Skip to main content

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/0.10.74")]
30
31pub mod assemble;
32pub mod cache;
33pub mod compile;
34pub mod deps;
35pub mod fetch;
36mod glibc;
37pub mod install;
38pub mod library;
39pub mod link;
40mod map;
41pub mod phase;
42pub mod preprocess;
43pub mod schedule;
44
45use std::fmt::Write as _;
46use std::io::Write as _;
47use std::path::PathBuf;
48
49use rucc_codegen::coverage::{self, Fired};
50use rucc_codegen::lowering::Lowerings;
51use rucc_codegen::pressure::Pressure;
52use rucc_pp::Dependency;
53use rucc_session::{
54    Compress, Control, Dumps, EmitKind, Hook, Options, Pic, PrefixMap, Preinclude, Protector,
55    SaveTemps, Session, Std, Wrapping, runtime,
56};
57use rucc_sysroot::{Manifest, Sysroot};
58use rucc_target::{ObjectFormat, Triple};
59use rucc_tuple::TargetTuple;
60
61use crate::link::LinkOptions;
62
63pub use crate::assemble::assemble;
64pub use crate::compile::{Artifact, Compiled, Temps, compile, compile_ir};
65pub use crate::phase::{ArchiveJob, Input, InputKind, Job, LinkJob, Output, Phase, Plan, Role};
66pub use crate::preprocess::{OsFileSystem, Preprocessed, preprocess};
67pub use crate::schedule::Jobs;
68
69/// The compiler's version, taken from the workspace manifest.
70pub const VERSION: &str = env!("CARGO_PKG_VERSION");
71
72/// What the command line asked for.
73#[derive(Debug, Clone, PartialEq, Eq)]
74pub enum Action {
75    /// Print usage and exit successfully.
76    Help,
77    /// Print the version and exit successfully.
78    Version,
79    /// Print one line and exit successfully, which is what the `-dump` and `-print` family do.
80    ///
81    /// A build system asks these before it compiles anything, and what it does with the answer
82    /// is paste it into a path or into another command line, so each one is a single line with
83    /// no decoration around it.
84    Print(String),
85    /// Print the resolved configuration and exit successfully.
86    PrintConfig(Box<Options>),
87    /// Print the passes the level will run and exit successfully.
88    PrintPipeline(Box<Options>),
89    /// Print the phase plan and the link line and exit successfully, which is `-###`.
90    PrintPlan {
91        /// The resolved options, which is what says what the link line is for.
92        opts: Box<Options>,
93        /// What to do to each input, and in what order.
94        plan: Box<Plan>,
95        /// What the command line said about linking.
96        link: Box<LinkOptions>,
97    },
98    /// `--fetch <tuple>`, which gets the sysroot this release pins for a target and installs it.
99    ///
100    /// The only action in this compiler that may run another program to move bytes onto the
101    /// machine, which is `spec/cross-compile/13-distribution.md` section 13.8's rule rather than a
102    /// property of how this happens to be written: a compilation has no branch that reaches it.
103    Fetch {
104        /// The artifact, from the table in [`rucc_sysroot::artifact`]. Resolved here rather than where the
105        /// work happens, so that a target nothing is pinned for is a refusal from the parser like
106        /// every other thing a command line can ask for and not have.
107        what: &'static rucc_sysroot::Pinned,
108        /// The target, which names the directory under the cache the tree is installed at and is
109        /// checked against the record inside the artifact.
110        target: TargetTuple,
111        /// Where the cache is, read where everything else that needs it reads it.
112        cache: PathBuf,
113    },
114    /// Compile the given inputs.
115    Compile {
116        /// The resolved options.
117        opts: Box<Options>,
118        /// What to do to each input, and in what order.
119        plan: Box<Plan>,
120        /// What the command line said about linking.
121        link: Box<LinkOptions>,
122        /// How many translation units to compile at once.
123        jobs: Jobs,
124        /// Whether `-v` asked for the plan to be printed while it runs.
125        verbose: bool,
126        /// What is worth saying about the command line before anything is compiled, printed as
127        /// warnings and once for the whole run rather than once per file.
128        ///
129        /// These are not diagnostics. A diagnostic is about a piece of source and has a span to
130        /// point at, and these are about the way two flags were combined, so there is nothing to
131        /// point at and nowhere below the driver that knows both halves. `-w` does not reach them
132        /// for the same reason it does not reach a refusal from the parser.
133        notes: Vec<String>,
134    },
135}
136
137/// Why a command line was rejected.
138#[derive(Debug, Clone, PartialEq, Eq)]
139pub struct CliError {
140    /// The message, lowercase and without a trailing period, in the same shape as any other
141    /// diagnostic.
142    pub message: String,
143}
144
145impl std::fmt::Display for CliError {
146    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
147        f.write_str(&self.message)
148    }
149}
150
151impl std::error::Error for CliError {}
152
153fn err(message: impl Into<String>) -> CliError {
154    CliError { message: message.into() }
155}
156
157/// The two halves of one prefix mapping flag's argument, where `flag` includes its trailing `=`.
158///
159/// The split is at the last `=` in what follows the flag, not the first, which is gcc's rule and
160/// the only one that lets a directory whose name contains an `=` be the old half. It also means
161/// `-fmacro-prefix-map=a=b=c` rewrites `a=b` to `c` rather than `a` to `b=c`, which looks like a
162/// trap until you notice the alternative traps the far more common case.
163fn rewrite<'a>(arg: &'a str, flag: &str) -> Result<(&'a str, &'a str), CliError> {
164    let rest = &arg[flag.len()..];
165    PrefixMap::split(rest).ok_or_else(|| {
166        let flag = flag.trim_end_matches('=');
167        err(format!(
168            "`{rest}` is not a rewrite for `{flag}`, which is an old prefix, an `=` and a new one"
169        ))
170    })
171}
172
173/// A question the command line asked instead of asking for a compilation.
174///
175/// These are answered after the loop rather than where they are read, because every one of them
176/// is about the target or about the library search and the last word on both is the end of the
177/// command line.
178enum Query {
179    /// `-dumpmachine`, the triple.
180    Machine,
181    /// `-dumpversion` and `-dumpfullversion`, which are the same three numbers here.
182    Version,
183    /// `-print-multiarch`, the directory name a distribution files this target under.
184    Multiarch,
185    /// `-print-search-dirs`, in the three lines GCC prints.
186    SearchDirs,
187    /// `-print-sysroot`, the root the headers and the libraries are read under.
188    Sysroot,
189    /// `-print-sysroot-provenance`, what is in that root and where each of it came from.
190    SysrootProvenance,
191    /// `-print-sysroot-digest`, the one number that names all of it.
192    SysrootDigest,
193    /// `-print-file-name=<name>`, the full path of a library file.
194    FileName(String),
195    /// `-print-prog-name=<name>`, the full path of a program.
196    ProgName(String),
197    /// `-print-libgcc-file-name`, which is `-print-file-name=libgcc.a` under another spelling.
198    Libgcc,
199}
200
201/// Usage text.
202///
203/// Deliberately short. `spec/04-driver-and-cli.md` puts the full flag reference in the
204/// manual page, because a `--help` nobody can read in one screen is a `--help` nobody reads.
205pub const USAGE: &str = "\
206rucc, an optimizing C compiler
207
208usage: rucc [options] file...
209
210options:
211  -c                     compile and assemble, do not link
212  -S                     compile only, emit assembly
213  -E                     preprocess only
214  -o <file>              write output to <file>, or to standard output for -
215  -D <name>[=<value>], -U <name>      define a macro, or undefine one after every -D
216  -I <dir>               add <dir> to the include search path
217  -iquote -isystem -idirafter <dir>   the other chains, -nostdinc drops ours
218  -I-, -iprefix <p>, -iwithprefix[before] <dir>   the older spellings of those
219  -include <file>, -imacros <file>    read <file> first, the second for its macros only
220  --sysroot=<dir>        look for the library's headers under <dir>, -isysroot too
221  -P, -dM                with -E: leave out the markers, or dump the macros
222  -M -MM -MD -MMD        write a make rule for the source, the last two compile as well
223  -MF <file> -MT <t> -MQ <t> -MP   where the rule goes, what it builds, targets with no recipe
224  -std=<dialect>         c89 through c23, and the gnu spellings
225  -fgnuc-version=<v>     the GCC release to claim, default 7.0.0
226  -x <lang>              treat later inputs as <lang>, or none to stop
227  -O<level>              optimize: 0, 1, 2, 3, s, z
228  -fsafety=<tier>        check memory safety: off, detect, enforce, kernel
229  -f[no-]sanitize=<what>   the negative is taken, the positive is refused by name
230  -f[no-]safety-subobject   a write has to stay inside the member it names
231  -f[no-]safety-restrict    two restrict pointers of one block may not meet
232  -f<pass> -fno-<pass> -fdump-ir=<what> -fopt-info[-<kind>][=FILE]
233  -fpass-fuel=<pass>=<n>, -fpass-fuel-global=<n>   stop a pass, or all of them, after n
234  -fdisable-<pass>[=<funcs>], -fenable-<pass>[=<funcs>]   run a pass on some functions only
235  -g -g0 -gdwarf-5, -fno-omit-frame-pointer, -mno-red-zone   debug info, frame pointer, red zone
236  -gz[=none|zlib|zlib-gnu|zstd] -gno-split-dwarf   compress debug sections, one file not two
237  -flto[=auto|jobserver|<n>] -fno-lto -ffat-lto-objects   read, and not done yet
238  -fprofile-use[=<path>] -fprofile-dir=<dir>   read too, where -fprofile-generate is refused
239  -f[no-]stack-protector[-strong|-all], -f[no-]stack-clash-protection, -fcf-protection=<edges>
240  -ffunction-sections -fdata-sections   a section per function or variable, for --gc-sections
241  -fvisibility=<what>    default, hidden, internal or protected, when nothing in the source said
242  -l<name>, -L <dir>, -B <dir>   link a library, where to look for one, where our own tools are
243  -fPIC -fpic -fPIE -fpie, -fno-common, -pipe   what it does anyway
244  -f[no-]strict-aliasing, -f[no-]delete-null-pointer-checks   what it assumes anyway
245  -static -shared -pie -no-pie -nostdlib -nostartfiles -nodefaultlibs -rdynamic -s   how to link
246  -Wl,<arg>, -Xlinker <arg>, -fuse-ld=<name>   hand an argument to the linker, or pick one
247  -Werror -pedantic -pedantic-errors -w -W[no-]system-headers   how much to say, and how fatal
248  -m64 -march= -mtune= -mcpu= -mabi= -mcmodel=   what machine to generate for
249  -pg -p, -mfentry -mno-fentry   call a profiler on the way in, and where that call goes
250  -fpatchable-function-entry=<n>[,<m>]   room at the top of every function to patch later
251  -fwrapv, -fwrapv-pointer, -fno-strict-overflow   signed or pointer overflow wraps
252  -ftrapv                signed overflow stops the program instead
253  -f[no-]signed-char, -f[no-]unsigned-char, -f[no-]short-enums   change the ABI
254  -ffp-contract=<how>    fuse a multiply and an addition: fast, on or off
255  -fexcess-precision=<how>, -f[no-]rounding-math, -f[no-]trapping-math   what may be folded
256  -ffile-prefix-map=<old>=<new>   rewrite that front of every path we put in the output
257  -fmacro-prefix-map= -fdebug-prefix-map= -fprofile-prefix-map=   the same, one output each
258  -pthread               build for more than one thread, and link the library for it
259  -dumpmachine -dumpversion -print-multiarch -print-search-dirs   what this compiler is
260  -print-file-name=<name> -print-prog-name=<name>   where a file or a program is
261  -print-sysroot         the root the headers and the libraries are read under
262  -print-sysroot-provenance   every input under it, where it came from and its licence
263  -print-sysroot-digest   the sha256 of that record, which names the whole sysroot in one line
264  --fetch <tuple>        get the sysroot this release pins for <tuple> and install it in the cache
265  --offline              never download anything, which a compilation never does anyway
266  -j[n]                  compile n translation units at once, default all
267  -v, -###               print each phase as it runs, or without running any
268  -save-temps[=cwd|obj], -time   keep the .i and the .s, say how long each step took
269  --target=<triple>      generate code for <triple>
270  --emit=<kind>          exe, obj, archive, asm, preprocessed, tast, ir, mir-final,
271                         safety-summary, type-granules
272  --print-config, --print-pipeline    print the configuration or the pipeline, and exit
273  --version              print the version and exit
274  -h, --help             print this message and exit
275
276See spec/04-driver-and-cli.md for the full flag reference.
277";
278
279/// The argument of a flag that may be joined to it or may be the next word.
280///
281/// `-DFOO` and `-D FOO` are the same thing, and `at` is where the flag's own letters end.
282fn joined_or_next(
283    arg: &str,
284    at: usize,
285    args: &[String],
286    i: &mut usize,
287) -> Result<String, CliError> {
288    if arg.len() > at {
289        return Ok(arg[at..].to_owned());
290    }
291    let next = args.get(*i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
292    *i += 1;
293    Ok(next.clone())
294}
295
296/// The smallest boundary a function is put on when the command line asked for no alignment at all.
297///
298/// Eight bytes, which is what gcc 16 gives `-fno-align-functions` on x86-64 and is a boundary every
299/// target this compiler has is happy with. It is not zero: a function still has to start somewhere
300/// an instruction may start, and the flag asks for the target's minimum rather than for none.
301const MIN_FUNC_ALIGN: u32 = 8;
302
303/// What `-falign-functions=N` asks for, as a power of two, or `None` for the target's own answer.
304///
305/// Zero and one both mean the default, which is gcc's reading of them, and everything else is
306/// rounded up to the next power of two, which is also gcc's: `-falign-functions=3` puts a function
307/// on a four byte boundary rather than being refused. Gives back `Err` shaped as an outer `None`
308/// only when the text is not a number, since that is the one thing gcc will not read either. A
309/// number larger than any alignment makes sense at is clamped rather than refused, for the same
310/// reason: this is a preference about speed and a build that wrote a silly one still deserves to
311/// compile.
312fn function_alignment(text: &str) -> Option<Option<u32>> {
313    // gcc takes `N:M:N2:M2`, where everything after the first number is about how far it is willing
314    // to go to reach the boundary. Only the boundary is answerable here, so the rest is read to
315    // check that it is numbers and then dropped.
316    let mut parts = text.split(':');
317    let first = parts.next()?;
318    if parts.any(|part| part.parse::<u64>().is_err()) {
319        return None;
320    }
321    let want: u64 = first.parse().ok()?;
322    if want <= 1 {
323        return Some(None);
324    }
325    let bytes = want.min(1 << 16).next_power_of_two();
326    Some(Some(u32::try_from(bytes).ok()?))
327}
328
329/// Every name that may follow `-fsanitize=`, which is gcc 16's list and three of this compiler's
330/// own.
331///
332/// The three are on it because `spec/07-types-and-semantics.md` section 7.7 already promises them:
333/// each undefined behaviour this compiler exploits is listed there with the check that detects it,
334/// and `alias`, `restrict` and `memory` are checks gcc has no spelling for. gcc refuses `memory`
335/// outright, since the sanitizer of that name is clang's. A name being here means it is a name
336/// rather than a typo, and nothing more than that: every one of them is refused after the loop,
337/// because none of them is implemented.
338///
339/// `all` is deliberately absent. gcc takes it only in the negative, so it is handled where each of
340/// those two spellings is read rather than by being on this list.
341const SANITIZERS: [&str; 34] = [
342    "address",
343    "kernel-address",
344    "hwaddress",
345    "kernel-hwaddress",
346    "pointer-compare",
347    "pointer-subtract",
348    "thread",
349    "leak",
350    "undefined",
351    "shift",
352    "shift-base",
353    "shift-exponent",
354    "integer-divide-by-zero",
355    "unreachable",
356    "vla-bound",
357    "null",
358    "return",
359    "signed-integer-overflow",
360    "bounds",
361    "bounds-strict",
362    "alignment",
363    "object-size",
364    "float-divide-by-zero",
365    "float-cast-overflow",
366    "nonnull-attribute",
367    "returns-nonnull-attribute",
368    "bool",
369    "enum",
370    "vptr",
371    "pointer-overflow",
372    "builtin",
373    "alias",
374    "restrict",
375    "memory",
376];
377
378/// Parses a command line, without the program name.
379///
380/// # Errors
381///
382/// Returns the message to print when the arguments do not name a compilation this compiler
383/// can attempt.
384pub fn parse_args(args: &[String]) -> Result<Action, CliError> {
385    let host = Triple::host()
386        .ok_or_else(|| err("this host is not a supported target and no --target was given"))?;
387    let mut opts = Options::new(host);
388    // Where the compiler is running, which is what `DW_AT_comp_dir` is and what a debugger joins a
389    // relative file name onto. Asked here rather than where the debug sections are written, because
390    // this is the one layer that is allowed to look at the process it is in, and because a command
391    // line that compiles four files should give the same answer for all four.
392    opts.working_dir = std::env::current_dir().ok().map(|dir| dir.to_string_lossy().into_owned());
393    let mut inputs: Vec<Input> = Vec::new();
394    let mut print_config = false;
395    let mut print_pipeline = false;
396    let mut print_plan = false;
397    let mut verbose = false;
398    let mut jobs = Jobs::default();
399    let mut nostdinc = false;
400    let mut sysroot: Option<PathBuf> = None;
401    // What the command line is worth warning about, filled in after the loop rather than during it,
402    // because every question of this kind is about two flags and the last word on both of them is
403    // the end of the loop.
404    let mut notes: Vec<String> = Vec::new();
405    // The whole ten field target, kept beside the three field one because `--target=` can pin a
406    // libc version and `Triple` has nowhere to put it. It decides `__GLIBC_MINOR__` and nothing
407    // else today, and `None` is a command line that named no target, which is this machine.
408    let mut pinned: Option<TargetTuple> = None;
409    let mut output = None;
410    let mut link = LinkOptions::default();
411    let mut query: Option<Query> = None;
412    // What `--fetch` named, and whether `--offline` forbade it. Both are weighed after the loop
413    // because either can be written after the other.
414    let mut fetch: Option<String> = None;
415    let mut offline = false;
416    let mut threads = false;
417    // Which sanitizers are still asked for by the end of the command line. Accumulated across the
418    // loop rather than answered where it was read, because `-fno-sanitize=` turns one off and a
419    // build that asks for a check and then takes it back has asked for nothing. What happens to a
420    // set that is not empty is decided after the loop.
421    let mut sanitizers: Vec<&str> = Vec::new();
422    // `-x` applies to inputs that come after it and stays in effect until the next one, which
423    // is why it is tracked across the loop rather than attached to a single argument.
424    let mut forced: Option<InputKind> = None;
425    // What `-iprefix` last said, stuck on the front of every later `-iwithprefix`. It applies to
426    // the flags after it and not the ones before, so a command line may set it more than once.
427    // GCC's default is its own installed header directory with the last component taken off,
428    // which is a path a cross compiler's build system knows and passes; there is no equivalent
429    // here, so with no `-iprefix` the prefix is nothing and `-iwithprefix` names a directory
430    // outright.
431    let mut iprefix = String::new();
432
433    let mut i = 0;
434    while i < args.len() {
435        let arg = args[i].as_str();
436        i += 1;
437        match arg {
438            "-h" | "--help" => return Ok(Action::Help),
439            "--version" => return Ok(Action::Version),
440            // The sysroot fetch, which is weighed after the loop rather than acted on here, because
441            // `--offline` written after it has to be able to forbid it. Both spellings, since a
442            // flag that takes a tuple gets written both ways and neither is a guess at what the
443            // other meant.
444            "--fetch" => {
445                let value = args
446                    .get(i)
447                    .ok_or_else(|| err("--fetch requires the target to get a sysroot for"))?;
448                i += 1;
449                fetch = Some(value.clone());
450            }
451            _ if arg.starts_with("--fetch=") => {
452                fetch = Some(arg["--fetch=".len()..].to_owned());
453            }
454            // Accepted on any command line and only ever read by the fetch, because an ordinary
455            // compile downloads nothing with or without it. So this flag takes nothing away today,
456            // which is the property section 13.2 asks for rather than an omission: a build that
457            // passes it is saying what it expects of this compiler, and what it expects is already
458            // true.
459            "--offline" => offline = true,
460            "--print-config" => print_config = true,
461            "--print-pipeline" => print_pipeline = true,
462            "-###" => print_plan = true,
463            "-v" => verbose = true,
464            // The files a compilation goes through, kept rather than thrown away. The bare
465            // spelling means `=obj` and not `=cwd`, which is not what the manual says and is what
466            // gcc 16 does; `SaveTemps::Object` carries the measurement.
467            "-save-temps" => opts.save_temps = SaveTemps::Object,
468            _ if arg.starts_with("-save-temps=") => {
469                opts.save_temps = arg["-save-temps=".len()..].parse().map_err(err)?;
470            }
471            // How long each step took. A misspelling of this is worth rejecting rather than
472            // ignoring, since a run that says nothing looks like a compilation that took no time.
473            "-time" => opts.time = true,
474            "-c" => opts.emit = EmitKind::Object,
475            "-S" => opts.emit = EmitKind::Asm,
476            "-E" => opts.emit = EmitKind::Preprocessed,
477            "-g" => opts.debug_info = true,
478            // GCC's own levels of how much debug information to write. Zero is none and every
479            // other number is some, and this compiler has one amount, so the numbers above zero
480            // all mean the same thing here. `-ggdb` is the same flag asking for whatever the
481            // debugger on the machine prefers, which is what we emit anyway.
482            "-g0" => opts.debug_info = false,
483            "-g1" | "-g2" | "-g3" | "-ggdb" | "-ggdb1" | "-ggdb2" | "-ggdb3" => {
484                opts.debug_info = true;
485            }
486            // The version of DWARF to write. We write DWARF 5 and nothing else, so a build that
487            // asks for another version is told rather than handed a file it cannot read.
488            "-gdwarf" | "-gdwarf-5" => opts.debug_info = true,
489            _ if arg.starts_with("-gdwarf-") => {
490                return Err(err(format!(
491                    "{arg}: this compiler writes DWARF 5 and no other version, see \
492                     spec/11-debug-info.md"
493                )));
494            }
495            // Whether the debug information goes in a file of its own beside the object. gcc
496            // writes that `.dwo` whether or not it found anything to put in it, which means a
497            // build system that declares the file as an output gets one and a make rule that
498            // depends on it fires. Refused for that reason rather than taken: section 4.1 takes a
499            // flag that changes nothing and refuses one that changes what is produced, and a file
500            // that does not appear is the plainest change of that kind there is. The negative
501            // spelling is taken, because putting it all in the object is what happens anyway.
502            "-gno-split-dwarf" => {}
503            "-gsplit-dwarf" => {
504                return Err(err(format!(
505                    "{arg}: this compiler writes no separate `.dwo` file, and a build that \
506                     expects one beside each object would wait for a file that never arrives, \
507                     see spec/11-debug-info.md"
508                )));
509            }
510            // How the debug sections are compressed. There are none yet, so every answer produces
511            // the same bytes and taking the flag promises nothing that is not kept. The value is
512            // still checked, because a typo in a distribution's flags is worth finding when the
513            // compiler reads it rather than when somebody later wonders why nothing got smaller.
514            // Bare `-gz` means `zlib`, which the manual leaves for the reader to discover.
515            "-gz" => opts.compress = Compress::Zlib,
516            _ if arg.starts_with("-gz=") => {
517                let how = &arg["-gz=".len()..];
518                opts.compress = how.parse().map_err(|()| {
519                    err(format!(
520                        "`{how}` is not a way to compress debug sections, which is none, zlib, \
521                         zlib-gnu or zstd"
522                    ))
523                })?;
524            }
525            "-Werror" => opts.warnings_are_errors = true,
526            // Nothing that is not fatal is said at all. Read at the one place a diagnostic goes
527            // through rather than here, so that a warning `-w` dropped is not counted either.
528            "-w" => opts.warnings = false,
529            // Off by default, the way gcc has it off. A header that came with the machine is not
530            // one the person compiling can change, so a warning about it is noise, and under
531            // `-Werror` it is a build that stops on a line nobody in the project wrote. Somebody
532            // porting a header does want to hear all of it, which is what the flag is for.
533            "-Wsystem-headers" => opts.system_header_warnings = true,
534            "-Wno-system-headers" => opts.system_header_warnings = false,
535            "-pedantic-errors" => {
536                opts.pedantic = true;
537                opts.warnings_are_errors = true;
538            }
539            "-P" => opts.line_markers = false,
540            // The dependency family, which section 4.4 calls required because every build system
541            // that generates its own makefiles asks for it. The two that end in `D` write a file
542            // beside the object and let the compilation happen, and the two that do not write to
543            // standard output and stop after it. Nothing here turns the system headers back on
544            // once a flag has turned them off, which is GCC's behaviour and is why `-MM -M` is
545            // `-MM`: the flag asking for fewer of them is the one with something to say.
546            "-M" => {
547                opts.deps.emit = true;
548                opts.deps.instead_of_compiling = true;
549            }
550            "-MM" => {
551                opts.deps.emit = true;
552                opts.deps.instead_of_compiling = true;
553                opts.deps.system_headers = false;
554            }
555            "-MD" => opts.deps.emit = true,
556            "-MMD" => {
557                opts.deps.emit = true;
558                opts.deps.system_headers = false;
559            }
560            "-MP" => opts.deps.phony = true,
561            // These three take a word and only in the separated form, which is how GCC spells
562            // them and how every build system writes them.
563            "-MF" | "-MT" | "-MQ" => {
564                let value =
565                    args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
566                i += 1;
567                match arg {
568                    "-MF" => opts.deps.file = Some(value.clone()),
569                    // The whole of the difference between the two. `-MT` is for a build that has
570                    // already escaped what it is passing, and `-MQ` is for one that has a name
571                    // and wants it to arrive as that name.
572                    "-MT" => opts.deps.targets.push(value.clone()),
573                    _ => opts.deps.targets.push(deps::escaped(value)),
574                }
575            }
576            // The questions a build system asks before it compiles anything. Answered after the
577            // loop, because each one is about the target or the library search and the command
578            // line has not finished saying what those are.
579            "-dumpmachine" => query = Some(Query::Machine),
580            "-dumpversion" | "-dumpfullversion" => query = Some(Query::Version),
581            "-print-multiarch" => query = Some(Query::Multiarch),
582            "-print-search-dirs" => query = Some(Query::SearchDirs),
583            "-print-sysroot" => query = Some(Query::Sysroot),
584            // Both spellings, because this one is ours rather than GCC's and our own documents
585            // write it both ways: section 13.5 of `spec/cross-compile/13-distribution.md` gives it
586            // two dashes like the other flags we invented, and document 12's table gives it one
587            // like the `-print-` family it sits in. A person who reads either and types what it
588            // says is right, so neither is refused.
589            "-print-sysroot-provenance" | "--print-sysroot-provenance" => {
590                query = Some(Query::SysrootProvenance);
591            }
592            "-print-sysroot-digest" | "--print-sysroot-digest" => {
593                query = Some(Query::SysrootDigest);
594            }
595            "-print-libgcc-file-name" => query = Some(Query::Libgcc),
596            _ if arg.starts_with("-print-file-name=") => {
597                query = Some(Query::FileName(arg["-print-file-name=".len()..].to_owned()));
598            }
599            _ if arg.starts_with("-print-prog-name=") => {
600                query = Some(Query::ProgName(arg["-print-prog-name=".len()..].to_owned()));
601            }
602            // A program built to run in more than one thread. On every platform this compiler
603            // targets that is a macro the library's headers read and one more library on the
604            // link line, and the library is added after the loop so that it lands after the
605            // objects that refer to it.
606            "-pthread" | "-pthreads" => {
607                opts.defines.push("_REENTRANT".to_owned());
608                threads = true;
609            }
610            "-ansi" => {
611                opts.std = Std::C89;
612                opts.gnu_extensions = false;
613            }
614            // `-Wpedantic` is the same flag under the name the `-W` family gives it, which is
615            // the spelling a build system that groups its warning flags tends to write.
616            "-pedantic" | "-Wpedantic" => opts.pedantic = true,
617            // Both directions, because a build that needs this for one directory turns it back
618            // off for the next one rather than leaving it on for the whole tree.
619            "-fpermissive" => opts.permissive = true,
620            "-fno-permissive" => opts.permissive = false,
621            "-ffreestanding" => opts.hosted = false,
622            "-fhosted" => opts.hosted = true,
623            "-fno-builtin" => opts.builtins = false,
624            "-fbuiltin" => opts.builtins = true,
625            // The C89 dialects are under GNU's reading whatever this says, so turning it off
626            // there is turning off something the dialect asked for, which is accepted and does
627            // nothing. gcc refuses that command line, and there is nothing it could have meant.
628            "-fgnu89-inline" => opts.gnu89_inline = true,
629            "-fno-gnu89-inline" => opts.gnu89_inline = false,
630            // Both directions of each, because a build system that wants one of these usually
631            // writes it beside the flag that turns it back off for one directory.
632            "-fno-omit-frame-pointer" => opts.frame_pointer = true,
633            "-fomit-frame-pointer" => opts.frame_pointer = false,
634            // Both directions again, for the same reason, and a third answer for a command line
635            // that wrote neither: see `reorder_blocks` in `rucc_session`.
636            "-freorder-blocks" => opts.reorder_blocks = Some(true),
637            "-fno-reorder-blocks" => opts.reorder_blocks = Some(false),
638            // gcc's name for the scheduler that runs after the registers are handed out, which is
639            // the only one rucc has: see `schedule_insns` in `rucc_session`. gcc also takes
640            // `-fschedule-insns` for the pass before allocation, and taking that one here would be
641            // a flag that says a pass ran when none did.
642            "-fschedule-insns2" => opts.schedule_insns = Some(true),
643            "-fno-schedule-insns2" => opts.schedule_insns = Some(false),
644            "-mno-red-zone" => opts.red_zone = false,
645            "-mred-zone" => opts.red_zone = true,
646            // Four flags rather than one with an argument, which is how gcc spells them and how
647            // every build line writes them. Last one wins, because a package build puts
648            // `-fstack-protector-strong` in its global flags and a directory that cannot have one
649            // turns it back off on the line after.
650            "-fno-stack-protector" | "-fno-stack-protector-all" | "-fno-stack-protector-strong" => {
651                opts.protector = Protector::None;
652            }
653            "-fstack-protector" => opts.protector = Protector::Buffers,
654            "-fstack-protector-strong" => opts.protector = Protector::Strong,
655            "-fstack-protector-all" => opts.protector = Protector::All,
656            // The other half of what a hardened build asks for, and it is a question about the
657            // frame rather than about the function, so it is a switch rather than a level.
658            "-fstack-clash-protection" => opts.stack_clash = true,
659            "-fno-stack-clash-protection" => opts.stack_clash = false,
660            // The third of them, and the one that is a question with an argument rather than a
661            // family of spellings, because what it asks about is which of the two edges of a
662            // control flow transfer is checked. Bare is both of them, which is what gcc does.
663            "-fcf-protection" => opts.control = Control::Full,
664            "-fno-cf-protection" => opts.control = Control::None,
665            // Two spellings of the same request, which is what gcc has as well. `-p` was the older
666            // profiler and `-pg` the one that also recorded who called whom, and on every platform
667            // this compiler targets there is now one hook and both ask for it.
668            "-pg" | "-p" => {
669                opts.profile = true;
670                link.profile = true;
671            }
672            // Accepted on their own and doing nothing on their own, which is gcc's behaviour: they
673            // say where the call goes and a command line that asked for no call has nowhere to put
674            // one. That matters because a build system that sets `-mfentry` globally and `-pg` per
675            // directory is a build system that would otherwise fail on every other directory.
676            "-mfentry" => opts.hook = Hook::Early,
677            "-mno-fentry" => opts.hook = Hook::Late,
678            // GCC drops its own include directory along with the system ones, because its
679            // headers are half of a pair with the library's and half a pair is worse than
680            // none. A build that passes this is supplying the whole set itself.
681            "-nostdinc" => nostdinc = true,
682            "-o" => {
683                output = Some(args.get(i).ok_or_else(|| err("-o requires an argument"))?.clone());
684                i += 1;
685            }
686            // The flags that take a directory only in the separated form. GCC spells them
687            // this way and nothing writes `-iquotedir`, so accepting the joined form would
688            // mean guessing at a path that starts with the flag's own letters.
689            // Apple's spelling of `--sysroot`, and the one its own build systems pass. The
690            // two mean the same thing here: the configured directories are under there rather
691            // than under the root.
692            "-isysroot" => {
693                let dir = args.get(i).ok_or_else(|| err("-isysroot requires an argument"))?;
694                i += 1;
695                sysroot = Some(PathBuf::from(dir));
696            }
697            "-iquote" | "-isystem" | "-idirafter" => {
698                let dir = args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
699                i += 1;
700                match arg {
701                    "-iquote" => opts.search.push_quote(dir.clone()),
702                    "-isystem" => opts.search.push_system(dir.clone()),
703                    _ => opts.search.push_after(dir.clone()),
704                }
705            }
706            "-iprefix" => {
707                iprefix = args.get(i).ok_or_else(|| err("-iprefix requires an argument"))?.clone();
708                i += 1;
709            }
710            // Where GCC puts these is not where its manual says it puts them, and this is the
711            // measured answer rather than the documented one: `-iwithprefix` lands in the
712            // `-isystem` slot and not the `-idirafter` slot, and `-iwithprefixbefore` lands in
713            // the `-I` slot. A cross build that uses them is relying on the behaviour, since
714            // that is the compiler it was developed against.
715            "-iwithprefix" | "-iwithprefixbefore" => {
716                let dir = args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
717                i += 1;
718                let dir = format!("{iprefix}{dir}");
719                if arg == "-iwithprefix" {
720                    opts.search.push_system(dir);
721                } else {
722                    opts.search.push_bracket(dir);
723                }
724            }
725            "-include" | "-imacros" => {
726                let name = args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
727                i += 1;
728                opts.preincludes
729                    .push(Preinclude { name: name.clone(), macros_only: arg == "-imacros" });
730            }
731            // The flag `-iquote` was introduced to replace, still passed by build systems old
732            // enough to predate the replacement. It is not a directory: it says that every `-I`
733            // so far is for quoted includes only, and that a quoted include stops looking next
734            // to the file that wrote it.
735            "-I-" => opts.search.split_quote_chain(),
736            "-x" => {
737                let lang = args.get(i).ok_or_else(|| err("-x requires an argument"))?;
738                i += 1;
739                forced = if lang == "none" {
740                    None
741                } else {
742                    Some(InputKind::from_x_arg(lang).map_err(|e| err(format!("{e}")))?)
743                };
744            }
745            // Not a GCC flag. spec/03-architecture.md section 3.5 compiles several
746            // translation units in one process rather than making the build system fork, and
747            // section 3.8's determinism check compares `-j1` against `-j16`, so the knob has
748            // to exist and has to be spelled the way `make` spells it.
749            // `-DFOO`, `-D FOO` and the same for `-U` and `-I`. Both forms are in wide use
750            // and a build system may produce either, so both are read here rather than
751            // being normalised by whatever generated the command line.
752            _ if arg.starts_with("-D") => {
753                let value = joined_or_next(arg, 2, args, &mut i)?;
754                opts.defines.push(value);
755            }
756            _ if arg.starts_with("-U") => {
757                let value = joined_or_next(arg, 2, args, &mut i)?;
758                opts.undefines.push(value);
759            }
760            _ if arg.starts_with("-I") => {
761                let dir = joined_or_next(arg, 2, args, &mut i)?;
762                opts.search.push_bracket(dir);
763            }
764            _ if arg.starts_with("-std=") => {
765                let name = &arg["-std=".len()..];
766                let (std, gnu) = Std::from_flag(name)
767                    .ok_or_else(|| err(format!("unknown dialect `{name}`, see --help")))?;
768                opts.std = std;
769                opts.gnu_extensions = gnu;
770            }
771            // Section 4.5. The claim decides which half of glibc's `sys/cdefs.h` we are
772            // handed, so a differential run that does not set it is comparing two compilers
773            // that believe they are different compilers.
774            // GCC packs these into one flag, so `-dDI` is two of them. Letters in the family
775            // that we have not written yet are accepted and ignored, because a dump is a
776            // debugging aid and a build that asks for one should still compile. A letter
777            // outside the family falls through to the unknown option error, which is what
778            // keeps `-dumpversion` from being read as a dump of nothing.
779            _ if Dumps::is_family(arg) => {
780                opts.dumps.add(&arg[2..]);
781            }
782            // One name at a time, which is what a build that means its own `memcpy` and the
783            // library's everything else writes. The name is not checked against a list, because
784            // the flag is about what the program means by a name and a program is allowed to mean
785            // something by a name this compiler has never heard of.
786            _ if arg.starts_with("-fno-builtin-") => {
787                opts.no_builtin.push(arg["-fno-builtin-".len()..].to_owned());
788            }
789            _ if arg.starts_with("-fgnuc-version=") => {
790                let v = &arg["-fgnuc-version=".len()..];
791                opts.gnuc = v.parse().map_err(err)?;
792            }
793            // spec/13-gnu-compat.md section 13.3 promises this flag an error that says why rather
794            // than the unknown option one, because a build reaching for it is asking for a feature
795            // and deserves to be told it is not coming rather than told the spelling is wrong.
796            // The negative form is what this compiler does anyway, so it is taken and dropped.
797            "-fnested-functions" => {
798                return Err(err(
799                    "nested functions are not supported: a call to one goes through a trampoline \
800                     written on the stack, which no target that enforces an unexecutable stack \
801                     allows",
802                ));
803            }
804            "-fno-nested-functions" => {}
805            // Which of the two links the output is for, which is a real difference and not a
806            // description of what happens anyway. Everything here is position independent either
807            // way, and what these decide is whether a name may be one another object defines or
808            // replaces, because a link that produces an executable puts every name in the same
809            // program and a link that produces a shared library does not.
810            //
811            // It matters that they are accepted at all, whatever they then do. Every autoconf and
812            // cmake build puts `-fPIC` on the compile line, so a compiler that rejects it cannot
813            // be the `CC` of a project that has a configure script, whatever else it can do. That
814            // is how this was found: building SQLite's test fixture stopped on it.
815            "-fPIC" | "-fpic" => opts.pic = Pic::Library,
816            // Not a synonym of the pair above, which is what they were treated as until #756. The
817            // library is the expensive answer and gcc makes it the one that has to be asked for,
818            // so this is also what nothing at all means.
819            "-fPIE" | "-fpie" => opts.pic = Pic::Executable,
820            // A different question from the pair above, and the one every distribution build of a
821            // shared library answers. `-fPIC` decides how an address is reached, and this decides
822            // whether the optimizer may believe a body it can see, because an exported name is one
823            // the dynamic linker may find another definition of first. On by default, which is
824            // gcc's arrangement and is the honest answer, and off is a promise the build makes and
825            // nothing checks.
826            "-fsemantic-interposition" => opts.interposition = true,
827            "-fno-semantic-interposition" => opts.interposition = false,
828            // Two requests rather than one, and the same table answers both, so what decides is
829            // whether either of them is standing. gcc arranges it the same way: the asynchronous
830            // one is the default here and it implies the other, and a line that asks for a table
831            // and against an asynchronous one gets a table.
832            "-fasynchronous-unwind-tables" => opts.async_unwind_tables = true,
833            "-fno-asynchronous-unwind-tables" => opts.async_unwind_tables = false,
834            "-funwind-tables" => opts.unwind_tables = true,
835            "-fno-unwind-tables" => opts.unwind_tables = false,
836            // The other direction is a request, not a description, and it is one this compiler
837            // cannot grant, so it gets the treatment section 13.3 asks for rather than the unknown
838            // option error. Answering it by carrying on would be answering a different question:
839            // the code would still be position independent, which is correct everywhere an
840            // ordinary program runs and is wrong in a kernel, where the flag is written precisely
841            // because there is no loader to fill a global offset table in.
842            "-fno-pic" | "-fno-pie" => {
843                return Err(err(
844                    "position dependent code is not supported: an address that may be in another \
845                     object is loaded out of the global offset table, and nothing here emits the \
846                     absolute form this asks for. Use -no-pie if what you meant was how to link",
847                ));
848            }
849            // A section per function and a section per variable, which is what makes
850            // `--gc-sections` able to drop anything: a linker can leave out a section nothing
851            // reaches and cannot leave out half of one. Both directions are taken, and the off
852            // one is the default rather than a refusal, since a build that writes it is asking
853            // for what happens anyway.
854            "-ffunction-sections" => opts.function_sections = true,
855            "-fno-function-sections" => opts.function_sections = false,
856            "-fdata-sections" => opts.data_sections = true,
857            "-fno-data-sections" => opts.data_sections = false,
858            // Another description of what this compiler does. A file scope declaration with no
859            // initializer is written into `.bss` as an ordinary defined symbol, not offered to the
860            // linker as a common one for it to merge, which is what `-fno-common` asks for and what
861            // gcc has done by default since 10. Nothing in the front end produces `Linkage::Common`
862            // at all.
863            "-fno-common" => {}
864            // What overflows rather than being undefined. Every one of these takes something away
865            // from the optimizer rather than asking it to do anything, which is why the negative
866            // spellings are the interesting ones and the positive spellings are the default.
867            //
868            // `-fno-strict-overflow` is both of the others, which is gcc's own reading of it: its
869            // help text for `-fstrict-overflow` says "negated as -fwrapv -fwrapv-pointer". So it is
870            // written here as the pair rather than kept as a third thing to test everywhere.
871            //
872            // `-ftrapv` is the exception and is the one that asks for something. It is the other
873            // answer to the question `-fwrapv` answers, so the two cannot both hold and each clears
874            // the other, which makes the last one on the command line the one that counts. That is
875            // gcc 16's behaviour and was measured rather than read: `-ftrapv -fwrapv` emits no
876            // checked calls and `-fwrapv -ftrapv` emits them. The positive spelling of the pointer
877            // question is left alone by both, because neither has anything to say about it.
878            "-fwrapv" => {
879                opts.wrapping.signed = true;
880                opts.wrapping.trap = false;
881            }
882            "-fno-wrapv" => opts.wrapping.signed = false,
883            "-fwrapv-pointer" => opts.wrapping.pointer = true,
884            "-fno-wrapv-pointer" => opts.wrapping.pointer = false,
885            "-fno-strict-overflow" => opts.wrapping = Wrapping::ALL,
886            // Which does not clear the checked one, because gcc does not: `-ftrapv
887            // -fstrict-overflow` still emits the calls. It says what is assumed and not what
888            // happens.
889            "-fstrict-overflow" => {
890                opts.wrapping.signed = false;
891                opts.wrapping.pointer = false;
892            }
893            "-ftrapv" => {
894                opts.wrapping.trap = true;
895                opts.wrapping.signed = false;
896            }
897            "-fno-trapv" => opts.wrapping.trap = false,
898            // The two flags that say what a plain `char` is, which is one question with two
899            // spellings each: gcc reads `-fno-signed-char` as `-funsigned-char` and
900            // `-fno-unsigned-char` as `-fsigned-char`, so there are four ways to write two
901            // answers and the last one written wins. Nothing is set until one of them is given,
902            // because the target's own ABI is the answer otherwise and it is not the same answer
903            // everywhere: x86-64 and Apple's arm64 are signed, Linux's arm64 is not.
904            "-fsigned-char" | "-fno-unsigned-char" => opts.char_signed = Some(true),
905            "-funsigned-char" | "-fno-signed-char" => opts.char_signed = Some(false),
906            // And the size of an enumeration, which is the other thing in this group that changes
907            // the ABI rather than the code.
908            "-fshort-enums" => opts.short_enums = true,
909            "-fno-short-enums" => opts.short_enums = false,
910            // And Microsoft's reading of an anonymous member, which changes the layout of every
911            // record that writes a tag on one. Nothing is set until one of them is given, because
912            // the target is the answer otherwise: gcc's mingw build has this on and its Linux
913            // build has it off.
914            "-fms-extensions" => opts.ms_extensions = Some(true),
915            "-fno-ms-extensions" => opts.ms_extensions = Some(false),
916            // And the request, which is the one that cannot be granted. It is a real difference and
917            // not a preference: two files each writing `int g;` link under `-fcommon` and are a
918            // duplicate definition without it, which is the whole reason the flag survives.
919            "-fcommon" => {
920                return Err(err(
921                    "a tentative definition is written into .bss as its own symbol here, and \
922                     nothing emits the common symbol this asks the linker to merge. Give the \
923                     variable a definition in one file and declare it extern in the others",
924                ));
925            }
926            // Both directions of this one are recorded, and what they decide is whether lowering
927            // names the type each access goes through. Turning it off is the front end leaving the
928            // name off rather than a pass being told to ignore one it can see, which is one
929            // condition in one place, and it is the reading that survives link time optimization:
930            // a unit built with the flag off keeps its own answer when its bodies end up in a
931            // module beside bodies that were not.
932            //
933            // Nothing in the pipeline reads those names yet. Layer 3 of the alias analysis does
934            // and is tested, and no pass at any level asks the alias analysis anything today, so
935            // no program compiles differently for having passed this. The flag is wired anyway,
936            // because the change that makes a pass ask is not the change anybody will remember to
937            // wire it in, and a flag that is taken and dropped once the names mean something is
938            // the miscompilation `spec/04-driver-and-cli.md` section 4.1 warns about in as many
939            // words.
940            "-fstrict-aliasing" => opts.strict_aliasing = true,
941            "-fno-strict-aliasing" => opts.strict_aliasing = false,
942            // The same shape of answer for the same reason, and the flag the kernel writes beside
943            // the one above it.
944            //
945            // Nothing here concludes that a pointer is not null from the fact that it was
946            // dereferenced. There is no such conclusion to draw from, because no pass records one:
947            // a load says where it read and nothing else, and a comparison against null is an
948            // ordinary comparison of two values the optimizer has no fact about. So a function
949            // that reads through a pointer and then tests it keeps the test, which is what the
950            // kernel wants and what `-fno-delete-null-pointer-checks` asks for, and what gcc has
951            // to be asked for because it draws the conclusion by default.
952            //
953            // `-fdelete-null-pointer-checks` is the request to draw it, and it goes the way
954            // `-fstrict-aliasing` does: assuming less than was asked for costs speed and not
955            // correctness, and `-O2` implies it, so refusing it would stop builds for nothing.
956            "-fdelete-null-pointer-checks" | "-fno-delete-null-pointer-checks" => {}
957            // The floating point group, which goes the same way and for the same reason, and which
958            // is worth writing out because the reason is easy to get backwards.
959            //
960            // Each of these has a restrictive spelling and a permissive one. The restrictive ones,
961            // `-frounding-math` and `-ftrapping-math`, say that the rounding mode may have been
962            // changed and that an exception raised by an operation may be looked at, so an
963            // arithmetic the compiler folds at compile time is an arithmetic whose rounding and
964            // whose exception the program does not get. Nothing here folds any floating point
965            // arithmetic in a function body: `0.1 + 0.2` is an `fadd` and `1.0 / 0.0` is a divide
966            // that runs, at every level. So both of those describe what already happens.
967            //
968            // The permissive ones are the other half, and they are licences rather than requests
969            // for an answer. `-fno-rounding-math` says the rounding mode is the default one and
970            // `-fno-trapping-math` says nothing looks at the exceptions, which together are
971            // permission to fold. Not folding is the conservative side of that permission and is
972            // what a program is entitled to whichever was written, so `-fno-rounding-math` costs
973            // speed and not correctness, which is the test section 4.1 puts a licence through.
974            "-frounding-math" | "-fno-rounding-math" => {}
975            // `-fno-trapping-math` is the one of the four that is kept, because there is one
976            // conversion this compiler does not fold and gcc folds under it, and the two answers
977            // differ. Converting a constant floating value to an integer type it does not fit in
978            // is undefined behaviour rather than a value: left to the hardware it is one
979            // instruction and the answer is the integer indefinite value, and folded it is the
980            // nearest end of the integer's range. Both compilers leave it to the instruction by
981            // default and gcc folds it under this flag, so a program built with it and compiled
982            // without it gets a different number rather than a slower one. `-ftrapping-math` is
983            // gcc's default, so a build spelling it out is asking for what it already has.
984            "-ftrapping-math" => opts.trapping_math = true,
985            "-fno-trapping-math" => opts.trapping_math = false,
986            // About temporary files rather than about code. There is nothing between the phases of
987            // one compilation here to write to a file in the first place.
988            "-pipe" => {}
989            // Nothing here writes colour, so all of these are the same answer, and it is the answer
990            // that costs nothing: the diagnostics come out plain either way and no build depends on
991            // an escape sequence being there. Taken rather than refused because cmake writes
992            // `-fdiagnostics-color=always` on every compile line when the generator is ninja, which
993            // makes this the second most common flag after `-fPIC` to stop a build over a question
994            // about how the text looks.
995            "-fdiagnostics-color" | "-fno-diagnostics-color" => {}
996            _ if arg.starts_with("-fdiagnostics-color=") => {}
997            // The link flags. None of them changes the compilation, which is why they are
998            // collected apart from `opts` and why `-lm` on a `-c` line is a note rather than an
999            // error: it is a thing said to a linker that is not going to run.
1000            "-static" => link.is_static = true,
1001            "-shared" => link.shared = true,
1002            "-pie" => link.pie = Some(true),
1003            "-no-pie" | "-nopie" => link.pie = Some(false),
1004            "-nostdlib" => link.no_stdlib = true,
1005            "-nostartfiles" => link.no_startfiles = true,
1006            "-nodefaultlibs" => link.no_defaultlibs = true,
1007            "-fno-builtins-lib" => link.no_builtins_lib = true,
1008            "-fbuiltins-lib" => link.no_builtins_lib = false,
1009            "-rdynamic" | "-export-dynamic" => link.export_dynamic = true,
1010            "-s" => link.strip = true,
1011            // Into the ordered input list rather than a list of its own, because a great many of
1012            // the linker's options are a bracket around the files after them and an option that
1013            // lost its place among them says nothing. `--whole-archive` is the one that found this.
1014            "-Xlinker" => {
1015                let next = args.get(i).ok_or_else(|| err("-Xlinker requires an argument"))?;
1016                i += 1;
1017                inputs.push(Input::linker(next));
1018            }
1019            _ if arg.starts_with("-Wl,") => {
1020                // Commas separate arguments rather than being part of one, which is what makes
1021                // `-Wl,-rpath,/opt/lib` two words to the linker and one word here.
1022                inputs.extend(arg["-Wl,".len()..].split(',').map(Input::linker));
1023            }
1024            _ if arg.starts_with("-fuse-ld=") => {
1025                link.use_ld = Some(arg["-fuse-ld=".len()..].to_owned());
1026            }
1027            _ if arg.starts_with("-l") && arg.len() > 2 => {
1028                inputs.push(Input::library(&arg[2..]));
1029            }
1030            "-l" => {
1031                let next = args.get(i).ok_or_else(|| err("-l requires an argument"))?;
1032                i += 1;
1033                inputs.push(Input::library(next));
1034            }
1035            _ if arg.starts_with("-L") => {
1036                link.search.push(PathBuf::from(joined_or_next(arg, 2, args, &mut i)?));
1037            }
1038            _ if arg.starts_with("-B") => {
1039                link.prefixes.push(PathBuf::from(joined_or_next(arg, 2, args, &mut i)?));
1040            }
1041            _ if arg.starts_with("-j") => {
1042                jobs = Jobs::parse(&arg[2..]).map_err(err)?;
1043            }
1044            _ if arg.starts_with("--sysroot=") => {
1045                sysroot = Some(PathBuf::from(&arg["--sysroot=".len()..]));
1046            }
1047            _ if arg.starts_with("--target=") => {
1048                let t = &arg["--target=".len()..];
1049                opts.target = t.parse().map_err(|e| err(format!("{e}")))?;
1050                // The same string again, as the model that has room for a libc version. A spelling
1051                // the three field parser took and this one does not is not an error, because the
1052                // one that decides what is compiled has already accepted it and the only thing
1053                // lost is a version nobody asked for.
1054                pinned = t.parse().ok();
1055            }
1056            _ if arg.starts_with("--emit=") => {
1057                let k = &arg["--emit=".len()..];
1058                opts.emit = k
1059                    .parse()
1060                    .map_err(|()| err(format!("unknown --emit kind `{k}`, see --help")))?;
1061            }
1062            // A bare `-O` is `-O1`, which is what GCC has and what a hand written makefile tends
1063            // to write. `-Og` is GCC's level for a build somebody is going to step through, and
1064            // it is `-O1` with the transformations that move code around left out; this compiler
1065            // has no such level yet, so it is the nearest one and `--print-pipeline` says what
1066            // that came to rather than the flag pretending otherwise.
1067            "-O" | "-Og" => opts.opt_level = rucc_session::OptLevel::O1,
1068            // The union of `-O3` and `-ffast-math`, and the second half of that changes what
1069            // floating point arithmetic means. Refused rather than taken as `-O3`, because a
1070            // build that asks for fast math and is quietly given ordinary arithmetic gets a
1071            // slower program than it asked for and a build that is given fast math it did not
1072            // ask for gets a wrong one.
1073            "-Ofast" => {
1074                return Err(err(
1075                    "-Ofast is -O3 with fast math, and fast math is not implemented, see \
1076                     spec/04-driver-and-cli.md section 4.6",
1077                ));
1078            }
1079            _ if arg.starts_with("-O") => {
1080                opts.opt_level = arg[2..]
1081                    .parse()
1082                    .map_err(|()| err(format!("unknown optimization level `{arg}`")))?;
1083            }
1084            // How far a multiply and an addition may be fused into one rounding. Before the
1085            // optimizer's `-f` family below for the reason the ones under it are, and kept rather
1086            // than dropped because it is the one flag in its group this compiler could act on: it
1087            // rides into the IR as an attribute on each function with a body, so the day the code
1088            // generator forms an `fma` it already knows which functions were given permission.
1089            // Nothing forms one today, under any value of this and under any `-march=`.
1090            _ if arg.starts_with("-ffp-contract=") => {
1091                let how = &arg["-ffp-contract=".len()..];
1092                opts.fp_contract = how.parse().map_err(|()| {
1093                    err(format!("`{how}` is not a contraction, which is fast, on or off"))
1094                })?;
1095            }
1096            // How much of an expression may be computed wider than it was written. The values are
1097            // gcc's and so is the refusal of anything else, and none of the three changes anything
1098            // here: an operation is computed in the type C says it is on every target this compiler
1099            // has a back end for, so `__FLT_EVAL_METHOD__` is 0 and `standard` is already what
1100            // happens. `fast` and `16` are permission to be wider, which is a licence this takes
1101            // and does not use, the same way the two above are. The flag is worth taking because
1102            // glibc's headers and a good deal of configure output write it, and because the answer
1103            // it asks about is one this compiler can state rather than guess at: there is no x87
1104            // target here, which is the machine the whole question was invented for.
1105            // Whether a local and a spilled value that are never both wanted may be the same bytes
1106            // of the frame. gcc's three values, and two of them mean the same thing here: what rucc
1107            // shares is a local whose address provably never leaves the function, which is narrower
1108            // than `named_vars` and narrower still than `all`, so both of them get it. `none` is
1109            // the one that changes anything, and it is the flag a program that reads a local
1110            // through a pointer it kept past the end of the block writes.
1111            _ if arg.starts_with("-fstack-reuse=") => {
1112                let how = &arg["-fstack-reuse=".len()..];
1113                opts.stack_reuse = match how {
1114                    "all" | "named_vars" => Some(true),
1115                    "none" => Some(false),
1116                    _ => {
1117                        return Err(err(format!(
1118                            "`{how}` is not a stack reuse, which is all, named_vars or none"
1119                        )));
1120                    }
1121                };
1122            }
1123            _ if arg.starts_with("-fexcess-precision=") => {
1124                let how = &arg["-fexcess-precision=".len()..];
1125                if !matches!(how, "16" | "fast" | "standard") {
1126                    return Err(err(format!(
1127                        "`{how}` is not an excess precision, which is 16, fast or standard"
1128                    )));
1129                }
1130            }
1131            // Which front of a path is rewritten before it reaches the output, which is how a
1132            // build gets the same bytes out of two different directories. The four spellings are
1133            // one flag each into three lists, and `-ffile-prefix-map=` is the three of them at
1134            // once. Only the macro list does anything today, because `__FILE__` is the only place
1135            // a path reaches the output: there is no DWARF and no profile data yet, so the other
1136            // two are recorded for the work that will read them. The argument splits at the last
1137            // `=` rather than the first, which is gcc's rule and is what lets a directory with an
1138            // `=` in its name be the old half.
1139            _ if arg.starts_with("-fmacro-prefix-map=") => {
1140                let (old, new) = rewrite(arg, "-fmacro-prefix-map=")?;
1141                opts.prefix_map.macros.push(old, new);
1142            }
1143            _ if arg.starts_with("-fdebug-prefix-map=") => {
1144                let (old, new) = rewrite(arg, "-fdebug-prefix-map=")?;
1145                opts.prefix_map.debug.push(old, new);
1146            }
1147            _ if arg.starts_with("-fprofile-prefix-map=") => {
1148                let (old, new) = rewrite(arg, "-fprofile-prefix-map=")?;
1149                opts.prefix_map.profile.push(old, new);
1150            }
1151            _ if arg.starts_with("-ffile-prefix-map=") => {
1152                let (old, new) = rewrite(arg, "-ffile-prefix-map=")?;
1153                opts.prefix_map.macros.push(old, new);
1154                opts.prefix_map.debug.push(old, new);
1155                opts.prefix_map.profile.push(old, new);
1156            }
1157            // A whole optimization rather than a flag, and the family is taken rather than
1158            // refused because of what ignoring it does. There is none of it here yet, so a build
1159            // that asks for it gets a program that is correct and slower than it could have been,
1160            // which is what section 4.1 means by a hint about speed and what every compilation at
1161            // `-O0` already is. The objects settle the rest of the argument: gcc's `-flto` object
1162            // holds the bytecode and no machine code at all, and every object here holds the code,
1163            // which is exactly what `-ffat-lto-objects` asks gcc for. So a build passing `-flto`
1164            // to this compiler gets objects that are more usable than the ones it asked for rather
1165            // than different ones. Every value is still checked against gcc's, because somebody
1166            // who wrote `-flto=thin` meant clang and had better hear about it here.
1167            "-flto" => opts.lto.requested = true,
1168            "-fno-lto" => opts.lto.requested = false,
1169            _ if arg.starts_with("-flto=") => {
1170                let how = &arg["-flto=".len()..];
1171                opts.lto.jobs = how.parse().map_err(|()| {
1172                    err(format!(
1173                        "`{how}` is not a number of link time jobs, which is auto, jobserver or a \
1174                         count above zero"
1175                    ))
1176                })?;
1177                opts.lto.requested = true;
1178            }
1179            _ if arg.starts_with("-flto-partition=") => {
1180                let how = &arg["-flto-partition=".len()..];
1181                opts.lto.partition = how.parse().map_err(|()| {
1182                    err(format!(
1183                        "`{how}` is not a partitioning model, which is balanced, 1to1, one, max \
1184                         or none"
1185                    ))
1186                })?;
1187            }
1188            _ if arg.starts_with("-flto-compression-level=") => {
1189                let how = &arg["-flto-compression-level=".len()..];
1190                let level =
1191                    how.parse::<u8>().ok().filter(|level| *level <= 19).ok_or_else(|| {
1192                        err(format!("`{how}` is not a compression level, 0 to 19"))
1193                    })?;
1194                opts.lto.compression = Some(level);
1195            }
1196            // Whether the object keeps its machine code as well as the bytecode. It always does
1197            // here, so the first of these describes what happens and the second asks for an object
1198            // with less in it, which is a smaller file and not a different program, so both are
1199            // taken.
1200            "-ffat-lto-objects" | "-fno-fat-lto-objects" => {}
1201            // Whether the linker is handed a plugin that does the link time work. The design in
1202            // `spec/09-optimizer.md` has this driver doing that work itself and never loading a
1203            // plugin into anybody, so neither answer is a question it has to hold.
1204            "-fuse-linker-plugin" | "-fno-use-linker-plugin" => {}
1205            // Reading a profile back. Taken for the reason the family above it is: nothing here
1206            // reads one, so a build that asks gets the program it would have got anyway, and gcc
1207            // itself produces a byte for byte identical object from `-fprofile-use` when there are
1208            // no counts beside the file. The path is recorded for the pass that will read it. The
1209            // warning gcc prints when it looked and found nothing is deliberately not copied,
1210            // because nothing here looks, and a warning about a file that was never opened would
1211            // fire on the builds that have a perfectly good profile as well as on the ones that
1212            // do not.
1213            "-fprofile-use" => opts.profile_data.requested = true,
1214            "-fno-profile-use" => opts.profile_data.requested = false,
1215            _ if arg.starts_with("-fprofile-use=") => {
1216                opts.profile_data.path = Some(arg["-fprofile-use=".len()..].to_string());
1217                opts.profile_data.requested = true;
1218            }
1219            _ if arg.starts_with("-fprofile-dir=") => {
1220                opts.profile_data.dir = Some(arg["-fprofile-dir=".len()..].to_string());
1221            }
1222            "-fprofile-abs-path" => opts.profile_data.absolute = true,
1223            "-fno-profile-abs-path" => opts.profile_data.absolute = false,
1224            "-fprofile-correction" => opts.profile_data.correction = true,
1225            "-fno-profile-correction" => opts.profile_data.correction = false,
1226            "-fprofile-partial-training" => opts.profile_data.partial_training = true,
1227            "-fno-profile-partial-training" => opts.profile_data.partial_training = false,
1228            // Writing the counts rather than reading them, which is refused rather than taken and
1229            // is the same line `-gsplit-dwarf` falls on the far side of. Ignoring these means a
1230            // file a build declared as an output never appears: the instrumented program writes a
1231            // `.gcda` as it exits and `-ftest-coverage` writes a `.gcno` beside the object, and a
1232            // two stage build that got neither would go on to optimize against no counts at all
1233            // and report coverage of nothing, with nothing along the way saying so. The objects
1234            // say the rest: gcc's `-fprofile-generate` object holds 375 bytes of code where a
1235            // plain one holds 71, and 296 bytes of counters that a plain one does not have, so
1236            // this is a flag that changes the output rather than a hint about speed.
1237            "-fprofile-arcs"
1238            | "--coverage"
1239            | "-fcondition-coverage"
1240            | "-fpath-coverage"
1241            | "-fprofile-generate" => {
1242                return Err(err(format!(
1243                    "{arg}: this compiler does not instrument for profiling, and a build that \
1244                     expects the counts a run of the instrumented program writes would optimize \
1245                     against nothing on its second pass, see spec/04-driver-and-cli.md"
1246                )));
1247            }
1248            _ if arg.starts_with("-fprofile-generate=") => {
1249                return Err(err(format!(
1250                    "{arg}: this compiler does not instrument for profiling, and a build that \
1251                     expects the counts a run of the instrumented program writes would optimize \
1252                     against nothing on its second pass, see spec/04-driver-and-cli.md"
1253                )));
1254            }
1255            "-ftest-coverage" => {
1256                return Err(err(format!(
1257                    "{arg}: this compiler writes no `.gcno` file beside the object, and a build \
1258                     that expects one would wait for a file that never arrives, see \
1259                     spec/04-driver-and-cli.md"
1260                )));
1261            }
1262            // The rest of the family describes instrumentation that is refused above, so what is
1263            // left to do with them is check them and drop them. They are checked because a
1264            // misspelling in a distribution's flags is worth finding here rather than on the day
1265            // the instrumentation lands, and dropped because there is nothing for an answer about
1266            // how a counter is written to be an answer about.
1267            _ if arg.starts_with("-fprofile-update=") => {
1268                let how = &arg["-fprofile-update=".len()..];
1269                if !matches!(how, "single" | "atomic" | "prefer-atomic") {
1270                    return Err(err(format!(
1271                        "`{how}` is not a profile update method, which is single, atomic or \
1272                         prefer-atomic"
1273                    )));
1274                }
1275            }
1276            _ if arg.starts_with("-fprofile-reproducible=") => {
1277                let how = &arg["-fprofile-reproducible=".len()..];
1278                if !matches!(how, "serial" | "parallel-runs" | "multithreaded") {
1279                    return Err(err(format!(
1280                        "`{how}` is not a profile reproducibility method, which is serial, \
1281                         parallel-runs or multithreaded"
1282                    )));
1283                }
1284            }
1285            "-fprofile-values" | "-fno-profile-values" | "-fprofile-info-section" => {}
1286            "-fno-test-coverage" | "-fno-profile-arcs" | "-fno-profile-generate" => {}
1287            _ if arg.starts_with("-fprofile-filter-files=")
1288                || arg.starts_with("-fprofile-exclude-files=")
1289                || arg.starts_with("-fprofile-note=") => {}
1290            // What every name gets when nothing in the source said, which the attribute in the
1291            // source overrides rather than the other way round. Before the optimizer's `-f`
1292            // family below for the reason the tier below it is.
1293            _ if arg.starts_with("-fvisibility=") => {
1294                let seen = &arg["-fvisibility=".len()..];
1295                opts.visibility = seen.parse().map_err(|()| {
1296                    err(format!(
1297                        "`{seen}` is not a visibility, which is default, hidden, internal or \
1298                         protected"
1299                    ))
1300                })?;
1301            }
1302            // Which edges of a control flow transfer are checked. Before the optimizer's `-f`
1303            // family below for the reason the two above it are, and last of the three so that the
1304            // bare spelling and the negative one are matched exactly rather than by this.
1305            _ if arg.starts_with("-fcf-protection=") => {
1306                let edges = &arg["-fcf-protection=".len()..];
1307                opts.control = edges.parse().map_err(|()| {
1308                    err(format!(
1309                        "`{edges}` is not a control flow protection, which is full, branch, \
1310                         return, none or check"
1311                    ))
1312                })?;
1313            }
1314            // How much room every function opens with for something to be written over later.
1315            // Before the optimizer's `-f` family below for the reason the ones above it are.
1316            _ if arg.starts_with("-fpatchable-function-entry=") => {
1317                let room = &arg["-fpatchable-function-entry=".len()..];
1318                opts.patchable = room.parse().map_err(|()| {
1319                    err(format!(
1320                        "`{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"
1321                    ))
1322                })?;
1323            }
1324            // The memory safety monitor, from section 15.4 of
1325            // `spec/safe-memory/15-integration.md`. Before the optimizer's `-f` family below,
1326            // because a pass that took the name `safety=detect` would otherwise be handed the
1327            // flag, and the tier is not a pass.
1328            _ if arg.starts_with("-fsafety=") => {
1329                let tier = &arg["-fsafety=".len()..];
1330                opts.safety = tier.parse().map_err(|()| {
1331                    err(format!(
1332                        "`{tier}` is not a safety tier, which is off, detect, enforce or kernel"
1333                    ))
1334                })?;
1335            }
1336            // Whether padding participates, from section 9.3 of document 09. Spelled out rather
1337            // than folded into the tier because it is a departure somebody who has read that
1338            // section makes, and the two defaults it describes are a property of what is being
1339            // built rather than of how much checking is wanted.
1340            _ if arg.starts_with("-fsafety-init=") => {
1341                let mode = &arg["-fsafety-init=".len()..];
1342                opts.padding = mode.parse().map_err(|()| {
1343                    err(format!("`{mode}` is not a padding mode, which is padding or nopadding"))
1344                })?;
1345            }
1346            // Row S4, from section 9.4 of document 09. A bare flag with no value, because the
1347            // strict form of that section needs a member id the front end does not name yet and
1348            // accepting the spelling for it would be accepting a promise this build cannot keep.
1349            // Before `-fno-` is looked at below, for the reason the tier is.
1350            "-fsafety-subobject" => opts.subobject = rucc_session::Subobject::Members,
1351            "-fno-safety-subobject" => opts.subobject = rucc_session::Subobject::Off,
1352            _ if arg.starts_with("-fsafety-subobject=") => {
1353                let form = &arg["-fsafety-subobject=".len()..];
1354                return Err(err(format!(
1355                    "`{form}` is not a form of -fsafety-subobject. The flag takes no value, and \
1356                     the strict form of section 9.4 is tamnd/rucc#967"
1357                )));
1358            }
1359            // Row Y8, from section 9.6 of document 09. A bare flag with no value, for the reason
1360            // the one above has none: there is one form of this check and a spelling that suggested
1361            // otherwise would be promising something. Before `-fno-` is looked at below, the same
1362            // way.
1363            "-fsafety-restrict" => opts.promise = rucc_session::Promise::Blocks,
1364            "-fno-safety-restrict" => opts.promise = rucc_session::Promise::Off,
1365            _ if arg.starts_with("-fsafety-restrict=") => {
1366                let form = &arg["-fsafety-restrict=".len()..];
1367                return Err(err(format!(
1368                    "`{form}` is not a form of -fsafety-restrict. The flag takes no value."
1369                )));
1370            }
1371            // Section 9.5's races, which take a value because the section gives them three modes
1372            // and the difference between two of them is which classes get reported rather than how
1373            // much is recorded. `-fno-` is the same as `=off` and is spelled out here for the same
1374            // reason the two above spell theirs out.
1375            _ if arg.starts_with("-fsafety-races=") => {
1376                let mode = &arg["-fsafety-races=".len()..];
1377                opts.races = mode.parse().map_err(|()| {
1378                    err(format!("`{mode}` is not a race mode, which is off, metadata or pointer"))
1379                })?;
1380            }
1381            "-fno-safety-races" => opts.races = rucc_session::Races::Off,
1382            // The sanitizers of document 12, which are checks at run time rather than a way of
1383            // generating the same program. Each name is held to gcc 16's list, and what is still
1384            // asked for by the end of the line is answered after the loop, so that a command line
1385            // which turns one on and then off again is a command line that asked for nothing.
1386            //
1387            // Before the optimizer's `-f` family below, for the reason the tier above it is.
1388            _ if arg.starts_with("-fsanitize=") => {
1389                for one in arg["-fsanitize=".len()..].split(',') {
1390                    if one == "all" {
1391                        // gcc takes `all` only in the negative, because turning every check on at
1392                        // once includes checks that contradict each other.
1393                        return Err(err(
1394                            "`-fsanitize=all` is not a gcc option, only `-fno-sanitize=all` is",
1395                        ));
1396                    }
1397                    if !SANITIZERS.contains(&one) {
1398                        return Err(err(format!(
1399                            "`{one}` is not a sanitizer, see spec/04-driver-and-cli.md section 4.7"
1400                        )));
1401                    }
1402                    if !sanitizers.contains(&one) {
1403                        sanitizers.push(one);
1404                    }
1405                }
1406            }
1407            _ if arg.starts_with("-fno-sanitize=") => {
1408                for one in arg["-fno-sanitize=".len()..].split(',') {
1409                    if one == "all" {
1410                        sanitizers.clear();
1411                        continue;
1412                    }
1413                    if !SANITIZERS.contains(&one) {
1414                        return Err(err(format!(
1415                            "`{one}` is not a sanitizer, see spec/04-driver-and-cli.md section 4.7"
1416                        )));
1417                    }
1418                    sanitizers.retain(|asked| *asked != one);
1419                }
1420            }
1421            // What a check does when it fires, and where the records about the checked objects go.
1422            // Each of them is an answer about the sanitizers refused after the loop, so there is
1423            // nothing left for them to change here. The names are still held to the list, because
1424            // a misspelling in a build's flags is worth finding when the compiler reads it.
1425            _ if arg.starts_with("-fsanitize-recover=")
1426                || arg.starts_with("-fno-sanitize-recover=")
1427                || arg.starts_with("-fsanitize-trap=")
1428                || arg.starts_with("-fno-sanitize-trap=") =>
1429            {
1430                // The guard above matched on a spelling that has an `=` in it, so the tail is
1431                // whatever follows the first one.
1432                let how = arg.split_once('=').map_or("", |(_, rest)| rest);
1433                for one in how.split(',') {
1434                    if one != "all" && !SANITIZERS.contains(&one) {
1435                        return Err(err(format!(
1436                            "`{one}` is not a sanitizer, see spec/04-driver-and-cli.md section 4.7"
1437                        )));
1438                    }
1439                }
1440            }
1441            "-fsanitize-undefined-trap-on-error"
1442            | "-fsanitize-address-use-after-scope"
1443            | "-fno-sanitize-address-use-after-scope" => {}
1444            _ if arg.starts_with("-fsanitize-sections=") => {}
1445            // Counting which edges a run reached, which is how a fuzzer knows an input was worth
1446            // keeping. Refused rather than dropped, because a fuzzer whose calls into
1447            // `__sanitizer_cov_*` were never generated runs blind and reports coverage of nothing,
1448            // and there is no point in the campaign where that announces itself.
1449            _ if arg.starts_with("-fsanitize-coverage=") => {
1450                let how = &arg["-fsanitize-coverage=".len()..];
1451                for one in how.split(',') {
1452                    if !matches!(one, "trace-pc" | "trace-cmp") {
1453                        return Err(err(format!(
1454                            "`{one}` is not a coverage instrumentation, which is trace-pc or \
1455                             trace-cmp"
1456                        )));
1457                    }
1458                }
1459                return Err(err(format!(
1460                    "{arg}: this compiler generates no coverage callbacks, and a fuzzer built \
1461                     with it would run without any feedback at all, see \
1462                     spec/04-driver-and-cli.md section 4.7"
1463                )));
1464            }
1465            // The optimizer's own flags, from section 9.10 of `spec/09-optimizer.md`. These come
1466            // after every `-f` the rest of the compiler answers to, so a pass can never take a
1467            // name that already means something else on the command line.
1468            _ if arg.starts_with("-fpass-fuel=") => {
1469                let (name, count) = arg["-fpass-fuel=".len()..]
1470                    .split_once('=')
1471                    .ok_or_else(|| err("-fpass-fuel= is spelled <pass>=<count>"))?;
1472                if rucc_opt::pass::find(name).is_none() {
1473                    return Err(err(format!(
1474                        "`{name}` is not a pass this compiler has, see --print-pipeline"
1475                    )));
1476                }
1477                let count: u32 = count
1478                    .parse()
1479                    .map_err(|_| err(format!("`{count}` is not a number of transformations")))?;
1480                opts.pass_fuel.push((name.to_owned(), count));
1481            }
1482            _ if arg.starts_with("-fpass-fuel-global=") => {
1483                let count = &arg["-fpass-fuel-global=".len()..];
1484                let count: u32 = count
1485                    .parse()
1486                    .map_err(|_| err(format!("`{count}` is not a number of transformations")))?;
1487                opts.pass_fuel_global = Some(count);
1488            }
1489            // Everything from `-fopt-info` to the end of the argument, which is optional
1490            // keywords joined by hyphens and an optional `=<file>`. Checked here rather than
1491            // where the remarks are printed, because by then the compilation somebody wanted
1492            // to hear about is over.
1493            _ if arg == "-fopt-info"
1494                || arg.starts_with("-fopt-info=")
1495                || arg.starts_with("-fopt-info-") =>
1496            {
1497                let rest = &arg["-fopt-info".len()..];
1498                let (kinds, file) = match rest.split_once('=') {
1499                    Some((kinds, file)) => (kinds, Some(file)),
1500                    None => (rest, None),
1501                };
1502                let kinds = kinds.strip_prefix('-').unwrap_or(kinds);
1503                rucc_opt::Wants::none().add(kinds).map_err(err)?;
1504                opts.opt_info.push(kinds.to_owned());
1505                if let Some(file) = file {
1506                    if file.is_empty() {
1507                        return Err(err("-fopt-info= was given no file to write to"));
1508                    }
1509                    opts.opt_info_file = Some(file.to_owned());
1510                }
1511            }
1512            _ if arg.starts_with("-fdump-ir=") => {
1513                // Checked here rather than where the dumps are taken, because the compilation
1514                // that would have been dumped is over by then.
1515                let spec = &arg["-fdump-ir=".len()..];
1516                rucc_opt::Dumps::default().add(spec).map_err(err)?;
1517                opts.dump_ir.push(spec.to_owned());
1518            }
1519            // Before the bare `-f<pass>` below, because a pass called `enable-something` would
1520            // otherwise take the flag away from the gate. Checked here rather than where the
1521            // pipeline reads it, for the reason that applies to all of these: a misspelled pass
1522            // name that quietly gated nothing looks exactly like a pass that is not the guilty
1523            // one, and a bisection would carry on past the thing it was looking for.
1524            _ if arg.starts_with("-fdisable-") || arg.starts_with("-fenable-") => {
1525                let on = arg.starts_with("-fenable-");
1526                let spec = &arg[if on { "-fenable-".len() } else { "-fdisable-".len() }..];
1527                rucc_opt::Gates::default().add(on, spec).map_err(err)?;
1528                opts.pass_gates.push((on, spec.to_owned()));
1529            }
1530            // gcc's spelling for a pass this compiler has under a shorter name. It goes above the
1531            // two arms below rather than into the pile of gcc pass names further down, because the
1532            // pass is here: dropping the flag would leave a build that asked for unrolling without
1533            // it, and refusing it stops the build outright, which is what libtommath's makefile
1534            // ran into. `-funroll-all-loops` is deliberately not in here: gcc's is the one that
1535            // unrolls without a trip count, which is a different and usually worse thing.
1536            "-funroll-loops" => opts.passes.push(("unroll".to_owned(), true)),
1537            "-fno-unroll-loops" => opts.passes.push(("unroll".to_owned(), false)),
1538            // Here rather than through the two arms below, because what this names is not a
1539            // `rucc_opt::Pass`. Section 34.6's propagation is a module at a time and everything in
1540            // the pass list is one function at a time. `-fipa-cp-clone` is deliberately not here:
1541            // gcc turns that one on at `-O3` and it is in the list of what M4 does not build.
1542            "-fipa-cp" => opts.passes.push((rucc_opt::ipcp::NAME.to_owned(), true)),
1543            "-fno-ipa-cp" => opts.passes.push((rucc_opt::ipcp::NAME.to_owned(), false)),
1544            // The other half of the same section, here for the same reason, and `-fipa-sra` in gcc
1545            // is the aggregate splitting as well as the parameter removal. Asking for it gets the
1546            // half that is built.
1547            "-fipa-sra" => opts.passes.push((rucc_opt::ipasra::NAME.to_owned(), true)),
1548            "-fno-ipa-sra" => opts.passes.push((rucc_opt::ipasra::NAME.to_owned(), false)),
1549            _ if arg.strip_prefix("-fno-").is_some_and(|n| rucc_opt::pass::find(n).is_some()) => {
1550                opts.passes.push((arg["-fno-".len()..].to_owned(), false));
1551            }
1552            _ if arg.strip_prefix("-f").is_some_and(|n| rucc_opt::pass::find(n).is_some()) => {
1553                opts.passes.push((arg["-f".len()..].to_owned(), true));
1554            }
1555            // The flags that name a pass of gcc's own. They arrive from the torture suite, where a
1556            // program reduced from a miscompilation usually names the pass that miscompiled it on
1557            // its `dg-options` line, and they arrive from hand written build files for the same
1558            // reason. Section 4.1 sorts a flag by what the output would be without it, and by that
1559            // rule these are one pile: a flag that turns one of gcc's passes on or off is asking
1560            // for a compiler that does not exist here, and the program it is attached to is a
1561            // correctness test that passes either way. Turning on a pass we do not have costs
1562            // speed, turning off a pass we do not have costs nothing, and neither changes what the
1563            // program computes.
1564            //
1565            // rucc's own pass names are matched above this, so `-fno-dce` turns off the dce this
1566            // compiler has rather than landing here, and the day one of these names becomes a pass
1567            // here it stops being taken and dropped without anybody editing this list.
1568            //
1569            // Two of them are prefixes rather than names, which is the one place this file takes a
1570            // family instead of a flag. gcc files its gimple passes under `-ftree-` and its
1571            // interprocedural passes under `-fipa-`, both namespaces are pass selection and
1572            // nothing else, and there is no member of either that changes the meaning of a program
1573            // that was already correct. The rest are written out one at a time, because they live
1574            // in the flat `-f` namespace where the neighbours do change meanings.
1575            _ if arg.starts_with("-ftree-") || arg.starts_with("-fno-tree-") => {}
1576            _ if arg.starts_with("-fipa-") || arg.starts_with("-fno-ipa-") => {}
1577            "-fexpensive-optimizations" | "-fno-expensive-optimizations" => {}
1578            "-fmodulo-sched" | "-fno-modulo-sched" => {}
1579            "-fvect-cost-model" | "-fno-vect-cost-model" => {}
1580            _ if arg.starts_with("-fvect-cost-model=") || arg.starts_with("-fsimd-cost-model=") => {
1581            }
1582            "-fearly-inlining" | "-fno-early-inlining" => {}
1583            "-finline"
1584            | "-fno-inline"
1585            | "-finline-functions"
1586            | "-fno-inline-functions"
1587            | "-finline-small-functions"
1588            | "-fno-inline-small-functions"
1589            | "-finline-functions-called-once"
1590            | "-fno-inline-functions-called-once" => {}
1591            "-foptimize-strlen" | "-fno-optimize-strlen" => {}
1592            "-fira-share-spill-slots" | "-fno-ira-share-spill-slots" => {}
1593            // Where a function starts, which is a thing this compiler already decides and so is a
1594            // request it can answer rather than one it has to drop. The bare form asks for the
1595            // target's default and the default here is the sixteen bytes gcc also gives, so it
1596            // says nothing; a number is a floor under every function that did not ask for more
1597            // itself; and the negative form asks for the smallest boundary the target has. gcc 16
1598            // rounds a number that is not a power of two up rather than refusing it, which is what
1599            // `=3` giving `.p2align 2` on x86-64 means, so this rounds too.
1600            "-falign-functions" => opts.align_functions = None,
1601            "-fno-align-functions" => opts.align_functions = Some(MIN_FUNC_ALIGN),
1602            _ if arg.starts_with("-falign-functions=") => {
1603                opts.align_functions = function_alignment(&arg["-falign-functions=".len()..])
1604                    .ok_or_else(|| {
1605                        err(format!("{arg}: the alignment has to be a number of bytes"))
1606                    })?;
1607            }
1608            // The other three of the family, which are about padding in front of a label inside a
1609            // body. This compiler writes none, and what they ask for is speed: a loop that starts
1610            // on a cache line boundary computes what a loop that does not computes. So they are
1611            // taken and dropped for the reason `-march=` is, and they are kept out of the arm
1612            // above because the question they ask is a different one and the day one of them is
1613            // answered it will be answered separately.
1614            _ if arg.starts_with("-falign-labels")
1615                || arg.starts_with("-falign-loops")
1616                || arg.starts_with("-falign-jumps")
1617                || arg.starts_with("-fno-align-labels")
1618                || arg.starts_with("-fno-align-loops")
1619                || arg.starts_with("-fno-align-jumps") => {}
1620            // The charset flags are not in that pile, because an encoding is a statement about
1621            // what the bytes of the source mean rather than about how fast the output is. The
1622            // preprocessor reads UTF-8 and has no converter, so the one name that describes what
1623            // already happens is taken and every other name is refused. Spelled without regard to
1624            // case and with both of the spellings iconv answers to, since a build writes whichever
1625            // one its author typed.
1626            _ if arg.starts_with("-finput-charset=") => {
1627                let name = &arg["-finput-charset=".len()..];
1628                if !name.eq_ignore_ascii_case("utf-8") && !name.eq_ignore_ascii_case("utf8") {
1629                    return Err(err(format!(
1630                        "-finput-charset={name}: the preprocessor reads UTF-8 and has no \
1631                         converter, so a file in another encoding would be read as though it were \
1632                         UTF-8 rather than converted",
1633                    )));
1634                }
1635            }
1636            // The three that come in on the same `dg-options` lines and are the other half of
1637            // section 4.1's rule, because each of them changes what the program does and not how
1638            // fast it does it. The negative form of each is what this compiler does anyway, so it
1639            // is taken and dropped, which is the shape `-fnested-functions` has above.
1640            "-ffast-math" => {
1641                return Err(err(
1642                    "-ffast-math is a licence to answer a floating point arithmetic differently \
1643                     from the way the source wrote it, and it is not one flag: it defines \
1644                     __FAST_MATH__, which a library header reads, and gcc links a startup file \
1645                     that puts the hardware in flush to zero mode for the whole process. Taking it \
1646                     and dropping it would change what other objects in the same program answer. \
1647                     -ffp-contract= and -fexcess-precision= are the parts of it this compiler has",
1648                ));
1649            }
1650            "-fno-fast-math" => {}
1651            "-fnon-call-exceptions" => {
1652                return Err(err(
1653                    "-fnon-call-exceptions is a promise that an instruction which is not a call \
1654                     can raise an exception the unwinder finds a handler for, and nothing here \
1655                     produces a landing pad for a trapping instruction. A program built without it \
1656                     would unwind past the handler it wrote",
1657                ));
1658            }
1659            "-fno-non-call-exceptions" => {}
1660            "-finstrument-functions" => {
1661                return Err(err(
1662                    "-finstrument-functions calls __cyg_profile_func_enter on entry to every \
1663                     function and __cyg_profile_func_exit on the way out, and nothing here emits \
1664                     either call. A program that asks for them usually counts them, so taking the \
1665                     flag and dropping it would turn a program that fails loudly into one that \
1666                     fails quietly",
1667                ));
1668            }
1669            "-fno-instrument-functions" => {}
1670            // The unstable options, spelled the way rustc spells them and carrying the same
1671            // promise, which is none: one of these may change or go away in any release. They are
1672            // measurements and debugging aids rather than things a build asks for, which is why
1673            // none of them is in the usage text and all of them are in section 4.11 of
1674            // `spec/04-driver-and-cli.md`.
1675            "-Zverify-each" => opts.verify_each = true,
1676            _ if arg.starts_with("-Zrule-coverage=") => {
1677                let file = &arg["-Zrule-coverage=".len()..];
1678                if file.is_empty() {
1679                    return Err(err("-Zrule-coverage= needs a file to write to"));
1680                }
1681                opts.rule_coverage = Some(file.to_owned());
1682            }
1683            _ if arg.starts_with("-Zcycle-accurate-model=") => {
1684                let value = &arg["-Zcycle-accurate-model=".len()..];
1685                opts.cycle_accurate_model = match value {
1686                    "yes" | "1" => Some(true),
1687                    "no" | "0" => Some(false),
1688                    _ => {
1689                        return Err(err("-Zcycle-accurate-model= takes yes or no"));
1690                    }
1691                };
1692            }
1693            _ if arg.starts_with("-Zlowering=") => {
1694                let file = &arg["-Zlowering=".len()..];
1695                if file.is_empty() {
1696                    return Err(err("-Zlowering= needs a file to write to"));
1697                }
1698                opts.lowering_dump = Some(file.to_owned());
1699            }
1700            _ if arg.starts_with("-Zregister-pressure=") => {
1701                let file = &arg["-Zregister-pressure=".len()..];
1702                if file.is_empty() {
1703                    return Err(err("-Zregister-pressure= needs a file to write to"));
1704                }
1705                opts.register_pressure = Some(file.to_owned());
1706            }
1707            _ if arg.starts_with("-Z") => {
1708                return Err(err(format!(
1709                    "`{arg}` is not an unstable option this compiler has, see \
1710                     spec/04-driver-and-cli.md section 4.11 for the ones it does"
1711                )));
1712            }
1713            // The word size, which is a statement about the target and is taken as one. A build
1714            // that says the size the target already has is saying nothing, and one that says the
1715            // other size is asking for a target this compiler does not have, which it is told
1716            // rather than being given the wrong one.
1717            "-m64" | "-m32" | "-mx32" => {
1718                let want: u32 = match arg {
1719                    "-m64" => 64,
1720                    _ => 32,
1721                };
1722                let have = rucc_target::TargetInfo::new(opts.target).pointer_width;
1723                if have != want {
1724                    return Err(err(format!(
1725                        "{arg} asks for a {want} bit target and {} is {have} bit, use \
1726                         --target= to name the one you mean",
1727                        opts.target
1728                    )));
1729                }
1730            }
1731            // Which processor in the family to generate for. This compiler emits the base
1732            // instruction set of the architecture and nothing above it, so a program built with
1733            // any of these runs on the machine that was named; it is a program that could have
1734            // been faster rather than a program that is wrong, which is what makes these safe to
1735            // take and ignore where a flag that changed the meaning of the code would not be.
1736            _ if arg.starts_with("-march=")
1737                || arg.starts_with("-mtune=")
1738                || arg.starts_with("-mcpu=") => {}
1739            // The calling convention, which is not safe to ignore. Taken when it names the one
1740            // the target already uses and refused otherwise.
1741            _ if arg.starts_with("-mabi=") => {
1742                let want = &arg["-mabi=".len()..];
1743                let have = match opts.target.arch {
1744                    rucc_target::Arch::X86_64 => "sysv",
1745                    rucc_target::Arch::Aarch64 => "lp64",
1746                    rucc_target::Arch::Riscv64 => "lp64d",
1747                };
1748                if want != have {
1749                    return Err(err(format!(
1750                        "{arg}: {} uses the {have} convention and this compiler has no other",
1751                        opts.target
1752                    )));
1753                }
1754            }
1755            // How far apart the pieces of the program may be. The small model is what we emit and
1756            // it is every hosted program's default; the kernel model is a different one and a
1757            // build that asks for it and does not get it links and then does not run.
1758            "-mcmodel=small" => {}
1759            _ if arg.starts_with("-mcmodel=") => {
1760                return Err(err(format!(
1761                    "{arg}: this compiler emits the small code model and no other, see \
1762                     spec/12-targets.md"
1763                )));
1764            }
1765            // GCC's own scripting language for how the driver builds a command line.
1766            // `spec/04-driver-and-cli.md` section 4.4 settles that we will not have it, so a
1767            // build reaching for it is told which flags do the same job.
1768            _ if arg.starts_with("-specs=") => {
1769                return Err(err(
1770                    "-specs= is not supported: the parts of it builds rely on are -B, -L, \
1771                     -nostdlib, -nostartfiles and -Wl,, see spec/04-driver-and-cli.md \
1772                     section 4.4",
1773                ));
1774            }
1775            // Arguments meant for a separate assembler or preprocessor, which this compiler does
1776            // not have: both are inside it and neither reads a command line. Refused rather than
1777            // dropped, because every one of these says something about the output and a build
1778            // that asked for `-Wa,--noexecstack` and was silently given an executable stack got
1779            // the opposite of what it asked for.
1780            _ if arg.starts_with("-Wa,") || arg.starts_with("-Wp,") => {
1781                return Err(err(format!(
1782                    "`{arg}` is an argument for a separate assembler or preprocessor, and both \
1783                     are inside this compiler rather than programs it runs"
1784                )));
1785            }
1786            "-Xassembler" | "-Xpreprocessor" => {
1787                return Err(err(format!(
1788                    "{arg} hands an argument to a separate assembler or preprocessor, and both \
1789                     are inside this compiler rather than programs it runs"
1790                )));
1791            }
1792            // Everything else in the `-W` family. `spec/04-driver-and-cli.md` section 4.1 has
1793            // this one as a rule about build systems rather than about warnings: autoconf finds
1794            // out whether a warning flag exists by passing it and looking at the exit status, so
1795            // a compiler that refuses one it has not heard of fails a configure script written
1796            // for a GCC newer than itself. The names are not checked against a list because this
1797            // compiler has no warning groups for a list to be of, which #485 is about.
1798            _ if arg.starts_with("-W") => {}
1799            // Flags that name something this compiler does not do and would not do differently
1800            // if it did. `-fno-ident` is about a comment in the output that we do not write
1801            // either way, and the others are about a way of ordering the compilation that has
1802            // been GCC's only way for twenty years. Section 4.1 asks for the list to be short
1803            // and for adding to it to be deliberate, which is why it is written out here.
1804            "-fno-ident"
1805            | "-fident"
1806            | "-funit-at-a-time"
1807            | "-fno-unit-at-a-time"
1808            | "-shared-libgcc"
1809            | "-static-libgcc" => {}
1810            _ if arg.starts_with('-') && arg.len() > 1 => {
1811                // Silently ignoring an unknown flag is how a build ends up not doing what
1812                // its author asked. spec/13-gnu-compat.md section 13.4 makes this an error
1813                // for the flags that change code generation, and the safe default until the
1814                // flag table is populated is to reject everything we do not know.
1815                return Err(err(format!("unknown option `{arg}`")));
1816            }
1817            _ => inputs.push(Input { path: arg.to_owned(), forced, role: Role::File }),
1818        }
1819    }
1820
1821    // The fetch, before anything that resolves a compilation, because `--fetch` does not describe
1822    // one. It is here rather than in the loop so that `--offline` can forbid it whichever order the
1823    // two were written in, and it is before the refusals below so that a command line asking for a
1824    // sysroot is not told about a sanitizer.
1825    if let Some(named) = fetch {
1826        return fetch_action(&named, offline, &inputs);
1827    }
1828
1829    // Last, so that it lands after every `-isystem` the command line gave. That is GCC's
1830    // order: a directory the user names outranks the compiler's own, and the compiler's own
1831    // outranks the library's. It is pushed after the loop rather than before it because
1832    // `SearchPath` appends within a group and the position is what the order is.
1833    // The same directory the headers were looked for under, because a sysroot is a statement
1834    // about a whole installation and not about half of one.
1835    // After the loop, because `-fno-sanitize=` can take back what an earlier flag asked for and a
1836    // command line that turns a check on and off again has asked for nothing. What is left is
1837    // refused rather than dropped, and it is the one place in this parser where the reason is not
1838    // that the output would differ. A sanitizer is a promise that the program is watched while it
1839    // runs, so a build that asks for one and is quietly given a program with no checks in it does
1840    // not get a slower program or a bigger file, it gets a test suite that passes for the wrong
1841    // reason. `-fsafety=` is the checking this compiler does have, and the message says so, because
1842    // somebody reaching for `-fsanitize=address` wants the nearest thing rather than a list of
1843    // options.
1844    if let Some(first) = sanitizers.first() {
1845        return Err(err(format!(
1846            "-fsanitize={first}: this compiler has no sanitizer instrumentation, and a build that \
1847             asked for one and got none would run its tests unchecked, see \
1848             spec/04-driver-and-cli.md section 4.7. `-fsafety=detect` is the memory checking this \
1849             compiler does have"
1850        )));
1851    }
1852    link.sysroot = sysroot.clone();
1853    // Where a sysroot for a target that is not this machine would be. Read once, here, rather than
1854    // inside the link line, because a link line that read the environment could only be tested on a
1855    // machine whose environment said the right thing, and the link line is the last thing that
1856    // touches a binary. `spec/cross-compile/13-distribution.md` section 13.2 owns the answer.
1857    link.cache = Some(cache::dir());
1858    // And the ten field spelling of the target, because the release on it decides two things the
1859    // three field one cannot say: whether a target that is this architecture is still a cross
1860    // compile, and which directory under the cache it is against. After the loop because the last
1861    // `--target=` on the command line is the one that counts.
1862    link.pinned = pinned;
1863    // After the loop rather than where `-pthread` was read, so that it lands after the objects
1864    // that refer to it. A static link takes the definitions it needs from a library when it
1865    // reaches it and not afterwards, so a library before the objects is a library that answers
1866    // nothing.
1867    if threads {
1868        inputs.push(Input::library("pthread"));
1869    }
1870    if let Some(query) = query {
1871        return Ok(Action::Print(answer(&query, &opts, &link)?));
1872    }
1873    // `-M` and `-MM` produce the rule and nothing else, so the run stops after phase 4 whatever
1874    // else the command line asked for. Read here rather than where the flag was, because a `-c`
1875    // written after it has to lose and the loop cannot know that until it has ended. The output
1876    // file is where the rule goes rather than where an object would have gone, and the last
1877    // phase being the preprocessor is what makes that true without a second rule for it.
1878    if opts.deps.instead_of_compiling {
1879        opts.emit = EmitKind::Preprocessed;
1880    }
1881    if !nostdinc {
1882        opts.search.push_system(runtime::DIR);
1883        // And the library's after ours, which is the other half of the same order. They go on
1884        // here rather than at the point `--target=` or `--sysroot=` was read because either
1885        // one changes the answer and the last word on both is the end of the loop.
1886        //
1887        // Which library's is the question `link::cross_sysroot` answers, and it is asked here so
1888        // that the headers and the libraries come from the same place. A target that is this
1889        // machine reads this machine's headers, and a target that is not reads the ones in the
1890        // sysroot for it rather than the ones next door.
1891        let cross = link::cross_sysroot(opts.target, &link);
1892        let kernel = link::cross_kernel(opts.target, &link);
1893        // And the version of those headers, which only the bundled tree has an answer for. A host
1894        // glibc and a tree the user named both define `__GLIBC_MINOR__` in their own `features.h`,
1895        // and a second definition with a different value is a warning on every file, so the
1896        // condition is the same one that chose the directories.
1897        if cross.is_some() {
1898            let target = pinned.unwrap_or_else(|| opts.target.tuple());
1899            opts.glibc_minor = rucc_sysroot::bundled_glibc_minor(target).map_err(|skew| {
1900                err(format!(
1901                    "{skew}; pin a release the tree has, or name a tree that has that one \
1902                     with --sysroot"
1903                ))
1904            })?;
1905        }
1906        let system =
1907            library::header_dirs(opts.target, sysroot.as_deref(), cross.as_ref(), kernel.as_ref());
1908        // The two licence walls of `spec/cross-compile/13-distribution.md` section 13.4, which are
1909        // the only way step 3 comes back with nothing on a hosted target. Section 8.6 asks for the
1910        // answer to name the licence and the lawful ways to get what is behind it, rather than
1911        // leaving a person with an `#include` that failed as though a directory had gone missing.
1912        //
1913        // It is left on the search path instead of refused here, because a program that includes
1914        // none of the library needs none of the SDK and section 8.6 is explicit that targeting the
1915        // platform has to keep working. So the reason waits until an include has actually failed,
1916        // which is the only moment it helps and the only moment it is true.
1917        //
1918        // The condition is that step 3 found nothing at all, so an `SDKROOT`, an `INCLUDE` or a mac
1919        // with Xcode on it all pass through untouched, and `-nostdinc` never reaches this block. A
1920        // `--sysroot` or `-isysroot` passes through as well, even when the tree it names turns out to
1921        // be empty or absent: somebody who wrote a path has already answered the question this
1922        // message asks, and answering it again over the top of a mistyped directory would hide the
1923        // mistake behind a licence notice.
1924        if system.is_empty() && sysroot.is_none() {
1925            let tuple = pinned.unwrap_or_else(|| opts.target.tuple());
1926            if let Some(wall) = rucc_sysroot::Wall::of(tuple) {
1927                opts.search.explain_missing_system(wall.no_headers(&tuple.to_canonical_string()));
1928            }
1929        }
1930        // And whether the tree somebody named is the release they asked for, which is the one
1931        // question left once the directories are settled and the only place both halves of it are
1932        // known. Only for a named tree, because that is the case where the release in the target
1933        // stops deciding anything, and `crate::glibc` is where the rest of the reasoning is.
1934        if sysroot.is_some() {
1935            notes.extend(glibc::skew(opts.target, pinned, &system));
1936        }
1937        for dir in system {
1938            opts.search.push_system(dir);
1939        }
1940    }
1941    // Once, here, rather than as each directory is pushed. A `-I` that names a system
1942    // directory has to lose to the system entry and the system entry is added last, so the
1943    // question cannot be answered until the whole path is known.
1944    opts.search.remove_duplicates();
1945
1946    // The target has to be resolved before the configuration is printed, so this check comes
1947    // after the loop rather than at the point `--print-config` was seen.
1948    if print_config {
1949        return Ok(Action::PrintConfig(Box::new(opts)));
1950    }
1951    if print_pipeline {
1952        return Ok(Action::PrintPipeline(Box::new(opts)));
1953    }
1954    let plan = Plan::new(&opts, &inputs, output.as_deref()).map_err(|e| err(e.message))?;
1955    if print_plan {
1956        return Ok(Action::PrintPlan {
1957            opts: Box::new(opts),
1958            plan: Box::new(plan),
1959            link: Box::new(link),
1960        });
1961    }
1962    Ok(Action::Compile {
1963        opts: Box::new(opts),
1964        plan: Box::new(plan),
1965        link: Box::new(link),
1966        jobs,
1967        verbose,
1968        notes,
1969    })
1970}
1971
1972/// What `--fetch <tuple>` asked for, or why it is not a thing that can be done.
1973///
1974/// The lookup happens here rather than at the point the bytes would move, so that a target this
1975/// release pins nothing for is a refusal from the parser and the only code that runs a downloader is
1976/// code that already knows what it is getting.
1977///
1978/// # Errors
1979///
1980/// [`CliError`] when `--offline` forbade it, when there are input files as well, when the tuple is
1981/// not a target this compiler knows, when its sysroot is behind one of section 13.4's licence walls,
1982/// and when this release pins no artifact for it.
1983fn fetch_action(named: &str, offline: bool, inputs: &[Input]) -> Result<Action, CliError> {
1984    // Not a precedence question. Section 13.2 says `--offline` forbids a fetch entirely, so a
1985    // command line that writes both has asked for two opposite things and the answer is to say so
1986    // rather than to pick one of them.
1987    if offline {
1988        return Err(err(
1989            "--fetch asks for a download and --offline forbids every download, so this command \
1990             line asks for two opposite things. Drop one of them: --offline is how a build says it \
1991             will not reach the network, and --fetch is the only thing in this compiler that does",
1992        ));
1993    }
1994    if let Some(first) = inputs.first() {
1995        return Err(err(format!(
1996            "--fetch gets a sysroot and compiles nothing, so `{}` on the same command line is an \
1997             input that nothing would read",
1998            first.path
1999        )));
2000    }
2001    let target: TargetTuple = named
2002        .parse()
2003        .map_err(|why| err(format!("--fetch {named}: {why}, so there is no sysroot to get")))?;
2004    // The canonical spelling, because that is what a row is named by and what the directory under
2005    // the cache is called, and a person is free to write a tuple the long way round.
2006    let tuple = target.to_canonical_string();
2007    // Before the table is consulted, because a target behind a licence wall is not a row that has not
2008    // been written yet. Section 13.4 is that no release pins one of these ever, so the message says
2009    // the licence and the two lawful ways rather than naming the producer that will publish the rest.
2010    if let Some(wall) = rucc_sysroot::Wall::of(target) {
2011        return Err(err(format!("--fetch {tuple}: {}", wall.no_fetch(&tuple))));
2012    }
2013    let Some(what) = rucc_sysroot::pinned_for(&tuple) else {
2014        return Err(err(unpinned(&tuple)));
2015    };
2016    Ok(Action::Fetch { what, target, cache: cache::dir() })
2017}
2018
2019/// Why there is nothing to fetch for a target, which is a different sentence when the table is
2020/// empty.
2021///
2022/// A release that pins nothing and a release that pins eleven targets and not this one are two
2023/// situations, and a message that did not tell them apart would send somebody looking for a typo in
2024/// their tuple when the answer is that this work is not finished.
2025fn unpinned(tuple: &str) -> String {
2026    let pinned = rucc_sysroot::pinned_targets();
2027    if pinned.is_empty() {
2028        return format!(
2029            "this release pins no sysroot for {tuple}, and it pins none for any target yet. A \
2030             sysroot is built and published by the producer in tamnd/rucc-cross, per \
2031             spec/cross-compile/13-distribution.md section 13.8, and a release of this compiler \
2032             names one by URL and by hash afterwards. Until then, pass --sysroot=<dir> to compile \
2033             against a tree you have already"
2034        );
2035    }
2036    format!(
2037        "this release pins no sysroot for {tuple}. What it pins is {}. Pass --sysroot=<dir> to \
2038         compile against a tree you have already",
2039        pinned.join(", ")
2040    )
2041}
2042
2043/// Gets the artifact and installs it, saying what each step did.
2044///
2045/// The steps are section 13.8's and so are the messages: the transport is somebody else's program
2046/// and the check is ours, so a person reading this wants to know which downloader ran, that the
2047/// bytes matched, how many files the record named and where the tree ended up. A fetch of something
2048/// that is already there says that instead and moves nothing.
2049fn fetch_sysroot(what: &rucc_sysroot::Pinned, target: TargetTuple, cache: &std::path::Path) -> i32 {
2050    let tuple = target.to_canonical_string();
2051    let archive = what.archive_in(cache);
2052    let say = |line: &str| println!("rucc: {tuple}: {line}");
2053    match fetch::fetch(what.url, what.sha256, &archive) {
2054        Ok(fetch::Fetched::AlreadyThere) => {
2055            say(&format!("{} is already here and matches the hash", archive.display()));
2056        }
2057        Ok(fetch::Fetched::Downloaded(by)) => {
2058            say(&format!("downloaded {} with {}", what.url, by.program()));
2059        }
2060        Err(why) => return complain(why),
2061    }
2062    match install::install(&archive, what.sha256, target, cache) {
2063        Ok(done) => {
2064            match &done.before {
2065                install::Before::Nothing => {
2066                    say(&format!("{} files installed at {}", done.files, done.root.display()));
2067                }
2068                install::Before::TheSame => {
2069                    say(&format!(
2070                        "the same sysroot is already at {}, so nothing moved",
2071                        done.root.display()
2072                    ));
2073                }
2074                install::Before::Different(was) => {
2075                    say(&format!(
2076                        "{} files installed at {}, over a tree whose record digested to {was}",
2077                        done.files,
2078                        done.root.display()
2079                    ));
2080                }
2081            }
2082            say(&format!("the record digests to {}", done.digest));
2083            0
2084        }
2085        Err(why) => complain(why),
2086    }
2087}
2088
2089/// What one of the `-dump` and `-print` flags prints.
2090///
2091/// GCC prints the name back unchanged when it cannot find the file a `-print` flag asked about,
2092/// which is what makes the answer safe to paste into a link line whether or not the file is
2093/// there, and this does the same.
2094fn answer(query: &Query, opts: &Options, link: &LinkOptions) -> Result<String, CliError> {
2095    let found = |name: &str| {
2096        link::find_in_search(link, opts.target, name)
2097            .map_or_else(|| name.to_owned(), |path| path.display().to_string())
2098    };
2099    Ok(match query {
2100        Query::Machine => opts.target.to_string(),
2101        Query::Version => VERSION.to_owned(),
2102        Query::Multiarch => link::multiarch(opts.target),
2103        // The three lines GCC prints, in its order and with its punctuation, because what reads
2104        // them is a script written against that shape. There is no installation directory to
2105        // report: this compiler is one binary that works wherever it is copied, and the headers
2106        // it ships are inside it, so `install` is where the binary is and nothing is under it.
2107        Query::SearchDirs => {
2108            let here = std::env::current_exe()
2109                .ok()
2110                .and_then(|p| p.parent().map(std::path::Path::to_path_buf))
2111                .unwrap_or_default();
2112            let list = |dirs: &[PathBuf]| {
2113                dirs.iter().map(|d| d.display().to_string()).collect::<Vec<_>>().join(":")
2114            };
2115            let libraries = link::search_dirs(link, opts.target);
2116            format!(
2117                "install: {}\nprograms: ={}\nlibraries: ={}",
2118                here.display(),
2119                list(&link.prefixes),
2120                list(&libraries)
2121            )
2122        }
2123        // The root the rest of the answers are under, which a build system asks for when it wants
2124        // to find a file itself rather than ask for one by name, and which is the first thing to
2125        // look at when a cross build read a header nobody expected. A native compile has no
2126        // sysroot and the answer is the empty line, which is what GCC prints when it was
2127        // configured without one. `--sysroot` wins over ours because it wins everywhere else.
2128        Query::Sysroot => {
2129            sysroot_root(opts, link).map(|root| root.display().to_string()).unwrap_or_default()
2130        }
2131        // Section 13.5 of `spec/cross-compile/13-distribution.md`: for every input that is not this
2132        // compiler's own code, what it is, where it was got, its hash, its licence and whether it
2133        // was bundled, generated or fetched. What is printed is the manifest the sysroot already
2134        // carries rather than a second format saying the same things, because the three uses 13.5
2135        // gives for this are a licence notice, a reproducibility check and a security audit, and all
2136        // three are somebody else parsing it. One format is one parser to write.
2137        // Read and rendered rather than copied out, so that what comes back is the format this
2138        // build understands. The last newline comes off because whatever prints an answer adds
2139        // one, the way it does for every other query here. Keeping it would put a blank line at
2140        // the end of the one answer that is a file somebody diffs against the file it came from.
2141        Query::SysrootProvenance => match sysroot_manifest(opts, link)? {
2142            Some(manifest) => manifest.render().trim_end_matches('\n').to_string(),
2143            None => String::new(),
2144        },
2145        // Section 13.2 of the same document, which asks for the hash of a cache directory's
2146        // contents in the directory's name. A name cannot carry one, because the path has to be
2147        // computable before anything has been read, by the producer about to write the files and by
2148        // the compiler about to read them, and neither has the contents when it asks. So the number
2149        // is here instead, and it is the sha256 of the record rather than of a walk of the tree,
2150        // which means `sha256sum` over the manifest answers the same thing.
2151        Query::SysrootDigest => match sysroot_manifest(opts, link)? {
2152            Some(manifest) => manifest.digest(),
2153            None => String::new(),
2154        },
2155        Query::FileName(name) => found(name),
2156        // The name GCC gives the library of routines a compiler's output calls that the C
2157        // library does not have. Ours is built in and there is no file, so the answer is the
2158        // name itself, which is what GCC prints when it cannot find one either.
2159        Query::Libgcc => found("libgcc.a"),
2160        // A program rather than a library: the linker and the archiver are the ones a build asks
2161        // about, and this compiler finds them on the path or under `-B` rather than shipping
2162        // them, so the name back is the honest answer unless a `-B` prefix holds one.
2163        Query::ProgName(name) => link
2164            .prefixes
2165            .iter()
2166            .map(|dir| dir.join(name))
2167            .find(|path| path.is_file())
2168            .map_or_else(|| name.clone(), |path| path.display().to_string()),
2169    })
2170}
2171
2172/// The root every sysroot answer is about.
2173///
2174/// One function rather than a copy in each, because the other flags exist to say what is inside the
2175/// tree this one names, and two answers that disagreed about which tree that is would be a
2176/// difference nobody would think to look for. `--sysroot` wins over ours because it wins everywhere
2177/// else.
2178fn sysroot_root(opts: &Options, link: &LinkOptions) -> Option<PathBuf> {
2179    link.sysroot
2180        .clone()
2181        .or_else(|| link::cross_sysroot(opts.target, link).map(|at| at.root().to_path_buf()))
2182}
2183
2184/// The record of the sysroot this command line reads, when there is one to read.
2185///
2186/// [`None`] covers two cases that both print nothing, and they are different things. A compile for
2187/// this machine has no sysroot at all, and a tree somebody laid out themselves and pointed
2188/// `--sysroot` at carries no manifest, so nothing here knows where any of it came from. Saying
2189/// nothing is the only honest answer to either, and a reader can tell it from a manifest with no
2190/// inputs in it because that one still has its header lines.
2191///
2192/// # Errors
2193///
2194/// A manifest this build cannot parse, and anything else that went wrong reading the file. Passing a
2195/// record we could not read on to whoever asked would make their parser the one that finds the
2196/// problem, and every use section 13.5 gives for these two flags is somebody else reading the
2197/// output.
2198fn sysroot_manifest(opts: &Options, link: &LinkOptions) -> Result<Option<Manifest>, CliError> {
2199    let Some(root) = sysroot_root(opts, link) else {
2200        return Ok(None);
2201    };
2202    let path = Sysroot::at(root, opts.target.tuple()).manifest_path();
2203    match std::fs::read_to_string(&path) {
2204        Ok(text) => Manifest::parse(&text)
2205            .map(Some)
2206            .map_err(|why| err(format!("{}: {why}", path.display()))),
2207        Err(why) if why.kind() == std::io::ErrorKind::NotFound => Ok(None),
2208        Err(why) => Err(err(format!("{}: {why}", path.display()))),
2209    }
2210}
2211
2212/// Renders the passes this level will run, in order, with what each one does.
2213///
2214/// The level is the whole of the answer unless a `-f` flag edited it, which is section 9.1 of
2215/// `spec/09-optimizer.md`: a level is a list somebody wrote down rather than something that
2216/// emerges from which flags happen to be set, and this is how that list is read.
2217#[must_use]
2218pub fn print_pipeline(opts: &Options) -> String {
2219    let mut settings = rucc_opt::Options::for_level(opts.opt_level);
2220    settings.toggles.clone_from(&opts.passes);
2221    settings.global_fuel = opts.pass_fuel_global;
2222    for (on, spec) in &opts.pass_gates {
2223        // Every spelling was checked while the arguments were parsed, so there is nothing here
2224        // this can refuse, and a listing is not the place to report it if there were.
2225        let _ = settings.gates.add(*on, spec);
2226    }
2227    rucc_opt::pipeline::print(&settings)
2228}
2229
2230/// Renders the resolved configuration.
2231///
2232/// One `key: value` per line, sorted by nothing in particular but fixed in order, because
2233/// this output is diffed across hosts in CI and a reordering would read as a change.
2234#[must_use]
2235pub fn print_config(opts: &Options) -> String {
2236    let sess = Session::new(opts.clone());
2237    let t = &sess.target;
2238    let mut out = String::new();
2239    let _ = writeln!(out, "version: {VERSION}");
2240    // The three field triple the driver was given rather than the ten field tuple it widens to,
2241    // because this output is what a build system reads to find out what it asked for. The tuple is
2242    // the compiler's model of the machine and this line is a receipt for a command line.
2243    let _ = writeln!(out, "target: {}", opts.target);
2244    let _ = writeln!(out, "arch: {}", opts.target.arch.as_str());
2245    let _ = writeln!(out, "os: {}", opts.target.os.as_str());
2246    let _ = writeln!(out, "env: {}", opts.target.env.as_str());
2247    let _ = writeln!(out, "object-format: {}", t.object_format.as_str());
2248    let _ = writeln!(out, "pointer-width: {}", t.pointer_width);
2249    let _ = writeln!(out, "long-width: {}", t.long_width);
2250    let _ = writeln!(out, "long-double-width: {}", t.long_double_width);
2251    let _ = writeln!(out, "endian: {}", if t.little_endian { "little" } else { "big" });
2252    let _ = writeln!(out, "char-signed: {}", t.char_is_signed);
2253    let _ = writeln!(out, "va-list: {}", t.va_list.map_or("none", |list| list.as_str()));
2254    // The register file as a count per class, which is enough to tell a target whose registers
2255    // are described from one whose are not without printing sixteen names nobody asked for.
2256    let regs: Vec<String> = t
2257        .regs
2258        .classes()
2259        .map(|(class, info)| format!("{} {}", info.name, t.regs.len(class)))
2260        .collect();
2261    let _ = writeln!(
2262        out,
2263        "registers: {}",
2264        if regs.is_empty() { "none".to_string() } else { regs.join(", ") }
2265    );
2266    // What the schedule was chosen with, which is a sentence rather than a name on purpose: two
2267    // runs of a benchmark that disagree are usually two models and not two compilers.
2268    let _ = writeln!(out, "timing-model: {}", t.timing.map_or("none", |timing| timing.model));
2269    let _ = writeln!(out, "opt-level: {}", sess.opts.opt_level);
2270    let _ = writeln!(out, "safety: {}", sess.opts.safety);
2271    let _ = writeln!(out, "emit: {}", sess.opts.emit.as_str());
2272    let _ = writeln!(out, "debug-info: {}", sess.opts.debug_info);
2273    let _ = writeln!(out, "frame-pointer: {}", sess.opts.frame_pointer);
2274    let _ = writeln!(out, "red-zone: {}", sess.opts.red_zone);
2275    let _ = writeln!(out, "stack-protector: {}", sess.opts.protector);
2276    let _ = writeln!(out, "stack-clash-protection: {}", sess.opts.stack_clash);
2277    let _ = writeln!(out, "cf-protection: {}", sess.opts.control);
2278    let _ = writeln!(out, "patchable-function-entry: {}", sess.opts.patchable);
2279    let _ = writeln!(out, "profile: {}", sess.opts.profile);
2280    let _ = writeln!(out, "profile-hook: {}", sess.opts.hook);
2281    // Last because it is the one key with more than one line under it, and the only one
2282    // whose value is a property of the machine rather than of the command line.
2283    for dir in sess.opts.search.dirs() {
2284        let system = if dir.is_system { " (system)" } else { "" };
2285        let _ = writeln!(out, "include: {}{system}", dir.path.display());
2286    }
2287    out
2288}
2289
2290/// The output name the make target is taken from, which is the `-o` argument or nothing.
2291///
2292/// A run that stops at the preprocessor has not named an object, whatever its `-o` says: under
2293/// `-E` that argument is the preprocessed text and under `-M` it is the rule itself, and neither
2294/// is a file `make` would rebuild by running this rule. GCC agrees and falls back to the source
2295/// name in both, which is why a `-MD -E -o out.i` writes `out.d` holding a rule for `a.o`. From
2296/// `-S` on the argument does name what the rule builds, and it is used as written.
2297fn deps_target_output<'a>(opts: &Options, plan: &'a Plan) -> Option<&'a str> {
2298    if opts.emit == EmitKind::Preprocessed { None } else { plan.output.as_deref() }
2299}
2300
2301/// Writes to a path the command line named rather than one the plan derived, where `-` is
2302/// standard output.
2303fn write_named(path: &str, bytes: &[u8]) -> Result<(), String> {
2304    if path == "-" {
2305        return write_out(&Output::Stdout, bytes);
2306    }
2307    write_out(&Output::File(path.to_owned()), bytes)
2308}
2309
2310/// Writes the make rule for one input, and reports whether it got there.
2311///
2312/// A rule with no file of its own goes where the compilation it replaced would have written,
2313/// which is what makes the usual makefile recipe work: `rucc -M $< -o $@` leaves the rule in
2314/// `$@`, and the same line with the `-o` left off puts it on standard output.
2315fn write_deps(
2316    opts: &Options,
2317    plan: &Plan,
2318    job: &Job,
2319    found: &[Dependency],
2320    stderr: &mut impl std::io::Write,
2321) -> bool {
2322    let targets = if opts.deps.targets.is_empty() {
2323        vec![deps::default_target(&job.input, deps_target_output(opts, plan))]
2324    } else {
2325        opts.deps.targets.clone()
2326    };
2327    let rule = deps::rule(&opts.deps, &targets, &job.input, found);
2328    // The file, on the other hand, is named after the `-o` in every mode that still has one to
2329    // spend, which is every mode except the two that spend it on the rule.
2330    let wrote = match deps::default_file(&opts.deps, &job.input, plan.output.as_deref()) {
2331        // A `-MF` on a run that had nowhere else to put the rule leaves the file the `-o`
2332        // named empty rather than absent, because a makefile that named it as a target of its
2333        // own is a makefile that will look for it.
2334        Some(path) => write_named(&path, rule.as_bytes()).and_then(|()| {
2335            if opts.deps.instead_of_compiling { write_out(&job.output, b"") } else { Ok(()) }
2336        }),
2337        None => write_out(&job.output, rule.as_bytes()),
2338    };
2339    if let Err(e) = wrote {
2340        let _ = writeln!(stderr, "rucc: error: {e}");
2341        return false;
2342    }
2343    true
2344}
2345
2346/// Runs phase 4 over every input that has one, and writes what came out.
2347///
2348/// One input that fails does not stop the others. A build that reports every file it could
2349/// not preprocess in one run is worth more than one that stops at the first, and the exit
2350/// status is still a failure either way.
2351fn preprocess_all(opts: &Options, plan: &Plan) -> i32 {
2352    let fs = OsFileSystem::new();
2353    let mut stderr = std::io::stderr().lock();
2354    let mut failed = false;
2355    for job in &plan.jobs {
2356        if !job.phases.first().is_some_and(|p| *p == Phase::Preprocess) {
2357            // An input that is already preprocessed, or an object file. GCC passes these
2358            // through untouched, and the plan has already said so in its notes.
2359            continue;
2360        }
2361        let started = std::time::Instant::now();
2362        let result = preprocess(opts, &job.input, &fs);
2363        if opts.time {
2364            say_time(&job.input, started.elapsed(), &mut stderr);
2365        }
2366        for message in &result.messages {
2367            let _ = writeln!(stderr, "{message}");
2368        }
2369        if result.failed() {
2370            failed = true;
2371            continue;
2372        }
2373        if opts.deps.emit {
2374            failed |= !write_deps(opts, plan, job, &result.deps, &mut stderr);
2375            // `-M` and `-MM` asked for the rule instead of the text, so there is nothing else
2376            // to write. The other two asked for both and fall through to the text below.
2377            if opts.deps.instead_of_compiling {
2378                continue;
2379            }
2380        }
2381        if let Err(e) = write_out(&job.output, result.text.as_bytes()) {
2382            let _ = writeln!(stderr, "rucc: error: {e}");
2383            failed = true;
2384        }
2385    }
2386    i32::from(failed)
2387}
2388
2389/// Whether this job is a file of assembly that has to be assembled and that nothing here assembles.
2390///
2391/// The phases rather than the kind, because there are two kinds of assembly input and one of them
2392/// is preprocessed first, and because an object file also has no compile phase and is not this: it
2393/// has no phases at all and goes to the linker as it is. A `.s` on a `-c` line has exactly
2394/// [`Phase::Assemble`] left, and a `.S` has the preprocessor in front of it, and neither has
2395/// anything the front end can do.
2396fn needs_an_assembler(job: &Job) -> bool {
2397    job.phases.contains(&Phase::Assemble) && !job.phases.contains(&Phase::Compile)
2398}
2399
2400/// Whether the preprocessor runs over it on the way in, which is the whole difference between the
2401/// two kinds of assembly input.
2402fn assembly_wants_cpp(job: &Job) -> bool {
2403    job.phases.contains(&Phase::Preprocess)
2404}
2405
2406/// Runs the front end over every input that has a compile phase, and writes what came out.
2407///
2408/// The same rule as [`preprocess_all`]: one input that fails does not stop the others, and the
2409/// exit status is a failure either way. An input that is already assembly or an object has no
2410/// compile phase and is passed over here, which the plan has already said in its notes.
2411fn compile_all(opts: &Options, plan: &Plan) -> i32 {
2412    let fs = OsFileSystem::new();
2413    let mut stderr = std::io::stderr().lock();
2414    let mut failed = false;
2415    let (mut remarks, ok) = Remarks::new(opts.opt_info_file.as_ref(), &mut stderr);
2416    failed |= !ok;
2417    let mut fired = Fired::new();
2418    let mut pressure = Pressure::new();
2419    let mut lowerings = Lowerings::new();
2420    for job in &plan.jobs {
2421        if !job.phases.contains(&Phase::Compile) && !needs_an_assembler(job) {
2422            continue;
2423        }
2424        // An input of IR is read back rather than compiled, since the C it came from is not
2425        // here any more. A file of assembly does not go through the front end at all and is
2426        // read by the assembler instead. Everything after this is the same for all three, so
2427        // the paths meet again at the messages and the file the result is written to.
2428        let started = std::time::Instant::now();
2429        let result = if needs_an_assembler(job) {
2430            assemble(opts, &job.input, assembly_wants_cpp(job), &fs)
2431        } else if job.kind == InputKind::Ir {
2432            compile_ir(opts, &job.input, &fs)
2433        } else {
2434            compile(opts, &job.input, &fs)
2435        };
2436        if opts.time {
2437            say_time(&job.input, started.elapsed(), &mut stderr);
2438        }
2439        fired.merge(&result.fired);
2440        pressure.merge(&result.pressure);
2441        lowerings.merge(&result.lowerings);
2442        failed |= !write_dumps(&job.input, &result.dumps, &mut stderr);
2443        failed |= !remarks.write(&result.remarks, &mut stderr);
2444        for message in &result.messages {
2445            let _ = writeln!(stderr, "{message}");
2446        }
2447        // Before the failure below, because a compilation that stopped in the back end is exactly
2448        // the one whose preprocessed source somebody wants to look at.
2449        failed |= !write_temps(job, &result.temps, &mut stderr);
2450        if result.failed() {
2451            failed = true;
2452            continue;
2453        }
2454        // `-MD` and `-MMD` write the rule beside the object and let the compilation happen, so
2455        // this is the one path where both files come out of the same run. An input of IR has no
2456        // dependencies to report and produces an empty list, which produces a rule naming only
2457        // itself, and that is the honest answer rather than a missing file.
2458        if opts.deps.emit {
2459            failed |= !write_deps(opts, plan, job, &result.deps, &mut stderr);
2460        }
2461        if let Err(e) = write_out(&job.output, result.artifact.bytes()) {
2462            let _ = writeln!(stderr, "rucc: error: {e}");
2463            failed = true;
2464        }
2465    }
2466    failed |= !write_coverage(opts, &fired, &mut stderr);
2467    failed |= !write_pressure(opts, &pressure, &mut stderr);
2468    failed |= !write_lowering(opts, &lowerings, &mut stderr);
2469    i32::from(failed)
2470}
2471
2472/// A directory for the object files only the link step ever sees, removed when it goes away.
2473///
2474/// `-c` writes its object where the user can see it and linking does not, which is the whole of
2475/// the difference: a `rucc a.c b.c` leaves an executable behind and nothing else, the same as
2476/// every other compiler. Removing them on drop rather than at the end of a function is so that a
2477/// link that failed leaves nothing behind either.
2478struct Scratch {
2479    /// Where the objects go.
2480    dir: PathBuf,
2481}
2482
2483impl Scratch {
2484    /// Makes one, under whatever the platform calls its temporary directory.
2485    ///
2486    /// The name carries the process id so that two compilers running at once do not share a
2487    /// directory, which they would otherwise do the moment two of them compiled a file of the
2488    /// same name.
2489    fn new() -> Result<Scratch, String> {
2490        let dir = std::env::temp_dir().join(format!("rucc-{}", std::process::id()));
2491        std::fs::create_dir_all(&dir).map_err(|e| format!("{}: {e}", dir.display()))?;
2492        Ok(Scratch { dir })
2493    }
2494}
2495
2496impl Drop for Scratch {
2497    fn drop(&mut self) {
2498        let _ = std::fs::remove_dir_all(&self.dir);
2499    }
2500}
2501
2502/// The link line the plan describes, for `-###`.
2503///
2504/// The names in it are the hints the plan carries rather than the temporaries a real compilation
2505/// would choose, because `-###` prints the line without having compiled anything and so has
2506/// nothing to point at. That also makes the printed line readable rather than naming a directory
2507/// that only exists while a compilation is running.
2508fn link_line(opts: &Options, link: &LinkOptions, job: &LinkJob) -> Result<String, link::Error> {
2509    let linker = link::find(opts.target, link)?;
2510    let args = link::line(opts.target, link, &job.inputs, &job.output)?;
2511    Ok(link::render(&linker, &args))
2512}
2513
2514/// Compiles everything, then links it.
2515///
2516/// The objects go in a directory that is removed afterwards, which is why this is not
2517/// [`compile_all`] followed by a link: the plan says an object feeding the linker is temporary
2518/// and does not say where, because where is a question that only has an answer once something is
2519/// running.
2520fn link_all(opts: &Options, plan: &Plan, link: &LinkOptions, verbose: bool) -> i32 {
2521    let Some(job) = &plan.link else {
2522        // Every path into here comes from a plan whose last phase is the link, and such a plan
2523        // has a link job. Saying so is cheaper than an unwrap that would have to be explained.
2524        let mut stderr = std::io::stderr().lock();
2525        let _ = writeln!(stderr, "rucc: error: there is nothing to link");
2526        return 1;
2527    };
2528    // Before anything is compiled, because a linker that is not on the machine is worth knowing
2529    // about in the second it takes to look rather than after the compilation.
2530    // And before that, whether this link has a line at all and whether what it reads is on the
2531    // machine. Both are answerable now, and a target whose sysroot has not been built is worth
2532    // saying so about before the compilation rather than after it.
2533    if let Err(why) = link::preflight(opts.target, link) {
2534        return complain(why);
2535    }
2536    let linker = match link::find(opts.target, link) {
2537        Ok(linker) => linker,
2538        Err(why) => return complain(why),
2539    };
2540    // And whether the one that was found can do this link, which for one linker and one target is
2541    // a question only the linker itself can answer. Here rather than inside the search, because
2542    // what it does is refuse rather than move on to the next candidate: nothing else in the list
2543    // links a produced Windows sysroot either.
2544    if let Err(why) = link::suitable(opts.target, &linker) {
2545        return complain(why);
2546    }
2547
2548    let scratch = match Scratch::new() {
2549        Ok(scratch) => scratch,
2550        Err(why) => return complain(format!("could not make a place for the object files: {why}")),
2551    };
2552
2553    let fs = OsFileSystem::new();
2554    let mut failed = false;
2555    // One per job, in job order, which is what lets the link line below be rebuilt with the real
2556    // paths in it: every job contributes exactly one file to the line and does so in this order.
2557    let mut produced: Vec<String> = Vec::with_capacity(plan.jobs.len());
2558    let mut fired = Fired::new();
2559    let mut pressure = Pressure::new();
2560    let mut lowerings = Lowerings::new();
2561    {
2562        let mut stderr = std::io::stderr().lock();
2563        let (mut remarks, ok) = Remarks::new(opts.opt_info_file.as_ref(), &mut stderr);
2564        failed |= !ok;
2565        for (at, job) in plan.jobs.iter().enumerate() {
2566            let out = match &job.output {
2567                Output::Temporary(hint) => {
2568                    // The index because two inputs in different directories can have the same
2569                    // name, and the two objects of `rucc a/x.c b/x.c` must not be one file.
2570                    scratch.dir.join(format!("{at}-{hint}")).display().to_string()
2571                }
2572                Output::File(path) => path.clone(),
2573                // A job feeding the linker never writes to standard output, since the plan gives
2574                // it a temporary. This is here so that the match is total rather than a panic.
2575                Output::Stdout => continue,
2576            };
2577            produced.push(out.clone());
2578            if !job.phases.contains(&Phase::Compile) && !needs_an_assembler(job) {
2579                continue;
2580            }
2581            let started = std::time::Instant::now();
2582            let result = if needs_an_assembler(job) {
2583                assemble(opts, &job.input, assembly_wants_cpp(job), &fs)
2584            } else if job.kind == InputKind::Ir {
2585                compile_ir(opts, &job.input, &fs)
2586            } else {
2587                compile(opts, &job.input, &fs)
2588            };
2589            if opts.time {
2590                say_time(&job.input, started.elapsed(), &mut stderr);
2591            }
2592            fired.merge(&result.fired);
2593            pressure.merge(&result.pressure);
2594            lowerings.merge(&result.lowerings);
2595            failed |= !write_dumps(&job.input, &result.dumps, &mut stderr);
2596            failed |= !remarks.write(&result.remarks, &mut stderr);
2597            for message in &result.messages {
2598                let _ = writeln!(stderr, "{message}");
2599            }
2600            failed |= !write_temps(job, &result.temps, &mut stderr);
2601            if result.failed() {
2602                failed = true;
2603                continue;
2604            }
2605            // A `-MD` on a command line that links writes the rule next to the executable and
2606            // names the executable as its target, since that is the file this source builds
2607            // here. The object it went through is in a temporary directory and is gone by the
2608            // time `make` reads any of this.
2609            if opts.deps.emit {
2610                failed |= !write_deps(opts, plan, job, &result.deps, &mut stderr);
2611            }
2612            if !matches!(result.artifact, Artifact::Object { .. }) {
2613                // Worth saying rather than writing whatever it is and letting the linker read it.
2614                // An empty file is a valid empty linker script, so a link handed one gets as far
2615                // as reporting every symbol of this file undefined, which is a page of messages
2616                // about something that went wrong here.
2617                let _ = writeln!(
2618                    stderr,
2619                    "rucc: internal error: {}: no object file was produced for the link",
2620                    job.input
2621                );
2622                failed = true;
2623                continue;
2624            }
2625            if let Err(e) = std::fs::write(&out, result.artifact.bytes()) {
2626                let _ = writeln!(stderr, "rucc: error: {out}: {e}");
2627                failed = true;
2628            }
2629        }
2630        failed |= !write_coverage(opts, &fired, &mut stderr);
2631        failed |= !write_pressure(opts, &pressure, &mut stderr);
2632        failed |= !write_lowering(opts, &lowerings, &mut stderr);
2633        failed |= !write_lowering(opts, &lowerings, &mut stderr);
2634    }
2635    if failed {
2636        // Nothing is linked from a compilation that did not finish. A linker run over the objects
2637        // that did compile would report every function of the file that did not as undefined,
2638        // which is a page of messages about a mistake already reported once.
2639        return 1;
2640    }
2641
2642    // The items in command line order with the temporaries filled in. A library and a word for the
2643    // linker contribute no job and pass through, and every file item takes the next job's real
2644    // output, which is what keeps whatever was written between two objects between them here.
2645    let mut outputs = produced.into_iter();
2646    let mut items = Vec::with_capacity(job.inputs.len());
2647    for item in &job.inputs {
2648        match item {
2649            link::Item::Library(name) => items.push(link::Item::Library(name.clone())),
2650            link::Item::Linker(arg) => items.push(link::Item::Linker(arg.clone())),
2651            link::Item::File(_) => match outputs.next() {
2652                Some(path) => items.push(link::Item::File(path)),
2653                None => return complain("the plan asks the linker for a file nothing produced"),
2654            },
2655        }
2656    }
2657
2658    let args = match link::line(opts.target, link, &items, &job.output) {
2659        Ok(args) => args,
2660        Err(why) => return complain(why),
2661    };
2662    if verbose {
2663        let mut stderr = std::io::stderr().lock();
2664        let _ = writeln!(stderr, "{}", link::render(&linker, &args));
2665    }
2666    let started = std::time::Instant::now();
2667    let ran = link::run(&linker, &args);
2668    if opts.time {
2669        // The one step of a compilation that really is another program, so this line is the same
2670        // measurement gcc's is and names the linker the way gcc names `collect2`.
2671        let mut stderr = std::io::stderr().lock();
2672        say_time(&linker.name, started.elapsed(), &mut stderr);
2673    }
2674    match ran {
2675        Ok(()) => 0,
2676        // The linker has already said what was wrong on its own error output, and repeating that
2677        // linking failed would only push its message further up the screen.
2678        Err(link::Error::Refused { .. }) => 1,
2679        Err(why) => complain(why),
2680    }
2681}
2682
2683/// Compiles everything and writes the objects into one static library.
2684///
2685/// No temporary directory and no second program. The objects never reach the file system at all:
2686/// they go from the compiler into the archive writer, which is both faster than writing a directory
2687/// of files for an `ar` to read back and the reason the symbol index can be written at all. A
2688/// member's index entries are the names the object writer says it wrote, and the only thing that
2689/// knows those is the run that wrote it.
2690///
2691/// `-save-temps` is the exception. It asked for the objects to be kept, the plan gave them names a
2692/// person can find, and they are written there as well as put in the archive.
2693fn archive_all(opts: &Options, plan: &Plan) -> i32 {
2694    let Some(job) = &plan.archive else {
2695        // Every path into here comes from a plan whose last phase is the archive, and such a plan
2696        // has an archive job. Saying so is cheaper than an unwrap that would have to be explained.
2697        return complain("there is nothing to put in an archive");
2698    };
2699    // Before anything is compiled, because a format this has no container for is worth knowing
2700    // about in the second it takes to look rather than after the whole compilation.
2701    let flavour = match opts.target.os.object_format() {
2702        ObjectFormat::Elf => rucc_archive::Flavour::Gnu,
2703        ObjectFormat::Coff => rucc_archive::Flavour::Coff,
2704        // Mach-O wants the BSD flavour, whose index is a different member under a different name,
2705        // and wasm has no archives of its own at all. Neither has an object writer either, so a
2706        // command line reaching this would have failed in the next step regardless.
2707        format @ (ObjectFormat::MachO | ObjectFormat::Wasm) => {
2708            return complain(format!(
2709                "there is no archive format for {} objects in this compiler yet",
2710                format.as_str()
2711            ));
2712        }
2713    };
2714
2715    let fs = OsFileSystem::new();
2716    let mut failed = false;
2717    let mut members: Vec<rucc_archive::Member> = Vec::with_capacity(plan.jobs.len());
2718    let mut names = job.members.iter();
2719    let mut fired = Fired::new();
2720    let mut pressure = Pressure::new();
2721    let mut lowerings = Lowerings::new();
2722    {
2723        let mut stderr = std::io::stderr().lock();
2724        let (mut remarks, ok) = Remarks::new(opts.opt_info_file.as_ref(), &mut stderr);
2725        failed |= !ok;
2726        for plan_job in &plan.jobs {
2727            // What the plan called this member. The two lists are walked together rather than the
2728            // name being worked out again here, so that what `-###` printed and what goes in the
2729            // file cannot come apart.
2730            let Some(member) = names.next() else {
2731                return complain("the plan asks the archive for a member nothing produced");
2732            };
2733            if !plan_job.phases.contains(&Phase::Compile) && !needs_an_assembler(plan_job) {
2734                // Neither something to compile nor something to assemble, so there is nothing to
2735                // put in, and an archive quietly missing a member is worse than a message.
2736                let _ = writeln!(
2737                    &mut stderr,
2738                    "rucc: error: {}: this compiler makes an archive out of what it compiles, and \
2739                     there is nothing here for it to do",
2740                    plan_job.input
2741                );
2742                failed = true;
2743                continue;
2744            }
2745            let started = std::time::Instant::now();
2746            let result = if needs_an_assembler(plan_job) {
2747                assemble(opts, &plan_job.input, assembly_wants_cpp(plan_job), &fs)
2748            } else if plan_job.kind == InputKind::Ir {
2749                compile_ir(opts, &plan_job.input, &fs)
2750            } else {
2751                compile(opts, &plan_job.input, &fs)
2752            };
2753            if opts.time {
2754                say_time(&plan_job.input, started.elapsed(), &mut stderr);
2755            }
2756            fired.merge(&result.fired);
2757            pressure.merge(&result.pressure);
2758            lowerings.merge(&result.lowerings);
2759            failed |= !write_dumps(&plan_job.input, &result.dumps, &mut stderr);
2760            failed |= !remarks.write(&result.remarks, &mut stderr);
2761            for message in &result.messages {
2762                let _ = writeln!(stderr, "{message}");
2763            }
2764            failed |= !write_temps(plan_job, &result.temps, &mut stderr);
2765            if result.failed() {
2766                failed = true;
2767                continue;
2768            }
2769            if opts.deps.emit {
2770                failed |= !write_deps(opts, plan, plan_job, &result.deps, &mut stderr);
2771            }
2772            let Artifact::Object { bytes, defines } = result.artifact else {
2773                let _ = writeln!(
2774                    stderr,
2775                    "rucc: internal error: {}: no object file was produced for the archive",
2776                    plan_job.input
2777                );
2778                failed = true;
2779                continue;
2780            };
2781            // Under `-save-temps` the plan gave the object a name a person can find, so it is
2782            // written there too. Otherwise it is only ever a member and never a file.
2783            if let Output::File(path) = &plan_job.output {
2784                if let Err(e) = std::fs::write(path, &bytes) {
2785                    let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2786                    failed = true;
2787                }
2788            }
2789            members.push(rucc_archive::Member { name: member.clone(), body: bytes, defines });
2790        }
2791        failed |= !write_coverage(opts, &fired, &mut stderr);
2792        failed |= !write_pressure(opts, &pressure, &mut stderr);
2793        failed |= !write_lowering(opts, &lowerings, &mut stderr);
2794        failed |= !write_lowering(opts, &lowerings, &mut stderr);
2795    }
2796    if failed {
2797        // Nothing is written from a compilation that did not finish, for the reason the link gives:
2798        // an archive missing the file that failed is one a link reports every name of as undefined,
2799        // which is a page of messages about a mistake already reported once.
2800        return 1;
2801    }
2802
2803    let bytes = match rucc_archive::write(flavour, &members) {
2804        Ok(bytes) => bytes,
2805        // Every one of these is a bug here rather than a program's mistake: the names came from the
2806        // object writer and the bodies came from this process.
2807        Err(why) => return complain(format!("the archive could not be written: {why}")),
2808    };
2809    match std::fs::write(&job.output, &bytes) {
2810        Ok(()) => 0,
2811        Err(e) => complain(format!("{}: {e}", job.output)),
2812    }
2813}
2814
2815/// Prints one driver level message and gives back the exit status that goes with it.
2816fn complain(why: impl std::fmt::Display) -> i32 {
2817    let mut stderr = std::io::stderr().lock();
2818    let _ = writeln!(stderr, "rucc: error: {why}");
2819    1
2820}
2821
2822/// Writes what `-Zrule-coverage=FILE` asked for, and says whether it could.
2823///
2824/// Once for the whole command line rather than once per input, because the question is which
2825/// lowering rules this run of the compiler reached and a file per input would leave the reader
2826/// unioning files to find out something one process already knew.
2827///
2828/// A file that could not be written is a failure and not a warning. What asks for this is a
2829/// measurement run, and a measurement that quietly did not happen is worse than one that stopped.
2830fn write_coverage(opts: &Options, fired: &Fired, stderr: &mut impl std::io::Write) -> bool {
2831    let Some(path) = &opts.rule_coverage else { return true };
2832    let Some(table) = coverage::table(opts.target.arch) else {
2833        let _ = writeln!(
2834            stderr,
2835            "rucc: error: there are no lowering rules for {} yet, so there is no coverage of them \
2836             to report",
2837            opts.target
2838        );
2839        return false;
2840    };
2841    match std::fs::write(path, fired.listing(table)) {
2842        Ok(()) => true,
2843        Err(e) => {
2844            let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2845            false
2846        }
2847    }
2848}
2849
2850/// Writes what `-Zregister-pressure=FILE` asked for, and says whether it could.
2851///
2852/// Once for the whole command line, for the reason [`write_coverage`] gives, and a file that could
2853/// not be written is a failure for the reason it gives too. There is no equivalent of the missing
2854/// rule table here, since every target this compiles for has an allocator, and a run that reached
2855/// no back end at all writes an empty listing rather than nothing: a measurement of a build that
2856/// produced no code is still an answer and it is the honest one.
2857fn write_pressure(opts: &Options, pressure: &Pressure, stderr: &mut impl std::io::Write) -> bool {
2858    let Some(path) = &opts.register_pressure else { return true };
2859    match std::fs::write(path, pressure.listing()) {
2860        Ok(()) => true,
2861        Err(e) => {
2862            let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2863            false
2864        }
2865    }
2866}
2867
2868/// Writes what `-Zlowering=FILE` asked for, and says whether it could.
2869///
2870/// Once for the whole command line, for the reason [`write_coverage`] gives, and a file that could
2871/// not be written is a failure for the reason it gives too. A run that reached no back end writes
2872/// an empty listing rather than nothing, the way [`write_pressure`] does and for the same reason.
2873fn write_lowering(opts: &Options, lowerings: &Lowerings, stderr: &mut impl std::io::Write) -> bool {
2874    let Some(path) = &opts.lowering_dump else { return true };
2875    match std::fs::write(path, lowerings.listing()) {
2876        Ok(()) => true,
2877        Err(e) => {
2878            let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2879            false
2880        }
2881    }
2882}
2883
2884/// Where the `-fopt-info` remarks go, and how much of the run has already gone there.
2885///
2886/// Standard error by default, and one file for the whole run when `-fopt-info=<file>` named one.
2887/// A file rather than the diagnostic stream is what a harness wants: the corpus in
2888/// `tamnd/rucc-corpus` matches a rejection against what the compiler said on standard error, and
2889/// a few thousand remarks mixed into that would bury it.
2890struct Remarks {
2891    /// The file, if there is one.
2892    file: Option<String>,
2893    /// Whether anything has been written to it yet, which decides between truncating and
2894    /// appending. One file holds the whole run rather than the last input in it.
2895    started: bool,
2896}
2897
2898impl Remarks {
2899    /// Prepares the destination, emptying the file if there is one.
2900    ///
2901    /// Emptied here rather than at the first remark, because a run where no pass had anything to
2902    /// say should leave an empty file and not yesterday's. An absent file and an empty one are
2903    /// different facts and something reading this will act on the difference.
2904    fn new(file: Option<&String>, stderr: &mut impl std::io::Write) -> (Self, bool) {
2905        let mut ok = true;
2906        if let Some(path) = file {
2907            if let Err(e) = std::fs::write(path, "") {
2908                let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2909                ok = false;
2910            }
2911        }
2912        (Self { file: file.cloned(), started: false }, ok)
2913    }
2914
2915    /// Writes one input's remarks, and says whether that worked.
2916    ///
2917    /// A file that cannot be written is a failure and not a warning, for the reason
2918    /// [`write_dumps`] gives: remarks that quietly did not arrive look exactly like a compilation
2919    /// where nothing happened.
2920    fn write(&mut self, text: &str, stderr: &mut impl std::io::Write) -> bool {
2921        if text.is_empty() {
2922            return true;
2923        }
2924        let Some(path) = &self.file else {
2925            let _ = write!(stderr, "{text}");
2926            return true;
2927        };
2928        let opened = std::fs::OpenOptions::new()
2929            .write(true)
2930            .append(self.started)
2931            .truncate(!self.started)
2932            .create(true)
2933            .open(path);
2934        self.started = true;
2935        let result =
2936            opened.and_then(|mut file| std::io::Write::write_all(&mut file, text.as_bytes()));
2937        if let Err(e) = result {
2938            let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2939            return false;
2940        }
2941        true
2942    }
2943}
2944
2945/// Writes what `-fdump-ir=` asked to see, one file per dump.
2946///
2947/// The name is the input file with the dump's own name and `.ir` after it, so a directory listing
2948/// after a run is the passes in the order they ran, per input. They go in the working directory
2949/// rather than beside the output, because a dump is something a person asked for at a prompt and
2950/// the working directory is where that person is.
2951///
2952/// A file that could not be written is a failure and not a warning, for the reason
2953/// [`write_coverage`] gives: what asked for this is somebody debugging a pass, and a dump that
2954/// quietly did not happen looks exactly like a pass that did not run.
2955fn write_dumps(input: &str, dumps: &[rucc_opt::Dump], stderr: &mut impl std::io::Write) -> bool {
2956    let stem = std::path::Path::new(input)
2957        .file_name()
2958        .map_or_else(|| input.to_owned(), |name| name.to_string_lossy().into_owned());
2959    let mut ok = true;
2960    for dump in dumps {
2961        let path = format!("{stem}.{}.ir", dump.name);
2962        if let Err(e) = std::fs::write(&path, &dump.text) {
2963            let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2964            ok = false;
2965        }
2966    }
2967    ok
2968}
2969
2970/// Writes the files `-save-temps` kept, which is nothing at all unless it was given.
2971///
2972/// A file that could not be written is a failure rather than a warning, for the reason
2973/// [`write_dumps`] gives: somebody asked for these by name, and one that quietly did not happen
2974/// looks like a compilation that never went through that step.
2975fn write_temps(job: &Job, temps: &Temps, stderr: &mut impl std::io::Write) -> bool {
2976    let mut ok = true;
2977    let kept = [(job.saved_text(), &temps.preprocessed), (job.saved_asm(), &temps.assembly)];
2978    for (path, text) in kept {
2979        // A step the compilation did not reach has nothing to keep, and a job that is not keeping
2980        // that step has nowhere to put it. Either way there is no file here.
2981        let (Some(path), Some(text)) = (path, text) else { continue };
2982        if let Err(e) = std::fs::write(&path, text) {
2983            let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2984            ok = false;
2985        }
2986    }
2987    ok
2988}
2989
2990/// One line of `-time`, which is what a step was called and how long it took.
2991///
2992/// GCC's two numbers are the user and the system time of a subprocess it ran. This compiler runs
2993/// no subprocess for anything but the link, so what is measured here is the wall clock of the
2994/// step and the second column is always zero. The shape of the line is kept because a person
2995/// reading it next to gcc's should not have to work out which column is which.
2996fn say_time(name: &str, took: std::time::Duration, stderr: &mut impl std::io::Write) {
2997    let _ = writeln!(stderr, "# {name} {:.2} {:.2}", took.as_secs_f64(), 0.0);
2998}
2999
3000/// Writes one job's result where the plan said it goes.
3001///
3002/// # Errors
3003///
3004/// Returns the message to print, which names the file when there is one, because "permission
3005/// denied" on its own does not say which file was refused.
3006fn write_out(output: &Output, bytes: &[u8]) -> Result<(), String> {
3007    match output {
3008        Output::Stdout => {
3009            let mut stdout = std::io::stdout().lock();
3010            stdout.write_all(bytes).map_err(|e| format!("writing to standard output: {e}"))
3011        }
3012        Output::File(path) | Output::Temporary(path) => {
3013            std::fs::write(path, bytes).map_err(|e| format!("{path}: {e}"))
3014        }
3015    }
3016}
3017
3018/// Runs the driver and returns the process exit code.
3019///
3020/// `args` excludes the program name. Output goes to `stdout` and errors to `stderr`, which
3021/// is the one place in the compiler that is true.
3022pub fn run(args: &[String]) -> i32 {
3023    match parse_args(args) {
3024        Ok(Action::Help) => {
3025            print!("{USAGE}");
3026            0
3027        }
3028        Ok(Action::Version) => {
3029            println!("rucc {VERSION}");
3030            0
3031        }
3032        Ok(Action::Print(line)) => {
3033            println!("{line}");
3034            0
3035        }
3036        Ok(Action::PrintConfig(opts)) => {
3037            print!("{}", print_config(&opts));
3038            0
3039        }
3040        Ok(Action::PrintPipeline(opts)) => {
3041            print!("{}", print_pipeline(&opts));
3042            0
3043        }
3044        Ok(Action::PrintPlan { opts, plan, link }) => {
3045            print!("{}", plan.render());
3046            // The line as it would be typed, which is the half of `-###` that section 4.3 says
3047            // arrives with the link. It is printed even when the linker is not on this machine,
3048            // because what a build wants from `-###` is what the compiler would do.
3049            if let Some(job) = &plan.link {
3050                match link_line(&opts, &link, job) {
3051                    Ok(line) => println!("{line}"),
3052                    Err(why) => {
3053                        let mut stderr = std::io::stderr().lock();
3054                        let _ = writeln!(stderr, "rucc: error: {why}");
3055                        return 1;
3056                    }
3057                }
3058            }
3059            0
3060        }
3061        Ok(Action::Fetch { what, target, cache }) => fetch_sysroot(what, target, &cache),
3062        Ok(Action::Compile { opts, plan, link, jobs, verbose, notes }) => {
3063            {
3064                let mut stderr = std::io::stderr().lock();
3065                // Before the plan rather than after it, because a note is about the command line
3066                // and the plan is what the command line was read as, so the reader wants the two
3067                // in that order.
3068                for note in &notes {
3069                    let _ = writeln!(stderr, "rucc: warning: {note}");
3070                }
3071                if verbose {
3072                    let _ = write!(stderr, "{}", plan.render());
3073                    let _ = writeln!(stderr, "workers: {}", jobs.count());
3074                }
3075            }
3076            if opts.emit == EmitKind::Preprocessed {
3077                return preprocess_all(&opts, &plan);
3078            }
3079            if opts.emit == EmitKind::Archive {
3080                return archive_all(&opts, &plan);
3081            }
3082            if opts.emit != EmitKind::Executable {
3083                return compile_all(&opts, &plan);
3084            }
3085            link_all(&opts, &plan, &link, verbose)
3086        }
3087        Err(e) => {
3088            let mut stderr = std::io::stderr().lock();
3089            let _ = writeln!(stderr, "rucc: error: {e}");
3090            let _ = writeln!(stderr, "rucc: note: run `rucc --help` for usage");
3091            1
3092        }
3093    }
3094}
3095
3096#[cfg(test)]
3097mod tests {
3098    use rucc_session::{
3099        Contract, GnucVersion, IncludeForm, LtoJobs, OptLevel, Partition, Patchable, Visibility,
3100    };
3101
3102    use super::*;
3103
3104    fn args(s: &[&str]) -> Vec<String> {
3105        s.iter().map(|x| (*x).to_owned()).collect()
3106    }
3107
3108    /// A target to write down where the host would otherwise decide, for the tests whose answer
3109    /// would be a different one on a different machine.
3110    ///
3111    /// Most of the tests here never name a target, which is right, because most of what the driver
3112    /// does with a command line is the same wherever it runs and a test that pinned one would be
3113    /// saying so in every case for the sake of the two that need it. The two that need it are the
3114    /// ones whose answer comes off the target rather than off the command line: the name an object
3115    /// gets, which is `a.o` here and `a.obj` on Windows, and whether Microsoft's reading of a
3116    /// nameless member is on, which is off here and on there. Both are the compiler being right, and
3117    /// a test that leaves the target to the host is asking a question with two correct answers.
3118    const LINUX: &str = "--target=x86_64-unknown-linux-gnu";
3119
3120    #[test]
3121    fn help_and_version_win_over_everything_else() {
3122        assert_eq!(parse_args(&args(&["-c", "--help", "x.c"])).unwrap(), Action::Help);
3123        assert_eq!(parse_args(&args(&["--version"])).unwrap(), Action::Version);
3124    }
3125
3126    fn compile(s: &[&str]) -> (Box<Options>, Box<Plan>) {
3127        match parse_args(&args(s)).expect("expected a compilation") {
3128            Action::Compile { opts, plan, .. } => (opts, plan),
3129            other => panic!("expected a compilation, got {other:?}"),
3130        }
3131    }
3132
3133    fn linking(s: &[&str]) -> (Box<LinkOptions>, Box<Plan>) {
3134        match parse_args(&args(s)).expect("expected a compilation") {
3135            Action::Compile { link, plan, .. } => (link, plan),
3136            other => panic!("expected a compilation, got {other:?}"),
3137        }
3138    }
3139
3140    fn notes(s: &[&str]) -> Vec<String> {
3141        match parse_args(&args(s)).expect("expected a compilation") {
3142            Action::Compile { notes, .. } => notes,
3143            other => panic!("expected a compilation, got {other:?}"),
3144        }
3145    }
3146
3147    /// The ordinary command line has nothing to say about itself, which is the property that makes
3148    /// a note worth reading when there is one.
3149    #[test]
3150    fn a_command_line_with_nothing_wrong_with_it_carries_no_notes() {
3151        assert_eq!(notes(&["-c", "a.c"]), Vec::<String>::new());
3152    }
3153
3154    /// A directory that is not there contributes nothing to the search path, so there is no tree to
3155    /// read a release out of and nothing to compare the pin against. Said as a test because this is
3156    /// the shape a hermetic machine takes: the probe reads the disk and every other machine has a
3157    /// different disk, so what can be asserted here is the silence.
3158    #[test]
3159    fn a_named_tree_that_is_not_on_the_machine_is_not_a_release_mismatch() {
3160        let said =
3161            notes(&["--target=x86_64-linux-gnu.2.28", "--sysroot=/nowhere-at-all", "-c", "a.c"]);
3162        assert_eq!(said, Vec::<String>::new());
3163    }
3164
3165    #[test]
3166    fn collects_inputs_and_flags() {
3167        let (opts, plan) = compile(&["-c", "-O2", "-g", "a.c", "b.c"]);
3168        let paths: Vec<&str> = plan.jobs.iter().map(|j| j.input.as_str()).collect();
3169        assert_eq!(paths, vec!["a.c", "b.c"]);
3170        assert_eq!(opts.opt_level, OptLevel::O2);
3171        assert_eq!(opts.emit, EmitKind::Object);
3172        assert!(opts.debug_info);
3173    }
3174
3175    /// The unstable options, which are spelled apart from everything else on purpose: what is
3176    /// under `-Z` promises nothing, and a build that reaches for one should have had to say so.
3177    #[test]
3178    fn an_unstable_option_is_taken_and_one_that_does_not_exist_is_refused() {
3179        let (opts, _) = compile(&["-c", "-Zrule-coverage=/tmp/rules.cov", "a.c"]);
3180        assert_eq!(opts.rule_coverage.as_deref(), Some("/tmp/rules.cov"));
3181
3182        let (plain, _) = compile(&["-c", "a.c"]);
3183        assert_eq!(plain.rule_coverage, None, "nothing is measured unless it was asked for");
3184
3185        assert!(parse_args(&args(&["-Zrule-coverage=", "a.c"])).is_err(), "a file with no name");
3186        let unknown = parse_args(&args(&["-Zwhat", "a.c"])).expect_err("there is no such option");
3187        assert!(unknown.message.contains("4.11"), "{}", unknown.message);
3188    }
3189
3190    /// The other measurement written to a file, which reads the same way and fails the same way.
3191    #[test]
3192    fn where_the_register_pressure_goes_is_asked_for_the_same_way() {
3193        let (opts, _) = compile(&["-c", "-O2", "-Zregister-pressure=/tmp/spills.txt", "a.c"]);
3194        assert_eq!(opts.register_pressure.as_deref(), Some("/tmp/spills.txt"));
3195
3196        let (plain, _) = compile(&["-c", "a.c"]);
3197        assert_eq!(plain.register_pressure, None, "nothing is measured unless it was asked for");
3198
3199        assert!(parse_args(&args(&["-Zregister-pressure=", "a.c"])).is_err(), "no file named");
3200    }
3201
3202    /// The third one, which says what the pre-selection lowering group did.
3203    #[test]
3204    fn where_the_lowering_dump_goes_is_asked_for_the_same_way() {
3205        let (opts, _) = compile(&["-c", "-O2", "-Zlowering=/tmp/lowering.txt", "a.c"]);
3206        assert_eq!(opts.lowering_dump.as_deref(), Some("/tmp/lowering.txt"));
3207
3208        let (plain, _) = compile(&["-c", "a.c"]);
3209        assert_eq!(plain.lowering_dump, None, "nothing is dumped unless it was asked for");
3210
3211        assert!(parse_args(&args(&["-Zlowering=", "a.c"])).is_err(), "no file named");
3212    }
3213
3214    /// Scheduling, which has the three way answer every optimization flag has: on, off, and
3215    /// nothing said, which is whatever the optimization level asks for. The name is gcc's, and
3216    /// gcc's has a two in it because gcc has a scheduler before allocation and one after and this
3217    /// is the one after.
3218    #[test]
3219    fn scheduling_can_be_turned_on_and_off_and_left_to_the_optimization_level() {
3220        let (on, _) = compile(&["-c", "-O0", "-fschedule-insns2", "a.c"]);
3221        assert_eq!(on.schedule_insns, Some(true));
3222
3223        let (off, _) = compile(&["-c", "-O2", "-fno-schedule-insns2", "a.c"]);
3224        assert_eq!(off.schedule_insns, Some(false));
3225
3226        let (quiet, _) = compile(&["-c", "-O2", "a.c"]);
3227        assert_eq!(quiet.schedule_insns, None, "nothing said, so the level decides");
3228        assert!(quiet.opt_level.schedules(), "and at this level the level says yes");
3229
3230        let (none, _) = compile(&["-c", "a.c"]);
3231        assert!(!none.opt_level.schedules(), "at no optimization it says no");
3232    }
3233
3234    /// Whether the timing model is worth holding an instruction back over, which is a `-Z` because
3235    /// it is a question about a target's description rather than about the program being compiled.
3236    #[test]
3237    fn whether_the_timing_model_is_cycle_accurate_can_be_overridden() {
3238        let (yes, _) = compile(&["-c", "-O2", "-Zcycle-accurate-model=yes", "a.c"]);
3239        assert_eq!(yes.cycle_accurate_model, Some(true));
3240
3241        let (no, _) = compile(&["-c", "-O2", "-Zcycle-accurate-model=no", "a.c"]);
3242        assert_eq!(no.cycle_accurate_model, Some(false));
3243
3244        let (plain, _) = compile(&["-c", "-O2", "a.c"]);
3245        assert_eq!(plain.cycle_accurate_model, None, "the target's own answer stands");
3246
3247        let bad = parse_args(&args(&["-Zcycle-accurate-model=maybe", "a.c"]))
3248            .expect_err("it takes yes or no");
3249        assert!(bad.message.contains("yes or no"), "{}", bad.message);
3250    }
3251
3252    #[test]
3253    fn a_bare_dash_o_means_o1_the_way_gcc_reads_it() {
3254        let (opts, _) = compile(&["-O", "a.c"]);
3255        assert_eq!(opts.opt_level, OptLevel::O1);
3256    }
3257
3258    #[test]
3259    fn dash_x_applies_to_later_inputs_only_and_none_stops_it() {
3260        let (_, plan) = compile(&["a.o", "-x", "c", "b.txt", "-x", "none", "c.o"]);
3261        assert_eq!(plan.jobs[0].kind, InputKind::LinkerInput);
3262        assert_eq!(plan.jobs[1].kind, InputKind::C);
3263        assert_eq!(plan.jobs[2].kind, InputKind::LinkerInput);
3264    }
3265
3266    #[test]
3267    fn dash_j_reaches_the_scheduler_and_defaults_to_the_machine() {
3268        let (_, _, jobs) = match parse_args(&args(&["-j4", "a.c"])).unwrap() {
3269            Action::Compile { opts, plan, jobs, .. } => (opts, plan, jobs),
3270            other => panic!("expected a compilation, got {other:?}"),
3271        };
3272        assert_eq!(jobs.count(), 4);
3273
3274        let default = match parse_args(&args(&["a.c"])).unwrap() {
3275            Action::Compile { jobs, .. } => jobs,
3276            other => panic!("expected a compilation, got {other:?}"),
3277        };
3278        assert_eq!(default, Jobs::available());
3279        assert!(parse_args(&args(&["-j0", "a.c"])).is_err());
3280    }
3281
3282    #[test]
3283    fn triple_hash_prints_the_plan_and_runs_nothing() {
3284        let a = parse_args(&args(&["-###", "-c", "a.c"])).unwrap();
3285        let Action::PrintPlan { plan, .. } = a else { panic!("expected a plan dump") };
3286        assert!(plan.render().contains("a.c: preprocess, compile, assemble -> a.o"));
3287    }
3288
3289    #[test]
3290    fn the_flag_that_keeps_the_intermediate_files_has_three_spellings_and_two_meanings() {
3291        // The bare one is `=obj` and not `=cwd`. gcc's manual says the opposite and gcc 16 does
3292        // this, and following the compiler is what makes a build that reads either of them find
3293        // the files where they are.
3294        assert_eq!(compile(&["-c", "-save-temps", "a.c"]).0.save_temps, SaveTemps::Object);
3295        assert_eq!(compile(&["-c", "-save-temps=obj", "a.c"]).0.save_temps, SaveTemps::Object);
3296        assert_eq!(compile(&["-c", "-save-temps=cwd", "a.c"]).0.save_temps, SaveTemps::Cwd);
3297        assert_eq!(compile(&["-c", "a.c"]).0.save_temps, SaveTemps::No);
3298        // The last one on the line decides, the way it does for every other flag with an
3299        // argument, and a keyword that is neither is fatal rather than ignored: a run that kept
3300        // nothing and said nothing looks exactly like one where the files were not produced.
3301        let (opts, _) = compile(&["-c", "-save-temps", "-save-temps=cwd", "a.c"]);
3302        assert_eq!(opts.save_temps, SaveTemps::Cwd);
3303        let e = parse_args(&args(&["-c", "-save-temps=nowhere", "a.c"])).unwrap_err();
3304        assert!(e.message.contains("accepted: cwd, obj"), "{}", e.message);
3305    }
3306
3307    #[test]
3308    fn the_flag_that_times_each_step_reaches_the_options_and_changes_nothing_else() {
3309        let (opts, plan) = compile(&["-c", "-time", "a.c"]);
3310        let (plain, without) = compile(&["-c", "a.c"]);
3311        assert!(opts.time);
3312        assert!(!plain.time);
3313        // Against the same line without the flag rather than against a spelling of the object's
3314        // name, since what the object is called is the host's business and this is not about that.
3315        assert_eq!(plan.jobs[0].output, without.jobs[0].output);
3316    }
3317
3318    #[test]
3319    fn dash_x_names_what_it_accepts_when_it_does_not_know_a_language() {
3320        let e = parse_args(&args(&["-x", "fortran", "a.c"])).unwrap_err();
3321        assert!(e.message.contains("assembler-with-cpp"), "{}", e.message);
3322    }
3323
3324    /// What `--fetch` says for a target this release pins nothing for, which is every target except
3325    /// the three windows-gnu ones today.
3326    #[test]
3327    fn a_fetch_of_a_target_nothing_is_pinned_for_says_so_rather_than_reaching_the_network() {
3328        let e = parse_args(&args(&["--fetch", "x86_64-linux-musl"])).unwrap_err();
3329        assert!(e.message.contains("pins no sysroot for x86_64-linux-musl"), "{}", e.message);
3330        // And what it does pin, because a release with some rows in the table and a release with
3331        // none are two situations and the second sentence is what tells them apart.
3332        assert!(e.message.contains("x86_64-windows-gnu"), "{}", e.message);
3333        // The joined spelling is the same flag.
3334        let joined = parse_args(&args(&["--fetch=x86_64-linux-musl"])).unwrap_err();
3335        assert_eq!(joined, e);
3336    }
3337
3338    /// The two targets a release will never pin, which is a different answer from the one above.
3339    ///
3340    /// Section 13.4. A person who reads "this release pins no sysroot yet" waits for a release that
3341    /// does, and no release of this compiler can ship either of these, so the message names the
3342    /// licence that decides it and what to do instead.
3343    #[test]
3344    fn a_fetch_of_a_target_behind_a_licence_wall_says_so_rather_than_saying_not_yet() {
3345        let e = parse_args(&args(&["--fetch", "aarch64-macos"])).unwrap_err();
3346        assert!(e.message.contains("Xcode licence"), "{}", e.message);
3347        assert!(e.message.contains("there never will be"), "{}", e.message);
3348        assert!(!e.message.contains("tamnd/rucc-cross"), "{}", e.message);
3349
3350        let e = parse_args(&args(&["--fetch", "x86_64-windows-msvc"])).unwrap_err();
3351        assert!(e.message.contains("redistributed"), "{}", e.message);
3352        // The way out of this one is a target rather than a download, and it is the default already.
3353        assert!(e.message.contains("mingw-w64"), "{}", e.message);
3354        // And the mingw-w64 target next to it is ours to ship and published, so the same flag has
3355        // something to get rather than a licence to explain.
3356        let action = parse_args(&args(&["--fetch", "x86_64-windows-gnu"])).expect("it is pinned");
3357        let Action::Fetch { what, .. } = action else { panic!("{action:?}") };
3358        assert_eq!(what.tuple, "x86_64-windows-gnu");
3359    }
3360
3361    /// An Apple target on a machine with no SDK, which is section 8.6's other host.
3362    ///
3363    /// Not run on a mac, where the SDK this is about is installed and the compile is the ordinary one
3364    /// that uses it. What the reason says is asserted in `rucc_sysroot::wall` and where it is printed
3365    /// is asserted in `rucc-pp`, so what is left here is that the driver works it out and leaves it
3366    /// where the preprocessor will find it, and that neither way past the wall leaves one behind.
3367    #[test]
3368    fn an_apple_target_with_no_sdk_anywhere_carries_the_licence_rather_than_a_missing_directory() {
3369        if cfg!(target_os = "macos") || std::env::var_os("SDKROOT").is_some() {
3370            return;
3371        }
3372        let (opts, _) = compile(&["--target=aarch64-macos", "-c", "a.c"]);
3373        let why = opts.search.missing_system().expect("the wall is the reason there are none");
3374        assert!(why.contains("aarch64-macos needs a macOS SDK"), "{why}");
3375        assert!(why.contains("Xcode licence"), "{why}");
3376        assert!(why.contains("-isysroot"), "{why}");
3377
3378        // A program that includes none of the library needs none of the SDK, which is what section
3379        // 8.6 means by being able to target the platform without one, so there is nothing to explain.
3380        let (opts, _) = compile(&["--target=aarch64-macos", "-nostdinc", "-c", "a.c"]);
3381        assert_eq!(opts.search.missing_system(), None);
3382        // And naming a path is the other way through, whether or not the path is there: a mistyped
3383        // directory is a mistake to report on its own terms rather than a licence to explain.
3384        let (opts, _) = compile(&["--target=aarch64-macos", "-isysroot", "/opt/sdk", "-c", "a.c"]);
3385        assert_eq!(opts.search.missing_system(), None);
3386    }
3387
3388    /// The same wall on the compile side of an MSVC target, where the way past it is a tuple.
3389    ///
3390    /// Not run on Windows, for the same reason the one above is not run on a mac: the wall stands in
3391    /// front of an SDK this machine does not have, and a Windows machine is the kind that does. The
3392    /// driver asks `vswhere` where Visual Studio is and takes the newest kit under it, so on a box
3393    /// with the build tools installed there are headers, no wall and nothing here to be about.
3394    /// `INCLUDE` is the other way a machine has one and is the other half of the guard, since a
3395    /// person can set that anywhere while Visual Studio is only found on the platform it runs on.
3396    #[test]
3397    fn an_msvc_target_with_no_sdk_named_says_which_environment_needs_nothing_installed() {
3398        if cfg!(target_os = "windows") || std::env::var_os("INCLUDE").is_some() {
3399            return;
3400        }
3401        let (opts, _) = compile(&["--target=x86_64-windows-msvc", "-c", "a.c"]);
3402        let why = opts.search.missing_system().expect("the wall is the reason there are none");
3403        assert!(why.contains("the Windows SDK and its universal CRT"), "{why}");
3404        assert!(why.contains("mingw-w64"), "{why}");
3405        // And the mingw-w64 target has its headers from us, so nothing is missing to explain.
3406        let (opts, _) = compile(&["--target=x86_64-windows-gnu", "-c", "a.c"]);
3407        assert_eq!(opts.search.missing_system(), None);
3408    }
3409
3410    #[test]
3411    fn a_fetch_with_no_target_and_a_fetch_of_a_tuple_that_is_not_one_both_say_which() {
3412        let e = parse_args(&args(&["--fetch"])).unwrap_err();
3413        assert!(e.message.contains("--fetch requires"), "{}", e.message);
3414        let e = parse_args(&args(&["--fetch", "sparc64-solaris-gnu"])).unwrap_err();
3415        assert!(e.message.contains("--fetch sparc64-solaris-gnu"), "{}", e.message);
3416        assert!(e.message.contains("no sysroot to get"), "{}", e.message);
3417    }
3418
3419    /// Both flags on one line ask for opposite things, in either order.
3420    #[test]
3421    fn a_fetch_and_offline_together_is_a_refusal_whichever_way_round_they_are_written() {
3422        for line in [
3423            vec!["--offline", "--fetch", "x86_64-linux-musl"],
3424            vec!["--fetch", "x86_64-linux-musl", "--offline"],
3425        ] {
3426            let e = parse_args(&args(&line)).unwrap_err();
3427            assert!(e.message.contains("two opposite things"), "{}", e.message);
3428        }
3429    }
3430
3431    #[test]
3432    fn a_fetch_does_not_compile_anything_and_says_so_when_it_is_handed_a_file() {
3433        let e = parse_args(&args(&["--fetch", "x86_64-linux-musl", "a.c"])).unwrap_err();
3434        assert!(e.message.contains("compiles nothing"), "{}", e.message);
3435        assert!(e.message.contains("a.c"), "{}", e.message);
3436    }
3437
3438    /// `--offline` on its own is accepted and changes nothing, because an ordinary compile
3439    /// downloads nothing with or without it. A build that passes it everywhere is the case this is
3440    /// for, and it must not lose the compilation it was passed beside.
3441    #[test]
3442    fn offline_on_a_compilation_is_the_same_compilation() {
3443        let (opts, plan) = compile(&["-c", "--offline", "a.c"]);
3444        let (plain, without) = compile(&["-c", "a.c"]);
3445        assert_eq!(opts.target, plain.target);
3446        assert_eq!(plan.jobs.len(), without.jobs.len());
3447        assert_eq!(plan.jobs[0].output, without.jobs[0].output);
3448    }
3449
3450    #[test]
3451    fn an_unknown_flag_is_an_error_rather_than_a_shrug() {
3452        let e = parse_args(&args(&["-fno-such-thing", "a.c"])).unwrap_err();
3453        assert!(e.message.contains("unknown option"), "{}", e.message);
3454    }
3455
3456    /// `-fpermissive` and the flag that turns it back off, which a build writes beside it when
3457    /// one directory needs the older rules and the rest of the tree does not.
3458    #[test]
3459    fn permissive_reads_in_both_directions_and_the_last_one_wins() {
3460        let (opts, _) = compile(&["-c", "a.c"]);
3461        assert!(!opts.permissive, "off unless it is asked for");
3462
3463        let (opts, _) = compile(&["-c", "-fpermissive", "a.c"]);
3464        assert!(opts.permissive);
3465
3466        let (opts, _) = compile(&["-c", "-fpermissive", "-fno-permissive", "a.c"]);
3467        assert!(!opts.permissive);
3468    }
3469
3470    #[test]
3471    fn asking_for_nested_functions_is_told_why_it_is_not_coming() {
3472        let e = parse_args(&args(&["-fnested-functions", "a.c"])).unwrap_err();
3473        assert!(e.message.contains("trampoline"), "{}", e.message);
3474        assert!(parse_args(&args(&["-fno-nested-functions", "a.c"])).is_ok());
3475    }
3476
3477    #[test]
3478    fn the_flag_every_configure_script_writes_is_taken() {
3479        // All four spellings, because a build writes whichever one its macros picked and a
3480        // compiler that takes three of them is a compiler that fails on the fourth.
3481        for flag in ["-fPIC", "-fpic", "-fPIE", "-fpie"] {
3482            let (opts, _) = compile(&["-c", flag, "a.c"]);
3483            assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3484        }
3485    }
3486
3487    #[test]
3488    fn a_table_is_written_unless_the_build_says_nothing_will_walk_it() {
3489        let (opts, _) = compile(&["-c", "a.c"]);
3490        assert!(opts.unwinds(), "the default is off");
3491        let (opts, _) = compile(&["-c", "-fno-asynchronous-unwind-tables", "a.c"]);
3492        assert!(!opts.unwinds(), "the build was not taken at its word");
3493        let (opts, _) = compile(&[
3494            "-c",
3495            "-fno-asynchronous-unwind-tables",
3496            "-fasynchronous-unwind-tables",
3497            "a.c",
3498        ]);
3499        assert!(opts.unwinds(), "the last flag did not win");
3500        // The weaker request, which the same table answers, so a line that asks for a table and
3501        // against an asynchronous one gets one. That is gcc's arrangement and it turns up when a
3502        // build turns the asynchronous one off globally and a directory asks for a table back.
3503        let (opts, _) =
3504            compile(&["-c", "-fno-asynchronous-unwind-tables", "-funwind-tables", "a.c"]);
3505        assert!(opts.unwinds(), "the weaker request was dropped");
3506        let (opts, _) = compile(&["-c", "-fno-unwind-tables", "a.c"]);
3507        assert!(opts.unwinds(), "the weaker negative turned off the stronger request");
3508        let (opts, _) =
3509            compile(&["-c", "-fno-unwind-tables", "-fno-asynchronous-unwind-tables", "a.c"]);
3510        assert!(!opts.unwinds(), "both were turned off and one stayed on");
3511    }
3512
3513    #[test]
3514    fn the_flags_that_describe_what_this_compiler_already_does_are_taken() {
3515        // Every one of these is on a real build line somewhere and every one of them was an
3516        // unknown option. What they have in common is that the answer rucc gives is the answer
3517        // they ask for, so there is nothing to implement and nothing to refuse.
3518        for flag in [
3519            "-fno-common",
3520            "-fstrict-aliasing",
3521            "-fno-strict-aliasing",
3522            "-fdelete-null-pointer-checks",
3523            "-fno-delete-null-pointer-checks",
3524            "-frounding-math",
3525            "-fno-rounding-math",
3526            "-fexcess-precision=standard",
3527            "-fexcess-precision=fast",
3528            "-fexcess-precision=16",
3529            "-pipe",
3530            "-fdiagnostics-color",
3531            "-fno-diagnostics-color",
3532            "-fdiagnostics-color=always",
3533            "-fdiagnostics-color=never",
3534            "-fdiagnostics-color=auto",
3535        ] {
3536            let (opts, _) = compile(&["-c", flag, "a.c"]);
3537            assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3538        }
3539    }
3540
3541    #[test]
3542    fn whether_an_exception_is_looked_at_is_kept_and_defaults_to_gccs_answer() {
3543        let (opts, _) = compile(&["-c", "a.c"]);
3544        assert!(opts.trapping_math, "the default was not gcc's");
3545        let (opts, _) = compile(&["-c", "-fno-trapping-math", "a.c"]);
3546        assert!(!opts.trapping_math);
3547        let (opts, _) = compile(&["-c", "-ftrapping-math", "a.c"]);
3548        assert!(opts.trapping_math, "spelling out the default turned it off");
3549        // The last one written wins, which is how a build line that inherits a flag from one
3550        // place and overrides it in another is read.
3551        let (opts, _) = compile(&["-c", "-fno-trapping-math", "-ftrapping-math", "a.c"]);
3552        assert!(opts.trapping_math);
3553    }
3554
3555    /// The flags a torture program writes on its own `dg-options` line, which is where most of
3556    /// these come from: a program reduced from a miscompilation names the pass that miscompiled
3557    /// it. Eighteen programs in the suite stopped on the driver before anything read them, and
3558    /// tamnd/rucc#1019 is the list.
3559    #[test]
3560    fn the_flags_that_name_a_pass_of_gccs_own_are_taken_and_dropped() {
3561        for flag in [
3562            "-fno-tree-ccp",
3563            "-fno-tree-dominator-opts",
3564            "-fno-tree-vrp",
3565            "-fno-tree-bit-ccp",
3566            "-fno-tree-coalesce-vars",
3567            "-ftree-vectorize",
3568            "-ftree-loop-distribution",
3569            "-fipa-pta",
3570            "-fmodulo-sched",
3571            "-fno-vect-cost-model",
3572            "-fvect-cost-model=unlimited",
3573            "-fsimd-cost-model=cheap",
3574            "-fexpensive-optimizations",
3575            "-fno-early-inlining",
3576            "-fno-inline",
3577            "-finline-functions",
3578            "-foptimize-strlen",
3579            "-fno-ira-share-spill-slots",
3580        ] {
3581            let (opts, _) = compile(&["-c", flag, "a.c"]);
3582            assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3583            assert!(opts.passes.is_empty(), "{flag} named a pass of gcc's and not one of ours");
3584        }
3585    }
3586
3587    /// The two namespaces are taken whole, so a name neither this test nor gcc 16 has heard of
3588    /// goes the same way as the ones above rather than stopping a build on the day gcc adds it.
3589    #[test]
3590    fn a_pass_name_in_either_family_is_taken_whether_or_not_it_is_one_gcc_has() {
3591        for flag in ["-ftree-no-such-pass", "-fno-ipa-no-such-pass"] {
3592            let (opts, _) = compile(&["-c", flag, "a.c"]);
3593            assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3594        }
3595    }
3596
3597    /// A pass this compiler has keeps its flag, since the arms that read the registry are above
3598    /// the family arms. `dce` is the one both compilers have a name for, and `execute/pr97421-2.c`
3599    /// is the program that writes it.
3600    #[test]
3601    fn a_pass_name_this_compiler_has_is_still_read_as_a_pass() {
3602        let (opts, _) = compile(&["-c", "-fno-dce", "a.c"]);
3603        assert_eq!(opts.passes, vec![("dce".to_owned(), false)]);
3604    }
3605
3606    /// gcc's name for the unroller reaches the unroller, in both directions. libtommath puts
3607    /// `-funroll-loops` in `CFLAGS` unconditionally, and before this it was an unknown option and
3608    /// the build stopped on its first file.
3609    #[test]
3610    fn the_gcc_spelling_of_the_unroller_turns_the_unroller_on_and_off() {
3611        let (opts, _) = compile(&["-c", "-funroll-loops", "a.c"]);
3612        assert_eq!(opts.passes, vec![("unroll".to_owned(), true)]);
3613        let (opts, _) = compile(&["-c", "-fno-unroll-loops", "a.c"]);
3614        assert_eq!(opts.passes, vec![("unroll".to_owned(), false)]);
3615    }
3616
3617    /// Where a function starts is a question this compiler answers, so the flag that asks about it
3618    /// is answered rather than dropped. femtolisp's Makefile writes the bare form on every compile
3619    /// of the project, and before this it was an unknown option and the build stopped on its first
3620    /// file. The numbers are gcc 16's, read off `-S` on x86-64: nothing and the bare form both
3621    /// give `.p2align 4`, `=32` gives 5, `=3` gives 2, and the negative form gives `.align 8`.
3622    #[test]
3623    fn the_alignment_of_a_function_is_a_request_this_compiler_can_answer() {
3624        let (opts, _) = compile(&["-c", "-falign-functions", "a.c"]);
3625        assert_eq!(opts.align_functions, None, "the bare form asks for the default");
3626
3627        let (opts, _) = compile(&["-c", "-falign-functions=32", "a.c"]);
3628        assert_eq!(opts.align_functions, Some(32));
3629
3630        let (opts, _) = compile(&["-c", "-falign-functions=3", "a.c"]);
3631        assert_eq!(opts.align_functions, Some(4), "rounded up rather than refused");
3632
3633        let (opts, _) = compile(&["-c", "-falign-functions=32:8", "a.c"]);
3634        assert_eq!(opts.align_functions, Some(32), "the boundary is the answerable half");
3635
3636        for flag in ["-falign-functions=0", "-falign-functions=1"] {
3637            let (opts, _) = compile(&["-c", flag, "a.c"]);
3638            assert_eq!(opts.align_functions, None, "{flag} means the default");
3639        }
3640
3641        let (opts, _) = compile(&["-c", "-fno-align-functions", "a.c"]);
3642        assert_eq!(opts.align_functions, Some(8), "the smallest boundary the target has");
3643
3644        // The last one on the line wins, which is how gcc reads a repeated flag.
3645        let (opts, _) = compile(&["-c", "-falign-functions=32", "-falign-functions", "a.c"]);
3646        assert_eq!(opts.align_functions, None);
3647
3648        let e = parse_args(&args(&["-c", "-falign-functions=big", "a.c"])).unwrap_err();
3649        assert!(e.message.contains("number of bytes"), "{}", e.message);
3650    }
3651
3652    /// The other three of the family are about padding inside a body, which nothing here writes,
3653    /// so they are taken and say nothing. Every spelling of each, since a build writes whichever
3654    /// one its author typed.
3655    #[test]
3656    fn the_alignment_flags_about_the_inside_of_a_body_are_taken_and_say_nothing() {
3657        for flag in [
3658            "-falign-labels",
3659            "-falign-loops",
3660            "-falign-jumps",
3661            "-falign-loops=16",
3662            "-falign-labels=32",
3663            "-fno-align-loops",
3664            "-fno-align-labels",
3665            "-fno-align-jumps",
3666        ] {
3667            let (opts, _) = compile(&["-c", flag, "a.c"]);
3668            assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3669            assert_eq!(opts.align_functions, None, "{flag} is not about where a function starts");
3670        }
3671    }
3672
3673    /// The encoding of the source is not a question about speed, so the one name that describes
3674    /// what the preprocessor does is taken and every other name is refused.
3675    #[test]
3676    fn the_input_charset_is_taken_when_it_names_the_one_that_is_read() {
3677        for flag in ["-finput-charset=utf-8", "-finput-charset=UTF-8", "-finput-charset=utf8"] {
3678            let (opts, _) = compile(&["-c", flag, "a.c"]);
3679            assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3680        }
3681
3682        let e = parse_args(&args(&["-c", "-finput-charset=latin1", "a.c"])).unwrap_err();
3683        assert!(e.message.contains("latin1"), "{}", e.message);
3684        assert!(e.message.contains("UTF-8"), "what is read is worth saying: {}", e.message);
3685    }
3686
3687    /// The other half of the same rule. Each of these changes what the program does rather than
3688    /// how fast it does it, so each is refused with the reason, and the negative of each is what
3689    /// happens anyway and is taken.
3690    #[test]
3691    fn the_three_that_change_the_answer_are_refused_and_their_negatives_are_taken() {
3692        for (flag, word) in [
3693            ("-ffast-math", "__FAST_MATH__"),
3694            ("-fnon-call-exceptions", "landing pad"),
3695            ("-finstrument-functions", "__cyg_profile_func_enter"),
3696        ] {
3697            let e = parse_args(&args(&["-c", flag, "a.c"])).unwrap_err();
3698            assert!(e.message.contains(word), "{flag}: {}", e.message);
3699            assert!(!e.message.contains("unknown option"), "{flag} deserves a reason");
3700
3701            let off = format!("-fno-{}", flag.trim_start_matches("-f"));
3702            let (opts, _) = compile(&["-c", &off, "a.c"]);
3703            assert_eq!(opts.emit, EmitKind::Object, "{off}");
3704        }
3705    }
3706
3707    #[test]
3708    fn asking_the_linker_to_merge_tentative_definitions_is_told_why_it_is_not_coming() {
3709        // The one of that family that is a request rather than a description, and it is a real
3710        // difference: two files each writing `int g;` link under it and do not without it.
3711        let e = parse_args(&args(&["-fcommon", "a.c"])).unwrap_err();
3712        assert!(e.message.contains(".bss"), "{}", e.message);
3713        assert!(e.message.contains("extern"), "the way out is worth saying: {}", e.message);
3714    }
3715
3716    #[test]
3717    fn asking_for_position_dependent_code_is_told_why_it_is_not_coming() {
3718        for flag in ["-fno-pic", "-fno-pie"] {
3719            let e = parse_args(&args(&[flag, "a.c"])).unwrap_err();
3720            assert!(e.message.contains("global offset table"), "{flag}: {}", e.message);
3721            // The one it may have meant, since the two are a letter apart and one of them is
3722            // about linking and is taken.
3723            assert!(e.message.contains("-no-pie"), "{flag}: {}", e.message);
3724        }
3725    }
3726
3727    #[test]
3728    fn an_unsupported_target_names_itself() {
3729        let e = parse_args(&args(&["--target=sparc64-linux-gnu", "a.c"])).unwrap_err();
3730        assert!(e.message.contains("sparc64"), "{}", e.message);
3731    }
3732
3733    #[test]
3734    fn no_inputs_is_an_error_but_print_config_needs_none() {
3735        assert!(parse_args(&args(&[])).is_err());
3736        assert!(matches!(parse_args(&args(&["--print-config"])), Ok(Action::PrintConfig(_))));
3737    }
3738
3739    #[test]
3740    fn print_config_reports_the_target_it_was_given_not_the_host() {
3741        let a = parse_args(&args(&["--print-config", "--target=riscv64-linux-musl"])).unwrap();
3742        let Action::PrintConfig(opts) = a else { panic!("expected a configuration dump") };
3743        let text = print_config(&opts);
3744        assert!(text.contains("target: riscv64-unknown-linux-musl"), "{text}");
3745        assert!(text.contains("char-signed: false"), "{text}");
3746        assert!(text.contains("object-format: elf"), "{text}");
3747        assert!(text.contains("va-list: void-pointer"), "{text}");
3748        // RISC-V has a register file and this compiler has not written it down yet, and the
3749        // dump says which of those two it is rather than leaving the line out.
3750        assert!(text.contains("registers: none"), "{text}");
3751        assert!(text.contains("timing-model: none"), "{text}");
3752    }
3753
3754    /// The model the schedule was chosen with, which is a receipt anybody comparing two runs of a
3755    /// benchmark needs: two numbers that disagree are usually two models and not two compilers.
3756    #[test]
3757    fn print_config_names_the_model_the_schedule_was_chosen_with() {
3758        let opts = Options::new("x86_64-unknown-linux-gnu".parse().unwrap());
3759        let text = print_config(&opts);
3760        let line = text.lines().find(|l| l.starts_with("timing-model:")).expect("the model");
3761        assert!(line.contains("Skylake"), "{line}");
3762        assert!(line.contains("published"), "a sentence saying where it came from: {line}");
3763    }
3764
3765    #[test]
3766    fn print_config_has_one_key_per_line_and_a_fixed_order() {
3767        let opts = Options::new("x86_64-unknown-linux-gnu".parse().unwrap());
3768        let text = print_config(&opts);
3769        let keys: Vec<&str> =
3770            text.lines().map(|l| l.split(':').next().unwrap_or_default()).collect();
3771        assert_eq!(keys[0], "version");
3772        assert_eq!(keys[1], "target");
3773        assert_eq!(keys.len(), 26);
3774        assert!(text.ends_with('\n'));
3775    }
3776
3777    #[test]
3778    fn the_safety_tier_is_read_off_the_command_line_and_a_wrong_one_is_refused() {
3779        let (opts, _) = compile(&["a.c"]);
3780        assert_eq!(opts.safety, rucc_session::Safety::Off);
3781
3782        for (flag, tier) in [
3783            ("-fsafety=detect", rucc_session::Safety::Detect),
3784            ("-fsafety=enforce", rucc_session::Safety::Enforce),
3785            ("-fsafety=kernel", rucc_session::Safety::Kernel),
3786            ("-fsafety=off", rucc_session::Safety::Off),
3787        ] {
3788            let (opts, _) = compile(&[flag, "a.c"]);
3789            assert_eq!(opts.safety, tier, "{flag}");
3790        }
3791
3792        // The last one wins, the way every other repeated flag on this command line does.
3793        let (opts, _) = compile(&["-fsafety=enforce", "-fsafety=off", "a.c"]);
3794        assert_eq!(opts.safety, rucc_session::Safety::Off);
3795
3796        // A misspelled tier is refused rather than ignored. Silently compiling without the
3797        // monitor a build asked for is the one failure mode this feature cannot have.
3798        let e = parse_args(&args(&["-fsafety=on", "a.c"])).unwrap_err();
3799        assert!(e.message.contains("is not a safety tier"), "{}", e.message);
3800        assert!(parse_args(&args(&["-fsafety", "a.c"])).is_err());
3801    }
3802
3803    #[test]
3804    fn the_padding_mode_is_read_off_the_command_line_and_a_wrong_one_is_refused() {
3805        // The default is the one section 9.3 of document 09 gives library code, which is that
3806        // padding does not participate, so a record filled a member at a time is not reported.
3807        let (opts, _) = compile(&["a.c"]);
3808        assert_eq!(opts.padding, rucc_session::Padding::Ignored);
3809
3810        let (opts, _) = compile(&["-fsafety=detect", "-fsafety-init=padding", "a.c"]);
3811        assert_eq!(opts.padding, rucc_session::Padding::Tracked);
3812
3813        let (opts, _) = compile(&["-fsafety-init=padding", "-fsafety-init=nopadding", "a.c"]);
3814        assert_eq!(opts.padding, rucc_session::Padding::Ignored);
3815
3816        // The tier is still a tier. A flag whose name starts the same way must not be eaten by
3817        // the one above it, which is the thing worth pinning about a pair of names like these.
3818        let (opts, _) = compile(&["-fsafety-init=padding", "a.c"]);
3819        assert_eq!(opts.safety, rucc_session::Safety::Off);
3820
3821        let e = parse_args(&args(&["-fsafety-init=some", "a.c"])).unwrap_err();
3822        assert!(e.message.contains("is not a padding mode"), "{}", e.message);
3823    }
3824
3825    #[test]
3826    fn whether_a_write_has_to_stay_inside_its_member_is_read_off_the_command_line() {
3827        // Off by default, because a store to allocated storage sets its effective type and C 6.5
3828        // lets a program reuse a buffer as something else. Row S4 is a build opting out of that.
3829        let (opts, _) = compile(&["a.c"]);
3830        assert_eq!(opts.subobject, rucc_session::Subobject::Off);
3831
3832        let (opts, _) = compile(&["-fsafety=detect", "-fsafety-subobject", "a.c"]);
3833        assert_eq!(opts.subobject, rucc_session::Subobject::Members);
3834
3835        let (opts, _) = compile(&["-fsafety-subobject", "-fno-safety-subobject", "a.c"]);
3836        assert_eq!(opts.subobject, rucc_session::Subobject::Off);
3837
3838        // It takes no value. The form that would take one is the strict reading of section 9.4,
3839        // which is not written yet, so say so rather than accept a spelling that does nothing.
3840        let e = parse_args(&args(&["-fsafety-subobject=strict", "a.c"])).unwrap_err();
3841        assert!(e.message.contains("tamnd/rucc#967"), "{}", e.message);
3842    }
3843
3844    #[test]
3845    fn whether_two_restrict_pointers_may_meet_is_read_off_the_command_line() {
3846        // Off by default, because the record a block keeps is the union of what each pointer
3847        // reached, so two pointers striding through one array without landing on the same byte are
3848        // reported and by the letter of the standard those are different objects. Row Y8 is a build
3849        // deciding it would rather know.
3850        let (opts, _) = compile(&["a.c"]);
3851        assert_eq!(opts.promise, rucc_session::Promise::Off);
3852
3853        let (opts, _) = compile(&["-fsafety=detect", "-fsafety-restrict", "a.c"]);
3854        assert_eq!(opts.promise, rucc_session::Promise::Blocks);
3855
3856        let (opts, _) = compile(&["-fsafety-restrict", "-fno-safety-restrict", "a.c"]);
3857        assert_eq!(opts.promise, rucc_session::Promise::Off);
3858
3859        // The tier is still a tier, which is the thing worth pinning about a pair of names where
3860        // one is the front of the other.
3861        let (opts, _) = compile(&["-fsafety-restrict", "a.c"]);
3862        assert_eq!(opts.safety, rucc_session::Safety::Off);
3863
3864        let e = parse_args(&args(&["-fsafety-restrict=blocks", "a.c"])).unwrap_err();
3865        assert!(e.message.contains("takes no value"), "{}", e.message);
3866    }
3867
3868    #[test]
3869    fn safety_races_takes_a_mode_and_defaults_to_watching_nothing() {
3870        // Three modes rather than a bare flag, because section 9.5 gives two answers that record
3871        // the same thing and report different classes, so a flag with no value could not say which
3872        // was wanted. Off by default for the reason on `rucc_session::Races`, which is not a cost
3873        // argument: this is the one plane where an edge nobody interposed costs a false report.
3874        let (opts, _) = compile(&["a.c"]);
3875        assert_eq!(opts.races, rucc_session::Races::Off);
3876
3877        let (opts, _) = compile(&["-fsafety-races=metadata", "a.c"]);
3878        assert_eq!(opts.races, rucc_session::Races::Metadata);
3879
3880        let (opts, _) = compile(&["-fsafety-races=pointer", "a.c"]);
3881        assert_eq!(opts.races, rucc_session::Races::Pointer);
3882
3883        // Last one wins, as it does for every other mode flag here.
3884        let (opts, _) = compile(&["-fsafety-races=pointer", "-fno-safety-races", "a.c"]);
3885        assert_eq!(opts.races, rucc_session::Races::Off);
3886
3887        let e = parse_args(&args(&["-fsafety-races=all", "a.c"])).unwrap_err();
3888        assert!(e.message.contains("off, metadata or pointer"), "{}", e.message);
3889    }
3890
3891    #[test]
3892    fn print_pipeline_answers_with_the_passes_the_level_asked_for() {
3893        let a = parse_args(&args(&["--print-pipeline", "-O2"])).unwrap();
3894        let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3895        let text = print_pipeline(&opts);
3896        assert!(text.starts_with("level: -O2\n"), "{text}");
3897        assert!(text.contains("fold"), "{text}");
3898
3899        let a = parse_args(&args(&["--print-pipeline"])).unwrap();
3900        let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3901        // Two passes run at `-O0` and neither is an optimization. The first moves what
3902        // `__builtin_expect` said onto the branch and takes the instruction away, so that nothing
3903        // past the optimizer has to know the instruction exists. The second removes code nothing
3904        // reaches. See issue 359.
3905        assert!(print_pipeline(&opts).contains("1: expect,"), "{}", print_pipeline(&opts));
3906        assert!(print_pipeline(&opts).contains("2: simplify-cfg,"), "{}", print_pipeline(&opts));
3907
3908        let a = parse_args(&args(&["--print-pipeline", "-fno-simplify-cfg"])).unwrap();
3909        let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3910        // The second turns off and the first does not, because nothing below the optimizer lowers
3911        // what it removes, so `-fno-expect` is a compile that stops rather than one that runs.
3912        let text = print_pipeline(&opts);
3913        assert!(text.contains("1: expect,"), "{text}");
3914        assert!(!text.contains("simplify-cfg"), "{text}");
3915    }
3916
3917    #[test]
3918    fn print_pipeline_takes_the_toggles_into_account() {
3919        let a = parse_args(&args(&["--print-pipeline", "-O2", "-fno-fold"])).unwrap();
3920        let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3921        let text = print_pipeline(&opts);
3922        // The one that was named is gone and the rest of the level is not, which is the whole
3923        // of what a toggle promises.
3924        assert!(!text.contains("fold"), "{text}");
3925        assert!(text.contains("dce"), "{text}");
3926
3927        // Every pass the compiler has, named off. Built from the registry rather than written
3928        // out, so a pass added later is turned off here too and this keeps testing the thing it
3929        // is about, which is that the toggles can empty a level down to the passes that are not
3930        // optional. Those are named, because a listing that is all of them is a level nobody
3931        // emptied and the assertion would pass while saying nothing.
3932        let mut off = vec!["--print-pipeline".to_owned(), "-O2".to_owned()];
3933        off.extend(rucc_opt::PASSES.iter().map(|p| format!("-fno-{}", p.name())));
3934        let spelled: Vec<&str> = off.iter().map(String::as_str).collect();
3935        let a = parse_args(&args(&spelled)).unwrap();
3936        let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3937        let text = print_pipeline(&opts);
3938        let left: Vec<&str> =
3939            rucc_opt::PASSES.iter().filter(|p| p.required()).map(|p| p.name()).collect();
3940        assert_eq!(left, vec!["expect"], "{text}");
3941        for (at, name) in left.iter().enumerate() {
3942            assert!(text.contains(&format!("{}: {name},", at + 1)), "{text}");
3943        }
3944        assert!(!text.contains("dce"), "{text}");
3945    }
3946
3947    #[test]
3948    fn print_pipeline_says_when_a_budget_will_stop_the_run_short() {
3949        let a = parse_args(&args(&["--print-pipeline", "-O2"])).unwrap();
3950        let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3951        assert!(!print_pipeline(&opts).contains("global fuel"));
3952
3953        let a = parse_args(&args(&["--print-pipeline", "-O2", "-fpass-fuel-global=4"])).unwrap();
3954        let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3955        let text = print_pipeline(&opts);
3956        // Because the listing is the answer to what this compilation will do, and a run that
3957        // stops after four rewrites is not doing what the level says it does.
3958        assert!(text.contains("global fuel: 4"), "{text}");
3959    }
3960
3961    /// A pass is turned on and off by its own name, and the order the flags were given in is
3962    /// kept, because the last spelling of a name is the one that decides.
3963    #[test]
3964    fn a_pass_is_named_by_dash_f_and_unnamed_by_dash_f_no() {
3965        let (opts, _) = compile(&["-c", "-O0", "-ffold", "-fno-fold", "-ffold", "a.c"]);
3966        assert_eq!(
3967            opts.passes,
3968            [("fold".to_owned(), true), ("fold".to_owned(), false), ("fold".to_owned(), true)]
3969        );
3970
3971        let e = parse_args(&args(&["-fno-such-pass", "a.c"])).unwrap_err();
3972        assert!(e.message.contains("unknown option"), "{}", e.message);
3973    }
3974
3975    #[test]
3976    fn pass_fuel_names_a_pass_and_a_count_and_refuses_anything_else() {
3977        let (opts, _) = compile(&["-c", "-O2", "-fpass-fuel=fold=3", "a.c"]);
3978        assert_eq!(opts.pass_fuel, [("fold".to_owned(), 3)]);
3979
3980        let e = parse_args(&args(&["-fpass-fuel=fold", "a.c"])).unwrap_err();
3981        assert!(e.message.contains("<pass>=<count>"), "{}", e.message);
3982        let e = parse_args(&args(&["-fpass-fuel=nosuch=3", "a.c"])).unwrap_err();
3983        assert!(e.message.contains("--print-pipeline"), "{}", e.message);
3984        let e = parse_args(&args(&["-fpass-fuel=fold=lots", "a.c"])).unwrap_err();
3985        assert!(e.message.contains("not a number"), "{}", e.message);
3986    }
3987
3988    #[test]
3989    fn global_pass_fuel_is_a_count_on_its_own_and_defaults_to_no_limit() {
3990        let (opts, _) = compile(&["-c", "-O2", "a.c"]);
3991        assert_eq!(opts.pass_fuel_global, None);
3992
3993        let (opts, _) = compile(&["-c", "-O2", "-fpass-fuel-global=12", "a.c"]);
3994        assert_eq!(opts.pass_fuel_global, Some(12));
3995        // And it is not the per pass flag with a longer name, so neither spelling swallows the
3996        // other.
3997        assert!(opts.pass_fuel.is_empty());
3998
3999        let e = parse_args(&args(&["-fpass-fuel-global=lots", "a.c"])).unwrap_err();
4000        assert!(e.message.contains("not a number"), "{}", e.message);
4001    }
4002
4003    #[test]
4004    fn a_gate_names_a_pass_and_optionally_the_functions_it_covers() {
4005        let (opts, _) = compile(&["-c", "-O2", "-fdisable-fold", "-fenable-fold=2-4,main", "a.c"]);
4006        assert_eq!(
4007            opts.pass_gates,
4008            [(false, "fold".to_owned()), (true, "fold=2-4,main".to_owned())],
4009            "the order is what decides, so it has to survive the parse"
4010        );
4011
4012        let e = parse_args(&args(&["-fdisable-nosuch", "a.c"])).unwrap_err();
4013        assert!(e.message.contains("--print-pipeline"), "{}", e.message);
4014        let e = parse_args(&args(&["-fenable-fold=9-2", "a.c"])).unwrap_err();
4015        assert!(e.message.contains("ends before it starts"), "{}", e.message);
4016        let e = parse_args(&args(&["-fdisable-fold=", "a.c"])).unwrap_err();
4017        assert!(e.message.contains("is empty"), "{}", e.message);
4018    }
4019
4020    #[test]
4021    fn the_pipeline_listing_says_which_passes_a_gate_touched() {
4022        let (opts, _) = compile(&["-c", "-O2", "-fdisable-fold=main", "a.c"]);
4023        let text = print_pipeline(&opts);
4024        assert!(text.contains("fold, "), "{text}");
4025        assert!(text.contains("[off for main]"), "{text}");
4026    }
4027
4028    /// The spelling is checked while the arguments are read, because a dump that names a pass
4029    /// this compiler does not have is a typo, and a typo found after the compilation has run is
4030    /// found too late to be any use.
4031    #[test]
4032    fn a_dump_is_checked_when_it_is_asked_for_rather_than_when_it_is_taken() {
4033        let (opts, _) = compile(&["-c", "-O2", "-fdump-ir=all", "-fdump-ir=after-fold", "a.c"]);
4034        assert_eq!(opts.dump_ir, ["all", "after-fold"]);
4035
4036        let e = parse_args(&args(&["-fdump-ir=after-nosuch", "a.c"])).unwrap_err();
4037        assert!(e.message.contains("nosuch"), "{}", e.message);
4038        assert!(parse_args(&args(&["-fdump-ir=sideways-fold", "a.c"])).is_err());
4039    }
4040
4041    /// Every spelling `-fopt-info` takes, and the one it does not.
4042    ///
4043    /// The keywords are checked here for the same reason a dump's pass name is: a person who
4044    /// misspelled one gets no output, and no output is also what a compilation where nothing
4045    /// happened looks like. Telling those two apart is the entire reason to reach for this flag.
4046    #[test]
4047    fn opt_info_takes_kinds_and_a_file_and_refuses_a_kind_it_does_not_have() {
4048        let (opts, _) = compile(&["-c", "-O2", "-fopt-info", "a.c"]);
4049        assert_eq!(opts.opt_info, [""], "a bare flag asks for the rewrites");
4050        assert_eq!(opts.opt_info_file, None, "and goes to standard error");
4051
4052        let (opts, _) = compile(&["-c", "-O2", "-fopt-info-missed-note", "a.c"]);
4053        assert_eq!(opts.opt_info, ["missed-note"]);
4054
4055        // Two flags add up rather than the second replacing the first, and the file is the last
4056        // one that named a file, which is how GCC treats both.
4057        let (opts, _) =
4058            compile(&["-c", "-O2", "-fopt-info-missed=one.txt", "-fopt-info-all=two.txt", "a.c"]);
4059        assert_eq!(opts.opt_info, ["missed", "all"]);
4060        assert_eq!(opts.opt_info_file.as_deref(), Some("two.txt"));
4061
4062        let e = parse_args(&args(&["-fopt-info-vectorized", "a.c"])).unwrap_err();
4063        assert!(e.message.contains("vectorized"), "{}", e.message);
4064        assert!(e.message.contains("`missed`"), "{}", e.message);
4065        let e = parse_args(&args(&["-fopt-info-missed=", "a.c"])).unwrap_err();
4066        assert!(e.message.contains("no file"), "{}", e.message);
4067    }
4068
4069    #[test]
4070    fn verify_each_is_unstable_and_off_unless_it_was_asked_for() {
4071        let (opts, _) = compile(&["-c", "-Zverify-each", "a.c"]);
4072        assert!(opts.verify_each);
4073        assert!(!USAGE.contains("verify-each"), "an unstable option stays out of the usage text");
4074    }
4075
4076    #[test]
4077    fn dash_o_needs_an_argument() {
4078        let e = parse_args(&args(&["a.c", "-o"])).unwrap_err();
4079        assert_eq!(e.message, "-o requires an argument");
4080    }
4081
4082    #[test]
4083    fn dash_d_and_dash_u_are_read_joined_or_separated_and_keep_their_order() {
4084        let (opts, _) = compile(&["-DFOO=1", "-D", "BAR", "-UBAZ", "-U", "QUX", "a.c"]);
4085        assert_eq!(opts.defines, ["FOO=1", "BAR"]);
4086        assert_eq!(opts.undefines, ["BAZ", "QUX"]);
4087    }
4088
4089    #[test]
4090    fn the_include_flags_land_on_the_chain_each_one_names() {
4091        // A sysroot with nothing under it, so that the library's own directories are the
4092        // same on every machine this test runs on, which is none of them.
4093        let (opts, _) = compile(&[
4094            "-Ii",
4095            "-iquote",
4096            "q",
4097            "-isystem",
4098            "sys",
4099            "-idirafter",
4100            "after",
4101            "--sysroot=/nowhere-at-all",
4102            "a.c",
4103        ]);
4104        let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
4105        // The compiler's own headers sit after every `-isystem` and before `-idirafter`,
4106        // which is where GCC puts its own: a directory the user named outranks ours.
4107        assert_eq!(dirs, ["q", "i", "sys", runtime::DIR, "after"]);
4108        assert!(!opts.search.dirs()[1].is_system);
4109        assert!(opts.search.dirs()[2].is_system);
4110    }
4111
4112    #[test]
4113    fn the_librarys_headers_come_after_the_compilers_own_and_go_away_with_them() {
4114        // Which machine this runs on decides what is on the path, so the test is about the
4115        // order rather than about the names: ours is on it, the library's follow it, and
4116        // `-nostdinc` is the one flag that takes both halves of the pair off at once.
4117        let (opts, _) = compile(&["a.c"]);
4118        let dirs = opts.search.dirs();
4119        let ours = dirs.iter().position(|d| d.path.to_str() == Some(runtime::DIR));
4120        assert_eq!(ours, Some(0), "{dirs:?}");
4121        assert!(dirs[1..].iter().all(|d| d.is_system), "{dirs:?}");
4122        let (bare, _) = compile(&["-nostdinc", "a.c"]);
4123        assert!(bare.search.dirs().is_empty(), "{:?}", bare.search.dirs());
4124    }
4125
4126    #[test]
4127    fn a_sysroot_moves_the_librarys_directories_and_nothing_else() {
4128        let (opts, _) = compile(&["-isystem", "sys", "--sysroot=/nowhere-at-all", "a.c"]);
4129        let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
4130        assert_eq!(dirs, ["sys", runtime::DIR]);
4131    }
4132
4133    #[test]
4134    fn a_cross_compile_reads_the_targets_own_headers_rather_than_the_ones_next_door() {
4135        // The target is not the machine this test runs on wherever it runs, so the answer is the
4136        // same on all of them: the libc's two include directories for that target, the kernel's
4137        // two, and nothing from here. A header read from here is the quiet failure of section 8.5, a
4138        // program that builds on the build machine and is wrong everywhere else.
4139        let (opts, _) = compile(&["--target=riscv64-linux-musl", "-c", "a.c"]);
4140        let dirs: Vec<&std::path::Path> =
4141            opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
4142        let root = cache::dir().join("sysroots").join("riscv64-linux-musl");
4143        let kernel = cache::dir().join("kernel-headers");
4144        assert_eq!(dirs.len(), 5, "{dirs:?}");
4145        assert_eq!(dirs[0], std::path::Path::new(runtime::DIR));
4146        assert_eq!(dirs[1], root.join("include").join("riscv64"));
4147        assert_eq!(dirs[2], root.join("include").join("generic"));
4148        // The kernel's, which are beside the sysroots rather than inside one, because every target
4149        // that shares an architecture reads the same files.
4150        assert_eq!(dirs[3], kernel.join("riscv"));
4151        assert_eq!(dirs[4], kernel.join("generic"));
4152    }
4153
4154    #[test]
4155    fn a_cross_compile_to_something_that_is_not_linux_reads_no_kernel_headers() {
4156        // The other side of the same answer. Windows has its own system headers and no `linux/` at
4157        // all, so the list is the libc's own and the question never arises, which is the `None` that
4158        // `link::cross_kernel` returns rather than a directory nothing would be found in.
4159        //
4160        // The libc's own is one directory rather than two here, because mingw-w64 publishes a single
4161        // header tree for every architecture and `Sysroot::splits_by_arch` says so.
4162        let (opts, _) = compile(&["--target=x86_64-pc-windows-gnu", "-c", "a.c"]);
4163        let dirs: Vec<&std::path::Path> =
4164            opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
4165        assert_eq!(dirs.len(), 2, "{dirs:?}");
4166        assert!(!dirs.iter().any(|dir| dir.ends_with("kernel-headers")), "{dirs:?}");
4167    }
4168
4169    #[test]
4170    fn the_glibc_version_macro_goes_with_the_bundled_tree_and_with_nothing_else() {
4171        // One tree serves every glibc release, so the release is what the target supplies, and the
4172        // condition is the same one that chose the directories. A host glibc and a tree somebody
4173        // named both define `__GLIBC_MINOR__` in their own `features.h`, and two definitions with
4174        // different values is a warning on every compilation of every file.
4175        //
4176        // The architecture is chosen against this machine's rather than written down, because the
4177        // bundled tree is only in effect for a target that is not this machine. The first version of
4178        // this test said x86_64-linux-gnu, which is a cross compile on a mac and this machine on a
4179        // Linux runner, so it passed here and failed there.
4180        let gnu = format!("--target={}-linux-gnu", cross_arch());
4181        let (bundled, _) = compile(&[&gnu, "-c", "a.c"]);
4182        assert_eq!(bundled.glibc_minor, Some(44));
4183        let pin = format!("{gnu}.2.28");
4184        let (pinned, _) = compile(&[&pin, "-c", "a.c"]);
4185        assert_eq!(pinned.glibc_minor, Some(28));
4186
4187        let (named, _) = compile(&[&gnu, "--sysroot=/nowhere-at-all", "-c", "a.c"]);
4188        assert_eq!(named.glibc_minor, None);
4189        let (none, _) = compile(&[&gnu, "-nostdinc", "-c", "a.c"]);
4190        assert_eq!(none.glibc_minor, None);
4191        let musl = format!("--target={}-linux-musl", cross_arch());
4192        let (musl, _) = compile(&[&musl, "-c", "a.c"]);
4193        assert_eq!(musl.glibc_minor, None);
4194
4195        // And this machine's own target gets nothing, whatever this machine is, because its headers
4196        // come from the machine and its own `features.h` defines the macro. On a glibc Linux box
4197        // that is the case this test had backwards; on a mac it is true for the other reason, which
4198        // is that Darwin is not a glibc target at all.
4199        if let Some(host) = Triple::host() {
4200            let native = format!("--target={}", host.tuple());
4201            let (native, _) = compile(&[&native, "-c", "a.c"]);
4202            assert_eq!(native.glibc_minor, None);
4203        }
4204    }
4205
4206    #[test]
4207    fn a_pinned_release_on_this_machines_own_target_reads_the_bundled_tree() {
4208        // The end to end half of the answer in `link::cross_for`. A release named for this machine's
4209        // own target is a cross compile, so the headers are the bundled tree's and the macro says
4210        // what was asked for rather than what this machine has.
4211        //
4212        // Only on a glibc box, because a release is a glibc release: a mac has no `__GLIBC_MINOR__`
4213        // to get wrong and nothing to pin. That makes this a test the Linux runners carry, which is
4214        // where the case lives.
4215        let Some(host) = Triple::host() else { return };
4216        if host.env != rucc_target::Env::Gnu {
4217            return;
4218        }
4219        let pin = format!("--target={}.2.28", host.tuple());
4220        let (opts, _) = compile(&[&pin, "-c", "a.c"]);
4221        assert_eq!(opts.glibc_minor, Some(28));
4222        let root = cache::dir().join("sysroots").join(format!("{}.2.28", host.tuple()));
4223        let dirs: Vec<&std::path::Path> =
4224            opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
4225        assert!(dirs.iter().any(|dir| dir.starts_with(&root)), "{dirs:?}");
4226        // And nothing of this machine's, which is the failure this was: a program compiled against
4227        // 2.44 declarations and told it was 2.28.
4228        assert!(!dirs.iter().any(|dir| *dir == std::path::Path::new("/usr/include")), "{dirs:?}");
4229    }
4230
4231    /// An architecture that is not this machine's, out of the three the driver has targets for.
4232    ///
4233    /// A test about the bundled sysroot has to name a target that is not the host, because a target
4234    /// that is the host reads the host's own headers and libraries. Asking which machine this is
4235    /// beats picking a row and hoping, and it is two lines.
4236    fn cross_arch() -> &'static str {
4237        match Triple::host().map(|host| host.arch) {
4238            Some(rucc_target::Arch::X86_64) => "aarch64",
4239            _ => "x86_64",
4240        }
4241    }
4242
4243    #[test]
4244    fn a_glibc_newer_than_the_bundled_tree_is_refused_by_name() {
4245        // Both versions in the message, because the two things a person can do about it are pin a
4246        // release the tree has and name a sysroot that has the one they asked for, and neither is a
4247        // choice they can make without knowing which release the tree is.
4248        //
4249        // Not this machine's architecture, for the reason the test above gives: the refusal is about
4250        // the bundled tree, and the bundled tree is not what a target that is this machine reads.
4251        let target = format!("--target={}-linux-gnu.2.99", cross_arch());
4252        let message = refused(&[&target, "-c", "a.c"]);
4253        assert!(message.contains("asked for glibc 2.99"), "{message}");
4254        assert!(message.contains("bundled headers are glibc 2.44"), "{message}");
4255        assert!(message.contains("--sysroot"), "{message}");
4256    }
4257
4258    #[test]
4259    fn a_sysroot_the_user_named_is_still_what_a_cross_compile_reads() {
4260        // The tree somebody assembled beats the one we would build, on the headers as on the
4261        // libraries. It is empty here, which is why the list comes out short: the directories under
4262        // it are checked for rather than assumed, and a tree that is not there offers nothing.
4263        let (opts, _) =
4264            compile(&["--target=riscv64-linux-musl", "--sysroot=/nowhere-at-all", "-c", "a.c"]);
4265        let dirs: Vec<&std::path::Path> =
4266            opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
4267        assert_eq!(dirs, [std::path::Path::new(runtime::DIR)]);
4268    }
4269
4270    #[test]
4271    fn dash_i_dash_moves_the_bracket_directories_into_the_quoted_chain() {
4272        let (opts, _) =
4273            compile(&["-Iinc1", "-iquote", "inc2", "-I-", "-Iinc3", "-nostdinc", "a.c"]);
4274        let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
4275        assert_eq!(dirs, ["inc1", "inc2", "inc3"]);
4276        // An angled include sees only what came after the flag.
4277        assert_eq!(opts.search.start(IncludeForm::Angled), 2);
4278        assert!(!opts.search.searches_current_dir());
4279    }
4280
4281    #[test]
4282    fn the_prefix_flags_stick_what_iprefix_said_on_the_front_of_what_follows_it() {
4283        let (opts, _) = compile(&[
4284            "-iprefix",
4285            "/tools/",
4286            "-iwithprefix",
4287            "late",
4288            "-iwithprefixbefore",
4289            "early",
4290            "-iprefix",
4291            "/other/",
4292            "-iwithprefix",
4293            "last",
4294            "-nostdinc",
4295            "a.c",
4296        ]);
4297        let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
4298        // `-iwithprefixbefore` is an `-I` and the other two are `-isystem`, which is where GCC
4299        // puts them rather than where its manual says it does.
4300        assert_eq!(dirs, ["/tools/early", "/tools/late", "/other/last"]);
4301        assert!(!opts.search.dirs()[0].is_system);
4302        assert!(opts.search.dirs()[1].is_system);
4303    }
4304
4305    #[test]
4306    fn the_files_named_on_the_command_line_keep_their_order_and_which_flag_named_them() {
4307        let (opts, _) =
4308            compile(&["-include", "one.h", "-imacros", "two.h", "-include", "3.h", "a.c"]);
4309        let names: Vec<&str> = opts.preincludes.iter().map(|p| p.name.as_str()).collect();
4310        assert_eq!(names, ["one.h", "two.h", "3.h"]);
4311        assert_eq!(opts.preincludes.iter().filter(|p| p.macros_only).count(), 1);
4312    }
4313
4314    #[test]
4315    fn nostdinc_takes_the_compilers_own_headers_off_the_path() {
4316        let (opts, _) = compile(&["-Ii", "-nostdinc", "a.c"]);
4317        let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
4318        assert_eq!(dirs, ["i"]);
4319    }
4320
4321    #[test]
4322    fn the_dialect_flags_set_the_language_and_the_extensions_separately() {
4323        let (opts, _) = compile(&["-std=gnu11", "a.c"]);
4324        assert_eq!(opts.std, Std::C11);
4325        assert!(opts.gnu_extensions);
4326
4327        let (opts, _) = compile(&["-std=iso9899:1999", "a.c"]);
4328        assert_eq!(opts.std, Std::C99);
4329        assert!(!opts.gnu_extensions);
4330
4331        let (opts, _) = compile(&["-ansi", "a.c"]);
4332        assert_eq!(opts.std, Std::C89);
4333        assert!(!opts.gnu_extensions);
4334
4335        let e = parse_args(&args(&["-std=c94jr", "a.c"])).unwrap_err();
4336        assert!(e.message.contains("unknown dialect"), "{}", e.message);
4337    }
4338
4339    #[test]
4340    fn the_dump_letters_are_a_family_and_everything_else_beginning_with_d_is_not() {
4341        let (opts, _) = compile(&["-dM", "a.c"]);
4342        assert!(opts.dumps.macros);
4343
4344        // Packed, the way GCC takes them, and a letter in the family we have not written yet
4345        // is accepted and does nothing rather than failing a build.
4346        let (opts, _) = compile(&["-dDM", "a.c"]);
4347        assert!(opts.dumps.macros);
4348        let (opts, _) = compile(&["-dD", "a.c"]);
4349        assert!(!opts.dumps.macros);
4350
4351        let (opts, _) = compile(&["a.c"]);
4352        assert!(!opts.dumps.any());
4353
4354        // `-dumpversion` is a different flag that happens to start the same way, and it is read
4355        // as itself rather than as a dump of nothing.
4356        assert_eq!(printed(&["-dumpversion", "a.c"]), VERSION);
4357    }
4358
4359    #[test]
4360    fn the_gcc_version_claimed_is_a_flag_and_the_short_spellings_are_the_ones_people_write() {
4361        let (opts, _) = compile(&["a.c"]);
4362        assert_eq!(
4363            opts.gnuc,
4364            GnucVersion { major: 7, minor: 0, patch: 0 },
4365            "the lowest claim a modern glibc gives its own declarations to"
4366        );
4367
4368        let (opts, _) = compile(&["-fgnuc-version=15.1.0", "a.c"]);
4369        assert_eq!(opts.gnuc, GnucVersion { major: 15, minor: 1, patch: 0 });
4370
4371        // A missing component is zero. `gcc -dumpversion` says `15` on a release with no
4372        // patchlevel and a harness that pastes that back has to be understood.
4373        let (opts, _) = compile(&["-fgnuc-version=15", "a.c"]);
4374        assert_eq!(opts.gnuc, GnucVersion { major: 15, minor: 0, patch: 0 });
4375
4376        let (opts, _) = compile(&["-fgnuc-version=13.2", "a.c"]);
4377        assert_eq!(opts.gnuc, GnucVersion { major: 13, minor: 2, patch: 0 });
4378
4379        let e = parse_args(&args(&["-fgnuc-version=15.x", "a.c"])).unwrap_err();
4380        assert!(e.message.contains("minor that is not a number"), "{}", e.message);
4381
4382        let e = parse_args(&args(&["-fgnuc-version=1.2.3.4", "a.c"])).unwrap_err();
4383        assert!(e.message.contains("more than three"), "{}", e.message);
4384    }
4385
4386    #[test]
4387    fn pedantic_has_two_spellings_and_is_not_the_same_knob_as_the_dialect() {
4388        let (opts, _) = compile(&["-std=c17", "-pedantic", "a.c"]);
4389        assert!(opts.pedantic);
4390        assert_eq!(opts.std, Std::C17);
4391
4392        // The `-W` family's name for it, which is what a build that groups its warning flags
4393        // tends to write.
4394        let (opts, _) = compile(&["-Wpedantic", "a.c"]);
4395        assert!(opts.pedantic);
4396
4397        let (opts, _) = compile(&["-std=c17", "a.c"]);
4398        assert!(!opts.pedantic, "a dialect on its own does not diagnose an extension");
4399    }
4400
4401    #[test]
4402    fn dash_p_and_dash_ffreestanding_reach_the_options() {
4403        let (opts, _) = compile(&["-E", "-P", "-ffreestanding", "a.c"]);
4404        assert!(!opts.line_markers);
4405        assert!(!opts.hosted);
4406        assert_eq!(opts.emit, EmitKind::Preprocessed);
4407    }
4408
4409    /// The two ways a build says it means its own function by a name the C library also has.
4410    ///
4411    /// `-fno-builtin` is all of them and `-fno-builtin-<name>` is one, and the second is what a
4412    /// build writes when it means its own `memcpy` and the library's everything else. The name is
4413    /// kept as it was written and not checked against anything, because a program is allowed to
4414    /// mean something by a name this compiler has never heard of.
4415    #[test]
4416    fn the_builtin_flags_are_read_in_both_directions_and_one_name_at_a_time() {
4417        let (opts, _) = compile(&["-c", "a.c"]);
4418        assert!(opts.builtins, "a library name means the library function by default");
4419        assert!(opts.no_builtin.is_empty());
4420
4421        let (opts, _) = compile(&["-c", "-fno-builtin", "a.c"]);
4422        assert!(!opts.builtins);
4423
4424        let (opts, _) = compile(&["-c", "-fno-builtin", "-fbuiltin", "a.c"]);
4425        assert!(opts.builtins, "the last mention decides");
4426
4427        let (opts, _) = compile(&["-c", "-fno-builtin-memcpy", "-fno-builtin-nonesuch", "a.c"]);
4428        assert!(opts.builtins, "one name is not the family");
4429        assert_eq!(opts.no_builtin, vec!["memcpy".to_owned(), "nonesuch".to_owned()]);
4430    }
4431
4432    /// `-fvisibility=`, which is on every cmake project that cares about which names it exports
4433    /// and which was refused as an unknown option until now.
4434    ///
4435    /// Four spellings and three answers. `internal` is hidden plus a promise about never taking
4436    /// the address across a component boundary, and nothing derives anything from that promise
4437    /// here, so it comes out as the weaker of the two rather than as a refusal that stops a build
4438    /// over a distinction this compiler does not make.
4439    #[test]
4440    fn visibility_takes_the_four_spellings_gcc_takes_and_refuses_the_rest() {
4441        let (opts, _) = compile(&["-c", "a.c"]);
4442        assert_eq!(opts.visibility, Visibility::Default, "exported unless something says not");
4443
4444        for (written, wanted) in [
4445            ("default", Visibility::Default),
4446            ("hidden", Visibility::Hidden),
4447            ("internal", Visibility::Hidden),
4448            ("protected", Visibility::Protected),
4449        ] {
4450            let (opts, _) = compile(&["-c", &format!("-fvisibility={written}"), "a.c"]);
4451            assert_eq!(opts.visibility, wanted, "{written}");
4452        }
4453
4454        // The last mention decides, which is what every other flag of this shape does and what a
4455        // build that turns something off for one directory relies on.
4456        let (opts, _) = compile(&["-c", "-fvisibility=hidden", "-fvisibility=default", "a.c"]);
4457        assert_eq!(opts.visibility, Visibility::Default, "the last mention decides");
4458
4459        // A spelling gcc does not take is refused rather than read as the default, because a
4460        // build that meant hidden and got exported is a library with the wrong interface and
4461        // nothing said about it anywhere.
4462        let failed = parse_args(&args(&["-fvisibility=none", "a.c"])).expect_err("refused");
4463        assert!(failed.to_string().contains("is not a visibility"), "{failed}");
4464    }
4465
4466    /// `-ffp-contract=`, which is the one flag in the floating point group that is kept rather than
4467    /// described, and the values are gcc 16's three.
4468    #[test]
4469    fn how_far_a_multiply_and_an_addition_may_be_fused_is_asked_for() {
4470        let (opts, _) = compile(&["-c", "a.c"]);
4471        assert_eq!(opts.fp_contract, Contract::Off, "a licence nobody granted is not assumed");
4472
4473        for (written, wanted) in
4474            [("off", Contract::Off), ("on", Contract::On), ("fast", Contract::Fast)]
4475        {
4476            let (opts, _) = compile(&["-c", &format!("-ffp-contract={written}"), "a.c"]);
4477            assert_eq!(opts.fp_contract, wanted, "{written}");
4478        }
4479
4480        let (opts, _) = compile(&["-c", "-ffp-contract=fast", "-ffp-contract=off", "a.c"]);
4481        assert_eq!(opts.fp_contract, Contract::Off, "the last mention decides");
4482
4483        // Refused rather than read as one of the three, because a build that asked for no fusing
4484        // and was given the default would be one whose numbers change and whose command line says
4485        // they should not. gcc refuses the same spellings and names the same three in its message.
4486        for bad in ["-ffp-contract=none", "-ffp-contract=", "-ffp-contract=Fast"] {
4487            let failed = parse_args(&args(&[bad, "a.c"])).expect_err("refused");
4488            assert!(failed.to_string().contains("is not a contraction"), "{bad}: {failed}");
4489        }
4490
4491        // And the other one that takes a value, which is taken and kept nowhere: every operation
4492        // here is computed in the type it was written in, so `standard` is what happens and the
4493        // other two are permission to do something this does not do.
4494        let failed = parse_args(&args(&["-fexcess-precision=long", "a.c"])).expect_err("refused");
4495        assert!(failed.to_string().contains("is not an excess precision"), "{failed}");
4496    }
4497
4498    /// The four prefix mapping flags, which are what a distribution passes to get the same bytes
4499    /// out of `/build/pkg-1.2` and out of `/home/someone/pkg-1.2`. Three lists rather than one
4500    /// because gcc has three, and `-ffile-prefix-map=` is the three of them at once.
4501    #[test]
4502    fn a_prefix_mapping_flag_goes_on_the_list_its_spelling_names() {
4503        let (opts, _) = compile(&["-c", "a.c"]);
4504        assert!(opts.prefix_map.macros.is_empty(), "nothing is rewritten unless it is asked for");
4505        assert!(opts.prefix_map.debug.is_empty(), "nor here");
4506        assert!(opts.prefix_map.profile.is_empty(), "nor here");
4507
4508        let (opts, _) = compile(&["-c", "-fmacro-prefix-map=/build=.", "a.c"]);
4509        assert_eq!(opts.prefix_map.macros.apply("/build/a.c"), "./a.c", "the one it names");
4510        assert!(opts.prefix_map.debug.is_empty(), "and not the two it does not");
4511
4512        let (opts, _) = compile(&["-c", "-fdebug-prefix-map=/build=.", "a.c"]);
4513        assert_eq!(opts.prefix_map.debug.apply("/build/a.c"), "./a.c", "the one it names");
4514        assert!(opts.prefix_map.macros.is_empty(), "and not the two it does not");
4515
4516        let (opts, _) = compile(&["-c", "-fprofile-prefix-map=/build=.", "a.c"]);
4517        assert_eq!(opts.prefix_map.profile.apply("/build/a.c"), "./a.c", "the one it names");
4518        assert!(opts.prefix_map.macros.is_empty(), "and not the two it does not");
4519
4520        let (opts, _) = compile(&["-c", "-ffile-prefix-map=/build=.", "a.c"]);
4521        for list in [&opts.prefix_map.macros, &opts.prefix_map.debug, &opts.prefix_map.profile] {
4522            assert_eq!(list.apply("/build/a.c"), "./a.c", "all three at once");
4523        }
4524
4525        // Every mention is kept and the last one that matches wins, unlike the flags above whose
4526        // last mention replaces the earlier ones. A build writes one of these per source root and
4527        // expects all of them to be in force, which is the whole point of a list.
4528        let (opts, _) =
4529            compile(&["-c", "-ffile-prefix-map=/a=one", "-ffile-prefix-map=/b=two", "a.c"]);
4530        assert_eq!(opts.prefix_map.macros.apply("/a/x.c"), "one/x.c", "the earlier one still acts");
4531        assert_eq!(opts.prefix_map.macros.apply("/b/x.c"), "two/x.c", "and so does the later one");
4532
4533        // An argument with no `=` is refused rather than ignored, because a build whose paths were
4534        // meant to be rewritten and were not is one that ships the build directory's name and says
4535        // nothing about it. gcc refuses the same thing.
4536        for bad in ["-fmacro-prefix-map=nope", "-ffile-prefix-map=", "-fdebug-prefix-map=/build"] {
4537            let failed = parse_args(&args(&[bad, "a.c"])).expect_err("refused");
4538            assert!(failed.to_string().contains("is not a rewrite for"), "{bad}: {failed}");
4539        }
4540    }
4541
4542    /// `-ffunction-sections` and `-fdata-sections`, which are what make `--gc-sections` able to
4543    /// drop anything: a linker can leave out a section nothing reaches and cannot leave out half of
4544    /// one. A kernel and an embedded image are both linked that way.
4545    ///
4546    /// Two flags rather than one because gcc has two, and a build that asks for one of them and not
4547    /// the other is a build that measured something: splitting the code is nearly free at link time
4548    /// and splitting the data can defeat the linker's ordering of what is next to what.
4549    #[test]
4550    fn a_section_per_function_and_a_section_per_variable_are_asked_for_one_at_a_time() {
4551        let (opts, _) = compile(&["-c", "a.c"]);
4552        assert!(!opts.function_sections, "one text section unless something says otherwise");
4553        assert!(!opts.data_sections);
4554
4555        let (opts, _) = compile(&["-c", "-ffunction-sections", "a.c"]);
4556        assert!(opts.function_sections);
4557        assert!(!opts.data_sections, "one flag is not the other");
4558
4559        let (opts, _) = compile(&["-c", "-fdata-sections", "a.c"]);
4560        assert!(opts.data_sections);
4561        assert!(!opts.function_sections);
4562
4563        // Both directions taken, and the off one is what happens anyway rather than a refusal,
4564        // since a build that writes it is asking for the default.
4565        let (opts, _) = compile(&[
4566            "-c",
4567            "-ffunction-sections",
4568            "-fno-function-sections",
4569            "-fdata-sections",
4570            "-fno-data-sections",
4571            "a.c",
4572        ]);
4573        assert!(!opts.function_sections, "the last mention decides");
4574        assert!(!opts.data_sections, "the last mention decides");
4575    }
4576
4577    /// `-fgnu89-inline`, which is off by default and is not implied by anything on the command
4578    /// line, since the dialect asks for GNU's reading further in rather than through this.
4579    #[test]
4580    fn gnu89_inline_is_off_until_it_is_asked_for_and_the_last_mention_decides() {
4581        let (opts, _) = compile(&["-c", "a.c"]);
4582        assert!(!opts.gnu89_inline, "C's reading of inline by default");
4583
4584        let (opts, _) = compile(&["-c", "-fgnu89-inline", "a.c"]);
4585        assert!(opts.gnu89_inline);
4586
4587        let (opts, _) = compile(&["-c", "-fgnu89-inline", "-fno-gnu89-inline", "a.c"]);
4588        assert!(!opts.gnu89_inline, "the last mention decides");
4589
4590        // The C89 dialects are under GNU's reading whether this was written or not, so the flag
4591        // stays off there and the dialect is what the checker and the macro set both ask. That is
4592        // also why `-std=c89 -fno-gnu89-inline` needs no diagnostic: it asks for the reading the
4593        // dialect already has. gcc refuses that command line, which is measured in the issue.
4594        let (opts, _) = compile(&["-c", "-std=c89", "a.c"]);
4595        assert!(!opts.gnu89_inline);
4596    }
4597
4598    /// Both spellings of both frame flags, since a build that wants one usually writes the
4599    /// other beside it for the one file that has to be compiled the ordinary way.
4600    #[test]
4601    fn the_two_frame_flags_are_read_in_both_directions() {
4602        let (opts, _) = compile(&["-c", "a.c"]);
4603        assert!(!opts.frame_pointer, "gcc omits it above -O0 and so does this");
4604        assert!(opts.red_zone, "the psABI has one and nothing said not to use it");
4605
4606        let (opts, _) = compile(&["-c", "-fno-omit-frame-pointer", "-mno-red-zone", "a.c"]);
4607        assert!(opts.frame_pointer);
4608        assert!(!opts.red_zone);
4609
4610        let (opts, _) = compile(&[
4611            "-c",
4612            "-fno-omit-frame-pointer",
4613            "-fomit-frame-pointer",
4614            "-mno-red-zone",
4615            "-mred-zone",
4616            "a.c",
4617        ]);
4618        assert!(!opts.frame_pointer, "the last one wins, as it does in gcc");
4619        assert!(opts.red_zone);
4620    }
4621
4622    /// Four flags rather than one with an argument, which is how gcc spells them, and the negative
4623    /// spelled three ways because a build that turns one off writes whichever it turned on.
4624    #[test]
4625    fn the_stack_protector_is_four_flags_and_the_last_one_wins() {
4626        let (opts, _) = compile(&["-c", "a.c"]);
4627        assert_eq!(opts.protector, Protector::None, "gcc protects nothing unless it was asked");
4628
4629        for (flag, want) in [
4630            ("-fstack-protector", Protector::Buffers),
4631            ("-fstack-protector-strong", Protector::Strong),
4632            ("-fstack-protector-all", Protector::All),
4633        ] {
4634            let (opts, _) = compile(&["-c", flag, "a.c"]);
4635            assert_eq!(opts.protector, want, "{flag}");
4636        }
4637
4638        // What a package build does: the strong one in the global flags and one directory that
4639        // cannot have a protector turning it off on the line after.
4640        for off in ["-fno-stack-protector", "-fno-stack-protector-strong"] {
4641            let (opts, _) = compile(&["-c", "-fstack-protector-strong", off, "a.c"]);
4642            assert_eq!(opts.protector, Protector::None, "{off}");
4643        }
4644        let (opts, _) = compile(&["-c", "-fno-stack-protector", "-fstack-protector-all", "a.c"]);
4645        assert_eq!(opts.protector, Protector::All, "the last one wins either way round");
4646    }
4647
4648    /// A switch rather than a level, because how a frame is taken is one question and which
4649    /// functions get a canary is another, and gcc spells it that way for the same reason.
4650    #[test]
4651    fn taking_a_frame_a_page_at_a_time_is_off_until_it_is_asked_for() {
4652        let (opts, _) = compile(&["-c", "a.c"]);
4653        assert!(!opts.stack_clash, "gcc takes a frame in one subtraction unless it was asked");
4654
4655        let (opts, _) = compile(&["-c", "-fstack-clash-protection", "a.c"]);
4656        assert!(opts.stack_clash);
4657
4658        // The same shape a package build uses for the protector: on in the global flags and off
4659        // for the one directory that cannot have it.
4660        let (opts, _) =
4661            compile(&["-c", "-fstack-clash-protection", "-fno-stack-clash-protection", "a.c"]);
4662        assert!(!opts.stack_clash);
4663        let (opts, _) =
4664            compile(&["-c", "-fno-stack-clash-protection", "-fstack-clash-protection", "a.c"]);
4665        assert!(opts.stack_clash, "the last one wins either way round");
4666
4667        // The two are independent, since one is about the frame and the other about the function.
4668        let (opts, _) =
4669            compile(&["-c", "-fstack-clash-protection", "-fstack-protector-strong", "a.c"]);
4670        assert!(opts.stack_clash);
4671        assert_eq!(opts.protector, Protector::Strong);
4672    }
4673
4674    /// One flag with an argument rather than a family of spellings, because what it asks about is
4675    /// which of the two edges of a control flow transfer is checked and the two are not separate
4676    /// questions to the hardware.
4677    #[test]
4678    fn which_control_flow_edges_are_checked_is_asked_for_by_name() {
4679        let (opts, _) = compile(&["-c", "a.c"]);
4680        assert_eq!(opts.control, Control::None, "gcc's default on the targets this compiler has");
4681
4682        for (arg, want) in [
4683            ("-fcf-protection", Control::Full),
4684            ("-fcf-protection=full", Control::Full),
4685            ("-fcf-protection=branch", Control::Branch),
4686            ("-fcf-protection=return", Control::Return),
4687            ("-fcf-protection=none", Control::None),
4688            ("-fcf-protection=check", Control::Check),
4689        ] {
4690            let (opts, _) = compile(&["-c", arg, "a.c"]);
4691            assert_eq!(opts.control, want, "{arg}");
4692        }
4693
4694        // The shape a package build uses: on in the global flags and off for the one directory
4695        // that cannot have it, whichever of the two spellings of off it reaches for.
4696        let (opts, _) = compile(&["-c", "-fcf-protection=full", "-fno-cf-protection", "a.c"]);
4697        assert_eq!(opts.control, Control::None);
4698        let (opts, _) = compile(&["-c", "-fno-cf-protection", "-fcf-protection=branch", "a.c"]);
4699        assert_eq!(opts.control, Control::Branch, "the last one wins either way round");
4700    }
4701
4702    /// The profiler is asked for by two spellings, and where its hook goes by two more.
4703    ///
4704    /// The two halves are separate on purpose. `-mfentry` on its own says where a call would go and
4705    /// asks for no call, which is what gcc does with it, and a build system that sets it globally
4706    /// and asks for the profile per directory needs that to be true rather than an error.
4707    ///
4708    /// The link is asserted alongside, because the flag changes it too and a build that compiled
4709    /// with it and linked without it is a program that calls the hook everywhere and never writes a
4710    /// profile.
4711    #[test]
4712    fn the_profiler_and_where_its_hook_goes_are_two_separate_questions() {
4713        let (opts, _) = compile(&["-c", "a.c"]);
4714        assert!(!opts.profile);
4715        assert_eq!(opts.hook, Hook::Platform, "neither was named, so the target decides");
4716
4717        for arg in ["-pg", "-p"] {
4718            let (opts, _) = compile(&["-c", arg, "a.c"]);
4719            assert!(opts.profile, "{arg}");
4720            let (link, _) = linking(&[arg, "a.c"]);
4721            assert!(link.profile, "{arg} changes the link as well");
4722        }
4723
4724        for (arg, want) in [("-mfentry", Hook::Early), ("-mno-fentry", Hook::Late)] {
4725            let (opts, _) = compile(&["-c", arg, "a.c"]);
4726            assert_eq!(opts.hook, want, "{arg}");
4727            assert!(!opts.profile, "{arg} asks for no call of its own");
4728        }
4729
4730        let (opts, _) = compile(&["-c", "-mfentry", "-mno-fentry", "-pg", "a.c"]);
4731        assert_eq!(opts.hook, Hook::Late, "the last one wins");
4732        assert!(opts.profile);
4733    }
4734
4735    /// How much room a patcher is promised, which is one number or two.
4736    ///
4737    /// A command line that did not ask is asserted alongside, because the flag has to be written to
4738    /// mean anything and a build that reserved room nobody asked for would grow every function in
4739    /// it for nothing.
4740    #[test]
4741    fn the_room_a_patcher_is_promised_is_a_number_of_bytes_and_where_they_go() {
4742        let (opts, _) = compile(&["-c", "a.c"]);
4743        assert_eq!(opts.patchable, Patchable::default());
4744        assert!(!opts.patchable.any(), "nothing is reserved unless it was asked for");
4745
4746        let (opts, _) = compile(&["-c", "-fpatchable-function-entry=16", "a.c"]);
4747        assert_eq!(opts.patchable, Patchable { total: 16, before: 0 });
4748
4749        let (opts, _) = compile(&["-c", "-fpatchable-function-entry=5,3", "a.c"]);
4750        assert_eq!(opts.patchable, Patchable { total: 5, before: 3 });
4751        assert_eq!(opts.patchable.after(), 2);
4752
4753        // The last one wins, which is what every other flag of this shape does and what a build
4754        // that adds one to a command line it did not write is relying on.
4755        let (opts, _) = compile(&[
4756            "-c",
4757            "-fpatchable-function-entry=5,3",
4758            "-fpatchable-function-entry=2",
4759            "a.c",
4760        ]);
4761        assert_eq!(opts.patchable, Patchable { total: 2, before: 0 });
4762    }
4763
4764    /// And a request nothing could satisfy is refused rather than rounded into one that can be.
4765    #[test]
4766    fn room_in_front_of_the_label_that_is_more_than_the_room_asked_for_is_refused() {
4767        for arg in ["-fpatchable-function-entry=1,2", "-fpatchable-function-entry=x"] {
4768            let e = parse_args(&args(&["-c", arg, "a.c"])).unwrap_err();
4769            assert!(e.message.contains("is not an amount of room to reserve"), "{}", e.message);
4770        }
4771    }
4772
4773    /// What wraps rather than being undefined, which is two questions and three flags.
4774    ///
4775    /// The older flag is the pair of the newer two, which is gcc's own reading of it, so a build
4776    /// that writes `-fno-strict-overflow` gets both and a build that writes one of the others gets
4777    /// only what it asked for.
4778    #[test]
4779    fn what_overflows_rather_than_being_undefined_is_asked_for_two_ways() {
4780        let (opts, _) = compile(&["-c", "a.c"]);
4781        assert_eq!(opts.wrapping, Wrapping::NONE, "nothing wraps unless it was asked for");
4782
4783        let (opts, _) = compile(&["-c", "-fwrapv", "a.c"]);
4784        assert_eq!(opts.wrapping, Wrapping { signed: true, pointer: false, trap: false });
4785
4786        let (opts, _) = compile(&["-c", "-fwrapv-pointer", "a.c"]);
4787        assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: true, trap: false });
4788
4789        let (opts, _) = compile(&["-c", "-fno-strict-overflow", "a.c"]);
4790        assert_eq!(opts.wrapping, Wrapping::ALL);
4791
4792        // And the last one wins, in both directions. A build that turns one of these on globally
4793        // and off for one directory is relying on that, and so is one that writes the pair and
4794        // then takes half of it back.
4795        let (opts, _) = compile(&["-c", "-fwrapv", "-fno-wrapv", "a.c"]);
4796        assert_eq!(opts.wrapping, Wrapping::NONE);
4797
4798        let (opts, _) = compile(&["-c", "-fno-strict-overflow", "-fstrict-overflow", "a.c"]);
4799        assert_eq!(opts.wrapping, Wrapping::NONE);
4800
4801        let (opts, _) = compile(&["-c", "-fno-strict-overflow", "-fno-wrapv-pointer", "a.c"]);
4802        assert_eq!(opts.wrapping, Wrapping { signed: true, pointer: false, trap: false });
4803    }
4804
4805    /// And the other answer to the signed question cannot be held at the same time as the first.
4806    ///
4807    /// A program cannot both wrap and stop, so writing both is writing a contradiction, and gcc
4808    /// resolves it by letting the last one win rather than by reporting anything. That was measured
4809    /// against gcc 16 rather than read out of the manual, which says nothing about it: `-ftrapv
4810    /// -fwrapv` emits no checked calls and `-fwrapv -ftrapv` emits them.
4811    #[test]
4812    fn a_signed_overflow_that_stops_is_the_other_answer_and_not_a_third_one() {
4813        let (opts, _) = compile(&["-c", "-ftrapv", "a.c"]);
4814        assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: false, trap: true });
4815
4816        let (opts, _) = compile(&["-c", "-fwrapv", "-ftrapv", "a.c"]);
4817        assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: false, trap: true });
4818
4819        let (opts, _) = compile(&["-c", "-ftrapv", "-fwrapv", "a.c"]);
4820        assert_eq!(opts.wrapping, Wrapping { signed: true, pointer: false, trap: false });
4821
4822        let (opts, _) = compile(&["-c", "-ftrapv", "-fno-strict-overflow", "a.c"]);
4823        assert_eq!(opts.wrapping, Wrapping::ALL);
4824
4825        let (opts, _) = compile(&["-c", "-ftrapv", "-fno-trapv", "a.c"]);
4826        assert_eq!(opts.wrapping, Wrapping::NONE);
4827
4828        // And the flag that says what may be assumed says nothing about what happens, so it leaves
4829        // this alone where it takes the wrapping away. gcc does the same.
4830        let (opts, _) = compile(&["-c", "-ftrapv", "-fstrict-overflow", "a.c"]);
4831        assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: false, trap: true });
4832    }
4833
4834    /// What a plain `char` is, which is four spellings of two answers and nothing by default.
4835    ///
4836    /// Nothing is the target's own answer and has to stay distinct from both of the others, since
4837    /// the same command line means a signed `char` on x86-64 and an unsigned one on Linux's arm64.
4838    /// The negative spellings are the other flag rather than a way of asking for the default, which
4839    /// was measured against gcc 16: `-fno-signed-char` defines `__CHAR_UNSIGNED__` and
4840    /// `-fno-unsigned-char` does not.
4841    #[test]
4842    fn the_signedness_of_a_plain_char_is_asked_for_in_four_ways() {
4843        let (opts, _) = compile(&["-c", "a.c"]);
4844        assert_eq!(opts.char_signed, None);
4845
4846        for flag in ["-fsigned-char", "-fno-unsigned-char"] {
4847            let (opts, _) = compile(&["-c", flag, "a.c"]);
4848            assert_eq!(opts.char_signed, Some(true), "{flag}");
4849        }
4850
4851        for flag in ["-funsigned-char", "-fno-signed-char"] {
4852            let (opts, _) = compile(&["-c", flag, "a.c"]);
4853            assert_eq!(opts.char_signed, Some(false), "{flag}");
4854        }
4855
4856        // And the last one wins, which is what a build that sets one globally and the other for a
4857        // directory relies on.
4858        let (opts, _) = compile(&["-c", "-funsigned-char", "-fsigned-char", "a.c"]);
4859        assert_eq!(opts.char_signed, Some(true));
4860
4861        // And what is asked for reaches the target, because that is what every other part of the
4862        // compiler asks. The triple is one whose own answer is the opposite, so a session that
4863        // ignored the flag would still read as signed here.
4864        let (opts, _) =
4865            compile(&["-c", "--target=aarch64-unknown-linux-gnu", "-fsigned-char", "a.c"]);
4866        assert!(Session::new(*opts).target.char_is_signed);
4867        let (opts, _) = compile(&["-c", "--target=aarch64-unknown-linux-gnu", "a.c"]);
4868        assert!(!Session::new(*opts).target.char_is_signed);
4869    }
4870
4871    /// And the size of an enumeration, which is one question with two spellings.
4872    #[test]
4873    fn the_smallest_enumeration_is_asked_for_and_taken_back() {
4874        let (opts, _) = compile(&["-c", "a.c"]);
4875        assert!(!opts.short_enums);
4876
4877        let (opts, _) = compile(&["-c", "-fshort-enums", "a.c"]);
4878        assert!(opts.short_enums);
4879
4880        let (opts, _) = compile(&["-c", "-fshort-enums", "-fno-short-enums", "a.c"]);
4881        assert!(!opts.short_enums);
4882
4883        let (opts, _) = compile(&["-c", "-fno-short-enums", "-fshort-enums", "a.c"]);
4884        assert!(opts.short_enums);
4885    }
4886
4887    /// And Microsoft's reading of an anonymous member, which the target answers where the command
4888    /// line said nothing. gcc's mingw build has it on and its Linux build has it off, so a header
4889    /// that closes a nameless union with a macro that expands to nothing is read the way the
4890    /// compiler that platform ships would read it.
4891    #[test]
4892    fn the_microsoft_reading_of_a_member_follows_the_target_until_it_is_asked_for() {
4893        // Named rather than left to the host, since the answer this asks for is the one a target
4894        // that is not Windows gives and on a Windows machine the host is not one of those.
4895        let (opts, _) = compile(&[LINUX, "-c", "a.c"]);
4896        assert!(!Session::new(*opts).ms_extensions());
4897
4898        let (opts, _) = compile(&["-c", "--target=x86_64-pc-windows-gnu", "a.c"]);
4899        assert!(Session::new(*opts).ms_extensions());
4900
4901        let (opts, _) = compile(&["-c", "-fms-extensions", "a.c"]);
4902        assert!(Session::new(*opts).ms_extensions());
4903
4904        let (opts, _) =
4905            compile(&["-c", "--target=x86_64-pc-windows-gnu", "-fno-ms-extensions", "a.c"]);
4906        assert!(!Session::new(*opts).ms_extensions());
4907    }
4908
4909    /// And a value nothing means is refused rather than taken for the nearest thing it looks like.
4910    ///
4911    /// `-fcf-protection=all` is the spelling somebody writes from memory, and a compiler that read
4912    /// it as `full` would be guessing, while one that let it fall through to the optimizer's `-f`
4913    /// family would report it as an unknown pass. Neither is the news the build wants.
4914    #[test]
4915    fn a_control_flow_protection_nothing_means_is_refused() {
4916        let e = parse_args(&args(&["-c", "-fcf-protection=all", "a.c"])).unwrap_err();
4917        assert!(e.message.contains("is not a control flow protection"), "{}", e.message);
4918        assert!(e.message.contains("full, branch, return, none or check"), "{}", e.message);
4919    }
4920
4921    #[test]
4922    fn the_link_flags_are_collected_apart_from_the_compilation() {
4923        let (link, _) = linking(&[
4924            "-static",
4925            "-nostartfiles",
4926            "-rdynamic",
4927            "-s",
4928            "-fuse-ld=mold",
4929            "-L/opt/lib",
4930            "-B",
4931            "/opt/tools",
4932            "a.c",
4933        ]);
4934        assert!(link.is_static);
4935        assert!(link.no_startfiles);
4936        assert!(link.export_dynamic);
4937        assert!(link.strip);
4938        assert_eq!(link.use_ld.as_deref(), Some("mold"));
4939        assert_eq!(link.search, vec![PathBuf::from("/opt/lib")]);
4940        assert_eq!(link.prefixes, vec![PathBuf::from("/opt/tools")]);
4941    }
4942
4943    #[test]
4944    fn a_comma_in_dash_wl_separates_two_arguments() {
4945        // The target is written down because the name of the object is derived from it, and `a.o`
4946        // on a Linux host is `a.obj` on a Windows one. What is under test is the splitting of the
4947        // argument, which has nothing to do with either.
4948        let (_, plan) = linking(&[LINUX, "-Wl,-rpath,/opt/lib", "-Xlinker", "--as-needed", "a.c"]);
4949        let link = plan.link.expect("expected a link step");
4950        assert_eq!(
4951            link.inputs,
4952            vec![
4953                link::Item::Linker("-rpath".into()),
4954                link::Item::Linker("/opt/lib".into()),
4955                link::Item::Linker("--as-needed".into()),
4956                link::Item::File("a.o".into()),
4957            ]
4958        );
4959    }
4960
4961    #[test]
4962    fn a_word_for_the_linker_keeps_its_place_among_the_files_too() {
4963        // What libtool writes around a set of convenience archives, and what #1279 was. Both words
4964        // are about the files between them, so the pair collected out of the line and appended to
4965        // the end is two options that bracket nothing and an archive that went in empty.
4966        let (_, plan) = linking(&[
4967            "--target=x86_64-unknown-linux-gnu",
4968            "a.c",
4969            "-Wl,--whole-archive",
4970            "libaesni.a",
4971            "-Wl,--no-whole-archive",
4972            "-lm",
4973        ]);
4974        let link = plan.link.expect("expected a link step");
4975        assert_eq!(
4976            link.inputs,
4977            vec![
4978                link::Item::File("a.o".into()),
4979                link::Item::Linker("--whole-archive".into()),
4980                link::Item::File("libaesni.a".into()),
4981                link::Item::Linker("--no-whole-archive".into()),
4982                link::Item::Library("m".into()),
4983            ]
4984        );
4985        // And it is not a job, because there is nothing to compile in a word for the linker.
4986        assert_eq!(plan.jobs.len(), 2);
4987    }
4988
4989    #[test]
4990    fn a_word_for_the_linker_on_a_dash_c_line_is_dropped_without_a_word() {
4991        // GCC says nothing about one either. `-Wl,` on a compile line is what a build system
4992        // writes when one variable holds the flags for both, and a note here would be a note on
4993        // every compile of every autotools project.
4994        let (_, plan) = linking(&["-c", "-Wl,--as-needed", "a.c"]);
4995        assert!(plan.link.is_none());
4996        assert!(plan.notes.is_empty(), "{:?}", plan.notes);
4997        assert_eq!(plan.jobs.len(), 1);
4998    }
4999
5000    #[test]
5001    fn a_library_keeps_its_place_between_the_objects() {
5002        // Link order is semantic: `-lm` written between two files resolves for the one before
5003        // it and not for the one after, so a library cannot be collected into a list of its own.
5004        // The target is named because the suffix of an object is the target's and this asserts
5005        // on the names: the same command line on a Windows host plans two `.obj` files.
5006        let (_, plan) = linking(&["--target=x86_64-unknown-linux-gnu", "a.c", "-lm", "b.c"]);
5007        let link = plan.link.expect("expected a link step");
5008        assert_eq!(
5009            link.inputs,
5010            vec![
5011                link::Item::File("a.o".into()),
5012                link::Item::Library("m".into()),
5013                link::Item::File("b.o".into()),
5014            ]
5015        );
5016        // And it is not a job, because there is nothing to compile in a library.
5017        assert_eq!(plan.jobs.len(), 2);
5018    }
5019
5020    #[test]
5021    fn a_library_on_a_dash_c_line_is_a_note_rather_than_an_error() {
5022        let (_, plan) = linking(&["-c", "-lm", "a.c"]);
5023        assert!(plan.link.is_none());
5024        assert!(plan.notes.iter().any(|n| n.contains("-lm")), "{:?}", plan.notes);
5025    }
5026
5027    #[test]
5028    fn the_sysroot_reaches_the_linker_as_well_as_the_headers() {
5029        let (link, _) = linking(&["--sysroot=/opt/root", "a.c"]);
5030        assert_eq!(link.sysroot, Some(PathBuf::from("/opt/root")));
5031    }
5032
5033    fn printed(s: &[&str]) -> String {
5034        match parse_args(&args(s)).expect("expected an answer") {
5035            Action::Print(line) => line,
5036            other => panic!("expected an answer, got {other:?}"),
5037        }
5038    }
5039
5040    fn refused(s: &[&str]) -> String {
5041        parse_args(&args(s)).expect_err("expected a refusal").message
5042    }
5043
5044    #[test]
5045    fn a_warning_flag_this_compiler_has_not_heard_of_is_taken_rather_than_refused() {
5046        // The rule in section 4.1, and the reason for it is autoconf: a configure script finds
5047        // out whether a warning flag exists by passing it and looking at the exit status, so a
5048        // compiler that refuses one it does not know fails a script written for a newer GCC.
5049        let (opts, _) = compile(&["-Wall", "-Wextra", "-Wno-format-truncation", "-c", "a.c"]);
5050        assert!(!opts.warnings_are_errors);
5051        assert!(opts.warnings);
5052        // The two spellings that do mean something are still read.
5053        let (opts, _) = compile(&["-Werror", "-c", "a.c"]);
5054        assert!(opts.warnings_are_errors);
5055        let (opts, _) = compile(&["-w", "-c", "a.c"]);
5056        assert!(!opts.warnings);
5057        // Off without being asked, the way gcc has it off, and both spellings are read.
5058        let (opts, _) = compile(&["-c", "a.c"]);
5059        assert!(!opts.system_header_warnings);
5060        let (opts, _) = compile(&["-Wsystem-headers", "-c", "a.c"]);
5061        assert!(opts.system_header_warnings);
5062        let (opts, _) = compile(&["-Wsystem-headers", "-Wno-system-headers", "-c", "a.c"]);
5063        assert!(!opts.system_header_warnings);
5064        let (opts, _) = compile(&["-pedantic-errors", "-c", "a.c"]);
5065        assert!(opts.pedantic && opts.warnings_are_errors);
5066    }
5067
5068    #[test]
5069    fn an_argument_for_a_separate_tool_is_refused_rather_than_dropped() {
5070        // Every one of these says something about the output, so the wrong answer is silence.
5071        assert!(refused(&["-Wa,--noexecstack", "-c", "a.c"]).contains("separate assembler"));
5072        assert!(refused(&["-Wp,-DX", "-c", "a.c"]).contains("separate assembler"));
5073        assert!(refused(&["-specs=/x", "a.c"]).contains("-specs= is not supported"));
5074        assert!(refused(&["-mcmodel=kernel", "-c", "a.c"]).contains("small code model"));
5075        assert!(refused(&["-gdwarf-4", "-c", "a.c"]).contains("DWARF 5"));
5076        assert!(refused(&["-Ofast", "-c", "a.c"]).contains("fast math"));
5077        // The word size the target does not have, which is a target this compiler was not asked
5078        // for rather than a flag it does not know.
5079        let no32 = refused(&["--target=x86_64-unknown-linux-gnu", "-m32", "-c", "a.c"]);
5080        assert!(no32.contains("32 bit target"), "{no32}");
5081    }
5082
5083    /// `-gz` and the two spellings of the split, which are the two questions about the shape of
5084    /// the debug output rather than about how much of it there is.
5085    ///
5086    /// Both answers here are about what happens when there is debug information to shape, and
5087    /// there is none yet, so what is being asserted is that the flags are read and remembered
5088    /// rather than that anything changed in the output. That is the whole of what taking them
5089    /// claims, and it is worth a test because the day `rucc-debug` writes a section this is where
5090    /// it comes to find out what the command line said.
5091    #[test]
5092    fn the_shape_of_the_debug_output_is_recorded_even_where_there_is_none_of_it() {
5093        let (opts, _) = compile(&["-c", "a.c"]);
5094        assert_eq!(opts.compress, Compress::None, "uncompressed unless somebody asks");
5095
5096        // Bare `-gz` is `-gz=zlib`, measured against gcc 16 rather than read out of the manual,
5097        // which describes the flag without ever saying which algorithm it picks.
5098        assert_eq!(compile(&["-gz", "-c", "a.c"]).0.compress, Compress::Zlib);
5099        for (spelling, want) in [
5100            ("none", Compress::None),
5101            ("zlib", Compress::Zlib),
5102            ("zlib-gnu", Compress::ZlibGnu),
5103            ("zstd", Compress::Zstd),
5104        ] {
5105            let (opts, _) = compile(&[&format!("-gz={spelling}"), "-c", "a.c"]);
5106            assert_eq!(opts.compress, want, "{spelling}");
5107        }
5108
5109        // A value nothing here has heard of is refused rather than rounded to the nearest one,
5110        // because a build that asked for `zstd` and quietly got `zlib` would ship a file its
5111        // reader may not understand and would have no way of finding out.
5112        for bad in ["-gz=gzip", "-gz="] {
5113            let failed = refused(&[bad, "-c", "a.c"]);
5114            assert!(failed.contains("is not a way to compress"), "{bad}: {failed}");
5115        }
5116
5117        // The split is refused in the direction that would have written a file and taken in the
5118        // direction that describes what happens. A build system that names the `.dwo` as an
5119        // output has to hear about it now rather than at the point the file is missing.
5120        let (opts, _) = compile(&["-gno-split-dwarf", "-g", "-c", "a.c"]);
5121        assert!(opts.debug_info, "the negative spelling says nothing about how much");
5122        let failed = refused(&["-gsplit-dwarf", "-c", "a.c"]);
5123        assert!(failed.contains(".dwo"), "the refusal names the file it would have written");
5124    }
5125
5126    /// The `-flto` family, which is the whole of an optimization this compiler does not do.
5127    ///
5128    /// Taken rather than refused because ignoring it gives a correct program that is slower than
5129    /// it could have been, which is section 4.1's hint about speed. The values are still held to
5130    /// gcc's, so a command line written for clang is told rather than quietly taken.
5131    #[test]
5132    fn the_link_time_family_is_read_and_checked_and_nothing_is_done_about_it() {
5133        let (opts, _) = compile(&["-c", "a.c"]);
5134        assert!(!opts.lto.requested, "nothing asks unless the command line does");
5135
5136        let (opts, _) = compile(&["-flto", "-c", "a.c"]);
5137        assert!(opts.lto.requested);
5138        assert_eq!(opts.lto.jobs, LtoJobs::One, "bare -flto is one process, the way gcc reads it");
5139
5140        // The last of the two directions wins, the same as every other pair of `-f` spellings.
5141        assert!(!compile(&["-flto", "-fno-lto", "-c", "a.c"]).0.lto.requested);
5142        assert!(compile(&["-fno-lto", "-flto", "-c", "a.c"]).0.lto.requested);
5143
5144        // A count is a count, and asking for one implies asking for the optimization.
5145        for (spelling, want) in [
5146            ("auto", LtoJobs::Auto),
5147            ("jobserver", LtoJobs::Jobserver),
5148            ("1", LtoJobs::One),
5149            ("8", LtoJobs::Count(8)),
5150        ] {
5151            let (opts, _) = compile(&[&format!("-flto={spelling}"), "-c", "a.c"]);
5152            assert_eq!(opts.lto.jobs, want, "{spelling}");
5153            assert!(opts.lto.requested, "{spelling} asks for it too");
5154        }
5155
5156        // gcc refuses a zero rather than reading it as `-fno-lto`, and `thin` is clang's spelling
5157        // of a question gcc answers with `-flto-partition=`, so somebody who wrote it meant a
5158        // different compiler and gets told so here rather than getting a serial link.
5159        for bad in ["-flto=0", "-flto=thin", "-flto=full", "-flto=-1"] {
5160            let failed = refused(&[bad, "-c", "a.c"]);
5161            assert!(failed.contains("link time jobs"), "{bad}: {failed}");
5162        }
5163
5164        // How the program is cut up before the work is spread over it.
5165        assert_eq!(compile(&["-c", "a.c"]).0.lto.partition, Partition::Balanced, "gcc's default");
5166        for (spelling, want) in [
5167            ("balanced", Partition::Balanced),
5168            ("1to1", Partition::OneToOne),
5169            ("one", Partition::One),
5170            ("max", Partition::Max),
5171            ("none", Partition::None),
5172        ] {
5173            let (opts, _) = compile(&[&format!("-flto-partition={spelling}"), "-c", "a.c"]);
5174            assert_eq!(opts.lto.partition, want, "{spelling}");
5175        }
5176        assert!(refused(&["-flto-partition=big", "-c", "a.c"]).contains("partitioning model"));
5177
5178        // And how hard the bytecode is compressed on its way into the object, which is zstd's
5179        // range of levels and is the range gcc checks an argument against.
5180        assert_eq!(compile(&["-c", "a.c"]).0.lto.compression, None, "whatever it does by default");
5181        assert_eq!(compile(&["-flto-compression-level=0", "-c", "a.c"]).0.lto.compression, Some(0));
5182        let (opts, _) = compile(&["-flto-compression-level=19", "-c", "a.c"]);
5183        assert_eq!(opts.lto.compression, Some(19));
5184        for bad in ["-flto-compression-level=20", "-flto-compression-level=-1"] {
5185            let failed = refused(&[bad, "-c", "a.c"]);
5186            assert!(failed.contains("compression level"), "{bad}: {failed}");
5187        }
5188
5189        // The two pairs that describe an arrangement rather than ask for one. Every object here
5190        // holds its machine code, so the fat spelling is what already happens and the other is a
5191        // smaller file rather than a different program, and the plugin pair is about a tool the
5192        // design in `spec/09-optimizer.md` never loads.
5193        for taken in [
5194            "-ffat-lto-objects",
5195            "-fno-fat-lto-objects",
5196            "-fuse-linker-plugin",
5197            "-fno-use-linker-plugin",
5198        ] {
5199            let (opts, _) = compile(&[taken, "-c", "a.c"]);
5200            assert!(!opts.lto.requested, "{taken} says nothing about whether to do it");
5201        }
5202    }
5203
5204    /// The profile family, which is the only one here that splits down the middle.
5205    ///
5206    /// Reading a profile is taken and writing one is refused, and the line between them is the one
5207    /// section 4.1 draws: ignoring a request to read the counts gives a correct program that is
5208    /// slower than it could have been, and ignoring a request to write them means a file the build
5209    /// declared as an output never appears.
5210    #[test]
5211    fn reading_a_profile_is_taken_and_writing_one_is_refused() {
5212        let (opts, _) = compile(&["-c", "a.c"]);
5213        assert!(!opts.profile_data.requested, "nothing asks unless the command line does");
5214        assert_eq!(opts.profile_data.path, None);
5215
5216        let (opts, _) = compile(&["-fprofile-use", "-c", "a.c"]);
5217        assert!(opts.profile_data.requested);
5218        assert_eq!(opts.profile_data.path, None, "beside the object, the way gcc looks");
5219
5220        let (opts, _) = compile(&["-fprofile-use=/counts", "-c", "a.c"]);
5221        assert!(opts.profile_data.requested, "naming a path asks for it too");
5222        assert_eq!(opts.profile_data.path.as_deref(), Some("/counts"));
5223
5224        // The last of the two directions wins, the same as every other pair of `-f` spellings.
5225        assert!(
5226            !compile(&["-fprofile-use", "-fno-profile-use", "-c", "a.c"]).0.profile_data.requested
5227        );
5228        assert!(
5229            compile(&["-fno-profile-use", "-fprofile-use", "-c", "a.c"]).0.profile_data.requested
5230        );
5231
5232        // The rest of the reading half, which is where the files are and three answers about what
5233        // to make of what is in them.
5234        let (opts, _) = compile(&[
5235            "-fprofile-dir=/build/profiles",
5236            "-fprofile-abs-path",
5237            "-fprofile-correction",
5238            "-fprofile-partial-training",
5239            "-c",
5240            "a.c",
5241        ]);
5242        assert_eq!(opts.profile_data.dir.as_deref(), Some("/build/profiles"));
5243        assert!(opts.profile_data.absolute);
5244        assert!(opts.profile_data.correction);
5245        assert!(opts.profile_data.partial_training);
5246
5247        // Writing one, which is refused by name. The first four instrument the program and the
5248        // last writes a file beside the object, and a build that got neither and no message would
5249        // go on to optimize against counts that were never gathered.
5250        for writing in [
5251            "-fprofile-generate",
5252            "-fprofile-generate=/build/profiles",
5253            "-fprofile-arcs",
5254            "--coverage",
5255            "-fcondition-coverage",
5256            "-fpath-coverage",
5257        ] {
5258            let failed = refused(&[writing, "-c", "a.c"]);
5259            assert!(failed.contains("instrument"), "{writing}: {failed}");
5260        }
5261        assert!(refused(&["-ftest-coverage", "-c", "a.c"]).contains(".gcno"), "it names the file");
5262
5263        // The negative spellings of the refused half are what already happens, so they are taken.
5264        for taken in ["-fno-profile-generate", "-fno-profile-arcs", "-fno-test-coverage"] {
5265            let (opts, _) = compile(&[taken, "-c", "a.c"]);
5266            assert!(!opts.profile_data.requested, "{taken} asks for nothing");
5267        }
5268
5269        // And the flags that describe the instrumentation that is refused above, which are checked
5270        // and dropped. Checked because a typo is worth finding here rather than on the day the
5271        // instrumentation lands.
5272        for taken in [
5273            "-fprofile-update=single",
5274            "-fprofile-update=atomic",
5275            "-fprofile-update=prefer-atomic",
5276            "-fprofile-reproducible=serial",
5277            "-fprofile-reproducible=parallel-runs",
5278            "-fprofile-reproducible=multithreaded",
5279            "-fprofile-values",
5280            "-fno-profile-values",
5281            "-fprofile-info-section",
5282            "-fprofile-filter-files=a.c",
5283            "-fprofile-exclude-files=b.c",
5284            "-fprofile-note=a.gcno",
5285        ] {
5286            let (opts, _) = compile(&[taken, "-c", "a.c"]);
5287            assert!(!opts.profile_data.requested, "{taken} says nothing about reading one");
5288        }
5289        assert!(refused(&["-fprofile-update=none", "-c", "a.c"]).contains("update method"));
5290        assert!(refused(&["-fprofile-reproducible=any", "-c", "a.c"]).contains("reproducibility"));
5291    }
5292
5293    /// The sanitizers, which are refused by name and are the one family refused for a reason that
5294    /// is not about the bytes.
5295    ///
5296    /// A sanitizer is a promise that the program is watched while it runs, so a build that asked
5297    /// for one and was quietly given a program with no checks in it gets a test suite that passes
5298    /// for the wrong reason rather than a slower program.
5299    #[test]
5300    fn a_sanitizer_that_is_still_asked_for_at_the_end_of_the_line_is_refused_by_name() {
5301        for asked in ["address", "undefined", "thread", "kernel-address", "leak", "memory"] {
5302            let failed = refused(&[&format!("-fsanitize={asked}"), "-c", "a.c"]);
5303            assert!(failed.contains(asked), "the refusal names what was asked for: {failed}");
5304            assert!(failed.contains("-fsafety=detect"), "and the nearest thing: {failed}");
5305        }
5306
5307        // A list is every name in it, and the first one still standing is the one named.
5308        let failed = refused(&["-fsanitize=address,undefined", "-c", "a.c"]);
5309        assert!(failed.contains("address"), "{failed}");
5310
5311        // A name that is not one, which is worth its own message: somebody who wrote `-fsanitize`
5312        // with a typo in it has a different problem from somebody who wrote a real one.
5313        for bad in ["-fsanitize=bogus", "-fsanitize=address,bogus", "-fno-sanitize=bogus"] {
5314            let failed = refused(&[bad, "-c", "a.c"]);
5315            assert!(failed.contains("is not a sanitizer"), "{bad}: {failed}");
5316        }
5317
5318        // gcc takes `all` only in the negative, and so does this.
5319        assert!(refused(&["-fsanitize=all", "-c", "a.c"]).contains("only `-fno-sanitize=all`"));
5320
5321        // Asking and then taking it back is asking for nothing, which is why the answer waits for
5322        // the end of the line. A build whose shared flags turn a check on and whose rule for one
5323        // file turns it off again compiles that file here.
5324        for pair in [
5325            ["-fsanitize=address", "-fno-sanitize=address"],
5326            ["-fsanitize=address,undefined", "-fno-sanitize=all"],
5327            ["-fsanitize=undefined", "-fno-sanitize=undefined"],
5328        ] {
5329            let (opts, _) = compile(&[pair[0], pair[1], "-c", "a.c"]);
5330            assert_eq!(opts.safety, rucc_session::Safety::Off, "{pair:?} asked for nothing");
5331        }
5332        // And the other order still asks, because the last word is the one that counts.
5333        assert!(!refused(&["-fno-sanitize=address", "-fsanitize=address", "-c", "a.c"]).is_empty());
5334
5335        // What a check does when it fires is an answer about checks that are refused, so there is
5336        // nothing left for it to change and it is taken.
5337        for taken in [
5338            "-fsanitize-recover=undefined",
5339            "-fno-sanitize-recover=all",
5340            "-fsanitize-trap=undefined",
5341            "-fno-sanitize-trap=all",
5342            "-fsanitize-undefined-trap-on-error",
5343            "-fsanitize-address-use-after-scope",
5344            "-fno-sanitize-address-use-after-scope",
5345            "-fsanitize-sections=.data",
5346        ] {
5347            let (opts, _) = compile(&[taken, "-c", "a.c"]);
5348            assert_eq!(opts.safety, rucc_session::Safety::Off, "{taken} asks for no checking");
5349        }
5350        assert!(refused(&["-fsanitize-recover=bogus", "-c", "a.c"]).contains("is not a sanitizer"));
5351
5352        // Coverage instrumentation is refused rather than dropped, because a fuzzer with no
5353        // feedback runs blind and never says so.
5354        let failed = refused(&["-fsanitize-coverage=trace-pc", "-c", "a.c"]);
5355        assert!(failed.contains("feedback"), "{failed}");
5356        let failed = refused(&["-fsanitize-coverage=trace-pc-guard", "-c", "a.c"]);
5357        assert!(failed.contains("trace-pc or trace-cmp"), "gcc takes two of them: {failed}");
5358    }
5359
5360    #[test]
5361    fn the_levels_gcc_spells_differently_are_the_levels_they_mean() {
5362        assert_eq!(compile(&["-O", "-c", "a.c"]).0.opt_level, OptLevel::O1);
5363        assert_eq!(compile(&["-Og", "-c", "a.c"]).0.opt_level, OptLevel::O1);
5364        assert_eq!(compile(&["-O2", "-c", "a.c"]).0.opt_level, OptLevel::O2);
5365    }
5366
5367    #[test]
5368    fn the_machine_flags_that_name_what_we_already_do_are_taken_and_the_rest_are_not() {
5369        let line = ["--target=x86_64-unknown-linux-gnu", "-m64", "-march=x86-64-v3"];
5370        let (opts, _) =
5371            compile(&[&line[..], &["-mtune=native", "-mabi=sysv", "-c", "a.c"]].concat());
5372        assert_eq!(opts.target.to_string(), "x86_64-unknown-linux-gnu");
5373        let wrong = refused(&["--target=x86_64-unknown-linux-gnu", "-mabi=ms", "-c", "a.c"]);
5374        assert!(wrong.contains("sysv convention"), "{wrong}");
5375    }
5376
5377    #[test]
5378    fn the_thread_flag_is_a_macro_and_a_library_and_the_library_goes_last() {
5379        let (opts, plan) = compile(&["-pthread", "-c", "a.c"]);
5380        assert!(opts.defines.iter().any(|d| d == "_REENTRANT"));
5381        // After the input, because a static link takes what it needs from a library when it
5382        // reaches it and not afterwards.
5383        let names: Vec<&str> = plan.jobs.iter().map(|j| j.input.as_str()).collect();
5384        assert_eq!(names, vec!["a.c"]);
5385    }
5386
5387    #[test]
5388    fn the_questions_a_build_system_asks_before_it_compiles_anything() {
5389        let target = "--target=x86_64-unknown-linux-gnu";
5390        assert_eq!(printed(&[target, "-dumpmachine"]), "x86_64-unknown-linux-gnu");
5391        assert_eq!(printed(&[target, "-dumpversion"]), VERSION);
5392        assert_eq!(printed(&[target, "-dumpfullversion"]), VERSION);
5393        assert_eq!(printed(&[target, "-print-multiarch"]), "x86_64-linux-gnu");
5394        // A name nothing holds comes back unchanged, which is GCC's rule and is what makes the
5395        // answer safe to paste into a link line whether or not the file is there.
5396        assert_eq!(printed(&[target, "-print-file-name=no-such-library.a"]), "no-such-library.a");
5397        assert_eq!(printed(&[target, "-print-prog-name=ld"]), "ld");
5398        let dirs = printed(&[target, "-print-search-dirs"]);
5399        assert!(dirs.starts_with("install: "), "{dirs}");
5400        assert!(dirs.contains("\nlibraries: ="), "{dirs}");
5401    }
5402
5403    #[test]
5404    fn the_sysroot_in_effect_is_the_one_the_command_line_named_or_the_one_for_the_target() {
5405        // A tree the user named is the answer whatever the target is, because it is the answer to
5406        // every other question too.
5407        assert_eq!(printed(&["--sysroot=/opt/cross", "-print-sysroot"]), "/opt/cross");
5408
5409        // A target that is no machine this suite runs on is read under the cache, and the answer is
5410        // the root rather than one of the directories under it, since what asks is looking for a
5411        // file of its own.
5412        let root = cache::dir().join("sysroots").join("riscv64-linux-musl");
5413        assert_eq!(
5414            printed(&["--target=riscv64-linux-musl", "-print-sysroot"]),
5415            root.display().to_string()
5416        );
5417
5418        // And a compile for this machine has no sysroot, which is the empty line GCC prints when it
5419        // was configured without one rather than a `/` that would be a claim about the filesystem.
5420        let host = Triple::host().expect("a host this compiler knows");
5421        assert_eq!(printed(&[&format!("--target={host}"), "-print-sysroot"]), "");
5422    }
5423
5424    #[test]
5425    fn the_provenance_of_a_sysroot_is_the_manifest_it_carries() {
5426        // Section 13.5 wants seven things per input and wants them machine readable, and the manifest
5427        // is the record that already has them, so the flag prints that rather than a second format.
5428        let manifest = "rucc sysroot manifest 3\n\
5429                        target\tx86_64-linux-musl\n\
5430                        kernel\t6.12\n\
5431                        include/generic/stdio.h\tmusl-1.2.5\t\
5432                        https://musl.libc.org/releases/musl-1.2.5.tar.gz\t\
5433                        0000000000000000000000000000000000000000000000000000000000000000\tmit\t\
5434                        bundled\n\
5435                        lib/libc.so\tmusl-1.2.5\t\
5436                        https://musl.libc.org/releases/musl-1.2.5.tar.gz\t\
5437                        1111111111111111111111111111111111111111111111111111111111111111\tmit\t\
5438                        generated\n";
5439        let tree = TempTree::new("provenance", &[("manifest", manifest)]);
5440        let sysroot = format!("--sysroot={}", tree.0.display());
5441        // The kernel line of tamnd/rucc#934 is in the answer without anything here naming it, because
5442        // the flag parses the record and renders it again rather than picking fields out of it. That
5443        // is the reason it prints a manifest and not a format of its own.
5444        //
5445        // The answer is the file without its last newline, because whatever prints it adds one. The
5446        // file is what somebody diffs the output against, so the two have to be the same bytes.
5447        assert_eq!(printed(&[&sysroot, "-print-sysroot-provenance"]) + "\n", manifest);
5448
5449        // A tree with no manifest in it is a tree somebody assembled themselves, and nothing here
5450        // knows where any of it came from. Saying nothing is the only honest answer, and a reader can
5451        // tell it from a manifest with no inputs because that one still has its two header lines.
5452        let bare = TempTree::new("provenance-bare", &[]);
5453        assert_eq!(
5454            printed(&[&format!("--sysroot={}", bare.0.display()), "-print-sysroot-provenance"]),
5455            ""
5456        );
5457
5458        // And a compile for this machine has no sysroot at all, which is the same empty answer
5459        // `-print-sysroot` gives for it.
5460        let host = Triple::host().expect("a host this compiler knows");
5461        assert_eq!(printed(&[&format!("--target={host}"), "-print-sysroot-provenance"]), "");
5462
5463        // And the other spelling, which section 13.5 is the document that writes.
5464        assert_eq!(printed(&[&sysroot, "--print-sysroot-provenance"]) + "\n", manifest);
5465
5466        // tamnd/rucc#1021. The digest of the same tree is the sha256 of that record, so it is one
5467        // line where the provenance is a few hundred, and it is checkable with `sha256sum` because
5468        // the bytes it is over are the bytes of the file. The number here is that hash of the
5469        // fixture above, computed by `sha256sum` rather than by this compiler.
5470        assert_eq!(
5471            printed(&[&sysroot, "-print-sysroot-digest"]),
5472            "d705ae6ebeafeb7fda4bd57cecc7882bf49784b17015664a09cfae25a1b2000a"
5473        );
5474        assert_eq!(
5475            printed(&[&sysroot, "--print-sysroot-digest"]),
5476            printed(&[&sysroot, "-print-sysroot-digest"])
5477        );
5478
5479        // And the two empty answers are empty here too, because a digest of nothing would read as a
5480        // claim about a sysroot rather than as the absence of one.
5481        assert_eq!(
5482            printed(&[&format!("--sysroot={}", bare.0.display()), "-print-sysroot-digest"]),
5483            ""
5484        );
5485        assert_eq!(printed(&[&format!("--target={host}"), "-print-sysroot-digest"]), "");
5486    }
5487
5488    #[test]
5489    fn a_manifest_this_build_cannot_read_is_refused_rather_than_printed() {
5490        // Passing a file we could not parse to whoever asked would make their parser the one that
5491        // finds the problem, and the three uses section 13.5 gives for this are all somebody else
5492        // parsing it.
5493        let tree = TempTree::new(
5494            "provenance-bad",
5495            &[("manifest", "rucc sysroot manifest 3\ntarget\tx86_64-linux-musl\nlib/libc.a\n")],
5496        );
5497        let message =
5498            refused(&[&format!("--sysroot={}", tree.0.display()), "-print-sysroot-provenance"]);
5499        assert!(message.contains("manifest"), "{message}");
5500        assert!(message.contains("1 fields where an input has six"), "{message}");
5501
5502        // The digest is refused for the same file and for a stronger reason: a hash of bytes this
5503        // build cannot read would be a number that names a record nobody can act on.
5504        let digest =
5505            refused(&[&format!("--sysroot={}", tree.0.display()), "-print-sysroot-digest"]);
5506        assert_eq!(digest, message);
5507    }
5508
5509    #[test]
5510    fn the_two_dependency_flags_that_stop_after_the_rule_stop_after_the_rule() {
5511        let (opts, _) = compile(&["-M", "a.c"]);
5512        assert!(opts.deps.emit && opts.deps.instead_of_compiling);
5513        assert!(opts.deps.system_headers, "plain -M lists them");
5514        assert_eq!(opts.emit, EmitKind::Preprocessed);
5515
5516        // Even where a later flag asked for something else, because the family is a mode and
5517        // the mode is what the run is for.
5518        let (opts, _) = compile(&["-M", "-c", "a.c"]);
5519        assert_eq!(opts.emit, EmitKind::Preprocessed);
5520
5521        let (opts, _) = compile(&["-MM", "a.c"]);
5522        assert!(!opts.deps.system_headers);
5523    }
5524
5525    #[test]
5526    fn the_two_that_end_in_d_leave_the_compilation_alone() {
5527        let (opts, _) = compile(&["-MD", "-c", "a.c"]);
5528        assert!(opts.deps.emit && !opts.deps.instead_of_compiling);
5529        assert!(opts.deps.system_headers);
5530        assert_eq!(opts.emit, EmitKind::Object);
5531
5532        let (opts, _) = compile(&["-MMD", "-c", "a.c"]);
5533        assert!(opts.deps.emit && !opts.deps.instead_of_compiling);
5534        assert!(!opts.deps.system_headers);
5535    }
5536
5537    #[test]
5538    fn nothing_puts_the_system_headers_back_once_a_flag_has_taken_them_out() {
5539        // GCC's rule, and not an oversight in it. The flag asking for fewer of them is read as
5540        // the answer, because the other one never asked the question.
5541        let (opts, _) = compile(&["-MM", "-M", "a.c"]);
5542        assert!(!opts.deps.system_headers);
5543        let (opts, _) = compile(&["-MD", "-MMD", "-c", "a.c"]);
5544        assert!(!opts.deps.system_headers);
5545        let (opts, _) = compile(&["-MMD", "-MD", "-c", "a.c"]);
5546        assert!(!opts.deps.system_headers);
5547    }
5548
5549    #[test]
5550    fn a_target_arrives_escaped_from_one_flag_and_untouched_from_the_other() {
5551        let (opts, _) = compile(&["-MM", "-MT", "a b.o", "-MQ", "a b.o", "a.c"]);
5552        assert_eq!(opts.deps.targets, vec!["a b.o".to_owned(), "a\\ b.o".to_owned()]);
5553    }
5554
5555    #[test]
5556    fn the_rest_of_the_family_is_a_file_and_a_switch() {
5557        let (opts, _) = compile(&["-MM", "-MF", "dep.d", "-MP", "a.c"]);
5558        assert_eq!(opts.deps.file.as_deref(), Some("dep.d"));
5559        assert!(opts.deps.phony);
5560
5561        for flag in ["-MF", "-MT", "-MQ"] {
5562            let e = parse_args(&args(&[flag])).unwrap_err();
5563            assert!(e.message.contains("requires an argument"), "{}", e.message);
5564        }
5565    }
5566
5567    /// A directory of sources for one test, removed when the test is done with it.
5568    struct TempTree(PathBuf);
5569
5570    impl Drop for TempTree {
5571        fn drop(&mut self) {
5572            let _ = std::fs::remove_dir_all(&self.0);
5573        }
5574    }
5575
5576    impl TempTree {
5577        fn new(name: &str, files: &[(&str, &str)]) -> TempTree {
5578            let dir = std::env::temp_dir().join(format!("rucc-deps-{}-{name}", std::process::id()));
5579            let _ = std::fs::remove_dir_all(&dir);
5580            std::fs::create_dir_all(&dir).expect("temporary directory should be writable");
5581            for (path, text) in files {
5582                let at = dir.join(path);
5583                if let Some(parent) = at.parent() {
5584                    std::fs::create_dir_all(parent).expect("creating a subdirectory should work");
5585                }
5586                std::fs::write(&at, text).expect("writing a temporary file should work");
5587            }
5588            TempTree(dir)
5589        }
5590
5591        fn path(&self, name: &str) -> String {
5592            self.0.join(name).to_string_lossy().into_owned()
5593        }
5594    }
5595
5596    #[test]
5597    fn the_rule_names_what_the_includes_found_and_names_each_of_them_once() {
5598        // End to end, because the list comes from the preprocessor and the format comes from
5599        // somewhere else, and a test of either half on its own would pass with the two of them
5600        // wired up backwards.
5601        let tree = TempTree::new(
5602            "found",
5603            &[
5604                ("a.c", "#include \"one.h\"\n#include \"two.h\"\nint main(void) { return X; }\n"),
5605                ("one.h", "#define X 0\n"),
5606                ("two.h", "#include \"one.h\"\n"),
5607            ],
5608        );
5609        let out = tree.path("dep.d");
5610        let code = run(&args(&["-MM", "-MF", &out, "-o", &tree.path("a.i"), &tree.path("a.c")]));
5611        assert_eq!(code, 0);
5612
5613        let text = std::fs::read_to_string(&out).expect("the rule should have been written");
5614        let names: Vec<&str> = text.split_whitespace().collect();
5615        // The target, the source, and each header once however many times it was reached.
5616        assert_eq!(names.first(), Some(&"a.o:"), "{text}");
5617        assert_eq!(names.iter().filter(|n| n.ends_with("one.h")).count(), 1, "{text}");
5618        assert_eq!(names.iter().filter(|n| n.ends_with("two.h")).count(), 1, "{text}");
5619        // And the `-o` went to the file the rule replaced, which is left empty rather than
5620        // absent because a makefile that named it as a target will look for it.
5621        assert_eq!(std::fs::read(tree.path("a.i")).expect("the output should exist"), b"");
5622    }
5623
5624    #[test]
5625    fn a_header_that_is_only_reached_under_a_guard_is_still_a_dependency() {
5626        // The multiple-include optimization means the second reach never opens the file. It is
5627        // still a file this translation unit was built from, so it is still in the rule.
5628        let tree = TempTree::new(
5629            "guarded",
5630            &[
5631                ("a.c", "#include \"g.h\"\n#include \"g.h\"\nint main(void) { return 0; }\n"),
5632                ("g.h", "#ifndef G\n#define G\n#endif\n"),
5633            ],
5634        );
5635        let out = tree.path("dep.d");
5636        let code = run(&args(&["-MM", "-MF", &out, "-o", &tree.path("a.i"), &tree.path("a.c")]));
5637        assert_eq!(code, 0);
5638        let text = std::fs::read_to_string(&out).expect("the rule should have been written");
5639        assert_eq!(text.split_whitespace().filter(|n| n.ends_with("g.h")).count(), 1, "{text}");
5640    }
5641
5642    #[test]
5643    fn every_imacros_file_is_read_before_every_include_file_whatever_order_they_were_written() {
5644        // Measured against GCC rather than read: the two flags the other way round produce the
5645        // same output byte for byte, so the command line order between the two families does not
5646        // decide anything and the order within one does. The `-include` file here can only see
5647        // the definition if the `-imacros` file that was written after it ran first.
5648        let tree = TempTree::new(
5649            "preinclude",
5650            &[
5651                ("a.c", "int main(void) { return 0; }\n"),
5652                ("i.h", "#ifdef FROM_MACROS\nint saw_it;\n#else\nint missed_it;\n#endif\n"),
5653                ("m.h", "#define FROM_MACROS 1\nint macros_text;\n"),
5654            ],
5655        );
5656        let out = tree.path("a.i");
5657        let code = run(&args(&[
5658            "-E",
5659            "-include",
5660            &tree.path("i.h"),
5661            "-imacros",
5662            &tree.path("m.h"),
5663            "-o",
5664            &out,
5665            &tree.path("a.c"),
5666        ]));
5667        assert_eq!(code, 0);
5668        let text = std::fs::read_to_string(&out).expect("the output should have been written");
5669        assert!(text.contains("saw_it"), "{text}");
5670        // And the text of the `-imacros` file is thrown away, which is the whole difference
5671        // between the two flags.
5672        assert!(!text.contains("macros_text"), "{text}");
5673    }
5674
5675    #[test]
5676    fn a_file_the_command_line_named_is_a_prerequisite_the_same_as_one_a_directive_named() {
5677        let tree = TempTree::new(
5678            "preinclude-deps",
5679            &[
5680                ("a.c", "int main(void) { return 0; }\n"),
5681                ("i.h", "int from_include;\n"),
5682                ("m.h", "#define M 1\n"),
5683            ],
5684        );
5685        let out = tree.path("dep.d");
5686        let code = run(&args(&[
5687            "-MM",
5688            "-MF",
5689            &out,
5690            "-include",
5691            &tree.path("i.h"),
5692            "-imacros",
5693            &tree.path("m.h"),
5694            "-o",
5695            &tree.path("a.i"),
5696            &tree.path("a.c"),
5697        ]));
5698        assert_eq!(code, 0);
5699        let text = std::fs::read_to_string(&out).expect("the rule should have been written");
5700        assert!(text.contains("i.h"), "{text}");
5701        assert!(text.contains("m.h"), "{text}");
5702    }
5703
5704    #[test]
5705    fn a_command_line_include_that_is_nowhere_on_the_path_is_an_error_and_not_a_warning() {
5706        // Including the directory of the source file, which is not on the path for these: the
5707        // command line was not written there, so a name in it is relative to where the compiler
5708        // was run rather than to where the source sits.
5709        let tree = TempTree::new(
5710            "preinclude-missing",
5711            &[("sub/a.c", "int main(void) { return 0; }\n"), ("sub/beside.h", "int x;\n")],
5712        );
5713        let code = run(&args(&["-E", "-include", "beside.h", "-o", "-", &tree.path("sub/a.c")]));
5714        assert_eq!(code, 1);
5715    }
5716
5717    #[test]
5718    fn a_command_line_that_links_names_the_executable_and_not_the_object_it_went_through() {
5719        // The object a link goes through is in a temporary directory and is gone before `make`
5720        // reads any of this, so the rule that named it would be a rule for a file that is never
5721        // there. The target and the file are both the `-o`, which is the executable.
5722        let (opts, plan) = compile(&["-MD", "sub/a.c", "-o", "prog"]);
5723        assert_eq!(plan.output.as_deref(), Some("prog"));
5724        assert_eq!(deps::default_target("sub/a.c", deps_target_output(&opts, &plan)), "prog");
5725        assert_eq!(
5726            deps::default_file(&opts.deps, "sub/a.c", plan.output.as_deref()).as_deref(),
5727            Some("prog.d")
5728        );
5729    }
5730
5731    #[test]
5732    fn the_plan_keeps_the_output_name_because_the_rule_is_written_from_it() {
5733        let (_, plan) = compile(&["-MMD", "-c", "sub/a.c", "-o", "obj/x.o"]);
5734        assert_eq!(plan.output.as_deref(), Some("obj/x.o"));
5735        let (_, plan) = compile(&["-MMD", "-c", "sub/a.c"]);
5736        assert_eq!(plan.output, None);
5737    }
5738
5739    #[test]
5740    fn usage_fits_on_a_screen() {
5741        // Not a style preference. A help text that scrolls is one nobody reads, and this is
5742        // the cheapest way to keep it honest as flags accumulate. The number goes up only when
5743        // a family of flags arrives that has nowhere to share a line, which the two pass gates
5744        // were and which the two fuel flags and `-fsafety=` now are, and it goes up by exactly
5745        // the lines that family took. The four it went up by last are the flags a build system
5746        // passes without being asked to: how much to say, what machine to generate for, threads,
5747        // and the questions `configure` asks before it compiles anything. The one it went up by
5748        // last is the second line of `--emit`, whose kinds are a family that has now outgrown
5749        // one line and has nowhere else to go. The two it went up by last are the dependency
5750        // family, which is eight flags that share nothing with anything above them. The one it
5751        // went up by last is the four spellings of position independent code, which every
5752        // configure script writes and which could only have shared the link line, and that line
5753        // is already four characters short of the limit. The two it went up by last are the rest
5754        // of the include family, which is six more flags that change where a header is looked for
5755        // and two that name a header outright. The one it went up by last is the pair that keeps
5756        // the intermediate files and times the steps, which belong next to the two flags above
5757        // them that are also about watching a compilation rather than changing one. The two it
5758        // went up by last are the section flags and the visibility flag, which are what a build
5759        // that cares about the size of what it ships and about which names it exports writes, and
5760        // the second of them was already taken and only missing from here. The one it went up by
5761        // last is the stack protector, which is four spellings of one question and which every
5762        // distribution puts on every command line it issues, so a build that reads this list
5763        // looking for it and does not find it has to go and read the specification instead. The one
5764        // it went up by last is the profiler, which is two spellings of the request and two of
5765        // where the call goes, and which is about watching a program run rather than about what is
5766        // generated, so it shares its subject with nothing above it. The one it went up by last is
5767        // the room a function opens with for something to be written over it later, which takes an
5768        // argument of its own shape and is what a kernel build asks for, so it fits beside the
5769        // profiler and nothing else. The one it went up by last is what overflows rather than being
5770        // undefined, which is three spellings of two questions and which a kernel build and a great
5771        // deal of code written before the standard settled both pass. The one it went up by last is
5772        // the other answer to the first of those questions, which could not share the line because
5773        // what it asks for is the opposite of what the flags on that line ask for. The one it went
5774        // up by last is the split of the line that lists what this compiler does anyway into that
5775        // and what it assumes anyway, which are two different claims that were sharing a line until
5776        // the second of them got a second flag and the line stopped fitting. The one it went up by
5777        // last is the three flags that change the ABI rather than the code, which have to be given
5778        // to every file in a program or none of them and which therefore belong somewhere a person
5779        // reading this list will see them. The one it went up by last is the floating point group,
5780        // which is two lines rather than one because the first of them is a choice this compiler
5781        // records and the rest are claims about what it does anyway, and putting a real setting on
5782        // the same line as three flags that change nothing would be misleading about both. The one
5783        // it went up by last is the flag that says a write has to stay inside the member it names,
5784        // which is a setting rather than a claim and so cannot share the line above it, that being
5785        // the one that picks a tier. The two it went up by last are the prefix mapping family,
5786        // which is four flags whose whole job is to keep a build's output the same from two
5787        // different directories, and which a person chasing a reproducible build comes here
5788        // looking for by name. The one it went up by last is how the debug sections are compressed
5789        // and whether they go in a file of their own, which are two questions about the shape of
5790        // the debug output, where the line above them is about how much of it there is. The one it
5791        // went up by last is the `restrict` contract, which is a setting for the same reason the
5792        // flag that keeps a write inside its member is and which is the check a person who has been
5793        // bitten by a vectorizer comes here looking for. The one it went up by last is link time
5794        // optimization, which is a whole optimization rather than a flag and which says so on its
5795        // own line, because a build that passes it and reads this looking for what it got is
5796        // asking a question no other line here answers. The one it went up by last is the sysroot,
5797        // which is the question somebody asks when a cross build read a file nobody expected, and
5798        // which has no room on the line above it because the answers there are a path each and this
5799        // one is the root all of them are under. The one it went up by last is what is inside that
5800        // root and where each of it came from, which is a question about a whole tree rather than
5801        // about a path and which is long enough on its own that it could not have shared a line with
5802        // anything. The one it went up by last is the profile family, which splits down the middle
5803        // where no other family here does, so the line has to name the half that is taken and the
5804        // half that is refused or it would be read as taking both. The one it went up by last is
5805        // the sanitizers, which are what somebody reaching for a checked build writes first and
5806        // which belong beside the tier that is the nearest thing here to what they asked for. The
5807        // one it went up by last is the digest of that record, which is the same tree as one number
5808        // and could not share the line above it because that line prints a few hundred lines and
5809        // this one prints sixty four characters, and a reader who wants the short answer is looking
5810        // for it by name rather than reading the long one. The one it went up by last is the
5811        // sysroot fetch, which is the only command here that gets something from somewhere else and
5812        // is therefore the one a person wants to have read before they run it rather than after.
5813        // And the flag beside it that forbids every download, which earns its line by being what a
5814        // build in a sealed environment passes and by meaning something even though an ordinary
5815        // compile downloads nothing either way.
5816        assert!(USAGE.lines().count() < 72, "usage text has grown past one screen");
5817    }
5818}