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/0.10.65")]
30
31pub mod assemble;
32pub mod cache;
33pub mod compile;
34pub mod deps;
35pub mod fetch;
36mod glibc;
37pub mod install;
38pub mod library;
39pub mod link;
40mod map;
41pub mod phase;
42pub mod preprocess;
43pub mod schedule;
44
45use std::fmt::Write as _;
46use std::io::Write as _;
47use std::path::PathBuf;
48
49use rucc_codegen::coverage::{self, Fired};
50use rucc_codegen::lowering::Lowerings;
51use rucc_codegen::pressure::Pressure;
52use rucc_pp::Dependency;
53use rucc_session::{
54 Compress, Control, Dumps, EmitKind, Hook, Options, Pic, PrefixMap, Preinclude, Protector,
55 SaveTemps, Session, Std, Wrapping, runtime,
56};
57use rucc_sysroot::{Manifest, Sysroot};
58use rucc_target::{ObjectFormat, Triple};
59use rucc_tuple::TargetTuple;
60
61use crate::link::LinkOptions;
62
63pub use crate::assemble::assemble;
64pub use crate::compile::{Artifact, Compiled, Temps, compile, compile_ir};
65pub use crate::phase::{ArchiveJob, Input, InputKind, Job, LinkJob, Output, Phase, Plan, Role};
66pub use crate::preprocess::{OsFileSystem, Preprocessed, preprocess};
67pub use crate::schedule::Jobs;
68
69/// The compiler's version, taken from the workspace manifest.
70pub const VERSION: &str = env!("CARGO_PKG_VERSION");
71
72/// What the command line asked for.
73#[derive(Debug, Clone, PartialEq, Eq)]
74pub enum Action {
75 /// Print usage and exit successfully.
76 Help,
77 /// Print the version and exit successfully.
78 Version,
79 /// Print one line and exit successfully, which is what the `-dump` and `-print` family do.
80 ///
81 /// A build system asks these before it compiles anything, and what it does with the answer
82 /// is paste it into a path or into another command line, so each one is a single line with
83 /// no decoration around it.
84 Print(String),
85 /// Print the resolved configuration and exit successfully.
86 PrintConfig(Box<Options>),
87 /// Print the passes the level will run and exit successfully.
88 PrintPipeline(Box<Options>),
89 /// Print the phase plan and the link line and exit successfully, which is `-###`.
90 PrintPlan {
91 /// The resolved options, which is what says what the link line is for.
92 opts: Box<Options>,
93 /// What to do to each input, and in what order.
94 plan: Box<Plan>,
95 /// What the command line said about linking.
96 link: Box<LinkOptions>,
97 },
98 /// `--fetch <tuple>`, which gets the sysroot this release pins for a target and installs it.
99 ///
100 /// The only action in this compiler that may run another program to move bytes onto the
101 /// machine, which is `spec/cross-compile/13-distribution.md` section 13.8's rule rather than a
102 /// property of how this happens to be written: a compilation has no branch that reaches it.
103 Fetch {
104 /// The artifact, from the table in [`rucc_sysroot::artifact`]. Resolved here rather than where the
105 /// work happens, so that a target nothing is pinned for is a refusal from the parser like
106 /// every other thing a command line can ask for and not have.
107 what: &'static rucc_sysroot::Pinned,
108 /// The target, which names the directory under the cache the tree is installed at and is
109 /// checked against the record inside the artifact.
110 target: TargetTuple,
111 /// Where the cache is, read where everything else that needs it reads it.
112 cache: PathBuf,
113 },
114 /// Compile the given inputs.
115 Compile {
116 /// The resolved options.
117 opts: Box<Options>,
118 /// What to do to each input, and in what order.
119 plan: Box<Plan>,
120 /// What the command line said about linking.
121 link: Box<LinkOptions>,
122 /// How many translation units to compile at once.
123 jobs: Jobs,
124 /// Whether `-v` asked for the plan to be printed while it runs.
125 verbose: bool,
126 /// What is worth saying about the command line before anything is compiled, printed as
127 /// warnings and once for the whole run rather than once per file.
128 ///
129 /// These are not diagnostics. A diagnostic is about a piece of source and has a span to
130 /// point at, and these are about the way two flags were combined, so there is nothing to
131 /// point at and nowhere below the driver that knows both halves. `-w` does not reach them
132 /// for the same reason it does not reach a refusal from the parser.
133 notes: Vec<String>,
134 },
135}
136
137/// Why a command line was rejected.
138#[derive(Debug, Clone, PartialEq, Eq)]
139pub struct CliError {
140 /// The message, lowercase and without a trailing period, in the same shape as any other
141 /// diagnostic.
142 pub message: String,
143}
144
145impl std::fmt::Display for CliError {
146 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
147 f.write_str(&self.message)
148 }
149}
150
151impl std::error::Error for CliError {}
152
153fn err(message: impl Into<String>) -> CliError {
154 CliError { message: message.into() }
155}
156
157/// The two halves of one prefix mapping flag's argument, where `flag` includes its trailing `=`.
158///
159/// The split is at the last `=` in what follows the flag, not the first, which is gcc's rule and
160/// the only one that lets a directory whose name contains an `=` be the old half. It also means
161/// `-fmacro-prefix-map=a=b=c` rewrites `a=b` to `c` rather than `a` to `b=c`, which looks like a
162/// trap until you notice the alternative traps the far more common case.
163fn rewrite<'a>(arg: &'a str, flag: &str) -> Result<(&'a str, &'a str), CliError> {
164 let rest = &arg[flag.len()..];
165 PrefixMap::split(rest).ok_or_else(|| {
166 let flag = flag.trim_end_matches('=');
167 err(format!(
168 "`{rest}` is not a rewrite for `{flag}`, which is an old prefix, an `=` and a new one"
169 ))
170 })
171}
172
173/// A question the command line asked instead of asking for a compilation.
174///
175/// These are answered after the loop rather than where they are read, because every one of them
176/// is about the target or about the library search and the last word on both is the end of the
177/// command line.
178enum Query {
179 /// `-dumpmachine`, the triple.
180 Machine,
181 /// `-dumpversion` and `-dumpfullversion`, which are the same three numbers here.
182 Version,
183 /// `-print-multiarch`, the directory name a distribution files this target under.
184 Multiarch,
185 /// `-print-search-dirs`, in the three lines GCC prints.
186 SearchDirs,
187 /// `-print-sysroot`, the root the headers and the libraries are read under.
188 Sysroot,
189 /// `-print-sysroot-provenance`, what is in that root and where each of it came from.
190 SysrootProvenance,
191 /// `-print-sysroot-digest`, the one number that names all of it.
192 SysrootDigest,
193 /// `-print-file-name=<name>`, the full path of a library file.
194 FileName(String),
195 /// `-print-prog-name=<name>`, the full path of a program.
196 ProgName(String),
197 /// `-print-libgcc-file-name`, which is `-print-file-name=libgcc.a` under another spelling.
198 Libgcc,
199}
200
201/// Usage text.
202///
203/// Deliberately short. `spec/04-driver-and-cli.md` puts the full flag reference in the
204/// manual page, because a `--help` nobody can read in one screen is a `--help` nobody reads.
205pub const USAGE: &str = "\
206rucc, an optimizing C compiler
207
208usage: rucc [options] file...
209
210options:
211 -c compile and assemble, do not link
212 -S compile only, emit assembly
213 -E preprocess only
214 -o <file> write output to <file>, or to standard output for -
215 -D <name>[=<value>], -U <name> define a macro, or undefine one after every -D
216 -I <dir> add <dir> to the include search path
217 -iquote -isystem -idirafter <dir> the other chains, -nostdinc drops ours
218 -I-, -iprefix <p>, -iwithprefix[before] <dir> the older spellings of those
219 -include <file>, -imacros <file> read <file> first, the second for its macros only
220 --sysroot=<dir> look for the library's headers under <dir>, -isysroot too
221 -P, -dM with -E: leave out the markers, or dump the macros
222 -M -MM -MD -MMD write a make rule for the source, the last two compile as well
223 -MF <file> -MT <t> -MQ <t> -MP where the rule goes, what it builds, targets with no recipe
224 -std=<dialect> c89 through c23, and the gnu spellings
225 -fgnuc-version=<v> the GCC release to claim, default 7.0.0
226 -x <lang> treat later inputs as <lang>, or none to stop
227 -O<level> optimize: 0, 1, 2, 3, s, z
228 -fsafety=<tier> check memory safety: off, detect, enforce, kernel
229 -f[no-]sanitize=<what> the negative is taken, the positive is refused by name
230 -f[no-]safety-subobject a write has to stay inside the member it names
231 -f[no-]safety-restrict two restrict pointers of one block may not meet
232 -f<pass> -fno-<pass> -fdump-ir=<what> -fopt-info[-<kind>][=FILE]
233 -fpass-fuel=<pass>=<n>, -fpass-fuel-global=<n> stop a pass, or all of them, after n
234 -fdisable-<pass>[=<funcs>], -fenable-<pass>[=<funcs>] run a pass on some functions only
235 -g -g0 -gdwarf-5, -fno-omit-frame-pointer, -mno-red-zone debug info, frame pointer, red zone
236 -gz[=none|zlib|zlib-gnu|zstd] -gno-split-dwarf compress debug sections, one file not two
237 -flto[=auto|jobserver|<n>] -fno-lto -ffat-lto-objects read, and not done yet
238 -fprofile-use[=<path>] -fprofile-dir=<dir> read too, where -fprofile-generate is refused
239 -f[no-]stack-protector[-strong|-all], -f[no-]stack-clash-protection, -fcf-protection=<edges>
240 -ffunction-sections -fdata-sections a section per function or variable, for --gc-sections
241 -fvisibility=<what> default, hidden, internal or protected, when nothing in the source said
242 -l<name>, -L <dir>, -B <dir> link a library, where to look for one, where our own tools are
243 -fPIC -fpic -fPIE -fpie, -fno-common, -pipe what it does anyway
244 -f[no-]strict-aliasing, -f[no-]delete-null-pointer-checks what it assumes anyway
245 -static -shared -pie -no-pie -nostdlib -nostartfiles -nodefaultlibs -rdynamic -s how to link
246 -Wl,<arg>, -Xlinker <arg>, -fuse-ld=<name> hand an argument to the linker, or pick one
247 -Werror -pedantic -pedantic-errors -w how much to say, and whether it is fatal
248 -m64 -march= -mtune= -mcpu= -mabi= -mcmodel= what machine to generate for
249 -pg -p, -mfentry -mno-fentry call a profiler on the way in, and where that call goes
250 -fpatchable-function-entry=<n>[,<m>] room at the top of every function to patch later
251 -fwrapv, -fwrapv-pointer, -fno-strict-overflow signed or pointer overflow wraps
252 -ftrapv signed overflow stops the program instead
253 -f[no-]signed-char, -f[no-]unsigned-char, -f[no-]short-enums change the ABI
254 -ffp-contract=<how> fuse a multiply and an addition: fast, on or off
255 -fexcess-precision=<how>, -f[no-]rounding-math, -f[no-]trapping-math what may be folded
256 -ffile-prefix-map=<old>=<new> rewrite that front of every path we put in the output
257 -fmacro-prefix-map= -fdebug-prefix-map= -fprofile-prefix-map= the same, one output each
258 -pthread build for more than one thread, and link the library for it
259 -dumpmachine -dumpversion -print-multiarch -print-search-dirs what this compiler is
260 -print-file-name=<name> -print-prog-name=<name> where a file or a program is
261 -print-sysroot the root the headers and the libraries are read under
262 -print-sysroot-provenance every input under it, where it came from and its licence
263 -print-sysroot-digest the sha256 of that record, which names the whole sysroot in one line
264 --fetch <tuple> get the sysroot this release pins for <tuple> and install it in the cache
265 --offline never download anything, which a compilation never does anyway
266 -j[n] compile n translation units at once, default all
267 -v, -### print each phase as it runs, or without running any
268 -save-temps[=cwd|obj], -time keep the .i and the .s, say how long each step took
269 --target=<triple> generate code for <triple>
270 --emit=<kind> exe, obj, archive, asm, preprocessed, tast, ir, mir-final,
271 safety-summary, type-granules
272 --print-config, --print-pipeline print the configuration or the pipeline, and exit
273 --version print the version and exit
274 -h, --help print this message and exit
275
276See spec/04-driver-and-cli.md for the full flag reference.
277";
278
279/// The argument of a flag that may be joined to it or may be the next word.
280///
281/// `-DFOO` and `-D FOO` are the same thing, and `at` is where the flag's own letters end.
282fn joined_or_next(
283 arg: &str,
284 at: usize,
285 args: &[String],
286 i: &mut usize,
287) -> Result<String, CliError> {
288 if arg.len() > at {
289 return Ok(arg[at..].to_owned());
290 }
291 let next = args.get(*i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
292 *i += 1;
293 Ok(next.clone())
294}
295
296/// Every name that may follow `-fsanitize=`, which is gcc 16's list and three of this compiler's
297/// own.
298///
299/// The three are on it because `spec/07-types-and-semantics.md` section 7.7 already promises them:
300/// each undefined behaviour this compiler exploits is listed there with the check that detects it,
301/// and `alias`, `restrict` and `memory` are checks gcc has no spelling for. gcc refuses `memory`
302/// outright, since the sanitizer of that name is clang's. A name being here means it is a name
303/// rather than a typo, and nothing more than that: every one of them is refused after the loop,
304/// because none of them is implemented.
305///
306/// `all` is deliberately absent. gcc takes it only in the negative, so it is handled where each of
307/// those two spellings is read rather than by being on this list.
308const SANITIZERS: [&str; 34] = [
309 "address",
310 "kernel-address",
311 "hwaddress",
312 "kernel-hwaddress",
313 "pointer-compare",
314 "pointer-subtract",
315 "thread",
316 "leak",
317 "undefined",
318 "shift",
319 "shift-base",
320 "shift-exponent",
321 "integer-divide-by-zero",
322 "unreachable",
323 "vla-bound",
324 "null",
325 "return",
326 "signed-integer-overflow",
327 "bounds",
328 "bounds-strict",
329 "alignment",
330 "object-size",
331 "float-divide-by-zero",
332 "float-cast-overflow",
333 "nonnull-attribute",
334 "returns-nonnull-attribute",
335 "bool",
336 "enum",
337 "vptr",
338 "pointer-overflow",
339 "builtin",
340 "alias",
341 "restrict",
342 "memory",
343];
344
345/// Parses a command line, without the program name.
346///
347/// # Errors
348///
349/// Returns the message to print when the arguments do not name a compilation this compiler
350/// can attempt.
351pub fn parse_args(args: &[String]) -> Result<Action, CliError> {
352 let host = Triple::host()
353 .ok_or_else(|| err("this host is not a supported target and no --target was given"))?;
354 let mut opts = Options::new(host);
355 let mut inputs: Vec<Input> = Vec::new();
356 let mut print_config = false;
357 let mut print_pipeline = false;
358 let mut print_plan = false;
359 let mut verbose = false;
360 let mut jobs = Jobs::default();
361 let mut nostdinc = false;
362 let mut sysroot: Option<PathBuf> = None;
363 // What the command line is worth warning about, filled in after the loop rather than during it,
364 // because every question of this kind is about two flags and the last word on both of them is
365 // the end of the loop.
366 let mut notes: Vec<String> = Vec::new();
367 // The whole ten field target, kept beside the three field one because `--target=` can pin a
368 // libc version and `Triple` has nowhere to put it. It decides `__GLIBC_MINOR__` and nothing
369 // else today, and `None` is a command line that named no target, which is this machine.
370 let mut pinned: Option<TargetTuple> = None;
371 let mut output = None;
372 let mut link = LinkOptions::default();
373 let mut query: Option<Query> = None;
374 // What `--fetch` named, and whether `--offline` forbade it. Both are weighed after the loop
375 // because either can be written after the other.
376 let mut fetch: Option<String> = None;
377 let mut offline = false;
378 let mut threads = false;
379 // Which sanitizers are still asked for by the end of the command line. Accumulated across the
380 // loop rather than answered where it was read, because `-fno-sanitize=` turns one off and a
381 // build that asks for a check and then takes it back has asked for nothing. What happens to a
382 // set that is not empty is decided after the loop.
383 let mut sanitizers: Vec<&str> = Vec::new();
384 // `-x` applies to inputs that come after it and stays in effect until the next one, which
385 // is why it is tracked across the loop rather than attached to a single argument.
386 let mut forced: Option<InputKind> = None;
387 // What `-iprefix` last said, stuck on the front of every later `-iwithprefix`. It applies to
388 // the flags after it and not the ones before, so a command line may set it more than once.
389 // GCC's default is its own installed header directory with the last component taken off,
390 // which is a path a cross compiler's build system knows and passes; there is no equivalent
391 // here, so with no `-iprefix` the prefix is nothing and `-iwithprefix` names a directory
392 // outright.
393 let mut iprefix = String::new();
394
395 let mut i = 0;
396 while i < args.len() {
397 let arg = args[i].as_str();
398 i += 1;
399 match arg {
400 "-h" | "--help" => return Ok(Action::Help),
401 "--version" => return Ok(Action::Version),
402 // The sysroot fetch, which is weighed after the loop rather than acted on here, because
403 // `--offline` written after it has to be able to forbid it. Both spellings, since a
404 // flag that takes a tuple gets written both ways and neither is a guess at what the
405 // other meant.
406 "--fetch" => {
407 let value = args
408 .get(i)
409 .ok_or_else(|| err("--fetch requires the target to get a sysroot for"))?;
410 i += 1;
411 fetch = Some(value.clone());
412 }
413 _ if arg.starts_with("--fetch=") => {
414 fetch = Some(arg["--fetch=".len()..].to_owned());
415 }
416 // Accepted on any command line and only ever read by the fetch, because an ordinary
417 // compile downloads nothing with or without it. So this flag takes nothing away today,
418 // which is the property section 13.2 asks for rather than an omission: a build that
419 // passes it is saying what it expects of this compiler, and what it expects is already
420 // true.
421 "--offline" => offline = true,
422 "--print-config" => print_config = true,
423 "--print-pipeline" => print_pipeline = true,
424 "-###" => print_plan = true,
425 "-v" => verbose = true,
426 // The files a compilation goes through, kept rather than thrown away. The bare
427 // spelling means `=obj` and not `=cwd`, which is not what the manual says and is what
428 // gcc 16 does; `SaveTemps::Object` carries the measurement.
429 "-save-temps" => opts.save_temps = SaveTemps::Object,
430 _ if arg.starts_with("-save-temps=") => {
431 opts.save_temps = arg["-save-temps=".len()..].parse().map_err(err)?;
432 }
433 // How long each step took. A misspelling of this is worth rejecting rather than
434 // ignoring, since a run that says nothing looks like a compilation that took no time.
435 "-time" => opts.time = true,
436 "-c" => opts.emit = EmitKind::Object,
437 "-S" => opts.emit = EmitKind::Asm,
438 "-E" => opts.emit = EmitKind::Preprocessed,
439 "-g" => opts.debug_info = true,
440 // GCC's own levels of how much debug information to write. Zero is none and every
441 // other number is some, and this compiler has one amount, so the numbers above zero
442 // all mean the same thing here. `-ggdb` is the same flag asking for whatever the
443 // debugger on the machine prefers, which is what we emit anyway.
444 "-g0" => opts.debug_info = false,
445 "-g1" | "-g2" | "-g3" | "-ggdb" | "-ggdb1" | "-ggdb2" | "-ggdb3" => {
446 opts.debug_info = true;
447 }
448 // The version of DWARF to write. We write DWARF 5 and nothing else, so a build that
449 // asks for another version is told rather than handed a file it cannot read.
450 "-gdwarf" | "-gdwarf-5" => opts.debug_info = true,
451 _ if arg.starts_with("-gdwarf-") => {
452 return Err(err(format!(
453 "{arg}: this compiler writes DWARF 5 and no other version, see \
454 spec/11-debug-info.md"
455 )));
456 }
457 // Whether the debug information goes in a file of its own beside the object. gcc
458 // writes that `.dwo` whether or not it found anything to put in it, which means a
459 // build system that declares the file as an output gets one and a make rule that
460 // depends on it fires. Refused for that reason rather than taken: section 4.1 takes a
461 // flag that changes nothing and refuses one that changes what is produced, and a file
462 // that does not appear is the plainest change of that kind there is. The negative
463 // spelling is taken, because putting it all in the object is what happens anyway.
464 "-gno-split-dwarf" => {}
465 "-gsplit-dwarf" => {
466 return Err(err(format!(
467 "{arg}: this compiler writes no separate `.dwo` file, and a build that \
468 expects one beside each object would wait for a file that never arrives, \
469 see spec/11-debug-info.md"
470 )));
471 }
472 // How the debug sections are compressed. There are none yet, so every answer produces
473 // the same bytes and taking the flag promises nothing that is not kept. The value is
474 // still checked, because a typo in a distribution's flags is worth finding when the
475 // compiler reads it rather than when somebody later wonders why nothing got smaller.
476 // Bare `-gz` means `zlib`, which the manual leaves for the reader to discover.
477 "-gz" => opts.compress = Compress::Zlib,
478 _ if arg.starts_with("-gz=") => {
479 let how = &arg["-gz=".len()..];
480 opts.compress = how.parse().map_err(|()| {
481 err(format!(
482 "`{how}` is not a way to compress debug sections, which is none, zlib, \
483 zlib-gnu or zstd"
484 ))
485 })?;
486 }
487 "-Werror" => opts.warnings_are_errors = true,
488 // Nothing that is not fatal is said at all. Read at the one place a diagnostic goes
489 // through rather than here, so that a warning `-w` dropped is not counted either.
490 "-w" => opts.warnings = false,
491 "-pedantic-errors" => {
492 opts.pedantic = true;
493 opts.warnings_are_errors = true;
494 }
495 "-P" => opts.line_markers = false,
496 // The dependency family, which section 4.4 calls required because every build system
497 // that generates its own makefiles asks for it. The two that end in `D` write a file
498 // beside the object and let the compilation happen, and the two that do not write to
499 // standard output and stop after it. Nothing here turns the system headers back on
500 // once a flag has turned them off, which is GCC's behaviour and is why `-MM -M` is
501 // `-MM`: the flag asking for fewer of them is the one with something to say.
502 "-M" => {
503 opts.deps.emit = true;
504 opts.deps.instead_of_compiling = true;
505 }
506 "-MM" => {
507 opts.deps.emit = true;
508 opts.deps.instead_of_compiling = true;
509 opts.deps.system_headers = false;
510 }
511 "-MD" => opts.deps.emit = true,
512 "-MMD" => {
513 opts.deps.emit = true;
514 opts.deps.system_headers = false;
515 }
516 "-MP" => opts.deps.phony = true,
517 // These three take a word and only in the separated form, which is how GCC spells
518 // them and how every build system writes them.
519 "-MF" | "-MT" | "-MQ" => {
520 let value =
521 args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
522 i += 1;
523 match arg {
524 "-MF" => opts.deps.file = Some(value.clone()),
525 // The whole of the difference between the two. `-MT` is for a build that has
526 // already escaped what it is passing, and `-MQ` is for one that has a name
527 // and wants it to arrive as that name.
528 "-MT" => opts.deps.targets.push(value.clone()),
529 _ => opts.deps.targets.push(deps::escaped(value)),
530 }
531 }
532 // The questions a build system asks before it compiles anything. Answered after the
533 // loop, because each one is about the target or the library search and the command
534 // line has not finished saying what those are.
535 "-dumpmachine" => query = Some(Query::Machine),
536 "-dumpversion" | "-dumpfullversion" => query = Some(Query::Version),
537 "-print-multiarch" => query = Some(Query::Multiarch),
538 "-print-search-dirs" => query = Some(Query::SearchDirs),
539 "-print-sysroot" => query = Some(Query::Sysroot),
540 // Both spellings, because this one is ours rather than GCC's and our own documents
541 // write it both ways: section 13.5 of `spec/cross-compile/13-distribution.md` gives it
542 // two dashes like the other flags we invented, and document 12's table gives it one
543 // like the `-print-` family it sits in. A person who reads either and types what it
544 // says is right, so neither is refused.
545 "-print-sysroot-provenance" | "--print-sysroot-provenance" => {
546 query = Some(Query::SysrootProvenance);
547 }
548 "-print-sysroot-digest" | "--print-sysroot-digest" => {
549 query = Some(Query::SysrootDigest);
550 }
551 "-print-libgcc-file-name" => query = Some(Query::Libgcc),
552 _ if arg.starts_with("-print-file-name=") => {
553 query = Some(Query::FileName(arg["-print-file-name=".len()..].to_owned()));
554 }
555 _ if arg.starts_with("-print-prog-name=") => {
556 query = Some(Query::ProgName(arg["-print-prog-name=".len()..].to_owned()));
557 }
558 // A program built to run in more than one thread. On every platform this compiler
559 // targets that is a macro the library's headers read and one more library on the
560 // link line, and the library is added after the loop so that it lands after the
561 // objects that refer to it.
562 "-pthread" | "-pthreads" => {
563 opts.defines.push("_REENTRANT".to_owned());
564 threads = true;
565 }
566 "-ansi" => {
567 opts.std = Std::C89;
568 opts.gnu_extensions = false;
569 }
570 // `-Wpedantic` is the same flag under the name the `-W` family gives it, which is
571 // the spelling a build system that groups its warning flags tends to write.
572 "-pedantic" | "-Wpedantic" => opts.pedantic = true,
573 // Both directions, because a build that needs this for one directory turns it back
574 // off for the next one rather than leaving it on for the whole tree.
575 "-fpermissive" => opts.permissive = true,
576 "-fno-permissive" => opts.permissive = false,
577 "-ffreestanding" => opts.hosted = false,
578 "-fhosted" => opts.hosted = true,
579 "-fno-builtin" => opts.builtins = false,
580 "-fbuiltin" => opts.builtins = true,
581 // The C89 dialects are under GNU's reading whatever this says, so turning it off
582 // there is turning off something the dialect asked for, which is accepted and does
583 // nothing. gcc refuses that command line, and there is nothing it could have meant.
584 "-fgnu89-inline" => opts.gnu89_inline = true,
585 "-fno-gnu89-inline" => opts.gnu89_inline = false,
586 // Both directions of each, because a build system that wants one of these usually
587 // writes it beside the flag that turns it back off for one directory.
588 "-fno-omit-frame-pointer" => opts.frame_pointer = true,
589 "-fomit-frame-pointer" => opts.frame_pointer = false,
590 // Both directions again, for the same reason, and a third answer for a command line
591 // that wrote neither: see `reorder_blocks` in `rucc_session`.
592 "-freorder-blocks" => opts.reorder_blocks = Some(true),
593 "-fno-reorder-blocks" => opts.reorder_blocks = Some(false),
594 // gcc's name for the scheduler that runs after the registers are handed out, which is
595 // the only one rucc has: see `schedule_insns` in `rucc_session`. gcc also takes
596 // `-fschedule-insns` for the pass before allocation, and taking that one here would be
597 // a flag that says a pass ran when none did.
598 "-fschedule-insns2" => opts.schedule_insns = Some(true),
599 "-fno-schedule-insns2" => opts.schedule_insns = Some(false),
600 "-mno-red-zone" => opts.red_zone = false,
601 "-mred-zone" => opts.red_zone = true,
602 // Four flags rather than one with an argument, which is how gcc spells them and how
603 // every build line writes them. Last one wins, because a package build puts
604 // `-fstack-protector-strong` in its global flags and a directory that cannot have one
605 // turns it back off on the line after.
606 "-fno-stack-protector" | "-fno-stack-protector-all" | "-fno-stack-protector-strong" => {
607 opts.protector = Protector::None;
608 }
609 "-fstack-protector" => opts.protector = Protector::Buffers,
610 "-fstack-protector-strong" => opts.protector = Protector::Strong,
611 "-fstack-protector-all" => opts.protector = Protector::All,
612 // The other half of what a hardened build asks for, and it is a question about the
613 // frame rather than about the function, so it is a switch rather than a level.
614 "-fstack-clash-protection" => opts.stack_clash = true,
615 "-fno-stack-clash-protection" => opts.stack_clash = false,
616 // The third of them, and the one that is a question with an argument rather than a
617 // family of spellings, because what it asks about is which of the two edges of a
618 // control flow transfer is checked. Bare is both of them, which is what gcc does.
619 "-fcf-protection" => opts.control = Control::Full,
620 "-fno-cf-protection" => opts.control = Control::None,
621 // Two spellings of the same request, which is what gcc has as well. `-p` was the older
622 // profiler and `-pg` the one that also recorded who called whom, and on every platform
623 // this compiler targets there is now one hook and both ask for it.
624 "-pg" | "-p" => {
625 opts.profile = true;
626 link.profile = true;
627 }
628 // Accepted on their own and doing nothing on their own, which is gcc's behaviour: they
629 // say where the call goes and a command line that asked for no call has nowhere to put
630 // one. That matters because a build system that sets `-mfentry` globally and `-pg` per
631 // directory is a build system that would otherwise fail on every other directory.
632 "-mfentry" => opts.hook = Hook::Early,
633 "-mno-fentry" => opts.hook = Hook::Late,
634 // GCC drops its own include directory along with the system ones, because its
635 // headers are half of a pair with the library's and half a pair is worse than
636 // none. A build that passes this is supplying the whole set itself.
637 "-nostdinc" => nostdinc = true,
638 "-o" => {
639 output = Some(args.get(i).ok_or_else(|| err("-o requires an argument"))?.clone());
640 i += 1;
641 }
642 // The flags that take a directory only in the separated form. GCC spells them
643 // this way and nothing writes `-iquotedir`, so accepting the joined form would
644 // mean guessing at a path that starts with the flag's own letters.
645 // Apple's spelling of `--sysroot`, and the one its own build systems pass. The
646 // two mean the same thing here: the configured directories are under there rather
647 // than under the root.
648 "-isysroot" => {
649 let dir = args.get(i).ok_or_else(|| err("-isysroot requires an argument"))?;
650 i += 1;
651 sysroot = Some(PathBuf::from(dir));
652 }
653 "-iquote" | "-isystem" | "-idirafter" => {
654 let dir = args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
655 i += 1;
656 match arg {
657 "-iquote" => opts.search.push_quote(dir.clone()),
658 "-isystem" => opts.search.push_system(dir.clone()),
659 _ => opts.search.push_after(dir.clone()),
660 }
661 }
662 "-iprefix" => {
663 iprefix = args.get(i).ok_or_else(|| err("-iprefix requires an argument"))?.clone();
664 i += 1;
665 }
666 // Where GCC puts these is not where its manual says it puts them, and this is the
667 // measured answer rather than the documented one: `-iwithprefix` lands in the
668 // `-isystem` slot and not the `-idirafter` slot, and `-iwithprefixbefore` lands in
669 // the `-I` slot. A cross build that uses them is relying on the behaviour, since
670 // that is the compiler it was developed against.
671 "-iwithprefix" | "-iwithprefixbefore" => {
672 let dir = args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
673 i += 1;
674 let dir = format!("{iprefix}{dir}");
675 if arg == "-iwithprefix" {
676 opts.search.push_system(dir);
677 } else {
678 opts.search.push_bracket(dir);
679 }
680 }
681 "-include" | "-imacros" => {
682 let name = args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
683 i += 1;
684 opts.preincludes
685 .push(Preinclude { name: name.clone(), macros_only: arg == "-imacros" });
686 }
687 // The flag `-iquote` was introduced to replace, still passed by build systems old
688 // enough to predate the replacement. It is not a directory: it says that every `-I`
689 // so far is for quoted includes only, and that a quoted include stops looking next
690 // to the file that wrote it.
691 "-I-" => opts.search.split_quote_chain(),
692 "-x" => {
693 let lang = args.get(i).ok_or_else(|| err("-x requires an argument"))?;
694 i += 1;
695 forced = if lang == "none" {
696 None
697 } else {
698 Some(InputKind::from_x_arg(lang).map_err(|e| err(format!("{e}")))?)
699 };
700 }
701 // Not a GCC flag. spec/03-architecture.md section 3.5 compiles several
702 // translation units in one process rather than making the build system fork, and
703 // section 3.8's determinism check compares `-j1` against `-j16`, so the knob has
704 // to exist and has to be spelled the way `make` spells it.
705 // `-DFOO`, `-D FOO` and the same for `-U` and `-I`. Both forms are in wide use
706 // and a build system may produce either, so both are read here rather than
707 // being normalised by whatever generated the command line.
708 _ if arg.starts_with("-D") => {
709 let value = joined_or_next(arg, 2, args, &mut i)?;
710 opts.defines.push(value);
711 }
712 _ if arg.starts_with("-U") => {
713 let value = joined_or_next(arg, 2, args, &mut i)?;
714 opts.undefines.push(value);
715 }
716 _ if arg.starts_with("-I") => {
717 let dir = joined_or_next(arg, 2, args, &mut i)?;
718 opts.search.push_bracket(dir);
719 }
720 _ if arg.starts_with("-std=") => {
721 let name = &arg["-std=".len()..];
722 let (std, gnu) = Std::from_flag(name)
723 .ok_or_else(|| err(format!("unknown dialect `{name}`, see --help")))?;
724 opts.std = std;
725 opts.gnu_extensions = gnu;
726 }
727 // Section 4.5. The claim decides which half of glibc's `sys/cdefs.h` we are
728 // handed, so a differential run that does not set it is comparing two compilers
729 // that believe they are different compilers.
730 // GCC packs these into one flag, so `-dDI` is two of them. Letters in the family
731 // that we have not written yet are accepted and ignored, because a dump is a
732 // debugging aid and a build that asks for one should still compile. A letter
733 // outside the family falls through to the unknown option error, which is what
734 // keeps `-dumpversion` from being read as a dump of nothing.
735 _ if Dumps::is_family(arg) => {
736 opts.dumps.add(&arg[2..]);
737 }
738 // One name at a time, which is what a build that means its own `memcpy` and the
739 // library's everything else writes. The name is not checked against a list, because
740 // the flag is about what the program means by a name and a program is allowed to mean
741 // something by a name this compiler has never heard of.
742 _ if arg.starts_with("-fno-builtin-") => {
743 opts.no_builtin.push(arg["-fno-builtin-".len()..].to_owned());
744 }
745 _ if arg.starts_with("-fgnuc-version=") => {
746 let v = &arg["-fgnuc-version=".len()..];
747 opts.gnuc = v.parse().map_err(err)?;
748 }
749 // spec/13-gnu-compat.md section 13.3 promises this flag an error that says why rather
750 // than the unknown option one, because a build reaching for it is asking for a feature
751 // and deserves to be told it is not coming rather than told the spelling is wrong.
752 // The negative form is what this compiler does anyway, so it is taken and dropped.
753 "-fnested-functions" => {
754 return Err(err(
755 "nested functions are not supported: a call to one goes through a trampoline \
756 written on the stack, which no target that enforces an unexecutable stack \
757 allows",
758 ));
759 }
760 "-fno-nested-functions" => {}
761 // Which of the two links the output is for, which is a real difference and not a
762 // description of what happens anyway. Everything here is position independent either
763 // way, and what these decide is whether a name may be one another object defines or
764 // replaces, because a link that produces an executable puts every name in the same
765 // program and a link that produces a shared library does not.
766 //
767 // It matters that they are accepted at all, whatever they then do. Every autoconf and
768 // cmake build puts `-fPIC` on the compile line, so a compiler that rejects it cannot
769 // be the `CC` of a project that has a configure script, whatever else it can do. That
770 // is how this was found: building SQLite's test fixture stopped on it.
771 "-fPIC" | "-fpic" => opts.pic = Pic::Library,
772 // Not a synonym of the pair above, which is what they were treated as until #756. The
773 // library is the expensive answer and gcc makes it the one that has to be asked for,
774 // so this is also what nothing at all means.
775 "-fPIE" | "-fpie" => opts.pic = Pic::Executable,
776 // A different question from the pair above, and the one every distribution build of a
777 // shared library answers. `-fPIC` decides how an address is reached, and this decides
778 // whether the optimizer may believe a body it can see, because an exported name is one
779 // the dynamic linker may find another definition of first. On by default, which is
780 // gcc's arrangement and is the honest answer, and off is a promise the build makes and
781 // nothing checks.
782 "-fsemantic-interposition" => opts.interposition = true,
783 "-fno-semantic-interposition" => opts.interposition = false,
784 // Two requests rather than one, and the same table answers both, so what decides is
785 // whether either of them is standing. gcc arranges it the same way: the asynchronous
786 // one is the default here and it implies the other, and a line that asks for a table
787 // and against an asynchronous one gets a table.
788 "-fasynchronous-unwind-tables" => opts.async_unwind_tables = true,
789 "-fno-asynchronous-unwind-tables" => opts.async_unwind_tables = false,
790 "-funwind-tables" => opts.unwind_tables = true,
791 "-fno-unwind-tables" => opts.unwind_tables = false,
792 // The other direction is a request, not a description, and it is one this compiler
793 // cannot grant, so it gets the treatment section 13.3 asks for rather than the unknown
794 // option error. Answering it by carrying on would be answering a different question:
795 // the code would still be position independent, which is correct everywhere an
796 // ordinary program runs and is wrong in a kernel, where the flag is written precisely
797 // because there is no loader to fill a global offset table in.
798 "-fno-pic" | "-fno-pie" => {
799 return Err(err(
800 "position dependent code is not supported: an address that may be in another \
801 object is loaded out of the global offset table, and nothing here emits the \
802 absolute form this asks for. Use -no-pie if what you meant was how to link",
803 ));
804 }
805 // A section per function and a section per variable, which is what makes
806 // `--gc-sections` able to drop anything: a linker can leave out a section nothing
807 // reaches and cannot leave out half of one. Both directions are taken, and the off
808 // one is the default rather than a refusal, since a build that writes it is asking
809 // for what happens anyway.
810 "-ffunction-sections" => opts.function_sections = true,
811 "-fno-function-sections" => opts.function_sections = false,
812 "-fdata-sections" => opts.data_sections = true,
813 "-fno-data-sections" => opts.data_sections = false,
814 // Another description of what this compiler does. A file scope declaration with no
815 // initializer is written into `.bss` as an ordinary defined symbol, not offered to the
816 // linker as a common one for it to merge, which is what `-fno-common` asks for and what
817 // gcc has done by default since 10. Nothing in the front end produces `Linkage::Common`
818 // at all.
819 "-fno-common" => {}
820 // What overflows rather than being undefined. Every one of these takes something away
821 // from the optimizer rather than asking it to do anything, which is why the negative
822 // spellings are the interesting ones and the positive spellings are the default.
823 //
824 // `-fno-strict-overflow` is both of the others, which is gcc's own reading of it: its
825 // help text for `-fstrict-overflow` says "negated as -fwrapv -fwrapv-pointer". So it is
826 // written here as the pair rather than kept as a third thing to test everywhere.
827 //
828 // `-ftrapv` is the exception and is the one that asks for something. It is the other
829 // answer to the question `-fwrapv` answers, so the two cannot both hold and each clears
830 // the other, which makes the last one on the command line the one that counts. That is
831 // gcc 16's behaviour and was measured rather than read: `-ftrapv -fwrapv` emits no
832 // checked calls and `-fwrapv -ftrapv` emits them. The positive spelling of the pointer
833 // question is left alone by both, because neither has anything to say about it.
834 "-fwrapv" => {
835 opts.wrapping.signed = true;
836 opts.wrapping.trap = false;
837 }
838 "-fno-wrapv" => opts.wrapping.signed = false,
839 "-fwrapv-pointer" => opts.wrapping.pointer = true,
840 "-fno-wrapv-pointer" => opts.wrapping.pointer = false,
841 "-fno-strict-overflow" => opts.wrapping = Wrapping::ALL,
842 // Which does not clear the checked one, because gcc does not: `-ftrapv
843 // -fstrict-overflow` still emits the calls. It says what is assumed and not what
844 // happens.
845 "-fstrict-overflow" => {
846 opts.wrapping.signed = false;
847 opts.wrapping.pointer = false;
848 }
849 "-ftrapv" => {
850 opts.wrapping.trap = true;
851 opts.wrapping.signed = false;
852 }
853 "-fno-trapv" => opts.wrapping.trap = false,
854 // The two flags that say what a plain `char` is, which is one question with two
855 // spellings each: gcc reads `-fno-signed-char` as `-funsigned-char` and
856 // `-fno-unsigned-char` as `-fsigned-char`, so there are four ways to write two
857 // answers and the last one written wins. Nothing is set until one of them is given,
858 // because the target's own ABI is the answer otherwise and it is not the same answer
859 // everywhere: x86-64 and Apple's arm64 are signed, Linux's arm64 is not.
860 "-fsigned-char" | "-fno-unsigned-char" => opts.char_signed = Some(true),
861 "-funsigned-char" | "-fno-signed-char" => opts.char_signed = Some(false),
862 // And the size of an enumeration, which is the other thing in this group that changes
863 // the ABI rather than the code.
864 "-fshort-enums" => opts.short_enums = true,
865 "-fno-short-enums" => opts.short_enums = false,
866 // And Microsoft's reading of an anonymous member, which changes the layout of every
867 // record that writes a tag on one. Nothing is set until one of them is given, because
868 // the target is the answer otherwise: gcc's mingw build has this on and its Linux
869 // build has it off.
870 "-fms-extensions" => opts.ms_extensions = Some(true),
871 "-fno-ms-extensions" => opts.ms_extensions = Some(false),
872 // And the request, which is the one that cannot be granted. It is a real difference and
873 // not a preference: two files each writing `int g;` link under `-fcommon` and are a
874 // duplicate definition without it, which is the whole reason the flag survives.
875 "-fcommon" => {
876 return Err(err(
877 "a tentative definition is written into .bss as its own symbol here, and \
878 nothing emits the common symbol this asks the linker to merge. Give the \
879 variable a definition in one file and declare it extern in the others",
880 ));
881 }
882 // Both directions of this one are recorded, and what they decide is whether lowering
883 // names the type each access goes through. Turning it off is the front end leaving the
884 // name off rather than a pass being told to ignore one it can see, which is one
885 // condition in one place, and it is the reading that survives link time optimization:
886 // a unit built with the flag off keeps its own answer when its bodies end up in a
887 // module beside bodies that were not.
888 //
889 // Nothing in the pipeline reads those names yet. Layer 3 of the alias analysis does
890 // and is tested, and no pass at any level asks the alias analysis anything today, so
891 // no program compiles differently for having passed this. The flag is wired anyway,
892 // because the change that makes a pass ask is not the change anybody will remember to
893 // wire it in, and a flag that is taken and dropped once the names mean something is
894 // the miscompilation `spec/04-driver-and-cli.md` section 4.1 warns about in as many
895 // words.
896 "-fstrict-aliasing" => opts.strict_aliasing = true,
897 "-fno-strict-aliasing" => opts.strict_aliasing = false,
898 // The same shape of answer for the same reason, and the flag the kernel writes beside
899 // the one above it.
900 //
901 // Nothing here concludes that a pointer is not null from the fact that it was
902 // dereferenced. There is no such conclusion to draw from, because no pass records one:
903 // a load says where it read and nothing else, and a comparison against null is an
904 // ordinary comparison of two values the optimizer has no fact about. So a function
905 // that reads through a pointer and then tests it keeps the test, which is what the
906 // kernel wants and what `-fno-delete-null-pointer-checks` asks for, and what gcc has
907 // to be asked for because it draws the conclusion by default.
908 //
909 // `-fdelete-null-pointer-checks` is the request to draw it, and it goes the way
910 // `-fstrict-aliasing` does: assuming less than was asked for costs speed and not
911 // correctness, and `-O2` implies it, so refusing it would stop builds for nothing.
912 "-fdelete-null-pointer-checks" | "-fno-delete-null-pointer-checks" => {}
913 // The floating point group, which goes the same way and for the same reason, and which
914 // is worth writing out because the reason is easy to get backwards.
915 //
916 // Each of these has a restrictive spelling and a permissive one. The restrictive ones,
917 // `-frounding-math` and `-ftrapping-math`, say that the rounding mode may have been
918 // changed and that an exception raised by an operation may be looked at, so an
919 // arithmetic the compiler folds at compile time is an arithmetic whose rounding and
920 // whose exception the program does not get. Nothing here folds any floating point
921 // arithmetic in a function body: `0.1 + 0.2` is an `fadd` and `1.0 / 0.0` is a divide
922 // that runs, at every level. So both of those describe what already happens.
923 //
924 // The permissive ones are the other half, and they are licences rather than requests
925 // for an answer. `-fno-rounding-math` says the rounding mode is the default one and
926 // `-fno-trapping-math` says nothing looks at the exceptions, which together are
927 // permission to fold. Not folding is the conservative side of that permission and is
928 // what a program is entitled to whichever was written, so `-fno-rounding-math` costs
929 // speed and not correctness, which is the test section 4.1 puts a licence through.
930 "-frounding-math" | "-fno-rounding-math" => {}
931 // `-fno-trapping-math` is the one of the four that is kept, because there is one
932 // conversion this compiler does not fold and gcc folds under it, and the two answers
933 // differ. Converting a constant floating value to an integer type it does not fit in
934 // is undefined behaviour rather than a value: left to the hardware it is one
935 // instruction and the answer is the integer indefinite value, and folded it is the
936 // nearest end of the integer's range. Both compilers leave it to the instruction by
937 // default and gcc folds it under this flag, so a program built with it and compiled
938 // without it gets a different number rather than a slower one. `-ftrapping-math` is
939 // gcc's default, so a build spelling it out is asking for what it already has.
940 "-ftrapping-math" => opts.trapping_math = true,
941 "-fno-trapping-math" => opts.trapping_math = false,
942 // About temporary files rather than about code. There is nothing between the phases of
943 // one compilation here to write to a file in the first place.
944 "-pipe" => {}
945 // Nothing here writes colour, so all of these are the same answer, and it is the answer
946 // that costs nothing: the diagnostics come out plain either way and no build depends on
947 // an escape sequence being there. Taken rather than refused because cmake writes
948 // `-fdiagnostics-color=always` on every compile line when the generator is ninja, which
949 // makes this the second most common flag after `-fPIC` to stop a build over a question
950 // about how the text looks.
951 "-fdiagnostics-color" | "-fno-diagnostics-color" => {}
952 _ if arg.starts_with("-fdiagnostics-color=") => {}
953 // The link flags. None of them changes the compilation, which is why they are
954 // collected apart from `opts` and why `-lm` on a `-c` line is a note rather than an
955 // error: it is a thing said to a linker that is not going to run.
956 "-static" => link.is_static = true,
957 "-shared" => link.shared = true,
958 "-pie" => link.pie = Some(true),
959 "-no-pie" | "-nopie" => link.pie = Some(false),
960 "-nostdlib" => link.no_stdlib = true,
961 "-nostartfiles" => link.no_startfiles = true,
962 "-nodefaultlibs" => link.no_defaultlibs = true,
963 "-fno-builtins-lib" => link.no_builtins_lib = true,
964 "-fbuiltins-lib" => link.no_builtins_lib = false,
965 "-rdynamic" | "-export-dynamic" => link.export_dynamic = true,
966 "-s" => link.strip = true,
967 // Into the ordered input list rather than a list of its own, because a great many of
968 // the linker's options are a bracket around the files after them and an option that
969 // lost its place among them says nothing. `--whole-archive` is the one that found this.
970 "-Xlinker" => {
971 let next = args.get(i).ok_or_else(|| err("-Xlinker requires an argument"))?;
972 i += 1;
973 inputs.push(Input::linker(next));
974 }
975 _ if arg.starts_with("-Wl,") => {
976 // Commas separate arguments rather than being part of one, which is what makes
977 // `-Wl,-rpath,/opt/lib` two words to the linker and one word here.
978 inputs.extend(arg["-Wl,".len()..].split(',').map(Input::linker));
979 }
980 _ if arg.starts_with("-fuse-ld=") => {
981 link.use_ld = Some(arg["-fuse-ld=".len()..].to_owned());
982 }
983 _ if arg.starts_with("-l") && arg.len() > 2 => {
984 inputs.push(Input::library(&arg[2..]));
985 }
986 "-l" => {
987 let next = args.get(i).ok_or_else(|| err("-l requires an argument"))?;
988 i += 1;
989 inputs.push(Input::library(next));
990 }
991 _ if arg.starts_with("-L") => {
992 link.search.push(PathBuf::from(joined_or_next(arg, 2, args, &mut i)?));
993 }
994 _ if arg.starts_with("-B") => {
995 link.prefixes.push(PathBuf::from(joined_or_next(arg, 2, args, &mut i)?));
996 }
997 _ if arg.starts_with("-j") => {
998 jobs = Jobs::parse(&arg[2..]).map_err(err)?;
999 }
1000 _ if arg.starts_with("--sysroot=") => {
1001 sysroot = Some(PathBuf::from(&arg["--sysroot=".len()..]));
1002 }
1003 _ if arg.starts_with("--target=") => {
1004 let t = &arg["--target=".len()..];
1005 opts.target = t.parse().map_err(|e| err(format!("{e}")))?;
1006 // The same string again, as the model that has room for a libc version. A spelling
1007 // the three field parser took and this one does not is not an error, because the
1008 // one that decides what is compiled has already accepted it and the only thing
1009 // lost is a version nobody asked for.
1010 pinned = t.parse().ok();
1011 }
1012 _ if arg.starts_with("--emit=") => {
1013 let k = &arg["--emit=".len()..];
1014 opts.emit = k
1015 .parse()
1016 .map_err(|()| err(format!("unknown --emit kind `{k}`, see --help")))?;
1017 }
1018 // A bare `-O` is `-O1`, which is what GCC has and what a hand written makefile tends
1019 // to write. `-Og` is GCC's level for a build somebody is going to step through, and
1020 // it is `-O1` with the transformations that move code around left out; this compiler
1021 // has no such level yet, so it is the nearest one and `--print-pipeline` says what
1022 // that came to rather than the flag pretending otherwise.
1023 "-O" | "-Og" => opts.opt_level = rucc_session::OptLevel::O1,
1024 // The union of `-O3` and `-ffast-math`, and the second half of that changes what
1025 // floating point arithmetic means. Refused rather than taken as `-O3`, because a
1026 // build that asks for fast math and is quietly given ordinary arithmetic gets a
1027 // slower program than it asked for and a build that is given fast math it did not
1028 // ask for gets a wrong one.
1029 "-Ofast" => {
1030 return Err(err(
1031 "-Ofast is -O3 with fast math, and fast math is not implemented, see \
1032 spec/04-driver-and-cli.md section 4.6",
1033 ));
1034 }
1035 _ if arg.starts_with("-O") => {
1036 opts.opt_level = arg[2..]
1037 .parse()
1038 .map_err(|()| err(format!("unknown optimization level `{arg}`")))?;
1039 }
1040 // How far a multiply and an addition may be fused into one rounding. Before the
1041 // optimizer's `-f` family below for the reason the ones under it are, and kept rather
1042 // than dropped because it is the one flag in its group this compiler could act on: it
1043 // rides into the IR as an attribute on each function with a body, so the day the code
1044 // generator forms an `fma` it already knows which functions were given permission.
1045 // Nothing forms one today, under any value of this and under any `-march=`.
1046 _ if arg.starts_with("-ffp-contract=") => {
1047 let how = &arg["-ffp-contract=".len()..];
1048 opts.fp_contract = how.parse().map_err(|()| {
1049 err(format!("`{how}` is not a contraction, which is fast, on or off"))
1050 })?;
1051 }
1052 // How much of an expression may be computed wider than it was written. The values are
1053 // gcc's and so is the refusal of anything else, and none of the three changes anything
1054 // here: an operation is computed in the type C says it is on every target this compiler
1055 // has a back end for, so `__FLT_EVAL_METHOD__` is 0 and `standard` is already what
1056 // happens. `fast` and `16` are permission to be wider, which is a licence this takes
1057 // and does not use, the same way the two above are. The flag is worth taking because
1058 // glibc's headers and a good deal of configure output write it, and because the answer
1059 // it asks about is one this compiler can state rather than guess at: there is no x87
1060 // target here, which is the machine the whole question was invented for.
1061 // Whether a local and a spilled value that are never both wanted may be the same bytes
1062 // of the frame. gcc's three values, and two of them mean the same thing here: what rucc
1063 // shares is a local whose address provably never leaves the function, which is narrower
1064 // than `named_vars` and narrower still than `all`, so both of them get it. `none` is
1065 // the one that changes anything, and it is the flag a program that reads a local
1066 // through a pointer it kept past the end of the block writes.
1067 _ if arg.starts_with("-fstack-reuse=") => {
1068 let how = &arg["-fstack-reuse=".len()..];
1069 opts.stack_reuse = match how {
1070 "all" | "named_vars" => Some(true),
1071 "none" => Some(false),
1072 _ => {
1073 return Err(err(format!(
1074 "`{how}` is not a stack reuse, which is all, named_vars or none"
1075 )));
1076 }
1077 };
1078 }
1079 _ if arg.starts_with("-fexcess-precision=") => {
1080 let how = &arg["-fexcess-precision=".len()..];
1081 if !matches!(how, "16" | "fast" | "standard") {
1082 return Err(err(format!(
1083 "`{how}` is not an excess precision, which is 16, fast or standard"
1084 )));
1085 }
1086 }
1087 // Which front of a path is rewritten before it reaches the output, which is how a
1088 // build gets the same bytes out of two different directories. The four spellings are
1089 // one flag each into three lists, and `-ffile-prefix-map=` is the three of them at
1090 // once. Only the macro list does anything today, because `__FILE__` is the only place
1091 // a path reaches the output: there is no DWARF and no profile data yet, so the other
1092 // two are recorded for the work that will read them. The argument splits at the last
1093 // `=` rather than the first, which is gcc's rule and is what lets a directory with an
1094 // `=` in its name be the old half.
1095 _ if arg.starts_with("-fmacro-prefix-map=") => {
1096 let (old, new) = rewrite(arg, "-fmacro-prefix-map=")?;
1097 opts.prefix_map.macros.push(old, new);
1098 }
1099 _ if arg.starts_with("-fdebug-prefix-map=") => {
1100 let (old, new) = rewrite(arg, "-fdebug-prefix-map=")?;
1101 opts.prefix_map.debug.push(old, new);
1102 }
1103 _ if arg.starts_with("-fprofile-prefix-map=") => {
1104 let (old, new) = rewrite(arg, "-fprofile-prefix-map=")?;
1105 opts.prefix_map.profile.push(old, new);
1106 }
1107 _ if arg.starts_with("-ffile-prefix-map=") => {
1108 let (old, new) = rewrite(arg, "-ffile-prefix-map=")?;
1109 opts.prefix_map.macros.push(old, new);
1110 opts.prefix_map.debug.push(old, new);
1111 opts.prefix_map.profile.push(old, new);
1112 }
1113 // A whole optimization rather than a flag, and the family is taken rather than
1114 // refused because of what ignoring it does. There is none of it here yet, so a build
1115 // that asks for it gets a program that is correct and slower than it could have been,
1116 // which is what section 4.1 means by a hint about speed and what every compilation at
1117 // `-O0` already is. The objects settle the rest of the argument: gcc's `-flto` object
1118 // holds the bytecode and no machine code at all, and every object here holds the code,
1119 // which is exactly what `-ffat-lto-objects` asks gcc for. So a build passing `-flto`
1120 // to this compiler gets objects that are more usable than the ones it asked for rather
1121 // than different ones. Every value is still checked against gcc's, because somebody
1122 // who wrote `-flto=thin` meant clang and had better hear about it here.
1123 "-flto" => opts.lto.requested = true,
1124 "-fno-lto" => opts.lto.requested = false,
1125 _ if arg.starts_with("-flto=") => {
1126 let how = &arg["-flto=".len()..];
1127 opts.lto.jobs = how.parse().map_err(|()| {
1128 err(format!(
1129 "`{how}` is not a number of link time jobs, which is auto, jobserver or a \
1130 count above zero"
1131 ))
1132 })?;
1133 opts.lto.requested = true;
1134 }
1135 _ if arg.starts_with("-flto-partition=") => {
1136 let how = &arg["-flto-partition=".len()..];
1137 opts.lto.partition = how.parse().map_err(|()| {
1138 err(format!(
1139 "`{how}` is not a partitioning model, which is balanced, 1to1, one, max \
1140 or none"
1141 ))
1142 })?;
1143 }
1144 _ if arg.starts_with("-flto-compression-level=") => {
1145 let how = &arg["-flto-compression-level=".len()..];
1146 let level =
1147 how.parse::<u8>().ok().filter(|level| *level <= 19).ok_or_else(|| {
1148 err(format!("`{how}` is not a compression level, 0 to 19"))
1149 })?;
1150 opts.lto.compression = Some(level);
1151 }
1152 // Whether the object keeps its machine code as well as the bytecode. It always does
1153 // here, so the first of these describes what happens and the second asks for an object
1154 // with less in it, which is a smaller file and not a different program, so both are
1155 // taken.
1156 "-ffat-lto-objects" | "-fno-fat-lto-objects" => {}
1157 // Whether the linker is handed a plugin that does the link time work. The design in
1158 // `spec/09-optimizer.md` has this driver doing that work itself and never loading a
1159 // plugin into anybody, so neither answer is a question it has to hold.
1160 "-fuse-linker-plugin" | "-fno-use-linker-plugin" => {}
1161 // Reading a profile back. Taken for the reason the family above it is: nothing here
1162 // reads one, so a build that asks gets the program it would have got anyway, and gcc
1163 // itself produces a byte for byte identical object from `-fprofile-use` when there are
1164 // no counts beside the file. The path is recorded for the pass that will read it. The
1165 // warning gcc prints when it looked and found nothing is deliberately not copied,
1166 // because nothing here looks, and a warning about a file that was never opened would
1167 // fire on the builds that have a perfectly good profile as well as on the ones that
1168 // do not.
1169 "-fprofile-use" => opts.profile_data.requested = true,
1170 "-fno-profile-use" => opts.profile_data.requested = false,
1171 _ if arg.starts_with("-fprofile-use=") => {
1172 opts.profile_data.path = Some(arg["-fprofile-use=".len()..].to_string());
1173 opts.profile_data.requested = true;
1174 }
1175 _ if arg.starts_with("-fprofile-dir=") => {
1176 opts.profile_data.dir = Some(arg["-fprofile-dir=".len()..].to_string());
1177 }
1178 "-fprofile-abs-path" => opts.profile_data.absolute = true,
1179 "-fno-profile-abs-path" => opts.profile_data.absolute = false,
1180 "-fprofile-correction" => opts.profile_data.correction = true,
1181 "-fno-profile-correction" => opts.profile_data.correction = false,
1182 "-fprofile-partial-training" => opts.profile_data.partial_training = true,
1183 "-fno-profile-partial-training" => opts.profile_data.partial_training = false,
1184 // Writing the counts rather than reading them, which is refused rather than taken and
1185 // is the same line `-gsplit-dwarf` falls on the far side of. Ignoring these means a
1186 // file a build declared as an output never appears: the instrumented program writes a
1187 // `.gcda` as it exits and `-ftest-coverage` writes a `.gcno` beside the object, and a
1188 // two stage build that got neither would go on to optimize against no counts at all
1189 // and report coverage of nothing, with nothing along the way saying so. The objects
1190 // say the rest: gcc's `-fprofile-generate` object holds 375 bytes of code where a
1191 // plain one holds 71, and 296 bytes of counters that a plain one does not have, so
1192 // this is a flag that changes the output rather than a hint about speed.
1193 "-fprofile-arcs"
1194 | "--coverage"
1195 | "-fcondition-coverage"
1196 | "-fpath-coverage"
1197 | "-fprofile-generate" => {
1198 return Err(err(format!(
1199 "{arg}: this compiler does not instrument for profiling, and a build that \
1200 expects the counts a run of the instrumented program writes would optimize \
1201 against nothing on its second pass, see spec/04-driver-and-cli.md"
1202 )));
1203 }
1204 _ if arg.starts_with("-fprofile-generate=") => {
1205 return Err(err(format!(
1206 "{arg}: this compiler does not instrument for profiling, and a build that \
1207 expects the counts a run of the instrumented program writes would optimize \
1208 against nothing on its second pass, see spec/04-driver-and-cli.md"
1209 )));
1210 }
1211 "-ftest-coverage" => {
1212 return Err(err(format!(
1213 "{arg}: this compiler writes no `.gcno` file beside the object, and a build \
1214 that expects one would wait for a file that never arrives, see \
1215 spec/04-driver-and-cli.md"
1216 )));
1217 }
1218 // The rest of the family describes instrumentation that is refused above, so what is
1219 // left to do with them is check them and drop them. They are checked because a
1220 // misspelling in a distribution's flags is worth finding here rather than on the day
1221 // the instrumentation lands, and dropped because there is nothing for an answer about
1222 // how a counter is written to be an answer about.
1223 _ if arg.starts_with("-fprofile-update=") => {
1224 let how = &arg["-fprofile-update=".len()..];
1225 if !matches!(how, "single" | "atomic" | "prefer-atomic") {
1226 return Err(err(format!(
1227 "`{how}` is not a profile update method, which is single, atomic or \
1228 prefer-atomic"
1229 )));
1230 }
1231 }
1232 _ if arg.starts_with("-fprofile-reproducible=") => {
1233 let how = &arg["-fprofile-reproducible=".len()..];
1234 if !matches!(how, "serial" | "parallel-runs" | "multithreaded") {
1235 return Err(err(format!(
1236 "`{how}` is not a profile reproducibility method, which is serial, \
1237 parallel-runs or multithreaded"
1238 )));
1239 }
1240 }
1241 "-fprofile-values" | "-fno-profile-values" | "-fprofile-info-section" => {}
1242 "-fno-test-coverage" | "-fno-profile-arcs" | "-fno-profile-generate" => {}
1243 _ if arg.starts_with("-fprofile-filter-files=")
1244 || arg.starts_with("-fprofile-exclude-files=")
1245 || arg.starts_with("-fprofile-note=") => {}
1246 // What every name gets when nothing in the source said, which the attribute in the
1247 // source overrides rather than the other way round. Before the optimizer's `-f`
1248 // family below for the reason the tier below it is.
1249 _ if arg.starts_with("-fvisibility=") => {
1250 let seen = &arg["-fvisibility=".len()..];
1251 opts.visibility = seen.parse().map_err(|()| {
1252 err(format!(
1253 "`{seen}` is not a visibility, which is default, hidden, internal or \
1254 protected"
1255 ))
1256 })?;
1257 }
1258 // Which edges of a control flow transfer are checked. Before the optimizer's `-f`
1259 // family below for the reason the two above it are, and last of the three so that the
1260 // bare spelling and the negative one are matched exactly rather than by this.
1261 _ if arg.starts_with("-fcf-protection=") => {
1262 let edges = &arg["-fcf-protection=".len()..];
1263 opts.control = edges.parse().map_err(|()| {
1264 err(format!(
1265 "`{edges}` is not a control flow protection, which is full, branch, \
1266 return, none or check"
1267 ))
1268 })?;
1269 }
1270 // How much room every function opens with for something to be written over later.
1271 // Before the optimizer's `-f` family below for the reason the ones above it are.
1272 _ if arg.starts_with("-fpatchable-function-entry=") => {
1273 let room = &arg["-fpatchable-function-entry=".len()..];
1274 opts.patchable = room.parse().map_err(|()| {
1275 err(format!(
1276 "`{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"
1277 ))
1278 })?;
1279 }
1280 // The memory safety monitor, from section 15.4 of
1281 // `spec/safe-memory/15-integration.md`. Before the optimizer's `-f` family below,
1282 // because a pass that took the name `safety=detect` would otherwise be handed the
1283 // flag, and the tier is not a pass.
1284 _ if arg.starts_with("-fsafety=") => {
1285 let tier = &arg["-fsafety=".len()..];
1286 opts.safety = tier.parse().map_err(|()| {
1287 err(format!(
1288 "`{tier}` is not a safety tier, which is off, detect, enforce or kernel"
1289 ))
1290 })?;
1291 }
1292 // Whether padding participates, from section 9.3 of document 09. Spelled out rather
1293 // than folded into the tier because it is a departure somebody who has read that
1294 // section makes, and the two defaults it describes are a property of what is being
1295 // built rather than of how much checking is wanted.
1296 _ if arg.starts_with("-fsafety-init=") => {
1297 let mode = &arg["-fsafety-init=".len()..];
1298 opts.padding = mode.parse().map_err(|()| {
1299 err(format!("`{mode}` is not a padding mode, which is padding or nopadding"))
1300 })?;
1301 }
1302 // Row S4, from section 9.4 of document 09. A bare flag with no value, because the
1303 // strict form of that section needs a member id the front end does not name yet and
1304 // accepting the spelling for it would be accepting a promise this build cannot keep.
1305 // Before `-fno-` is looked at below, for the reason the tier is.
1306 "-fsafety-subobject" => opts.subobject = rucc_session::Subobject::Members,
1307 "-fno-safety-subobject" => opts.subobject = rucc_session::Subobject::Off,
1308 _ if arg.starts_with("-fsafety-subobject=") => {
1309 let form = &arg["-fsafety-subobject=".len()..];
1310 return Err(err(format!(
1311 "`{form}` is not a form of -fsafety-subobject. The flag takes no value, and \
1312 the strict form of section 9.4 is tamnd/rucc#967"
1313 )));
1314 }
1315 // Row Y8, from section 9.6 of document 09. A bare flag with no value, for the reason
1316 // the one above has none: there is one form of this check and a spelling that suggested
1317 // otherwise would be promising something. Before `-fno-` is looked at below, the same
1318 // way.
1319 "-fsafety-restrict" => opts.promise = rucc_session::Promise::Blocks,
1320 "-fno-safety-restrict" => opts.promise = rucc_session::Promise::Off,
1321 _ if arg.starts_with("-fsafety-restrict=") => {
1322 let form = &arg["-fsafety-restrict=".len()..];
1323 return Err(err(format!(
1324 "`{form}` is not a form of -fsafety-restrict. The flag takes no value."
1325 )));
1326 }
1327 // Section 9.5's races, which take a value because the section gives them three modes
1328 // and the difference between two of them is which classes get reported rather than how
1329 // much is recorded. `-fno-` is the same as `=off` and is spelled out here for the same
1330 // reason the two above spell theirs out.
1331 _ if arg.starts_with("-fsafety-races=") => {
1332 let mode = &arg["-fsafety-races=".len()..];
1333 opts.races = mode.parse().map_err(|()| {
1334 err(format!("`{mode}` is not a race mode, which is off, metadata or pointer"))
1335 })?;
1336 }
1337 "-fno-safety-races" => opts.races = rucc_session::Races::Off,
1338 // The sanitizers of document 12, which are checks at run time rather than a way of
1339 // generating the same program. Each name is held to gcc 16's list, and what is still
1340 // asked for by the end of the line is answered after the loop, so that a command line
1341 // which turns one on and then off again is a command line that asked for nothing.
1342 //
1343 // Before the optimizer's `-f` family below, for the reason the tier above it is.
1344 _ if arg.starts_with("-fsanitize=") => {
1345 for one in arg["-fsanitize=".len()..].split(',') {
1346 if one == "all" {
1347 // gcc takes `all` only in the negative, because turning every check on at
1348 // once includes checks that contradict each other.
1349 return Err(err(
1350 "`-fsanitize=all` is not a gcc option, only `-fno-sanitize=all` is",
1351 ));
1352 }
1353 if !SANITIZERS.contains(&one) {
1354 return Err(err(format!(
1355 "`{one}` is not a sanitizer, see spec/04-driver-and-cli.md section 4.7"
1356 )));
1357 }
1358 if !sanitizers.contains(&one) {
1359 sanitizers.push(one);
1360 }
1361 }
1362 }
1363 _ if arg.starts_with("-fno-sanitize=") => {
1364 for one in arg["-fno-sanitize=".len()..].split(',') {
1365 if one == "all" {
1366 sanitizers.clear();
1367 continue;
1368 }
1369 if !SANITIZERS.contains(&one) {
1370 return Err(err(format!(
1371 "`{one}` is not a sanitizer, see spec/04-driver-and-cli.md section 4.7"
1372 )));
1373 }
1374 sanitizers.retain(|asked| *asked != one);
1375 }
1376 }
1377 // What a check does when it fires, and where the records about the checked objects go.
1378 // Each of them is an answer about the sanitizers refused after the loop, so there is
1379 // nothing left for them to change here. The names are still held to the list, because
1380 // a misspelling in a build's flags is worth finding when the compiler reads it.
1381 _ if arg.starts_with("-fsanitize-recover=")
1382 || arg.starts_with("-fno-sanitize-recover=")
1383 || arg.starts_with("-fsanitize-trap=")
1384 || arg.starts_with("-fno-sanitize-trap=") =>
1385 {
1386 // The guard above matched on a spelling that has an `=` in it, so the tail is
1387 // whatever follows the first one.
1388 let how = arg.split_once('=').map_or("", |(_, rest)| rest);
1389 for one in how.split(',') {
1390 if one != "all" && !SANITIZERS.contains(&one) {
1391 return Err(err(format!(
1392 "`{one}` is not a sanitizer, see spec/04-driver-and-cli.md section 4.7"
1393 )));
1394 }
1395 }
1396 }
1397 "-fsanitize-undefined-trap-on-error"
1398 | "-fsanitize-address-use-after-scope"
1399 | "-fno-sanitize-address-use-after-scope" => {}
1400 _ if arg.starts_with("-fsanitize-sections=") => {}
1401 // Counting which edges a run reached, which is how a fuzzer knows an input was worth
1402 // keeping. Refused rather than dropped, because a fuzzer whose calls into
1403 // `__sanitizer_cov_*` were never generated runs blind and reports coverage of nothing,
1404 // and there is no point in the campaign where that announces itself.
1405 _ if arg.starts_with("-fsanitize-coverage=") => {
1406 let how = &arg["-fsanitize-coverage=".len()..];
1407 for one in how.split(',') {
1408 if !matches!(one, "trace-pc" | "trace-cmp") {
1409 return Err(err(format!(
1410 "`{one}` is not a coverage instrumentation, which is trace-pc or \
1411 trace-cmp"
1412 )));
1413 }
1414 }
1415 return Err(err(format!(
1416 "{arg}: this compiler generates no coverage callbacks, and a fuzzer built \
1417 with it would run without any feedback at all, see \
1418 spec/04-driver-and-cli.md section 4.7"
1419 )));
1420 }
1421 // The optimizer's own flags, from section 9.10 of `spec/09-optimizer.md`. These come
1422 // after every `-f` the rest of the compiler answers to, so a pass can never take a
1423 // name that already means something else on the command line.
1424 _ if arg.starts_with("-fpass-fuel=") => {
1425 let (name, count) = arg["-fpass-fuel=".len()..]
1426 .split_once('=')
1427 .ok_or_else(|| err("-fpass-fuel= is spelled <pass>=<count>"))?;
1428 if rucc_opt::pass::find(name).is_none() {
1429 return Err(err(format!(
1430 "`{name}` is not a pass this compiler has, see --print-pipeline"
1431 )));
1432 }
1433 let count: u32 = count
1434 .parse()
1435 .map_err(|_| err(format!("`{count}` is not a number of transformations")))?;
1436 opts.pass_fuel.push((name.to_owned(), count));
1437 }
1438 _ if arg.starts_with("-fpass-fuel-global=") => {
1439 let count = &arg["-fpass-fuel-global=".len()..];
1440 let count: u32 = count
1441 .parse()
1442 .map_err(|_| err(format!("`{count}` is not a number of transformations")))?;
1443 opts.pass_fuel_global = Some(count);
1444 }
1445 // Everything from `-fopt-info` to the end of the argument, which is optional
1446 // keywords joined by hyphens and an optional `=<file>`. Checked here rather than
1447 // where the remarks are printed, because by then the compilation somebody wanted
1448 // to hear about is over.
1449 _ if arg == "-fopt-info"
1450 || arg.starts_with("-fopt-info=")
1451 || arg.starts_with("-fopt-info-") =>
1452 {
1453 let rest = &arg["-fopt-info".len()..];
1454 let (kinds, file) = match rest.split_once('=') {
1455 Some((kinds, file)) => (kinds, Some(file)),
1456 None => (rest, None),
1457 };
1458 let kinds = kinds.strip_prefix('-').unwrap_or(kinds);
1459 rucc_opt::Wants::none().add(kinds).map_err(err)?;
1460 opts.opt_info.push(kinds.to_owned());
1461 if let Some(file) = file {
1462 if file.is_empty() {
1463 return Err(err("-fopt-info= was given no file to write to"));
1464 }
1465 opts.opt_info_file = Some(file.to_owned());
1466 }
1467 }
1468 _ if arg.starts_with("-fdump-ir=") => {
1469 // Checked here rather than where the dumps are taken, because the compilation
1470 // that would have been dumped is over by then.
1471 let spec = &arg["-fdump-ir=".len()..];
1472 rucc_opt::Dumps::default().add(spec).map_err(err)?;
1473 opts.dump_ir.push(spec.to_owned());
1474 }
1475 // Before the bare `-f<pass>` below, because a pass called `enable-something` would
1476 // otherwise take the flag away from the gate. Checked here rather than where the
1477 // pipeline reads it, for the reason that applies to all of these: a misspelled pass
1478 // name that quietly gated nothing looks exactly like a pass that is not the guilty
1479 // one, and a bisection would carry on past the thing it was looking for.
1480 _ if arg.starts_with("-fdisable-") || arg.starts_with("-fenable-") => {
1481 let on = arg.starts_with("-fenable-");
1482 let spec = &arg[if on { "-fenable-".len() } else { "-fdisable-".len() }..];
1483 rucc_opt::Gates::default().add(on, spec).map_err(err)?;
1484 opts.pass_gates.push((on, spec.to_owned()));
1485 }
1486 // gcc's spelling for a pass this compiler has under a shorter name. It goes above the
1487 // two arms below rather than into the pile of gcc pass names further down, because the
1488 // pass is here: dropping the flag would leave a build that asked for unrolling without
1489 // it, and refusing it stops the build outright, which is what libtommath's makefile
1490 // ran into. `-funroll-all-loops` is deliberately not in here: gcc's is the one that
1491 // unrolls without a trip count, which is a different and usually worse thing.
1492 "-funroll-loops" => opts.passes.push(("unroll".to_owned(), true)),
1493 "-fno-unroll-loops" => opts.passes.push(("unroll".to_owned(), false)),
1494 _ if arg.strip_prefix("-fno-").is_some_and(|n| rucc_opt::pass::find(n).is_some()) => {
1495 opts.passes.push((arg["-fno-".len()..].to_owned(), false));
1496 }
1497 _ if arg.strip_prefix("-f").is_some_and(|n| rucc_opt::pass::find(n).is_some()) => {
1498 opts.passes.push((arg["-f".len()..].to_owned(), true));
1499 }
1500 // The flags that name a pass of gcc's own. They arrive from the torture suite, where a
1501 // program reduced from a miscompilation usually names the pass that miscompiled it on
1502 // its `dg-options` line, and they arrive from hand written build files for the same
1503 // reason. Section 4.1 sorts a flag by what the output would be without it, and by that
1504 // rule these are one pile: a flag that turns one of gcc's passes on or off is asking
1505 // for a compiler that does not exist here, and the program it is attached to is a
1506 // correctness test that passes either way. Turning on a pass we do not have costs
1507 // speed, turning off a pass we do not have costs nothing, and neither changes what the
1508 // program computes.
1509 //
1510 // rucc's own pass names are matched above this, so `-fno-dce` turns off the dce this
1511 // compiler has rather than landing here, and the day one of these names becomes a pass
1512 // here it stops being taken and dropped without anybody editing this list.
1513 //
1514 // Two of them are prefixes rather than names, which is the one place this file takes a
1515 // family instead of a flag. gcc files its gimple passes under `-ftree-` and its
1516 // interprocedural passes under `-fipa-`, both namespaces are pass selection and
1517 // nothing else, and there is no member of either that changes the meaning of a program
1518 // that was already correct. The rest are written out one at a time, because they live
1519 // in the flat `-f` namespace where the neighbours do change meanings.
1520 _ if arg.starts_with("-ftree-") || arg.starts_with("-fno-tree-") => {}
1521 _ if arg.starts_with("-fipa-") || arg.starts_with("-fno-ipa-") => {}
1522 "-fexpensive-optimizations" | "-fno-expensive-optimizations" => {}
1523 "-fmodulo-sched" | "-fno-modulo-sched" => {}
1524 "-fvect-cost-model" | "-fno-vect-cost-model" => {}
1525 _ if arg.starts_with("-fvect-cost-model=") || arg.starts_with("-fsimd-cost-model=") => {
1526 }
1527 "-fearly-inlining" | "-fno-early-inlining" => {}
1528 "-finline"
1529 | "-fno-inline"
1530 | "-finline-functions"
1531 | "-fno-inline-functions"
1532 | "-finline-small-functions"
1533 | "-fno-inline-small-functions"
1534 | "-finline-functions-called-once"
1535 | "-fno-inline-functions-called-once" => {}
1536 "-foptimize-strlen" | "-fno-optimize-strlen" => {}
1537 "-fira-share-spill-slots" | "-fno-ira-share-spill-slots" => {}
1538 // The charset flags are not in that pile, because an encoding is a statement about
1539 // what the bytes of the source mean rather than about how fast the output is. The
1540 // preprocessor reads UTF-8 and has no converter, so the one name that describes what
1541 // already happens is taken and every other name is refused. Spelled without regard to
1542 // case and with both of the spellings iconv answers to, since a build writes whichever
1543 // one its author typed.
1544 _ if arg.starts_with("-finput-charset=") => {
1545 let name = &arg["-finput-charset=".len()..];
1546 if !name.eq_ignore_ascii_case("utf-8") && !name.eq_ignore_ascii_case("utf8") {
1547 return Err(err(format!(
1548 "-finput-charset={name}: the preprocessor reads UTF-8 and has no \
1549 converter, so a file in another encoding would be read as though it were \
1550 UTF-8 rather than converted",
1551 )));
1552 }
1553 }
1554 // The three that come in on the same `dg-options` lines and are the other half of
1555 // section 4.1's rule, because each of them changes what the program does and not how
1556 // fast it does it. The negative form of each is what this compiler does anyway, so it
1557 // is taken and dropped, which is the shape `-fnested-functions` has above.
1558 "-ffast-math" => {
1559 return Err(err(
1560 "-ffast-math is a licence to answer a floating point arithmetic differently \
1561 from the way the source wrote it, and it is not one flag: it defines \
1562 __FAST_MATH__, which a library header reads, and gcc links a startup file \
1563 that puts the hardware in flush to zero mode for the whole process. Taking it \
1564 and dropping it would change what other objects in the same program answer. \
1565 -ffp-contract= and -fexcess-precision= are the parts of it this compiler has",
1566 ));
1567 }
1568 "-fno-fast-math" => {}
1569 "-fnon-call-exceptions" => {
1570 return Err(err(
1571 "-fnon-call-exceptions is a promise that an instruction which is not a call \
1572 can raise an exception the unwinder finds a handler for, and nothing here \
1573 produces a landing pad for a trapping instruction. A program built without it \
1574 would unwind past the handler it wrote",
1575 ));
1576 }
1577 "-fno-non-call-exceptions" => {}
1578 "-finstrument-functions" => {
1579 return Err(err(
1580 "-finstrument-functions calls __cyg_profile_func_enter on entry to every \
1581 function and __cyg_profile_func_exit on the way out, and nothing here emits \
1582 either call. A program that asks for them usually counts them, so taking the \
1583 flag and dropping it would turn a program that fails loudly into one that \
1584 fails quietly",
1585 ));
1586 }
1587 "-fno-instrument-functions" => {}
1588 // The unstable options, spelled the way rustc spells them and carrying the same
1589 // promise, which is none: one of these may change or go away in any release. They are
1590 // measurements and debugging aids rather than things a build asks for, which is why
1591 // none of them is in the usage text and all of them are in section 4.11 of
1592 // `spec/04-driver-and-cli.md`.
1593 "-Zverify-each" => opts.verify_each = true,
1594 _ if arg.starts_with("-Zrule-coverage=") => {
1595 let file = &arg["-Zrule-coverage=".len()..];
1596 if file.is_empty() {
1597 return Err(err("-Zrule-coverage= needs a file to write to"));
1598 }
1599 opts.rule_coverage = Some(file.to_owned());
1600 }
1601 _ if arg.starts_with("-Zcycle-accurate-model=") => {
1602 let value = &arg["-Zcycle-accurate-model=".len()..];
1603 opts.cycle_accurate_model = match value {
1604 "yes" | "1" => Some(true),
1605 "no" | "0" => Some(false),
1606 _ => {
1607 return Err(err("-Zcycle-accurate-model= takes yes or no"));
1608 }
1609 };
1610 }
1611 _ if arg.starts_with("-Zlowering=") => {
1612 let file = &arg["-Zlowering=".len()..];
1613 if file.is_empty() {
1614 return Err(err("-Zlowering= needs a file to write to"));
1615 }
1616 opts.lowering_dump = Some(file.to_owned());
1617 }
1618 _ if arg.starts_with("-Zregister-pressure=") => {
1619 let file = &arg["-Zregister-pressure=".len()..];
1620 if file.is_empty() {
1621 return Err(err("-Zregister-pressure= needs a file to write to"));
1622 }
1623 opts.register_pressure = Some(file.to_owned());
1624 }
1625 _ if arg.starts_with("-Z") => {
1626 return Err(err(format!(
1627 "`{arg}` is not an unstable option this compiler has, see \
1628 spec/04-driver-and-cli.md section 4.11 for the ones it does"
1629 )));
1630 }
1631 // The word size, which is a statement about the target and is taken as one. A build
1632 // that says the size the target already has is saying nothing, and one that says the
1633 // other size is asking for a target this compiler does not have, which it is told
1634 // rather than being given the wrong one.
1635 "-m64" | "-m32" | "-mx32" => {
1636 let want: u32 = match arg {
1637 "-m64" => 64,
1638 _ => 32,
1639 };
1640 let have = rucc_target::TargetInfo::new(opts.target).pointer_width;
1641 if have != want {
1642 return Err(err(format!(
1643 "{arg} asks for a {want} bit target and {} is {have} bit, use \
1644 --target= to name the one you mean",
1645 opts.target
1646 )));
1647 }
1648 }
1649 // Which processor in the family to generate for. This compiler emits the base
1650 // instruction set of the architecture and nothing above it, so a program built with
1651 // any of these runs on the machine that was named; it is a program that could have
1652 // been faster rather than a program that is wrong, which is what makes these safe to
1653 // take and ignore where a flag that changed the meaning of the code would not be.
1654 _ if arg.starts_with("-march=")
1655 || arg.starts_with("-mtune=")
1656 || arg.starts_with("-mcpu=") => {}
1657 // The calling convention, which is not safe to ignore. Taken when it names the one
1658 // the target already uses and refused otherwise.
1659 _ if arg.starts_with("-mabi=") => {
1660 let want = &arg["-mabi=".len()..];
1661 let have = match opts.target.arch {
1662 rucc_target::Arch::X86_64 => "sysv",
1663 rucc_target::Arch::Aarch64 => "lp64",
1664 rucc_target::Arch::Riscv64 => "lp64d",
1665 };
1666 if want != have {
1667 return Err(err(format!(
1668 "{arg}: {} uses the {have} convention and this compiler has no other",
1669 opts.target
1670 )));
1671 }
1672 }
1673 // How far apart the pieces of the program may be. The small model is what we emit and
1674 // it is every hosted program's default; the kernel model is a different one and a
1675 // build that asks for it and does not get it links and then does not run.
1676 "-mcmodel=small" => {}
1677 _ if arg.starts_with("-mcmodel=") => {
1678 return Err(err(format!(
1679 "{arg}: this compiler emits the small code model and no other, see \
1680 spec/12-targets.md"
1681 )));
1682 }
1683 // GCC's own scripting language for how the driver builds a command line.
1684 // `spec/04-driver-and-cli.md` section 4.4 settles that we will not have it, so a
1685 // build reaching for it is told which flags do the same job.
1686 _ if arg.starts_with("-specs=") => {
1687 return Err(err(
1688 "-specs= is not supported: the parts of it builds rely on are -B, -L, \
1689 -nostdlib, -nostartfiles and -Wl,, see spec/04-driver-and-cli.md \
1690 section 4.4",
1691 ));
1692 }
1693 // Arguments meant for a separate assembler or preprocessor, which this compiler does
1694 // not have: both are inside it and neither reads a command line. Refused rather than
1695 // dropped, because every one of these says something about the output and a build
1696 // that asked for `-Wa,--noexecstack` and was silently given an executable stack got
1697 // the opposite of what it asked for.
1698 _ if arg.starts_with("-Wa,") || arg.starts_with("-Wp,") => {
1699 return Err(err(format!(
1700 "`{arg}` is an argument for a separate assembler or preprocessor, and both \
1701 are inside this compiler rather than programs it runs"
1702 )));
1703 }
1704 "-Xassembler" | "-Xpreprocessor" => {
1705 return Err(err(format!(
1706 "{arg} hands an argument to a separate assembler or preprocessor, and both \
1707 are inside this compiler rather than programs it runs"
1708 )));
1709 }
1710 // Everything else in the `-W` family. `spec/04-driver-and-cli.md` section 4.1 has
1711 // this one as a rule about build systems rather than about warnings: autoconf finds
1712 // out whether a warning flag exists by passing it and looking at the exit status, so
1713 // a compiler that refuses one it has not heard of fails a configure script written
1714 // for a GCC newer than itself. The names are not checked against a list because this
1715 // compiler has no warning groups for a list to be of, which #485 is about.
1716 _ if arg.starts_with("-W") => {}
1717 // Flags that name something this compiler does not do and would not do differently
1718 // if it did. `-fno-ident` is about a comment in the output that we do not write
1719 // either way, and the others are about a way of ordering the compilation that has
1720 // been GCC's only way for twenty years. Section 4.1 asks for the list to be short
1721 // and for adding to it to be deliberate, which is why it is written out here.
1722 "-fno-ident"
1723 | "-fident"
1724 | "-funit-at-a-time"
1725 | "-fno-unit-at-a-time"
1726 | "-shared-libgcc"
1727 | "-static-libgcc" => {}
1728 _ if arg.starts_with('-') && arg.len() > 1 => {
1729 // Silently ignoring an unknown flag is how a build ends up not doing what
1730 // its author asked. spec/13-gnu-compat.md section 13.4 makes this an error
1731 // for the flags that change code generation, and the safe default until the
1732 // flag table is populated is to reject everything we do not know.
1733 return Err(err(format!("unknown option `{arg}`")));
1734 }
1735 _ => inputs.push(Input { path: arg.to_owned(), forced, role: Role::File }),
1736 }
1737 }
1738
1739 // The fetch, before anything that resolves a compilation, because `--fetch` does not describe
1740 // one. It is here rather than in the loop so that `--offline` can forbid it whichever order the
1741 // two were written in, and it is before the refusals below so that a command line asking for a
1742 // sysroot is not told about a sanitizer.
1743 if let Some(named) = fetch {
1744 return fetch_action(&named, offline, &inputs);
1745 }
1746
1747 // Last, so that it lands after every `-isystem` the command line gave. That is GCC's
1748 // order: a directory the user names outranks the compiler's own, and the compiler's own
1749 // outranks the library's. It is pushed after the loop rather than before it because
1750 // `SearchPath` appends within a group and the position is what the order is.
1751 // The same directory the headers were looked for under, because a sysroot is a statement
1752 // about a whole installation and not about half of one.
1753 // After the loop, because `-fno-sanitize=` can take back what an earlier flag asked for and a
1754 // command line that turns a check on and off again has asked for nothing. What is left is
1755 // refused rather than dropped, and it is the one place in this parser where the reason is not
1756 // that the output would differ. A sanitizer is a promise that the program is watched while it
1757 // runs, so a build that asks for one and is quietly given a program with no checks in it does
1758 // not get a slower program or a bigger file, it gets a test suite that passes for the wrong
1759 // reason. `-fsafety=` is the checking this compiler does have, and the message says so, because
1760 // somebody reaching for `-fsanitize=address` wants the nearest thing rather than a list of
1761 // options.
1762 if let Some(first) = sanitizers.first() {
1763 return Err(err(format!(
1764 "-fsanitize={first}: this compiler has no sanitizer instrumentation, and a build that \
1765 asked for one and got none would run its tests unchecked, see \
1766 spec/04-driver-and-cli.md section 4.7. `-fsafety=detect` is the memory checking this \
1767 compiler does have"
1768 )));
1769 }
1770 link.sysroot = sysroot.clone();
1771 // Where a sysroot for a target that is not this machine would be. Read once, here, rather than
1772 // inside the link line, because a link line that read the environment could only be tested on a
1773 // machine whose environment said the right thing, and the link line is the last thing that
1774 // touches a binary. `spec/cross-compile/13-distribution.md` section 13.2 owns the answer.
1775 link.cache = Some(cache::dir());
1776 // And the ten field spelling of the target, because the release on it decides two things the
1777 // three field one cannot say: whether a target that is this architecture is still a cross
1778 // compile, and which directory under the cache it is against. After the loop because the last
1779 // `--target=` on the command line is the one that counts.
1780 link.pinned = pinned;
1781 // After the loop rather than where `-pthread` was read, so that it lands after the objects
1782 // that refer to it. A static link takes the definitions it needs from a library when it
1783 // reaches it and not afterwards, so a library before the objects is a library that answers
1784 // nothing.
1785 if threads {
1786 inputs.push(Input::library("pthread"));
1787 }
1788 if let Some(query) = query {
1789 return Ok(Action::Print(answer(&query, &opts, &link)?));
1790 }
1791 // `-M` and `-MM` produce the rule and nothing else, so the run stops after phase 4 whatever
1792 // else the command line asked for. Read here rather than where the flag was, because a `-c`
1793 // written after it has to lose and the loop cannot know that until it has ended. The output
1794 // file is where the rule goes rather than where an object would have gone, and the last
1795 // phase being the preprocessor is what makes that true without a second rule for it.
1796 if opts.deps.instead_of_compiling {
1797 opts.emit = EmitKind::Preprocessed;
1798 }
1799 if !nostdinc {
1800 opts.search.push_system(runtime::DIR);
1801 // And the library's after ours, which is the other half of the same order. They go on
1802 // here rather than at the point `--target=` or `--sysroot=` was read because either
1803 // one changes the answer and the last word on both is the end of the loop.
1804 //
1805 // Which library's is the question `link::cross_sysroot` answers, and it is asked here so
1806 // that the headers and the libraries come from the same place. A target that is this
1807 // machine reads this machine's headers, and a target that is not reads the ones in the
1808 // sysroot for it rather than the ones next door.
1809 let cross = link::cross_sysroot(opts.target, &link);
1810 let kernel = link::cross_kernel(opts.target, &link);
1811 // And the version of those headers, which only the bundled tree has an answer for. A host
1812 // glibc and a tree the user named both define `__GLIBC_MINOR__` in their own `features.h`,
1813 // and a second definition with a different value is a warning on every file, so the
1814 // condition is the same one that chose the directories.
1815 if cross.is_some() {
1816 let target = pinned.unwrap_or_else(|| opts.target.tuple());
1817 opts.glibc_minor = rucc_sysroot::bundled_glibc_minor(target).map_err(|skew| {
1818 err(format!(
1819 "{skew}; pin a release the tree has, or name a tree that has that one \
1820 with --sysroot"
1821 ))
1822 })?;
1823 }
1824 let system =
1825 library::header_dirs(opts.target, sysroot.as_deref(), cross.as_ref(), kernel.as_ref());
1826 // The two licence walls of `spec/cross-compile/13-distribution.md` section 13.4, which are
1827 // the only way step 3 comes back with nothing on a hosted target. Section 8.6 asks for the
1828 // answer to name the licence and the lawful ways to get what is behind it, rather than
1829 // leaving a person with an `#include` that failed as though a directory had gone missing.
1830 //
1831 // It is left on the search path instead of refused here, because a program that includes
1832 // none of the library needs none of the SDK and section 8.6 is explicit that targeting the
1833 // platform has to keep working. So the reason waits until an include has actually failed,
1834 // which is the only moment it helps and the only moment it is true.
1835 //
1836 // The condition is that step 3 found nothing at all, so an `SDKROOT`, an `INCLUDE` or a mac
1837 // with Xcode on it all pass through untouched, and `-nostdinc` never reaches this block. A
1838 // `--sysroot` or `-isysroot` passes through as well, even when the tree it names turns out to
1839 // be empty or absent: somebody who wrote a path has already answered the question this
1840 // message asks, and answering it again over the top of a mistyped directory would hide the
1841 // mistake behind a licence notice.
1842 if system.is_empty() && sysroot.is_none() {
1843 let tuple = pinned.unwrap_or_else(|| opts.target.tuple());
1844 if let Some(wall) = rucc_sysroot::Wall::of(tuple) {
1845 opts.search.explain_missing_system(wall.no_headers(&tuple.to_canonical_string()));
1846 }
1847 }
1848 // And whether the tree somebody named is the release they asked for, which is the one
1849 // question left once the directories are settled and the only place both halves of it are
1850 // known. Only for a named tree, because that is the case where the release in the target
1851 // stops deciding anything, and `crate::glibc` is where the rest of the reasoning is.
1852 if sysroot.is_some() {
1853 notes.extend(glibc::skew(opts.target, pinned, &system));
1854 }
1855 for dir in system {
1856 opts.search.push_system(dir);
1857 }
1858 }
1859 // Once, here, rather than as each directory is pushed. A `-I` that names a system
1860 // directory has to lose to the system entry and the system entry is added last, so the
1861 // question cannot be answered until the whole path is known.
1862 opts.search.remove_duplicates();
1863
1864 // The target has to be resolved before the configuration is printed, so this check comes
1865 // after the loop rather than at the point `--print-config` was seen.
1866 if print_config {
1867 return Ok(Action::PrintConfig(Box::new(opts)));
1868 }
1869 if print_pipeline {
1870 return Ok(Action::PrintPipeline(Box::new(opts)));
1871 }
1872 let plan = Plan::new(&opts, &inputs, output.as_deref()).map_err(|e| err(e.message))?;
1873 if print_plan {
1874 return Ok(Action::PrintPlan {
1875 opts: Box::new(opts),
1876 plan: Box::new(plan),
1877 link: Box::new(link),
1878 });
1879 }
1880 Ok(Action::Compile {
1881 opts: Box::new(opts),
1882 plan: Box::new(plan),
1883 link: Box::new(link),
1884 jobs,
1885 verbose,
1886 notes,
1887 })
1888}
1889
1890/// What `--fetch <tuple>` asked for, or why it is not a thing that can be done.
1891///
1892/// The lookup happens here rather than at the point the bytes would move, so that a target this
1893/// release pins nothing for is a refusal from the parser and the only code that runs a downloader is
1894/// code that already knows what it is getting.
1895///
1896/// # Errors
1897///
1898/// [`CliError`] when `--offline` forbade it, when there are input files as well, when the tuple is
1899/// not a target this compiler knows, when its sysroot is behind one of section 13.4's licence walls,
1900/// and when this release pins no artifact for it.
1901fn fetch_action(named: &str, offline: bool, inputs: &[Input]) -> Result<Action, CliError> {
1902 // Not a precedence question. Section 13.2 says `--offline` forbids a fetch entirely, so a
1903 // command line that writes both has asked for two opposite things and the answer is to say so
1904 // rather than to pick one of them.
1905 if offline {
1906 return Err(err(
1907 "--fetch asks for a download and --offline forbids every download, so this command \
1908 line asks for two opposite things. Drop one of them: --offline is how a build says it \
1909 will not reach the network, and --fetch is the only thing in this compiler that does",
1910 ));
1911 }
1912 if let Some(first) = inputs.first() {
1913 return Err(err(format!(
1914 "--fetch gets a sysroot and compiles nothing, so `{}` on the same command line is an \
1915 input that nothing would read",
1916 first.path
1917 )));
1918 }
1919 let target: TargetTuple = named
1920 .parse()
1921 .map_err(|why| err(format!("--fetch {named}: {why}, so there is no sysroot to get")))?;
1922 // The canonical spelling, because that is what a row is named by and what the directory under
1923 // the cache is called, and a person is free to write a tuple the long way round.
1924 let tuple = target.to_canonical_string();
1925 // Before the table is consulted, because a target behind a licence wall is not a row that has not
1926 // been written yet. Section 13.4 is that no release pins one of these ever, so the message says
1927 // the licence and the two lawful ways rather than naming the producer that will publish the rest.
1928 if let Some(wall) = rucc_sysroot::Wall::of(target) {
1929 return Err(err(format!("--fetch {tuple}: {}", wall.no_fetch(&tuple))));
1930 }
1931 let Some(what) = rucc_sysroot::pinned_for(&tuple) else {
1932 return Err(err(unpinned(&tuple)));
1933 };
1934 Ok(Action::Fetch { what, target, cache: cache::dir() })
1935}
1936
1937/// Why there is nothing to fetch for a target, which is a different sentence when the table is
1938/// empty.
1939///
1940/// A release that pins nothing and a release that pins eleven targets and not this one are two
1941/// situations, and a message that did not tell them apart would send somebody looking for a typo in
1942/// their tuple when the answer is that this work is not finished.
1943fn unpinned(tuple: &str) -> String {
1944 let pinned = rucc_sysroot::pinned_targets();
1945 if pinned.is_empty() {
1946 return format!(
1947 "this release pins no sysroot for {tuple}, and it pins none for any target yet. A \
1948 sysroot is built and published by the producer in tamnd/rucc-cross, per \
1949 spec/cross-compile/13-distribution.md section 13.8, and a release of this compiler \
1950 names one by URL and by hash afterwards. Until then, pass --sysroot=<dir> to compile \
1951 against a tree you have already"
1952 );
1953 }
1954 format!(
1955 "this release pins no sysroot for {tuple}. What it pins is {}. Pass --sysroot=<dir> to \
1956 compile against a tree you have already",
1957 pinned.join(", ")
1958 )
1959}
1960
1961/// Gets the artifact and installs it, saying what each step did.
1962///
1963/// The steps are section 13.8's and so are the messages: the transport is somebody else's program
1964/// and the check is ours, so a person reading this wants to know which downloader ran, that the
1965/// bytes matched, how many files the record named and where the tree ended up. A fetch of something
1966/// that is already there says that instead and moves nothing.
1967fn fetch_sysroot(what: &rucc_sysroot::Pinned, target: TargetTuple, cache: &std::path::Path) -> i32 {
1968 let tuple = target.to_canonical_string();
1969 let archive = what.archive_in(cache);
1970 let say = |line: &str| println!("rucc: {tuple}: {line}");
1971 match fetch::fetch(what.url, what.sha256, &archive) {
1972 Ok(fetch::Fetched::AlreadyThere) => {
1973 say(&format!("{} is already here and matches the hash", archive.display()));
1974 }
1975 Ok(fetch::Fetched::Downloaded(by)) => {
1976 say(&format!("downloaded {} with {}", what.url, by.program()));
1977 }
1978 Err(why) => return complain(why),
1979 }
1980 match install::install(&archive, what.sha256, target, cache) {
1981 Ok(done) => {
1982 match &done.before {
1983 install::Before::Nothing => {
1984 say(&format!("{} files installed at {}", done.files, done.root.display()));
1985 }
1986 install::Before::TheSame => {
1987 say(&format!(
1988 "the same sysroot is already at {}, so nothing moved",
1989 done.root.display()
1990 ));
1991 }
1992 install::Before::Different(was) => {
1993 say(&format!(
1994 "{} files installed at {}, over a tree whose record digested to {was}",
1995 done.files,
1996 done.root.display()
1997 ));
1998 }
1999 }
2000 say(&format!("the record digests to {}", done.digest));
2001 0
2002 }
2003 Err(why) => complain(why),
2004 }
2005}
2006
2007/// What one of the `-dump` and `-print` flags prints.
2008///
2009/// GCC prints the name back unchanged when it cannot find the file a `-print` flag asked about,
2010/// which is what makes the answer safe to paste into a link line whether or not the file is
2011/// there, and this does the same.
2012fn answer(query: &Query, opts: &Options, link: &LinkOptions) -> Result<String, CliError> {
2013 let found = |name: &str| {
2014 link::find_in_search(link, opts.target, name)
2015 .map_or_else(|| name.to_owned(), |path| path.display().to_string())
2016 };
2017 Ok(match query {
2018 Query::Machine => opts.target.to_string(),
2019 Query::Version => VERSION.to_owned(),
2020 Query::Multiarch => link::multiarch(opts.target),
2021 // The three lines GCC prints, in its order and with its punctuation, because what reads
2022 // them is a script written against that shape. There is no installation directory to
2023 // report: this compiler is one binary that works wherever it is copied, and the headers
2024 // it ships are inside it, so `install` is where the binary is and nothing is under it.
2025 Query::SearchDirs => {
2026 let here = std::env::current_exe()
2027 .ok()
2028 .and_then(|p| p.parent().map(std::path::Path::to_path_buf))
2029 .unwrap_or_default();
2030 let list = |dirs: &[PathBuf]| {
2031 dirs.iter().map(|d| d.display().to_string()).collect::<Vec<_>>().join(":")
2032 };
2033 let libraries = link::search_dirs(link, opts.target);
2034 format!(
2035 "install: {}\nprograms: ={}\nlibraries: ={}",
2036 here.display(),
2037 list(&link.prefixes),
2038 list(&libraries)
2039 )
2040 }
2041 // The root the rest of the answers are under, which a build system asks for when it wants
2042 // to find a file itself rather than ask for one by name, and which is the first thing to
2043 // look at when a cross build read a header nobody expected. A native compile has no
2044 // sysroot and the answer is the empty line, which is what GCC prints when it was
2045 // configured without one. `--sysroot` wins over ours because it wins everywhere else.
2046 Query::Sysroot => {
2047 sysroot_root(opts, link).map(|root| root.display().to_string()).unwrap_or_default()
2048 }
2049 // Section 13.5 of `spec/cross-compile/13-distribution.md`: for every input that is not this
2050 // compiler's own code, what it is, where it was got, its hash, its licence and whether it
2051 // was bundled, generated or fetched. What is printed is the manifest the sysroot already
2052 // carries rather than a second format saying the same things, because the three uses 13.5
2053 // gives for this are a licence notice, a reproducibility check and a security audit, and all
2054 // three are somebody else parsing it. One format is one parser to write.
2055 // Read and rendered rather than copied out, so that what comes back is the format this
2056 // build understands. The last newline comes off because whatever prints an answer adds
2057 // one, the way it does for every other query here. Keeping it would put a blank line at
2058 // the end of the one answer that is a file somebody diffs against the file it came from.
2059 Query::SysrootProvenance => match sysroot_manifest(opts, link)? {
2060 Some(manifest) => manifest.render().trim_end_matches('\n').to_string(),
2061 None => String::new(),
2062 },
2063 // Section 13.2 of the same document, which asks for the hash of a cache directory's
2064 // contents in the directory's name. A name cannot carry one, because the path has to be
2065 // computable before anything has been read, by the producer about to write the files and by
2066 // the compiler about to read them, and neither has the contents when it asks. So the number
2067 // is here instead, and it is the sha256 of the record rather than of a walk of the tree,
2068 // which means `sha256sum` over the manifest answers the same thing.
2069 Query::SysrootDigest => match sysroot_manifest(opts, link)? {
2070 Some(manifest) => manifest.digest(),
2071 None => String::new(),
2072 },
2073 Query::FileName(name) => found(name),
2074 // The name GCC gives the library of routines a compiler's output calls that the C
2075 // library does not have. Ours is built in and there is no file, so the answer is the
2076 // name itself, which is what GCC prints when it cannot find one either.
2077 Query::Libgcc => found("libgcc.a"),
2078 // A program rather than a library: the linker and the archiver are the ones a build asks
2079 // about, and this compiler finds them on the path or under `-B` rather than shipping
2080 // them, so the name back is the honest answer unless a `-B` prefix holds one.
2081 Query::ProgName(name) => link
2082 .prefixes
2083 .iter()
2084 .map(|dir| dir.join(name))
2085 .find(|path| path.is_file())
2086 .map_or_else(|| name.clone(), |path| path.display().to_string()),
2087 })
2088}
2089
2090/// The root every sysroot answer is about.
2091///
2092/// One function rather than a copy in each, because the other flags exist to say what is inside the
2093/// tree this one names, and two answers that disagreed about which tree that is would be a
2094/// difference nobody would think to look for. `--sysroot` wins over ours because it wins everywhere
2095/// else.
2096fn sysroot_root(opts: &Options, link: &LinkOptions) -> Option<PathBuf> {
2097 link.sysroot
2098 .clone()
2099 .or_else(|| link::cross_sysroot(opts.target, link).map(|at| at.root().to_path_buf()))
2100}
2101
2102/// The record of the sysroot this command line reads, when there is one to read.
2103///
2104/// [`None`] covers two cases that both print nothing, and they are different things. A compile for
2105/// this machine has no sysroot at all, and a tree somebody laid out themselves and pointed
2106/// `--sysroot` at carries no manifest, so nothing here knows where any of it came from. Saying
2107/// nothing is the only honest answer to either, and a reader can tell it from a manifest with no
2108/// inputs in it because that one still has its header lines.
2109///
2110/// # Errors
2111///
2112/// A manifest this build cannot parse, and anything else that went wrong reading the file. Passing a
2113/// record we could not read on to whoever asked would make their parser the one that finds the
2114/// problem, and every use section 13.5 gives for these two flags is somebody else reading the
2115/// output.
2116fn sysroot_manifest(opts: &Options, link: &LinkOptions) -> Result<Option<Manifest>, CliError> {
2117 let Some(root) = sysroot_root(opts, link) else {
2118 return Ok(None);
2119 };
2120 let path = Sysroot::at(root, opts.target.tuple()).manifest_path();
2121 match std::fs::read_to_string(&path) {
2122 Ok(text) => Manifest::parse(&text)
2123 .map(Some)
2124 .map_err(|why| err(format!("{}: {why}", path.display()))),
2125 Err(why) if why.kind() == std::io::ErrorKind::NotFound => Ok(None),
2126 Err(why) => Err(err(format!("{}: {why}", path.display()))),
2127 }
2128}
2129
2130/// Renders the passes this level will run, in order, with what each one does.
2131///
2132/// The level is the whole of the answer unless a `-f` flag edited it, which is section 9.1 of
2133/// `spec/09-optimizer.md`: a level is a list somebody wrote down rather than something that
2134/// emerges from which flags happen to be set, and this is how that list is read.
2135#[must_use]
2136pub fn print_pipeline(opts: &Options) -> String {
2137 let mut settings = rucc_opt::Options::for_level(opts.opt_level);
2138 settings.toggles.clone_from(&opts.passes);
2139 settings.global_fuel = opts.pass_fuel_global;
2140 for (on, spec) in &opts.pass_gates {
2141 // Every spelling was checked while the arguments were parsed, so there is nothing here
2142 // this can refuse, and a listing is not the place to report it if there were.
2143 let _ = settings.gates.add(*on, spec);
2144 }
2145 rucc_opt::pipeline::print(&settings)
2146}
2147
2148/// Renders the resolved configuration.
2149///
2150/// One `key: value` per line, sorted by nothing in particular but fixed in order, because
2151/// this output is diffed across hosts in CI and a reordering would read as a change.
2152#[must_use]
2153pub fn print_config(opts: &Options) -> String {
2154 let sess = Session::new(opts.clone());
2155 let t = &sess.target;
2156 let mut out = String::new();
2157 let _ = writeln!(out, "version: {VERSION}");
2158 // The three field triple the driver was given rather than the ten field tuple it widens to,
2159 // because this output is what a build system reads to find out what it asked for. The tuple is
2160 // the compiler's model of the machine and this line is a receipt for a command line.
2161 let _ = writeln!(out, "target: {}", opts.target);
2162 let _ = writeln!(out, "arch: {}", opts.target.arch.as_str());
2163 let _ = writeln!(out, "os: {}", opts.target.os.as_str());
2164 let _ = writeln!(out, "env: {}", opts.target.env.as_str());
2165 let _ = writeln!(out, "object-format: {}", t.object_format.as_str());
2166 let _ = writeln!(out, "pointer-width: {}", t.pointer_width);
2167 let _ = writeln!(out, "long-width: {}", t.long_width);
2168 let _ = writeln!(out, "long-double-width: {}", t.long_double_width);
2169 let _ = writeln!(out, "endian: {}", if t.little_endian { "little" } else { "big" });
2170 let _ = writeln!(out, "char-signed: {}", t.char_is_signed);
2171 let _ = writeln!(out, "va-list: {}", t.va_list.map_or("none", |list| list.as_str()));
2172 // The register file as a count per class, which is enough to tell a target whose registers
2173 // are described from one whose are not without printing sixteen names nobody asked for.
2174 let regs: Vec<String> = t
2175 .regs
2176 .classes()
2177 .map(|(class, info)| format!("{} {}", info.name, t.regs.len(class)))
2178 .collect();
2179 let _ = writeln!(
2180 out,
2181 "registers: {}",
2182 if regs.is_empty() { "none".to_string() } else { regs.join(", ") }
2183 );
2184 // What the schedule was chosen with, which is a sentence rather than a name on purpose: two
2185 // runs of a benchmark that disagree are usually two models and not two compilers.
2186 let _ = writeln!(out, "timing-model: {}", t.timing.map_or("none", |timing| timing.model));
2187 let _ = writeln!(out, "opt-level: {}", sess.opts.opt_level);
2188 let _ = writeln!(out, "safety: {}", sess.opts.safety);
2189 let _ = writeln!(out, "emit: {}", sess.opts.emit.as_str());
2190 let _ = writeln!(out, "debug-info: {}", sess.opts.debug_info);
2191 let _ = writeln!(out, "frame-pointer: {}", sess.opts.frame_pointer);
2192 let _ = writeln!(out, "red-zone: {}", sess.opts.red_zone);
2193 let _ = writeln!(out, "stack-protector: {}", sess.opts.protector);
2194 let _ = writeln!(out, "stack-clash-protection: {}", sess.opts.stack_clash);
2195 let _ = writeln!(out, "cf-protection: {}", sess.opts.control);
2196 let _ = writeln!(out, "patchable-function-entry: {}", sess.opts.patchable);
2197 let _ = writeln!(out, "profile: {}", sess.opts.profile);
2198 let _ = writeln!(out, "profile-hook: {}", sess.opts.hook);
2199 // Last because it is the one key with more than one line under it, and the only one
2200 // whose value is a property of the machine rather than of the command line.
2201 for dir in sess.opts.search.dirs() {
2202 let system = if dir.is_system { " (system)" } else { "" };
2203 let _ = writeln!(out, "include: {}{system}", dir.path.display());
2204 }
2205 out
2206}
2207
2208/// The output name the make target is taken from, which is the `-o` argument or nothing.
2209///
2210/// A run that stops at the preprocessor has not named an object, whatever its `-o` says: under
2211/// `-E` that argument is the preprocessed text and under `-M` it is the rule itself, and neither
2212/// is a file `make` would rebuild by running this rule. GCC agrees and falls back to the source
2213/// name in both, which is why a `-MD -E -o out.i` writes `out.d` holding a rule for `a.o`. From
2214/// `-S` on the argument does name what the rule builds, and it is used as written.
2215fn deps_target_output<'a>(opts: &Options, plan: &'a Plan) -> Option<&'a str> {
2216 if opts.emit == EmitKind::Preprocessed { None } else { plan.output.as_deref() }
2217}
2218
2219/// Writes to a path the command line named rather than one the plan derived, where `-` is
2220/// standard output.
2221fn write_named(path: &str, bytes: &[u8]) -> Result<(), String> {
2222 if path == "-" {
2223 return write_out(&Output::Stdout, bytes);
2224 }
2225 write_out(&Output::File(path.to_owned()), bytes)
2226}
2227
2228/// Writes the make rule for one input, and reports whether it got there.
2229///
2230/// A rule with no file of its own goes where the compilation it replaced would have written,
2231/// which is what makes the usual makefile recipe work: `rucc -M $< -o $@` leaves the rule in
2232/// `$@`, and the same line with the `-o` left off puts it on standard output.
2233fn write_deps(
2234 opts: &Options,
2235 plan: &Plan,
2236 job: &Job,
2237 found: &[Dependency],
2238 stderr: &mut impl std::io::Write,
2239) -> bool {
2240 let targets = if opts.deps.targets.is_empty() {
2241 vec![deps::default_target(&job.input, deps_target_output(opts, plan))]
2242 } else {
2243 opts.deps.targets.clone()
2244 };
2245 let rule = deps::rule(&opts.deps, &targets, &job.input, found);
2246 // The file, on the other hand, is named after the `-o` in every mode that still has one to
2247 // spend, which is every mode except the two that spend it on the rule.
2248 let wrote = match deps::default_file(&opts.deps, &job.input, plan.output.as_deref()) {
2249 // A `-MF` on a run that had nowhere else to put the rule leaves the file the `-o`
2250 // named empty rather than absent, because a makefile that named it as a target of its
2251 // own is a makefile that will look for it.
2252 Some(path) => write_named(&path, rule.as_bytes()).and_then(|()| {
2253 if opts.deps.instead_of_compiling { write_out(&job.output, b"") } else { Ok(()) }
2254 }),
2255 None => write_out(&job.output, rule.as_bytes()),
2256 };
2257 if let Err(e) = wrote {
2258 let _ = writeln!(stderr, "rucc: error: {e}");
2259 return false;
2260 }
2261 true
2262}
2263
2264/// Runs phase 4 over every input that has one, and writes what came out.
2265///
2266/// One input that fails does not stop the others. A build that reports every file it could
2267/// not preprocess in one run is worth more than one that stops at the first, and the exit
2268/// status is still a failure either way.
2269fn preprocess_all(opts: &Options, plan: &Plan) -> i32 {
2270 let fs = OsFileSystem::new();
2271 let mut stderr = std::io::stderr().lock();
2272 let mut failed = false;
2273 for job in &plan.jobs {
2274 if !job.phases.first().is_some_and(|p| *p == Phase::Preprocess) {
2275 // An input that is already preprocessed, or an object file. GCC passes these
2276 // through untouched, and the plan has already said so in its notes.
2277 continue;
2278 }
2279 let started = std::time::Instant::now();
2280 let result = preprocess(opts, &job.input, &fs);
2281 if opts.time {
2282 say_time(&job.input, started.elapsed(), &mut stderr);
2283 }
2284 for message in &result.messages {
2285 let _ = writeln!(stderr, "{message}");
2286 }
2287 if result.failed() {
2288 failed = true;
2289 continue;
2290 }
2291 if opts.deps.emit {
2292 failed |= !write_deps(opts, plan, job, &result.deps, &mut stderr);
2293 // `-M` and `-MM` asked for the rule instead of the text, so there is nothing else
2294 // to write. The other two asked for both and fall through to the text below.
2295 if opts.deps.instead_of_compiling {
2296 continue;
2297 }
2298 }
2299 if let Err(e) = write_out(&job.output, result.text.as_bytes()) {
2300 let _ = writeln!(stderr, "rucc: error: {e}");
2301 failed = true;
2302 }
2303 }
2304 i32::from(failed)
2305}
2306
2307/// Whether this job is a file of assembly that has to be assembled and that nothing here assembles.
2308///
2309/// The phases rather than the kind, because there are two kinds of assembly input and one of them
2310/// is preprocessed first, and because an object file also has no compile phase and is not this: it
2311/// has no phases at all and goes to the linker as it is. A `.s` on a `-c` line has exactly
2312/// [`Phase::Assemble`] left, and a `.S` has the preprocessor in front of it, and neither has
2313/// anything the front end can do.
2314fn needs_an_assembler(job: &Job) -> bool {
2315 job.phases.contains(&Phase::Assemble) && !job.phases.contains(&Phase::Compile)
2316}
2317
2318/// Whether the preprocessor runs over it on the way in, which is the whole difference between the
2319/// two kinds of assembly input.
2320fn assembly_wants_cpp(job: &Job) -> bool {
2321 job.phases.contains(&Phase::Preprocess)
2322}
2323
2324/// Runs the front end over every input that has a compile phase, and writes what came out.
2325///
2326/// The same rule as [`preprocess_all`]: one input that fails does not stop the others, and the
2327/// exit status is a failure either way. An input that is already assembly or an object has no
2328/// compile phase and is passed over here, which the plan has already said in its notes.
2329fn compile_all(opts: &Options, plan: &Plan) -> i32 {
2330 let fs = OsFileSystem::new();
2331 let mut stderr = std::io::stderr().lock();
2332 let mut failed = false;
2333 let (mut remarks, ok) = Remarks::new(opts.opt_info_file.as_ref(), &mut stderr);
2334 failed |= !ok;
2335 let mut fired = Fired::new();
2336 let mut pressure = Pressure::new();
2337 let mut lowerings = Lowerings::new();
2338 for job in &plan.jobs {
2339 if !job.phases.contains(&Phase::Compile) && !needs_an_assembler(job) {
2340 continue;
2341 }
2342 // An input of IR is read back rather than compiled, since the C it came from is not
2343 // here any more. A file of assembly does not go through the front end at all and is
2344 // read by the assembler instead. Everything after this is the same for all three, so
2345 // the paths meet again at the messages and the file the result is written to.
2346 let started = std::time::Instant::now();
2347 let result = if needs_an_assembler(job) {
2348 assemble(opts, &job.input, assembly_wants_cpp(job), &fs)
2349 } else if job.kind == InputKind::Ir {
2350 compile_ir(opts, &job.input, &fs)
2351 } else {
2352 compile(opts, &job.input, &fs)
2353 };
2354 if opts.time {
2355 say_time(&job.input, started.elapsed(), &mut stderr);
2356 }
2357 fired.merge(&result.fired);
2358 pressure.merge(&result.pressure);
2359 lowerings.merge(&result.lowerings);
2360 failed |= !write_dumps(&job.input, &result.dumps, &mut stderr);
2361 failed |= !remarks.write(&result.remarks, &mut stderr);
2362 for message in &result.messages {
2363 let _ = writeln!(stderr, "{message}");
2364 }
2365 // Before the failure below, because a compilation that stopped in the back end is exactly
2366 // the one whose preprocessed source somebody wants to look at.
2367 failed |= !write_temps(job, &result.temps, &mut stderr);
2368 if result.failed() {
2369 failed = true;
2370 continue;
2371 }
2372 // `-MD` and `-MMD` write the rule beside the object and let the compilation happen, so
2373 // this is the one path where both files come out of the same run. An input of IR has no
2374 // dependencies to report and produces an empty list, which produces a rule naming only
2375 // itself, and that is the honest answer rather than a missing file.
2376 if opts.deps.emit {
2377 failed |= !write_deps(opts, plan, job, &result.deps, &mut stderr);
2378 }
2379 if let Err(e) = write_out(&job.output, result.artifact.bytes()) {
2380 let _ = writeln!(stderr, "rucc: error: {e}");
2381 failed = true;
2382 }
2383 }
2384 failed |= !write_coverage(opts, &fired, &mut stderr);
2385 failed |= !write_pressure(opts, &pressure, &mut stderr);
2386 failed |= !write_lowering(opts, &lowerings, &mut stderr);
2387 i32::from(failed)
2388}
2389
2390/// A directory for the object files only the link step ever sees, removed when it goes away.
2391///
2392/// `-c` writes its object where the user can see it and linking does not, which is the whole of
2393/// the difference: a `rucc a.c b.c` leaves an executable behind and nothing else, the same as
2394/// every other compiler. Removing them on drop rather than at the end of a function is so that a
2395/// link that failed leaves nothing behind either.
2396struct Scratch {
2397 /// Where the objects go.
2398 dir: PathBuf,
2399}
2400
2401impl Scratch {
2402 /// Makes one, under whatever the platform calls its temporary directory.
2403 ///
2404 /// The name carries the process id so that two compilers running at once do not share a
2405 /// directory, which they would otherwise do the moment two of them compiled a file of the
2406 /// same name.
2407 fn new() -> Result<Scratch, String> {
2408 let dir = std::env::temp_dir().join(format!("rucc-{}", std::process::id()));
2409 std::fs::create_dir_all(&dir).map_err(|e| format!("{}: {e}", dir.display()))?;
2410 Ok(Scratch { dir })
2411 }
2412}
2413
2414impl Drop for Scratch {
2415 fn drop(&mut self) {
2416 let _ = std::fs::remove_dir_all(&self.dir);
2417 }
2418}
2419
2420/// The link line the plan describes, for `-###`.
2421///
2422/// The names in it are the hints the plan carries rather than the temporaries a real compilation
2423/// would choose, because `-###` prints the line without having compiled anything and so has
2424/// nothing to point at. That also makes the printed line readable rather than naming a directory
2425/// that only exists while a compilation is running.
2426fn link_line(opts: &Options, link: &LinkOptions, job: &LinkJob) -> Result<String, link::Error> {
2427 let linker = link::find(opts.target, link)?;
2428 let args = link::line(opts.target, link, &job.inputs, &job.output)?;
2429 Ok(link::render(&linker, &args))
2430}
2431
2432/// Compiles everything, then links it.
2433///
2434/// The objects go in a directory that is removed afterwards, which is why this is not
2435/// [`compile_all`] followed by a link: the plan says an object feeding the linker is temporary
2436/// and does not say where, because where is a question that only has an answer once something is
2437/// running.
2438fn link_all(opts: &Options, plan: &Plan, link: &LinkOptions, verbose: bool) -> i32 {
2439 let Some(job) = &plan.link else {
2440 // Every path into here comes from a plan whose last phase is the link, and such a plan
2441 // has a link job. Saying so is cheaper than an unwrap that would have to be explained.
2442 let mut stderr = std::io::stderr().lock();
2443 let _ = writeln!(stderr, "rucc: error: there is nothing to link");
2444 return 1;
2445 };
2446 // Before anything is compiled, because a linker that is not on the machine is worth knowing
2447 // about in the second it takes to look rather than after the compilation.
2448 // And before that, whether this link has a line at all and whether what it reads is on the
2449 // machine. Both are answerable now, and a target whose sysroot has not been built is worth
2450 // saying so about before the compilation rather than after it.
2451 if let Err(why) = link::preflight(opts.target, link) {
2452 return complain(why);
2453 }
2454 let linker = match link::find(opts.target, link) {
2455 Ok(linker) => linker,
2456 Err(why) => return complain(why),
2457 };
2458
2459 let scratch = match Scratch::new() {
2460 Ok(scratch) => scratch,
2461 Err(why) => return complain(format!("could not make a place for the object files: {why}")),
2462 };
2463
2464 let fs = OsFileSystem::new();
2465 let mut failed = false;
2466 // One per job, in job order, which is what lets the link line below be rebuilt with the real
2467 // paths in it: every job contributes exactly one file to the line and does so in this order.
2468 let mut produced: Vec<String> = Vec::with_capacity(plan.jobs.len());
2469 let mut fired = Fired::new();
2470 let mut pressure = Pressure::new();
2471 let mut lowerings = Lowerings::new();
2472 {
2473 let mut stderr = std::io::stderr().lock();
2474 let (mut remarks, ok) = Remarks::new(opts.opt_info_file.as_ref(), &mut stderr);
2475 failed |= !ok;
2476 for (at, job) in plan.jobs.iter().enumerate() {
2477 let out = match &job.output {
2478 Output::Temporary(hint) => {
2479 // The index because two inputs in different directories can have the same
2480 // name, and the two objects of `rucc a/x.c b/x.c` must not be one file.
2481 scratch.dir.join(format!("{at}-{hint}")).display().to_string()
2482 }
2483 Output::File(path) => path.clone(),
2484 // A job feeding the linker never writes to standard output, since the plan gives
2485 // it a temporary. This is here so that the match is total rather than a panic.
2486 Output::Stdout => continue,
2487 };
2488 produced.push(out.clone());
2489 if !job.phases.contains(&Phase::Compile) && !needs_an_assembler(job) {
2490 continue;
2491 }
2492 let started = std::time::Instant::now();
2493 let result = if needs_an_assembler(job) {
2494 assemble(opts, &job.input, assembly_wants_cpp(job), &fs)
2495 } else if job.kind == InputKind::Ir {
2496 compile_ir(opts, &job.input, &fs)
2497 } else {
2498 compile(opts, &job.input, &fs)
2499 };
2500 if opts.time {
2501 say_time(&job.input, started.elapsed(), &mut stderr);
2502 }
2503 fired.merge(&result.fired);
2504 pressure.merge(&result.pressure);
2505 lowerings.merge(&result.lowerings);
2506 failed |= !write_dumps(&job.input, &result.dumps, &mut stderr);
2507 failed |= !remarks.write(&result.remarks, &mut stderr);
2508 for message in &result.messages {
2509 let _ = writeln!(stderr, "{message}");
2510 }
2511 failed |= !write_temps(job, &result.temps, &mut stderr);
2512 if result.failed() {
2513 failed = true;
2514 continue;
2515 }
2516 // A `-MD` on a command line that links writes the rule next to the executable and
2517 // names the executable as its target, since that is the file this source builds
2518 // here. The object it went through is in a temporary directory and is gone by the
2519 // time `make` reads any of this.
2520 if opts.deps.emit {
2521 failed |= !write_deps(opts, plan, job, &result.deps, &mut stderr);
2522 }
2523 if !matches!(result.artifact, Artifact::Object { .. }) {
2524 // Worth saying rather than writing whatever it is and letting the linker read it.
2525 // An empty file is a valid empty linker script, so a link handed one gets as far
2526 // as reporting every symbol of this file undefined, which is a page of messages
2527 // about something that went wrong here.
2528 let _ = writeln!(
2529 stderr,
2530 "rucc: internal error: {}: no object file was produced for the link",
2531 job.input
2532 );
2533 failed = true;
2534 continue;
2535 }
2536 if let Err(e) = std::fs::write(&out, result.artifact.bytes()) {
2537 let _ = writeln!(stderr, "rucc: error: {out}: {e}");
2538 failed = true;
2539 }
2540 }
2541 failed |= !write_coverage(opts, &fired, &mut stderr);
2542 failed |= !write_pressure(opts, &pressure, &mut stderr);
2543 failed |= !write_lowering(opts, &lowerings, &mut stderr);
2544 failed |= !write_lowering(opts, &lowerings, &mut stderr);
2545 }
2546 if failed {
2547 // Nothing is linked from a compilation that did not finish. A linker run over the objects
2548 // that did compile would report every function of the file that did not as undefined,
2549 // which is a page of messages about a mistake already reported once.
2550 return 1;
2551 }
2552
2553 // The items in command line order with the temporaries filled in. A library and a word for the
2554 // linker contribute no job and pass through, and every file item takes the next job's real
2555 // output, which is what keeps whatever was written between two objects between them here.
2556 let mut outputs = produced.into_iter();
2557 let mut items = Vec::with_capacity(job.inputs.len());
2558 for item in &job.inputs {
2559 match item {
2560 link::Item::Library(name) => items.push(link::Item::Library(name.clone())),
2561 link::Item::Linker(arg) => items.push(link::Item::Linker(arg.clone())),
2562 link::Item::File(_) => match outputs.next() {
2563 Some(path) => items.push(link::Item::File(path)),
2564 None => return complain("the plan asks the linker for a file nothing produced"),
2565 },
2566 }
2567 }
2568
2569 let args = match link::line(opts.target, link, &items, &job.output) {
2570 Ok(args) => args,
2571 Err(why) => return complain(why),
2572 };
2573 if verbose {
2574 let mut stderr = std::io::stderr().lock();
2575 let _ = writeln!(stderr, "{}", link::render(&linker, &args));
2576 }
2577 let started = std::time::Instant::now();
2578 let ran = link::run(&linker, &args);
2579 if opts.time {
2580 // The one step of a compilation that really is another program, so this line is the same
2581 // measurement gcc's is and names the linker the way gcc names `collect2`.
2582 let mut stderr = std::io::stderr().lock();
2583 say_time(&linker.name, started.elapsed(), &mut stderr);
2584 }
2585 match ran {
2586 Ok(()) => 0,
2587 // The linker has already said what was wrong on its own error output, and repeating that
2588 // linking failed would only push its message further up the screen.
2589 Err(link::Error::Refused { .. }) => 1,
2590 Err(why) => complain(why),
2591 }
2592}
2593
2594/// Compiles everything and writes the objects into one static library.
2595///
2596/// No temporary directory and no second program. The objects never reach the file system at all:
2597/// they go from the compiler into the archive writer, which is both faster than writing a directory
2598/// of files for an `ar` to read back and the reason the symbol index can be written at all. A
2599/// member's index entries are the names the object writer says it wrote, and the only thing that
2600/// knows those is the run that wrote it.
2601///
2602/// `-save-temps` is the exception. It asked for the objects to be kept, the plan gave them names a
2603/// person can find, and they are written there as well as put in the archive.
2604fn archive_all(opts: &Options, plan: &Plan) -> i32 {
2605 let Some(job) = &plan.archive else {
2606 // Every path into here comes from a plan whose last phase is the archive, and such a plan
2607 // has an archive job. Saying so is cheaper than an unwrap that would have to be explained.
2608 return complain("there is nothing to put in an archive");
2609 };
2610 // Before anything is compiled, because a format this has no container for is worth knowing
2611 // about in the second it takes to look rather than after the whole compilation.
2612 let flavour = match opts.target.os.object_format() {
2613 ObjectFormat::Elf => rucc_archive::Flavour::Gnu,
2614 ObjectFormat::Coff => rucc_archive::Flavour::Coff,
2615 // Mach-O wants the BSD flavour, whose index is a different member under a different name,
2616 // and wasm has no archives of its own at all. Neither has an object writer either, so a
2617 // command line reaching this would have failed in the next step regardless.
2618 format @ (ObjectFormat::MachO | ObjectFormat::Wasm) => {
2619 return complain(format!(
2620 "there is no archive format for {} objects in this compiler yet",
2621 format.as_str()
2622 ));
2623 }
2624 };
2625
2626 let fs = OsFileSystem::new();
2627 let mut failed = false;
2628 let mut members: Vec<rucc_archive::Member> = Vec::with_capacity(plan.jobs.len());
2629 let mut names = job.members.iter();
2630 let mut fired = Fired::new();
2631 let mut pressure = Pressure::new();
2632 let mut lowerings = Lowerings::new();
2633 {
2634 let mut stderr = std::io::stderr().lock();
2635 let (mut remarks, ok) = Remarks::new(opts.opt_info_file.as_ref(), &mut stderr);
2636 failed |= !ok;
2637 for plan_job in &plan.jobs {
2638 // What the plan called this member. The two lists are walked together rather than the
2639 // name being worked out again here, so that what `-###` printed and what goes in the
2640 // file cannot come apart.
2641 let Some(member) = names.next() else {
2642 return complain("the plan asks the archive for a member nothing produced");
2643 };
2644 if !plan_job.phases.contains(&Phase::Compile) && !needs_an_assembler(plan_job) {
2645 // Neither something to compile nor something to assemble, so there is nothing to
2646 // put in, and an archive quietly missing a member is worse than a message.
2647 let _ = writeln!(
2648 &mut stderr,
2649 "rucc: error: {}: this compiler makes an archive out of what it compiles, and \
2650 there is nothing here for it to do",
2651 plan_job.input
2652 );
2653 failed = true;
2654 continue;
2655 }
2656 let started = std::time::Instant::now();
2657 let result = if needs_an_assembler(plan_job) {
2658 assemble(opts, &plan_job.input, assembly_wants_cpp(plan_job), &fs)
2659 } else if plan_job.kind == InputKind::Ir {
2660 compile_ir(opts, &plan_job.input, &fs)
2661 } else {
2662 compile(opts, &plan_job.input, &fs)
2663 };
2664 if opts.time {
2665 say_time(&plan_job.input, started.elapsed(), &mut stderr);
2666 }
2667 fired.merge(&result.fired);
2668 pressure.merge(&result.pressure);
2669 lowerings.merge(&result.lowerings);
2670 failed |= !write_dumps(&plan_job.input, &result.dumps, &mut stderr);
2671 failed |= !remarks.write(&result.remarks, &mut stderr);
2672 for message in &result.messages {
2673 let _ = writeln!(stderr, "{message}");
2674 }
2675 failed |= !write_temps(plan_job, &result.temps, &mut stderr);
2676 if result.failed() {
2677 failed = true;
2678 continue;
2679 }
2680 if opts.deps.emit {
2681 failed |= !write_deps(opts, plan, plan_job, &result.deps, &mut stderr);
2682 }
2683 let Artifact::Object { bytes, defines } = result.artifact else {
2684 let _ = writeln!(
2685 stderr,
2686 "rucc: internal error: {}: no object file was produced for the archive",
2687 plan_job.input
2688 );
2689 failed = true;
2690 continue;
2691 };
2692 // Under `-save-temps` the plan gave the object a name a person can find, so it is
2693 // written there too. Otherwise it is only ever a member and never a file.
2694 if let Output::File(path) = &plan_job.output {
2695 if let Err(e) = std::fs::write(path, &bytes) {
2696 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2697 failed = true;
2698 }
2699 }
2700 members.push(rucc_archive::Member { name: member.clone(), body: bytes, defines });
2701 }
2702 failed |= !write_coverage(opts, &fired, &mut stderr);
2703 failed |= !write_pressure(opts, &pressure, &mut stderr);
2704 failed |= !write_lowering(opts, &lowerings, &mut stderr);
2705 failed |= !write_lowering(opts, &lowerings, &mut stderr);
2706 }
2707 if failed {
2708 // Nothing is written from a compilation that did not finish, for the reason the link gives:
2709 // an archive missing the file that failed is one a link reports every name of as undefined,
2710 // which is a page of messages about a mistake already reported once.
2711 return 1;
2712 }
2713
2714 let bytes = match rucc_archive::write(flavour, &members) {
2715 Ok(bytes) => bytes,
2716 // Every one of these is a bug here rather than a program's mistake: the names came from the
2717 // object writer and the bodies came from this process.
2718 Err(why) => return complain(format!("the archive could not be written: {why}")),
2719 };
2720 match std::fs::write(&job.output, &bytes) {
2721 Ok(()) => 0,
2722 Err(e) => complain(format!("{}: {e}", job.output)),
2723 }
2724}
2725
2726/// Prints one driver level message and gives back the exit status that goes with it.
2727fn complain(why: impl std::fmt::Display) -> i32 {
2728 let mut stderr = std::io::stderr().lock();
2729 let _ = writeln!(stderr, "rucc: error: {why}");
2730 1
2731}
2732
2733/// Writes what `-Zrule-coverage=FILE` asked for, and says whether it could.
2734///
2735/// Once for the whole command line rather than once per input, because the question is which
2736/// lowering rules this run of the compiler reached and a file per input would leave the reader
2737/// unioning files to find out something one process already knew.
2738///
2739/// A file that could not be written is a failure and not a warning. What asks for this is a
2740/// measurement run, and a measurement that quietly did not happen is worse than one that stopped.
2741fn write_coverage(opts: &Options, fired: &Fired, stderr: &mut impl std::io::Write) -> bool {
2742 let Some(path) = &opts.rule_coverage else { return true };
2743 let Some(table) = coverage::table(opts.target.arch) else {
2744 let _ = writeln!(
2745 stderr,
2746 "rucc: error: there are no lowering rules for {} yet, so there is no coverage of them \
2747 to report",
2748 opts.target
2749 );
2750 return false;
2751 };
2752 match std::fs::write(path, fired.listing(table)) {
2753 Ok(()) => true,
2754 Err(e) => {
2755 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2756 false
2757 }
2758 }
2759}
2760
2761/// Writes what `-Zregister-pressure=FILE` asked for, and says whether it could.
2762///
2763/// Once for the whole command line, for the reason [`write_coverage`] gives, and a file that could
2764/// not be written is a failure for the reason it gives too. There is no equivalent of the missing
2765/// rule table here, since every target this compiles for has an allocator, and a run that reached
2766/// no back end at all writes an empty listing rather than nothing: a measurement of a build that
2767/// produced no code is still an answer and it is the honest one.
2768fn write_pressure(opts: &Options, pressure: &Pressure, stderr: &mut impl std::io::Write) -> bool {
2769 let Some(path) = &opts.register_pressure else { return true };
2770 match std::fs::write(path, pressure.listing()) {
2771 Ok(()) => true,
2772 Err(e) => {
2773 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2774 false
2775 }
2776 }
2777}
2778
2779/// Writes what `-Zlowering=FILE` asked for, and says whether it could.
2780///
2781/// Once for the whole command line, for the reason [`write_coverage`] gives, and a file that could
2782/// not be written is a failure for the reason it gives too. A run that reached no back end writes
2783/// an empty listing rather than nothing, the way [`write_pressure`] does and for the same reason.
2784fn write_lowering(opts: &Options, lowerings: &Lowerings, stderr: &mut impl std::io::Write) -> bool {
2785 let Some(path) = &opts.lowering_dump else { return true };
2786 match std::fs::write(path, lowerings.listing()) {
2787 Ok(()) => true,
2788 Err(e) => {
2789 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2790 false
2791 }
2792 }
2793}
2794
2795/// Where the `-fopt-info` remarks go, and how much of the run has already gone there.
2796///
2797/// Standard error by default, and one file for the whole run when `-fopt-info=<file>` named one.
2798/// A file rather than the diagnostic stream is what a harness wants: the corpus in
2799/// `tamnd/rucc-corpus` matches a rejection against what the compiler said on standard error, and
2800/// a few thousand remarks mixed into that would bury it.
2801struct Remarks {
2802 /// The file, if there is one.
2803 file: Option<String>,
2804 /// Whether anything has been written to it yet, which decides between truncating and
2805 /// appending. One file holds the whole run rather than the last input in it.
2806 started: bool,
2807}
2808
2809impl Remarks {
2810 /// Prepares the destination, emptying the file if there is one.
2811 ///
2812 /// Emptied here rather than at the first remark, because a run where no pass had anything to
2813 /// say should leave an empty file and not yesterday's. An absent file and an empty one are
2814 /// different facts and something reading this will act on the difference.
2815 fn new(file: Option<&String>, stderr: &mut impl std::io::Write) -> (Self, bool) {
2816 let mut ok = true;
2817 if let Some(path) = file {
2818 if let Err(e) = std::fs::write(path, "") {
2819 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2820 ok = false;
2821 }
2822 }
2823 (Self { file: file.cloned(), started: false }, ok)
2824 }
2825
2826 /// Writes one input's remarks, and says whether that worked.
2827 ///
2828 /// A file that cannot be written is a failure and not a warning, for the reason
2829 /// [`write_dumps`] gives: remarks that quietly did not arrive look exactly like a compilation
2830 /// where nothing happened.
2831 fn write(&mut self, text: &str, stderr: &mut impl std::io::Write) -> bool {
2832 if text.is_empty() {
2833 return true;
2834 }
2835 let Some(path) = &self.file else {
2836 let _ = write!(stderr, "{text}");
2837 return true;
2838 };
2839 let opened = std::fs::OpenOptions::new()
2840 .write(true)
2841 .append(self.started)
2842 .truncate(!self.started)
2843 .create(true)
2844 .open(path);
2845 self.started = true;
2846 let result =
2847 opened.and_then(|mut file| std::io::Write::write_all(&mut file, text.as_bytes()));
2848 if let Err(e) = result {
2849 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2850 return false;
2851 }
2852 true
2853 }
2854}
2855
2856/// Writes what `-fdump-ir=` asked to see, one file per dump.
2857///
2858/// The name is the input file with the dump's own name and `.ir` after it, so a directory listing
2859/// after a run is the passes in the order they ran, per input. They go in the working directory
2860/// rather than beside the output, because a dump is something a person asked for at a prompt and
2861/// the working directory is where that person is.
2862///
2863/// A file that could not be written is a failure and not a warning, for the reason
2864/// [`write_coverage`] gives: what asked for this is somebody debugging a pass, and a dump that
2865/// quietly did not happen looks exactly like a pass that did not run.
2866fn write_dumps(input: &str, dumps: &[rucc_opt::Dump], stderr: &mut impl std::io::Write) -> bool {
2867 let stem = std::path::Path::new(input)
2868 .file_name()
2869 .map_or_else(|| input.to_owned(), |name| name.to_string_lossy().into_owned());
2870 let mut ok = true;
2871 for dump in dumps {
2872 let path = format!("{stem}.{}.ir", dump.name);
2873 if let Err(e) = std::fs::write(&path, &dump.text) {
2874 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2875 ok = false;
2876 }
2877 }
2878 ok
2879}
2880
2881/// Writes the files `-save-temps` kept, which is nothing at all unless it was given.
2882///
2883/// A file that could not be written is a failure rather than a warning, for the reason
2884/// [`write_dumps`] gives: somebody asked for these by name, and one that quietly did not happen
2885/// looks like a compilation that never went through that step.
2886fn write_temps(job: &Job, temps: &Temps, stderr: &mut impl std::io::Write) -> bool {
2887 let mut ok = true;
2888 let kept = [(job.saved_text(), &temps.preprocessed), (job.saved_asm(), &temps.assembly)];
2889 for (path, text) in kept {
2890 // A step the compilation did not reach has nothing to keep, and a job that is not keeping
2891 // that step has nowhere to put it. Either way there is no file here.
2892 let (Some(path), Some(text)) = (path, text) else { continue };
2893 if let Err(e) = std::fs::write(&path, text) {
2894 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2895 ok = false;
2896 }
2897 }
2898 ok
2899}
2900
2901/// One line of `-time`, which is what a step was called and how long it took.
2902///
2903/// GCC's two numbers are the user and the system time of a subprocess it ran. This compiler runs
2904/// no subprocess for anything but the link, so what is measured here is the wall clock of the
2905/// step and the second column is always zero. The shape of the line is kept because a person
2906/// reading it next to gcc's should not have to work out which column is which.
2907fn say_time(name: &str, took: std::time::Duration, stderr: &mut impl std::io::Write) {
2908 let _ = writeln!(stderr, "# {name} {:.2} {:.2}", took.as_secs_f64(), 0.0);
2909}
2910
2911/// Writes one job's result where the plan said it goes.
2912///
2913/// # Errors
2914///
2915/// Returns the message to print, which names the file when there is one, because "permission
2916/// denied" on its own does not say which file was refused.
2917fn write_out(output: &Output, bytes: &[u8]) -> Result<(), String> {
2918 match output {
2919 Output::Stdout => {
2920 let mut stdout = std::io::stdout().lock();
2921 stdout.write_all(bytes).map_err(|e| format!("writing to standard output: {e}"))
2922 }
2923 Output::File(path) | Output::Temporary(path) => {
2924 std::fs::write(path, bytes).map_err(|e| format!("{path}: {e}"))
2925 }
2926 }
2927}
2928
2929/// Runs the driver and returns the process exit code.
2930///
2931/// `args` excludes the program name. Output goes to `stdout` and errors to `stderr`, which
2932/// is the one place in the compiler that is true.
2933pub fn run(args: &[String]) -> i32 {
2934 match parse_args(args) {
2935 Ok(Action::Help) => {
2936 print!("{USAGE}");
2937 0
2938 }
2939 Ok(Action::Version) => {
2940 println!("rucc {VERSION}");
2941 0
2942 }
2943 Ok(Action::Print(line)) => {
2944 println!("{line}");
2945 0
2946 }
2947 Ok(Action::PrintConfig(opts)) => {
2948 print!("{}", print_config(&opts));
2949 0
2950 }
2951 Ok(Action::PrintPipeline(opts)) => {
2952 print!("{}", print_pipeline(&opts));
2953 0
2954 }
2955 Ok(Action::PrintPlan { opts, plan, link }) => {
2956 print!("{}", plan.render());
2957 // The line as it would be typed, which is the half of `-###` that section 4.3 says
2958 // arrives with the link. It is printed even when the linker is not on this machine,
2959 // because what a build wants from `-###` is what the compiler would do.
2960 if let Some(job) = &plan.link {
2961 match link_line(&opts, &link, job) {
2962 Ok(line) => println!("{line}"),
2963 Err(why) => {
2964 let mut stderr = std::io::stderr().lock();
2965 let _ = writeln!(stderr, "rucc: error: {why}");
2966 return 1;
2967 }
2968 }
2969 }
2970 0
2971 }
2972 Ok(Action::Fetch { what, target, cache }) => fetch_sysroot(what, target, &cache),
2973 Ok(Action::Compile { opts, plan, link, jobs, verbose, notes }) => {
2974 {
2975 let mut stderr = std::io::stderr().lock();
2976 // Before the plan rather than after it, because a note is about the command line
2977 // and the plan is what the command line was read as, so the reader wants the two
2978 // in that order.
2979 for note in ¬es {
2980 let _ = writeln!(stderr, "rucc: warning: {note}");
2981 }
2982 if verbose {
2983 let _ = write!(stderr, "{}", plan.render());
2984 let _ = writeln!(stderr, "workers: {}", jobs.count());
2985 }
2986 }
2987 if opts.emit == EmitKind::Preprocessed {
2988 return preprocess_all(&opts, &plan);
2989 }
2990 if opts.emit == EmitKind::Archive {
2991 return archive_all(&opts, &plan);
2992 }
2993 if opts.emit != EmitKind::Executable {
2994 return compile_all(&opts, &plan);
2995 }
2996 link_all(&opts, &plan, &link, verbose)
2997 }
2998 Err(e) => {
2999 let mut stderr = std::io::stderr().lock();
3000 let _ = writeln!(stderr, "rucc: error: {e}");
3001 let _ = writeln!(stderr, "rucc: note: run `rucc --help` for usage");
3002 1
3003 }
3004 }
3005}
3006
3007#[cfg(test)]
3008mod tests {
3009 use rucc_session::{
3010 Contract, GnucVersion, IncludeForm, LtoJobs, OptLevel, Partition, Patchable, Visibility,
3011 };
3012
3013 use super::*;
3014
3015 fn args(s: &[&str]) -> Vec<String> {
3016 s.iter().map(|x| (*x).to_owned()).collect()
3017 }
3018
3019 /// A target to write down where the host would otherwise decide, for the tests whose answer
3020 /// would be a different one on a different machine.
3021 ///
3022 /// Most of the tests here never name a target, which is right, because most of what the driver
3023 /// does with a command line is the same wherever it runs and a test that pinned one would be
3024 /// saying so in every case for the sake of the two that need it. The two that need it are the
3025 /// ones whose answer comes off the target rather than off the command line: the name an object
3026 /// gets, which is `a.o` here and `a.obj` on Windows, and whether Microsoft's reading of a
3027 /// nameless member is on, which is off here and on there. Both are the compiler being right, and
3028 /// a test that leaves the target to the host is asking a question with two correct answers.
3029 const LINUX: &str = "--target=x86_64-unknown-linux-gnu";
3030
3031 #[test]
3032 fn help_and_version_win_over_everything_else() {
3033 assert_eq!(parse_args(&args(&["-c", "--help", "x.c"])).unwrap(), Action::Help);
3034 assert_eq!(parse_args(&args(&["--version"])).unwrap(), Action::Version);
3035 }
3036
3037 fn compile(s: &[&str]) -> (Box<Options>, Box<Plan>) {
3038 match parse_args(&args(s)).expect("expected a compilation") {
3039 Action::Compile { opts, plan, .. } => (opts, plan),
3040 other => panic!("expected a compilation, got {other:?}"),
3041 }
3042 }
3043
3044 fn linking(s: &[&str]) -> (Box<LinkOptions>, Box<Plan>) {
3045 match parse_args(&args(s)).expect("expected a compilation") {
3046 Action::Compile { link, plan, .. } => (link, plan),
3047 other => panic!("expected a compilation, got {other:?}"),
3048 }
3049 }
3050
3051 fn notes(s: &[&str]) -> Vec<String> {
3052 match parse_args(&args(s)).expect("expected a compilation") {
3053 Action::Compile { notes, .. } => notes,
3054 other => panic!("expected a compilation, got {other:?}"),
3055 }
3056 }
3057
3058 /// The ordinary command line has nothing to say about itself, which is the property that makes
3059 /// a note worth reading when there is one.
3060 #[test]
3061 fn a_command_line_with_nothing_wrong_with_it_carries_no_notes() {
3062 assert_eq!(notes(&["-c", "a.c"]), Vec::<String>::new());
3063 }
3064
3065 /// A directory that is not there contributes nothing to the search path, so there is no tree to
3066 /// read a release out of and nothing to compare the pin against. Said as a test because this is
3067 /// the shape a hermetic machine takes: the probe reads the disk and every other machine has a
3068 /// different disk, so what can be asserted here is the silence.
3069 #[test]
3070 fn a_named_tree_that_is_not_on_the_machine_is_not_a_release_mismatch() {
3071 let said =
3072 notes(&["--target=x86_64-linux-gnu.2.28", "--sysroot=/nowhere-at-all", "-c", "a.c"]);
3073 assert_eq!(said, Vec::<String>::new());
3074 }
3075
3076 #[test]
3077 fn collects_inputs_and_flags() {
3078 let (opts, plan) = compile(&["-c", "-O2", "-g", "a.c", "b.c"]);
3079 let paths: Vec<&str> = plan.jobs.iter().map(|j| j.input.as_str()).collect();
3080 assert_eq!(paths, vec!["a.c", "b.c"]);
3081 assert_eq!(opts.opt_level, OptLevel::O2);
3082 assert_eq!(opts.emit, EmitKind::Object);
3083 assert!(opts.debug_info);
3084 }
3085
3086 /// The unstable options, which are spelled apart from everything else on purpose: what is
3087 /// under `-Z` promises nothing, and a build that reaches for one should have had to say so.
3088 #[test]
3089 fn an_unstable_option_is_taken_and_one_that_does_not_exist_is_refused() {
3090 let (opts, _) = compile(&["-c", "-Zrule-coverage=/tmp/rules.cov", "a.c"]);
3091 assert_eq!(opts.rule_coverage.as_deref(), Some("/tmp/rules.cov"));
3092
3093 let (plain, _) = compile(&["-c", "a.c"]);
3094 assert_eq!(plain.rule_coverage, None, "nothing is measured unless it was asked for");
3095
3096 assert!(parse_args(&args(&["-Zrule-coverage=", "a.c"])).is_err(), "a file with no name");
3097 let unknown = parse_args(&args(&["-Zwhat", "a.c"])).expect_err("there is no such option");
3098 assert!(unknown.message.contains("4.11"), "{}", unknown.message);
3099 }
3100
3101 /// The other measurement written to a file, which reads the same way and fails the same way.
3102 #[test]
3103 fn where_the_register_pressure_goes_is_asked_for_the_same_way() {
3104 let (opts, _) = compile(&["-c", "-O2", "-Zregister-pressure=/tmp/spills.txt", "a.c"]);
3105 assert_eq!(opts.register_pressure.as_deref(), Some("/tmp/spills.txt"));
3106
3107 let (plain, _) = compile(&["-c", "a.c"]);
3108 assert_eq!(plain.register_pressure, None, "nothing is measured unless it was asked for");
3109
3110 assert!(parse_args(&args(&["-Zregister-pressure=", "a.c"])).is_err(), "no file named");
3111 }
3112
3113 /// The third one, which says what the pre-selection lowering group did.
3114 #[test]
3115 fn where_the_lowering_dump_goes_is_asked_for_the_same_way() {
3116 let (opts, _) = compile(&["-c", "-O2", "-Zlowering=/tmp/lowering.txt", "a.c"]);
3117 assert_eq!(opts.lowering_dump.as_deref(), Some("/tmp/lowering.txt"));
3118
3119 let (plain, _) = compile(&["-c", "a.c"]);
3120 assert_eq!(plain.lowering_dump, None, "nothing is dumped unless it was asked for");
3121
3122 assert!(parse_args(&args(&["-Zlowering=", "a.c"])).is_err(), "no file named");
3123 }
3124
3125 /// Scheduling, which has the three way answer every optimization flag has: on, off, and
3126 /// nothing said, which is whatever the optimization level asks for. The name is gcc's, and
3127 /// gcc's has a two in it because gcc has a scheduler before allocation and one after and this
3128 /// is the one after.
3129 #[test]
3130 fn scheduling_can_be_turned_on_and_off_and_left_to_the_optimization_level() {
3131 let (on, _) = compile(&["-c", "-O0", "-fschedule-insns2", "a.c"]);
3132 assert_eq!(on.schedule_insns, Some(true));
3133
3134 let (off, _) = compile(&["-c", "-O2", "-fno-schedule-insns2", "a.c"]);
3135 assert_eq!(off.schedule_insns, Some(false));
3136
3137 let (quiet, _) = compile(&["-c", "-O2", "a.c"]);
3138 assert_eq!(quiet.schedule_insns, None, "nothing said, so the level decides");
3139 assert!(quiet.opt_level.schedules(), "and at this level the level says yes");
3140
3141 let (none, _) = compile(&["-c", "a.c"]);
3142 assert!(!none.opt_level.schedules(), "at no optimization it says no");
3143 }
3144
3145 /// Whether the timing model is worth holding an instruction back over, which is a `-Z` because
3146 /// it is a question about a target's description rather than about the program being compiled.
3147 #[test]
3148 fn whether_the_timing_model_is_cycle_accurate_can_be_overridden() {
3149 let (yes, _) = compile(&["-c", "-O2", "-Zcycle-accurate-model=yes", "a.c"]);
3150 assert_eq!(yes.cycle_accurate_model, Some(true));
3151
3152 let (no, _) = compile(&["-c", "-O2", "-Zcycle-accurate-model=no", "a.c"]);
3153 assert_eq!(no.cycle_accurate_model, Some(false));
3154
3155 let (plain, _) = compile(&["-c", "-O2", "a.c"]);
3156 assert_eq!(plain.cycle_accurate_model, None, "the target's own answer stands");
3157
3158 let bad = parse_args(&args(&["-Zcycle-accurate-model=maybe", "a.c"]))
3159 .expect_err("it takes yes or no");
3160 assert!(bad.message.contains("yes or no"), "{}", bad.message);
3161 }
3162
3163 #[test]
3164 fn a_bare_dash_o_means_o1_the_way_gcc_reads_it() {
3165 let (opts, _) = compile(&["-O", "a.c"]);
3166 assert_eq!(opts.opt_level, OptLevel::O1);
3167 }
3168
3169 #[test]
3170 fn dash_x_applies_to_later_inputs_only_and_none_stops_it() {
3171 let (_, plan) = compile(&["a.o", "-x", "c", "b.txt", "-x", "none", "c.o"]);
3172 assert_eq!(plan.jobs[0].kind, InputKind::LinkerInput);
3173 assert_eq!(plan.jobs[1].kind, InputKind::C);
3174 assert_eq!(plan.jobs[2].kind, InputKind::LinkerInput);
3175 }
3176
3177 #[test]
3178 fn dash_j_reaches_the_scheduler_and_defaults_to_the_machine() {
3179 let (_, _, jobs) = match parse_args(&args(&["-j4", "a.c"])).unwrap() {
3180 Action::Compile { opts, plan, jobs, .. } => (opts, plan, jobs),
3181 other => panic!("expected a compilation, got {other:?}"),
3182 };
3183 assert_eq!(jobs.count(), 4);
3184
3185 let default = match parse_args(&args(&["a.c"])).unwrap() {
3186 Action::Compile { jobs, .. } => jobs,
3187 other => panic!("expected a compilation, got {other:?}"),
3188 };
3189 assert_eq!(default, Jobs::available());
3190 assert!(parse_args(&args(&["-j0", "a.c"])).is_err());
3191 }
3192
3193 #[test]
3194 fn triple_hash_prints_the_plan_and_runs_nothing() {
3195 let a = parse_args(&args(&["-###", "-c", "a.c"])).unwrap();
3196 let Action::PrintPlan { plan, .. } = a else { panic!("expected a plan dump") };
3197 assert!(plan.render().contains("a.c: preprocess, compile, assemble -> a.o"));
3198 }
3199
3200 #[test]
3201 fn the_flag_that_keeps_the_intermediate_files_has_three_spellings_and_two_meanings() {
3202 // The bare one is `=obj` and not `=cwd`. gcc's manual says the opposite and gcc 16 does
3203 // this, and following the compiler is what makes a build that reads either of them find
3204 // the files where they are.
3205 assert_eq!(compile(&["-c", "-save-temps", "a.c"]).0.save_temps, SaveTemps::Object);
3206 assert_eq!(compile(&["-c", "-save-temps=obj", "a.c"]).0.save_temps, SaveTemps::Object);
3207 assert_eq!(compile(&["-c", "-save-temps=cwd", "a.c"]).0.save_temps, SaveTemps::Cwd);
3208 assert_eq!(compile(&["-c", "a.c"]).0.save_temps, SaveTemps::No);
3209 // The last one on the line decides, the way it does for every other flag with an
3210 // argument, and a keyword that is neither is fatal rather than ignored: a run that kept
3211 // nothing and said nothing looks exactly like one where the files were not produced.
3212 let (opts, _) = compile(&["-c", "-save-temps", "-save-temps=cwd", "a.c"]);
3213 assert_eq!(opts.save_temps, SaveTemps::Cwd);
3214 let e = parse_args(&args(&["-c", "-save-temps=nowhere", "a.c"])).unwrap_err();
3215 assert!(e.message.contains("accepted: cwd, obj"), "{}", e.message);
3216 }
3217
3218 #[test]
3219 fn the_flag_that_times_each_step_reaches_the_options_and_changes_nothing_else() {
3220 let (opts, plan) = compile(&["-c", "-time", "a.c"]);
3221 let (plain, without) = compile(&["-c", "a.c"]);
3222 assert!(opts.time);
3223 assert!(!plain.time);
3224 // Against the same line without the flag rather than against a spelling of the object's
3225 // name, since what the object is called is the host's business and this is not about that.
3226 assert_eq!(plan.jobs[0].output, without.jobs[0].output);
3227 }
3228
3229 #[test]
3230 fn dash_x_names_what_it_accepts_when_it_does_not_know_a_language() {
3231 let e = parse_args(&args(&["-x", "fortran", "a.c"])).unwrap_err();
3232 assert!(e.message.contains("assembler-with-cpp"), "{}", e.message);
3233 }
3234
3235 /// What `--fetch` says for a target this release pins nothing for, which is every target except
3236 /// the three windows-gnu ones today.
3237 #[test]
3238 fn a_fetch_of_a_target_nothing_is_pinned_for_says_so_rather_than_reaching_the_network() {
3239 let e = parse_args(&args(&["--fetch", "x86_64-linux-musl"])).unwrap_err();
3240 assert!(e.message.contains("pins no sysroot for x86_64-linux-musl"), "{}", e.message);
3241 // And what it does pin, because a release with some rows in the table and a release with
3242 // none are two situations and the second sentence is what tells them apart.
3243 assert!(e.message.contains("x86_64-windows-gnu"), "{}", e.message);
3244 // The joined spelling is the same flag.
3245 let joined = parse_args(&args(&["--fetch=x86_64-linux-musl"])).unwrap_err();
3246 assert_eq!(joined, e);
3247 }
3248
3249 /// The two targets a release will never pin, which is a different answer from the one above.
3250 ///
3251 /// Section 13.4. A person who reads "this release pins no sysroot yet" waits for a release that
3252 /// does, and no release of this compiler can ship either of these, so the message names the
3253 /// licence that decides it and what to do instead.
3254 #[test]
3255 fn a_fetch_of_a_target_behind_a_licence_wall_says_so_rather_than_saying_not_yet() {
3256 let e = parse_args(&args(&["--fetch", "aarch64-macos"])).unwrap_err();
3257 assert!(e.message.contains("Xcode licence"), "{}", e.message);
3258 assert!(e.message.contains("there never will be"), "{}", e.message);
3259 assert!(!e.message.contains("tamnd/rucc-cross"), "{}", e.message);
3260
3261 let e = parse_args(&args(&["--fetch", "x86_64-windows-msvc"])).unwrap_err();
3262 assert!(e.message.contains("redistributed"), "{}", e.message);
3263 // The way out of this one is a target rather than a download, and it is the default already.
3264 assert!(e.message.contains("mingw-w64"), "{}", e.message);
3265 // And the mingw-w64 target next to it is ours to ship and published, so the same flag has
3266 // something to get rather than a licence to explain.
3267 let action = parse_args(&args(&["--fetch", "x86_64-windows-gnu"])).expect("it is pinned");
3268 let Action::Fetch { what, .. } = action else { panic!("{action:?}") };
3269 assert_eq!(what.tuple, "x86_64-windows-gnu");
3270 }
3271
3272 /// An Apple target on a machine with no SDK, which is section 8.6's other host.
3273 ///
3274 /// Not run on a mac, where the SDK this is about is installed and the compile is the ordinary one
3275 /// that uses it. What the reason says is asserted in `rucc_sysroot::wall` and where it is printed
3276 /// is asserted in `rucc-pp`, so what is left here is that the driver works it out and leaves it
3277 /// where the preprocessor will find it, and that neither way past the wall leaves one behind.
3278 #[test]
3279 fn an_apple_target_with_no_sdk_anywhere_carries_the_licence_rather_than_a_missing_directory() {
3280 if cfg!(target_os = "macos") || std::env::var_os("SDKROOT").is_some() {
3281 return;
3282 }
3283 let (opts, _) = compile(&["--target=aarch64-macos", "-c", "a.c"]);
3284 let why = opts.search.missing_system().expect("the wall is the reason there are none");
3285 assert!(why.contains("aarch64-macos needs a macOS SDK"), "{why}");
3286 assert!(why.contains("Xcode licence"), "{why}");
3287 assert!(why.contains("-isysroot"), "{why}");
3288
3289 // A program that includes none of the library needs none of the SDK, which is what section
3290 // 8.6 means by being able to target the platform without one, so there is nothing to explain.
3291 let (opts, _) = compile(&["--target=aarch64-macos", "-nostdinc", "-c", "a.c"]);
3292 assert_eq!(opts.search.missing_system(), None);
3293 // And naming a path is the other way through, whether or not the path is there: a mistyped
3294 // directory is a mistake to report on its own terms rather than a licence to explain.
3295 let (opts, _) = compile(&["--target=aarch64-macos", "-isysroot", "/opt/sdk", "-c", "a.c"]);
3296 assert_eq!(opts.search.missing_system(), None);
3297 }
3298
3299 /// The same wall on the compile side of an MSVC target, where the way past it is a tuple.
3300 ///
3301 /// Not run on Windows, for the same reason the one above is not run on a mac: the wall stands in
3302 /// front of an SDK this machine does not have, and a Windows machine is the kind that does. The
3303 /// driver asks `vswhere` where Visual Studio is and takes the newest kit under it, so on a box
3304 /// with the build tools installed there are headers, no wall and nothing here to be about.
3305 /// `INCLUDE` is the other way a machine has one and is the other half of the guard, since a
3306 /// person can set that anywhere while Visual Studio is only found on the platform it runs on.
3307 #[test]
3308 fn an_msvc_target_with_no_sdk_named_says_which_environment_needs_nothing_installed() {
3309 if cfg!(target_os = "windows") || std::env::var_os("INCLUDE").is_some() {
3310 return;
3311 }
3312 let (opts, _) = compile(&["--target=x86_64-windows-msvc", "-c", "a.c"]);
3313 let why = opts.search.missing_system().expect("the wall is the reason there are none");
3314 assert!(why.contains("the Windows SDK and its universal CRT"), "{why}");
3315 assert!(why.contains("mingw-w64"), "{why}");
3316 // And the mingw-w64 target has its headers from us, so nothing is missing to explain.
3317 let (opts, _) = compile(&["--target=x86_64-windows-gnu", "-c", "a.c"]);
3318 assert_eq!(opts.search.missing_system(), None);
3319 }
3320
3321 #[test]
3322 fn a_fetch_with_no_target_and_a_fetch_of_a_tuple_that_is_not_one_both_say_which() {
3323 let e = parse_args(&args(&["--fetch"])).unwrap_err();
3324 assert!(e.message.contains("--fetch requires"), "{}", e.message);
3325 let e = parse_args(&args(&["--fetch", "sparc64-solaris-gnu"])).unwrap_err();
3326 assert!(e.message.contains("--fetch sparc64-solaris-gnu"), "{}", e.message);
3327 assert!(e.message.contains("no sysroot to get"), "{}", e.message);
3328 }
3329
3330 /// Both flags on one line ask for opposite things, in either order.
3331 #[test]
3332 fn a_fetch_and_offline_together_is_a_refusal_whichever_way_round_they_are_written() {
3333 for line in [
3334 vec!["--offline", "--fetch", "x86_64-linux-musl"],
3335 vec!["--fetch", "x86_64-linux-musl", "--offline"],
3336 ] {
3337 let e = parse_args(&args(&line)).unwrap_err();
3338 assert!(e.message.contains("two opposite things"), "{}", e.message);
3339 }
3340 }
3341
3342 #[test]
3343 fn a_fetch_does_not_compile_anything_and_says_so_when_it_is_handed_a_file() {
3344 let e = parse_args(&args(&["--fetch", "x86_64-linux-musl", "a.c"])).unwrap_err();
3345 assert!(e.message.contains("compiles nothing"), "{}", e.message);
3346 assert!(e.message.contains("a.c"), "{}", e.message);
3347 }
3348
3349 /// `--offline` on its own is accepted and changes nothing, because an ordinary compile
3350 /// downloads nothing with or without it. A build that passes it everywhere is the case this is
3351 /// for, and it must not lose the compilation it was passed beside.
3352 #[test]
3353 fn offline_on_a_compilation_is_the_same_compilation() {
3354 let (opts, plan) = compile(&["-c", "--offline", "a.c"]);
3355 let (plain, without) = compile(&["-c", "a.c"]);
3356 assert_eq!(opts.target, plain.target);
3357 assert_eq!(plan.jobs.len(), without.jobs.len());
3358 assert_eq!(plan.jobs[0].output, without.jobs[0].output);
3359 }
3360
3361 #[test]
3362 fn an_unknown_flag_is_an_error_rather_than_a_shrug() {
3363 let e = parse_args(&args(&["-fno-such-thing", "a.c"])).unwrap_err();
3364 assert!(e.message.contains("unknown option"), "{}", e.message);
3365 }
3366
3367 /// `-fpermissive` and the flag that turns it back off, which a build writes beside it when
3368 /// one directory needs the older rules and the rest of the tree does not.
3369 #[test]
3370 fn permissive_reads_in_both_directions_and_the_last_one_wins() {
3371 let (opts, _) = compile(&["-c", "a.c"]);
3372 assert!(!opts.permissive, "off unless it is asked for");
3373
3374 let (opts, _) = compile(&["-c", "-fpermissive", "a.c"]);
3375 assert!(opts.permissive);
3376
3377 let (opts, _) = compile(&["-c", "-fpermissive", "-fno-permissive", "a.c"]);
3378 assert!(!opts.permissive);
3379 }
3380
3381 #[test]
3382 fn asking_for_nested_functions_is_told_why_it_is_not_coming() {
3383 let e = parse_args(&args(&["-fnested-functions", "a.c"])).unwrap_err();
3384 assert!(e.message.contains("trampoline"), "{}", e.message);
3385 assert!(parse_args(&args(&["-fno-nested-functions", "a.c"])).is_ok());
3386 }
3387
3388 #[test]
3389 fn the_flag_every_configure_script_writes_is_taken() {
3390 // All four spellings, because a build writes whichever one its macros picked and a
3391 // compiler that takes three of them is a compiler that fails on the fourth.
3392 for flag in ["-fPIC", "-fpic", "-fPIE", "-fpie"] {
3393 let (opts, _) = compile(&["-c", flag, "a.c"]);
3394 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3395 }
3396 }
3397
3398 #[test]
3399 fn a_table_is_written_unless_the_build_says_nothing_will_walk_it() {
3400 let (opts, _) = compile(&["-c", "a.c"]);
3401 assert!(opts.unwinds(), "the default is off");
3402 let (opts, _) = compile(&["-c", "-fno-asynchronous-unwind-tables", "a.c"]);
3403 assert!(!opts.unwinds(), "the build was not taken at its word");
3404 let (opts, _) = compile(&[
3405 "-c",
3406 "-fno-asynchronous-unwind-tables",
3407 "-fasynchronous-unwind-tables",
3408 "a.c",
3409 ]);
3410 assert!(opts.unwinds(), "the last flag did not win");
3411 // The weaker request, which the same table answers, so a line that asks for a table and
3412 // against an asynchronous one gets one. That is gcc's arrangement and it turns up when a
3413 // build turns the asynchronous one off globally and a directory asks for a table back.
3414 let (opts, _) =
3415 compile(&["-c", "-fno-asynchronous-unwind-tables", "-funwind-tables", "a.c"]);
3416 assert!(opts.unwinds(), "the weaker request was dropped");
3417 let (opts, _) = compile(&["-c", "-fno-unwind-tables", "a.c"]);
3418 assert!(opts.unwinds(), "the weaker negative turned off the stronger request");
3419 let (opts, _) =
3420 compile(&["-c", "-fno-unwind-tables", "-fno-asynchronous-unwind-tables", "a.c"]);
3421 assert!(!opts.unwinds(), "both were turned off and one stayed on");
3422 }
3423
3424 #[test]
3425 fn the_flags_that_describe_what_this_compiler_already_does_are_taken() {
3426 // Every one of these is on a real build line somewhere and every one of them was an
3427 // unknown option. What they have in common is that the answer rucc gives is the answer
3428 // they ask for, so there is nothing to implement and nothing to refuse.
3429 for flag in [
3430 "-fno-common",
3431 "-fstrict-aliasing",
3432 "-fno-strict-aliasing",
3433 "-fdelete-null-pointer-checks",
3434 "-fno-delete-null-pointer-checks",
3435 "-frounding-math",
3436 "-fno-rounding-math",
3437 "-fexcess-precision=standard",
3438 "-fexcess-precision=fast",
3439 "-fexcess-precision=16",
3440 "-pipe",
3441 "-fdiagnostics-color",
3442 "-fno-diagnostics-color",
3443 "-fdiagnostics-color=always",
3444 "-fdiagnostics-color=never",
3445 "-fdiagnostics-color=auto",
3446 ] {
3447 let (opts, _) = compile(&["-c", flag, "a.c"]);
3448 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3449 }
3450 }
3451
3452 #[test]
3453 fn whether_an_exception_is_looked_at_is_kept_and_defaults_to_gccs_answer() {
3454 let (opts, _) = compile(&["-c", "a.c"]);
3455 assert!(opts.trapping_math, "the default was not gcc's");
3456 let (opts, _) = compile(&["-c", "-fno-trapping-math", "a.c"]);
3457 assert!(!opts.trapping_math);
3458 let (opts, _) = compile(&["-c", "-ftrapping-math", "a.c"]);
3459 assert!(opts.trapping_math, "spelling out the default turned it off");
3460 // The last one written wins, which is how a build line that inherits a flag from one
3461 // place and overrides it in another is read.
3462 let (opts, _) = compile(&["-c", "-fno-trapping-math", "-ftrapping-math", "a.c"]);
3463 assert!(opts.trapping_math);
3464 }
3465
3466 /// The flags a torture program writes on its own `dg-options` line, which is where most of
3467 /// these come from: a program reduced from a miscompilation names the pass that miscompiled
3468 /// it. Eighteen programs in the suite stopped on the driver before anything read them, and
3469 /// tamnd/rucc#1019 is the list.
3470 #[test]
3471 fn the_flags_that_name_a_pass_of_gccs_own_are_taken_and_dropped() {
3472 for flag in [
3473 "-fno-tree-ccp",
3474 "-fno-tree-dominator-opts",
3475 "-fno-tree-vrp",
3476 "-fno-tree-bit-ccp",
3477 "-fno-tree-coalesce-vars",
3478 "-ftree-vectorize",
3479 "-ftree-loop-distribution",
3480 "-fno-ipa-cp",
3481 "-fipa-pta",
3482 "-fmodulo-sched",
3483 "-fno-vect-cost-model",
3484 "-fvect-cost-model=unlimited",
3485 "-fsimd-cost-model=cheap",
3486 "-fexpensive-optimizations",
3487 "-fno-early-inlining",
3488 "-fno-inline",
3489 "-finline-functions",
3490 "-foptimize-strlen",
3491 "-fno-ira-share-spill-slots",
3492 ] {
3493 let (opts, _) = compile(&["-c", flag, "a.c"]);
3494 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3495 assert!(opts.passes.is_empty(), "{flag} named a pass of gcc's and not one of ours");
3496 }
3497 }
3498
3499 /// The two namespaces are taken whole, so a name neither this test nor gcc 16 has heard of
3500 /// goes the same way as the ones above rather than stopping a build on the day gcc adds it.
3501 #[test]
3502 fn a_pass_name_in_either_family_is_taken_whether_or_not_it_is_one_gcc_has() {
3503 for flag in ["-ftree-no-such-pass", "-fno-ipa-no-such-pass"] {
3504 let (opts, _) = compile(&["-c", flag, "a.c"]);
3505 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3506 }
3507 }
3508
3509 /// A pass this compiler has keeps its flag, since the arms that read the registry are above
3510 /// the family arms. `dce` is the one both compilers have a name for, and `execute/pr97421-2.c`
3511 /// is the program that writes it.
3512 #[test]
3513 fn a_pass_name_this_compiler_has_is_still_read_as_a_pass() {
3514 let (opts, _) = compile(&["-c", "-fno-dce", "a.c"]);
3515 assert_eq!(opts.passes, vec![("dce".to_owned(), false)]);
3516 }
3517
3518 /// gcc's name for the unroller reaches the unroller, in both directions. libtommath puts
3519 /// `-funroll-loops` in `CFLAGS` unconditionally, and before this it was an unknown option and
3520 /// the build stopped on its first file.
3521 #[test]
3522 fn the_gcc_spelling_of_the_unroller_turns_the_unroller_on_and_off() {
3523 let (opts, _) = compile(&["-c", "-funroll-loops", "a.c"]);
3524 assert_eq!(opts.passes, vec![("unroll".to_owned(), true)]);
3525 let (opts, _) = compile(&["-c", "-fno-unroll-loops", "a.c"]);
3526 assert_eq!(opts.passes, vec![("unroll".to_owned(), false)]);
3527 }
3528
3529 /// The encoding of the source is not a question about speed, so the one name that describes
3530 /// what the preprocessor does is taken and every other name is refused.
3531 #[test]
3532 fn the_input_charset_is_taken_when_it_names_the_one_that_is_read() {
3533 for flag in ["-finput-charset=utf-8", "-finput-charset=UTF-8", "-finput-charset=utf8"] {
3534 let (opts, _) = compile(&["-c", flag, "a.c"]);
3535 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3536 }
3537
3538 let e = parse_args(&args(&["-c", "-finput-charset=latin1", "a.c"])).unwrap_err();
3539 assert!(e.message.contains("latin1"), "{}", e.message);
3540 assert!(e.message.contains("UTF-8"), "what is read is worth saying: {}", e.message);
3541 }
3542
3543 /// The other half of the same rule. Each of these changes what the program does rather than
3544 /// how fast it does it, so each is refused with the reason, and the negative of each is what
3545 /// happens anyway and is taken.
3546 #[test]
3547 fn the_three_that_change_the_answer_are_refused_and_their_negatives_are_taken() {
3548 for (flag, word) in [
3549 ("-ffast-math", "__FAST_MATH__"),
3550 ("-fnon-call-exceptions", "landing pad"),
3551 ("-finstrument-functions", "__cyg_profile_func_enter"),
3552 ] {
3553 let e = parse_args(&args(&["-c", flag, "a.c"])).unwrap_err();
3554 assert!(e.message.contains(word), "{flag}: {}", e.message);
3555 assert!(!e.message.contains("unknown option"), "{flag} deserves a reason");
3556
3557 let off = format!("-fno-{}", flag.trim_start_matches("-f"));
3558 let (opts, _) = compile(&["-c", &off, "a.c"]);
3559 assert_eq!(opts.emit, EmitKind::Object, "{off}");
3560 }
3561 }
3562
3563 #[test]
3564 fn asking_the_linker_to_merge_tentative_definitions_is_told_why_it_is_not_coming() {
3565 // The one of that family that is a request rather than a description, and it is a real
3566 // difference: two files each writing `int g;` link under it and do not without it.
3567 let e = parse_args(&args(&["-fcommon", "a.c"])).unwrap_err();
3568 assert!(e.message.contains(".bss"), "{}", e.message);
3569 assert!(e.message.contains("extern"), "the way out is worth saying: {}", e.message);
3570 }
3571
3572 #[test]
3573 fn asking_for_position_dependent_code_is_told_why_it_is_not_coming() {
3574 for flag in ["-fno-pic", "-fno-pie"] {
3575 let e = parse_args(&args(&[flag, "a.c"])).unwrap_err();
3576 assert!(e.message.contains("global offset table"), "{flag}: {}", e.message);
3577 // The one it may have meant, since the two are a letter apart and one of them is
3578 // about linking and is taken.
3579 assert!(e.message.contains("-no-pie"), "{flag}: {}", e.message);
3580 }
3581 }
3582
3583 #[test]
3584 fn an_unsupported_target_names_itself() {
3585 let e = parse_args(&args(&["--target=sparc64-linux-gnu", "a.c"])).unwrap_err();
3586 assert!(e.message.contains("sparc64"), "{}", e.message);
3587 }
3588
3589 #[test]
3590 fn no_inputs_is_an_error_but_print_config_needs_none() {
3591 assert!(parse_args(&args(&[])).is_err());
3592 assert!(matches!(parse_args(&args(&["--print-config"])), Ok(Action::PrintConfig(_))));
3593 }
3594
3595 #[test]
3596 fn print_config_reports_the_target_it_was_given_not_the_host() {
3597 let a = parse_args(&args(&["--print-config", "--target=riscv64-linux-musl"])).unwrap();
3598 let Action::PrintConfig(opts) = a else { panic!("expected a configuration dump") };
3599 let text = print_config(&opts);
3600 assert!(text.contains("target: riscv64-unknown-linux-musl"), "{text}");
3601 assert!(text.contains("char-signed: false"), "{text}");
3602 assert!(text.contains("object-format: elf"), "{text}");
3603 assert!(text.contains("va-list: void-pointer"), "{text}");
3604 // RISC-V has a register file and this compiler has not written it down yet, and the
3605 // dump says which of those two it is rather than leaving the line out.
3606 assert!(text.contains("registers: none"), "{text}");
3607 assert!(text.contains("timing-model: none"), "{text}");
3608 }
3609
3610 /// The model the schedule was chosen with, which is a receipt anybody comparing two runs of a
3611 /// benchmark needs: two numbers that disagree are usually two models and not two compilers.
3612 #[test]
3613 fn print_config_names_the_model_the_schedule_was_chosen_with() {
3614 let opts = Options::new("x86_64-unknown-linux-gnu".parse().unwrap());
3615 let text = print_config(&opts);
3616 let line = text.lines().find(|l| l.starts_with("timing-model:")).expect("the model");
3617 assert!(line.contains("Skylake"), "{line}");
3618 assert!(line.contains("published"), "a sentence saying where it came from: {line}");
3619 }
3620
3621 #[test]
3622 fn print_config_has_one_key_per_line_and_a_fixed_order() {
3623 let opts = Options::new("x86_64-unknown-linux-gnu".parse().unwrap());
3624 let text = print_config(&opts);
3625 let keys: Vec<&str> =
3626 text.lines().map(|l| l.split(':').next().unwrap_or_default()).collect();
3627 assert_eq!(keys[0], "version");
3628 assert_eq!(keys[1], "target");
3629 assert_eq!(keys.len(), 26);
3630 assert!(text.ends_with('\n'));
3631 }
3632
3633 #[test]
3634 fn the_safety_tier_is_read_off_the_command_line_and_a_wrong_one_is_refused() {
3635 let (opts, _) = compile(&["a.c"]);
3636 assert_eq!(opts.safety, rucc_session::Safety::Off);
3637
3638 for (flag, tier) in [
3639 ("-fsafety=detect", rucc_session::Safety::Detect),
3640 ("-fsafety=enforce", rucc_session::Safety::Enforce),
3641 ("-fsafety=kernel", rucc_session::Safety::Kernel),
3642 ("-fsafety=off", rucc_session::Safety::Off),
3643 ] {
3644 let (opts, _) = compile(&[flag, "a.c"]);
3645 assert_eq!(opts.safety, tier, "{flag}");
3646 }
3647
3648 // The last one wins, the way every other repeated flag on this command line does.
3649 let (opts, _) = compile(&["-fsafety=enforce", "-fsafety=off", "a.c"]);
3650 assert_eq!(opts.safety, rucc_session::Safety::Off);
3651
3652 // A misspelled tier is refused rather than ignored. Silently compiling without the
3653 // monitor a build asked for is the one failure mode this feature cannot have.
3654 let e = parse_args(&args(&["-fsafety=on", "a.c"])).unwrap_err();
3655 assert!(e.message.contains("is not a safety tier"), "{}", e.message);
3656 assert!(parse_args(&args(&["-fsafety", "a.c"])).is_err());
3657 }
3658
3659 #[test]
3660 fn the_padding_mode_is_read_off_the_command_line_and_a_wrong_one_is_refused() {
3661 // The default is the one section 9.3 of document 09 gives library code, which is that
3662 // padding does not participate, so a record filled a member at a time is not reported.
3663 let (opts, _) = compile(&["a.c"]);
3664 assert_eq!(opts.padding, rucc_session::Padding::Ignored);
3665
3666 let (opts, _) = compile(&["-fsafety=detect", "-fsafety-init=padding", "a.c"]);
3667 assert_eq!(opts.padding, rucc_session::Padding::Tracked);
3668
3669 let (opts, _) = compile(&["-fsafety-init=padding", "-fsafety-init=nopadding", "a.c"]);
3670 assert_eq!(opts.padding, rucc_session::Padding::Ignored);
3671
3672 // The tier is still a tier. A flag whose name starts the same way must not be eaten by
3673 // the one above it, which is the thing worth pinning about a pair of names like these.
3674 let (opts, _) = compile(&["-fsafety-init=padding", "a.c"]);
3675 assert_eq!(opts.safety, rucc_session::Safety::Off);
3676
3677 let e = parse_args(&args(&["-fsafety-init=some", "a.c"])).unwrap_err();
3678 assert!(e.message.contains("is not a padding mode"), "{}", e.message);
3679 }
3680
3681 #[test]
3682 fn whether_a_write_has_to_stay_inside_its_member_is_read_off_the_command_line() {
3683 // Off by default, because a store to allocated storage sets its effective type and C 6.5
3684 // lets a program reuse a buffer as something else. Row S4 is a build opting out of that.
3685 let (opts, _) = compile(&["a.c"]);
3686 assert_eq!(opts.subobject, rucc_session::Subobject::Off);
3687
3688 let (opts, _) = compile(&["-fsafety=detect", "-fsafety-subobject", "a.c"]);
3689 assert_eq!(opts.subobject, rucc_session::Subobject::Members);
3690
3691 let (opts, _) = compile(&["-fsafety-subobject", "-fno-safety-subobject", "a.c"]);
3692 assert_eq!(opts.subobject, rucc_session::Subobject::Off);
3693
3694 // It takes no value. The form that would take one is the strict reading of section 9.4,
3695 // which is not written yet, so say so rather than accept a spelling that does nothing.
3696 let e = parse_args(&args(&["-fsafety-subobject=strict", "a.c"])).unwrap_err();
3697 assert!(e.message.contains("tamnd/rucc#967"), "{}", e.message);
3698 }
3699
3700 #[test]
3701 fn whether_two_restrict_pointers_may_meet_is_read_off_the_command_line() {
3702 // Off by default, because the record a block keeps is the union of what each pointer
3703 // reached, so two pointers striding through one array without landing on the same byte are
3704 // reported and by the letter of the standard those are different objects. Row Y8 is a build
3705 // deciding it would rather know.
3706 let (opts, _) = compile(&["a.c"]);
3707 assert_eq!(opts.promise, rucc_session::Promise::Off);
3708
3709 let (opts, _) = compile(&["-fsafety=detect", "-fsafety-restrict", "a.c"]);
3710 assert_eq!(opts.promise, rucc_session::Promise::Blocks);
3711
3712 let (opts, _) = compile(&["-fsafety-restrict", "-fno-safety-restrict", "a.c"]);
3713 assert_eq!(opts.promise, rucc_session::Promise::Off);
3714
3715 // The tier is still a tier, which is the thing worth pinning about a pair of names where
3716 // one is the front of the other.
3717 let (opts, _) = compile(&["-fsafety-restrict", "a.c"]);
3718 assert_eq!(opts.safety, rucc_session::Safety::Off);
3719
3720 let e = parse_args(&args(&["-fsafety-restrict=blocks", "a.c"])).unwrap_err();
3721 assert!(e.message.contains("takes no value"), "{}", e.message);
3722 }
3723
3724 #[test]
3725 fn safety_races_takes_a_mode_and_defaults_to_watching_nothing() {
3726 // Three modes rather than a bare flag, because section 9.5 gives two answers that record
3727 // the same thing and report different classes, so a flag with no value could not say which
3728 // was wanted. Off by default for the reason on `rucc_session::Races`, which is not a cost
3729 // argument: this is the one plane where an edge nobody interposed costs a false report.
3730 let (opts, _) = compile(&["a.c"]);
3731 assert_eq!(opts.races, rucc_session::Races::Off);
3732
3733 let (opts, _) = compile(&["-fsafety-races=metadata", "a.c"]);
3734 assert_eq!(opts.races, rucc_session::Races::Metadata);
3735
3736 let (opts, _) = compile(&["-fsafety-races=pointer", "a.c"]);
3737 assert_eq!(opts.races, rucc_session::Races::Pointer);
3738
3739 // Last one wins, as it does for every other mode flag here.
3740 let (opts, _) = compile(&["-fsafety-races=pointer", "-fno-safety-races", "a.c"]);
3741 assert_eq!(opts.races, rucc_session::Races::Off);
3742
3743 let e = parse_args(&args(&["-fsafety-races=all", "a.c"])).unwrap_err();
3744 assert!(e.message.contains("off, metadata or pointer"), "{}", e.message);
3745 }
3746
3747 #[test]
3748 fn print_pipeline_answers_with_the_passes_the_level_asked_for() {
3749 let a = parse_args(&args(&["--print-pipeline", "-O2"])).unwrap();
3750 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3751 let text = print_pipeline(&opts);
3752 assert!(text.starts_with("level: -O2\n"), "{text}");
3753 assert!(text.contains("fold"), "{text}");
3754
3755 let a = parse_args(&args(&["--print-pipeline"])).unwrap();
3756 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3757 // Two passes run at `-O0` and neither is an optimization. The first moves what
3758 // `__builtin_expect` said onto the branch and takes the instruction away, so that nothing
3759 // past the optimizer has to know the instruction exists. The second removes code nothing
3760 // reaches. See issue 359.
3761 assert!(print_pipeline(&opts).contains("1: expect,"), "{}", print_pipeline(&opts));
3762 assert!(print_pipeline(&opts).contains("2: simplify-cfg,"), "{}", print_pipeline(&opts));
3763
3764 let a = parse_args(&args(&["--print-pipeline", "-fno-simplify-cfg"])).unwrap();
3765 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3766 // The second turns off and the first does not, because nothing below the optimizer lowers
3767 // what it removes, so `-fno-expect` is a compile that stops rather than one that runs.
3768 let text = print_pipeline(&opts);
3769 assert!(text.contains("1: expect,"), "{text}");
3770 assert!(!text.contains("simplify-cfg"), "{text}");
3771 }
3772
3773 #[test]
3774 fn print_pipeline_takes_the_toggles_into_account() {
3775 let a = parse_args(&args(&["--print-pipeline", "-O2", "-fno-fold"])).unwrap();
3776 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3777 let text = print_pipeline(&opts);
3778 // The one that was named is gone and the rest of the level is not, which is the whole
3779 // of what a toggle promises.
3780 assert!(!text.contains("fold"), "{text}");
3781 assert!(text.contains("dce"), "{text}");
3782
3783 // Every pass the compiler has, named off. Built from the registry rather than written
3784 // out, so a pass added later is turned off here too and this keeps testing the thing it
3785 // is about, which is that the toggles can empty a level down to the passes that are not
3786 // optional. Those are named, because a listing that is all of them is a level nobody
3787 // emptied and the assertion would pass while saying nothing.
3788 let mut off = vec!["--print-pipeline".to_owned(), "-O2".to_owned()];
3789 off.extend(rucc_opt::PASSES.iter().map(|p| format!("-fno-{}", p.name())));
3790 let spelled: Vec<&str> = off.iter().map(String::as_str).collect();
3791 let a = parse_args(&args(&spelled)).unwrap();
3792 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3793 let text = print_pipeline(&opts);
3794 let left: Vec<&str> =
3795 rucc_opt::PASSES.iter().filter(|p| p.required()).map(|p| p.name()).collect();
3796 assert_eq!(left, vec!["expect"], "{text}");
3797 for (at, name) in left.iter().enumerate() {
3798 assert!(text.contains(&format!("{}: {name},", at + 1)), "{text}");
3799 }
3800 assert!(!text.contains("dce"), "{text}");
3801 }
3802
3803 #[test]
3804 fn print_pipeline_says_when_a_budget_will_stop_the_run_short() {
3805 let a = parse_args(&args(&["--print-pipeline", "-O2"])).unwrap();
3806 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3807 assert!(!print_pipeline(&opts).contains("global fuel"));
3808
3809 let a = parse_args(&args(&["--print-pipeline", "-O2", "-fpass-fuel-global=4"])).unwrap();
3810 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3811 let text = print_pipeline(&opts);
3812 // Because the listing is the answer to what this compilation will do, and a run that
3813 // stops after four rewrites is not doing what the level says it does.
3814 assert!(text.contains("global fuel: 4"), "{text}");
3815 }
3816
3817 /// A pass is turned on and off by its own name, and the order the flags were given in is
3818 /// kept, because the last spelling of a name is the one that decides.
3819 #[test]
3820 fn a_pass_is_named_by_dash_f_and_unnamed_by_dash_f_no() {
3821 let (opts, _) = compile(&["-c", "-O0", "-ffold", "-fno-fold", "-ffold", "a.c"]);
3822 assert_eq!(
3823 opts.passes,
3824 [("fold".to_owned(), true), ("fold".to_owned(), false), ("fold".to_owned(), true)]
3825 );
3826
3827 let e = parse_args(&args(&["-fno-such-pass", "a.c"])).unwrap_err();
3828 assert!(e.message.contains("unknown option"), "{}", e.message);
3829 }
3830
3831 #[test]
3832 fn pass_fuel_names_a_pass_and_a_count_and_refuses_anything_else() {
3833 let (opts, _) = compile(&["-c", "-O2", "-fpass-fuel=fold=3", "a.c"]);
3834 assert_eq!(opts.pass_fuel, [("fold".to_owned(), 3)]);
3835
3836 let e = parse_args(&args(&["-fpass-fuel=fold", "a.c"])).unwrap_err();
3837 assert!(e.message.contains("<pass>=<count>"), "{}", e.message);
3838 let e = parse_args(&args(&["-fpass-fuel=nosuch=3", "a.c"])).unwrap_err();
3839 assert!(e.message.contains("--print-pipeline"), "{}", e.message);
3840 let e = parse_args(&args(&["-fpass-fuel=fold=lots", "a.c"])).unwrap_err();
3841 assert!(e.message.contains("not a number"), "{}", e.message);
3842 }
3843
3844 #[test]
3845 fn global_pass_fuel_is_a_count_on_its_own_and_defaults_to_no_limit() {
3846 let (opts, _) = compile(&["-c", "-O2", "a.c"]);
3847 assert_eq!(opts.pass_fuel_global, None);
3848
3849 let (opts, _) = compile(&["-c", "-O2", "-fpass-fuel-global=12", "a.c"]);
3850 assert_eq!(opts.pass_fuel_global, Some(12));
3851 // And it is not the per pass flag with a longer name, so neither spelling swallows the
3852 // other.
3853 assert!(opts.pass_fuel.is_empty());
3854
3855 let e = parse_args(&args(&["-fpass-fuel-global=lots", "a.c"])).unwrap_err();
3856 assert!(e.message.contains("not a number"), "{}", e.message);
3857 }
3858
3859 #[test]
3860 fn a_gate_names_a_pass_and_optionally_the_functions_it_covers() {
3861 let (opts, _) = compile(&["-c", "-O2", "-fdisable-fold", "-fenable-fold=2-4,main", "a.c"]);
3862 assert_eq!(
3863 opts.pass_gates,
3864 [(false, "fold".to_owned()), (true, "fold=2-4,main".to_owned())],
3865 "the order is what decides, so it has to survive the parse"
3866 );
3867
3868 let e = parse_args(&args(&["-fdisable-nosuch", "a.c"])).unwrap_err();
3869 assert!(e.message.contains("--print-pipeline"), "{}", e.message);
3870 let e = parse_args(&args(&["-fenable-fold=9-2", "a.c"])).unwrap_err();
3871 assert!(e.message.contains("ends before it starts"), "{}", e.message);
3872 let e = parse_args(&args(&["-fdisable-fold=", "a.c"])).unwrap_err();
3873 assert!(e.message.contains("is empty"), "{}", e.message);
3874 }
3875
3876 #[test]
3877 fn the_pipeline_listing_says_which_passes_a_gate_touched() {
3878 let (opts, _) = compile(&["-c", "-O2", "-fdisable-fold=main", "a.c"]);
3879 let text = print_pipeline(&opts);
3880 assert!(text.contains("fold, "), "{text}");
3881 assert!(text.contains("[off for main]"), "{text}");
3882 }
3883
3884 /// The spelling is checked while the arguments are read, because a dump that names a pass
3885 /// this compiler does not have is a typo, and a typo found after the compilation has run is
3886 /// found too late to be any use.
3887 #[test]
3888 fn a_dump_is_checked_when_it_is_asked_for_rather_than_when_it_is_taken() {
3889 let (opts, _) = compile(&["-c", "-O2", "-fdump-ir=all", "-fdump-ir=after-fold", "a.c"]);
3890 assert_eq!(opts.dump_ir, ["all", "after-fold"]);
3891
3892 let e = parse_args(&args(&["-fdump-ir=after-nosuch", "a.c"])).unwrap_err();
3893 assert!(e.message.contains("nosuch"), "{}", e.message);
3894 assert!(parse_args(&args(&["-fdump-ir=sideways-fold", "a.c"])).is_err());
3895 }
3896
3897 /// Every spelling `-fopt-info` takes, and the one it does not.
3898 ///
3899 /// The keywords are checked here for the same reason a dump's pass name is: a person who
3900 /// misspelled one gets no output, and no output is also what a compilation where nothing
3901 /// happened looks like. Telling those two apart is the entire reason to reach for this flag.
3902 #[test]
3903 fn opt_info_takes_kinds_and_a_file_and_refuses_a_kind_it_does_not_have() {
3904 let (opts, _) = compile(&["-c", "-O2", "-fopt-info", "a.c"]);
3905 assert_eq!(opts.opt_info, [""], "a bare flag asks for the rewrites");
3906 assert_eq!(opts.opt_info_file, None, "and goes to standard error");
3907
3908 let (opts, _) = compile(&["-c", "-O2", "-fopt-info-missed-note", "a.c"]);
3909 assert_eq!(opts.opt_info, ["missed-note"]);
3910
3911 // Two flags add up rather than the second replacing the first, and the file is the last
3912 // one that named a file, which is how GCC treats both.
3913 let (opts, _) =
3914 compile(&["-c", "-O2", "-fopt-info-missed=one.txt", "-fopt-info-all=two.txt", "a.c"]);
3915 assert_eq!(opts.opt_info, ["missed", "all"]);
3916 assert_eq!(opts.opt_info_file.as_deref(), Some("two.txt"));
3917
3918 let e = parse_args(&args(&["-fopt-info-vectorized", "a.c"])).unwrap_err();
3919 assert!(e.message.contains("vectorized"), "{}", e.message);
3920 assert!(e.message.contains("`missed`"), "{}", e.message);
3921 let e = parse_args(&args(&["-fopt-info-missed=", "a.c"])).unwrap_err();
3922 assert!(e.message.contains("no file"), "{}", e.message);
3923 }
3924
3925 #[test]
3926 fn verify_each_is_unstable_and_off_unless_it_was_asked_for() {
3927 let (opts, _) = compile(&["-c", "-Zverify-each", "a.c"]);
3928 assert!(opts.verify_each);
3929 assert!(!USAGE.contains("verify-each"), "an unstable option stays out of the usage text");
3930 }
3931
3932 #[test]
3933 fn dash_o_needs_an_argument() {
3934 let e = parse_args(&args(&["a.c", "-o"])).unwrap_err();
3935 assert_eq!(e.message, "-o requires an argument");
3936 }
3937
3938 #[test]
3939 fn dash_d_and_dash_u_are_read_joined_or_separated_and_keep_their_order() {
3940 let (opts, _) = compile(&["-DFOO=1", "-D", "BAR", "-UBAZ", "-U", "QUX", "a.c"]);
3941 assert_eq!(opts.defines, ["FOO=1", "BAR"]);
3942 assert_eq!(opts.undefines, ["BAZ", "QUX"]);
3943 }
3944
3945 #[test]
3946 fn the_include_flags_land_on_the_chain_each_one_names() {
3947 // A sysroot with nothing under it, so that the library's own directories are the
3948 // same on every machine this test runs on, which is none of them.
3949 let (opts, _) = compile(&[
3950 "-Ii",
3951 "-iquote",
3952 "q",
3953 "-isystem",
3954 "sys",
3955 "-idirafter",
3956 "after",
3957 "--sysroot=/nowhere-at-all",
3958 "a.c",
3959 ]);
3960 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
3961 // The compiler's own headers sit after every `-isystem` and before `-idirafter`,
3962 // which is where GCC puts its own: a directory the user named outranks ours.
3963 assert_eq!(dirs, ["q", "i", "sys", runtime::DIR, "after"]);
3964 assert!(!opts.search.dirs()[1].is_system);
3965 assert!(opts.search.dirs()[2].is_system);
3966 }
3967
3968 #[test]
3969 fn the_librarys_headers_come_after_the_compilers_own_and_go_away_with_them() {
3970 // Which machine this runs on decides what is on the path, so the test is about the
3971 // order rather than about the names: ours is on it, the library's follow it, and
3972 // `-nostdinc` is the one flag that takes both halves of the pair off at once.
3973 let (opts, _) = compile(&["a.c"]);
3974 let dirs = opts.search.dirs();
3975 let ours = dirs.iter().position(|d| d.path.to_str() == Some(runtime::DIR));
3976 assert_eq!(ours, Some(0), "{dirs:?}");
3977 assert!(dirs[1..].iter().all(|d| d.is_system), "{dirs:?}");
3978 let (bare, _) = compile(&["-nostdinc", "a.c"]);
3979 assert!(bare.search.dirs().is_empty(), "{:?}", bare.search.dirs());
3980 }
3981
3982 #[test]
3983 fn a_sysroot_moves_the_librarys_directories_and_nothing_else() {
3984 let (opts, _) = compile(&["-isystem", "sys", "--sysroot=/nowhere-at-all", "a.c"]);
3985 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
3986 assert_eq!(dirs, ["sys", runtime::DIR]);
3987 }
3988
3989 #[test]
3990 fn a_cross_compile_reads_the_targets_own_headers_rather_than_the_ones_next_door() {
3991 // The target is not the machine this test runs on wherever it runs, so the answer is the
3992 // same on all of them: the libc's two include directories for that target, the kernel's
3993 // two, and nothing from here. A header read from here is the quiet failure of section 8.5, a
3994 // program that builds on the build machine and is wrong everywhere else.
3995 let (opts, _) = compile(&["--target=riscv64-linux-musl", "-c", "a.c"]);
3996 let dirs: Vec<&std::path::Path> =
3997 opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
3998 let root = cache::dir().join("sysroots").join("riscv64-linux-musl");
3999 let kernel = cache::dir().join("kernel-headers");
4000 assert_eq!(dirs.len(), 5, "{dirs:?}");
4001 assert_eq!(dirs[0], std::path::Path::new(runtime::DIR));
4002 assert_eq!(dirs[1], root.join("include").join("riscv64"));
4003 assert_eq!(dirs[2], root.join("include").join("generic"));
4004 // The kernel's, which are beside the sysroots rather than inside one, because every target
4005 // that shares an architecture reads the same files.
4006 assert_eq!(dirs[3], kernel.join("riscv"));
4007 assert_eq!(dirs[4], kernel.join("generic"));
4008 }
4009
4010 #[test]
4011 fn a_cross_compile_to_something_that_is_not_linux_reads_no_kernel_headers() {
4012 // The other side of the same answer. Windows has its own system headers and no `linux/` at
4013 // all, so the list is the libc's own and the question never arises, which is the `None` that
4014 // `link::cross_kernel` returns rather than a directory nothing would be found in.
4015 //
4016 // The libc's own is one directory rather than two here, because mingw-w64 publishes a single
4017 // header tree for every architecture and `Sysroot::splits_by_arch` says so.
4018 let (opts, _) = compile(&["--target=x86_64-pc-windows-gnu", "-c", "a.c"]);
4019 let dirs: Vec<&std::path::Path> =
4020 opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
4021 assert_eq!(dirs.len(), 2, "{dirs:?}");
4022 assert!(!dirs.iter().any(|dir| dir.ends_with("kernel-headers")), "{dirs:?}");
4023 }
4024
4025 #[test]
4026 fn the_glibc_version_macro_goes_with_the_bundled_tree_and_with_nothing_else() {
4027 // One tree serves every glibc release, so the release is what the target supplies, and the
4028 // condition is the same one that chose the directories. A host glibc and a tree somebody
4029 // named both define `__GLIBC_MINOR__` in their own `features.h`, and two definitions with
4030 // different values is a warning on every compilation of every file.
4031 //
4032 // The architecture is chosen against this machine's rather than written down, because the
4033 // bundled tree is only in effect for a target that is not this machine. The first version of
4034 // this test said x86_64-linux-gnu, which is a cross compile on a mac and this machine on a
4035 // Linux runner, so it passed here and failed there.
4036 let gnu = format!("--target={}-linux-gnu", cross_arch());
4037 let (bundled, _) = compile(&[&gnu, "-c", "a.c"]);
4038 assert_eq!(bundled.glibc_minor, Some(44));
4039 let pin = format!("{gnu}.2.28");
4040 let (pinned, _) = compile(&[&pin, "-c", "a.c"]);
4041 assert_eq!(pinned.glibc_minor, Some(28));
4042
4043 let (named, _) = compile(&[&gnu, "--sysroot=/nowhere-at-all", "-c", "a.c"]);
4044 assert_eq!(named.glibc_minor, None);
4045 let (none, _) = compile(&[&gnu, "-nostdinc", "-c", "a.c"]);
4046 assert_eq!(none.glibc_minor, None);
4047 let musl = format!("--target={}-linux-musl", cross_arch());
4048 let (musl, _) = compile(&[&musl, "-c", "a.c"]);
4049 assert_eq!(musl.glibc_minor, None);
4050
4051 // And this machine's own target gets nothing, whatever this machine is, because its headers
4052 // come from the machine and its own `features.h` defines the macro. On a glibc Linux box
4053 // that is the case this test had backwards; on a mac it is true for the other reason, which
4054 // is that Darwin is not a glibc target at all.
4055 if let Some(host) = Triple::host() {
4056 let native = format!("--target={}", host.tuple());
4057 let (native, _) = compile(&[&native, "-c", "a.c"]);
4058 assert_eq!(native.glibc_minor, None);
4059 }
4060 }
4061
4062 #[test]
4063 fn a_pinned_release_on_this_machines_own_target_reads_the_bundled_tree() {
4064 // The end to end half of the answer in `link::cross_for`. A release named for this machine's
4065 // own target is a cross compile, so the headers are the bundled tree's and the macro says
4066 // what was asked for rather than what this machine has.
4067 //
4068 // Only on a glibc box, because a release is a glibc release: a mac has no `__GLIBC_MINOR__`
4069 // to get wrong and nothing to pin. That makes this a test the Linux runners carry, which is
4070 // where the case lives.
4071 let Some(host) = Triple::host() else { return };
4072 if host.env != rucc_target::Env::Gnu {
4073 return;
4074 }
4075 let pin = format!("--target={}.2.28", host.tuple());
4076 let (opts, _) = compile(&[&pin, "-c", "a.c"]);
4077 assert_eq!(opts.glibc_minor, Some(28));
4078 let root = cache::dir().join("sysroots").join(format!("{}.2.28", host.tuple()));
4079 let dirs: Vec<&std::path::Path> =
4080 opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
4081 assert!(dirs.iter().any(|dir| dir.starts_with(&root)), "{dirs:?}");
4082 // And nothing of this machine's, which is the failure this was: a program compiled against
4083 // 2.44 declarations and told it was 2.28.
4084 assert!(!dirs.iter().any(|dir| *dir == std::path::Path::new("/usr/include")), "{dirs:?}");
4085 }
4086
4087 /// An architecture that is not this machine's, out of the three the driver has targets for.
4088 ///
4089 /// A test about the bundled sysroot has to name a target that is not the host, because a target
4090 /// that is the host reads the host's own headers and libraries. Asking which machine this is
4091 /// beats picking a row and hoping, and it is two lines.
4092 fn cross_arch() -> &'static str {
4093 match Triple::host().map(|host| host.arch) {
4094 Some(rucc_target::Arch::X86_64) => "aarch64",
4095 _ => "x86_64",
4096 }
4097 }
4098
4099 #[test]
4100 fn a_glibc_newer_than_the_bundled_tree_is_refused_by_name() {
4101 // Both versions in the message, because the two things a person can do about it are pin a
4102 // release the tree has and name a sysroot that has the one they asked for, and neither is a
4103 // choice they can make without knowing which release the tree is.
4104 //
4105 // Not this machine's architecture, for the reason the test above gives: the refusal is about
4106 // the bundled tree, and the bundled tree is not what a target that is this machine reads.
4107 let target = format!("--target={}-linux-gnu.2.99", cross_arch());
4108 let message = refused(&[&target, "-c", "a.c"]);
4109 assert!(message.contains("asked for glibc 2.99"), "{message}");
4110 assert!(message.contains("bundled headers are glibc 2.44"), "{message}");
4111 assert!(message.contains("--sysroot"), "{message}");
4112 }
4113
4114 #[test]
4115 fn a_sysroot_the_user_named_is_still_what_a_cross_compile_reads() {
4116 // The tree somebody assembled beats the one we would build, on the headers as on the
4117 // libraries. It is empty here, which is why the list comes out short: the directories under
4118 // it are checked for rather than assumed, and a tree that is not there offers nothing.
4119 let (opts, _) =
4120 compile(&["--target=riscv64-linux-musl", "--sysroot=/nowhere-at-all", "-c", "a.c"]);
4121 let dirs: Vec<&std::path::Path> =
4122 opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
4123 assert_eq!(dirs, [std::path::Path::new(runtime::DIR)]);
4124 }
4125
4126 #[test]
4127 fn dash_i_dash_moves_the_bracket_directories_into_the_quoted_chain() {
4128 let (opts, _) =
4129 compile(&["-Iinc1", "-iquote", "inc2", "-I-", "-Iinc3", "-nostdinc", "a.c"]);
4130 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
4131 assert_eq!(dirs, ["inc1", "inc2", "inc3"]);
4132 // An angled include sees only what came after the flag.
4133 assert_eq!(opts.search.start(IncludeForm::Angled), 2);
4134 assert!(!opts.search.searches_current_dir());
4135 }
4136
4137 #[test]
4138 fn the_prefix_flags_stick_what_iprefix_said_on_the_front_of_what_follows_it() {
4139 let (opts, _) = compile(&[
4140 "-iprefix",
4141 "/tools/",
4142 "-iwithprefix",
4143 "late",
4144 "-iwithprefixbefore",
4145 "early",
4146 "-iprefix",
4147 "/other/",
4148 "-iwithprefix",
4149 "last",
4150 "-nostdinc",
4151 "a.c",
4152 ]);
4153 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
4154 // `-iwithprefixbefore` is an `-I` and the other two are `-isystem`, which is where GCC
4155 // puts them rather than where its manual says it does.
4156 assert_eq!(dirs, ["/tools/early", "/tools/late", "/other/last"]);
4157 assert!(!opts.search.dirs()[0].is_system);
4158 assert!(opts.search.dirs()[1].is_system);
4159 }
4160
4161 #[test]
4162 fn the_files_named_on_the_command_line_keep_their_order_and_which_flag_named_them() {
4163 let (opts, _) =
4164 compile(&["-include", "one.h", "-imacros", "two.h", "-include", "3.h", "a.c"]);
4165 let names: Vec<&str> = opts.preincludes.iter().map(|p| p.name.as_str()).collect();
4166 assert_eq!(names, ["one.h", "two.h", "3.h"]);
4167 assert_eq!(opts.preincludes.iter().filter(|p| p.macros_only).count(), 1);
4168 }
4169
4170 #[test]
4171 fn nostdinc_takes_the_compilers_own_headers_off_the_path() {
4172 let (opts, _) = compile(&["-Ii", "-nostdinc", "a.c"]);
4173 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
4174 assert_eq!(dirs, ["i"]);
4175 }
4176
4177 #[test]
4178 fn the_dialect_flags_set_the_language_and_the_extensions_separately() {
4179 let (opts, _) = compile(&["-std=gnu11", "a.c"]);
4180 assert_eq!(opts.std, Std::C11);
4181 assert!(opts.gnu_extensions);
4182
4183 let (opts, _) = compile(&["-std=iso9899:1999", "a.c"]);
4184 assert_eq!(opts.std, Std::C99);
4185 assert!(!opts.gnu_extensions);
4186
4187 let (opts, _) = compile(&["-ansi", "a.c"]);
4188 assert_eq!(opts.std, Std::C89);
4189 assert!(!opts.gnu_extensions);
4190
4191 let e = parse_args(&args(&["-std=c94jr", "a.c"])).unwrap_err();
4192 assert!(e.message.contains("unknown dialect"), "{}", e.message);
4193 }
4194
4195 #[test]
4196 fn the_dump_letters_are_a_family_and_everything_else_beginning_with_d_is_not() {
4197 let (opts, _) = compile(&["-dM", "a.c"]);
4198 assert!(opts.dumps.macros);
4199
4200 // Packed, the way GCC takes them, and a letter in the family we have not written yet
4201 // is accepted and does nothing rather than failing a build.
4202 let (opts, _) = compile(&["-dDM", "a.c"]);
4203 assert!(opts.dumps.macros);
4204 let (opts, _) = compile(&["-dD", "a.c"]);
4205 assert!(!opts.dumps.macros);
4206
4207 let (opts, _) = compile(&["a.c"]);
4208 assert!(!opts.dumps.any());
4209
4210 // `-dumpversion` is a different flag that happens to start the same way, and it is read
4211 // as itself rather than as a dump of nothing.
4212 assert_eq!(printed(&["-dumpversion", "a.c"]), VERSION);
4213 }
4214
4215 #[test]
4216 fn the_gcc_version_claimed_is_a_flag_and_the_short_spellings_are_the_ones_people_write() {
4217 let (opts, _) = compile(&["a.c"]);
4218 assert_eq!(
4219 opts.gnuc,
4220 GnucVersion { major: 7, minor: 0, patch: 0 },
4221 "the lowest claim a modern glibc gives its own declarations to"
4222 );
4223
4224 let (opts, _) = compile(&["-fgnuc-version=15.1.0", "a.c"]);
4225 assert_eq!(opts.gnuc, GnucVersion { major: 15, minor: 1, patch: 0 });
4226
4227 // A missing component is zero. `gcc -dumpversion` says `15` on a release with no
4228 // patchlevel and a harness that pastes that back has to be understood.
4229 let (opts, _) = compile(&["-fgnuc-version=15", "a.c"]);
4230 assert_eq!(opts.gnuc, GnucVersion { major: 15, minor: 0, patch: 0 });
4231
4232 let (opts, _) = compile(&["-fgnuc-version=13.2", "a.c"]);
4233 assert_eq!(opts.gnuc, GnucVersion { major: 13, minor: 2, patch: 0 });
4234
4235 let e = parse_args(&args(&["-fgnuc-version=15.x", "a.c"])).unwrap_err();
4236 assert!(e.message.contains("minor that is not a number"), "{}", e.message);
4237
4238 let e = parse_args(&args(&["-fgnuc-version=1.2.3.4", "a.c"])).unwrap_err();
4239 assert!(e.message.contains("more than three"), "{}", e.message);
4240 }
4241
4242 #[test]
4243 fn pedantic_has_two_spellings_and_is_not_the_same_knob_as_the_dialect() {
4244 let (opts, _) = compile(&["-std=c17", "-pedantic", "a.c"]);
4245 assert!(opts.pedantic);
4246 assert_eq!(opts.std, Std::C17);
4247
4248 // The `-W` family's name for it, which is what a build that groups its warning flags
4249 // tends to write.
4250 let (opts, _) = compile(&["-Wpedantic", "a.c"]);
4251 assert!(opts.pedantic);
4252
4253 let (opts, _) = compile(&["-std=c17", "a.c"]);
4254 assert!(!opts.pedantic, "a dialect on its own does not diagnose an extension");
4255 }
4256
4257 #[test]
4258 fn dash_p_and_dash_ffreestanding_reach_the_options() {
4259 let (opts, _) = compile(&["-E", "-P", "-ffreestanding", "a.c"]);
4260 assert!(!opts.line_markers);
4261 assert!(!opts.hosted);
4262 assert_eq!(opts.emit, EmitKind::Preprocessed);
4263 }
4264
4265 /// The two ways a build says it means its own function by a name the C library also has.
4266 ///
4267 /// `-fno-builtin` is all of them and `-fno-builtin-<name>` is one, and the second is what a
4268 /// build writes when it means its own `memcpy` and the library's everything else. The name is
4269 /// kept as it was written and not checked against anything, because a program is allowed to
4270 /// mean something by a name this compiler has never heard of.
4271 #[test]
4272 fn the_builtin_flags_are_read_in_both_directions_and_one_name_at_a_time() {
4273 let (opts, _) = compile(&["-c", "a.c"]);
4274 assert!(opts.builtins, "a library name means the library function by default");
4275 assert!(opts.no_builtin.is_empty());
4276
4277 let (opts, _) = compile(&["-c", "-fno-builtin", "a.c"]);
4278 assert!(!opts.builtins);
4279
4280 let (opts, _) = compile(&["-c", "-fno-builtin", "-fbuiltin", "a.c"]);
4281 assert!(opts.builtins, "the last mention decides");
4282
4283 let (opts, _) = compile(&["-c", "-fno-builtin-memcpy", "-fno-builtin-nonesuch", "a.c"]);
4284 assert!(opts.builtins, "one name is not the family");
4285 assert_eq!(opts.no_builtin, vec!["memcpy".to_owned(), "nonesuch".to_owned()]);
4286 }
4287
4288 /// `-fvisibility=`, which is on every cmake project that cares about which names it exports
4289 /// and which was refused as an unknown option until now.
4290 ///
4291 /// Four spellings and three answers. `internal` is hidden plus a promise about never taking
4292 /// the address across a component boundary, and nothing derives anything from that promise
4293 /// here, so it comes out as the weaker of the two rather than as a refusal that stops a build
4294 /// over a distinction this compiler does not make.
4295 #[test]
4296 fn visibility_takes_the_four_spellings_gcc_takes_and_refuses_the_rest() {
4297 let (opts, _) = compile(&["-c", "a.c"]);
4298 assert_eq!(opts.visibility, Visibility::Default, "exported unless something says not");
4299
4300 for (written, wanted) in [
4301 ("default", Visibility::Default),
4302 ("hidden", Visibility::Hidden),
4303 ("internal", Visibility::Hidden),
4304 ("protected", Visibility::Protected),
4305 ] {
4306 let (opts, _) = compile(&["-c", &format!("-fvisibility={written}"), "a.c"]);
4307 assert_eq!(opts.visibility, wanted, "{written}");
4308 }
4309
4310 // The last mention decides, which is what every other flag of this shape does and what a
4311 // build that turns something off for one directory relies on.
4312 let (opts, _) = compile(&["-c", "-fvisibility=hidden", "-fvisibility=default", "a.c"]);
4313 assert_eq!(opts.visibility, Visibility::Default, "the last mention decides");
4314
4315 // A spelling gcc does not take is refused rather than read as the default, because a
4316 // build that meant hidden and got exported is a library with the wrong interface and
4317 // nothing said about it anywhere.
4318 let failed = parse_args(&args(&["-fvisibility=none", "a.c"])).expect_err("refused");
4319 assert!(failed.to_string().contains("is not a visibility"), "{failed}");
4320 }
4321
4322 /// `-ffp-contract=`, which is the one flag in the floating point group that is kept rather than
4323 /// described, and the values are gcc 16's three.
4324 #[test]
4325 fn how_far_a_multiply_and_an_addition_may_be_fused_is_asked_for() {
4326 let (opts, _) = compile(&["-c", "a.c"]);
4327 assert_eq!(opts.fp_contract, Contract::Off, "a licence nobody granted is not assumed");
4328
4329 for (written, wanted) in
4330 [("off", Contract::Off), ("on", Contract::On), ("fast", Contract::Fast)]
4331 {
4332 let (opts, _) = compile(&["-c", &format!("-ffp-contract={written}"), "a.c"]);
4333 assert_eq!(opts.fp_contract, wanted, "{written}");
4334 }
4335
4336 let (opts, _) = compile(&["-c", "-ffp-contract=fast", "-ffp-contract=off", "a.c"]);
4337 assert_eq!(opts.fp_contract, Contract::Off, "the last mention decides");
4338
4339 // Refused rather than read as one of the three, because a build that asked for no fusing
4340 // and was given the default would be one whose numbers change and whose command line says
4341 // they should not. gcc refuses the same spellings and names the same three in its message.
4342 for bad in ["-ffp-contract=none", "-ffp-contract=", "-ffp-contract=Fast"] {
4343 let failed = parse_args(&args(&[bad, "a.c"])).expect_err("refused");
4344 assert!(failed.to_string().contains("is not a contraction"), "{bad}: {failed}");
4345 }
4346
4347 // And the other one that takes a value, which is taken and kept nowhere: every operation
4348 // here is computed in the type it was written in, so `standard` is what happens and the
4349 // other two are permission to do something this does not do.
4350 let failed = parse_args(&args(&["-fexcess-precision=long", "a.c"])).expect_err("refused");
4351 assert!(failed.to_string().contains("is not an excess precision"), "{failed}");
4352 }
4353
4354 /// The four prefix mapping flags, which are what a distribution passes to get the same bytes
4355 /// out of `/build/pkg-1.2` and out of `/home/someone/pkg-1.2`. Three lists rather than one
4356 /// because gcc has three, and `-ffile-prefix-map=` is the three of them at once.
4357 #[test]
4358 fn a_prefix_mapping_flag_goes_on_the_list_its_spelling_names() {
4359 let (opts, _) = compile(&["-c", "a.c"]);
4360 assert!(opts.prefix_map.macros.is_empty(), "nothing is rewritten unless it is asked for");
4361 assert!(opts.prefix_map.debug.is_empty(), "nor here");
4362 assert!(opts.prefix_map.profile.is_empty(), "nor here");
4363
4364 let (opts, _) = compile(&["-c", "-fmacro-prefix-map=/build=.", "a.c"]);
4365 assert_eq!(opts.prefix_map.macros.apply("/build/a.c"), "./a.c", "the one it names");
4366 assert!(opts.prefix_map.debug.is_empty(), "and not the two it does not");
4367
4368 let (opts, _) = compile(&["-c", "-fdebug-prefix-map=/build=.", "a.c"]);
4369 assert_eq!(opts.prefix_map.debug.apply("/build/a.c"), "./a.c", "the one it names");
4370 assert!(opts.prefix_map.macros.is_empty(), "and not the two it does not");
4371
4372 let (opts, _) = compile(&["-c", "-fprofile-prefix-map=/build=.", "a.c"]);
4373 assert_eq!(opts.prefix_map.profile.apply("/build/a.c"), "./a.c", "the one it names");
4374 assert!(opts.prefix_map.macros.is_empty(), "and not the two it does not");
4375
4376 let (opts, _) = compile(&["-c", "-ffile-prefix-map=/build=.", "a.c"]);
4377 for list in [&opts.prefix_map.macros, &opts.prefix_map.debug, &opts.prefix_map.profile] {
4378 assert_eq!(list.apply("/build/a.c"), "./a.c", "all three at once");
4379 }
4380
4381 // Every mention is kept and the last one that matches wins, unlike the flags above whose
4382 // last mention replaces the earlier ones. A build writes one of these per source root and
4383 // expects all of them to be in force, which is the whole point of a list.
4384 let (opts, _) =
4385 compile(&["-c", "-ffile-prefix-map=/a=one", "-ffile-prefix-map=/b=two", "a.c"]);
4386 assert_eq!(opts.prefix_map.macros.apply("/a/x.c"), "one/x.c", "the earlier one still acts");
4387 assert_eq!(opts.prefix_map.macros.apply("/b/x.c"), "two/x.c", "and so does the later one");
4388
4389 // An argument with no `=` is refused rather than ignored, because a build whose paths were
4390 // meant to be rewritten and were not is one that ships the build directory's name and says
4391 // nothing about it. gcc refuses the same thing.
4392 for bad in ["-fmacro-prefix-map=nope", "-ffile-prefix-map=", "-fdebug-prefix-map=/build"] {
4393 let failed = parse_args(&args(&[bad, "a.c"])).expect_err("refused");
4394 assert!(failed.to_string().contains("is not a rewrite for"), "{bad}: {failed}");
4395 }
4396 }
4397
4398 /// `-ffunction-sections` and `-fdata-sections`, which are what make `--gc-sections` able to
4399 /// drop anything: a linker can leave out a section nothing reaches and cannot leave out half of
4400 /// one. A kernel and an embedded image are both linked that way.
4401 ///
4402 /// Two flags rather than one because gcc has two, and a build that asks for one of them and not
4403 /// the other is a build that measured something: splitting the code is nearly free at link time
4404 /// and splitting the data can defeat the linker's ordering of what is next to what.
4405 #[test]
4406 fn a_section_per_function_and_a_section_per_variable_are_asked_for_one_at_a_time() {
4407 let (opts, _) = compile(&["-c", "a.c"]);
4408 assert!(!opts.function_sections, "one text section unless something says otherwise");
4409 assert!(!opts.data_sections);
4410
4411 let (opts, _) = compile(&["-c", "-ffunction-sections", "a.c"]);
4412 assert!(opts.function_sections);
4413 assert!(!opts.data_sections, "one flag is not the other");
4414
4415 let (opts, _) = compile(&["-c", "-fdata-sections", "a.c"]);
4416 assert!(opts.data_sections);
4417 assert!(!opts.function_sections);
4418
4419 // Both directions taken, and the off one is what happens anyway rather than a refusal,
4420 // since a build that writes it is asking for the default.
4421 let (opts, _) = compile(&[
4422 "-c",
4423 "-ffunction-sections",
4424 "-fno-function-sections",
4425 "-fdata-sections",
4426 "-fno-data-sections",
4427 "a.c",
4428 ]);
4429 assert!(!opts.function_sections, "the last mention decides");
4430 assert!(!opts.data_sections, "the last mention decides");
4431 }
4432
4433 /// `-fgnu89-inline`, which is off by default and is not implied by anything on the command
4434 /// line, since the dialect asks for GNU's reading further in rather than through this.
4435 #[test]
4436 fn gnu89_inline_is_off_until_it_is_asked_for_and_the_last_mention_decides() {
4437 let (opts, _) = compile(&["-c", "a.c"]);
4438 assert!(!opts.gnu89_inline, "C's reading of inline by default");
4439
4440 let (opts, _) = compile(&["-c", "-fgnu89-inline", "a.c"]);
4441 assert!(opts.gnu89_inline);
4442
4443 let (opts, _) = compile(&["-c", "-fgnu89-inline", "-fno-gnu89-inline", "a.c"]);
4444 assert!(!opts.gnu89_inline, "the last mention decides");
4445
4446 // The C89 dialects are under GNU's reading whether this was written or not, so the flag
4447 // stays off there and the dialect is what the checker and the macro set both ask. That is
4448 // also why `-std=c89 -fno-gnu89-inline` needs no diagnostic: it asks for the reading the
4449 // dialect already has. gcc refuses that command line, which is measured in the issue.
4450 let (opts, _) = compile(&["-c", "-std=c89", "a.c"]);
4451 assert!(!opts.gnu89_inline);
4452 }
4453
4454 /// Both spellings of both frame flags, since a build that wants one usually writes the
4455 /// other beside it for the one file that has to be compiled the ordinary way.
4456 #[test]
4457 fn the_two_frame_flags_are_read_in_both_directions() {
4458 let (opts, _) = compile(&["-c", "a.c"]);
4459 assert!(!opts.frame_pointer, "gcc omits it above -O0 and so does this");
4460 assert!(opts.red_zone, "the psABI has one and nothing said not to use it");
4461
4462 let (opts, _) = compile(&["-c", "-fno-omit-frame-pointer", "-mno-red-zone", "a.c"]);
4463 assert!(opts.frame_pointer);
4464 assert!(!opts.red_zone);
4465
4466 let (opts, _) = compile(&[
4467 "-c",
4468 "-fno-omit-frame-pointer",
4469 "-fomit-frame-pointer",
4470 "-mno-red-zone",
4471 "-mred-zone",
4472 "a.c",
4473 ]);
4474 assert!(!opts.frame_pointer, "the last one wins, as it does in gcc");
4475 assert!(opts.red_zone);
4476 }
4477
4478 /// Four flags rather than one with an argument, which is how gcc spells them, and the negative
4479 /// spelled three ways because a build that turns one off writes whichever it turned on.
4480 #[test]
4481 fn the_stack_protector_is_four_flags_and_the_last_one_wins() {
4482 let (opts, _) = compile(&["-c", "a.c"]);
4483 assert_eq!(opts.protector, Protector::None, "gcc protects nothing unless it was asked");
4484
4485 for (flag, want) in [
4486 ("-fstack-protector", Protector::Buffers),
4487 ("-fstack-protector-strong", Protector::Strong),
4488 ("-fstack-protector-all", Protector::All),
4489 ] {
4490 let (opts, _) = compile(&["-c", flag, "a.c"]);
4491 assert_eq!(opts.protector, want, "{flag}");
4492 }
4493
4494 // What a package build does: the strong one in the global flags and one directory that
4495 // cannot have a protector turning it off on the line after.
4496 for off in ["-fno-stack-protector", "-fno-stack-protector-strong"] {
4497 let (opts, _) = compile(&["-c", "-fstack-protector-strong", off, "a.c"]);
4498 assert_eq!(opts.protector, Protector::None, "{off}");
4499 }
4500 let (opts, _) = compile(&["-c", "-fno-stack-protector", "-fstack-protector-all", "a.c"]);
4501 assert_eq!(opts.protector, Protector::All, "the last one wins either way round");
4502 }
4503
4504 /// A switch rather than a level, because how a frame is taken is one question and which
4505 /// functions get a canary is another, and gcc spells it that way for the same reason.
4506 #[test]
4507 fn taking_a_frame_a_page_at_a_time_is_off_until_it_is_asked_for() {
4508 let (opts, _) = compile(&["-c", "a.c"]);
4509 assert!(!opts.stack_clash, "gcc takes a frame in one subtraction unless it was asked");
4510
4511 let (opts, _) = compile(&["-c", "-fstack-clash-protection", "a.c"]);
4512 assert!(opts.stack_clash);
4513
4514 // The same shape a package build uses for the protector: on in the global flags and off
4515 // for the one directory that cannot have it.
4516 let (opts, _) =
4517 compile(&["-c", "-fstack-clash-protection", "-fno-stack-clash-protection", "a.c"]);
4518 assert!(!opts.stack_clash);
4519 let (opts, _) =
4520 compile(&["-c", "-fno-stack-clash-protection", "-fstack-clash-protection", "a.c"]);
4521 assert!(opts.stack_clash, "the last one wins either way round");
4522
4523 // The two are independent, since one is about the frame and the other about the function.
4524 let (opts, _) =
4525 compile(&["-c", "-fstack-clash-protection", "-fstack-protector-strong", "a.c"]);
4526 assert!(opts.stack_clash);
4527 assert_eq!(opts.protector, Protector::Strong);
4528 }
4529
4530 /// One flag with an argument rather than a family of spellings, because what it asks about is
4531 /// which of the two edges of a control flow transfer is checked and the two are not separate
4532 /// questions to the hardware.
4533 #[test]
4534 fn which_control_flow_edges_are_checked_is_asked_for_by_name() {
4535 let (opts, _) = compile(&["-c", "a.c"]);
4536 assert_eq!(opts.control, Control::None, "gcc's default on the targets this compiler has");
4537
4538 for (arg, want) in [
4539 ("-fcf-protection", Control::Full),
4540 ("-fcf-protection=full", Control::Full),
4541 ("-fcf-protection=branch", Control::Branch),
4542 ("-fcf-protection=return", Control::Return),
4543 ("-fcf-protection=none", Control::None),
4544 ("-fcf-protection=check", Control::Check),
4545 ] {
4546 let (opts, _) = compile(&["-c", arg, "a.c"]);
4547 assert_eq!(opts.control, want, "{arg}");
4548 }
4549
4550 // The shape a package build uses: on in the global flags and off for the one directory
4551 // that cannot have it, whichever of the two spellings of off it reaches for.
4552 let (opts, _) = compile(&["-c", "-fcf-protection=full", "-fno-cf-protection", "a.c"]);
4553 assert_eq!(opts.control, Control::None);
4554 let (opts, _) = compile(&["-c", "-fno-cf-protection", "-fcf-protection=branch", "a.c"]);
4555 assert_eq!(opts.control, Control::Branch, "the last one wins either way round");
4556 }
4557
4558 /// The profiler is asked for by two spellings, and where its hook goes by two more.
4559 ///
4560 /// The two halves are separate on purpose. `-mfentry` on its own says where a call would go and
4561 /// asks for no call, which is what gcc does with it, and a build system that sets it globally
4562 /// and asks for the profile per directory needs that to be true rather than an error.
4563 ///
4564 /// The link is asserted alongside, because the flag changes it too and a build that compiled
4565 /// with it and linked without it is a program that calls the hook everywhere and never writes a
4566 /// profile.
4567 #[test]
4568 fn the_profiler_and_where_its_hook_goes_are_two_separate_questions() {
4569 let (opts, _) = compile(&["-c", "a.c"]);
4570 assert!(!opts.profile);
4571 assert_eq!(opts.hook, Hook::Platform, "neither was named, so the target decides");
4572
4573 for arg in ["-pg", "-p"] {
4574 let (opts, _) = compile(&["-c", arg, "a.c"]);
4575 assert!(opts.profile, "{arg}");
4576 let (link, _) = linking(&[arg, "a.c"]);
4577 assert!(link.profile, "{arg} changes the link as well");
4578 }
4579
4580 for (arg, want) in [("-mfentry", Hook::Early), ("-mno-fentry", Hook::Late)] {
4581 let (opts, _) = compile(&["-c", arg, "a.c"]);
4582 assert_eq!(opts.hook, want, "{arg}");
4583 assert!(!opts.profile, "{arg} asks for no call of its own");
4584 }
4585
4586 let (opts, _) = compile(&["-c", "-mfentry", "-mno-fentry", "-pg", "a.c"]);
4587 assert_eq!(opts.hook, Hook::Late, "the last one wins");
4588 assert!(opts.profile);
4589 }
4590
4591 /// How much room a patcher is promised, which is one number or two.
4592 ///
4593 /// A command line that did not ask is asserted alongside, because the flag has to be written to
4594 /// mean anything and a build that reserved room nobody asked for would grow every function in
4595 /// it for nothing.
4596 #[test]
4597 fn the_room_a_patcher_is_promised_is_a_number_of_bytes_and_where_they_go() {
4598 let (opts, _) = compile(&["-c", "a.c"]);
4599 assert_eq!(opts.patchable, Patchable::default());
4600 assert!(!opts.patchable.any(), "nothing is reserved unless it was asked for");
4601
4602 let (opts, _) = compile(&["-c", "-fpatchable-function-entry=16", "a.c"]);
4603 assert_eq!(opts.patchable, Patchable { total: 16, before: 0 });
4604
4605 let (opts, _) = compile(&["-c", "-fpatchable-function-entry=5,3", "a.c"]);
4606 assert_eq!(opts.patchable, Patchable { total: 5, before: 3 });
4607 assert_eq!(opts.patchable.after(), 2);
4608
4609 // The last one wins, which is what every other flag of this shape does and what a build
4610 // that adds one to a command line it did not write is relying on.
4611 let (opts, _) = compile(&[
4612 "-c",
4613 "-fpatchable-function-entry=5,3",
4614 "-fpatchable-function-entry=2",
4615 "a.c",
4616 ]);
4617 assert_eq!(opts.patchable, Patchable { total: 2, before: 0 });
4618 }
4619
4620 /// And a request nothing could satisfy is refused rather than rounded into one that can be.
4621 #[test]
4622 fn room_in_front_of_the_label_that_is_more_than_the_room_asked_for_is_refused() {
4623 for arg in ["-fpatchable-function-entry=1,2", "-fpatchable-function-entry=x"] {
4624 let e = parse_args(&args(&["-c", arg, "a.c"])).unwrap_err();
4625 assert!(e.message.contains("is not an amount of room to reserve"), "{}", e.message);
4626 }
4627 }
4628
4629 /// What wraps rather than being undefined, which is two questions and three flags.
4630 ///
4631 /// The older flag is the pair of the newer two, which is gcc's own reading of it, so a build
4632 /// that writes `-fno-strict-overflow` gets both and a build that writes one of the others gets
4633 /// only what it asked for.
4634 #[test]
4635 fn what_overflows_rather_than_being_undefined_is_asked_for_two_ways() {
4636 let (opts, _) = compile(&["-c", "a.c"]);
4637 assert_eq!(opts.wrapping, Wrapping::NONE, "nothing wraps unless it was asked for");
4638
4639 let (opts, _) = compile(&["-c", "-fwrapv", "a.c"]);
4640 assert_eq!(opts.wrapping, Wrapping { signed: true, pointer: false, trap: false });
4641
4642 let (opts, _) = compile(&["-c", "-fwrapv-pointer", "a.c"]);
4643 assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: true, trap: false });
4644
4645 let (opts, _) = compile(&["-c", "-fno-strict-overflow", "a.c"]);
4646 assert_eq!(opts.wrapping, Wrapping::ALL);
4647
4648 // And the last one wins, in both directions. A build that turns one of these on globally
4649 // and off for one directory is relying on that, and so is one that writes the pair and
4650 // then takes half of it back.
4651 let (opts, _) = compile(&["-c", "-fwrapv", "-fno-wrapv", "a.c"]);
4652 assert_eq!(opts.wrapping, Wrapping::NONE);
4653
4654 let (opts, _) = compile(&["-c", "-fno-strict-overflow", "-fstrict-overflow", "a.c"]);
4655 assert_eq!(opts.wrapping, Wrapping::NONE);
4656
4657 let (opts, _) = compile(&["-c", "-fno-strict-overflow", "-fno-wrapv-pointer", "a.c"]);
4658 assert_eq!(opts.wrapping, Wrapping { signed: true, pointer: false, trap: false });
4659 }
4660
4661 /// And the other answer to the signed question cannot be held at the same time as the first.
4662 ///
4663 /// A program cannot both wrap and stop, so writing both is writing a contradiction, and gcc
4664 /// resolves it by letting the last one win rather than by reporting anything. That was measured
4665 /// against gcc 16 rather than read out of the manual, which says nothing about it: `-ftrapv
4666 /// -fwrapv` emits no checked calls and `-fwrapv -ftrapv` emits them.
4667 #[test]
4668 fn a_signed_overflow_that_stops_is_the_other_answer_and_not_a_third_one() {
4669 let (opts, _) = compile(&["-c", "-ftrapv", "a.c"]);
4670 assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: false, trap: true });
4671
4672 let (opts, _) = compile(&["-c", "-fwrapv", "-ftrapv", "a.c"]);
4673 assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: false, trap: true });
4674
4675 let (opts, _) = compile(&["-c", "-ftrapv", "-fwrapv", "a.c"]);
4676 assert_eq!(opts.wrapping, Wrapping { signed: true, pointer: false, trap: false });
4677
4678 let (opts, _) = compile(&["-c", "-ftrapv", "-fno-strict-overflow", "a.c"]);
4679 assert_eq!(opts.wrapping, Wrapping::ALL);
4680
4681 let (opts, _) = compile(&["-c", "-ftrapv", "-fno-trapv", "a.c"]);
4682 assert_eq!(opts.wrapping, Wrapping::NONE);
4683
4684 // And the flag that says what may be assumed says nothing about what happens, so it leaves
4685 // this alone where it takes the wrapping away. gcc does the same.
4686 let (opts, _) = compile(&["-c", "-ftrapv", "-fstrict-overflow", "a.c"]);
4687 assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: false, trap: true });
4688 }
4689
4690 /// What a plain `char` is, which is four spellings of two answers and nothing by default.
4691 ///
4692 /// Nothing is the target's own answer and has to stay distinct from both of the others, since
4693 /// the same command line means a signed `char` on x86-64 and an unsigned one on Linux's arm64.
4694 /// The negative spellings are the other flag rather than a way of asking for the default, which
4695 /// was measured against gcc 16: `-fno-signed-char` defines `__CHAR_UNSIGNED__` and
4696 /// `-fno-unsigned-char` does not.
4697 #[test]
4698 fn the_signedness_of_a_plain_char_is_asked_for_in_four_ways() {
4699 let (opts, _) = compile(&["-c", "a.c"]);
4700 assert_eq!(opts.char_signed, None);
4701
4702 for flag in ["-fsigned-char", "-fno-unsigned-char"] {
4703 let (opts, _) = compile(&["-c", flag, "a.c"]);
4704 assert_eq!(opts.char_signed, Some(true), "{flag}");
4705 }
4706
4707 for flag in ["-funsigned-char", "-fno-signed-char"] {
4708 let (opts, _) = compile(&["-c", flag, "a.c"]);
4709 assert_eq!(opts.char_signed, Some(false), "{flag}");
4710 }
4711
4712 // And the last one wins, which is what a build that sets one globally and the other for a
4713 // directory relies on.
4714 let (opts, _) = compile(&["-c", "-funsigned-char", "-fsigned-char", "a.c"]);
4715 assert_eq!(opts.char_signed, Some(true));
4716
4717 // And what is asked for reaches the target, because that is what every other part of the
4718 // compiler asks. The triple is one whose own answer is the opposite, so a session that
4719 // ignored the flag would still read as signed here.
4720 let (opts, _) =
4721 compile(&["-c", "--target=aarch64-unknown-linux-gnu", "-fsigned-char", "a.c"]);
4722 assert!(Session::new(*opts).target.char_is_signed);
4723 let (opts, _) = compile(&["-c", "--target=aarch64-unknown-linux-gnu", "a.c"]);
4724 assert!(!Session::new(*opts).target.char_is_signed);
4725 }
4726
4727 /// And the size of an enumeration, which is one question with two spellings.
4728 #[test]
4729 fn the_smallest_enumeration_is_asked_for_and_taken_back() {
4730 let (opts, _) = compile(&["-c", "a.c"]);
4731 assert!(!opts.short_enums);
4732
4733 let (opts, _) = compile(&["-c", "-fshort-enums", "a.c"]);
4734 assert!(opts.short_enums);
4735
4736 let (opts, _) = compile(&["-c", "-fshort-enums", "-fno-short-enums", "a.c"]);
4737 assert!(!opts.short_enums);
4738
4739 let (opts, _) = compile(&["-c", "-fno-short-enums", "-fshort-enums", "a.c"]);
4740 assert!(opts.short_enums);
4741 }
4742
4743 /// And Microsoft's reading of an anonymous member, which the target answers where the command
4744 /// line said nothing. gcc's mingw build has it on and its Linux build has it off, so a header
4745 /// that closes a nameless union with a macro that expands to nothing is read the way the
4746 /// compiler that platform ships would read it.
4747 #[test]
4748 fn the_microsoft_reading_of_a_member_follows_the_target_until_it_is_asked_for() {
4749 // Named rather than left to the host, since the answer this asks for is the one a target
4750 // that is not Windows gives and on a Windows machine the host is not one of those.
4751 let (opts, _) = compile(&[LINUX, "-c", "a.c"]);
4752 assert!(!Session::new(*opts).ms_extensions());
4753
4754 let (opts, _) = compile(&["-c", "--target=x86_64-pc-windows-gnu", "a.c"]);
4755 assert!(Session::new(*opts).ms_extensions());
4756
4757 let (opts, _) = compile(&["-c", "-fms-extensions", "a.c"]);
4758 assert!(Session::new(*opts).ms_extensions());
4759
4760 let (opts, _) =
4761 compile(&["-c", "--target=x86_64-pc-windows-gnu", "-fno-ms-extensions", "a.c"]);
4762 assert!(!Session::new(*opts).ms_extensions());
4763 }
4764
4765 /// And a value nothing means is refused rather than taken for the nearest thing it looks like.
4766 ///
4767 /// `-fcf-protection=all` is the spelling somebody writes from memory, and a compiler that read
4768 /// it as `full` would be guessing, while one that let it fall through to the optimizer's `-f`
4769 /// family would report it as an unknown pass. Neither is the news the build wants.
4770 #[test]
4771 fn a_control_flow_protection_nothing_means_is_refused() {
4772 let e = parse_args(&args(&["-c", "-fcf-protection=all", "a.c"])).unwrap_err();
4773 assert!(e.message.contains("is not a control flow protection"), "{}", e.message);
4774 assert!(e.message.contains("full, branch, return, none or check"), "{}", e.message);
4775 }
4776
4777 #[test]
4778 fn the_link_flags_are_collected_apart_from_the_compilation() {
4779 let (link, _) = linking(&[
4780 "-static",
4781 "-nostartfiles",
4782 "-rdynamic",
4783 "-s",
4784 "-fuse-ld=mold",
4785 "-L/opt/lib",
4786 "-B",
4787 "/opt/tools",
4788 "a.c",
4789 ]);
4790 assert!(link.is_static);
4791 assert!(link.no_startfiles);
4792 assert!(link.export_dynamic);
4793 assert!(link.strip);
4794 assert_eq!(link.use_ld.as_deref(), Some("mold"));
4795 assert_eq!(link.search, vec![PathBuf::from("/opt/lib")]);
4796 assert_eq!(link.prefixes, vec![PathBuf::from("/opt/tools")]);
4797 }
4798
4799 #[test]
4800 fn a_comma_in_dash_wl_separates_two_arguments() {
4801 // The target is written down because the name of the object is derived from it, and `a.o`
4802 // on a Linux host is `a.obj` on a Windows one. What is under test is the splitting of the
4803 // argument, which has nothing to do with either.
4804 let (_, plan) = linking(&[LINUX, "-Wl,-rpath,/opt/lib", "-Xlinker", "--as-needed", "a.c"]);
4805 let link = plan.link.expect("expected a link step");
4806 assert_eq!(
4807 link.inputs,
4808 vec![
4809 link::Item::Linker("-rpath".into()),
4810 link::Item::Linker("/opt/lib".into()),
4811 link::Item::Linker("--as-needed".into()),
4812 link::Item::File("a.o".into()),
4813 ]
4814 );
4815 }
4816
4817 #[test]
4818 fn a_word_for_the_linker_keeps_its_place_among_the_files_too() {
4819 // What libtool writes around a set of convenience archives, and what #1279 was. Both words
4820 // are about the files between them, so the pair collected out of the line and appended to
4821 // the end is two options that bracket nothing and an archive that went in empty.
4822 let (_, plan) = linking(&[
4823 "--target=x86_64-unknown-linux-gnu",
4824 "a.c",
4825 "-Wl,--whole-archive",
4826 "libaesni.a",
4827 "-Wl,--no-whole-archive",
4828 "-lm",
4829 ]);
4830 let link = plan.link.expect("expected a link step");
4831 assert_eq!(
4832 link.inputs,
4833 vec![
4834 link::Item::File("a.o".into()),
4835 link::Item::Linker("--whole-archive".into()),
4836 link::Item::File("libaesni.a".into()),
4837 link::Item::Linker("--no-whole-archive".into()),
4838 link::Item::Library("m".into()),
4839 ]
4840 );
4841 // And it is not a job, because there is nothing to compile in a word for the linker.
4842 assert_eq!(plan.jobs.len(), 2);
4843 }
4844
4845 #[test]
4846 fn a_word_for_the_linker_on_a_dash_c_line_is_dropped_without_a_word() {
4847 // GCC says nothing about one either. `-Wl,` on a compile line is what a build system
4848 // writes when one variable holds the flags for both, and a note here would be a note on
4849 // every compile of every autotools project.
4850 let (_, plan) = linking(&["-c", "-Wl,--as-needed", "a.c"]);
4851 assert!(plan.link.is_none());
4852 assert!(plan.notes.is_empty(), "{:?}", plan.notes);
4853 assert_eq!(plan.jobs.len(), 1);
4854 }
4855
4856 #[test]
4857 fn a_library_keeps_its_place_between_the_objects() {
4858 // Link order is semantic: `-lm` written between two files resolves for the one before
4859 // it and not for the one after, so a library cannot be collected into a list of its own.
4860 // The target is named because the suffix of an object is the target's and this asserts
4861 // on the names: the same command line on a Windows host plans two `.obj` files.
4862 let (_, plan) = linking(&["--target=x86_64-unknown-linux-gnu", "a.c", "-lm", "b.c"]);
4863 let link = plan.link.expect("expected a link step");
4864 assert_eq!(
4865 link.inputs,
4866 vec![
4867 link::Item::File("a.o".into()),
4868 link::Item::Library("m".into()),
4869 link::Item::File("b.o".into()),
4870 ]
4871 );
4872 // And it is not a job, because there is nothing to compile in a library.
4873 assert_eq!(plan.jobs.len(), 2);
4874 }
4875
4876 #[test]
4877 fn a_library_on_a_dash_c_line_is_a_note_rather_than_an_error() {
4878 let (_, plan) = linking(&["-c", "-lm", "a.c"]);
4879 assert!(plan.link.is_none());
4880 assert!(plan.notes.iter().any(|n| n.contains("-lm")), "{:?}", plan.notes);
4881 }
4882
4883 #[test]
4884 fn the_sysroot_reaches_the_linker_as_well_as_the_headers() {
4885 let (link, _) = linking(&["--sysroot=/opt/root", "a.c"]);
4886 assert_eq!(link.sysroot, Some(PathBuf::from("/opt/root")));
4887 }
4888
4889 fn printed(s: &[&str]) -> String {
4890 match parse_args(&args(s)).expect("expected an answer") {
4891 Action::Print(line) => line,
4892 other => panic!("expected an answer, got {other:?}"),
4893 }
4894 }
4895
4896 fn refused(s: &[&str]) -> String {
4897 parse_args(&args(s)).expect_err("expected a refusal").message
4898 }
4899
4900 #[test]
4901 fn a_warning_flag_this_compiler_has_not_heard_of_is_taken_rather_than_refused() {
4902 // The rule in section 4.1, and the reason for it is autoconf: a configure script finds
4903 // out whether a warning flag exists by passing it and looking at the exit status, so a
4904 // compiler that refuses one it does not know fails a script written for a newer GCC.
4905 let (opts, _) = compile(&["-Wall", "-Wextra", "-Wno-format-truncation", "-c", "a.c"]);
4906 assert!(!opts.warnings_are_errors);
4907 assert!(opts.warnings);
4908 // The two spellings that do mean something are still read.
4909 let (opts, _) = compile(&["-Werror", "-c", "a.c"]);
4910 assert!(opts.warnings_are_errors);
4911 let (opts, _) = compile(&["-w", "-c", "a.c"]);
4912 assert!(!opts.warnings);
4913 let (opts, _) = compile(&["-pedantic-errors", "-c", "a.c"]);
4914 assert!(opts.pedantic && opts.warnings_are_errors);
4915 }
4916
4917 #[test]
4918 fn an_argument_for_a_separate_tool_is_refused_rather_than_dropped() {
4919 // Every one of these says something about the output, so the wrong answer is silence.
4920 assert!(refused(&["-Wa,--noexecstack", "-c", "a.c"]).contains("separate assembler"));
4921 assert!(refused(&["-Wp,-DX", "-c", "a.c"]).contains("separate assembler"));
4922 assert!(refused(&["-specs=/x", "a.c"]).contains("-specs= is not supported"));
4923 assert!(refused(&["-mcmodel=kernel", "-c", "a.c"]).contains("small code model"));
4924 assert!(refused(&["-gdwarf-4", "-c", "a.c"]).contains("DWARF 5"));
4925 assert!(refused(&["-Ofast", "-c", "a.c"]).contains("fast math"));
4926 // The word size the target does not have, which is a target this compiler was not asked
4927 // for rather than a flag it does not know.
4928 let no32 = refused(&["--target=x86_64-unknown-linux-gnu", "-m32", "-c", "a.c"]);
4929 assert!(no32.contains("32 bit target"), "{no32}");
4930 }
4931
4932 /// `-gz` and the two spellings of the split, which are the two questions about the shape of
4933 /// the debug output rather than about how much of it there is.
4934 ///
4935 /// Both answers here are about what happens when there is debug information to shape, and
4936 /// there is none yet, so what is being asserted is that the flags are read and remembered
4937 /// rather than that anything changed in the output. That is the whole of what taking them
4938 /// claims, and it is worth a test because the day `rucc-debug` writes a section this is where
4939 /// it comes to find out what the command line said.
4940 #[test]
4941 fn the_shape_of_the_debug_output_is_recorded_even_where_there_is_none_of_it() {
4942 let (opts, _) = compile(&["-c", "a.c"]);
4943 assert_eq!(opts.compress, Compress::None, "uncompressed unless somebody asks");
4944
4945 // Bare `-gz` is `-gz=zlib`, measured against gcc 16 rather than read out of the manual,
4946 // which describes the flag without ever saying which algorithm it picks.
4947 assert_eq!(compile(&["-gz", "-c", "a.c"]).0.compress, Compress::Zlib);
4948 for (spelling, want) in [
4949 ("none", Compress::None),
4950 ("zlib", Compress::Zlib),
4951 ("zlib-gnu", Compress::ZlibGnu),
4952 ("zstd", Compress::Zstd),
4953 ] {
4954 let (opts, _) = compile(&[&format!("-gz={spelling}"), "-c", "a.c"]);
4955 assert_eq!(opts.compress, want, "{spelling}");
4956 }
4957
4958 // A value nothing here has heard of is refused rather than rounded to the nearest one,
4959 // because a build that asked for `zstd` and quietly got `zlib` would ship a file its
4960 // reader may not understand and would have no way of finding out.
4961 for bad in ["-gz=gzip", "-gz="] {
4962 let failed = refused(&[bad, "-c", "a.c"]);
4963 assert!(failed.contains("is not a way to compress"), "{bad}: {failed}");
4964 }
4965
4966 // The split is refused in the direction that would have written a file and taken in the
4967 // direction that describes what happens. A build system that names the `.dwo` as an
4968 // output has to hear about it now rather than at the point the file is missing.
4969 let (opts, _) = compile(&["-gno-split-dwarf", "-g", "-c", "a.c"]);
4970 assert!(opts.debug_info, "the negative spelling says nothing about how much");
4971 let failed = refused(&["-gsplit-dwarf", "-c", "a.c"]);
4972 assert!(failed.contains(".dwo"), "the refusal names the file it would have written");
4973 }
4974
4975 /// The `-flto` family, which is the whole of an optimization this compiler does not do.
4976 ///
4977 /// Taken rather than refused because ignoring it gives a correct program that is slower than
4978 /// it could have been, which is section 4.1's hint about speed. The values are still held to
4979 /// gcc's, so a command line written for clang is told rather than quietly taken.
4980 #[test]
4981 fn the_link_time_family_is_read_and_checked_and_nothing_is_done_about_it() {
4982 let (opts, _) = compile(&["-c", "a.c"]);
4983 assert!(!opts.lto.requested, "nothing asks unless the command line does");
4984
4985 let (opts, _) = compile(&["-flto", "-c", "a.c"]);
4986 assert!(opts.lto.requested);
4987 assert_eq!(opts.lto.jobs, LtoJobs::One, "bare -flto is one process, the way gcc reads it");
4988
4989 // The last of the two directions wins, the same as every other pair of `-f` spellings.
4990 assert!(!compile(&["-flto", "-fno-lto", "-c", "a.c"]).0.lto.requested);
4991 assert!(compile(&["-fno-lto", "-flto", "-c", "a.c"]).0.lto.requested);
4992
4993 // A count is a count, and asking for one implies asking for the optimization.
4994 for (spelling, want) in [
4995 ("auto", LtoJobs::Auto),
4996 ("jobserver", LtoJobs::Jobserver),
4997 ("1", LtoJobs::One),
4998 ("8", LtoJobs::Count(8)),
4999 ] {
5000 let (opts, _) = compile(&[&format!("-flto={spelling}"), "-c", "a.c"]);
5001 assert_eq!(opts.lto.jobs, want, "{spelling}");
5002 assert!(opts.lto.requested, "{spelling} asks for it too");
5003 }
5004
5005 // gcc refuses a zero rather than reading it as `-fno-lto`, and `thin` is clang's spelling
5006 // of a question gcc answers with `-flto-partition=`, so somebody who wrote it meant a
5007 // different compiler and gets told so here rather than getting a serial link.
5008 for bad in ["-flto=0", "-flto=thin", "-flto=full", "-flto=-1"] {
5009 let failed = refused(&[bad, "-c", "a.c"]);
5010 assert!(failed.contains("link time jobs"), "{bad}: {failed}");
5011 }
5012
5013 // How the program is cut up before the work is spread over it.
5014 assert_eq!(compile(&["-c", "a.c"]).0.lto.partition, Partition::Balanced, "gcc's default");
5015 for (spelling, want) in [
5016 ("balanced", Partition::Balanced),
5017 ("1to1", Partition::OneToOne),
5018 ("one", Partition::One),
5019 ("max", Partition::Max),
5020 ("none", Partition::None),
5021 ] {
5022 let (opts, _) = compile(&[&format!("-flto-partition={spelling}"), "-c", "a.c"]);
5023 assert_eq!(opts.lto.partition, want, "{spelling}");
5024 }
5025 assert!(refused(&["-flto-partition=big", "-c", "a.c"]).contains("partitioning model"));
5026
5027 // And how hard the bytecode is compressed on its way into the object, which is zstd's
5028 // range of levels and is the range gcc checks an argument against.
5029 assert_eq!(compile(&["-c", "a.c"]).0.lto.compression, None, "whatever it does by default");
5030 assert_eq!(compile(&["-flto-compression-level=0", "-c", "a.c"]).0.lto.compression, Some(0));
5031 let (opts, _) = compile(&["-flto-compression-level=19", "-c", "a.c"]);
5032 assert_eq!(opts.lto.compression, Some(19));
5033 for bad in ["-flto-compression-level=20", "-flto-compression-level=-1"] {
5034 let failed = refused(&[bad, "-c", "a.c"]);
5035 assert!(failed.contains("compression level"), "{bad}: {failed}");
5036 }
5037
5038 // The two pairs that describe an arrangement rather than ask for one. Every object here
5039 // holds its machine code, so the fat spelling is what already happens and the other is a
5040 // smaller file rather than a different program, and the plugin pair is about a tool the
5041 // design in `spec/09-optimizer.md` never loads.
5042 for taken in [
5043 "-ffat-lto-objects",
5044 "-fno-fat-lto-objects",
5045 "-fuse-linker-plugin",
5046 "-fno-use-linker-plugin",
5047 ] {
5048 let (opts, _) = compile(&[taken, "-c", "a.c"]);
5049 assert!(!opts.lto.requested, "{taken} says nothing about whether to do it");
5050 }
5051 }
5052
5053 /// The profile family, which is the only one here that splits down the middle.
5054 ///
5055 /// Reading a profile is taken and writing one is refused, and the line between them is the one
5056 /// section 4.1 draws: ignoring a request to read the counts gives a correct program that is
5057 /// slower than it could have been, and ignoring a request to write them means a file the build
5058 /// declared as an output never appears.
5059 #[test]
5060 fn reading_a_profile_is_taken_and_writing_one_is_refused() {
5061 let (opts, _) = compile(&["-c", "a.c"]);
5062 assert!(!opts.profile_data.requested, "nothing asks unless the command line does");
5063 assert_eq!(opts.profile_data.path, None);
5064
5065 let (opts, _) = compile(&["-fprofile-use", "-c", "a.c"]);
5066 assert!(opts.profile_data.requested);
5067 assert_eq!(opts.profile_data.path, None, "beside the object, the way gcc looks");
5068
5069 let (opts, _) = compile(&["-fprofile-use=/counts", "-c", "a.c"]);
5070 assert!(opts.profile_data.requested, "naming a path asks for it too");
5071 assert_eq!(opts.profile_data.path.as_deref(), Some("/counts"));
5072
5073 // The last of the two directions wins, the same as every other pair of `-f` spellings.
5074 assert!(
5075 !compile(&["-fprofile-use", "-fno-profile-use", "-c", "a.c"]).0.profile_data.requested
5076 );
5077 assert!(
5078 compile(&["-fno-profile-use", "-fprofile-use", "-c", "a.c"]).0.profile_data.requested
5079 );
5080
5081 // The rest of the reading half, which is where the files are and three answers about what
5082 // to make of what is in them.
5083 let (opts, _) = compile(&[
5084 "-fprofile-dir=/build/profiles",
5085 "-fprofile-abs-path",
5086 "-fprofile-correction",
5087 "-fprofile-partial-training",
5088 "-c",
5089 "a.c",
5090 ]);
5091 assert_eq!(opts.profile_data.dir.as_deref(), Some("/build/profiles"));
5092 assert!(opts.profile_data.absolute);
5093 assert!(opts.profile_data.correction);
5094 assert!(opts.profile_data.partial_training);
5095
5096 // Writing one, which is refused by name. The first four instrument the program and the
5097 // last writes a file beside the object, and a build that got neither and no message would
5098 // go on to optimize against counts that were never gathered.
5099 for writing in [
5100 "-fprofile-generate",
5101 "-fprofile-generate=/build/profiles",
5102 "-fprofile-arcs",
5103 "--coverage",
5104 "-fcondition-coverage",
5105 "-fpath-coverage",
5106 ] {
5107 let failed = refused(&[writing, "-c", "a.c"]);
5108 assert!(failed.contains("instrument"), "{writing}: {failed}");
5109 }
5110 assert!(refused(&["-ftest-coverage", "-c", "a.c"]).contains(".gcno"), "it names the file");
5111
5112 // The negative spellings of the refused half are what already happens, so they are taken.
5113 for taken in ["-fno-profile-generate", "-fno-profile-arcs", "-fno-test-coverage"] {
5114 let (opts, _) = compile(&[taken, "-c", "a.c"]);
5115 assert!(!opts.profile_data.requested, "{taken} asks for nothing");
5116 }
5117
5118 // And the flags that describe the instrumentation that is refused above, which are checked
5119 // and dropped. Checked because a typo is worth finding here rather than on the day the
5120 // instrumentation lands.
5121 for taken in [
5122 "-fprofile-update=single",
5123 "-fprofile-update=atomic",
5124 "-fprofile-update=prefer-atomic",
5125 "-fprofile-reproducible=serial",
5126 "-fprofile-reproducible=parallel-runs",
5127 "-fprofile-reproducible=multithreaded",
5128 "-fprofile-values",
5129 "-fno-profile-values",
5130 "-fprofile-info-section",
5131 "-fprofile-filter-files=a.c",
5132 "-fprofile-exclude-files=b.c",
5133 "-fprofile-note=a.gcno",
5134 ] {
5135 let (opts, _) = compile(&[taken, "-c", "a.c"]);
5136 assert!(!opts.profile_data.requested, "{taken} says nothing about reading one");
5137 }
5138 assert!(refused(&["-fprofile-update=none", "-c", "a.c"]).contains("update method"));
5139 assert!(refused(&["-fprofile-reproducible=any", "-c", "a.c"]).contains("reproducibility"));
5140 }
5141
5142 /// The sanitizers, which are refused by name and are the one family refused for a reason that
5143 /// is not about the bytes.
5144 ///
5145 /// A sanitizer is a promise that the program is watched while it runs, so a build that asked
5146 /// for one and was quietly given a program with no checks in it gets a test suite that passes
5147 /// for the wrong reason rather than a slower program.
5148 #[test]
5149 fn a_sanitizer_that_is_still_asked_for_at_the_end_of_the_line_is_refused_by_name() {
5150 for asked in ["address", "undefined", "thread", "kernel-address", "leak", "memory"] {
5151 let failed = refused(&[&format!("-fsanitize={asked}"), "-c", "a.c"]);
5152 assert!(failed.contains(asked), "the refusal names what was asked for: {failed}");
5153 assert!(failed.contains("-fsafety=detect"), "and the nearest thing: {failed}");
5154 }
5155
5156 // A list is every name in it, and the first one still standing is the one named.
5157 let failed = refused(&["-fsanitize=address,undefined", "-c", "a.c"]);
5158 assert!(failed.contains("address"), "{failed}");
5159
5160 // A name that is not one, which is worth its own message: somebody who wrote `-fsanitize`
5161 // with a typo in it has a different problem from somebody who wrote a real one.
5162 for bad in ["-fsanitize=bogus", "-fsanitize=address,bogus", "-fno-sanitize=bogus"] {
5163 let failed = refused(&[bad, "-c", "a.c"]);
5164 assert!(failed.contains("is not a sanitizer"), "{bad}: {failed}");
5165 }
5166
5167 // gcc takes `all` only in the negative, and so does this.
5168 assert!(refused(&["-fsanitize=all", "-c", "a.c"]).contains("only `-fno-sanitize=all`"));
5169
5170 // Asking and then taking it back is asking for nothing, which is why the answer waits for
5171 // the end of the line. A build whose shared flags turn a check on and whose rule for one
5172 // file turns it off again compiles that file here.
5173 for pair in [
5174 ["-fsanitize=address", "-fno-sanitize=address"],
5175 ["-fsanitize=address,undefined", "-fno-sanitize=all"],
5176 ["-fsanitize=undefined", "-fno-sanitize=undefined"],
5177 ] {
5178 let (opts, _) = compile(&[pair[0], pair[1], "-c", "a.c"]);
5179 assert_eq!(opts.safety, rucc_session::Safety::Off, "{pair:?} asked for nothing");
5180 }
5181 // And the other order still asks, because the last word is the one that counts.
5182 assert!(!refused(&["-fno-sanitize=address", "-fsanitize=address", "-c", "a.c"]).is_empty());
5183
5184 // What a check does when it fires is an answer about checks that are refused, so there is
5185 // nothing left for it to change and it is taken.
5186 for taken in [
5187 "-fsanitize-recover=undefined",
5188 "-fno-sanitize-recover=all",
5189 "-fsanitize-trap=undefined",
5190 "-fno-sanitize-trap=all",
5191 "-fsanitize-undefined-trap-on-error",
5192 "-fsanitize-address-use-after-scope",
5193 "-fno-sanitize-address-use-after-scope",
5194 "-fsanitize-sections=.data",
5195 ] {
5196 let (opts, _) = compile(&[taken, "-c", "a.c"]);
5197 assert_eq!(opts.safety, rucc_session::Safety::Off, "{taken} asks for no checking");
5198 }
5199 assert!(refused(&["-fsanitize-recover=bogus", "-c", "a.c"]).contains("is not a sanitizer"));
5200
5201 // Coverage instrumentation is refused rather than dropped, because a fuzzer with no
5202 // feedback runs blind and never says so.
5203 let failed = refused(&["-fsanitize-coverage=trace-pc", "-c", "a.c"]);
5204 assert!(failed.contains("feedback"), "{failed}");
5205 let failed = refused(&["-fsanitize-coverage=trace-pc-guard", "-c", "a.c"]);
5206 assert!(failed.contains("trace-pc or trace-cmp"), "gcc takes two of them: {failed}");
5207 }
5208
5209 #[test]
5210 fn the_levels_gcc_spells_differently_are_the_levels_they_mean() {
5211 assert_eq!(compile(&["-O", "-c", "a.c"]).0.opt_level, OptLevel::O1);
5212 assert_eq!(compile(&["-Og", "-c", "a.c"]).0.opt_level, OptLevel::O1);
5213 assert_eq!(compile(&["-O2", "-c", "a.c"]).0.opt_level, OptLevel::O2);
5214 }
5215
5216 #[test]
5217 fn the_machine_flags_that_name_what_we_already_do_are_taken_and_the_rest_are_not() {
5218 let line = ["--target=x86_64-unknown-linux-gnu", "-m64", "-march=x86-64-v3"];
5219 let (opts, _) =
5220 compile(&[&line[..], &["-mtune=native", "-mabi=sysv", "-c", "a.c"]].concat());
5221 assert_eq!(opts.target.to_string(), "x86_64-unknown-linux-gnu");
5222 let wrong = refused(&["--target=x86_64-unknown-linux-gnu", "-mabi=ms", "-c", "a.c"]);
5223 assert!(wrong.contains("sysv convention"), "{wrong}");
5224 }
5225
5226 #[test]
5227 fn the_thread_flag_is_a_macro_and_a_library_and_the_library_goes_last() {
5228 let (opts, plan) = compile(&["-pthread", "-c", "a.c"]);
5229 assert!(opts.defines.iter().any(|d| d == "_REENTRANT"));
5230 // After the input, because a static link takes what it needs from a library when it
5231 // reaches it and not afterwards.
5232 let names: Vec<&str> = plan.jobs.iter().map(|j| j.input.as_str()).collect();
5233 assert_eq!(names, vec!["a.c"]);
5234 }
5235
5236 #[test]
5237 fn the_questions_a_build_system_asks_before_it_compiles_anything() {
5238 let target = "--target=x86_64-unknown-linux-gnu";
5239 assert_eq!(printed(&[target, "-dumpmachine"]), "x86_64-unknown-linux-gnu");
5240 assert_eq!(printed(&[target, "-dumpversion"]), VERSION);
5241 assert_eq!(printed(&[target, "-dumpfullversion"]), VERSION);
5242 assert_eq!(printed(&[target, "-print-multiarch"]), "x86_64-linux-gnu");
5243 // A name nothing holds comes back unchanged, which is GCC's rule and is what makes the
5244 // answer safe to paste into a link line whether or not the file is there.
5245 assert_eq!(printed(&[target, "-print-file-name=no-such-library.a"]), "no-such-library.a");
5246 assert_eq!(printed(&[target, "-print-prog-name=ld"]), "ld");
5247 let dirs = printed(&[target, "-print-search-dirs"]);
5248 assert!(dirs.starts_with("install: "), "{dirs}");
5249 assert!(dirs.contains("\nlibraries: ="), "{dirs}");
5250 }
5251
5252 #[test]
5253 fn the_sysroot_in_effect_is_the_one_the_command_line_named_or_the_one_for_the_target() {
5254 // A tree the user named is the answer whatever the target is, because it is the answer to
5255 // every other question too.
5256 assert_eq!(printed(&["--sysroot=/opt/cross", "-print-sysroot"]), "/opt/cross");
5257
5258 // A target that is no machine this suite runs on is read under the cache, and the answer is
5259 // the root rather than one of the directories under it, since what asks is looking for a
5260 // file of its own.
5261 let root = cache::dir().join("sysroots").join("riscv64-linux-musl");
5262 assert_eq!(
5263 printed(&["--target=riscv64-linux-musl", "-print-sysroot"]),
5264 root.display().to_string()
5265 );
5266
5267 // And a compile for this machine has no sysroot, which is the empty line GCC prints when it
5268 // was configured without one rather than a `/` that would be a claim about the filesystem.
5269 let host = Triple::host().expect("a host this compiler knows");
5270 assert_eq!(printed(&[&format!("--target={host}"), "-print-sysroot"]), "");
5271 }
5272
5273 #[test]
5274 fn the_provenance_of_a_sysroot_is_the_manifest_it_carries() {
5275 // Section 13.5 wants seven things per input and wants them machine readable, and the manifest
5276 // is the record that already has them, so the flag prints that rather than a second format.
5277 let manifest = "rucc sysroot manifest 3\n\
5278 target\tx86_64-linux-musl\n\
5279 kernel\t6.12\n\
5280 include/generic/stdio.h\tmusl-1.2.5\t\
5281 https://musl.libc.org/releases/musl-1.2.5.tar.gz\t\
5282 0000000000000000000000000000000000000000000000000000000000000000\tmit\t\
5283 bundled\n\
5284 lib/libc.so\tmusl-1.2.5\t\
5285 https://musl.libc.org/releases/musl-1.2.5.tar.gz\t\
5286 1111111111111111111111111111111111111111111111111111111111111111\tmit\t\
5287 generated\n";
5288 let tree = TempTree::new("provenance", &[("manifest", manifest)]);
5289 let sysroot = format!("--sysroot={}", tree.0.display());
5290 // The kernel line of tamnd/rucc#934 is in the answer without anything here naming it, because
5291 // the flag parses the record and renders it again rather than picking fields out of it. That
5292 // is the reason it prints a manifest and not a format of its own.
5293 //
5294 // The answer is the file without its last newline, because whatever prints it adds one. The
5295 // file is what somebody diffs the output against, so the two have to be the same bytes.
5296 assert_eq!(printed(&[&sysroot, "-print-sysroot-provenance"]) + "\n", manifest);
5297
5298 // A tree with no manifest in it is a tree somebody assembled themselves, and nothing here
5299 // knows where any of it came from. Saying nothing is the only honest answer, and a reader can
5300 // tell it from a manifest with no inputs because that one still has its two header lines.
5301 let bare = TempTree::new("provenance-bare", &[]);
5302 assert_eq!(
5303 printed(&[&format!("--sysroot={}", bare.0.display()), "-print-sysroot-provenance"]),
5304 ""
5305 );
5306
5307 // And a compile for this machine has no sysroot at all, which is the same empty answer
5308 // `-print-sysroot` gives for it.
5309 let host = Triple::host().expect("a host this compiler knows");
5310 assert_eq!(printed(&[&format!("--target={host}"), "-print-sysroot-provenance"]), "");
5311
5312 // And the other spelling, which section 13.5 is the document that writes.
5313 assert_eq!(printed(&[&sysroot, "--print-sysroot-provenance"]) + "\n", manifest);
5314
5315 // tamnd/rucc#1021. The digest of the same tree is the sha256 of that record, so it is one
5316 // line where the provenance is a few hundred, and it is checkable with `sha256sum` because
5317 // the bytes it is over are the bytes of the file. The number here is that hash of the
5318 // fixture above, computed by `sha256sum` rather than by this compiler.
5319 assert_eq!(
5320 printed(&[&sysroot, "-print-sysroot-digest"]),
5321 "d705ae6ebeafeb7fda4bd57cecc7882bf49784b17015664a09cfae25a1b2000a"
5322 );
5323 assert_eq!(
5324 printed(&[&sysroot, "--print-sysroot-digest"]),
5325 printed(&[&sysroot, "-print-sysroot-digest"])
5326 );
5327
5328 // And the two empty answers are empty here too, because a digest of nothing would read as a
5329 // claim about a sysroot rather than as the absence of one.
5330 assert_eq!(
5331 printed(&[&format!("--sysroot={}", bare.0.display()), "-print-sysroot-digest"]),
5332 ""
5333 );
5334 assert_eq!(printed(&[&format!("--target={host}"), "-print-sysroot-digest"]), "");
5335 }
5336
5337 #[test]
5338 fn a_manifest_this_build_cannot_read_is_refused_rather_than_printed() {
5339 // Passing a file we could not parse to whoever asked would make their parser the one that
5340 // finds the problem, and the three uses section 13.5 gives for this are all somebody else
5341 // parsing it.
5342 let tree = TempTree::new(
5343 "provenance-bad",
5344 &[("manifest", "rucc sysroot manifest 3\ntarget\tx86_64-linux-musl\nlib/libc.a\n")],
5345 );
5346 let message =
5347 refused(&[&format!("--sysroot={}", tree.0.display()), "-print-sysroot-provenance"]);
5348 assert!(message.contains("manifest"), "{message}");
5349 assert!(message.contains("1 fields where an input has six"), "{message}");
5350
5351 // The digest is refused for the same file and for a stronger reason: a hash of bytes this
5352 // build cannot read would be a number that names a record nobody can act on.
5353 let digest =
5354 refused(&[&format!("--sysroot={}", tree.0.display()), "-print-sysroot-digest"]);
5355 assert_eq!(digest, message);
5356 }
5357
5358 #[test]
5359 fn the_two_dependency_flags_that_stop_after_the_rule_stop_after_the_rule() {
5360 let (opts, _) = compile(&["-M", "a.c"]);
5361 assert!(opts.deps.emit && opts.deps.instead_of_compiling);
5362 assert!(opts.deps.system_headers, "plain -M lists them");
5363 assert_eq!(opts.emit, EmitKind::Preprocessed);
5364
5365 // Even where a later flag asked for something else, because the family is a mode and
5366 // the mode is what the run is for.
5367 let (opts, _) = compile(&["-M", "-c", "a.c"]);
5368 assert_eq!(opts.emit, EmitKind::Preprocessed);
5369
5370 let (opts, _) = compile(&["-MM", "a.c"]);
5371 assert!(!opts.deps.system_headers);
5372 }
5373
5374 #[test]
5375 fn the_two_that_end_in_d_leave_the_compilation_alone() {
5376 let (opts, _) = compile(&["-MD", "-c", "a.c"]);
5377 assert!(opts.deps.emit && !opts.deps.instead_of_compiling);
5378 assert!(opts.deps.system_headers);
5379 assert_eq!(opts.emit, EmitKind::Object);
5380
5381 let (opts, _) = compile(&["-MMD", "-c", "a.c"]);
5382 assert!(opts.deps.emit && !opts.deps.instead_of_compiling);
5383 assert!(!opts.deps.system_headers);
5384 }
5385
5386 #[test]
5387 fn nothing_puts_the_system_headers_back_once_a_flag_has_taken_them_out() {
5388 // GCC's rule, and not an oversight in it. The flag asking for fewer of them is read as
5389 // the answer, because the other one never asked the question.
5390 let (opts, _) = compile(&["-MM", "-M", "a.c"]);
5391 assert!(!opts.deps.system_headers);
5392 let (opts, _) = compile(&["-MD", "-MMD", "-c", "a.c"]);
5393 assert!(!opts.deps.system_headers);
5394 let (opts, _) = compile(&["-MMD", "-MD", "-c", "a.c"]);
5395 assert!(!opts.deps.system_headers);
5396 }
5397
5398 #[test]
5399 fn a_target_arrives_escaped_from_one_flag_and_untouched_from_the_other() {
5400 let (opts, _) = compile(&["-MM", "-MT", "a b.o", "-MQ", "a b.o", "a.c"]);
5401 assert_eq!(opts.deps.targets, vec!["a b.o".to_owned(), "a\\ b.o".to_owned()]);
5402 }
5403
5404 #[test]
5405 fn the_rest_of_the_family_is_a_file_and_a_switch() {
5406 let (opts, _) = compile(&["-MM", "-MF", "dep.d", "-MP", "a.c"]);
5407 assert_eq!(opts.deps.file.as_deref(), Some("dep.d"));
5408 assert!(opts.deps.phony);
5409
5410 for flag in ["-MF", "-MT", "-MQ"] {
5411 let e = parse_args(&args(&[flag])).unwrap_err();
5412 assert!(e.message.contains("requires an argument"), "{}", e.message);
5413 }
5414 }
5415
5416 /// A directory of sources for one test, removed when the test is done with it.
5417 struct TempTree(PathBuf);
5418
5419 impl Drop for TempTree {
5420 fn drop(&mut self) {
5421 let _ = std::fs::remove_dir_all(&self.0);
5422 }
5423 }
5424
5425 impl TempTree {
5426 fn new(name: &str, files: &[(&str, &str)]) -> TempTree {
5427 let dir = std::env::temp_dir().join(format!("rucc-deps-{}-{name}", std::process::id()));
5428 let _ = std::fs::remove_dir_all(&dir);
5429 std::fs::create_dir_all(&dir).expect("temporary directory should be writable");
5430 for (path, text) in files {
5431 let at = dir.join(path);
5432 if let Some(parent) = at.parent() {
5433 std::fs::create_dir_all(parent).expect("creating a subdirectory should work");
5434 }
5435 std::fs::write(&at, text).expect("writing a temporary file should work");
5436 }
5437 TempTree(dir)
5438 }
5439
5440 fn path(&self, name: &str) -> String {
5441 self.0.join(name).to_string_lossy().into_owned()
5442 }
5443 }
5444
5445 #[test]
5446 fn the_rule_names_what_the_includes_found_and_names_each_of_them_once() {
5447 // End to end, because the list comes from the preprocessor and the format comes from
5448 // somewhere else, and a test of either half on its own would pass with the two of them
5449 // wired up backwards.
5450 let tree = TempTree::new(
5451 "found",
5452 &[
5453 ("a.c", "#include \"one.h\"\n#include \"two.h\"\nint main(void) { return X; }\n"),
5454 ("one.h", "#define X 0\n"),
5455 ("two.h", "#include \"one.h\"\n"),
5456 ],
5457 );
5458 let out = tree.path("dep.d");
5459 let code = run(&args(&["-MM", "-MF", &out, "-o", &tree.path("a.i"), &tree.path("a.c")]));
5460 assert_eq!(code, 0);
5461
5462 let text = std::fs::read_to_string(&out).expect("the rule should have been written");
5463 let names: Vec<&str> = text.split_whitespace().collect();
5464 // The target, the source, and each header once however many times it was reached.
5465 assert_eq!(names.first(), Some(&"a.o:"), "{text}");
5466 assert_eq!(names.iter().filter(|n| n.ends_with("one.h")).count(), 1, "{text}");
5467 assert_eq!(names.iter().filter(|n| n.ends_with("two.h")).count(), 1, "{text}");
5468 // And the `-o` went to the file the rule replaced, which is left empty rather than
5469 // absent because a makefile that named it as a target will look for it.
5470 assert_eq!(std::fs::read(tree.path("a.i")).expect("the output should exist"), b"");
5471 }
5472
5473 #[test]
5474 fn a_header_that_is_only_reached_under_a_guard_is_still_a_dependency() {
5475 // The multiple-include optimization means the second reach never opens the file. It is
5476 // still a file this translation unit was built from, so it is still in the rule.
5477 let tree = TempTree::new(
5478 "guarded",
5479 &[
5480 ("a.c", "#include \"g.h\"\n#include \"g.h\"\nint main(void) { return 0; }\n"),
5481 ("g.h", "#ifndef G\n#define G\n#endif\n"),
5482 ],
5483 );
5484 let out = tree.path("dep.d");
5485 let code = run(&args(&["-MM", "-MF", &out, "-o", &tree.path("a.i"), &tree.path("a.c")]));
5486 assert_eq!(code, 0);
5487 let text = std::fs::read_to_string(&out).expect("the rule should have been written");
5488 assert_eq!(text.split_whitespace().filter(|n| n.ends_with("g.h")).count(), 1, "{text}");
5489 }
5490
5491 #[test]
5492 fn every_imacros_file_is_read_before_every_include_file_whatever_order_they_were_written() {
5493 // Measured against GCC rather than read: the two flags the other way round produce the
5494 // same output byte for byte, so the command line order between the two families does not
5495 // decide anything and the order within one does. The `-include` file here can only see
5496 // the definition if the `-imacros` file that was written after it ran first.
5497 let tree = TempTree::new(
5498 "preinclude",
5499 &[
5500 ("a.c", "int main(void) { return 0; }\n"),
5501 ("i.h", "#ifdef FROM_MACROS\nint saw_it;\n#else\nint missed_it;\n#endif\n"),
5502 ("m.h", "#define FROM_MACROS 1\nint macros_text;\n"),
5503 ],
5504 );
5505 let out = tree.path("a.i");
5506 let code = run(&args(&[
5507 "-E",
5508 "-include",
5509 &tree.path("i.h"),
5510 "-imacros",
5511 &tree.path("m.h"),
5512 "-o",
5513 &out,
5514 &tree.path("a.c"),
5515 ]));
5516 assert_eq!(code, 0);
5517 let text = std::fs::read_to_string(&out).expect("the output should have been written");
5518 assert!(text.contains("saw_it"), "{text}");
5519 // And the text of the `-imacros` file is thrown away, which is the whole difference
5520 // between the two flags.
5521 assert!(!text.contains("macros_text"), "{text}");
5522 }
5523
5524 #[test]
5525 fn a_file_the_command_line_named_is_a_prerequisite_the_same_as_one_a_directive_named() {
5526 let tree = TempTree::new(
5527 "preinclude-deps",
5528 &[
5529 ("a.c", "int main(void) { return 0; }\n"),
5530 ("i.h", "int from_include;\n"),
5531 ("m.h", "#define M 1\n"),
5532 ],
5533 );
5534 let out = tree.path("dep.d");
5535 let code = run(&args(&[
5536 "-MM",
5537 "-MF",
5538 &out,
5539 "-include",
5540 &tree.path("i.h"),
5541 "-imacros",
5542 &tree.path("m.h"),
5543 "-o",
5544 &tree.path("a.i"),
5545 &tree.path("a.c"),
5546 ]));
5547 assert_eq!(code, 0);
5548 let text = std::fs::read_to_string(&out).expect("the rule should have been written");
5549 assert!(text.contains("i.h"), "{text}");
5550 assert!(text.contains("m.h"), "{text}");
5551 }
5552
5553 #[test]
5554 fn a_command_line_include_that_is_nowhere_on_the_path_is_an_error_and_not_a_warning() {
5555 // Including the directory of the source file, which is not on the path for these: the
5556 // command line was not written there, so a name in it is relative to where the compiler
5557 // was run rather than to where the source sits.
5558 let tree = TempTree::new(
5559 "preinclude-missing",
5560 &[("sub/a.c", "int main(void) { return 0; }\n"), ("sub/beside.h", "int x;\n")],
5561 );
5562 let code = run(&args(&["-E", "-include", "beside.h", "-o", "-", &tree.path("sub/a.c")]));
5563 assert_eq!(code, 1);
5564 }
5565
5566 #[test]
5567 fn a_command_line_that_links_names_the_executable_and_not_the_object_it_went_through() {
5568 // The object a link goes through is in a temporary directory and is gone before `make`
5569 // reads any of this, so the rule that named it would be a rule for a file that is never
5570 // there. The target and the file are both the `-o`, which is the executable.
5571 let (opts, plan) = compile(&["-MD", "sub/a.c", "-o", "prog"]);
5572 assert_eq!(plan.output.as_deref(), Some("prog"));
5573 assert_eq!(deps::default_target("sub/a.c", deps_target_output(&opts, &plan)), "prog");
5574 assert_eq!(
5575 deps::default_file(&opts.deps, "sub/a.c", plan.output.as_deref()).as_deref(),
5576 Some("prog.d")
5577 );
5578 }
5579
5580 #[test]
5581 fn the_plan_keeps_the_output_name_because_the_rule_is_written_from_it() {
5582 let (_, plan) = compile(&["-MMD", "-c", "sub/a.c", "-o", "obj/x.o"]);
5583 assert_eq!(plan.output.as_deref(), Some("obj/x.o"));
5584 let (_, plan) = compile(&["-MMD", "-c", "sub/a.c"]);
5585 assert_eq!(plan.output, None);
5586 }
5587
5588 #[test]
5589 fn usage_fits_on_a_screen() {
5590 // Not a style preference. A help text that scrolls is one nobody reads, and this is
5591 // the cheapest way to keep it honest as flags accumulate. The number goes up only when
5592 // a family of flags arrives that has nowhere to share a line, which the two pass gates
5593 // were and which the two fuel flags and `-fsafety=` now are, and it goes up by exactly
5594 // the lines that family took. The four it went up by last are the flags a build system
5595 // passes without being asked to: how much to say, what machine to generate for, threads,
5596 // and the questions `configure` asks before it compiles anything. The one it went up by
5597 // last is the second line of `--emit`, whose kinds are a family that has now outgrown
5598 // one line and has nowhere else to go. The two it went up by last are the dependency
5599 // family, which is eight flags that share nothing with anything above them. The one it
5600 // went up by last is the four spellings of position independent code, which every
5601 // configure script writes and which could only have shared the link line, and that line
5602 // is already four characters short of the limit. The two it went up by last are the rest
5603 // of the include family, which is six more flags that change where a header is looked for
5604 // and two that name a header outright. The one it went up by last is the pair that keeps
5605 // the intermediate files and times the steps, which belong next to the two flags above
5606 // them that are also about watching a compilation rather than changing one. The two it
5607 // went up by last are the section flags and the visibility flag, which are what a build
5608 // that cares about the size of what it ships and about which names it exports writes, and
5609 // the second of them was already taken and only missing from here. The one it went up by
5610 // last is the stack protector, which is four spellings of one question and which every
5611 // distribution puts on every command line it issues, so a build that reads this list
5612 // looking for it and does not find it has to go and read the specification instead. The one
5613 // it went up by last is the profiler, which is two spellings of the request and two of
5614 // where the call goes, and which is about watching a program run rather than about what is
5615 // generated, so it shares its subject with nothing above it. The one it went up by last is
5616 // the room a function opens with for something to be written over it later, which takes an
5617 // argument of its own shape and is what a kernel build asks for, so it fits beside the
5618 // profiler and nothing else. The one it went up by last is what overflows rather than being
5619 // undefined, which is three spellings of two questions and which a kernel build and a great
5620 // deal of code written before the standard settled both pass. The one it went up by last is
5621 // the other answer to the first of those questions, which could not share the line because
5622 // what it asks for is the opposite of what the flags on that line ask for. The one it went
5623 // up by last is the split of the line that lists what this compiler does anyway into that
5624 // and what it assumes anyway, which are two different claims that were sharing a line until
5625 // the second of them got a second flag and the line stopped fitting. The one it went up by
5626 // last is the three flags that change the ABI rather than the code, which have to be given
5627 // to every file in a program or none of them and which therefore belong somewhere a person
5628 // reading this list will see them. The one it went up by last is the floating point group,
5629 // which is two lines rather than one because the first of them is a choice this compiler
5630 // records and the rest are claims about what it does anyway, and putting a real setting on
5631 // the same line as three flags that change nothing would be misleading about both. The one
5632 // it went up by last is the flag that says a write has to stay inside the member it names,
5633 // which is a setting rather than a claim and so cannot share the line above it, that being
5634 // the one that picks a tier. The two it went up by last are the prefix mapping family,
5635 // which is four flags whose whole job is to keep a build's output the same from two
5636 // different directories, and which a person chasing a reproducible build comes here
5637 // looking for by name. The one it went up by last is how the debug sections are compressed
5638 // and whether they go in a file of their own, which are two questions about the shape of
5639 // the debug output, where the line above them is about how much of it there is. The one it
5640 // went up by last is the `restrict` contract, which is a setting for the same reason the
5641 // flag that keeps a write inside its member is and which is the check a person who has been
5642 // bitten by a vectorizer comes here looking for. The one it went up by last is link time
5643 // optimization, which is a whole optimization rather than a flag and which says so on its
5644 // own line, because a build that passes it and reads this looking for what it got is
5645 // asking a question no other line here answers. The one it went up by last is the sysroot,
5646 // which is the question somebody asks when a cross build read a file nobody expected, and
5647 // which has no room on the line above it because the answers there are a path each and this
5648 // one is the root all of them are under. The one it went up by last is what is inside that
5649 // root and where each of it came from, which is a question about a whole tree rather than
5650 // about a path and which is long enough on its own that it could not have shared a line with
5651 // anything. The one it went up by last is the profile family, which splits down the middle
5652 // where no other family here does, so the line has to name the half that is taken and the
5653 // half that is refused or it would be read as taking both. The one it went up by last is
5654 // the sanitizers, which are what somebody reaching for a checked build writes first and
5655 // which belong beside the tier that is the nearest thing here to what they asked for. The
5656 // one it went up by last is the digest of that record, which is the same tree as one number
5657 // and could not share the line above it because that line prints a few hundred lines and
5658 // this one prints sixty four characters, and a reader who wants the short answer is looking
5659 // for it by name rather than reading the long one. The one it went up by last is the
5660 // sysroot fetch, which is the only command here that gets something from somewhere else and
5661 // is therefore the one a person wants to have read before they run it rather than after.
5662 // And the flag beside it that forbids every download, which earns its line by being what a
5663 // build in a sealed environment passes and by meaning something even though an ordinary
5664 // compile downloads nothing either way.
5665 assert!(USAGE.lines().count() < 72, "usage text has grown past one screen");
5666 }
5667}