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/0.10.21")]
30
31pub mod cache;
32pub mod compile;
33pub mod deps;
34pub mod library;
35pub mod link;
36mod map;
37pub mod phase;
38pub mod preprocess;
39pub mod schedule;
40
41use std::fmt::Write as _;
42use std::io::Write as _;
43use std::path::PathBuf;
44
45use rucc_codegen::coverage::{self, Fired};
46use rucc_codegen::pressure::Pressure;
47use rucc_pp::Dependency;
48use rucc_session::{
49 Compress, Control, Dumps, EmitKind, Hook, Options, Pic, PrefixMap, Preinclude, Protector,
50 SaveTemps, Session, Std, Wrapping, runtime,
51};
52use rucc_sysroot::{Manifest, Sysroot};
53use rucc_target::Triple;
54
55use crate::link::LinkOptions;
56
57pub use crate::compile::{Artifact, Compiled, Temps, compile, compile_ir};
58pub use crate::phase::{Input, InputKind, Job, LinkJob, Output, Phase, Plan};
59pub use crate::preprocess::{OsFileSystem, Preprocessed, preprocess};
60pub use crate::schedule::Jobs;
61
62/// The compiler's version, taken from the workspace manifest.
63pub const VERSION: &str = env!("CARGO_PKG_VERSION");
64
65/// What the command line asked for.
66#[derive(Debug, Clone, PartialEq, Eq)]
67pub enum Action {
68 /// Print usage and exit successfully.
69 Help,
70 /// Print the version and exit successfully.
71 Version,
72 /// Print one line and exit successfully, which is what the `-dump` and `-print` family do.
73 ///
74 /// A build system asks these before it compiles anything, and what it does with the answer
75 /// is paste it into a path or into another command line, so each one is a single line with
76 /// no decoration around it.
77 Print(String),
78 /// Print the resolved configuration and exit successfully.
79 PrintConfig(Box<Options>),
80 /// Print the passes the level will run and exit successfully.
81 PrintPipeline(Box<Options>),
82 /// Print the phase plan and the link line and exit successfully, which is `-###`.
83 PrintPlan {
84 /// The resolved options, which is what says what the link line is for.
85 opts: Box<Options>,
86 /// What to do to each input, and in what order.
87 plan: Box<Plan>,
88 /// What the command line said about linking.
89 link: Box<LinkOptions>,
90 },
91 /// Compile the given inputs.
92 Compile {
93 /// The resolved options.
94 opts: Box<Options>,
95 /// What to do to each input, and in what order.
96 plan: Box<Plan>,
97 /// What the command line said about linking.
98 link: Box<LinkOptions>,
99 /// How many translation units to compile at once.
100 jobs: Jobs,
101 /// Whether `-v` asked for the plan to be printed while it runs.
102 verbose: bool,
103 },
104}
105
106/// Why a command line was rejected.
107#[derive(Debug, Clone, PartialEq, Eq)]
108pub struct CliError {
109 /// The message, lowercase and without a trailing period, in the same shape as any other
110 /// diagnostic.
111 pub message: String,
112}
113
114impl std::fmt::Display for CliError {
115 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
116 f.write_str(&self.message)
117 }
118}
119
120impl std::error::Error for CliError {}
121
122fn err(message: impl Into<String>) -> CliError {
123 CliError { message: message.into() }
124}
125
126/// The two halves of one prefix mapping flag's argument, where `flag` includes its trailing `=`.
127///
128/// The split is at the last `=` in what follows the flag, not the first, which is gcc's rule and
129/// the only one that lets a directory whose name contains an `=` be the old half. It also means
130/// `-fmacro-prefix-map=a=b=c` rewrites `a=b` to `c` rather than `a` to `b=c`, which looks like a
131/// trap until you notice the alternative traps the far more common case.
132fn rewrite<'a>(arg: &'a str, flag: &str) -> Result<(&'a str, &'a str), CliError> {
133 let rest = &arg[flag.len()..];
134 PrefixMap::split(rest).ok_or_else(|| {
135 let flag = flag.trim_end_matches('=');
136 err(format!(
137 "`{rest}` is not a rewrite for `{flag}`, which is an old prefix, an `=` and a new one"
138 ))
139 })
140}
141
142/// A question the command line asked instead of asking for a compilation.
143///
144/// These are answered after the loop rather than where they are read, because every one of them
145/// is about the target or about the library search and the last word on both is the end of the
146/// command line.
147enum Query {
148 /// `-dumpmachine`, the triple.
149 Machine,
150 /// `-dumpversion` and `-dumpfullversion`, which are the same three numbers here.
151 Version,
152 /// `-print-multiarch`, the directory name a distribution files this target under.
153 Multiarch,
154 /// `-print-search-dirs`, in the three lines GCC prints.
155 SearchDirs,
156 /// `-print-sysroot`, the root the headers and the libraries are read under.
157 Sysroot,
158 /// `-print-sysroot-provenance`, what is in that root and where each of it came from.
159 SysrootProvenance,
160 /// `-print-file-name=<name>`, the full path of a library file.
161 FileName(String),
162 /// `-print-prog-name=<name>`, the full path of a program.
163 ProgName(String),
164 /// `-print-libgcc-file-name`, which is `-print-file-name=libgcc.a` under another spelling.
165 Libgcc,
166}
167
168/// Usage text.
169///
170/// Deliberately short. `spec/04-driver-and-cli.md` puts the full flag reference in the
171/// manual page, because a `--help` nobody can read in one screen is a `--help` nobody reads.
172pub const USAGE: &str = "\
173rucc, an optimizing C compiler
174
175usage: rucc [options] file...
176
177options:
178 -c compile and assemble, do not link
179 -S compile only, emit assembly
180 -E preprocess only
181 -o <file> write output to <file>, or to standard output for -
182 -D <name>[=<value>], -U <name> define a macro, or undefine one after every -D
183 -I <dir> add <dir> to the include search path
184 -iquote -isystem -idirafter <dir> the other chains, -nostdinc drops ours
185 -I-, -iprefix <p>, -iwithprefix[before] <dir> the older spellings of those
186 -include <file>, -imacros <file> read <file> first, the second for its macros only
187 --sysroot=<dir> look for the library's headers under <dir>, -isysroot too
188 -P, -dM with -E: leave out the markers, or dump the macros
189 -M -MM -MD -MMD write a make rule for the source, the last two compile as well
190 -MF <file> -MT <t> -MQ <t> -MP where the rule goes, what it builds, targets with no recipe
191 -std=<dialect> c89 through c23, and the gnu spellings
192 -fgnuc-version=<v> the GCC release to claim, default 7.0.0
193 -x <lang> treat later inputs as <lang>, or none to stop
194 -O<level> optimize: 0, 1, 2, 3, s, z
195 -fsafety=<tier> check memory safety: off, detect, enforce, kernel
196 -f[no-]sanitize=<what> the negative is taken, the positive is refused by name
197 -f[no-]safety-subobject a write has to stay inside the member it names
198 -f[no-]safety-restrict two restrict pointers of one block may not meet
199 -f<pass> -fno-<pass> -fdump-ir=<what> -fopt-info[-<kind>][=FILE]
200 -fpass-fuel=<pass>=<n>, -fpass-fuel-global=<n> stop a pass, or all of them, after n
201 -fdisable-<pass>[=<funcs>], -fenable-<pass>[=<funcs>] run a pass on some functions only
202 -g -g0 -gdwarf-5, -fno-omit-frame-pointer, -mno-red-zone debug info, frame pointer, red zone
203 -gz[=none|zlib|zlib-gnu|zstd] -gno-split-dwarf compress debug sections, one file not two
204 -flto[=auto|jobserver|<n>] -fno-lto -ffat-lto-objects read, and not done yet
205 -fprofile-use[=<path>] -fprofile-dir=<dir> read too, where -fprofile-generate is refused
206 -f[no-]stack-protector[-strong|-all], -f[no-]stack-clash-protection, -fcf-protection=<edges>
207 -ffunction-sections -fdata-sections a section per function or variable, for --gc-sections
208 -fvisibility=<what> default, hidden, internal or protected, when nothing in the source said
209 -l<name>, -L <dir>, -B <dir> link a library, where to look for one, where our own tools are
210 -fPIC -fpic -fPIE -fpie, -fno-common, -pipe what it does anyway
211 -f[no-]strict-aliasing, -f[no-]delete-null-pointer-checks what it assumes anyway
212 -static -shared -pie -no-pie -nostdlib -nostartfiles -nodefaultlibs -rdynamic -s how to link
213 -Wl,<arg>, -Xlinker <arg>, -fuse-ld=<name> hand an argument to the linker, or pick one
214 -Werror -pedantic -pedantic-errors -w how much to say, and whether it is fatal
215 -m64 -march= -mtune= -mcpu= -mabi= -mcmodel= what machine to generate for
216 -pg -p, -mfentry -mno-fentry call a profiler on the way in, and where that call goes
217 -fpatchable-function-entry=<n>[,<m>] room at the top of every function to patch later
218 -fwrapv, -fwrapv-pointer, -fno-strict-overflow signed or pointer overflow wraps
219 -ftrapv signed overflow stops the program instead
220 -f[no-]signed-char, -f[no-]unsigned-char, -f[no-]short-enums change the ABI
221 -ffp-contract=<how> fuse a multiply and an addition: fast, on or off
222 -fexcess-precision=<how>, -f[no-]rounding-math, -f[no-]trapping-math what it does anyway
223 -ffile-prefix-map=<old>=<new> rewrite that front of every path we put in the output
224 -fmacro-prefix-map= -fdebug-prefix-map= -fprofile-prefix-map= the same, one output each
225 -pthread build for more than one thread, and link the library for it
226 -dumpmachine -dumpversion -print-multiarch -print-search-dirs what this compiler is
227 -print-file-name=<name> -print-prog-name=<name> where a file or a program is
228 -print-sysroot the root the headers and the libraries are read under
229 -print-sysroot-provenance every input under it, where it came from and its licence
230 -j[n] compile n translation units at once, default all
231 -v, -### print each phase as it runs, or without running any
232 -save-temps[=cwd|obj], -time keep the .i and the .s, say how long each step took
233 --target=<triple> generate code for <triple>
234 --emit=<kind> exe, obj, asm, preprocessed, tast, ir, mir-final,
235 safety-summary, type-granules
236 --print-config, --print-pipeline print the configuration or the pipeline, and exit
237 --version print the version and exit
238 -h, --help print this message and exit
239
240See spec/04-driver-and-cli.md for the full flag reference.
241";
242
243/// The argument of a flag that may be joined to it or may be the next word.
244///
245/// `-DFOO` and `-D FOO` are the same thing, and `at` is where the flag's own letters end.
246fn joined_or_next(
247 arg: &str,
248 at: usize,
249 args: &[String],
250 i: &mut usize,
251) -> Result<String, CliError> {
252 if arg.len() > at {
253 return Ok(arg[at..].to_owned());
254 }
255 let next = args.get(*i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
256 *i += 1;
257 Ok(next.clone())
258}
259
260/// Every name that may follow `-fsanitize=`, which is gcc 16's list and three of this compiler's
261/// own.
262///
263/// The three are on it because `spec/07-types-and-semantics.md` section 7.7 already promises them:
264/// each undefined behaviour this compiler exploits is listed there with the check that detects it,
265/// and `alias`, `restrict` and `memory` are checks gcc has no spelling for. gcc refuses `memory`
266/// outright, since the sanitizer of that name is clang's. A name being here means it is a name
267/// rather than a typo, and nothing more than that: every one of them is refused after the loop,
268/// because none of them is implemented.
269///
270/// `all` is deliberately absent. gcc takes it only in the negative, so it is handled where each of
271/// those two spellings is read rather than by being on this list.
272const SANITIZERS: [&str; 34] = [
273 "address",
274 "kernel-address",
275 "hwaddress",
276 "kernel-hwaddress",
277 "pointer-compare",
278 "pointer-subtract",
279 "thread",
280 "leak",
281 "undefined",
282 "shift",
283 "shift-base",
284 "shift-exponent",
285 "integer-divide-by-zero",
286 "unreachable",
287 "vla-bound",
288 "null",
289 "return",
290 "signed-integer-overflow",
291 "bounds",
292 "bounds-strict",
293 "alignment",
294 "object-size",
295 "float-divide-by-zero",
296 "float-cast-overflow",
297 "nonnull-attribute",
298 "returns-nonnull-attribute",
299 "bool",
300 "enum",
301 "vptr",
302 "pointer-overflow",
303 "builtin",
304 "alias",
305 "restrict",
306 "memory",
307];
308
309/// Parses a command line, without the program name.
310///
311/// # Errors
312///
313/// Returns the message to print when the arguments do not name a compilation this compiler
314/// can attempt.
315pub fn parse_args(args: &[String]) -> Result<Action, CliError> {
316 let host = Triple::host()
317 .ok_or_else(|| err("this host is not a supported target and no --target was given"))?;
318 let mut opts = Options::new(host);
319 let mut inputs: Vec<Input> = Vec::new();
320 let mut print_config = false;
321 let mut print_pipeline = false;
322 let mut print_plan = false;
323 let mut verbose = false;
324 let mut jobs = Jobs::default();
325 let mut nostdinc = false;
326 let mut sysroot: Option<PathBuf> = None;
327 // The whole ten field target, kept beside the three field one because `--target=` can pin a
328 // libc version and `Triple` has nowhere to put it. It decides `__GLIBC_MINOR__` and nothing
329 // else today, and `None` is a command line that named no target, which is this machine.
330 let mut pinned: Option<rucc_tuple::TargetTuple> = None;
331 let mut output = None;
332 let mut link = LinkOptions::default();
333 let mut query: Option<Query> = None;
334 let mut threads = false;
335 // Which sanitizers are still asked for by the end of the command line. Accumulated across the
336 // loop rather than answered where it was read, because `-fno-sanitize=` turns one off and a
337 // build that asks for a check and then takes it back has asked for nothing. What happens to a
338 // set that is not empty is decided after the loop.
339 let mut sanitizers: Vec<&str> = Vec::new();
340 // `-x` applies to inputs that come after it and stays in effect until the next one, which
341 // is why it is tracked across the loop rather than attached to a single argument.
342 let mut forced: Option<InputKind> = None;
343 // What `-iprefix` last said, stuck on the front of every later `-iwithprefix`. It applies to
344 // the flags after it and not the ones before, so a command line may set it more than once.
345 // GCC's default is its own installed header directory with the last component taken off,
346 // which is a path a cross compiler's build system knows and passes; there is no equivalent
347 // here, so with no `-iprefix` the prefix is nothing and `-iwithprefix` names a directory
348 // outright.
349 let mut iprefix = String::new();
350
351 let mut i = 0;
352 while i < args.len() {
353 let arg = args[i].as_str();
354 i += 1;
355 match arg {
356 "-h" | "--help" => return Ok(Action::Help),
357 "--version" => return Ok(Action::Version),
358 "--print-config" => print_config = true,
359 "--print-pipeline" => print_pipeline = true,
360 "-###" => print_plan = true,
361 "-v" => verbose = true,
362 // The files a compilation goes through, kept rather than thrown away. The bare
363 // spelling means `=obj` and not `=cwd`, which is not what the manual says and is what
364 // gcc 16 does; `SaveTemps::Object` carries the measurement.
365 "-save-temps" => opts.save_temps = SaveTemps::Object,
366 _ if arg.starts_with("-save-temps=") => {
367 opts.save_temps = arg["-save-temps=".len()..].parse().map_err(err)?;
368 }
369 // How long each step took. A misspelling of this is worth rejecting rather than
370 // ignoring, since a run that says nothing looks like a compilation that took no time.
371 "-time" => opts.time = true,
372 "-c" => opts.emit = EmitKind::Object,
373 "-S" => opts.emit = EmitKind::Asm,
374 "-E" => opts.emit = EmitKind::Preprocessed,
375 "-g" => opts.debug_info = true,
376 // GCC's own levels of how much debug information to write. Zero is none and every
377 // other number is some, and this compiler has one amount, so the numbers above zero
378 // all mean the same thing here. `-ggdb` is the same flag asking for whatever the
379 // debugger on the machine prefers, which is what we emit anyway.
380 "-g0" => opts.debug_info = false,
381 "-g1" | "-g2" | "-g3" | "-ggdb" | "-ggdb1" | "-ggdb2" | "-ggdb3" => {
382 opts.debug_info = true;
383 }
384 // The version of DWARF to write. We write DWARF 5 and nothing else, so a build that
385 // asks for another version is told rather than handed a file it cannot read.
386 "-gdwarf" | "-gdwarf-5" => opts.debug_info = true,
387 _ if arg.starts_with("-gdwarf-") => {
388 return Err(err(format!(
389 "{arg}: this compiler writes DWARF 5 and no other version, see \
390 spec/11-debug-info.md"
391 )));
392 }
393 // Whether the debug information goes in a file of its own beside the object. gcc
394 // writes that `.dwo` whether or not it found anything to put in it, which means a
395 // build system that declares the file as an output gets one and a make rule that
396 // depends on it fires. Refused for that reason rather than taken: section 4.1 takes a
397 // flag that changes nothing and refuses one that changes what is produced, and a file
398 // that does not appear is the plainest change of that kind there is. The negative
399 // spelling is taken, because putting it all in the object is what happens anyway.
400 "-gno-split-dwarf" => {}
401 "-gsplit-dwarf" => {
402 return Err(err(format!(
403 "{arg}: this compiler writes no separate `.dwo` file, and a build that \
404 expects one beside each object would wait for a file that never arrives, \
405 see spec/11-debug-info.md"
406 )));
407 }
408 // How the debug sections are compressed. There are none yet, so every answer produces
409 // the same bytes and taking the flag promises nothing that is not kept. The value is
410 // still checked, because a typo in a distribution's flags is worth finding when the
411 // compiler reads it rather than when somebody later wonders why nothing got smaller.
412 // Bare `-gz` means `zlib`, which the manual leaves for the reader to discover.
413 "-gz" => opts.compress = Compress::Zlib,
414 _ if arg.starts_with("-gz=") => {
415 let how = &arg["-gz=".len()..];
416 opts.compress = how.parse().map_err(|()| {
417 err(format!(
418 "`{how}` is not a way to compress debug sections, which is none, zlib, \
419 zlib-gnu or zstd"
420 ))
421 })?;
422 }
423 "-Werror" => opts.warnings_are_errors = true,
424 // Nothing that is not fatal is said at all. Read at the one place a diagnostic goes
425 // through rather than here, so that a warning `-w` dropped is not counted either.
426 "-w" => opts.warnings = false,
427 "-pedantic-errors" => {
428 opts.pedantic = true;
429 opts.warnings_are_errors = true;
430 }
431 "-P" => opts.line_markers = false,
432 // The dependency family, which section 4.4 calls required because every build system
433 // that generates its own makefiles asks for it. The two that end in `D` write a file
434 // beside the object and let the compilation happen, and the two that do not write to
435 // standard output and stop after it. Nothing here turns the system headers back on
436 // once a flag has turned them off, which is GCC's behaviour and is why `-MM -M` is
437 // `-MM`: the flag asking for fewer of them is the one with something to say.
438 "-M" => {
439 opts.deps.emit = true;
440 opts.deps.instead_of_compiling = true;
441 }
442 "-MM" => {
443 opts.deps.emit = true;
444 opts.deps.instead_of_compiling = true;
445 opts.deps.system_headers = false;
446 }
447 "-MD" => opts.deps.emit = true,
448 "-MMD" => {
449 opts.deps.emit = true;
450 opts.deps.system_headers = false;
451 }
452 "-MP" => opts.deps.phony = true,
453 // These three take a word and only in the separated form, which is how GCC spells
454 // them and how every build system writes them.
455 "-MF" | "-MT" | "-MQ" => {
456 let value =
457 args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
458 i += 1;
459 match arg {
460 "-MF" => opts.deps.file = Some(value.clone()),
461 // The whole of the difference between the two. `-MT` is for a build that has
462 // already escaped what it is passing, and `-MQ` is for one that has a name
463 // and wants it to arrive as that name.
464 "-MT" => opts.deps.targets.push(value.clone()),
465 _ => opts.deps.targets.push(deps::escaped(value)),
466 }
467 }
468 // The questions a build system asks before it compiles anything. Answered after the
469 // loop, because each one is about the target or the library search and the command
470 // line has not finished saying what those are.
471 "-dumpmachine" => query = Some(Query::Machine),
472 "-dumpversion" | "-dumpfullversion" => query = Some(Query::Version),
473 "-print-multiarch" => query = Some(Query::Multiarch),
474 "-print-search-dirs" => query = Some(Query::SearchDirs),
475 "-print-sysroot" => query = Some(Query::Sysroot),
476 // Both spellings, because this one is ours rather than GCC's and our own documents
477 // write it both ways: section 13.5 of `spec/cross-compile/13-distribution.md` gives it
478 // two dashes like the other flags we invented, and document 12's table gives it one
479 // like the `-print-` family it sits in. A person who reads either and types what it
480 // says is right, so neither is refused.
481 "-print-sysroot-provenance" | "--print-sysroot-provenance" => {
482 query = Some(Query::SysrootProvenance);
483 }
484 "-print-libgcc-file-name" => query = Some(Query::Libgcc),
485 _ if arg.starts_with("-print-file-name=") => {
486 query = Some(Query::FileName(arg["-print-file-name=".len()..].to_owned()));
487 }
488 _ if arg.starts_with("-print-prog-name=") => {
489 query = Some(Query::ProgName(arg["-print-prog-name=".len()..].to_owned()));
490 }
491 // A program built to run in more than one thread. On every platform this compiler
492 // targets that is a macro the library's headers read and one more library on the
493 // link line, and the library is added after the loop so that it lands after the
494 // objects that refer to it.
495 "-pthread" | "-pthreads" => {
496 opts.defines.push("_REENTRANT".to_owned());
497 threads = true;
498 }
499 "-ansi" => {
500 opts.std = Std::C89;
501 opts.gnu_extensions = false;
502 }
503 // `-Wpedantic` is the same flag under the name the `-W` family gives it, which is
504 // the spelling a build system that groups its warning flags tends to write.
505 "-pedantic" | "-Wpedantic" => opts.pedantic = true,
506 // Both directions, because a build that needs this for one directory turns it back
507 // off for the next one rather than leaving it on for the whole tree.
508 "-fpermissive" => opts.permissive = true,
509 "-fno-permissive" => opts.permissive = false,
510 "-ffreestanding" => opts.hosted = false,
511 "-fhosted" => opts.hosted = true,
512 "-fno-builtin" => opts.builtins = false,
513 "-fbuiltin" => opts.builtins = true,
514 // The C89 dialects are under GNU's reading whatever this says, so turning it off
515 // there is turning off something the dialect asked for, which is accepted and does
516 // nothing. gcc refuses that command line, and there is nothing it could have meant.
517 "-fgnu89-inline" => opts.gnu89_inline = true,
518 "-fno-gnu89-inline" => opts.gnu89_inline = false,
519 // Both directions of each, because a build system that wants one of these usually
520 // writes it beside the flag that turns it back off for one directory.
521 "-fno-omit-frame-pointer" => opts.frame_pointer = true,
522 "-fomit-frame-pointer" => opts.frame_pointer = false,
523 "-mno-red-zone" => opts.red_zone = false,
524 "-mred-zone" => opts.red_zone = true,
525 // Four flags rather than one with an argument, which is how gcc spells them and how
526 // every build line writes them. Last one wins, because a package build puts
527 // `-fstack-protector-strong` in its global flags and a directory that cannot have one
528 // turns it back off on the line after.
529 "-fno-stack-protector" | "-fno-stack-protector-all" | "-fno-stack-protector-strong" => {
530 opts.protector = Protector::None;
531 }
532 "-fstack-protector" => opts.protector = Protector::Buffers,
533 "-fstack-protector-strong" => opts.protector = Protector::Strong,
534 "-fstack-protector-all" => opts.protector = Protector::All,
535 // The other half of what a hardened build asks for, and it is a question about the
536 // frame rather than about the function, so it is a switch rather than a level.
537 "-fstack-clash-protection" => opts.stack_clash = true,
538 "-fno-stack-clash-protection" => opts.stack_clash = false,
539 // The third of them, and the one that is a question with an argument rather than a
540 // family of spellings, because what it asks about is which of the two edges of a
541 // control flow transfer is checked. Bare is both of them, which is what gcc does.
542 "-fcf-protection" => opts.control = Control::Full,
543 "-fno-cf-protection" => opts.control = Control::None,
544 // Two spellings of the same request, which is what gcc has as well. `-p` was the older
545 // profiler and `-pg` the one that also recorded who called whom, and on every platform
546 // this compiler targets there is now one hook and both ask for it.
547 "-pg" | "-p" => {
548 opts.profile = true;
549 link.profile = true;
550 }
551 // Accepted on their own and doing nothing on their own, which is gcc's behaviour: they
552 // say where the call goes and a command line that asked for no call has nowhere to put
553 // one. That matters because a build system that sets `-mfentry` globally and `-pg` per
554 // directory is a build system that would otherwise fail on every other directory.
555 "-mfentry" => opts.hook = Hook::Early,
556 "-mno-fentry" => opts.hook = Hook::Late,
557 // GCC drops its own include directory along with the system ones, because its
558 // headers are half of a pair with the library's and half a pair is worse than
559 // none. A build that passes this is supplying the whole set itself.
560 "-nostdinc" => nostdinc = true,
561 "-o" => {
562 output = Some(args.get(i).ok_or_else(|| err("-o requires an argument"))?.clone());
563 i += 1;
564 }
565 // The flags that take a directory only in the separated form. GCC spells them
566 // this way and nothing writes `-iquotedir`, so accepting the joined form would
567 // mean guessing at a path that starts with the flag's own letters.
568 // Apple's spelling of `--sysroot`, and the one its own build systems pass. The
569 // two mean the same thing here: the configured directories are under there rather
570 // than under the root.
571 "-isysroot" => {
572 let dir = args.get(i).ok_or_else(|| err("-isysroot requires an argument"))?;
573 i += 1;
574 sysroot = Some(PathBuf::from(dir));
575 }
576 "-iquote" | "-isystem" | "-idirafter" => {
577 let dir = args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
578 i += 1;
579 match arg {
580 "-iquote" => opts.search.push_quote(dir.clone()),
581 "-isystem" => opts.search.push_system(dir.clone()),
582 _ => opts.search.push_after(dir.clone()),
583 }
584 }
585 "-iprefix" => {
586 iprefix = args.get(i).ok_or_else(|| err("-iprefix requires an argument"))?.clone();
587 i += 1;
588 }
589 // Where GCC puts these is not where its manual says it puts them, and this is the
590 // measured answer rather than the documented one: `-iwithprefix` lands in the
591 // `-isystem` slot and not the `-idirafter` slot, and `-iwithprefixbefore` lands in
592 // the `-I` slot. A cross build that uses them is relying on the behaviour, since
593 // that is the compiler it was developed against.
594 "-iwithprefix" | "-iwithprefixbefore" => {
595 let dir = args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
596 i += 1;
597 let dir = format!("{iprefix}{dir}");
598 if arg == "-iwithprefix" {
599 opts.search.push_system(dir);
600 } else {
601 opts.search.push_bracket(dir);
602 }
603 }
604 "-include" | "-imacros" => {
605 let name = args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
606 i += 1;
607 opts.preincludes
608 .push(Preinclude { name: name.clone(), macros_only: arg == "-imacros" });
609 }
610 // The flag `-iquote` was introduced to replace, still passed by build systems old
611 // enough to predate the replacement. It is not a directory: it says that every `-I`
612 // so far is for quoted includes only, and that a quoted include stops looking next
613 // to the file that wrote it.
614 "-I-" => opts.search.split_quote_chain(),
615 "-x" => {
616 let lang = args.get(i).ok_or_else(|| err("-x requires an argument"))?;
617 i += 1;
618 forced = if lang == "none" {
619 None
620 } else {
621 Some(InputKind::from_x_arg(lang).map_err(|e| err(format!("{e}")))?)
622 };
623 }
624 // Not a GCC flag. spec/03-architecture.md section 3.5 compiles several
625 // translation units in one process rather than making the build system fork, and
626 // section 3.8's determinism check compares `-j1` against `-j16`, so the knob has
627 // to exist and has to be spelled the way `make` spells it.
628 // `-DFOO`, `-D FOO` and the same for `-U` and `-I`. Both forms are in wide use
629 // and a build system may produce either, so both are read here rather than
630 // being normalised by whatever generated the command line.
631 _ if arg.starts_with("-D") => {
632 let value = joined_or_next(arg, 2, args, &mut i)?;
633 opts.defines.push(value);
634 }
635 _ if arg.starts_with("-U") => {
636 let value = joined_or_next(arg, 2, args, &mut i)?;
637 opts.undefines.push(value);
638 }
639 _ if arg.starts_with("-I") => {
640 let dir = joined_or_next(arg, 2, args, &mut i)?;
641 opts.search.push_bracket(dir);
642 }
643 _ if arg.starts_with("-std=") => {
644 let name = &arg["-std=".len()..];
645 let (std, gnu) = Std::from_flag(name)
646 .ok_or_else(|| err(format!("unknown dialect `{name}`, see --help")))?;
647 opts.std = std;
648 opts.gnu_extensions = gnu;
649 }
650 // Section 4.5. The claim decides which half of glibc's `sys/cdefs.h` we are
651 // handed, so a differential run that does not set it is comparing two compilers
652 // that believe they are different compilers.
653 // GCC packs these into one flag, so `-dDI` is two of them. Letters in the family
654 // that we have not written yet are accepted and ignored, because a dump is a
655 // debugging aid and a build that asks for one should still compile. A letter
656 // outside the family falls through to the unknown option error, which is what
657 // keeps `-dumpversion` from being read as a dump of nothing.
658 _ if Dumps::is_family(arg) => {
659 opts.dumps.add(&arg[2..]);
660 }
661 // One name at a time, which is what a build that means its own `memcpy` and the
662 // library's everything else writes. The name is not checked against a list, because
663 // the flag is about what the program means by a name and a program is allowed to mean
664 // something by a name this compiler has never heard of.
665 _ if arg.starts_with("-fno-builtin-") => {
666 opts.no_builtin.push(arg["-fno-builtin-".len()..].to_owned());
667 }
668 _ if arg.starts_with("-fgnuc-version=") => {
669 let v = &arg["-fgnuc-version=".len()..];
670 opts.gnuc = v.parse().map_err(err)?;
671 }
672 // spec/13-gnu-compat.md section 13.3 promises this flag an error that says why rather
673 // than the unknown option one, because a build reaching for it is asking for a feature
674 // and deserves to be told it is not coming rather than told the spelling is wrong.
675 // The negative form is what this compiler does anyway, so it is taken and dropped.
676 "-fnested-functions" => {
677 return Err(err(
678 "nested functions are not supported: a call to one goes through a trampoline \
679 written on the stack, which no target that enforces an unexecutable stack \
680 allows",
681 ));
682 }
683 "-fno-nested-functions" => {}
684 // Which of the two links the output is for, which is a real difference and not a
685 // description of what happens anyway. Everything here is position independent either
686 // way, and what these decide is whether a name may be one another object defines or
687 // replaces, because a link that produces an executable puts every name in the same
688 // program and a link that produces a shared library does not.
689 //
690 // It matters that they are accepted at all, whatever they then do. Every autoconf and
691 // cmake build puts `-fPIC` on the compile line, so a compiler that rejects it cannot
692 // be the `CC` of a project that has a configure script, whatever else it can do. That
693 // is how this was found: building SQLite's test fixture stopped on it.
694 "-fPIC" | "-fpic" => opts.pic = Pic::Library,
695 // Not a synonym of the pair above, which is what they were treated as until #756. The
696 // library is the expensive answer and gcc makes it the one that has to be asked for,
697 // so this is also what nothing at all means.
698 "-fPIE" | "-fpie" => opts.pic = Pic::Executable,
699 // A different question from the pair above, and the one every distribution build of a
700 // shared library answers. `-fPIC` decides how an address is reached, and this decides
701 // whether the optimizer may believe a body it can see, because an exported name is one
702 // the dynamic linker may find another definition of first. On by default, which is
703 // gcc's arrangement and is the honest answer, and off is a promise the build makes and
704 // nothing checks.
705 "-fsemantic-interposition" => opts.interposition = true,
706 "-fno-semantic-interposition" => opts.interposition = false,
707 // Two requests rather than one, and the same table answers both, so what decides is
708 // whether either of them is standing. gcc arranges it the same way: the asynchronous
709 // one is the default here and it implies the other, and a line that asks for a table
710 // and against an asynchronous one gets a table.
711 "-fasynchronous-unwind-tables" => opts.async_unwind_tables = true,
712 "-fno-asynchronous-unwind-tables" => opts.async_unwind_tables = false,
713 "-funwind-tables" => opts.unwind_tables = true,
714 "-fno-unwind-tables" => opts.unwind_tables = false,
715 // The other direction is a request, not a description, and it is one this compiler
716 // cannot grant, so it gets the treatment section 13.3 asks for rather than the unknown
717 // option error. Answering it by carrying on would be answering a different question:
718 // the code would still be position independent, which is correct everywhere an
719 // ordinary program runs and is wrong in a kernel, where the flag is written precisely
720 // because there is no loader to fill a global offset table in.
721 "-fno-pic" | "-fno-pie" => {
722 return Err(err(
723 "position dependent code is not supported: an address that may be in another \
724 object is loaded out of the global offset table, and nothing here emits the \
725 absolute form this asks for. Use -no-pie if what you meant was how to link",
726 ));
727 }
728 // A section per function and a section per variable, which is what makes
729 // `--gc-sections` able to drop anything: a linker can leave out a section nothing
730 // reaches and cannot leave out half of one. Both directions are taken, and the off
731 // one is the default rather than a refusal, since a build that writes it is asking
732 // for what happens anyway.
733 "-ffunction-sections" => opts.function_sections = true,
734 "-fno-function-sections" => opts.function_sections = false,
735 "-fdata-sections" => opts.data_sections = true,
736 "-fno-data-sections" => opts.data_sections = false,
737 // Another description of what this compiler does. A file scope declaration with no
738 // initializer is written into `.bss` as an ordinary defined symbol, not offered to the
739 // linker as a common one for it to merge, which is what `-fno-common` asks for and what
740 // gcc has done by default since 10. Nothing in the front end produces `Linkage::Common`
741 // at all.
742 "-fno-common" => {}
743 // What overflows rather than being undefined. Every one of these takes something away
744 // from the optimizer rather than asking it to do anything, which is why the negative
745 // spellings are the interesting ones and the positive spellings are the default.
746 //
747 // `-fno-strict-overflow` is both of the others, which is gcc's own reading of it: its
748 // help text for `-fstrict-overflow` says "negated as -fwrapv -fwrapv-pointer". So it is
749 // written here as the pair rather than kept as a third thing to test everywhere.
750 //
751 // `-ftrapv` is the exception and is the one that asks for something. It is the other
752 // answer to the question `-fwrapv` answers, so the two cannot both hold and each clears
753 // the other, which makes the last one on the command line the one that counts. That is
754 // gcc 16's behaviour and was measured rather than read: `-ftrapv -fwrapv` emits no
755 // checked calls and `-fwrapv -ftrapv` emits them. The positive spelling of the pointer
756 // question is left alone by both, because neither has anything to say about it.
757 "-fwrapv" => {
758 opts.wrapping.signed = true;
759 opts.wrapping.trap = false;
760 }
761 "-fno-wrapv" => opts.wrapping.signed = false,
762 "-fwrapv-pointer" => opts.wrapping.pointer = true,
763 "-fno-wrapv-pointer" => opts.wrapping.pointer = false,
764 "-fno-strict-overflow" => opts.wrapping = Wrapping::ALL,
765 // Which does not clear the checked one, because gcc does not: `-ftrapv
766 // -fstrict-overflow` still emits the calls. It says what is assumed and not what
767 // happens.
768 "-fstrict-overflow" => {
769 opts.wrapping.signed = false;
770 opts.wrapping.pointer = false;
771 }
772 "-ftrapv" => {
773 opts.wrapping.trap = true;
774 opts.wrapping.signed = false;
775 }
776 "-fno-trapv" => opts.wrapping.trap = false,
777 // The two flags that say what a plain `char` is, which is one question with two
778 // spellings each: gcc reads `-fno-signed-char` as `-funsigned-char` and
779 // `-fno-unsigned-char` as `-fsigned-char`, so there are four ways to write two
780 // answers and the last one written wins. Nothing is set until one of them is given,
781 // because the target's own ABI is the answer otherwise and it is not the same answer
782 // everywhere: x86-64 and Apple's arm64 are signed, Linux's arm64 is not.
783 "-fsigned-char" | "-fno-unsigned-char" => opts.char_signed = Some(true),
784 "-funsigned-char" | "-fno-signed-char" => opts.char_signed = Some(false),
785 // And the size of an enumeration, which is the other thing in this group that changes
786 // the ABI rather than the code.
787 "-fshort-enums" => opts.short_enums = true,
788 "-fno-short-enums" => opts.short_enums = false,
789 // And the request, which is the one that cannot be granted. It is a real difference and
790 // not a preference: two files each writing `int g;` link under `-fcommon` and are a
791 // duplicate definition without it, which is the whole reason the flag survives.
792 "-fcommon" => {
793 return Err(err(
794 "a tentative definition is written into .bss as its own symbol here, and \
795 nothing emits the common symbol this asks the linker to merge. Give the \
796 variable a definition in one file and declare it extern in the others",
797 ));
798 }
799 // Both directions of this one are recorded, and what they decide is whether lowering
800 // names the type each access goes through. Turning it off is the front end leaving the
801 // name off rather than a pass being told to ignore one it can see, which is one
802 // condition in one place, and it is the reading that survives link time optimization:
803 // a unit built with the flag off keeps its own answer when its bodies end up in a
804 // module beside bodies that were not.
805 //
806 // Nothing in the pipeline reads those names yet. Layer 3 of the alias analysis does
807 // and is tested, and no pass at any level asks the alias analysis anything today, so
808 // no program compiles differently for having passed this. The flag is wired anyway,
809 // because the change that makes a pass ask is not the change anybody will remember to
810 // wire it in, and a flag that is taken and dropped once the names mean something is
811 // the miscompilation `spec/04-driver-and-cli.md` section 4.1 warns about in as many
812 // words.
813 "-fstrict-aliasing" => opts.strict_aliasing = true,
814 "-fno-strict-aliasing" => opts.strict_aliasing = false,
815 // The same shape of answer for the same reason, and the flag the kernel writes beside
816 // the one above it.
817 //
818 // Nothing here concludes that a pointer is not null from the fact that it was
819 // dereferenced. There is no such conclusion to draw from, because no pass records one:
820 // a load says where it read and nothing else, and a comparison against null is an
821 // ordinary comparison of two values the optimizer has no fact about. So a function
822 // that reads through a pointer and then tests it keeps the test, which is what the
823 // kernel wants and what `-fno-delete-null-pointer-checks` asks for, and what gcc has
824 // to be asked for because it draws the conclusion by default.
825 //
826 // `-fdelete-null-pointer-checks` is the request to draw it, and it goes the way
827 // `-fstrict-aliasing` does: assuming less than was asked for costs speed and not
828 // correctness, and `-O2` implies it, so refusing it would stop builds for nothing.
829 "-fdelete-null-pointer-checks" | "-fno-delete-null-pointer-checks" => {}
830 // The floating point group, which goes the same way and for the same reason, and which
831 // is worth writing out because the reason is easy to get backwards.
832 //
833 // Each of these has a restrictive spelling and a permissive one. The restrictive ones,
834 // `-frounding-math` and `-ftrapping-math`, say that the rounding mode may have been
835 // changed and that an exception raised by an operation may be looked at, so an
836 // arithmetic the compiler folds at compile time is an arithmetic whose rounding and
837 // whose exception the program does not get. Nothing here folds any floating point
838 // arithmetic in a function body: `0.1 + 0.2` is an `fadd` and `1.0 / 0.0` is a divide
839 // that runs, at every level. So both of those describe what already happens.
840 //
841 // The permissive ones are the other half, and they are licences rather than requests
842 // for an answer. `-fno-rounding-math` says the rounding mode is the default one and
843 // `-fno-trapping-math` says nothing looks at the exceptions, which together are
844 // permission to fold. Not folding is the conservative side of that permission and is
845 // what a program is entitled to whichever was written, so the flag costs speed and not
846 // correctness, which is the test section 4.1 puts a licence through. `-ftrapping-math`
847 // is also gcc's default, so a build spelling it out is a build asking for what it
848 // already has.
849 "-frounding-math" | "-fno-rounding-math" => {}
850 "-ftrapping-math" | "-fno-trapping-math" => {}
851 // About temporary files rather than about code. There is nothing between the phases of
852 // one compilation here to write to a file in the first place.
853 "-pipe" => {}
854 // Nothing here writes colour, so all of these are the same answer, and it is the answer
855 // that costs nothing: the diagnostics come out plain either way and no build depends on
856 // an escape sequence being there. Taken rather than refused because cmake writes
857 // `-fdiagnostics-color=always` on every compile line when the generator is ninja, which
858 // makes this the second most common flag after `-fPIC` to stop a build over a question
859 // about how the text looks.
860 "-fdiagnostics-color" | "-fno-diagnostics-color" => {}
861 _ if arg.starts_with("-fdiagnostics-color=") => {}
862 // The link flags. None of them changes the compilation, which is why they are
863 // collected apart from `opts` and why `-lm` on a `-c` line is a note rather than an
864 // error: it is a thing said to a linker that is not going to run.
865 "-static" => link.is_static = true,
866 "-shared" => link.shared = true,
867 "-pie" => link.pie = Some(true),
868 "-no-pie" | "-nopie" => link.pie = Some(false),
869 "-nostdlib" => link.no_stdlib = true,
870 "-nostartfiles" => link.no_startfiles = true,
871 "-nodefaultlibs" => link.no_defaultlibs = true,
872 "-fno-builtins-lib" => link.no_builtins_lib = true,
873 "-fbuiltins-lib" => link.no_builtins_lib = false,
874 "-rdynamic" | "-export-dynamic" => link.export_dynamic = true,
875 "-s" => link.strip = true,
876 "-Xlinker" => {
877 let next = args.get(i).ok_or_else(|| err("-Xlinker requires an argument"))?;
878 i += 1;
879 link.passthrough.push(next.clone());
880 }
881 _ if arg.starts_with("-Wl,") => {
882 // Commas separate arguments rather than being part of one, which is what makes
883 // `-Wl,-rpath,/opt/lib` two words to the linker and one word here.
884 link.passthrough.extend(arg["-Wl,".len()..].split(',').map(str::to_owned));
885 }
886 _ if arg.starts_with("-fuse-ld=") => {
887 link.use_ld = Some(arg["-fuse-ld=".len()..].to_owned());
888 }
889 _ if arg.starts_with("-l") && arg.len() > 2 => {
890 inputs.push(Input::library(&arg[2..]));
891 }
892 "-l" => {
893 let next = args.get(i).ok_or_else(|| err("-l requires an argument"))?;
894 i += 1;
895 inputs.push(Input::library(next));
896 }
897 _ if arg.starts_with("-L") => {
898 link.search.push(PathBuf::from(joined_or_next(arg, 2, args, &mut i)?));
899 }
900 _ if arg.starts_with("-B") => {
901 link.prefixes.push(PathBuf::from(joined_or_next(arg, 2, args, &mut i)?));
902 }
903 _ if arg.starts_with("-j") => {
904 jobs = Jobs::parse(&arg[2..]).map_err(err)?;
905 }
906 _ if arg.starts_with("--sysroot=") => {
907 sysroot = Some(PathBuf::from(&arg["--sysroot=".len()..]));
908 }
909 _ if arg.starts_with("--target=") => {
910 let t = &arg["--target=".len()..];
911 opts.target = t.parse().map_err(|e| err(format!("{e}")))?;
912 // The same string again, as the model that has room for a libc version. A spelling
913 // the three field parser took and this one does not is not an error, because the
914 // one that decides what is compiled has already accepted it and the only thing
915 // lost is a version nobody asked for.
916 pinned = t.parse().ok();
917 }
918 _ if arg.starts_with("--emit=") => {
919 let k = &arg["--emit=".len()..];
920 opts.emit = k
921 .parse()
922 .map_err(|()| err(format!("unknown --emit kind `{k}`, see --help")))?;
923 }
924 // A bare `-O` is `-O1`, which is what GCC has and what a hand written makefile tends
925 // to write. `-Og` is GCC's level for a build somebody is going to step through, and
926 // it is `-O1` with the transformations that move code around left out; this compiler
927 // has no such level yet, so it is the nearest one and `--print-pipeline` says what
928 // that came to rather than the flag pretending otherwise.
929 "-O" | "-Og" => opts.opt_level = rucc_session::OptLevel::O1,
930 // The union of `-O3` and `-ffast-math`, and the second half of that changes what
931 // floating point arithmetic means. Refused rather than taken as `-O3`, because a
932 // build that asks for fast math and is quietly given ordinary arithmetic gets a
933 // slower program than it asked for and a build that is given fast math it did not
934 // ask for gets a wrong one.
935 "-Ofast" => {
936 return Err(err(
937 "-Ofast is -O3 with fast math, and fast math is not implemented, see \
938 spec/04-driver-and-cli.md section 4.6",
939 ));
940 }
941 _ if arg.starts_with("-O") => {
942 opts.opt_level = arg[2..]
943 .parse()
944 .map_err(|()| err(format!("unknown optimization level `{arg}`")))?;
945 }
946 // How far a multiply and an addition may be fused into one rounding. Before the
947 // optimizer's `-f` family below for the reason the ones under it are, and kept rather
948 // than dropped because it is the one flag in its group this compiler could act on: it
949 // rides into the IR as an attribute on each function with a body, so the day the code
950 // generator forms an `fma` it already knows which functions were given permission.
951 // Nothing forms one today, under any value of this and under any `-march=`.
952 _ if arg.starts_with("-ffp-contract=") => {
953 let how = &arg["-ffp-contract=".len()..];
954 opts.fp_contract = how.parse().map_err(|()| {
955 err(format!("`{how}` is not a contraction, which is fast, on or off"))
956 })?;
957 }
958 // How much of an expression may be computed wider than it was written. The values are
959 // gcc's and so is the refusal of anything else, and none of the three changes anything
960 // here: an operation is computed in the type C says it is on every target this compiler
961 // has a back end for, so `__FLT_EVAL_METHOD__` is 0 and `standard` is already what
962 // happens. `fast` and `16` are permission to be wider, which is a licence this takes
963 // and does not use, the same way the two above are. The flag is worth taking because
964 // glibc's headers and a good deal of configure output write it, and because the answer
965 // it asks about is one this compiler can state rather than guess at: there is no x87
966 // target here, which is the machine the whole question was invented for.
967 _ if arg.starts_with("-fexcess-precision=") => {
968 let how = &arg["-fexcess-precision=".len()..];
969 if !matches!(how, "16" | "fast" | "standard") {
970 return Err(err(format!(
971 "`{how}` is not an excess precision, which is 16, fast or standard"
972 )));
973 }
974 }
975 // Which front of a path is rewritten before it reaches the output, which is how a
976 // build gets the same bytes out of two different directories. The four spellings are
977 // one flag each into three lists, and `-ffile-prefix-map=` is the three of them at
978 // once. Only the macro list does anything today, because `__FILE__` is the only place
979 // a path reaches the output: there is no DWARF and no profile data yet, so the other
980 // two are recorded for the work that will read them. The argument splits at the last
981 // `=` rather than the first, which is gcc's rule and is what lets a directory with an
982 // `=` in its name be the old half.
983 _ if arg.starts_with("-fmacro-prefix-map=") => {
984 let (old, new) = rewrite(arg, "-fmacro-prefix-map=")?;
985 opts.prefix_map.macros.push(old, new);
986 }
987 _ if arg.starts_with("-fdebug-prefix-map=") => {
988 let (old, new) = rewrite(arg, "-fdebug-prefix-map=")?;
989 opts.prefix_map.debug.push(old, new);
990 }
991 _ if arg.starts_with("-fprofile-prefix-map=") => {
992 let (old, new) = rewrite(arg, "-fprofile-prefix-map=")?;
993 opts.prefix_map.profile.push(old, new);
994 }
995 _ if arg.starts_with("-ffile-prefix-map=") => {
996 let (old, new) = rewrite(arg, "-ffile-prefix-map=")?;
997 opts.prefix_map.macros.push(old, new);
998 opts.prefix_map.debug.push(old, new);
999 opts.prefix_map.profile.push(old, new);
1000 }
1001 // A whole optimization rather than a flag, and the family is taken rather than
1002 // refused because of what ignoring it does. There is none of it here yet, so a build
1003 // that asks for it gets a program that is correct and slower than it could have been,
1004 // which is what section 4.1 means by a hint about speed and what every compilation at
1005 // `-O0` already is. The objects settle the rest of the argument: gcc's `-flto` object
1006 // holds the bytecode and no machine code at all, and every object here holds the code,
1007 // which is exactly what `-ffat-lto-objects` asks gcc for. So a build passing `-flto`
1008 // to this compiler gets objects that are more usable than the ones it asked for rather
1009 // than different ones. Every value is still checked against gcc's, because somebody
1010 // who wrote `-flto=thin` meant clang and had better hear about it here.
1011 "-flto" => opts.lto.requested = true,
1012 "-fno-lto" => opts.lto.requested = false,
1013 _ if arg.starts_with("-flto=") => {
1014 let how = &arg["-flto=".len()..];
1015 opts.lto.jobs = how.parse().map_err(|()| {
1016 err(format!(
1017 "`{how}` is not a number of link time jobs, which is auto, jobserver or a \
1018 count above zero"
1019 ))
1020 })?;
1021 opts.lto.requested = true;
1022 }
1023 _ if arg.starts_with("-flto-partition=") => {
1024 let how = &arg["-flto-partition=".len()..];
1025 opts.lto.partition = how.parse().map_err(|()| {
1026 err(format!(
1027 "`{how}` is not a partitioning model, which is balanced, 1to1, one, max \
1028 or none"
1029 ))
1030 })?;
1031 }
1032 _ if arg.starts_with("-flto-compression-level=") => {
1033 let how = &arg["-flto-compression-level=".len()..];
1034 let level =
1035 how.parse::<u8>().ok().filter(|level| *level <= 19).ok_or_else(|| {
1036 err(format!("`{how}` is not a compression level, 0 to 19"))
1037 })?;
1038 opts.lto.compression = Some(level);
1039 }
1040 // Whether the object keeps its machine code as well as the bytecode. It always does
1041 // here, so the first of these describes what happens and the second asks for an object
1042 // with less in it, which is a smaller file and not a different program, so both are
1043 // taken.
1044 "-ffat-lto-objects" | "-fno-fat-lto-objects" => {}
1045 // Whether the linker is handed a plugin that does the link time work. The design in
1046 // `spec/09-optimizer.md` has this driver doing that work itself and never loading a
1047 // plugin into anybody, so neither answer is a question it has to hold.
1048 "-fuse-linker-plugin" | "-fno-use-linker-plugin" => {}
1049 // Reading a profile back. Taken for the reason the family above it is: nothing here
1050 // reads one, so a build that asks gets the program it would have got anyway, and gcc
1051 // itself produces a byte for byte identical object from `-fprofile-use` when there are
1052 // no counts beside the file. The path is recorded for the pass that will read it. The
1053 // warning gcc prints when it looked and found nothing is deliberately not copied,
1054 // because nothing here looks, and a warning about a file that was never opened would
1055 // fire on the builds that have a perfectly good profile as well as on the ones that
1056 // do not.
1057 "-fprofile-use" => opts.profile_data.requested = true,
1058 "-fno-profile-use" => opts.profile_data.requested = false,
1059 _ if arg.starts_with("-fprofile-use=") => {
1060 opts.profile_data.path = Some(arg["-fprofile-use=".len()..].to_string());
1061 opts.profile_data.requested = true;
1062 }
1063 _ if arg.starts_with("-fprofile-dir=") => {
1064 opts.profile_data.dir = Some(arg["-fprofile-dir=".len()..].to_string());
1065 }
1066 "-fprofile-abs-path" => opts.profile_data.absolute = true,
1067 "-fno-profile-abs-path" => opts.profile_data.absolute = false,
1068 "-fprofile-correction" => opts.profile_data.correction = true,
1069 "-fno-profile-correction" => opts.profile_data.correction = false,
1070 "-fprofile-partial-training" => opts.profile_data.partial_training = true,
1071 "-fno-profile-partial-training" => opts.profile_data.partial_training = false,
1072 // Writing the counts rather than reading them, which is refused rather than taken and
1073 // is the same line `-gsplit-dwarf` falls on the far side of. Ignoring these means a
1074 // file a build declared as an output never appears: the instrumented program writes a
1075 // `.gcda` as it exits and `-ftest-coverage` writes a `.gcno` beside the object, and a
1076 // two stage build that got neither would go on to optimize against no counts at all
1077 // and report coverage of nothing, with nothing along the way saying so. The objects
1078 // say the rest: gcc's `-fprofile-generate` object holds 375 bytes of code where a
1079 // plain one holds 71, and 296 bytes of counters that a plain one does not have, so
1080 // this is a flag that changes the output rather than a hint about speed.
1081 "-fprofile-arcs"
1082 | "--coverage"
1083 | "-fcondition-coverage"
1084 | "-fpath-coverage"
1085 | "-fprofile-generate" => {
1086 return Err(err(format!(
1087 "{arg}: this compiler does not instrument for profiling, and a build that \
1088 expects the counts a run of the instrumented program writes would optimize \
1089 against nothing on its second pass, see spec/04-driver-and-cli.md"
1090 )));
1091 }
1092 _ if arg.starts_with("-fprofile-generate=") => {
1093 return Err(err(format!(
1094 "{arg}: this compiler does not instrument for profiling, and a build that \
1095 expects the counts a run of the instrumented program writes would optimize \
1096 against nothing on its second pass, see spec/04-driver-and-cli.md"
1097 )));
1098 }
1099 "-ftest-coverage" => {
1100 return Err(err(format!(
1101 "{arg}: this compiler writes no `.gcno` file beside the object, and a build \
1102 that expects one would wait for a file that never arrives, see \
1103 spec/04-driver-and-cli.md"
1104 )));
1105 }
1106 // The rest of the family describes instrumentation that is refused above, so what is
1107 // left to do with them is check them and drop them. They are checked because a
1108 // misspelling in a distribution's flags is worth finding here rather than on the day
1109 // the instrumentation lands, and dropped because there is nothing for an answer about
1110 // how a counter is written to be an answer about.
1111 _ if arg.starts_with("-fprofile-update=") => {
1112 let how = &arg["-fprofile-update=".len()..];
1113 if !matches!(how, "single" | "atomic" | "prefer-atomic") {
1114 return Err(err(format!(
1115 "`{how}` is not a profile update method, which is single, atomic or \
1116 prefer-atomic"
1117 )));
1118 }
1119 }
1120 _ if arg.starts_with("-fprofile-reproducible=") => {
1121 let how = &arg["-fprofile-reproducible=".len()..];
1122 if !matches!(how, "serial" | "parallel-runs" | "multithreaded") {
1123 return Err(err(format!(
1124 "`{how}` is not a profile reproducibility method, which is serial, \
1125 parallel-runs or multithreaded"
1126 )));
1127 }
1128 }
1129 "-fprofile-values" | "-fno-profile-values" | "-fprofile-info-section" => {}
1130 "-fno-test-coverage" | "-fno-profile-arcs" | "-fno-profile-generate" => {}
1131 _ if arg.starts_with("-fprofile-filter-files=")
1132 || arg.starts_with("-fprofile-exclude-files=")
1133 || arg.starts_with("-fprofile-note=") => {}
1134 // What every name gets when nothing in the source said, which the attribute in the
1135 // source overrides rather than the other way round. Before the optimizer's `-f`
1136 // family below for the reason the tier below it is.
1137 _ if arg.starts_with("-fvisibility=") => {
1138 let seen = &arg["-fvisibility=".len()..];
1139 opts.visibility = seen.parse().map_err(|()| {
1140 err(format!(
1141 "`{seen}` is not a visibility, which is default, hidden, internal or \
1142 protected"
1143 ))
1144 })?;
1145 }
1146 // Which edges of a control flow transfer are checked. Before the optimizer's `-f`
1147 // family below for the reason the two above it are, and last of the three so that the
1148 // bare spelling and the negative one are matched exactly rather than by this.
1149 _ if arg.starts_with("-fcf-protection=") => {
1150 let edges = &arg["-fcf-protection=".len()..];
1151 opts.control = edges.parse().map_err(|()| {
1152 err(format!(
1153 "`{edges}` is not a control flow protection, which is full, branch, \
1154 return, none or check"
1155 ))
1156 })?;
1157 }
1158 // How much room every function opens with for something to be written over later.
1159 // Before the optimizer's `-f` family below for the reason the ones above it are.
1160 _ if arg.starts_with("-fpatchable-function-entry=") => {
1161 let room = &arg["-fpatchable-function-entry=".len()..];
1162 opts.patchable = room.parse().map_err(|()| {
1163 err(format!(
1164 "`{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"
1165 ))
1166 })?;
1167 }
1168 // The memory safety monitor, from section 15.4 of
1169 // `spec/safe-memory/15-integration.md`. Before the optimizer's `-f` family below,
1170 // because a pass that took the name `safety=detect` would otherwise be handed the
1171 // flag, and the tier is not a pass.
1172 _ if arg.starts_with("-fsafety=") => {
1173 let tier = &arg["-fsafety=".len()..];
1174 opts.safety = tier.parse().map_err(|()| {
1175 err(format!(
1176 "`{tier}` is not a safety tier, which is off, detect, enforce or kernel"
1177 ))
1178 })?;
1179 }
1180 // Whether padding participates, from section 9.3 of document 09. Spelled out rather
1181 // than folded into the tier because it is a departure somebody who has read that
1182 // section makes, and the two defaults it describes are a property of what is being
1183 // built rather than of how much checking is wanted.
1184 _ if arg.starts_with("-fsafety-init=") => {
1185 let mode = &arg["-fsafety-init=".len()..];
1186 opts.padding = mode.parse().map_err(|()| {
1187 err(format!("`{mode}` is not a padding mode, which is padding or nopadding"))
1188 })?;
1189 }
1190 // Row S4, from section 9.4 of document 09. A bare flag with no value, because the
1191 // strict form of that section needs a member id the front end does not name yet and
1192 // accepting the spelling for it would be accepting a promise this build cannot keep.
1193 // Before `-fno-` is looked at below, for the reason the tier is.
1194 "-fsafety-subobject" => opts.subobject = rucc_session::Subobject::Members,
1195 "-fno-safety-subobject" => opts.subobject = rucc_session::Subobject::Off,
1196 _ if arg.starts_with("-fsafety-subobject=") => {
1197 let form = &arg["-fsafety-subobject=".len()..];
1198 return Err(err(format!(
1199 "`{form}` is not a form of -fsafety-subobject. The flag takes no value, and \
1200 the strict form of section 9.4 is tamnd/rucc#967"
1201 )));
1202 }
1203 // Row Y8, from section 9.6 of document 09. A bare flag with no value, for the reason
1204 // the one above has none: there is one form of this check and a spelling that suggested
1205 // otherwise would be promising something. Before `-fno-` is looked at below, the same
1206 // way.
1207 "-fsafety-restrict" => opts.promise = rucc_session::Promise::Blocks,
1208 "-fno-safety-restrict" => opts.promise = rucc_session::Promise::Off,
1209 _ if arg.starts_with("-fsafety-restrict=") => {
1210 let form = &arg["-fsafety-restrict=".len()..];
1211 return Err(err(format!(
1212 "`{form}` is not a form of -fsafety-restrict. The flag takes no value."
1213 )));
1214 }
1215 // The sanitizers of document 12, which are checks at run time rather than a way of
1216 // generating the same program. Each name is held to gcc 16's list, and what is still
1217 // asked for by the end of the line is answered after the loop, so that a command line
1218 // which turns one on and then off again is a command line that asked for nothing.
1219 //
1220 // Before the optimizer's `-f` family below, for the reason the tier above it is.
1221 _ if arg.starts_with("-fsanitize=") => {
1222 for one in arg["-fsanitize=".len()..].split(',') {
1223 if one == "all" {
1224 // gcc takes `all` only in the negative, because turning every check on at
1225 // once includes checks that contradict each other.
1226 return Err(err(
1227 "`-fsanitize=all` is not a gcc option, only `-fno-sanitize=all` is",
1228 ));
1229 }
1230 if !SANITIZERS.contains(&one) {
1231 return Err(err(format!(
1232 "`{one}` is not a sanitizer, see spec/04-driver-and-cli.md section 4.7"
1233 )));
1234 }
1235 if !sanitizers.contains(&one) {
1236 sanitizers.push(one);
1237 }
1238 }
1239 }
1240 _ if arg.starts_with("-fno-sanitize=") => {
1241 for one in arg["-fno-sanitize=".len()..].split(',') {
1242 if one == "all" {
1243 sanitizers.clear();
1244 continue;
1245 }
1246 if !SANITIZERS.contains(&one) {
1247 return Err(err(format!(
1248 "`{one}` is not a sanitizer, see spec/04-driver-and-cli.md section 4.7"
1249 )));
1250 }
1251 sanitizers.retain(|asked| *asked != one);
1252 }
1253 }
1254 // What a check does when it fires, and where the records about the checked objects go.
1255 // Each of them is an answer about the sanitizers refused after the loop, so there is
1256 // nothing left for them to change here. The names are still held to the list, because
1257 // a misspelling in a build's flags is worth finding when the compiler reads it.
1258 _ if arg.starts_with("-fsanitize-recover=")
1259 || arg.starts_with("-fno-sanitize-recover=")
1260 || arg.starts_with("-fsanitize-trap=")
1261 || arg.starts_with("-fno-sanitize-trap=") =>
1262 {
1263 // The guard above matched on a spelling that has an `=` in it, so the tail is
1264 // whatever follows the first one.
1265 let how = arg.split_once('=').map_or("", |(_, rest)| rest);
1266 for one in how.split(',') {
1267 if one != "all" && !SANITIZERS.contains(&one) {
1268 return Err(err(format!(
1269 "`{one}` is not a sanitizer, see spec/04-driver-and-cli.md section 4.7"
1270 )));
1271 }
1272 }
1273 }
1274 "-fsanitize-undefined-trap-on-error"
1275 | "-fsanitize-address-use-after-scope"
1276 | "-fno-sanitize-address-use-after-scope" => {}
1277 _ if arg.starts_with("-fsanitize-sections=") => {}
1278 // Counting which edges a run reached, which is how a fuzzer knows an input was worth
1279 // keeping. Refused rather than dropped, because a fuzzer whose calls into
1280 // `__sanitizer_cov_*` were never generated runs blind and reports coverage of nothing,
1281 // and there is no point in the campaign where that announces itself.
1282 _ if arg.starts_with("-fsanitize-coverage=") => {
1283 let how = &arg["-fsanitize-coverage=".len()..];
1284 for one in how.split(',') {
1285 if !matches!(one, "trace-pc" | "trace-cmp") {
1286 return Err(err(format!(
1287 "`{one}` is not a coverage instrumentation, which is trace-pc or \
1288 trace-cmp"
1289 )));
1290 }
1291 }
1292 return Err(err(format!(
1293 "{arg}: this compiler generates no coverage callbacks, and a fuzzer built \
1294 with it would run without any feedback at all, see \
1295 spec/04-driver-and-cli.md section 4.7"
1296 )));
1297 }
1298 // The optimizer's own flags, from section 9.10 of `spec/09-optimizer.md`. These come
1299 // after every `-f` the rest of the compiler answers to, so a pass can never take a
1300 // name that already means something else on the command line.
1301 _ if arg.starts_with("-fpass-fuel=") => {
1302 let (name, count) = arg["-fpass-fuel=".len()..]
1303 .split_once('=')
1304 .ok_or_else(|| err("-fpass-fuel= is spelled <pass>=<count>"))?;
1305 if rucc_opt::pass::find(name).is_none() {
1306 return Err(err(format!(
1307 "`{name}` is not a pass this compiler has, see --print-pipeline"
1308 )));
1309 }
1310 let count: u32 = count
1311 .parse()
1312 .map_err(|_| err(format!("`{count}` is not a number of transformations")))?;
1313 opts.pass_fuel.push((name.to_owned(), count));
1314 }
1315 _ if arg.starts_with("-fpass-fuel-global=") => {
1316 let count = &arg["-fpass-fuel-global=".len()..];
1317 let count: u32 = count
1318 .parse()
1319 .map_err(|_| err(format!("`{count}` is not a number of transformations")))?;
1320 opts.pass_fuel_global = Some(count);
1321 }
1322 // Everything from `-fopt-info` to the end of the argument, which is optional
1323 // keywords joined by hyphens and an optional `=<file>`. Checked here rather than
1324 // where the remarks are printed, because by then the compilation somebody wanted
1325 // to hear about is over.
1326 _ if arg == "-fopt-info"
1327 || arg.starts_with("-fopt-info=")
1328 || arg.starts_with("-fopt-info-") =>
1329 {
1330 let rest = &arg["-fopt-info".len()..];
1331 let (kinds, file) = match rest.split_once('=') {
1332 Some((kinds, file)) => (kinds, Some(file)),
1333 None => (rest, None),
1334 };
1335 let kinds = kinds.strip_prefix('-').unwrap_or(kinds);
1336 rucc_opt::Wants::none().add(kinds).map_err(err)?;
1337 opts.opt_info.push(kinds.to_owned());
1338 if let Some(file) = file {
1339 if file.is_empty() {
1340 return Err(err("-fopt-info= was given no file to write to"));
1341 }
1342 opts.opt_info_file = Some(file.to_owned());
1343 }
1344 }
1345 _ if arg.starts_with("-fdump-ir=") => {
1346 // Checked here rather than where the dumps are taken, because the compilation
1347 // that would have been dumped is over by then.
1348 let spec = &arg["-fdump-ir=".len()..];
1349 rucc_opt::Dumps::default().add(spec).map_err(err)?;
1350 opts.dump_ir.push(spec.to_owned());
1351 }
1352 // Before the bare `-f<pass>` below, because a pass called `enable-something` would
1353 // otherwise take the flag away from the gate. Checked here rather than where the
1354 // pipeline reads it, for the reason that applies to all of these: a misspelled pass
1355 // name that quietly gated nothing looks exactly like a pass that is not the guilty
1356 // one, and a bisection would carry on past the thing it was looking for.
1357 _ if arg.starts_with("-fdisable-") || arg.starts_with("-fenable-") => {
1358 let on = arg.starts_with("-fenable-");
1359 let spec = &arg[if on { "-fenable-".len() } else { "-fdisable-".len() }..];
1360 rucc_opt::Gates::default().add(on, spec).map_err(err)?;
1361 opts.pass_gates.push((on, spec.to_owned()));
1362 }
1363 _ if arg.strip_prefix("-fno-").is_some_and(|n| rucc_opt::pass::find(n).is_some()) => {
1364 opts.passes.push((arg["-fno-".len()..].to_owned(), false));
1365 }
1366 _ if arg.strip_prefix("-f").is_some_and(|n| rucc_opt::pass::find(n).is_some()) => {
1367 opts.passes.push((arg["-f".len()..].to_owned(), true));
1368 }
1369 // The unstable options, spelled the way rustc spells them and carrying the same
1370 // promise, which is none: one of these may change or go away in any release. They are
1371 // measurements and debugging aids rather than things a build asks for, which is why
1372 // none of them is in the usage text and all of them are in section 4.11 of
1373 // `spec/04-driver-and-cli.md`.
1374 "-Zverify-each" => opts.verify_each = true,
1375 _ if arg.starts_with("-Zrule-coverage=") => {
1376 let file = &arg["-Zrule-coverage=".len()..];
1377 if file.is_empty() {
1378 return Err(err("-Zrule-coverage= needs a file to write to"));
1379 }
1380 opts.rule_coverage = Some(file.to_owned());
1381 }
1382 _ if arg.starts_with("-Zregister-pressure=") => {
1383 let file = &arg["-Zregister-pressure=".len()..];
1384 if file.is_empty() {
1385 return Err(err("-Zregister-pressure= needs a file to write to"));
1386 }
1387 opts.register_pressure = Some(file.to_owned());
1388 }
1389 _ if arg.starts_with("-Z") => {
1390 return Err(err(format!(
1391 "`{arg}` is not an unstable option this compiler has, see \
1392 spec/04-driver-and-cli.md section 4.11 for the ones it does"
1393 )));
1394 }
1395 // The word size, which is a statement about the target and is taken as one. A build
1396 // that says the size the target already has is saying nothing, and one that says the
1397 // other size is asking for a target this compiler does not have, which it is told
1398 // rather than being given the wrong one.
1399 "-m64" | "-m32" | "-mx32" => {
1400 let want: u32 = match arg {
1401 "-m64" => 64,
1402 _ => 32,
1403 };
1404 let have = rucc_target::TargetInfo::new(opts.target).pointer_width;
1405 if have != want {
1406 return Err(err(format!(
1407 "{arg} asks for a {want} bit target and {} is {have} bit, use \
1408 --target= to name the one you mean",
1409 opts.target
1410 )));
1411 }
1412 }
1413 // Which processor in the family to generate for. This compiler emits the base
1414 // instruction set of the architecture and nothing above it, so a program built with
1415 // any of these runs on the machine that was named; it is a program that could have
1416 // been faster rather than a program that is wrong, which is what makes these safe to
1417 // take and ignore where a flag that changed the meaning of the code would not be.
1418 _ if arg.starts_with("-march=")
1419 || arg.starts_with("-mtune=")
1420 || arg.starts_with("-mcpu=") => {}
1421 // The calling convention, which is not safe to ignore. Taken when it names the one
1422 // the target already uses and refused otherwise.
1423 _ if arg.starts_with("-mabi=") => {
1424 let want = &arg["-mabi=".len()..];
1425 let have = match opts.target.arch {
1426 rucc_target::Arch::X86_64 => "sysv",
1427 rucc_target::Arch::Aarch64 => "lp64",
1428 rucc_target::Arch::Riscv64 => "lp64d",
1429 };
1430 if want != have {
1431 return Err(err(format!(
1432 "{arg}: {} uses the {have} convention and this compiler has no other",
1433 opts.target
1434 )));
1435 }
1436 }
1437 // How far apart the pieces of the program may be. The small model is what we emit and
1438 // it is every hosted program's default; the kernel model is a different one and a
1439 // build that asks for it and does not get it links and then does not run.
1440 "-mcmodel=small" => {}
1441 _ if arg.starts_with("-mcmodel=") => {
1442 return Err(err(format!(
1443 "{arg}: this compiler emits the small code model and no other, see \
1444 spec/12-targets.md"
1445 )));
1446 }
1447 // GCC's own scripting language for how the driver builds a command line.
1448 // `spec/04-driver-and-cli.md` section 4.4 settles that we will not have it, so a
1449 // build reaching for it is told which flags do the same job.
1450 _ if arg.starts_with("-specs=") => {
1451 return Err(err(
1452 "-specs= is not supported: the parts of it builds rely on are -B, -L, \
1453 -nostdlib, -nostartfiles and -Wl,, see spec/04-driver-and-cli.md \
1454 section 4.4",
1455 ));
1456 }
1457 // Arguments meant for a separate assembler or preprocessor, which this compiler does
1458 // not have: both are inside it and neither reads a command line. Refused rather than
1459 // dropped, because every one of these says something about the output and a build
1460 // that asked for `-Wa,--noexecstack` and was silently given an executable stack got
1461 // the opposite of what it asked for.
1462 _ if arg.starts_with("-Wa,") || arg.starts_with("-Wp,") => {
1463 return Err(err(format!(
1464 "`{arg}` is an argument for a separate assembler or preprocessor, and both \
1465 are inside this compiler rather than programs it runs"
1466 )));
1467 }
1468 "-Xassembler" | "-Xpreprocessor" => {
1469 return Err(err(format!(
1470 "{arg} hands an argument to a separate assembler or preprocessor, and both \
1471 are inside this compiler rather than programs it runs"
1472 )));
1473 }
1474 // Everything else in the `-W` family. `spec/04-driver-and-cli.md` section 4.1 has
1475 // this one as a rule about build systems rather than about warnings: autoconf finds
1476 // out whether a warning flag exists by passing it and looking at the exit status, so
1477 // a compiler that refuses one it has not heard of fails a configure script written
1478 // for a GCC newer than itself. The names are not checked against a list because this
1479 // compiler has no warning groups for a list to be of, which #485 is about.
1480 _ if arg.starts_with("-W") => {}
1481 // Flags that name something this compiler does not do and would not do differently
1482 // if it did. `-fno-ident` is about a comment in the output that we do not write
1483 // either way, and the others are about a way of ordering the compilation that has
1484 // been GCC's only way for twenty years. Section 4.1 asks for the list to be short
1485 // and for adding to it to be deliberate, which is why it is written out here.
1486 "-fno-ident"
1487 | "-fident"
1488 | "-funit-at-a-time"
1489 | "-fno-unit-at-a-time"
1490 | "-shared-libgcc"
1491 | "-static-libgcc" => {}
1492 _ if arg.starts_with('-') && arg.len() > 1 => {
1493 // Silently ignoring an unknown flag is how a build ends up not doing what
1494 // its author asked. spec/13-gnu-compat.md section 13.4 makes this an error
1495 // for the flags that change code generation, and the safe default until the
1496 // flag table is populated is to reject everything we do not know.
1497 return Err(err(format!("unknown option `{arg}`")));
1498 }
1499 _ => inputs.push(Input { path: arg.to_owned(), forced, library: false }),
1500 }
1501 }
1502
1503 // Last, so that it lands after every `-isystem` the command line gave. That is GCC's
1504 // order: a directory the user names outranks the compiler's own, and the compiler's own
1505 // outranks the library's. It is pushed after the loop rather than before it because
1506 // `SearchPath` appends within a group and the position is what the order is.
1507 // The same directory the headers were looked for under, because a sysroot is a statement
1508 // about a whole installation and not about half of one.
1509 // After the loop, because `-fno-sanitize=` can take back what an earlier flag asked for and a
1510 // command line that turns a check on and off again has asked for nothing. What is left is
1511 // refused rather than dropped, and it is the one place in this parser where the reason is not
1512 // that the output would differ. A sanitizer is a promise that the program is watched while it
1513 // runs, so a build that asks for one and is quietly given a program with no checks in it does
1514 // not get a slower program or a bigger file, it gets a test suite that passes for the wrong
1515 // reason. `-fsafety=` is the checking this compiler does have, and the message says so, because
1516 // somebody reaching for `-fsanitize=address` wants the nearest thing rather than a list of
1517 // options.
1518 if let Some(first) = sanitizers.first() {
1519 return Err(err(format!(
1520 "-fsanitize={first}: this compiler has no sanitizer instrumentation, and a build that \
1521 asked for one and got none would run its tests unchecked, see \
1522 spec/04-driver-and-cli.md section 4.7. `-fsafety=detect` is the memory checking this \
1523 compiler does have"
1524 )));
1525 }
1526 link.sysroot = sysroot.clone();
1527 // Where a sysroot for a target that is not this machine would be. Read once, here, rather than
1528 // inside the link line, because a link line that read the environment could only be tested on a
1529 // machine whose environment said the right thing, and the link line is the last thing that
1530 // touches a binary. `spec/cross-compile/13-distribution.md` section 13.2 owns the answer.
1531 link.cache = Some(cache::dir());
1532 // And the ten field spelling of the target, because the release on it decides two things the
1533 // three field one cannot say: whether a target that is this architecture is still a cross
1534 // compile, and which directory under the cache it is against. After the loop because the last
1535 // `--target=` on the command line is the one that counts.
1536 link.pinned = pinned;
1537 // After the loop rather than where `-pthread` was read, so that it lands after the objects
1538 // that refer to it. A static link takes the definitions it needs from a library when it
1539 // reaches it and not afterwards, so a library before the objects is a library that answers
1540 // nothing.
1541 if threads {
1542 inputs.push(Input::library("pthread"));
1543 }
1544 if let Some(query) = query {
1545 return Ok(Action::Print(answer(&query, &opts, &link)?));
1546 }
1547 // `-M` and `-MM` produce the rule and nothing else, so the run stops after phase 4 whatever
1548 // else the command line asked for. Read here rather than where the flag was, because a `-c`
1549 // written after it has to lose and the loop cannot know that until it has ended. The output
1550 // file is where the rule goes rather than where an object would have gone, and the last
1551 // phase being the preprocessor is what makes that true without a second rule for it.
1552 if opts.deps.instead_of_compiling {
1553 opts.emit = EmitKind::Preprocessed;
1554 }
1555 if !nostdinc {
1556 opts.search.push_system(runtime::DIR);
1557 // And the library's after ours, which is the other half of the same order. They go on
1558 // here rather than at the point `--target=` or `--sysroot=` was read because either
1559 // one changes the answer and the last word on both is the end of the loop.
1560 //
1561 // Which library's is the question `link::cross_sysroot` answers, and it is asked here so
1562 // that the headers and the libraries come from the same place. A target that is this
1563 // machine reads this machine's headers, and a target that is not reads the ones in the
1564 // sysroot for it rather than the ones next door.
1565 let cross = link::cross_sysroot(opts.target, &link);
1566 let kernel = link::cross_kernel(opts.target, &link);
1567 // And the version of those headers, which only the bundled tree has an answer for. A host
1568 // glibc and a tree the user named both define `__GLIBC_MINOR__` in their own `features.h`,
1569 // and a second definition with a different value is a warning on every file, so the
1570 // condition is the same one that chose the directories.
1571 if cross.is_some() {
1572 let target = pinned.unwrap_or_else(|| opts.target.tuple());
1573 opts.glibc_minor = rucc_sysroot::bundled_glibc_minor(target).map_err(|skew| {
1574 err(format!(
1575 "{skew}; pin a release the tree has, or name a tree that has that one \
1576 with --sysroot"
1577 ))
1578 })?;
1579 }
1580 for dir in
1581 library::header_dirs(opts.target, sysroot.as_deref(), cross.as_ref(), kernel.as_ref())
1582 {
1583 opts.search.push_system(dir);
1584 }
1585 }
1586 // Once, here, rather than as each directory is pushed. A `-I` that names a system
1587 // directory has to lose to the system entry and the system entry is added last, so the
1588 // question cannot be answered until the whole path is known.
1589 opts.search.remove_duplicates();
1590
1591 // The target has to be resolved before the configuration is printed, so this check comes
1592 // after the loop rather than at the point `--print-config` was seen.
1593 if print_config {
1594 return Ok(Action::PrintConfig(Box::new(opts)));
1595 }
1596 if print_pipeline {
1597 return Ok(Action::PrintPipeline(Box::new(opts)));
1598 }
1599 let plan = Plan::new(&opts, &inputs, output.as_deref()).map_err(|e| err(e.message))?;
1600 if print_plan {
1601 return Ok(Action::PrintPlan {
1602 opts: Box::new(opts),
1603 plan: Box::new(plan),
1604 link: Box::new(link),
1605 });
1606 }
1607 Ok(Action::Compile {
1608 opts: Box::new(opts),
1609 plan: Box::new(plan),
1610 link: Box::new(link),
1611 jobs,
1612 verbose,
1613 })
1614}
1615
1616/// What one of the `-dump` and `-print` flags prints.
1617///
1618/// GCC prints the name back unchanged when it cannot find the file a `-print` flag asked about,
1619/// which is what makes the answer safe to paste into a link line whether or not the file is
1620/// there, and this does the same.
1621fn answer(query: &Query, opts: &Options, link: &LinkOptions) -> Result<String, CliError> {
1622 let found = |name: &str| {
1623 link::find_in_search(link, opts.target, name)
1624 .map_or_else(|| name.to_owned(), |path| path.display().to_string())
1625 };
1626 Ok(match query {
1627 Query::Machine => opts.target.to_string(),
1628 Query::Version => VERSION.to_owned(),
1629 Query::Multiarch => link::multiarch(opts.target),
1630 // The three lines GCC prints, in its order and with its punctuation, because what reads
1631 // them is a script written against that shape. There is no installation directory to
1632 // report: this compiler is one binary that works wherever it is copied, and the headers
1633 // it ships are inside it, so `install` is where the binary is and nothing is under it.
1634 Query::SearchDirs => {
1635 let here = std::env::current_exe()
1636 .ok()
1637 .and_then(|p| p.parent().map(std::path::Path::to_path_buf))
1638 .unwrap_or_default();
1639 let list = |dirs: &[PathBuf]| {
1640 dirs.iter().map(|d| d.display().to_string()).collect::<Vec<_>>().join(":")
1641 };
1642 let libraries = link::search_dirs(link, opts.target);
1643 format!(
1644 "install: {}\nprograms: ={}\nlibraries: ={}",
1645 here.display(),
1646 list(&link.prefixes),
1647 list(&libraries)
1648 )
1649 }
1650 // The root the rest of the answers are under, which a build system asks for when it wants
1651 // to find a file itself rather than ask for one by name, and which is the first thing to
1652 // look at when a cross build read a header nobody expected. A native compile has no
1653 // sysroot and the answer is the empty line, which is what GCC prints when it was
1654 // configured without one. `--sysroot` wins over ours because it wins everywhere else.
1655 Query::Sysroot => {
1656 sysroot_root(opts, link).map(|root| root.display().to_string()).unwrap_or_default()
1657 }
1658 // Section 13.5 of `spec/cross-compile/13-distribution.md`: for every input that is not this
1659 // compiler's own code, what it is, where it was got, its hash, its licence and whether it
1660 // was bundled, generated or fetched. What is printed is the manifest the sysroot already
1661 // carries rather than a second format saying the same things, because the three uses 13.5
1662 // gives for this are a licence notice, a reproducibility check and a security audit, and all
1663 // three are somebody else parsing it. One format is one parser to write.
1664 Query::SysrootProvenance => {
1665 let Some(root) = sysroot_root(opts, link) else {
1666 return Ok(String::new());
1667 };
1668 let path = Sysroot::at(root, opts.target.tuple()).manifest_path();
1669 match std::fs::read_to_string(&path) {
1670 // Read and rendered rather than copied out, so that what comes back is the format
1671 // this build understands. A file this build cannot read is a file whose lines it
1672 // cannot vouch for, and printing it anyway would pass the problem to whoever parses
1673 // the output next.
1674 Ok(text) => Manifest::parse(&text)
1675 .map_err(|why| err(format!("{}: {why}", path.display())))?
1676 .render(),
1677 // A tree with no manifest in it is a tree somebody laid out themselves and pointed
1678 // `--sysroot` at, and nothing here knows where any of it came from. The answer is
1679 // nothing, which a reader can tell apart from a manifest with no inputs in it
1680 // because that one still has its header line.
1681 Err(why) if why.kind() == std::io::ErrorKind::NotFound => String::new(),
1682 Err(why) => return Err(err(format!("{}: {why}", path.display()))),
1683 }
1684 }
1685 Query::FileName(name) => found(name),
1686 // The name GCC gives the library of routines a compiler's output calls that the C
1687 // library does not have. Ours is built in and there is no file, so the answer is the
1688 // name itself, which is what GCC prints when it cannot find one either.
1689 Query::Libgcc => found("libgcc.a"),
1690 // A program rather than a library: the linker and the archiver are the ones a build asks
1691 // about, and this compiler finds them on the path or under `-B` rather than shipping
1692 // them, so the name back is the honest answer unless a `-B` prefix holds one.
1693 Query::ProgName(name) => link
1694 .prefixes
1695 .iter()
1696 .map(|dir| dir.join(name))
1697 .find(|path| path.is_file())
1698 .map_or_else(|| name.clone(), |path| path.display().to_string()),
1699 })
1700}
1701
1702/// The root both of the sysroot answers are about.
1703///
1704/// One function rather than a copy in each, because the second flag exists to say what is inside the
1705/// tree the first one names, and two answers that disagreed about which tree that is would be a
1706/// difference nobody would think to look for. `--sysroot` wins over ours because it wins everywhere
1707/// else.
1708fn sysroot_root(opts: &Options, link: &LinkOptions) -> Option<PathBuf> {
1709 link.sysroot
1710 .clone()
1711 .or_else(|| link::cross_sysroot(opts.target, link).map(|at| at.root().to_path_buf()))
1712}
1713
1714/// Renders the passes this level will run, in order, with what each one does.
1715///
1716/// The level is the whole of the answer unless a `-f` flag edited it, which is section 9.1 of
1717/// `spec/09-optimizer.md`: a level is a list somebody wrote down rather than something that
1718/// emerges from which flags happen to be set, and this is how that list is read.
1719#[must_use]
1720pub fn print_pipeline(opts: &Options) -> String {
1721 let mut settings = rucc_opt::Options::for_level(opts.opt_level);
1722 settings.toggles.clone_from(&opts.passes);
1723 settings.global_fuel = opts.pass_fuel_global;
1724 for (on, spec) in &opts.pass_gates {
1725 // Every spelling was checked while the arguments were parsed, so there is nothing here
1726 // this can refuse, and a listing is not the place to report it if there were.
1727 let _ = settings.gates.add(*on, spec);
1728 }
1729 rucc_opt::pipeline::print(&settings)
1730}
1731
1732/// Renders the resolved configuration.
1733///
1734/// One `key: value` per line, sorted by nothing in particular but fixed in order, because
1735/// this output is diffed across hosts in CI and a reordering would read as a change.
1736#[must_use]
1737pub fn print_config(opts: &Options) -> String {
1738 let sess = Session::new(opts.clone());
1739 let t = &sess.target;
1740 let mut out = String::new();
1741 let _ = writeln!(out, "version: {VERSION}");
1742 // The three field triple the driver was given rather than the ten field tuple it widens to,
1743 // because this output is what a build system reads to find out what it asked for. The tuple is
1744 // the compiler's model of the machine and this line is a receipt for a command line.
1745 let _ = writeln!(out, "target: {}", opts.target);
1746 let _ = writeln!(out, "arch: {}", opts.target.arch.as_str());
1747 let _ = writeln!(out, "os: {}", opts.target.os.as_str());
1748 let _ = writeln!(out, "env: {}", opts.target.env.as_str());
1749 let _ = writeln!(out, "object-format: {}", t.object_format.as_str());
1750 let _ = writeln!(out, "pointer-width: {}", t.pointer_width);
1751 let _ = writeln!(out, "long-width: {}", t.long_width);
1752 let _ = writeln!(out, "long-double-width: {}", t.long_double_width);
1753 let _ = writeln!(out, "endian: {}", if t.little_endian { "little" } else { "big" });
1754 let _ = writeln!(out, "char-signed: {}", t.char_is_signed);
1755 let _ = writeln!(out, "va-list: {}", t.va_list.map_or("none", |list| list.as_str()));
1756 // The register file as a count per class, which is enough to tell a target whose registers
1757 // are described from one whose are not without printing sixteen names nobody asked for.
1758 let regs: Vec<String> = t
1759 .regs
1760 .classes()
1761 .map(|(class, info)| format!("{} {}", info.name, t.regs.len(class)))
1762 .collect();
1763 let _ = writeln!(
1764 out,
1765 "registers: {}",
1766 if regs.is_empty() { "none".to_string() } else { regs.join(", ") }
1767 );
1768 let _ = writeln!(out, "opt-level: {}", sess.opts.opt_level);
1769 let _ = writeln!(out, "safety: {}", sess.opts.safety);
1770 let _ = writeln!(out, "emit: {}", sess.opts.emit.as_str());
1771 let _ = writeln!(out, "debug-info: {}", sess.opts.debug_info);
1772 let _ = writeln!(out, "frame-pointer: {}", sess.opts.frame_pointer);
1773 let _ = writeln!(out, "red-zone: {}", sess.opts.red_zone);
1774 let _ = writeln!(out, "stack-protector: {}", sess.opts.protector);
1775 let _ = writeln!(out, "stack-clash-protection: {}", sess.opts.stack_clash);
1776 let _ = writeln!(out, "cf-protection: {}", sess.opts.control);
1777 let _ = writeln!(out, "patchable-function-entry: {}", sess.opts.patchable);
1778 let _ = writeln!(out, "profile: {}", sess.opts.profile);
1779 let _ = writeln!(out, "profile-hook: {}", sess.opts.hook);
1780 // Last because it is the one key with more than one line under it, and the only one
1781 // whose value is a property of the machine rather than of the command line.
1782 for dir in sess.opts.search.dirs() {
1783 let system = if dir.is_system { " (system)" } else { "" };
1784 let _ = writeln!(out, "include: {}{system}", dir.path.display());
1785 }
1786 out
1787}
1788
1789/// The output name the make target is taken from, which is the `-o` argument or nothing.
1790///
1791/// A run that stops at the preprocessor has not named an object, whatever its `-o` says: under
1792/// `-E` that argument is the preprocessed text and under `-M` it is the rule itself, and neither
1793/// is a file `make` would rebuild by running this rule. GCC agrees and falls back to the source
1794/// name in both, which is why a `-MD -E -o out.i` writes `out.d` holding a rule for `a.o`. From
1795/// `-S` on the argument does name what the rule builds, and it is used as written.
1796fn deps_target_output<'a>(opts: &Options, plan: &'a Plan) -> Option<&'a str> {
1797 if opts.emit == EmitKind::Preprocessed { None } else { plan.output.as_deref() }
1798}
1799
1800/// Writes to a path the command line named rather than one the plan derived, where `-` is
1801/// standard output.
1802fn write_named(path: &str, bytes: &[u8]) -> Result<(), String> {
1803 if path == "-" {
1804 return write_out(&Output::Stdout, bytes);
1805 }
1806 write_out(&Output::File(path.to_owned()), bytes)
1807}
1808
1809/// Writes the make rule for one input, and reports whether it got there.
1810///
1811/// A rule with no file of its own goes where the compilation it replaced would have written,
1812/// which is what makes the usual makefile recipe work: `rucc -M $< -o $@` leaves the rule in
1813/// `$@`, and the same line with the `-o` left off puts it on standard output.
1814fn write_deps(
1815 opts: &Options,
1816 plan: &Plan,
1817 job: &Job,
1818 found: &[Dependency],
1819 stderr: &mut impl std::io::Write,
1820) -> bool {
1821 let targets = if opts.deps.targets.is_empty() {
1822 vec![deps::default_target(&job.input, deps_target_output(opts, plan))]
1823 } else {
1824 opts.deps.targets.clone()
1825 };
1826 let rule = deps::rule(&opts.deps, &targets, &job.input, found);
1827 // The file, on the other hand, is named after the `-o` in every mode that still has one to
1828 // spend, which is every mode except the two that spend it on the rule.
1829 let wrote = match deps::default_file(&opts.deps, &job.input, plan.output.as_deref()) {
1830 // A `-MF` on a run that had nowhere else to put the rule leaves the file the `-o`
1831 // named empty rather than absent, because a makefile that named it as a target of its
1832 // own is a makefile that will look for it.
1833 Some(path) => write_named(&path, rule.as_bytes()).and_then(|()| {
1834 if opts.deps.instead_of_compiling { write_out(&job.output, b"") } else { Ok(()) }
1835 }),
1836 None => write_out(&job.output, rule.as_bytes()),
1837 };
1838 if let Err(e) = wrote {
1839 let _ = writeln!(stderr, "rucc: error: {e}");
1840 return false;
1841 }
1842 true
1843}
1844
1845/// Runs phase 4 over every input that has one, and writes what came out.
1846///
1847/// One input that fails does not stop the others. A build that reports every file it could
1848/// not preprocess in one run is worth more than one that stops at the first, and the exit
1849/// status is still a failure either way.
1850fn preprocess_all(opts: &Options, plan: &Plan) -> i32 {
1851 let fs = OsFileSystem::new();
1852 let mut stderr = std::io::stderr().lock();
1853 let mut failed = false;
1854 for job in &plan.jobs {
1855 if !job.phases.first().is_some_and(|p| *p == Phase::Preprocess) {
1856 // An input that is already preprocessed, or an object file. GCC passes these
1857 // through untouched, and the plan has already said so in its notes.
1858 continue;
1859 }
1860 let started = std::time::Instant::now();
1861 let result = preprocess(opts, &job.input, &fs);
1862 if opts.time {
1863 say_time(&job.input, started.elapsed(), &mut stderr);
1864 }
1865 for message in &result.messages {
1866 let _ = writeln!(stderr, "{message}");
1867 }
1868 if result.failed() {
1869 failed = true;
1870 continue;
1871 }
1872 if opts.deps.emit {
1873 failed |= !write_deps(opts, plan, job, &result.deps, &mut stderr);
1874 // `-M` and `-MM` asked for the rule instead of the text, so there is nothing else
1875 // to write. The other two asked for both and fall through to the text below.
1876 if opts.deps.instead_of_compiling {
1877 continue;
1878 }
1879 }
1880 if let Err(e) = write_out(&job.output, result.text.as_bytes()) {
1881 let _ = writeln!(stderr, "rucc: error: {e}");
1882 failed = true;
1883 }
1884 }
1885 i32::from(failed)
1886}
1887
1888/// Runs the front end over every input that has a compile phase, and writes what came out.
1889///
1890/// The same rule as [`preprocess_all`]: one input that fails does not stop the others, and the
1891/// exit status is a failure either way. An input that is already assembly or an object has no
1892/// compile phase and is passed over here, which the plan has already said in its notes.
1893fn compile_all(opts: &Options, plan: &Plan) -> i32 {
1894 let fs = OsFileSystem::new();
1895 let mut stderr = std::io::stderr().lock();
1896 let mut failed = false;
1897 let (mut remarks, ok) = Remarks::new(opts.opt_info_file.as_ref(), &mut stderr);
1898 failed |= !ok;
1899 let mut fired = Fired::new();
1900 let mut pressure = Pressure::new();
1901 for job in &plan.jobs {
1902 if !job.phases.contains(&Phase::Compile) {
1903 continue;
1904 }
1905 // An input of IR is read back rather than compiled, since the C it came from is not
1906 // here any more. Everything after this is the same, so the two paths meet again at the
1907 // messages and the file the result is written to.
1908 let started = std::time::Instant::now();
1909 let result = if job.kind == InputKind::Ir {
1910 compile_ir(opts, &job.input, &fs)
1911 } else {
1912 compile(opts, &job.input, &fs)
1913 };
1914 if opts.time {
1915 say_time(&job.input, started.elapsed(), &mut stderr);
1916 }
1917 fired.merge(&result.fired);
1918 pressure.merge(&result.pressure);
1919 failed |= !write_dumps(&job.input, &result.dumps, &mut stderr);
1920 failed |= !remarks.write(&result.remarks, &mut stderr);
1921 for message in &result.messages {
1922 let _ = writeln!(stderr, "{message}");
1923 }
1924 // Before the failure below, because a compilation that stopped in the back end is exactly
1925 // the one whose preprocessed source somebody wants to look at.
1926 failed |= !write_temps(job, &result.temps, &mut stderr);
1927 if result.failed() {
1928 failed = true;
1929 continue;
1930 }
1931 // `-MD` and `-MMD` write the rule beside the object and let the compilation happen, so
1932 // this is the one path where both files come out of the same run. An input of IR has no
1933 // dependencies to report and produces an empty list, which produces a rule naming only
1934 // itself, and that is the honest answer rather than a missing file.
1935 if opts.deps.emit {
1936 failed |= !write_deps(opts, plan, job, &result.deps, &mut stderr);
1937 }
1938 if let Err(e) = write_out(&job.output, result.artifact.bytes()) {
1939 let _ = writeln!(stderr, "rucc: error: {e}");
1940 failed = true;
1941 }
1942 }
1943 failed |= !write_coverage(opts, &fired, &mut stderr);
1944 failed |= !write_pressure(opts, &pressure, &mut stderr);
1945 i32::from(failed)
1946}
1947
1948/// A directory for the object files only the link step ever sees, removed when it goes away.
1949///
1950/// `-c` writes its object where the user can see it and linking does not, which is the whole of
1951/// the difference: a `rucc a.c b.c` leaves an executable behind and nothing else, the same as
1952/// every other compiler. Removing them on drop rather than at the end of a function is so that a
1953/// link that failed leaves nothing behind either.
1954struct Scratch {
1955 /// Where the objects go.
1956 dir: PathBuf,
1957}
1958
1959impl Scratch {
1960 /// Makes one, under whatever the platform calls its temporary directory.
1961 ///
1962 /// The name carries the process id so that two compilers running at once do not share a
1963 /// directory, which they would otherwise do the moment two of them compiled a file of the
1964 /// same name.
1965 fn new() -> Result<Scratch, String> {
1966 let dir = std::env::temp_dir().join(format!("rucc-{}", std::process::id()));
1967 std::fs::create_dir_all(&dir).map_err(|e| format!("{}: {e}", dir.display()))?;
1968 Ok(Scratch { dir })
1969 }
1970}
1971
1972impl Drop for Scratch {
1973 fn drop(&mut self) {
1974 let _ = std::fs::remove_dir_all(&self.dir);
1975 }
1976}
1977
1978/// The link line the plan describes, for `-###`.
1979///
1980/// The names in it are the hints the plan carries rather than the temporaries a real compilation
1981/// would choose, because `-###` prints the line without having compiled anything and so has
1982/// nothing to point at. That also makes the printed line readable rather than naming a directory
1983/// that only exists while a compilation is running.
1984fn link_line(opts: &Options, link: &LinkOptions, job: &LinkJob) -> Result<String, link::Error> {
1985 let linker = link::find(opts.target, link)?;
1986 let args = link::line(opts.target, link, &job.inputs, &job.output)?;
1987 Ok(link::render(&linker, &args))
1988}
1989
1990/// Compiles everything, then links it.
1991///
1992/// The objects go in a directory that is removed afterwards, which is why this is not
1993/// [`compile_all`] followed by a link: the plan says an object feeding the linker is temporary
1994/// and does not say where, because where is a question that only has an answer once something is
1995/// running.
1996fn link_all(opts: &Options, plan: &Plan, link: &LinkOptions, verbose: bool) -> i32 {
1997 let Some(job) = &plan.link else {
1998 // Every path into here comes from a plan whose last phase is the link, and such a plan
1999 // has a link job. Saying so is cheaper than an unwrap that would have to be explained.
2000 let mut stderr = std::io::stderr().lock();
2001 let _ = writeln!(stderr, "rucc: error: there is nothing to link");
2002 return 1;
2003 };
2004 // Before anything is compiled, because a linker that is not on the machine is worth knowing
2005 // about in the second it takes to look rather than after the compilation.
2006 // And before that, whether this link has a line at all and whether what it reads is on the
2007 // machine. Both are answerable now, and a target whose sysroot has not been built is worth
2008 // saying so about before the compilation rather than after it.
2009 if let Err(why) = link::preflight(opts.target, link) {
2010 return complain(why);
2011 }
2012 let linker = match link::find(opts.target, link) {
2013 Ok(linker) => linker,
2014 Err(why) => return complain(why),
2015 };
2016
2017 let scratch = match Scratch::new() {
2018 Ok(scratch) => scratch,
2019 Err(why) => return complain(format!("could not make a place for the object files: {why}")),
2020 };
2021
2022 let fs = OsFileSystem::new();
2023 let mut failed = false;
2024 // One per job, in job order, which is what lets the link line below be rebuilt with the real
2025 // paths in it: every job contributes exactly one file to the line and does so in this order.
2026 let mut produced: Vec<String> = Vec::with_capacity(plan.jobs.len());
2027 let mut fired = Fired::new();
2028 let mut pressure = Pressure::new();
2029 {
2030 let mut stderr = std::io::stderr().lock();
2031 let (mut remarks, ok) = Remarks::new(opts.opt_info_file.as_ref(), &mut stderr);
2032 failed |= !ok;
2033 for (at, job) in plan.jobs.iter().enumerate() {
2034 let out = match &job.output {
2035 Output::Temporary(hint) => {
2036 // The index because two inputs in different directories can have the same
2037 // name, and the two objects of `rucc a/x.c b/x.c` must not be one file.
2038 scratch.dir.join(format!("{at}-{hint}")).display().to_string()
2039 }
2040 Output::File(path) => path.clone(),
2041 // A job feeding the linker never writes to standard output, since the plan gives
2042 // it a temporary. This is here so that the match is total rather than a panic.
2043 Output::Stdout => continue,
2044 };
2045 produced.push(out.clone());
2046 if !job.phases.contains(&Phase::Compile) {
2047 continue;
2048 }
2049 let started = std::time::Instant::now();
2050 let result = if job.kind == InputKind::Ir {
2051 compile_ir(opts, &job.input, &fs)
2052 } else {
2053 compile(opts, &job.input, &fs)
2054 };
2055 if opts.time {
2056 say_time(&job.input, started.elapsed(), &mut stderr);
2057 }
2058 fired.merge(&result.fired);
2059 pressure.merge(&result.pressure);
2060 failed |= !write_dumps(&job.input, &result.dumps, &mut stderr);
2061 failed |= !remarks.write(&result.remarks, &mut stderr);
2062 for message in &result.messages {
2063 let _ = writeln!(stderr, "{message}");
2064 }
2065 failed |= !write_temps(job, &result.temps, &mut stderr);
2066 if result.failed() {
2067 failed = true;
2068 continue;
2069 }
2070 // A `-MD` on a command line that links writes the rule next to the executable and
2071 // names the executable as its target, since that is the file this source builds
2072 // here. The object it went through is in a temporary directory and is gone by the
2073 // time `make` reads any of this.
2074 if opts.deps.emit {
2075 failed |= !write_deps(opts, plan, job, &result.deps, &mut stderr);
2076 }
2077 if !matches!(result.artifact, Artifact::Object(_)) {
2078 // Worth saying rather than writing whatever it is and letting the linker read it.
2079 // An empty file is a valid empty linker script, so a link handed one gets as far
2080 // as reporting every symbol of this file undefined, which is a page of messages
2081 // about something that went wrong here.
2082 let _ = writeln!(
2083 stderr,
2084 "rucc: internal error: {}: no object file was produced for the link",
2085 job.input
2086 );
2087 failed = true;
2088 continue;
2089 }
2090 if let Err(e) = std::fs::write(&out, result.artifact.bytes()) {
2091 let _ = writeln!(stderr, "rucc: error: {out}: {e}");
2092 failed = true;
2093 }
2094 }
2095 failed |= !write_coverage(opts, &fired, &mut stderr);
2096 failed |= !write_pressure(opts, &pressure, &mut stderr);
2097 }
2098 if failed {
2099 // Nothing is linked from a compilation that did not finish. A linker run over the objects
2100 // that did compile would report every function of the file that did not as undefined,
2101 // which is a page of messages about a mistake already reported once.
2102 return 1;
2103 }
2104
2105 // The items in command line order with the temporaries filled in. A library contributes no
2106 // job and passes through, and every file item takes the next job's real output, which is
2107 // what keeps a library that was written between two objects between them here.
2108 let mut outputs = produced.into_iter();
2109 let mut items = Vec::with_capacity(job.inputs.len());
2110 for item in &job.inputs {
2111 match item {
2112 link::Item::Library(name) => items.push(link::Item::Library(name.clone())),
2113 link::Item::File(_) => match outputs.next() {
2114 Some(path) => items.push(link::Item::File(path)),
2115 None => return complain("the plan asks the linker for a file nothing produced"),
2116 },
2117 }
2118 }
2119
2120 let args = match link::line(opts.target, link, &items, &job.output) {
2121 Ok(args) => args,
2122 Err(why) => return complain(why),
2123 };
2124 if verbose {
2125 let mut stderr = std::io::stderr().lock();
2126 let _ = writeln!(stderr, "{}", link::render(&linker, &args));
2127 }
2128 let started = std::time::Instant::now();
2129 let ran = link::run(&linker, &args);
2130 if opts.time {
2131 // The one step of a compilation that really is another program, so this line is the same
2132 // measurement gcc's is and names the linker the way gcc names `collect2`.
2133 let mut stderr = std::io::stderr().lock();
2134 say_time(&linker.name, started.elapsed(), &mut stderr);
2135 }
2136 match ran {
2137 Ok(()) => 0,
2138 // The linker has already said what was wrong on its own error output, and repeating that
2139 // linking failed would only push its message further up the screen.
2140 Err(link::Error::Refused { .. }) => 1,
2141 Err(why) => complain(why),
2142 }
2143}
2144
2145/// Prints one driver level message and gives back the exit status that goes with it.
2146fn complain(why: impl std::fmt::Display) -> i32 {
2147 let mut stderr = std::io::stderr().lock();
2148 let _ = writeln!(stderr, "rucc: error: {why}");
2149 1
2150}
2151
2152/// Writes what `-Zrule-coverage=FILE` asked for, and says whether it could.
2153///
2154/// Once for the whole command line rather than once per input, because the question is which
2155/// lowering rules this run of the compiler reached and a file per input would leave the reader
2156/// unioning files to find out something one process already knew.
2157///
2158/// A file that could not be written is a failure and not a warning. What asks for this is a
2159/// measurement run, and a measurement that quietly did not happen is worse than one that stopped.
2160fn write_coverage(opts: &Options, fired: &Fired, stderr: &mut impl std::io::Write) -> bool {
2161 let Some(path) = &opts.rule_coverage else { return true };
2162 let Some(table) = coverage::table(opts.target.arch) else {
2163 let _ = writeln!(
2164 stderr,
2165 "rucc: error: there are no lowering rules for {} yet, so there is no coverage of them \
2166 to report",
2167 opts.target
2168 );
2169 return false;
2170 };
2171 match std::fs::write(path, fired.listing(table)) {
2172 Ok(()) => true,
2173 Err(e) => {
2174 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2175 false
2176 }
2177 }
2178}
2179
2180/// Writes what `-Zregister-pressure=FILE` asked for, and says whether it could.
2181///
2182/// Once for the whole command line, for the reason [`write_coverage`] gives, and a file that could
2183/// not be written is a failure for the reason it gives too. There is no equivalent of the missing
2184/// rule table here, since every target this compiles for has an allocator, and a run that reached
2185/// no back end at all writes an empty listing rather than nothing: a measurement of a build that
2186/// produced no code is still an answer and it is the honest one.
2187fn write_pressure(opts: &Options, pressure: &Pressure, stderr: &mut impl std::io::Write) -> bool {
2188 let Some(path) = &opts.register_pressure else { return true };
2189 match std::fs::write(path, pressure.listing()) {
2190 Ok(()) => true,
2191 Err(e) => {
2192 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2193 false
2194 }
2195 }
2196}
2197
2198/// Where the `-fopt-info` remarks go, and how much of the run has already gone there.
2199///
2200/// Standard error by default, and one file for the whole run when `-fopt-info=<file>` named one.
2201/// A file rather than the diagnostic stream is what a harness wants: the corpus in
2202/// `tamnd/rucc-corpus` matches a rejection against what the compiler said on standard error, and
2203/// a few thousand remarks mixed into that would bury it.
2204struct Remarks {
2205 /// The file, if there is one.
2206 file: Option<String>,
2207 /// Whether anything has been written to it yet, which decides between truncating and
2208 /// appending. One file holds the whole run rather than the last input in it.
2209 started: bool,
2210}
2211
2212impl Remarks {
2213 /// Prepares the destination, emptying the file if there is one.
2214 ///
2215 /// Emptied here rather than at the first remark, because a run where no pass had anything to
2216 /// say should leave an empty file and not yesterday's. An absent file and an empty one are
2217 /// different facts and something reading this will act on the difference.
2218 fn new(file: Option<&String>, stderr: &mut impl std::io::Write) -> (Self, bool) {
2219 let mut ok = true;
2220 if let Some(path) = file {
2221 if let Err(e) = std::fs::write(path, "") {
2222 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2223 ok = false;
2224 }
2225 }
2226 (Self { file: file.cloned(), started: false }, ok)
2227 }
2228
2229 /// Writes one input's remarks, and says whether that worked.
2230 ///
2231 /// A file that cannot be written is a failure and not a warning, for the reason
2232 /// [`write_dumps`] gives: remarks that quietly did not arrive look exactly like a compilation
2233 /// where nothing happened.
2234 fn write(&mut self, text: &str, stderr: &mut impl std::io::Write) -> bool {
2235 if text.is_empty() {
2236 return true;
2237 }
2238 let Some(path) = &self.file else {
2239 let _ = write!(stderr, "{text}");
2240 return true;
2241 };
2242 let opened = std::fs::OpenOptions::new()
2243 .write(true)
2244 .append(self.started)
2245 .truncate(!self.started)
2246 .create(true)
2247 .open(path);
2248 self.started = true;
2249 let result =
2250 opened.and_then(|mut file| std::io::Write::write_all(&mut file, text.as_bytes()));
2251 if let Err(e) = result {
2252 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2253 return false;
2254 }
2255 true
2256 }
2257}
2258
2259/// Writes what `-fdump-ir=` asked to see, one file per dump.
2260///
2261/// The name is the input file with the dump's own name and `.ir` after it, so a directory listing
2262/// after a run is the passes in the order they ran, per input. They go in the working directory
2263/// rather than beside the output, because a dump is something a person asked for at a prompt and
2264/// the working directory is where that person is.
2265///
2266/// A file that could not be written is a failure and not a warning, for the reason
2267/// [`write_coverage`] gives: what asked for this is somebody debugging a pass, and a dump that
2268/// quietly did not happen looks exactly like a pass that did not run.
2269fn write_dumps(input: &str, dumps: &[rucc_opt::Dump], stderr: &mut impl std::io::Write) -> bool {
2270 let stem = std::path::Path::new(input)
2271 .file_name()
2272 .map_or_else(|| input.to_owned(), |name| name.to_string_lossy().into_owned());
2273 let mut ok = true;
2274 for dump in dumps {
2275 let path = format!("{stem}.{}.ir", dump.name);
2276 if let Err(e) = std::fs::write(&path, &dump.text) {
2277 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2278 ok = false;
2279 }
2280 }
2281 ok
2282}
2283
2284/// Writes the files `-save-temps` kept, which is nothing at all unless it was given.
2285///
2286/// A file that could not be written is a failure rather than a warning, for the reason
2287/// [`write_dumps`] gives: somebody asked for these by name, and one that quietly did not happen
2288/// looks like a compilation that never went through that step.
2289fn write_temps(job: &Job, temps: &Temps, stderr: &mut impl std::io::Write) -> bool {
2290 let mut ok = true;
2291 let kept = [(job.saved_text(), &temps.preprocessed), (job.saved_asm(), &temps.assembly)];
2292 for (path, text) in kept {
2293 // A step the compilation did not reach has nothing to keep, and a job that is not keeping
2294 // that step has nowhere to put it. Either way there is no file here.
2295 let (Some(path), Some(text)) = (path, text) else { continue };
2296 if let Err(e) = std::fs::write(&path, text) {
2297 let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2298 ok = false;
2299 }
2300 }
2301 ok
2302}
2303
2304/// One line of `-time`, which is what a step was called and how long it took.
2305///
2306/// GCC's two numbers are the user and the system time of a subprocess it ran. This compiler runs
2307/// no subprocess for anything but the link, so what is measured here is the wall clock of the
2308/// step and the second column is always zero. The shape of the line is kept because a person
2309/// reading it next to gcc's should not have to work out which column is which.
2310fn say_time(name: &str, took: std::time::Duration, stderr: &mut impl std::io::Write) {
2311 let _ = writeln!(stderr, "# {name} {:.2} {:.2}", took.as_secs_f64(), 0.0);
2312}
2313
2314/// Writes one job's result where the plan said it goes.
2315///
2316/// # Errors
2317///
2318/// Returns the message to print, which names the file when there is one, because "permission
2319/// denied" on its own does not say which file was refused.
2320fn write_out(output: &Output, bytes: &[u8]) -> Result<(), String> {
2321 match output {
2322 Output::Stdout => {
2323 let mut stdout = std::io::stdout().lock();
2324 stdout.write_all(bytes).map_err(|e| format!("writing to standard output: {e}"))
2325 }
2326 Output::File(path) | Output::Temporary(path) => {
2327 std::fs::write(path, bytes).map_err(|e| format!("{path}: {e}"))
2328 }
2329 }
2330}
2331
2332/// Runs the driver and returns the process exit code.
2333///
2334/// `args` excludes the program name. Output goes to `stdout` and errors to `stderr`, which
2335/// is the one place in the compiler that is true.
2336pub fn run(args: &[String]) -> i32 {
2337 match parse_args(args) {
2338 Ok(Action::Help) => {
2339 print!("{USAGE}");
2340 0
2341 }
2342 Ok(Action::Version) => {
2343 println!("rucc {VERSION}");
2344 0
2345 }
2346 Ok(Action::Print(line)) => {
2347 println!("{line}");
2348 0
2349 }
2350 Ok(Action::PrintConfig(opts)) => {
2351 print!("{}", print_config(&opts));
2352 0
2353 }
2354 Ok(Action::PrintPipeline(opts)) => {
2355 print!("{}", print_pipeline(&opts));
2356 0
2357 }
2358 Ok(Action::PrintPlan { opts, plan, link }) => {
2359 print!("{}", plan.render());
2360 // The line as it would be typed, which is the half of `-###` that section 4.3 says
2361 // arrives with the link. It is printed even when the linker is not on this machine,
2362 // because what a build wants from `-###` is what the compiler would do.
2363 if let Some(job) = &plan.link {
2364 match link_line(&opts, &link, job) {
2365 Ok(line) => println!("{line}"),
2366 Err(why) => {
2367 let mut stderr = std::io::stderr().lock();
2368 let _ = writeln!(stderr, "rucc: error: {why}");
2369 return 1;
2370 }
2371 }
2372 }
2373 0
2374 }
2375 Ok(Action::Compile { opts, plan, link, jobs, verbose }) => {
2376 {
2377 let mut stderr = std::io::stderr().lock();
2378 if verbose {
2379 let _ = write!(stderr, "{}", plan.render());
2380 let _ = writeln!(stderr, "workers: {}", jobs.count());
2381 }
2382 }
2383 if opts.emit == EmitKind::Preprocessed {
2384 return preprocess_all(&opts, &plan);
2385 }
2386 if opts.emit != EmitKind::Executable {
2387 return compile_all(&opts, &plan);
2388 }
2389 link_all(&opts, &plan, &link, verbose)
2390 }
2391 Err(e) => {
2392 let mut stderr = std::io::stderr().lock();
2393 let _ = writeln!(stderr, "rucc: error: {e}");
2394 let _ = writeln!(stderr, "rucc: note: run `rucc --help` for usage");
2395 1
2396 }
2397 }
2398}
2399
2400#[cfg(test)]
2401mod tests {
2402 use rucc_session::{
2403 Contract, GnucVersion, IncludeForm, LtoJobs, OptLevel, Partition, Patchable, Visibility,
2404 };
2405
2406 use super::*;
2407
2408 fn args(s: &[&str]) -> Vec<String> {
2409 s.iter().map(|x| (*x).to_owned()).collect()
2410 }
2411
2412 #[test]
2413 fn help_and_version_win_over_everything_else() {
2414 assert_eq!(parse_args(&args(&["-c", "--help", "x.c"])).unwrap(), Action::Help);
2415 assert_eq!(parse_args(&args(&["--version"])).unwrap(), Action::Version);
2416 }
2417
2418 fn compile(s: &[&str]) -> (Box<Options>, Box<Plan>) {
2419 match parse_args(&args(s)).expect("expected a compilation") {
2420 Action::Compile { opts, plan, .. } => (opts, plan),
2421 other => panic!("expected a compilation, got {other:?}"),
2422 }
2423 }
2424
2425 fn linking(s: &[&str]) -> (Box<LinkOptions>, Box<Plan>) {
2426 match parse_args(&args(s)).expect("expected a compilation") {
2427 Action::Compile { link, plan, .. } => (link, plan),
2428 other => panic!("expected a compilation, got {other:?}"),
2429 }
2430 }
2431
2432 #[test]
2433 fn collects_inputs_and_flags() {
2434 let (opts, plan) = compile(&["-c", "-O2", "-g", "a.c", "b.c"]);
2435 let paths: Vec<&str> = plan.jobs.iter().map(|j| j.input.as_str()).collect();
2436 assert_eq!(paths, vec!["a.c", "b.c"]);
2437 assert_eq!(opts.opt_level, OptLevel::O2);
2438 assert_eq!(opts.emit, EmitKind::Object);
2439 assert!(opts.debug_info);
2440 }
2441
2442 /// The unstable options, which are spelled apart from everything else on purpose: what is
2443 /// under `-Z` promises nothing, and a build that reaches for one should have had to say so.
2444 #[test]
2445 fn an_unstable_option_is_taken_and_one_that_does_not_exist_is_refused() {
2446 let (opts, _) = compile(&["-c", "-Zrule-coverage=/tmp/rules.cov", "a.c"]);
2447 assert_eq!(opts.rule_coverage.as_deref(), Some("/tmp/rules.cov"));
2448
2449 let (plain, _) = compile(&["-c", "a.c"]);
2450 assert_eq!(plain.rule_coverage, None, "nothing is measured unless it was asked for");
2451
2452 assert!(parse_args(&args(&["-Zrule-coverage=", "a.c"])).is_err(), "a file with no name");
2453 let unknown = parse_args(&args(&["-Zwhat", "a.c"])).expect_err("there is no such option");
2454 assert!(unknown.message.contains("4.11"), "{}", unknown.message);
2455 }
2456
2457 /// The other measurement written to a file, which reads the same way and fails the same way.
2458 #[test]
2459 fn where_the_register_pressure_goes_is_asked_for_the_same_way() {
2460 let (opts, _) = compile(&["-c", "-O2", "-Zregister-pressure=/tmp/spills.txt", "a.c"]);
2461 assert_eq!(opts.register_pressure.as_deref(), Some("/tmp/spills.txt"));
2462
2463 let (plain, _) = compile(&["-c", "a.c"]);
2464 assert_eq!(plain.register_pressure, None, "nothing is measured unless it was asked for");
2465
2466 assert!(parse_args(&args(&["-Zregister-pressure=", "a.c"])).is_err(), "no file named");
2467 }
2468
2469 #[test]
2470 fn a_bare_dash_o_means_o1_the_way_gcc_reads_it() {
2471 let (opts, _) = compile(&["-O", "a.c"]);
2472 assert_eq!(opts.opt_level, OptLevel::O1);
2473 }
2474
2475 #[test]
2476 fn dash_x_applies_to_later_inputs_only_and_none_stops_it() {
2477 let (_, plan) = compile(&["a.o", "-x", "c", "b.txt", "-x", "none", "c.o"]);
2478 assert_eq!(plan.jobs[0].kind, InputKind::LinkerInput);
2479 assert_eq!(plan.jobs[1].kind, InputKind::C);
2480 assert_eq!(plan.jobs[2].kind, InputKind::LinkerInput);
2481 }
2482
2483 #[test]
2484 fn dash_j_reaches_the_scheduler_and_defaults_to_the_machine() {
2485 let (_, _, jobs) = match parse_args(&args(&["-j4", "a.c"])).unwrap() {
2486 Action::Compile { opts, plan, jobs, .. } => (opts, plan, jobs),
2487 other => panic!("expected a compilation, got {other:?}"),
2488 };
2489 assert_eq!(jobs.count(), 4);
2490
2491 let default = match parse_args(&args(&["a.c"])).unwrap() {
2492 Action::Compile { jobs, .. } => jobs,
2493 other => panic!("expected a compilation, got {other:?}"),
2494 };
2495 assert_eq!(default, Jobs::available());
2496 assert!(parse_args(&args(&["-j0", "a.c"])).is_err());
2497 }
2498
2499 #[test]
2500 fn triple_hash_prints_the_plan_and_runs_nothing() {
2501 let a = parse_args(&args(&["-###", "-c", "a.c"])).unwrap();
2502 let Action::PrintPlan { plan, .. } = a else { panic!("expected a plan dump") };
2503 assert!(plan.render().contains("a.c: preprocess, compile, assemble -> a.o"));
2504 }
2505
2506 #[test]
2507 fn the_flag_that_keeps_the_intermediate_files_has_three_spellings_and_two_meanings() {
2508 // The bare one is `=obj` and not `=cwd`. gcc's manual says the opposite and gcc 16 does
2509 // this, and following the compiler is what makes a build that reads either of them find
2510 // the files where they are.
2511 assert_eq!(compile(&["-c", "-save-temps", "a.c"]).0.save_temps, SaveTemps::Object);
2512 assert_eq!(compile(&["-c", "-save-temps=obj", "a.c"]).0.save_temps, SaveTemps::Object);
2513 assert_eq!(compile(&["-c", "-save-temps=cwd", "a.c"]).0.save_temps, SaveTemps::Cwd);
2514 assert_eq!(compile(&["-c", "a.c"]).0.save_temps, SaveTemps::No);
2515 // The last one on the line decides, the way it does for every other flag with an
2516 // argument, and a keyword that is neither is fatal rather than ignored: a run that kept
2517 // nothing and said nothing looks exactly like one where the files were not produced.
2518 let (opts, _) = compile(&["-c", "-save-temps", "-save-temps=cwd", "a.c"]);
2519 assert_eq!(opts.save_temps, SaveTemps::Cwd);
2520 let e = parse_args(&args(&["-c", "-save-temps=nowhere", "a.c"])).unwrap_err();
2521 assert!(e.message.contains("accepted: cwd, obj"), "{}", e.message);
2522 }
2523
2524 #[test]
2525 fn the_flag_that_times_each_step_reaches_the_options_and_changes_nothing_else() {
2526 let (opts, plan) = compile(&["-c", "-time", "a.c"]);
2527 let (plain, without) = compile(&["-c", "a.c"]);
2528 assert!(opts.time);
2529 assert!(!plain.time);
2530 // Against the same line without the flag rather than against a spelling of the object's
2531 // name, since what the object is called is the host's business and this is not about that.
2532 assert_eq!(plan.jobs[0].output, without.jobs[0].output);
2533 }
2534
2535 #[test]
2536 fn dash_x_names_what_it_accepts_when_it_does_not_know_a_language() {
2537 let e = parse_args(&args(&["-x", "fortran", "a.c"])).unwrap_err();
2538 assert!(e.message.contains("assembler-with-cpp"), "{}", e.message);
2539 }
2540
2541 #[test]
2542 fn an_unknown_flag_is_an_error_rather_than_a_shrug() {
2543 let e = parse_args(&args(&["-fno-such-thing", "a.c"])).unwrap_err();
2544 assert!(e.message.contains("unknown option"), "{}", e.message);
2545 }
2546
2547 /// `-fpermissive` and the flag that turns it back off, which a build writes beside it when
2548 /// one directory needs the older rules and the rest of the tree does not.
2549 #[test]
2550 fn permissive_reads_in_both_directions_and_the_last_one_wins() {
2551 let (opts, _) = compile(&["-c", "a.c"]);
2552 assert!(!opts.permissive, "off unless it is asked for");
2553
2554 let (opts, _) = compile(&["-c", "-fpermissive", "a.c"]);
2555 assert!(opts.permissive);
2556
2557 let (opts, _) = compile(&["-c", "-fpermissive", "-fno-permissive", "a.c"]);
2558 assert!(!opts.permissive);
2559 }
2560
2561 #[test]
2562 fn asking_for_nested_functions_is_told_why_it_is_not_coming() {
2563 let e = parse_args(&args(&["-fnested-functions", "a.c"])).unwrap_err();
2564 assert!(e.message.contains("trampoline"), "{}", e.message);
2565 assert!(parse_args(&args(&["-fno-nested-functions", "a.c"])).is_ok());
2566 }
2567
2568 #[test]
2569 fn the_flag_every_configure_script_writes_is_taken() {
2570 // All four spellings, because a build writes whichever one its macros picked and a
2571 // compiler that takes three of them is a compiler that fails on the fourth.
2572 for flag in ["-fPIC", "-fpic", "-fPIE", "-fpie"] {
2573 let (opts, _) = compile(&["-c", flag, "a.c"]);
2574 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
2575 }
2576 }
2577
2578 #[test]
2579 fn a_table_is_written_unless_the_build_says_nothing_will_walk_it() {
2580 let (opts, _) = compile(&["-c", "a.c"]);
2581 assert!(opts.unwinds(), "the default is off");
2582 let (opts, _) = compile(&["-c", "-fno-asynchronous-unwind-tables", "a.c"]);
2583 assert!(!opts.unwinds(), "the build was not taken at its word");
2584 let (opts, _) = compile(&[
2585 "-c",
2586 "-fno-asynchronous-unwind-tables",
2587 "-fasynchronous-unwind-tables",
2588 "a.c",
2589 ]);
2590 assert!(opts.unwinds(), "the last flag did not win");
2591 // The weaker request, which the same table answers, so a line that asks for a table and
2592 // against an asynchronous one gets one. That is gcc's arrangement and it turns up when a
2593 // build turns the asynchronous one off globally and a directory asks for a table back.
2594 let (opts, _) =
2595 compile(&["-c", "-fno-asynchronous-unwind-tables", "-funwind-tables", "a.c"]);
2596 assert!(opts.unwinds(), "the weaker request was dropped");
2597 let (opts, _) = compile(&["-c", "-fno-unwind-tables", "a.c"]);
2598 assert!(opts.unwinds(), "the weaker negative turned off the stronger request");
2599 let (opts, _) =
2600 compile(&["-c", "-fno-unwind-tables", "-fno-asynchronous-unwind-tables", "a.c"]);
2601 assert!(!opts.unwinds(), "both were turned off and one stayed on");
2602 }
2603
2604 #[test]
2605 fn the_flags_that_describe_what_this_compiler_already_does_are_taken() {
2606 // Every one of these is on a real build line somewhere and every one of them was an
2607 // unknown option. What they have in common is that the answer rucc gives is the answer
2608 // they ask for, so there is nothing to implement and nothing to refuse.
2609 for flag in [
2610 "-fno-common",
2611 "-fstrict-aliasing",
2612 "-fno-strict-aliasing",
2613 "-fdelete-null-pointer-checks",
2614 "-fno-delete-null-pointer-checks",
2615 "-frounding-math",
2616 "-fno-rounding-math",
2617 "-ftrapping-math",
2618 "-fno-trapping-math",
2619 "-fexcess-precision=standard",
2620 "-fexcess-precision=fast",
2621 "-fexcess-precision=16",
2622 "-pipe",
2623 "-fdiagnostics-color",
2624 "-fno-diagnostics-color",
2625 "-fdiagnostics-color=always",
2626 "-fdiagnostics-color=never",
2627 "-fdiagnostics-color=auto",
2628 ] {
2629 let (opts, _) = compile(&["-c", flag, "a.c"]);
2630 assert_eq!(opts.emit, EmitKind::Object, "{flag}");
2631 }
2632 }
2633
2634 #[test]
2635 fn asking_the_linker_to_merge_tentative_definitions_is_told_why_it_is_not_coming() {
2636 // The one of that family that is a request rather than a description, and it is a real
2637 // difference: two files each writing `int g;` link under it and do not without it.
2638 let e = parse_args(&args(&["-fcommon", "a.c"])).unwrap_err();
2639 assert!(e.message.contains(".bss"), "{}", e.message);
2640 assert!(e.message.contains("extern"), "the way out is worth saying: {}", e.message);
2641 }
2642
2643 #[test]
2644 fn asking_for_position_dependent_code_is_told_why_it_is_not_coming() {
2645 for flag in ["-fno-pic", "-fno-pie"] {
2646 let e = parse_args(&args(&[flag, "a.c"])).unwrap_err();
2647 assert!(e.message.contains("global offset table"), "{flag}: {}", e.message);
2648 // The one it may have meant, since the two are a letter apart and one of them is
2649 // about linking and is taken.
2650 assert!(e.message.contains("-no-pie"), "{flag}: {}", e.message);
2651 }
2652 }
2653
2654 #[test]
2655 fn an_unsupported_target_names_itself() {
2656 let e = parse_args(&args(&["--target=sparc64-linux-gnu", "a.c"])).unwrap_err();
2657 assert!(e.message.contains("sparc64"), "{}", e.message);
2658 }
2659
2660 #[test]
2661 fn no_inputs_is_an_error_but_print_config_needs_none() {
2662 assert!(parse_args(&args(&[])).is_err());
2663 assert!(matches!(parse_args(&args(&["--print-config"])), Ok(Action::PrintConfig(_))));
2664 }
2665
2666 #[test]
2667 fn print_config_reports_the_target_it_was_given_not_the_host() {
2668 let a = parse_args(&args(&["--print-config", "--target=riscv64-linux-musl"])).unwrap();
2669 let Action::PrintConfig(opts) = a else { panic!("expected a configuration dump") };
2670 let text = print_config(&opts);
2671 assert!(text.contains("target: riscv64-unknown-linux-musl"), "{text}");
2672 assert!(text.contains("char-signed: false"), "{text}");
2673 assert!(text.contains("object-format: elf"), "{text}");
2674 assert!(text.contains("va-list: void-pointer"), "{text}");
2675 // RISC-V has a register file and this compiler has not written it down yet, and the
2676 // dump says which of those two it is rather than leaving the line out.
2677 assert!(text.contains("registers: none"), "{text}");
2678 }
2679
2680 #[test]
2681 fn print_config_has_one_key_per_line_and_a_fixed_order() {
2682 let opts = Options::new("x86_64-unknown-linux-gnu".parse().unwrap());
2683 let text = print_config(&opts);
2684 let keys: Vec<&str> =
2685 text.lines().map(|l| l.split(':').next().unwrap_or_default()).collect();
2686 assert_eq!(keys[0], "version");
2687 assert_eq!(keys[1], "target");
2688 assert_eq!(keys.len(), 25);
2689 assert!(text.ends_with('\n'));
2690 }
2691
2692 #[test]
2693 fn the_safety_tier_is_read_off_the_command_line_and_a_wrong_one_is_refused() {
2694 let (opts, _) = compile(&["a.c"]);
2695 assert_eq!(opts.safety, rucc_session::Safety::Off);
2696
2697 for (flag, tier) in [
2698 ("-fsafety=detect", rucc_session::Safety::Detect),
2699 ("-fsafety=enforce", rucc_session::Safety::Enforce),
2700 ("-fsafety=kernel", rucc_session::Safety::Kernel),
2701 ("-fsafety=off", rucc_session::Safety::Off),
2702 ] {
2703 let (opts, _) = compile(&[flag, "a.c"]);
2704 assert_eq!(opts.safety, tier, "{flag}");
2705 }
2706
2707 // The last one wins, the way every other repeated flag on this command line does.
2708 let (opts, _) = compile(&["-fsafety=enforce", "-fsafety=off", "a.c"]);
2709 assert_eq!(opts.safety, rucc_session::Safety::Off);
2710
2711 // A misspelled tier is refused rather than ignored. Silently compiling without the
2712 // monitor a build asked for is the one failure mode this feature cannot have.
2713 let e = parse_args(&args(&["-fsafety=on", "a.c"])).unwrap_err();
2714 assert!(e.message.contains("is not a safety tier"), "{}", e.message);
2715 assert!(parse_args(&args(&["-fsafety", "a.c"])).is_err());
2716 }
2717
2718 #[test]
2719 fn the_padding_mode_is_read_off_the_command_line_and_a_wrong_one_is_refused() {
2720 // The default is the one section 9.3 of document 09 gives library code, which is that
2721 // padding does not participate, so a record filled a member at a time is not reported.
2722 let (opts, _) = compile(&["a.c"]);
2723 assert_eq!(opts.padding, rucc_session::Padding::Ignored);
2724
2725 let (opts, _) = compile(&["-fsafety=detect", "-fsafety-init=padding", "a.c"]);
2726 assert_eq!(opts.padding, rucc_session::Padding::Tracked);
2727
2728 let (opts, _) = compile(&["-fsafety-init=padding", "-fsafety-init=nopadding", "a.c"]);
2729 assert_eq!(opts.padding, rucc_session::Padding::Ignored);
2730
2731 // The tier is still a tier. A flag whose name starts the same way must not be eaten by
2732 // the one above it, which is the thing worth pinning about a pair of names like these.
2733 let (opts, _) = compile(&["-fsafety-init=padding", "a.c"]);
2734 assert_eq!(opts.safety, rucc_session::Safety::Off);
2735
2736 let e = parse_args(&args(&["-fsafety-init=some", "a.c"])).unwrap_err();
2737 assert!(e.message.contains("is not a padding mode"), "{}", e.message);
2738 }
2739
2740 #[test]
2741 fn whether_a_write_has_to_stay_inside_its_member_is_read_off_the_command_line() {
2742 // Off by default, because a store to allocated storage sets its effective type and C 6.5
2743 // lets a program reuse a buffer as something else. Row S4 is a build opting out of that.
2744 let (opts, _) = compile(&["a.c"]);
2745 assert_eq!(opts.subobject, rucc_session::Subobject::Off);
2746
2747 let (opts, _) = compile(&["-fsafety=detect", "-fsafety-subobject", "a.c"]);
2748 assert_eq!(opts.subobject, rucc_session::Subobject::Members);
2749
2750 let (opts, _) = compile(&["-fsafety-subobject", "-fno-safety-subobject", "a.c"]);
2751 assert_eq!(opts.subobject, rucc_session::Subobject::Off);
2752
2753 // It takes no value. The form that would take one is the strict reading of section 9.4,
2754 // which is not written yet, so say so rather than accept a spelling that does nothing.
2755 let e = parse_args(&args(&["-fsafety-subobject=strict", "a.c"])).unwrap_err();
2756 assert!(e.message.contains("tamnd/rucc#967"), "{}", e.message);
2757 }
2758
2759 #[test]
2760 fn whether_two_restrict_pointers_may_meet_is_read_off_the_command_line() {
2761 // Off by default, because the record a block keeps is the union of what each pointer
2762 // reached, so two pointers striding through one array without landing on the same byte are
2763 // reported and by the letter of the standard those are different objects. Row Y8 is a build
2764 // deciding it would rather know.
2765 let (opts, _) = compile(&["a.c"]);
2766 assert_eq!(opts.promise, rucc_session::Promise::Off);
2767
2768 let (opts, _) = compile(&["-fsafety=detect", "-fsafety-restrict", "a.c"]);
2769 assert_eq!(opts.promise, rucc_session::Promise::Blocks);
2770
2771 let (opts, _) = compile(&["-fsafety-restrict", "-fno-safety-restrict", "a.c"]);
2772 assert_eq!(opts.promise, rucc_session::Promise::Off);
2773
2774 // The tier is still a tier, which is the thing worth pinning about a pair of names where
2775 // one is the front of the other.
2776 let (opts, _) = compile(&["-fsafety-restrict", "a.c"]);
2777 assert_eq!(opts.safety, rucc_session::Safety::Off);
2778
2779 let e = parse_args(&args(&["-fsafety-restrict=blocks", "a.c"])).unwrap_err();
2780 assert!(e.message.contains("takes no value"), "{}", e.message);
2781 }
2782
2783 #[test]
2784 fn print_pipeline_answers_with_the_passes_the_level_asked_for() {
2785 let a = parse_args(&args(&["--print-pipeline", "-O2"])).unwrap();
2786 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
2787 let text = print_pipeline(&opts);
2788 assert!(text.starts_with("level: -O2\n"), "{text}");
2789 assert!(text.contains("fold"), "{text}");
2790
2791 let a = parse_args(&args(&["--print-pipeline"])).unwrap();
2792 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
2793 // One pass runs at `-O0` and it is the one that removes code nothing reaches, which is
2794 // not an optimization. See issue 359.
2795 assert!(print_pipeline(&opts).contains("1: simplify-cfg,"), "{}", print_pipeline(&opts));
2796
2797 let a = parse_args(&args(&["--print-pipeline", "-fno-simplify-cfg"])).unwrap();
2798 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
2799 // And with that one turned off there is nothing left, which the dump says rather than
2800 // printing an empty list.
2801 assert!(print_pipeline(&opts).contains("no passes"), "{}", print_pipeline(&opts));
2802 }
2803
2804 #[test]
2805 fn print_pipeline_takes_the_toggles_into_account() {
2806 let a = parse_args(&args(&["--print-pipeline", "-O2", "-fno-fold"])).unwrap();
2807 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
2808 let text = print_pipeline(&opts);
2809 // The one that was named is gone and the rest of the level is not, which is the whole
2810 // of what a toggle promises.
2811 assert!(!text.contains("fold"), "{text}");
2812 assert!(text.contains("dce"), "{text}");
2813
2814 // Every pass the compiler has, named off. Built from the registry rather than written
2815 // out, so a pass added later is turned off here too and this keeps testing the thing it
2816 // is about, which is that the toggles can empty a level.
2817 let mut off = vec!["--print-pipeline".to_owned(), "-O2".to_owned()];
2818 off.extend(rucc_opt::PASSES.iter().map(|p| format!("-fno-{}", p.name())));
2819 let spelled: Vec<&str> = off.iter().map(String::as_str).collect();
2820 let a = parse_args(&args(&spelled)).unwrap();
2821 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
2822 assert!(print_pipeline(&opts).contains("no passes"), "{}", print_pipeline(&opts));
2823 }
2824
2825 #[test]
2826 fn print_pipeline_says_when_a_budget_will_stop_the_run_short() {
2827 let a = parse_args(&args(&["--print-pipeline", "-O2"])).unwrap();
2828 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
2829 assert!(!print_pipeline(&opts).contains("global fuel"));
2830
2831 let a = parse_args(&args(&["--print-pipeline", "-O2", "-fpass-fuel-global=4"])).unwrap();
2832 let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
2833 let text = print_pipeline(&opts);
2834 // Because the listing is the answer to what this compilation will do, and a run that
2835 // stops after four rewrites is not doing what the level says it does.
2836 assert!(text.contains("global fuel: 4"), "{text}");
2837 }
2838
2839 /// A pass is turned on and off by its own name, and the order the flags were given in is
2840 /// kept, because the last spelling of a name is the one that decides.
2841 #[test]
2842 fn a_pass_is_named_by_dash_f_and_unnamed_by_dash_f_no() {
2843 let (opts, _) = compile(&["-c", "-O0", "-ffold", "-fno-fold", "-ffold", "a.c"]);
2844 assert_eq!(
2845 opts.passes,
2846 [("fold".to_owned(), true), ("fold".to_owned(), false), ("fold".to_owned(), true)]
2847 );
2848
2849 let e = parse_args(&args(&["-fno-such-pass", "a.c"])).unwrap_err();
2850 assert!(e.message.contains("unknown option"), "{}", e.message);
2851 }
2852
2853 #[test]
2854 fn pass_fuel_names_a_pass_and_a_count_and_refuses_anything_else() {
2855 let (opts, _) = compile(&["-c", "-O2", "-fpass-fuel=fold=3", "a.c"]);
2856 assert_eq!(opts.pass_fuel, [("fold".to_owned(), 3)]);
2857
2858 let e = parse_args(&args(&["-fpass-fuel=fold", "a.c"])).unwrap_err();
2859 assert!(e.message.contains("<pass>=<count>"), "{}", e.message);
2860 let e = parse_args(&args(&["-fpass-fuel=nosuch=3", "a.c"])).unwrap_err();
2861 assert!(e.message.contains("--print-pipeline"), "{}", e.message);
2862 let e = parse_args(&args(&["-fpass-fuel=fold=lots", "a.c"])).unwrap_err();
2863 assert!(e.message.contains("not a number"), "{}", e.message);
2864 }
2865
2866 #[test]
2867 fn global_pass_fuel_is_a_count_on_its_own_and_defaults_to_no_limit() {
2868 let (opts, _) = compile(&["-c", "-O2", "a.c"]);
2869 assert_eq!(opts.pass_fuel_global, None);
2870
2871 let (opts, _) = compile(&["-c", "-O2", "-fpass-fuel-global=12", "a.c"]);
2872 assert_eq!(opts.pass_fuel_global, Some(12));
2873 // And it is not the per pass flag with a longer name, so neither spelling swallows the
2874 // other.
2875 assert!(opts.pass_fuel.is_empty());
2876
2877 let e = parse_args(&args(&["-fpass-fuel-global=lots", "a.c"])).unwrap_err();
2878 assert!(e.message.contains("not a number"), "{}", e.message);
2879 }
2880
2881 #[test]
2882 fn a_gate_names_a_pass_and_optionally_the_functions_it_covers() {
2883 let (opts, _) = compile(&["-c", "-O2", "-fdisable-fold", "-fenable-fold=2-4,main", "a.c"]);
2884 assert_eq!(
2885 opts.pass_gates,
2886 [(false, "fold".to_owned()), (true, "fold=2-4,main".to_owned())],
2887 "the order is what decides, so it has to survive the parse"
2888 );
2889
2890 let e = parse_args(&args(&["-fdisable-nosuch", "a.c"])).unwrap_err();
2891 assert!(e.message.contains("--print-pipeline"), "{}", e.message);
2892 let e = parse_args(&args(&["-fenable-fold=9-2", "a.c"])).unwrap_err();
2893 assert!(e.message.contains("ends before it starts"), "{}", e.message);
2894 let e = parse_args(&args(&["-fdisable-fold=", "a.c"])).unwrap_err();
2895 assert!(e.message.contains("is empty"), "{}", e.message);
2896 }
2897
2898 #[test]
2899 fn the_pipeline_listing_says_which_passes_a_gate_touched() {
2900 let (opts, _) = compile(&["-c", "-O2", "-fdisable-fold=main", "a.c"]);
2901 let text = print_pipeline(&opts);
2902 assert!(text.contains("fold, "), "{text}");
2903 assert!(text.contains("[off for main]"), "{text}");
2904 }
2905
2906 /// The spelling is checked while the arguments are read, because a dump that names a pass
2907 /// this compiler does not have is a typo, and a typo found after the compilation has run is
2908 /// found too late to be any use.
2909 #[test]
2910 fn a_dump_is_checked_when_it_is_asked_for_rather_than_when_it_is_taken() {
2911 let (opts, _) = compile(&["-c", "-O2", "-fdump-ir=all", "-fdump-ir=after-fold", "a.c"]);
2912 assert_eq!(opts.dump_ir, ["all", "after-fold"]);
2913
2914 let e = parse_args(&args(&["-fdump-ir=after-nosuch", "a.c"])).unwrap_err();
2915 assert!(e.message.contains("nosuch"), "{}", e.message);
2916 assert!(parse_args(&args(&["-fdump-ir=sideways-fold", "a.c"])).is_err());
2917 }
2918
2919 /// Every spelling `-fopt-info` takes, and the one it does not.
2920 ///
2921 /// The keywords are checked here for the same reason a dump's pass name is: a person who
2922 /// misspelled one gets no output, and no output is also what a compilation where nothing
2923 /// happened looks like. Telling those two apart is the entire reason to reach for this flag.
2924 #[test]
2925 fn opt_info_takes_kinds_and_a_file_and_refuses_a_kind_it_does_not_have() {
2926 let (opts, _) = compile(&["-c", "-O2", "-fopt-info", "a.c"]);
2927 assert_eq!(opts.opt_info, [""], "a bare flag asks for the rewrites");
2928 assert_eq!(opts.opt_info_file, None, "and goes to standard error");
2929
2930 let (opts, _) = compile(&["-c", "-O2", "-fopt-info-missed-note", "a.c"]);
2931 assert_eq!(opts.opt_info, ["missed-note"]);
2932
2933 // Two flags add up rather than the second replacing the first, and the file is the last
2934 // one that named a file, which is how GCC treats both.
2935 let (opts, _) =
2936 compile(&["-c", "-O2", "-fopt-info-missed=one.txt", "-fopt-info-all=two.txt", "a.c"]);
2937 assert_eq!(opts.opt_info, ["missed", "all"]);
2938 assert_eq!(opts.opt_info_file.as_deref(), Some("two.txt"));
2939
2940 let e = parse_args(&args(&["-fopt-info-vectorized", "a.c"])).unwrap_err();
2941 assert!(e.message.contains("vectorized"), "{}", e.message);
2942 assert!(e.message.contains("`missed`"), "{}", e.message);
2943 let e = parse_args(&args(&["-fopt-info-missed=", "a.c"])).unwrap_err();
2944 assert!(e.message.contains("no file"), "{}", e.message);
2945 }
2946
2947 #[test]
2948 fn verify_each_is_unstable_and_off_unless_it_was_asked_for() {
2949 let (opts, _) = compile(&["-c", "-Zverify-each", "a.c"]);
2950 assert!(opts.verify_each);
2951 assert!(!USAGE.contains("verify-each"), "an unstable option stays out of the usage text");
2952 }
2953
2954 #[test]
2955 fn dash_o_needs_an_argument() {
2956 let e = parse_args(&args(&["a.c", "-o"])).unwrap_err();
2957 assert_eq!(e.message, "-o requires an argument");
2958 }
2959
2960 #[test]
2961 fn dash_d_and_dash_u_are_read_joined_or_separated_and_keep_their_order() {
2962 let (opts, _) = compile(&["-DFOO=1", "-D", "BAR", "-UBAZ", "-U", "QUX", "a.c"]);
2963 assert_eq!(opts.defines, ["FOO=1", "BAR"]);
2964 assert_eq!(opts.undefines, ["BAZ", "QUX"]);
2965 }
2966
2967 #[test]
2968 fn the_include_flags_land_on_the_chain_each_one_names() {
2969 // A sysroot with nothing under it, so that the library's own directories are the
2970 // same on every machine this test runs on, which is none of them.
2971 let (opts, _) = compile(&[
2972 "-Ii",
2973 "-iquote",
2974 "q",
2975 "-isystem",
2976 "sys",
2977 "-idirafter",
2978 "after",
2979 "--sysroot=/nowhere-at-all",
2980 "a.c",
2981 ]);
2982 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
2983 // The compiler's own headers sit after every `-isystem` and before `-idirafter`,
2984 // which is where GCC puts its own: a directory the user named outranks ours.
2985 assert_eq!(dirs, ["q", "i", "sys", runtime::DIR, "after"]);
2986 assert!(!opts.search.dirs()[1].is_system);
2987 assert!(opts.search.dirs()[2].is_system);
2988 }
2989
2990 #[test]
2991 fn the_librarys_headers_come_after_the_compilers_own_and_go_away_with_them() {
2992 // Which machine this runs on decides what is on the path, so the test is about the
2993 // order rather than about the names: ours is on it, the library's follow it, and
2994 // `-nostdinc` is the one flag that takes both halves of the pair off at once.
2995 let (opts, _) = compile(&["a.c"]);
2996 let dirs = opts.search.dirs();
2997 let ours = dirs.iter().position(|d| d.path.to_str() == Some(runtime::DIR));
2998 assert_eq!(ours, Some(0), "{dirs:?}");
2999 assert!(dirs[1..].iter().all(|d| d.is_system), "{dirs:?}");
3000 let (bare, _) = compile(&["-nostdinc", "a.c"]);
3001 assert!(bare.search.dirs().is_empty(), "{:?}", bare.search.dirs());
3002 }
3003
3004 #[test]
3005 fn a_sysroot_moves_the_librarys_directories_and_nothing_else() {
3006 let (opts, _) = compile(&["-isystem", "sys", "--sysroot=/nowhere-at-all", "a.c"]);
3007 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
3008 assert_eq!(dirs, ["sys", runtime::DIR]);
3009 }
3010
3011 #[test]
3012 fn a_cross_compile_reads_the_targets_own_headers_rather_than_the_ones_next_door() {
3013 // The target is not the machine this test runs on wherever it runs, so the answer is the
3014 // same on all of them: the libc's two include directories for that target, the kernel's
3015 // two, and nothing from here. A header read from here is the quiet failure of section 8.5, a
3016 // program that builds on the build machine and is wrong everywhere else.
3017 let (opts, _) = compile(&["--target=riscv64-linux-musl", "-c", "a.c"]);
3018 let dirs: Vec<&std::path::Path> =
3019 opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
3020 let root = cache::dir().join("sysroots").join("riscv64-linux-musl");
3021 let kernel = cache::dir().join("kernel-headers");
3022 assert_eq!(dirs.len(), 5, "{dirs:?}");
3023 assert_eq!(dirs[0], std::path::Path::new(runtime::DIR));
3024 assert_eq!(dirs[1], root.join("include").join("riscv64"));
3025 assert_eq!(dirs[2], root.join("include").join("generic"));
3026 // The kernel's, which are beside the sysroots rather than inside one, because every target
3027 // that shares an architecture reads the same files.
3028 assert_eq!(dirs[3], kernel.join("riscv"));
3029 assert_eq!(dirs[4], kernel.join("generic"));
3030 }
3031
3032 #[test]
3033 fn a_cross_compile_to_something_that_is_not_linux_reads_no_kernel_headers() {
3034 // The other side of the same answer. Windows has its own system headers and no `linux/` at
3035 // all, so the list is the libc's two and the question never arises, which is the `None` that
3036 // `link::cross_kernel` returns rather than a directory nothing would be found in.
3037 let (opts, _) = compile(&["--target=x86_64-pc-windows-gnu", "-c", "a.c"]);
3038 let dirs: Vec<&std::path::Path> =
3039 opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
3040 assert_eq!(dirs.len(), 3, "{dirs:?}");
3041 assert!(!dirs.iter().any(|dir| dir.ends_with("kernel-headers")), "{dirs:?}");
3042 }
3043
3044 #[test]
3045 fn the_glibc_version_macro_goes_with_the_bundled_tree_and_with_nothing_else() {
3046 // One tree serves every glibc release, so the release is what the target supplies, and the
3047 // condition is the same one that chose the directories. A host glibc and a tree somebody
3048 // named both define `__GLIBC_MINOR__` in their own `features.h`, and two definitions with
3049 // different values is a warning on every compilation of every file.
3050 //
3051 // The architecture is chosen against this machine's rather than written down, because the
3052 // bundled tree is only in effect for a target that is not this machine. The first version of
3053 // this test said x86_64-linux-gnu, which is a cross compile on a mac and this machine on a
3054 // Linux runner, so it passed here and failed there.
3055 let gnu = format!("--target={}-linux-gnu", cross_arch());
3056 let (bundled, _) = compile(&[&gnu, "-c", "a.c"]);
3057 assert_eq!(bundled.glibc_minor, Some(44));
3058 let pin = format!("{gnu}.2.28");
3059 let (pinned, _) = compile(&[&pin, "-c", "a.c"]);
3060 assert_eq!(pinned.glibc_minor, Some(28));
3061
3062 let (named, _) = compile(&[&gnu, "--sysroot=/nowhere-at-all", "-c", "a.c"]);
3063 assert_eq!(named.glibc_minor, None);
3064 let (none, _) = compile(&[&gnu, "-nostdinc", "-c", "a.c"]);
3065 assert_eq!(none.glibc_minor, None);
3066 let musl = format!("--target={}-linux-musl", cross_arch());
3067 let (musl, _) = compile(&[&musl, "-c", "a.c"]);
3068 assert_eq!(musl.glibc_minor, None);
3069
3070 // And this machine's own target gets nothing, whatever this machine is, because its headers
3071 // come from the machine and its own `features.h` defines the macro. On a glibc Linux box
3072 // that is the case this test had backwards; on a mac it is true for the other reason, which
3073 // is that Darwin is not a glibc target at all.
3074 if let Some(host) = Triple::host() {
3075 let native = format!("--target={}", host.tuple());
3076 let (native, _) = compile(&[&native, "-c", "a.c"]);
3077 assert_eq!(native.glibc_minor, None);
3078 }
3079 }
3080
3081 #[test]
3082 fn a_pinned_release_on_this_machines_own_target_reads_the_bundled_tree() {
3083 // The end to end half of the answer in `link::cross_for`. A release named for this machine's
3084 // own target is a cross compile, so the headers are the bundled tree's and the macro says
3085 // what was asked for rather than what this machine has.
3086 //
3087 // Only on a glibc box, because a release is a glibc release: a mac has no `__GLIBC_MINOR__`
3088 // to get wrong and nothing to pin. That makes this a test the Linux runners carry, which is
3089 // where the case lives.
3090 let Some(host) = Triple::host() else { return };
3091 if host.env != rucc_target::Env::Gnu {
3092 return;
3093 }
3094 let pin = format!("--target={}.2.28", host.tuple());
3095 let (opts, _) = compile(&[&pin, "-c", "a.c"]);
3096 assert_eq!(opts.glibc_minor, Some(28));
3097 let root = cache::dir().join("sysroots").join(format!("{}.2.28", host.tuple()));
3098 let dirs: Vec<&std::path::Path> =
3099 opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
3100 assert!(dirs.iter().any(|dir| dir.starts_with(&root)), "{dirs:?}");
3101 // And nothing of this machine's, which is the failure this was: a program compiled against
3102 // 2.44 declarations and told it was 2.28.
3103 assert!(!dirs.iter().any(|dir| *dir == std::path::Path::new("/usr/include")), "{dirs:?}");
3104 }
3105
3106 /// An architecture that is not this machine's, out of the three the driver has targets for.
3107 ///
3108 /// A test about the bundled sysroot has to name a target that is not the host, because a target
3109 /// that is the host reads the host's own headers and libraries. Asking which machine this is
3110 /// beats picking a row and hoping, and it is two lines.
3111 fn cross_arch() -> &'static str {
3112 match Triple::host().map(|host| host.arch) {
3113 Some(rucc_target::Arch::X86_64) => "aarch64",
3114 _ => "x86_64",
3115 }
3116 }
3117
3118 #[test]
3119 fn a_glibc_newer_than_the_bundled_tree_is_refused_by_name() {
3120 // Both versions in the message, because the two things a person can do about it are pin a
3121 // release the tree has and name a sysroot that has the one they asked for, and neither is a
3122 // choice they can make without knowing which release the tree is.
3123 //
3124 // Not this machine's architecture, for the reason the test above gives: the refusal is about
3125 // the bundled tree, and the bundled tree is not what a target that is this machine reads.
3126 let target = format!("--target={}-linux-gnu.2.99", cross_arch());
3127 let message = refused(&[&target, "-c", "a.c"]);
3128 assert!(message.contains("asked for glibc 2.99"), "{message}");
3129 assert!(message.contains("bundled headers are glibc 2.44"), "{message}");
3130 assert!(message.contains("--sysroot"), "{message}");
3131 }
3132
3133 #[test]
3134 fn a_sysroot_the_user_named_is_still_what_a_cross_compile_reads() {
3135 // The tree somebody assembled beats the one we would build, on the headers as on the
3136 // libraries. It is empty here, which is why the list comes out short: the directories under
3137 // it are checked for rather than assumed, and a tree that is not there offers nothing.
3138 let (opts, _) =
3139 compile(&["--target=riscv64-linux-musl", "--sysroot=/nowhere-at-all", "-c", "a.c"]);
3140 let dirs: Vec<&std::path::Path> =
3141 opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
3142 assert_eq!(dirs, [std::path::Path::new(runtime::DIR)]);
3143 }
3144
3145 #[test]
3146 fn dash_i_dash_moves_the_bracket_directories_into_the_quoted_chain() {
3147 let (opts, _) =
3148 compile(&["-Iinc1", "-iquote", "inc2", "-I-", "-Iinc3", "-nostdinc", "a.c"]);
3149 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
3150 assert_eq!(dirs, ["inc1", "inc2", "inc3"]);
3151 // An angled include sees only what came after the flag.
3152 assert_eq!(opts.search.start(IncludeForm::Angled), 2);
3153 assert!(!opts.search.searches_current_dir());
3154 }
3155
3156 #[test]
3157 fn the_prefix_flags_stick_what_iprefix_said_on_the_front_of_what_follows_it() {
3158 let (opts, _) = compile(&[
3159 "-iprefix",
3160 "/tools/",
3161 "-iwithprefix",
3162 "late",
3163 "-iwithprefixbefore",
3164 "early",
3165 "-iprefix",
3166 "/other/",
3167 "-iwithprefix",
3168 "last",
3169 "-nostdinc",
3170 "a.c",
3171 ]);
3172 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
3173 // `-iwithprefixbefore` is an `-I` and the other two are `-isystem`, which is where GCC
3174 // puts them rather than where its manual says it does.
3175 assert_eq!(dirs, ["/tools/early", "/tools/late", "/other/last"]);
3176 assert!(!opts.search.dirs()[0].is_system);
3177 assert!(opts.search.dirs()[1].is_system);
3178 }
3179
3180 #[test]
3181 fn the_files_named_on_the_command_line_keep_their_order_and_which_flag_named_them() {
3182 let (opts, _) =
3183 compile(&["-include", "one.h", "-imacros", "two.h", "-include", "3.h", "a.c"]);
3184 let names: Vec<&str> = opts.preincludes.iter().map(|p| p.name.as_str()).collect();
3185 assert_eq!(names, ["one.h", "two.h", "3.h"]);
3186 assert_eq!(opts.preincludes.iter().filter(|p| p.macros_only).count(), 1);
3187 }
3188
3189 #[test]
3190 fn nostdinc_takes_the_compilers_own_headers_off_the_path() {
3191 let (opts, _) = compile(&["-Ii", "-nostdinc", "a.c"]);
3192 let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
3193 assert_eq!(dirs, ["i"]);
3194 }
3195
3196 #[test]
3197 fn the_dialect_flags_set_the_language_and_the_extensions_separately() {
3198 let (opts, _) = compile(&["-std=gnu11", "a.c"]);
3199 assert_eq!(opts.std, Std::C11);
3200 assert!(opts.gnu_extensions);
3201
3202 let (opts, _) = compile(&["-std=iso9899:1999", "a.c"]);
3203 assert_eq!(opts.std, Std::C99);
3204 assert!(!opts.gnu_extensions);
3205
3206 let (opts, _) = compile(&["-ansi", "a.c"]);
3207 assert_eq!(opts.std, Std::C89);
3208 assert!(!opts.gnu_extensions);
3209
3210 let e = parse_args(&args(&["-std=c94jr", "a.c"])).unwrap_err();
3211 assert!(e.message.contains("unknown dialect"), "{}", e.message);
3212 }
3213
3214 #[test]
3215 fn the_dump_letters_are_a_family_and_everything_else_beginning_with_d_is_not() {
3216 let (opts, _) = compile(&["-dM", "a.c"]);
3217 assert!(opts.dumps.macros);
3218
3219 // Packed, the way GCC takes them, and a letter in the family we have not written yet
3220 // is accepted and does nothing rather than failing a build.
3221 let (opts, _) = compile(&["-dDM", "a.c"]);
3222 assert!(opts.dumps.macros);
3223 let (opts, _) = compile(&["-dD", "a.c"]);
3224 assert!(!opts.dumps.macros);
3225
3226 let (opts, _) = compile(&["a.c"]);
3227 assert!(!opts.dumps.any());
3228
3229 // `-dumpversion` is a different flag that happens to start the same way, and it is read
3230 // as itself rather than as a dump of nothing.
3231 assert_eq!(printed(&["-dumpversion", "a.c"]), VERSION);
3232 }
3233
3234 #[test]
3235 fn the_gcc_version_claimed_is_a_flag_and_the_short_spellings_are_the_ones_people_write() {
3236 let (opts, _) = compile(&["a.c"]);
3237 assert_eq!(
3238 opts.gnuc,
3239 GnucVersion { major: 7, minor: 0, patch: 0 },
3240 "the lowest claim a modern glibc gives its own declarations to"
3241 );
3242
3243 let (opts, _) = compile(&["-fgnuc-version=15.1.0", "a.c"]);
3244 assert_eq!(opts.gnuc, GnucVersion { major: 15, minor: 1, patch: 0 });
3245
3246 // A missing component is zero. `gcc -dumpversion` says `15` on a release with no
3247 // patchlevel and a harness that pastes that back has to be understood.
3248 let (opts, _) = compile(&["-fgnuc-version=15", "a.c"]);
3249 assert_eq!(opts.gnuc, GnucVersion { major: 15, minor: 0, patch: 0 });
3250
3251 let (opts, _) = compile(&["-fgnuc-version=13.2", "a.c"]);
3252 assert_eq!(opts.gnuc, GnucVersion { major: 13, minor: 2, patch: 0 });
3253
3254 let e = parse_args(&args(&["-fgnuc-version=15.x", "a.c"])).unwrap_err();
3255 assert!(e.message.contains("minor that is not a number"), "{}", e.message);
3256
3257 let e = parse_args(&args(&["-fgnuc-version=1.2.3.4", "a.c"])).unwrap_err();
3258 assert!(e.message.contains("more than three"), "{}", e.message);
3259 }
3260
3261 #[test]
3262 fn pedantic_has_two_spellings_and_is_not_the_same_knob_as_the_dialect() {
3263 let (opts, _) = compile(&["-std=c17", "-pedantic", "a.c"]);
3264 assert!(opts.pedantic);
3265 assert_eq!(opts.std, Std::C17);
3266
3267 // The `-W` family's name for it, which is what a build that groups its warning flags
3268 // tends to write.
3269 let (opts, _) = compile(&["-Wpedantic", "a.c"]);
3270 assert!(opts.pedantic);
3271
3272 let (opts, _) = compile(&["-std=c17", "a.c"]);
3273 assert!(!opts.pedantic, "a dialect on its own does not diagnose an extension");
3274 }
3275
3276 #[test]
3277 fn dash_p_and_dash_ffreestanding_reach_the_options() {
3278 let (opts, _) = compile(&["-E", "-P", "-ffreestanding", "a.c"]);
3279 assert!(!opts.line_markers);
3280 assert!(!opts.hosted);
3281 assert_eq!(opts.emit, EmitKind::Preprocessed);
3282 }
3283
3284 /// The two ways a build says it means its own function by a name the C library also has.
3285 ///
3286 /// `-fno-builtin` is all of them and `-fno-builtin-<name>` is one, and the second is what a
3287 /// build writes when it means its own `memcpy` and the library's everything else. The name is
3288 /// kept as it was written and not checked against anything, because a program is allowed to
3289 /// mean something by a name this compiler has never heard of.
3290 #[test]
3291 fn the_builtin_flags_are_read_in_both_directions_and_one_name_at_a_time() {
3292 let (opts, _) = compile(&["-c", "a.c"]);
3293 assert!(opts.builtins, "a library name means the library function by default");
3294 assert!(opts.no_builtin.is_empty());
3295
3296 let (opts, _) = compile(&["-c", "-fno-builtin", "a.c"]);
3297 assert!(!opts.builtins);
3298
3299 let (opts, _) = compile(&["-c", "-fno-builtin", "-fbuiltin", "a.c"]);
3300 assert!(opts.builtins, "the last mention decides");
3301
3302 let (opts, _) = compile(&["-c", "-fno-builtin-memcpy", "-fno-builtin-nonesuch", "a.c"]);
3303 assert!(opts.builtins, "one name is not the family");
3304 assert_eq!(opts.no_builtin, vec!["memcpy".to_owned(), "nonesuch".to_owned()]);
3305 }
3306
3307 /// `-fvisibility=`, which is on every cmake project that cares about which names it exports
3308 /// and which was refused as an unknown option until now.
3309 ///
3310 /// Four spellings and three answers. `internal` is hidden plus a promise about never taking
3311 /// the address across a component boundary, and nothing derives anything from that promise
3312 /// here, so it comes out as the weaker of the two rather than as a refusal that stops a build
3313 /// over a distinction this compiler does not make.
3314 #[test]
3315 fn visibility_takes_the_four_spellings_gcc_takes_and_refuses_the_rest() {
3316 let (opts, _) = compile(&["-c", "a.c"]);
3317 assert_eq!(opts.visibility, Visibility::Default, "exported unless something says not");
3318
3319 for (written, wanted) in [
3320 ("default", Visibility::Default),
3321 ("hidden", Visibility::Hidden),
3322 ("internal", Visibility::Hidden),
3323 ("protected", Visibility::Protected),
3324 ] {
3325 let (opts, _) = compile(&["-c", &format!("-fvisibility={written}"), "a.c"]);
3326 assert_eq!(opts.visibility, wanted, "{written}");
3327 }
3328
3329 // The last mention decides, which is what every other flag of this shape does and what a
3330 // build that turns something off for one directory relies on.
3331 let (opts, _) = compile(&["-c", "-fvisibility=hidden", "-fvisibility=default", "a.c"]);
3332 assert_eq!(opts.visibility, Visibility::Default, "the last mention decides");
3333
3334 // A spelling gcc does not take is refused rather than read as the default, because a
3335 // build that meant hidden and got exported is a library with the wrong interface and
3336 // nothing said about it anywhere.
3337 let failed = parse_args(&args(&["-fvisibility=none", "a.c"])).expect_err("refused");
3338 assert!(failed.to_string().contains("is not a visibility"), "{failed}");
3339 }
3340
3341 /// `-ffp-contract=`, which is the one flag in the floating point group that is kept rather than
3342 /// described, and the values are gcc 16's three.
3343 #[test]
3344 fn how_far_a_multiply_and_an_addition_may_be_fused_is_asked_for() {
3345 let (opts, _) = compile(&["-c", "a.c"]);
3346 assert_eq!(opts.fp_contract, Contract::Off, "a licence nobody granted is not assumed");
3347
3348 for (written, wanted) in
3349 [("off", Contract::Off), ("on", Contract::On), ("fast", Contract::Fast)]
3350 {
3351 let (opts, _) = compile(&["-c", &format!("-ffp-contract={written}"), "a.c"]);
3352 assert_eq!(opts.fp_contract, wanted, "{written}");
3353 }
3354
3355 let (opts, _) = compile(&["-c", "-ffp-contract=fast", "-ffp-contract=off", "a.c"]);
3356 assert_eq!(opts.fp_contract, Contract::Off, "the last mention decides");
3357
3358 // Refused rather than read as one of the three, because a build that asked for no fusing
3359 // and was given the default would be one whose numbers change and whose command line says
3360 // they should not. gcc refuses the same spellings and names the same three in its message.
3361 for bad in ["-ffp-contract=none", "-ffp-contract=", "-ffp-contract=Fast"] {
3362 let failed = parse_args(&args(&[bad, "a.c"])).expect_err("refused");
3363 assert!(failed.to_string().contains("is not a contraction"), "{bad}: {failed}");
3364 }
3365
3366 // And the other one that takes a value, which is taken and kept nowhere: every operation
3367 // here is computed in the type it was written in, so `standard` is what happens and the
3368 // other two are permission to do something this does not do.
3369 let failed = parse_args(&args(&["-fexcess-precision=long", "a.c"])).expect_err("refused");
3370 assert!(failed.to_string().contains("is not an excess precision"), "{failed}");
3371 }
3372
3373 /// The four prefix mapping flags, which are what a distribution passes to get the same bytes
3374 /// out of `/build/pkg-1.2` and out of `/home/someone/pkg-1.2`. Three lists rather than one
3375 /// because gcc has three, and `-ffile-prefix-map=` is the three of them at once.
3376 #[test]
3377 fn a_prefix_mapping_flag_goes_on_the_list_its_spelling_names() {
3378 let (opts, _) = compile(&["-c", "a.c"]);
3379 assert!(opts.prefix_map.macros.is_empty(), "nothing is rewritten unless it is asked for");
3380 assert!(opts.prefix_map.debug.is_empty(), "nor here");
3381 assert!(opts.prefix_map.profile.is_empty(), "nor here");
3382
3383 let (opts, _) = compile(&["-c", "-fmacro-prefix-map=/build=.", "a.c"]);
3384 assert_eq!(opts.prefix_map.macros.apply("/build/a.c"), "./a.c", "the one it names");
3385 assert!(opts.prefix_map.debug.is_empty(), "and not the two it does not");
3386
3387 let (opts, _) = compile(&["-c", "-fdebug-prefix-map=/build=.", "a.c"]);
3388 assert_eq!(opts.prefix_map.debug.apply("/build/a.c"), "./a.c", "the one it names");
3389 assert!(opts.prefix_map.macros.is_empty(), "and not the two it does not");
3390
3391 let (opts, _) = compile(&["-c", "-fprofile-prefix-map=/build=.", "a.c"]);
3392 assert_eq!(opts.prefix_map.profile.apply("/build/a.c"), "./a.c", "the one it names");
3393 assert!(opts.prefix_map.macros.is_empty(), "and not the two it does not");
3394
3395 let (opts, _) = compile(&["-c", "-ffile-prefix-map=/build=.", "a.c"]);
3396 for list in [&opts.prefix_map.macros, &opts.prefix_map.debug, &opts.prefix_map.profile] {
3397 assert_eq!(list.apply("/build/a.c"), "./a.c", "all three at once");
3398 }
3399
3400 // Every mention is kept and the last one that matches wins, unlike the flags above whose
3401 // last mention replaces the earlier ones. A build writes one of these per source root and
3402 // expects all of them to be in force, which is the whole point of a list.
3403 let (opts, _) =
3404 compile(&["-c", "-ffile-prefix-map=/a=one", "-ffile-prefix-map=/b=two", "a.c"]);
3405 assert_eq!(opts.prefix_map.macros.apply("/a/x.c"), "one/x.c", "the earlier one still acts");
3406 assert_eq!(opts.prefix_map.macros.apply("/b/x.c"), "two/x.c", "and so does the later one");
3407
3408 // An argument with no `=` is refused rather than ignored, because a build whose paths were
3409 // meant to be rewritten and were not is one that ships the build directory's name and says
3410 // nothing about it. gcc refuses the same thing.
3411 for bad in ["-fmacro-prefix-map=nope", "-ffile-prefix-map=", "-fdebug-prefix-map=/build"] {
3412 let failed = parse_args(&args(&[bad, "a.c"])).expect_err("refused");
3413 assert!(failed.to_string().contains("is not a rewrite for"), "{bad}: {failed}");
3414 }
3415 }
3416
3417 /// `-ffunction-sections` and `-fdata-sections`, which are what make `--gc-sections` able to
3418 /// drop anything: a linker can leave out a section nothing reaches and cannot leave out half of
3419 /// one. A kernel and an embedded image are both linked that way.
3420 ///
3421 /// Two flags rather than one because gcc has two, and a build that asks for one of them and not
3422 /// the other is a build that measured something: splitting the code is nearly free at link time
3423 /// and splitting the data can defeat the linker's ordering of what is next to what.
3424 #[test]
3425 fn a_section_per_function_and_a_section_per_variable_are_asked_for_one_at_a_time() {
3426 let (opts, _) = compile(&["-c", "a.c"]);
3427 assert!(!opts.function_sections, "one text section unless something says otherwise");
3428 assert!(!opts.data_sections);
3429
3430 let (opts, _) = compile(&["-c", "-ffunction-sections", "a.c"]);
3431 assert!(opts.function_sections);
3432 assert!(!opts.data_sections, "one flag is not the other");
3433
3434 let (opts, _) = compile(&["-c", "-fdata-sections", "a.c"]);
3435 assert!(opts.data_sections);
3436 assert!(!opts.function_sections);
3437
3438 // Both directions taken, and the off one is what happens anyway rather than a refusal,
3439 // since a build that writes it is asking for the default.
3440 let (opts, _) = compile(&[
3441 "-c",
3442 "-ffunction-sections",
3443 "-fno-function-sections",
3444 "-fdata-sections",
3445 "-fno-data-sections",
3446 "a.c",
3447 ]);
3448 assert!(!opts.function_sections, "the last mention decides");
3449 assert!(!opts.data_sections, "the last mention decides");
3450 }
3451
3452 /// `-fgnu89-inline`, which is off by default and is not implied by anything on the command
3453 /// line, since the dialect asks for GNU's reading further in rather than through this.
3454 #[test]
3455 fn gnu89_inline_is_off_until_it_is_asked_for_and_the_last_mention_decides() {
3456 let (opts, _) = compile(&["-c", "a.c"]);
3457 assert!(!opts.gnu89_inline, "C's reading of inline by default");
3458
3459 let (opts, _) = compile(&["-c", "-fgnu89-inline", "a.c"]);
3460 assert!(opts.gnu89_inline);
3461
3462 let (opts, _) = compile(&["-c", "-fgnu89-inline", "-fno-gnu89-inline", "a.c"]);
3463 assert!(!opts.gnu89_inline, "the last mention decides");
3464
3465 // The C89 dialects are under GNU's reading whether this was written or not, so the flag
3466 // stays off there and the dialect is what the checker and the macro set both ask. That is
3467 // also why `-std=c89 -fno-gnu89-inline` needs no diagnostic: it asks for the reading the
3468 // dialect already has. gcc refuses that command line, which is measured in the issue.
3469 let (opts, _) = compile(&["-c", "-std=c89", "a.c"]);
3470 assert!(!opts.gnu89_inline);
3471 }
3472
3473 /// Both spellings of both frame flags, since a build that wants one usually writes the
3474 /// other beside it for the one file that has to be compiled the ordinary way.
3475 #[test]
3476 fn the_two_frame_flags_are_read_in_both_directions() {
3477 let (opts, _) = compile(&["-c", "a.c"]);
3478 assert!(!opts.frame_pointer, "gcc omits it above -O0 and so does this");
3479 assert!(opts.red_zone, "the psABI has one and nothing said not to use it");
3480
3481 let (opts, _) = compile(&["-c", "-fno-omit-frame-pointer", "-mno-red-zone", "a.c"]);
3482 assert!(opts.frame_pointer);
3483 assert!(!opts.red_zone);
3484
3485 let (opts, _) = compile(&[
3486 "-c",
3487 "-fno-omit-frame-pointer",
3488 "-fomit-frame-pointer",
3489 "-mno-red-zone",
3490 "-mred-zone",
3491 "a.c",
3492 ]);
3493 assert!(!opts.frame_pointer, "the last one wins, as it does in gcc");
3494 assert!(opts.red_zone);
3495 }
3496
3497 /// Four flags rather than one with an argument, which is how gcc spells them, and the negative
3498 /// spelled three ways because a build that turns one off writes whichever it turned on.
3499 #[test]
3500 fn the_stack_protector_is_four_flags_and_the_last_one_wins() {
3501 let (opts, _) = compile(&["-c", "a.c"]);
3502 assert_eq!(opts.protector, Protector::None, "gcc protects nothing unless it was asked");
3503
3504 for (flag, want) in [
3505 ("-fstack-protector", Protector::Buffers),
3506 ("-fstack-protector-strong", Protector::Strong),
3507 ("-fstack-protector-all", Protector::All),
3508 ] {
3509 let (opts, _) = compile(&["-c", flag, "a.c"]);
3510 assert_eq!(opts.protector, want, "{flag}");
3511 }
3512
3513 // What a package build does: the strong one in the global flags and one directory that
3514 // cannot have a protector turning it off on the line after.
3515 for off in ["-fno-stack-protector", "-fno-stack-protector-strong"] {
3516 let (opts, _) = compile(&["-c", "-fstack-protector-strong", off, "a.c"]);
3517 assert_eq!(opts.protector, Protector::None, "{off}");
3518 }
3519 let (opts, _) = compile(&["-c", "-fno-stack-protector", "-fstack-protector-all", "a.c"]);
3520 assert_eq!(opts.protector, Protector::All, "the last one wins either way round");
3521 }
3522
3523 /// A switch rather than a level, because how a frame is taken is one question and which
3524 /// functions get a canary is another, and gcc spells it that way for the same reason.
3525 #[test]
3526 fn taking_a_frame_a_page_at_a_time_is_off_until_it_is_asked_for() {
3527 let (opts, _) = compile(&["-c", "a.c"]);
3528 assert!(!opts.stack_clash, "gcc takes a frame in one subtraction unless it was asked");
3529
3530 let (opts, _) = compile(&["-c", "-fstack-clash-protection", "a.c"]);
3531 assert!(opts.stack_clash);
3532
3533 // The same shape a package build uses for the protector: on in the global flags and off
3534 // for the one directory that cannot have it.
3535 let (opts, _) =
3536 compile(&["-c", "-fstack-clash-protection", "-fno-stack-clash-protection", "a.c"]);
3537 assert!(!opts.stack_clash);
3538 let (opts, _) =
3539 compile(&["-c", "-fno-stack-clash-protection", "-fstack-clash-protection", "a.c"]);
3540 assert!(opts.stack_clash, "the last one wins either way round");
3541
3542 // The two are independent, since one is about the frame and the other about the function.
3543 let (opts, _) =
3544 compile(&["-c", "-fstack-clash-protection", "-fstack-protector-strong", "a.c"]);
3545 assert!(opts.stack_clash);
3546 assert_eq!(opts.protector, Protector::Strong);
3547 }
3548
3549 /// One flag with an argument rather than a family of spellings, because what it asks about is
3550 /// which of the two edges of a control flow transfer is checked and the two are not separate
3551 /// questions to the hardware.
3552 #[test]
3553 fn which_control_flow_edges_are_checked_is_asked_for_by_name() {
3554 let (opts, _) = compile(&["-c", "a.c"]);
3555 assert_eq!(opts.control, Control::None, "gcc's default on the targets this compiler has");
3556
3557 for (arg, want) in [
3558 ("-fcf-protection", Control::Full),
3559 ("-fcf-protection=full", Control::Full),
3560 ("-fcf-protection=branch", Control::Branch),
3561 ("-fcf-protection=return", Control::Return),
3562 ("-fcf-protection=none", Control::None),
3563 ("-fcf-protection=check", Control::Check),
3564 ] {
3565 let (opts, _) = compile(&["-c", arg, "a.c"]);
3566 assert_eq!(opts.control, want, "{arg}");
3567 }
3568
3569 // The shape a package build uses: on in the global flags and off for the one directory
3570 // that cannot have it, whichever of the two spellings of off it reaches for.
3571 let (opts, _) = compile(&["-c", "-fcf-protection=full", "-fno-cf-protection", "a.c"]);
3572 assert_eq!(opts.control, Control::None);
3573 let (opts, _) = compile(&["-c", "-fno-cf-protection", "-fcf-protection=branch", "a.c"]);
3574 assert_eq!(opts.control, Control::Branch, "the last one wins either way round");
3575 }
3576
3577 /// The profiler is asked for by two spellings, and where its hook goes by two more.
3578 ///
3579 /// The two halves are separate on purpose. `-mfentry` on its own says where a call would go and
3580 /// asks for no call, which is what gcc does with it, and a build system that sets it globally
3581 /// and asks for the profile per directory needs that to be true rather than an error.
3582 ///
3583 /// The link is asserted alongside, because the flag changes it too and a build that compiled
3584 /// with it and linked without it is a program that calls the hook everywhere and never writes a
3585 /// profile.
3586 #[test]
3587 fn the_profiler_and_where_its_hook_goes_are_two_separate_questions() {
3588 let (opts, _) = compile(&["-c", "a.c"]);
3589 assert!(!opts.profile);
3590 assert_eq!(opts.hook, Hook::Platform, "neither was named, so the target decides");
3591
3592 for arg in ["-pg", "-p"] {
3593 let (opts, _) = compile(&["-c", arg, "a.c"]);
3594 assert!(opts.profile, "{arg}");
3595 let (link, _) = linking(&[arg, "a.c"]);
3596 assert!(link.profile, "{arg} changes the link as well");
3597 }
3598
3599 for (arg, want) in [("-mfentry", Hook::Early), ("-mno-fentry", Hook::Late)] {
3600 let (opts, _) = compile(&["-c", arg, "a.c"]);
3601 assert_eq!(opts.hook, want, "{arg}");
3602 assert!(!opts.profile, "{arg} asks for no call of its own");
3603 }
3604
3605 let (opts, _) = compile(&["-c", "-mfentry", "-mno-fentry", "-pg", "a.c"]);
3606 assert_eq!(opts.hook, Hook::Late, "the last one wins");
3607 assert!(opts.profile);
3608 }
3609
3610 /// How much room a patcher is promised, which is one number or two.
3611 ///
3612 /// A command line that did not ask is asserted alongside, because the flag has to be written to
3613 /// mean anything and a build that reserved room nobody asked for would grow every function in
3614 /// it for nothing.
3615 #[test]
3616 fn the_room_a_patcher_is_promised_is_a_number_of_bytes_and_where_they_go() {
3617 let (opts, _) = compile(&["-c", "a.c"]);
3618 assert_eq!(opts.patchable, Patchable::default());
3619 assert!(!opts.patchable.any(), "nothing is reserved unless it was asked for");
3620
3621 let (opts, _) = compile(&["-c", "-fpatchable-function-entry=16", "a.c"]);
3622 assert_eq!(opts.patchable, Patchable { total: 16, before: 0 });
3623
3624 let (opts, _) = compile(&["-c", "-fpatchable-function-entry=5,3", "a.c"]);
3625 assert_eq!(opts.patchable, Patchable { total: 5, before: 3 });
3626 assert_eq!(opts.patchable.after(), 2);
3627
3628 // The last one wins, which is what every other flag of this shape does and what a build
3629 // that adds one to a command line it did not write is relying on.
3630 let (opts, _) = compile(&[
3631 "-c",
3632 "-fpatchable-function-entry=5,3",
3633 "-fpatchable-function-entry=2",
3634 "a.c",
3635 ]);
3636 assert_eq!(opts.patchable, Patchable { total: 2, before: 0 });
3637 }
3638
3639 /// And a request nothing could satisfy is refused rather than rounded into one that can be.
3640 #[test]
3641 fn room_in_front_of_the_label_that_is_more_than_the_room_asked_for_is_refused() {
3642 for arg in ["-fpatchable-function-entry=1,2", "-fpatchable-function-entry=x"] {
3643 let e = parse_args(&args(&["-c", arg, "a.c"])).unwrap_err();
3644 assert!(e.message.contains("is not an amount of room to reserve"), "{}", e.message);
3645 }
3646 }
3647
3648 /// What wraps rather than being undefined, which is two questions and three flags.
3649 ///
3650 /// The older flag is the pair of the newer two, which is gcc's own reading of it, so a build
3651 /// that writes `-fno-strict-overflow` gets both and a build that writes one of the others gets
3652 /// only what it asked for.
3653 #[test]
3654 fn what_overflows_rather_than_being_undefined_is_asked_for_two_ways() {
3655 let (opts, _) = compile(&["-c", "a.c"]);
3656 assert_eq!(opts.wrapping, Wrapping::NONE, "nothing wraps unless it was asked for");
3657
3658 let (opts, _) = compile(&["-c", "-fwrapv", "a.c"]);
3659 assert_eq!(opts.wrapping, Wrapping { signed: true, pointer: false, trap: false });
3660
3661 let (opts, _) = compile(&["-c", "-fwrapv-pointer", "a.c"]);
3662 assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: true, trap: false });
3663
3664 let (opts, _) = compile(&["-c", "-fno-strict-overflow", "a.c"]);
3665 assert_eq!(opts.wrapping, Wrapping::ALL);
3666
3667 // And the last one wins, in both directions. A build that turns one of these on globally
3668 // and off for one directory is relying on that, and so is one that writes the pair and
3669 // then takes half of it back.
3670 let (opts, _) = compile(&["-c", "-fwrapv", "-fno-wrapv", "a.c"]);
3671 assert_eq!(opts.wrapping, Wrapping::NONE);
3672
3673 let (opts, _) = compile(&["-c", "-fno-strict-overflow", "-fstrict-overflow", "a.c"]);
3674 assert_eq!(opts.wrapping, Wrapping::NONE);
3675
3676 let (opts, _) = compile(&["-c", "-fno-strict-overflow", "-fno-wrapv-pointer", "a.c"]);
3677 assert_eq!(opts.wrapping, Wrapping { signed: true, pointer: false, trap: false });
3678 }
3679
3680 /// And the other answer to the signed question cannot be held at the same time as the first.
3681 ///
3682 /// A program cannot both wrap and stop, so writing both is writing a contradiction, and gcc
3683 /// resolves it by letting the last one win rather than by reporting anything. That was measured
3684 /// against gcc 16 rather than read out of the manual, which says nothing about it: `-ftrapv
3685 /// -fwrapv` emits no checked calls and `-fwrapv -ftrapv` emits them.
3686 #[test]
3687 fn a_signed_overflow_that_stops_is_the_other_answer_and_not_a_third_one() {
3688 let (opts, _) = compile(&["-c", "-ftrapv", "a.c"]);
3689 assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: false, trap: true });
3690
3691 let (opts, _) = compile(&["-c", "-fwrapv", "-ftrapv", "a.c"]);
3692 assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: false, trap: true });
3693
3694 let (opts, _) = compile(&["-c", "-ftrapv", "-fwrapv", "a.c"]);
3695 assert_eq!(opts.wrapping, Wrapping { signed: true, pointer: false, trap: false });
3696
3697 let (opts, _) = compile(&["-c", "-ftrapv", "-fno-strict-overflow", "a.c"]);
3698 assert_eq!(opts.wrapping, Wrapping::ALL);
3699
3700 let (opts, _) = compile(&["-c", "-ftrapv", "-fno-trapv", "a.c"]);
3701 assert_eq!(opts.wrapping, Wrapping::NONE);
3702
3703 // And the flag that says what may be assumed says nothing about what happens, so it leaves
3704 // this alone where it takes the wrapping away. gcc does the same.
3705 let (opts, _) = compile(&["-c", "-ftrapv", "-fstrict-overflow", "a.c"]);
3706 assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: false, trap: true });
3707 }
3708
3709 /// What a plain `char` is, which is four spellings of two answers and nothing by default.
3710 ///
3711 /// Nothing is the target's own answer and has to stay distinct from both of the others, since
3712 /// the same command line means a signed `char` on x86-64 and an unsigned one on Linux's arm64.
3713 /// The negative spellings are the other flag rather than a way of asking for the default, which
3714 /// was measured against gcc 16: `-fno-signed-char` defines `__CHAR_UNSIGNED__` and
3715 /// `-fno-unsigned-char` does not.
3716 #[test]
3717 fn the_signedness_of_a_plain_char_is_asked_for_in_four_ways() {
3718 let (opts, _) = compile(&["-c", "a.c"]);
3719 assert_eq!(opts.char_signed, None);
3720
3721 for flag in ["-fsigned-char", "-fno-unsigned-char"] {
3722 let (opts, _) = compile(&["-c", flag, "a.c"]);
3723 assert_eq!(opts.char_signed, Some(true), "{flag}");
3724 }
3725
3726 for flag in ["-funsigned-char", "-fno-signed-char"] {
3727 let (opts, _) = compile(&["-c", flag, "a.c"]);
3728 assert_eq!(opts.char_signed, Some(false), "{flag}");
3729 }
3730
3731 // And the last one wins, which is what a build that sets one globally and the other for a
3732 // directory relies on.
3733 let (opts, _) = compile(&["-c", "-funsigned-char", "-fsigned-char", "a.c"]);
3734 assert_eq!(opts.char_signed, Some(true));
3735
3736 // And what is asked for reaches the target, because that is what every other part of the
3737 // compiler asks. The triple is one whose own answer is the opposite, so a session that
3738 // ignored the flag would still read as signed here.
3739 let (opts, _) =
3740 compile(&["-c", "--target=aarch64-unknown-linux-gnu", "-fsigned-char", "a.c"]);
3741 assert!(Session::new(*opts).target.char_is_signed);
3742 let (opts, _) = compile(&["-c", "--target=aarch64-unknown-linux-gnu", "a.c"]);
3743 assert!(!Session::new(*opts).target.char_is_signed);
3744 }
3745
3746 /// And the size of an enumeration, which is one question with two spellings.
3747 #[test]
3748 fn the_smallest_enumeration_is_asked_for_and_taken_back() {
3749 let (opts, _) = compile(&["-c", "a.c"]);
3750 assert!(!opts.short_enums);
3751
3752 let (opts, _) = compile(&["-c", "-fshort-enums", "a.c"]);
3753 assert!(opts.short_enums);
3754
3755 let (opts, _) = compile(&["-c", "-fshort-enums", "-fno-short-enums", "a.c"]);
3756 assert!(!opts.short_enums);
3757
3758 let (opts, _) = compile(&["-c", "-fno-short-enums", "-fshort-enums", "a.c"]);
3759 assert!(opts.short_enums);
3760 }
3761
3762 /// And a value nothing means is refused rather than taken for the nearest thing it looks like.
3763 ///
3764 /// `-fcf-protection=all` is the spelling somebody writes from memory, and a compiler that read
3765 /// it as `full` would be guessing, while one that let it fall through to the optimizer's `-f`
3766 /// family would report it as an unknown pass. Neither is the news the build wants.
3767 #[test]
3768 fn a_control_flow_protection_nothing_means_is_refused() {
3769 let e = parse_args(&args(&["-c", "-fcf-protection=all", "a.c"])).unwrap_err();
3770 assert!(e.message.contains("is not a control flow protection"), "{}", e.message);
3771 assert!(e.message.contains("full, branch, return, none or check"), "{}", e.message);
3772 }
3773
3774 #[test]
3775 fn the_link_flags_are_collected_apart_from_the_compilation() {
3776 let (link, _) = linking(&[
3777 "-static",
3778 "-nostartfiles",
3779 "-rdynamic",
3780 "-s",
3781 "-fuse-ld=mold",
3782 "-L/opt/lib",
3783 "-B",
3784 "/opt/tools",
3785 "a.c",
3786 ]);
3787 assert!(link.is_static);
3788 assert!(link.no_startfiles);
3789 assert!(link.export_dynamic);
3790 assert!(link.strip);
3791 assert_eq!(link.use_ld.as_deref(), Some("mold"));
3792 assert_eq!(link.search, vec![PathBuf::from("/opt/lib")]);
3793 assert_eq!(link.prefixes, vec![PathBuf::from("/opt/tools")]);
3794 }
3795
3796 #[test]
3797 fn a_comma_in_dash_wl_separates_two_arguments() {
3798 let (link, _) = linking(&["-Wl,-rpath,/opt/lib", "-Xlinker", "--as-needed", "a.c"]);
3799 assert_eq!(link.passthrough, vec!["-rpath", "/opt/lib", "--as-needed"]);
3800 }
3801
3802 #[test]
3803 fn a_library_keeps_its_place_between_the_objects() {
3804 // Link order is semantic: `-lm` written between two files resolves for the one before
3805 // it and not for the one after, so a library cannot be collected into a list of its own.
3806 // The target is named because the suffix of an object is the target's and this asserts
3807 // on the names: the same command line on a Windows host plans two `.obj` files.
3808 let (_, plan) = linking(&["--target=x86_64-unknown-linux-gnu", "a.c", "-lm", "b.c"]);
3809 let link = plan.link.expect("expected a link step");
3810 assert_eq!(
3811 link.inputs,
3812 vec![
3813 link::Item::File("a.o".into()),
3814 link::Item::Library("m".into()),
3815 link::Item::File("b.o".into()),
3816 ]
3817 );
3818 // And it is not a job, because there is nothing to compile in a library.
3819 assert_eq!(plan.jobs.len(), 2);
3820 }
3821
3822 #[test]
3823 fn a_library_on_a_dash_c_line_is_a_note_rather_than_an_error() {
3824 let (_, plan) = linking(&["-c", "-lm", "a.c"]);
3825 assert!(plan.link.is_none());
3826 assert!(plan.notes.iter().any(|n| n.contains("-lm")), "{:?}", plan.notes);
3827 }
3828
3829 #[test]
3830 fn the_sysroot_reaches_the_linker_as_well_as_the_headers() {
3831 let (link, _) = linking(&["--sysroot=/opt/root", "a.c"]);
3832 assert_eq!(link.sysroot, Some(PathBuf::from("/opt/root")));
3833 }
3834
3835 fn printed(s: &[&str]) -> String {
3836 match parse_args(&args(s)).expect("expected an answer") {
3837 Action::Print(line) => line,
3838 other => panic!("expected an answer, got {other:?}"),
3839 }
3840 }
3841
3842 fn refused(s: &[&str]) -> String {
3843 parse_args(&args(s)).expect_err("expected a refusal").message
3844 }
3845
3846 #[test]
3847 fn a_warning_flag_this_compiler_has_not_heard_of_is_taken_rather_than_refused() {
3848 // The rule in section 4.1, and the reason for it is autoconf: a configure script finds
3849 // out whether a warning flag exists by passing it and looking at the exit status, so a
3850 // compiler that refuses one it does not know fails a script written for a newer GCC.
3851 let (opts, _) = compile(&["-Wall", "-Wextra", "-Wno-format-truncation", "-c", "a.c"]);
3852 assert!(!opts.warnings_are_errors);
3853 assert!(opts.warnings);
3854 // The two spellings that do mean something are still read.
3855 let (opts, _) = compile(&["-Werror", "-c", "a.c"]);
3856 assert!(opts.warnings_are_errors);
3857 let (opts, _) = compile(&["-w", "-c", "a.c"]);
3858 assert!(!opts.warnings);
3859 let (opts, _) = compile(&["-pedantic-errors", "-c", "a.c"]);
3860 assert!(opts.pedantic && opts.warnings_are_errors);
3861 }
3862
3863 #[test]
3864 fn an_argument_for_a_separate_tool_is_refused_rather_than_dropped() {
3865 // Every one of these says something about the output, so the wrong answer is silence.
3866 assert!(refused(&["-Wa,--noexecstack", "-c", "a.c"]).contains("separate assembler"));
3867 assert!(refused(&["-Wp,-DX", "-c", "a.c"]).contains("separate assembler"));
3868 assert!(refused(&["-specs=/x", "a.c"]).contains("-specs= is not supported"));
3869 assert!(refused(&["-mcmodel=kernel", "-c", "a.c"]).contains("small code model"));
3870 assert!(refused(&["-gdwarf-4", "-c", "a.c"]).contains("DWARF 5"));
3871 assert!(refused(&["-Ofast", "-c", "a.c"]).contains("fast math"));
3872 // The word size the target does not have, which is a target this compiler was not asked
3873 // for rather than a flag it does not know.
3874 let no32 = refused(&["--target=x86_64-unknown-linux-gnu", "-m32", "-c", "a.c"]);
3875 assert!(no32.contains("32 bit target"), "{no32}");
3876 }
3877
3878 /// `-gz` and the two spellings of the split, which are the two questions about the shape of
3879 /// the debug output rather than about how much of it there is.
3880 ///
3881 /// Both answers here are about what happens when there is debug information to shape, and
3882 /// there is none yet, so what is being asserted is that the flags are read and remembered
3883 /// rather than that anything changed in the output. That is the whole of what taking them
3884 /// claims, and it is worth a test because the day `rucc-debug` writes a section this is where
3885 /// it comes to find out what the command line said.
3886 #[test]
3887 fn the_shape_of_the_debug_output_is_recorded_even_where_there_is_none_of_it() {
3888 let (opts, _) = compile(&["-c", "a.c"]);
3889 assert_eq!(opts.compress, Compress::None, "uncompressed unless somebody asks");
3890
3891 // Bare `-gz` is `-gz=zlib`, measured against gcc 16 rather than read out of the manual,
3892 // which describes the flag without ever saying which algorithm it picks.
3893 assert_eq!(compile(&["-gz", "-c", "a.c"]).0.compress, Compress::Zlib);
3894 for (spelling, want) in [
3895 ("none", Compress::None),
3896 ("zlib", Compress::Zlib),
3897 ("zlib-gnu", Compress::ZlibGnu),
3898 ("zstd", Compress::Zstd),
3899 ] {
3900 let (opts, _) = compile(&[&format!("-gz={spelling}"), "-c", "a.c"]);
3901 assert_eq!(opts.compress, want, "{spelling}");
3902 }
3903
3904 // A value nothing here has heard of is refused rather than rounded to the nearest one,
3905 // because a build that asked for `zstd` and quietly got `zlib` would ship a file its
3906 // reader may not understand and would have no way of finding out.
3907 for bad in ["-gz=gzip", "-gz="] {
3908 let failed = refused(&[bad, "-c", "a.c"]);
3909 assert!(failed.contains("is not a way to compress"), "{bad}: {failed}");
3910 }
3911
3912 // The split is refused in the direction that would have written a file and taken in the
3913 // direction that describes what happens. A build system that names the `.dwo` as an
3914 // output has to hear about it now rather than at the point the file is missing.
3915 let (opts, _) = compile(&["-gno-split-dwarf", "-g", "-c", "a.c"]);
3916 assert!(opts.debug_info, "the negative spelling says nothing about how much");
3917 let failed = refused(&["-gsplit-dwarf", "-c", "a.c"]);
3918 assert!(failed.contains(".dwo"), "the refusal names the file it would have written");
3919 }
3920
3921 /// The `-flto` family, which is the whole of an optimization this compiler does not do.
3922 ///
3923 /// Taken rather than refused because ignoring it gives a correct program that is slower than
3924 /// it could have been, which is section 4.1's hint about speed. The values are still held to
3925 /// gcc's, so a command line written for clang is told rather than quietly taken.
3926 #[test]
3927 fn the_link_time_family_is_read_and_checked_and_nothing_is_done_about_it() {
3928 let (opts, _) = compile(&["-c", "a.c"]);
3929 assert!(!opts.lto.requested, "nothing asks unless the command line does");
3930
3931 let (opts, _) = compile(&["-flto", "-c", "a.c"]);
3932 assert!(opts.lto.requested);
3933 assert_eq!(opts.lto.jobs, LtoJobs::One, "bare -flto is one process, the way gcc reads it");
3934
3935 // The last of the two directions wins, the same as every other pair of `-f` spellings.
3936 assert!(!compile(&["-flto", "-fno-lto", "-c", "a.c"]).0.lto.requested);
3937 assert!(compile(&["-fno-lto", "-flto", "-c", "a.c"]).0.lto.requested);
3938
3939 // A count is a count, and asking for one implies asking for the optimization.
3940 for (spelling, want) in [
3941 ("auto", LtoJobs::Auto),
3942 ("jobserver", LtoJobs::Jobserver),
3943 ("1", LtoJobs::One),
3944 ("8", LtoJobs::Count(8)),
3945 ] {
3946 let (opts, _) = compile(&[&format!("-flto={spelling}"), "-c", "a.c"]);
3947 assert_eq!(opts.lto.jobs, want, "{spelling}");
3948 assert!(opts.lto.requested, "{spelling} asks for it too");
3949 }
3950
3951 // gcc refuses a zero rather than reading it as `-fno-lto`, and `thin` is clang's spelling
3952 // of a question gcc answers with `-flto-partition=`, so somebody who wrote it meant a
3953 // different compiler and gets told so here rather than getting a serial link.
3954 for bad in ["-flto=0", "-flto=thin", "-flto=full", "-flto=-1"] {
3955 let failed = refused(&[bad, "-c", "a.c"]);
3956 assert!(failed.contains("link time jobs"), "{bad}: {failed}");
3957 }
3958
3959 // How the program is cut up before the work is spread over it.
3960 assert_eq!(compile(&["-c", "a.c"]).0.lto.partition, Partition::Balanced, "gcc's default");
3961 for (spelling, want) in [
3962 ("balanced", Partition::Balanced),
3963 ("1to1", Partition::OneToOne),
3964 ("one", Partition::One),
3965 ("max", Partition::Max),
3966 ("none", Partition::None),
3967 ] {
3968 let (opts, _) = compile(&[&format!("-flto-partition={spelling}"), "-c", "a.c"]);
3969 assert_eq!(opts.lto.partition, want, "{spelling}");
3970 }
3971 assert!(refused(&["-flto-partition=big", "-c", "a.c"]).contains("partitioning model"));
3972
3973 // And how hard the bytecode is compressed on its way into the object, which is zstd's
3974 // range of levels and is the range gcc checks an argument against.
3975 assert_eq!(compile(&["-c", "a.c"]).0.lto.compression, None, "whatever it does by default");
3976 assert_eq!(compile(&["-flto-compression-level=0", "-c", "a.c"]).0.lto.compression, Some(0));
3977 let (opts, _) = compile(&["-flto-compression-level=19", "-c", "a.c"]);
3978 assert_eq!(opts.lto.compression, Some(19));
3979 for bad in ["-flto-compression-level=20", "-flto-compression-level=-1"] {
3980 let failed = refused(&[bad, "-c", "a.c"]);
3981 assert!(failed.contains("compression level"), "{bad}: {failed}");
3982 }
3983
3984 // The two pairs that describe an arrangement rather than ask for one. Every object here
3985 // holds its machine code, so the fat spelling is what already happens and the other is a
3986 // smaller file rather than a different program, and the plugin pair is about a tool the
3987 // design in `spec/09-optimizer.md` never loads.
3988 for taken in [
3989 "-ffat-lto-objects",
3990 "-fno-fat-lto-objects",
3991 "-fuse-linker-plugin",
3992 "-fno-use-linker-plugin",
3993 ] {
3994 let (opts, _) = compile(&[taken, "-c", "a.c"]);
3995 assert!(!opts.lto.requested, "{taken} says nothing about whether to do it");
3996 }
3997 }
3998
3999 /// The profile family, which is the only one here that splits down the middle.
4000 ///
4001 /// Reading a profile is taken and writing one is refused, and the line between them is the one
4002 /// section 4.1 draws: ignoring a request to read the counts gives a correct program that is
4003 /// slower than it could have been, and ignoring a request to write them means a file the build
4004 /// declared as an output never appears.
4005 #[test]
4006 fn reading_a_profile_is_taken_and_writing_one_is_refused() {
4007 let (opts, _) = compile(&["-c", "a.c"]);
4008 assert!(!opts.profile_data.requested, "nothing asks unless the command line does");
4009 assert_eq!(opts.profile_data.path, None);
4010
4011 let (opts, _) = compile(&["-fprofile-use", "-c", "a.c"]);
4012 assert!(opts.profile_data.requested);
4013 assert_eq!(opts.profile_data.path, None, "beside the object, the way gcc looks");
4014
4015 let (opts, _) = compile(&["-fprofile-use=/counts", "-c", "a.c"]);
4016 assert!(opts.profile_data.requested, "naming a path asks for it too");
4017 assert_eq!(opts.profile_data.path.as_deref(), Some("/counts"));
4018
4019 // The last of the two directions wins, the same as every other pair of `-f` spellings.
4020 assert!(
4021 !compile(&["-fprofile-use", "-fno-profile-use", "-c", "a.c"]).0.profile_data.requested
4022 );
4023 assert!(
4024 compile(&["-fno-profile-use", "-fprofile-use", "-c", "a.c"]).0.profile_data.requested
4025 );
4026
4027 // The rest of the reading half, which is where the files are and three answers about what
4028 // to make of what is in them.
4029 let (opts, _) = compile(&[
4030 "-fprofile-dir=/build/profiles",
4031 "-fprofile-abs-path",
4032 "-fprofile-correction",
4033 "-fprofile-partial-training",
4034 "-c",
4035 "a.c",
4036 ]);
4037 assert_eq!(opts.profile_data.dir.as_deref(), Some("/build/profiles"));
4038 assert!(opts.profile_data.absolute);
4039 assert!(opts.profile_data.correction);
4040 assert!(opts.profile_data.partial_training);
4041
4042 // Writing one, which is refused by name. The first four instrument the program and the
4043 // last writes a file beside the object, and a build that got neither and no message would
4044 // go on to optimize against counts that were never gathered.
4045 for writing in [
4046 "-fprofile-generate",
4047 "-fprofile-generate=/build/profiles",
4048 "-fprofile-arcs",
4049 "--coverage",
4050 "-fcondition-coverage",
4051 "-fpath-coverage",
4052 ] {
4053 let failed = refused(&[writing, "-c", "a.c"]);
4054 assert!(failed.contains("instrument"), "{writing}: {failed}");
4055 }
4056 assert!(refused(&["-ftest-coverage", "-c", "a.c"]).contains(".gcno"), "it names the file");
4057
4058 // The negative spellings of the refused half are what already happens, so they are taken.
4059 for taken in ["-fno-profile-generate", "-fno-profile-arcs", "-fno-test-coverage"] {
4060 let (opts, _) = compile(&[taken, "-c", "a.c"]);
4061 assert!(!opts.profile_data.requested, "{taken} asks for nothing");
4062 }
4063
4064 // And the flags that describe the instrumentation that is refused above, which are checked
4065 // and dropped. Checked because a typo is worth finding here rather than on the day the
4066 // instrumentation lands.
4067 for taken in [
4068 "-fprofile-update=single",
4069 "-fprofile-update=atomic",
4070 "-fprofile-update=prefer-atomic",
4071 "-fprofile-reproducible=serial",
4072 "-fprofile-reproducible=parallel-runs",
4073 "-fprofile-reproducible=multithreaded",
4074 "-fprofile-values",
4075 "-fno-profile-values",
4076 "-fprofile-info-section",
4077 "-fprofile-filter-files=a.c",
4078 "-fprofile-exclude-files=b.c",
4079 "-fprofile-note=a.gcno",
4080 ] {
4081 let (opts, _) = compile(&[taken, "-c", "a.c"]);
4082 assert!(!opts.profile_data.requested, "{taken} says nothing about reading one");
4083 }
4084 assert!(refused(&["-fprofile-update=none", "-c", "a.c"]).contains("update method"));
4085 assert!(refused(&["-fprofile-reproducible=any", "-c", "a.c"]).contains("reproducibility"));
4086 }
4087
4088 /// The sanitizers, which are refused by name and are the one family refused for a reason that
4089 /// is not about the bytes.
4090 ///
4091 /// A sanitizer is a promise that the program is watched while it runs, so a build that asked
4092 /// for one and was quietly given a program with no checks in it gets a test suite that passes
4093 /// for the wrong reason rather than a slower program.
4094 #[test]
4095 fn a_sanitizer_that_is_still_asked_for_at_the_end_of_the_line_is_refused_by_name() {
4096 for asked in ["address", "undefined", "thread", "kernel-address", "leak", "memory"] {
4097 let failed = refused(&[&format!("-fsanitize={asked}"), "-c", "a.c"]);
4098 assert!(failed.contains(asked), "the refusal names what was asked for: {failed}");
4099 assert!(failed.contains("-fsafety=detect"), "and the nearest thing: {failed}");
4100 }
4101
4102 // A list is every name in it, and the first one still standing is the one named.
4103 let failed = refused(&["-fsanitize=address,undefined", "-c", "a.c"]);
4104 assert!(failed.contains("address"), "{failed}");
4105
4106 // A name that is not one, which is worth its own message: somebody who wrote `-fsanitize`
4107 // with a typo in it has a different problem from somebody who wrote a real one.
4108 for bad in ["-fsanitize=bogus", "-fsanitize=address,bogus", "-fno-sanitize=bogus"] {
4109 let failed = refused(&[bad, "-c", "a.c"]);
4110 assert!(failed.contains("is not a sanitizer"), "{bad}: {failed}");
4111 }
4112
4113 // gcc takes `all` only in the negative, and so does this.
4114 assert!(refused(&["-fsanitize=all", "-c", "a.c"]).contains("only `-fno-sanitize=all`"));
4115
4116 // Asking and then taking it back is asking for nothing, which is why the answer waits for
4117 // the end of the line. A build whose shared flags turn a check on and whose rule for one
4118 // file turns it off again compiles that file here.
4119 for pair in [
4120 ["-fsanitize=address", "-fno-sanitize=address"],
4121 ["-fsanitize=address,undefined", "-fno-sanitize=all"],
4122 ["-fsanitize=undefined", "-fno-sanitize=undefined"],
4123 ] {
4124 let (opts, _) = compile(&[pair[0], pair[1], "-c", "a.c"]);
4125 assert_eq!(opts.safety, rucc_session::Safety::Off, "{pair:?} asked for nothing");
4126 }
4127 // And the other order still asks, because the last word is the one that counts.
4128 assert!(!refused(&["-fno-sanitize=address", "-fsanitize=address", "-c", "a.c"]).is_empty());
4129
4130 // What a check does when it fires is an answer about checks that are refused, so there is
4131 // nothing left for it to change and it is taken.
4132 for taken in [
4133 "-fsanitize-recover=undefined",
4134 "-fno-sanitize-recover=all",
4135 "-fsanitize-trap=undefined",
4136 "-fno-sanitize-trap=all",
4137 "-fsanitize-undefined-trap-on-error",
4138 "-fsanitize-address-use-after-scope",
4139 "-fno-sanitize-address-use-after-scope",
4140 "-fsanitize-sections=.data",
4141 ] {
4142 let (opts, _) = compile(&[taken, "-c", "a.c"]);
4143 assert_eq!(opts.safety, rucc_session::Safety::Off, "{taken} asks for no checking");
4144 }
4145 assert!(refused(&["-fsanitize-recover=bogus", "-c", "a.c"]).contains("is not a sanitizer"));
4146
4147 // Coverage instrumentation is refused rather than dropped, because a fuzzer with no
4148 // feedback runs blind and never says so.
4149 let failed = refused(&["-fsanitize-coverage=trace-pc", "-c", "a.c"]);
4150 assert!(failed.contains("feedback"), "{failed}");
4151 let failed = refused(&["-fsanitize-coverage=trace-pc-guard", "-c", "a.c"]);
4152 assert!(failed.contains("trace-pc or trace-cmp"), "gcc takes two of them: {failed}");
4153 }
4154
4155 #[test]
4156 fn the_levels_gcc_spells_differently_are_the_levels_they_mean() {
4157 assert_eq!(compile(&["-O", "-c", "a.c"]).0.opt_level, OptLevel::O1);
4158 assert_eq!(compile(&["-Og", "-c", "a.c"]).0.opt_level, OptLevel::O1);
4159 assert_eq!(compile(&["-O2", "-c", "a.c"]).0.opt_level, OptLevel::O2);
4160 }
4161
4162 #[test]
4163 fn the_machine_flags_that_name_what_we_already_do_are_taken_and_the_rest_are_not() {
4164 let line = ["--target=x86_64-unknown-linux-gnu", "-m64", "-march=x86-64-v3"];
4165 let (opts, _) =
4166 compile(&[&line[..], &["-mtune=native", "-mabi=sysv", "-c", "a.c"]].concat());
4167 assert_eq!(opts.target.to_string(), "x86_64-unknown-linux-gnu");
4168 let wrong = refused(&["--target=x86_64-unknown-linux-gnu", "-mabi=ms", "-c", "a.c"]);
4169 assert!(wrong.contains("sysv convention"), "{wrong}");
4170 }
4171
4172 #[test]
4173 fn the_thread_flag_is_a_macro_and_a_library_and_the_library_goes_last() {
4174 let (opts, plan) = compile(&["-pthread", "-c", "a.c"]);
4175 assert!(opts.defines.iter().any(|d| d == "_REENTRANT"));
4176 // After the input, because a static link takes what it needs from a library when it
4177 // reaches it and not afterwards.
4178 let names: Vec<&str> = plan.jobs.iter().map(|j| j.input.as_str()).collect();
4179 assert_eq!(names, vec!["a.c"]);
4180 }
4181
4182 #[test]
4183 fn the_questions_a_build_system_asks_before_it_compiles_anything() {
4184 let target = "--target=x86_64-unknown-linux-gnu";
4185 assert_eq!(printed(&[target, "-dumpmachine"]), "x86_64-unknown-linux-gnu");
4186 assert_eq!(printed(&[target, "-dumpversion"]), VERSION);
4187 assert_eq!(printed(&[target, "-dumpfullversion"]), VERSION);
4188 assert_eq!(printed(&[target, "-print-multiarch"]), "x86_64-linux-gnu");
4189 // A name nothing holds comes back unchanged, which is GCC's rule and is what makes the
4190 // answer safe to paste into a link line whether or not the file is there.
4191 assert_eq!(printed(&[target, "-print-file-name=no-such-library.a"]), "no-such-library.a");
4192 assert_eq!(printed(&[target, "-print-prog-name=ld"]), "ld");
4193 let dirs = printed(&[target, "-print-search-dirs"]);
4194 assert!(dirs.starts_with("install: "), "{dirs}");
4195 assert!(dirs.contains("\nlibraries: ="), "{dirs}");
4196 }
4197
4198 #[test]
4199 fn the_sysroot_in_effect_is_the_one_the_command_line_named_or_the_one_for_the_target() {
4200 // A tree the user named is the answer whatever the target is, because it is the answer to
4201 // every other question too.
4202 assert_eq!(printed(&["--sysroot=/opt/cross", "-print-sysroot"]), "/opt/cross");
4203
4204 // A target that is no machine this suite runs on is read under the cache, and the answer is
4205 // the root rather than one of the directories under it, since what asks is looking for a
4206 // file of its own.
4207 let root = cache::dir().join("sysroots").join("riscv64-linux-musl");
4208 assert_eq!(
4209 printed(&["--target=riscv64-linux-musl", "-print-sysroot"]),
4210 root.display().to_string()
4211 );
4212
4213 // And a compile for this machine has no sysroot, which is the empty line GCC prints when it
4214 // was configured without one rather than a `/` that would be a claim about the filesystem.
4215 let host = Triple::host().expect("a host this compiler knows");
4216 assert_eq!(printed(&[&format!("--target={host}"), "-print-sysroot"]), "");
4217 }
4218
4219 #[test]
4220 fn the_provenance_of_a_sysroot_is_the_manifest_it_carries() {
4221 // Section 13.5 wants seven things per input and wants them machine readable, and the manifest
4222 // is the record that already has them, so the flag prints that rather than a second format.
4223 let manifest = "rucc sysroot manifest 2\n\
4224 target\tx86_64-linux-musl\n\
4225 include/generic/stdio.h\tmusl-1.2.5\t\
4226 https://musl.libc.org/releases/musl-1.2.5.tar.gz\t\
4227 0000000000000000000000000000000000000000000000000000000000000000\tmit\t\
4228 bundled\n\
4229 lib/libc.so\tmusl-1.2.5\t\
4230 https://musl.libc.org/releases/musl-1.2.5.tar.gz\t\
4231 1111111111111111111111111111111111111111111111111111111111111111\tmit\t\
4232 generated\n";
4233 let tree = TempTree::new("provenance", &[("manifest", manifest)]);
4234 let sysroot = format!("--sysroot={}", tree.0.display());
4235 assert_eq!(printed(&[&sysroot, "-print-sysroot-provenance"]), manifest);
4236
4237 // A tree with no manifest in it is a tree somebody assembled themselves, and nothing here
4238 // knows where any of it came from. Saying nothing is the only honest answer, and a reader can
4239 // tell it from a manifest with no inputs because that one still has its two header lines.
4240 let bare = TempTree::new("provenance-bare", &[]);
4241 assert_eq!(
4242 printed(&[&format!("--sysroot={}", bare.0.display()), "-print-sysroot-provenance"]),
4243 ""
4244 );
4245
4246 // And a compile for this machine has no sysroot at all, which is the same empty answer
4247 // `-print-sysroot` gives for it.
4248 let host = Triple::host().expect("a host this compiler knows");
4249 assert_eq!(printed(&[&format!("--target={host}"), "-print-sysroot-provenance"]), "");
4250
4251 // And the other spelling, which section 13.5 is the document that writes.
4252 assert_eq!(printed(&[&sysroot, "--print-sysroot-provenance"]), manifest);
4253 }
4254
4255 #[test]
4256 fn a_manifest_this_build_cannot_read_is_refused_rather_than_printed() {
4257 // Passing a file we could not parse to whoever asked would make their parser the one that
4258 // finds the problem, and the three uses section 13.5 gives for this are all somebody else
4259 // parsing it.
4260 let tree = TempTree::new(
4261 "provenance-bad",
4262 &[("manifest", "rucc sysroot manifest 2\ntarget\tx86_64-linux-musl\nlib/libc.a\n")],
4263 );
4264 let message =
4265 refused(&[&format!("--sysroot={}", tree.0.display()), "-print-sysroot-provenance"]);
4266 assert!(message.contains("manifest"), "{message}");
4267 assert!(message.contains("1 fields where an input has six"), "{message}");
4268 }
4269
4270 #[test]
4271 fn the_two_dependency_flags_that_stop_after_the_rule_stop_after_the_rule() {
4272 let (opts, _) = compile(&["-M", "a.c"]);
4273 assert!(opts.deps.emit && opts.deps.instead_of_compiling);
4274 assert!(opts.deps.system_headers, "plain -M lists them");
4275 assert_eq!(opts.emit, EmitKind::Preprocessed);
4276
4277 // Even where a later flag asked for something else, because the family is a mode and
4278 // the mode is what the run is for.
4279 let (opts, _) = compile(&["-M", "-c", "a.c"]);
4280 assert_eq!(opts.emit, EmitKind::Preprocessed);
4281
4282 let (opts, _) = compile(&["-MM", "a.c"]);
4283 assert!(!opts.deps.system_headers);
4284 }
4285
4286 #[test]
4287 fn the_two_that_end_in_d_leave_the_compilation_alone() {
4288 let (opts, _) = compile(&["-MD", "-c", "a.c"]);
4289 assert!(opts.deps.emit && !opts.deps.instead_of_compiling);
4290 assert!(opts.deps.system_headers);
4291 assert_eq!(opts.emit, EmitKind::Object);
4292
4293 let (opts, _) = compile(&["-MMD", "-c", "a.c"]);
4294 assert!(opts.deps.emit && !opts.deps.instead_of_compiling);
4295 assert!(!opts.deps.system_headers);
4296 }
4297
4298 #[test]
4299 fn nothing_puts_the_system_headers_back_once_a_flag_has_taken_them_out() {
4300 // GCC's rule, and not an oversight in it. The flag asking for fewer of them is read as
4301 // the answer, because the other one never asked the question.
4302 let (opts, _) = compile(&["-MM", "-M", "a.c"]);
4303 assert!(!opts.deps.system_headers);
4304 let (opts, _) = compile(&["-MD", "-MMD", "-c", "a.c"]);
4305 assert!(!opts.deps.system_headers);
4306 let (opts, _) = compile(&["-MMD", "-MD", "-c", "a.c"]);
4307 assert!(!opts.deps.system_headers);
4308 }
4309
4310 #[test]
4311 fn a_target_arrives_escaped_from_one_flag_and_untouched_from_the_other() {
4312 let (opts, _) = compile(&["-MM", "-MT", "a b.o", "-MQ", "a b.o", "a.c"]);
4313 assert_eq!(opts.deps.targets, vec!["a b.o".to_owned(), "a\\ b.o".to_owned()]);
4314 }
4315
4316 #[test]
4317 fn the_rest_of_the_family_is_a_file_and_a_switch() {
4318 let (opts, _) = compile(&["-MM", "-MF", "dep.d", "-MP", "a.c"]);
4319 assert_eq!(opts.deps.file.as_deref(), Some("dep.d"));
4320 assert!(opts.deps.phony);
4321
4322 for flag in ["-MF", "-MT", "-MQ"] {
4323 let e = parse_args(&args(&[flag])).unwrap_err();
4324 assert!(e.message.contains("requires an argument"), "{}", e.message);
4325 }
4326 }
4327
4328 /// A directory of sources for one test, removed when the test is done with it.
4329 struct TempTree(PathBuf);
4330
4331 impl Drop for TempTree {
4332 fn drop(&mut self) {
4333 let _ = std::fs::remove_dir_all(&self.0);
4334 }
4335 }
4336
4337 impl TempTree {
4338 fn new(name: &str, files: &[(&str, &str)]) -> TempTree {
4339 let dir = std::env::temp_dir().join(format!("rucc-deps-{}-{name}", std::process::id()));
4340 let _ = std::fs::remove_dir_all(&dir);
4341 std::fs::create_dir_all(&dir).expect("temporary directory should be writable");
4342 for (path, text) in files {
4343 let at = dir.join(path);
4344 if let Some(parent) = at.parent() {
4345 std::fs::create_dir_all(parent).expect("creating a subdirectory should work");
4346 }
4347 std::fs::write(&at, text).expect("writing a temporary file should work");
4348 }
4349 TempTree(dir)
4350 }
4351
4352 fn path(&self, name: &str) -> String {
4353 self.0.join(name).to_string_lossy().into_owned()
4354 }
4355 }
4356
4357 #[test]
4358 fn the_rule_names_what_the_includes_found_and_names_each_of_them_once() {
4359 // End to end, because the list comes from the preprocessor and the format comes from
4360 // somewhere else, and a test of either half on its own would pass with the two of them
4361 // wired up backwards.
4362 let tree = TempTree::new(
4363 "found",
4364 &[
4365 ("a.c", "#include \"one.h\"\n#include \"two.h\"\nint main(void) { return X; }\n"),
4366 ("one.h", "#define X 0\n"),
4367 ("two.h", "#include \"one.h\"\n"),
4368 ],
4369 );
4370 let out = tree.path("dep.d");
4371 let code = run(&args(&["-MM", "-MF", &out, "-o", &tree.path("a.i"), &tree.path("a.c")]));
4372 assert_eq!(code, 0);
4373
4374 let text = std::fs::read_to_string(&out).expect("the rule should have been written");
4375 let names: Vec<&str> = text.split_whitespace().collect();
4376 // The target, the source, and each header once however many times it was reached.
4377 assert_eq!(names.first(), Some(&"a.o:"), "{text}");
4378 assert_eq!(names.iter().filter(|n| n.ends_with("one.h")).count(), 1, "{text}");
4379 assert_eq!(names.iter().filter(|n| n.ends_with("two.h")).count(), 1, "{text}");
4380 // And the `-o` went to the file the rule replaced, which is left empty rather than
4381 // absent because a makefile that named it as a target will look for it.
4382 assert_eq!(std::fs::read(tree.path("a.i")).expect("the output should exist"), b"");
4383 }
4384
4385 #[test]
4386 fn a_header_that_is_only_reached_under_a_guard_is_still_a_dependency() {
4387 // The multiple-include optimization means the second reach never opens the file. It is
4388 // still a file this translation unit was built from, so it is still in the rule.
4389 let tree = TempTree::new(
4390 "guarded",
4391 &[
4392 ("a.c", "#include \"g.h\"\n#include \"g.h\"\nint main(void) { return 0; }\n"),
4393 ("g.h", "#ifndef G\n#define G\n#endif\n"),
4394 ],
4395 );
4396 let out = tree.path("dep.d");
4397 let code = run(&args(&["-MM", "-MF", &out, "-o", &tree.path("a.i"), &tree.path("a.c")]));
4398 assert_eq!(code, 0);
4399 let text = std::fs::read_to_string(&out).expect("the rule should have been written");
4400 assert_eq!(text.split_whitespace().filter(|n| n.ends_with("g.h")).count(), 1, "{text}");
4401 }
4402
4403 #[test]
4404 fn every_imacros_file_is_read_before_every_include_file_whatever_order_they_were_written() {
4405 // Measured against GCC rather than read: the two flags the other way round produce the
4406 // same output byte for byte, so the command line order between the two families does not
4407 // decide anything and the order within one does. The `-include` file here can only see
4408 // the definition if the `-imacros` file that was written after it ran first.
4409 let tree = TempTree::new(
4410 "preinclude",
4411 &[
4412 ("a.c", "int main(void) { return 0; }\n"),
4413 ("i.h", "#ifdef FROM_MACROS\nint saw_it;\n#else\nint missed_it;\n#endif\n"),
4414 ("m.h", "#define FROM_MACROS 1\nint macros_text;\n"),
4415 ],
4416 );
4417 let out = tree.path("a.i");
4418 let code = run(&args(&[
4419 "-E",
4420 "-include",
4421 &tree.path("i.h"),
4422 "-imacros",
4423 &tree.path("m.h"),
4424 "-o",
4425 &out,
4426 &tree.path("a.c"),
4427 ]));
4428 assert_eq!(code, 0);
4429 let text = std::fs::read_to_string(&out).expect("the output should have been written");
4430 assert!(text.contains("saw_it"), "{text}");
4431 // And the text of the `-imacros` file is thrown away, which is the whole difference
4432 // between the two flags.
4433 assert!(!text.contains("macros_text"), "{text}");
4434 }
4435
4436 #[test]
4437 fn a_file_the_command_line_named_is_a_prerequisite_the_same_as_one_a_directive_named() {
4438 let tree = TempTree::new(
4439 "preinclude-deps",
4440 &[
4441 ("a.c", "int main(void) { return 0; }\n"),
4442 ("i.h", "int from_include;\n"),
4443 ("m.h", "#define M 1\n"),
4444 ],
4445 );
4446 let out = tree.path("dep.d");
4447 let code = run(&args(&[
4448 "-MM",
4449 "-MF",
4450 &out,
4451 "-include",
4452 &tree.path("i.h"),
4453 "-imacros",
4454 &tree.path("m.h"),
4455 "-o",
4456 &tree.path("a.i"),
4457 &tree.path("a.c"),
4458 ]));
4459 assert_eq!(code, 0);
4460 let text = std::fs::read_to_string(&out).expect("the rule should have been written");
4461 assert!(text.contains("i.h"), "{text}");
4462 assert!(text.contains("m.h"), "{text}");
4463 }
4464
4465 #[test]
4466 fn a_command_line_include_that_is_nowhere_on_the_path_is_an_error_and_not_a_warning() {
4467 // Including the directory of the source file, which is not on the path for these: the
4468 // command line was not written there, so a name in it is relative to where the compiler
4469 // was run rather than to where the source sits.
4470 let tree = TempTree::new(
4471 "preinclude-missing",
4472 &[("sub/a.c", "int main(void) { return 0; }\n"), ("sub/beside.h", "int x;\n")],
4473 );
4474 let code = run(&args(&["-E", "-include", "beside.h", "-o", "-", &tree.path("sub/a.c")]));
4475 assert_eq!(code, 1);
4476 }
4477
4478 #[test]
4479 fn a_command_line_that_links_names_the_executable_and_not_the_object_it_went_through() {
4480 // The object a link goes through is in a temporary directory and is gone before `make`
4481 // reads any of this, so the rule that named it would be a rule for a file that is never
4482 // there. The target and the file are both the `-o`, which is the executable.
4483 let (opts, plan) = compile(&["-MD", "sub/a.c", "-o", "prog"]);
4484 assert_eq!(plan.output.as_deref(), Some("prog"));
4485 assert_eq!(deps::default_target("sub/a.c", deps_target_output(&opts, &plan)), "prog");
4486 assert_eq!(
4487 deps::default_file(&opts.deps, "sub/a.c", plan.output.as_deref()).as_deref(),
4488 Some("prog.d")
4489 );
4490 }
4491
4492 #[test]
4493 fn the_plan_keeps_the_output_name_because_the_rule_is_written_from_it() {
4494 let (_, plan) = compile(&["-MMD", "-c", "sub/a.c", "-o", "obj/x.o"]);
4495 assert_eq!(plan.output.as_deref(), Some("obj/x.o"));
4496 let (_, plan) = compile(&["-MMD", "-c", "sub/a.c"]);
4497 assert_eq!(plan.output, None);
4498 }
4499
4500 #[test]
4501 fn usage_fits_on_a_screen() {
4502 // Not a style preference. A help text that scrolls is one nobody reads, and this is
4503 // the cheapest way to keep it honest as flags accumulate. The number goes up only when
4504 // a family of flags arrives that has nowhere to share a line, which the two pass gates
4505 // were and which the two fuel flags and `-fsafety=` now are, and it goes up by exactly
4506 // the lines that family took. The four it went up by last are the flags a build system
4507 // passes without being asked to: how much to say, what machine to generate for, threads,
4508 // and the questions `configure` asks before it compiles anything. The one it went up by
4509 // last is the second line of `--emit`, whose kinds are a family that has now outgrown
4510 // one line and has nowhere else to go. The two it went up by last are the dependency
4511 // family, which is eight flags that share nothing with anything above them. The one it
4512 // went up by last is the four spellings of position independent code, which every
4513 // configure script writes and which could only have shared the link line, and that line
4514 // is already four characters short of the limit. The two it went up by last are the rest
4515 // of the include family, which is six more flags that change where a header is looked for
4516 // and two that name a header outright. The one it went up by last is the pair that keeps
4517 // the intermediate files and times the steps, which belong next to the two flags above
4518 // them that are also about watching a compilation rather than changing one. The two it
4519 // went up by last are the section flags and the visibility flag, which are what a build
4520 // that cares about the size of what it ships and about which names it exports writes, and
4521 // the second of them was already taken and only missing from here. The one it went up by
4522 // last is the stack protector, which is four spellings of one question and which every
4523 // distribution puts on every command line it issues, so a build that reads this list
4524 // looking for it and does not find it has to go and read the specification instead. The one
4525 // it went up by last is the profiler, which is two spellings of the request and two of
4526 // where the call goes, and which is about watching a program run rather than about what is
4527 // generated, so it shares its subject with nothing above it. The one it went up by last is
4528 // the room a function opens with for something to be written over it later, which takes an
4529 // argument of its own shape and is what a kernel build asks for, so it fits beside the
4530 // profiler and nothing else. The one it went up by last is what overflows rather than being
4531 // undefined, which is three spellings of two questions and which a kernel build and a great
4532 // deal of code written before the standard settled both pass. The one it went up by last is
4533 // the other answer to the first of those questions, which could not share the line because
4534 // what it asks for is the opposite of what the flags on that line ask for. The one it went
4535 // up by last is the split of the line that lists what this compiler does anyway into that
4536 // and what it assumes anyway, which are two different claims that were sharing a line until
4537 // the second of them got a second flag and the line stopped fitting. The one it went up by
4538 // last is the three flags that change the ABI rather than the code, which have to be given
4539 // to every file in a program or none of them and which therefore belong somewhere a person
4540 // reading this list will see them. The one it went up by last is the floating point group,
4541 // which is two lines rather than one because the first of them is a choice this compiler
4542 // records and the rest are claims about what it does anyway, and putting a real setting on
4543 // the same line as three flags that change nothing would be misleading about both. The one
4544 // it went up by last is the flag that says a write has to stay inside the member it names,
4545 // which is a setting rather than a claim and so cannot share the line above it, that being
4546 // the one that picks a tier. The two it went up by last are the prefix mapping family,
4547 // which is four flags whose whole job is to keep a build's output the same from two
4548 // different directories, and which a person chasing a reproducible build comes here
4549 // looking for by name. The one it went up by last is how the debug sections are compressed
4550 // and whether they go in a file of their own, which are two questions about the shape of
4551 // the debug output, where the line above them is about how much of it there is. The one it
4552 // went up by last is the `restrict` contract, which is a setting for the same reason the
4553 // flag that keeps a write inside its member is and which is the check a person who has been
4554 // bitten by a vectorizer comes here looking for. The one it went up by last is link time
4555 // optimization, which is a whole optimization rather than a flag and which says so on its
4556 // own line, because a build that passes it and reads this looking for what it got is
4557 // asking a question no other line here answers. The one it went up by last is the sysroot,
4558 // which is the question somebody asks when a cross build read a file nobody expected, and
4559 // which has no room on the line above it because the answers there are a path each and this
4560 // one is the root all of them are under. The one it went up by last is what is inside that
4561 // root and where each of it came from, which is a question about a whole tree rather than
4562 // about a path and which is long enough on its own that it could not have shared a line with
4563 // anything. The one it went up by last is the profile family, which splits down the middle
4564 // where no other family here does, so the line has to name the half that is taken and the
4565 // half that is refused or it would be read as taking both. The one it went up by last is
4566 // the sanitizers, which are what somebody reaching for a checked build writes first and
4567 // which belong beside the tier that is the nearest thing here to what they asked for.
4568 assert!(USAGE.lines().count() < 69, "usage text has grown past one screen");
4569 }
4570}