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/0.10.47")]
30
31pub mod cache;
32pub mod compile;
33pub mod deps;
34pub mod fetch;
35mod glibc;
36pub mod install;
37pub mod library;
38pub mod link;
39mod map;
40pub mod phase;
41pub mod preprocess;
42pub mod schedule;
43
44use std::fmt::Write as _;
45use std::io::Write as _;
46use std::path::PathBuf;
47
48use rucc_codegen::coverage::{self, Fired};
49use rucc_codegen::lowering::Lowerings;
50use rucc_codegen::pressure::Pressure;
51use rucc_pp::Dependency;
52use rucc_session::{
53 Compress, Control, Dumps, EmitKind, Hook, Options, Pic, PrefixMap, Preinclude, Protector,
54 SaveTemps, Session, Std, Wrapping, runtime,
55};
56use rucc_sysroot::{Manifest, Sysroot};
57use rucc_target::{ObjectFormat, Triple};
58use rucc_tuple::TargetTuple;
59
60use crate::link::LinkOptions;
61
62pub use crate::compile::{Artifact, Compiled, Temps, compile, compile_ir};
63pub use crate::phase::{ArchiveJob, Input, InputKind, Job, LinkJob, Output, Phase, Plan};
64pub use crate::preprocess::{OsFileSystem, Preprocessed, preprocess};
65pub use crate::schedule::Jobs;
66
67/// The compiler's version, taken from the workspace manifest.
68pub const VERSION: &str = env!("CARGO_PKG_VERSION");
69
70/// What the command line asked for.
71#[derive(Debug, Clone, PartialEq, Eq)]
72pub enum Action {
73 /// Print usage and exit successfully.
74 Help,
75 /// Print the version and exit successfully.
76 Version,
77 /// Print one line and exit successfully, which is what the `-dump` and `-print` family do.
78 ///
79 /// A build system asks these before it compiles anything, and what it does with the answer
80 /// is paste it into a path or into another command line, so each one is a single line with
81 /// no decoration around it.
82 Print(String),
83 /// Print the resolved configuration and exit successfully.
84 PrintConfig(Box<Options>),
85 /// Print the passes the level will run and exit successfully.
86 PrintPipeline(Box<Options>),
87 /// Print the phase plan and the link line and exit successfully, which is `-###`.
88 PrintPlan {
89 /// The resolved options, which is what says what the link line is for.
90 opts: Box<Options>,
91 /// What to do to each input, and in what order.
92 plan: Box<Plan>,
93 /// What the command line said about linking.
94 link: Box<LinkOptions>,
95 },
96 /// `--fetch <tuple>`, which gets the sysroot this release pins for a target and installs it.
97 ///
98 /// The only action in this compiler that may run another program to move bytes onto the
99 /// machine, which is `spec/cross-compile/13-distribution.md` section 13.8's rule rather than a
100 /// property of how this happens to be written: a compilation has no branch that reaches it.
101 Fetch {
102 /// The artifact, from the table in [`rucc_sysroot::artifact`]. Resolved here rather than where the
103 /// work happens, so that a target nothing is pinned for is a refusal from the parser like
104 /// every other thing a command line can ask for and not have.
105 what: &'static rucc_sysroot::Pinned,
106 /// The target, which names the directory under the cache the tree is installed at and is
107 /// checked against the record inside the artifact.
108 target: TargetTuple,
109 /// Where the cache is, read where everything else that needs it reads it.
110 cache: PathBuf,
111 },
112 /// Compile the given inputs.
113 Compile {
114 /// The resolved options.
115 opts: Box<Options>,
116 /// What to do to each input, and in what order.
117 plan: Box<Plan>,
118 /// What the command line said about linking.
119 link: Box<LinkOptions>,
120 /// How many translation units to compile at once.
121 jobs: Jobs,
122 /// Whether `-v` asked for the plan to be printed while it runs.
123 verbose: bool,
124 /// What is worth saying about the command line before anything is compiled, printed as
125 /// warnings and once for the whole run rather than once per file.
126 ///
127 /// These are not diagnostics. A diagnostic is about a piece of source and has a span to
128 /// point at, and these are about the way two flags were combined, so there is nothing to
129 /// point at and nowhere below the driver that knows both halves. `-w` does not reach them
130 /// for the same reason it does not reach a refusal from the parser.
131 notes: Vec<String>,
132 },
133}
134
135/// Why a command line was rejected.
136#[derive(Debug, Clone, PartialEq, Eq)]
137pub struct CliError {
138 /// The message, lowercase and without a trailing period, in the same shape as any other
139 /// diagnostic.
140 pub message: String,
141}
142
143impl std::fmt::Display for CliError {
144 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
145 f.write_str(&self.message)
146 }
147}
148
149impl std::error::Error for CliError {}
150
151fn err(message: impl Into<String>) -> CliError {
152 CliError { message: message.into() }
153}
154
155/// The two halves of one prefix mapping flag's argument, where `flag` includes its trailing `=`.
156///
157/// The split is at the last `=` in what follows the flag, not the first, which is gcc's rule and
158/// the only one that lets a directory whose name contains an `=` be the old half. It also means
159/// `-fmacro-prefix-map=a=b=c` rewrites `a=b` to `c` rather than `a` to `b=c`, which looks like a
160/// trap until you notice the alternative traps the far more common case.
161fn rewrite<'a>(arg: &'a str, flag: &str) -> Result<(&'a str, &'a str), CliError> {
162 let rest = &arg[flag.len()..];
163 PrefixMap::split(rest).ok_or_else(|| {
164 let flag = flag.trim_end_matches('=');
165 err(format!(
166 "`{rest}` is not a rewrite for `{flag}`, which is an old prefix, an `=` and a new one"
167 ))
168 })
169}
170
171/// A question the command line asked instead of asking for a compilation.
172///
173/// These are answered after the loop rather than where they are read, because every one of them
174/// is about the target or about the library search and the last word on both is the end of the
175/// command line.
176enum Query {
177 /// `-dumpmachine`, the triple.
178 Machine,
179 /// `-dumpversion` and `-dumpfullversion`, which are the same three numbers here.
180 Version,
181 /// `-print-multiarch`, the directory name a distribution files this target under.
182 Multiarch,
183 /// `-print-search-dirs`, in the three lines GCC prints.
184 SearchDirs,
185 /// `-print-sysroot`, the root the headers and the libraries are read under.
186 Sysroot,
187 /// `-print-sysroot-provenance`, what is in that root and where each of it came from.
188 SysrootProvenance,
189 /// `-print-sysroot-digest`, the one number that names all of it.
190 SysrootDigest,
191 /// `-print-file-name=<name>`, the full path of a library file.
192 FileName(String),
193 /// `-print-prog-name=<name>`, the full path of a program.
194 ProgName(String),
195 /// `-print-libgcc-file-name`, which is `-print-file-name=libgcc.a` under another spelling.
196 Libgcc,
197}
198
199/// Usage text.
200///
201/// Deliberately short. `spec/04-driver-and-cli.md` puts the full flag reference in the
202/// manual page, because a `--help` nobody can read in one screen is a `--help` nobody reads.
203pub const USAGE: &str = "\
204rucc, an optimizing C compiler
205
206usage: rucc [options] file...
207
208options:
209 -c compile and assemble, do not link
210 -S compile only, emit assembly
211 -E preprocess only
212 -o <file> write output to <file>, or to standard output for -
213 -D <name>[=<value>], -U <name> define a macro, or undefine one after every -D
214 -I <dir> add <dir> to the include search path
215 -iquote -isystem -idirafter <dir> the other chains, -nostdinc drops ours
216 -I-, -iprefix <p>, -iwithprefix[before] <dir> the older spellings of those
217 -include <file>, -imacros <file> read <file> first, the second for its macros only
218 --sysroot=<dir> look for the library's headers under <dir>, -isysroot too
219 -P, -dM with -E: leave out the markers, or dump the macros
220 -M -MM -MD -MMD write a make rule for the source, the last two compile as well
221 -MF <file> -MT <t> -MQ <t> -MP where the rule goes, what it builds, targets with no recipe
222 -std=<dialect> c89 through c23, and the gnu spellings
223 -fgnuc-version=<v> the GCC release to claim, default 7.0.0
224 -x <lang> treat later inputs as <lang>, or none to stop
225 -O<level> optimize: 0, 1, 2, 3, s, z
226 -fsafety=<tier> check memory safety: off, detect, enforce, kernel
227 -f[no-]sanitize=<what> the negative is taken, the positive is refused by name
228 -f[no-]safety-subobject a write has to stay inside the member it names
229 -f[no-]safety-restrict two restrict pointers of one block may not meet
230 -f<pass> -fno-<pass> -fdump-ir=<what> -fopt-info[-<kind>][=FILE]
231 -fpass-fuel=<pass>=<n>, -fpass-fuel-global=<n> stop a pass, or all of them, after n
232 -fdisable-<pass>[=<funcs>], -fenable-<pass>[=<funcs>] run a pass on some functions only
233 -g -g0 -gdwarf-5, -fno-omit-frame-pointer, -mno-red-zone debug info, frame pointer, red zone
234 -gz[=none|zlib|zlib-gnu|zstd] -gno-split-dwarf compress debug sections, one file not two
235 -flto[=auto|jobserver|<n>] -fno-lto -ffat-lto-objects read, and not done yet
236 -fprofile-use[=<path>] -fprofile-dir=<dir> read too, where -fprofile-generate is refused
237 -f[no-]stack-protector[-strong|-all], -f[no-]stack-clash-protection, -fcf-protection=<edges>
238 -ffunction-sections -fdata-sections a section per function or variable, for --gc-sections
239 -fvisibility=<what> default, hidden, internal or protected, when nothing in the source said
240 -l<name>, -L <dir>, -B <dir> link a library, where to look for one, where our own tools are
241 -fPIC -fpic -fPIE -fpie, -fno-common, -pipe what it does anyway
242 -f[no-]strict-aliasing, -f[no-]delete-null-pointer-checks what it assumes anyway
243 -static -shared -pie -no-pie -nostdlib -nostartfiles -nodefaultlibs -rdynamic -s how to link
244 -Wl,<arg>, -Xlinker <arg>, -fuse-ld=<name> hand an argument to the linker, or pick one
245 -Werror -pedantic -pedantic-errors -w how much to say, and whether it is fatal
246 -m64 -march= -mtune= -mcpu= -mabi= -mcmodel= what machine to generate for
247 -pg -p, -mfentry -mno-fentry call a profiler on the way in, and where that call goes
248 -fpatchable-function-entry=<n>[,<m>] room at the top of every function to patch later
249 -fwrapv, -fwrapv-pointer, -fno-strict-overflow signed or pointer overflow wraps
250 -ftrapv signed overflow stops the program instead
251 -f[no-]signed-char, -f[no-]unsigned-char, -f[no-]short-enums change the ABI
252 -ffp-contract=<how> fuse a multiply and an addition: fast, on or off
253 -fexcess-precision=<how>, -f[no-]rounding-math, -f[no-]trapping-math what may be folded
254 -ffile-prefix-map=<old>=<new> rewrite that front of every path we put in the output
255 -fmacro-prefix-map= -fdebug-prefix-map= -fprofile-prefix-map= the same, one output each
256 -pthread build for more than one thread, and link the library for it
257 -dumpmachine -dumpversion -print-multiarch -print-search-dirs what this compiler is
258 -print-file-name=<name> -print-prog-name=<name> where a file or a program is
259 -print-sysroot the root the headers and the libraries are read under
260 -print-sysroot-provenance every input under it, where it came from and its licence
261 -print-sysroot-digest the sha256 of that record, which names the whole sysroot in one line
262 --fetch <tuple> get the sysroot this release pins for <tuple> and install it in the cache
263 --offline never download anything, which a compilation never does anyway
264 -j[n] compile n translation units at once, default all
265 -v, -### print each phase as it runs, or without running any
266 -save-temps[=cwd|obj], -time keep the .i and the .s, say how long each step took
267 --target=<triple> generate code for <triple>
268 --emit=<kind> exe, obj, archive, asm, preprocessed, tast, ir, mir-final,
269 safety-summary, type-granules
270 --print-config, --print-pipeline print the configuration or the pipeline, and exit
271 --version print the version and exit
272 -h, --help print this message and exit
273
274See spec/04-driver-and-cli.md for the full flag reference.
275";
276
277/// The argument of a flag that may be joined to it or may be the next word.
278///
279/// `-DFOO` and `-D FOO` are the same thing, and `at` is where the flag's own letters end.
280fn joined_or_next(
281 arg: &str,
282 at: usize,
283 args: &[String],
284 i: &mut usize,
285) -> Result<String, CliError> {
286 if arg.len() > at {
287 return Ok(arg[at..].to_owned());
288 }
289 let next = args.get(*i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
290 *i += 1;
291 Ok(next.clone())
292}
293
294/// Every name that may follow `-fsanitize=`, which is gcc 16's list and three of this compiler's
295/// own.
296///
297/// The three are on it because `spec/07-types-and-semantics.md` section 7.7 already promises them:
298/// each undefined behaviour this compiler exploits is listed there with the check that detects it,
299/// and `alias`, `restrict` and `memory` are checks gcc has no spelling for. gcc refuses `memory`
300/// outright, since the sanitizer of that name is clang's. A name being here means it is a name
301/// rather than a typo, and nothing more than that: every one of them is refused after the loop,
302/// because none of them is implemented.
303///
304/// `all` is deliberately absent. gcc takes it only in the negative, so it is handled where each of
305/// those two spellings is read rather than by being on this list.
306const SANITIZERS: [&str; 34] = [
307 "address",
308 "kernel-address",
309 "hwaddress",
310 "kernel-hwaddress",
311 "pointer-compare",
312 "pointer-subtract",
313 "thread",
314 "leak",
315 "undefined",
316 "shift",
317 "shift-base",
318 "shift-exponent",
319 "integer-divide-by-zero",
320 "unreachable",
321 "vla-bound",
322 "null",
323 "return",
324 "signed-integer-overflow",
325 "bounds",
326 "bounds-strict",
327 "alignment",
328 "object-size",
329 "float-divide-by-zero",
330 "float-cast-overflow",
331 "nonnull-attribute",
332 "returns-nonnull-attribute",
333 "bool",
334 "enum",
335 "vptr",
336 "pointer-overflow",
337 "builtin",
338 "alias",
339 "restrict",
340 "memory",
341];
342
343/// Parses a command line, without the program name.
344///
345/// # Errors
346///
347/// Returns the message to print when the arguments do not name a compilation this compiler
348/// can attempt.
349pub fn parse_args(args: &[String]) -> Result<Action, CliError> {
350 let host = Triple::host()
351 .ok_or_else(|| err("this host is not a supported target and no --target was given"))?;
352 let mut opts = Options::new(host);
353 let mut inputs: Vec<Input> = Vec::new();
354 let mut print_config = false;
355 let mut print_pipeline = false;
356 let mut print_plan = false;
357 let mut verbose = false;
358 let mut jobs = Jobs::default();
359 let mut nostdinc = false;
360 let mut sysroot: Option<PathBuf> = None;
361 // What the command line is worth warning about, filled in after the loop rather than during it,
362 // because every question of this kind is about two flags and the last word on both of them is
363 // the end of the loop.
364 let mut notes: Vec<String> = Vec::new();
365 // The whole ten field target, kept beside the three field one because `--target=` can pin a
366 // libc version and `Triple` has nowhere to put it. It decides `__GLIBC_MINOR__` and nothing
367 // else today, and `None` is a command line that named no target, which is this machine.
368 let mut pinned: Option<TargetTuple> = None;
369 let mut output = None;
370 let mut link = LinkOptions::default();
371 let mut query: Option<Query> = None;
372 // What `--fetch` named, and whether `--offline` forbade it. Both are weighed after the loop
373 // because either can be written after the other.
374 let mut fetch: Option<String> = None;
375 let mut offline = false;
376 let mut threads = false;
377 // Which sanitizers are still asked for by the end of the command line. Accumulated across the
378 // loop rather than answered where it was read, because `-fno-sanitize=` turns one off and a
379 // build that asks for a check and then takes it back has asked for nothing. What happens to a
380 // set that is not empty is decided after the loop.
381 let mut sanitizers: Vec<&str> = Vec::new();
382 // `-x` applies to inputs that come after it and stays in effect until the next one, which
383 // is why it is tracked across the loop rather than attached to a single argument.
384 let mut forced: Option<InputKind> = None;
385 // What `-iprefix` last said, stuck on the front of every later `-iwithprefix`. It applies to
386 // the flags after it and not the ones before, so a command line may set it more than once.
387 // GCC's default is its own installed header directory with the last component taken off,
388 // which is a path a cross compiler's build system knows and passes; there is no equivalent
389 // here, so with no `-iprefix` the prefix is nothing and `-iwithprefix` names a directory
390 // outright.
391 let mut iprefix = String::new();
392
393 let mut i = 0;
394 while i < args.len() {
395 let arg = args[i].as_str();
396 i += 1;
397 match arg {
398 "-h" | "--help" => return Ok(Action::Help),
399 "--version" => return Ok(Action::Version),
400 // The sysroot fetch, which is weighed after the loop rather than acted on here, because
401 // `--offline` written after it has to be able to forbid it. Both spellings, since a
402 // flag that takes a tuple gets written both ways and neither is a guess at what the
403 // other meant.
404 "--fetch" => {
405 let value = args
406 .get(i)
407 .ok_or_else(|| err("--fetch requires the target to get a sysroot for"))?;
408 i += 1;
409 fetch = Some(value.clone());
410 }
411 _ if arg.starts_with("--fetch=") => {
412 fetch = Some(arg["--fetch=".len()..].to_owned());
413 }
414 // Accepted on any command line and only ever read by the fetch, because an ordinary
415 // compile downloads nothing with or without it. So this flag takes nothing away today,
416 // which is the property section 13.2 asks for rather than an omission: a build that
417 // passes it is saying what it expects of this compiler, and what it expects is already
418 // true.
419 "--offline" => offline = true,
420 "--print-config" => print_config = true,
421 "--print-pipeline" => print_pipeline = true,
422 "-###" => print_plan = true,
423 "-v" => verbose = true,
424 // The files a compilation goes through, kept rather than thrown away. The bare
425 // spelling means `=obj` and not `=cwd`, which is not what the manual says and is what
426 // gcc 16 does; `SaveTemps::Object` carries the measurement.
427 "-save-temps" => opts.save_temps = SaveTemps::Object,
428 _ if arg.starts_with("-save-temps=") => {
429 opts.save_temps = arg["-save-temps=".len()..].parse().map_err(err)?;
430 }
431 // How long each step took. A misspelling of this is worth rejecting rather than
432 // ignoring, since a run that says nothing looks like a compilation that took no time.
433 "-time" => opts.time = true,
434 "-c" => opts.emit = EmitKind::Object,
435 "-S" => opts.emit = EmitKind::Asm,
436 "-E" => opts.emit = EmitKind::Preprocessed,
437 "-g" => opts.debug_info = true,
438 // GCC's own levels of how much debug information to write. Zero is none and every
439 // other number is some, and this compiler has one amount, so the numbers above zero
440 // all mean the same thing here. `-ggdb` is the same flag asking for whatever the
441 // debugger on the machine prefers, which is what we emit anyway.
442 "-g0" => opts.debug_info = false,
443 "-g1" | "-g2" | "-g3" | "-ggdb" | "-ggdb1" | "-ggdb2" | "-ggdb3" => {
444 opts.debug_info = true;
445 }
446 // The version of DWARF to write. We write DWARF 5 and nothing else, so a build that
447 // asks for another version is told rather than handed a file it cannot read.
448 "-gdwarf" | "-gdwarf-5" => opts.debug_info = true,
449 _ if arg.starts_with("-gdwarf-") => {
450 return Err(err(format!(
451 "{arg}: this compiler writes DWARF 5 and no other version, see \
452 spec/11-debug-info.md"
453 )));
454 }
455 // Whether the debug information goes in a file of its own beside the object. gcc
456 // writes that `.dwo` whether or not it found anything to put in it, which means a
457 // build system that declares the file as an output gets one and a make rule that
458 // depends on it fires. Refused for that reason rather than taken: section 4.1 takes a
459 // flag that changes nothing and refuses one that changes what is produced, and a file
460 // that does not appear is the plainest change of that kind there is. The negative
461 // spelling is taken, because putting it all in the object is what happens anyway.
462 "-gno-split-dwarf" => {}
463 "-gsplit-dwarf" => {
464 return Err(err(format!(
465 "{arg}: this compiler writes no separate `.dwo` file, and a build that \
466 expects one beside each object would wait for a file that never arrives, \
467 see spec/11-debug-info.md"
468 )));
469 }
470 // How the debug sections are compressed. There are none yet, so every answer produces
471 // the same bytes and taking the flag promises nothing that is not kept. The value is
472 // still checked, because a typo in a distribution's flags is worth finding when the
473 // compiler reads it rather than when somebody later wonders why nothing got smaller.
474 // Bare `-gz` means `zlib`, which the manual leaves for the reader to discover.
475 "-gz" => opts.compress = Compress::Zlib,
476 _ if arg.starts_with("-gz=") => {
477 let how = &arg["-gz=".len()..];
478 opts.compress = how.parse().map_err(|()| {
479 err(format!(
480 "`{how}` is not a way to compress debug sections, which is none, zlib, \
481 zlib-gnu or zstd"
482 ))
483 })?;
484 }
485 "-Werror" => opts.warnings_are_errors = true,
486 // Nothing that is not fatal is said at all. Read at the one place a diagnostic goes
487 // through rather than here, so that a warning `-w` dropped is not counted either.
488 "-w" => opts.warnings = false,
489 "-pedantic-errors" => {
490 opts.pedantic = true;
491 opts.warnings_are_errors = true;
492 }
493 "-P" => opts.line_markers = false,
494 // The dependency family, which section 4.4 calls required because every build system
495 // that generates its own makefiles asks for it. The two that end in `D` write a file
496 // beside the object and let the compilation happen, and the two that do not write to
497 // standard output and stop after it. Nothing here turns the system headers back on
498 // once a flag has turned them off, which is GCC's behaviour and is why `-MM -M` is
499 // `-MM`: the flag asking for fewer of them is the one with something to say.
500 "-M" => {
501 opts.deps.emit = true;
502 opts.deps.instead_of_compiling = true;
503 }
504 "-MM" => {
505 opts.deps.emit = true;
506 opts.deps.instead_of_compiling = true;
507 opts.deps.system_headers = false;
508 }
509 "-MD" => opts.deps.emit = true,
510 "-MMD" => {
511 opts.deps.emit = true;
512 opts.deps.system_headers = false;
513 }
514 "-MP" => opts.deps.phony = true,
515 // These three take a word and only in the separated form, which is how GCC spells
516 // them and how every build system writes them.
517 "-MF" | "-MT" | "-MQ" => {
518 let value =
519 args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
520 i += 1;
521 match arg {
522 "-MF" => opts.deps.file = Some(value.clone()),
523 // The whole of the difference between the two. `-MT` is for a build that has
524 // already escaped what it is passing, and `-MQ` is for one that has a name
525 // and wants it to arrive as that name.
526 "-MT" => opts.deps.targets.push(value.clone()),
527 _ => opts.deps.targets.push(deps::escaped(value)),
528 }
529 }
530 // The questions a build system asks before it compiles anything. Answered after the
531 // loop, because each one is about the target or the library search and the command
532 // line has not finished saying what those are.
533 "-dumpmachine" => query = Some(Query::Machine),
534 "-dumpversion" | "-dumpfullversion" => query = Some(Query::Version),
535 "-print-multiarch" => query = Some(Query::Multiarch),
536 "-print-search-dirs" => query = Some(Query::SearchDirs),
537 "-print-sysroot" => query = Some(Query::Sysroot),
538 // Both spellings, because this one is ours rather than GCC's and our own documents
539 // write it both ways: section 13.5 of `spec/cross-compile/13-distribution.md` gives it
540 // two dashes like the other flags we invented, and document 12's table gives it one
541 // like the `-print-` family it sits in. A person who reads either and types what it
542 // says is right, so neither is refused.
543 "-print-sysroot-provenance" | "--print-sysroot-provenance" => {
544 query = Some(Query::SysrootProvenance);
545 }
546 "-print-sysroot-digest" | "--print-sysroot-digest" => {
547 query = Some(Query::SysrootDigest);
548 }
549 "-print-libgcc-file-name" => query = Some(Query::Libgcc),
550 _ if arg.starts_with("-print-file-name=") => {
551 query = Some(Query::FileName(arg["-print-file-name=".len()..].to_owned()));
552 }
553 _ if arg.starts_with("-print-prog-name=") => {
554 query = Some(Query::ProgName(arg["-print-prog-name=".len()..].to_owned()));
555 }
556 // A program built to run in more than one thread. On every platform this compiler
557 // targets that is a macro the library's headers read and one more library on the
558 // link line, and the library is added after the loop so that it lands after the
559 // objects that refer to it.
560 "-pthread" | "-pthreads" => {
561 opts.defines.push("_REENTRANT".to_owned());
562 threads = true;
563 }
564 "-ansi" => {
565 opts.std = Std::C89;
566 opts.gnu_extensions = false;
567 }
568 // `-Wpedantic` is the same flag under the name the `-W` family gives it, which is
569 // the spelling a build system that groups its warning flags tends to write.
570 "-pedantic" | "-Wpedantic" => opts.pedantic = true,
571 // Both directions, because a build that needs this for one directory turns it back
572 // off for the next one rather than leaving it on for the whole tree.
573 "-fpermissive" => opts.permissive = true,
574 "-fno-permissive" => opts.permissive = false,
575 "-ffreestanding" => opts.hosted = false,
576 "-fhosted" => opts.hosted = true,
577 "-fno-builtin" => opts.builtins = false,
578 "-fbuiltin" => opts.builtins = true,
579 // The C89 dialects are under GNU's reading whatever this says, so turning it off
580 // there is turning off something the dialect asked for, which is accepted and does
581 // nothing. gcc refuses that command line, and there is nothing it could have meant.
582 "-fgnu89-inline" => opts.gnu89_inline = true,
583 "-fno-gnu89-inline" => opts.gnu89_inline = false,
584 // Both directions of each, because a build system that wants one of these usually
585 // writes it beside the flag that turns it back off for one directory.
586 "-fno-omit-frame-pointer" => opts.frame_pointer = true,
587 "-fomit-frame-pointer" => opts.frame_pointer = false,
588 // Both directions again, for the same reason, and a third answer for a command line
589 // that wrote neither: see `reorder_blocks` in `rucc_session`.
590 "-freorder-blocks" => opts.reorder_blocks = Some(true),
591 "-fno-reorder-blocks" => opts.reorder_blocks = Some(false),
592 // gcc's name for the scheduler that runs after the registers are handed out, which is
593 // the only one rucc has: see `schedule_insns` in `rucc_session`. gcc also takes
594 // `-fschedule-insns` for the pass before allocation, and taking that one here would be
595 // a flag that says a pass ran when none did.
596 "-fschedule-insns2" => opts.schedule_insns = Some(true),
597 "-fno-schedule-insns2" => opts.schedule_insns = Some(false),
598 "-mno-red-zone" => opts.red_zone = false,
599 "-mred-zone" => opts.red_zone = true,
600 // Four flags rather than one with an argument, which is how gcc spells them and how
601 // every build line writes them. Last one wins, because a package build puts
602 // `-fstack-protector-strong` in its global flags and a directory that cannot have one
603 // turns it back off on the line after.
604 "-fno-stack-protector" | "-fno-stack-protector-all" | "-fno-stack-protector-strong" => {
605 opts.protector = Protector::None;
606 }
607 "-fstack-protector" => opts.protector = Protector::Buffers,
608 "-fstack-protector-strong" => opts.protector = Protector::Strong,
609 "-fstack-protector-all" => opts.protector = Protector::All,
610 // The other half of what a hardened build asks for, and it is a question about the
611 // frame rather than about the function, so it is a switch rather than a level.
612 "-fstack-clash-protection" => opts.stack_clash = true,
613 "-fno-stack-clash-protection" => opts.stack_clash = false,
614 // The third of them, and the one that is a question with an argument rather than a
615 // family of spellings, because what it asks about is which of the two edges of a
616 // control flow transfer is checked. Bare is both of them, which is what gcc does.
617 "-fcf-protection" => opts.control = Control::Full,
618 "-fno-cf-protection" => opts.control = Control::None,
619 // Two spellings of the same request, which is what gcc has as well. `-p` was the older
620 // profiler and `-pg` the one that also recorded who called whom, and on every platform
621 // this compiler targets there is now one hook and both ask for it.
622 "-pg" | "-p" => {
623 opts.profile = true;
624 link.profile = true;
625 }
626 // Accepted on their own and doing nothing on their own, which is gcc's behaviour: they
627 // say where the call goes and a command line that asked for no call has nowhere to put
628 // one. That matters because a build system that sets `-mfentry` globally and `-pg` per
629 // directory is a build system that would otherwise fail on every other directory.
630 "-mfentry" => opts.hook = Hook::Early,
631 "-mno-fentry" => opts.hook = Hook::Late,
632 // GCC drops its own include directory along with the system ones, because its
633 // headers are half of a pair with the library's and half a pair is worse than
634 // none. A build that passes this is supplying the whole set itself.
635 "-nostdinc" => nostdinc = true,
636 "-o" => {
637 output = Some(args.get(i).ok_or_else(|| err("-o requires an argument"))?.clone());
638 i += 1;
639 }
640 // The flags that take a directory only in the separated form. GCC spells them
641 // this way and nothing writes `-iquotedir`, so accepting the joined form would
642 // mean guessing at a path that starts with the flag's own letters.
643 // Apple's spelling of `--sysroot`, and the one its own build systems pass. The
644 // two mean the same thing here: the configured directories are under there rather
645 // than under the root.
646 "-isysroot" => {
647 let dir = args.get(i).ok_or_else(|| err("-isysroot requires an argument"))?;
648 i += 1;
649 sysroot = Some(PathBuf::from(dir));
650 }
651 "-iquote" | "-isystem" | "-idirafter" => {
652 let dir = args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
653 i += 1;
654 match arg {
655 "-iquote" => opts.search.push_quote(dir.clone()),
656 "-isystem" => opts.search.push_system(dir.clone()),
657 _ => opts.search.push_after(dir.clone()),
658 }
659 }
660 "-iprefix" => {
661 iprefix = args.get(i).ok_or_else(|| err("-iprefix requires an argument"))?.clone();
662 i += 1;
663 }
664 // Where GCC puts these is not where its manual says it puts them, and this is the
665 // measured answer rather than the documented one: `-iwithprefix` lands in the
666 // `-isystem` slot and not the `-idirafter` slot, and `-iwithprefixbefore` lands in
667 // the `-I` slot. A cross build that uses them is relying on the behaviour, since
668 // that is the compiler it was developed against.
669 "-iwithprefix" | "-iwithprefixbefore" => {
670 let dir = args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
671 i += 1;
672 let dir = format!("{iprefix}{dir}");
673 if arg == "-iwithprefix" {
674 opts.search.push_system(dir);
675 } else {
676 opts.search.push_bracket(dir);
677 }
678 }
679 "-include" | "-imacros" => {
680 let name = args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
681 i += 1;
682 opts.preincludes
683 .push(Preinclude { name: name.clone(), macros_only: arg == "-imacros" });
684 }
685 // The flag `-iquote` was introduced to replace, still passed by build systems old
686 // enough to predate the replacement. It is not a directory: it says that every `-I`
687 // so far is for quoted includes only, and that a quoted include stops looking next
688 // to the file that wrote it.
689 "-I-" => opts.search.split_quote_chain(),
690 "-x" => {
691 let lang = args.get(i).ok_or_else(|| err("-x requires an argument"))?;
692 i += 1;
693 forced = if lang == "none" {
694 None
695 } else {
696 Some(InputKind::from_x_arg(lang).map_err(|e| err(format!("{e}")))?)
697 };
698 }
699 // Not a GCC flag. spec/03-architecture.md section 3.5 compiles several
700 // translation units in one process rather than making the build system fork, and
701 // section 3.8's determinism check compares `-j1` against `-j16`, so the knob has
702 // to exist and has to be spelled the way `make` spells it.
703 // `-DFOO`, `-D FOO` and the same for `-U` and `-I`. Both forms are in wide use
704 // and a build system may produce either, so both are read here rather than
705 // being normalised by whatever generated the command line.
706 _ if arg.starts_with("-D") => {
707 let value = joined_or_next(arg, 2, args, &mut i)?;
708 opts.defines.push(value);
709 }
710 _ if arg.starts_with("-U") => {
711 let value = joined_or_next(arg, 2, args, &mut i)?;
712 opts.undefines.push(value);
713 }
714 _ if arg.starts_with("-I") => {
715 let dir = joined_or_next(arg, 2, args, &mut i)?;
716 opts.search.push_bracket(dir);
717 }
718 _ if arg.starts_with("-std=") => {
719 let name = &arg["-std=".len()..];
720 let (std, gnu) = Std::from_flag(name)
721 .ok_or_else(|| err(format!("unknown dialect `{name}`, see --help")))?;
722 opts.std = std;
723 opts.gnu_extensions = gnu;
724 }
725 // Section 4.5. The claim decides which half of glibc's `sys/cdefs.h` we are
726 // handed, so a differential run that does not set it is comparing two compilers
727 // that believe they are different compilers.
728 // GCC packs these into one flag, so `-dDI` is two of them. Letters in the family
729 // that we have not written yet are accepted and ignored, because a dump is a
730 // debugging aid and a build that asks for one should still compile. A letter
731 // outside the family falls through to the unknown option error, which is what
732 // keeps `-dumpversion` from being read as a dump of nothing.
733 _ if Dumps::is_family(arg) => {
734 opts.dumps.add(&arg[2..]);
735 }
736 // One name at a time, which is what a build that means its own `memcpy` and the
737 // library's everything else writes. The name is not checked against a list, because
738 // the flag is about what the program means by a name and a program is allowed to mean
739 // something by a name this compiler has never heard of.
740 _ if arg.starts_with("-fno-builtin-") => {
741 opts.no_builtin.push(arg["-fno-builtin-".len()..].to_owned());
742 }
743 _ if arg.starts_with("-fgnuc-version=") => {
744 let v = &arg["-fgnuc-version=".len()..];
745 opts.gnuc = v.parse().map_err(err)?;
746 }
747 // spec/13-gnu-compat.md section 13.3 promises this flag an error that says why rather
748 // than the unknown option one, because a build reaching for it is asking for a feature
749 // and deserves to be told it is not coming rather than told the spelling is wrong.
750 // The negative form is what this compiler does anyway, so it is taken and dropped.
751 "-fnested-functions" => {
752 return Err(err(
753 "nested functions are not supported: a call to one goes through a trampoline \
754 written on the stack, which no target that enforces an unexecutable stack \
755 allows",
756 ));
757 }
758 "-fno-nested-functions" => {}
759 // Which of the two links the output is for, which is a real difference and not a
760 // description of what happens anyway. Everything here is position independent either
761 // way, and what these decide is whether a name may be one another object defines or
762 // replaces, because a link that produces an executable puts every name in the same
763 // program and a link that produces a shared library does not.
764 //
765 // It matters that they are accepted at all, whatever they then do. Every autoconf and
766 // cmake build puts `-fPIC` on the compile line, so a compiler that rejects it cannot
767 // be the `CC` of a project that has a configure script, whatever else it can do. That
768 // is how this was found: building SQLite's test fixture stopped on it.
769 "-fPIC" | "-fpic" => opts.pic = Pic::Library,
770 // Not a synonym of the pair above, which is what they were treated as until #756. The
771 // library is the expensive answer and gcc makes it the one that has to be asked for,
772 // so this is also what nothing at all means.
773 "-fPIE" | "-fpie" => opts.pic = Pic::Executable,
774 // A different question from the pair above, and the one every distribution build of a
775 // shared library answers. `-fPIC` decides how an address is reached, and this decides
776 // whether the optimizer may believe a body it can see, because an exported name is one
777 // the dynamic linker may find another definition of first. On by default, which is
778 // gcc's arrangement and is the honest answer, and off is a promise the build makes and
779 // nothing checks.
780 "-fsemantic-interposition" => opts.interposition = true,
781 "-fno-semantic-interposition" => opts.interposition = false,
782 // Two requests rather than one, and the same table answers both, so what decides is
783 // whether either of them is standing. gcc arranges it the same way: the asynchronous
784 // one is the default here and it implies the other, and a line that asks for a table
785 // and against an asynchronous one gets a table.
786 "-fasynchronous-unwind-tables" => opts.async_unwind_tables = true,
787 "-fno-asynchronous-unwind-tables" => opts.async_unwind_tables = false,
788 "-funwind-tables" => opts.unwind_tables = true,
789 "-fno-unwind-tables" => opts.unwind_tables = false,
790 // The other direction is a request, not a description, and it is one this compiler
791 // cannot grant, so it gets the treatment section 13.3 asks for rather than the unknown
792 // option error. Answering it by carrying on would be answering a different question:
793 // the code would still be position independent, which is correct everywhere an
794 // ordinary program runs and is wrong in a kernel, where the flag is written precisely
795 // because there is no loader to fill a global offset table in.
796 "-fno-pic" | "-fno-pie" => {
797 return Err(err(
798 "position dependent code is not supported: an address that may be in another \
799 object is loaded out of the global offset table, and nothing here emits the \
800 absolute form this asks for. Use -no-pie if what you meant was how to link",
801 ));
802 }
803 // A section per function and a section per variable, which is what makes
804 // `--gc-sections` able to drop anything: a linker can leave out a section nothing
805 // reaches and cannot leave out half of one. Both directions are taken, and the off
806 // one is the default rather than a refusal, since a build that writes it is asking
807 // for what happens anyway.
808 "-ffunction-sections" => opts.function_sections = true,
809 "-fno-function-sections" => opts.function_sections = false,
810 "-fdata-sections" => opts.data_sections = true,
811 "-fno-data-sections" => opts.data_sections = false,
812 // Another description of what this compiler does. A file scope declaration with no
813 // initializer is written into `.bss` as an ordinary defined symbol, not offered to the
814 // linker as a common one for it to merge, which is what `-fno-common` asks for and what
815 // gcc has done by default since 10. Nothing in the front end produces `Linkage::Common`
816 // at all.
817 "-fno-common" => {}
818 // What overflows rather than being undefined. Every one of these takes something away
819 // from the optimizer rather than asking it to do anything, which is why the negative
820 // spellings are the interesting ones and the positive spellings are the default.
821 //
822 // `-fno-strict-overflow` is both of the others, which is gcc's own reading of it: its
823 // help text for `-fstrict-overflow` says "negated as -fwrapv -fwrapv-pointer". So it is
824 // written here as the pair rather than kept as a third thing to test everywhere.
825 //
826 // `-ftrapv` is the exception and is the one that asks for something. It is the other
827 // answer to the question `-fwrapv` answers, so the two cannot both hold and each clears
828 // the other, which makes the last one on the command line the one that counts. That is
829 // gcc 16's behaviour and was measured rather than read: `-ftrapv -fwrapv` emits no
830 // checked calls and `-fwrapv -ftrapv` emits them. The positive spelling of the pointer
831 // question is left alone by both, because neither has anything to say about it.
832 "-fwrapv" => {
833 opts.wrapping.signed = true;
834 opts.wrapping.trap = false;
835 }
836 "-fno-wrapv" => opts.wrapping.signed = false,
837 "-fwrapv-pointer" => opts.wrapping.pointer = true,
838 "-fno-wrapv-pointer" => opts.wrapping.pointer = false,
839 "-fno-strict-overflow" => opts.wrapping = Wrapping::ALL,
840 // Which does not clear the checked one, because gcc does not: `-ftrapv
841 // -fstrict-overflow` still emits the calls. It says what is assumed and not what
842 // happens.
843 "-fstrict-overflow" => {
844 opts.wrapping.signed = false;
845 opts.wrapping.pointer = false;
846 }
847 "-ftrapv" => {
848 opts.wrapping.trap = true;
849 opts.wrapping.signed = false;
850 }
851 "-fno-trapv" => opts.wrapping.trap = false,
852 // The two flags that say what a plain `char` is, which is one question with two
853 // spellings each: gcc reads `-fno-signed-char` as `-funsigned-char` and
854 // `-fno-unsigned-char` as `-fsigned-char`, so there are four ways to write two
855 // answers and the last one written wins. Nothing is set until one of them is given,
856 // because the target's own ABI is the answer otherwise and it is not the same answer
857 // everywhere: x86-64 and Apple's arm64 are signed, Linux's arm64 is not.
858 "-fsigned-char" | "-fno-unsigned-char" => opts.char_signed = Some(true),
859 "-funsigned-char" | "-fno-signed-char" => opts.char_signed = Some(false),
860 // And the size of an enumeration, which is the other thing in this group that changes
861 // the ABI rather than the code.
862 "-fshort-enums" => opts.short_enums = true,
863 "-fno-short-enums" => opts.short_enums = false,
864 // And the request, which is the one that cannot be granted. It is a real difference and
865 // not a preference: two files each writing `int g;` link under `-fcommon` and are a
866 // duplicate definition without it, which is the whole reason the flag survives.
867 "-fcommon" => {
868 return Err(err(
869 "a tentative definition is written into .bss as its own symbol here, and \
870 nothing emits the common symbol this asks the linker to merge. Give the \
871 variable a definition in one file and declare it extern in the others",
872 ));
873 }
874 // Both directions of this one are recorded, and what they decide is whether lowering
875 // names the type each access goes through. Turning it off is the front end leaving the
876 // name off rather than a pass being told to ignore one it can see, which is one
877 // condition in one place, and it is the reading that survives link time optimization:
878 // a unit built with the flag off keeps its own answer when its bodies end up in a
879 // module beside bodies that were not.
880 //
881 // Nothing in the pipeline reads those names yet. Layer 3 of the alias analysis does
882 // and is tested, and no pass at any level asks the alias analysis anything today, so
883 // no program compiles differently for having passed this. The flag is wired anyway,
884 // because the change that makes a pass ask is not the change anybody will remember to
885 // wire it in, and a flag that is taken and dropped once the names mean something is
886 // the miscompilation `spec/04-driver-and-cli.md` section 4.1 warns about in as many
887 // words.
888 "-fstrict-aliasing" => opts.strict_aliasing = true,
889 "-fno-strict-aliasing" => opts.strict_aliasing = false,
890 // The same shape of answer for the same reason, and the flag the kernel writes beside
891 // the one above it.
892 //
893 // Nothing here concludes that a pointer is not null from the fact that it was
894 // dereferenced. There is no such conclusion to draw from, because no pass records one:
895 // a load says where it read and nothing else, and a comparison against null is an
896 // ordinary comparison of two values the optimizer has no fact about. So a function
897 // that reads through a pointer and then tests it keeps the test, which is what the
898 // kernel wants and what `-fno-delete-null-pointer-checks` asks for, and what gcc has
899 // to be asked for because it draws the conclusion by default.
900 //
901 // `-fdelete-null-pointer-checks` is the request to draw it, and it goes the way
902 // `-fstrict-aliasing` does: assuming less than was asked for costs speed and not
903 // correctness, and `-O2` implies it, so refusing it would stop builds for nothing.
904 "-fdelete-null-pointer-checks" | "-fno-delete-null-pointer-checks" => {}
905 // The floating point group, which goes the same way and for the same reason, and which
906 // is worth writing out because the reason is easy to get backwards.
907 //
908 // Each of these has a restrictive spelling and a permissive one. The restrictive ones,
909 // `-frounding-math` and `-ftrapping-math`, say that the rounding mode may have been
910 // changed and that an exception raised by an operation may be looked at, so an
911 // arithmetic the compiler folds at compile time is an arithmetic whose rounding and
912 // whose exception the program does not get. Nothing here folds any floating point
913 // arithmetic in a function body: `0.1 + 0.2` is an `fadd` and `1.0 / 0.0` is a divide
914 // that runs, at every level. So both of those describe what already happens.
915 //
916 // The permissive ones are the other half, and they are licences rather than requests
917 // for an answer. `-fno-rounding-math` says the rounding mode is the default one and
918 // `-fno-trapping-math` says nothing looks at the exceptions, which together are
919 // permission to fold. Not folding is the conservative side of that permission and is
920 // what a program is entitled to whichever was written, so `-fno-rounding-math` costs
921 // speed and not correctness, which is the test section 4.1 puts a licence through.
922 "-frounding-math" | "-fno-rounding-math" => {}
923 // `-fno-trapping-math` is the one of the four that is kept, because there is one
924 // conversion this compiler does not fold and gcc folds under it, and the two answers
925 // differ. Converting a constant floating value to an integer type it does not fit in
926 // is undefined behaviour rather than a value: left to the hardware it is one
927 // instruction and the answer is the integer indefinite value, and folded it is the
928 // nearest end of the integer's range. Both compilers leave it to the instruction by
929 // default and gcc folds it under this flag, so a program built with it and compiled
930 // without it gets a different number rather than a slower one. `-ftrapping-math` is
931 // gcc's default, so a build spelling it out is asking for what it already has.
932 "-ftrapping-math" => opts.trapping_math = true,
933 "-fno-trapping-math" => opts.trapping_math = false,
934 // About temporary files rather than about code. There is nothing between the phases of
935 // one compilation here to write to a file in the first place.
936 "-pipe" => {}
937 // Nothing here writes colour, so all of these are the same answer, and it is the answer
938 // that costs nothing: the diagnostics come out plain either way and no build depends on
939 // an escape sequence being there. Taken rather than refused because cmake writes
940 // `-fdiagnostics-color=always` on every compile line when the generator is ninja, which
941 // makes this the second most common flag after `-fPIC` to stop a build over a question
942 // about how the text looks.
943 "-fdiagnostics-color" | "-fno-diagnostics-color" => {}
944 _ if arg.starts_with("-fdiagnostics-color=") => {}
945 // The link flags. None of them changes the compilation, which is why they are
946 // collected apart from `opts` and why `-lm` on a `-c` line is a note rather than an
947 // error: it is a thing said to a linker that is not going to run.
948 "-static" => link.is_static = true,
949 "-shared" => link.shared = true,
950 "-pie" => link.pie = Some(true),
951 "-no-pie" | "-nopie" => link.pie = Some(false),
952 "-nostdlib" => link.no_stdlib = true,
953 "-nostartfiles" => link.no_startfiles = true,
954 "-nodefaultlibs" => link.no_defaultlibs = true,
955 "-fno-builtins-lib" => link.no_builtins_lib = true,
956 "-fbuiltins-lib" => link.no_builtins_lib = false,
957 "-rdynamic" | "-export-dynamic" => link.export_dynamic = true,
958 "-s" => link.strip = true,
959 "-Xlinker" => {
960 let next = args.get(i).ok_or_else(|| err("-Xlinker requires an argument"))?;
961 i += 1;
962 link.passthrough.push(next.clone());
963 }
964 _ if arg.starts_with("-Wl,") => {
965 // Commas separate arguments rather than being part of one, which is what makes
966 // `-Wl,-rpath,/opt/lib` two words to the linker and one word here.
967 link.passthrough.extend(arg["-Wl,".len()..].split(',').map(str::to_owned));
968 }
969 _ if arg.starts_with("-fuse-ld=") => {
970 link.use_ld = Some(arg["-fuse-ld=".len()..].to_owned());
971 }
972 _ if arg.starts_with("-l") && arg.len() > 2 => {
973 inputs.push(Input::library(&arg[2..]));
974 }
975 "-l" => {
976 let next = args.get(i).ok_or_else(|| err("-l requires an argument"))?;
977 i += 1;
978 inputs.push(Input::library(next));
979 }
980 _ if arg.starts_with("-L") => {
981 link.search.push(PathBuf::from(joined_or_next(arg, 2, args, &mut i)?));
982 }
983 _ if arg.starts_with("-B") => {
984 link.prefixes.push(PathBuf::from(joined_or_next(arg, 2, args, &mut i)?));
985 }
986 _ if arg.starts_with("-j") => {
987 jobs = Jobs::parse(&arg[2..]).map_err(err)?;
988 }
989 _ if arg.starts_with("--sysroot=") => {
990 sysroot = Some(PathBuf::from(&arg["--sysroot=".len()..]));
991 }
992 _ if arg.starts_with("--target=") => {
993 let t = &arg["--target=".len()..];
994 opts.target = t.parse().map_err(|e| err(format!("{e}")))?;
995 // The same string again, as the model that has room for a libc version. A spelling
996 // the three field parser took and this one does not is not an error, because the
997 // one that decides what is compiled has already accepted it and the only thing
998 // lost is a version nobody asked for.
999 pinned = t.parse().ok();
1000 }
1001 _ if arg.starts_with("--emit=") => {
1002 let k = &arg["--emit=".len()..];
1003 opts.emit = k
1004 .parse()
1005 .map_err(|()| err(format!("unknown --emit kind `{k}`, see --help")))?;
1006 }
1007 // A bare `-O` is `-O1`, which is what GCC has and what a hand written makefile tends
1008 // to write. `-Og` is GCC's level for a build somebody is going to step through, and
1009 // it is `-O1` with the transformations that move code around left out; this compiler
1010 // has no such level yet, so it is the nearest one and `--print-pipeline` says what
1011 // that came to rather than the flag pretending otherwise.
1012 "-O" | "-Og" => opts.opt_level = rucc_session::OptLevel::O1,
1013 // The union of `-O3` and `-ffast-math`, and the second half of that changes what
1014 // floating point arithmetic means. Refused rather than taken as `-O3`, because a
1015 // build that asks for fast math and is quietly given ordinary arithmetic gets a
1016 // slower program than it asked for and a build that is given fast math it did not
1017 // ask for gets a wrong one.
1018 "-Ofast" => {
1019 return Err(err(
1020 "-Ofast is -O3 with fast math, and fast math is not implemented, see \
1021 spec/04-driver-and-cli.md section 4.6",
1022 ));
1023 }
1024 _ if arg.starts_with("-O") => {
1025 opts.opt_level = arg[2..]
1026 .parse()
1027 .map_err(|()| err(format!("unknown optimization level `{arg}`")))?;
1028 }
1029 // How far a multiply and an addition may be fused into one rounding. Before the
1030 // optimizer's `-f` family below for the reason the ones under it are, and kept rather
1031 // than dropped because it is the one flag in its group this compiler could act on: it
1032 // rides into the IR as an attribute on each function with a body, so the day the code
1033 // generator forms an `fma` it already knows which functions were given permission.
1034 // Nothing forms one today, under any value of this and under any `-march=`.
1035 _ if arg.starts_with("-ffp-contract=") => {
1036 let how = &arg["-ffp-contract=".len()..];
1037 opts.fp_contract = how.parse().map_err(|()| {
1038 err(format!("`{how}` is not a contraction, which is fast, on or off"))
1039 })?;
1040 }
1041 // How much of an expression may be computed wider than it was written. The values are
1042 // gcc's and so is the refusal of anything else, and none of the three changes anything
1043 // here: an operation is computed in the type C says it is on every target this compiler
1044 // has a back end for, so `__FLT_EVAL_METHOD__` is 0 and `standard` is already what
1045 // happens. `fast` and `16` are permission to be wider, which is a licence this takes
1046 // and does not use, the same way the two above are. The flag is worth taking because
1047 // glibc's headers and a good deal of configure output write it, and because the answer
1048 // it asks about is one this compiler can state rather than guess at: there is no x87
1049 // target here, which is the machine the whole question was invented for.
1050 // Whether a local and a spilled value that are never both wanted may be the same bytes
1051 // of the frame. gcc's three values, and two of them mean the same thing here: what rucc
1052 // shares is a local whose address provably never leaves the function, which is narrower
1053 // than `named_vars` and narrower still than `all`, so both of them get it. `none` is
1054 // the one that changes anything, and it is the flag a program that reads a local
1055 // through a pointer it kept past the end of the block writes.
1056 _ if arg.starts_with("-fstack-reuse=") => {
1057 let how = &arg["-fstack-reuse=".len()..];
1058 opts.stack_reuse = match how {
1059 "all" | "named_vars" => Some(true),
1060 "none" => Some(false),
1061 _ => {
1062 return Err(err(format!(
1063 "`{how}` is not a stack reuse, which is all, named_vars or none"
1064 )));
1065 }
1066 };
1067 }
1068 _ if arg.starts_with("-fexcess-precision=") => {
1069 let how = &arg["-fexcess-precision=".len()..];
1070 if !matches!(how, "16" | "fast" | "standard") {
1071 return Err(err(format!(
1072 "`{how}` is not an excess precision, which is 16, fast or standard"
1073 )));
1074 }
1075 }
1076 // Which front of a path is rewritten before it reaches the output, which is how a
1077 // build gets the same bytes out of two different directories. The four spellings are
1078 // one flag each into three lists, and `-ffile-prefix-map=` is the three of them at
1079 // once. Only the macro list does anything today, because `__FILE__` is the only place
1080 // a path reaches the output: there is no DWARF and no profile data yet, so the other
1081 // two are recorded for the work that will read them. The argument splits at the last
1082 // `=` rather than the first, which is gcc's rule and is what lets a directory with an
1083 // `=` in its name be the old half.
1084 _ if arg.starts_with("-fmacro-prefix-map=") => {
1085 let (old, new) = rewrite(arg, "-fmacro-prefix-map=")?;
1086 opts.prefix_map.macros.push(old, new);
1087 }
1088 _ if arg.starts_with("-fdebug-prefix-map=") => {
1089 let (old, new) = rewrite(arg, "-fdebug-prefix-map=")?;
1090 opts.prefix_map.debug.push(old, new);
1091 }
1092 _ if arg.starts_with("-fprofile-prefix-map=") => {
1093 let (old, new) = rewrite(arg, "-fprofile-prefix-map=")?;
1094 opts.prefix_map.profile.push(old, new);
1095 }
1096 _ if arg.starts_with("-ffile-prefix-map=") => {
1097 let (old, new) = rewrite(arg, "-ffile-prefix-map=")?;
1098 opts.prefix_map.macros.push(old, new);
1099 opts.prefix_map.debug.push(old, new);
1100 opts.prefix_map.profile.push(old, new);
1101 }
1102 // A whole optimization rather than a flag, and the family is taken rather than
1103 // refused because of what ignoring it does. There is none of it here yet, so a build
1104 // that asks for it gets a program that is correct and slower than it could have been,
1105 // which is what section 4.1 means by a hint about speed and what every compilation at
1106 // `-O0` already is. The objects settle the rest of the argument: gcc's `-flto` object
1107 // holds the bytecode and no machine code at all, and every object here holds the code,
1108 // which is exactly what `-ffat-lto-objects` asks gcc for. So a build passing `-flto`
1109 // to this compiler gets objects that are more usable than the ones it asked for rather
1110 // than different ones. Every value is still checked against gcc's, because somebody
1111 // who wrote `-flto=thin` meant clang and had better hear about it here.
1112 "-flto" => opts.lto.requested = true,
1113 "-fno-lto" => opts.lto.requested = false,
1114 _ if arg.starts_with("-flto=") => {
1115 let how = &arg["-flto=".len()..];
1116 opts.lto.jobs = how.parse().map_err(|()| {
1117 err(format!(
1118 "`{how}` is not a number of link time jobs, which is auto, jobserver or a \
1119 count above zero"
1120 ))
1121 })?;
1122 opts.lto.requested = true;
1123 }
1124 _ if arg.starts_with("-flto-partition=") => {
1125 let how = &arg["-flto-partition=".len()..];
1126 opts.lto.partition = how.parse().map_err(|()| {
1127 err(format!(
1128 "`{how}` is not a partitioning model, which is balanced, 1to1, one, max \
1129 or none"
1130 ))
1131 })?;
1132 }
1133 _ if arg.starts_with("-flto-compression-level=") => {
1134 let how = &arg["-flto-compression-level=".len()..];
1135 let level =
1136 how.parse::<u8>().ok().filter(|level| *level <= 19).ok_or_else(|| {
1137 err(format!("`{how}` is not a compression level, 0 to 19"))
1138 })?;
1139 opts.lto.compression = Some(level);
1140 }
1141 // Whether the object keeps its machine code as well as the bytecode. It always does
1142 // here, so the first of these describes what happens and the second asks for an object
1143 // with less in it, which is a smaller file and not a different program, so both are
1144 // taken.
1145 "-ffat-lto-objects" | "-fno-fat-lto-objects" => {}
1146 // Whether the linker is handed a plugin that does the link time work. The design in
1147 // `spec/09-optimizer.md` has this driver doing that work itself and never loading a
1148 // plugin into anybody, so neither answer is a question it has to hold.
1149 "-fuse-linker-plugin" | "-fno-use-linker-plugin" => {}
1150 // Reading a profile back. Taken for the reason the family above it is: nothing here
1151 // reads one, so a build that asks gets the program it would have got anyway, and gcc
1152 // itself produces a byte for byte identical object from `-fprofile-use` when there are
1153 // no counts beside the file. The path is recorded for the pass that will read it. The
1154 // warning gcc prints when it looked and found nothing is deliberately not copied,
1155 // because nothing here looks, and a warning about a file that was never opened would
1156 // fire on the builds that have a perfectly good profile as well as on the ones that
1157 // do not.
1158 "-fprofile-use" => opts.profile_data.requested = true,
1159 "-fno-profile-use" => opts.profile_data.requested = false,
1160 _ if arg.starts_with("-fprofile-use=") => {
1161 opts.profile_data.path = Some(arg["-fprofile-use=".len()..].to_string());
1162 opts.profile_data.requested = true;
1163 }
1164 _ if arg.starts_with("-fprofile-dir=") => {
1165 opts.profile_data.dir = Some(arg["-fprofile-dir=".len()..].to_string());
1166 }
1167 "-fprofile-abs-path" => opts.profile_data.absolute = true,
1168 "-fno-profile-abs-path" => opts.profile_data.absolute = false,
1169 "-fprofile-correction" => opts.profile_data.correction = true,
1170 "-fno-profile-correction" => opts.profile_data.correction = false,
1171 "-fprofile-partial-training" => opts.profile_data.partial_training = true,
1172 "-fno-profile-partial-training" => opts.profile_data.partial_training = false,
1173 // Writing the counts rather than reading them, which is refused rather than taken and
1174 // is the same line `-gsplit-dwarf` falls on the far side of. Ignoring these means a
1175 // file a build declared as an output never appears: the instrumented program writes a
1176 // `.gcda` as it exits and `-ftest-coverage` writes a `.gcno` beside the object, and a
1177 // two stage build that got neither would go on to optimize against no counts at all
1178 // and report coverage of nothing, with nothing along the way saying so. The objects
1179 // say the rest: gcc's `-fprofile-generate` object holds 375 bytes of code where a
1180 // plain one holds 71, and 296 bytes of counters that a plain one does not have, so
1181 // this is a flag that changes the output rather than a hint about speed.
1182 "-fprofile-arcs"
1183 | "--coverage"
1184 | "-fcondition-coverage"
1185 | "-fpath-coverage"
1186 | "-fprofile-generate" => {
1187 return Err(err(format!(
1188 "{arg}: this compiler does not instrument for profiling, and a build that \
1189 expects the counts a run of the instrumented program writes would optimize \
1190 against nothing on its second pass, see spec/04-driver-and-cli.md"
1191 )));
1192 }
1193 _ if arg.starts_with("-fprofile-generate=") => {
1194 return Err(err(format!(
1195 "{arg}: this compiler does not instrument for profiling, and a build that \
1196 expects the counts a run of the instrumented program writes would optimize \
1197 against nothing on its second pass, see spec/04-driver-and-cli.md"
1198 )));
1199 }
1200 "-ftest-coverage" => {
1201 return Err(err(format!(
1202 "{arg}: this compiler writes no `.gcno` file beside the object, and a build \
1203 that expects one would wait for a file that never arrives, see \
1204 spec/04-driver-and-cli.md"
1205 )));
1206 }
1207 // The rest of the family describes instrumentation that is refused above, so what is
1208 // left to do with them is check them and drop them. They are checked because a
1209 // misspelling in a distribution's flags is worth finding here rather than on the day
1210 // the instrumentation lands, and dropped because there is nothing for an answer about
1211 // how a counter is written to be an answer about.
1212 _ if arg.starts_with("-fprofile-update=") => {
1213 let how = &arg["-fprofile-update=".len()..];
1214 if !matches!(how, "single" | "atomic" | "prefer-atomic") {
1215 return Err(err(format!(
1216 "`{how}` is not a profile update method, which is single, atomic or \
1217 prefer-atomic"
1218 )));
1219 }
1220 }
1221 _ if arg.starts_with("-fprofile-reproducible=") => {
1222 let how = &arg["-fprofile-reproducible=".len()..];
1223 if !matches!(how, "serial" | "parallel-runs" | "multithreaded") {
1224 return Err(err(format!(
1225 "`{how}` is not a profile reproducibility method, which is serial, \
1226 parallel-runs or multithreaded"
1227 )));
1228 }
1229 }
1230 "-fprofile-values" | "-fno-profile-values" | "-fprofile-info-section" => {}
1231 "-fno-test-coverage" | "-fno-profile-arcs" | "-fno-profile-generate" => {}
1232 _ if arg.starts_with("-fprofile-filter-files=")
1233 || arg.starts_with("-fprofile-exclude-files=")
1234 || arg.starts_with("-fprofile-note=") => {}
1235 // What every name gets when nothing in the source said, which the attribute in the
1236 // source overrides rather than the other way round. Before the optimizer's `-f`
1237 // family below for the reason the tier below it is.
1238 _ if arg.starts_with("-fvisibility=") => {
1239 let seen = &arg["-fvisibility=".len()..];
1240 opts.visibility = seen.parse().map_err(|()| {
1241 err(format!(
1242 "`{seen}` is not a visibility, which is default, hidden, internal or \
1243 protected"
1244 ))
1245 })?;
1246 }
1247 // Which edges of a control flow transfer are checked. Before the optimizer's `-f`
1248 // family below for the reason the two above it are, and last of the three so that the
1249 // bare spelling and the negative one are matched exactly rather than by this.
1250 _ if arg.starts_with("-fcf-protection=") => {
1251 let edges = &arg["-fcf-protection=".len()..];
1252 opts.control = edges.parse().map_err(|()| {
1253 err(format!(
1254 "`{edges}` is not a control flow protection, which is full, branch, \
1255 return, none or check"
1256 ))
1257 })?;
1258 }
1259 // How much room every function opens with for something to be written over later.
1260 // Before the optimizer's `-f` family below for the reason the ones above it are.
1261 _ if arg.starts_with("-fpatchable-function-entry=") => {
1262 let room = &arg["-fpatchable-function-entry=".len()..];
1263 opts.patchable = room.parse().map_err(|()| {
1264 err(format!(
1265 "`{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"
1266 ))
1267 })?;
1268 }
1269 // The memory safety monitor, from section 15.4 of
1270 // `spec/safe-memory/15-integration.md`. Before the optimizer's `-f` family below,
1271 // because a pass that took the name `safety=detect` would otherwise be handed the
1272 // flag, and the tier is not a pass.
1273 _ if arg.starts_with("-fsafety=") => {
1274 let tier = &arg["-fsafety=".len()..];
1275 opts.safety = tier.parse().map_err(|()| {
1276 err(format!(
1277 "`{tier}` is not a safety tier, which is off, detect, enforce or kernel"
1278 ))
1279 })?;
1280 }
1281 // Whether padding participates, from section 9.3 of document 09. Spelled out rather
1282 // than folded into the tier because it is a departure somebody who has read that
1283 // section makes, and the two defaults it describes are a property of what is being
1284 // built rather than of how much checking is wanted.
1285 _ if arg.starts_with("-fsafety-init=") => {
1286 let mode = &arg["-fsafety-init=".len()..];
1287 opts.padding = mode.parse().map_err(|()| {
1288 err(format!("`{mode}` is not a padding mode, which is padding or nopadding"))
1289 })?;
1290 }
1291 // Row S4, from section 9.4 of document 09. A bare flag with no value, because the
1292 // strict form of that section needs a member id the front end does not name yet and
1293 // accepting the spelling for it would be accepting a promise this build cannot keep.
1294 // Before `-fno-` is looked at below, for the reason the tier is.
1295 "-fsafety-subobject" => opts.subobject = rucc_session::Subobject::Members,
1296 "-fno-safety-subobject" => opts.subobject = rucc_session::Subobject::Off,
1297 _ if arg.starts_with("-fsafety-subobject=") => {
1298 let form = &arg["-fsafety-subobject=".len()..];
1299 return Err(err(format!(
1300 "`{form}` is not a form of -fsafety-subobject. The flag takes no value, and \
1301 the strict form of section 9.4 is tamnd/rucc#967"
1302 )));
1303 }
1304 // Row Y8, from section 9.6 of document 09. A bare flag with no value, for the reason
1305 // the one above has none: there is one form of this check and a spelling that suggested
1306 // otherwise would be promising something. Before `-fno-` is looked at below, the same
1307 // way.
1308 "-fsafety-restrict" => opts.promise = rucc_session::Promise::Blocks,
1309 "-fno-safety-restrict" => opts.promise = rucc_session::Promise::Off,
1310 _ if arg.starts_with("-fsafety-restrict=") => {
1311 let form = &arg["-fsafety-restrict=".len()..];
1312 return Err(err(format!(
1313 "`{form}` is not a form of -fsafety-restrict. The flag takes no value."
1314 )));
1315 }
1316 // Section 9.5's races, which take a value because the section gives them three modes
1317 // and the difference between two of them is which classes get reported rather than how
1318 // much is recorded. `-fno-` is the same as `=off` and is spelled out here for the same
1319 // reason the two above spell theirs out.
1320 _ if arg.starts_with("-fsafety-races=") => {
1321 let mode = &arg["-fsafety-races=".len()..];
1322 opts.races = mode.parse().map_err(|()| {
1323 err(format!("`{mode}` is not a race mode, which is off, metadata or pointer"))
1324 })?;
1325 }
1326 "-fno-safety-races" => opts.races = rucc_session::Races::Off,
1327 // The sanitizers of document 12, which are checks at run time rather than a way of
1328 // generating the same program. Each name is held to gcc 16's list, and what is still
1329 // asked for by the end of the line is answered after the loop, so that a command line
1330 // which turns one on and then off again is a command line that asked for nothing.
1331 //
1332 // Before the optimizer's `-f` family below, for the reason the tier above it is.
1333 _ if arg.starts_with("-fsanitize=") => {
1334 for one in arg["-fsanitize=".len()..].split(',') {
1335 if one == "all" {
1336 // gcc takes `all` only in the negative, because turning every check on at
1337 // once includes checks that contradict each other.
1338 return Err(err(
1339 "`-fsanitize=all` is not a gcc option, only `-fno-sanitize=all` is",
1340 ));
1341 }
1342 if !SANITIZERS.contains(&one) {
1343 return Err(err(format!(
1344 "`{one}` is not a sanitizer, see spec/04-driver-and-cli.md section 4.7"
1345 )));
1346 }
1347 if !sanitizers.contains(&one) {
1348 sanitizers.push(one);
1349 }
1350 }
1351 }
1352 _ if arg.starts_with("-fno-sanitize=") => {
1353 for one in arg["-fno-sanitize=".len()..].split(',') {
1354 if one == "all" {
1355 sanitizers.clear();
1356 continue;
1357 }
1358 if !SANITIZERS.contains(&one) {
1359 return Err(err(format!(
1360 "`{one}` is not a sanitizer, see spec/04-driver-and-cli.md section 4.7"
1361 )));
1362 }
1363 sanitizers.retain(|asked| *asked != one);
1364 }
1365 }
1366 // What a check does when it fires, and where the records about the checked objects go.
1367 // Each of them is an answer about the sanitizers refused after the loop, so there is
1368 // nothing left for them to change here. The names are still held to the list, because
1369 // a misspelling in a build's flags is worth finding when the compiler reads it.
1370 _ if arg.starts_with("-fsanitize-recover=")
1371 || arg.starts_with("-fno-sanitize-recover=")
1372 || arg.starts_with("-fsanitize-trap=")
1373 || arg.starts_with("-fno-sanitize-trap=") =>
1374 {
1375 // The guard above matched on a spelling that has an `=` in it, so the tail is
1376 // whatever follows the first one.
1377 let how = arg.split_once('=').map_or("", |(_, rest)| rest);
1378 for one in how.split(',') {
1379 if one != "all" && !SANITIZERS.contains(&one) {
1380 return Err(err(format!(
1381 "`{one}` is not a sanitizer, see spec/04-driver-and-cli.md section 4.7"
1382 )));
1383 }
1384 }
1385 }
1386 "-fsanitize-undefined-trap-on-error"
1387 | "-fsanitize-address-use-after-scope"
1388 | "-fno-sanitize-address-use-after-scope" => {}
1389 _ if arg.starts_with("-fsanitize-sections=") => {}
1390 // Counting which edges a run reached, which is how a fuzzer knows an input was worth
1391 // keeping. Refused rather than dropped, because a fuzzer whose calls into
1392 // `__sanitizer_cov_*` were never generated runs blind and reports coverage of nothing,
1393 // and there is no point in the campaign where that announces itself.
1394 _ if arg.starts_with("-fsanitize-coverage=") => {
1395 let how = &arg["-fsanitize-coverage=".len()..];
1396 for one in how.split(',') {
1397 if !matches!(one, "trace-pc" | "trace-cmp") {
1398 return Err(err(format!(
1399 "`{one}` is not a coverage instrumentation, which is trace-pc or \
1400 trace-cmp"
1401 )));
1402 }
1403 }
1404 return Err(err(format!(
1405 "{arg}: this compiler generates no coverage callbacks, and a fuzzer built \
1406 with it would run without any feedback at all, see \
1407 spec/04-driver-and-cli.md section 4.7"
1408 )));
1409 }
1410 // The optimizer's own flags, from section 9.10 of `spec/09-optimizer.md`. These come
1411 // after every `-f` the rest of the compiler answers to, so a pass can never take a
1412 // name that already means something else on the command line.
1413 _ if arg.starts_with("-fpass-fuel=") => {
1414 let (name, count) = arg["-fpass-fuel=".len()..]
1415 .split_once('=')
1416 .ok_or_else(|| err("-fpass-fuel= is spelled <pass>=<count>"))?;
1417 if rucc_opt::pass::find(name).is_none() {
1418 return Err(err(format!(
1419 "`{name}` is not a pass this compiler has, see --print-pipeline"
1420 )));
1421 }
1422 let count: u32 = count
1423 .parse()
1424 .map_err(|_| err(format!("`{count}` is not a number of transformations")))?;
1425 opts.pass_fuel.push((name.to_owned(), count));
1426 }
1427 _ if arg.starts_with("-fpass-fuel-global=") => {
1428 let count = &arg["-fpass-fuel-global=".len()..];
1429 let count: u32 = count
1430 .parse()
1431 .map_err(|_| err(format!("`{count}` is not a number of transformations")))?;
1432 opts.pass_fuel_global = Some(count);
1433 }
1434 // Everything from `-fopt-info` to the end of the argument, which is optional
1435 // keywords joined by hyphens and an optional `=<file>`. Checked here rather than
1436 // where the remarks are printed, because by then the compilation somebody wanted
1437 // to hear about is over.
1438 _ if arg == "-fopt-info"
1439 || arg.starts_with("-fopt-info=")
1440 || arg.starts_with("-fopt-info-") =>
1441 {
1442 let rest = &arg["-fopt-info".len()..];
1443 let (kinds, file) = match rest.split_once('=') {
1444 Some((kinds, file)) => (kinds, Some(file)),
1445 None => (rest, None),
1446 };
1447 let kinds = kinds.strip_prefix('-').unwrap_or(kinds);
1448 rucc_opt::Wants::none().add(kinds).map_err(err)?;
1449 opts.opt_info.push(kinds.to_owned());
1450 if let Some(file) = file {
1451 if file.is_empty() {
1452 return Err(err("-fopt-info= was given no file to write to"));
1453 }
1454 opts.opt_info_file = Some(file.to_owned());
1455 }
1456 }
1457 _ if arg.starts_with("-fdump-ir=") => {
1458 // Checked here rather than where the dumps are taken, because the compilation
1459 // that would have been dumped is over by then.
1460 let spec = &arg["-fdump-ir=".len()..];
1461 rucc_opt::Dumps::default().add(spec).map_err(err)?;
1462 opts.dump_ir.push(spec.to_owned());
1463 }
1464 // Before the bare `-f<pass>` below, because a pass called `enable-something` would
1465 // otherwise take the flag away from the gate. Checked here rather than where the
1466 // pipeline reads it, for the reason that applies to all of these: a misspelled pass
1467 // name that quietly gated nothing looks exactly like a pass that is not the guilty
1468 // one, and a bisection would carry on past the thing it was looking for.
1469 _ if arg.starts_with("-fdisable-") || arg.starts_with("-fenable-") => {
1470 let on = arg.starts_with("-fenable-");
1471 let spec = &arg[if on { "-fenable-".len() } else { "-fdisable-".len() }..];
1472 rucc_opt::Gates::default().add(on, spec).map_err(err)?;
1473 opts.pass_gates.push((on, spec.to_owned()));
1474 }
1475 // gcc's spelling for a pass this compiler has under a shorter name. It goes above the
1476 // two arms below rather than into the pile of gcc pass names further down, because the
1477 // pass is here: dropping the flag would leave a build that asked for unrolling without
1478 // it, and refusing it stops the build outright, which is what libtommath's makefile
1479 // ran into. `-funroll-all-loops` is deliberately not in here: gcc's is the one that
1480 // unrolls without a trip count, which is a different and usually worse thing.
1481 "-funroll-loops" => opts.passes.push(("unroll".to_owned(), true)),
1482 "-fno-unroll-loops" => opts.passes.push(("unroll".to_owned(), false)),
1483 _ if arg.strip_prefix("-fno-").is_some_and(|n| rucc_opt::pass::find(n).is_some()) => {
1484 opts.passes.push((arg["-fno-".len()..].to_owned(), false));
1485 }
1486 _ if arg.strip_prefix("-f").is_some_and(|n| rucc_opt::pass::find(n).is_some()) => {
1487 opts.passes.push((arg["-f".len()..].to_owned(), true));
1488 }
1489 // The flags that name a pass of gcc's own. They arrive from the torture suite, where a
1490 // program reduced from a miscompilation usually names the pass that miscompiled it on
1491 // its `dg-options` line, and they arrive from hand written build files for the same
1492 // reason. Section 4.1 sorts a flag by what the output would be without it, and by that
1493 // rule these are one pile: a flag that turns one of gcc's passes on or off is asking
1494 // for a compiler that does not exist here, and the program it is attached to is a
1495 // correctness test that passes either way. Turning on a pass we do not have costs
1496 // speed, turning off a pass we do not have costs nothing, and neither changes what the
1497 // program computes.
1498 //
1499 // rucc's own pass names are matched above this, so `-fno-dce` turns off the dce this
1500 // compiler has rather than landing here, and the day one of these names becomes a pass
1501 // here it stops being taken and dropped without anybody editing this list.
1502 //
1503 // Two of them are prefixes rather than names, which is the one place this file takes a
1504 // family instead of a flag. gcc files its gimple passes under `-ftree-` and its
1505 // interprocedural passes under `-fipa-`, both namespaces are pass selection and
1506 // nothing else, and there is no member of either that changes the meaning of a program
1507 // that was already correct. The rest are written out one at a time, because they live
1508 // in the flat `-f` namespace where the neighbours do change meanings.
1509 _ if arg.starts_with("-ftree-") || arg.starts_with("-fno-tree-") => {}
1510 _ if arg.starts_with("-fipa-") || arg.starts_with("-fno-ipa-") => {}
1511 "-fexpensive-optimizations" | "-fno-expensive-optimizations" => {}
1512 "-fmodulo-sched" | "-fno-modulo-sched" => {}
1513 "-fvect-cost-model" | "-fno-vect-cost-model" => {}
1514 _ if arg.starts_with("-fvect-cost-model=") || arg.starts_with("-fsimd-cost-model=") => {
1515 }
1516 "-fearly-inlining" | "-fno-early-inlining" => {}
1517 "-finline"
1518 | "-fno-inline"
1519 | "-finline-functions"
1520 | "-fno-inline-functions"
1521 | "-finline-small-functions"
1522 | "-fno-inline-small-functions"
1523 | "-finline-functions-called-once"
1524 | "-fno-inline-functions-called-once" => {}
1525 "-foptimize-strlen" | "-fno-optimize-strlen" => {}
1526 "-fira-share-spill-slots" | "-fno-ira-share-spill-slots" => {}
1527 // The charset flags are not in that pile, because an encoding is a statement about
1528 // what the bytes of the source mean rather than about how fast the output is. The
1529 // preprocessor reads UTF-8 and has no converter, so the one name that describes what
1530 // already happens is taken and every other name is refused. Spelled without regard to
1531 // case and with both of the spellings iconv answers to, since a build writes whichever
1532 // one its author typed.
1533 _ if arg.starts_with("-finput-charset=") => {
1534 let name = &arg["-finput-charset=".len()..];
1535 if !name.eq_ignore_ascii_case("utf-8") && !name.eq_ignore_ascii_case("utf8") {
1536 return Err(err(format!(
1537 "-finput-charset={name}: the preprocessor reads UTF-8 and has no \
1538 converter, so a file in another encoding would be read as though it were \
1539 UTF-8 rather than converted",
1540 )));
1541 }
1542 }
1543 // The three that come in on the same `dg-options` lines and are the other half of
1544 // section 4.1's rule, because each of them changes what the program does and not how
1545 // fast it does it. The negative form of each is what this compiler does anyway, so it
1546 // is taken and dropped, which is the shape `-fnested-functions` has above.
1547 "-ffast-math" => {
1548 return Err(err(
1549 "-ffast-math is a licence to answer a floating point arithmetic differently \
1550 from the way the source wrote it, and it is not one flag: it defines \
1551 __FAST_MATH__, which a library header reads, and gcc links a startup file \
1552 that puts the hardware in flush to zero mode for the whole process. Taking it \
1553 and dropping it would change what other objects in the same program answer. \
1554 -ffp-contract= and -fexcess-precision= are the parts of it this compiler has",
1555 ));
1556 }
1557 "-fno-fast-math" => {}
1558 "-fnon-call-exceptions" => {
1559 return Err(err(
1560 "-fnon-call-exceptions is a promise that an instruction which is not a call \
1561 can raise an exception the unwinder finds a handler for, and nothing here \
1562 produces a landing pad for a trapping instruction. A program built without it \
1563 would unwind past the handler it wrote",
1564 ));
1565 }
1566 "-fno-non-call-exceptions" => {}
1567 "-finstrument-functions" => {
1568 return Err(err(
1569 "-finstrument-functions calls __cyg_profile_func_enter on entry to every \
1570 function and __cyg_profile_func_exit on the way out, and nothing here emits \
1571 either call. A program that asks for them usually counts them, so taking the \
1572 flag and dropping it would turn a program that fails loudly into one that \
1573 fails quietly",
1574 ));
1575 }
1576 "-fno-instrument-functions" => {}
1577 // The unstable options, spelled the way rustc spells them and carrying the same
1578 // promise, which is none: one of these may change or go away in any release. They are
1579 // measurements and debugging aids rather than things a build asks for, which is why
1580 // none of them is in the usage text and all of them are in section 4.11 of
1581 // `spec/04-driver-and-cli.md`.
1582 "-Zverify-each" => opts.verify_each = true,
1583 _ if arg.starts_with("-Zrule-coverage=") => {
1584 let file = &arg["-Zrule-coverage=".len()..];
1585 if file.is_empty() {
1586 return Err(err("-Zrule-coverage= needs a file to write to"));
1587 }
1588 opts.rule_coverage = Some(file.to_owned());
1589 }
1590 _ if arg.starts_with("-Zcycle-accurate-model=") => {
1591 let value = &arg["-Zcycle-accurate-model=".len()..];
1592 opts.cycle_accurate_model = match value {
1593 "yes" | "1" => Some(true),
1594 "no" | "0" => Some(false),
1595 _ => {
1596 return Err(err("-Zcycle-accurate-model= takes yes or no"));
1597 }
1598 };
1599 }
1600 _ if arg.starts_with("-Zlowering=") => {
1601 let file = &arg["-Zlowering=".len()..];
1602 if file.is_empty() {
1603 return Err(err("-Zlowering= needs a file to write to"));
1604 }
1605 opts.lowering_dump = Some(file.to_owned());
1606 }
1607 _ if arg.starts_with("-Zregister-pressure=") => {
1608 let file = &arg["-Zregister-pressure=".len()..];
1609 if file.is_empty() {
1610 return Err(err("-Zregister-pressure= needs a file to write to"));
1611 }
1612 opts.register_pressure = Some(file.to_owned());
1613 }
1614 _ if arg.starts_with("-Z") => {
1615 return Err(err(format!(
1616 "`{arg}` is not an unstable option this compiler has, see \
1617 spec/04-driver-and-cli.md section 4.11 for the ones it does"
1618 )));
1619 }
1620 // The word size, which is a statement about the target and is taken as one. A build
1621 // that says the size the target already has is saying nothing, and one that says the
1622 // other size is asking for a target this compiler does not have, which it is told
1623 // rather than being given the wrong one.
1624 "-m64" | "-m32" | "-mx32" => {
1625 let want: u32 = match arg {
1626 "-m64" => 64,
1627 _ => 32,
1628 };
1629 let have = rucc_target::TargetInfo::new(opts.target).pointer_width;
1630 if have != want {
1631 return Err(err(format!(
1632 "{arg} asks for a {want} bit target and {} is {have} bit, use \
1633 --target= to name the one you mean",
1634 opts.target
1635 )));
1636 }
1637 }
1638 // Which processor in the family to generate for. This compiler emits the base
1639 // instruction set of the architecture and nothing above it, so a program built with
1640 // any of these runs on the machine that was named; it is a program that could have
1641 // been faster rather than a program that is wrong, which is what makes these safe to
1642 // take and ignore where a flag that changed the meaning of the code would not be.
1643 _ if arg.starts_with("-march=")
1644 || arg.starts_with("-mtune=")
1645 || arg.starts_with("-mcpu=") => {}
1646 // The calling convention, which is not safe to ignore. Taken when it names the one
1647 // the target already uses and refused otherwise.
1648 _ if arg.starts_with("-mabi=") => {
1649 let want = &arg["-mabi=".len()..];
1650 let have = match opts.target.arch {
1651 rucc_target::Arch::X86_64 => "sysv",
1652 rucc_target::Arch::Aarch64 => "lp64",
1653 rucc_target::Arch::Riscv64 => "lp64d",
1654 };
1655 if want != have {
1656 return Err(err(format!(
1657 "{arg}: {} uses the {have} convention and this compiler has no other",
1658 opts.target
1659 )));
1660 }
1661 }
1662 // How far apart the pieces of the program may be. The small model is what we emit and
1663 // it is every hosted program's default; the kernel model is a different one and a
1664 // build that asks for it and does not get it links and then does not run.
1665 "-mcmodel=small" => {}
1666 _ if arg.starts_with("-mcmodel=") => {
1667 return Err(err(format!(
1668 "{arg}: this compiler emits the small code model and no other, see \
1669 spec/12-targets.md"
1670 )));
1671 }
1672 // GCC's own scripting language for how the driver builds a command line.
1673 // `spec/04-driver-and-cli.md` section 4.4 settles that we will not have it, so a
1674 // build reaching for it is told which flags do the same job.
1675 _ if arg.starts_with("-specs=") => {
1676 return Err(err(
1677 "-specs= is not supported: the parts of it builds rely on are -B, -L, \
1678 -nostdlib, -nostartfiles and -Wl,, see spec/04-driver-and-cli.md \
1679 section 4.4",
1680 ));
1681 }
1682 // Arguments meant for a separate assembler or preprocessor, which this compiler does
1683 // not have: both are inside it and neither reads a command line. Refused rather than
1684 // dropped, because every one of these says something about the output and a build
1685 // that asked for `-Wa,--noexecstack` and was silently given an executable stack got
1686 // the opposite of what it asked for.
1687 _ if arg.starts_with("-Wa,") || arg.starts_with("-Wp,") => {
1688 return Err(err(format!(
1689 "`{arg}` is an argument for a separate assembler or preprocessor, and both \
1690 are inside this compiler rather than programs it runs"
1691 )));
1692 }
1693 "-Xassembler" | "-Xpreprocessor" => {
1694 return Err(err(format!(
1695 "{arg} hands an argument to a separate assembler or preprocessor, and both \
1696 are inside this compiler rather than programs it runs"
1697 )));
1698 }
1699 // Everything else in the `-W` family. `spec/04-driver-and-cli.md` section 4.1 has
1700 // this one as a rule about build systems rather than about warnings: autoconf finds
1701 // out whether a warning flag exists by passing it and looking at the exit status, so
1702 // a compiler that refuses one it has not heard of fails a configure script written
1703 // for a GCC newer than itself. The names are not checked against a list because this
1704 // compiler has no warning groups for a list to be of, which #485 is about.
1705 _ if arg.starts_with("-W") => {}
1706 // Flags that name something this compiler does not do and would not do differently
1707 // if it did. `-fno-ident` is about a comment in the output that we do not write
1708 // either way, and the others are about a way of ordering the compilation that has
1709 // been GCC's only way for twenty years. Section 4.1 asks for the list to be short
1710 // and for adding to it to be deliberate, which is why it is written out here.
1711 "-fno-ident"
1712 | "-fident"
1713 | "-funit-at-a-time"
1714 | "-fno-unit-at-a-time"
1715 | "-shared-libgcc"
1716 | "-static-libgcc" => {}
1717 _ if arg.starts_with('-') && arg.len() > 1 => {
1718 // Silently ignoring an unknown flag is how a build ends up not doing what
1719 // its author asked. spec/13-gnu-compat.md section 13.4 makes this an error
1720 // for the flags that change code generation, and the safe default until the
1721 // flag table is populated is to reject everything we do not know.
1722 return Err(err(format!("unknown option `{arg}`")));
1723 }
1724 _ => inputs.push(Input { path: arg.to_owned(), forced, library: false }),
1725 }
1726 }
1727
1728 // The fetch, before anything that resolves a compilation, because `--fetch` does not describe
1729 // one. It is here rather than in the loop so that `--offline` can forbid it whichever order the
1730 // two were written in, and it is before the refusals below so that a command line asking for a
1731 // sysroot is not told about a sanitizer.
1732 if let Some(named) = fetch {
1733 return fetch_action(&named, offline, &inputs);
1734 }
1735
1736 // Last, so that it lands after every `-isystem` the command line gave. That is GCC's
1737 // order: a directory the user names outranks the compiler's own, and the compiler's own
1738 // outranks the library's. It is pushed after the loop rather than before it because
1739 // `SearchPath` appends within a group and the position is what the order is.
1740 // The same directory the headers were looked for under, because a sysroot is a statement
1741 // about a whole installation and not about half of one.
1742 // After the loop, because `-fno-sanitize=` can take back what an earlier flag asked for and a
1743 // command line that turns a check on and off again has asked for nothing. What is left is
1744 // refused rather than dropped, and it is the one place in this parser where the reason is not
1745 // that the output would differ. A sanitizer is a promise that the program is watched while it
1746 // runs, so a build that asks for one and is quietly given a program with no checks in it does
1747 // not get a slower program or a bigger file, it gets a test suite that passes for the wrong
1748 // reason. `-fsafety=` is the checking this compiler does have, and the message says so, because
1749 // somebody reaching for `-fsanitize=address` wants the nearest thing rather than a list of
1750 // options.
1751 if let Some(first) = sanitizers.first() {
1752 return Err(err(format!(
1753 "-fsanitize={first}: this compiler has no sanitizer instrumentation, and a build that \
1754 asked for one and got none would run its tests unchecked, see \
1755 spec/04-driver-and-cli.md section 4.7. `-fsafety=detect` is the memory checking this \
1756 compiler does have"
1757 )));
1758 }
1759 link.sysroot = sysroot.clone();
1760 // Where a sysroot for a target that is not this machine would be. Read once, here, rather than
1761 // inside the link line, because a link line that read the environment could only be tested on a
1762 // machine whose environment said the right thing, and the link line is the last thing that
1763 // touches a binary. `spec/cross-compile/13-distribution.md` section 13.2 owns the answer.
1764 link.cache = Some(cache::dir());
1765 // And the ten field spelling of the target, because the release on it decides two things the
1766 // three field one cannot say: whether a target that is this architecture is still a cross
1767 // compile, and which directory under the cache it is against. After the loop because the last
1768 // `--target=` on the command line is the one that counts.
1769 link.pinned = pinned;
1770 // After the loop rather than where `-pthread` was read, so that it lands after the objects
1771 // that refer to it. A static link takes the definitions it needs from a library when it
1772 // reaches it and not afterwards, so a library before the objects is a library that answers
1773 // nothing.
1774 if threads {
1775 inputs.push(Input::library("pthread"));
1776 }
1777 if let Some(query) = query {
1778 return Ok(Action::Print(answer(&query, &opts, &link)?));
1779 }
1780 // `-M` and `-MM` produce the rule and nothing else, so the run stops after phase 4 whatever
1781 // else the command line asked for. Read here rather than where the flag was, because a `-c`
1782 // written after it has to lose and the loop cannot know that until it has ended. The output
1783 // file is where the rule goes rather than where an object would have gone, and the last
1784 // phase being the preprocessor is what makes that true without a second rule for it.
1785 if opts.deps.instead_of_compiling {
1786 opts.emit = EmitKind::Preprocessed;
1787 }
1788 if !nostdinc {
1789 opts.search.push_system(runtime::DIR);
1790 // And the library's after ours, which is the other half of the same order. They go on
1791 // here rather than at the point `--target=` or `--sysroot=` was read because either
1792 // one changes the answer and the last word on both is the end of the loop.
1793 //
1794 // Which library's is the question `link::cross_sysroot` answers, and it is asked here so
1795 // that the headers and the libraries come from the same place. A target that is this
1796 // machine reads this machine's headers, and a target that is not reads the ones in the
1797 // sysroot for it rather than the ones next door.
1798 let cross = link::cross_sysroot(opts.target, &link);
1799 let kernel = link::cross_kernel(opts.target, &link);
1800 // And the version of those headers, which only the bundled tree has an answer for. A host
1801 // glibc and a tree the user named both define `__GLIBC_MINOR__` in their own `features.h`,
1802 // and a second definition with a different value is a warning on every file, so the
1803 // condition is the same one that chose the directories.
1804 if cross.is_some() {
1805 let target = pinned.unwrap_or_else(|| opts.target.tuple());
1806 opts.glibc_minor = rucc_sysroot::bundled_glibc_minor(target).map_err(|skew| {
1807 err(format!(
1808 "{skew}; pin a release the tree has, or name a tree that has that one \
1809 with --sysroot"
1810 ))
1811 })?;
1812 }
1813 let system =
1814 library::header_dirs(opts.target, sysroot.as_deref(), cross.as_ref(), kernel.as_ref());
1815 // The two licence walls of `spec/cross-compile/13-distribution.md` section 13.4, which are
1816 // the only way step 3 comes back with nothing on a hosted target. Section 8.6 asks for the
1817 // answer to name the licence and the lawful ways to get what is behind it, rather than
1818 // leaving a person with an `#include` that failed as though a directory had gone missing.
1819 //
1820 // It is left on the search path instead of refused here, because a program that includes
1821 // none of the library needs none of the SDK and section 8.6 is explicit that targeting the
1822 // platform has to keep working. So the reason waits until an include has actually failed,
1823 // which is the only moment it helps and the only moment it is true.
1824 //
1825 // The condition is that step 3 found nothing at all, so an `SDKROOT`, an `INCLUDE` or a mac
1826 // with Xcode on it all pass through untouched, and `-nostdinc` never reaches this block. A
1827 // `--sysroot` or `-isysroot` passes through as well, even when the tree it names turns out to
1828 // be empty or absent: somebody who wrote a path has already answered the question this
1829 // message asks, and answering it again over the top of a mistyped directory would hide the
1830 // mistake behind a licence notice.
1831 if system.is_empty() && sysroot.is_none() {
1832 let tuple = pinned.unwrap_or_else(|| opts.target.tuple());
1833 if let Some(wall) = rucc_sysroot::Wall::of(tuple) {
1834 opts.search.explain_missing_system(wall.no_headers(&tuple.to_canonical_string()));
1835 }
1836 }
1837 // And whether the tree somebody named is the release they asked for, which is the one
1838 // question left once the directories are settled and the only place both halves of it are
1839 // known. Only for a named tree, because that is the case where the release in the target
1840 // stops deciding anything, and `crate::glibc` is where the rest of the reasoning is.
1841 if sysroot.is_some() {
1842 notes.extend(glibc::skew(opts.target, pinned, &system));
1843 }
1844 for dir in system {
1845 opts.search.push_system(dir);
1846 }
1847 }
1848 // Once, here, rather than as each directory is pushed. A `-I` that names a system
1849 // directory has to lose to the system entry and the system entry is added last, so the
1850 // question cannot be answered until the whole path is known.
1851 opts.search.remove_duplicates();
1852
1853 // The target has to be resolved before the configuration is printed, so this check comes
1854 // after the loop rather than at the point `--print-config` was seen.
1855 if print_config {
1856 return Ok(Action::PrintConfig(Box::new(opts)));
1857 }
1858 if print_pipeline {
1859 return Ok(Action::PrintPipeline(Box::new(opts)));
1860 }
1861 let plan = Plan::new(&opts, &inputs, output.as_deref()).map_err(|e| err(e.message))?;
1862 if print_plan {
1863 return Ok(Action::PrintPlan {
1864 opts: Box::new(opts),
1865 plan: Box::new(plan),
1866 link: Box::new(link),
1867 });
1868 }
1869 Ok(Action::Compile {
1870 opts: Box::new(opts),
1871 plan: Box::new(plan),
1872 link: Box::new(link),
1873 jobs,
1874 verbose,
1875 notes,
1876 })
1877}
1878
1879/// What `--fetch <tuple>` asked for, or why it is not a thing that can be done.
1880///
1881/// The lookup happens here rather than at the point the bytes would move, so that a target this
1882/// release pins nothing for is a refusal from the parser and the only code that runs a downloader is
1883/// code that already knows what it is getting.
1884///
1885/// # Errors
1886///
1887/// [`CliError`] when `--offline` forbade it, when there are input files as well, when the tuple is
1888/// not a target this compiler knows, when its sysroot is behind one of section 13.4's licence walls,
1889/// and when this release pins no artifact for it.
1890fn fetch_action(named: &str, offline: bool, inputs: &[Input]) -> Result<Action, CliError> {
1891 // Not a precedence question. Section 13.2 says `--offline` forbids a fetch entirely, so a
1892 // command line that writes both has asked for two opposite things and the answer is to say so
1893 // rather than to pick one of them.
1894 if offline {
1895 return Err(err(
1896 "--fetch asks for a download and --offline forbids every download, so this command \
1897 line asks for two opposite things. Drop one of them: --offline is how a build says it \
1898 will not reach the network, and --fetch is the only thing in this compiler that does",
1899 ));
1900 }
1901 if let Some(first) = inputs.first() {
1902 return Err(err(format!(
1903 "--fetch gets a sysroot and compiles nothing, so `{}` on the same command line is an \
1904 input that nothing would read",
1905 first.path
1906 )));
1907 }
1908 let target: TargetTuple = named
1909 .parse()
1910 .map_err(|why| err(format!("--fetch {named}: {why}, so there is no sysroot to get")))?;
1911 // The canonical spelling, because that is what a row is named by and what the directory under
1912 // the cache is called, and a person is free to write a tuple the long way round.
1913 let tuple = target.to_canonical_string();
1914 // Before the table is consulted, because a target behind a licence wall is not a row that has not
1915 // been written yet. Section 13.4 is that no release pins one of these ever, so the message says
1916 // the licence and the two lawful ways rather than naming the producer that will publish the rest.
1917 if let Some(wall) = rucc_sysroot::Wall::of(target) {
1918 return Err(err(format!("--fetch {tuple}: {}", wall.no_fetch(&tuple))));
1919 }
1920 let Some(what) = rucc_sysroot::pinned_for(&tuple) else {
1921 return Err(err(unpinned(&tuple)));
1922 };
1923 Ok(Action::Fetch { what, target, cache: cache::dir() })
1924}
1925
1926/// Why there is nothing to fetch for a target, which is a different sentence while the table is
1927/// empty.
1928///
1929/// A release that pins nothing and a release that pins eleven targets and not this one are two
1930/// situations, and a message that did not tell them apart would send somebody looking for a typo in
1931/// their tuple when the answer is that this work is not finished.
1932fn unpinned(tuple: &str) -> String {
1933 let pinned = rucc_sysroot::pinned_targets();
1934 if pinned.is_empty() {
1935 return format!(
1936 "this release pins no sysroot for {tuple}, and it pins none for any target yet. A \
1937 sysroot is built and published by the producer in tamnd/rucc-cross, per \
1938 spec/cross-compile/13-distribution.md section 13.8, and a release of this compiler \
1939 names one by URL and by hash afterwards. Until then, pass --sysroot=<dir> to compile \
1940 against a tree you have already"
1941 );
1942 }
1943 format!(
1944 "this release pins no sysroot for {tuple}. What it pins is {}. Pass --sysroot=<dir> to \
1945 compile against a tree you have already",
1946 pinned.join(", ")
1947 )
1948}
1949
1950/// Gets the artifact and installs it, saying what each step did.
1951///
1952/// The steps are section 13.8's and so are the messages: the transport is somebody else's program
1953/// and the check is ours, so a person reading this wants to know which downloader ran, that the
1954/// bytes matched, how many files the record named and where the tree ended up. A fetch of something
1955/// that is already there says that instead and moves nothing.
1956fn fetch_sysroot(what: &rucc_sysroot::Pinned, target: TargetTuple, cache: &std::path::Path) -> i32 {
1957 let tuple = target.to_canonical_string();
1958 let archive = what.archive_in(cache);
1959 let say = |line: &str| println!("rucc: {tuple}: {line}");
1960 match fetch::fetch(what.url, what.sha256, &archive) {
1961 Ok(fetch::Fetched::AlreadyThere) => {
1962 say(&format!("{} is already here and matches the hash", archive.display()));
1963 }
1964 Ok(fetch::Fetched::Downloaded(by)) => {
1965 say(&format!("downloaded {} with {}", what.url, by.program()));
1966 }
1967 Err(why) => return complain(why),
1968 }
1969 match install::install(&archive, what.sha256, target, cache) {
1970 Ok(done) => {
1971 match &done.before {
1972 install::Before::Nothing => {
1973 say(&format!("{} files installed at {}", done.files, done.root.display()));
1974 }
1975 install::Before::TheSame => {
1976 say(&format!(
1977 "the same sysroot is already at {}, so nothing moved",
1978 done.root.display()
1979 ));
1980 }
1981 install::Before::Different(was) => {
1982 say(&format!(
1983 "{} files installed at {}, over a tree whose record digested to {was}",
1984 done.files,
1985 done.root.display()
1986 ));
1987 }
1988 }
1989 say(&format!("the record digests to {}", done.digest));
1990 0
1991 }
1992 Err(why) => complain(why),
1993 }
1994}
1995
1996/// What one of the `-dump` and `-print` flags prints.
1997///
1998/// GCC prints the name back unchanged when it cannot find the file a `-print` flag asked about,
1999/// which is what makes the answer safe to paste into a link line whether or not the file is
2000/// there, and this does the same.
2001fn answer(query: &Query, opts: &Options, link: &LinkOptions) -> Result<String, CliError> {
2002 let found = |name: &str| {
2003 link::find_in_search(link, opts.target, name)
2004 .map_or_else(|| name.to_owned(), |path| path.display().to_string())
2005 };
2006 Ok(match query {
2007 Query::Machine => opts.target.to_string(),
2008 Query::Version => VERSION.to_owned(),
2009 Query::Multiarch => link::multiarch(opts.target),
2010 // The three lines GCC prints, in its order and with its punctuation, because what reads
2011 // them is a script written against that shape. There is no installation directory to
2012 // report: this compiler is one binary that works wherever it is copied, and the headers
2013 // it ships are inside it, so `install` is where the binary is and nothing is under it.
2014 Query::SearchDirs => {
2015 let here = std::env::current_exe()
2016 .ok()
2017 .and_then(|p| p.parent().map(std::path::Path::to_path_buf))
2018 .unwrap_or_default();
2019 let list = |dirs: &[PathBuf]| {
2020 dirs.iter().map(|d| d.display().to_string()).collect::<Vec<_>>().join(":")
2021 };
2022 let libraries = link::search_dirs(link, opts.target);
2023 format!(
2024 "install: {}\nprograms: ={}\nlibraries: ={}",
2025 here.display(),
2026 list(&link.prefixes),
2027 list(&libraries)
2028 )
2029 }
2030 // The root the rest of the answers are under, which a build system asks for when it wants
2031 // to find a file itself rather than ask for one by name, and which is the first thing to
2032 // look at when a cross build read a header nobody expected. A native compile has no
2033 // sysroot and the answer is the empty line, which is what GCC prints when it was
2034 // configured without one. `--sysroot` wins over ours because it wins everywhere else.
2035 Query::Sysroot => {
2036 sysroot_root(opts, link).map(|root| root.display().to_string()).unwrap_or_default()
2037 }
2038 // Section 13.5 of `spec/cross-compile/13-distribution.md`: for every input that is not this
2039 // compiler's own code, what it is, where it was got, its hash, its licence and whether it
2040 // was bundled, generated or fetched. What is printed is the manifest the sysroot already
2041 // carries rather than a second format saying the same things, because the three uses 13.5
2042 // gives for this are a licence notice, a reproducibility check and a security audit, and all
2043 // three are somebody else parsing it. One format is one parser to write.
2044 // Read and rendered rather than copied out, so that what comes back is the format this
2045 // build understands. The last newline comes off because whatever prints an answer adds
2046 // one, the way it does for every other query here. Keeping it would put a blank line at
2047 // the end of the one answer that is a file somebody diffs against the file it came from.
2048 Query::SysrootProvenance => match sysroot_manifest(opts, link)? {
2049 Some(manifest) => manifest.render().trim_end_matches('\n').to_string(),
2050 None => String::new(),
2051 },
2052 // Section 13.2 of the same document, which asks for the hash of a cache directory's
2053 // contents in the directory's name. A name cannot carry one, because the path has to be
2054 // computable before anything has been read, by the producer about to write the files and by
2055 // the compiler about to read them, and neither has the contents when it asks. So the number
2056 // is here instead, and it is the sha256 of the record rather than of a walk of the tree,
2057 // which means `sha256sum` over the manifest answers the same thing.
2058 Query::SysrootDigest => match sysroot_manifest(opts, link)? {
2059 Some(manifest) => manifest.digest(),
2060 None => String::new(),
2061 },
2062 Query::FileName(name) => found(name),
2063 // The name GCC gives the library of routines a compiler's output calls that the C
2064 // library does not have. Ours is built in and there is no file, so the answer is the
2065 // name itself, which is what GCC prints when it cannot find one either.
2066 Query::Libgcc => found("libgcc.a"),
2067 // A program rather than a library: the linker and the archiver are the ones a build asks
2068 // about, and this compiler finds them on the path or under `-B` rather than shipping
2069 // them, so the name back is the honest answer unless a `-B` prefix holds one.
2070 Query::ProgName(name) => link
2071 .prefixes
2072 .iter()
2073 .map(|dir| dir.join(name))
2074 .find(|path| path.is_file())
2075 .map_or_else(|| name.clone(), |path| path.display().to_string()),
2076 })
2077}
2078
2079/// The root every sysroot answer is about.
2080///
2081/// One function rather than a copy in each, because the other flags exist to say what is inside the
2082/// tree this one names, and two answers that disagreed about which tree that is would be a
2083/// difference nobody would think to look for. `--sysroot` wins over ours because it wins everywhere
2084/// else.
2085fn sysroot_root(opts: &Options, link: &LinkOptions) -> Option<PathBuf> {
2086 link.sysroot
2087 .clone()
2088 .or_else(|| link::cross_sysroot(opts.target, link).map(|at| at.root().to_path_buf()))
2089}
2090
2091/// The record of the sysroot this command line reads, when there is one to read.
2092///
2093/// [`None`] covers two cases that both print nothing, and they are different things. A compile for
2094/// this machine has no sysroot at all, and a tree somebody laid out themselves and pointed
2095/// `--sysroot` at carries no manifest, so nothing here knows where any of it came from. Saying
2096/// nothing is the only honest answer to either, and a reader can tell it from a manifest with no
2097/// inputs in it because that one still has its header lines.
2098///
2099/// # Errors
2100///
2101/// A manifest this build cannot parse, and anything else that went wrong reading the file. Passing a
2102/// record we could not read on to whoever asked would make their parser the one that finds the
2103/// problem, and every use section 13.5 gives for these two flags is somebody else reading the
2104/// output.
2105fn sysroot_manifest(opts: &Options, link: &LinkOptions) -> Result<Option<Manifest>, CliError> {
2106 let Some(root) = sysroot_root(opts, link) else {
2107 return Ok(None);
2108 };
2109 let path = Sysroot::at(root, opts.target.tuple()).manifest_path();
2110 match std::fs::read_to_string(&path) {
2111 Ok(text) => Manifest::parse(&text)
2112 .map(Some)
2113 .map_err(|why| err(format!("{}: {why}", path.display()))),
2114 Err(why) if why.kind() == std::io::ErrorKind::NotFound => Ok(None),
2115 Err(why) => Err(err(format!("{}: {why}", path.display()))),
2116 }
2117}
2118
2119/// Renders the passes this level will run, in order, with what each one does.
2120///
2121/// The level is the whole of the answer unless a `-f` flag edited it, which is section 9.1 of
2122/// `spec/09-optimizer.md`: a level is a list somebody wrote down rather than something that
2123/// emerges from which flags happen to be set, and this is how that list is read.
2124#[must_use]
2125pub fn print_pipeline(opts: &Options) -> String {
2126 let mut settings = rucc_opt::Options::for_level(opts.opt_level);
2127 settings.toggles.clone_from(&opts.passes);
2128 settings.global_fuel = opts.pass_fuel_global;
2129 for (on, spec) in &opts.pass_gates {
2130 // Every spelling was checked while the arguments were parsed, so there is nothing here
2131 // this can refuse, and a listing is not the place to report it if there were.
2132 let _ = settings.gates.add(*on, spec);
2133 }
2134 rucc_opt::pipeline::print(&settings)
2135}
2136
2137/// Renders the resolved configuration.
2138///
2139/// One `key: value` per line, sorted by nothing in particular but fixed in order, because
2140/// this output is diffed across hosts in CI and a reordering would read as a change.
2141#[must_use]
2142pub fn print_config(opts: &Options) -> String {
2143 let sess = Session::new(opts.clone());
2144 let t = &sess.target;
2145 let mut out = String::new();
2146 let _ = writeln!(out, "version: {VERSION}");
2147 // The three field triple the driver was given rather than the ten field tuple it widens to,
2148 // because this output is what a build system reads to find out what it asked for. The tuple is
2149 // the compiler's model of the machine and this line is a receipt for a command line.
2150 let _ = writeln!(out, "target: {}", opts.target);
2151 let _ = writeln!(out, "arch: {}", opts.target.arch.as_str());
2152 let _ = writeln!(out, "os: {}", opts.target.os.as_str());
2153 let _ = writeln!(out, "env: {}", opts.target.env.as_str());
2154 let _ = writeln!(out, "object-format: {}", t.object_format.as_str());
2155 let _ = writeln!(out, "pointer-width: {}", t.pointer_width);
2156 let _ = writeln!(out, "long-width: {}", t.long_width);
2157 let _ = writeln!(out, "long-double-width: {}", t.long_double_width);
2158 let _ = writeln!(out, "endian: {}", if t.little_endian { "little" } else { "big" });
2159 let _ = writeln!(out, "char-signed: {}", t.char_is_signed);
2160 let _ = writeln!(out, "va-list: {}", t.va_list.map_or("none", |list| list.as_str()));
2161 // The register file as a count per class, which is enough to tell a target whose registers
2162 // are described from one whose are not without printing sixteen names nobody asked for.
2163 let regs: Vec<String> = t
2164 .regs
2165 .classes()
2166 .map(|(class, info)| format!("{} {}", info.name, t.regs.len(class)))
2167 .collect();
2168 let _ = writeln!(
2169 out,
2170 "registers: {}",
2171 if regs.is_empty() { "none".to_string() } else { regs.join(", ") }
2172 );
2173 // What the schedule was chosen with, which is a sentence rather than a name on purpose: two
2174 // runs of a benchmark that disagree are usually two models and not two compilers.
2175 let _ = writeln!(out, "timing-model: {}", t.timing.map_or("none", |timing| timing.model));
2176 let _ = writeln!(out, "opt-level: {}", sess.opts.opt_level);
2177 let _ = writeln!(out, "safety: {}", sess.opts.safety);
2178 let _ = writeln!(out, "emit: {}", sess.opts.emit.as_str());
2179 let _ = writeln!(out, "debug-info: {}", sess.opts.debug_info);
2180 let _ = writeln!(out, "frame-pointer: {}", sess.opts.frame_pointer);
2181 let _ = writeln!(out, "red-zone: {}", sess.opts.red_zone);
2182 let _ = writeln!(out, "stack-protector: {}", sess.opts.protector);
2183 let _ = writeln!(out, "stack-clash-protection: {}", sess.opts.stack_clash);
2184 let _ = writeln!(out, "cf-protection: {}", sess.opts.control);
2185 let _ = writeln!(out, "patchable-function-entry: {}", sess.opts.patchable);
2186 let _ = writeln!(out, "profile: {}", sess.opts.profile);
2187 let _ = writeln!(out, "profile-hook: {}", sess.opts.hook);
2188 // Last because it is the one key with more than one line under it, and the only one
2189 // whose value is a property of the machine rather than of the command line.
2190 for dir in sess.opts.search.dirs() {
2191 let system = if dir.is_system { " (system)" } else { "" };
2192 let _ = writeln!(out, "include: {}{system}", dir.path.display());
2193 }
2194 out
2195}
2196
2197/// The output name the make target is taken from, which is the `-o` argument or nothing.
2198///
2199/// A run that stops at the preprocessor has not named an object, whatever its `-o` says: under
2200/// `-E` that argument is the preprocessed text and under `-M` it is the rule itself, and neither
2201/// is a file `make` would rebuild by running this rule. GCC agrees and falls back to the source
2202/// name in both, which is why a `-MD -E -o out.i` writes `out.d` holding a rule for `a.o`. From
2203/// `-S` on the argument does name what the rule builds, and it is used as written.
2204fn deps_target_output<'a>(opts: &Options, plan: &'a Plan) -> Option<&'a str> {
2205 if opts.emit == EmitKind::Preprocessed { None } else { plan.output.as_deref() }
2206}
2207
2208/// Writes to a path the command line named rather than one the plan derived, where `-` is
2209/// standard output.
2210fn write_named(path: &str, bytes: &[u8]) -> Result<(), String> {
2211 if path == "-" {
2212 return write_out(&Output::Stdout, bytes);
2213 }
2214 write_out(&Output::File(path.to_owned()), bytes)
2215}
2216
2217/// Writes the make rule for one input, and reports whether it got there.
2218///
2219/// A rule with no file of its own goes where the compilation it replaced would have written,
2220/// which is what makes the usual makefile recipe work: `rucc -M $< -o $@` leaves the rule in
2221/// `$@`, and the same line with the `-o` left off puts it on standard output.
2222fn write_deps(
2223 opts: &Options,
2224 plan: &Plan,
2225 job: &Job,
2226 found: &[Dependency],
2227 stderr: &mut impl std::io::Write,
2228) -> bool {
2229 let targets = if opts.deps.targets.is_empty() {
2230 vec![deps::default_target(&job.input, deps_target_output(opts, plan))]
2231 } else {
2232 opts.deps.targets.clone()
2233 };
2234 let rule = deps::rule(&opts.deps, &targets, &job.input, found);
2235 // The file, on the other hand, is named after the `-o` in every mode that still has one to
2236 // spend, which is every mode except the two that spend it on the rule.
2237 let wrote = match deps::default_file(&opts.deps, &job.input, plan.output.as_deref()) {
2238 // A `-MF` on a run that had nowhere else to put the rule leaves the file the `-o`
2239 // named empty rather than absent, because a makefile that named it as a target of its
2240 // own is a makefile that will look for it.
2241 Some(path) => write_named(&path, rule.as_bytes()).and_then(|()| {
2242 if opts.deps.instead_of_compiling { write_out(&job.output, b"") } else { Ok(()) }
2243 }),
2244 None => write_out(&job.output, rule.as_bytes()),
2245 };
2246 if let Err(e) = wrote {
2247 let _ = writeln!(stderr, "rucc: error: {e}");
2248 return false;
2249 }
2250 true
2251}
2252
2253/// Runs phase 4 over every input that has one, and writes what came out.
2254///
2255/// One input that fails does not stop the others. A build that reports every file it could
2256/// not preprocess in one run is worth more than one that stops at the first, and the exit
2257/// status is still a failure either way.
2258fn preprocess_all(opts: &Options, plan: &Plan) -> i32 {
2259 let fs = OsFileSystem::new();
2260 let mut stderr = std::io::stderr().lock();
2261 let mut failed = false;
2262 for job in &plan.jobs {
2263 if !job.phases.first().is_some_and(|p| *p == Phase::Preprocess) {
2264 // An input that is already preprocessed, or an object file. GCC passes these
2265 // through untouched, and the plan has already said so in its notes.
2266 continue;
2267 }
2268 let started = std::time::Instant::now();
2269 let result = preprocess(opts, &job.input, &fs);
2270 if opts.time {
2271 say_time(&job.input, started.elapsed(), &mut stderr);
2272 }
2273 for message in &result.messages {
2274 let _ = writeln!(stderr, "{message}");
2275 }
2276 if result.failed() {
2277 failed = true;
2278 continue;
2279 }
2280 if opts.deps.emit {
2281 failed |= !write_deps(opts, plan, job, &result.deps, &mut stderr);
2282 // `-M` and `-MM` asked for the rule instead of the text, so there is nothing else
2283 // to write. The other two asked for both and fall through to the text below.
2284 if opts.deps.instead_of_compiling {
2285 continue;
2286 }
2287 }
2288 if let Err(e) = write_out(&job.output, result.text.as_bytes()) {
2289 let _ = writeln!(stderr, "rucc: error: {e}");
2290 failed = true;
2291 }
2292 }
2293 i32::from(failed)
2294}
2295
2296/// Runs the front end over every input that has a compile phase, and writes what came out.
2297///
2298/// The same rule as [`preprocess_all`]: one input that fails does not stop the others, and the
2299/// exit status is a failure either way. An input that is already assembly or an object has no
2300/// compile phase and is passed over here, which the plan has already said in its notes.
2301fn compile_all(opts: &Options, plan: &Plan) -> i32 {
2302 let fs = OsFileSystem::new();
2303 let mut stderr = std::io::stderr().lock();
2304 let mut failed = false;
2305 let (mut remarks, ok) = Remarks::new(opts.opt_info_file.as_ref(), &mut stderr);
2306 failed |= !ok;
2307 let mut fired = Fired::new();
2308 let mut pressure = Pressure::new();
2309 let mut lowerings = Lowerings::new();
2310 for job in &plan.jobs {
2311 if !job.phases.contains(&Phase::Compile) {
2312 continue;
2313 }
2314 // An input of IR is read back rather than compiled, since the C it came from is not
2315 // here any more. Everything after this is the same, so the two paths meet again at the
2316 // messages and the file the result is written to.
2317 let started = std::time::Instant::now();
2318 let result = if job.kind == InputKind::Ir {
2319 compile_ir(opts, &job.input, &fs)
2320 } else {
2321 compile(opts, &job.input, &fs)
2322 };
2323 if opts.time {
2324 say_time(&job.input, started.elapsed(), &mut stderr);
2325 }
2326 fired.merge(&result.fired);
2327 pressure.merge(&result.pressure);
2328 lowerings.merge(&result.lowerings);
2329 failed |= !write_dumps(&job.input, &result.dumps, &mut stderr);
2330 failed |= !remarks.write(&result.remarks, &mut stderr);
2331 for message in &result.messages {
2332 let _ = writeln!(stderr, "{message}");
2333 }
2334 // Before the failure below, because a compilation that stopped in the back end is exactly
2335 // the one whose preprocessed source somebody wants to look at.
2336 failed |= !write_temps(job, &result.temps, &mut stderr);
2337 if result.failed() {
2338 failed = true;
2339 continue;
2340 }
2341 // `-MD` and `-MMD` write the rule beside the object and let the compilation happen, so
2342 // this is the one path where both files come out of the same run. An input of IR has no
2343 // dependencies to report and produces an empty list, which produces a rule naming only
2344 // itself, and that is the honest answer rather than a missing file.
2345 if opts.deps.emit {
2346 failed |= !write_deps(opts, plan, job, &result.deps, &mut stderr);
2347 }
2348 if let Err(e) = write_out(&job.output, result.artifact.bytes()) {
2349 let _ = writeln!(stderr, "rucc: error: {e}");
2350 failed = true;
2351 }
2352 }
2353 failed |= !write_coverage(opts, &fired, &mut stderr);
2354 failed |= !write_pressure(opts, &pressure, &mut stderr);
2355 failed |= !write_lowering(opts, &lowerings, &mut stderr);
2356 i32::from(failed)
2357}
2358
2359/// A directory for the object files only the link step ever sees, removed when it goes away.
2360///
2361/// `-c` writes its object where the user can see it and linking does not, which is the whole of
2362/// the difference: a `rucc a.c b.c` leaves an executable behind and nothing else, the same as
2363/// every other compiler. Removing them on drop rather than at the end of a function is so that a
2364/// link that failed leaves nothing behind either.
2365struct Scratch {
2366 /// Where the objects go.
2367 dir: PathBuf,
2368}
2369
2370impl Scratch {
2371 /// Makes one, under whatever the platform calls its temporary directory.
2372 ///
2373 /// The name carries the process id so that two compilers running at once do not share a
2374 /// directory, which they would otherwise do the moment two of them compiled a file of the
2375 /// same name.
2376 fn new() -> Result<Scratch, String> {
2377 let dir = std::env::temp_dir().join(format!("rucc-{}", std::process::id()));
2378 std::fs::create_dir_all(&dir).map_err(|e| format!("{}: {e}", dir.display()))?;
2379 Ok(Scratch { dir })
2380 }
2381}
2382
2383impl Drop for Scratch {
2384 fn drop(&mut self) {
2385 let _ = std::fs::remove_dir_all(&self.dir);
2386 }
2387}
2388
2389/// The link line the plan describes, for `-###`.
2390///
2391/// The names in it are the hints the plan carries rather than the temporaries a real compilation
2392/// would choose, because `-###` prints the line without having compiled anything and so has
2393/// nothing to point at. That also makes the printed line readable rather than naming a directory
2394/// that only exists while a compilation is running.
2395fn link_line(opts: &Options, link: &LinkOptions, job: &LinkJob) -> Result<String, link::Error> {
2396 let linker = link::find(opts.target, link)?;
2397 let args = link::line(opts.target, link, &job.inputs, &job.output)?;
2398 Ok(link::render(&linker, &args))
2399}
2400
2401/// Compiles everything, then links it.
2402///
2403/// The objects go in a directory that is removed afterwards, which is why this is not
2404/// [`compile_all`] followed by a link: the plan says an object feeding the linker is temporary
2405/// and does not say where, because where is a question that only has an answer once something is
2406/// running.
2407fn link_all(opts: &Options, plan: &Plan, link: &LinkOptions, verbose: bool) -> i32 {
2408 let Some(job) = &plan.link else {
2409 // Every path into here comes from a plan whose last phase is the link, and such a plan
2410 // has a link job. Saying so is cheaper than an unwrap that would have to be explained.
2411 let mut stderr = std::io::stderr().lock();
2412 let _ = writeln!(stderr, "rucc: error: there is nothing to link");
2413 return 1;
2414 };
2415 // Before anything is compiled, because a linker that is not on the machine is worth knowing
2416 // about in the second it takes to look rather than after the compilation.
2417 // And before that, whether this link has a line at all and whether what it reads is on the
2418 // machine. Both are answerable now, and a target whose sysroot has not been built is worth
2419 // saying so about before the compilation rather than after it.
2420 if let Err(why) = link::preflight(opts.target, link) {
2421 return complain(why);
2422 }
2423 let linker = match link::find(opts.target, link) {
2424 Ok(linker) => linker,
2425 Err(why) => return complain(why),
2426 };
2427
2428 let scratch = match Scratch::new() {
2429 Ok(scratch) => scratch,
2430 Err(why) => return complain(format!("could not make a place for the object files: {why}")),
2431 };
2432
2433 let fs = OsFileSystem::new();
2434 let mut failed = false;
2435 // One per job, in job order, which is what lets the link line below be rebuilt with the real
2436 // paths in it: every job contributes exactly one file to the line and does so in this order.
2437 let mut produced: Vec<String> = Vec::with_capacity(plan.jobs.len());
2438 let mut fired = Fired::new();
2439 let mut pressure = Pressure::new();
2440 let mut lowerings = Lowerings::new();
2441 {
2442 let mut stderr = std::io::stderr().lock();
2443 let (mut remarks, ok) = Remarks::new(opts.opt_info_file.as_ref(), &mut stderr);
2444 failed |= !ok;
2445 for (at, job) in plan.jobs.iter().enumerate() {
2446 let out = match &job.output {
2447 Output::Temporary(hint) => {
2448 // The index because two inputs in different directories can have the same
2449 // name, and the two objects of `rucc a/x.c b/x.c` must not be one file.
2450 scratch.dir.join(format!("{at}-{hint}")).display().to_string()
2451 }
2452 Output::File(path) => path.clone(),
2453 // A job feeding the linker never writes to standard output, since the plan gives
2454 // it a temporary. This is here so that the match is total rather than a panic.
2455 Output::Stdout => continue,
2456 };
2457 produced.push(out.clone());
2458 if !job.phases.contains(&Phase::Compile) {
2459 continue;
2460 }
2461 let started = std::time::Instant::now();
2462 let result = if job.kind == InputKind::Ir {
2463 compile_ir(opts, &job.input, &fs)
2464 } else {
2465 compile(opts, &job.input, &fs)
2466 };
2467 if opts.time {
2468 say_time(&job.input, started.elapsed(), &mut stderr);
2469 }
2470 fired.merge(&result.fired);
2471 pressure.merge(&result.pressure);
2472 lowerings.merge(&result.lowerings);
2473 lowerings.merge(&result.lowerings);
2474 failed |= !write_dumps(&job.input, &result.dumps, &mut stderr);
2475 failed |= !remarks.write(&result.remarks, &mut stderr);
2476 for message in &result.messages {
2477 let _ = writeln!(stderr, "{message}");
2478 }
2479 failed |= !write_temps(job, &result.temps, &mut stderr);
2480 if result.failed() {
2481 failed = true;
2482 continue;
2483 }
2484 // A `-MD` on a command line that links writes the rule next to the executable and
2485 // names the executable as its target, since that is the file this source builds
2486 // here. The object it went through is in a temporary directory and is gone by the
2487 // time `make` reads any of this.
2488 if opts.deps.emit {
2489 failed |= !write_deps(opts, plan, job, &result.deps, &mut stderr);
2490 }
2491 if !matches!(result.artifact, Artifact::Object { .. }) {
2492 // Worth saying rather than writing whatever it is and letting the linker read it.
2493 // An empty file is a valid empty linker script, so a link handed one gets as far
2494 // as reporting every symbol of this file undefined, which is a page of messages
2495 // about something that went wrong here.
2496 let _ = writeln!(
2497 stderr,
2498 "rucc: internal error: {}: no object file was produced for the link",
2499 job.input
2500 );
2501 failed = true;
2502 continue;
2503 }
2504 if let Err(e) = std::fs::write(&out, result.artifact.bytes()) {
2505 let _ = writeln!(stderr, "rucc: error: {out}: {e}");
2506 failed = true;
2507 }
2508 }
2509 failed |= !write_coverage(opts, &fired, &mut stderr);
2510 failed |= !write_pressure(opts, &pressure, &mut stderr);
2511 failed |= !write_lowering(opts, &lowerings, &mut stderr);
2512 failed |= !write_lowering(opts, &lowerings, &mut stderr);
2513 }
2514 if failed {
2515 // Nothing is linked from a compilation that did not finish. A linker run over the objects
2516 // that did compile would report every function of the file that did not as undefined,
2517 // which is a page of messages about a mistake already reported once.
2518 return 1;
2519 }
2520
2521 // The items in command line order with the temporaries filled in. A library contributes no
2522 // job and passes through, and every file item takes the next job's real output, which is
2523 // what keeps a library that was written between two objects between them here.
2524 let mut outputs = produced.into_iter();
2525 let mut items = Vec::with_capacity(job.inputs.len());
2526 for item in &job.inputs {
2527 match item {
2528 link::Item::Library(name) => items.push(link::Item::Library(name.clone())),
2529 link::Item::File(_) => match outputs.next() {
2530 Some(path) => items.push(link::Item::File(path)),
2531 None => return complain("the plan asks the linker for a file nothing produced"),
2532 },
2533 }
2534 }
2535
2536 let args = match link::line(opts.target, link, &items, &job.output) {
2537 Ok(args) => args,
2538 Err(why) => return complain(why),
2539 };
2540 if verbose {
2541 let mut stderr = std::io::stderr().lock();
2542 let _ = writeln!(stderr, "{}", link::render(&linker, &args));
2543 }
2544 let started = std::time::Instant::now();
2545 let ran = link::run(&linker, &args);
2546 if opts.time {
2547 // The one step of a compilation that really is another program, so this line is the same
2548 // measurement gcc's is and names the linker the way gcc names `collect2`.
2549 let mut stderr = std::io::stderr().lock();
2550 say_time(&linker.name, started.elapsed(), &mut stderr);
2551 }
2552 match ran {
2553 Ok(()) => 0,
2554 // The linker has already said what was wrong on its own error output, and repeating that
2555 // linking failed would only push its message further up the screen.
2556 Err(link::Error::Refused { .. }) => 1,
2557 Err(why) => complain(why),
2558 }
2559}
2560
2561/// Compiles everything and writes the objects into one static library.
2562///
2563/// No temporary directory and no second program. The objects never reach the file system at all:
2564/// they go from the compiler into the archive writer, which is both faster than writing a directory
2565/// of files for an `ar` to read back and the reason the symbol index can be written at all. A
2566/// member's index entries are the names the object writer says it wrote, and the only thing that
2567/// knows those is the run that wrote it.
2568///
2569/// `-save-temps` is the exception. It asked for the objects to be kept, the plan gave them names a
2570/// person can find, and they are written there as well as put in the archive.
2571fn archive_all(opts: &Options, plan: &Plan) -> i32 {
2572 let Some(job) = &plan.archive else {
2573 // Every path into here comes from a plan whose last phase is the archive, and such a plan
2574 // has an archive job. Saying so is cheaper than an unwrap that would have to be explained.
2575 return complain("there is nothing to put in an archive");
2576 };
2577 // Before anything is compiled, because a format this has no container for is worth knowing
2578 // about in the second it takes to look rather than after the whole compilation.
2579 let flavour = match opts.target.os.object_format() {
2580 ObjectFormat::Elf => rucc_archive::Flavour::Gnu,
2581 ObjectFormat::Coff => rucc_archive::Flavour::Coff,
2582 // Mach-O wants the BSD flavour, whose index is a different member under a different name,
2583 // and wasm has no archives of its own at all. Neither has an object writer either, so a
2584 // command line reaching this would have failed in the next step regardless.
2585 format @ (ObjectFormat::MachO | ObjectFormat::Wasm) => {
2586 return complain(format!(
2587 "there is no archive format for {} objects in this compiler yet",
2588 format.as_str()
2589 ));
2590 }
2591 };
2592
2593 let fs = OsFileSystem::new();
2594 let mut failed = false;
2595 let mut members: Vec<rucc_archive::Member> = Vec::with_capacity(plan.jobs.len());
2596 let mut names = job.members.iter();
2597 let mut fired = Fired::new();
2598 let mut pressure = Pressure::new();
2599 let mut lowerings = Lowerings::new();
2600 {
2601 let mut stderr = std::io::stderr().lock();
2602 let (mut remarks, ok) = Remarks::new(opts.opt_info_file.as_ref(), &mut stderr);
2603 failed |= !ok;
2604 for plan_job in &plan.jobs {
2605 // What the plan called this member. The two lists are walked together rather than the
2606 // name being worked out again here, so that what `-###` printed and what goes in the
2607 // file cannot come apart.
2608 let Some(member) = names.next() else {
2609 return complain("the plan asks the archive for a member nothing produced");
2610 };
2611 if !plan_job.phases.contains(&Phase::Compile) {
2612 // Assembly, which enters the pipeline after the compile phase. There is no
2613 // assembler for a file of text in this compiler, so there is no object to put in,
2614 // and an archive quietly missing one is worse than a message about it.
2615 let _ = writeln!(
2616 &mut stderr,
2617 "rucc: error: {}: this compiler has no assembler for a file of assembly yet, \
2618 so it cannot go into an archive",
2619 plan_job.input
2620 );
2621 failed = true;
2622 continue;
2623 }
2624 let started = std::time::Instant::now();
2625 let result = if plan_job.kind == InputKind::Ir {
2626 compile_ir(opts, &plan_job.input, &fs)
2627 } else {
2628 compile(opts, &plan_job.input, &fs)
2629 };
2630 if opts.time {
2631 say_time(&plan_job.input, started.elapsed(), &mut stderr);
2632 }
2633 fired.merge(&result.fired);
2634 pressure.merge(&result.pressure);
2635 lowerings.merge(&result.lowerings);
2636 lowerings.merge(&result.lowerings);
2637 failed |= !write_dumps(&plan_job.input, &result.dumps, &mut stderr);
2638 failed |= !remarks.write(&result.remarks, &mut stderr);
2639 for message in &result.messages {
2640 let _ = writeln!(stderr, "{message}");
2641 }
2642 failed |= !write_temps(plan_job, &result.temps, &mut stderr);
2643 if result.failed() {
2644 failed = true;
2645 continue;
2646 }
2647 if opts.deps.emit {
2648 failed |= !write_deps(opts, plan, plan_job, &result.deps, &mut stderr);
2649 }
2650 let Artifact::Object { bytes, defines } = result.artifact else {
2651 let _ = writeln!(
2652 stderr,
2653 "rucc: internal error: {}: no object file was produced for the archive",
2654 plan_job.input
2655 );
2656 failed = true;
2657 continue;
2658 };
2659 // Under `-save-temps` the plan gave the object a name a person can find, so it is
2660 // written there too. Otherwise it is only ever a member and never a file.
2661 if let Output::File(path) = &plan_job.output {
2662 if let Err(e) = std::fs::write(path, &bytes) {
2663 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2664 failed = true;
2665 }
2666 }
2667 members.push(rucc_archive::Member { name: member.clone(), body: bytes, defines });
2668 }
2669 failed |= !write_coverage(opts, &fired, &mut stderr);
2670 failed |= !write_pressure(opts, &pressure, &mut stderr);
2671 failed |= !write_lowering(opts, &lowerings, &mut stderr);
2672 failed |= !write_lowering(opts, &lowerings, &mut stderr);
2673 }
2674 if failed {
2675 // Nothing is written from a compilation that did not finish, for the reason the link gives:
2676 // an archive missing the file that failed is one a link reports every name of as undefined,
2677 // which is a page of messages about a mistake already reported once.
2678 return 1;
2679 }
2680
2681 let bytes = match rucc_archive::write(flavour, &members) {
2682 Ok(bytes) => bytes,
2683 // Every one of these is a bug here rather than a program's mistake: the names came from the
2684 // object writer and the bodies came from this process.
2685 Err(why) => return complain(format!("the archive could not be written: {why}")),
2686 };
2687 match std::fs::write(&job.output, &bytes) {
2688 Ok(()) => 0,
2689 Err(e) => complain(format!("{}: {e}", job.output)),
2690 }
2691}
2692
2693/// Prints one driver level message and gives back the exit status that goes with it.
2694fn complain(why: impl std::fmt::Display) -> i32 {
2695 let mut stderr = std::io::stderr().lock();
2696 let _ = writeln!(stderr, "rucc: error: {why}");
2697 1
2698}
2699
2700/// Writes what `-Zrule-coverage=FILE` asked for, and says whether it could.
2701///
2702/// Once for the whole command line rather than once per input, because the question is which
2703/// lowering rules this run of the compiler reached and a file per input would leave the reader
2704/// unioning files to find out something one process already knew.
2705///
2706/// A file that could not be written is a failure and not a warning. What asks for this is a
2707/// measurement run, and a measurement that quietly did not happen is worse than one that stopped.
2708fn write_coverage(opts: &Options, fired: &Fired, stderr: &mut impl std::io::Write) -> bool {
2709 let Some(path) = &opts.rule_coverage else { return true };
2710 let Some(table) = coverage::table(opts.target.arch) else {
2711 let _ = writeln!(
2712 stderr,
2713 "rucc: error: there are no lowering rules for {} yet, so there is no coverage of them \
2714 to report",
2715 opts.target
2716 );
2717 return false;
2718 };
2719 match std::fs::write(path, fired.listing(table)) {
2720 Ok(()) => true,
2721 Err(e) => {
2722 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2723 false
2724 }
2725 }
2726}
2727
2728/// Writes what `-Zregister-pressure=FILE` asked for, and says whether it could.
2729///
2730/// Once for the whole command line, for the reason [`write_coverage`] gives, and a file that could
2731/// not be written is a failure for the reason it gives too. There is no equivalent of the missing
2732/// rule table here, since every target this compiles for has an allocator, and a run that reached
2733/// no back end at all writes an empty listing rather than nothing: a measurement of a build that
2734/// produced no code is still an answer and it is the honest one.
2735fn write_pressure(opts: &Options, pressure: &Pressure, stderr: &mut impl std::io::Write) -> bool {
2736 let Some(path) = &opts.register_pressure else { return true };
2737 match std::fs::write(path, pressure.listing()) {
2738 Ok(()) => true,
2739 Err(e) => {
2740 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2741 false
2742 }
2743 }
2744}
2745
2746/// Writes what `-Zlowering=FILE` asked for, and says whether it could.
2747///
2748/// Once for the whole command line, for the reason [`write_coverage`] gives, and a file that could
2749/// not be written is a failure for the reason it gives too. A run that reached no back end writes
2750/// an empty listing rather than nothing, the way [`write_pressure`] does and for the same reason.
2751fn write_lowering(opts: &Options, lowerings: &Lowerings, stderr: &mut impl std::io::Write) -> bool {
2752 let Some(path) = &opts.lowering_dump else { return true };
2753 match std::fs::write(path, lowerings.listing()) {
2754 Ok(()) => true,
2755 Err(e) => {
2756 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2757 false
2758 }
2759 }
2760}
2761
2762/// Where the `-fopt-info` remarks go, and how much of the run has already gone there.
2763///
2764/// Standard error by default, and one file for the whole run when `-fopt-info=<file>` named one.
2765/// A file rather than the diagnostic stream is what a harness wants: the corpus in
2766/// `tamnd/rucc-corpus` matches a rejection against what the compiler said on standard error, and
2767/// a few thousand remarks mixed into that would bury it.
2768struct Remarks {
2769 /// The file, if there is one.
2770 file: Option<String>,
2771 /// Whether anything has been written to it yet, which decides between truncating and
2772 /// appending. One file holds the whole run rather than the last input in it.
2773 started: bool,
2774}
2775
2776impl Remarks {
2777 /// Prepares the destination, emptying the file if there is one.
2778 ///
2779 /// Emptied here rather than at the first remark, because a run where no pass had anything to
2780 /// say should leave an empty file and not yesterday's. An absent file and an empty one are
2781 /// different facts and something reading this will act on the difference.
2782 fn new(file: Option<&String>, stderr: &mut impl std::io::Write) -> (Self, bool) {
2783 let mut ok = true;
2784 if let Some(path) = file {
2785 if let Err(e) = std::fs::write(path, "") {
2786 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2787 ok = false;
2788 }
2789 }
2790 (Self { file: file.cloned(), started: false }, ok)
2791 }
2792
2793 /// Writes one input's remarks, and says whether that worked.
2794 ///
2795 /// A file that cannot be written is a failure and not a warning, for the reason
2796 /// [`write_dumps`] gives: remarks that quietly did not arrive look exactly like a compilation
2797 /// where nothing happened.
2798 fn write(&mut self, text: &str, stderr: &mut impl std::io::Write) -> bool {
2799 if text.is_empty() {
2800 return true;
2801 }
2802 let Some(path) = &self.file else {
2803 let _ = write!(stderr, "{text}");
2804 return true;
2805 };
2806 let opened = std::fs::OpenOptions::new()
2807 .write(true)
2808 .append(self.started)
2809 .truncate(!self.started)
2810 .create(true)
2811 .open(path);
2812 self.started = true;
2813 let result =
2814 opened.and_then(|mut file| std::io::Write::write_all(&mut file, text.as_bytes()));
2815 if let Err(e) = result {
2816 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2817 return false;
2818 }
2819 true
2820 }
2821}
2822
2823/// Writes what `-fdump-ir=` asked to see, one file per dump.
2824///
2825/// The name is the input file with the dump's own name and `.ir` after it, so a directory listing
2826/// after a run is the passes in the order they ran, per input. They go in the working directory
2827/// rather than beside the output, because a dump is something a person asked for at a prompt and
2828/// the working directory is where that person is.
2829///
2830/// A file that could not be written is a failure and not a warning, for the reason
2831/// [`write_coverage`] gives: what asked for this is somebody debugging a pass, and a dump that
2832/// quietly did not happen looks exactly like a pass that did not run.
2833fn write_dumps(input: &str, dumps: &[rucc_opt::Dump], stderr: &mut impl std::io::Write) -> bool {
2834 let stem = std::path::Path::new(input)
2835 .file_name()
2836 .map_or_else(|| input.to_owned(), |name| name.to_string_lossy().into_owned());
2837 let mut ok = true;
2838 for dump in dumps {
2839 let path = format!("{stem}.{}.ir", dump.name);
2840 if let Err(e) = std::fs::write(&path, &dump.text) {
2841 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2842 ok = false;
2843 }
2844 }
2845 ok
2846}
2847
2848/// Writes the files `-save-temps` kept, which is nothing at all unless it was given.
2849///
2850/// A file that could not be written is a failure rather than a warning, for the reason
2851/// [`write_dumps`] gives: somebody asked for these by name, and one that quietly did not happen
2852/// looks like a compilation that never went through that step.
2853fn write_temps(job: &Job, temps: &Temps, stderr: &mut impl std::io::Write) -> bool {
2854 let mut ok = true;
2855 let kept = [(job.saved_text(), &temps.preprocessed), (job.saved_asm(), &temps.assembly)];
2856 for (path, text) in kept {
2857 // A step the compilation did not reach has nothing to keep, and a job that is not keeping
2858 // that step has nowhere to put it. Either way there is no file here.
2859 let (Some(path), Some(text)) = (path, text) else { continue };
2860 if let Err(e) = std::fs::write(&path, text) {
2861 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2862 ok = false;
2863 }
2864 }
2865 ok
2866}
2867
2868/// One line of `-time`, which is what a step was called and how long it took.
2869///
2870/// GCC's two numbers are the user and the system time of a subprocess it ran. This compiler runs
2871/// no subprocess for anything but the link, so what is measured here is the wall clock of the
2872/// step and the second column is always zero. The shape of the line is kept because a person
2873/// reading it next to gcc's should not have to work out which column is which.
2874fn say_time(name: &str, took: std::time::Duration, stderr: &mut impl std::io::Write) {
2875 let _ = writeln!(stderr, "# {name} {:.2} {:.2}", took.as_secs_f64(), 0.0);
2876}
2877
2878/// Writes one job's result where the plan said it goes.
2879///
2880/// # Errors
2881///
2882/// Returns the message to print, which names the file when there is one, because "permission
2883/// denied" on its own does not say which file was refused.
2884fn write_out(output: &Output, bytes: &[u8]) -> Result<(), String> {
2885 match output {
2886 Output::Stdout => {
2887 let mut stdout = std::io::stdout().lock();
2888 stdout.write_all(bytes).map_err(|e| format!("writing to standard output: {e}"))
2889 }
2890 Output::File(path) | Output::Temporary(path) => {
2891 std::fs::write(path, bytes).map_err(|e| format!("{path}: {e}"))
2892 }
2893 }
2894}
2895
2896/// Runs the driver and returns the process exit code.
2897///
2898/// `args` excludes the program name. Output goes to `stdout` and errors to `stderr`, which
2899/// is the one place in the compiler that is true.
2900pub fn run(args: &[String]) -> i32 {
2901 match parse_args(args) {
2902 Ok(Action::Help) => {
2903 print!("{USAGE}");
2904 0
2905 }
2906 Ok(Action::Version) => {
2907 println!("rucc {VERSION}");
2908 0
2909 }
2910 Ok(Action::Print(line)) => {
2911 println!("{line}");
2912 0
2913 }
2914 Ok(Action::PrintConfig(opts)) => {
2915 print!("{}", print_config(&opts));
2916 0
2917 }
2918 Ok(Action::PrintPipeline(opts)) => {
2919 print!("{}", print_pipeline(&opts));
2920 0
2921 }
2922 Ok(Action::PrintPlan { opts, plan, link }) => {
2923 print!("{}", plan.render());
2924 // The line as it would be typed, which is the half of `-###` that section 4.3 says
2925 // arrives with the link. It is printed even when the linker is not on this machine,
2926 // because what a build wants from `-###` is what the compiler would do.
2927 if let Some(job) = &plan.link {
2928 match link_line(&opts, &link, job) {
2929 Ok(line) => println!("{line}"),
2930 Err(why) => {
2931 let mut stderr = std::io::stderr().lock();
2932 let _ = writeln!(stderr, "rucc: error: {why}");
2933 return 1;
2934 }
2935 }
2936 }
2937 0
2938 }
2939 Ok(Action::Fetch { what, target, cache }) => fetch_sysroot(what, target, &cache),
2940 Ok(Action::Compile { opts, plan, link, jobs, verbose, notes }) => {
2941 {
2942 let mut stderr = std::io::stderr().lock();
2943 // Before the plan rather than after it, because a note is about the command line
2944 // and the plan is what the command line was read as, so the reader wants the two
2945 // in that order.
2946 for note in ¬es {
2947 let _ = writeln!(stderr, "rucc: warning: {note}");
2948 }
2949 if verbose {
2950 let _ = write!(stderr, "{}", plan.render());
2951 let _ = writeln!(stderr, "workers: {}", jobs.count());
2952 }
2953 }
2954 if opts.emit == EmitKind::Preprocessed {
2955 return preprocess_all(&opts, &plan);
2956 }
2957 if opts.emit == EmitKind::Archive {
2958 return archive_all(&opts, &plan);
2959 }
2960 if opts.emit != EmitKind::Executable {
2961 return compile_all(&opts, &plan);
2962 }
2963 link_all(&opts, &plan, &link, verbose)
2964 }
2965 Err(e) => {
2966 let mut stderr = std::io::stderr().lock();
2967 let _ = writeln!(stderr, "rucc: error: {e}");
2968 let _ = writeln!(stderr, "rucc: note: run `rucc --help` for usage");
2969 1
2970 }
2971 }
2972}
2973
2974#[cfg(test)]
2975mod tests {
2976 use rucc_session::{
2977 Contract, GnucVersion, IncludeForm, LtoJobs, OptLevel, Partition, Patchable, Visibility,
2978 };
2979
2980 use super::*;
2981
2982 fn args(s: &[&str]) -> Vec<String> {
2983 s.iter().map(|x| (*x).to_owned()).collect()
2984 }
2985
2986 #[test]
2987 fn help_and_version_win_over_everything_else() {
2988 assert_eq!(parse_args(&args(&["-c", "--help", "x.c"])).unwrap(), Action::Help);
2989 assert_eq!(parse_args(&args(&["--version"])).unwrap(), Action::Version);
2990 }
2991
2992 fn compile(s: &[&str]) -> (Box<Options>, Box<Plan>) {
2993 match parse_args(&args(s)).expect("expected a compilation") {
2994 Action::Compile { opts, plan, .. } => (opts, plan),
2995 other => panic!("expected a compilation, got {other:?}"),
2996 }
2997 }
2998
2999 fn linking(s: &[&str]) -> (Box<LinkOptions>, Box<Plan>) {
3000 match parse_args(&args(s)).expect("expected a compilation") {
3001 Action::Compile { link, plan, .. } => (link, plan),
3002 other => panic!("expected a compilation, got {other:?}"),
3003 }
3004 }
3005
3006 fn notes(s: &[&str]) -> Vec<String> {
3007 match parse_args(&args(s)).expect("expected a compilation") {
3008 Action::Compile { notes, .. } => notes,
3009 other => panic!("expected a compilation, got {other:?}"),
3010 }
3011 }
3012
3013 /// The ordinary command line has nothing to say about itself, which is the property that makes
3014 /// a note worth reading when there is one.
3015 #[test]
3016 fn a_command_line_with_nothing_wrong_with_it_carries_no_notes() {
3017 assert_eq!(notes(&["-c", "a.c"]), Vec::<String>::new());
3018 }
3019
3020 /// A directory that is not there contributes nothing to the search path, so there is no tree to
3021 /// read a release out of and nothing to compare the pin against. Said as a test because this is
3022 /// the shape a hermetic machine takes: the probe reads the disk and every other machine has a
3023 /// different disk, so what can be asserted here is the silence.
3024 #[test]
3025 fn a_named_tree_that_is_not_on_the_machine_is_not_a_release_mismatch() {
3026 let said =
3027 notes(&["--target=x86_64-linux-gnu.2.28", "--sysroot=/nowhere-at-all", "-c", "a.c"]);
3028 assert_eq!(said, Vec::<String>::new());
3029 }
3030
3031 #[test]
3032 fn collects_inputs_and_flags() {
3033 let (opts, plan) = compile(&["-c", "-O2", "-g", "a.c", "b.c"]);
3034 let paths: Vec<&str> = plan.jobs.iter().map(|j| j.input.as_str()).collect();
3035 assert_eq!(paths, vec!["a.c", "b.c"]);
3036 assert_eq!(opts.opt_level, OptLevel::O2);
3037 assert_eq!(opts.emit, EmitKind::Object);
3038 assert!(opts.debug_info);
3039 }
3040
3041 /// The unstable options, which are spelled apart from everything else on purpose: what is
3042 /// under `-Z` promises nothing, and a build that reaches for one should have had to say so.
3043 #[test]
3044 fn an_unstable_option_is_taken_and_one_that_does_not_exist_is_refused() {
3045 let (opts, _) = compile(&["-c", "-Zrule-coverage=/tmp/rules.cov", "a.c"]);
3046 assert_eq!(opts.rule_coverage.as_deref(), Some("/tmp/rules.cov"));
3047
3048 let (plain, _) = compile(&["-c", "a.c"]);
3049 assert_eq!(plain.rule_coverage, None, "nothing is measured unless it was asked for");
3050
3051 assert!(parse_args(&args(&["-Zrule-coverage=", "a.c"])).is_err(), "a file with no name");
3052 let unknown = parse_args(&args(&["-Zwhat", "a.c"])).expect_err("there is no such option");
3053 assert!(unknown.message.contains("4.11"), "{}", unknown.message);
3054 }
3055
3056 /// The other measurement written to a file, which reads the same way and fails the same way.
3057 #[test]
3058 fn where_the_register_pressure_goes_is_asked_for_the_same_way() {
3059 let (opts, _) = compile(&["-c", "-O2", "-Zregister-pressure=/tmp/spills.txt", "a.c"]);
3060 assert_eq!(opts.register_pressure.as_deref(), Some("/tmp/spills.txt"));
3061
3062 let (plain, _) = compile(&["-c", "a.c"]);
3063 assert_eq!(plain.register_pressure, None, "nothing is measured unless it was asked for");
3064
3065 assert!(parse_args(&args(&["-Zregister-pressure=", "a.c"])).is_err(), "no file named");
3066 }
3067
3068 /// The third one, which says what the pre-selection lowering group did.
3069 #[test]
3070 fn where_the_lowering_dump_goes_is_asked_for_the_same_way() {
3071 let (opts, _) = compile(&["-c", "-O2", "-Zlowering=/tmp/lowering.txt", "a.c"]);
3072 assert_eq!(opts.lowering_dump.as_deref(), Some("/tmp/lowering.txt"));
3073
3074 let (plain, _) = compile(&["-c", "a.c"]);
3075 assert_eq!(plain.lowering_dump, None, "nothing is dumped unless it was asked for");
3076
3077 assert!(parse_args(&args(&["-Zlowering=", "a.c"])).is_err(), "no file named");
3078 }
3079
3080 /// Scheduling, which has the three way answer every optimization flag has: on, off, and
3081 /// nothing said, which is whatever the optimization level asks for. The name is gcc's, and
3082 /// gcc's has a two in it because gcc has a scheduler before allocation and one after and this
3083 /// is the one after.
3084 #[test]
3085 fn scheduling_can_be_turned_on_and_off_and_left_to_the_optimization_level() {
3086 let (on, _) = compile(&["-c", "-O0", "-fschedule-insns2", "a.c"]);
3087 assert_eq!(on.schedule_insns, Some(true));
3088
3089 let (off, _) = compile(&["-c", "-O2", "-fno-schedule-insns2", "a.c"]);
3090 assert_eq!(off.schedule_insns, Some(false));
3091
3092 let (quiet, _) = compile(&["-c", "-O2", "a.c"]);
3093 assert_eq!(quiet.schedule_insns, None, "nothing said, so the level decides");
3094 assert!(quiet.opt_level.schedules(), "and at this level the level says yes");
3095
3096 let (none, _) = compile(&["-c", "a.c"]);
3097 assert!(!none.opt_level.schedules(), "at no optimization it says no");
3098 }
3099
3100 /// Whether the timing model is worth holding an instruction back over, which is a `-Z` because
3101 /// it is a question about a target's description rather than about the program being compiled.
3102 #[test]
3103 fn whether_the_timing_model_is_cycle_accurate_can_be_overridden() {
3104 let (yes, _) = compile(&["-c", "-O2", "-Zcycle-accurate-model=yes", "a.c"]);
3105 assert_eq!(yes.cycle_accurate_model, Some(true));
3106
3107 let (no, _) = compile(&["-c", "-O2", "-Zcycle-accurate-model=no", "a.c"]);
3108 assert_eq!(no.cycle_accurate_model, Some(false));
3109
3110 let (plain, _) = compile(&["-c", "-O2", "a.c"]);
3111 assert_eq!(plain.cycle_accurate_model, None, "the target's own answer stands");
3112
3113 let bad = parse_args(&args(&["-Zcycle-accurate-model=maybe", "a.c"]))
3114 .expect_err("it takes yes or no");
3115 assert!(bad.message.contains("yes or no"), "{}", bad.message);
3116 }
3117
3118 #[test]
3119 fn a_bare_dash_o_means_o1_the_way_gcc_reads_it() {
3120 let (opts, _) = compile(&["-O", "a.c"]);
3121 assert_eq!(opts.opt_level, OptLevel::O1);
3122 }
3123
3124 #[test]
3125 fn dash_x_applies_to_later_inputs_only_and_none_stops_it() {
3126 let (_, plan) = compile(&["a.o", "-x", "c", "b.txt", "-x", "none", "c.o"]);
3127 assert_eq!(plan.jobs[0].kind, InputKind::LinkerInput);
3128 assert_eq!(plan.jobs[1].kind, InputKind::C);
3129 assert_eq!(plan.jobs[2].kind, InputKind::LinkerInput);
3130 }
3131
3132 #[test]
3133 fn dash_j_reaches_the_scheduler_and_defaults_to_the_machine() {
3134 let (_, _, jobs) = match parse_args(&args(&["-j4", "a.c"])).unwrap() {
3135 Action::Compile { opts, plan, jobs, .. } => (opts, plan, jobs),
3136 other => panic!("expected a compilation, got {other:?}"),
3137 };
3138 assert_eq!(jobs.count(), 4);
3139
3140 let default = match parse_args(&args(&["a.c"])).unwrap() {
3141 Action::Compile { jobs, .. } => jobs,
3142 other => panic!("expected a compilation, got {other:?}"),
3143 };
3144 assert_eq!(default, Jobs::available());
3145 assert!(parse_args(&args(&["-j0", "a.c"])).is_err());
3146 }
3147
3148 #[test]
3149 fn triple_hash_prints_the_plan_and_runs_nothing() {
3150 let a = parse_args(&args(&["-###", "-c", "a.c"])).unwrap();
3151 let Action::PrintPlan { plan, .. } = a else { panic!("expected a plan dump") };
3152 assert!(plan.render().contains("a.c: preprocess, compile, assemble -> a.o"));
3153 }
3154
3155 #[test]
3156 fn the_flag_that_keeps_the_intermediate_files_has_three_spellings_and_two_meanings() {
3157 // The bare one is `=obj` and not `=cwd`. gcc's manual says the opposite and gcc 16 does
3158 // this, and following the compiler is what makes a build that reads either of them find
3159 // the files where they are.
3160 assert_eq!(compile(&["-c", "-save-temps", "a.c"]).0.save_temps, SaveTemps::Object);
3161 assert_eq!(compile(&["-c", "-save-temps=obj", "a.c"]).0.save_temps, SaveTemps::Object);
3162 assert_eq!(compile(&["-c", "-save-temps=cwd", "a.c"]).0.save_temps, SaveTemps::Cwd);
3163 assert_eq!(compile(&["-c", "a.c"]).0.save_temps, SaveTemps::No);
3164 // The last one on the line decides, the way it does for every other flag with an
3165 // argument, and a keyword that is neither is fatal rather than ignored: a run that kept
3166 // nothing and said nothing looks exactly like one where the files were not produced.
3167 let (opts, _) = compile(&["-c", "-save-temps", "-save-temps=cwd", "a.c"]);
3168 assert_eq!(opts.save_temps, SaveTemps::Cwd);
3169 let e = parse_args(&args(&["-c", "-save-temps=nowhere", "a.c"])).unwrap_err();
3170 assert!(e.message.contains("accepted: cwd, obj"), "{}", e.message);
3171 }
3172
3173 #[test]
3174 fn the_flag_that_times_each_step_reaches_the_options_and_changes_nothing_else() {
3175 let (opts, plan) = compile(&["-c", "-time", "a.c"]);
3176 let (plain, without) = compile(&["-c", "a.c"]);
3177 assert!(opts.time);
3178 assert!(!plain.time);
3179 // Against the same line without the flag rather than against a spelling of the object's
3180 // name, since what the object is called is the host's business and this is not about that.
3181 assert_eq!(plan.jobs[0].output, without.jobs[0].output);
3182 }
3183
3184 #[test]
3185 fn dash_x_names_what_it_accepts_when_it_does_not_know_a_language() {
3186 let e = parse_args(&args(&["-x", "fortran", "a.c"])).unwrap_err();
3187 assert!(e.message.contains("assembler-with-cpp"), "{}", e.message);
3188 }
3189
3190 /// What `--fetch` says while the table in [`artifact`] has no rows in it, which is what every
3191 /// run of it says today and is the reason the message distinguishes the two cases.
3192 #[test]
3193 fn a_fetch_of_a_target_nothing_is_pinned_for_says_so_rather_than_reaching_the_network() {
3194 let e = parse_args(&args(&["--fetch", "x86_64-linux-musl"])).unwrap_err();
3195 assert!(e.message.contains("pins no sysroot for x86_64-linux-musl"), "{}", e.message);
3196 // And where one comes from, because the answer is not on this machine.
3197 assert!(e.message.contains("tamnd/rucc-cross"), "{}", e.message);
3198 // The joined spelling is the same flag.
3199 let joined = parse_args(&args(&["--fetch=x86_64-linux-musl"])).unwrap_err();
3200 assert_eq!(joined, e);
3201 }
3202
3203 /// The two targets a release will never pin, which is a different answer from the one above.
3204 ///
3205 /// Section 13.4. A person who reads "this release pins no sysroot yet" waits for a release that
3206 /// does, and no release of this compiler can ship either of these, so the message names the
3207 /// licence that decides it and what to do instead.
3208 #[test]
3209 fn a_fetch_of_a_target_behind_a_licence_wall_says_so_rather_than_saying_not_yet() {
3210 let e = parse_args(&args(&["--fetch", "aarch64-macos"])).unwrap_err();
3211 assert!(e.message.contains("Xcode licence"), "{}", e.message);
3212 assert!(e.message.contains("there never will be"), "{}", e.message);
3213 assert!(!e.message.contains("tamnd/rucc-cross"), "{}", e.message);
3214
3215 let e = parse_args(&args(&["--fetch", "x86_64-windows-msvc"])).unwrap_err();
3216 assert!(e.message.contains("redistributed"), "{}", e.message);
3217 // The way out of this one is a target rather than a download, and it is the default already.
3218 assert!(e.message.contains("mingw-w64"), "{}", e.message);
3219 // And the mingw-w64 target next to it is an ordinary unpinned target.
3220 let e = parse_args(&args(&["--fetch", "x86_64-windows-gnu"])).unwrap_err();
3221 assert!(e.message.contains("pins no sysroot"), "{}", e.message);
3222 }
3223
3224 /// An Apple target on a machine with no SDK, which is section 8.6's other host.
3225 ///
3226 /// Not run on a mac, where the SDK this is about is installed and the compile is the ordinary one
3227 /// that uses it. What the reason says is asserted in `rucc_sysroot::wall` and where it is printed
3228 /// is asserted in `rucc-pp`, so what is left here is that the driver works it out and leaves it
3229 /// where the preprocessor will find it, and that neither way past the wall leaves one behind.
3230 #[test]
3231 fn an_apple_target_with_no_sdk_anywhere_carries_the_licence_rather_than_a_missing_directory() {
3232 if cfg!(target_os = "macos") || std::env::var_os("SDKROOT").is_some() {
3233 return;
3234 }
3235 let (opts, _) = compile(&["--target=aarch64-macos", "-c", "a.c"]);
3236 let why = opts.search.missing_system().expect("the wall is the reason there are none");
3237 assert!(why.contains("aarch64-macos needs a macOS SDK"), "{why}");
3238 assert!(why.contains("Xcode licence"), "{why}");
3239 assert!(why.contains("-isysroot"), "{why}");
3240
3241 // A program that includes none of the library needs none of the SDK, which is what section
3242 // 8.6 means by being able to target the platform without one, so there is nothing to explain.
3243 let (opts, _) = compile(&["--target=aarch64-macos", "-nostdinc", "-c", "a.c"]);
3244 assert_eq!(opts.search.missing_system(), None);
3245 // And naming a path is the other way through, whether or not the path is there: a mistyped
3246 // directory is a mistake to report on its own terms rather than a licence to explain.
3247 let (opts, _) = compile(&["--target=aarch64-macos", "-isysroot", "/opt/sdk", "-c", "a.c"]);
3248 assert_eq!(opts.search.missing_system(), None);
3249 }
3250
3251 /// The same wall on the compile side of an MSVC target, where the way past it is a tuple.
3252 #[test]
3253 fn an_msvc_target_with_no_sdk_named_says_which_environment_needs_nothing_installed() {
3254 if std::env::var_os("INCLUDE").is_some() {
3255 return;
3256 }
3257 let (opts, _) = compile(&["--target=x86_64-windows-msvc", "-c", "a.c"]);
3258 let why = opts.search.missing_system().expect("the wall is the reason there are none");
3259 assert!(why.contains("the Windows SDK and its universal CRT"), "{why}");
3260 assert!(why.contains("mingw-w64"), "{why}");
3261 // And the mingw-w64 target has its headers from us, so nothing is missing to explain.
3262 let (opts, _) = compile(&["--target=x86_64-windows-gnu", "-c", "a.c"]);
3263 assert_eq!(opts.search.missing_system(), None);
3264 }
3265
3266 #[test]
3267 fn a_fetch_with_no_target_and_a_fetch_of_a_tuple_that_is_not_one_both_say_which() {
3268 let e = parse_args(&args(&["--fetch"])).unwrap_err();
3269 assert!(e.message.contains("--fetch requires"), "{}", e.message);
3270 let e = parse_args(&args(&["--fetch", "sparc64-solaris-gnu"])).unwrap_err();
3271 assert!(e.message.contains("--fetch sparc64-solaris-gnu"), "{}", e.message);
3272 assert!(e.message.contains("no sysroot to get"), "{}", e.message);
3273 }
3274
3275 /// Both flags on one line ask for opposite things, in either order.
3276 #[test]
3277 fn a_fetch_and_offline_together_is_a_refusal_whichever_way_round_they_are_written() {
3278 for line in [
3279 vec!["--offline", "--fetch", "x86_64-linux-musl"],
3280 vec!["--fetch", "x86_64-linux-musl", "--offline"],
3281 ] {
3282 let e = parse_args(&args(&line)).unwrap_err();
3283 assert!(e.message.contains("two opposite things"), "{}", e.message);
3284 }
3285 }
3286
3287 #[test]
3288 fn a_fetch_does_not_compile_anything_and_says_so_when_it_is_handed_a_file() {
3289 let e = parse_args(&args(&["--fetch", "x86_64-linux-musl", "a.c"])).unwrap_err();
3290 assert!(e.message.contains("compiles nothing"), "{}", e.message);
3291 assert!(e.message.contains("a.c"), "{}", e.message);
3292 }
3293
3294 /// `--offline` on its own is accepted and changes nothing, because an ordinary compile
3295 /// downloads nothing with or without it. A build that passes it everywhere is the case this is
3296 /// for, and it must not lose the compilation it was passed beside.
3297 #[test]
3298 fn offline_on_a_compilation_is_the_same_compilation() {
3299 let (opts, plan) = compile(&["-c", "--offline", "a.c"]);
3300 let (plain, without) = compile(&["-c", "a.c"]);
3301 assert_eq!(opts.target, plain.target);
3302 assert_eq!(plan.jobs.len(), without.jobs.len());
3303 assert_eq!(plan.jobs[0].output, without.jobs[0].output);
3304 }
3305
3306 #[test]
3307 fn an_unknown_flag_is_an_error_rather_than_a_shrug() {
3308 let e = parse_args(&args(&["-fno-such-thing", "a.c"])).unwrap_err();
3309 assert!(e.message.contains("unknown option"), "{}", e.message);
3310 }
3311
3312 /// `-fpermissive` and the flag that turns it back off, which a build writes beside it when
3313 /// one directory needs the older rules and the rest of the tree does not.
3314 #[test]
3315 fn permissive_reads_in_both_directions_and_the_last_one_wins() {
3316 let (opts, _) = compile(&["-c", "a.c"]);
3317 assert!(!opts.permissive, "off unless it is asked for");
3318
3319 let (opts, _) = compile(&["-c", "-fpermissive", "a.c"]);
3320 assert!(opts.permissive);
3321
3322 let (opts, _) = compile(&["-c", "-fpermissive", "-fno-permissive", "a.c"]);
3323 assert!(!opts.permissive);
3324 }
3325
3326 #[test]
3327 fn asking_for_nested_functions_is_told_why_it_is_not_coming() {
3328 let e = parse_args(&args(&["-fnested-functions", "a.c"])).unwrap_err();
3329 assert!(e.message.contains("trampoline"), "{}", e.message);
3330 assert!(parse_args(&args(&["-fno-nested-functions", "a.c"])).is_ok());
3331 }
3332
3333 #[test]
3334 fn the_flag_every_configure_script_writes_is_taken() {
3335 // All four spellings, because a build writes whichever one its macros picked and a
3336 // compiler that takes three of them is a compiler that fails on the fourth.
3337 for flag in ["-fPIC", "-fpic", "-fPIE", "-fpie"] {
3338 let (opts, _) = compile(&["-c", flag, "a.c"]);
3339 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3340 }
3341 }
3342
3343 #[test]
3344 fn a_table_is_written_unless_the_build_says_nothing_will_walk_it() {
3345 let (opts, _) = compile(&["-c", "a.c"]);
3346 assert!(opts.unwinds(), "the default is off");
3347 let (opts, _) = compile(&["-c", "-fno-asynchronous-unwind-tables", "a.c"]);
3348 assert!(!opts.unwinds(), "the build was not taken at its word");
3349 let (opts, _) = compile(&[
3350 "-c",
3351 "-fno-asynchronous-unwind-tables",
3352 "-fasynchronous-unwind-tables",
3353 "a.c",
3354 ]);
3355 assert!(opts.unwinds(), "the last flag did not win");
3356 // The weaker request, which the same table answers, so a line that asks for a table and
3357 // against an asynchronous one gets one. That is gcc's arrangement and it turns up when a
3358 // build turns the asynchronous one off globally and a directory asks for a table back.
3359 let (opts, _) =
3360 compile(&["-c", "-fno-asynchronous-unwind-tables", "-funwind-tables", "a.c"]);
3361 assert!(opts.unwinds(), "the weaker request was dropped");
3362 let (opts, _) = compile(&["-c", "-fno-unwind-tables", "a.c"]);
3363 assert!(opts.unwinds(), "the weaker negative turned off the stronger request");
3364 let (opts, _) =
3365 compile(&["-c", "-fno-unwind-tables", "-fno-asynchronous-unwind-tables", "a.c"]);
3366 assert!(!opts.unwinds(), "both were turned off and one stayed on");
3367 }
3368
3369 #[test]
3370 fn the_flags_that_describe_what_this_compiler_already_does_are_taken() {
3371 // Every one of these is on a real build line somewhere and every one of them was an
3372 // unknown option. What they have in common is that the answer rucc gives is the answer
3373 // they ask for, so there is nothing to implement and nothing to refuse.
3374 for flag in [
3375 "-fno-common",
3376 "-fstrict-aliasing",
3377 "-fno-strict-aliasing",
3378 "-fdelete-null-pointer-checks",
3379 "-fno-delete-null-pointer-checks",
3380 "-frounding-math",
3381 "-fno-rounding-math",
3382 "-fexcess-precision=standard",
3383 "-fexcess-precision=fast",
3384 "-fexcess-precision=16",
3385 "-pipe",
3386 "-fdiagnostics-color",
3387 "-fno-diagnostics-color",
3388 "-fdiagnostics-color=always",
3389 "-fdiagnostics-color=never",
3390 "-fdiagnostics-color=auto",
3391 ] {
3392 let (opts, _) = compile(&["-c", flag, "a.c"]);
3393 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3394 }
3395 }
3396
3397 #[test]
3398 fn whether_an_exception_is_looked_at_is_kept_and_defaults_to_gccs_answer() {
3399 let (opts, _) = compile(&["-c", "a.c"]);
3400 assert!(opts.trapping_math, "the default was not gcc's");
3401 let (opts, _) = compile(&["-c", "-fno-trapping-math", "a.c"]);
3402 assert!(!opts.trapping_math);
3403 let (opts, _) = compile(&["-c", "-ftrapping-math", "a.c"]);
3404 assert!(opts.trapping_math, "spelling out the default turned it off");
3405 // The last one written wins, which is how a build line that inherits a flag from one
3406 // place and overrides it in another is read.
3407 let (opts, _) = compile(&["-c", "-fno-trapping-math", "-ftrapping-math", "a.c"]);
3408 assert!(opts.trapping_math);
3409 }
3410
3411 /// The flags a torture program writes on its own `dg-options` line, which is where most of
3412 /// these come from: a program reduced from a miscompilation names the pass that miscompiled
3413 /// it. Eighteen programs in the suite stopped on the driver before anything read them, and
3414 /// tamnd/rucc#1019 is the list.
3415 #[test]
3416 fn the_flags_that_name_a_pass_of_gccs_own_are_taken_and_dropped() {
3417 for flag in [
3418 "-fno-tree-ccp",
3419 "-fno-tree-dominator-opts",
3420 "-fno-tree-vrp",
3421 "-fno-tree-bit-ccp",
3422 "-fno-tree-coalesce-vars",
3423 "-ftree-vectorize",
3424 "-ftree-loop-distribution",
3425 "-fno-ipa-cp",
3426 "-fipa-pta",
3427 "-fmodulo-sched",
3428 "-fno-vect-cost-model",
3429 "-fvect-cost-model=unlimited",
3430 "-fsimd-cost-model=cheap",
3431 "-fexpensive-optimizations",
3432 "-fno-early-inlining",
3433 "-fno-inline",
3434 "-finline-functions",
3435 "-foptimize-strlen",
3436 "-fno-ira-share-spill-slots",
3437 ] {
3438 let (opts, _) = compile(&["-c", flag, "a.c"]);
3439 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3440 assert!(opts.passes.is_empty(), "{flag} named a pass of gcc's and not one of ours");
3441 }
3442 }
3443
3444 /// The two namespaces are taken whole, so a name neither this test nor gcc 16 has heard of
3445 /// goes the same way as the ones above rather than stopping a build on the day gcc adds it.
3446 #[test]
3447 fn a_pass_name_in_either_family_is_taken_whether_or_not_it_is_one_gcc_has() {
3448 for flag in ["-ftree-no-such-pass", "-fno-ipa-no-such-pass"] {
3449 let (opts, _) = compile(&["-c", flag, "a.c"]);
3450 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3451 }
3452 }
3453
3454 /// A pass this compiler has keeps its flag, since the arms that read the registry are above
3455 /// the family arms. `dce` is the one both compilers have a name for, and `execute/pr97421-2.c`
3456 /// is the program that writes it.
3457 #[test]
3458 fn a_pass_name_this_compiler_has_is_still_read_as_a_pass() {
3459 let (opts, _) = compile(&["-c", "-fno-dce", "a.c"]);
3460 assert_eq!(opts.passes, vec![("dce".to_owned(), false)]);
3461 }
3462
3463 /// gcc's name for the unroller reaches the unroller, in both directions. libtommath puts
3464 /// `-funroll-loops` in `CFLAGS` unconditionally, and before this it was an unknown option and
3465 /// the build stopped on its first file.
3466 #[test]
3467 fn the_gcc_spelling_of_the_unroller_turns_the_unroller_on_and_off() {
3468 let (opts, _) = compile(&["-c", "-funroll-loops", "a.c"]);
3469 assert_eq!(opts.passes, vec![("unroll".to_owned(), true)]);
3470 let (opts, _) = compile(&["-c", "-fno-unroll-loops", "a.c"]);
3471 assert_eq!(opts.passes, vec![("unroll".to_owned(), false)]);
3472 }
3473
3474 /// The encoding of the source is not a question about speed, so the one name that describes
3475 /// what the preprocessor does is taken and every other name is refused.
3476 #[test]
3477 fn the_input_charset_is_taken_when_it_names_the_one_that_is_read() {
3478 for flag in ["-finput-charset=utf-8", "-finput-charset=UTF-8", "-finput-charset=utf8"] {
3479 let (opts, _) = compile(&["-c", flag, "a.c"]);
3480 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3481 }
3482
3483 let e = parse_args(&args(&["-c", "-finput-charset=latin1", "a.c"])).unwrap_err();
3484 assert!(e.message.contains("latin1"), "{}", e.message);
3485 assert!(e.message.contains("UTF-8"), "what is read is worth saying: {}", e.message);
3486 }
3487
3488 /// The other half of the same rule. Each of these changes what the program does rather than
3489 /// how fast it does it, so each is refused with the reason, and the negative of each is what
3490 /// happens anyway and is taken.
3491 #[test]
3492 fn the_three_that_change_the_answer_are_refused_and_their_negatives_are_taken() {
3493 for (flag, word) in [
3494 ("-ffast-math", "__FAST_MATH__"),
3495 ("-fnon-call-exceptions", "landing pad"),
3496 ("-finstrument-functions", "__cyg_profile_func_enter"),
3497 ] {
3498 let e = parse_args(&args(&["-c", flag, "a.c"])).unwrap_err();
3499 assert!(e.message.contains(word), "{flag}: {}", e.message);
3500 assert!(!e.message.contains("unknown option"), "{flag} deserves a reason");
3501
3502 let off = format!("-fno-{}", flag.trim_start_matches("-f"));
3503 let (opts, _) = compile(&["-c", &off, "a.c"]);
3504 assert_eq!(opts.emit, EmitKind::Object, "{off}");
3505 }
3506 }
3507
3508 #[test]
3509 fn asking_the_linker_to_merge_tentative_definitions_is_told_why_it_is_not_coming() {
3510 // The one of that family that is a request rather than a description, and it is a real
3511 // difference: two files each writing `int g;` link under it and do not without it.
3512 let e = parse_args(&args(&["-fcommon", "a.c"])).unwrap_err();
3513 assert!(e.message.contains(".bss"), "{}", e.message);
3514 assert!(e.message.contains("extern"), "the way out is worth saying: {}", e.message);
3515 }
3516
3517 #[test]
3518 fn asking_for_position_dependent_code_is_told_why_it_is_not_coming() {
3519 for flag in ["-fno-pic", "-fno-pie"] {
3520 let e = parse_args(&args(&[flag, "a.c"])).unwrap_err();
3521 assert!(e.message.contains("global offset table"), "{flag}: {}", e.message);
3522 // The one it may have meant, since the two are a letter apart and one of them is
3523 // about linking and is taken.
3524 assert!(e.message.contains("-no-pie"), "{flag}: {}", e.message);
3525 }
3526 }
3527
3528 #[test]
3529 fn an_unsupported_target_names_itself() {
3530 let e = parse_args(&args(&["--target=sparc64-linux-gnu", "a.c"])).unwrap_err();
3531 assert!(e.message.contains("sparc64"), "{}", e.message);
3532 }
3533
3534 #[test]
3535 fn no_inputs_is_an_error_but_print_config_needs_none() {
3536 assert!(parse_args(&args(&[])).is_err());
3537 assert!(matches!(parse_args(&args(&["--print-config"])), Ok(Action::PrintConfig(_))));
3538 }
3539
3540 #[test]
3541 fn print_config_reports_the_target_it_was_given_not_the_host() {
3542 let a = parse_args(&args(&["--print-config", "--target=riscv64-linux-musl"])).unwrap();
3543 let Action::PrintConfig(opts) = a else { panic!("expected a configuration dump") };
3544 let text = print_config(&opts);
3545 assert!(text.contains("target: riscv64-unknown-linux-musl"), "{text}");
3546 assert!(text.contains("char-signed: false"), "{text}");
3547 assert!(text.contains("object-format: elf"), "{text}");
3548 assert!(text.contains("va-list: void-pointer"), "{text}");
3549 // RISC-V has a register file and this compiler has not written it down yet, and the
3550 // dump says which of those two it is rather than leaving the line out.
3551 assert!(text.contains("registers: none"), "{text}");
3552 assert!(text.contains("timing-model: none"), "{text}");
3553 }
3554
3555 /// The model the schedule was chosen with, which is a receipt anybody comparing two runs of a
3556 /// benchmark needs: two numbers that disagree are usually two models and not two compilers.
3557 #[test]
3558 fn print_config_names_the_model_the_schedule_was_chosen_with() {
3559 let opts = Options::new("x86_64-unknown-linux-gnu".parse().unwrap());
3560 let text = print_config(&opts);
3561 let line = text.lines().find(|l| l.starts_with("timing-model:")).expect("the model");
3562 assert!(line.contains("Skylake"), "{line}");
3563 assert!(line.contains("published"), "a sentence saying where it came from: {line}");
3564 }
3565
3566 #[test]
3567 fn print_config_has_one_key_per_line_and_a_fixed_order() {
3568 let opts = Options::new("x86_64-unknown-linux-gnu".parse().unwrap());
3569 let text = print_config(&opts);
3570 let keys: Vec<&str> =
3571 text.lines().map(|l| l.split(':').next().unwrap_or_default()).collect();
3572 assert_eq!(keys[0], "version");
3573 assert_eq!(keys[1], "target");
3574 assert_eq!(keys.len(), 26);
3575 assert!(text.ends_with('\n'));
3576 }
3577
3578 #[test]
3579 fn the_safety_tier_is_read_off_the_command_line_and_a_wrong_one_is_refused() {
3580 let (opts, _) = compile(&["a.c"]);
3581 assert_eq!(opts.safety, rucc_session::Safety::Off);
3582
3583 for (flag, tier) in [
3584 ("-fsafety=detect", rucc_session::Safety::Detect),
3585 ("-fsafety=enforce", rucc_session::Safety::Enforce),
3586 ("-fsafety=kernel", rucc_session::Safety::Kernel),
3587 ("-fsafety=off", rucc_session::Safety::Off),
3588 ] {
3589 let (opts, _) = compile(&[flag, "a.c"]);
3590 assert_eq!(opts.safety, tier, "{flag}");
3591 }
3592
3593 // The last one wins, the way every other repeated flag on this command line does.
3594 let (opts, _) = compile(&["-fsafety=enforce", "-fsafety=off", "a.c"]);
3595 assert_eq!(opts.safety, rucc_session::Safety::Off);
3596
3597 // A misspelled tier is refused rather than ignored. Silently compiling without the
3598 // monitor a build asked for is the one failure mode this feature cannot have.
3599 let e = parse_args(&args(&["-fsafety=on", "a.c"])).unwrap_err();
3600 assert!(e.message.contains("is not a safety tier"), "{}", e.message);
3601 assert!(parse_args(&args(&["-fsafety", "a.c"])).is_err());
3602 }
3603
3604 #[test]
3605 fn the_padding_mode_is_read_off_the_command_line_and_a_wrong_one_is_refused() {
3606 // The default is the one section 9.3 of document 09 gives library code, which is that
3607 // padding does not participate, so a record filled a member at a time is not reported.
3608 let (opts, _) = compile(&["a.c"]);
3609 assert_eq!(opts.padding, rucc_session::Padding::Ignored);
3610
3611 let (opts, _) = compile(&["-fsafety=detect", "-fsafety-init=padding", "a.c"]);
3612 assert_eq!(opts.padding, rucc_session::Padding::Tracked);
3613
3614 let (opts, _) = compile(&["-fsafety-init=padding", "-fsafety-init=nopadding", "a.c"]);
3615 assert_eq!(opts.padding, rucc_session::Padding::Ignored);
3616
3617 // The tier is still a tier. A flag whose name starts the same way must not be eaten by
3618 // the one above it, which is the thing worth pinning about a pair of names like these.
3619 let (opts, _) = compile(&["-fsafety-init=padding", "a.c"]);
3620 assert_eq!(opts.safety, rucc_session::Safety::Off);
3621
3622 let e = parse_args(&args(&["-fsafety-init=some", "a.c"])).unwrap_err();
3623 assert!(e.message.contains("is not a padding mode"), "{}", e.message);
3624 }
3625
3626 #[test]
3627 fn whether_a_write_has_to_stay_inside_its_member_is_read_off_the_command_line() {
3628 // Off by default, because a store to allocated storage sets its effective type and C 6.5
3629 // lets a program reuse a buffer as something else. Row S4 is a build opting out of that.
3630 let (opts, _) = compile(&["a.c"]);
3631 assert_eq!(opts.subobject, rucc_session::Subobject::Off);
3632
3633 let (opts, _) = compile(&["-fsafety=detect", "-fsafety-subobject", "a.c"]);
3634 assert_eq!(opts.subobject, rucc_session::Subobject::Members);
3635
3636 let (opts, _) = compile(&["-fsafety-subobject", "-fno-safety-subobject", "a.c"]);
3637 assert_eq!(opts.subobject, rucc_session::Subobject::Off);
3638
3639 // It takes no value. The form that would take one is the strict reading of section 9.4,
3640 // which is not written yet, so say so rather than accept a spelling that does nothing.
3641 let e = parse_args(&args(&["-fsafety-subobject=strict", "a.c"])).unwrap_err();
3642 assert!(e.message.contains("tamnd/rucc#967"), "{}", e.message);
3643 }
3644
3645 #[test]
3646 fn whether_two_restrict_pointers_may_meet_is_read_off_the_command_line() {
3647 // Off by default, because the record a block keeps is the union of what each pointer
3648 // reached, so two pointers striding through one array without landing on the same byte are
3649 // reported and by the letter of the standard those are different objects. Row Y8 is a build
3650 // deciding it would rather know.
3651 let (opts, _) = compile(&["a.c"]);
3652 assert_eq!(opts.promise, rucc_session::Promise::Off);
3653
3654 let (opts, _) = compile(&["-fsafety=detect", "-fsafety-restrict", "a.c"]);
3655 assert_eq!(opts.promise, rucc_session::Promise::Blocks);
3656
3657 let (opts, _) = compile(&["-fsafety-restrict", "-fno-safety-restrict", "a.c"]);
3658 assert_eq!(opts.promise, rucc_session::Promise::Off);
3659
3660 // The tier is still a tier, which is the thing worth pinning about a pair of names where
3661 // one is the front of the other.
3662 let (opts, _) = compile(&["-fsafety-restrict", "a.c"]);
3663 assert_eq!(opts.safety, rucc_session::Safety::Off);
3664
3665 let e = parse_args(&args(&["-fsafety-restrict=blocks", "a.c"])).unwrap_err();
3666 assert!(e.message.contains("takes no value"), "{}", e.message);
3667 }
3668
3669 #[test]
3670 fn safety_races_takes_a_mode_and_defaults_to_watching_nothing() {
3671 // Three modes rather than a bare flag, because section 9.5 gives two answers that record
3672 // the same thing and report different classes, so a flag with no value could not say which
3673 // was wanted. Off by default for the reason on `rucc_session::Races`, which is not a cost
3674 // argument: this is the one plane where an edge nobody interposed costs a false report.
3675 let (opts, _) = compile(&["a.c"]);
3676 assert_eq!(opts.races, rucc_session::Races::Off);
3677
3678 let (opts, _) = compile(&["-fsafety-races=metadata", "a.c"]);
3679 assert_eq!(opts.races, rucc_session::Races::Metadata);
3680
3681 let (opts, _) = compile(&["-fsafety-races=pointer", "a.c"]);
3682 assert_eq!(opts.races, rucc_session::Races::Pointer);
3683
3684 // Last one wins, as it does for every other mode flag here.
3685 let (opts, _) = compile(&["-fsafety-races=pointer", "-fno-safety-races", "a.c"]);
3686 assert_eq!(opts.races, rucc_session::Races::Off);
3687
3688 let e = parse_args(&args(&["-fsafety-races=all", "a.c"])).unwrap_err();
3689 assert!(e.message.contains("off, metadata or pointer"), "{}", e.message);
3690 }
3691
3692 #[test]
3693 fn print_pipeline_answers_with_the_passes_the_level_asked_for() {
3694 let a = parse_args(&args(&["--print-pipeline", "-O2"])).unwrap();
3695 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3696 let text = print_pipeline(&opts);
3697 assert!(text.starts_with("level: -O2\n"), "{text}");
3698 assert!(text.contains("fold"), "{text}");
3699
3700 let a = parse_args(&args(&["--print-pipeline"])).unwrap();
3701 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3702 // Two passes run at `-O0` and neither is an optimization. The first moves what
3703 // `__builtin_expect` said onto the branch and takes the instruction away, so that nothing
3704 // past the optimizer has to know the instruction exists. The second removes code nothing
3705 // reaches. See issue 359.
3706 assert!(print_pipeline(&opts).contains("1: expect,"), "{}", print_pipeline(&opts));
3707 assert!(print_pipeline(&opts).contains("2: simplify-cfg,"), "{}", print_pipeline(&opts));
3708
3709 let a = parse_args(&args(&["--print-pipeline", "-fno-simplify-cfg"])).unwrap();
3710 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3711 // The second turns off and the first does not, because nothing below the optimizer lowers
3712 // what it removes, so `-fno-expect` is a compile that stops rather than one that runs.
3713 let text = print_pipeline(&opts);
3714 assert!(text.contains("1: expect,"), "{text}");
3715 assert!(!text.contains("simplify-cfg"), "{text}");
3716 }
3717
3718 #[test]
3719 fn print_pipeline_takes_the_toggles_into_account() {
3720 let a = parse_args(&args(&["--print-pipeline", "-O2", "-fno-fold"])).unwrap();
3721 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3722 let text = print_pipeline(&opts);
3723 // The one that was named is gone and the rest of the level is not, which is the whole
3724 // of what a toggle promises.
3725 assert!(!text.contains("fold"), "{text}");
3726 assert!(text.contains("dce"), "{text}");
3727
3728 // Every pass the compiler has, named off. Built from the registry rather than written
3729 // out, so a pass added later is turned off here too and this keeps testing the thing it
3730 // is about, which is that the toggles can empty a level down to the passes that are not
3731 // optional. Those are named, because a listing that is all of them is a level nobody
3732 // emptied and the assertion would pass while saying nothing.
3733 let mut off = vec!["--print-pipeline".to_owned(), "-O2".to_owned()];
3734 off.extend(rucc_opt::PASSES.iter().map(|p| format!("-fno-{}", p.name())));
3735 let spelled: Vec<&str> = off.iter().map(String::as_str).collect();
3736 let a = parse_args(&args(&spelled)).unwrap();
3737 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3738 let text = print_pipeline(&opts);
3739 let left: Vec<&str> =
3740 rucc_opt::PASSES.iter().filter(|p| p.required()).map(|p| p.name()).collect();
3741 assert_eq!(left, vec!["expect"], "{text}");
3742 for (at, name) in left.iter().enumerate() {
3743 assert!(text.contains(&format!("{}: {name},", at + 1)), "{text}");
3744 }
3745 assert!(!text.contains("dce"), "{text}");
3746 }
3747
3748 #[test]
3749 fn print_pipeline_says_when_a_budget_will_stop_the_run_short() {
3750 let a = parse_args(&args(&["--print-pipeline", "-O2"])).unwrap();
3751 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3752 assert!(!print_pipeline(&opts).contains("global fuel"));
3753
3754 let a = parse_args(&args(&["--print-pipeline", "-O2", "-fpass-fuel-global=4"])).unwrap();
3755 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3756 let text = print_pipeline(&opts);
3757 // Because the listing is the answer to what this compilation will do, and a run that
3758 // stops after four rewrites is not doing what the level says it does.
3759 assert!(text.contains("global fuel: 4"), "{text}");
3760 }
3761
3762 /// A pass is turned on and off by its own name, and the order the flags were given in is
3763 /// kept, because the last spelling of a name is the one that decides.
3764 #[test]
3765 fn a_pass_is_named_by_dash_f_and_unnamed_by_dash_f_no() {
3766 let (opts, _) = compile(&["-c", "-O0", "-ffold", "-fno-fold", "-ffold", "a.c"]);
3767 assert_eq!(
3768 opts.passes,
3769 [("fold".to_owned(), true), ("fold".to_owned(), false), ("fold".to_owned(), true)]
3770 );
3771
3772 let e = parse_args(&args(&["-fno-such-pass", "a.c"])).unwrap_err();
3773 assert!(e.message.contains("unknown option"), "{}", e.message);
3774 }
3775
3776 #[test]
3777 fn pass_fuel_names_a_pass_and_a_count_and_refuses_anything_else() {
3778 let (opts, _) = compile(&["-c", "-O2", "-fpass-fuel=fold=3", "a.c"]);
3779 assert_eq!(opts.pass_fuel, [("fold".to_owned(), 3)]);
3780
3781 let e = parse_args(&args(&["-fpass-fuel=fold", "a.c"])).unwrap_err();
3782 assert!(e.message.contains("<pass>=<count>"), "{}", e.message);
3783 let e = parse_args(&args(&["-fpass-fuel=nosuch=3", "a.c"])).unwrap_err();
3784 assert!(e.message.contains("--print-pipeline"), "{}", e.message);
3785 let e = parse_args(&args(&["-fpass-fuel=fold=lots", "a.c"])).unwrap_err();
3786 assert!(e.message.contains("not a number"), "{}", e.message);
3787 }
3788
3789 #[test]
3790 fn global_pass_fuel_is_a_count_on_its_own_and_defaults_to_no_limit() {
3791 let (opts, _) = compile(&["-c", "-O2", "a.c"]);
3792 assert_eq!(opts.pass_fuel_global, None);
3793
3794 let (opts, _) = compile(&["-c", "-O2", "-fpass-fuel-global=12", "a.c"]);
3795 assert_eq!(opts.pass_fuel_global, Some(12));
3796 // And it is not the per pass flag with a longer name, so neither spelling swallows the
3797 // other.
3798 assert!(opts.pass_fuel.is_empty());
3799
3800 let e = parse_args(&args(&["-fpass-fuel-global=lots", "a.c"])).unwrap_err();
3801 assert!(e.message.contains("not a number"), "{}", e.message);
3802 }
3803
3804 #[test]
3805 fn a_gate_names_a_pass_and_optionally_the_functions_it_covers() {
3806 let (opts, _) = compile(&["-c", "-O2", "-fdisable-fold", "-fenable-fold=2-4,main", "a.c"]);
3807 assert_eq!(
3808 opts.pass_gates,
3809 [(false, "fold".to_owned()), (true, "fold=2-4,main".to_owned())],
3810 "the order is what decides, so it has to survive the parse"
3811 );
3812
3813 let e = parse_args(&args(&["-fdisable-nosuch", "a.c"])).unwrap_err();
3814 assert!(e.message.contains("--print-pipeline"), "{}", e.message);
3815 let e = parse_args(&args(&["-fenable-fold=9-2", "a.c"])).unwrap_err();
3816 assert!(e.message.contains("ends before it starts"), "{}", e.message);
3817 let e = parse_args(&args(&["-fdisable-fold=", "a.c"])).unwrap_err();
3818 assert!(e.message.contains("is empty"), "{}", e.message);
3819 }
3820
3821 #[test]
3822 fn the_pipeline_listing_says_which_passes_a_gate_touched() {
3823 let (opts, _) = compile(&["-c", "-O2", "-fdisable-fold=main", "a.c"]);
3824 let text = print_pipeline(&opts);
3825 assert!(text.contains("fold, "), "{text}");
3826 assert!(text.contains("[off for main]"), "{text}");
3827 }
3828
3829 /// The spelling is checked while the arguments are read, because a dump that names a pass
3830 /// this compiler does not have is a typo, and a typo found after the compilation has run is
3831 /// found too late to be any use.
3832 #[test]
3833 fn a_dump_is_checked_when_it_is_asked_for_rather_than_when_it_is_taken() {
3834 let (opts, _) = compile(&["-c", "-O2", "-fdump-ir=all", "-fdump-ir=after-fold", "a.c"]);
3835 assert_eq!(opts.dump_ir, ["all", "after-fold"]);
3836
3837 let e = parse_args(&args(&["-fdump-ir=after-nosuch", "a.c"])).unwrap_err();
3838 assert!(e.message.contains("nosuch"), "{}", e.message);
3839 assert!(parse_args(&args(&["-fdump-ir=sideways-fold", "a.c"])).is_err());
3840 }
3841
3842 /// Every spelling `-fopt-info` takes, and the one it does not.
3843 ///
3844 /// The keywords are checked here for the same reason a dump's pass name is: a person who
3845 /// misspelled one gets no output, and no output is also what a compilation where nothing
3846 /// happened looks like. Telling those two apart is the entire reason to reach for this flag.
3847 #[test]
3848 fn opt_info_takes_kinds_and_a_file_and_refuses_a_kind_it_does_not_have() {
3849 let (opts, _) = compile(&["-c", "-O2", "-fopt-info", "a.c"]);
3850 assert_eq!(opts.opt_info, [""], "a bare flag asks for the rewrites");
3851 assert_eq!(opts.opt_info_file, None, "and goes to standard error");
3852
3853 let (opts, _) = compile(&["-c", "-O2", "-fopt-info-missed-note", "a.c"]);
3854 assert_eq!(opts.opt_info, ["missed-note"]);
3855
3856 // Two flags add up rather than the second replacing the first, and the file is the last
3857 // one that named a file, which is how GCC treats both.
3858 let (opts, _) =
3859 compile(&["-c", "-O2", "-fopt-info-missed=one.txt", "-fopt-info-all=two.txt", "a.c"]);
3860 assert_eq!(opts.opt_info, ["missed", "all"]);
3861 assert_eq!(opts.opt_info_file.as_deref(), Some("two.txt"));
3862
3863 let e = parse_args(&args(&["-fopt-info-vectorized", "a.c"])).unwrap_err();
3864 assert!(e.message.contains("vectorized"), "{}", e.message);
3865 assert!(e.message.contains("`missed`"), "{}", e.message);
3866 let e = parse_args(&args(&["-fopt-info-missed=", "a.c"])).unwrap_err();
3867 assert!(e.message.contains("no file"), "{}", e.message);
3868 }
3869
3870 #[test]
3871 fn verify_each_is_unstable_and_off_unless_it_was_asked_for() {
3872 let (opts, _) = compile(&["-c", "-Zverify-each", "a.c"]);
3873 assert!(opts.verify_each);
3874 assert!(!USAGE.contains("verify-each"), "an unstable option stays out of the usage text");
3875 }
3876
3877 #[test]
3878 fn dash_o_needs_an_argument() {
3879 let e = parse_args(&args(&["a.c", "-o"])).unwrap_err();
3880 assert_eq!(e.message, "-o requires an argument");
3881 }
3882
3883 #[test]
3884 fn dash_d_and_dash_u_are_read_joined_or_separated_and_keep_their_order() {
3885 let (opts, _) = compile(&["-DFOO=1", "-D", "BAR", "-UBAZ", "-U", "QUX", "a.c"]);
3886 assert_eq!(opts.defines, ["FOO=1", "BAR"]);
3887 assert_eq!(opts.undefines, ["BAZ", "QUX"]);
3888 }
3889
3890 #[test]
3891 fn the_include_flags_land_on_the_chain_each_one_names() {
3892 // A sysroot with nothing under it, so that the library's own directories are the
3893 // same on every machine this test runs on, which is none of them.
3894 let (opts, _) = compile(&[
3895 "-Ii",
3896 "-iquote",
3897 "q",
3898 "-isystem",
3899 "sys",
3900 "-idirafter",
3901 "after",
3902 "--sysroot=/nowhere-at-all",
3903 "a.c",
3904 ]);
3905 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
3906 // The compiler's own headers sit after every `-isystem` and before `-idirafter`,
3907 // which is where GCC puts its own: a directory the user named outranks ours.
3908 assert_eq!(dirs, ["q", "i", "sys", runtime::DIR, "after"]);
3909 assert!(!opts.search.dirs()[1].is_system);
3910 assert!(opts.search.dirs()[2].is_system);
3911 }
3912
3913 #[test]
3914 fn the_librarys_headers_come_after_the_compilers_own_and_go_away_with_them() {
3915 // Which machine this runs on decides what is on the path, so the test is about the
3916 // order rather than about the names: ours is on it, the library's follow it, and
3917 // `-nostdinc` is the one flag that takes both halves of the pair off at once.
3918 let (opts, _) = compile(&["a.c"]);
3919 let dirs = opts.search.dirs();
3920 let ours = dirs.iter().position(|d| d.path.to_str() == Some(runtime::DIR));
3921 assert_eq!(ours, Some(0), "{dirs:?}");
3922 assert!(dirs[1..].iter().all(|d| d.is_system), "{dirs:?}");
3923 let (bare, _) = compile(&["-nostdinc", "a.c"]);
3924 assert!(bare.search.dirs().is_empty(), "{:?}", bare.search.dirs());
3925 }
3926
3927 #[test]
3928 fn a_sysroot_moves_the_librarys_directories_and_nothing_else() {
3929 let (opts, _) = compile(&["-isystem", "sys", "--sysroot=/nowhere-at-all", "a.c"]);
3930 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
3931 assert_eq!(dirs, ["sys", runtime::DIR]);
3932 }
3933
3934 #[test]
3935 fn a_cross_compile_reads_the_targets_own_headers_rather_than_the_ones_next_door() {
3936 // The target is not the machine this test runs on wherever it runs, so the answer is the
3937 // same on all of them: the libc's two include directories for that target, the kernel's
3938 // two, and nothing from here. A header read from here is the quiet failure of section 8.5, a
3939 // program that builds on the build machine and is wrong everywhere else.
3940 let (opts, _) = compile(&["--target=riscv64-linux-musl", "-c", "a.c"]);
3941 let dirs: Vec<&std::path::Path> =
3942 opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
3943 let root = cache::dir().join("sysroots").join("riscv64-linux-musl");
3944 let kernel = cache::dir().join("kernel-headers");
3945 assert_eq!(dirs.len(), 5, "{dirs:?}");
3946 assert_eq!(dirs[0], std::path::Path::new(runtime::DIR));
3947 assert_eq!(dirs[1], root.join("include").join("riscv64"));
3948 assert_eq!(dirs[2], root.join("include").join("generic"));
3949 // The kernel's, which are beside the sysroots rather than inside one, because every target
3950 // that shares an architecture reads the same files.
3951 assert_eq!(dirs[3], kernel.join("riscv"));
3952 assert_eq!(dirs[4], kernel.join("generic"));
3953 }
3954
3955 #[test]
3956 fn a_cross_compile_to_something_that_is_not_linux_reads_no_kernel_headers() {
3957 // The other side of the same answer. Windows has its own system headers and no `linux/` at
3958 // all, so the list is the libc's two and the question never arises, which is the `None` that
3959 // `link::cross_kernel` returns rather than a directory nothing would be found in.
3960 let (opts, _) = compile(&["--target=x86_64-pc-windows-gnu", "-c", "a.c"]);
3961 let dirs: Vec<&std::path::Path> =
3962 opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
3963 assert_eq!(dirs.len(), 3, "{dirs:?}");
3964 assert!(!dirs.iter().any(|dir| dir.ends_with("kernel-headers")), "{dirs:?}");
3965 }
3966
3967 #[test]
3968 fn the_glibc_version_macro_goes_with_the_bundled_tree_and_with_nothing_else() {
3969 // One tree serves every glibc release, so the release is what the target supplies, and the
3970 // condition is the same one that chose the directories. A host glibc and a tree somebody
3971 // named both define `__GLIBC_MINOR__` in their own `features.h`, and two definitions with
3972 // different values is a warning on every compilation of every file.
3973 //
3974 // The architecture is chosen against this machine's rather than written down, because the
3975 // bundled tree is only in effect for a target that is not this machine. The first version of
3976 // this test said x86_64-linux-gnu, which is a cross compile on a mac and this machine on a
3977 // Linux runner, so it passed here and failed there.
3978 let gnu = format!("--target={}-linux-gnu", cross_arch());
3979 let (bundled, _) = compile(&[&gnu, "-c", "a.c"]);
3980 assert_eq!(bundled.glibc_minor, Some(44));
3981 let pin = format!("{gnu}.2.28");
3982 let (pinned, _) = compile(&[&pin, "-c", "a.c"]);
3983 assert_eq!(pinned.glibc_minor, Some(28));
3984
3985 let (named, _) = compile(&[&gnu, "--sysroot=/nowhere-at-all", "-c", "a.c"]);
3986 assert_eq!(named.glibc_minor, None);
3987 let (none, _) = compile(&[&gnu, "-nostdinc", "-c", "a.c"]);
3988 assert_eq!(none.glibc_minor, None);
3989 let musl = format!("--target={}-linux-musl", cross_arch());
3990 let (musl, _) = compile(&[&musl, "-c", "a.c"]);
3991 assert_eq!(musl.glibc_minor, None);
3992
3993 // And this machine's own target gets nothing, whatever this machine is, because its headers
3994 // come from the machine and its own `features.h` defines the macro. On a glibc Linux box
3995 // that is the case this test had backwards; on a mac it is true for the other reason, which
3996 // is that Darwin is not a glibc target at all.
3997 if let Some(host) = Triple::host() {
3998 let native = format!("--target={}", host.tuple());
3999 let (native, _) = compile(&[&native, "-c", "a.c"]);
4000 assert_eq!(native.glibc_minor, None);
4001 }
4002 }
4003
4004 #[test]
4005 fn a_pinned_release_on_this_machines_own_target_reads_the_bundled_tree() {
4006 // The end to end half of the answer in `link::cross_for`. A release named for this machine's
4007 // own target is a cross compile, so the headers are the bundled tree's and the macro says
4008 // what was asked for rather than what this machine has.
4009 //
4010 // Only on a glibc box, because a release is a glibc release: a mac has no `__GLIBC_MINOR__`
4011 // to get wrong and nothing to pin. That makes this a test the Linux runners carry, which is
4012 // where the case lives.
4013 let Some(host) = Triple::host() else { return };
4014 if host.env != rucc_target::Env::Gnu {
4015 return;
4016 }
4017 let pin = format!("--target={}.2.28", host.tuple());
4018 let (opts, _) = compile(&[&pin, "-c", "a.c"]);
4019 assert_eq!(opts.glibc_minor, Some(28));
4020 let root = cache::dir().join("sysroots").join(format!("{}.2.28", host.tuple()));
4021 let dirs: Vec<&std::path::Path> =
4022 opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
4023 assert!(dirs.iter().any(|dir| dir.starts_with(&root)), "{dirs:?}");
4024 // And nothing of this machine's, which is the failure this was: a program compiled against
4025 // 2.44 declarations and told it was 2.28.
4026 assert!(!dirs.iter().any(|dir| *dir == std::path::Path::new("/usr/include")), "{dirs:?}");
4027 }
4028
4029 /// An architecture that is not this machine's, out of the three the driver has targets for.
4030 ///
4031 /// A test about the bundled sysroot has to name a target that is not the host, because a target
4032 /// that is the host reads the host's own headers and libraries. Asking which machine this is
4033 /// beats picking a row and hoping, and it is two lines.
4034 fn cross_arch() -> &'static str {
4035 match Triple::host().map(|host| host.arch) {
4036 Some(rucc_target::Arch::X86_64) => "aarch64",
4037 _ => "x86_64",
4038 }
4039 }
4040
4041 #[test]
4042 fn a_glibc_newer_than_the_bundled_tree_is_refused_by_name() {
4043 // Both versions in the message, because the two things a person can do about it are pin a
4044 // release the tree has and name a sysroot that has the one they asked for, and neither is a
4045 // choice they can make without knowing which release the tree is.
4046 //
4047 // Not this machine's architecture, for the reason the test above gives: the refusal is about
4048 // the bundled tree, and the bundled tree is not what a target that is this machine reads.
4049 let target = format!("--target={}-linux-gnu.2.99", cross_arch());
4050 let message = refused(&[&target, "-c", "a.c"]);
4051 assert!(message.contains("asked for glibc 2.99"), "{message}");
4052 assert!(message.contains("bundled headers are glibc 2.44"), "{message}");
4053 assert!(message.contains("--sysroot"), "{message}");
4054 }
4055
4056 #[test]
4057 fn a_sysroot_the_user_named_is_still_what_a_cross_compile_reads() {
4058 // The tree somebody assembled beats the one we would build, on the headers as on the
4059 // libraries. It is empty here, which is why the list comes out short: the directories under
4060 // it are checked for rather than assumed, and a tree that is not there offers nothing.
4061 let (opts, _) =
4062 compile(&["--target=riscv64-linux-musl", "--sysroot=/nowhere-at-all", "-c", "a.c"]);
4063 let dirs: Vec<&std::path::Path> =
4064 opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
4065 assert_eq!(dirs, [std::path::Path::new(runtime::DIR)]);
4066 }
4067
4068 #[test]
4069 fn dash_i_dash_moves_the_bracket_directories_into_the_quoted_chain() {
4070 let (opts, _) =
4071 compile(&["-Iinc1", "-iquote", "inc2", "-I-", "-Iinc3", "-nostdinc", "a.c"]);
4072 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
4073 assert_eq!(dirs, ["inc1", "inc2", "inc3"]);
4074 // An angled include sees only what came after the flag.
4075 assert_eq!(opts.search.start(IncludeForm::Angled), 2);
4076 assert!(!opts.search.searches_current_dir());
4077 }
4078
4079 #[test]
4080 fn the_prefix_flags_stick_what_iprefix_said_on_the_front_of_what_follows_it() {
4081 let (opts, _) = compile(&[
4082 "-iprefix",
4083 "/tools/",
4084 "-iwithprefix",
4085 "late",
4086 "-iwithprefixbefore",
4087 "early",
4088 "-iprefix",
4089 "/other/",
4090 "-iwithprefix",
4091 "last",
4092 "-nostdinc",
4093 "a.c",
4094 ]);
4095 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
4096 // `-iwithprefixbefore` is an `-I` and the other two are `-isystem`, which is where GCC
4097 // puts them rather than where its manual says it does.
4098 assert_eq!(dirs, ["/tools/early", "/tools/late", "/other/last"]);
4099 assert!(!opts.search.dirs()[0].is_system);
4100 assert!(opts.search.dirs()[1].is_system);
4101 }
4102
4103 #[test]
4104 fn the_files_named_on_the_command_line_keep_their_order_and_which_flag_named_them() {
4105 let (opts, _) =
4106 compile(&["-include", "one.h", "-imacros", "two.h", "-include", "3.h", "a.c"]);
4107 let names: Vec<&str> = opts.preincludes.iter().map(|p| p.name.as_str()).collect();
4108 assert_eq!(names, ["one.h", "two.h", "3.h"]);
4109 assert_eq!(opts.preincludes.iter().filter(|p| p.macros_only).count(), 1);
4110 }
4111
4112 #[test]
4113 fn nostdinc_takes_the_compilers_own_headers_off_the_path() {
4114 let (opts, _) = compile(&["-Ii", "-nostdinc", "a.c"]);
4115 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
4116 assert_eq!(dirs, ["i"]);
4117 }
4118
4119 #[test]
4120 fn the_dialect_flags_set_the_language_and_the_extensions_separately() {
4121 let (opts, _) = compile(&["-std=gnu11", "a.c"]);
4122 assert_eq!(opts.std, Std::C11);
4123 assert!(opts.gnu_extensions);
4124
4125 let (opts, _) = compile(&["-std=iso9899:1999", "a.c"]);
4126 assert_eq!(opts.std, Std::C99);
4127 assert!(!opts.gnu_extensions);
4128
4129 let (opts, _) = compile(&["-ansi", "a.c"]);
4130 assert_eq!(opts.std, Std::C89);
4131 assert!(!opts.gnu_extensions);
4132
4133 let e = parse_args(&args(&["-std=c94jr", "a.c"])).unwrap_err();
4134 assert!(e.message.contains("unknown dialect"), "{}", e.message);
4135 }
4136
4137 #[test]
4138 fn the_dump_letters_are_a_family_and_everything_else_beginning_with_d_is_not() {
4139 let (opts, _) = compile(&["-dM", "a.c"]);
4140 assert!(opts.dumps.macros);
4141
4142 // Packed, the way GCC takes them, and a letter in the family we have not written yet
4143 // is accepted and does nothing rather than failing a build.
4144 let (opts, _) = compile(&["-dDM", "a.c"]);
4145 assert!(opts.dumps.macros);
4146 let (opts, _) = compile(&["-dD", "a.c"]);
4147 assert!(!opts.dumps.macros);
4148
4149 let (opts, _) = compile(&["a.c"]);
4150 assert!(!opts.dumps.any());
4151
4152 // `-dumpversion` is a different flag that happens to start the same way, and it is read
4153 // as itself rather than as a dump of nothing.
4154 assert_eq!(printed(&["-dumpversion", "a.c"]), VERSION);
4155 }
4156
4157 #[test]
4158 fn the_gcc_version_claimed_is_a_flag_and_the_short_spellings_are_the_ones_people_write() {
4159 let (opts, _) = compile(&["a.c"]);
4160 assert_eq!(
4161 opts.gnuc,
4162 GnucVersion { major: 7, minor: 0, patch: 0 },
4163 "the lowest claim a modern glibc gives its own declarations to"
4164 );
4165
4166 let (opts, _) = compile(&["-fgnuc-version=15.1.0", "a.c"]);
4167 assert_eq!(opts.gnuc, GnucVersion { major: 15, minor: 1, patch: 0 });
4168
4169 // A missing component is zero. `gcc -dumpversion` says `15` on a release with no
4170 // patchlevel and a harness that pastes that back has to be understood.
4171 let (opts, _) = compile(&["-fgnuc-version=15", "a.c"]);
4172 assert_eq!(opts.gnuc, GnucVersion { major: 15, minor: 0, patch: 0 });
4173
4174 let (opts, _) = compile(&["-fgnuc-version=13.2", "a.c"]);
4175 assert_eq!(opts.gnuc, GnucVersion { major: 13, minor: 2, patch: 0 });
4176
4177 let e = parse_args(&args(&["-fgnuc-version=15.x", "a.c"])).unwrap_err();
4178 assert!(e.message.contains("minor that is not a number"), "{}", e.message);
4179
4180 let e = parse_args(&args(&["-fgnuc-version=1.2.3.4", "a.c"])).unwrap_err();
4181 assert!(e.message.contains("more than three"), "{}", e.message);
4182 }
4183
4184 #[test]
4185 fn pedantic_has_two_spellings_and_is_not_the_same_knob_as_the_dialect() {
4186 let (opts, _) = compile(&["-std=c17", "-pedantic", "a.c"]);
4187 assert!(opts.pedantic);
4188 assert_eq!(opts.std, Std::C17);
4189
4190 // The `-W` family's name for it, which is what a build that groups its warning flags
4191 // tends to write.
4192 let (opts, _) = compile(&["-Wpedantic", "a.c"]);
4193 assert!(opts.pedantic);
4194
4195 let (opts, _) = compile(&["-std=c17", "a.c"]);
4196 assert!(!opts.pedantic, "a dialect on its own does not diagnose an extension");
4197 }
4198
4199 #[test]
4200 fn dash_p_and_dash_ffreestanding_reach_the_options() {
4201 let (opts, _) = compile(&["-E", "-P", "-ffreestanding", "a.c"]);
4202 assert!(!opts.line_markers);
4203 assert!(!opts.hosted);
4204 assert_eq!(opts.emit, EmitKind::Preprocessed);
4205 }
4206
4207 /// The two ways a build says it means its own function by a name the C library also has.
4208 ///
4209 /// `-fno-builtin` is all of them and `-fno-builtin-<name>` is one, and the second is what a
4210 /// build writes when it means its own `memcpy` and the library's everything else. The name is
4211 /// kept as it was written and not checked against anything, because a program is allowed to
4212 /// mean something by a name this compiler has never heard of.
4213 #[test]
4214 fn the_builtin_flags_are_read_in_both_directions_and_one_name_at_a_time() {
4215 let (opts, _) = compile(&["-c", "a.c"]);
4216 assert!(opts.builtins, "a library name means the library function by default");
4217 assert!(opts.no_builtin.is_empty());
4218
4219 let (opts, _) = compile(&["-c", "-fno-builtin", "a.c"]);
4220 assert!(!opts.builtins);
4221
4222 let (opts, _) = compile(&["-c", "-fno-builtin", "-fbuiltin", "a.c"]);
4223 assert!(opts.builtins, "the last mention decides");
4224
4225 let (opts, _) = compile(&["-c", "-fno-builtin-memcpy", "-fno-builtin-nonesuch", "a.c"]);
4226 assert!(opts.builtins, "one name is not the family");
4227 assert_eq!(opts.no_builtin, vec!["memcpy".to_owned(), "nonesuch".to_owned()]);
4228 }
4229
4230 /// `-fvisibility=`, which is on every cmake project that cares about which names it exports
4231 /// and which was refused as an unknown option until now.
4232 ///
4233 /// Four spellings and three answers. `internal` is hidden plus a promise about never taking
4234 /// the address across a component boundary, and nothing derives anything from that promise
4235 /// here, so it comes out as the weaker of the two rather than as a refusal that stops a build
4236 /// over a distinction this compiler does not make.
4237 #[test]
4238 fn visibility_takes_the_four_spellings_gcc_takes_and_refuses_the_rest() {
4239 let (opts, _) = compile(&["-c", "a.c"]);
4240 assert_eq!(opts.visibility, Visibility::Default, "exported unless something says not");
4241
4242 for (written, wanted) in [
4243 ("default", Visibility::Default),
4244 ("hidden", Visibility::Hidden),
4245 ("internal", Visibility::Hidden),
4246 ("protected", Visibility::Protected),
4247 ] {
4248 let (opts, _) = compile(&["-c", &format!("-fvisibility={written}"), "a.c"]);
4249 assert_eq!(opts.visibility, wanted, "{written}");
4250 }
4251
4252 // The last mention decides, which is what every other flag of this shape does and what a
4253 // build that turns something off for one directory relies on.
4254 let (opts, _) = compile(&["-c", "-fvisibility=hidden", "-fvisibility=default", "a.c"]);
4255 assert_eq!(opts.visibility, Visibility::Default, "the last mention decides");
4256
4257 // A spelling gcc does not take is refused rather than read as the default, because a
4258 // build that meant hidden and got exported is a library with the wrong interface and
4259 // nothing said about it anywhere.
4260 let failed = parse_args(&args(&["-fvisibility=none", "a.c"])).expect_err("refused");
4261 assert!(failed.to_string().contains("is not a visibility"), "{failed}");
4262 }
4263
4264 /// `-ffp-contract=`, which is the one flag in the floating point group that is kept rather than
4265 /// described, and the values are gcc 16's three.
4266 #[test]
4267 fn how_far_a_multiply_and_an_addition_may_be_fused_is_asked_for() {
4268 let (opts, _) = compile(&["-c", "a.c"]);
4269 assert_eq!(opts.fp_contract, Contract::Off, "a licence nobody granted is not assumed");
4270
4271 for (written, wanted) in
4272 [("off", Contract::Off), ("on", Contract::On), ("fast", Contract::Fast)]
4273 {
4274 let (opts, _) = compile(&["-c", &format!("-ffp-contract={written}"), "a.c"]);
4275 assert_eq!(opts.fp_contract, wanted, "{written}");
4276 }
4277
4278 let (opts, _) = compile(&["-c", "-ffp-contract=fast", "-ffp-contract=off", "a.c"]);
4279 assert_eq!(opts.fp_contract, Contract::Off, "the last mention decides");
4280
4281 // Refused rather than read as one of the three, because a build that asked for no fusing
4282 // and was given the default would be one whose numbers change and whose command line says
4283 // they should not. gcc refuses the same spellings and names the same three in its message.
4284 for bad in ["-ffp-contract=none", "-ffp-contract=", "-ffp-contract=Fast"] {
4285 let failed = parse_args(&args(&[bad, "a.c"])).expect_err("refused");
4286 assert!(failed.to_string().contains("is not a contraction"), "{bad}: {failed}");
4287 }
4288
4289 // And the other one that takes a value, which is taken and kept nowhere: every operation
4290 // here is computed in the type it was written in, so `standard` is what happens and the
4291 // other two are permission to do something this does not do.
4292 let failed = parse_args(&args(&["-fexcess-precision=long", "a.c"])).expect_err("refused");
4293 assert!(failed.to_string().contains("is not an excess precision"), "{failed}");
4294 }
4295
4296 /// The four prefix mapping flags, which are what a distribution passes to get the same bytes
4297 /// out of `/build/pkg-1.2` and out of `/home/someone/pkg-1.2`. Three lists rather than one
4298 /// because gcc has three, and `-ffile-prefix-map=` is the three of them at once.
4299 #[test]
4300 fn a_prefix_mapping_flag_goes_on_the_list_its_spelling_names() {
4301 let (opts, _) = compile(&["-c", "a.c"]);
4302 assert!(opts.prefix_map.macros.is_empty(), "nothing is rewritten unless it is asked for");
4303 assert!(opts.prefix_map.debug.is_empty(), "nor here");
4304 assert!(opts.prefix_map.profile.is_empty(), "nor here");
4305
4306 let (opts, _) = compile(&["-c", "-fmacro-prefix-map=/build=.", "a.c"]);
4307 assert_eq!(opts.prefix_map.macros.apply("/build/a.c"), "./a.c", "the one it names");
4308 assert!(opts.prefix_map.debug.is_empty(), "and not the two it does not");
4309
4310 let (opts, _) = compile(&["-c", "-fdebug-prefix-map=/build=.", "a.c"]);
4311 assert_eq!(opts.prefix_map.debug.apply("/build/a.c"), "./a.c", "the one it names");
4312 assert!(opts.prefix_map.macros.is_empty(), "and not the two it does not");
4313
4314 let (opts, _) = compile(&["-c", "-fprofile-prefix-map=/build=.", "a.c"]);
4315 assert_eq!(opts.prefix_map.profile.apply("/build/a.c"), "./a.c", "the one it names");
4316 assert!(opts.prefix_map.macros.is_empty(), "and not the two it does not");
4317
4318 let (opts, _) = compile(&["-c", "-ffile-prefix-map=/build=.", "a.c"]);
4319 for list in [&opts.prefix_map.macros, &opts.prefix_map.debug, &opts.prefix_map.profile] {
4320 assert_eq!(list.apply("/build/a.c"), "./a.c", "all three at once");
4321 }
4322
4323 // Every mention is kept and the last one that matches wins, unlike the flags above whose
4324 // last mention replaces the earlier ones. A build writes one of these per source root and
4325 // expects all of them to be in force, which is the whole point of a list.
4326 let (opts, _) =
4327 compile(&["-c", "-ffile-prefix-map=/a=one", "-ffile-prefix-map=/b=two", "a.c"]);
4328 assert_eq!(opts.prefix_map.macros.apply("/a/x.c"), "one/x.c", "the earlier one still acts");
4329 assert_eq!(opts.prefix_map.macros.apply("/b/x.c"), "two/x.c", "and so does the later one");
4330
4331 // An argument with no `=` is refused rather than ignored, because a build whose paths were
4332 // meant to be rewritten and were not is one that ships the build directory's name and says
4333 // nothing about it. gcc refuses the same thing.
4334 for bad in ["-fmacro-prefix-map=nope", "-ffile-prefix-map=", "-fdebug-prefix-map=/build"] {
4335 let failed = parse_args(&args(&[bad, "a.c"])).expect_err("refused");
4336 assert!(failed.to_string().contains("is not a rewrite for"), "{bad}: {failed}");
4337 }
4338 }
4339
4340 /// `-ffunction-sections` and `-fdata-sections`, which are what make `--gc-sections` able to
4341 /// drop anything: a linker can leave out a section nothing reaches and cannot leave out half of
4342 /// one. A kernel and an embedded image are both linked that way.
4343 ///
4344 /// Two flags rather than one because gcc has two, and a build that asks for one of them and not
4345 /// the other is a build that measured something: splitting the code is nearly free at link time
4346 /// and splitting the data can defeat the linker's ordering of what is next to what.
4347 #[test]
4348 fn a_section_per_function_and_a_section_per_variable_are_asked_for_one_at_a_time() {
4349 let (opts, _) = compile(&["-c", "a.c"]);
4350 assert!(!opts.function_sections, "one text section unless something says otherwise");
4351 assert!(!opts.data_sections);
4352
4353 let (opts, _) = compile(&["-c", "-ffunction-sections", "a.c"]);
4354 assert!(opts.function_sections);
4355 assert!(!opts.data_sections, "one flag is not the other");
4356
4357 let (opts, _) = compile(&["-c", "-fdata-sections", "a.c"]);
4358 assert!(opts.data_sections);
4359 assert!(!opts.function_sections);
4360
4361 // Both directions taken, and the off one is what happens anyway rather than a refusal,
4362 // since a build that writes it is asking for the default.
4363 let (opts, _) = compile(&[
4364 "-c",
4365 "-ffunction-sections",
4366 "-fno-function-sections",
4367 "-fdata-sections",
4368 "-fno-data-sections",
4369 "a.c",
4370 ]);
4371 assert!(!opts.function_sections, "the last mention decides");
4372 assert!(!opts.data_sections, "the last mention decides");
4373 }
4374
4375 /// `-fgnu89-inline`, which is off by default and is not implied by anything on the command
4376 /// line, since the dialect asks for GNU's reading further in rather than through this.
4377 #[test]
4378 fn gnu89_inline_is_off_until_it_is_asked_for_and_the_last_mention_decides() {
4379 let (opts, _) = compile(&["-c", "a.c"]);
4380 assert!(!opts.gnu89_inline, "C's reading of inline by default");
4381
4382 let (opts, _) = compile(&["-c", "-fgnu89-inline", "a.c"]);
4383 assert!(opts.gnu89_inline);
4384
4385 let (opts, _) = compile(&["-c", "-fgnu89-inline", "-fno-gnu89-inline", "a.c"]);
4386 assert!(!opts.gnu89_inline, "the last mention decides");
4387
4388 // The C89 dialects are under GNU's reading whether this was written or not, so the flag
4389 // stays off there and the dialect is what the checker and the macro set both ask. That is
4390 // also why `-std=c89 -fno-gnu89-inline` needs no diagnostic: it asks for the reading the
4391 // dialect already has. gcc refuses that command line, which is measured in the issue.
4392 let (opts, _) = compile(&["-c", "-std=c89", "a.c"]);
4393 assert!(!opts.gnu89_inline);
4394 }
4395
4396 /// Both spellings of both frame flags, since a build that wants one usually writes the
4397 /// other beside it for the one file that has to be compiled the ordinary way.
4398 #[test]
4399 fn the_two_frame_flags_are_read_in_both_directions() {
4400 let (opts, _) = compile(&["-c", "a.c"]);
4401 assert!(!opts.frame_pointer, "gcc omits it above -O0 and so does this");
4402 assert!(opts.red_zone, "the psABI has one and nothing said not to use it");
4403
4404 let (opts, _) = compile(&["-c", "-fno-omit-frame-pointer", "-mno-red-zone", "a.c"]);
4405 assert!(opts.frame_pointer);
4406 assert!(!opts.red_zone);
4407
4408 let (opts, _) = compile(&[
4409 "-c",
4410 "-fno-omit-frame-pointer",
4411 "-fomit-frame-pointer",
4412 "-mno-red-zone",
4413 "-mred-zone",
4414 "a.c",
4415 ]);
4416 assert!(!opts.frame_pointer, "the last one wins, as it does in gcc");
4417 assert!(opts.red_zone);
4418 }
4419
4420 /// Four flags rather than one with an argument, which is how gcc spells them, and the negative
4421 /// spelled three ways because a build that turns one off writes whichever it turned on.
4422 #[test]
4423 fn the_stack_protector_is_four_flags_and_the_last_one_wins() {
4424 let (opts, _) = compile(&["-c", "a.c"]);
4425 assert_eq!(opts.protector, Protector::None, "gcc protects nothing unless it was asked");
4426
4427 for (flag, want) in [
4428 ("-fstack-protector", Protector::Buffers),
4429 ("-fstack-protector-strong", Protector::Strong),
4430 ("-fstack-protector-all", Protector::All),
4431 ] {
4432 let (opts, _) = compile(&["-c", flag, "a.c"]);
4433 assert_eq!(opts.protector, want, "{flag}");
4434 }
4435
4436 // What a package build does: the strong one in the global flags and one directory that
4437 // cannot have a protector turning it off on the line after.
4438 for off in ["-fno-stack-protector", "-fno-stack-protector-strong"] {
4439 let (opts, _) = compile(&["-c", "-fstack-protector-strong", off, "a.c"]);
4440 assert_eq!(opts.protector, Protector::None, "{off}");
4441 }
4442 let (opts, _) = compile(&["-c", "-fno-stack-protector", "-fstack-protector-all", "a.c"]);
4443 assert_eq!(opts.protector, Protector::All, "the last one wins either way round");
4444 }
4445
4446 /// A switch rather than a level, because how a frame is taken is one question and which
4447 /// functions get a canary is another, and gcc spells it that way for the same reason.
4448 #[test]
4449 fn taking_a_frame_a_page_at_a_time_is_off_until_it_is_asked_for() {
4450 let (opts, _) = compile(&["-c", "a.c"]);
4451 assert!(!opts.stack_clash, "gcc takes a frame in one subtraction unless it was asked");
4452
4453 let (opts, _) = compile(&["-c", "-fstack-clash-protection", "a.c"]);
4454 assert!(opts.stack_clash);
4455
4456 // The same shape a package build uses for the protector: on in the global flags and off
4457 // for the one directory that cannot have it.
4458 let (opts, _) =
4459 compile(&["-c", "-fstack-clash-protection", "-fno-stack-clash-protection", "a.c"]);
4460 assert!(!opts.stack_clash);
4461 let (opts, _) =
4462 compile(&["-c", "-fno-stack-clash-protection", "-fstack-clash-protection", "a.c"]);
4463 assert!(opts.stack_clash, "the last one wins either way round");
4464
4465 // The two are independent, since one is about the frame and the other about the function.
4466 let (opts, _) =
4467 compile(&["-c", "-fstack-clash-protection", "-fstack-protector-strong", "a.c"]);
4468 assert!(opts.stack_clash);
4469 assert_eq!(opts.protector, Protector::Strong);
4470 }
4471
4472 /// One flag with an argument rather than a family of spellings, because what it asks about is
4473 /// which of the two edges of a control flow transfer is checked and the two are not separate
4474 /// questions to the hardware.
4475 #[test]
4476 fn which_control_flow_edges_are_checked_is_asked_for_by_name() {
4477 let (opts, _) = compile(&["-c", "a.c"]);
4478 assert_eq!(opts.control, Control::None, "gcc's default on the targets this compiler has");
4479
4480 for (arg, want) in [
4481 ("-fcf-protection", Control::Full),
4482 ("-fcf-protection=full", Control::Full),
4483 ("-fcf-protection=branch", Control::Branch),
4484 ("-fcf-protection=return", Control::Return),
4485 ("-fcf-protection=none", Control::None),
4486 ("-fcf-protection=check", Control::Check),
4487 ] {
4488 let (opts, _) = compile(&["-c", arg, "a.c"]);
4489 assert_eq!(opts.control, want, "{arg}");
4490 }
4491
4492 // The shape a package build uses: on in the global flags and off for the one directory
4493 // that cannot have it, whichever of the two spellings of off it reaches for.
4494 let (opts, _) = compile(&["-c", "-fcf-protection=full", "-fno-cf-protection", "a.c"]);
4495 assert_eq!(opts.control, Control::None);
4496 let (opts, _) = compile(&["-c", "-fno-cf-protection", "-fcf-protection=branch", "a.c"]);
4497 assert_eq!(opts.control, Control::Branch, "the last one wins either way round");
4498 }
4499
4500 /// The profiler is asked for by two spellings, and where its hook goes by two more.
4501 ///
4502 /// The two halves are separate on purpose. `-mfentry` on its own says where a call would go and
4503 /// asks for no call, which is what gcc does with it, and a build system that sets it globally
4504 /// and asks for the profile per directory needs that to be true rather than an error.
4505 ///
4506 /// The link is asserted alongside, because the flag changes it too and a build that compiled
4507 /// with it and linked without it is a program that calls the hook everywhere and never writes a
4508 /// profile.
4509 #[test]
4510 fn the_profiler_and_where_its_hook_goes_are_two_separate_questions() {
4511 let (opts, _) = compile(&["-c", "a.c"]);
4512 assert!(!opts.profile);
4513 assert_eq!(opts.hook, Hook::Platform, "neither was named, so the target decides");
4514
4515 for arg in ["-pg", "-p"] {
4516 let (opts, _) = compile(&["-c", arg, "a.c"]);
4517 assert!(opts.profile, "{arg}");
4518 let (link, _) = linking(&[arg, "a.c"]);
4519 assert!(link.profile, "{arg} changes the link as well");
4520 }
4521
4522 for (arg, want) in [("-mfentry", Hook::Early), ("-mno-fentry", Hook::Late)] {
4523 let (opts, _) = compile(&["-c", arg, "a.c"]);
4524 assert_eq!(opts.hook, want, "{arg}");
4525 assert!(!opts.profile, "{arg} asks for no call of its own");
4526 }
4527
4528 let (opts, _) = compile(&["-c", "-mfentry", "-mno-fentry", "-pg", "a.c"]);
4529 assert_eq!(opts.hook, Hook::Late, "the last one wins");
4530 assert!(opts.profile);
4531 }
4532
4533 /// How much room a patcher is promised, which is one number or two.
4534 ///
4535 /// A command line that did not ask is asserted alongside, because the flag has to be written to
4536 /// mean anything and a build that reserved room nobody asked for would grow every function in
4537 /// it for nothing.
4538 #[test]
4539 fn the_room_a_patcher_is_promised_is_a_number_of_bytes_and_where_they_go() {
4540 let (opts, _) = compile(&["-c", "a.c"]);
4541 assert_eq!(opts.patchable, Patchable::default());
4542 assert!(!opts.patchable.any(), "nothing is reserved unless it was asked for");
4543
4544 let (opts, _) = compile(&["-c", "-fpatchable-function-entry=16", "a.c"]);
4545 assert_eq!(opts.patchable, Patchable { total: 16, before: 0 });
4546
4547 let (opts, _) = compile(&["-c", "-fpatchable-function-entry=5,3", "a.c"]);
4548 assert_eq!(opts.patchable, Patchable { total: 5, before: 3 });
4549 assert_eq!(opts.patchable.after(), 2);
4550
4551 // The last one wins, which is what every other flag of this shape does and what a build
4552 // that adds one to a command line it did not write is relying on.
4553 let (opts, _) = compile(&[
4554 "-c",
4555 "-fpatchable-function-entry=5,3",
4556 "-fpatchable-function-entry=2",
4557 "a.c",
4558 ]);
4559 assert_eq!(opts.patchable, Patchable { total: 2, before: 0 });
4560 }
4561
4562 /// And a request nothing could satisfy is refused rather than rounded into one that can be.
4563 #[test]
4564 fn room_in_front_of_the_label_that_is_more_than_the_room_asked_for_is_refused() {
4565 for arg in ["-fpatchable-function-entry=1,2", "-fpatchable-function-entry=x"] {
4566 let e = parse_args(&args(&["-c", arg, "a.c"])).unwrap_err();
4567 assert!(e.message.contains("is not an amount of room to reserve"), "{}", e.message);
4568 }
4569 }
4570
4571 /// What wraps rather than being undefined, which is two questions and three flags.
4572 ///
4573 /// The older flag is the pair of the newer two, which is gcc's own reading of it, so a build
4574 /// that writes `-fno-strict-overflow` gets both and a build that writes one of the others gets
4575 /// only what it asked for.
4576 #[test]
4577 fn what_overflows_rather_than_being_undefined_is_asked_for_two_ways() {
4578 let (opts, _) = compile(&["-c", "a.c"]);
4579 assert_eq!(opts.wrapping, Wrapping::NONE, "nothing wraps unless it was asked for");
4580
4581 let (opts, _) = compile(&["-c", "-fwrapv", "a.c"]);
4582 assert_eq!(opts.wrapping, Wrapping { signed: true, pointer: false, trap: false });
4583
4584 let (opts, _) = compile(&["-c", "-fwrapv-pointer", "a.c"]);
4585 assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: true, trap: false });
4586
4587 let (opts, _) = compile(&["-c", "-fno-strict-overflow", "a.c"]);
4588 assert_eq!(opts.wrapping, Wrapping::ALL);
4589
4590 // And the last one wins, in both directions. A build that turns one of these on globally
4591 // and off for one directory is relying on that, and so is one that writes the pair and
4592 // then takes half of it back.
4593 let (opts, _) = compile(&["-c", "-fwrapv", "-fno-wrapv", "a.c"]);
4594 assert_eq!(opts.wrapping, Wrapping::NONE);
4595
4596 let (opts, _) = compile(&["-c", "-fno-strict-overflow", "-fstrict-overflow", "a.c"]);
4597 assert_eq!(opts.wrapping, Wrapping::NONE);
4598
4599 let (opts, _) = compile(&["-c", "-fno-strict-overflow", "-fno-wrapv-pointer", "a.c"]);
4600 assert_eq!(opts.wrapping, Wrapping { signed: true, pointer: false, trap: false });
4601 }
4602
4603 /// And the other answer to the signed question cannot be held at the same time as the first.
4604 ///
4605 /// A program cannot both wrap and stop, so writing both is writing a contradiction, and gcc
4606 /// resolves it by letting the last one win rather than by reporting anything. That was measured
4607 /// against gcc 16 rather than read out of the manual, which says nothing about it: `-ftrapv
4608 /// -fwrapv` emits no checked calls and `-fwrapv -ftrapv` emits them.
4609 #[test]
4610 fn a_signed_overflow_that_stops_is_the_other_answer_and_not_a_third_one() {
4611 let (opts, _) = compile(&["-c", "-ftrapv", "a.c"]);
4612 assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: false, trap: true });
4613
4614 let (opts, _) = compile(&["-c", "-fwrapv", "-ftrapv", "a.c"]);
4615 assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: false, trap: true });
4616
4617 let (opts, _) = compile(&["-c", "-ftrapv", "-fwrapv", "a.c"]);
4618 assert_eq!(opts.wrapping, Wrapping { signed: true, pointer: false, trap: false });
4619
4620 let (opts, _) = compile(&["-c", "-ftrapv", "-fno-strict-overflow", "a.c"]);
4621 assert_eq!(opts.wrapping, Wrapping::ALL);
4622
4623 let (opts, _) = compile(&["-c", "-ftrapv", "-fno-trapv", "a.c"]);
4624 assert_eq!(opts.wrapping, Wrapping::NONE);
4625
4626 // And the flag that says what may be assumed says nothing about what happens, so it leaves
4627 // this alone where it takes the wrapping away. gcc does the same.
4628 let (opts, _) = compile(&["-c", "-ftrapv", "-fstrict-overflow", "a.c"]);
4629 assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: false, trap: true });
4630 }
4631
4632 /// What a plain `char` is, which is four spellings of two answers and nothing by default.
4633 ///
4634 /// Nothing is the target's own answer and has to stay distinct from both of the others, since
4635 /// the same command line means a signed `char` on x86-64 and an unsigned one on Linux's arm64.
4636 /// The negative spellings are the other flag rather than a way of asking for the default, which
4637 /// was measured against gcc 16: `-fno-signed-char` defines `__CHAR_UNSIGNED__` and
4638 /// `-fno-unsigned-char` does not.
4639 #[test]
4640 fn the_signedness_of_a_plain_char_is_asked_for_in_four_ways() {
4641 let (opts, _) = compile(&["-c", "a.c"]);
4642 assert_eq!(opts.char_signed, None);
4643
4644 for flag in ["-fsigned-char", "-fno-unsigned-char"] {
4645 let (opts, _) = compile(&["-c", flag, "a.c"]);
4646 assert_eq!(opts.char_signed, Some(true), "{flag}");
4647 }
4648
4649 for flag in ["-funsigned-char", "-fno-signed-char"] {
4650 let (opts, _) = compile(&["-c", flag, "a.c"]);
4651 assert_eq!(opts.char_signed, Some(false), "{flag}");
4652 }
4653
4654 // And the last one wins, which is what a build that sets one globally and the other for a
4655 // directory relies on.
4656 let (opts, _) = compile(&["-c", "-funsigned-char", "-fsigned-char", "a.c"]);
4657 assert_eq!(opts.char_signed, Some(true));
4658
4659 // And what is asked for reaches the target, because that is what every other part of the
4660 // compiler asks. The triple is one whose own answer is the opposite, so a session that
4661 // ignored the flag would still read as signed here.
4662 let (opts, _) =
4663 compile(&["-c", "--target=aarch64-unknown-linux-gnu", "-fsigned-char", "a.c"]);
4664 assert!(Session::new(*opts).target.char_is_signed);
4665 let (opts, _) = compile(&["-c", "--target=aarch64-unknown-linux-gnu", "a.c"]);
4666 assert!(!Session::new(*opts).target.char_is_signed);
4667 }
4668
4669 /// And the size of an enumeration, which is one question with two spellings.
4670 #[test]
4671 fn the_smallest_enumeration_is_asked_for_and_taken_back() {
4672 let (opts, _) = compile(&["-c", "a.c"]);
4673 assert!(!opts.short_enums);
4674
4675 let (opts, _) = compile(&["-c", "-fshort-enums", "a.c"]);
4676 assert!(opts.short_enums);
4677
4678 let (opts, _) = compile(&["-c", "-fshort-enums", "-fno-short-enums", "a.c"]);
4679 assert!(!opts.short_enums);
4680
4681 let (opts, _) = compile(&["-c", "-fno-short-enums", "-fshort-enums", "a.c"]);
4682 assert!(opts.short_enums);
4683 }
4684
4685 /// And a value nothing means is refused rather than taken for the nearest thing it looks like.
4686 ///
4687 /// `-fcf-protection=all` is the spelling somebody writes from memory, and a compiler that read
4688 /// it as `full` would be guessing, while one that let it fall through to the optimizer's `-f`
4689 /// family would report it as an unknown pass. Neither is the news the build wants.
4690 #[test]
4691 fn a_control_flow_protection_nothing_means_is_refused() {
4692 let e = parse_args(&args(&["-c", "-fcf-protection=all", "a.c"])).unwrap_err();
4693 assert!(e.message.contains("is not a control flow protection"), "{}", e.message);
4694 assert!(e.message.contains("full, branch, return, none or check"), "{}", e.message);
4695 }
4696
4697 #[test]
4698 fn the_link_flags_are_collected_apart_from_the_compilation() {
4699 let (link, _) = linking(&[
4700 "-static",
4701 "-nostartfiles",
4702 "-rdynamic",
4703 "-s",
4704 "-fuse-ld=mold",
4705 "-L/opt/lib",
4706 "-B",
4707 "/opt/tools",
4708 "a.c",
4709 ]);
4710 assert!(link.is_static);
4711 assert!(link.no_startfiles);
4712 assert!(link.export_dynamic);
4713 assert!(link.strip);
4714 assert_eq!(link.use_ld.as_deref(), Some("mold"));
4715 assert_eq!(link.search, vec![PathBuf::from("/opt/lib")]);
4716 assert_eq!(link.prefixes, vec![PathBuf::from("/opt/tools")]);
4717 }
4718
4719 #[test]
4720 fn a_comma_in_dash_wl_separates_two_arguments() {
4721 let (link, _) = linking(&["-Wl,-rpath,/opt/lib", "-Xlinker", "--as-needed", "a.c"]);
4722 assert_eq!(link.passthrough, vec!["-rpath", "/opt/lib", "--as-needed"]);
4723 }
4724
4725 #[test]
4726 fn a_library_keeps_its_place_between_the_objects() {
4727 // Link order is semantic: `-lm` written between two files resolves for the one before
4728 // it and not for the one after, so a library cannot be collected into a list of its own.
4729 // The target is named because the suffix of an object is the target's and this asserts
4730 // on the names: the same command line on a Windows host plans two `.obj` files.
4731 let (_, plan) = linking(&["--target=x86_64-unknown-linux-gnu", "a.c", "-lm", "b.c"]);
4732 let link = plan.link.expect("expected a link step");
4733 assert_eq!(
4734 link.inputs,
4735 vec![
4736 link::Item::File("a.o".into()),
4737 link::Item::Library("m".into()),
4738 link::Item::File("b.o".into()),
4739 ]
4740 );
4741 // And it is not a job, because there is nothing to compile in a library.
4742 assert_eq!(plan.jobs.len(), 2);
4743 }
4744
4745 #[test]
4746 fn a_library_on_a_dash_c_line_is_a_note_rather_than_an_error() {
4747 let (_, plan) = linking(&["-c", "-lm", "a.c"]);
4748 assert!(plan.link.is_none());
4749 assert!(plan.notes.iter().any(|n| n.contains("-lm")), "{:?}", plan.notes);
4750 }
4751
4752 #[test]
4753 fn the_sysroot_reaches_the_linker_as_well_as_the_headers() {
4754 let (link, _) = linking(&["--sysroot=/opt/root", "a.c"]);
4755 assert_eq!(link.sysroot, Some(PathBuf::from("/opt/root")));
4756 }
4757
4758 fn printed(s: &[&str]) -> String {
4759 match parse_args(&args(s)).expect("expected an answer") {
4760 Action::Print(line) => line,
4761 other => panic!("expected an answer, got {other:?}"),
4762 }
4763 }
4764
4765 fn refused(s: &[&str]) -> String {
4766 parse_args(&args(s)).expect_err("expected a refusal").message
4767 }
4768
4769 #[test]
4770 fn a_warning_flag_this_compiler_has_not_heard_of_is_taken_rather_than_refused() {
4771 // The rule in section 4.1, and the reason for it is autoconf: a configure script finds
4772 // out whether a warning flag exists by passing it and looking at the exit status, so a
4773 // compiler that refuses one it does not know fails a script written for a newer GCC.
4774 let (opts, _) = compile(&["-Wall", "-Wextra", "-Wno-format-truncation", "-c", "a.c"]);
4775 assert!(!opts.warnings_are_errors);
4776 assert!(opts.warnings);
4777 // The two spellings that do mean something are still read.
4778 let (opts, _) = compile(&["-Werror", "-c", "a.c"]);
4779 assert!(opts.warnings_are_errors);
4780 let (opts, _) = compile(&["-w", "-c", "a.c"]);
4781 assert!(!opts.warnings);
4782 let (opts, _) = compile(&["-pedantic-errors", "-c", "a.c"]);
4783 assert!(opts.pedantic && opts.warnings_are_errors);
4784 }
4785
4786 #[test]
4787 fn an_argument_for_a_separate_tool_is_refused_rather_than_dropped() {
4788 // Every one of these says something about the output, so the wrong answer is silence.
4789 assert!(refused(&["-Wa,--noexecstack", "-c", "a.c"]).contains("separate assembler"));
4790 assert!(refused(&["-Wp,-DX", "-c", "a.c"]).contains("separate assembler"));
4791 assert!(refused(&["-specs=/x", "a.c"]).contains("-specs= is not supported"));
4792 assert!(refused(&["-mcmodel=kernel", "-c", "a.c"]).contains("small code model"));
4793 assert!(refused(&["-gdwarf-4", "-c", "a.c"]).contains("DWARF 5"));
4794 assert!(refused(&["-Ofast", "-c", "a.c"]).contains("fast math"));
4795 // The word size the target does not have, which is a target this compiler was not asked
4796 // for rather than a flag it does not know.
4797 let no32 = refused(&["--target=x86_64-unknown-linux-gnu", "-m32", "-c", "a.c"]);
4798 assert!(no32.contains("32 bit target"), "{no32}");
4799 }
4800
4801 /// `-gz` and the two spellings of the split, which are the two questions about the shape of
4802 /// the debug output rather than about how much of it there is.
4803 ///
4804 /// Both answers here are about what happens when there is debug information to shape, and
4805 /// there is none yet, so what is being asserted is that the flags are read and remembered
4806 /// rather than that anything changed in the output. That is the whole of what taking them
4807 /// claims, and it is worth a test because the day `rucc-debug` writes a section this is where
4808 /// it comes to find out what the command line said.
4809 #[test]
4810 fn the_shape_of_the_debug_output_is_recorded_even_where_there_is_none_of_it() {
4811 let (opts, _) = compile(&["-c", "a.c"]);
4812 assert_eq!(opts.compress, Compress::None, "uncompressed unless somebody asks");
4813
4814 // Bare `-gz` is `-gz=zlib`, measured against gcc 16 rather than read out of the manual,
4815 // which describes the flag without ever saying which algorithm it picks.
4816 assert_eq!(compile(&["-gz", "-c", "a.c"]).0.compress, Compress::Zlib);
4817 for (spelling, want) in [
4818 ("none", Compress::None),
4819 ("zlib", Compress::Zlib),
4820 ("zlib-gnu", Compress::ZlibGnu),
4821 ("zstd", Compress::Zstd),
4822 ] {
4823 let (opts, _) = compile(&[&format!("-gz={spelling}"), "-c", "a.c"]);
4824 assert_eq!(opts.compress, want, "{spelling}");
4825 }
4826
4827 // A value nothing here has heard of is refused rather than rounded to the nearest one,
4828 // because a build that asked for `zstd` and quietly got `zlib` would ship a file its
4829 // reader may not understand and would have no way of finding out.
4830 for bad in ["-gz=gzip", "-gz="] {
4831 let failed = refused(&[bad, "-c", "a.c"]);
4832 assert!(failed.contains("is not a way to compress"), "{bad}: {failed}");
4833 }
4834
4835 // The split is refused in the direction that would have written a file and taken in the
4836 // direction that describes what happens. A build system that names the `.dwo` as an
4837 // output has to hear about it now rather than at the point the file is missing.
4838 let (opts, _) = compile(&["-gno-split-dwarf", "-g", "-c", "a.c"]);
4839 assert!(opts.debug_info, "the negative spelling says nothing about how much");
4840 let failed = refused(&["-gsplit-dwarf", "-c", "a.c"]);
4841 assert!(failed.contains(".dwo"), "the refusal names the file it would have written");
4842 }
4843
4844 /// The `-flto` family, which is the whole of an optimization this compiler does not do.
4845 ///
4846 /// Taken rather than refused because ignoring it gives a correct program that is slower than
4847 /// it could have been, which is section 4.1's hint about speed. The values are still held to
4848 /// gcc's, so a command line written for clang is told rather than quietly taken.
4849 #[test]
4850 fn the_link_time_family_is_read_and_checked_and_nothing_is_done_about_it() {
4851 let (opts, _) = compile(&["-c", "a.c"]);
4852 assert!(!opts.lto.requested, "nothing asks unless the command line does");
4853
4854 let (opts, _) = compile(&["-flto", "-c", "a.c"]);
4855 assert!(opts.lto.requested);
4856 assert_eq!(opts.lto.jobs, LtoJobs::One, "bare -flto is one process, the way gcc reads it");
4857
4858 // The last of the two directions wins, the same as every other pair of `-f` spellings.
4859 assert!(!compile(&["-flto", "-fno-lto", "-c", "a.c"]).0.lto.requested);
4860 assert!(compile(&["-fno-lto", "-flto", "-c", "a.c"]).0.lto.requested);
4861
4862 // A count is a count, and asking for one implies asking for the optimization.
4863 for (spelling, want) in [
4864 ("auto", LtoJobs::Auto),
4865 ("jobserver", LtoJobs::Jobserver),
4866 ("1", LtoJobs::One),
4867 ("8", LtoJobs::Count(8)),
4868 ] {
4869 let (opts, _) = compile(&[&format!("-flto={spelling}"), "-c", "a.c"]);
4870 assert_eq!(opts.lto.jobs, want, "{spelling}");
4871 assert!(opts.lto.requested, "{spelling} asks for it too");
4872 }
4873
4874 // gcc refuses a zero rather than reading it as `-fno-lto`, and `thin` is clang's spelling
4875 // of a question gcc answers with `-flto-partition=`, so somebody who wrote it meant a
4876 // different compiler and gets told so here rather than getting a serial link.
4877 for bad in ["-flto=0", "-flto=thin", "-flto=full", "-flto=-1"] {
4878 let failed = refused(&[bad, "-c", "a.c"]);
4879 assert!(failed.contains("link time jobs"), "{bad}: {failed}");
4880 }
4881
4882 // How the program is cut up before the work is spread over it.
4883 assert_eq!(compile(&["-c", "a.c"]).0.lto.partition, Partition::Balanced, "gcc's default");
4884 for (spelling, want) in [
4885 ("balanced", Partition::Balanced),
4886 ("1to1", Partition::OneToOne),
4887 ("one", Partition::One),
4888 ("max", Partition::Max),
4889 ("none", Partition::None),
4890 ] {
4891 let (opts, _) = compile(&[&format!("-flto-partition={spelling}"), "-c", "a.c"]);
4892 assert_eq!(opts.lto.partition, want, "{spelling}");
4893 }
4894 assert!(refused(&["-flto-partition=big", "-c", "a.c"]).contains("partitioning model"));
4895
4896 // And how hard the bytecode is compressed on its way into the object, which is zstd's
4897 // range of levels and is the range gcc checks an argument against.
4898 assert_eq!(compile(&["-c", "a.c"]).0.lto.compression, None, "whatever it does by default");
4899 assert_eq!(compile(&["-flto-compression-level=0", "-c", "a.c"]).0.lto.compression, Some(0));
4900 let (opts, _) = compile(&["-flto-compression-level=19", "-c", "a.c"]);
4901 assert_eq!(opts.lto.compression, Some(19));
4902 for bad in ["-flto-compression-level=20", "-flto-compression-level=-1"] {
4903 let failed = refused(&[bad, "-c", "a.c"]);
4904 assert!(failed.contains("compression level"), "{bad}: {failed}");
4905 }
4906
4907 // The two pairs that describe an arrangement rather than ask for one. Every object here
4908 // holds its machine code, so the fat spelling is what already happens and the other is a
4909 // smaller file rather than a different program, and the plugin pair is about a tool the
4910 // design in `spec/09-optimizer.md` never loads.
4911 for taken in [
4912 "-ffat-lto-objects",
4913 "-fno-fat-lto-objects",
4914 "-fuse-linker-plugin",
4915 "-fno-use-linker-plugin",
4916 ] {
4917 let (opts, _) = compile(&[taken, "-c", "a.c"]);
4918 assert!(!opts.lto.requested, "{taken} says nothing about whether to do it");
4919 }
4920 }
4921
4922 /// The profile family, which is the only one here that splits down the middle.
4923 ///
4924 /// Reading a profile is taken and writing one is refused, and the line between them is the one
4925 /// section 4.1 draws: ignoring a request to read the counts gives a correct program that is
4926 /// slower than it could have been, and ignoring a request to write them means a file the build
4927 /// declared as an output never appears.
4928 #[test]
4929 fn reading_a_profile_is_taken_and_writing_one_is_refused() {
4930 let (opts, _) = compile(&["-c", "a.c"]);
4931 assert!(!opts.profile_data.requested, "nothing asks unless the command line does");
4932 assert_eq!(opts.profile_data.path, None);
4933
4934 let (opts, _) = compile(&["-fprofile-use", "-c", "a.c"]);
4935 assert!(opts.profile_data.requested);
4936 assert_eq!(opts.profile_data.path, None, "beside the object, the way gcc looks");
4937
4938 let (opts, _) = compile(&["-fprofile-use=/counts", "-c", "a.c"]);
4939 assert!(opts.profile_data.requested, "naming a path asks for it too");
4940 assert_eq!(opts.profile_data.path.as_deref(), Some("/counts"));
4941
4942 // The last of the two directions wins, the same as every other pair of `-f` spellings.
4943 assert!(
4944 !compile(&["-fprofile-use", "-fno-profile-use", "-c", "a.c"]).0.profile_data.requested
4945 );
4946 assert!(
4947 compile(&["-fno-profile-use", "-fprofile-use", "-c", "a.c"]).0.profile_data.requested
4948 );
4949
4950 // The rest of the reading half, which is where the files are and three answers about what
4951 // to make of what is in them.
4952 let (opts, _) = compile(&[
4953 "-fprofile-dir=/build/profiles",
4954 "-fprofile-abs-path",
4955 "-fprofile-correction",
4956 "-fprofile-partial-training",
4957 "-c",
4958 "a.c",
4959 ]);
4960 assert_eq!(opts.profile_data.dir.as_deref(), Some("/build/profiles"));
4961 assert!(opts.profile_data.absolute);
4962 assert!(opts.profile_data.correction);
4963 assert!(opts.profile_data.partial_training);
4964
4965 // Writing one, which is refused by name. The first four instrument the program and the
4966 // last writes a file beside the object, and a build that got neither and no message would
4967 // go on to optimize against counts that were never gathered.
4968 for writing in [
4969 "-fprofile-generate",
4970 "-fprofile-generate=/build/profiles",
4971 "-fprofile-arcs",
4972 "--coverage",
4973 "-fcondition-coverage",
4974 "-fpath-coverage",
4975 ] {
4976 let failed = refused(&[writing, "-c", "a.c"]);
4977 assert!(failed.contains("instrument"), "{writing}: {failed}");
4978 }
4979 assert!(refused(&["-ftest-coverage", "-c", "a.c"]).contains(".gcno"), "it names the file");
4980
4981 // The negative spellings of the refused half are what already happens, so they are taken.
4982 for taken in ["-fno-profile-generate", "-fno-profile-arcs", "-fno-test-coverage"] {
4983 let (opts, _) = compile(&[taken, "-c", "a.c"]);
4984 assert!(!opts.profile_data.requested, "{taken} asks for nothing");
4985 }
4986
4987 // And the flags that describe the instrumentation that is refused above, which are checked
4988 // and dropped. Checked because a typo is worth finding here rather than on the day the
4989 // instrumentation lands.
4990 for taken in [
4991 "-fprofile-update=single",
4992 "-fprofile-update=atomic",
4993 "-fprofile-update=prefer-atomic",
4994 "-fprofile-reproducible=serial",
4995 "-fprofile-reproducible=parallel-runs",
4996 "-fprofile-reproducible=multithreaded",
4997 "-fprofile-values",
4998 "-fno-profile-values",
4999 "-fprofile-info-section",
5000 "-fprofile-filter-files=a.c",
5001 "-fprofile-exclude-files=b.c",
5002 "-fprofile-note=a.gcno",
5003 ] {
5004 let (opts, _) = compile(&[taken, "-c", "a.c"]);
5005 assert!(!opts.profile_data.requested, "{taken} says nothing about reading one");
5006 }
5007 assert!(refused(&["-fprofile-update=none", "-c", "a.c"]).contains("update method"));
5008 assert!(refused(&["-fprofile-reproducible=any", "-c", "a.c"]).contains("reproducibility"));
5009 }
5010
5011 /// The sanitizers, which are refused by name and are the one family refused for a reason that
5012 /// is not about the bytes.
5013 ///
5014 /// A sanitizer is a promise that the program is watched while it runs, so a build that asked
5015 /// for one and was quietly given a program with no checks in it gets a test suite that passes
5016 /// for the wrong reason rather than a slower program.
5017 #[test]
5018 fn a_sanitizer_that_is_still_asked_for_at_the_end_of_the_line_is_refused_by_name() {
5019 for asked in ["address", "undefined", "thread", "kernel-address", "leak", "memory"] {
5020 let failed = refused(&[&format!("-fsanitize={asked}"), "-c", "a.c"]);
5021 assert!(failed.contains(asked), "the refusal names what was asked for: {failed}");
5022 assert!(failed.contains("-fsafety=detect"), "and the nearest thing: {failed}");
5023 }
5024
5025 // A list is every name in it, and the first one still standing is the one named.
5026 let failed = refused(&["-fsanitize=address,undefined", "-c", "a.c"]);
5027 assert!(failed.contains("address"), "{failed}");
5028
5029 // A name that is not one, which is worth its own message: somebody who wrote `-fsanitize`
5030 // with a typo in it has a different problem from somebody who wrote a real one.
5031 for bad in ["-fsanitize=bogus", "-fsanitize=address,bogus", "-fno-sanitize=bogus"] {
5032 let failed = refused(&[bad, "-c", "a.c"]);
5033 assert!(failed.contains("is not a sanitizer"), "{bad}: {failed}");
5034 }
5035
5036 // gcc takes `all` only in the negative, and so does this.
5037 assert!(refused(&["-fsanitize=all", "-c", "a.c"]).contains("only `-fno-sanitize=all`"));
5038
5039 // Asking and then taking it back is asking for nothing, which is why the answer waits for
5040 // the end of the line. A build whose shared flags turn a check on and whose rule for one
5041 // file turns it off again compiles that file here.
5042 for pair in [
5043 ["-fsanitize=address", "-fno-sanitize=address"],
5044 ["-fsanitize=address,undefined", "-fno-sanitize=all"],
5045 ["-fsanitize=undefined", "-fno-sanitize=undefined"],
5046 ] {
5047 let (opts, _) = compile(&[pair[0], pair[1], "-c", "a.c"]);
5048 assert_eq!(opts.safety, rucc_session::Safety::Off, "{pair:?} asked for nothing");
5049 }
5050 // And the other order still asks, because the last word is the one that counts.
5051 assert!(!refused(&["-fno-sanitize=address", "-fsanitize=address", "-c", "a.c"]).is_empty());
5052
5053 // What a check does when it fires is an answer about checks that are refused, so there is
5054 // nothing left for it to change and it is taken.
5055 for taken in [
5056 "-fsanitize-recover=undefined",
5057 "-fno-sanitize-recover=all",
5058 "-fsanitize-trap=undefined",
5059 "-fno-sanitize-trap=all",
5060 "-fsanitize-undefined-trap-on-error",
5061 "-fsanitize-address-use-after-scope",
5062 "-fno-sanitize-address-use-after-scope",
5063 "-fsanitize-sections=.data",
5064 ] {
5065 let (opts, _) = compile(&[taken, "-c", "a.c"]);
5066 assert_eq!(opts.safety, rucc_session::Safety::Off, "{taken} asks for no checking");
5067 }
5068 assert!(refused(&["-fsanitize-recover=bogus", "-c", "a.c"]).contains("is not a sanitizer"));
5069
5070 // Coverage instrumentation is refused rather than dropped, because a fuzzer with no
5071 // feedback runs blind and never says so.
5072 let failed = refused(&["-fsanitize-coverage=trace-pc", "-c", "a.c"]);
5073 assert!(failed.contains("feedback"), "{failed}");
5074 let failed = refused(&["-fsanitize-coverage=trace-pc-guard", "-c", "a.c"]);
5075 assert!(failed.contains("trace-pc or trace-cmp"), "gcc takes two of them: {failed}");
5076 }
5077
5078 #[test]
5079 fn the_levels_gcc_spells_differently_are_the_levels_they_mean() {
5080 assert_eq!(compile(&["-O", "-c", "a.c"]).0.opt_level, OptLevel::O1);
5081 assert_eq!(compile(&["-Og", "-c", "a.c"]).0.opt_level, OptLevel::O1);
5082 assert_eq!(compile(&["-O2", "-c", "a.c"]).0.opt_level, OptLevel::O2);
5083 }
5084
5085 #[test]
5086 fn the_machine_flags_that_name_what_we_already_do_are_taken_and_the_rest_are_not() {
5087 let line = ["--target=x86_64-unknown-linux-gnu", "-m64", "-march=x86-64-v3"];
5088 let (opts, _) =
5089 compile(&[&line[..], &["-mtune=native", "-mabi=sysv", "-c", "a.c"]].concat());
5090 assert_eq!(opts.target.to_string(), "x86_64-unknown-linux-gnu");
5091 let wrong = refused(&["--target=x86_64-unknown-linux-gnu", "-mabi=ms", "-c", "a.c"]);
5092 assert!(wrong.contains("sysv convention"), "{wrong}");
5093 }
5094
5095 #[test]
5096 fn the_thread_flag_is_a_macro_and_a_library_and_the_library_goes_last() {
5097 let (opts, plan) = compile(&["-pthread", "-c", "a.c"]);
5098 assert!(opts.defines.iter().any(|d| d == "_REENTRANT"));
5099 // After the input, because a static link takes what it needs from a library when it
5100 // reaches it and not afterwards.
5101 let names: Vec<&str> = plan.jobs.iter().map(|j| j.input.as_str()).collect();
5102 assert_eq!(names, vec!["a.c"]);
5103 }
5104
5105 #[test]
5106 fn the_questions_a_build_system_asks_before_it_compiles_anything() {
5107 let target = "--target=x86_64-unknown-linux-gnu";
5108 assert_eq!(printed(&[target, "-dumpmachine"]), "x86_64-unknown-linux-gnu");
5109 assert_eq!(printed(&[target, "-dumpversion"]), VERSION);
5110 assert_eq!(printed(&[target, "-dumpfullversion"]), VERSION);
5111 assert_eq!(printed(&[target, "-print-multiarch"]), "x86_64-linux-gnu");
5112 // A name nothing holds comes back unchanged, which is GCC's rule and is what makes the
5113 // answer safe to paste into a link line whether or not the file is there.
5114 assert_eq!(printed(&[target, "-print-file-name=no-such-library.a"]), "no-such-library.a");
5115 assert_eq!(printed(&[target, "-print-prog-name=ld"]), "ld");
5116 let dirs = printed(&[target, "-print-search-dirs"]);
5117 assert!(dirs.starts_with("install: "), "{dirs}");
5118 assert!(dirs.contains("\nlibraries: ="), "{dirs}");
5119 }
5120
5121 #[test]
5122 fn the_sysroot_in_effect_is_the_one_the_command_line_named_or_the_one_for_the_target() {
5123 // A tree the user named is the answer whatever the target is, because it is the answer to
5124 // every other question too.
5125 assert_eq!(printed(&["--sysroot=/opt/cross", "-print-sysroot"]), "/opt/cross");
5126
5127 // A target that is no machine this suite runs on is read under the cache, and the answer is
5128 // the root rather than one of the directories under it, since what asks is looking for a
5129 // file of its own.
5130 let root = cache::dir().join("sysroots").join("riscv64-linux-musl");
5131 assert_eq!(
5132 printed(&["--target=riscv64-linux-musl", "-print-sysroot"]),
5133 root.display().to_string()
5134 );
5135
5136 // And a compile for this machine has no sysroot, which is the empty line GCC prints when it
5137 // was configured without one rather than a `/` that would be a claim about the filesystem.
5138 let host = Triple::host().expect("a host this compiler knows");
5139 assert_eq!(printed(&[&format!("--target={host}"), "-print-sysroot"]), "");
5140 }
5141
5142 #[test]
5143 fn the_provenance_of_a_sysroot_is_the_manifest_it_carries() {
5144 // Section 13.5 wants seven things per input and wants them machine readable, and the manifest
5145 // is the record that already has them, so the flag prints that rather than a second format.
5146 let manifest = "rucc sysroot manifest 3\n\
5147 target\tx86_64-linux-musl\n\
5148 kernel\t6.12\n\
5149 include/generic/stdio.h\tmusl-1.2.5\t\
5150 https://musl.libc.org/releases/musl-1.2.5.tar.gz\t\
5151 0000000000000000000000000000000000000000000000000000000000000000\tmit\t\
5152 bundled\n\
5153 lib/libc.so\tmusl-1.2.5\t\
5154 https://musl.libc.org/releases/musl-1.2.5.tar.gz\t\
5155 1111111111111111111111111111111111111111111111111111111111111111\tmit\t\
5156 generated\n";
5157 let tree = TempTree::new("provenance", &[("manifest", manifest)]);
5158 let sysroot = format!("--sysroot={}", tree.0.display());
5159 // The kernel line of tamnd/rucc#934 is in the answer without anything here naming it, because
5160 // the flag parses the record and renders it again rather than picking fields out of it. That
5161 // is the reason it prints a manifest and not a format of its own.
5162 //
5163 // The answer is the file without its last newline, because whatever prints it adds one. The
5164 // file is what somebody diffs the output against, so the two have to be the same bytes.
5165 assert_eq!(printed(&[&sysroot, "-print-sysroot-provenance"]) + "\n", manifest);
5166
5167 // A tree with no manifest in it is a tree somebody assembled themselves, and nothing here
5168 // knows where any of it came from. Saying nothing is the only honest answer, and a reader can
5169 // tell it from a manifest with no inputs because that one still has its two header lines.
5170 let bare = TempTree::new("provenance-bare", &[]);
5171 assert_eq!(
5172 printed(&[&format!("--sysroot={}", bare.0.display()), "-print-sysroot-provenance"]),
5173 ""
5174 );
5175
5176 // And a compile for this machine has no sysroot at all, which is the same empty answer
5177 // `-print-sysroot` gives for it.
5178 let host = Triple::host().expect("a host this compiler knows");
5179 assert_eq!(printed(&[&format!("--target={host}"), "-print-sysroot-provenance"]), "");
5180
5181 // And the other spelling, which section 13.5 is the document that writes.
5182 assert_eq!(printed(&[&sysroot, "--print-sysroot-provenance"]) + "\n", manifest);
5183
5184 // tamnd/rucc#1021. The digest of the same tree is the sha256 of that record, so it is one
5185 // line where the provenance is a few hundred, and it is checkable with `sha256sum` because
5186 // the bytes it is over are the bytes of the file. The number here is that hash of the
5187 // fixture above, computed by `sha256sum` rather than by this compiler.
5188 assert_eq!(
5189 printed(&[&sysroot, "-print-sysroot-digest"]),
5190 "d705ae6ebeafeb7fda4bd57cecc7882bf49784b17015664a09cfae25a1b2000a"
5191 );
5192 assert_eq!(
5193 printed(&[&sysroot, "--print-sysroot-digest"]),
5194 printed(&[&sysroot, "-print-sysroot-digest"])
5195 );
5196
5197 // And the two empty answers are empty here too, because a digest of nothing would read as a
5198 // claim about a sysroot rather than as the absence of one.
5199 assert_eq!(
5200 printed(&[&format!("--sysroot={}", bare.0.display()), "-print-sysroot-digest"]),
5201 ""
5202 );
5203 assert_eq!(printed(&[&format!("--target={host}"), "-print-sysroot-digest"]), "");
5204 }
5205
5206 #[test]
5207 fn a_manifest_this_build_cannot_read_is_refused_rather_than_printed() {
5208 // Passing a file we could not parse to whoever asked would make their parser the one that
5209 // finds the problem, and the three uses section 13.5 gives for this are all somebody else
5210 // parsing it.
5211 let tree = TempTree::new(
5212 "provenance-bad",
5213 &[("manifest", "rucc sysroot manifest 3\ntarget\tx86_64-linux-musl\nlib/libc.a\n")],
5214 );
5215 let message =
5216 refused(&[&format!("--sysroot={}", tree.0.display()), "-print-sysroot-provenance"]);
5217 assert!(message.contains("manifest"), "{message}");
5218 assert!(message.contains("1 fields where an input has six"), "{message}");
5219
5220 // The digest is refused for the same file and for a stronger reason: a hash of bytes this
5221 // build cannot read would be a number that names a record nobody can act on.
5222 let digest =
5223 refused(&[&format!("--sysroot={}", tree.0.display()), "-print-sysroot-digest"]);
5224 assert_eq!(digest, message);
5225 }
5226
5227 #[test]
5228 fn the_two_dependency_flags_that_stop_after_the_rule_stop_after_the_rule() {
5229 let (opts, _) = compile(&["-M", "a.c"]);
5230 assert!(opts.deps.emit && opts.deps.instead_of_compiling);
5231 assert!(opts.deps.system_headers, "plain -M lists them");
5232 assert_eq!(opts.emit, EmitKind::Preprocessed);
5233
5234 // Even where a later flag asked for something else, because the family is a mode and
5235 // the mode is what the run is for.
5236 let (opts, _) = compile(&["-M", "-c", "a.c"]);
5237 assert_eq!(opts.emit, EmitKind::Preprocessed);
5238
5239 let (opts, _) = compile(&["-MM", "a.c"]);
5240 assert!(!opts.deps.system_headers);
5241 }
5242
5243 #[test]
5244 fn the_two_that_end_in_d_leave_the_compilation_alone() {
5245 let (opts, _) = compile(&["-MD", "-c", "a.c"]);
5246 assert!(opts.deps.emit && !opts.deps.instead_of_compiling);
5247 assert!(opts.deps.system_headers);
5248 assert_eq!(opts.emit, EmitKind::Object);
5249
5250 let (opts, _) = compile(&["-MMD", "-c", "a.c"]);
5251 assert!(opts.deps.emit && !opts.deps.instead_of_compiling);
5252 assert!(!opts.deps.system_headers);
5253 }
5254
5255 #[test]
5256 fn nothing_puts_the_system_headers_back_once_a_flag_has_taken_them_out() {
5257 // GCC's rule, and not an oversight in it. The flag asking for fewer of them is read as
5258 // the answer, because the other one never asked the question.
5259 let (opts, _) = compile(&["-MM", "-M", "a.c"]);
5260 assert!(!opts.deps.system_headers);
5261 let (opts, _) = compile(&["-MD", "-MMD", "-c", "a.c"]);
5262 assert!(!opts.deps.system_headers);
5263 let (opts, _) = compile(&["-MMD", "-MD", "-c", "a.c"]);
5264 assert!(!opts.deps.system_headers);
5265 }
5266
5267 #[test]
5268 fn a_target_arrives_escaped_from_one_flag_and_untouched_from_the_other() {
5269 let (opts, _) = compile(&["-MM", "-MT", "a b.o", "-MQ", "a b.o", "a.c"]);
5270 assert_eq!(opts.deps.targets, vec!["a b.o".to_owned(), "a\\ b.o".to_owned()]);
5271 }
5272
5273 #[test]
5274 fn the_rest_of_the_family_is_a_file_and_a_switch() {
5275 let (opts, _) = compile(&["-MM", "-MF", "dep.d", "-MP", "a.c"]);
5276 assert_eq!(opts.deps.file.as_deref(), Some("dep.d"));
5277 assert!(opts.deps.phony);
5278
5279 for flag in ["-MF", "-MT", "-MQ"] {
5280 let e = parse_args(&args(&[flag])).unwrap_err();
5281 assert!(e.message.contains("requires an argument"), "{}", e.message);
5282 }
5283 }
5284
5285 /// A directory of sources for one test, removed when the test is done with it.
5286 struct TempTree(PathBuf);
5287
5288 impl Drop for TempTree {
5289 fn drop(&mut self) {
5290 let _ = std::fs::remove_dir_all(&self.0);
5291 }
5292 }
5293
5294 impl TempTree {
5295 fn new(name: &str, files: &[(&str, &str)]) -> TempTree {
5296 let dir = std::env::temp_dir().join(format!("rucc-deps-{}-{name}", std::process::id()));
5297 let _ = std::fs::remove_dir_all(&dir);
5298 std::fs::create_dir_all(&dir).expect("temporary directory should be writable");
5299 for (path, text) in files {
5300 let at = dir.join(path);
5301 if let Some(parent) = at.parent() {
5302 std::fs::create_dir_all(parent).expect("creating a subdirectory should work");
5303 }
5304 std::fs::write(&at, text).expect("writing a temporary file should work");
5305 }
5306 TempTree(dir)
5307 }
5308
5309 fn path(&self, name: &str) -> String {
5310 self.0.join(name).to_string_lossy().into_owned()
5311 }
5312 }
5313
5314 #[test]
5315 fn the_rule_names_what_the_includes_found_and_names_each_of_them_once() {
5316 // End to end, because the list comes from the preprocessor and the format comes from
5317 // somewhere else, and a test of either half on its own would pass with the two of them
5318 // wired up backwards.
5319 let tree = TempTree::new(
5320 "found",
5321 &[
5322 ("a.c", "#include \"one.h\"\n#include \"two.h\"\nint main(void) { return X; }\n"),
5323 ("one.h", "#define X 0\n"),
5324 ("two.h", "#include \"one.h\"\n"),
5325 ],
5326 );
5327 let out = tree.path("dep.d");
5328 let code = run(&args(&["-MM", "-MF", &out, "-o", &tree.path("a.i"), &tree.path("a.c")]));
5329 assert_eq!(code, 0);
5330
5331 let text = std::fs::read_to_string(&out).expect("the rule should have been written");
5332 let names: Vec<&str> = text.split_whitespace().collect();
5333 // The target, the source, and each header once however many times it was reached.
5334 assert_eq!(names.first(), Some(&"a.o:"), "{text}");
5335 assert_eq!(names.iter().filter(|n| n.ends_with("one.h")).count(), 1, "{text}");
5336 assert_eq!(names.iter().filter(|n| n.ends_with("two.h")).count(), 1, "{text}");
5337 // And the `-o` went to the file the rule replaced, which is left empty rather than
5338 // absent because a makefile that named it as a target will look for it.
5339 assert_eq!(std::fs::read(tree.path("a.i")).expect("the output should exist"), b"");
5340 }
5341
5342 #[test]
5343 fn a_header_that_is_only_reached_under_a_guard_is_still_a_dependency() {
5344 // The multiple-include optimization means the second reach never opens the file. It is
5345 // still a file this translation unit was built from, so it is still in the rule.
5346 let tree = TempTree::new(
5347 "guarded",
5348 &[
5349 ("a.c", "#include \"g.h\"\n#include \"g.h\"\nint main(void) { return 0; }\n"),
5350 ("g.h", "#ifndef G\n#define G\n#endif\n"),
5351 ],
5352 );
5353 let out = tree.path("dep.d");
5354 let code = run(&args(&["-MM", "-MF", &out, "-o", &tree.path("a.i"), &tree.path("a.c")]));
5355 assert_eq!(code, 0);
5356 let text = std::fs::read_to_string(&out).expect("the rule should have been written");
5357 assert_eq!(text.split_whitespace().filter(|n| n.ends_with("g.h")).count(), 1, "{text}");
5358 }
5359
5360 #[test]
5361 fn every_imacros_file_is_read_before_every_include_file_whatever_order_they_were_written() {
5362 // Measured against GCC rather than read: the two flags the other way round produce the
5363 // same output byte for byte, so the command line order between the two families does not
5364 // decide anything and the order within one does. The `-include` file here can only see
5365 // the definition if the `-imacros` file that was written after it ran first.
5366 let tree = TempTree::new(
5367 "preinclude",
5368 &[
5369 ("a.c", "int main(void) { return 0; }\n"),
5370 ("i.h", "#ifdef FROM_MACROS\nint saw_it;\n#else\nint missed_it;\n#endif\n"),
5371 ("m.h", "#define FROM_MACROS 1\nint macros_text;\n"),
5372 ],
5373 );
5374 let out = tree.path("a.i");
5375 let code = run(&args(&[
5376 "-E",
5377 "-include",
5378 &tree.path("i.h"),
5379 "-imacros",
5380 &tree.path("m.h"),
5381 "-o",
5382 &out,
5383 &tree.path("a.c"),
5384 ]));
5385 assert_eq!(code, 0);
5386 let text = std::fs::read_to_string(&out).expect("the output should have been written");
5387 assert!(text.contains("saw_it"), "{text}");
5388 // And the text of the `-imacros` file is thrown away, which is the whole difference
5389 // between the two flags.
5390 assert!(!text.contains("macros_text"), "{text}");
5391 }
5392
5393 #[test]
5394 fn a_file_the_command_line_named_is_a_prerequisite_the_same_as_one_a_directive_named() {
5395 let tree = TempTree::new(
5396 "preinclude-deps",
5397 &[
5398 ("a.c", "int main(void) { return 0; }\n"),
5399 ("i.h", "int from_include;\n"),
5400 ("m.h", "#define M 1\n"),
5401 ],
5402 );
5403 let out = tree.path("dep.d");
5404 let code = run(&args(&[
5405 "-MM",
5406 "-MF",
5407 &out,
5408 "-include",
5409 &tree.path("i.h"),
5410 "-imacros",
5411 &tree.path("m.h"),
5412 "-o",
5413 &tree.path("a.i"),
5414 &tree.path("a.c"),
5415 ]));
5416 assert_eq!(code, 0);
5417 let text = std::fs::read_to_string(&out).expect("the rule should have been written");
5418 assert!(text.contains("i.h"), "{text}");
5419 assert!(text.contains("m.h"), "{text}");
5420 }
5421
5422 #[test]
5423 fn a_command_line_include_that_is_nowhere_on_the_path_is_an_error_and_not_a_warning() {
5424 // Including the directory of the source file, which is not on the path for these: the
5425 // command line was not written there, so a name in it is relative to where the compiler
5426 // was run rather than to where the source sits.
5427 let tree = TempTree::new(
5428 "preinclude-missing",
5429 &[("sub/a.c", "int main(void) { return 0; }\n"), ("sub/beside.h", "int x;\n")],
5430 );
5431 let code = run(&args(&["-E", "-include", "beside.h", "-o", "-", &tree.path("sub/a.c")]));
5432 assert_eq!(code, 1);
5433 }
5434
5435 #[test]
5436 fn a_command_line_that_links_names_the_executable_and_not_the_object_it_went_through() {
5437 // The object a link goes through is in a temporary directory and is gone before `make`
5438 // reads any of this, so the rule that named it would be a rule for a file that is never
5439 // there. The target and the file are both the `-o`, which is the executable.
5440 let (opts, plan) = compile(&["-MD", "sub/a.c", "-o", "prog"]);
5441 assert_eq!(plan.output.as_deref(), Some("prog"));
5442 assert_eq!(deps::default_target("sub/a.c", deps_target_output(&opts, &plan)), "prog");
5443 assert_eq!(
5444 deps::default_file(&opts.deps, "sub/a.c", plan.output.as_deref()).as_deref(),
5445 Some("prog.d")
5446 );
5447 }
5448
5449 #[test]
5450 fn the_plan_keeps_the_output_name_because_the_rule_is_written_from_it() {
5451 let (_, plan) = compile(&["-MMD", "-c", "sub/a.c", "-o", "obj/x.o"]);
5452 assert_eq!(plan.output.as_deref(), Some("obj/x.o"));
5453 let (_, plan) = compile(&["-MMD", "-c", "sub/a.c"]);
5454 assert_eq!(plan.output, None);
5455 }
5456
5457 #[test]
5458 fn usage_fits_on_a_screen() {
5459 // Not a style preference. A help text that scrolls is one nobody reads, and this is
5460 // the cheapest way to keep it honest as flags accumulate. The number goes up only when
5461 // a family of flags arrives that has nowhere to share a line, which the two pass gates
5462 // were and which the two fuel flags and `-fsafety=` now are, and it goes up by exactly
5463 // the lines that family took. The four it went up by last are the flags a build system
5464 // passes without being asked to: how much to say, what machine to generate for, threads,
5465 // and the questions `configure` asks before it compiles anything. The one it went up by
5466 // last is the second line of `--emit`, whose kinds are a family that has now outgrown
5467 // one line and has nowhere else to go. The two it went up by last are the dependency
5468 // family, which is eight flags that share nothing with anything above them. The one it
5469 // went up by last is the four spellings of position independent code, which every
5470 // configure script writes and which could only have shared the link line, and that line
5471 // is already four characters short of the limit. The two it went up by last are the rest
5472 // of the include family, which is six more flags that change where a header is looked for
5473 // and two that name a header outright. The one it went up by last is the pair that keeps
5474 // the intermediate files and times the steps, which belong next to the two flags above
5475 // them that are also about watching a compilation rather than changing one. The two it
5476 // went up by last are the section flags and the visibility flag, which are what a build
5477 // that cares about the size of what it ships and about which names it exports writes, and
5478 // the second of them was already taken and only missing from here. The one it went up by
5479 // last is the stack protector, which is four spellings of one question and which every
5480 // distribution puts on every command line it issues, so a build that reads this list
5481 // looking for it and does not find it has to go and read the specification instead. The one
5482 // it went up by last is the profiler, which is two spellings of the request and two of
5483 // where the call goes, and which is about watching a program run rather than about what is
5484 // generated, so it shares its subject with nothing above it. The one it went up by last is
5485 // the room a function opens with for something to be written over it later, which takes an
5486 // argument of its own shape and is what a kernel build asks for, so it fits beside the
5487 // profiler and nothing else. The one it went up by last is what overflows rather than being
5488 // undefined, which is three spellings of two questions and which a kernel build and a great
5489 // deal of code written before the standard settled both pass. The one it went up by last is
5490 // the other answer to the first of those questions, which could not share the line because
5491 // what it asks for is the opposite of what the flags on that line ask for. The one it went
5492 // up by last is the split of the line that lists what this compiler does anyway into that
5493 // and what it assumes anyway, which are two different claims that were sharing a line until
5494 // the second of them got a second flag and the line stopped fitting. The one it went up by
5495 // last is the three flags that change the ABI rather than the code, which have to be given
5496 // to every file in a program or none of them and which therefore belong somewhere a person
5497 // reading this list will see them. The one it went up by last is the floating point group,
5498 // which is two lines rather than one because the first of them is a choice this compiler
5499 // records and the rest are claims about what it does anyway, and putting a real setting on
5500 // the same line as three flags that change nothing would be misleading about both. The one
5501 // it went up by last is the flag that says a write has to stay inside the member it names,
5502 // which is a setting rather than a claim and so cannot share the line above it, that being
5503 // the one that picks a tier. The two it went up by last are the prefix mapping family,
5504 // which is four flags whose whole job is to keep a build's output the same from two
5505 // different directories, and which a person chasing a reproducible build comes here
5506 // looking for by name. The one it went up by last is how the debug sections are compressed
5507 // and whether they go in a file of their own, which are two questions about the shape of
5508 // the debug output, where the line above them is about how much of it there is. The one it
5509 // went up by last is the `restrict` contract, which is a setting for the same reason the
5510 // flag that keeps a write inside its member is and which is the check a person who has been
5511 // bitten by a vectorizer comes here looking for. The one it went up by last is link time
5512 // optimization, which is a whole optimization rather than a flag and which says so on its
5513 // own line, because a build that passes it and reads this looking for what it got is
5514 // asking a question no other line here answers. The one it went up by last is the sysroot,
5515 // which is the question somebody asks when a cross build read a file nobody expected, and
5516 // which has no room on the line above it because the answers there are a path each and this
5517 // one is the root all of them are under. The one it went up by last is what is inside that
5518 // root and where each of it came from, which is a question about a whole tree rather than
5519 // about a path and which is long enough on its own that it could not have shared a line with
5520 // anything. The one it went up by last is the profile family, which splits down the middle
5521 // where no other family here does, so the line has to name the half that is taken and the
5522 // half that is refused or it would be read as taking both. The one it went up by last is
5523 // the sanitizers, which are what somebody reaching for a checked build writes first and
5524 // which belong beside the tier that is the nearest thing here to what they asked for. The
5525 // one it went up by last is the digest of that record, which is the same tree as one number
5526 // and could not share the line above it because that line prints a few hundred lines and
5527 // this one prints sixty four characters, and a reader who wants the short answer is looking
5528 // for it by name rather than reading the long one. The one it went up by last is the
5529 // sysroot fetch, which is the only command here that gets something from somewhere else and
5530 // is therefore the one a person wants to have read before they run it rather than after.
5531 // And the flag beside it that forbids every download, which earns its line by being what a
5532 // build in a sealed environment passes and by meaning something even though an ordinary
5533 // compile downloads nothing either way.
5534 assert!(USAGE.lines().count() < 72, "usage text has grown past one screen");
5535 }
5536}