13.4, which are
1808 // the only way step 3 comes back with nothing on a hosted target. Section 8.6 asks for the
1809 // answer to name the licence and the lawful ways to get what is behind it, rather than
1810 // leaving a person with an `#include` that failed as though a directory had gone missing.
1811 //
1812 // It is left on the search path instead of refused here, because a program that includes
1813 // none of the library needs none of the SDK and section 8.6 is explicit that targeting the
1814 // platform has to keep working. So the reason waits until an include has actually failed,
1815 // which is the only moment it helps and the only moment it is true.
1816 //
1817 // The condition is that step 3 found nothing at all, so an `SDKROOT`, an `INCLUDE` or a mac
1818 // with Xcode on it all pass through untouched, and `-nostdinc` never reaches this block. A
1819 // `--sysroot` or `-isysroot` passes through as well, even when the tree it names turns out to
1820 // be empty or absent: somebody who wrote a path has already answered the question this
1821 // message asks, and answering it again over the top of a mistyped directory would hide the
1822 // mistake behind a licence notice.
1823 if system.is_empty() && sysroot.is_none() {
1824 let tuple = pinned.unwrap_or_else(|| opts.target.tuple());
1825 if let Some(wall) = rucc_sysroot::Wall::of(tuple) {
1826 opts.search.explain_missing_system(wall.no_headers(&tuple.to_canonical_string()));
1827 }
1828 }
1829 // And whether the tree somebody named is the release they asked for, which is the one
1830 // question left once the directories are settled and the only place both halves of it are
1831 // known. Only for a named tree, because that is the case where the release in the target
1832 // stops deciding anything, and `crate::glibc` is where the rest of the reasoning is.
1833 if sysroot.is_some() {
1834 notes.extend(glibc::skew(opts.target, pinned, &system));
1835 }
1836 for dir in system {
1837 opts.search.push_system(dir);
1838 }
1839 }
1840 // Once, here, rather than as each directory is pushed. A `-I` that names a system
1841 // directory has to lose to the system entry and the system entry is added last, so the
1842 // question cannot be answered until the whole path is known.
1843 opts.search.remove_duplicates();
1844
1845 // The target has to be resolved before the configuration is printed, so this check comes
1846 // after the loop rather than at the point `--print-config` was seen.
1847 if print_config {
1848 return Ok(Action::PrintConfig(Box::new(opts)));
1849 }
1850 if print_pipeline {
1851 return Ok(Action::PrintPipeline(Box::new(opts)));
1852 }
1853 let plan = Plan::new(&opts, &inputs, output.as_deref()).map_err(|e| err(e.message))?;
1854 if print_plan {
1855 return Ok(Action::PrintPlan {
1856 opts: Box::new(opts),
1857 plan: Box::new(plan),
1858 link: Box::new(link),
1859 });
1860 }
1861 Ok(Action::Compile {
1862 opts: Box::new(opts),
1863 plan: Box::new(plan),
1864 link: Box::new(link),
1865 jobs,
1866 verbose,
1867 notes,
1868 })
1869}
1870
1871/// What `--fetch <tuple>` asked for, or why it is not a thing that can be done.
1872///
1873/// The lookup happens here rather than at the point the bytes would move, so that a target this
1874/// release pins nothing for is a refusal from the parser and the only code that runs a downloader is
1875/// code that already knows what it is getting.
1876///
1877/// # Errors
1878///
1879/// [`CliError`] when `--offline` forbade it, when there are input files as well, when the tuple is
1880/// not a target this compiler knows, when its sysroot is behind one of section 13.4's licence walls,
1881/// and when this release pins no artifact for it.
1882fn fetch_action(named: &str, offline: bool, inputs: &[Input]) -> Result<Action, CliError> {
1883 // Not a precedence question. Section 13.2 says `--offline` forbids a fetch entirely, so a
1884 // command line that writes both has asked for two opposite things and the answer is to say so
1885 // rather than to pick one of them.
1886 if offline {
1887 return Err(err(
1888 "--fetch asks for a download and --offline forbids every download, so this command \
1889 line asks for two opposite things. Drop one of them: --offline is how a build says it \
1890 will not reach the network, and --fetch is the only thing in this compiler that does",
1891 ));
1892 }
1893 if let Some(first) = inputs.first() {
1894 return Err(err(format!(
1895 "--fetch gets a sysroot and compiles nothing, so `{}` on the same command line is an \
1896 input that nothing would read",
1897 first.path
1898 )));
1899 }
1900 let target: TargetTuple = named
1901 .parse()
1902 .map_err(|why| err(format!("--fetch {named}: {why}, so there is no sysroot to get")))?;
1903 // The canonical spelling, because that is what a row is named by and what the directory under
1904 // the cache is called, and a person is free to write a tuple the long way round.
1905 let tuple = target.to_canonical_string();
1906 // Before the table is consulted, because a target behind a licence wall is not a row that has not
1907 // been written yet. Section 13.4 is that no release pins one of these ever, so the message says
1908 // the licence and the two lawful ways rather than naming the producer that will publish the rest.
1909 if let Some(wall) = rucc_sysroot::Wall::of(target) {
1910 return Err(err(format!("--fetch {tuple}: {}", wall.no_fetch(&tuple))));
1911 }
1912 let Some(what) = rucc_sysroot::pinned_for(&tuple) else {
1913 return Err(err(unpinned(&tuple)));
1914 };
1915 Ok(Action::Fetch { what, target, cache: cache::dir() })
1916}
1917
1918/// Why there is nothing to fetch for a target, which is a different sentence while the table is
1919/// empty.
1920///
1921/// A release that pins nothing and a release that pins eleven targets and not this one are two
1922/// situations, and a message that did not tell them apart would send somebody looking for a typo in
1923/// their tuple when the answer is that this work is not finished.
1924fn unpinned(tuple: &str) -> String {
1925 let pinned = rucc_sysroot::pinned_targets();
1926 if pinned.is_empty() {
1927 return format!(
1928 "this release pins no sysroot for {tuple}, and it pins none for any target yet. A \
1929 sysroot is built and published by the producer in tamnd/rucc-cross, per \
1930 spec/cross-compile/13-distribution.md section 13.8, and a release of this compiler \
1931 names one by URL and by hash afterwards. Until then, pass --sysroot=<dir> to compile \
1932 against a tree you have already"
1933 );
1934 }
1935 format!(
1936 "this release pins no sysroot for {tuple}. What it pins is {}. Pass --sysroot=<dir> to \
1937 compile against a tree you have already",
1938 pinned.join(", ")
1939 )
1940}
1941
1942/// Gets the artifact and installs it, saying what each step did.
1943///
1944/// The steps are section 13.8's and so are the messages: the transport is somebody else's program
1945/// and the check is ours, so a person reading this wants to know which downloader ran, that the
1946/// bytes matched, how many files the record named and where the tree ended up. A fetch of something
1947/// that is already there says that instead and moves nothing.
1948fn fetch_sysroot(what: &rucc_sysroot::Pinned, target: TargetTuple, cache: &std::path::Path) -> i32 {
1949 let tuple = target.to_canonical_string();
1950 let archive = what.archive_in(cache);
1951 let say = |line: &str| println!("rucc: {tuple}: {line}");
1952 match fetch::fetch(what.url, what.sha256, &archive) {
1953 Ok(fetch::Fetched::AlreadyThere) => {
1954 say(&format!("{} is already here and matches the hash", archive.display()));
1955 }
1956 Ok(fetch::Fetched::Downloaded(by)) => {
1957 say(&format!("downloaded {} with {}", what.url, by.program()));
1958 }
1959 Err(why) => return complain(why),
1960 }
1961 match install::install(&archive, what.sha256, target, cache) {
1962 Ok(done) => {
1963 match &done.before {
1964 install::Before::Nothing => {
1965 say(&format!("{} files installed at {}", done.files, done.root.display()));
1966 }
1967 install::Before::TheSame => {
1968 say(&format!(
1969 "the same sysroot is already at {}, so nothing moved",
1970 done.root.display()
1971 ));
1972 }
1973 install::Before::Different(was) => {
1974 say(&format!(
1975 "{} files installed at {}, over a tree whose record digested to {was}",
1976 done.files,
1977 done.root.display()
1978 ));
1979 }
1980 }
1981 say(&format!("the record digests to {}", done.digest));
1982 0
1983 }
1984 Err(why) => complain(why),
1985 }
1986}
1987
1988/// What one of the `-dump` and `-print` flags prints.
1989///
1990/// GCC prints the name back unchanged when it cannot find the file a `-print` flag asked about,
1991/// which is what makes the answer safe to paste into a link line whether or not the file is
1992/// there, and this does the same.
1993fn answer(query: &Query, opts: &Options, link: &LinkOptions) -> Result<String, CliError> {
1994 let found = |name: &str| {
1995 link::find_in_search(link, opts.target, name)
1996 .map_or_else(|| name.to_owned(), |path| path.display().to_string())
1997 };
1998 Ok(match query {
1999 Query::Machine => opts.target.to_string(),
2000 Query::Version => VERSION.to_owned(),
2001 Query::Multiarch => link::multiarch(opts.target),
2002 // The three lines GCC prints, in its order and with its punctuation, because what reads
2003 // them is a script written against that shape. There is no installation directory to
2004 // report: this compiler is one binary that works wherever it is copied, and the headers
2005 // it ships are inside it, so `install` is where the binary is and nothing is under it.
2006 Query::SearchDirs => {
2007 let here = std::env::current_exe()
2008 .ok()
2009 .and_then(|p| p.parent().map(std::path::Path::to_path_buf))
2010 .unwrap_or_default();
2011 let list = |dirs: &[PathBuf]| {
2012 dirs.iter().map(|d| d.display().to_string()).collect::<Vec<_>>().join(":")
2013 };
2014 let libraries = link::search_dirs(link, opts.target);
2015 format!(
2016 "install: {}\nprograms: ={}\nlibraries: ={}",
2017 here.display(),
2018 list(&link.prefixes),
2019 list(&libraries)
2020 )
2021 }
2022 // The root the rest of the answers are under, which a build system asks for when it wants
2023 // to find a file itself rather than ask for one by name, and which is the first thing to
2024 // look at when a cross build read a header nobody expected. A native compile has no
2025 // sysroot and the answer is the empty line, which is what GCC prints when it was
2026 // configured without one. `--sysroot` wins over ours because it wins everywhere else.
2027 Query::Sysroot => {
2028 sysroot_root(opts, link).map(|root| root.display().to_string()).unwrap_or_default()
2029 }
2030 // Section 13.5 of `spec/cross-compile/13-distribution.md`: for every input that is not this
2031 // compiler's own code, what it is, where it was got, its hash, its licence and whether it
2032 // was bundled, generated or fetched. What is printed is the manifest the sysroot already
2033 // carries rather than a second format saying the same things, because the three uses 13.5
2034 // gives for this are a licence notice, a reproducibility check and a security audit, and all
2035 // three are somebody else parsing it. One format is one parser to write.
2036 // Read and rendered rather than copied out, so that what comes back is the format this
2037 // build understands. The last newline comes off because whatever prints an answer adds
2038 // one, the way it does for every other query here. Keeping it would put a blank line at
2039 // the end of the one answer that is a file somebody diffs against the file it came from.
2040 Query::SysrootProvenance => match sysroot_manifest(opts, link)? {
2041 Some(manifest) => manifest.render().trim_end_matches('\n').to_string(),
2042 None => String::new(),
2043 },
2044 // Section 13.2 of the same document, which asks for the hash of a cache directory's
2045 // contents in the directory's name. A name cannot carry one, because the path has to be
2046 // computable before anything has been read, by the producer about to write the files and by
2047 // the compiler about to read them, and neither has the contents when it asks. So the number
2048 // is here instead, and it is the sha256 of the record rather than of a walk of the tree,
2049 // which means `sha256sum` over the manifest answers the same thing.
2050 Query::SysrootDigest => match sysroot_manifest(opts, link)? {
2051 Some(manifest) => manifest.digest(),
2052 None => String::new(),
2053 },
2054 Query::FileName(name) => found(name),
2055 // The name GCC gives the library of routines a compiler's output calls that the C
2056 // library does not have. Ours is built in and there is no file, so the answer is the
2057 // name itself, which is what GCC prints when it cannot find one either.
2058 Query::Libgcc => found("libgcc.a"),
2059 // A program rather than a library: the linker and the archiver are the ones a build asks
2060 // about, and this compiler finds them on the path or under `-B` rather than shipping
2061 // them, so the name back is the honest answer unless a `-B` prefix holds one.
2062 Query::ProgName(name) => link
2063 .prefixes
2064 .iter()
2065 .map(|dir| dir.join(name))
2066 .find(|path| path.is_file())
2067 .map_or_else(|| name.clone(), |path| path.display().to_string()),
2068 })
2069}
2070
2071/// The root every sysroot answer is about.
2072///
2073/// One function rather than a copy in each, because the other flags exist to say what is inside the
2074/// tree this one names, and two answers that disagreed about which tree that is would be a
2075/// difference nobody would think to look for. `--sysroot` wins over ours because it wins everywhere
2076/// else.
2077fn sysroot_root(opts: &Options, link: &LinkOptions) -> Option<PathBuf> {
2078 link.sysroot
2079 .clone()
2080 .or_else(|| link::cross_sysroot(opts.target, link).map(|at| at.root().to_path_buf()))
2081}
2082
2083/// The record of the sysroot this command line reads, when there is one to read.
2084///
2085/// [`None`] covers two cases that both print nothing, and they are different things. A compile for
2086/// this machine has no sysroot at all, and a tree somebody laid out themselves and pointed
2087/// `--sysroot` at carries no manifest, so nothing here knows where any of it came from. Saying
2088/// nothing is the only honest answer to either, and a reader can tell it from a manifest with no
2089/// inputs in it because that one still has its header lines.
2090///
2091/// # Errors
2092///
2093/// A manifest this build cannot parse, and anything else that went wrong reading the file. Passing a
2094/// record we could not read on to whoever asked would make their parser the one that finds the
2095/// problem, and every use section 13.5 gives for these two flags is somebody else reading the
2096/// output.
2097fn sysroot_manifest(opts: &Options, link: &LinkOptions) -> Result<Option<Manifest>, CliError> {
2098 let Some(root) = sysroot_root(opts, link) else {
2099 return Ok(None);
2100 };
2101 let path = Sysroot::at(root, opts.target.tuple()).manifest_path();
2102 match std::fs::read_to_string(&path) {
2103 Ok(text) => Manifest::parse(&text)
2104 .map(Some)
2105 .map_err(|why| err(format!("{}: {why}", path.display()))),
2106 Err(why) if why.kind() == std::io::ErrorKind::NotFound => Ok(None),
2107 Err(why) => Err(err(format!("{}: {why}", path.display()))),
2108 }
2109}
2110
2111/// Renders the passes this level will run, in order, with what each one does.
2112///
2113/// The level is the whole of the answer unless a `-f` flag edited it, which is section 9.1 of
2114/// `spec/09-optimizer.md`: a level is a list somebody wrote down rather than something that
2115/// emerges from which flags happen to be set, and this is how that list is read.
2116#[must_use]
2117pub fn print_pipeline(opts: &Options) -> String {
2118 let mut settings = rucc_opt::Options::for_level(opts.opt_level);
2119 settings.toggles.clone_from(&opts.passes);
2120 settings.global_fuel = opts.pass_fuel_global;
2121 for (on, spec) in &opts.pass_gates {
2122 // Every spelling was checked while the arguments were parsed, so there is nothing here
2123 // this can refuse, and a listing is not the place to report it if there were.
2124 let _ = settings.gates.add(*on, spec);
2125 }
2126 rucc_opt::pipeline::print(&settings)
2127}
2128
2129/// Renders the resolved configuration.
2130///
2131/// One `key: value` per line, sorted by nothing in particular but fixed in order, because
2132/// this output is diffed across hosts in CI and a reordering would read as a change.
2133#[must_use]
2134pub fn print_config(opts: &Options) -> String {
2135 let sess = Session::new(opts.clone());
2136 let t = &sess.target;
2137 let mut out = String::new();
2138 let _ = writeln!(out, "version: {VERSION}");
2139 // The three field triple the driver was given rather than the ten field tuple it widens to,
2140 // because this output is what a build system reads to find out what it asked for. The tuple is
2141 // the compiler's model of the machine and this line is a receipt for a command line.
2142 let _ = writeln!(out, "target: {}", opts.target);
2143 let _ = writeln!(out, "arch: {}", opts.target.arch.as_str());
2144 let _ = writeln!(out, "os: {}", opts.target.os.as_str());
2145 let _ = writeln!(out, "env: {}", opts.target.env.as_str());
2146 let _ = writeln!(out, "object-format: {}", t.object_format.as_str());
2147 let _ = writeln!(out, "pointer-width: {}", t.pointer_width);
2148 let _ = writeln!(out, "long-width: {}", t.long_width);
2149 let _ = writeln!(out, "long-double-width: {}", t.long_double_width);
2150 let _ = writeln!(out, "endian: {}", if t.little_endian { "little" } else { "big" });
2151 let _ = writeln!(out, "char-signed: {}", t.char_is_signed);
2152 let _ = writeln!(out, "va-list: {}", t.va_list.map_or("none", |list| list.as_str()));
2153 // The register file as a count per class, which is enough to tell a target whose registers
2154 // are described from one whose are not without printing sixteen names nobody asked for.
2155 let regs: Vec<String> = t
2156 .regs
2157 .classes()
2158 .map(|(class, info)| format!("{} {}", info.name, t.regs.len(class)))
2159 .collect();
2160 let _ = writeln!(
2161 out,
2162 "registers: {}",
2163 if regs.is_empty() { "none".to_string() } else { regs.join(", ") }
2164 );
2165 // What the schedule was chosen with, which is a sentence rather than a name on purpose: two
2166 // runs of a benchmark that disagree are usually two models and not two compilers.
2167 let _ = writeln!(out, "timing-model: {}", t.timing.map_or("none", |timing| timing.model));
2168 let _ = writeln!(out, "opt-level: {}", sess.opts.opt_level);
2169 let _ = writeln!(out, "safety: {}", sess.opts.safety);
2170 let _ = writeln!(out, "emit: {}", sess.opts.emit.as_str());
2171 let _ = writeln!(out, "debug-info: {}", sess.opts.debug_info);
2172 let _ = writeln!(out, "frame-pointer: {}", sess.opts.frame_pointer);
2173 let _ = writeln!(out, "red-zone: {}", sess.opts.red_zone);
2174 let _ = writeln!(out, "stack-protector: {}", sess.opts.protector);
2175 let _ = writeln!(out, "stack-clash-protection: {}", sess.opts.stack_clash);
2176 let _ = writeln!(out, "cf-protection: {}", sess.opts.control);
2177 let _ = writeln!(out, "patchable-function-entry: {}", sess.opts.patchable);
2178 let _ = writeln!(out, "profile: {}", sess.opts.profile);
2179 let _ = writeln!(out, "profile-hook: {}", sess.opts.hook);
2180 // Last because it is the one key with more than one line under it, and the only one
2181 // whose value is a property of the machine rather than of the command line.
2182 for dir in sess.opts.search.dirs() {
2183 let system = if dir.is_system { " (system)" } else { "" };
2184 let _ = writeln!(out, "include: {}{system}", dir.path.display());
2185 }
2186 out
2187}
2188
2189/// The output name the make target is taken from, which is the `-o` argument or nothing.
2190///
2191/// A run that stops at the preprocessor has not named an object, whatever its `-o` says: under
2192/// `-E` that argument is the preprocessed text and under `-M` it is the rule itself, and neither
2193/// is a file `make` would rebuild by running this rule. GCC agrees and falls back to the source
2194/// name in both, which is why a `-MD -E -o out.i` writes `out.d` holding a rule for `a.o`. From
2195/// `-S` on the argument does name what the rule builds, and it is used as written.
2196fn deps_target_output<'a>(opts: &Options, plan: &'a Plan) -> Option<&'a str> {
2197 if opts.emit == EmitKind::Preprocessed { None } else { plan.output.as_deref() }
2198}
2199
2200/// Writes to a path the command line named rather than one the plan derived, where `-` is
2201/// standard output.
2202fn write_named(path: &str, bytes: &[u8]) -> Result<(), String> {
2203 if path == "-" {
2204 return write_out(&Output::Stdout, bytes);
2205 }
2206 write_out(&Output::File(path.to_owned()), bytes)
2207}
2208
2209/// Writes the make rule for one input, and reports whether it got there.
2210///
2211/// A rule with no file of its own goes where the compilation it replaced would have written,
2212/// which is what makes the usual makefile recipe work: `rucc -M $< -o $@` leaves the rule in
2213/// `$@`, and the same line with the `-o` left off puts it on standard output.
2214fn write_deps(
2215 opts: &Options,
2216 plan: &Plan,
2217 job: &Job,
2218 found: &[Dependency],
2219 stderr: &mut impl std::io::Write,
2220) -> bool {
2221 let targets = if opts.deps.targets.is_empty() {
2222 vec![deps::default_target(&job.input, deps_target_output(opts, plan))]
2223 } else {
2224 opts.deps.targets.clone()
2225 };
2226 let rule = deps::rule(&opts.deps, &targets, &job.input, found);
2227 // The file, on the other hand, is named after the `-o` in every mode that still has one to
2228 // spend, which is every mode except the two that spend it on the rule.
2229 let wrote = match deps::default_file(&opts.deps, &job.input, plan.output.as_deref()) {
2230 // A `-MF` on a run that had nowhere else to put the rule leaves the file the `-o`
2231 // named empty rather than absent, because a makefile that named it as a target of its
2232 // own is a makefile that will look for it.
2233 Some(path) => write_named(&path, rule.as_bytes()).and_then(|()| {
2234 if opts.deps.instead_of_compiling { write_out(&job.output, b"") } else { Ok(()) }
2235 }),
2236 None => write_out(&job.output, rule.as_bytes()),
2237 };
2238 if let Err(e) = wrote {
2239 let _ = writeln!(stderr, "rucc: error: {e}");
2240 return false;
2241 }
2242 true
2243}
2244
2245/// Runs phase 4 over every input that has one, and writes what came out.
2246///
2247/// One input that fails does not stop the others. A build that reports every file it could
2248/// not preprocess in one run is worth more than one that stops at the first, and the exit
2249/// status is still a failure either way.
2250fn preprocess_all(opts: &Options, plan: &Plan) -> i32 {
2251 let fs = OsFileSystem::new();
2252 let mut stderr = std::io::stderr().lock();
2253 let mut failed = false;
2254 for job in &plan.jobs {
2255 if !job.phases.first().is_some_and(|p| *p == Phase::Preprocess) {
2256 // An input that is already preprocessed, or an object file. GCC passes these
2257 // through untouched, and the plan has already said so in its notes.
2258 continue;
2259 }
2260 let started = std::time::Instant::now();
2261 let result = preprocess(opts, &job.input, &fs);
2262 if opts.time {
2263 say_time(&job.input, started.elapsed(), &mut stderr);
2264 }
2265 for message in &result.messages {
2266 let _ = writeln!(stderr, "{message}");
2267 }
2268 if result.failed() {
2269 failed = true;
2270 continue;
2271 }
2272 if opts.deps.emit {
2273 failed |= !write_deps(opts, plan, job, &result.deps, &mut stderr);
2274 // `-M` and `-MM` asked for the rule instead of the text, so there is nothing else
2275 // to write. The other two asked for both and fall through to the text below.
2276 if opts.deps.instead_of_compiling {
2277 continue;
2278 }
2279 }
2280 if let Err(e) = write_out(&job.output, result.text.as_bytes()) {
2281 let _ = writeln!(stderr, "rucc: error: {e}");
2282 failed = true;
2283 }
2284 }
2285 i32::from(failed)
2286}
2287
2288/// Runs the front end over every input that has a compile phase, and writes what came out.
2289///
2290/// The same rule as [`preprocess_all`]: one input that fails does not stop the others, and the
2291/// exit status is a failure either way. An input that is already assembly or an object has no
2292/// compile phase and is passed over here, which the plan has already said in its notes.
2293fn compile_all(opts: &Options, plan: &Plan) -> i32 {
2294 let fs = OsFileSystem::new();
2295 let mut stderr = std::io::stderr().lock();
2296 let mut failed = false;
2297 let (mut remarks, ok) = Remarks::new(opts.opt_info_file.as_ref(), &mut stderr);
2298 failed |= !ok;
2299 let mut fired = Fired::new();
2300 let mut pressure = Pressure::new();
2301 for job in &plan.jobs {
2302 if !job.phases.contains(&Phase::Compile) {
2303 continue;
2304 }
2305 // An input of IR is read back rather than compiled, since the C it came from is not
2306 // here any more. Everything after this is the same, so the two paths meet again at the
2307 // messages and the file the result is written to.
2308 let started = std::time::Instant::now();
2309 let result = if job.kind == InputKind::Ir {
2310 compile_ir(opts, &job.input, &fs)
2311 } else {
2312 compile(opts, &job.input, &fs)
2313 };
2314 if opts.time {
2315 say_time(&job.input, started.elapsed(), &mut stderr);
2316 }
2317 fired.merge(&result.fired);
2318 pressure.merge(&result.pressure);
2319 failed |= !write_dumps(&job.input, &result.dumps, &mut stderr);
2320 failed |= !remarks.write(&result.remarks, &mut stderr);
2321 for message in &result.messages {
2322 let _ = writeln!(stderr, "{message}");
2323 }
2324 // Before the failure below, because a compilation that stopped in the back end is exactly
2325 // the one whose preprocessed source somebody wants to look at.
2326 failed |= !write_temps(job, &result.temps, &mut stderr);
2327 if result.failed() {
2328 failed = true;
2329 continue;
2330 }
2331 // `-MD` and `-MMD` write the rule beside the object and let the compilation happen, so
2332 // this is the one path where both files come out of the same run. An input of IR has no
2333 // dependencies to report and produces an empty list, which produces a rule naming only
2334 // itself, and that is the honest answer rather than a missing file.
2335 if opts.deps.emit {
2336 failed |= !write_deps(opts, plan, job, &result.deps, &mut stderr);
2337 }
2338 if let Err(e) = write_out(&job.output, result.artifact.bytes()) {
2339 let _ = writeln!(stderr, "rucc: error: {e}");
2340 failed = true;
2341 }
2342 }
2343 failed |= !write_coverage(opts, &fired, &mut stderr);
2344 failed |= !write_pressure(opts, &pressure, &mut stderr);
2345 i32::from(failed)
2346}
2347
2348/// A directory for the object files only the link step ever sees, removed when it goes away.
2349///
2350/// `-c` writes its object where the user can see it and linking does not, which is the whole of
2351/// the difference: a `rucc a.c b.c` leaves an executable behind and nothing else, the same as
2352/// every other compiler. Removing them on drop rather than at the end of a function is so that a
2353/// link that failed leaves nothing behind either.
2354struct Scratch {
2355 /// Where the objects go.
2356 dir: PathBuf,
2357}
2358
2359impl Scratch {
2360 /// Makes one, under whatever the platform calls its temporary directory.
2361 ///
2362 /// The name carries the process id so that two compilers running at once do not share a
2363 /// directory, which they would otherwise do the moment two of them compiled a file of the
2364 /// same name.
2365 fn new() -> Result<Scratch, String> {
2366 let dir = std::env::temp_dir().join(format!("rucc-{}", std::process::id()));
2367 std::fs::create_dir_all(&dir).map_err(|e| format!("{}: {e}", dir.display()))?;
2368 Ok(Scratch { dir })
2369 }
2370}
2371
2372impl Drop for Scratch {
2373 fn drop(&mut self) {
2374 let _ = std::fs::remove_dir_all(&self.dir);
2375 }
2376}
2377
2378/// The link line the plan describes, for `-###`.
2379///
2380/// The names in it are the hints the plan carries rather than the temporaries a real compilation
2381/// would choose, because `-###` prints the line without having compiled anything and so has
2382/// nothing to point at. That also makes the printed line readable rather than naming a directory
2383/// that only exists while a compilation is running.
2384fn link_line(opts: &Options, link: &LinkOptions, job: &LinkJob) -> Result<String, link::Error> {
2385 let linker = link::find(opts.target, link)?;
2386 let args = link::line(opts.target, link, &job.inputs, &job.output)?;
2387 Ok(link::render(&linker, &args))
2388}
2389
2390/// Compiles everything, then links it.
2391///
2392/// The objects go in a directory that is removed afterwards, which is why this is not
2393/// [`compile_all`] followed by a link: the plan says an object feeding the linker is temporary
2394/// and does not say where, because where is a question that only has an answer once something is
2395/// running.
2396fn link_all(opts: &Options, plan: &Plan, link: &LinkOptions, verbose: bool) -> i32 {
2397 let Some(job) = &plan.link else {
2398 // Every path into here comes from a plan whose last phase is the link, and such a plan
2399 // has a link job. Saying so is cheaper than an unwrap that would have to be explained.
2400 let mut stderr = std::io::stderr().lock();
2401 let _ = writeln!(stderr, "rucc: error: there is nothing to link");
2402 return 1;
2403 };
2404 // Before anything is compiled, because a linker that is not on the machine is worth knowing
2405 // about in the second it takes to look rather than after the compilation.
2406 // And before that, whether this link has a line at all and whether what it reads is on the
2407 // machine. Both are answerable now, and a target whose sysroot has not been built is worth
2408 // saying so about before the compilation rather than after it.
2409 if let Err(why) = link::preflight(opts.target, link) {
2410 return complain(why);
2411 }
2412 let linker = match link::find(opts.target, link) {
2413 Ok(linker) => linker,
2414 Err(why) => return complain(why),
2415 };
2416
2417 let scratch = match Scratch::new() {
2418 Ok(scratch) => scratch,
2419 Err(why) => return complain(format!("could not make a place for the object files: {why}")),
2420 };
2421
2422 let fs = OsFileSystem::new();
2423 let mut failed = false;
2424 // One per job, in job order, which is what lets the link line below be rebuilt with the real
2425 // paths in it: every job contributes exactly one file to the line and does so in this order.
2426 let mut produced: Vec<String> = Vec::with_capacity(plan.jobs.len());
2427 let mut fired = Fired::new();
2428 let mut pressure = Pressure::new();
2429 {
2430 let mut stderr = std::io::stderr().lock();
2431 let (mut remarks, ok) = Remarks::new(opts.opt_info_file.as_ref(), &mut stderr);
2432 failed |= !ok;
2433 for (at, job) in plan.jobs.iter().enumerate() {
2434 let out = match &job.output {
2435 Output::Temporary(hint) => {
2436 // The index because two inputs in different directories can have the same
2437 // name, and the two objects of `rucc a/x.c b/x.c` must not be one file.
2438 scratch.dir.join(format!("{at}-{hint}")).display().to_string()
2439 }
2440 Output::File(path) => path.clone(),
2441 // A job feeding the linker never writes to standard output, since the plan gives
2442 // it a temporary. This is here so that the match is total rather than a panic.
2443 Output::Stdout => continue,
2444 };
2445 produced.push(out.clone());
2446 if !job.phases.contains(&Phase::Compile) {
2447 continue;
2448 }
2449 let started = std::time::Instant::now();
2450 let result = if job.kind == InputKind::Ir {
2451 compile_ir(opts, &job.input, &fs)
2452 } else {
2453 compile(opts, &job.input, &fs)
2454 };
2455 if opts.time {
2456 say_time(&job.input, started.elapsed(), &mut stderr);
2457 }
2458 fired.merge(&result.fired);
2459 pressure.merge(&result.pressure);
2460 failed |= !write_dumps(&job.input, &result.dumps, &mut stderr);
2461 failed |= !remarks.write(&result.remarks, &mut stderr);
2462 for message in &result.messages {
2463 let _ = writeln!(stderr, "{message}");
2464 }
2465 failed |= !write_temps(job, &result.temps, &mut stderr);
2466 if result.failed() {
2467 failed = true;
2468 continue;
2469 }
2470 // A `-MD` on a command line that links writes the rule next to the executable and
2471 // names the executable as its target, since that is the file this source builds
2472 // here. The object it went through is in a temporary directory and is gone by the
2473 // time `make` reads any of this.
2474 if opts.deps.emit {
2475 failed |= !write_deps(opts, plan, job, &result.deps, &mut stderr);
2476 }
2477 if !matches!(result.artifact, Artifact::Object { .. }) {
2478 // Worth saying rather than writing whatever it is and letting the linker read it.
2479 // An empty file is a valid empty linker script, so a link handed one gets as far
2480 // as reporting every symbol of this file undefined, which is a page of messages
2481 // about something that went wrong here.
2482 let _ = writeln!(
2483 stderr,
2484 "rucc: internal error: {}: no object file was produced for the link",
2485 job.input
2486 );
2487 failed = true;
2488 continue;
2489 }
2490 if let Err(e) = std::fs::write(&out, result.artifact.bytes()) {
2491 let _ = writeln!(stderr, "rucc: error: {out}: {e}");
2492 failed = true;
2493 }
2494 }
2495 failed |= !write_coverage(opts, &fired, &mut stderr);
2496 failed |= !write_pressure(opts, &pressure, &mut stderr);
2497 }
2498 if failed {
2499 // Nothing is linked from a compilation that did not finish. A linker run over the objects
2500 // that did compile would report every function of the file that did not as undefined,
2501 // which is a page of messages about a mistake already reported once.
2502 return 1;
2503 }
2504
2505 // The items in command line order with the temporaries filled in. A library contributes no
2506 // job and passes through, and every file item takes the next job's real output, which is
2507 // what keeps a library that was written between two objects between them here.
2508 let mut outputs = produced.into_iter();
2509 let mut items = Vec::with_capacity(job.inputs.len());
2510 for item in &job.inputs {
2511 match item {
2512 link::Item::Library(name) => items.push(link::Item::Library(name.clone())),
2513 link::Item::File(_) => match outputs.next() {
2514 Some(path) => items.push(link::Item::File(path)),
2515 None => return complain("the plan asks the linker for a file nothing produced"),
2516 },
2517 }
2518 }
2519
2520 let args = match link::line(opts.target, link, &items, &job.output) {
2521 Ok(args) => args,
2522 Err(why) => return complain(why),
2523 };
2524 if verbose {
2525 let mut stderr = std::io::stderr().lock();
2526 let _ = writeln!(stderr, "{}", link::render(&linker, &args));
2527 }
2528 let started = std::time::Instant::now();
2529 let ran = link::run(&linker, &args);
2530 if opts.time {
2531 // The one step of a compilation that really is another program, so this line is the same
2532 // measurement gcc's is and names the linker the way gcc names `collect2`.
2533 let mut stderr = std::io::stderr().lock();
2534 say_time(&linker.name, started.elapsed(), &mut stderr);
2535 }
2536 match ran {
2537 Ok(()) => 0,
2538 // The linker has already said what was wrong on its own error output, and repeating that
2539 // linking failed would only push its message further up the screen.
2540 Err(link::Error::Refused { .. }) => 1,
2541 Err(why) => complain(why),
2542 }
2543}
2544
2545/// Compiles everything and writes the objects into one static library.
2546///
2547/// No temporary directory and no second program. The objects never reach the file system at all:
2548/// they go from the compiler into the archive writer, which is both faster than writing a directory
2549/// of files for an `ar` to read back and the reason the symbol index can be written at all. A
2550/// member's index entries are the names the object writer says it wrote, and the only thing that
2551/// knows those is the run that wrote it.
2552///
2553/// `-save-temps` is the exception. It asked for the objects to be kept, the plan gave them names a
2554/// person can find, and they are written there as well as put in the archive.
2555fn archive_all(opts: &Options, plan: &Plan) -> i32 {
2556 let Some(job) = &plan.archive else {
2557 // Every path into here comes from a plan whose last phase is the archive, and such a plan
2558 // has an archive job. Saying so is cheaper than an unwrap that would have to be explained.
2559 return complain("there is nothing to put in an archive");
2560 };
2561 // Before anything is compiled, because a format this has no container for is worth knowing
2562 // about in the second it takes to look rather than after the whole compilation.
2563 let flavour = match opts.target.os.object_format() {
2564 ObjectFormat::Elf => rucc_archive::Flavour::Gnu,
2565 ObjectFormat::Coff => rucc_archive::Flavour::Coff,
2566 // Mach-O wants the BSD flavour, whose index is a different member under a different name,
2567 // and wasm has no archives of its own at all. Neither has an object writer either, so a
2568 // command line reaching this would have failed in the next step regardless.
2569 format @ (ObjectFormat::MachO | ObjectFormat::Wasm) => {
2570 return complain(format!(
2571 "there is no archive format for {} objects in this compiler yet",
2572 format.as_str()
2573 ));
2574 }
2575 };
2576
2577 let fs = OsFileSystem::new();
2578 let mut failed = false;
2579 let mut members: Vec<rucc_archive::Member> = Vec::with_capacity(plan.jobs.len());
2580 let mut names = job.members.iter();
2581 let mut fired = Fired::new();
2582 let mut pressure = Pressure::new();
2583 {
2584 let mut stderr = std::io::stderr().lock();
2585 let (mut remarks, ok) = Remarks::new(opts.opt_info_file.as_ref(), &mut stderr);
2586 failed |= !ok;
2587 for plan_job in &plan.jobs {
2588 // What the plan called this member. The two lists are walked together rather than the
2589 // name being worked out again here, so that what `-###` printed and what goes in the
2590 // file cannot come apart.
2591 let Some(member) = names.next() else {
2592 return complain("the plan asks the archive for a member nothing produced");
2593 };
2594 if !plan_job.phases.contains(&Phase::Compile) {
2595 // Assembly, which enters the pipeline after the compile phase. There is no
2596 // assembler for a file of text in this compiler, so there is no object to put in,
2597 // and an archive quietly missing one is worse than a message about it.
2598 let _ = writeln!(
2599 &mut stderr,
2600 "rucc: error: {}: this compiler has no assembler for a file of assembly yet, \
2601 so it cannot go into an archive",
2602 plan_job.input
2603 );
2604 failed = true;
2605 continue;
2606 }
2607 let started = std::time::Instant::now();
2608 let result = if plan_job.kind == InputKind::Ir {
2609 compile_ir(opts, &plan_job.input, &fs)
2610 } else {
2611 compile(opts, &plan_job.input, &fs)
2612 };
2613 if opts.time {
2614 say_time(&plan_job.input, started.elapsed(), &mut stderr);
2615 }
2616 fired.merge(&result.fired);
2617 pressure.merge(&result.pressure);
2618 failed |= !write_dumps(&plan_job.input, &result.dumps, &mut stderr);
2619 failed |= !remarks.write(&result.remarks, &mut stderr);
2620 for message in &result.messages {
2621 let _ = writeln!(stderr, "{message}");
2622 }
2623 failed |= !write_temps(plan_job, &result.temps, &mut stderr);
2624 if result.failed() {
2625 failed = true;
2626 continue;
2627 }
2628 if opts.deps.emit {
2629 failed |= !write_deps(opts, plan, plan_job, &result.deps, &mut stderr);
2630 }
2631 let Artifact::Object { bytes, defines } = result.artifact else {
2632 let _ = writeln!(
2633 stderr,
2634 "rucc: internal error: {}: no object file was produced for the archive",
2635 plan_job.input
2636 );
2637 failed = true;
2638 continue;
2639 };
2640 // Under `-save-temps` the plan gave the object a name a person can find, so it is
2641 // written there too. Otherwise it is only ever a member and never a file.
2642 if let Output::File(path) = &plan_job.output {
2643 if let Err(e) = std::fs::write(path, &bytes) {
2644 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2645 failed = true;
2646 }
2647 }
2648 members.push(rucc_archive::Member { name: member.clone(), body: bytes, defines });
2649 }
2650 failed |= !write_coverage(opts, &fired, &mut stderr);
2651 failed |= !write_pressure(opts, &pressure, &mut stderr);
2652 }
2653 if failed {
2654 // Nothing is written from a compilation that did not finish, for the reason the link gives:
2655 // an archive missing the file that failed is one a link reports every name of as undefined,
2656 // which is a page of messages about a mistake already reported once.
2657 return 1;
2658 }
2659
2660 let bytes = match rucc_archive::write(flavour, &members) {
2661 Ok(bytes) => bytes,
2662 // Every one of these is a bug here rather than a program's mistake: the names came from the
2663 // object writer and the bodies came from this process.
2664 Err(why) => return complain(format!("the archive could not be written: {why}")),
2665 };
2666 match std::fs::write(&job.output, &bytes) {
2667 Ok(()) => 0,
2668 Err(e) => complain(format!("{}: {e}", job.output)),
2669 }
2670}
2671
2672/// Prints one driver level message and gives back the exit status that goes with it.
2673fn complain(why: impl std::fmt::Display) -> i32 {
2674 let mut stderr = std::io::stderr().lock();
2675 let _ = writeln!(stderr, "rucc: error: {why}");
2676 1
2677}
2678
2679/// Writes what `-Zrule-coverage=FILE` asked for, and says whether it could.
2680///
2681/// Once for the whole command line rather than once per input, because the question is which
2682/// lowering rules this run of the compiler reached and a file per input would leave the reader
2683/// unioning files to find out something one process already knew.
2684///
2685/// A file that could not be written is a failure and not a warning. What asks for this is a
2686/// measurement run, and a measurement that quietly did not happen is worse than one that stopped.
2687fn write_coverage(opts: &Options, fired: &Fired, stderr: &mut impl std::io::Write) -> bool {
2688 let Some(path) = &opts.rule_coverage else { return true };
2689 let Some(table) = coverage::table(opts.target.arch) else {
2690 let _ = writeln!(
2691 stderr,
2692 "rucc: error: there are no lowering rules for {} yet, so there is no coverage of them \
2693 to report",
2694 opts.target
2695 );
2696 return false;
2697 };
2698 match std::fs::write(path, fired.listing(table)) {
2699 Ok(()) => true,
2700 Err(e) => {
2701 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2702 false
2703 }
2704 }
2705}
2706
2707/// Writes what `-Zregister-pressure=FILE` asked for, and says whether it could.
2708///
2709/// Once for the whole command line, for the reason [`write_coverage`] gives, and a file that could
2710/// not be written is a failure for the reason it gives too. There is no equivalent of the missing
2711/// rule table here, since every target this compiles for has an allocator, and a run that reached
2712/// no back end at all writes an empty listing rather than nothing: a measurement of a build that
2713/// produced no code is still an answer and it is the honest one.
2714fn write_pressure(opts: &Options, pressure: &Pressure, stderr: &mut impl std::io::Write) -> bool {
2715 let Some(path) = &opts.register_pressure else { return true };
2716 match std::fs::write(path, pressure.listing()) {
2717 Ok(()) => true,
2718 Err(e) => {
2719 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2720 false
2721 }
2722 }
2723}
2724
2725/// Where the `-fopt-info` remarks go, and how much of the run has already gone there.
2726///
2727/// Standard error by default, and one file for the whole run when `-fopt-info=<file>` named one.
2728/// A file rather than the diagnostic stream is what a harness wants: the corpus in
2729/// `tamnd/rucc-corpus` matches a rejection against what the compiler said on standard error, and
2730/// a few thousand remarks mixed into that would bury it.
2731struct Remarks {
2732 /// The file, if there is one.
2733 file: Option<String>,
2734 /// Whether anything has been written to it yet, which decides between truncating and
2735 /// appending. One file holds the whole run rather than the last input in it.
2736 started: bool,
2737}
2738
2739impl Remarks {
2740 /// Prepares the destination, emptying the file if there is one.
2741 ///
2742 /// Emptied here rather than at the first remark, because a run where no pass had anything to
2743 /// say should leave an empty file and not yesterday's. An absent file and an empty one are
2744 /// different facts and something reading this will act on the difference.
2745 fn new(file: Option<&String>, stderr: &mut impl std::io::Write) -> (Self, bool) {
2746 let mut ok = true;
2747 if let Some(path) = file {
2748 if let Err(e) = std::fs::write(path, "") {
2749 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2750 ok = false;
2751 }
2752 }
2753 (Self { file: file.cloned(), started: false }, ok)
2754 }
2755
2756 /// Writes one input's remarks, and says whether that worked.
2757 ///
2758 /// A file that cannot be written is a failure and not a warning, for the reason
2759 /// [`write_dumps`] gives: remarks that quietly did not arrive look exactly like a compilation
2760 /// where nothing happened.
2761 fn write(&mut self, text: &str, stderr: &mut impl std::io::Write) -> bool {
2762 if text.is_empty() {
2763 return true;
2764 }
2765 let Some(path) = &self.file else {
2766 let _ = write!(stderr, "{text}");
2767 return true;
2768 };
2769 let opened = std::fs::OpenOptions::new()
2770 .write(true)
2771 .append(self.started)
2772 .truncate(!self.started)
2773 .create(true)
2774 .open(path);
2775 self.started = true;
2776 let result =
2777 opened.and_then(|mut file| std::io::Write::write_all(&mut file, text.as_bytes()));
2778 if let Err(e) = result {
2779 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2780 return false;
2781 }
2782 true
2783 }
2784}
2785
2786/// Writes what `-fdump-ir=` asked to see, one file per dump.
2787///
2788/// The name is the input file with the dump's own name and `.ir` after it, so a directory listing
2789/// after a run is the passes in the order they ran, per input. They go in the working directory
2790/// rather than beside the output, because a dump is something a person asked for at a prompt and
2791/// the working directory is where that person is.
2792///
2793/// A file that could not be written is a failure and not a warning, for the reason
2794/// [`write_coverage`] gives: what asked for this is somebody debugging a pass, and a dump that
2795/// quietly did not happen looks exactly like a pass that did not run.
2796fn write_dumps(input: &str, dumps: &[rucc_opt::Dump], stderr: &mut impl std::io::Write) -> bool {
2797 let stem = std::path::Path::new(input)
2798 .file_name()
2799 .map_or_else(|| input.to_owned(), |name| name.to_string_lossy().into_owned());
2800 let mut ok = true;
2801 for dump in dumps {
2802 let path = format!("{stem}.{}.ir", dump.name);
2803 if let Err(e) = std::fs::write(&path, &dump.text) {
2804 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2805 ok = false;
2806 }
2807 }
2808 ok
2809}
2810
2811/// Writes the files `-save-temps` kept, which is nothing at all unless it was given.
2812///
2813/// A file that could not be written is a failure rather than a warning, for the reason
2814/// [`write_dumps`] gives: somebody asked for these by name, and one that quietly did not happen
2815/// looks like a compilation that never went through that step.
2816fn write_temps(job: &Job, temps: &Temps, stderr: &mut impl std::io::Write) -> bool {
2817 let mut ok = true;
2818 let kept = [(job.saved_text(), &temps.preprocessed), (job.saved_asm(), &temps.assembly)];
2819 for (path, text) in kept {
2820 // A step the compilation did not reach has nothing to keep, and a job that is not keeping
2821 // that step has nowhere to put it. Either way there is no file here.
2822 let (Some(path), Some(text)) = (path, text) else { continue };
2823 if let Err(e) = std::fs::write(&path, text) {
2824 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2825 ok = false;
2826 }
2827 }
2828 ok
2829}
2830
2831/// One line of `-time`, which is what a step was called and how long it took.
2832///
2833/// GCC's two numbers are the user and the system time of a subprocess it ran. This compiler runs
2834/// no subprocess for anything but the link, so what is measured here is the wall clock of the
2835/// step and the second column is always zero. The shape of the line is kept because a person
2836/// reading it next to gcc's should not have to work out which column is which.
2837fn say_time(name: &str, took: std::time::Duration, stderr: &mut impl std::io::Write) {
2838 let _ = writeln!(stderr, "# {name} {:.2} {:.2}", took.as_secs_f64(), 0.0);
2839}
2840
2841/// Writes one job's result where the plan said it goes.
2842///
2843/// # Errors
2844///
2845/// Returns the message to print, which names the file when there is one, because "permission
2846/// denied" on its own does not say which file was refused.
2847fn write_out(output: &Output, bytes: &[u8]) -> Result<(), String> {
2848 match output {
2849 Output::Stdout => {
2850 let mut stdout = std::io::stdout().lock();
2851 stdout.write_all(bytes).map_err(|e| format!("writing to standard output: {e}"))
2852 }
2853 Output::File(path) | Output::Temporary(path) => {
2854 std::fs::write(path, bytes).map_err(|e| format!("{path}: {e}"))
2855 }
2856 }
2857}
2858
2859/// Runs the driver and returns the process exit code.
2860///
2861/// `args` excludes the program name. Output goes to `stdout` and errors to `stderr`, which
2862/// is the one place in the compiler that is true.
2863pub fn run(args: &[String]) -> i32 {
2864 match parse_args(args) {
2865 Ok(Action::Help) => {
2866 print!("{USAGE}");
2867 0
2868 }
2869 Ok(Action::Version) => {
2870 println!("rucc {VERSION}");
2871 0
2872 }
2873 Ok(Action::Print(line)) => {
2874 println!("{line}");
2875 0
2876 }
2877 Ok(Action::PrintConfig(opts)) => {
2878 print!("{}", print_config(&opts));
2879 0
2880 }
2881 Ok(Action::PrintPipeline(opts)) => {
2882 print!("{}", print_pipeline(&opts));
2883 0
2884 }
2885 Ok(Action::PrintPlan { opts, plan, link }) => {
2886 print!("{}", plan.render());
2887 // The line as it would be typed, which is the half of `-###` that section 4.3 says
2888 // arrives with the link. It is printed even when the linker is not on this machine,
2889 // because what a build wants from `-###` is what the compiler would do.
2890 if let Some(job) = &plan.link {
2891 match link_line(&opts, &link, job) {
2892 Ok(line) => println!("{line}"),
2893 Err(why) => {
2894 let mut stderr = std::io::stderr().lock();
2895 let _ = writeln!(stderr, "rucc: error: {why}");
2896 return 1;
2897 }
2898 }
2899 }
2900 0
2901 }
2902 Ok(Action::Fetch { what, target, cache }) => fetch_sysroot(what, target, &cache),
2903 Ok(Action::Compile { opts, plan, link, jobs, verbose, notes }) => {
2904 {
2905 let mut stderr = std::io::stderr().lock();
2906 // Before the plan rather than after it, because a note is about the command line
2907 // and the plan is what the command line was read as, so the reader wants the two
2908 // in that order.
2909 for note in ¬es {
2910 let _ = writeln!(stderr, "rucc: warning: {note}");
2911 }
2912 if verbose {
2913 let _ = write!(stderr, "{}", plan.render());
2914 let _ = writeln!(stderr, "workers: {}", jobs.count());
2915 }
2916 }
2917 if opts.emit == EmitKind::Preprocessed {
2918 return preprocess_all(&opts, &plan);
2919 }
2920 if opts.emit == EmitKind::Archive {
2921 return archive_all(&opts, &plan);
2922 }
2923 if opts.emit != EmitKind::Executable {
2924 return compile_all(&opts, &plan);
2925 }
2926 link_all(&opts, &plan, &link, verbose)
2927 }
2928 Err(e) => {
2929 let mut stderr = std::io::stderr().lock();
2930 let _ = writeln!(stderr, "rucc: error: {e}");
2931 let _ = writeln!(stderr, "rucc: note: run `rucc --help` for usage");
2932 1
2933 }
2934 }
2935}
2936
2937#[cfg(test)]
2938mod tests {
2939 use rucc_session::{
2940 Contract, GnucVersion, IncludeForm, LtoJobs, OptLevel, Partition, Patchable, Visibility,
2941 };
2942
2943 use super::*;
2944
2945 fn args(s: &[&str]) -> Vec<String> {
2946 s.iter().map(|x| (*x).to_owned()).collect()
2947 }
2948
2949 #[test]
2950 fn help_and_version_win_over_everything_else() {
2951 assert_eq!(parse_args(&args(&["-c", "--help", "x.c"])).unwrap(), Action::Help);
2952 assert_eq!(parse_args(&args(&["--version"])).unwrap(), Action::Version);
2953 }
2954
2955 fn compile(s: &[&str]) -> (Box<Options>, Box<Plan>) {
2956 match parse_args(&args(s)).expect("expected a compilation") {
2957 Action::Compile { opts, plan, .. } => (opts, plan),
2958 other => panic!("expected a compilation, got {other:?}"),
2959 }
2960 }
2961
2962 fn linking(s: &[&str]) -> (Box<LinkOptions>, Box<Plan>) {
2963 match parse_args(&args(s)).expect("expected a compilation") {
2964 Action::Compile { link, plan, .. } => (link, plan),
2965 other => panic!("expected a compilation, got {other:?}"),
2966 }
2967 }
2968
2969 fn notes(s: &[&str]) -> Vec<String> {
2970 match parse_args(&args(s)).expect("expected a compilation") {
2971 Action::Compile { notes, .. } => notes,
2972 other => panic!("expected a compilation, got {other:?}"),
2973 }
2974 }
2975
2976 /// The ordinary command line has nothing to say about itself, which is the property that makes
2977 /// a note worth reading when there is one.
2978 #[test]
2979 fn a_command_line_with_nothing_wrong_with_it_carries_no_notes() {
2980 assert_eq!(notes(&["-c", "a.c"]), Vec::<String>::new());
2981 }
2982
2983 /// A directory that is not there contributes nothing to the search path, so there is no tree to
2984 /// read a release out of and nothing to compare the pin against. Said as a test because this is
2985 /// the shape a hermetic machine takes: the probe reads the disk and every other machine has a
2986 /// different disk, so what can be asserted here is the silence.
2987 #[test]
2988 fn a_named_tree_that_is_not_on_the_machine_is_not_a_release_mismatch() {
2989 let said =
2990 notes(&["--target=x86_64-linux-gnu.2.28", "--sysroot=/nowhere-at-all", "-c", "a.c"]);
2991 assert_eq!(said, Vec::<String>::new());
2992 }
2993
2994 #[test]
2995 fn collects_inputs_and_flags() {
2996 let (opts, plan) = compile(&["-c", "-O2", "-g", "a.c", "b.c"]);
2997 let paths: Vec<&str> = plan.jobs.iter().map(|j| j.input.as_str()).collect();
2998 assert_eq!(paths, vec!["a.c", "b.c"]);
2999 assert_eq!(opts.opt_level, OptLevel::O2);
3000 assert_eq!(opts.emit, EmitKind::Object);
3001 assert!(opts.debug_info);
3002 }
3003
3004 /// The unstable options, which are spelled apart from everything else on purpose: what is
3005 /// under `-Z` promises nothing, and a build that reaches for one should have had to say so.
3006 #[test]
3007 fn an_unstable_option_is_taken_and_one_that_does_not_exist_is_refused() {
3008 let (opts, _) = compile(&["-c", "-Zrule-coverage=/tmp/rules.cov", "a.c"]);
3009 assert_eq!(opts.rule_coverage.as_deref(), Some("/tmp/rules.cov"));
3010
3011 let (plain, _) = compile(&["-c", "a.c"]);
3012 assert_eq!(plain.rule_coverage, None, "nothing is measured unless it was asked for");
3013
3014 assert!(parse_args(&args(&["-Zrule-coverage=", "a.c"])).is_err(), "a file with no name");
3015 let unknown = parse_args(&args(&["-Zwhat", "a.c"])).expect_err("there is no such option");
3016 assert!(unknown.message.contains("4.11"), "{}", unknown.message);
3017 }
3018
3019 /// The other measurement written to a file, which reads the same way and fails the same way.
3020 #[test]
3021 fn where_the_register_pressure_goes_is_asked_for_the_same_way() {
3022 let (opts, _) = compile(&["-c", "-O2", "-Zregister-pressure=/tmp/spills.txt", "a.c"]);
3023 assert_eq!(opts.register_pressure.as_deref(), Some("/tmp/spills.txt"));
3024
3025 let (plain, _) = compile(&["-c", "a.c"]);
3026 assert_eq!(plain.register_pressure, None, "nothing is measured unless it was asked for");
3027
3028 assert!(parse_args(&args(&["-Zregister-pressure=", "a.c"])).is_err(), "no file named");
3029 }
3030
3031 /// Scheduling, which has the three way answer every optimization flag has: on, off, and
3032 /// nothing said, which is whatever the optimization level asks for. The name is gcc's, and
3033 /// gcc's has a two in it because gcc has a scheduler before allocation and one after and this
3034 /// is the one after.
3035 #[test]
3036 fn scheduling_can_be_turned_on_and_off_and_left_to_the_optimization_level() {
3037 let (on, _) = compile(&["-c", "-O0", "-fschedule-insns2", "a.c"]);
3038 assert_eq!(on.schedule_insns, Some(true));
3039
3040 let (off, _) = compile(&["-c", "-O2", "-fno-schedule-insns2", "a.c"]);
3041 assert_eq!(off.schedule_insns, Some(false));
3042
3043 let (quiet, _) = compile(&["-c", "-O2", "a.c"]);
3044 assert_eq!(quiet.schedule_insns, None, "nothing said, so the level decides");
3045 assert!(quiet.opt_level.schedules(), "and at this level the level says yes");
3046
3047 let (none, _) = compile(&["-c", "a.c"]);
3048 assert!(!none.opt_level.schedules(), "at no optimization it says no");
3049 }
3050
3051 /// Whether the timing model is worth holding an instruction back over, which is a `-Z` because
3052 /// it is a question about a target's description rather than about the program being compiled.
3053 #[test]
3054 fn whether_the_timing_model_is_cycle_accurate_can_be_overridden() {
3055 let (yes, _) = compile(&["-c", "-O2", "-Zcycle-accurate-model=yes", "a.c"]);
3056 assert_eq!(yes.cycle_accurate_model, Some(true));
3057
3058 let (no, _) = compile(&["-c", "-O2", "-Zcycle-accurate-model=no", "a.c"]);
3059 assert_eq!(no.cycle_accurate_model, Some(false));
3060
3061 let (plain, _) = compile(&["-c", "-O2", "a.c"]);
3062 assert_eq!(plain.cycle_accurate_model, None, "the target's own answer stands");
3063
3064 let bad = parse_args(&args(&["-Zcycle-accurate-model=maybe", "a.c"]))
3065 .expect_err("it takes yes or no");
3066 assert!(bad.message.contains("yes or no"), "{}", bad.message);
3067 }
3068
3069 #[test]
3070 fn a_bare_dash_o_means_o1_the_way_gcc_reads_it() {
3071 let (opts, _) = compile(&["-O", "a.c"]);
3072 assert_eq!(opts.opt_level, OptLevel::O1);
3073 }
3074
3075 #[test]
3076 fn dash_x_applies_to_later_inputs_only_and_none_stops_it() {
3077 let (_, plan) = compile(&["a.o", "-x", "c", "b.txt", "-x", "none", "c.o"]);
3078 assert_eq!(plan.jobs[0].kind, InputKind::LinkerInput);
3079 assert_eq!(plan.jobs[1].kind, InputKind::C);
3080 assert_eq!(plan.jobs[2].kind, InputKind::LinkerInput);
3081 }
3082
3083 #[test]
3084 fn dash_j_reaches_the_scheduler_and_defaults_to_the_machine() {
3085 let (_, _, jobs) = match parse_args(&args(&["-j4", "a.c"])).unwrap() {
3086 Action::Compile { opts, plan, jobs, .. } => (opts, plan, jobs),
3087 other => panic!("expected a compilation, got {other:?}"),
3088 };
3089 assert_eq!(jobs.count(), 4);
3090
3091 let default = match parse_args(&args(&["a.c"])).unwrap() {
3092 Action::Compile { jobs, .. } => jobs,
3093 other => panic!("expected a compilation, got {other:?}"),
3094 };
3095 assert_eq!(default, Jobs::available());
3096 assert!(parse_args(&args(&["-j0", "a.c"])).is_err());
3097 }
3098
3099 #[test]
3100 fn triple_hash_prints_the_plan_and_runs_nothing() {
3101 let a = parse_args(&args(&["-###", "-c", "a.c"])).unwrap();
3102 let Action::PrintPlan { plan, .. } = a else { panic!("expected a plan dump") };
3103 assert!(plan.render().contains("a.c: preprocess, compile, assemble -> a.o"));
3104 }
3105
3106 #[test]
3107 fn the_flag_that_keeps_the_intermediate_files_has_three_spellings_and_two_meanings() {
3108 // The bare one is `=obj` and not `=cwd`. gcc's manual says the opposite and gcc 16 does
3109 // this, and following the compiler is what makes a build that reads either of them find
3110 // the files where they are.
3111 assert_eq!(compile(&["-c", "-save-temps", "a.c"]).0.save_temps, SaveTemps::Object);
3112 assert_eq!(compile(&["-c", "-save-temps=obj", "a.c"]).0.save_temps, SaveTemps::Object);
3113 assert_eq!(compile(&["-c", "-save-temps=cwd", "a.c"]).0.save_temps, SaveTemps::Cwd);
3114 assert_eq!(compile(&["-c", "a.c"]).0.save_temps, SaveTemps::No);
3115 // The last one on the line decides, the way it does for every other flag with an
3116 // argument, and a keyword that is neither is fatal rather than ignored: a run that kept
3117 // nothing and said nothing looks exactly like one where the files were not produced.
3118 let (opts, _) = compile(&["-c", "-save-temps", "-save-temps=cwd", "a.c"]);
3119 assert_eq!(opts.save_temps, SaveTemps::Cwd);
3120 let e = parse_args(&args(&["-c", "-save-temps=nowhere", "a.c"])).unwrap_err();
3121 assert!(e.message.contains("accepted: cwd, obj"), "{}", e.message);
3122 }
3123
3124 #[test]
3125 fn the_flag_that_times_each_step_reaches_the_options_and_changes_nothing_else() {
3126 let (opts, plan) = compile(&["-c", "-time", "a.c"]);
3127 let (plain, without) = compile(&["-c", "a.c"]);
3128 assert!(opts.time);
3129 assert!(!plain.time);
3130 // Against the same line without the flag rather than against a spelling of the object's
3131 // name, since what the object is called is the host's business and this is not about that.
3132 assert_eq!(plan.jobs[0].output, without.jobs[0].output);
3133 }
3134
3135 #[test]
3136 fn dash_x_names_what_it_accepts_when_it_does_not_know_a_language() {
3137 let e = parse_args(&args(&["-x", "fortran", "a.c"])).unwrap_err();
3138 assert!(e.message.contains("assembler-with-cpp"), "{}", e.message);
3139 }
3140
3141 /// What `--fetch` says while the table in [`artifact`] has no rows in it, which is what every
3142 /// run of it says today and is the reason the message distinguishes the two cases.
3143 #[test]
3144 fn a_fetch_of_a_target_nothing_is_pinned_for_says_so_rather_than_reaching_the_network() {
3145 let e = parse_args(&args(&["--fetch", "x86_64-linux-musl"])).unwrap_err();
3146 assert!(e.message.contains("pins no sysroot for x86_64-linux-musl"), "{}", e.message);
3147 // And where one comes from, because the answer is not on this machine.
3148 assert!(e.message.contains("tamnd/rucc-cross"), "{}", e.message);
3149 // The joined spelling is the same flag.
3150 let joined = parse_args(&args(&["--fetch=x86_64-linux-musl"])).unwrap_err();
3151 assert_eq!(joined, e);
3152 }
3153
3154 /// The two targets a release will never pin, which is a different answer from the one above.
3155 ///
3156 /// Section 13.4. A person who reads "this release pins no sysroot yet" waits for a release that
3157 /// does, and no release of this compiler can ship either of these, so the message names the
3158 /// licence that decides it and what to do instead.
3159 #[test]
3160 fn a_fetch_of_a_target_behind_a_licence_wall_says_so_rather_than_saying_not_yet() {
3161 let e = parse_args(&args(&["--fetch", "aarch64-macos"])).unwrap_err();
3162 assert!(e.message.contains("Xcode licence"), "{}", e.message);
3163 assert!(e.message.contains("there never will be"), "{}", e.message);
3164 assert!(!e.message.contains("tamnd/rucc-cross"), "{}", e.message);
3165
3166 let e = parse_args(&args(&["--fetch", "x86_64-windows-msvc"])).unwrap_err();
3167 assert!(e.message.contains("redistributed"), "{}", e.message);
3168 // The way out of this one is a target rather than a download, and it is the default already.
3169 assert!(e.message.contains("mingw-w64"), "{}", e.message);
3170 // And the mingw-w64 target next to it is an ordinary unpinned target.
3171 let e = parse_args(&args(&["--fetch", "x86_64-windows-gnu"])).unwrap_err();
3172 assert!(e.message.contains("pins no sysroot"), "{}", e.message);
3173 }
3174
3175 /// An Apple target on a machine with no SDK, which is section 8.6's other host.
3176 ///
3177 /// Not run on a mac, where the SDK this is about is installed and the compile is the ordinary one
3178 /// that uses it. What the reason says is asserted in `rucc_sysroot::wall` and where it is printed
3179 /// is asserted in `rucc-pp`, so what is left here is that the driver works it out and leaves it
3180 /// where the preprocessor will find it, and that neither way past the wall leaves one behind.
3181 #[test]
3182 fn an_apple_target_with_no_sdk_anywhere_carries_the_licence_rather_than_a_missing_directory() {
3183 if cfg!(target_os = "macos") || std::env::var_os("SDKROOT").is_some() {
3184 return;
3185 }
3186 let (opts, _) = compile(&["--target=aarch64-macos", "-c", "a.c"]);
3187 let why = opts.search.missing_system().expect("the wall is the reason there are none");
3188 assert!(why.contains("aarch64-macos needs a macOS SDK"), "{why}");
3189 assert!(why.contains("Xcode licence"), "{why}");
3190 assert!(why.contains("-isysroot"), "{why}");
3191
3192 // A program that includes none of the library needs none of the SDK, which is what section
3193 // 8.6 means by being able to target the platform without one, so there is nothing to explain.
3194 let (opts, _) = compile(&["--target=aarch64-macos", "-nostdinc", "-c", "a.c"]);
3195 assert_eq!(opts.search.missing_system(), None);
3196 // And naming a path is the other way through, whether or not the path is there: a mistyped
3197 // directory is a mistake to report on its own terms rather than a licence to explain.
3198 let (opts, _) = compile(&["--target=aarch64-macos", "-isysroot", "/opt/sdk", "-c", "a.c"]);
3199 assert_eq!(opts.search.missing_system(), None);
3200 }
3201
3202 /// The same wall on the compile side of an MSVC target, where the way past it is a tuple.
3203 #[test]
3204 fn an_msvc_target_with_no_sdk_named_says_which_environment_needs_nothing_installed() {
3205 if std::env::var_os("INCLUDE").is_some() {
3206 return;
3207 }
3208 let (opts, _) = compile(&["--target=x86_64-windows-msvc", "-c", "a.c"]);
3209 let why = opts.search.missing_system().expect("the wall is the reason there are none");
3210 assert!(why.contains("the Windows SDK and its universal CRT"), "{why}");
3211 assert!(why.contains("mingw-w64"), "{why}");
3212 // And the mingw-w64 target has its headers from us, so nothing is missing to explain.
3213 let (opts, _) = compile(&["--target=x86_64-windows-gnu", "-c", "a.c"]);
3214 assert_eq!(opts.search.missing_system(), None);
3215 }
3216
3217 #[test]
3218 fn a_fetch_with_no_target_and_a_fetch_of_a_tuple_that_is_not_one_both_say_which() {
3219 let e = parse_args(&args(&["--fetch"])).unwrap_err();
3220 assert!(e.message.contains("--fetch requires"), "{}", e.message);
3221 let e = parse_args(&args(&["--fetch", "sparc64-solaris-gnu"])).unwrap_err();
3222 assert!(e.message.contains("--fetch sparc64-solaris-gnu"), "{}", e.message);
3223 assert!(e.message.contains("no sysroot to get"), "{}", e.message);
3224 }
3225
3226 /// Both flags on one line ask for opposite things, in either order.
3227 #[test]
3228 fn a_fetch_and_offline_together_is_a_refusal_whichever_way_round_they_are_written() {
3229 for line in [
3230 vec!["--offline", "--fetch", "x86_64-linux-musl"],
3231 vec!["--fetch", "x86_64-linux-musl", "--offline"],
3232 ] {
3233 let e = parse_args(&args(&line)).unwrap_err();
3234 assert!(e.message.contains("two opposite things"), "{}", e.message);
3235 }
3236 }
3237
3238 #[test]
3239 fn a_fetch_does_not_compile_anything_and_says_so_when_it_is_handed_a_file() {
3240 let e = parse_args(&args(&["--fetch", "x86_64-linux-musl", "a.c"])).unwrap_err();
3241 assert!(e.message.contains("compiles nothing"), "{}", e.message);
3242 assert!(e.message.contains("a.c"), "{}", e.message);
3243 }
3244
3245 /// `--offline` on its own is accepted and changes nothing, because an ordinary compile
3246 /// downloads nothing with or without it. A build that passes it everywhere is the case this is
3247 /// for, and it must not lose the compilation it was passed beside.
3248 #[test]
3249 fn offline_on_a_compilation_is_the_same_compilation() {
3250 let (opts, plan) = compile(&["-c", "--offline", "a.c"]);
3251 let (plain, without) = compile(&["-c", "a.c"]);
3252 assert_eq!(opts.target, plain.target);
3253 assert_eq!(plan.jobs.len(), without.jobs.len());
3254 assert_eq!(plan.jobs[0].output, without.jobs[0].output);
3255 }
3256
3257 #[test]
3258 fn an_unknown_flag_is_an_error_rather_than_a_shrug() {
3259 let e = parse_args(&args(&["-fno-such-thing", "a.c"])).unwrap_err();
3260 assert!(e.message.contains("unknown option"), "{}", e.message);
3261 }
3262
3263 /// `-fpermissive` and the flag that turns it back off, which a build writes beside it when
3264 /// one directory needs the older rules and the rest of the tree does not.
3265 #[test]
3266 fn permissive_reads_in_both_directions_and_the_last_one_wins() {
3267 let (opts, _) = compile(&["-c", "a.c"]);
3268 assert!(!opts.permissive, "off unless it is asked for");
3269
3270 let (opts, _) = compile(&["-c", "-fpermissive", "a.c"]);
3271 assert!(opts.permissive);
3272
3273 let (opts, _) = compile(&["-c", "-fpermissive", "-fno-permissive", "a.c"]);
3274 assert!(!opts.permissive);
3275 }
3276
3277 #[test]
3278 fn asking_for_nested_functions_is_told_why_it_is_not_coming() {
3279 let e = parse_args(&args(&["-fnested-functions", "a.c"])).unwrap_err();
3280 assert!(e.message.contains("trampoline"), "{}", e.message);
3281 assert!(parse_args(&args(&["-fno-nested-functions", "a.c"])).is_ok());
3282 }
3283
3284 #[test]
3285 fn the_flag_every_configure_script_writes_is_taken() {
3286 // All four spellings, because a build writes whichever one its macros picked and a
3287 // compiler that takes three of them is a compiler that fails on the fourth.
3288 for flag in ["-fPIC", "-fpic", "-fPIE", "-fpie"] {
3289 let (opts, _) = compile(&["-c", flag, "a.c"]);
3290 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3291 }
3292 }
3293
3294 #[test]
3295 fn a_table_is_written_unless_the_build_says_nothing_will_walk_it() {
3296 let (opts, _) = compile(&["-c", "a.c"]);
3297 assert!(opts.unwinds(), "the default is off");
3298 let (opts, _) = compile(&["-c", "-fno-asynchronous-unwind-tables", "a.c"]);
3299 assert!(!opts.unwinds(), "the build was not taken at its word");
3300 let (opts, _) = compile(&[
3301 "-c",
3302 "-fno-asynchronous-unwind-tables",
3303 "-fasynchronous-unwind-tables",
3304 "a.c",
3305 ]);
3306 assert!(opts.unwinds(), "the last flag did not win");
3307 // The weaker request, which the same table answers, so a line that asks for a table and
3308 // against an asynchronous one gets one. That is gcc's arrangement and it turns up when a
3309 // build turns the asynchronous one off globally and a directory asks for a table back.
3310 let (opts, _) =
3311 compile(&["-c", "-fno-asynchronous-unwind-tables", "-funwind-tables", "a.c"]);
3312 assert!(opts.unwinds(), "the weaker request was dropped");
3313 let (opts, _) = compile(&["-c", "-fno-unwind-tables", "a.c"]);
3314 assert!(opts.unwinds(), "the weaker negative turned off the stronger request");
3315 let (opts, _) =
3316 compile(&["-c", "-fno-unwind-tables", "-fno-asynchronous-unwind-tables", "a.c"]);
3317 assert!(!opts.unwinds(), "both were turned off and one stayed on");
3318 }
3319
3320 #[test]
3321 fn the_flags_that_describe_what_this_compiler_already_does_are_taken() {
3322 // Every one of these is on a real build line somewhere and every one of them was an
3323 // unknown option. What they have in common is that the answer rucc gives is the answer
3324 // they ask for, so there is nothing to implement and nothing to refuse.
3325 for flag in [
3326 "-fno-common",
3327 "-fstrict-aliasing",
3328 "-fno-strict-aliasing",
3329 "-fdelete-null-pointer-checks",
3330 "-fno-delete-null-pointer-checks",
3331 "-frounding-math",
3332 "-fno-rounding-math",
3333 "-fexcess-precision=standard",
3334 "-fexcess-precision=fast",
3335 "-fexcess-precision=16",
3336 "-pipe",
3337 "-fdiagnostics-color",
3338 "-fno-diagnostics-color",
3339 "-fdiagnostics-color=always",
3340 "-fdiagnostics-color=never",
3341 "-fdiagnostics-color=auto",
3342 ] {
3343 let (opts, _) = compile(&["-c", flag, "a.c"]);
3344 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3345 }
3346 }
3347
3348 #[test]
3349 fn whether_an_exception_is_looked_at_is_kept_and_defaults_to_gccs_answer() {
3350 let (opts, _) = compile(&["-c", "a.c"]);
3351 assert!(opts.trapping_math, "the default was not gcc's");
3352 let (opts, _) = compile(&["-c", "-fno-trapping-math", "a.c"]);
3353 assert!(!opts.trapping_math);
3354 let (opts, _) = compile(&["-c", "-ftrapping-math", "a.c"]);
3355 assert!(opts.trapping_math, "spelling out the default turned it off");
3356 // The last one written wins, which is how a build line that inherits a flag from one
3357 // place and overrides it in another is read.
3358 let (opts, _) = compile(&["-c", "-fno-trapping-math", "-ftrapping-math", "a.c"]);
3359 assert!(opts.trapping_math);
3360 }
3361
3362 /// The flags a torture program writes on its own `dg-options` line, which is where most of
3363 /// these come from: a program reduced from a miscompilation names the pass that miscompiled
3364 /// it. Eighteen programs in the suite stopped on the driver before anything read them, and
3365 /// tamnd/rucc#1019 is the list.
3366 #[test]
3367 fn the_flags_that_name_a_pass_of_gccs_own_are_taken_and_dropped() {
3368 for flag in [
3369 "-fno-tree-ccp",
3370 "-fno-tree-dominator-opts",
3371 "-fno-tree-vrp",
3372 "-fno-tree-bit-ccp",
3373 "-fno-tree-coalesce-vars",
3374 "-ftree-vectorize",
3375 "-ftree-loop-distribution",
3376 "-fno-ipa-cp",
3377 "-fipa-pta",
3378 "-fmodulo-sched",
3379 "-fno-vect-cost-model",
3380 "-fvect-cost-model=unlimited",
3381 "-fsimd-cost-model=cheap",
3382 "-fexpensive-optimizations",
3383 "-fno-early-inlining",
3384 "-fno-inline",
3385 "-finline-functions",
3386 "-foptimize-strlen",
3387 "-fno-ira-share-spill-slots",
3388 ] {
3389 let (opts, _) = compile(&["-c", flag, "a.c"]);
3390 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3391 assert!(opts.passes.is_empty(), "{flag} named a pass of gcc's and not one of ours");
3392 }
3393 }
3394
3395 /// The two namespaces are taken whole, so a name neither this test nor gcc 16 has heard of
3396 /// goes the same way as the ones above rather than stopping a build on the day gcc adds it.
3397 #[test]
3398 fn a_pass_name_in_either_family_is_taken_whether_or_not_it_is_one_gcc_has() {
3399 for flag in ["-ftree-no-such-pass", "-fno-ipa-no-such-pass"] {
3400 let (opts, _) = compile(&["-c", flag, "a.c"]);
3401 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3402 }
3403 }
3404
3405 /// A pass this compiler has keeps its flag, since the arms that read the registry are above
3406 /// the family arms. `dce` is the one both compilers have a name for, and `execute/pr97421-2.c`
3407 /// is the program that writes it.
3408 #[test]
3409 fn a_pass_name_this_compiler_has_is_still_read_as_a_pass() {
3410 let (opts, _) = compile(&["-c", "-fno-dce", "a.c"]);
3411 assert_eq!(opts.passes, vec![("dce".to_owned(), false)]);
3412 }
3413
3414 /// gcc's name for the unroller reaches the unroller, in both directions. libtommath puts
3415 /// `-funroll-loops` in `CFLAGS` unconditionally, and before this it was an unknown option and
3416 /// the build stopped on its first file.
3417 #[test]
3418 fn the_gcc_spelling_of_the_unroller_turns_the_unroller_on_and_off() {
3419 let (opts, _) = compile(&["-c", "-funroll-loops", "a.c"]);
3420 assert_eq!(opts.passes, vec![("unroll".to_owned(), true)]);
3421 let (opts, _) = compile(&["-c", "-fno-unroll-loops", "a.c"]);
3422 assert_eq!(opts.passes, vec![("unroll".to_owned(), false)]);
3423 }
3424
3425 /// The encoding of the source is not a question about speed, so the one name that describes
3426 /// what the preprocessor does is taken and every other name is refused.
3427 #[test]
3428 fn the_input_charset_is_taken_when_it_names_the_one_that_is_read() {
3429 for flag in ["-finput-charset=utf-8", "-finput-charset=UTF-8", "-finput-charset=utf8"] {
3430 let (opts, _) = compile(&["-c", flag, "a.c"]);
3431 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3432 }
3433
3434 let e = parse_args(&args(&["-c", "-finput-charset=latin1", "a.c"])).unwrap_err();
3435 assert!(e.message.contains("latin1"), "{}", e.message);
3436 assert!(e.message.contains("UTF-8"), "what is read is worth saying: {}", e.message);
3437 }
3438
3439 /// The other half of the same rule. Each of these changes what the program does rather than
3440 /// how fast it does it, so each is refused with the reason, and the negative of each is what
3441 /// happens anyway and is taken.
3442 #[test]
3443 fn the_three_that_change_the_answer_are_refused_and_their_negatives_are_taken() {
3444 for (flag, word) in [
3445 ("-ffast-math", "__FAST_MATH__"),
3446 ("-fnon-call-exceptions", "landing pad"),
3447 ("-finstrument-functions", "__cyg_profile_func_enter"),
3448 ] {
3449 let e = parse_args(&args(&["-c", flag, "a.c"])).unwrap_err();
3450 assert!(e.message.contains(word), "{flag}: {}", e.message);
3451 assert!(!e.message.contains("unknown option"), "{flag} deserves a reason");
3452
3453 let off = format!("-fno-{}", flag.trim_start_matches("-f"));
3454 let (opts, _) = compile(&["-c", &off, "a.c"]);
3455 assert_eq!(opts.emit, EmitKind::Object, "{off}");
3456 }
3457 }
3458
3459 #[test]
3460 fn asking_the_linker_to_merge_tentative_definitions_is_told_why_it_is_not_coming() {
3461 // The one of that family that is a request rather than a description, and it is a real
3462 // difference: two files each writing `int g;` link under it and do not without it.
3463 let e = parse_args(&args(&["-fcommon", "a.c"])).unwrap_err();
3464 assert!(e.message.contains(".bss"), "{}", e.message);
3465 assert!(e.message.contains("extern"), "the way out is worth saying: {}", e.message);
3466 }
3467
3468 #[test]
3469 fn asking_for_position_dependent_code_is_told_why_it_is_not_coming() {
3470 for flag in ["-fno-pic", "-fno-pie"] {
3471 let e = parse_args(&args(&[flag, "a.c"])).unwrap_err();
3472 assert!(e.message.contains("global offset table"), "{flag}: {}", e.message);
3473 // The one it may have meant, since the two are a letter apart and one of them is
3474 // about linking and is taken.
3475 assert!(e.message.contains("-no-pie"), "{flag}: {}", e.message);
3476 }
3477 }
3478
3479 #[test]
3480 fn an_unsupported_target_names_itself() {
3481 let e = parse_args(&args(&["--target=sparc64-linux-gnu", "a.c"])).unwrap_err();
3482 assert!(e.message.contains("sparc64"), "{}", e.message);
3483 }
3484
3485 #[test]
3486 fn no_inputs_is_an_error_but_print_config_needs_none() {
3487 assert!(parse_args(&args(&[])).is_err());
3488 assert!(matches!(parse_args(&args(&["--print-config"])), Ok(Action::PrintConfig(_))));
3489 }
3490
3491 #[test]
3492 fn print_config_reports_the_target_it_was_given_not_the_host() {
3493 let a = parse_args(&args(&["--print-config", "--target=riscv64-linux-musl"])).unwrap();
3494 let Action::PrintConfig(opts) = a else { panic!("expected a configuration dump") };
3495 let text = print_config(&opts);
3496 assert!(text.contains("target: riscv64-unknown-linux-musl"), "{text}");
3497 assert!(text.contains("char-signed: false"), "{text}");
3498 assert!(text.contains("object-format: elf"), "{text}");
3499 assert!(text.contains("va-list: void-pointer"), "{text}");
3500 // RISC-V has a register file and this compiler has not written it down yet, and the
3501 // dump says which of those two it is rather than leaving the line out.
3502 assert!(text.contains("registers: none"), "{text}");
3503 assert!(text.contains("timing-model: none"), "{text}");
3504 }
3505
3506 /// The model the schedule was chosen with, which is a receipt anybody comparing two runs of a
3507 /// benchmark needs: two numbers that disagree are usually two models and not two compilers.
3508 #[test]
3509 fn print_config_names_the_model_the_schedule_was_chosen_with() {
3510 let opts = Options::new("x86_64-unknown-linux-gnu".parse().unwrap());
3511 let text = print_config(&opts);
3512 let line = text.lines().find(|l| l.starts_with("timing-model:")).expect("the model");
3513 assert!(line.contains("Skylake"), "{line}");
3514 assert!(line.contains("published"), "a sentence saying where it came from: {line}");
3515 }
3516
3517 #[test]
3518 fn print_config_has_one_key_per_line_and_a_fixed_order() {
3519 let opts = Options::new("x86_64-unknown-linux-gnu".parse().unwrap());
3520 let text = print_config(&opts);
3521 let keys: Vec<&str> =
3522 text.lines().map(|l| l.split(':').next().unwrap_or_default()).collect();
3523 assert_eq!(keys[0], "version");
3524 assert_eq!(keys[1], "target");
3525 assert_eq!(keys.len(), 26);
3526 assert!(text.ends_with('\n'));
3527 }
3528
3529 #[test]
3530 fn the_safety_tier_is_read_off_the_command_line_and_a_wrong_one_is_refused() {
3531 let (opts, _) = compile(&["a.c"]);
3532 assert_eq!(opts.safety, rucc_session::Safety::Off);
3533
3534 for (flag, tier) in [
3535 ("-fsafety=detect", rucc_session::Safety::Detect),
3536 ("-fsafety=enforce", rucc_session::Safety::Enforce),
3537 ("-fsafety=kernel", rucc_session::Safety::Kernel),
3538 ("-fsafety=off", rucc_session::Safety::Off),
3539 ] {
3540 let (opts, _) = compile(&[flag, "a.c"]);
3541 assert_eq!(opts.safety, tier, "{flag}");
3542 }
3543
3544 // The last one wins, the way every other repeated flag on this command line does.
3545 let (opts, _) = compile(&["-fsafety=enforce", "-fsafety=off", "a.c"]);
3546 assert_eq!(opts.safety, rucc_session::Safety::Off);
3547
3548 // A misspelled tier is refused rather than ignored. Silently compiling without the
3549 // monitor a build asked for is the one failure mode this feature cannot have.
3550 let e = parse_args(&args(&["-fsafety=on", "a.c"])).unwrap_err();
3551 assert!(e.message.contains("is not a safety tier"), "{}", e.message);
3552 assert!(parse_args(&args(&["-fsafety", "a.c"])).is_err());
3553 }
3554
3555 #[test]
3556 fn the_padding_mode_is_read_off_the_command_line_and_a_wrong_one_is_refused() {
3557 // The default is the one section 9.3 of document 09 gives library code, which is that
3558 // padding does not participate, so a record filled a member at a time is not reported.
3559 let (opts, _) = compile(&["a.c"]);
3560 assert_eq!(opts.padding, rucc_session::Padding::Ignored);
3561
3562 let (opts, _) = compile(&["-fsafety=detect", "-fsafety-init=padding", "a.c"]);
3563 assert_eq!(opts.padding, rucc_session::Padding::Tracked);
3564
3565 let (opts, _) = compile(&["-fsafety-init=padding", "-fsafety-init=nopadding", "a.c"]);
3566 assert_eq!(opts.padding, rucc_session::Padding::Ignored);
3567
3568 // The tier is still a tier. A flag whose name starts the same way must not be eaten by
3569 // the one above it, which is the thing worth pinning about a pair of names like these.
3570 let (opts, _) = compile(&["-fsafety-init=padding", "a.c"]);
3571 assert_eq!(opts.safety, rucc_session::Safety::Off);
3572
3573 let e = parse_args(&args(&["-fsafety-init=some", "a.c"])).unwrap_err();
3574 assert!(e.message.contains("is not a padding mode"), "{}", e.message);
3575 }
3576
3577 #[test]
3578 fn whether_a_write_has_to_stay_inside_its_member_is_read_off_the_command_line() {
3579 // Off by default, because a store to allocated storage sets its effective type and C 6.5
3580 // lets a program reuse a buffer as something else. Row S4 is a build opting out of that.
3581 let (opts, _) = compile(&["a.c"]);
3582 assert_eq!(opts.subobject, rucc_session::Subobject::Off);
3583
3584 let (opts, _) = compile(&["-fsafety=detect", "-fsafety-subobject", "a.c"]);
3585 assert_eq!(opts.subobject, rucc_session::Subobject::Members);
3586
3587 let (opts, _) = compile(&["-fsafety-subobject", "-fno-safety-subobject", "a.c"]);
3588 assert_eq!(opts.subobject, rucc_session::Subobject::Off);
3589
3590 // It takes no value. The form that would take one is the strict reading of section 9.4,
3591 // which is not written yet, so say so rather than accept a spelling that does nothing.
3592 let e = parse_args(&args(&["-fsafety-subobject=strict", "a.c"])).unwrap_err();
3593 assert!(e.message.contains("tamnd/rucc#967"), "{}", e.message);
3594 }
3595
3596 #[test]
3597 fn whether_two_restrict_pointers_may_meet_is_read_off_the_command_line() {
3598 // Off by default, because the record a block keeps is the union of what each pointer
3599 // reached, so two pointers striding through one array without landing on the same byte are
3600 // reported and by the letter of the standard those are different objects. Row Y8 is a build
3601 // deciding it would rather know.
3602 let (opts, _) = compile(&["a.c"]);
3603 assert_eq!(opts.promise, rucc_session::Promise::Off);
3604
3605 let (opts, _) = compile(&["-fsafety=detect", "-fsafety-restrict", "a.c"]);
3606 assert_eq!(opts.promise, rucc_session::Promise::Blocks);
3607
3608 let (opts, _) = compile(&["-fsafety-restrict", "-fno-safety-restrict", "a.c"]);
3609 assert_eq!(opts.promise, rucc_session::Promise::Off);
3610
3611 // The tier is still a tier, which is the thing worth pinning about a pair of names where
3612 // one is the front of the other.
3613 let (opts, _) = compile(&["-fsafety-restrict", "a.c"]);
3614 assert_eq!(opts.safety, rucc_session::Safety::Off);
3615
3616 let e = parse_args(&args(&["-fsafety-restrict=blocks", "a.c"])).unwrap_err();
3617 assert!(e.message.contains("takes no value"), "{}", e.message);
3618 }
3619
3620 #[test]
3621 fn safety_races_takes_a_mode_and_defaults_to_watching_nothing() {
3622 // Three modes rather than a bare flag, because section 9.5 gives two answers that record
3623 // the same thing and report different classes, so a flag with no value could not say which
3624 // was wanted. Off by default for the reason on `rucc_session::Races`, which is not a cost
3625 // argument: this is the one plane where an edge nobody interposed costs a false report.
3626 let (opts, _) = compile(&["a.c"]);
3627 assert_eq!(opts.races, rucc_session::Races::Off);
3628
3629 let (opts, _) = compile(&["-fsafety-races=metadata", "a.c"]);
3630 assert_eq!(opts.races, rucc_session::Races::Metadata);
3631
3632 let (opts, _) = compile(&["-fsafety-races=pointer", "a.c"]);
3633 assert_eq!(opts.races, rucc_session::Races::Pointer);
3634
3635 // Last one wins, as it does for every other mode flag here.
3636 let (opts, _) = compile(&["-fsafety-races=pointer", "-fno-safety-races", "a.c"]);
3637 assert_eq!(opts.races, rucc_session::Races::Off);
3638
3639 let e = parse_args(&args(&["-fsafety-races=all", "a.c"])).unwrap_err();
3640 assert!(e.message.contains("off, metadata or pointer"), "{}", e.message);
3641 }
3642
3643 #[test]
3644 fn print_pipeline_answers_with_the_passes_the_level_asked_for() {
3645 let a = parse_args(&args(&["--print-pipeline", "-O2"])).unwrap();
3646 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3647 let text = print_pipeline(&opts);
3648 assert!(text.starts_with("level: -O2\n"), "{text}");
3649 assert!(text.contains("fold"), "{text}");
3650
3651 let a = parse_args(&args(&["--print-pipeline"])).unwrap();
3652 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3653 // Two passes run at `-O0` and neither is an optimization. The first moves what
3654 // `__builtin_expect` said onto the branch and takes the instruction away, so that nothing
3655 // past the optimizer has to know the instruction exists. The second removes code nothing
3656 // reaches. See issue 359.
3657 assert!(print_pipeline(&opts).contains("1: expect,"), "{}", print_pipeline(&opts));
3658 assert!(print_pipeline(&opts).contains("2: simplify-cfg,"), "{}", print_pipeline(&opts));
3659
3660 let a = parse_args(&args(&["--print-pipeline", "-fno-simplify-cfg"])).unwrap();
3661 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3662 // The second turns off and the first does not, because nothing below the optimizer lowers
3663 // what it removes, so `-fno-expect` is a compile that stops rather than one that runs.
3664 let text = print_pipeline(&opts);
3665 assert!(text.contains("1: expect,"), "{text}");
3666 assert!(!text.contains("simplify-cfg"), "{text}");
3667 }
3668
3669 #[test]
3670 fn print_pipeline_takes_the_toggles_into_account() {
3671 let a = parse_args(&args(&["--print-pipeline", "-O2", "-fno-fold"])).unwrap();
3672 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3673 let text = print_pipeline(&opts);
3674 // The one that was named is gone and the rest of the level is not, which is the whole
3675 // of what a toggle promises.
3676 assert!(!text.contains("fold"), "{text}");
3677 assert!(text.contains("dce"), "{text}");
3678
3679 // Every pass the compiler has, named off. Built from the registry rather than written
3680 // out, so a pass added later is turned off here too and this keeps testing the thing it
3681 // is about, which is that the toggles can empty a level down to the passes that are not
3682 // optional. Those are named, because a listing that is all of them is a level nobody
3683 // emptied and the assertion would pass while saying nothing.
3684 let mut off = vec!["--print-pipeline".to_owned(), "-O2".to_owned()];
3685 off.extend(rucc_opt::PASSES.iter().map(|p| format!("-fno-{}", p.name())));
3686 let spelled: Vec<&str> = off.iter().map(String::as_str).collect();
3687 let a = parse_args(&args(&spelled)).unwrap();
3688 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3689 let text = print_pipeline(&opts);
3690 let left: Vec<&str> =
3691 rucc_opt::PASSES.iter().filter(|p| p.required()).map(|p| p.name()).collect();
3692 assert_eq!(left, vec!["expect"], "{text}");
3693 for (at, name) in left.iter().enumerate() {
3694 assert!(text.contains(&format!("{}: {name},", at + 1)), "{text}");
3695 }
3696 assert!(!text.contains("dce"), "{text}");
3697 }
3698
3699 #[test]
3700 fn print_pipeline_says_when_a_budget_will_stop_the_run_short() {
3701 let a = parse_args(&args(&["--print-pipeline", "-O2"])).unwrap();
3702 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3703 assert!(!print_pipeline(&opts).contains("global fuel"));
3704
3705 let a = parse_args(&args(&["--print-pipeline", "-O2", "-fpass-fuel-global=4"])).unwrap();
3706 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3707 let text = print_pipeline(&opts);
3708 // Because the listing is the answer to what this compilation will do, and a run that
3709 // stops after four rewrites is not doing what the level says it does.
3710 assert!(text.contains("global fuel: 4"), "{text}");
3711 }
3712
3713 /// A pass is turned on and off by its own name, and the order the flags were given in is
3714 /// kept, because the last spelling of a name is the one that decides.
3715 #[test]
3716 fn a_pass_is_named_by_dash_f_and_unnamed_by_dash_f_no() {
3717 let (opts, _) = compile(&["-c", "-O0", "-ffold", "-fno-fold", "-ffold", "a.c"]);
3718 assert_eq!(
3719 opts.passes,
3720 [("fold".to_owned(), true), ("fold".to_owned(), false), ("fold".to_owned(), true)]
3721 );
3722
3723 let e = parse_args(&args(&["-fno-such-pass", "a.c"])).unwrap_err();
3724 assert!(e.message.contains("unknown option"), "{}", e.message);
3725 }
3726
3727 #[test]
3728 fn pass_fuel_names_a_pass_and_a_count_and_refuses_anything_else() {
3729 let (opts, _) = compile(&["-c", "-O2", "-fpass-fuel=fold=3", "a.c"]);
3730 assert_eq!(opts.pass_fuel, [("fold".to_owned(), 3)]);
3731
3732 let e = parse_args(&args(&["-fpass-fuel=fold", "a.c"])).unwrap_err();
3733 assert!(e.message.contains("<pass>=<count>"), "{}", e.message);
3734 let e = parse_args(&args(&["-fpass-fuel=nosuch=3", "a.c"])).unwrap_err();
3735 assert!(e.message.contains("--print-pipeline"), "{}", e.message);
3736 let e = parse_args(&args(&["-fpass-fuel=fold=lots", "a.c"])).unwrap_err();
3737 assert!(e.message.contains("not a number"), "{}", e.message);
3738 }
3739
3740 #[test]
3741 fn global_pass_fuel_is_a_count_on_its_own_and_defaults_to_no_limit() {
3742 let (opts, _) = compile(&["-c", "-O2", "a.c"]);
3743 assert_eq!(opts.pass_fuel_global, None);
3744
3745 let (opts, _) = compile(&["-c", "-O2", "-fpass-fuel-global=12", "a.c"]);
3746 assert_eq!(opts.pass_fuel_global, Some(12));
3747 // And it is not the per pass flag with a longer name, so neither spelling swallows the
3748 // other.
3749 assert!(opts.pass_fuel.is_empty());
3750
3751 let e = parse_args(&args(&["-fpass-fuel-global=lots", "a.c"])).unwrap_err();
3752 assert!(e.message.contains("not a number"), "{}", e.message);
3753 }
3754
3755 #[test]
3756 fn a_gate_names_a_pass_and_optionally_the_functions_it_covers() {
3757 let (opts, _) = compile(&["-c", "-O2", "-fdisable-fold", "-fenable-fold=2-4,main", "a.c"]);
3758 assert_eq!(
3759 opts.pass_gates,
3760 [(false, "fold".to_owned()), (true, "fold=2-4,main".to_owned())],
3761 "the order is what decides, so it has to survive the parse"
3762 );
3763
3764 let e = parse_args(&args(&["-fdisable-nosuch", "a.c"])).unwrap_err();
3765 assert!(e.message.contains("--print-pipeline"), "{}", e.message);
3766 let e = parse_args(&args(&["-fenable-fold=9-2", "a.c"])).unwrap_err();
3767 assert!(e.message.contains("ends before it starts"), "{}", e.message);
3768 let e = parse_args(&args(&["-fdisable-fold=", "a.c"])).unwrap_err();
3769 assert!(e.message.contains("is empty"), "{}", e.message);
3770 }
3771
3772 #[test]
3773 fn the_pipeline_listing_says_which_passes_a_gate_touched() {
3774 let (opts, _) = compile(&["-c", "-O2", "-fdisable-fold=main", "a.c"]);
3775 let text = print_pipeline(&opts);
3776 assert!(text.contains("fold, "), "{text}");
3777 assert!(text.contains("[off for main]"), "{text}");
3778 }
3779
3780 /// The spelling is checked while the arguments are read, because a dump that names a pass
3781 /// this compiler does not have is a typo, and a typo found after the compilation has run is
3782 /// found too late to be any use.
3783 #[test]
3784 fn a_dump_is_checked_when_it_is_asked_for_rather_than_when_it_is_taken() {
3785 let (opts, _) = compile(&["-c", "-O2", "-fdump-ir=all", "-fdump-ir=after-fold", "a.c"]);
3786 assert_eq!(opts.dump_ir, ["all", "after-fold"]);
3787
3788 let e = parse_args(&args(&["-fdump-ir=after-nosuch", "a.c"])).unwrap_err();
3789 assert!(e.message.contains("nosuch"), "{}", e.message);
3790 assert!(parse_args(&args(&["-fdump-ir=sideways-fold", "a.c"])).is_err());
3791 }
3792
3793 /// Every spelling `-fopt-info` takes, and the one it does not.
3794 ///
3795 /// The keywords are checked here for the same reason a dump's pass name is: a person who
3796 /// misspelled one gets no output, and no output is also what a compilation where nothing
3797 /// happened looks like. Telling those two apart is the entire reason to reach for this flag.
3798 #[test]
3799 fn opt_info_takes_kinds_and_a_file_and_refuses_a_kind_it_does_not_have() {
3800 let (opts, _) = compile(&["-c", "-O2", "-fopt-info", "a.c"]);
3801 assert_eq!(opts.opt_info, [""], "a bare flag asks for the rewrites");
3802 assert_eq!(opts.opt_info_file, None, "and goes to standard error");
3803
3804 let (opts, _) = compile(&["-c", "-O2", "-fopt-info-missed-note", "a.c"]);
3805 assert_eq!(opts.opt_info, ["missed-note"]);
3806
3807 // Two flags add up rather than the second replacing the first, and the file is the last
3808 // one that named a file, which is how GCC treats both.
3809 let (opts, _) =
3810 compile(&["-c", "-O2", "-fopt-info-missed=one.txt", "-fopt-info-all=two.txt", "a.c"]);
3811 assert_eq!(opts.opt_info, ["missed", "all"]);
3812 assert_eq!(opts.opt_info_file.as_deref(), Some("two.txt"));
3813
3814 let e = parse_args(&args(&["-fopt-info-vectorized", "a.c"])).unwrap_err();
3815 assert!(e.message.contains("vectorized"), "{}", e.message);
3816 assert!(e.message.contains("`missed`"), "{}", e.message);
3817 let e = parse_args(&args(&["-fopt-info-missed=", "a.c"])).unwrap_err();
3818 assert!(e.message.contains("no file"), "{}", e.message);
3819 }
3820
3821 #[test]
3822 fn verify_each_is_unstable_and_off_unless_it_was_asked_for() {
3823 let (opts, _) = compile(&["-c", "-Zverify-each", "a.c"]);
3824 assert!(opts.verify_each);
3825 assert!(!USAGE.contains("verify-each"), "an unstable option stays out of the usage text");
3826 }
3827
3828 #[test]
3829 fn dash_o_needs_an_argument() {
3830 let e = parse_args(&args(&["a.c", "-o"])).unwrap_err();
3831 assert_eq!(e.message, "-o requires an argument");
3832 }
3833
3834 #[test]
3835 fn dash_d_and_dash_u_are_read_joined_or_separated_and_keep_their_order() {
3836 let (opts, _) = compile(&["-DFOO=1", "-D", "BAR", "-UBAZ", "-U", "QUX", "a.c"]);
3837 assert_eq!(opts.defines, ["FOO=1", "BAR"]);
3838 assert_eq!(opts.undefines, ["BAZ", "QUX"]);
3839 }
3840
3841 #[test]
3842 fn the_include_flags_land_on_the_chain_each_one_names() {
3843 // A sysroot with nothing under it, so that the library's own directories are the
3844 // same on every machine this test runs on, which is none of them.
3845 let (opts, _) = compile(&[
3846 "-Ii",
3847 "-iquote",
3848 "q",
3849 "-isystem",
3850 "sys",
3851 "-idirafter",
3852 "after",
3853 "--sysroot=/nowhere-at-all",
3854 "a.c",
3855 ]);
3856 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
3857 // The compiler's own headers sit after every `-isystem` and before `-idirafter`,
3858 // which is where GCC puts its own: a directory the user named outranks ours.
3859 assert_eq!(dirs, ["q", "i", "sys", runtime::DIR, "after"]);
3860 assert!(!opts.search.dirs()[1].is_system);
3861 assert!(opts.search.dirs()[2].is_system);
3862 }
3863
3864 #[test]
3865 fn the_librarys_headers_come_after_the_compilers_own_and_go_away_with_them() {
3866 // Which machine this runs on decides what is on the path, so the test is about the
3867 // order rather than about the names: ours is on it, the library's follow it, and
3868 // `-nostdinc` is the one flag that takes both halves of the pair off at once.
3869 let (opts, _) = compile(&["a.c"]);
3870 let dirs = opts.search.dirs();
3871 let ours = dirs.iter().position(|d| d.path.to_str() == Some(runtime::DIR));
3872 assert_eq!(ours, Some(0), "{dirs:?}");
3873 assert!(dirs[1..].iter().all(|d| d.is_system), "{dirs:?}");
3874 let (bare, _) = compile(&["-nostdinc", "a.c"]);
3875 assert!(bare.search.dirs().is_empty(), "{:?}", bare.search.dirs());
3876 }
3877
3878 #[test]
3879 fn a_sysroot_moves_the_librarys_directories_and_nothing_else() {
3880 let (opts, _) = compile(&["-isystem", "sys", "--sysroot=/nowhere-at-all", "a.c"]);
3881 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
3882 assert_eq!(dirs, ["sys", runtime::DIR]);
3883 }
3884
3885 #[test]
3886 fn a_cross_compile_reads_the_targets_own_headers_rather_than_the_ones_next_door() {
3887 // The target is not the machine this test runs on wherever it runs, so the answer is the
3888 // same on all of them: the libc's two include directories for that target, the kernel's
3889 // two, and nothing from here. A header read from here is the quiet failure of section 8.5, a
3890 // program that builds on the build machine and is wrong everywhere else.
3891 let (opts, _) = compile(&["--target=riscv64-linux-musl", "-c", "a.c"]);
3892 let dirs: Vec<&std::path::Path> =
3893 opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
3894 let root = cache::dir().join("sysroots").join("riscv64-linux-musl");
3895 let kernel = cache::dir().join("kernel-headers");
3896 assert_eq!(dirs.len(), 5, "{dirs:?}");
3897 assert_eq!(dirs[0], std::path::Path::new(runtime::DIR));
3898 assert_eq!(dirs[1], root.join("include").join("riscv64"));
3899 assert_eq!(dirs[2], root.join("include").join("generic"));
3900 // The kernel's, which are beside the sysroots rather than inside one, because every target
3901 // that shares an architecture reads the same files.
3902 assert_eq!(dirs[3], kernel.join("riscv"));
3903 assert_eq!(dirs[4], kernel.join("generic"));
3904 }
3905
3906 #[test]
3907 fn a_cross_compile_to_something_that_is_not_linux_reads_no_kernel_headers() {
3908 // The other side of the same answer. Windows has its own system headers and no `linux/` at
3909 // all, so the list is the libc's two and the question never arises, which is the `None` that
3910 // `link::cross_kernel` returns rather than a directory nothing would be found in.
3911 let (opts, _) = compile(&["--target=x86_64-pc-windows-gnu", "-c", "a.c"]);
3912 let dirs: Vec<&std::path::Path> =
3913 opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
3914 assert_eq!(dirs.len(), 3, "{dirs:?}");
3915 assert!(!dirs.iter().any(|dir| dir.ends_with("kernel-headers")), "{dirs:?}");
3916 }
3917
3918 #[test]
3919 fn the_glibc_version_macro_goes_with_the_bundled_tree_and_with_nothing_else() {
3920 // One tree serves every glibc release, so the release is what the target supplies, and the
3921 // condition is the same one that chose the directories. A host glibc and a tree somebody
3922 // named both define `__GLIBC_MINOR__` in their own `features.h`, and two definitions with
3923 // different values is a warning on every compilation of every file.
3924 //
3925 // The architecture is chosen against this machine's rather than written down, because the
3926 // bundled tree is only in effect for a target that is not this machine. The first version of
3927 // this test said x86_64-linux-gnu, which is a cross compile on a mac and this machine on a
3928 // Linux runner, so it passed here and failed there.
3929 let gnu = format!("--target={}-linux-gnu", cross_arch());
3930 let (bundled, _) = compile(&[&gnu, "-c", "a.c"]);
3931 assert_eq!(bundled.glibc_minor, Some(44));
3932 let pin = format!("{gnu}.2.28");
3933 let (pinned, _) = compile(&[&pin, "-c", "a.c"]);
3934 assert_eq!(pinned.glibc_minor, Some(28));
3935
3936 let (named, _) = compile(&[&gnu, "--sysroot=/nowhere-at-all", "-c", "a.c"]);
3937 assert_eq!(named.glibc_minor, None);
3938 let (none, _) = compile(&[&gnu, "-nostdinc", "-c", "a.c"]);
3939 assert_eq!(none.glibc_minor, None);
3940 let musl = format!("--target={}-linux-musl", cross_arch());
3941 let (musl, _) = compile(&[&musl, "-c", "a.c"]);
3942 assert_eq!(musl.glibc_minor, None);
3943
3944 // And this machine's own target gets nothing, whatever this machine is, because its headers
3945 // come from the machine and its own `features.h` defines the macro. On a glibc Linux box
3946 // that is the case this test had backwards; on a mac it is true for the other reason, which
3947 // is that Darwin is not a glibc target at all.
3948 if let Some(host) = Triple::host() {
3949 let native = format!("--target={}", host.tuple());
3950 let (native, _) = compile(&[&native, "-c", "a.c"]);
3951 assert_eq!(native.glibc_minor, None);
3952 }
3953 }
3954
3955 #[test]
3956 fn a_pinned_release_on_this_machines_own_target_reads_the_bundled_tree() {
3957 // The end to end half of the answer in `link::cross_for`. A release named for this machine's
3958 // own target is a cross compile, so the headers are the bundled tree's and the macro says
3959 // what was asked for rather than what this machine has.
3960 //
3961 // Only on a glibc box, because a release is a glibc release: a mac has no `__GLIBC_MINOR__`
3962 // to get wrong and nothing to pin. That makes this a test the Linux runners carry, which is
3963 // where the case lives.
3964 let Some(host) = Triple::host() else { return };
3965 if host.env != rucc_target::Env::Gnu {
3966 return;
3967 }
3968 let pin = format!("--target={}.2.28", host.tuple());
3969 let (opts, _) = compile(&[&pin, "-c", "a.c"]);
3970 assert_eq!(opts.glibc_minor, Some(28));
3971 let root = cache::dir().join("sysroots").join(format!("{}.2.28", host.tuple()));
3972 let dirs: Vec<&std::path::Path> =
3973 opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
3974 assert!(dirs.iter().any(|dir| dir.starts_with(&root)), "{dirs:?}");
3975 // And nothing of this machine's, which is the failure this was: a program compiled against
3976 // 2.44 declarations and told it was 2.28.
3977 assert!(!dirs.iter().any(|dir| *dir == std::path::Path::new("/usr/include")), "{dirs:?}");
3978 }
3979
3980 /// An architecture that is not this machine's, out of the three the driver has targets for.
3981 ///
3982 /// A test about the bundled sysroot has to name a target that is not the host, because a target
3983 /// that is the host reads the host's own headers and libraries. Asking which machine this is
3984 /// beats picking a row and hoping, and it is two lines.
3985 fn cross_arch() -> &'static str {
3986 match Triple::host().map(|host| host.arch) {
3987 Some(rucc_target::Arch::X86_64) => "aarch64",
3988 _ => "x86_64",
3989 }
3990 }
3991
3992 #[test]
3993 fn a_glibc_newer_than_the_bundled_tree_is_refused_by_name() {
3994 // Both versions in the message, because the two things a person can do about it are pin a
3995 // release the tree has and name a sysroot that has the one they asked for, and neither is a
3996 // choice they can make without knowing which release the tree is.
3997 //
3998 // Not this machine's architecture, for the reason the test above gives: the refusal is about
3999 // the bundled tree, and the bundled tree is not what a target that is this machine reads.
4000 let target = format!("--target={}-linux-gnu.2.99", cross_arch());
4001 let message = refused(&[&target, "-c", "a.c"]);
4002 assert!(message.contains("asked for glibc 2.99"), "{message}");
4003 assert!(message.contains("bundled headers are glibc 2.44"), "{message}");
4004 assert!(message.contains("--sysroot"), "{message}");
4005 }
4006
4007 #[test]
4008 fn a_sysroot_the_user_named_is_still_what_a_cross_compile_reads() {
4009 // The tree somebody assembled beats the one we would build, on the headers as on the
4010 // libraries. It is empty here, which is why the list comes out short: the directories under
4011 // it are checked for rather than assumed, and a tree that is not there offers nothing.
4012 let (opts, _) =
4013 compile(&["--target=riscv64-linux-musl", "--sysroot=/nowhere-at-all", "-c", "a.c"]);
4014 let dirs: Vec<&std::path::Path> =
4015 opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
4016 assert_eq!(dirs, [std::path::Path::new(runtime::DIR)]);
4017 }
4018
4019 #[test]
4020 fn dash_i_dash_moves_the_bracket_directories_into_the_quoted_chain() {
4021 let (opts, _) =
4022 compile(&["-Iinc1", "-iquote", "inc2", "-I-", "-Iinc3", "-nostdinc", "a.c"]);
4023 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
4024 assert_eq!(dirs, ["inc1", "inc2", "inc3"]);
4025 // An angled include sees only what came after the flag.
4026 assert_eq!(opts.search.start(IncludeForm::Angled), 2);
4027 assert!(!opts.search.searches_current_dir());
4028 }
4029
4030 #[test]
4031 fn the_prefix_flags_stick_what_iprefix_said_on_the_front_of_what_follows_it() {
4032 let (opts, _) = compile(&[
4033 "-iprefix",
4034 "/tools/",
4035 "-iwithprefix",
4036 "late",
4037 "-iwithprefixbefore",
4038 "early",
4039 "-iprefix",
4040 "/other/",
4041 "-iwithprefix",
4042 "last",
4043 "-nostdinc",
4044 "a.c",
4045 ]);
4046 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
4047 // `-iwithprefixbefore` is an `-I` and the other two are `-isystem`, which is where GCC
4048 // puts them rather than where its manual says it does.
4049 assert_eq!(dirs, ["/tools/early", "/tools/late", "/other/last"]);
4050 assert!(!opts.search.dirs()[0].is_system);
4051 assert!(opts.search.dirs()[1].is_system);
4052 }
4053
4054 #[test]
4055 fn the_files_named_on_the_command_line_keep_their_order_and_which_flag_named_them() {
4056 let (opts, _) =
4057 compile(&["-include", "one.h", "-imacros", "two.h", "-include", "3.h", "a.c"]);
4058 let names: Vec<&str> = opts.preincludes.iter().map(|p| p.name.as_str()).collect();
4059 assert_eq!(names, ["one.h", "two.h", "3.h"]);
4060 assert_eq!(opts.preincludes.iter().filter(|p| p.macros_only).count(), 1);
4061 }
4062
4063 #[test]
4064 fn nostdinc_takes_the_compilers_own_headers_off_the_path() {
4065 let (opts, _) = compile(&["-Ii", "-nostdinc", "a.c"]);
4066 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
4067 assert_eq!(dirs, ["i"]);
4068 }
4069
4070 #[test]
4071 fn the_dialect_flags_set_the_language_and_the_extensions_separately() {
4072 let (opts, _) = compile(&["-std=gnu11", "a.c"]);
4073 assert_eq!(opts.std, Std::C11);
4074 assert!(opts.gnu_extensions);
4075
4076 let (opts, _) = compile(&["-std=iso9899:1999", "a.c"]);
4077 assert_eq!(opts.std, Std::C99);
4078 assert!(!opts.gnu_extensions);
4079
4080 let (opts, _) = compile(&["-ansi", "a.c"]);
4081 assert_eq!(opts.std, Std::C89);
4082 assert!(!opts.gnu_extensions);
4083
4084 let e = parse_args(&args(&["-std=c94jr", "a.c"])).unwrap_err();
4085 assert!(e.message.contains("unknown dialect"), "{}", e.message);
4086 }
4087
4088 #[test]
4089 fn the_dump_letters_are_a_family_and_everything_else_beginning_with_d_is_not() {
4090 let (opts, _) = compile(&["-dM", "a.c"]);
4091 assert!(opts.dumps.macros);
4092
4093 // Packed, the way GCC takes them, and a letter in the family we have not written yet
4094 // is accepted and does nothing rather than failing a build.
4095 let (opts, _) = compile(&["-dDM", "a.c"]);
4096 assert!(opts.dumps.macros);
4097 let (opts, _) = compile(&["-dD", "a.c"]);
4098 assert!(!opts.dumps.macros);
4099
4100 let (opts, _) = compile(&["a.c"]);
4101 assert!(!opts.dumps.any());
4102
4103 // `-dumpversion` is a different flag that happens to start the same way, and it is read
4104 // as itself rather than as a dump of nothing.
4105 assert_eq!(printed(&["-dumpversion", "a.c"]), VERSION);
4106 }
4107
4108 #[test]
4109 fn the_gcc_version_claimed_is_a_flag_and_the_short_spellings_are_the_ones_people_write() {
4110 let (opts, _) = compile(&["a.c"]);
4111 assert_eq!(
4112 opts.gnuc,
4113 GnucVersion { major: 7, minor: 0, patch: 0 },
4114 "the lowest claim a modern glibc gives its own declarations to"
4115 );
4116
4117 let (opts, _) = compile(&["-fgnuc-version=15.1.0", "a.c"]);
4118 assert_eq!(opts.gnuc, GnucVersion { major: 15, minor: 1, patch: 0 });
4119
4120 // A missing component is zero. `gcc -dumpversion` says `15` on a release with no
4121 // patchlevel and a harness that pastes that back has to be understood.
4122 let (opts, _) = compile(&["-fgnuc-version=15", "a.c"]);
4123 assert_eq!(opts.gnuc, GnucVersion { major: 15, minor: 0, patch: 0 });
4124
4125 let (opts, _) = compile(&["-fgnuc-version=13.2", "a.c"]);
4126 assert_eq!(opts.gnuc, GnucVersion { major: 13, minor: 2, patch: 0 });
4127
4128 let e = parse_args(&args(&["-fgnuc-version=15.x", "a.c"])).unwrap_err();
4129 assert!(e.message.contains("minor that is not a number"), "{}", e.message);
4130
4131 let e = parse_args(&args(&["-fgnuc-version=1.2.3.4", "a.c"])).unwrap_err();
4132 assert!(e.message.contains("more than three"), "{}", e.message);
4133 }
4134
4135 #[test]
4136 fn pedantic_has_two_spellings_and_is_not_the_same_knob_as_the_dialect() {
4137 let (opts, _) = compile(&["-std=c17", "-pedantic", "a.c"]);
4138 assert!(opts.pedantic);
4139 assert_eq!(opts.std, Std::C17);
4140
4141 // The `-W` family's name for it, which is what a build that groups its warning flags
4142 // tends to write.
4143 let (opts, _) = compile(&["-Wpedantic", "a.c"]);
4144 assert!(opts.pedantic);
4145
4146 let (opts, _) = compile(&["-std=c17", "a.c"]);
4147 assert!(!opts.pedantic, "a dialect on its own does not diagnose an extension");
4148 }
4149
4150 #[test]
4151 fn dash_p_and_dash_ffreestanding_reach_the_options() {
4152 let (opts, _) = compile(&["-E", "-P", "-ffreestanding", "a.c"]);
4153 assert!(!opts.line_markers);
4154 assert!(!opts.hosted);
4155 assert_eq!(opts.emit, EmitKind::Preprocessed);
4156 }
4157
4158 /// The two ways a build says it means its own function by a name the C library also has.
4159 ///
4160 /// `-fno-builtin` is all of them and `-fno-builtin-<name>` is one, and the second is what a
4161 /// build writes when it means its own `memcpy` and the library's everything else. The name is
4162 /// kept as it was written and not checked against anything, because a program is allowed to
4163 /// mean something by a name this compiler has never heard of.
4164 #[test]
4165 fn the_builtin_flags_are_read_in_both_directions_and_one_name_at_a_time() {
4166 let (opts, _) = compile(&["-c", "a.c"]);
4167 assert!(opts.builtins, "a library name means the library function by default");
4168 assert!(opts.no_builtin.is_empty());
4169
4170 let (opts, _) = compile(&["-c", "-fno-builtin", "a.c"]);
4171 assert!(!opts.builtins);
4172
4173 let (opts, _) = compile(&["-c", "-fno-builtin", "-fbuiltin", "a.c"]);
4174 assert!(opts.builtins, "the last mention decides");
4175
4176 let (opts, _) = compile(&["-c", "-fno-builtin-memcpy", "-fno-builtin-nonesuch", "a.c"]);
4177 assert!(opts.builtins, "one name is not the family");
4178 assert_eq!(opts.no_builtin, vec!["memcpy".to_owned(), "nonesuch".to_owned()]);
4179 }
4180
4181 /// `-fvisibility=`, which is on every cmake project that cares about which names it exports
4182 /// and which was refused as an unknown option until now.
4183 ///
4184 /// Four spellings and three answers. `internal` is hidden plus a promise about never taking
4185 /// the address across a component boundary, and nothing derives anything from that promise
4186 /// here, so it comes out as the weaker of the two rather than as a refusal that stops a build
4187 /// over a distinction this compiler does not make.
4188 #[test]
4189 fn visibility_takes_the_four_spellings_gcc_takes_and_refuses_the_rest() {
4190 let (opts, _) = compile(&["-c", "a.c"]);
4191 assert_eq!(opts.visibility, Visibility::Default, "exported unless something says not");
4192
4193 for (written, wanted) in [
4194 ("default", Visibility::Default),
4195 ("hidden", Visibility::Hidden),
4196 ("internal", Visibility::Hidden),
4197 ("protected", Visibility::Protected),
4198 ] {
4199 let (opts, _) = compile(&["-c", &format!("-fvisibility={written}"), "a.c"]);
4200 assert_eq!(opts.visibility, wanted, "{written}");
4201 }
4202
4203 // The last mention decides, which is what every other flag of this shape does and what a
4204 // build that turns something off for one directory relies on.
4205 let (opts, _) = compile(&["-c", "-fvisibility=hidden", "-fvisibility=default", "a.c"]);
4206 assert_eq!(opts.visibility, Visibility::Default, "the last mention decides");
4207
4208 // A spelling gcc does not take is refused rather than read as the default, because a
4209 // build that meant hidden and got exported is a library with the wrong interface and
4210 // nothing said about it anywhere.
4211 let failed = parse_args(&args(&["-fvisibility=none", "a.c"])).expect_err("refused");
4212 assert!(failed.to_string().contains("is not a visibility"), "{failed}");
4213 }
4214
4215 /// `-ffp-contract=`, which is the one flag in the floating point group that is kept rather than
4216 /// described, and the values are gcc 16's three.
4217 #[test]
4218 fn how_far_a_multiply_and_an_addition_may_be_fused_is_asked_for() {
4219 let (opts, _) = compile(&["-c", "a.c"]);
4220 assert_eq!(opts.fp_contract, Contract::Off, "a licence nobody granted is not assumed");
4221
4222 for (written, wanted) in
4223 [("off", Contract::Off), ("on", Contract::On), ("fast", Contract::Fast)]
4224 {
4225 let (opts, _) = compile(&["-c", &format!("-ffp-contract={written}"), "a.c"]);
4226 assert_eq!(opts.fp_contract, wanted, "{written}");
4227 }
4228
4229 let (opts, _) = compile(&["-c", "-ffp-contract=fast", "-ffp-contract=off", "a.c"]);
4230 assert_eq!(opts.fp_contract, Contract::Off, "the last mention decides");
4231
4232 // Refused rather than read as one of the three, because a build that asked for no fusing
4233 // and was given the default would be one whose numbers change and whose command line says
4234 // they should not. gcc refuses the same spellings and names the same three in its message.
4235 for bad in ["-ffp-contract=none", "-ffp-contract=", "-ffp-contract=Fast"] {
4236 let failed = parse_args(&args(&[bad, "a.c"])).expect_err("refused");
4237 assert!(failed.to_string().contains("is not a contraction"), "{bad}: {failed}");
4238 }
4239
4240 // And the other one that takes a value, which is taken and kept nowhere: every operation
4241 // here is computed in the type it was written in, so `standard` is what happens and the
4242 // other two are permission to do something this does not do.
4243 let failed = parse_args(&args(&["-fexcess-precision=long", "a.c"])).expect_err("refused");
4244 assert!(failed.to_string().contains("is not an excess precision"), "{failed}");
4245 }
4246
4247 /// The four prefix mapping flags, which are what a distribution passes to get the same bytes
4248 /// out of `/build/pkg-1.2` and out of `/home/someone/pkg-1.2`. Three lists rather than one
4249 /// because gcc has three, and `-ffile-prefix-map=` is the three of them at once.
4250 #[test]
4251 fn a_prefix_mapping_flag_goes_on_the_list_its_spelling_names() {
4252 let (opts, _) = compile(&["-c", "a.c"]);
4253 assert!(opts.prefix_map.macros.is_empty(), "nothing is rewritten unless it is asked for");
4254 assert!(opts.prefix_map.debug.is_empty(), "nor here");
4255 assert!(opts.prefix_map.profile.is_empty(), "nor here");
4256
4257 let (opts, _) = compile(&["-c", "-fmacro-prefix-map=/build=.", "a.c"]);
4258 assert_eq!(opts.prefix_map.macros.apply("/build/a.c"), "./a.c", "the one it names");
4259 assert!(opts.prefix_map.debug.is_empty(), "and not the two it does not");
4260
4261 let (opts, _) = compile(&["-c", "-fdebug-prefix-map=/build=.", "a.c"]);
4262 assert_eq!(opts.prefix_map.debug.apply("/build/a.c"), "./a.c", "the one it names");
4263 assert!(opts.prefix_map.macros.is_empty(), "and not the two it does not");
4264
4265 let (opts, _) = compile(&["-c", "-fprofile-prefix-map=/build=.", "a.c"]);
4266 assert_eq!(opts.prefix_map.profile.apply("/build/a.c"), "./a.c", "the one it names");
4267 assert!(opts.prefix_map.macros.is_empty(), "and not the two it does not");
4268
4269 let (opts, _) = compile(&["-c", "-ffile-prefix-map=/build=.", "a.c"]);
4270 for list in [&opts.prefix_map.macros, &opts.prefix_map.debug, &opts.prefix_map.profile] {
4271 assert_eq!(list.apply("/build/a.c"), "./a.c", "all three at once");
4272 }
4273
4274 // Every mention is kept and the last one that matches wins, unlike the flags above whose
4275 // last mention replaces the earlier ones. A build writes one of these per source root and
4276 // expects all of them to be in force, which is the whole point of a list.
4277 let (opts, _) =
4278 compile(&["-c", "-ffile-prefix-map=/a=one", "-ffile-prefix-map=/b=two", "a.c"]);
4279 assert_eq!(opts.prefix_map.macros.apply("/a/x.c"), "one/x.c", "the earlier one still acts");
4280 assert_eq!(opts.prefix_map.macros.apply("/b/x.c"), "two/x.c", "and so does the later one");
4281
4282 // An argument with no `=` is refused rather than ignored, because a build whose paths were
4283 // meant to be rewritten and were not is one that ships the build directory's name and says
4284 // nothing about it. gcc refuses the same thing.
4285 for bad in ["-fmacro-prefix-map=nope", "-ffile-prefix-map=", "-fdebug-prefix-map=/build"] {
4286 let failed = parse_args(&args(&[bad, "a.c"])).expect_err("refused");
4287 assert!(failed.to_string().contains("is not a rewrite for"), "{bad}: {failed}");
4288 }
4289 }
4290
4291 /// `-ffunction-sections` and `-fdata-sections`, which are what make `--gc-sections` able to
4292 /// drop anything: a linker can leave out a section nothing reaches and cannot leave out half of
4293 /// one. A kernel and an embedded image are both linked that way.
4294 ///
4295 /// Two flags rather than one because gcc has two, and a build that asks for one of them and not
4296 /// the other is a build that measured something: splitting the code is nearly free at link time
4297 /// and splitting the data can defeat the linker's ordering of what is next to what.
4298 #[test]
4299 fn a_section_per_function_and_a_section_per_variable_are_asked_for_one_at_a_time() {
4300 let (opts, _) = compile(&["-c", "a.c"]);
4301 assert!(!opts.function_sections, "one text section unless something says otherwise");
4302 assert!(!opts.data_sections);
4303
4304 let (opts, _) = compile(&["-c", "-ffunction-sections", "a.c"]);
4305 assert!(opts.function_sections);
4306 assert!(!opts.data_sections, "one flag is not the other");
4307
4308 let (opts, _) = compile(&["-c", "-fdata-sections", "a.c"]);
4309 assert!(opts.data_sections);
4310 assert!(!opts.function_sections);
4311
4312 // Both directions taken, and the off one is what happens anyway rather than a refusal,
4313 // since a build that writes it is asking for the default.
4314 let (opts, _) = compile(&[
4315 "-c",
4316 "-ffunction-sections",
4317 "-fno-function-sections",
4318 "-fdata-sections",
4319 "-fno-data-sections",
4320 "a.c",
4321 ]);
4322 assert!(!opts.function_sections, "the last mention decides");
4323 assert!(!opts.data_sections, "the last mention decides");
4324 }
4325
4326 /// `-fgnu89-inline`, which is off by default and is not implied by anything on the command
4327 /// line, since the dialect asks for GNU's reading further in rather than through this.
4328 #[test]
4329 fn gnu89_inline_is_off_until_it_is_asked_for_and_the_last_mention_decides() {
4330 let (opts, _) = compile(&["-c", "a.c"]);
4331 assert!(!opts.gnu89_inline, "C's reading of inline by default");
4332
4333 let (opts, _) = compile(&["-c", "-fgnu89-inline", "a.c"]);
4334 assert!(opts.gnu89_inline);
4335
4336 let (opts, _) = compile(&["-c", "-fgnu89-inline", "-fno-gnu89-inline", "a.c"]);
4337 assert!(!opts.gnu89_inline, "the last mention decides");
4338
4339 // The C89 dialects are under GNU's reading whether this was written or not, so the flag
4340 // stays off there and the dialect is what the checker and the macro set both ask. That is
4341 // also why `-std=c89 -fno-gnu89-inline` needs no diagnostic: it asks for the reading the
4342 // dialect already has. gcc refuses that command line, which is measured in the issue.
4343 let (opts, _) = compile(&["-c", "-std=c89", "a.c"]);
4344 assert!(!opts.gnu89_inline);
4345 }
4346
4347 /// Both spellings of both frame flags, since a build that wants one usually writes the
4348 /// other beside it for the one file that has to be compiled the ordinary way.
4349 #[test]
4350 fn the_two_frame_flags_are_read_in_both_directions() {
4351 let (opts, _) = compile(&["-c", "a.c"]);
4352 assert!(!opts.frame_pointer, "gcc omits it above -O0 and so does this");
4353 assert!(opts.red_zone, "the psABI has one and nothing said not to use it");
4354
4355 let (opts, _) = compile(&["-c", "-fno-omit-frame-pointer", "-mno-red-zone", "a.c"]);
4356 assert!(opts.frame_pointer);
4357 assert!(!opts.red_zone);
4358
4359 let (opts, _) = compile(&[
4360 "-c",
4361 "-fno-omit-frame-pointer",
4362 "-fomit-frame-pointer",
4363 "-mno-red-zone",
4364 "-mred-zone",
4365 "a.c",
4366 ]);
4367 assert!(!opts.frame_pointer, "the last one wins, as it does in gcc");
4368 assert!(opts.red_zone);
4369 }
4370
4371 /// Four flags rather than one with an argument, which is how gcc spells them, and the negative
4372 /// spelled three ways because a build that turns one off writes whichever it turned on.
4373 #[test]
4374 fn the_stack_protector_is_four_flags_and_the_last_one_wins() {
4375 let (opts, _) = compile(&["-c", "a.c"]);
4376 assert_eq!(opts.protector, Protector::None, "gcc protects nothing unless it was asked");
4377
4378 for (flag, want) in [
4379 ("-fstack-protector", Protector::Buffers),
4380 ("-fstack-protector-strong", Protector::Strong),
4381 ("-fstack-protector-all", Protector::All),
4382 ] {
4383 let (opts, _) = compile(&["-c", flag, "a.c"]);
4384 assert_eq!(opts.protector, want, "{flag}");
4385 }
4386
4387 // What a package build does: the strong one in the global flags and one directory that
4388 // cannot have a protector turning it off on the line after.
4389 for off in ["-fno-stack-protector", "-fno-stack-protector-strong"] {
4390 let (opts, _) = compile(&["-c", "-fstack-protector-strong", off, "a.c"]);
4391 assert_eq!(opts.protector, Protector::None, "{off}");
4392 }
4393 let (opts, _) = compile(&["-c", "-fno-stack-protector", "-fstack-protector-all", "a.c"]);
4394 assert_eq!(opts.protector, Protector::All, "the last one wins either way round");
4395 }
4396
4397 /// A switch rather than a level, because how a frame is taken is one question and which
4398 /// functions get a canary is another, and gcc spells it that way for the same reason.
4399 #[test]
4400 fn taking_a_frame_a_page_at_a_time_is_off_until_it_is_asked_for() {
4401 let (opts, _) = compile(&["-c", "a.c"]);
4402 assert!(!opts.stack_clash, "gcc takes a frame in one subtraction unless it was asked");
4403
4404 let (opts, _) = compile(&["-c", "-fstack-clash-protection", "a.c"]);
4405 assert!(opts.stack_clash);
4406
4407 // The same shape a package build uses for the protector: on in the global flags and off
4408 // for the one directory that cannot have it.
4409 let (opts, _) =
4410 compile(&["-c", "-fstack-clash-protection", "-fno-stack-clash-protection", "a.c"]);
4411 assert!(!opts.stack_clash);
4412 let (opts, _) =
4413 compile(&["-c", "-fno-stack-clash-protection", "-fstack-clash-protection", "a.c"]);
4414 assert!(opts.stack_clash, "the last one wins either way round");
4415
4416 // The two are independent, since one is about the frame and the other about the function.
4417 let (opts, _) =
4418 compile(&["-c", "-fstack-clash-protection", "-fstack-protector-strong", "a.c"]);
4419 assert!(opts.stack_clash);
4420 assert_eq!(opts.protector, Protector::Strong);
4421 }
4422
4423 /// One flag with an argument rather than a family of spellings, because what it asks about is
4424 /// which of the two edges of a control flow transfer is checked and the two are not separate
4425 /// questions to the hardware.
4426 #[test]
4427 fn which_control_flow_edges_are_checked_is_asked_for_by_name() {
4428 let (opts, _) = compile(&["-c", "a.c"]);
4429 assert_eq!(opts.control, Control::None, "gcc's default on the targets this compiler has");
4430
4431 for (arg, want) in [
4432 ("-fcf-protection", Control::Full),
4433 ("-fcf-protection=full", Control::Full),
4434 ("-fcf-protection=branch", Control::Branch),
4435 ("-fcf-protection=return", Control::Return),
4436 ("-fcf-protection=none", Control::None),
4437 ("-fcf-protection=check", Control::Check),
4438 ] {
4439 let (opts, _) = compile(&["-c", arg, "a.c"]);
4440 assert_eq!(opts.control, want, "{arg}");
4441 }
4442
4443 // The shape a package build uses: on in the global flags and off for the one directory
4444 // that cannot have it, whichever of the two spellings of off it reaches for.
4445 let (opts, _) = compile(&["-c", "-fcf-protection=full", "-fno-cf-protection", "a.c"]);
4446 assert_eq!(opts.control, Control::None);
4447 let (opts, _) = compile(&["-c", "-fno-cf-protection", "-fcf-protection=branch", "a.c"]);
4448 assert_eq!(opts.control, Control::Branch, "the last one wins either way round");
4449 }
4450
4451 /// The profiler is asked for by two spellings, and where its hook goes by two more.
4452 ///
4453 /// The two halves are separate on purpose. `-mfentry` on its own says where a call would go and
4454 /// asks for no call, which is what gcc does with it, and a build system that sets it globally
4455 /// and asks for the profile per directory needs that to be true rather than an error.
4456 ///
4457 /// The link is asserted alongside, because the flag changes it too and a build that compiled
4458 /// with it and linked without it is a program that calls the hook everywhere and never writes a
4459 /// profile.
4460 #[test]
4461 fn the_profiler_and_where_its_hook_goes_are_two_separate_questions() {
4462 let (opts, _) = compile(&["-c", "a.c"]);
4463 assert!(!opts.profile);
4464 assert_eq!(opts.hook, Hook::Platform, "neither was named, so the target decides");
4465
4466 for arg in ["-pg", "-p"] {
4467 let (opts, _) = compile(&["-c", arg, "a.c"]);
4468 assert!(opts.profile, "{arg}");
4469 let (link, _) = linking(&[arg, "a.c"]);
4470 assert!(link.profile, "{arg} changes the link as well");
4471 }
4472
4473 for (arg, want) in [("-mfentry", Hook::Early), ("-mno-fentry", Hook::Late)] {
4474 let (opts, _) = compile(&["-c", arg, "a.c"]);
4475 assert_eq!(opts.hook, want, "{arg}");
4476 assert!(!opts.profile, "{arg} asks for no call of its own");
4477 }
4478
4479 let (opts, _) = compile(&["-c", "-mfentry", "-mno-fentry", "-pg", "a.c"]);
4480 assert_eq!(opts.hook, Hook::Late, "the last one wins");
4481 assert!(opts.profile);
4482 }
4483
4484 /// How much room a patcher is promised, which is one number or two.
4485 ///
4486 /// A command line that did not ask is asserted alongside, because the flag has to be written to
4487 /// mean anything and a build that reserved room nobody asked for would grow every function in
4488 /// it for nothing.
4489 #[test]
4490 fn the_room_a_patcher_is_promised_is_a_number_of_bytes_and_where_they_go() {
4491 let (opts, _) = compile(&["-c", "a.c"]);
4492 assert_eq!(opts.patchable, Patchable::default());
4493 assert!(!opts.patchable.any(), "nothing is reserved unless it was asked for");
4494
4495 let (opts, _) = compile(&["-c", "-fpatchable-function-entry=16", "a.c"]);
4496 assert_eq!(opts.patchable, Patchable { total: 16, before: 0 });
4497
4498 let (opts, _) = compile(&["-c", "-fpatchable-function-entry=5,3", "a.c"]);
4499 assert_eq!(opts.patchable, Patchable { total: 5, before: 3 });
4500 assert_eq!(opts.patchable.after(), 2);
4501
4502 // The last one wins, which is what every other flag of this shape does and what a build
4503 // that adds one to a command line it did not write is relying on.
4504 let (opts, _) = compile(&[
4505 "-c",
4506 "-fpatchable-function-entry=5,3",
4507 "-fpatchable-function-entry=2",
4508 "a.c",
4509 ]);
4510 assert_eq!(opts.patchable, Patchable { total: 2, before: 0 });
4511 }
4512
4513 /// And a request nothing could satisfy is refused rather than rounded into one that can be.
4514 #[test]
4515 fn room_in_front_of_the_label_that_is_more_than_the_room_asked_for_is_refused() {
4516 for arg in ["-fpatchable-function-entry=1,2", "-fpatchable-function-entry=x"] {
4517 let e = parse_args(&args(&["-c", arg, "a.c"])).unwrap_err();
4518 assert!(e.message.contains("is not an amount of room to reserve"), "{}", e.message);
4519 }
4520 }
4521
4522 /// What wraps rather than being undefined, which is two questions and three flags.
4523 ///
4524 /// The older flag is the pair of the newer two, which is gcc's own reading of it, so a build
4525 /// that writes `-fno-strict-overflow` gets both and a build that writes one of the others gets
4526 /// only what it asked for.
4527 #[test]
4528 fn what_overflows_rather_than_being_undefined_is_asked_for_two_ways() {
4529 let (opts, _) = compile(&["-c", "a.c"]);
4530 assert_eq!(opts.wrapping, Wrapping::NONE, "nothing wraps unless it was asked for");
4531
4532 let (opts, _) = compile(&["-c", "-fwrapv", "a.c"]);
4533 assert_eq!(opts.wrapping, Wrapping { signed: true, pointer: false, trap: false });
4534
4535 let (opts, _) = compile(&["-c", "-fwrapv-pointer", "a.c"]);
4536 assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: true, trap: false });
4537
4538 let (opts, _) = compile(&["-c", "-fno-strict-overflow", "a.c"]);
4539 assert_eq!(opts.wrapping, Wrapping::ALL);
4540
4541 // And the last one wins, in both directions. A build that turns one of these on globally
4542 // and off for one directory is relying on that, and so is one that writes the pair and
4543 // then takes half of it back.
4544 let (opts, _) = compile(&["-c", "-fwrapv", "-fno-wrapv", "a.c"]);
4545 assert_eq!(opts.wrapping, Wrapping::NONE);
4546
4547 let (opts, _) = compile(&["-c", "-fno-strict-overflow", "-fstrict-overflow", "a.c"]);
4548 assert_eq!(opts.wrapping, Wrapping::NONE);
4549
4550 let (opts, _) = compile(&["-c", "-fno-strict-overflow", "-fno-wrapv-pointer", "a.c"]);
4551 assert_eq!(opts.wrapping, Wrapping { signed: true, pointer: false, trap: false });
4552 }
4553
4554 /// And the other answer to the signed question cannot be held at the same time as the first.
4555 ///
4556 /// A program cannot both wrap and stop, so writing both is writing a contradiction, and gcc
4557 /// resolves it by letting the last one win rather than by reporting anything. That was measured
4558 /// against gcc 16 rather than read out of the manual, which says nothing about it: `-ftrapv
4559 /// -fwrapv` emits no checked calls and `-fwrapv -ftrapv` emits them.
4560 #[test]
4561 fn a_signed_overflow_that_stops_is_the_other_answer_and_not_a_third_one() {
4562 let (opts, _) = compile(&["-c", "-ftrapv", "a.c"]);
4563 assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: false, trap: true });
4564
4565 let (opts, _) = compile(&["-c", "-fwrapv", "-ftrapv", "a.c"]);
4566 assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: false, trap: true });
4567
4568 let (opts, _) = compile(&["-c", "-ftrapv", "-fwrapv", "a.c"]);
4569 assert_eq!(opts.wrapping, Wrapping { signed: true, pointer: false, trap: false });
4570
4571 let (opts, _) = compile(&["-c", "-ftrapv", "-fno-strict-overflow", "a.c"]);
4572 assert_eq!(opts.wrapping, Wrapping::ALL);
4573
4574 let (opts, _) = compile(&["-c", "-ftrapv", "-fno-trapv", "a.c"]);
4575 assert_eq!(opts.wrapping, Wrapping::NONE);
4576
4577 // And the flag that says what may be assumed says nothing about what happens, so it leaves
4578 // this alone where it takes the wrapping away. gcc does the same.
4579 let (opts, _) = compile(&["-c", "-ftrapv", "-fstrict-overflow", "a.c"]);
4580 assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: false, trap: true });
4581 }
4582
4583 /// What a plain `char` is, which is four spellings of two answers and nothing by default.
4584 ///
4585 /// Nothing is the target's own answer and has to stay distinct from both of the others, since
4586 /// the same command line means a signed `char` on x86-64 and an unsigned one on Linux's arm64.
4587 /// The negative spellings are the other flag rather than a way of asking for the default, which
4588 /// was measured against gcc 16: `-fno-signed-char` defines `__CHAR_UNSIGNED__` and
4589 /// `-fno-unsigned-char` does not.
4590 #[test]
4591 fn the_signedness_of_a_plain_char_is_asked_for_in_four_ways() {
4592 let (opts, _) = compile(&["-c", "a.c"]);
4593 assert_eq!(opts.char_signed, None);
4594
4595 for flag in ["-fsigned-char", "-fno-unsigned-char"] {
4596 let (opts, _) = compile(&["-c", flag, "a.c"]);
4597 assert_eq!(opts.char_signed, Some(true), "{flag}");
4598 }
4599
4600 for flag in ["-funsigned-char", "-fno-signed-char"] {
4601 let (opts, _) = compile(&["-c", flag, "a.c"]);
4602 assert_eq!(opts.char_signed, Some(false), "{flag}");
4603 }
4604
4605 // And the last one wins, which is what a build that sets one globally and the other for a
4606 // directory relies on.
4607 let (opts, _) = compile(&["-c", "-funsigned-char", "-fsigned-char", "a.c"]);
4608 assert_eq!(opts.char_signed, Some(true));
4609
4610 // And what is asked for reaches the target, because that is what every other part of the
4611 // compiler asks. The triple is one whose own answer is the opposite, so a session that
4612 // ignored the flag would still read as signed here.
4613 let (opts, _) =
4614 compile(&["-c", "--target=aarch64-unknown-linux-gnu", "-fsigned-char", "a.c"]);
4615 assert!(Session::new(*opts).target.char_is_signed);
4616 let (opts, _) = compile(&["-c", "--target=aarch64-unknown-linux-gnu", "a.c"]);
4617 assert!(!Session::new(*opts).target.char_is_signed);
4618 }
4619
4620 /// And the size of an enumeration, which is one question with two spellings.
4621 #[test]
4622 fn the_smallest_enumeration_is_asked_for_and_taken_back() {
4623 let (opts, _) = compile(&["-c", "a.c"]);
4624 assert!(!opts.short_enums);
4625
4626 let (opts, _) = compile(&["-c", "-fshort-enums", "a.c"]);
4627 assert!(opts.short_enums);
4628
4629 let (opts, _) = compile(&["-c", "-fshort-enums", "-fno-short-enums", "a.c"]);
4630 assert!(!opts.short_enums);
4631
4632 let (opts, _) = compile(&["-c", "-fno-short-enums", "-fshort-enums", "a.c"]);
4633 assert!(opts.short_enums);
4634 }
4635
4636 /// And a value nothing means is refused rather than taken for the nearest thing it looks like.
4637 ///
4638 /// `-fcf-protection=all` is the spelling somebody writes from memory, and a compiler that read
4639 /// it as `full` would be guessing, while one that let it fall through to the optimizer's `-f`
4640 /// family would report it as an unknown pass. Neither is the news the build wants.
4641 #[test]
4642 fn a_control_flow_protection_nothing_means_is_refused() {
4643 let e = parse_args(&args(&["-c", "-fcf-protection=all", "a.c"])).unwrap_err();
4644 assert!(e.message.contains("is not a control flow protection"), "{}", e.message);
4645 assert!(e.message.contains("full, branch, return, none or check"), "{}", e.message);
4646 }
4647
4648 #[test]
4649 fn the_link_flags_are_collected_apart_from_the_compilation() {
4650 let (link, _) = linking(&[
4651 "-static",
4652 "-nostartfiles",
4653 "-rdynamic",
4654 "-s",
4655 "-fuse-ld=mold",
4656 "-L/opt/lib",
4657 "-B",
4658 "/opt/tools",
4659 "a.c",
4660 ]);
4661 assert!(link.is_static);
4662 assert!(link.no_startfiles);
4663 assert!(link.export_dynamic);
4664 assert!(link.strip);
4665 assert_eq!(link.use_ld.as_deref(), Some("mold"));
4666 assert_eq!(link.search, vec![PathBuf::from("/opt/lib")]);
4667 assert_eq!(link.prefixes, vec![PathBuf::from("/opt/tools")]);
4668 }
4669
4670 #[test]
4671 fn a_comma_in_dash_wl_separates_two_arguments() {
4672 let (link, _) = linking(&["-Wl,-rpath,/opt/lib", "-Xlinker", "--as-needed", "a.c"]);
4673 assert_eq!(link.passthrough, vec!["-rpath", "/opt/lib", "--as-needed"]);
4674 }
4675
4676 #[test]
4677 fn a_library_keeps_its_place_between_the_objects() {
4678 // Link order is semantic: `-lm` written between two files resolves for the one before
4679 // it and not for the one after, so a library cannot be collected into a list of its own.
4680 // The target is named because the suffix of an object is the target's and this asserts
4681 // on the names: the same command line on a Windows host plans two `.obj` files.
4682 let (_, plan) = linking(&["--target=x86_64-unknown-linux-gnu", "a.c", "-lm", "b.c"]);
4683 let link = plan.link.expect("expected a link step");
4684 assert_eq!(
4685 link.inputs,
4686 vec![
4687 link::Item::File("a.o".into()),
4688 link::Item::Library("m".into()),
4689 link::Item::File("b.o".into()),
4690 ]
4691 );
4692 // And it is not a job, because there is nothing to compile in a library.
4693 assert_eq!(plan.jobs.len(), 2);
4694 }
4695
4696 #[test]
4697 fn a_library_on_a_dash_c_line_is_a_note_rather_than_an_error() {
4698 let (_, plan) = linking(&["-c", "-lm", "a.c"]);
4699 assert!(plan.link.is_none());
4700 assert!(plan.notes.iter().any(|n| n.contains("-lm")), "{:?}", plan.notes);
4701 }
4702
4703 #[test]
4704 fn the_sysroot_reaches_the_linker_as_well_as_the_headers() {
4705 let (link, _) = linking(&["--sysroot=/opt/root", "a.c"]);
4706 assert_eq!(link.sysroot, Some(PathBuf::from("/opt/root")));
4707 }
4708
4709 fn printed(s: &[&str]) -> String {
4710 match parse_args(&args(s)).expect("expected an answer") {
4711 Action::Print(line) => line,
4712 other => panic!("expected an answer, got {other:?}"),
4713 }
4714 }
4715
4716 fn refused(s: &[&str]) -> String {
4717 parse_args(&args(s)).expect_err("expected a refusal").message
4718 }
4719
4720 #[test]
4721 fn a_warning_flag_this_compiler_has_not_heard_of_is_taken_rather_than_refused() {
4722 // The rule in section 4.1, and the reason for it is autoconf: a configure script finds
4723 // out whether a warning flag exists by passing it and looking at the exit status, so a
4724 // compiler that refuses one it does not know fails a script written for a newer GCC.
4725 let (opts, _) = compile(&["-Wall", "-Wextra", "-Wno-format-truncation", "-c", "a.c"]);
4726 assert!(!opts.warnings_are_errors);
4727 assert!(opts.warnings);
4728 // The two spellings that do mean something are still read.
4729 let (opts, _) = compile(&["-Werror", "-c", "a.c"]);
4730 assert!(opts.warnings_are_errors);
4731 let (opts, _) = compile(&["-w", "-c", "a.c"]);
4732 assert!(!opts.warnings);
4733 let (opts, _) = compile(&["-pedantic-errors", "-c", "a.c"]);
4734 assert!(opts.pedantic && opts.warnings_are_errors);
4735 }
4736
4737 #[test]
4738 fn an_argument_for_a_separate_tool_is_refused_rather_than_dropped() {
4739 // Every one of these says something about the output, so the wrong answer is silence.
4740 assert!(refused(&["-Wa,--noexecstack", "-c", "a.c"]).contains("separate assembler"));
4741 assert!(refused(&["-Wp,-DX", "-c", "a.c"]).contains("separate assembler"));
4742 assert!(refused(&["-specs=/x", "a.c"]).contains("-specs= is not supported"));
4743 assert!(refused(&["-mcmodel=kernel", "-c", "a.c"]).contains("small code model"));
4744 assert!(refused(&["-gdwarf-4", "-c", "a.c"]).contains("DWARF 5"));
4745 assert!(refused(&["-Ofast", "-c", "a.c"]).contains("fast math"));
4746 // The word size the target does not have, which is a target this compiler was not asked
4747 // for rather than a flag it does not know.
4748 let no32 = refused(&["--target=x86_64-unknown-linux-gnu", "-m32", "-c", "a.c"]);
4749 assert!(no32.contains("32 bit target"), "{no32}");
4750 }
4751
4752 /// `-gz` and the two spellings of the split, which are the two questions about the shape of
4753 /// the debug output rather than about how much of it there is.
4754 ///
4755 /// Both answers here are about what happens when there is debug information to shape, and
4756 /// there is none yet, so what is being asserted is that the flags are read and remembered
4757 /// rather than that anything changed in the output. That is the whole of what taking them
4758 /// claims, and it is worth a test because the day `rucc-debug` writes a section this is where
4759 /// it comes to find out what the command line said.
4760 #[test]
4761 fn the_shape_of_the_debug_output_is_recorded_even_where_there_is_none_of_it() {
4762 let (opts, _) = compile(&["-c", "a.c"]);
4763 assert_eq!(opts.compress, Compress::None, "uncompressed unless somebody asks");
4764
4765 // Bare `-gz` is `-gz=zlib`, measured against gcc 16 rather than read out of the manual,
4766 // which describes the flag without ever saying which algorithm it picks.
4767 assert_eq!(compile(&["-gz", "-c", "a.c"]).0.compress, Compress::Zlib);
4768 for (spelling, want) in [
4769 ("none", Compress::None),
4770 ("zlib", Compress::Zlib),
4771 ("zlib-gnu", Compress::ZlibGnu),
4772 ("zstd", Compress::Zstd),
4773 ] {
4774 let (opts, _) = compile(&[&format!("-gz={spelling}"), "-c", "a.c"]);
4775 assert_eq!(opts.compress, want, "{spelling}");
4776 }
4777
4778 // A value nothing here has heard of is refused rather than rounded to the nearest one,
4779 // because a build that asked for `zstd` and quietly got `zlib` would ship a file its
4780 // reader may not understand and would have no way of finding out.
4781 for bad in ["-gz=gzip", "-gz="] {
4782 let failed = refused(&[bad, "-c", "a.c"]);
4783 assert!(failed.contains("is not a way to compress"), "{bad}: {failed}");
4784 }
4785
4786 // The split is refused in the direction that would have written a file and taken in the
4787 // direction that describes what happens. A build system that names the `.dwo` as an
4788 // output has to hear about it now rather than at the point the file is missing.
4789 let (opts, _) = compile(&["-gno-split-dwarf", "-g", "-c", "a.c"]);
4790 assert!(opts.debug_info, "the negative spelling says nothing about how much");
4791 let failed = refused(&["-gsplit-dwarf", "-c", "a.c"]);
4792 assert!(failed.contains(".dwo"), "the refusal names the file it would have written");
4793 }
4794
4795 /// The `-flto` family, which is the whole of an optimization this compiler does not do.
4796 ///
4797 /// Taken rather than refused because ignoring it gives a correct program that is slower than
4798 /// it could have been, which is section 4.1's hint about speed. The values are still held to
4799 /// gcc's, so a command line written for clang is told rather than quietly taken.
4800 #[test]
4801 fn the_link_time_family_is_read_and_checked_and_nothing_is_done_about_it() {
4802 let (opts, _) = compile(&["-c", "a.c"]);
4803 assert!(!opts.lto.requested, "nothing asks unless the command line does");
4804
4805 let (opts, _) = compile(&["-flto", "-c", "a.c"]);
4806 assert!(opts.lto.requested);
4807 assert_eq!(opts.lto.jobs, LtoJobs::One, "bare -flto is one process, the way gcc reads it");
4808
4809 // The last of the two directions wins, the same as every other pair of `-f` spellings.
4810 assert!(!compile(&["-flto", "-fno-lto", "-c", "a.c"]).0.lto.requested);
4811 assert!(compile(&["-fno-lto", "-flto", "-c", "a.c"]).0.lto.requested);
4812
4813 // A count is a count, and asking for one implies asking for the optimization.
4814 for (spelling, want) in [
4815 ("auto", LtoJobs::Auto),
4816 ("jobserver", LtoJobs::Jobserver),
4817 ("1", LtoJobs::One),
4818 ("8", LtoJobs::Count(8)),
4819 ] {
4820 let (opts, _) = compile(&[&format!("-flto={spelling}"), "-c", "a.c"]);
4821 assert_eq!(opts.lto.jobs, want, "{spelling}");
4822 assert!(opts.lto.requested, "{spelling} asks for it too");
4823 }
4824
4825 // gcc refuses a zero rather than reading it as `-fno-lto`, and `thin` is clang's spelling
4826 // of a question gcc answers with `-flto-partition=`, so somebody who wrote it meant a
4827 // different compiler and gets told so here rather than getting a serial link.
4828 for bad in ["-flto=0", "-flto=thin", "-flto=full", "-flto=-1"] {
4829 let failed = refused(&[bad, "-c", "a.c"]);
4830 assert!(failed.contains("link time jobs"), "{bad}: {failed}");
4831 }
4832
4833 // How the program is cut up before the work is spread over it.
4834 assert_eq!(compile(&["-c", "a.c"]).0.lto.partition, Partition::Balanced, "gcc's default");
4835 for (spelling, want) in [
4836 ("balanced", Partition::Balanced),
4837 ("1to1", Partition::OneToOne),
4838 ("one", Partition::One),
4839 ("max", Partition::Max),
4840 ("none", Partition::None),
4841 ] {
4842 let (opts, _) = compile(&[&format!("-flto-partition={spelling}"), "-c", "a.c"]);
4843 assert_eq!(opts.lto.partition, want, "{spelling}");
4844 }
4845 assert!(refused(&["-flto-partition=big", "-c", "a.c"]).contains("partitioning model"));
4846
4847 // And how hard the bytecode is compressed on its way into the object, which is zstd's
4848 // range of levels and is the range gcc checks an argument against.
4849 assert_eq!(compile(&["-c", "a.c"]).0.lto.compression, None, "whatever it does by default");
4850 assert_eq!(compile(&["-flto-compression-level=0", "-c", "a.c"]).0.lto.compression, Some(0));
4851 let (opts, _) = compile(&["-flto-compression-level=19", "-c", "a.c"]);
4852 assert_eq!(opts.lto.compression, Some(19));
4853 for bad in ["-flto-compression-level=20", "-flto-compression-level=-1"] {
4854 let failed = refused(&[bad, "-c", "a.c"]);
4855 assert!(failed.contains("compression level"), "{bad}: {failed}");
4856 }
4857
4858 // The two pairs that describe an arrangement rather than ask for one. Every object here
4859 // holds its machine code, so the fat spelling is what already happens and the other is a
4860 // smaller file rather than a different program, and the plugin pair is about a tool the
4861 // design in `spec/09-optimizer.md` never loads.
4862 for taken in [
4863 "-ffat-lto-objects",
4864 "-fno-fat-lto-objects",
4865 "-fuse-linker-plugin",
4866 "-fno-use-linker-plugin",
4867 ] {
4868 let (opts, _) = compile(&[taken, "-c", "a.c"]);
4869 assert!(!opts.lto.requested, "{taken} says nothing about whether to do it");
4870 }
4871 }
4872
4873 /// The profile family, which is the only one here that splits down the middle.
4874 ///
4875 /// Reading a profile is taken and writing one is refused, and the line between them is the one
4876 /// section 4.1 draws: ignoring a request to read the counts gives a correct program that is
4877 /// slower than it could have been, and ignoring a request to write them means a file the build
4878 /// declared as an output never appears.
4879 #[test]
4880 fn reading_a_profile_is_taken_and_writing_one_is_refused() {
4881 let (opts, _) = compile(&["-c", "a.c"]);
4882 assert!(!opts.profile_data.requested, "nothing asks unless the command line does");
4883 assert_eq!(opts.profile_data.path, None);
4884
4885 let (opts, _) = compile(&["-fprofile-use", "-c", "a.c"]);
4886 assert!(opts.profile_data.requested);
4887 assert_eq!(opts.profile_data.path, None, "beside the object, the way gcc looks");
4888
4889 let (opts, _) = compile(&["-fprofile-use=/counts", "-c", "a.c"]);
4890 assert!(opts.profile_data.requested, "naming a path asks for it too");
4891 assert_eq!(opts.profile_data.path.as_deref(), Some("/counts"));
4892
4893 // The last of the two directions wins, the same as every other pair of `-f` spellings.
4894 assert!(
4895 !compile(&["-fprofile-use", "-fno-profile-use", "-c", "a.c"]).0.profile_data.requested
4896 );
4897 assert!(
4898 compile(&["-fno-profile-use", "-fprofile-use", "-c", "a.c"]).0.profile_data.requested
4899 );
4900
4901 // The rest of the reading half, which is where the files are and three answers about what
4902 // to make of what is in them.
4903 let (opts, _) = compile(&[
4904 "-fprofile-dir=/build/profiles",
4905 "-fprofile-abs-path",
4906 "-fprofile-correction",
4907 "-fprofile-partial-training",
4908 "-c",
4909 "a.c",
4910 ]);
4911 assert_eq!(opts.profile_data.dir.as_deref(), Some("/build/profiles"));
4912 assert!(opts.profile_data.absolute);
4913 assert!(opts.profile_data.correction);
4914 assert!(opts.profile_data.partial_training);
4915
4916 // Writing one, which is refused by name. The first four instrument the program and the
4917 // last writes a file beside the object, and a build that got neither and no message would
4918 // go on to optimize against counts that were never gathered.
4919 for writing in [
4920 "-fprofile-generate",
4921 "-fprofile-generate=/build/profiles",
4922 "-fprofile-arcs",
4923 "--coverage",
4924 "-fcondition-coverage",
4925 "-fpath-coverage",
4926 ] {
4927 let failed = refused(&[writing, "-c", "a.c"]);
4928 assert!(failed.contains("instrument"), "{writing}: {failed}");
4929 }
4930 assert!(refused(&["-ftest-coverage", "-c", "a.c"]).contains(".gcno"), "it names the file");
4931
4932 // The negative spellings of the refused half are what already happens, so they are taken.
4933 for taken in ["-fno-profile-generate", "-fno-profile-arcs", "-fno-test-coverage"] {
4934 let (opts, _) = compile(&[taken, "-c", "a.c"]);
4935 assert!(!opts.profile_data.requested, "{taken} asks for nothing");
4936 }
4937
4938 // And the flags that describe the instrumentation that is refused above, which are checked
4939 // and dropped. Checked because a typo is worth finding here rather than on the day the
4940 // instrumentation lands.
4941 for taken in [
4942 "-fprofile-update=single",
4943 "-fprofile-update=atomic",
4944 "-fprofile-update=prefer-atomic",
4945 "-fprofile-reproducible=serial",
4946 "-fprofile-reproducible=parallel-runs",
4947 "-fprofile-reproducible=multithreaded",
4948 "-fprofile-values",
4949 "-fno-profile-values",
4950 "-fprofile-info-section",
4951 "-fprofile-filter-files=a.c",
4952 "-fprofile-exclude-files=b.c",
4953 "-fprofile-note=a.gcno",
4954 ] {
4955 let (opts, _) = compile(&[taken, "-c", "a.c"]);
4956 assert!(!opts.profile_data.requested, "{taken} says nothing about reading one");
4957 }
4958 assert!(refused(&["-fprofile-update=none", "-c", "a.c"]).contains("update method"));
4959 assert!(refused(&["-fprofile-reproducible=any", "-c", "a.c"]).contains("reproducibility"));
4960 }
4961
4962 /// The sanitizers, which are refused by name and are the one family refused for a reason that
4963 /// is not about the bytes.
4964 ///
4965 /// A sanitizer is a promise that the program is watched while it runs, so a build that asked
4966 /// for one and was quietly given a program with no checks in it gets a test suite that passes
4967 /// for the wrong reason rather than a slower program.
4968 #[test]
4969 fn a_sanitizer_that_is_still_asked_for_at_the_end_of_the_line_is_refused_by_name() {
4970 for asked in ["address", "undefined", "thread", "kernel-address", "leak", "memory"] {
4971 let failed = refused(&[&format!("-fsanitize={asked}"), "-c", "a.c"]);
4972 assert!(failed.contains(asked), "the refusal names what was asked for: {failed}");
4973 assert!(failed.contains("-fsafety=detect"), "and the nearest thing: {failed}");
4974 }
4975
4976 // A list is every name in it, and the first one still standing is the one named.
4977 let failed = refused(&["-fsanitize=address,undefined", "-c", "a.c"]);
4978 assert!(failed.contains("address"), "{failed}");
4979
4980 // A name that is not one, which is worth its own message: somebody who wrote `-fsanitize`
4981 // with a typo in it has a different problem from somebody who wrote a real one.
4982 for bad in ["-fsanitize=bogus", "-fsanitize=address,bogus", "-fno-sanitize=bogus"] {
4983 let failed = refused(&[bad, "-c", "a.c"]);
4984 assert!(failed.contains("is not a sanitizer"), "{bad}: {failed}");
4985 }
4986
4987 // gcc takes `all` only in the negative, and so does this.
4988 assert!(refused(&["-fsanitize=all", "-c", "a.c"]).contains("only `-fno-sanitize=all`"));
4989
4990 // Asking and then taking it back is asking for nothing, which is why the answer waits for
4991 // the end of the line. A build whose shared flags turn a check on and whose rule for one
4992 // file turns it off again compiles that file here.
4993 for pair in [
4994 ["-fsanitize=address", "-fno-sanitize=address"],
4995 ["-fsanitize=address,undefined", "-fno-sanitize=all"],
4996 ["-fsanitize=undefined", "-fno-sanitize=undefined"],
4997 ] {
4998 let (opts, _) = compile(&[pair[0], pair[1], "-c", "a.c"]);
4999 assert_eq!(opts.safety, rucc_session::Safety::Off, "{pair:?} asked for nothing");
5000 }
5001 // And the other order still asks, because the last word is the one that counts.
5002 assert!(!refused(&["-fno-sanitize=address", "-fsanitize=address", "-c", "a.c"]).is_empty());
5003
5004 // What a check does when it fires is an answer about checks that are refused, so there is
5005 // nothing left for it to change and it is taken.
5006 for taken in [
5007 "-fsanitize-recover=undefined",
5008 "-fno-sanitize-recover=all",
5009 "-fsanitize-trap=undefined",
5010 "-fno-sanitize-trap=all",
5011 "-fsanitize-undefined-trap-on-error",
5012 "-fsanitize-address-use-after-scope",
5013 "-fno-sanitize-address-use-after-scope",
5014 "-fsanitize-sections=.data",
5015 ] {
5016 let (opts, _) = compile(&[taken, "-c", "a.c"]);
5017 assert_eq!(opts.safety, rucc_session::Safety::Off, "{taken} asks for no checking");
5018 }
5019 assert!(refused(&["-fsanitize-recover=bogus", "-c", "a.c"]).contains("is not a sanitizer"));
5020
5021 // Coverage instrumentation is refused rather than dropped, because a fuzzer with no
5022 // feedback runs blind and never says so.
5023 let failed = refused(&["-fsanitize-coverage=trace-pc", "-c", "a.c"]);
5024 assert!(failed.contains("feedback"), "{failed}");
5025 let failed = refused(&["-fsanitize-coverage=trace-pc-guard", "-c", "a.c"]);
5026 assert!(failed.contains("trace-pc or trace-cmp"), "gcc takes two of them: {failed}");
5027 }
5028
5029 #[test]
5030 fn the_levels_gcc_spells_differently_are_the_levels_they_mean() {
5031 assert_eq!(compile(&["-O", "-c", "a.c"]).0.opt_level, OptLevel::O1);
5032 assert_eq!(compile(&["-Og", "-c", "a.c"]).0.opt_level, OptLevel::O1);
5033 assert_eq!(compile(&["-O2", "-c", "a.c"]).0.opt_level, OptLevel::O2);
5034 }
5035
5036 #[test]
5037 fn the_machine_flags_that_name_what_we_already_do_are_taken_and_the_rest_are_not() {
5038 let line = ["--target=x86_64-unknown-linux-gnu", "-m64", "-march=x86-64-v3"];
5039 let (opts, _) =
5040 compile(&[&line[..], &["-mtune=native", "-mabi=sysv", "-c", "a.c"]].concat());
5041 assert_eq!(opts.target.to_string(), "x86_64-unknown-linux-gnu");
5042 let wrong = refused(&["--target=x86_64-unknown-linux-gnu", "-mabi=ms", "-c", "a.c"]);
5043 assert!(wrong.contains("sysv convention"), "{wrong}");
5044 }
5045
5046 #[test]
5047 fn the_thread_flag_is_a_macro_and_a_library_and_the_library_goes_last() {
5048 let (opts, plan) = compile(&["-pthread", "-c", "a.c"]);
5049 assert!(opts.defines.iter().any(|d| d == "_REENTRANT"));
5050 // After the input, because a static link takes what it needs from a library when it
5051 // reaches it and not afterwards.
5052 let names: Vec<&str> = plan.jobs.iter().map(|j| j.input.as_str()).collect();
5053 assert_eq!(names, vec!["a.c"]);
5054 }
5055
5056 #[test]
5057 fn the_questions_a_build_system_asks_before_it_compiles_anything() {
5058 let target = "--target=x86_64-unknown-linux-gnu";
5059 assert_eq!(printed(&[target, "-dumpmachine"]), "x86_64-unknown-linux-gnu");
5060 assert_eq!(printed(&[target, "-dumpversion"]), VERSION);
5061 assert_eq!(printed(&[target, "-dumpfullversion"]), VERSION);
5062 assert_eq!(printed(&[target, "-print-multiarch"]), "x86_64-linux-gnu");
5063 // A name nothing holds comes back unchanged, which is GCC's rule and is what makes the
5064 // answer safe to paste into a link line whether or not the file is there.
5065 assert_eq!(printed(&[target, "-print-file-name=no-such-library.a"]), "no-such-library.a");
5066 assert_eq!(printed(&[target, "-print-prog-name=ld"]), "ld");
5067 let dirs = printed(&[target, "-print-search-dirs"]);
5068 assert!(dirs.starts_with("install: "), "{dirs}");
5069 assert!(dirs.contains("\nlibraries: ="), "{dirs}");
5070 }
5071
5072 #[test]
5073 fn the_sysroot_in_effect_is_the_one_the_command_line_named_or_the_one_for_the_target() {
5074 // A tree the user named is the answer whatever the target is, because it is the answer to
5075 // every other question too.
5076 assert_eq!(printed(&["--sysroot=/opt/cross", "-print-sysroot"]), "/opt/cross");
5077
5078 // A target that is no machine this suite runs on is read under the cache, and the answer is
5079 // the root rather than one of the directories under it, since what asks is looking for a
5080 // file of its own.
5081 let root = cache::dir().join("sysroots").join("riscv64-linux-musl");
5082 assert_eq!(
5083 printed(&["--target=riscv64-linux-musl", "-print-sysroot"]),
5084 root.display().to_string()
5085 );
5086
5087 // And a compile for this machine has no sysroot, which is the empty line GCC prints when it
5088 // was configured without one rather than a `/` that would be a claim about the filesystem.
5089 let host = Triple::host().expect("a host this compiler knows");
5090 assert_eq!(printed(&[&format!("--target={host}"), "-print-sysroot"]), "");
5091 }
5092
5093 #[test]
5094 fn the_provenance_of_a_sysroot_is_the_manifest_it_carries() {
5095 // Section 13.5 wants seven things per input and wants them machine readable, and the manifest
5096 // is the record that already has them, so the flag prints that rather than a second format.
5097 let manifest = "rucc sysroot manifest 3\n\
5098 target\tx86_64-linux-musl\n\
5099 kernel\t6.12\n\
5100 include/generic/stdio.h\tmusl-1.2.5\t\
5101 https://musl.libc.org/releases/musl-1.2.5.tar.gz\t\
5102 0000000000000000000000000000000000000000000000000000000000000000\tmit\t\
5103 bundled\n\
5104 lib/libc.so\tmusl-1.2.5\t\
5105 https://musl.libc.org/releases/musl-1.2.5.tar.gz\t\
5106 1111111111111111111111111111111111111111111111111111111111111111\tmit\t\
5107 generated\n";
5108 let tree = TempTree::new("provenance", &[("manifest", manifest)]);
5109 let sysroot = format!("--sysroot={}", tree.0.display());
5110 // The kernel line of tamnd/rucc#934 is in the answer without anything here naming it, because
5111 // the flag parses the record and renders it again rather than picking fields out of it. That
5112 // is the reason it prints a manifest and not a format of its own.
5113 //
5114 // The answer is the file without its last newline, because whatever prints it adds one. The
5115 // file is what somebody diffs the output against, so the two have to be the same bytes.
5116 assert_eq!(printed(&[&sysroot, "-print-sysroot-provenance"]) + "\n", manifest);
5117
5118 // A tree with no manifest in it is a tree somebody assembled themselves, and nothing here
5119 // knows where any of it came from. Saying nothing is the only honest answer, and a reader can
5120 // tell it from a manifest with no inputs because that one still has its two header lines.
5121 let bare = TempTree::new("provenance-bare", &[]);
5122 assert_eq!(
5123 printed(&[&format!("--sysroot={}", bare.0.display()), "-print-sysroot-provenance"]),
5124 ""
5125 );
5126
5127 // And a compile for this machine has no sysroot at all, which is the same empty answer
5128 // `-print-sysroot` gives for it.
5129 let host = Triple::host().expect("a host this compiler knows");
5130 assert_eq!(printed(&[&format!("--target={host}"), "-print-sysroot-provenance"]), "");
5131
5132 // And the other spelling, which section 13.5 is the document that writes.
5133 assert_eq!(printed(&[&sysroot, "--print-sysroot-provenance"]) + "\n", manifest);
5134
5135 // tamnd/rucc#1021. The digest of the same tree is the sha256 of that record, so it is one
5136 // line where the provenance is a few hundred, and it is checkable with `sha256sum` because
5137 // the bytes it is over are the bytes of the file. The number here is that hash of the
5138 // fixture above, computed by `sha256sum` rather than by this compiler.
5139 assert_eq!(
5140 printed(&[&sysroot, "-print-sysroot-digest"]),
5141 "d705ae6ebeafeb7fda4bd57cecc7882bf49784b17015664a09cfae25a1b2000a"
5142 );
5143 assert_eq!(
5144 printed(&[&sysroot, "--print-sysroot-digest"]),
5145 printed(&[&sysroot, "-print-sysroot-digest"])
5146 );
5147
5148 // And the two empty answers are empty here too, because a digest of nothing would read as a
5149 // claim about a sysroot rather than as the absence of one.
5150 assert_eq!(
5151 printed(&[&format!("--sysroot={}", bare.0.display()), "-print-sysroot-digest"]),
5152 ""
5153 );
5154 assert_eq!(printed(&[&format!("--target={host}"), "-print-sysroot-digest"]), "");
5155 }
5156
5157 #[test]
5158 fn a_manifest_this_build_cannot_read_is_refused_rather_than_printed() {
5159 // Passing a file we could not parse to whoever asked would make their parser the one that
5160 // finds the problem, and the three uses section 13.5 gives for this are all somebody else
5161 // parsing it.
5162 let tree = TempTree::new(
5163 "provenance-bad",
5164 &[("manifest", "rucc sysroot manifest 3\ntarget\tx86_64-linux-musl\nlib/libc.a\n")],
5165 );
5166 let message =
5167 refused(&[&format!("--sysroot={}", tree.0.display()), "-print-sysroot-provenance"]);
5168 assert!(message.contains("manifest"), "{message}");
5169 assert!(message.contains("1 fields where an input has six"), "{message}");
5170
5171 // The digest is refused for the same file and for a stronger reason: a hash of bytes this
5172 // build cannot read would be a number that names a record nobody can act on.
5173 let digest =
5174 refused(&[&format!("--sysroot={}", tree.0.display()), "-print-sysroot-digest"]);
5175 assert_eq!(digest, message);
5176 }
5177
5178 #[test]
5179 fn the_two_dependency_flags_that_stop_after_the_rule_stop_after_the_rule() {
5180 let (opts, _) = compile(&["-M", "a.c"]);
5181 assert!(opts.deps.emit && opts.deps.instead_of_compiling);
5182 assert!(opts.deps.system_headers, "plain -M lists them");
5183 assert_eq!(opts.emit, EmitKind::Preprocessed);
5184
5185 // Even where a later flag asked for something else, because the family is a mode and
5186 // the mode is what the run is for.
5187 let (opts, _) = compile(&["-M", "-c", "a.c"]);
5188 assert_eq!(opts.emit, EmitKind::Preprocessed);
5189
5190 let (opts, _) = compile(&["-MM", "a.c"]);
5191 assert!(!opts.deps.system_headers);
5192 }
5193
5194 #[test]
5195 fn the_two_that_end_in_d_leave_the_compilation_alone() {
5196 let (opts, _) = compile(&["-MD", "-c", "a.c"]);
5197 assert!(opts.deps.emit && !opts.deps.instead_of_compiling);
5198 assert!(opts.deps.system_headers);
5199 assert_eq!(opts.emit, EmitKind::Object);
5200
5201 let (opts, _) = compile(&["-MMD", "-c", "a.c"]);
5202 assert!(opts.deps.emit && !opts.deps.instead_of_compiling);
5203 assert!(!opts.deps.system_headers);
5204 }
5205
5206 #[test]
5207 fn nothing_puts_the_system_headers_back_once_a_flag_has_taken_them_out() {
5208 // GCC's rule, and not an oversight in it. The flag asking for fewer of them is read as
5209 // the answer, because the other one never asked the question.
5210 let (opts, _) = compile(&["-MM", "-M", "a.c"]);
5211 assert!(!opts.deps.system_headers);
5212 let (opts, _) = compile(&["-MD", "-MMD", "-c", "a.c"]);
5213 assert!(!opts.deps.system_headers);
5214 let (opts, _) = compile(&["-MMD", "-MD", "-c", "a.c"]);
5215 assert!(!opts.deps.system_headers);
5216 }
5217
5218 #[test]
5219 fn a_target_arrives_escaped_from_one_flag_and_untouched_from_the_other() {
5220 let (opts, _) = compile(&["-MM", "-MT", "a b.o", "-MQ", "a b.o", "a.c"]);
5221 assert_eq!(opts.deps.targets, vec!["a b.o".to_owned(), "a\\ b.o".to_owned()]);
5222 }
5223
5224 #[test]
5225 fn the_rest_of_the_family_is_a_file_and_a_switch() {
5226 let (opts, _) = compile(&["-MM", "-MF", "dep.d", "-MP", "a.c"]);
5227 assert_eq!(opts.deps.file.as_deref(), Some("dep.d"));
5228 assert!(opts.deps.phony);
5229
5230 for flag in ["-MF", "-MT", "-MQ"] {
5231 let e = parse_args(&args(&[flag])).unwrap_err();
5232 assert!(e.message.contains("requires an argument"), "{}", e.message);
5233 }
5234 }
5235
5236 /// A directory of sources for one test, removed when the test is done with it.
5237 struct TempTree(PathBuf);
5238
5239 impl Drop for TempTree {
5240 fn drop(&mut self) {
5241 let _ = std::fs::remove_dir_all(&self.0);
5242 }
5243 }
5244
5245 impl TempTree {
5246 fn new(name: &str, files: &[(&str, &str)]) -> TempTree {
5247 let dir = std::env::temp_dir().join(format!("rucc-deps-{}-{name}", std::process::id()));
5248 let _ = std::fs::remove_dir_all(&dir);
5249 std::fs::create_dir_all(&dir).expect("temporary directory should be writable");
5250 for (path, text) in files {
5251 let at = dir.join(path);
5252 if let Some(parent) = at.parent() {
5253 std::fs::create_dir_all(parent).expect("creating a subdirectory should work");
5254 }
5255 std::fs::write(&at, text).expect("writing a temporary file should work");
5256 }
5257 TempTree(dir)
5258 }
5259
5260 fn path(&self, name: &str) -> String {
5261 self.0.join(name).to_string_lossy().into_owned()
5262 }
5263 }
5264
5265 #[test]
5266 fn the_rule_names_what_the_includes_found_and_names_each_of_them_once() {
5267 // End to end, because the list comes from the preprocessor and the format comes from
5268 // somewhere else, and a test of either half on its own would pass with the two of them
5269 // wired up backwards.
5270 let tree = TempTree::new(
5271 "found",
5272 &[
5273 ("a.c", "#include \"one.h\"\n#include \"two.h\"\nint main(void) { return X; }\n"),
5274 ("one.h", "#define X 0\n"),
5275 ("two.h", "#include \"one.h\"\n"),
5276 ],
5277 );
5278 let out = tree.path("dep.d");
5279 let code = run(&args(&["-MM", "-MF", &out, "-o", &tree.path("a.i"), &tree.path("a.c")]));
5280 assert_eq!(code, 0);
5281
5282 let text = std::fs::read_to_string(&out).expect("the rule should have been written");
5283 let names: Vec<&str> = text.split_whitespace().collect();
5284 // The target, the source, and each header once however many times it was reached.
5285 assert_eq!(names.first(), Some(&"a.o:"), "{text}");
5286 assert_eq!(names.iter().filter(|n| n.ends_with("one.h")).count(), 1, "{text}");
5287 assert_eq!(names.iter().filter(|n| n.ends_with("two.h")).count(), 1, "{text}");
5288 // And the `-o` went to the file the rule replaced, which is left empty rather than
5289 // absent because a makefile that named it as a target will look for it.
5290 assert_eq!(std::fs::read(tree.path("a.i")).expect("the output should exist"), b"");
5291 }
5292
5293 #[test]
5294 fn a_header_that_is_only_reached_under_a_guard_is_still_a_dependency() {
5295 // The multiple-include optimization means the second reach never opens the file. It is
5296 // still a file this translation unit was built from, so it is still in the rule.
5297 let tree = TempTree::new(
5298 "guarded",
5299 &[
5300 ("a.c", "#include \"g.h\"\n#include \"g.h\"\nint main(void) { return 0; }\n"),
5301 ("g.h", "#ifndef G\n#define G\n#endif\n"),
5302 ],
5303 );
5304 let out = tree.path("dep.d");
5305 let code = run(&args(&["-MM", "-MF", &out, "-o", &tree.path("a.i"), &tree.path("a.c")]));
5306 assert_eq!(code, 0);
5307 let text = std::fs::read_to_string(&out).expect("the rule should have been written");
5308 assert_eq!(text.split_whitespace().filter(|n| n.ends_with("g.h")).count(), 1, "{text}");
5309 }
5310
5311 #[test]
5312 fn every_imacros_file_is_read_before_every_include_file_whatever_order_they_were_written() {
5313 // Measured against GCC rather than read: the two flags the other way round produce the
5314 // same output byte for byte, so the command line order between the two families does not
5315 // decide anything and the order within one does. The `-include` file here can only see
5316 // the definition if the `-imacros` file that was written after it ran first.
5317 let tree = TempTree::new(
5318 "preinclude",
5319 &[
5320 ("a.c", "int main(void) { return 0; }\n"),
5321 ("i.h", "#ifdef FROM_MACROS\nint saw_it;\n#else\nint missed_it;\n#endif\n"),
5322 ("m.h", "#define FROM_MACROS 1\nint macros_text;\n"),
5323 ],
5324 );
5325 let out = tree.path("a.i");
5326 let code = run(&args(&[
5327 "-E",
5328 "-include",
5329 &tree.path("i.h"),
5330 "-imacros",
5331 &tree.path("m.h"),
5332 "-o",
5333 &out,
5334 &tree.path("a.c"),
5335 ]));
5336 assert_eq!(code, 0);
5337 let text = std::fs::read_to_string(&out).expect("the output should have been written");
5338 assert!(text.contains("saw_it"), "{text}");
5339 // And the text of the `-imacros` file is thrown away, which is the whole difference
5340 // between the two flags.
5341 assert!(!text.contains("macros_text"), "{text}");
5342 }
5343
5344 #[test]
5345 fn a_file_the_command_line_named_is_a_prerequisite_the_same_as_one_a_directive_named() {
5346 let tree = TempTree::new(
5347 "preinclude-deps",
5348 &[
5349 ("a.c", "int main(void) { return 0; }\n"),
5350 ("i.h", "int from_include;\n"),
5351 ("m.h", "#define M 1\n"),
5352 ],
5353 );
5354 let out = tree.path("dep.d");
5355 let code = run(&args(&[
5356 "-MM",
5357 "-MF",
5358 &out,
5359 "-include",
5360 &tree.path("i.h"),
5361 "-imacros",
5362 &tree.path("m.h"),
5363 "-o",
5364 &tree.path("a.i"),
5365 &tree.path("a.c"),
5366 ]));
5367 assert_eq!(code, 0);
5368 let text = std::fs::read_to_string(&out).expect("the rule should have been written");
5369 assert!(text.contains("i.h"), "{text}");
5370 assert!(text.contains("m.h"), "{text}");
5371 }
5372
5373 #[test]
5374 fn a_command_line_include_that_is_nowhere_on_the_path_is_an_error_and_not_a_warning() {
5375 // Including the directory of the source file, which is not on the path for these: the
5376 // command line was not written there, so a name in it is relative to where the compiler
5377 // was run rather than to where the source sits.
5378 let tree = TempTree::new(
5379 "preinclude-missing",
5380 &[("sub/a.c", "int main(void) { return 0; }\n"), ("sub/beside.h", "int x;\n")],
5381 );
5382 let code = run(&args(&["-E", "-include", "beside.h", "-o", "-", &tree.path("sub/a.c")]));
5383 assert_eq!(code, 1);
5384 }
5385
5386 #[test]
5387 fn a_command_line_that_links_names_the_executable_and_not_the_object_it_went_through() {
5388 // The object a link goes through is in a temporary directory and is gone before `make`
5389 // reads any of this, so the rule that named it would be a rule for a file that is never
5390 // there. The target and the file are both the `-o`, which is the executable.
5391 let (opts, plan) = compile(&["-MD", "sub/a.c", "-o", "prog"]);
5392 assert_eq!(plan.output.as_deref(), Some("prog"));
5393 assert_eq!(deps::default_target("sub/a.c", deps_target_output(&opts, &plan)), "prog");
5394 assert_eq!(
5395 deps::default_file(&opts.deps, "sub/a.c", plan.output.as_deref()).as_deref(),
5396 Some("prog.d")
5397 );
5398 }
5399
5400 #[test]
5401 fn the_plan_keeps_the_output_name_because_the_rule_is_written_from_it() {
5402 let (_, plan) = compile(&["-MMD", "-c", "sub/a.c", "-o", "obj/x.o"]);
5403 assert_eq!(plan.output.as_deref(), Some("obj/x.o"));
5404 let (_, plan) = compile(&["-MMD", "-c", "sub/a.c"]);
5405 assert_eq!(plan.output, None);
5406 }
5407
5408 #[test]
5409 fn usage_fits_on_a_screen() {
5410 // Not a style preference. A help text that scrolls is one nobody reads, and this is
5411 // the cheapest way to keep it honest as flags accumulate. The number goes up only when
5412 // a family of flags arrives that has nowhere to share a line, which the two pass gates
5413 // were and which the two fuel flags and `-fsafety=` now are, and it goes up by exactly
5414 // the lines that family took. The four it went up by last are the flags a build system
5415 // passes without being asked to: how much to say, what machine to generate for, threads,
5416 // and the questions `configure` asks before it compiles anything. The one it went up by
5417 // last is the second line of `--emit`, whose kinds are a family that has now outgrown
5418 // one line and has nowhere else to go. The two it went up by last are the dependency
5419 // family, which is eight flags that share nothing with anything above them. The one it
5420 // went up by last is the four spellings of position independent code, which every
5421 // configure script writes and which could only have shared the link line, and that line
5422 // is already four characters short of the limit. The two it went up by last are the rest
5423 // of the include family, which is six more flags that change where a header is looked for
5424 // and two that name a header outright. The one it went up by last is the pair that keeps
5425 // the intermediate files and times the steps, which belong next to the two flags above
5426 // them that are also about watching a compilation rather than changing one. The two it
5427 // went up by last are the section flags and the visibility flag, which are what a build
5428 // that cares about the size of what it ships and about which names it exports writes, and
5429 // the second of them was already taken and only missing from here. The one it went up by
5430 // last is the stack protector, which is four spellings of one question and which every
5431 // distribution puts on every command line it issues, so a build that reads this list
5432 // looking for it and does not find it has to go and read the specification instead. The one
5433 // it went up by last is the profiler, which is two spellings of the request and two of
5434 // where the call goes, and which is about watching a program run rather than about what is
5435 // generated, so it shares its subject with nothing above it. The one it went up by last is
5436 // the room a function opens with for something to be written over it later, which takes an
5437 // argument of its own shape and is what a kernel build asks for, so it fits beside the
5438 // profiler and nothing else. The one it went up by last is what overflows rather than being
5439 // undefined, which is three spellings of two questions and which a kernel build and a great
5440 // deal of code written before the standard settled both pass. The one it went up by last is
5441 // the other answer to the first of those questions, which could not share the line because
5442 // what it asks for is the opposite of what the flags on that line ask for. The one it went
5443 // up by last is the split of the line that lists what this compiler does anyway into that
5444 // and what it assumes anyway, which are two different claims that were sharing a line until
5445 // the second of them got a second flag and the line stopped fitting. The one it went up by
5446 // last is the three flags that change the ABI rather than the code, which have to be given
5447 // to every file in a program or none of them and which therefore belong somewhere a person
5448 // reading this list will see them. The one it went up by last is the floating point group,
5449 // which is two lines rather than one because the first of them is a choice this compiler
5450 // records and the rest are claims about what it does anyway, and putting a real setting on
5451 // the same line as three flags that change nothing would be misleading about both. The one
5452 // it went up by last is the flag that says a write has to stay inside the member it names,
5453 // which is a setting rather than a claim and so cannot share the line above it, that being
5454 // the one that picks a tier. The two it went up by last are the prefix mapping family,
5455 // which is four flags whose whole job is to keep a build's output the same from two
5456 // different directories, and which a person chasing a reproducible build comes here
5457 // looking for by name. The one it went up by last is how the debug sections are compressed
5458 // and whether they go in a file of their own, which are two questions about the shape of
5459 // the debug output, where the line above them is about how much of it there is. The one it
5460 // went up by last is the `restrict` contract, which is a setting for the same reason the
5461 // flag that keeps a write inside its member is and which is the check a person who has been
5462 // bitten by a vectorizer comes here looking for. The one it went up by last is link time
5463 // optimization, which is a whole optimization rather than a flag and which says so on its
5464 // own line, because a build that passes it and reads this looking for what it got is
5465 // asking a question no other line here answers. The one it went up by last is the sysroot,
5466 // which is the question somebody asks when a cross build read a file nobody expected, and
5467 // which has no room on the line above it because the answers there are a path each and this
5468 // one is the root all of them are under. The one it went up by last is what is inside that
5469 // root and where each of it came from, which is a question about a whole tree rather than
5470 // about a path and which is long enough on its own that it could not have shared a line with
5471 // anything. The one it went up by last is the profile family, which splits down the middle
5472 // where no other family here does, so the line has to name the half that is taken and the
5473 // half that is refused or it would be read as taking both. The one it went up by last is
5474 // the sanitizers, which are what somebody reaching for a checked build writes first and
5475 // which belong beside the tier that is the nearest thing here to what they asked for. The
5476 // one it went up by last is the digest of that record, which is the same tree as one number
5477 // and could not share the line above it because that line prints a few hundred lines and
5478 // this one prints sixty four characters, and a reader who wants the short answer is looking
5479 // for it by name rather than reading the long one. The one it went up by last is the
5480 // sysroot fetch, which is the only command here that gets something from somewhere else and
5481 // is therefore the one a person wants to have read before they run it rather than after.
5482 // And the flag beside it that forbids every download, which earns its line by being what a
5483 // build in a sealed environment passes and by meaning something even though an ordinary
5484 // compile downloads nothing either way.
5485 assert!(USAGE.lines().count() < 72, "usage text has grown past one screen");
5486 }
5487}