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/0.10.76")]
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;
44mod shapes;
45
46use std::fmt::Write as _;
47use std::io::Write as _;
48use std::path::PathBuf;
49
50use rucc_codegen::coverage::{self, Fired};
51use rucc_codegen::lowering::Lowerings;
52use rucc_codegen::pressure::Pressure;
53use rucc_pp::Dependency;
54use rucc_session::{
55    Compress, Control, Dumps, EmitKind, Hook, Options, Pic, PrefixMap, Preinclude, Protector,
56    SaveTemps, Session, Std, Wrapping, runtime,
57};
58use rucc_sysroot::{Manifest, Sysroot};
59use rucc_target::{ObjectFormat, Triple};
60use rucc_tuple::TargetTuple;
61
62use crate::link::LinkOptions;
63
64pub use crate::assemble::assemble;
65pub use crate::compile::{Artifact, Compiled, Temps, compile, compile_ir};
66pub use crate::phase::{ArchiveJob, Input, InputKind, Job, LinkJob, Output, Phase, Plan, Role};
67pub use crate::preprocess::{OsFileSystem, Preprocessed, preprocess};
68pub use crate::schedule::Jobs;
69
70/// The compiler's version, taken from the workspace manifest.
71pub const VERSION: &str = env!("CARGO_PKG_VERSION");
72
73/// What the command line asked for.
74#[derive(Debug, Clone, PartialEq, Eq)]
75pub enum Action {
76    /// Print usage and exit successfully.
77    Help,
78    /// Print the version and exit successfully.
79    Version,
80    /// Print one line and exit successfully, which is what the `-dump` and `-print` family do.
81    ///
82    /// A build system asks these before it compiles anything, and what it does with the answer
83    /// is paste it into a path or into another command line, so each one is a single line with
84    /// no decoration around it.
85    Print(String),
86    /// Print the resolved configuration and exit successfully.
87    PrintConfig(Box<Options>),
88    /// Print the passes the level will run and exit successfully.
89    PrintPipeline(Box<Options>),
90    /// Print the phase plan and the link line and exit successfully, which is `-###`.
91    PrintPlan {
92        /// The resolved options, which is what says what the link line is for.
93        opts: Box<Options>,
94        /// What to do to each input, and in what order.
95        plan: Box<Plan>,
96        /// What the command line said about linking.
97        link: Box<LinkOptions>,
98    },
99    /// `--fetch <tuple>`, which gets the sysroot this release pins for a target and installs it.
100    ///
101    /// The only action in this compiler that may run another program to move bytes onto the
102    /// machine, which is `spec/cross-compile/13-distribution.md` section 13.8's rule rather than a
103    /// property of how this happens to be written: a compilation has no branch that reaches it.
104    Fetch {
105        /// The artifact, from the table in [`rucc_sysroot::artifact`]. Resolved here rather than where the
106        /// work happens, so that a target nothing is pinned for is a refusal from the parser like
107        /// every other thing a command line can ask for and not have.
108        what: &'static rucc_sysroot::Pinned,
109        /// The target, which names the directory under the cache the tree is installed at and is
110        /// checked against the record inside the artifact.
111        target: TargetTuple,
112        /// Where the cache is, read where everything else that needs it reads it.
113        cache: PathBuf,
114    },
115    /// Compile the given inputs.
116    Compile {
117        /// The resolved options.
118        opts: Box<Options>,
119        /// What to do to each input, and in what order.
120        plan: Box<Plan>,
121        /// What the command line said about linking.
122        link: Box<LinkOptions>,
123        /// How many translation units to compile at once.
124        jobs: Jobs,
125        /// Whether `-v` asked for the plan to be printed while it runs.
126        verbose: bool,
127        /// What is worth saying about the command line before anything is compiled, printed as
128        /// warnings and once for the whole run rather than once per file.
129        ///
130        /// These are not diagnostics. A diagnostic is about a piece of source and has a span to
131        /// point at, and these are about the way two flags were combined, so there is nothing to
132        /// point at and nowhere below the driver that knows both halves. `-w` does not reach them
133        /// for the same reason it does not reach a refusal from the parser.
134        notes: Vec<String>,
135    },
136}
137
138/// Why a command line was rejected.
139#[derive(Debug, Clone, PartialEq, Eq)]
140pub struct CliError {
141    /// The message, lowercase and without a trailing period, in the same shape as any other
142    /// diagnostic.
143    pub message: String,
144}
145
146impl std::fmt::Display for CliError {
147    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
148        f.write_str(&self.message)
149    }
150}
151
152impl std::error::Error for CliError {}
153
154fn err(message: impl Into<String>) -> CliError {
155    CliError { message: message.into() }
156}
157
158/// The two halves of one prefix mapping flag's argument, where `flag` includes its trailing `=`.
159///
160/// The split is at the last `=` in what follows the flag, not the first, which is gcc's rule and
161/// the only one that lets a directory whose name contains an `=` be the old half. It also means
162/// `-fmacro-prefix-map=a=b=c` rewrites `a=b` to `c` rather than `a` to `b=c`, which looks like a
163/// trap until you notice the alternative traps the far more common case.
164fn rewrite<'a>(arg: &'a str, flag: &str) -> Result<(&'a str, &'a str), CliError> {
165    let rest = &arg[flag.len()..];
166    PrefixMap::split(rest).ok_or_else(|| {
167        let flag = flag.trim_end_matches('=');
168        err(format!(
169            "`{rest}` is not a rewrite for `{flag}`, which is an old prefix, an `=` and a new one"
170        ))
171    })
172}
173
174/// A question the command line asked instead of asking for a compilation.
175///
176/// These are answered after the loop rather than where they are read, because every one of them
177/// is about the target or about the library search and the last word on both is the end of the
178/// command line.
179enum Query {
180    /// `-dumpmachine`, the triple.
181    Machine,
182    /// `-dumpversion` and `-dumpfullversion`, which are the same three numbers here.
183    Version,
184    /// `-print-multiarch`, the directory name a distribution files this target under.
185    Multiarch,
186    /// `-print-search-dirs`, in the three lines GCC prints.
187    SearchDirs,
188    /// `-print-sysroot`, the root the headers and the libraries are read under.
189    Sysroot,
190    /// `-print-sysroot-provenance`, what is in that root and where each of it came from.
191    SysrootProvenance,
192    /// `-print-sysroot-digest`, the one number that names all of it.
193    SysrootDigest,
194    /// `-print-file-name=<name>`, the full path of a library file.
195    FileName(String),
196    /// `-print-prog-name=<name>`, the full path of a program.
197    ProgName(String),
198    /// `-print-libgcc-file-name`, which is `-print-file-name=libgcc.a` under another spelling.
199    Libgcc,
200}
201
202/// Usage text.
203///
204/// Deliberately short. `spec/04-driver-and-cli.md` puts the full flag reference in the
205/// manual page, because a `--help` nobody can read in one screen is a `--help` nobody reads.
206pub const USAGE: &str = "\
207rucc, an optimizing C compiler
208
209usage: rucc [options] file...
210
211options:
212  -c                     compile and assemble, do not link
213  -S                     compile only, emit assembly
214  -E                     preprocess only
215  -o <file>              write output to <file>, or to standard output for -
216  -D <name>[=<value>], -U <name>      define a macro, or undefine one after every -D
217  -I <dir>               add <dir> to the include search path
218  -iquote -isystem -idirafter <dir>   the other chains, -nostdinc drops ours
219  -I-, -iprefix <p>, -iwithprefix[before] <dir>   the older spellings of those
220  -include <file>, -imacros <file>    read <file> first, the second for its macros only
221  --sysroot=<dir>        look for the library's headers under <dir>, -isysroot too
222  -P, -dM                with -E: leave out the markers, or dump the macros
223  -M -MM -MD -MMD        write a make rule for the source, the last two compile as well
224  -MF <file> -MT <t> -MQ <t> -MP   where the rule goes, what it builds, targets with no recipe
225  -std=<dialect>         c89 through c23, and the gnu spellings
226  -fgnuc-version=<v>     the GCC release to claim, default 16.0.0
227  -x <lang>              treat later inputs as <lang>, or none to stop
228  -O<level>              optimize: 0, 1, 2, 3, s, z
229  -fsafety=<tier>        check memory safety: off, detect, enforce, kernel
230  -f[no-]sanitize=<what>   the negative is taken, the positive is refused by name
231  -f[no-]safety-subobject   a write has to stay inside the member it names
232  -f[no-]safety-restrict    two restrict pointers of one block may not meet
233  -f<pass> -fno-<pass> -fdump-ir=<what> -fopt-info[-<kind>][=FILE]
234  -fpass-fuel=<pass>=<n>, -fpass-fuel-global=<n>   stop a pass, or all of them, after n
235  -fdisable-<pass>[=<funcs>], -fenable-<pass>[=<funcs>]   run a pass on some functions only
236  -g -g0 -gdwarf-5, -fno-omit-frame-pointer, -mno-red-zone   debug info, frame pointer, red zone
237  -gz[=none|zlib|zlib-gnu|zstd] -gno-split-dwarf   compress debug sections, one file not two
238  -flto[=auto|jobserver|<n>] -fno-lto -ffat-lto-objects   read, and not done yet
239  -fprofile-use[=<path>] -fprofile-dir=<dir>   read too, where -fprofile-generate is refused
240  -f[no-]stack-protector[-strong|-all], -f[no-]stack-clash-protection, -fcf-protection=<edges>
241  -ffunction-sections -fdata-sections   a section per function or variable, for --gc-sections
242  -fvisibility=<what>    default, hidden, internal or protected, when nothing in the source said
243  -l<name>, -L <dir>, -B <dir>   link a library, where to look for one, where our own tools are
244  -fPIC -fpic -fPIE -fpie, -fno-common, -pipe   what it does anyway
245  -f[no-]strict-aliasing, -f[no-]delete-null-pointer-checks   what it assumes anyway
246  -static -shared -pie -no-pie -nostdlib -nostartfiles -nodefaultlibs -rdynamic -s   how to link
247  -Wl,<arg>, -Xlinker <arg>, -fuse-ld=<name>   hand an argument to the linker, or pick one
248  -Werror -pedantic -pedantic-errors -w -W[no-]system-headers   how much to say, and how fatal
249  -m64 -march= -mtune= -mcpu= -mabi= -mcmodel=   what machine to generate for
250  -pg -p, -mfentry -mno-fentry   call a profiler on the way in, and where that call goes
251  -fpatchable-function-entry=<n>[,<m>]   room at the top of every function to patch later
252  -fwrapv, -fwrapv-pointer, -fno-strict-overflow   signed or pointer overflow wraps
253  -ftrapv                signed overflow stops the program instead
254  -f[no-]signed-char, -f[no-]unsigned-char, -f[no-]short-enums   change the ABI
255  -ffp-contract=<how>    fuse a multiply and an addition: fast, on or off
256  -fexcess-precision=<how>, -f[no-]rounding-math, -f[no-]trapping-math   what may be folded
257  -ffile-prefix-map=<old>=<new>   rewrite that front of every path we put in the output
258  -fmacro-prefix-map= -fdebug-prefix-map= -fprofile-prefix-map=   the same, one output each
259  -pthread               build for more than one thread, and link the library for it
260  -dumpmachine -dumpversion -print-multiarch -print-search-dirs   what this compiler is
261  -print-file-name=<name> -print-prog-name=<name>   where a file or a program is
262  -print-sysroot         the root the headers and the libraries are read under
263  -print-sysroot-provenance   every input under it, where it came from and its licence
264  -print-sysroot-digest   the sha256 of that record, which names the whole sysroot in one line
265  --fetch <tuple>        get the sysroot this release pins for <tuple> and install it in the cache
266  --offline              never download anything, which a compilation never does anyway
267  -j[n]                  compile n translation units at once, default all
268  -v, -###               print each phase as it runs, or without running any
269  -save-temps[=cwd|obj], -time   keep the .i and the .s, say how long each step took
270  --target=<triple>      generate code for <triple>
271  --emit=<kind>          exe, obj, archive, asm, preprocessed, tast, ir, mir-final,
272                         safety-summary, type-granules
273  --print-config, --print-pipeline    print the configuration or the pipeline, and exit
274  --version              print the version and exit
275  -h, --help             print this message and exit
276
277See spec/04-driver-and-cli.md for the full flag reference.
278";
279
280/// The argument of a flag that may be joined to it or may be the next word.
281///
282/// `-DFOO` and `-D FOO` are the same thing, and `at` is where the flag's own letters end.
283fn joined_or_next(
284    arg: &str,
285    at: usize,
286    args: &[String],
287    i: &mut usize,
288) -> Result<String, CliError> {
289    if arg.len() > at {
290        return Ok(arg[at..].to_owned());
291    }
292    let next = args.get(*i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
293    *i += 1;
294    Ok(next.clone())
295}
296
297/// The smallest boundary a function is put on when the command line asked for no alignment at all.
298///
299/// Eight bytes, which is what gcc 16 gives `-fno-align-functions` on x86-64 and is a boundary every
300/// target this compiler has is happy with. It is not zero: a function still has to start somewhere
301/// an instruction may start, and the flag asks for the target's minimum rather than for none.
302const MIN_FUNC_ALIGN: u32 = 8;
303
304/// What `-falign-functions=N` asks for, as a power of two, or `None` for the target's own answer.
305///
306/// Zero and one both mean the default, which is gcc's reading of them, and everything else is
307/// rounded up to the next power of two, which is also gcc's: `-falign-functions=3` puts a function
308/// on a four byte boundary rather than being refused. Gives back `Err` shaped as an outer `None`
309/// only when the text is not a number, since that is the one thing gcc will not read either. A
310/// number larger than any alignment makes sense at is clamped rather than refused, for the same
311/// reason: this is a preference about speed and a build that wrote a silly one still deserves to
312/// compile.
313fn function_alignment(text: &str) -> Option<Option<u32>> {
314    // gcc takes `N:M:N2:M2`, where everything after the first number is about how far it is willing
315    // to go to reach the boundary. Only the boundary is answerable here, so the rest is read to
316    // check that it is numbers and then dropped.
317    let mut parts = text.split(':');
318    let first = parts.next()?;
319    if parts.any(|part| part.parse::<u64>().is_err()) {
320        return None;
321    }
322    let want: u64 = first.parse().ok()?;
323    if want <= 1 {
324        return Some(None);
325    }
326    let bytes = want.min(1 << 16).next_power_of_two();
327    Some(Some(u32::try_from(bytes).ok()?))
328}
329
330/// Every name that may follow `-fsanitize=`, which is gcc 16's list and three of this compiler's
331/// own.
332///
333/// The three are on it because `spec/07-types-and-semantics.md` section 7.7 already promises them:
334/// each undefined behaviour this compiler exploits is listed there with the check that detects it,
335/// and `alias`, `restrict` and `memory` are checks gcc has no spelling for. gcc refuses `memory`
336/// outright, since the sanitizer of that name is clang's. A name being here means it is a name
337/// rather than a typo, and nothing more than that: every one of them is refused after the loop,
338/// because none of them is implemented.
339///
340/// `all` is deliberately absent. gcc takes it only in the negative, so it is handled where each of
341/// those two spellings is read rather than by being on this list.
342const SANITIZERS: [&str; 34] = [
343    "address",
344    "kernel-address",
345    "hwaddress",
346    "kernel-hwaddress",
347    "pointer-compare",
348    "pointer-subtract",
349    "thread",
350    "leak",
351    "undefined",
352    "shift",
353    "shift-base",
354    "shift-exponent",
355    "integer-divide-by-zero",
356    "unreachable",
357    "vla-bound",
358    "null",
359    "return",
360    "signed-integer-overflow",
361    "bounds",
362    "bounds-strict",
363    "alignment",
364    "object-size",
365    "float-divide-by-zero",
366    "float-cast-overflow",
367    "nonnull-attribute",
368    "returns-nonnull-attribute",
369    "bool",
370    "enum",
371    "vptr",
372    "pointer-overflow",
373    "builtin",
374    "alias",
375    "restrict",
376    "memory",
377];
378
379/// Parses a command line, without the program name.
380///
381/// # Errors
382///
383/// Returns the message to print when the arguments do not name a compilation this compiler
384/// can attempt.
385pub fn parse_args(args: &[String]) -> Result<Action, CliError> {
386    let host = Triple::host()
387        .ok_or_else(|| err("this host is not a supported target and no --target was given"))?;
388    let mut opts = Options::new(host);
389    // Where the compiler is running, which is what `DW_AT_comp_dir` is and what a debugger joins a
390    // relative file name onto. Asked here rather than where the debug sections are written, because
391    // this is the one layer that is allowed to look at the process it is in, and because a command
392    // line that compiles four files should give the same answer for all four.
393    opts.working_dir = std::env::current_dir().ok().map(|dir| dir.to_string_lossy().into_owned());
394    let mut inputs: Vec<Input> = Vec::new();
395    let mut print_config = false;
396    let mut print_pipeline = false;
397    let mut print_plan = false;
398    let mut verbose = false;
399    let mut jobs = Jobs::default();
400    let mut nostdinc = false;
401    let mut sysroot: Option<PathBuf> = None;
402    // What the command line is worth warning about, filled in after the loop rather than during it,
403    // because every question of this kind is about two flags and the last word on both of them is
404    // the end of the loop.
405    let mut notes: Vec<String> = Vec::new();
406    // The whole ten field target, kept beside the three field one because `--target=` can pin a
407    // libc version and `Triple` has nowhere to put it. It decides `__GLIBC_MINOR__` and nothing
408    // else today, and `None` is a command line that named no target, which is this machine.
409    let mut pinned: Option<TargetTuple> = None;
410    let mut output = None;
411    let mut link = LinkOptions::default();
412    let mut query: Option<Query> = None;
413    // What `--fetch` named, and whether `--offline` forbade it. Both are weighed after the loop
414    // because either can be written after the other.
415    let mut fetch: Option<String> = None;
416    let mut offline = false;
417    let mut threads = false;
418    // Which sanitizers are still asked for by the end of the command line. Accumulated across the
419    // loop rather than answered where it was read, because `-fno-sanitize=` turns one off and a
420    // build that asks for a check and then takes it back has asked for nothing. What happens to a
421    // set that is not empty is decided after the loop.
422    let mut sanitizers: Vec<&str> = Vec::new();
423    // `-x` applies to inputs that come after it and stays in effect until the next one, which
424    // is why it is tracked across the loop rather than attached to a single argument.
425    let mut forced: Option<InputKind> = None;
426    // What `-iprefix` last said, stuck on the front of every later `-iwithprefix`. It applies to
427    // the flags after it and not the ones before, so a command line may set it more than once.
428    // GCC's default is its own installed header directory with the last component taken off,
429    // which is a path a cross compiler's build system knows and passes; there is no equivalent
430    // here, so with no `-iprefix` the prefix is nothing and `-iwithprefix` names a directory
431    // outright.
432    let mut iprefix = String::new();
433
434    let mut i = 0;
435    while i < args.len() {
436        let arg = args[i].as_str();
437        i += 1;
438        match arg {
439            "-h" | "--help" => return Ok(Action::Help),
440            "--version" => return Ok(Action::Version),
441            // The sysroot fetch, which is weighed after the loop rather than acted on here, because
442            // `--offline` written after it has to be able to forbid it. Both spellings, since a
443            // flag that takes a tuple gets written both ways and neither is a guess at what the
444            // other meant.
445            "--fetch" => {
446                let value = args
447                    .get(i)
448                    .ok_or_else(|| err("--fetch requires the target to get a sysroot for"))?;
449                i += 1;
450                fetch = Some(value.clone());
451            }
452            _ if arg.starts_with("--fetch=") => {
453                fetch = Some(arg["--fetch=".len()..].to_owned());
454            }
455            // Accepted on any command line and only ever read by the fetch, because an ordinary
456            // compile downloads nothing with or without it. So this flag takes nothing away today,
457            // which is the property section 13.2 asks for rather than an omission: a build that
458            // passes it is saying what it expects of this compiler, and what it expects is already
459            // true.
460            "--offline" => offline = true,
461            "--print-config" => print_config = true,
462            "--print-pipeline" => print_pipeline = true,
463            "-###" => print_plan = true,
464            "-v" => verbose = true,
465            // The files a compilation goes through, kept rather than thrown away. The bare
466            // spelling means `=obj` and not `=cwd`, which is not what the manual says and is what
467            // gcc 16 does; `SaveTemps::Object` carries the measurement.
468            "-save-temps" => opts.save_temps = SaveTemps::Object,
469            _ if arg.starts_with("-save-temps=") => {
470                opts.save_temps = arg["-save-temps=".len()..].parse().map_err(err)?;
471            }
472            // How long each step took. A misspelling of this is worth rejecting rather than
473            // ignoring, since a run that says nothing looks like a compilation that took no time.
474            "-time" => opts.time = true,
475            "-c" => opts.emit = EmitKind::Object,
476            "-S" => opts.emit = EmitKind::Asm,
477            "-E" => opts.emit = EmitKind::Preprocessed,
478            "-g" => opts.debug_info = true,
479            // GCC's own levels of how much debug information to write. Zero is none and every
480            // other number is some, and this compiler has one amount, so the numbers above zero
481            // all mean the same thing here. `-ggdb` is the same flag asking for whatever the
482            // debugger on the machine prefers, which is what we emit anyway.
483            "-g0" => opts.debug_info = false,
484            "-g1" | "-g2" | "-g3" | "-ggdb" | "-ggdb1" | "-ggdb2" | "-ggdb3" => {
485                opts.debug_info = true;
486            }
487            // The version of DWARF to write. We write DWARF 5 and nothing else, so a build that
488            // asks for another version is told rather than handed a file it cannot read.
489            "-gdwarf" | "-gdwarf-5" => opts.debug_info = true,
490            _ if arg.starts_with("-gdwarf-") => {
491                return Err(err(format!(
492                    "{arg}: this compiler writes DWARF 5 and no other version, see \
493                     spec/11-debug-info.md"
494                )));
495            }
496            // Whether the debug information goes in a file of its own beside the object. gcc
497            // writes that `.dwo` whether or not it found anything to put in it, which means a
498            // build system that declares the file as an output gets one and a make rule that
499            // depends on it fires. Refused for that reason rather than taken: section 4.1 takes a
500            // flag that changes nothing and refuses one that changes what is produced, and a file
501            // that does not appear is the plainest change of that kind there is. The negative
502            // spelling is taken, because putting it all in the object is what happens anyway.
503            "-gno-split-dwarf" => {}
504            "-gsplit-dwarf" => {
505                return Err(err(format!(
506                    "{arg}: this compiler writes no separate `.dwo` file, and a build that \
507                     expects one beside each object would wait for a file that never arrives, \
508                     see spec/11-debug-info.md"
509                )));
510            }
511            // How the debug sections are compressed. There are none yet, so every answer produces
512            // the same bytes and taking the flag promises nothing that is not kept. The value is
513            // still checked, because a typo in a distribution's flags is worth finding when the
514            // compiler reads it rather than when somebody later wonders why nothing got smaller.
515            // Bare `-gz` means `zlib`, which the manual leaves for the reader to discover.
516            "-gz" => opts.compress = Compress::Zlib,
517            _ if arg.starts_with("-gz=") => {
518                let how = &arg["-gz=".len()..];
519                opts.compress = how.parse().map_err(|()| {
520                    err(format!(
521                        "`{how}` is not a way to compress debug sections, which is none, zlib, \
522                         zlib-gnu or zstd"
523                    ))
524                })?;
525            }
526            "-Werror" => opts.warnings_are_errors = true,
527            // Nothing that is not fatal is said at all. Read at the one place a diagnostic goes
528            // through rather than here, so that a warning `-w` dropped is not counted either.
529            "-w" => opts.warnings = false,
530            // Off by default, the way gcc has it off. A header that came with the machine is not
531            // one the person compiling can change, so a warning about it is noise, and under
532            // `-Werror` it is a build that stops on a line nobody in the project wrote. Somebody
533            // porting a header does want to hear all of it, which is what the flag is for.
534            "-Wsystem-headers" => opts.system_header_warnings = true,
535            "-Wno-system-headers" => opts.system_header_warnings = false,
536            "-pedantic-errors" => {
537                opts.pedantic = true;
538                opts.warnings_are_errors = true;
539            }
540            "-P" => opts.line_markers = false,
541            // The dependency family, which section 4.4 calls required because every build system
542            // that generates its own makefiles asks for it. The two that end in `D` write a file
543            // beside the object and let the compilation happen, and the two that do not write to
544            // standard output and stop after it. Nothing here turns the system headers back on
545            // once a flag has turned them off, which is GCC's behaviour and is why `-MM -M` is
546            // `-MM`: the flag asking for fewer of them is the one with something to say.
547            "-M" => {
548                opts.deps.emit = true;
549                opts.deps.instead_of_compiling = true;
550            }
551            "-MM" => {
552                opts.deps.emit = true;
553                opts.deps.instead_of_compiling = true;
554                opts.deps.system_headers = false;
555            }
556            "-MD" => opts.deps.emit = true,
557            "-MMD" => {
558                opts.deps.emit = true;
559                opts.deps.system_headers = false;
560            }
561            "-MP" => opts.deps.phony = true,
562            // These three take a word and only in the separated form, which is how GCC spells
563            // them and how every build system writes them.
564            "-MF" | "-MT" | "-MQ" => {
565                let value =
566                    args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
567                i += 1;
568                match arg {
569                    "-MF" => opts.deps.file = Some(value.clone()),
570                    // The whole of the difference between the two. `-MT` is for a build that has
571                    // already escaped what it is passing, and `-MQ` is for one that has a name
572                    // and wants it to arrive as that name.
573                    "-MT" => opts.deps.targets.push(value.clone()),
574                    _ => opts.deps.targets.push(deps::escaped(value)),
575                }
576            }
577            // The questions a build system asks before it compiles anything. Answered after the
578            // loop, because each one is about the target or the library search and the command
579            // line has not finished saying what those are.
580            "-dumpmachine" => query = Some(Query::Machine),
581            "-dumpversion" | "-dumpfullversion" => query = Some(Query::Version),
582            "-print-multiarch" => query = Some(Query::Multiarch),
583            "-print-search-dirs" => query = Some(Query::SearchDirs),
584            "-print-sysroot" => query = Some(Query::Sysroot),
585            // Both spellings, because this one is ours rather than GCC's and our own documents
586            // write it both ways: section 13.5 of `spec/cross-compile/13-distribution.md` gives it
587            // two dashes like the other flags we invented, and document 12's table gives it one
588            // like the `-print-` family it sits in. A person who reads either and types what it
589            // says is right, so neither is refused.
590            "-print-sysroot-provenance" | "--print-sysroot-provenance" => {
591                query = Some(Query::SysrootProvenance);
592            }
593            "-print-sysroot-digest" | "--print-sysroot-digest" => {
594                query = Some(Query::SysrootDigest);
595            }
596            "-print-libgcc-file-name" => query = Some(Query::Libgcc),
597            _ if arg.starts_with("-print-file-name=") => {
598                query = Some(Query::FileName(arg["-print-file-name=".len()..].to_owned()));
599            }
600            _ if arg.starts_with("-print-prog-name=") => {
601                query = Some(Query::ProgName(arg["-print-prog-name=".len()..].to_owned()));
602            }
603            // A program built to run in more than one thread. On every platform this compiler
604            // targets that is a macro the library's headers read and one more library on the
605            // link line, and the library is added after the loop so that it lands after the
606            // objects that refer to it.
607            "-pthread" | "-pthreads" => {
608                opts.defines.push("_REENTRANT".to_owned());
609                threads = true;
610            }
611            "-ansi" => {
612                opts.std = Std::C89;
613                opts.gnu_extensions = false;
614            }
615            // `-Wpedantic` is the same flag under the name the `-W` family gives it, which is
616            // the spelling a build system that groups its warning flags tends to write.
617            "-pedantic" | "-Wpedantic" => opts.pedantic = true,
618            // Both directions, because a build that needs this for one directory turns it back
619            // off for the next one rather than leaving it on for the whole tree.
620            "-fpermissive" => opts.permissive = true,
621            "-fno-permissive" => opts.permissive = false,
622            "-ffreestanding" => opts.hosted = false,
623            "-fhosted" => opts.hosted = true,
624            "-fno-builtin" => opts.builtins = false,
625            "-fbuiltin" => opts.builtins = true,
626            // The C89 dialects are under GNU's reading whatever this says, so turning it off
627            // there is turning off something the dialect asked for, which is accepted and does
628            // nothing. gcc refuses that command line, and there is nothing it could have meant.
629            "-fgnu89-inline" => opts.gnu89_inline = true,
630            "-fno-gnu89-inline" => opts.gnu89_inline = false,
631            // Both directions of each, because a build system that wants one of these usually
632            // writes it beside the flag that turns it back off for one directory.
633            "-fno-omit-frame-pointer" => opts.frame_pointer = true,
634            "-fomit-frame-pointer" => opts.frame_pointer = false,
635            // Both directions again, for the same reason, and a third answer for a command line
636            // that wrote neither: see `reorder_blocks` in `rucc_session`.
637            "-freorder-blocks" => opts.reorder_blocks = Some(true),
638            "-fno-reorder-blocks" => opts.reorder_blocks = Some(false),
639            // gcc's name for the scheduler that runs after the registers are handed out, which is
640            // the only one rucc has: see `schedule_insns` in `rucc_session`. gcc also takes
641            // `-fschedule-insns` for the pass before allocation, and taking that one here would be
642            // a flag that says a pass ran when none did.
643            "-fschedule-insns2" => opts.schedule_insns = Some(true),
644            "-fno-schedule-insns2" => opts.schedule_insns = Some(false),
645            "-mno-red-zone" => opts.red_zone = false,
646            "-mred-zone" => opts.red_zone = true,
647            // Four flags rather than one with an argument, which is how gcc spells them and how
648            // every build line writes them. Last one wins, because a package build puts
649            // `-fstack-protector-strong` in its global flags and a directory that cannot have one
650            // turns it back off on the line after.
651            "-fno-stack-protector" | "-fno-stack-protector-all" | "-fno-stack-protector-strong" => {
652                opts.protector = Protector::None;
653            }
654            "-fstack-protector" => opts.protector = Protector::Buffers,
655            "-fstack-protector-strong" => opts.protector = Protector::Strong,
656            "-fstack-protector-all" => opts.protector = Protector::All,
657            // The other half of what a hardened build asks for, and it is a question about the
658            // frame rather than about the function, so it is a switch rather than a level.
659            "-fstack-clash-protection" => opts.stack_clash = true,
660            "-fno-stack-clash-protection" => opts.stack_clash = false,
661            // The third of them, and the one that is a question with an argument rather than a
662            // family of spellings, because what it asks about is which of the two edges of a
663            // control flow transfer is checked. Bare is both of them, which is what gcc does.
664            "-fcf-protection" => opts.control = Control::Full,
665            "-fno-cf-protection" => opts.control = Control::None,
666            // Two spellings of the same request, which is what gcc has as well. `-p` was the older
667            // profiler and `-pg` the one that also recorded who called whom, and on every platform
668            // this compiler targets there is now one hook and both ask for it.
669            "-pg" | "-p" => {
670                opts.profile = true;
671                link.profile = true;
672            }
673            // Accepted on their own and doing nothing on their own, which is gcc's behaviour: they
674            // say where the call goes and a command line that asked for no call has nowhere to put
675            // one. That matters because a build system that sets `-mfentry` globally and `-pg` per
676            // directory is a build system that would otherwise fail on every other directory.
677            "-mfentry" => opts.hook = Hook::Early,
678            "-mno-fentry" => opts.hook = Hook::Late,
679            // GCC drops its own include directory along with the system ones, because its
680            // headers are half of a pair with the library's and half a pair is worse than
681            // none. A build that passes this is supplying the whole set itself.
682            "-nostdinc" => nostdinc = true,
683            "-o" => {
684                output = Some(args.get(i).ok_or_else(|| err("-o requires an argument"))?.clone());
685                i += 1;
686            }
687            // The flags that take a directory only in the separated form. GCC spells them
688            // this way and nothing writes `-iquotedir`, so accepting the joined form would
689            // mean guessing at a path that starts with the flag's own letters.
690            // Apple's spelling of `--sysroot`, and the one its own build systems pass. The
691            // two mean the same thing here: the configured directories are under there rather
692            // than under the root.
693            "-isysroot" => {
694                let dir = args.get(i).ok_or_else(|| err("-isysroot requires an argument"))?;
695                i += 1;
696                sysroot = Some(PathBuf::from(dir));
697            }
698            "-iquote" | "-isystem" | "-idirafter" => {
699                let dir = args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
700                i += 1;
701                match arg {
702                    "-iquote" => opts.search.push_quote(dir.clone()),
703                    "-isystem" => opts.search.push_system(dir.clone()),
704                    _ => opts.search.push_after(dir.clone()),
705                }
706            }
707            "-iprefix" => {
708                iprefix = args.get(i).ok_or_else(|| err("-iprefix requires an argument"))?.clone();
709                i += 1;
710            }
711            // Where GCC puts these is not where its manual says it puts them, and this is the
712            // measured answer rather than the documented one: `-iwithprefix` lands in the
713            // `-isystem` slot and not the `-idirafter` slot, and `-iwithprefixbefore` lands in
714            // the `-I` slot. A cross build that uses them is relying on the behaviour, since
715            // that is the compiler it was developed against.
716            "-iwithprefix" | "-iwithprefixbefore" => {
717                let dir = args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
718                i += 1;
719                let dir = format!("{iprefix}{dir}");
720                if arg == "-iwithprefix" {
721                    opts.search.push_system(dir);
722                } else {
723                    opts.search.push_bracket(dir);
724                }
725            }
726            "-include" | "-imacros" => {
727                let name = args.get(i).ok_or_else(|| err(format!("{arg} requires an argument")))?;
728                i += 1;
729                opts.preincludes
730                    .push(Preinclude { name: name.clone(), macros_only: arg == "-imacros" });
731            }
732            // The flag `-iquote` was introduced to replace, still passed by build systems old
733            // enough to predate the replacement. It is not a directory: it says that every `-I`
734            // so far is for quoted includes only, and that a quoted include stops looking next
735            // to the file that wrote it.
736            "-I-" => opts.search.split_quote_chain(),
737            "-x" => {
738                let lang = args.get(i).ok_or_else(|| err("-x requires an argument"))?;
739                i += 1;
740                forced = if lang == "none" {
741                    None
742                } else {
743                    Some(InputKind::from_x_arg(lang).map_err(|e| err(format!("{e}")))?)
744                };
745            }
746            // Not a GCC flag. spec/03-architecture.md section 3.5 compiles several
747            // translation units in one process rather than making the build system fork, and
748            // section 3.8's determinism check compares `-j1` against `-j16`, so the knob has
749            // to exist and has to be spelled the way `make` spells it.
750            // `-DFOO`, `-D FOO` and the same for `-U` and `-I`. Both forms are in wide use
751            // and a build system may produce either, so both are read here rather than
752            // being normalised by whatever generated the command line.
753            _ if arg.starts_with("-D") => {
754                let value = joined_or_next(arg, 2, args, &mut i)?;
755                opts.defines.push(value);
756            }
757            _ if arg.starts_with("-U") => {
758                let value = joined_or_next(arg, 2, args, &mut i)?;
759                opts.undefines.push(value);
760            }
761            _ if arg.starts_with("-I") => {
762                let dir = joined_or_next(arg, 2, args, &mut i)?;
763                opts.search.push_bracket(dir);
764            }
765            _ if arg.starts_with("-std=") => {
766                let name = &arg["-std=".len()..];
767                let (std, gnu) = Std::from_flag(name)
768                    .ok_or_else(|| err(format!("unknown dialect `{name}`, see --help")))?;
769                opts.std = std;
770                opts.gnu_extensions = gnu;
771            }
772            // Section 4.5. The claim decides which half of glibc's `sys/cdefs.h` we are
773            // handed, so a differential run that does not set it is comparing two compilers
774            // that believe they are different compilers.
775            // GCC packs these into one flag, so `-dDI` is two of them. Letters in the family
776            // that we have not written yet are accepted and ignored, because a dump is a
777            // debugging aid and a build that asks for one should still compile. A letter
778            // outside the family falls through to the unknown option error, which is what
779            // keeps `-dumpversion` from being read as a dump of nothing.
780            _ if Dumps::is_family(arg) => {
781                opts.dumps.add(&arg[2..]);
782            }
783            // One name at a time, which is what a build that means its own `memcpy` and the
784            // library's everything else writes. The name is not checked against a list, because
785            // the flag is about what the program means by a name and a program is allowed to mean
786            // something by a name this compiler has never heard of.
787            _ if arg.starts_with("-fno-builtin-") => {
788                opts.no_builtin.push(arg["-fno-builtin-".len()..].to_owned());
789            }
790            _ if arg.starts_with("-fgnuc-version=") => {
791                let v = &arg["-fgnuc-version=".len()..];
792                opts.gnuc = v.parse().map_err(err)?;
793            }
794            // spec/13-gnu-compat.md section 13.3 promises this flag an error that says why rather
795            // than the unknown option one, because a build reaching for it is asking for a feature
796            // and deserves to be told it is not coming rather than told the spelling is wrong.
797            // The negative form is what this compiler does anyway, so it is taken and dropped.
798            "-fnested-functions" => {
799                return Err(err(
800                    "nested functions are not supported: a call to one goes through a trampoline \
801                     written on the stack, which no target that enforces an unexecutable stack \
802                     allows",
803                ));
804            }
805            "-fno-nested-functions" => {}
806            // Which of the two links the output is for, which is a real difference and not a
807            // description of what happens anyway. Everything here is position independent either
808            // way, and what these decide is whether a name may be one another object defines or
809            // replaces, because a link that produces an executable puts every name in the same
810            // program and a link that produces a shared library does not.
811            //
812            // It matters that they are accepted at all, whatever they then do. Every autoconf and
813            // cmake build puts `-fPIC` on the compile line, so a compiler that rejects it cannot
814            // be the `CC` of a project that has a configure script, whatever else it can do. That
815            // is how this was found: building SQLite's test fixture stopped on it.
816            "-fPIC" | "-fpic" => opts.pic = Pic::Library,
817            // Not a synonym of the pair above, which is what they were treated as until #756. The
818            // library is the expensive answer and gcc makes it the one that has to be asked for,
819            // so this is also what nothing at all means.
820            "-fPIE" | "-fpie" => opts.pic = Pic::Executable,
821            // A different question from the pair above, and the one every distribution build of a
822            // shared library answers. `-fPIC` decides how an address is reached, and this decides
823            // whether the optimizer may believe a body it can see, because an exported name is one
824            // the dynamic linker may find another definition of first. On by default, which is
825            // gcc's arrangement and is the honest answer, and off is a promise the build makes and
826            // nothing checks.
827            "-fsemantic-interposition" => opts.interposition = true,
828            "-fno-semantic-interposition" => opts.interposition = false,
829            // Two requests rather than one, and the same table answers both, so what decides is
830            // whether either of them is standing. gcc arranges it the same way: the asynchronous
831            // one is the default here and it implies the other, and a line that asks for a table
832            // and against an asynchronous one gets a table.
833            "-fasynchronous-unwind-tables" => opts.async_unwind_tables = true,
834            "-fno-asynchronous-unwind-tables" => opts.async_unwind_tables = false,
835            "-funwind-tables" => opts.unwind_tables = true,
836            "-fno-unwind-tables" => opts.unwind_tables = false,
837            // The other direction is a request, not a description, and it is one this compiler
838            // cannot grant, so it gets the treatment section 13.3 asks for rather than the unknown
839            // option error. Answering it by carrying on would be answering a different question:
840            // the code would still be position independent, which is correct everywhere an
841            // ordinary program runs and is wrong in a kernel, where the flag is written precisely
842            // because there is no loader to fill a global offset table in.
843            "-fno-pic" | "-fno-pie" => {
844                return Err(err(
845                    "position dependent code is not supported: an address that may be in another \
846                     object is loaded out of the global offset table, and nothing here emits the \
847                     absolute form this asks for. Use -no-pie if what you meant was how to link",
848                ));
849            }
850            // A section per function and a section per variable, which is what makes
851            // `--gc-sections` able to drop anything: a linker can leave out a section nothing
852            // reaches and cannot leave out half of one. Both directions are taken, and the off
853            // one is the default rather than a refusal, since a build that writes it is asking
854            // for what happens anyway.
855            "-ffunction-sections" => opts.function_sections = true,
856            "-fno-function-sections" => opts.function_sections = false,
857            "-fdata-sections" => opts.data_sections = true,
858            "-fno-data-sections" => opts.data_sections = false,
859            // Another description of what this compiler does. A file scope declaration with no
860            // initializer is written into `.bss` as an ordinary defined symbol, not offered to the
861            // linker as a common one for it to merge, which is what `-fno-common` asks for and what
862            // gcc has done by default since 10. Nothing in the front end produces `Linkage::Common`
863            // at all.
864            "-fno-common" => {}
865            // What overflows rather than being undefined. Every one of these takes something away
866            // from the optimizer rather than asking it to do anything, which is why the negative
867            // spellings are the interesting ones and the positive spellings are the default.
868            //
869            // `-fno-strict-overflow` is both of the others, which is gcc's own reading of it: its
870            // help text for `-fstrict-overflow` says "negated as -fwrapv -fwrapv-pointer". So it is
871            // written here as the pair rather than kept as a third thing to test everywhere.
872            //
873            // `-ftrapv` is the exception and is the one that asks for something. It is the other
874            // answer to the question `-fwrapv` answers, so the two cannot both hold and each clears
875            // the other, which makes the last one on the command line the one that counts. That is
876            // gcc 16's behaviour and was measured rather than read: `-ftrapv -fwrapv` emits no
877            // checked calls and `-fwrapv -ftrapv` emits them. The positive spelling of the pointer
878            // question is left alone by both, because neither has anything to say about it.
879            "-fwrapv" => {
880                opts.wrapping.signed = true;
881                opts.wrapping.trap = false;
882            }
883            "-fno-wrapv" => opts.wrapping.signed = false,
884            "-fwrapv-pointer" => opts.wrapping.pointer = true,
885            "-fno-wrapv-pointer" => opts.wrapping.pointer = false,
886            "-fno-strict-overflow" => opts.wrapping = Wrapping::ALL,
887            // Which does not clear the checked one, because gcc does not: `-ftrapv
888            // -fstrict-overflow` still emits the calls. It says what is assumed and not what
889            // happens.
890            "-fstrict-overflow" => {
891                opts.wrapping.signed = false;
892                opts.wrapping.pointer = false;
893            }
894            "-ftrapv" => {
895                opts.wrapping.trap = true;
896                opts.wrapping.signed = false;
897            }
898            "-fno-trapv" => opts.wrapping.trap = false,
899            // The two flags that say what a plain `char` is, which is one question with two
900            // spellings each: gcc reads `-fno-signed-char` as `-funsigned-char` and
901            // `-fno-unsigned-char` as `-fsigned-char`, so there are four ways to write two
902            // answers and the last one written wins. Nothing is set until one of them is given,
903            // because the target's own ABI is the answer otherwise and it is not the same answer
904            // everywhere: x86-64 and Apple's arm64 are signed, Linux's arm64 is not.
905            "-fsigned-char" | "-fno-unsigned-char" => opts.char_signed = Some(true),
906            "-funsigned-char" | "-fno-signed-char" => opts.char_signed = Some(false),
907            // And the size of an enumeration, which is the other thing in this group that changes
908            // the ABI rather than the code.
909            "-fshort-enums" => opts.short_enums = true,
910            "-fno-short-enums" => opts.short_enums = false,
911            // And Microsoft's reading of an anonymous member, which changes the layout of every
912            // record that writes a tag on one. Nothing is set until one of them is given, because
913            // the target is the answer otherwise: gcc's mingw build has this on and its Linux
914            // build has it off.
915            "-fms-extensions" => opts.ms_extensions = Some(true),
916            "-fno-ms-extensions" => opts.ms_extensions = Some(false),
917            // And the request, which is the one that cannot be granted. It is a real difference and
918            // not a preference: two files each writing `int g;` link under `-fcommon` and are a
919            // duplicate definition without it, which is the whole reason the flag survives.
920            "-fcommon" => {
921                return Err(err(
922                    "a tentative definition is written into .bss as its own symbol here, and \
923                     nothing emits the common symbol this asks the linker to merge. Give the \
924                     variable a definition in one file and declare it extern in the others",
925                ));
926            }
927            // Both directions of this one are recorded, and what they decide is whether lowering
928            // names the type each access goes through. Turning it off is the front end leaving the
929            // name off rather than a pass being told to ignore one it can see, which is one
930            // condition in one place, and it is the reading that survives link time optimization:
931            // a unit built with the flag off keeps its own answer when its bodies end up in a
932            // module beside bodies that were not.
933            //
934            // Nothing in the pipeline reads those names yet. Layer 3 of the alias analysis does
935            // and is tested, and no pass at any level asks the alias analysis anything today, so
936            // no program compiles differently for having passed this. The flag is wired anyway,
937            // because the change that makes a pass ask is not the change anybody will remember to
938            // wire it in, and a flag that is taken and dropped once the names mean something is
939            // the miscompilation `spec/04-driver-and-cli.md` section 4.1 warns about in as many
940            // words.
941            "-fstrict-aliasing" => opts.strict_aliasing = true,
942            "-fno-strict-aliasing" => opts.strict_aliasing = false,
943            // The same shape of answer for the same reason, and the flag the kernel writes beside
944            // the one above it.
945            //
946            // Nothing here concludes that a pointer is not null from the fact that it was
947            // dereferenced. There is no such conclusion to draw from, because no pass records one:
948            // a load says where it read and nothing else, and a comparison against null is an
949            // ordinary comparison of two values the optimizer has no fact about. So a function
950            // that reads through a pointer and then tests it keeps the test, which is what the
951            // kernel wants and what `-fno-delete-null-pointer-checks` asks for, and what gcc has
952            // to be asked for because it draws the conclusion by default.
953            //
954            // `-fdelete-null-pointer-checks` is the request to draw it, and it goes the way
955            // `-fstrict-aliasing` does: assuming less than was asked for costs speed and not
956            // correctness, and `-O2` implies it, so refusing it would stop builds for nothing.
957            "-fdelete-null-pointer-checks" | "-fno-delete-null-pointer-checks" => {}
958            // The floating point group, which goes the same way and for the same reason, and which
959            // is worth writing out because the reason is easy to get backwards.
960            //
961            // Each of these has a restrictive spelling and a permissive one. The restrictive ones,
962            // `-frounding-math` and `-ftrapping-math`, say that the rounding mode may have been
963            // changed and that an exception raised by an operation may be looked at, so an
964            // arithmetic the compiler folds at compile time is an arithmetic whose rounding and
965            // whose exception the program does not get. Nothing here folds any floating point
966            // arithmetic in a function body: `0.1 + 0.2` is an `fadd` and `1.0 / 0.0` is a divide
967            // that runs, at every level. So both of those describe what already happens.
968            //
969            // The permissive ones are the other half, and they are licences rather than requests
970            // for an answer. `-fno-rounding-math` says the rounding mode is the default one and
971            // `-fno-trapping-math` says nothing looks at the exceptions, which together are
972            // permission to fold. Not folding is the conservative side of that permission and is
973            // what a program is entitled to whichever was written, so `-fno-rounding-math` costs
974            // speed and not correctness, which is the test section 4.1 puts a licence through.
975            "-frounding-math" | "-fno-rounding-math" => {}
976            // `-fno-trapping-math` is the one of the four that is kept, because there is one
977            // conversion this compiler does not fold and gcc folds under it, and the two answers
978            // differ. Converting a constant floating value to an integer type it does not fit in
979            // is undefined behaviour rather than a value: left to the hardware it is one
980            // instruction and the answer is the integer indefinite value, and folded it is the
981            // nearest end of the integer's range. Both compilers leave it to the instruction by
982            // default and gcc folds it under this flag, so a program built with it and compiled
983            // without it gets a different number rather than a slower one. `-ftrapping-math` is
984            // gcc's default, so a build spelling it out is asking for what it already has.
985            "-ftrapping-math" => opts.trapping_math = true,
986            "-fno-trapping-math" => opts.trapping_math = false,
987            // About temporary files rather than about code. There is nothing between the phases of
988            // one compilation here to write to a file in the first place.
989            "-pipe" => {}
990            // Nothing here writes colour, so all of these are the same answer, and it is the answer
991            // that costs nothing: the diagnostics come out plain either way and no build depends on
992            // an escape sequence being there. Taken rather than refused because cmake writes
993            // `-fdiagnostics-color=always` on every compile line when the generator is ninja, which
994            // makes this the second most common flag after `-fPIC` to stop a build over a question
995            // about how the text looks.
996            "-fdiagnostics-color" | "-fno-diagnostics-color" => {}
997            _ if arg.starts_with("-fdiagnostics-color=") => {}
998            // The link flags. None of them changes the compilation, which is why they are
999            // collected apart from `opts` and why `-lm` on a `-c` line is a note rather than an
1000            // error: it is a thing said to a linker that is not going to run.
1001            "-static" => link.is_static = true,
1002            "-shared" => link.shared = true,
1003            "-pie" => link.pie = Some(true),
1004            "-no-pie" | "-nopie" => link.pie = Some(false),
1005            "-nostdlib" => link.no_stdlib = true,
1006            "-nostartfiles" => link.no_startfiles = true,
1007            "-nodefaultlibs" => link.no_defaultlibs = true,
1008            "-fno-builtins-lib" => link.no_builtins_lib = true,
1009            "-fbuiltins-lib" => link.no_builtins_lib = false,
1010            "-rdynamic" | "-export-dynamic" => link.export_dynamic = true,
1011            "-s" => link.strip = true,
1012            // Into the ordered input list rather than a list of its own, because a great many of
1013            // the linker's options are a bracket around the files after them and an option that
1014            // lost its place among them says nothing. `--whole-archive` is the one that found this.
1015            "-Xlinker" => {
1016                let next = args.get(i).ok_or_else(|| err("-Xlinker requires an argument"))?;
1017                i += 1;
1018                inputs.push(Input::linker(next));
1019            }
1020            _ if arg.starts_with("-Wl,") => {
1021                // Commas separate arguments rather than being part of one, which is what makes
1022                // `-Wl,-rpath,/opt/lib` two words to the linker and one word here.
1023                inputs.extend(arg["-Wl,".len()..].split(',').map(Input::linker));
1024            }
1025            _ if arg.starts_with("-fuse-ld=") => {
1026                link.use_ld = Some(arg["-fuse-ld=".len()..].to_owned());
1027            }
1028            _ if arg.starts_with("-l") && arg.len() > 2 => {
1029                inputs.push(Input::library(&arg[2..]));
1030            }
1031            "-l" => {
1032                let next = args.get(i).ok_or_else(|| err("-l requires an argument"))?;
1033                i += 1;
1034                inputs.push(Input::library(next));
1035            }
1036            _ if arg.starts_with("-L") => {
1037                link.search.push(PathBuf::from(joined_or_next(arg, 2, args, &mut i)?));
1038            }
1039            _ if arg.starts_with("-B") => {
1040                link.prefixes.push(PathBuf::from(joined_or_next(arg, 2, args, &mut i)?));
1041            }
1042            _ if arg.starts_with("-j") => {
1043                jobs = Jobs::parse(&arg[2..]).map_err(err)?;
1044            }
1045            _ if arg.starts_with("--sysroot=") => {
1046                sysroot = Some(PathBuf::from(&arg["--sysroot=".len()..]));
1047            }
1048            _ if arg.starts_with("--target=") => {
1049                let t = &arg["--target=".len()..];
1050                opts.target = t.parse().map_err(|e| err(format!("{e}")))?;
1051                // The same string again, as the model that has room for a libc version. A spelling
1052                // the three field parser took and this one does not is not an error, because the
1053                // one that decides what is compiled has already accepted it and the only thing
1054                // lost is a version nobody asked for.
1055                pinned = t.parse().ok();
1056            }
1057            _ if arg.starts_with("--emit=") => {
1058                let k = &arg["--emit=".len()..];
1059                opts.emit = k
1060                    .parse()
1061                    .map_err(|()| err(format!("unknown --emit kind `{k}`, see --help")))?;
1062            }
1063            // A bare `-O` is `-O1`, which is what GCC has and what a hand written makefile tends
1064            // to write. `-Og` is GCC's level for a build somebody is going to step through, and
1065            // it is `-O1` with the transformations that move code around left out; this compiler
1066            // has no such level yet, so it is the nearest one and `--print-pipeline` says what
1067            // that came to rather than the flag pretending otherwise.
1068            "-O" | "-Og" => opts.opt_level = rucc_session::OptLevel::O1,
1069            // The union of `-O3` and `-ffast-math`, and the second half of that changes what
1070            // floating point arithmetic means. Refused rather than taken as `-O3`, because a
1071            // build that asks for fast math and is quietly given ordinary arithmetic gets a
1072            // slower program than it asked for and a build that is given fast math it did not
1073            // ask for gets a wrong one.
1074            "-Ofast" => {
1075                return Err(err(
1076                    "-Ofast is -O3 with fast math, and fast math is not implemented, see \
1077                     spec/04-driver-and-cli.md section 4.6",
1078                ));
1079            }
1080            _ if arg.starts_with("-O") => {
1081                opts.opt_level = arg[2..]
1082                    .parse()
1083                    .map_err(|()| err(format!("unknown optimization level `{arg}`")))?;
1084            }
1085            // How far a multiply and an addition may be fused into one rounding. Before the
1086            // optimizer's `-f` family below for the reason the ones under it are, and kept rather
1087            // than dropped because it is the one flag in its group this compiler could act on: it
1088            // rides into the IR as an attribute on each function with a body, so the day the code
1089            // generator forms an `fma` it already knows which functions were given permission.
1090            // Nothing forms one today, under any value of this and under any `-march=`.
1091            _ if arg.starts_with("-ffp-contract=") => {
1092                let how = &arg["-ffp-contract=".len()..];
1093                opts.fp_contract = how.parse().map_err(|()| {
1094                    err(format!("`{how}` is not a contraction, which is fast, on or off"))
1095                })?;
1096            }
1097            // How much of an expression may be computed wider than it was written. The values are
1098            // gcc's and so is the refusal of anything else, and none of the three changes anything
1099            // here: an operation is computed in the type C says it is on every target this compiler
1100            // has a back end for, so `__FLT_EVAL_METHOD__` is 0 and `standard` is already what
1101            // happens. `fast` and `16` are permission to be wider, which is a licence this takes
1102            // and does not use, the same way the two above are. The flag is worth taking because
1103            // glibc's headers and a good deal of configure output write it, and because the answer
1104            // it asks about is one this compiler can state rather than guess at: there is no x87
1105            // target here, which is the machine the whole question was invented for.
1106            // Whether a local and a spilled value that are never both wanted may be the same bytes
1107            // of the frame. gcc's three values, and two of them mean the same thing here: what rucc
1108            // shares is a local whose address provably never leaves the function, which is narrower
1109            // than `named_vars` and narrower still than `all`, so both of them get it. `none` is
1110            // the one that changes anything, and it is the flag a program that reads a local
1111            // through a pointer it kept past the end of the block writes.
1112            _ if arg.starts_with("-fstack-reuse=") => {
1113                let how = &arg["-fstack-reuse=".len()..];
1114                opts.stack_reuse = match how {
1115                    "all" | "named_vars" => Some(true),
1116                    "none" => Some(false),
1117                    _ => {
1118                        return Err(err(format!(
1119                            "`{how}` is not a stack reuse, which is all, named_vars or none"
1120                        )));
1121                    }
1122                };
1123            }
1124            _ if arg.starts_with("-fexcess-precision=") => {
1125                let how = &arg["-fexcess-precision=".len()..];
1126                if !matches!(how, "16" | "fast" | "standard") {
1127                    return Err(err(format!(
1128                        "`{how}` is not an excess precision, which is 16, fast or standard"
1129                    )));
1130                }
1131            }
1132            // Which front of a path is rewritten before it reaches the output, which is how a
1133            // build gets the same bytes out of two different directories. The four spellings are
1134            // one flag each into three lists, and `-ffile-prefix-map=` is the three of them at
1135            // once. Only the macro list does anything today, because `__FILE__` is the only place
1136            // a path reaches the output: there is no DWARF and no profile data yet, so the other
1137            // two are recorded for the work that will read them. The argument splits at the last
1138            // `=` rather than the first, which is gcc's rule and is what lets a directory with an
1139            // `=` in its name be the old half.
1140            _ if arg.starts_with("-fmacro-prefix-map=") => {
1141                let (old, new) = rewrite(arg, "-fmacro-prefix-map=")?;
1142                opts.prefix_map.macros.push(old, new);
1143            }
1144            _ if arg.starts_with("-fdebug-prefix-map=") => {
1145                let (old, new) = rewrite(arg, "-fdebug-prefix-map=")?;
1146                opts.prefix_map.debug.push(old, new);
1147            }
1148            _ if arg.starts_with("-fprofile-prefix-map=") => {
1149                let (old, new) = rewrite(arg, "-fprofile-prefix-map=")?;
1150                opts.prefix_map.profile.push(old, new);
1151            }
1152            _ if arg.starts_with("-ffile-prefix-map=") => {
1153                let (old, new) = rewrite(arg, "-ffile-prefix-map=")?;
1154                opts.prefix_map.macros.push(old, new);
1155                opts.prefix_map.debug.push(old, new);
1156                opts.prefix_map.profile.push(old, new);
1157            }
1158            // A whole optimization rather than a flag, and the family is taken rather than
1159            // refused because of what ignoring it does. There is none of it here yet, so a build
1160            // that asks for it gets a program that is correct and slower than it could have been,
1161            // which is what section 4.1 means by a hint about speed and what every compilation at
1162            // `-O0` already is. The objects settle the rest of the argument: gcc's `-flto` object
1163            // holds the bytecode and no machine code at all, and every object here holds the code,
1164            // which is exactly what `-ffat-lto-objects` asks gcc for. So a build passing `-flto`
1165            // to this compiler gets objects that are more usable than the ones it asked for rather
1166            // than different ones. Every value is still checked against gcc's, because somebody
1167            // who wrote `-flto=thin` meant clang and had better hear about it here.
1168            "-flto" => opts.lto.requested = true,
1169            "-fno-lto" => opts.lto.requested = false,
1170            _ if arg.starts_with("-flto=") => {
1171                let how = &arg["-flto=".len()..];
1172                opts.lto.jobs = how.parse().map_err(|()| {
1173                    err(format!(
1174                        "`{how}` is not a number of link time jobs, which is auto, jobserver or a \
1175                         count above zero"
1176                    ))
1177                })?;
1178                opts.lto.requested = true;
1179            }
1180            _ if arg.starts_with("-flto-partition=") => {
1181                let how = &arg["-flto-partition=".len()..];
1182                opts.lto.partition = how.parse().map_err(|()| {
1183                    err(format!(
1184                        "`{how}` is not a partitioning model, which is balanced, 1to1, one, max \
1185                         or none"
1186                    ))
1187                })?;
1188            }
1189            _ if arg.starts_with("-flto-compression-level=") => {
1190                let how = &arg["-flto-compression-level=".len()..];
1191                let level =
1192                    how.parse::<u8>().ok().filter(|level| *level <= 19).ok_or_else(|| {
1193                        err(format!("`{how}` is not a compression level, 0 to 19"))
1194                    })?;
1195                opts.lto.compression = Some(level);
1196            }
1197            // Whether the object keeps its machine code as well as the bytecode. It always does
1198            // here, so the first of these describes what happens and the second asks for an object
1199            // with less in it, which is a smaller file and not a different program, so both are
1200            // taken.
1201            "-ffat-lto-objects" | "-fno-fat-lto-objects" => {}
1202            // Whether the linker is handed a plugin that does the link time work. The design in
1203            // `spec/09-optimizer.md` has this driver doing that work itself and never loading a
1204            // plugin into anybody, so neither answer is a question it has to hold.
1205            "-fuse-linker-plugin" | "-fno-use-linker-plugin" => {}
1206            // Reading a profile back. Taken for the reason the family above it is: nothing here
1207            // reads one, so a build that asks gets the program it would have got anyway, and gcc
1208            // itself produces a byte for byte identical object from `-fprofile-use` when there are
1209            // no counts beside the file. The path is recorded for the pass that will read it. The
1210            // warning gcc prints when it looked and found nothing is deliberately not copied,
1211            // because nothing here looks, and a warning about a file that was never opened would
1212            // fire on the builds that have a perfectly good profile as well as on the ones that
1213            // do not.
1214            "-fprofile-use" => opts.profile_data.requested = true,
1215            "-fno-profile-use" => opts.profile_data.requested = false,
1216            _ if arg.starts_with("-fprofile-use=") => {
1217                opts.profile_data.path = Some(arg["-fprofile-use=".len()..].to_string());
1218                opts.profile_data.requested = true;
1219            }
1220            _ if arg.starts_with("-fprofile-dir=") => {
1221                opts.profile_data.dir = Some(arg["-fprofile-dir=".len()..].to_string());
1222            }
1223            "-fprofile-abs-path" => opts.profile_data.absolute = true,
1224            "-fno-profile-abs-path" => opts.profile_data.absolute = false,
1225            "-fprofile-correction" => opts.profile_data.correction = true,
1226            "-fno-profile-correction" => opts.profile_data.correction = false,
1227            "-fprofile-partial-training" => opts.profile_data.partial_training = true,
1228            "-fno-profile-partial-training" => opts.profile_data.partial_training = false,
1229            // Writing the counts rather than reading them, which is refused rather than taken and
1230            // is the same line `-gsplit-dwarf` falls on the far side of. Ignoring these means a
1231            // file a build declared as an output never appears: the instrumented program writes a
1232            // `.gcda` as it exits and `-ftest-coverage` writes a `.gcno` beside the object, and a
1233            // two stage build that got neither would go on to optimize against no counts at all
1234            // and report coverage of nothing, with nothing along the way saying so. The objects
1235            // say the rest: gcc's `-fprofile-generate` object holds 375 bytes of code where a
1236            // plain one holds 71, and 296 bytes of counters that a plain one does not have, so
1237            // this is a flag that changes the output rather than a hint about speed.
1238            "-fprofile-arcs"
1239            | "--coverage"
1240            | "-fcondition-coverage"
1241            | "-fpath-coverage"
1242            | "-fprofile-generate" => {
1243                return Err(err(format!(
1244                    "{arg}: this compiler does not instrument for profiling, and a build that \
1245                     expects the counts a run of the instrumented program writes would optimize \
1246                     against nothing on its second pass, see spec/04-driver-and-cli.md"
1247                )));
1248            }
1249            _ if arg.starts_with("-fprofile-generate=") => {
1250                return Err(err(format!(
1251                    "{arg}: this compiler does not instrument for profiling, and a build that \
1252                     expects the counts a run of the instrumented program writes would optimize \
1253                     against nothing on its second pass, see spec/04-driver-and-cli.md"
1254                )));
1255            }
1256            "-ftest-coverage" => {
1257                return Err(err(format!(
1258                    "{arg}: this compiler writes no `.gcno` file beside the object, and a build \
1259                     that expects one would wait for a file that never arrives, see \
1260                     spec/04-driver-and-cli.md"
1261                )));
1262            }
1263            // The rest of the family describes instrumentation that is refused above, so what is
1264            // left to do with them is check them and drop them. They are checked because a
1265            // misspelling in a distribution's flags is worth finding here rather than on the day
1266            // the instrumentation lands, and dropped because there is nothing for an answer about
1267            // how a counter is written to be an answer about.
1268            _ if arg.starts_with("-fprofile-update=") => {
1269                let how = &arg["-fprofile-update=".len()..];
1270                if !matches!(how, "single" | "atomic" | "prefer-atomic") {
1271                    return Err(err(format!(
1272                        "`{how}` is not a profile update method, which is single, atomic or \
1273                         prefer-atomic"
1274                    )));
1275                }
1276            }
1277            _ if arg.starts_with("-fprofile-reproducible=") => {
1278                let how = &arg["-fprofile-reproducible=".len()..];
1279                if !matches!(how, "serial" | "parallel-runs" | "multithreaded") {
1280                    return Err(err(format!(
1281                        "`{how}` is not a profile reproducibility method, which is serial, \
1282                         parallel-runs or multithreaded"
1283                    )));
1284                }
1285            }
1286            "-fprofile-values" | "-fno-profile-values" | "-fprofile-info-section" => {}
1287            "-fno-test-coverage" | "-fno-profile-arcs" | "-fno-profile-generate" => {}
1288            _ if arg.starts_with("-fprofile-filter-files=")
1289                || arg.starts_with("-fprofile-exclude-files=")
1290                || arg.starts_with("-fprofile-note=") => {}
1291            // What every name gets when nothing in the source said, which the attribute in the
1292            // source overrides rather than the other way round. Before the optimizer's `-f`
1293            // family below for the reason the tier below it is.
1294            _ if arg.starts_with("-fvisibility=") => {
1295                let seen = &arg["-fvisibility=".len()..];
1296                opts.visibility = seen.parse().map_err(|()| {
1297                    err(format!(
1298                        "`{seen}` is not a visibility, which is default, hidden, internal or \
1299                         protected"
1300                    ))
1301                })?;
1302            }
1303            // Which edges of a control flow transfer are checked. Before the optimizer's `-f`
1304            // family below for the reason the two above it are, and last of the three so that the
1305            // bare spelling and the negative one are matched exactly rather than by this.
1306            _ if arg.starts_with("-fcf-protection=") => {
1307                let edges = &arg["-fcf-protection=".len()..];
1308                opts.control = edges.parse().map_err(|()| {
1309                    err(format!(
1310                        "`{edges}` is not a control flow protection, which is full, branch, \
1311                         return, none or check"
1312                    ))
1313                })?;
1314            }
1315            // How much room every function opens with for something to be written over later.
1316            // Before the optimizer's `-f` family below for the reason the ones above it are.
1317            _ if arg.starts_with("-fpatchable-function-entry=") => {
1318                let room = &arg["-fpatchable-function-entry=".len()..];
1319                opts.patchable = room.parse().map_err(|()| {
1320                    err(format!(
1321                        "`{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"
1322                    ))
1323                })?;
1324            }
1325            // The memory safety monitor, from section 15.4 of
1326            // `spec/safe-memory/15-integration.md`. Before the optimizer's `-f` family below,
1327            // because a pass that took the name `safety=detect` would otherwise be handed the
1328            // flag, and the tier is not a pass.
1329            _ if arg.starts_with("-fsafety=") => {
1330                let tier = &arg["-fsafety=".len()..];
1331                opts.safety = tier.parse().map_err(|()| {
1332                    err(format!(
1333                        "`{tier}` is not a safety tier, which is off, detect, enforce or kernel"
1334                    ))
1335                })?;
1336            }
1337            // Whether padding participates, from section 9.3 of document 09. Spelled out rather
1338            // than folded into the tier because it is a departure somebody who has read that
1339            // section makes, and the two defaults it describes are a property of what is being
1340            // built rather than of how much checking is wanted.
1341            _ if arg.starts_with("-fsafety-init=") => {
1342                let mode = &arg["-fsafety-init=".len()..];
1343                opts.padding = mode.parse().map_err(|()| {
1344                    err(format!("`{mode}` is not a padding mode, which is padding or nopadding"))
1345                })?;
1346            }
1347            // Row S4, from section 9.4 of document 09. A bare flag with no value, because the
1348            // strict form of that section needs a member id the front end does not name yet and
1349            // accepting the spelling for it would be accepting a promise this build cannot keep.
1350            // Before `-fno-` is looked at below, for the reason the tier is.
1351            "-fsafety-subobject" => opts.subobject = rucc_session::Subobject::Members,
1352            "-fno-safety-subobject" => opts.subobject = rucc_session::Subobject::Off,
1353            _ if arg.starts_with("-fsafety-subobject=") => {
1354                let form = &arg["-fsafety-subobject=".len()..];
1355                return Err(err(format!(
1356                    "`{form}` is not a form of -fsafety-subobject. The flag takes no value, and \
1357                     the strict form of section 9.4 is tamnd/rucc#967"
1358                )));
1359            }
1360            // Row Y8, from section 9.6 of document 09. A bare flag with no value, for the reason
1361            // the one above has none: there is one form of this check and a spelling that suggested
1362            // otherwise would be promising something. Before `-fno-` is looked at below, the same
1363            // way.
1364            "-fsafety-restrict" => opts.promise = rucc_session::Promise::Blocks,
1365            "-fno-safety-restrict" => opts.promise = rucc_session::Promise::Off,
1366            _ if arg.starts_with("-fsafety-restrict=") => {
1367                let form = &arg["-fsafety-restrict=".len()..];
1368                return Err(err(format!(
1369                    "`{form}` is not a form of -fsafety-restrict. The flag takes no value."
1370                )));
1371            }
1372            // Section 9.5's races, which take a value because the section gives them three modes
1373            // and the difference between two of them is which classes get reported rather than how
1374            // much is recorded. `-fno-` is the same as `=off` and is spelled out here for the same
1375            // reason the two above spell theirs out.
1376            _ if arg.starts_with("-fsafety-races=") => {
1377                let mode = &arg["-fsafety-races=".len()..];
1378                opts.races = mode.parse().map_err(|()| {
1379                    err(format!("`{mode}` is not a race mode, which is off, metadata or pointer"))
1380                })?;
1381            }
1382            "-fno-safety-races" => opts.races = rucc_session::Races::Off,
1383            // The sanitizers of document 12, which are checks at run time rather than a way of
1384            // generating the same program. Each name is held to gcc 16's list, and what is still
1385            // asked for by the end of the line is answered after the loop, so that a command line
1386            // which turns one on and then off again is a command line that asked for nothing.
1387            //
1388            // Before the optimizer's `-f` family below, for the reason the tier above it is.
1389            _ if arg.starts_with("-fsanitize=") => {
1390                for one in arg["-fsanitize=".len()..].split(',') {
1391                    if one == "all" {
1392                        // gcc takes `all` only in the negative, because turning every check on at
1393                        // once includes checks that contradict each other.
1394                        return Err(err(
1395                            "`-fsanitize=all` is not a gcc option, only `-fno-sanitize=all` is",
1396                        ));
1397                    }
1398                    if !SANITIZERS.contains(&one) {
1399                        return Err(err(format!(
1400                            "`{one}` is not a sanitizer, see spec/04-driver-and-cli.md section 4.7"
1401                        )));
1402                    }
1403                    if !sanitizers.contains(&one) {
1404                        sanitizers.push(one);
1405                    }
1406                }
1407            }
1408            _ if arg.starts_with("-fno-sanitize=") => {
1409                for one in arg["-fno-sanitize=".len()..].split(',') {
1410                    if one == "all" {
1411                        sanitizers.clear();
1412                        continue;
1413                    }
1414                    if !SANITIZERS.contains(&one) {
1415                        return Err(err(format!(
1416                            "`{one}` is not a sanitizer, see spec/04-driver-and-cli.md section 4.7"
1417                        )));
1418                    }
1419                    sanitizers.retain(|asked| *asked != one);
1420                }
1421            }
1422            // What a check does when it fires, and where the records about the checked objects go.
1423            // Each of them is an answer about the sanitizers refused after the loop, so there is
1424            // nothing left for them to change here. The names are still held to the list, because
1425            // a misspelling in a build's flags is worth finding when the compiler reads it.
1426            _ if arg.starts_with("-fsanitize-recover=")
1427                || arg.starts_with("-fno-sanitize-recover=")
1428                || arg.starts_with("-fsanitize-trap=")
1429                || arg.starts_with("-fno-sanitize-trap=") =>
1430            {
1431                // The guard above matched on a spelling that has an `=` in it, so the tail is
1432                // whatever follows the first one.
1433                let how = arg.split_once('=').map_or("", |(_, rest)| rest);
1434                for one in how.split(',') {
1435                    if one != "all" && !SANITIZERS.contains(&one) {
1436                        return Err(err(format!(
1437                            "`{one}` is not a sanitizer, see spec/04-driver-and-cli.md section 4.7"
1438                        )));
1439                    }
1440                }
1441            }
1442            "-fsanitize-undefined-trap-on-error"
1443            | "-fsanitize-address-use-after-scope"
1444            | "-fno-sanitize-address-use-after-scope" => {}
1445            _ if arg.starts_with("-fsanitize-sections=") => {}
1446            // Counting which edges a run reached, which is how a fuzzer knows an input was worth
1447            // keeping. Refused rather than dropped, because a fuzzer whose calls into
1448            // `__sanitizer_cov_*` were never generated runs blind and reports coverage of nothing,
1449            // and there is no point in the campaign where that announces itself.
1450            _ if arg.starts_with("-fsanitize-coverage=") => {
1451                let how = &arg["-fsanitize-coverage=".len()..];
1452                for one in how.split(',') {
1453                    if !matches!(one, "trace-pc" | "trace-cmp") {
1454                        return Err(err(format!(
1455                            "`{one}` is not a coverage instrumentation, which is trace-pc or \
1456                             trace-cmp"
1457                        )));
1458                    }
1459                }
1460                return Err(err(format!(
1461                    "{arg}: this compiler generates no coverage callbacks, and a fuzzer built \
1462                     with it would run without any feedback at all, see \
1463                     spec/04-driver-and-cli.md section 4.7"
1464                )));
1465            }
1466            // The optimizer's own flags, from section 9.10 of `spec/09-optimizer.md`. These come
1467            // after every `-f` the rest of the compiler answers to, so a pass can never take a
1468            // name that already means something else on the command line.
1469            _ if arg.starts_with("-fpass-fuel=") => {
1470                let (name, count) = arg["-fpass-fuel=".len()..]
1471                    .split_once('=')
1472                    .ok_or_else(|| err("-fpass-fuel= is spelled <pass>=<count>"))?;
1473                if rucc_opt::pass::find(name).is_none() {
1474                    return Err(err(format!(
1475                        "`{name}` is not a pass this compiler has, see --print-pipeline"
1476                    )));
1477                }
1478                let count: u32 = count
1479                    .parse()
1480                    .map_err(|_| err(format!("`{count}` is not a number of transformations")))?;
1481                opts.pass_fuel.push((name.to_owned(), count));
1482            }
1483            _ if arg.starts_with("-fpass-fuel-global=") => {
1484                let count = &arg["-fpass-fuel-global=".len()..];
1485                let count: u32 = count
1486                    .parse()
1487                    .map_err(|_| err(format!("`{count}` is not a number of transformations")))?;
1488                opts.pass_fuel_global = Some(count);
1489            }
1490            // Everything from `-fopt-info` to the end of the argument, which is optional
1491            // keywords joined by hyphens and an optional `=<file>`. Checked here rather than
1492            // where the remarks are printed, because by then the compilation somebody wanted
1493            // to hear about is over.
1494            _ if arg == "-fopt-info"
1495                || arg.starts_with("-fopt-info=")
1496                || arg.starts_with("-fopt-info-") =>
1497            {
1498                let rest = &arg["-fopt-info".len()..];
1499                let (kinds, file) = match rest.split_once('=') {
1500                    Some((kinds, file)) => (kinds, Some(file)),
1501                    None => (rest, None),
1502                };
1503                let kinds = kinds.strip_prefix('-').unwrap_or(kinds);
1504                rucc_opt::Wants::none().add(kinds).map_err(err)?;
1505                opts.opt_info.push(kinds.to_owned());
1506                if let Some(file) = file {
1507                    if file.is_empty() {
1508                        return Err(err("-fopt-info= was given no file to write to"));
1509                    }
1510                    opts.opt_info_file = Some(file.to_owned());
1511                }
1512            }
1513            _ if arg.starts_with("-fdump-ir=") => {
1514                // Checked here rather than where the dumps are taken, because the compilation
1515                // that would have been dumped is over by then.
1516                let spec = &arg["-fdump-ir=".len()..];
1517                rucc_opt::Dumps::default().add(spec).map_err(err)?;
1518                opts.dump_ir.push(spec.to_owned());
1519            }
1520            // Before the bare `-f<pass>` below, because a pass called `enable-something` would
1521            // otherwise take the flag away from the gate. Checked here rather than where the
1522            // pipeline reads it, for the reason that applies to all of these: a misspelled pass
1523            // name that quietly gated nothing looks exactly like a pass that is not the guilty
1524            // one, and a bisection would carry on past the thing it was looking for.
1525            _ if arg.starts_with("-fdisable-") || arg.starts_with("-fenable-") => {
1526                let on = arg.starts_with("-fenable-");
1527                let spec = &arg[if on { "-fenable-".len() } else { "-fdisable-".len() }..];
1528                rucc_opt::Gates::default().add(on, spec).map_err(err)?;
1529                opts.pass_gates.push((on, spec.to_owned()));
1530            }
1531            // gcc's spelling for a pass this compiler has under a shorter name. It goes above the
1532            // two arms below rather than into the pile of gcc pass names further down, because the
1533            // pass is here: dropping the flag would leave a build that asked for unrolling without
1534            // it, and refusing it stops the build outright, which is what libtommath's makefile
1535            // ran into. `-funroll-all-loops` is deliberately not in here: gcc's is the one that
1536            // unrolls without a trip count, which is a different and usually worse thing.
1537            "-funroll-loops" => opts.passes.push(("unroll".to_owned(), true)),
1538            "-fno-unroll-loops" => opts.passes.push(("unroll".to_owned(), false)),
1539            // Here rather than through the two arms below, because what this names is not a
1540            // `rucc_opt::Pass`. Section 34.6's propagation is a module at a time and everything in
1541            // the pass list is one function at a time. `-fipa-cp-clone` is deliberately not here:
1542            // gcc turns that one on at `-O3` and it is in the list of what M4 does not build.
1543            "-fipa-cp" => opts.passes.push((rucc_opt::ipcp::NAME.to_owned(), true)),
1544            "-fno-ipa-cp" => opts.passes.push((rucc_opt::ipcp::NAME.to_owned(), false)),
1545            // The other half of the same section, here for the same reason, and `-fipa-sra` in gcc
1546            // is the aggregate splitting as well as the parameter removal. Asking for it gets the
1547            // half that is built.
1548            "-fipa-sra" => opts.passes.push((rucc_opt::ipasra::NAME.to_owned(), true)),
1549            "-fno-ipa-sra" => opts.passes.push((rucc_opt::ipasra::NAME.to_owned(), false)),
1550            // And the printf family fold, which is a module at a time for the same reason and so is
1551            // not a `rucc_opt::Pass` either. gcc has no flag of its own for this one, since
1552            // `-fno-builtin` already turns it off along with everything else the standard names
1553            // mean. This spelling is for taking one thing away during a bisection without taking
1554            // the rest of section 20.1 away with it.
1555            "-flibcall" => opts.passes.push((rucc_opt::libcall::NAME.to_owned(), true)),
1556            "-fno-libcall" => opts.passes.push((rucc_opt::libcall::NAME.to_owned(), false)),
1557            _ if arg.strip_prefix("-fno-").is_some_and(|n| rucc_opt::pass::find(n).is_some()) => {
1558                opts.passes.push((arg["-fno-".len()..].to_owned(), false));
1559            }
1560            _ if arg.strip_prefix("-f").is_some_and(|n| rucc_opt::pass::find(n).is_some()) => {
1561                opts.passes.push((arg["-f".len()..].to_owned(), true));
1562            }
1563            // The flags that name a pass of gcc's own. They arrive from the torture suite, where a
1564            // program reduced from a miscompilation usually names the pass that miscompiled it on
1565            // its `dg-options` line, and they arrive from hand written build files for the same
1566            // reason. Section 4.1 sorts a flag by what the output would be without it, and by that
1567            // rule these are one pile: a flag that turns one of gcc's passes on or off is asking
1568            // for a compiler that does not exist here, and the program it is attached to is a
1569            // correctness test that passes either way. Turning on a pass we do not have costs
1570            // speed, turning off a pass we do not have costs nothing, and neither changes what the
1571            // program computes.
1572            //
1573            // rucc's own pass names are matched above this, so `-fno-dce` turns off the dce this
1574            // compiler has rather than landing here, and the day one of these names becomes a pass
1575            // here it stops being taken and dropped without anybody editing this list.
1576            //
1577            // Two of them are prefixes rather than names, which is the one place this file takes a
1578            // family instead of a flag. gcc files its gimple passes under `-ftree-` and its
1579            // interprocedural passes under `-fipa-`, both namespaces are pass selection and
1580            // nothing else, and there is no member of either that changes the meaning of a program
1581            // that was already correct. The rest are written out one at a time, because they live
1582            // in the flat `-f` namespace where the neighbours do change meanings.
1583            _ if arg.starts_with("-ftree-") || arg.starts_with("-fno-tree-") => {}
1584            _ if arg.starts_with("-fipa-") || arg.starts_with("-fno-ipa-") => {}
1585            "-fexpensive-optimizations" | "-fno-expensive-optimizations" => {}
1586            "-fmodulo-sched" | "-fno-modulo-sched" => {}
1587            "-fvect-cost-model" | "-fno-vect-cost-model" => {}
1588            _ if arg.starts_with("-fvect-cost-model=") || arg.starts_with("-fsimd-cost-model=") => {
1589            }
1590            "-fearly-inlining" | "-fno-early-inlining" => {}
1591            "-finline"
1592            | "-fno-inline"
1593            | "-finline-functions"
1594            | "-fno-inline-functions"
1595            | "-finline-small-functions"
1596            | "-fno-inline-small-functions"
1597            | "-finline-functions-called-once"
1598            | "-fno-inline-functions-called-once" => {}
1599            "-foptimize-strlen" | "-fno-optimize-strlen" => {}
1600            "-fira-share-spill-slots" | "-fno-ira-share-spill-slots" => {}
1601            // Where a function starts, which is a thing this compiler already decides and so is a
1602            // request it can answer rather than one it has to drop. The bare form asks for the
1603            // target's default and the default here is the sixteen bytes gcc also gives, so it
1604            // says nothing; a number is a floor under every function that did not ask for more
1605            // itself; and the negative form asks for the smallest boundary the target has. gcc 16
1606            // rounds a number that is not a power of two up rather than refusing it, which is what
1607            // `=3` giving `.p2align 2` on x86-64 means, so this rounds too.
1608            "-falign-functions" => opts.align_functions = None,
1609            "-fno-align-functions" => opts.align_functions = Some(MIN_FUNC_ALIGN),
1610            _ if arg.starts_with("-falign-functions=") => {
1611                opts.align_functions = function_alignment(&arg["-falign-functions=".len()..])
1612                    .ok_or_else(|| {
1613                        err(format!("{arg}: the alignment has to be a number of bytes"))
1614                    })?;
1615            }
1616            // The other three of the family, which are about padding in front of a label inside a
1617            // body. This compiler writes none, and what they ask for is speed: a loop that starts
1618            // on a cache line boundary computes what a loop that does not computes. So they are
1619            // taken and dropped for the reason `-march=` is, and they are kept out of the arm
1620            // above because the question they ask is a different one and the day one of them is
1621            // answered it will be answered separately.
1622            _ if arg.starts_with("-falign-labels")
1623                || arg.starts_with("-falign-loops")
1624                || arg.starts_with("-falign-jumps")
1625                || arg.starts_with("-fno-align-labels")
1626                || arg.starts_with("-fno-align-loops")
1627                || arg.starts_with("-fno-align-jumps") => {}
1628            // The charset flags are not in that pile, because an encoding is a statement about
1629            // what the bytes of the source mean rather than about how fast the output is. The
1630            // preprocessor reads UTF-8 and has no converter, so the one name that describes what
1631            // already happens is taken and every other name is refused. Spelled without regard to
1632            // case and with both of the spellings iconv answers to, since a build writes whichever
1633            // one its author typed.
1634            _ if arg.starts_with("-finput-charset=") => {
1635                let name = &arg["-finput-charset=".len()..];
1636                if !name.eq_ignore_ascii_case("utf-8") && !name.eq_ignore_ascii_case("utf8") {
1637                    return Err(err(format!(
1638                        "-finput-charset={name}: the preprocessor reads UTF-8 and has no \
1639                         converter, so a file in another encoding would be read as though it were \
1640                         UTF-8 rather than converted",
1641                    )));
1642                }
1643            }
1644            // The three that come in on the same `dg-options` lines and are the other half of
1645            // section 4.1's rule, because each of them changes what the program does and not how
1646            // fast it does it. The negative form of each is what this compiler does anyway, so it
1647            // is taken and dropped, which is the shape `-fnested-functions` has above.
1648            "-ffast-math" => {
1649                return Err(err(
1650                    "-ffast-math is a licence to answer a floating point arithmetic differently \
1651                     from the way the source wrote it, and it is not one flag: it defines \
1652                     __FAST_MATH__, which a library header reads, and gcc links a startup file \
1653                     that puts the hardware in flush to zero mode for the whole process. Taking it \
1654                     and dropping it would change what other objects in the same program answer. \
1655                     -ffp-contract= and -fexcess-precision= are the parts of it this compiler has",
1656                ));
1657            }
1658            "-fno-fast-math" => {}
1659            "-fnon-call-exceptions" => {
1660                return Err(err(
1661                    "-fnon-call-exceptions is a promise that an instruction which is not a call \
1662                     can raise an exception the unwinder finds a handler for, and nothing here \
1663                     produces a landing pad for a trapping instruction. A program built without it \
1664                     would unwind past the handler it wrote",
1665                ));
1666            }
1667            "-fno-non-call-exceptions" => {}
1668            "-finstrument-functions" => {
1669                return Err(err(
1670                    "-finstrument-functions calls __cyg_profile_func_enter on entry to every \
1671                     function and __cyg_profile_func_exit on the way out, and nothing here emits \
1672                     either call. A program that asks for them usually counts them, so taking the \
1673                     flag and dropping it would turn a program that fails loudly into one that \
1674                     fails quietly",
1675                ));
1676            }
1677            "-fno-instrument-functions" => {}
1678            // The unstable options, spelled the way rustc spells them and carrying the same
1679            // promise, which is none: one of these may change or go away in any release. They are
1680            // measurements and debugging aids rather than things a build asks for, which is why
1681            // none of them is in the usage text and all of them are in section 4.11 of
1682            // `spec/04-driver-and-cli.md`.
1683            "-Zverify-each" => opts.verify_each = true,
1684            _ if arg.starts_with("-Zrule-coverage=") => {
1685                let file = &arg["-Zrule-coverage=".len()..];
1686                if file.is_empty() {
1687                    return Err(err("-Zrule-coverage= needs a file to write to"));
1688                }
1689                opts.rule_coverage = Some(file.to_owned());
1690            }
1691            _ if arg.starts_with("-Zcycle-accurate-model=") => {
1692                let value = &arg["-Zcycle-accurate-model=".len()..];
1693                opts.cycle_accurate_model = match value {
1694                    "yes" | "1" => Some(true),
1695                    "no" | "0" => Some(false),
1696                    _ => {
1697                        return Err(err("-Zcycle-accurate-model= takes yes or no"));
1698                    }
1699                };
1700            }
1701            _ if arg.starts_with("-Zlowering=") => {
1702                let file = &arg["-Zlowering=".len()..];
1703                if file.is_empty() {
1704                    return Err(err("-Zlowering= needs a file to write to"));
1705                }
1706                opts.lowering_dump = Some(file.to_owned());
1707            }
1708            _ if arg.starts_with("-Zregister-pressure=") => {
1709                let file = &arg["-Zregister-pressure=".len()..];
1710                if file.is_empty() {
1711                    return Err(err("-Zregister-pressure= needs a file to write to"));
1712                }
1713                opts.register_pressure = Some(file.to_owned());
1714            }
1715            _ if arg.starts_with("-Z") => {
1716                return Err(err(format!(
1717                    "`{arg}` is not an unstable option this compiler has, see \
1718                     spec/04-driver-and-cli.md section 4.11 for the ones it does"
1719                )));
1720            }
1721            // The word size, which is a statement about the target and is taken as one. A build
1722            // that says the size the target already has is saying nothing, and one that says the
1723            // other size is asking for a target this compiler does not have, which it is told
1724            // rather than being given the wrong one.
1725            "-m64" | "-m32" | "-mx32" => {
1726                let want: u32 = match arg {
1727                    "-m64" => 64,
1728                    _ => 32,
1729                };
1730                let have = rucc_target::TargetInfo::new(opts.target).pointer_width;
1731                if have != want {
1732                    return Err(err(format!(
1733                        "{arg} asks for a {want} bit target and {} is {have} bit, use \
1734                         --target= to name the one you mean",
1735                        opts.target
1736                    )));
1737                }
1738            }
1739            // Which processor in the family to generate for. This compiler emits the base
1740            // instruction set of the architecture and nothing above it, so a program built with
1741            // any of these runs on the machine that was named; it is a program that could have
1742            // been faster rather than a program that is wrong, which is what makes these safe to
1743            // take and ignore where a flag that changed the meaning of the code would not be.
1744            _ if arg.starts_with("-march=")
1745                || arg.starts_with("-mtune=")
1746                || arg.starts_with("-mcpu=") => {}
1747            // The calling convention, which is not safe to ignore. Taken when it names the one
1748            // the target already uses and refused otherwise.
1749            _ if arg.starts_with("-mabi=") => {
1750                let want = &arg["-mabi=".len()..];
1751                let have = match opts.target.arch {
1752                    rucc_target::Arch::X86_64 => "sysv",
1753                    rucc_target::Arch::Aarch64 => "lp64",
1754                    rucc_target::Arch::Riscv64 => "lp64d",
1755                };
1756                if want != have {
1757                    return Err(err(format!(
1758                        "{arg}: {} uses the {have} convention and this compiler has no other",
1759                        opts.target
1760                    )));
1761                }
1762            }
1763            // How far apart the pieces of the program may be. The small model is what we emit and
1764            // it is every hosted program's default; the kernel model is a different one and a
1765            // build that asks for it and does not get it links and then does not run.
1766            "-mcmodel=small" => {}
1767            _ if arg.starts_with("-mcmodel=") => {
1768                return Err(err(format!(
1769                    "{arg}: this compiler emits the small code model and no other, see \
1770                     spec/12-targets.md"
1771                )));
1772            }
1773            // GCC's own scripting language for how the driver builds a command line.
1774            // `spec/04-driver-and-cli.md` section 4.4 settles that we will not have it, so a
1775            // build reaching for it is told which flags do the same job.
1776            _ if arg.starts_with("-specs=") => {
1777                return Err(err(
1778                    "-specs= is not supported: the parts of it builds rely on are -B, -L, \
1779                     -nostdlib, -nostartfiles and -Wl,, see spec/04-driver-and-cli.md \
1780                     section 4.4",
1781                ));
1782            }
1783            // Arguments meant for a separate assembler or preprocessor, which this compiler does
1784            // not have: both are inside it and neither reads a command line. Refused rather than
1785            // dropped, because every one of these says something about the output and a build
1786            // that asked for `-Wa,--noexecstack` and was silently given an executable stack got
1787            // the opposite of what it asked for.
1788            _ if arg.starts_with("-Wa,") || arg.starts_with("-Wp,") => {
1789                return Err(err(format!(
1790                    "`{arg}` is an argument for a separate assembler or preprocessor, and both \
1791                     are inside this compiler rather than programs it runs"
1792                )));
1793            }
1794            "-Xassembler" | "-Xpreprocessor" => {
1795                return Err(err(format!(
1796                    "{arg} hands an argument to a separate assembler or preprocessor, and both \
1797                     are inside this compiler rather than programs it runs"
1798                )));
1799            }
1800            // Everything else in the `-W` family. `spec/04-driver-and-cli.md` section 4.1 has
1801            // this one as a rule about build systems rather than about warnings: autoconf finds
1802            // out whether a warning flag exists by passing it and looking at the exit status, so
1803            // a compiler that refuses one it has not heard of fails a configure script written
1804            // for a GCC newer than itself. The names are not checked against a list because this
1805            // compiler has no warning groups for a list to be of, which #485 is about.
1806            _ if arg.starts_with("-W") => {}
1807            // Flags that name something this compiler does not do and would not do differently
1808            // if it did. `-fno-ident` is about a comment in the output that we do not write
1809            // either way, and the others are about a way of ordering the compilation that has
1810            // been GCC's only way for twenty years. Section 4.1 asks for the list to be short
1811            // and for adding to it to be deliberate, which is why it is written out here.
1812            "-fno-ident"
1813            | "-fident"
1814            | "-funit-at-a-time"
1815            | "-fno-unit-at-a-time"
1816            | "-shared-libgcc"
1817            | "-static-libgcc" => {}
1818            _ if arg.starts_with('-') && arg.len() > 1 => {
1819                // Silently ignoring an unknown flag is how a build ends up not doing what
1820                // its author asked. spec/13-gnu-compat.md section 13.4 makes this an error
1821                // for the flags that change code generation, and the safe default until the
1822                // flag table is populated is to reject everything we do not know.
1823                return Err(err(format!("unknown option `{arg}`")));
1824            }
1825            _ => inputs.push(Input { path: arg.to_owned(), forced, role: Role::File }),
1826        }
1827    }
1828
1829    // The fetch, before anything that resolves a compilation, because `--fetch` does not describe
1830    // one. It is here rather than in the loop so that `--offline` can forbid it whichever order the
1831    // two were written in, and it is before the refusals below so that a command line asking for a
1832    // sysroot is not told about a sanitizer.
1833    if let Some(named) = fetch {
1834        return fetch_action(&named, offline, &inputs);
1835    }
1836
1837    // Last, so that it lands after every `-isystem` the command line gave. That is GCC's
1838    // order: a directory the user names outranks the compiler's own, and the compiler's own
1839    // outranks the library's. It is pushed after the loop rather than before it because
1840    // `SearchPath` appends within a group and the position is what the order is.
1841    // The same directory the headers were looked for under, because a sysroot is a statement
1842    // about a whole installation and not about half of one.
1843    // After the loop, because `-fno-sanitize=` can take back what an earlier flag asked for and a
1844    // command line that turns a check on and off again has asked for nothing. What is left is
1845    // refused rather than dropped, and it is the one place in this parser where the reason is not
1846    // that the output would differ. A sanitizer is a promise that the program is watched while it
1847    // runs, so a build that asks for one and is quietly given a program with no checks in it does
1848    // not get a slower program or a bigger file, it gets a test suite that passes for the wrong
1849    // reason. `-fsafety=` is the checking this compiler does have, and the message says so, because
1850    // somebody reaching for `-fsanitize=address` wants the nearest thing rather than a list of
1851    // options.
1852    if let Some(first) = sanitizers.first() {
1853        return Err(err(format!(
1854            "-fsanitize={first}: this compiler has no sanitizer instrumentation, and a build that \
1855             asked for one and got none would run its tests unchecked, see \
1856             spec/04-driver-and-cli.md section 4.7. `-fsafety=detect` is the memory checking this \
1857             compiler does have"
1858        )));
1859    }
1860    link.sysroot = sysroot.clone();
1861    // Where a sysroot for a target that is not this machine would be. Read once, here, rather than
1862    // inside the link line, because a link line that read the environment could only be tested on a
1863    // machine whose environment said the right thing, and the link line is the last thing that
1864    // touches a binary. `spec/cross-compile/13-distribution.md` section 13.2 owns the answer.
1865    link.cache = Some(cache::dir());
1866    // And the ten field spelling of the target, because the release on it decides two things the
1867    // three field one cannot say: whether a target that is this architecture is still a cross
1868    // compile, and which directory under the cache it is against. After the loop because the last
1869    // `--target=` on the command line is the one that counts.
1870    link.pinned = pinned;
1871    // After the loop rather than where `-pthread` was read, so that it lands after the objects
1872    // that refer to it. A static link takes the definitions it needs from a library when it
1873    // reaches it and not afterwards, so a library before the objects is a library that answers
1874    // nothing.
1875    if threads {
1876        inputs.push(Input::library("pthread"));
1877    }
1878    if let Some(query) = query {
1879        return Ok(Action::Print(answer(&query, &opts, &link)?));
1880    }
1881    // `-M` and `-MM` produce the rule and nothing else, so the run stops after phase 4 whatever
1882    // else the command line asked for. Read here rather than where the flag was, because a `-c`
1883    // written after it has to lose and the loop cannot know that until it has ended. The output
1884    // file is where the rule goes rather than where an object would have gone, and the last
1885    // phase being the preprocessor is what makes that true without a second rule for it.
1886    if opts.deps.instead_of_compiling {
1887        opts.emit = EmitKind::Preprocessed;
1888    }
1889    if !nostdinc {
1890        opts.search.push_system(runtime::DIR);
1891        // And the library's after ours, which is the other half of the same order. They go on
1892        // here rather than at the point `--target=` or `--sysroot=` was read because either
1893        // one changes the answer and the last word on both is the end of the loop.
1894        //
1895        // Which library's is the question `link::cross_sysroot` answers, and it is asked here so
1896        // that the headers and the libraries come from the same place. A target that is this
1897        // machine reads this machine's headers, and a target that is not reads the ones in the
1898        // sysroot for it rather than the ones next door.
1899        let cross = link::cross_sysroot(opts.target, &link);
1900        let kernel = link::cross_kernel(opts.target, &link);
1901        // And the version of those headers, which only the bundled tree has an answer for. A host
1902        // glibc and a tree the user named both define `__GLIBC_MINOR__` in their own `features.h`,
1903        // and a second definition with a different value is a warning on every file, so the
1904        // condition is the same one that chose the directories.
1905        if cross.is_some() {
1906            let target = pinned.unwrap_or_else(|| opts.target.tuple());
1907            opts.glibc_minor = rucc_sysroot::bundled_glibc_minor(target).map_err(|skew| {
1908                err(format!(
1909                    "{skew}; pin a release the tree has, or name a tree that has that one \
1910                     with --sysroot"
1911                ))
1912            })?;
1913        }
1914        let system =
1915            library::header_dirs(opts.target, sysroot.as_deref(), cross.as_ref(), kernel.as_ref());
1916        // The two licence walls of `spec/cross-compile/13-distribution.md` section 13.4, which are
1917        // the only way step 3 comes back with nothing on a hosted target. Section 8.6 asks for the
1918        // answer to name the licence and the lawful ways to get what is behind it, rather than
1919        // leaving a person with an `#include` that failed as though a directory had gone missing.
1920        //
1921        // It is left on the search path instead of refused here, because a program that includes
1922        // none of the library needs none of the SDK and section 8.6 is explicit that targeting the
1923        // platform has to keep working. So the reason waits until an include has actually failed,
1924        // which is the only moment it helps and the only moment it is true.
1925        //
1926        // The condition is that step 3 found nothing at all, so an `SDKROOT`, an `INCLUDE` or a mac
1927        // with Xcode on it all pass through untouched, and `-nostdinc` never reaches this block. A
1928        // `--sysroot` or `-isysroot` passes through as well, even when the tree it names turns out to
1929        // be empty or absent: somebody who wrote a path has already answered the question this
1930        // message asks, and answering it again over the top of a mistyped directory would hide the
1931        // mistake behind a licence notice.
1932        if system.is_empty() && sysroot.is_none() {
1933            let tuple = pinned.unwrap_or_else(|| opts.target.tuple());
1934            if let Some(wall) = rucc_sysroot::Wall::of(tuple) {
1935                opts.search.explain_missing_system(wall.no_headers(&tuple.to_canonical_string()));
1936            }
1937        }
1938        // And whether the tree somebody named is the release they asked for, which is the one
1939        // question left once the directories are settled and the only place both halves of it are
1940        // known. Only for a named tree, because that is the case where the release in the target
1941        // stops deciding anything, and `crate::glibc` is where the rest of the reasoning is.
1942        if sysroot.is_some() {
1943            notes.extend(glibc::skew(opts.target, pinned, &system));
1944        }
1945        for dir in system {
1946            opts.search.push_system(dir);
1947        }
1948    }
1949    // Once, here, rather than as each directory is pushed. A `-I` that names a system
1950    // directory has to lose to the system entry and the system entry is added last, so the
1951    // question cannot be answered until the whole path is known.
1952    opts.search.remove_duplicates();
1953
1954    // The target has to be resolved before the configuration is printed, so this check comes
1955    // after the loop rather than at the point `--print-config` was seen.
1956    if print_config {
1957        return Ok(Action::PrintConfig(Box::new(opts)));
1958    }
1959    if print_pipeline {
1960        return Ok(Action::PrintPipeline(Box::new(opts)));
1961    }
1962    let plan = Plan::new(&opts, &inputs, output.as_deref()).map_err(|e| err(e.message))?;
1963    if print_plan {
1964        return Ok(Action::PrintPlan {
1965            opts: Box::new(opts),
1966            plan: Box::new(plan),
1967            link: Box::new(link),
1968        });
1969    }
1970    Ok(Action::Compile {
1971        opts: Box::new(opts),
1972        plan: Box::new(plan),
1973        link: Box::new(link),
1974        jobs,
1975        verbose,
1976        notes,
1977    })
1978}
1979
1980/// What `--fetch <tuple>` asked for, or why it is not a thing that can be done.
1981///
1982/// The lookup happens here rather than at the point the bytes would move, so that a target this
1983/// release pins nothing for is a refusal from the parser and the only code that runs a downloader is
1984/// code that already knows what it is getting.
1985///
1986/// # Errors
1987///
1988/// [`CliError`] when `--offline` forbade it, when there are input files as well, when the tuple is
1989/// not a target this compiler knows, when its sysroot is behind one of section 13.4's licence walls,
1990/// and when this release pins no artifact for it.
1991fn fetch_action(named: &str, offline: bool, inputs: &[Input]) -> Result<Action, CliError> {
1992    // Not a precedence question. Section 13.2 says `--offline` forbids a fetch entirely, so a
1993    // command line that writes both has asked for two opposite things and the answer is to say so
1994    // rather than to pick one of them.
1995    if offline {
1996        return Err(err(
1997            "--fetch asks for a download and --offline forbids every download, so this command \
1998             line asks for two opposite things. Drop one of them: --offline is how a build says it \
1999             will not reach the network, and --fetch is the only thing in this compiler that does",
2000        ));
2001    }
2002    if let Some(first) = inputs.first() {
2003        return Err(err(format!(
2004            "--fetch gets a sysroot and compiles nothing, so `{}` on the same command line is an \
2005             input that nothing would read",
2006            first.path
2007        )));
2008    }
2009    let target: TargetTuple = named
2010        .parse()
2011        .map_err(|why| err(format!("--fetch {named}: {why}, so there is no sysroot to get")))?;
2012    // The canonical spelling, because that is what a row is named by and what the directory under
2013    // the cache is called, and a person is free to write a tuple the long way round.
2014    let tuple = target.to_canonical_string();
2015    // Before the table is consulted, because a target behind a licence wall is not a row that has not
2016    // been written yet. Section 13.4 is that no release pins one of these ever, so the message says
2017    // the licence and the two lawful ways rather than naming the producer that will publish the rest.
2018    if let Some(wall) = rucc_sysroot::Wall::of(target) {
2019        return Err(err(format!("--fetch {tuple}: {}", wall.no_fetch(&tuple))));
2020    }
2021    let Some(what) = rucc_sysroot::pinned_for(&tuple) else {
2022        return Err(err(unpinned(&tuple)));
2023    };
2024    Ok(Action::Fetch { what, target, cache: cache::dir() })
2025}
2026
2027/// Why there is nothing to fetch for a target, which is a different sentence when the table is
2028/// empty.
2029///
2030/// A release that pins nothing and a release that pins eleven targets and not this one are two
2031/// situations, and a message that did not tell them apart would send somebody looking for a typo in
2032/// their tuple when the answer is that this work is not finished.
2033fn unpinned(tuple: &str) -> String {
2034    let pinned = rucc_sysroot::pinned_targets();
2035    if pinned.is_empty() {
2036        return format!(
2037            "this release pins no sysroot for {tuple}, and it pins none for any target yet. A \
2038             sysroot is built and published by the producer in tamnd/rucc-cross, per \
2039             spec/cross-compile/13-distribution.md section 13.8, and a release of this compiler \
2040             names one by URL and by hash afterwards. Until then, pass --sysroot=<dir> to compile \
2041             against a tree you have already"
2042        );
2043    }
2044    format!(
2045        "this release pins no sysroot for {tuple}. What it pins is {}. Pass --sysroot=<dir> to \
2046         compile against a tree you have already",
2047        pinned.join(", ")
2048    )
2049}
2050
2051/// Gets the artifact and installs it, saying what each step did.
2052///
2053/// The steps are section 13.8's and so are the messages: the transport is somebody else's program
2054/// and the check is ours, so a person reading this wants to know which downloader ran, that the
2055/// bytes matched, how many files the record named and where the tree ended up. A fetch of something
2056/// that is already there says that instead and moves nothing.
2057fn fetch_sysroot(what: &rucc_sysroot::Pinned, target: TargetTuple, cache: &std::path::Path) -> i32 {
2058    let tuple = target.to_canonical_string();
2059    let archive = what.archive_in(cache);
2060    let say = |line: &str| println!("rucc: {tuple}: {line}");
2061    match fetch::fetch(what.url, what.sha256, &archive) {
2062        Ok(fetch::Fetched::AlreadyThere) => {
2063            say(&format!("{} is already here and matches the hash", archive.display()));
2064        }
2065        Ok(fetch::Fetched::Downloaded(by)) => {
2066            say(&format!("downloaded {} with {}", what.url, by.program()));
2067        }
2068        Err(why) => return complain(why),
2069    }
2070    match install::install(&archive, what.sha256, target, cache) {
2071        Ok(done) => {
2072            match &done.before {
2073                install::Before::Nothing => {
2074                    say(&format!("{} files installed at {}", done.files, done.root.display()));
2075                }
2076                install::Before::TheSame => {
2077                    say(&format!(
2078                        "the same sysroot is already at {}, so nothing moved",
2079                        done.root.display()
2080                    ));
2081                }
2082                install::Before::Different(was) => {
2083                    say(&format!(
2084                        "{} files installed at {}, over a tree whose record digested to {was}",
2085                        done.files,
2086                        done.root.display()
2087                    ));
2088                }
2089            }
2090            say(&format!("the record digests to {}", done.digest));
2091            0
2092        }
2093        Err(why) => complain(why),
2094    }
2095}
2096
2097/// What one of the `-dump` and `-print` flags prints.
2098///
2099/// GCC prints the name back unchanged when it cannot find the file a `-print` flag asked about,
2100/// which is what makes the answer safe to paste into a link line whether or not the file is
2101/// there, and this does the same.
2102fn answer(query: &Query, opts: &Options, link: &LinkOptions) -> Result<String, CliError> {
2103    let found = |name: &str| {
2104        link::find_in_search(link, opts.target, name)
2105            .map_or_else(|| name.to_owned(), |path| path.display().to_string())
2106    };
2107    Ok(match query {
2108        Query::Machine => opts.target.to_string(),
2109        Query::Version => VERSION.to_owned(),
2110        Query::Multiarch => link::multiarch(opts.target),
2111        // The three lines GCC prints, in its order and with its punctuation, because what reads
2112        // them is a script written against that shape. There is no installation directory to
2113        // report: this compiler is one binary that works wherever it is copied, and the headers
2114        // it ships are inside it, so `install` is where the binary is and nothing is under it.
2115        Query::SearchDirs => {
2116            let here = std::env::current_exe()
2117                .ok()
2118                .and_then(|p| p.parent().map(std::path::Path::to_path_buf))
2119                .unwrap_or_default();
2120            let list = |dirs: &[PathBuf]| {
2121                dirs.iter().map(|d| d.display().to_string()).collect::<Vec<_>>().join(":")
2122            };
2123            let libraries = link::search_dirs(link, opts.target);
2124            format!(
2125                "install: {}\nprograms: ={}\nlibraries: ={}",
2126                here.display(),
2127                list(&link.prefixes),
2128                list(&libraries)
2129            )
2130        }
2131        // The root the rest of the answers are under, which a build system asks for when it wants
2132        // to find a file itself rather than ask for one by name, and which is the first thing to
2133        // look at when a cross build read a header nobody expected. A native compile has no
2134        // sysroot and the answer is the empty line, which is what GCC prints when it was
2135        // configured without one. `--sysroot` wins over ours because it wins everywhere else.
2136        Query::Sysroot => {
2137            sysroot_root(opts, link).map(|root| root.display().to_string()).unwrap_or_default()
2138        }
2139        // Section 13.5 of `spec/cross-compile/13-distribution.md`: for every input that is not this
2140        // compiler's own code, what it is, where it was got, its hash, its licence and whether it
2141        // was bundled, generated or fetched. What is printed is the manifest the sysroot already
2142        // carries rather than a second format saying the same things, because the three uses 13.5
2143        // gives for this are a licence notice, a reproducibility check and a security audit, and all
2144        // three are somebody else parsing it. One format is one parser to write.
2145        // Read and rendered rather than copied out, so that what comes back is the format this
2146        // build understands. The last newline comes off because whatever prints an answer adds
2147        // one, the way it does for every other query here. Keeping it would put a blank line at
2148        // the end of the one answer that is a file somebody diffs against the file it came from.
2149        Query::SysrootProvenance => match sysroot_manifest(opts, link)? {
2150            Some(manifest) => manifest.render().trim_end_matches('\n').to_string(),
2151            None => String::new(),
2152        },
2153        // Section 13.2 of the same document, which asks for the hash of a cache directory's
2154        // contents in the directory's name. A name cannot carry one, because the path has to be
2155        // computable before anything has been read, by the producer about to write the files and by
2156        // the compiler about to read them, and neither has the contents when it asks. So the number
2157        // is here instead, and it is the sha256 of the record rather than of a walk of the tree,
2158        // which means `sha256sum` over the manifest answers the same thing.
2159        Query::SysrootDigest => match sysroot_manifest(opts, link)? {
2160            Some(manifest) => manifest.digest(),
2161            None => String::new(),
2162        },
2163        Query::FileName(name) => found(name),
2164        // The name GCC gives the library of routines a compiler's output calls that the C
2165        // library does not have. Ours is built in and there is no file, so the answer is the
2166        // name itself, which is what GCC prints when it cannot find one either.
2167        Query::Libgcc => found("libgcc.a"),
2168        // A program rather than a library: the linker and the archiver are the ones a build asks
2169        // about, and this compiler finds them on the path or under `-B` rather than shipping
2170        // them, so the name back is the honest answer unless a `-B` prefix holds one.
2171        Query::ProgName(name) => link
2172            .prefixes
2173            .iter()
2174            .map(|dir| dir.join(name))
2175            .find(|path| path.is_file())
2176            .map_or_else(|| name.clone(), |path| path.display().to_string()),
2177    })
2178}
2179
2180/// The root every sysroot answer is about.
2181///
2182/// One function rather than a copy in each, because the other flags exist to say what is inside the
2183/// tree this one names, and two answers that disagreed about which tree that is would be a
2184/// difference nobody would think to look for. `--sysroot` wins over ours because it wins everywhere
2185/// else.
2186fn sysroot_root(opts: &Options, link: &LinkOptions) -> Option<PathBuf> {
2187    link.sysroot
2188        .clone()
2189        .or_else(|| link::cross_sysroot(opts.target, link).map(|at| at.root().to_path_buf()))
2190}
2191
2192/// The record of the sysroot this command line reads, when there is one to read.
2193///
2194/// [`None`] covers two cases that both print nothing, and they are different things. A compile for
2195/// this machine has no sysroot at all, and a tree somebody laid out themselves and pointed
2196/// `--sysroot` at carries no manifest, so nothing here knows where any of it came from. Saying
2197/// nothing is the only honest answer to either, and a reader can tell it from a manifest with no
2198/// inputs in it because that one still has its header lines.
2199///
2200/// # Errors
2201///
2202/// A manifest this build cannot parse, and anything else that went wrong reading the file. Passing a
2203/// record we could not read on to whoever asked would make their parser the one that finds the
2204/// problem, and every use section 13.5 gives for these two flags is somebody else reading the
2205/// output.
2206fn sysroot_manifest(opts: &Options, link: &LinkOptions) -> Result<Option<Manifest>, CliError> {
2207    let Some(root) = sysroot_root(opts, link) else {
2208        return Ok(None);
2209    };
2210    let path = Sysroot::at(root, opts.target.tuple()).manifest_path();
2211    match std::fs::read_to_string(&path) {
2212        Ok(text) => Manifest::parse(&text)
2213            .map(Some)
2214            .map_err(|why| err(format!("{}: {why}", path.display()))),
2215        Err(why) if why.kind() == std::io::ErrorKind::NotFound => Ok(None),
2216        Err(why) => Err(err(format!("{}: {why}", path.display()))),
2217    }
2218}
2219
2220/// Renders the passes this level will run, in order, with what each one does.
2221///
2222/// The level is the whole of the answer unless a `-f` flag edited it, which is section 9.1 of
2223/// `spec/09-optimizer.md`: a level is a list somebody wrote down rather than something that
2224/// emerges from which flags happen to be set, and this is how that list is read.
2225#[must_use]
2226pub fn print_pipeline(opts: &Options) -> String {
2227    let mut settings = rucc_opt::Options::for_level(opts.opt_level);
2228    settings.toggles.clone_from(&opts.passes);
2229    settings.global_fuel = opts.pass_fuel_global;
2230    for (on, spec) in &opts.pass_gates {
2231        // Every spelling was checked while the arguments were parsed, so there is nothing here
2232        // this can refuse, and a listing is not the place to report it if there were.
2233        let _ = settings.gates.add(*on, spec);
2234    }
2235    rucc_opt::pipeline::print(&settings)
2236}
2237
2238/// Renders the resolved configuration.
2239///
2240/// One `key: value` per line, sorted by nothing in particular but fixed in order, because
2241/// this output is diffed across hosts in CI and a reordering would read as a change.
2242#[must_use]
2243pub fn print_config(opts: &Options) -> String {
2244    let sess = Session::new(opts.clone());
2245    let t = &sess.target;
2246    let mut out = String::new();
2247    let _ = writeln!(out, "version: {VERSION}");
2248    // The three field triple the driver was given rather than the ten field tuple it widens to,
2249    // because this output is what a build system reads to find out what it asked for. The tuple is
2250    // the compiler's model of the machine and this line is a receipt for a command line.
2251    let _ = writeln!(out, "target: {}", opts.target);
2252    let _ = writeln!(out, "arch: {}", opts.target.arch.as_str());
2253    let _ = writeln!(out, "os: {}", opts.target.os.as_str());
2254    let _ = writeln!(out, "env: {}", opts.target.env.as_str());
2255    let _ = writeln!(out, "object-format: {}", t.object_format.as_str());
2256    let _ = writeln!(out, "pointer-width: {}", t.pointer_width);
2257    let _ = writeln!(out, "long-width: {}", t.long_width);
2258    let _ = writeln!(out, "long-double-width: {}", t.long_double_width);
2259    let _ = writeln!(out, "endian: {}", if t.little_endian { "little" } else { "big" });
2260    let _ = writeln!(out, "char-signed: {}", t.char_is_signed);
2261    let _ = writeln!(out, "va-list: {}", t.va_list.map_or("none", |list| list.as_str()));
2262    // The register file as a count per class, which is enough to tell a target whose registers
2263    // are described from one whose are not without printing sixteen names nobody asked for.
2264    let regs: Vec<String> = t
2265        .regs
2266        .classes()
2267        .map(|(class, info)| format!("{} {}", info.name, t.regs.len(class)))
2268        .collect();
2269    let _ = writeln!(
2270        out,
2271        "registers: {}",
2272        if regs.is_empty() { "none".to_string() } else { regs.join(", ") }
2273    );
2274    // What the schedule was chosen with, which is a sentence rather than a name on purpose: two
2275    // runs of a benchmark that disagree are usually two models and not two compilers.
2276    let _ = writeln!(out, "timing-model: {}", t.timing.map_or("none", |timing| timing.model));
2277    let _ = writeln!(out, "opt-level: {}", sess.opts.opt_level);
2278    let _ = writeln!(out, "safety: {}", sess.opts.safety);
2279    let _ = writeln!(out, "emit: {}", sess.opts.emit.as_str());
2280    let _ = writeln!(out, "debug-info: {}", sess.opts.debug_info);
2281    let _ = writeln!(out, "frame-pointer: {}", sess.opts.frame_pointer);
2282    let _ = writeln!(out, "red-zone: {}", sess.opts.red_zone);
2283    let _ = writeln!(out, "stack-protector: {}", sess.opts.protector);
2284    let _ = writeln!(out, "stack-clash-protection: {}", sess.opts.stack_clash);
2285    let _ = writeln!(out, "cf-protection: {}", sess.opts.control);
2286    let _ = writeln!(out, "patchable-function-entry: {}", sess.opts.patchable);
2287    let _ = writeln!(out, "profile: {}", sess.opts.profile);
2288    let _ = writeln!(out, "profile-hook: {}", sess.opts.hook);
2289    // Last because it is the one key with more than one line under it, and the only one
2290    // whose value is a property of the machine rather than of the command line.
2291    for dir in sess.opts.search.dirs() {
2292        let system = if dir.is_system { " (system)" } else { "" };
2293        let _ = writeln!(out, "include: {}{system}", dir.path.display());
2294    }
2295    out
2296}
2297
2298/// The output name the make target is taken from, which is the `-o` argument or nothing.
2299///
2300/// A run that stops at the preprocessor has not named an object, whatever its `-o` says: under
2301/// `-E` that argument is the preprocessed text and under `-M` it is the rule itself, and neither
2302/// is a file `make` would rebuild by running this rule. GCC agrees and falls back to the source
2303/// name in both, which is why a `-MD -E -o out.i` writes `out.d` holding a rule for `a.o`. From
2304/// `-S` on the argument does name what the rule builds, and it is used as written.
2305fn deps_target_output<'a>(opts: &Options, plan: &'a Plan) -> Option<&'a str> {
2306    if opts.emit == EmitKind::Preprocessed { None } else { plan.output.as_deref() }
2307}
2308
2309/// Writes to a path the command line named rather than one the plan derived, where `-` is
2310/// standard output.
2311fn write_named(path: &str, bytes: &[u8]) -> Result<(), String> {
2312    if path == "-" {
2313        return write_out(&Output::Stdout, bytes);
2314    }
2315    write_out(&Output::File(path.to_owned()), bytes)
2316}
2317
2318/// Writes the make rule for one input, and reports whether it got there.
2319///
2320/// A rule with no file of its own goes where the compilation it replaced would have written,
2321/// which is what makes the usual makefile recipe work: `rucc -M $< -o $@` leaves the rule in
2322/// `$@`, and the same line with the `-o` left off puts it on standard output.
2323fn write_deps(
2324    opts: &Options,
2325    plan: &Plan,
2326    job: &Job,
2327    found: &[Dependency],
2328    stderr: &mut impl std::io::Write,
2329) -> bool {
2330    let targets = if opts.deps.targets.is_empty() {
2331        vec![deps::default_target(&job.input, deps_target_output(opts, plan))]
2332    } else {
2333        opts.deps.targets.clone()
2334    };
2335    let rule = deps::rule(&opts.deps, &targets, &job.input, found);
2336    // The file, on the other hand, is named after the `-o` in every mode that still has one to
2337    // spend, which is every mode except the two that spend it on the rule.
2338    let wrote = match deps::default_file(&opts.deps, &job.input, plan.output.as_deref()) {
2339        // A `-MF` on a run that had nowhere else to put the rule leaves the file the `-o`
2340        // named empty rather than absent, because a makefile that named it as a target of its
2341        // own is a makefile that will look for it.
2342        Some(path) => write_named(&path, rule.as_bytes()).and_then(|()| {
2343            if opts.deps.instead_of_compiling { write_out(&job.output, b"") } else { Ok(()) }
2344        }),
2345        None => write_out(&job.output, rule.as_bytes()),
2346    };
2347    if let Err(e) = wrote {
2348        let _ = writeln!(stderr, "rucc: error: {e}");
2349        return false;
2350    }
2351    true
2352}
2353
2354/// Runs phase 4 over every input that has one, and writes what came out.
2355///
2356/// One input that fails does not stop the others. A build that reports every file it could
2357/// not preprocess in one run is worth more than one that stops at the first, and the exit
2358/// status is still a failure either way.
2359fn preprocess_all(opts: &Options, plan: &Plan) -> i32 {
2360    let fs = OsFileSystem::new();
2361    let mut stderr = std::io::stderr().lock();
2362    let mut failed = false;
2363    for job in &plan.jobs {
2364        if !job.phases.first().is_some_and(|p| *p == Phase::Preprocess) {
2365            // An input that is already preprocessed, or an object file. GCC passes these
2366            // through untouched, and the plan has already said so in its notes.
2367            continue;
2368        }
2369        let started = std::time::Instant::now();
2370        let result = preprocess(opts, &job.input, &fs);
2371        if opts.time {
2372            say_time(&job.input, started.elapsed(), &mut stderr);
2373        }
2374        for message in &result.messages {
2375            let _ = writeln!(stderr, "{message}");
2376        }
2377        if result.failed() {
2378            failed = true;
2379            continue;
2380        }
2381        if opts.deps.emit {
2382            failed |= !write_deps(opts, plan, job, &result.deps, &mut stderr);
2383            // `-M` and `-MM` asked for the rule instead of the text, so there is nothing else
2384            // to write. The other two asked for both and fall through to the text below.
2385            if opts.deps.instead_of_compiling {
2386                continue;
2387            }
2388        }
2389        if let Err(e) = write_out(&job.output, result.text.as_bytes()) {
2390            let _ = writeln!(stderr, "rucc: error: {e}");
2391            failed = true;
2392        }
2393    }
2394    i32::from(failed)
2395}
2396
2397/// Whether this job is a file of assembly that has to be assembled and that nothing here assembles.
2398///
2399/// The phases rather than the kind, because there are two kinds of assembly input and one of them
2400/// is preprocessed first, and because an object file also has no compile phase and is not this: it
2401/// has no phases at all and goes to the linker as it is. A `.s` on a `-c` line has exactly
2402/// [`Phase::Assemble`] left, and a `.S` has the preprocessor in front of it, and neither has
2403/// anything the front end can do.
2404fn needs_an_assembler(job: &Job) -> bool {
2405    job.phases.contains(&Phase::Assemble) && !job.phases.contains(&Phase::Compile)
2406}
2407
2408/// Whether the preprocessor runs over it on the way in, which is the whole difference between the
2409/// two kinds of assembly input.
2410fn assembly_wants_cpp(job: &Job) -> bool {
2411    job.phases.contains(&Phase::Preprocess)
2412}
2413
2414/// Runs the front end over every input that has a compile phase, and writes what came out.
2415///
2416/// The same rule as [`preprocess_all`]: one input that fails does not stop the others, and the
2417/// exit status is a failure either way. An input that is already assembly or an object has no
2418/// compile phase and is passed over here, which the plan has already said in its notes.
2419fn compile_all(opts: &Options, plan: &Plan) -> i32 {
2420    let fs = OsFileSystem::new();
2421    let mut stderr = std::io::stderr().lock();
2422    let mut failed = false;
2423    let (mut remarks, ok) = Remarks::new(opts.opt_info_file.as_ref(), &mut stderr);
2424    failed |= !ok;
2425    let mut fired = Fired::new();
2426    let mut pressure = Pressure::new();
2427    let mut lowerings = Lowerings::new();
2428    for job in &plan.jobs {
2429        if !job.phases.contains(&Phase::Compile) && !needs_an_assembler(job) {
2430            continue;
2431        }
2432        // An input of IR is read back rather than compiled, since the C it came from is not
2433        // here any more. A file of assembly does not go through the front end at all and is
2434        // read by the assembler instead. Everything after this is the same for all three, so
2435        // the paths meet again at the messages and the file the result is written to.
2436        let started = std::time::Instant::now();
2437        let result = if needs_an_assembler(job) {
2438            assemble(opts, &job.input, assembly_wants_cpp(job), &fs)
2439        } else if job.kind == InputKind::Ir {
2440            compile_ir(opts, &job.input, &fs)
2441        } else {
2442            compile(opts, &job.input, &fs)
2443        };
2444        if opts.time {
2445            say_time(&job.input, started.elapsed(), &mut stderr);
2446        }
2447        fired.merge(&result.fired);
2448        pressure.merge(&result.pressure);
2449        lowerings.merge(&result.lowerings);
2450        failed |= !write_dumps(&job.input, &result.dumps, &mut stderr);
2451        failed |= !remarks.write(&result.remarks, &mut stderr);
2452        for message in &result.messages {
2453            let _ = writeln!(stderr, "{message}");
2454        }
2455        // Before the failure below, because a compilation that stopped in the back end is exactly
2456        // the one whose preprocessed source somebody wants to look at.
2457        failed |= !write_temps(job, &result.temps, &mut stderr);
2458        if result.failed() {
2459            failed = true;
2460            continue;
2461        }
2462        // `-MD` and `-MMD` write the rule beside the object and let the compilation happen, so
2463        // this is the one path where both files come out of the same run. An input of IR has no
2464        // dependencies to report and produces an empty list, which produces a rule naming only
2465        // itself, and that is the honest answer rather than a missing file.
2466        if opts.deps.emit {
2467            failed |= !write_deps(opts, plan, job, &result.deps, &mut stderr);
2468        }
2469        if let Err(e) = write_out(&job.output, result.artifact.bytes()) {
2470            let _ = writeln!(stderr, "rucc: error: {e}");
2471            failed = true;
2472        }
2473    }
2474    failed |= !write_coverage(opts, &fired, &mut stderr);
2475    failed |= !write_pressure(opts, &pressure, &mut stderr);
2476    failed |= !write_lowering(opts, &lowerings, &mut stderr);
2477    i32::from(failed)
2478}
2479
2480/// A directory for the object files only the link step ever sees, removed when it goes away.
2481///
2482/// `-c` writes its object where the user can see it and linking does not, which is the whole of
2483/// the difference: a `rucc a.c b.c` leaves an executable behind and nothing else, the same as
2484/// every other compiler. Removing them on drop rather than at the end of a function is so that a
2485/// link that failed leaves nothing behind either.
2486struct Scratch {
2487    /// Where the objects go.
2488    dir: PathBuf,
2489}
2490
2491impl Scratch {
2492    /// Makes one, under whatever the platform calls its temporary directory.
2493    ///
2494    /// The name carries the process id so that two compilers running at once do not share a
2495    /// directory, which they would otherwise do the moment two of them compiled a file of the
2496    /// same name.
2497    fn new() -> Result<Scratch, String> {
2498        let dir = std::env::temp_dir().join(format!("rucc-{}", std::process::id()));
2499        std::fs::create_dir_all(&dir).map_err(|e| format!("{}: {e}", dir.display()))?;
2500        Ok(Scratch { dir })
2501    }
2502}
2503
2504impl Drop for Scratch {
2505    fn drop(&mut self) {
2506        let _ = std::fs::remove_dir_all(&self.dir);
2507    }
2508}
2509
2510/// The link line the plan describes, for `-###`.
2511///
2512/// The names in it are the hints the plan carries rather than the temporaries a real compilation
2513/// would choose, because `-###` prints the line without having compiled anything and so has
2514/// nothing to point at. That also makes the printed line readable rather than naming a directory
2515/// that only exists while a compilation is running.
2516fn link_line(opts: &Options, link: &LinkOptions, job: &LinkJob) -> Result<String, link::Error> {
2517    let linker = link::find(opts.target, link)?;
2518    let args = link::line(opts.target, link, &job.inputs, &job.output)?;
2519    Ok(link::render(&linker, &args))
2520}
2521
2522/// Compiles everything, then links it.
2523///
2524/// The objects go in a directory that is removed afterwards, which is why this is not
2525/// [`compile_all`] followed by a link: the plan says an object feeding the linker is temporary
2526/// and does not say where, because where is a question that only has an answer once something is
2527/// running.
2528fn link_all(opts: &Options, plan: &Plan, link: &LinkOptions, verbose: bool) -> i32 {
2529    let Some(job) = &plan.link else {
2530        // Every path into here comes from a plan whose last phase is the link, and such a plan
2531        // has a link job. Saying so is cheaper than an unwrap that would have to be explained.
2532        let mut stderr = std::io::stderr().lock();
2533        let _ = writeln!(stderr, "rucc: error: there is nothing to link");
2534        return 1;
2535    };
2536    // Before anything is compiled, because a linker that is not on the machine is worth knowing
2537    // about in the second it takes to look rather than after the compilation.
2538    // And before that, whether this link has a line at all and whether what it reads is on the
2539    // machine. Both are answerable now, and a target whose sysroot has not been built is worth
2540    // saying so about before the compilation rather than after it.
2541    if let Err(why) = link::preflight(opts.target, link) {
2542        return complain(why);
2543    }
2544    let linker = match link::find(opts.target, link) {
2545        Ok(linker) => linker,
2546        Err(why) => return complain(why),
2547    };
2548    // And whether the one that was found can do this link, which for one linker and one target is
2549    // a question only the linker itself can answer. Here rather than inside the search, because
2550    // what it does is refuse rather than move on to the next candidate: nothing else in the list
2551    // links a produced Windows sysroot either.
2552    if let Err(why) = link::suitable(opts.target, &linker) {
2553        return complain(why);
2554    }
2555
2556    let scratch = match Scratch::new() {
2557        Ok(scratch) => scratch,
2558        Err(why) => return complain(format!("could not make a place for the object files: {why}")),
2559    };
2560
2561    let fs = OsFileSystem::new();
2562    let mut failed = false;
2563    // One per job, in job order, which is what lets the link line below be rebuilt with the real
2564    // paths in it: every job contributes exactly one file to the line and does so in this order.
2565    let mut produced: Vec<String> = Vec::with_capacity(plan.jobs.len());
2566    let mut fired = Fired::new();
2567    let mut pressure = Pressure::new();
2568    let mut lowerings = Lowerings::new();
2569    {
2570        let mut stderr = std::io::stderr().lock();
2571        let (mut remarks, ok) = Remarks::new(opts.opt_info_file.as_ref(), &mut stderr);
2572        failed |= !ok;
2573        for (at, job) in plan.jobs.iter().enumerate() {
2574            let out = match &job.output {
2575                Output::Temporary(hint) => {
2576                    // The index because two inputs in different directories can have the same
2577                    // name, and the two objects of `rucc a/x.c b/x.c` must not be one file.
2578                    scratch.dir.join(format!("{at}-{hint}")).display().to_string()
2579                }
2580                Output::File(path) => path.clone(),
2581                // A job feeding the linker never writes to standard output, since the plan gives
2582                // it a temporary. This is here so that the match is total rather than a panic.
2583                Output::Stdout => continue,
2584            };
2585            produced.push(out.clone());
2586            if !job.phases.contains(&Phase::Compile) && !needs_an_assembler(job) {
2587                continue;
2588            }
2589            let started = std::time::Instant::now();
2590            let result = if needs_an_assembler(job) {
2591                assemble(opts, &job.input, assembly_wants_cpp(job), &fs)
2592            } else if job.kind == InputKind::Ir {
2593                compile_ir(opts, &job.input, &fs)
2594            } else {
2595                compile(opts, &job.input, &fs)
2596            };
2597            if opts.time {
2598                say_time(&job.input, started.elapsed(), &mut stderr);
2599            }
2600            fired.merge(&result.fired);
2601            pressure.merge(&result.pressure);
2602            lowerings.merge(&result.lowerings);
2603            failed |= !write_dumps(&job.input, &result.dumps, &mut stderr);
2604            failed |= !remarks.write(&result.remarks, &mut stderr);
2605            for message in &result.messages {
2606                let _ = writeln!(stderr, "{message}");
2607            }
2608            failed |= !write_temps(job, &result.temps, &mut stderr);
2609            if result.failed() {
2610                failed = true;
2611                continue;
2612            }
2613            // A `-MD` on a command line that links writes the rule next to the executable and
2614            // names the executable as its target, since that is the file this source builds
2615            // here. The object it went through is in a temporary directory and is gone by the
2616            // time `make` reads any of this.
2617            if opts.deps.emit {
2618                failed |= !write_deps(opts, plan, job, &result.deps, &mut stderr);
2619            }
2620            if !matches!(result.artifact, Artifact::Object { .. }) {
2621                // Worth saying rather than writing whatever it is and letting the linker read it.
2622                // An empty file is a valid empty linker script, so a link handed one gets as far
2623                // as reporting every symbol of this file undefined, which is a page of messages
2624                // about something that went wrong here.
2625                let _ = writeln!(
2626                    stderr,
2627                    "rucc: internal error: {}: no object file was produced for the link",
2628                    job.input
2629                );
2630                failed = true;
2631                continue;
2632            }
2633            if let Err(e) = std::fs::write(&out, result.artifact.bytes()) {
2634                let _ = writeln!(stderr, "rucc: error: {out}: {e}");
2635                failed = true;
2636            }
2637        }
2638        failed |= !write_coverage(opts, &fired, &mut stderr);
2639        failed |= !write_pressure(opts, &pressure, &mut stderr);
2640        failed |= !write_lowering(opts, &lowerings, &mut stderr);
2641        failed |= !write_lowering(opts, &lowerings, &mut stderr);
2642    }
2643    if failed {
2644        // Nothing is linked from a compilation that did not finish. A linker run over the objects
2645        // that did compile would report every function of the file that did not as undefined,
2646        // which is a page of messages about a mistake already reported once.
2647        return 1;
2648    }
2649
2650    // The items in command line order with the temporaries filled in. A library and a word for the
2651    // linker contribute no job and pass through, and every file item takes the next job's real
2652    // output, which is what keeps whatever was written between two objects between them here.
2653    let mut outputs = produced.into_iter();
2654    let mut items = Vec::with_capacity(job.inputs.len());
2655    for item in &job.inputs {
2656        match item {
2657            link::Item::Library(name) => items.push(link::Item::Library(name.clone())),
2658            link::Item::Linker(arg) => items.push(link::Item::Linker(arg.clone())),
2659            link::Item::File(_) => match outputs.next() {
2660                Some(path) => items.push(link::Item::File(path)),
2661                None => return complain("the plan asks the linker for a file nothing produced"),
2662            },
2663        }
2664    }
2665
2666    let args = match link::line(opts.target, link, &items, &job.output) {
2667        Ok(args) => args,
2668        Err(why) => return complain(why),
2669    };
2670    if verbose {
2671        let mut stderr = std::io::stderr().lock();
2672        let _ = writeln!(stderr, "{}", link::render(&linker, &args));
2673    }
2674    let started = std::time::Instant::now();
2675    let ran = link::run(&linker, &args);
2676    if opts.time {
2677        // The one step of a compilation that really is another program, so this line is the same
2678        // measurement gcc's is and names the linker the way gcc names `collect2`.
2679        let mut stderr = std::io::stderr().lock();
2680        say_time(&linker.name, started.elapsed(), &mut stderr);
2681    }
2682    match ran {
2683        Ok(()) => 0,
2684        // The linker has already said what was wrong on its own error output, and repeating that
2685        // linking failed would only push its message further up the screen.
2686        Err(link::Error::Refused { .. }) => 1,
2687        Err(why) => complain(why),
2688    }
2689}
2690
2691/// Compiles everything and writes the objects into one static library.
2692///
2693/// No temporary directory and no second program. The objects never reach the file system at all:
2694/// they go from the compiler into the archive writer, which is both faster than writing a directory
2695/// of files for an `ar` to read back and the reason the symbol index can be written at all. A
2696/// member's index entries are the names the object writer says it wrote, and the only thing that
2697/// knows those is the run that wrote it.
2698///
2699/// `-save-temps` is the exception. It asked for the objects to be kept, the plan gave them names a
2700/// person can find, and they are written there as well as put in the archive.
2701fn archive_all(opts: &Options, plan: &Plan) -> i32 {
2702    let Some(job) = &plan.archive else {
2703        // Every path into here comes from a plan whose last phase is the archive, and such a plan
2704        // has an archive job. Saying so is cheaper than an unwrap that would have to be explained.
2705        return complain("there is nothing to put in an archive");
2706    };
2707    // Before anything is compiled, because a format this has no container for is worth knowing
2708    // about in the second it takes to look rather than after the whole compilation.
2709    let flavour = match opts.target.os.object_format() {
2710        ObjectFormat::Elf => rucc_archive::Flavour::Gnu,
2711        ObjectFormat::Coff => rucc_archive::Flavour::Coff,
2712        // Mach-O wants the BSD flavour, whose index is a different member under a different name,
2713        // and wasm has no archives of its own at all. Neither has an object writer either, so a
2714        // command line reaching this would have failed in the next step regardless.
2715        format @ (ObjectFormat::MachO | ObjectFormat::Wasm) => {
2716            return complain(format!(
2717                "there is no archive format for {} objects in this compiler yet",
2718                format.as_str()
2719            ));
2720        }
2721    };
2722
2723    let fs = OsFileSystem::new();
2724    let mut failed = false;
2725    let mut members: Vec<rucc_archive::Member> = Vec::with_capacity(plan.jobs.len());
2726    let mut names = job.members.iter();
2727    let mut fired = Fired::new();
2728    let mut pressure = Pressure::new();
2729    let mut lowerings = Lowerings::new();
2730    {
2731        let mut stderr = std::io::stderr().lock();
2732        let (mut remarks, ok) = Remarks::new(opts.opt_info_file.as_ref(), &mut stderr);
2733        failed |= !ok;
2734        for plan_job in &plan.jobs {
2735            // What the plan called this member. The two lists are walked together rather than the
2736            // name being worked out again here, so that what `-###` printed and what goes in the
2737            // file cannot come apart.
2738            let Some(member) = names.next() else {
2739                return complain("the plan asks the archive for a member nothing produced");
2740            };
2741            if !plan_job.phases.contains(&Phase::Compile) && !needs_an_assembler(plan_job) {
2742                // Neither something to compile nor something to assemble, so there is nothing to
2743                // put in, and an archive quietly missing a member is worse than a message.
2744                let _ = writeln!(
2745                    &mut stderr,
2746                    "rucc: error: {}: this compiler makes an archive out of what it compiles, and \
2747                     there is nothing here for it to do",
2748                    plan_job.input
2749                );
2750                failed = true;
2751                continue;
2752            }
2753            let started = std::time::Instant::now();
2754            let result = if needs_an_assembler(plan_job) {
2755                assemble(opts, &plan_job.input, assembly_wants_cpp(plan_job), &fs)
2756            } else if plan_job.kind == InputKind::Ir {
2757                compile_ir(opts, &plan_job.input, &fs)
2758            } else {
2759                compile(opts, &plan_job.input, &fs)
2760            };
2761            if opts.time {
2762                say_time(&plan_job.input, started.elapsed(), &mut stderr);
2763            }
2764            fired.merge(&result.fired);
2765            pressure.merge(&result.pressure);
2766            lowerings.merge(&result.lowerings);
2767            failed |= !write_dumps(&plan_job.input, &result.dumps, &mut stderr);
2768            failed |= !remarks.write(&result.remarks, &mut stderr);
2769            for message in &result.messages {
2770                let _ = writeln!(stderr, "{message}");
2771            }
2772            failed |= !write_temps(plan_job, &result.temps, &mut stderr);
2773            if result.failed() {
2774                failed = true;
2775                continue;
2776            }
2777            if opts.deps.emit {
2778                failed |= !write_deps(opts, plan, plan_job, &result.deps, &mut stderr);
2779            }
2780            let Artifact::Object { bytes, defines } = result.artifact else {
2781                let _ = writeln!(
2782                    stderr,
2783                    "rucc: internal error: {}: no object file was produced for the archive",
2784                    plan_job.input
2785                );
2786                failed = true;
2787                continue;
2788            };
2789            // Under `-save-temps` the plan gave the object a name a person can find, so it is
2790            // written there too. Otherwise it is only ever a member and never a file.
2791            if let Output::File(path) = &plan_job.output {
2792                if let Err(e) = std::fs::write(path, &bytes) {
2793                    let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2794                    failed = true;
2795                }
2796            }
2797            members.push(rucc_archive::Member { name: member.clone(), body: bytes, defines });
2798        }
2799        failed |= !write_coverage(opts, &fired, &mut stderr);
2800        failed |= !write_pressure(opts, &pressure, &mut stderr);
2801        failed |= !write_lowering(opts, &lowerings, &mut stderr);
2802        failed |= !write_lowering(opts, &lowerings, &mut stderr);
2803    }
2804    if failed {
2805        // Nothing is written from a compilation that did not finish, for the reason the link gives:
2806        // an archive missing the file that failed is one a link reports every name of as undefined,
2807        // which is a page of messages about a mistake already reported once.
2808        return 1;
2809    }
2810
2811    let bytes = match rucc_archive::write(flavour, &members) {
2812        Ok(bytes) => bytes,
2813        // Every one of these is a bug here rather than a program's mistake: the names came from the
2814        // object writer and the bodies came from this process.
2815        Err(why) => return complain(format!("the archive could not be written: {why}")),
2816    };
2817    match std::fs::write(&job.output, &bytes) {
2818        Ok(()) => 0,
2819        Err(e) => complain(format!("{}: {e}", job.output)),
2820    }
2821}
2822
2823/// Prints one driver level message and gives back the exit status that goes with it.
2824fn complain(why: impl std::fmt::Display) -> i32 {
2825    let mut stderr = std::io::stderr().lock();
2826    let _ = writeln!(stderr, "rucc: error: {why}");
2827    1
2828}
2829
2830/// Writes what `-Zrule-coverage=FILE` asked for, and says whether it could.
2831///
2832/// Once for the whole command line rather than once per input, because the question is which
2833/// lowering rules this run of the compiler reached and a file per input would leave the reader
2834/// unioning files to find out something one process already knew.
2835///
2836/// A file that could not be written is a failure and not a warning. What asks for this is a
2837/// measurement run, and a measurement that quietly did not happen is worse than one that stopped.
2838fn write_coverage(opts: &Options, fired: &Fired, stderr: &mut impl std::io::Write) -> bool {
2839    let Some(path) = &opts.rule_coverage else { return true };
2840    let Some(table) = coverage::table(opts.target.arch) else {
2841        let _ = writeln!(
2842            stderr,
2843            "rucc: error: there are no lowering rules for {} yet, so there is no coverage of them \
2844             to report",
2845            opts.target
2846        );
2847        return false;
2848    };
2849    match std::fs::write(path, fired.listing(table)) {
2850        Ok(()) => true,
2851        Err(e) => {
2852            let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2853            false
2854        }
2855    }
2856}
2857
2858/// Writes what `-Zregister-pressure=FILE` asked for, and says whether it could.
2859///
2860/// Once for the whole command line, for the reason [`write_coverage`] gives, and a file that could
2861/// not be written is a failure for the reason it gives too. There is no equivalent of the missing
2862/// rule table here, since every target this compiles for has an allocator, and a run that reached
2863/// no back end at all writes an empty listing rather than nothing: a measurement of a build that
2864/// produced no code is still an answer and it is the honest one.
2865fn write_pressure(opts: &Options, pressure: &Pressure, stderr: &mut impl std::io::Write) -> bool {
2866    let Some(path) = &opts.register_pressure else { return true };
2867    match std::fs::write(path, pressure.listing()) {
2868        Ok(()) => true,
2869        Err(e) => {
2870            let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2871            false
2872        }
2873    }
2874}
2875
2876/// Writes what `-Zlowering=FILE` asked for, and says whether it could.
2877///
2878/// Once for the whole command line, for the reason [`write_coverage`] gives, and a file that could
2879/// not be written is a failure for the reason it gives too. A run that reached no back end writes
2880/// an empty listing rather than nothing, the way [`write_pressure`] does and for the same reason.
2881fn write_lowering(opts: &Options, lowerings: &Lowerings, stderr: &mut impl std::io::Write) -> bool {
2882    let Some(path) = &opts.lowering_dump else { return true };
2883    match std::fs::write(path, lowerings.listing()) {
2884        Ok(()) => true,
2885        Err(e) => {
2886            let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2887            false
2888        }
2889    }
2890}
2891
2892/// Where the `-fopt-info` remarks go, and how much of the run has already gone there.
2893///
2894/// Standard error by default, and one file for the whole run when `-fopt-info=<file>` named one.
2895/// A file rather than the diagnostic stream is what a harness wants: the corpus in
2896/// `tamnd/rucc-corpus` matches a rejection against what the compiler said on standard error, and
2897/// a few thousand remarks mixed into that would bury it.
2898struct Remarks {
2899    /// The file, if there is one.
2900    file: Option<String>,
2901    /// Whether anything has been written to it yet, which decides between truncating and
2902    /// appending. One file holds the whole run rather than the last input in it.
2903    started: bool,
2904}
2905
2906impl Remarks {
2907    /// Prepares the destination, emptying the file if there is one.
2908    ///
2909    /// Emptied here rather than at the first remark, because a run where no pass had anything to
2910    /// say should leave an empty file and not yesterday's. An absent file and an empty one are
2911    /// different facts and something reading this will act on the difference.
2912    fn new(file: Option<&String>, stderr: &mut impl std::io::Write) -> (Self, bool) {
2913        let mut ok = true;
2914        if let Some(path) = file {
2915            if let Err(e) = std::fs::write(path, "") {
2916                let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2917                ok = false;
2918            }
2919        }
2920        (Self { file: file.cloned(), started: false }, ok)
2921    }
2922
2923    /// Writes one input's remarks, and says whether that worked.
2924    ///
2925    /// A file that cannot be written is a failure and not a warning, for the reason
2926    /// [`write_dumps`] gives: remarks that quietly did not arrive look exactly like a compilation
2927    /// where nothing happened.
2928    fn write(&mut self, text: &str, stderr: &mut impl std::io::Write) -> bool {
2929        if text.is_empty() {
2930            return true;
2931        }
2932        let Some(path) = &self.file else {
2933            let _ = write!(stderr, "{text}");
2934            return true;
2935        };
2936        let opened = std::fs::OpenOptions::new()
2937            .write(true)
2938            .append(self.started)
2939            .truncate(!self.started)
2940            .create(true)
2941            .open(path);
2942        self.started = true;
2943        let result =
2944            opened.and_then(|mut file| std::io::Write::write_all(&mut file, text.as_bytes()));
2945        if let Err(e) = result {
2946            let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2947            return false;
2948        }
2949        true
2950    }
2951}
2952
2953/// Writes what `-fdump-ir=` asked to see, one file per dump.
2954///
2955/// The name is the input file with the dump's own name and `.ir` after it, so a directory listing
2956/// after a run is the passes in the order they ran, per input. They go in the working directory
2957/// rather than beside the output, because a dump is something a person asked for at a prompt and
2958/// the working directory is where that person is.
2959///
2960/// A file that could not be written is a failure and not a warning, for the reason
2961/// [`write_coverage`] gives: what asked for this is somebody debugging a pass, and a dump that
2962/// quietly did not happen looks exactly like a pass that did not run.
2963fn write_dumps(input: &str, dumps: &[rucc_opt::Dump], stderr: &mut impl std::io::Write) -> bool {
2964    let stem = std::path::Path::new(input)
2965        .file_name()
2966        .map_or_else(|| input.to_owned(), |name| name.to_string_lossy().into_owned());
2967    let mut ok = true;
2968    for dump in dumps {
2969        let path = format!("{stem}.{}.ir", dump.name);
2970        if let Err(e) = std::fs::write(&path, &dump.text) {
2971            let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2972            ok = false;
2973        }
2974    }
2975    ok
2976}
2977
2978/// Writes the files `-save-temps` kept, which is nothing at all unless it was given.
2979///
2980/// A file that could not be written is a failure rather than a warning, for the reason
2981/// [`write_dumps`] gives: somebody asked for these by name, and one that quietly did not happen
2982/// looks like a compilation that never went through that step.
2983fn write_temps(job: &Job, temps: &Temps, stderr: &mut impl std::io::Write) -> bool {
2984    let mut ok = true;
2985    let kept = [(job.saved_text(), &temps.preprocessed), (job.saved_asm(), &temps.assembly)];
2986    for (path, text) in kept {
2987        // A step the compilation did not reach has nothing to keep, and a job that is not keeping
2988        // that step has nowhere to put it. Either way there is no file here.
2989        let (Some(path), Some(text)) = (path, text) else { continue };
2990        if let Err(e) = std::fs::write(&path, text) {
2991            let _ = writeln!(stderr, "rucc: error: {path}: {e}");
2992            ok = false;
2993        }
2994    }
2995    ok
2996}
2997
2998/// One line of `-time`, which is what a step was called and how long it took.
2999///
3000/// GCC's two numbers are the user and the system time of a subprocess it ran. This compiler runs
3001/// no subprocess for anything but the link, so what is measured here is the wall clock of the
3002/// step and the second column is always zero. The shape of the line is kept because a person
3003/// reading it next to gcc's should not have to work out which column is which.
3004fn say_time(name: &str, took: std::time::Duration, stderr: &mut impl std::io::Write) {
3005    let _ = writeln!(stderr, "# {name} {:.2} {:.2}", took.as_secs_f64(), 0.0);
3006}
3007
3008/// Writes one job's result where the plan said it goes.
3009///
3010/// # Errors
3011///
3012/// Returns the message to print, which names the file when there is one, because "permission
3013/// denied" on its own does not say which file was refused.
3014fn write_out(output: &Output, bytes: &[u8]) -> Result<(), String> {
3015    match output {
3016        Output::Stdout => {
3017            let mut stdout = std::io::stdout().lock();
3018            stdout.write_all(bytes).map_err(|e| format!("writing to standard output: {e}"))
3019        }
3020        Output::File(path) | Output::Temporary(path) => {
3021            std::fs::write(path, bytes).map_err(|e| format!("{path}: {e}"))
3022        }
3023    }
3024}
3025
3026/// Runs the driver and returns the process exit code.
3027///
3028/// `args` excludes the program name. Output goes to `stdout` and errors to `stderr`, which
3029/// is the one place in the compiler that is true.
3030pub fn run(args: &[String]) -> i32 {
3031    match parse_args(args) {
3032        Ok(Action::Help) => {
3033            print!("{USAGE}");
3034            0
3035        }
3036        Ok(Action::Version) => {
3037            println!("rucc {VERSION}");
3038            0
3039        }
3040        Ok(Action::Print(line)) => {
3041            println!("{line}");
3042            0
3043        }
3044        Ok(Action::PrintConfig(opts)) => {
3045            print!("{}", print_config(&opts));
3046            0
3047        }
3048        Ok(Action::PrintPipeline(opts)) => {
3049            print!("{}", print_pipeline(&opts));
3050            0
3051        }
3052        Ok(Action::PrintPlan { opts, plan, link }) => {
3053            print!("{}", plan.render());
3054            // The line as it would be typed, which is the half of `-###` that section 4.3 says
3055            // arrives with the link. It is printed even when the linker is not on this machine,
3056            // because what a build wants from `-###` is what the compiler would do.
3057            if let Some(job) = &plan.link {
3058                match link_line(&opts, &link, job) {
3059                    Ok(line) => println!("{line}"),
3060                    Err(why) => {
3061                        let mut stderr = std::io::stderr().lock();
3062                        let _ = writeln!(stderr, "rucc: error: {why}");
3063                        return 1;
3064                    }
3065                }
3066            }
3067            0
3068        }
3069        Ok(Action::Fetch { what, target, cache }) => fetch_sysroot(what, target, &cache),
3070        Ok(Action::Compile { opts, plan, link, jobs, verbose, notes }) => {
3071            {
3072                let mut stderr = std::io::stderr().lock();
3073                // Before the plan rather than after it, because a note is about the command line
3074                // and the plan is what the command line was read as, so the reader wants the two
3075                // in that order.
3076                for note in &notes {
3077                    let _ = writeln!(stderr, "rucc: warning: {note}");
3078                }
3079                if verbose {
3080                    let _ = write!(stderr, "{}", plan.render());
3081                    let _ = writeln!(stderr, "workers: {}", jobs.count());
3082                }
3083            }
3084            if opts.emit == EmitKind::Preprocessed {
3085                return preprocess_all(&opts, &plan);
3086            }
3087            if opts.emit == EmitKind::Archive {
3088                return archive_all(&opts, &plan);
3089            }
3090            if opts.emit != EmitKind::Executable {
3091                return compile_all(&opts, &plan);
3092            }
3093            link_all(&opts, &plan, &link, verbose)
3094        }
3095        Err(e) => {
3096            let mut stderr = std::io::stderr().lock();
3097            let _ = writeln!(stderr, "rucc: error: {e}");
3098            let _ = writeln!(stderr, "rucc: note: run `rucc --help` for usage");
3099            1
3100        }
3101    }
3102}
3103
3104#[cfg(test)]
3105mod tests {
3106    use rucc_session::{
3107        Contract, GnucVersion, IncludeForm, LtoJobs, OptLevel, Partition, Patchable, Visibility,
3108    };
3109
3110    use super::*;
3111
3112    fn args(s: &[&str]) -> Vec<String> {
3113        s.iter().map(|x| (*x).to_owned()).collect()
3114    }
3115
3116    /// A target to write down where the host would otherwise decide, for the tests whose answer
3117    /// would be a different one on a different machine.
3118    ///
3119    /// Most of the tests here never name a target, which is right, because most of what the driver
3120    /// does with a command line is the same wherever it runs and a test that pinned one would be
3121    /// saying so in every case for the sake of the two that need it. The two that need it are the
3122    /// ones whose answer comes off the target rather than off the command line: the name an object
3123    /// gets, which is `a.o` here and `a.obj` on Windows, and whether Microsoft's reading of a
3124    /// nameless member is on, which is off here and on there. Both are the compiler being right, and
3125    /// a test that leaves the target to the host is asking a question with two correct answers.
3126    const LINUX: &str = "--target=x86_64-unknown-linux-gnu";
3127
3128    #[test]
3129    fn help_and_version_win_over_everything_else() {
3130        assert_eq!(parse_args(&args(&["-c", "--help", "x.c"])).unwrap(), Action::Help);
3131        assert_eq!(parse_args(&args(&["--version"])).unwrap(), Action::Version);
3132    }
3133
3134    fn compile(s: &[&str]) -> (Box<Options>, Box<Plan>) {
3135        match parse_args(&args(s)).expect("expected a compilation") {
3136            Action::Compile { opts, plan, .. } => (opts, plan),
3137            other => panic!("expected a compilation, got {other:?}"),
3138        }
3139    }
3140
3141    fn linking(s: &[&str]) -> (Box<LinkOptions>, Box<Plan>) {
3142        match parse_args(&args(s)).expect("expected a compilation") {
3143            Action::Compile { link, plan, .. } => (link, plan),
3144            other => panic!("expected a compilation, got {other:?}"),
3145        }
3146    }
3147
3148    fn notes(s: &[&str]) -> Vec<String> {
3149        match parse_args(&args(s)).expect("expected a compilation") {
3150            Action::Compile { notes, .. } => notes,
3151            other => panic!("expected a compilation, got {other:?}"),
3152        }
3153    }
3154
3155    /// The ordinary command line has nothing to say about itself, which is the property that makes
3156    /// a note worth reading when there is one.
3157    #[test]
3158    fn a_command_line_with_nothing_wrong_with_it_carries_no_notes() {
3159        assert_eq!(notes(&["-c", "a.c"]), Vec::<String>::new());
3160    }
3161
3162    /// A directory that is not there contributes nothing to the search path, so there is no tree to
3163    /// read a release out of and nothing to compare the pin against. Said as a test because this is
3164    /// the shape a hermetic machine takes: the probe reads the disk and every other machine has a
3165    /// different disk, so what can be asserted here is the silence.
3166    #[test]
3167    fn a_named_tree_that_is_not_on_the_machine_is_not_a_release_mismatch() {
3168        let said =
3169            notes(&["--target=x86_64-linux-gnu.2.28", "--sysroot=/nowhere-at-all", "-c", "a.c"]);
3170        assert_eq!(said, Vec::<String>::new());
3171    }
3172
3173    #[test]
3174    fn collects_inputs_and_flags() {
3175        let (opts, plan) = compile(&["-c", "-O2", "-g", "a.c", "b.c"]);
3176        let paths: Vec<&str> = plan.jobs.iter().map(|j| j.input.as_str()).collect();
3177        assert_eq!(paths, vec!["a.c", "b.c"]);
3178        assert_eq!(opts.opt_level, OptLevel::O2);
3179        assert_eq!(opts.emit, EmitKind::Object);
3180        assert!(opts.debug_info);
3181    }
3182
3183    /// The unstable options, which are spelled apart from everything else on purpose: what is
3184    /// under `-Z` promises nothing, and a build that reaches for one should have had to say so.
3185    #[test]
3186    fn an_unstable_option_is_taken_and_one_that_does_not_exist_is_refused() {
3187        let (opts, _) = compile(&["-c", "-Zrule-coverage=/tmp/rules.cov", "a.c"]);
3188        assert_eq!(opts.rule_coverage.as_deref(), Some("/tmp/rules.cov"));
3189
3190        let (plain, _) = compile(&["-c", "a.c"]);
3191        assert_eq!(plain.rule_coverage, None, "nothing is measured unless it was asked for");
3192
3193        assert!(parse_args(&args(&["-Zrule-coverage=", "a.c"])).is_err(), "a file with no name");
3194        let unknown = parse_args(&args(&["-Zwhat", "a.c"])).expect_err("there is no such option");
3195        assert!(unknown.message.contains("4.11"), "{}", unknown.message);
3196    }
3197
3198    /// The other measurement written to a file, which reads the same way and fails the same way.
3199    #[test]
3200    fn where_the_register_pressure_goes_is_asked_for_the_same_way() {
3201        let (opts, _) = compile(&["-c", "-O2", "-Zregister-pressure=/tmp/spills.txt", "a.c"]);
3202        assert_eq!(opts.register_pressure.as_deref(), Some("/tmp/spills.txt"));
3203
3204        let (plain, _) = compile(&["-c", "a.c"]);
3205        assert_eq!(plain.register_pressure, None, "nothing is measured unless it was asked for");
3206
3207        assert!(parse_args(&args(&["-Zregister-pressure=", "a.c"])).is_err(), "no file named");
3208    }
3209
3210    /// The third one, which says what the pre-selection lowering group did.
3211    #[test]
3212    fn where_the_lowering_dump_goes_is_asked_for_the_same_way() {
3213        let (opts, _) = compile(&["-c", "-O2", "-Zlowering=/tmp/lowering.txt", "a.c"]);
3214        assert_eq!(opts.lowering_dump.as_deref(), Some("/tmp/lowering.txt"));
3215
3216        let (plain, _) = compile(&["-c", "a.c"]);
3217        assert_eq!(plain.lowering_dump, None, "nothing is dumped unless it was asked for");
3218
3219        assert!(parse_args(&args(&["-Zlowering=", "a.c"])).is_err(), "no file named");
3220    }
3221
3222    /// Scheduling, which has the three way answer every optimization flag has: on, off, and
3223    /// nothing said, which is whatever the optimization level asks for. The name is gcc's, and
3224    /// gcc's has a two in it because gcc has a scheduler before allocation and one after and this
3225    /// is the one after.
3226    #[test]
3227    fn scheduling_can_be_turned_on_and_off_and_left_to_the_optimization_level() {
3228        let (on, _) = compile(&["-c", "-O0", "-fschedule-insns2", "a.c"]);
3229        assert_eq!(on.schedule_insns, Some(true));
3230
3231        let (off, _) = compile(&["-c", "-O2", "-fno-schedule-insns2", "a.c"]);
3232        assert_eq!(off.schedule_insns, Some(false));
3233
3234        let (quiet, _) = compile(&["-c", "-O2", "a.c"]);
3235        assert_eq!(quiet.schedule_insns, None, "nothing said, so the level decides");
3236        assert!(quiet.opt_level.schedules(), "and at this level the level says yes");
3237
3238        let (none, _) = compile(&["-c", "a.c"]);
3239        assert!(!none.opt_level.schedules(), "at no optimization it says no");
3240    }
3241
3242    /// Whether the timing model is worth holding an instruction back over, which is a `-Z` because
3243    /// it is a question about a target's description rather than about the program being compiled.
3244    #[test]
3245    fn whether_the_timing_model_is_cycle_accurate_can_be_overridden() {
3246        let (yes, _) = compile(&["-c", "-O2", "-Zcycle-accurate-model=yes", "a.c"]);
3247        assert_eq!(yes.cycle_accurate_model, Some(true));
3248
3249        let (no, _) = compile(&["-c", "-O2", "-Zcycle-accurate-model=no", "a.c"]);
3250        assert_eq!(no.cycle_accurate_model, Some(false));
3251
3252        let (plain, _) = compile(&["-c", "-O2", "a.c"]);
3253        assert_eq!(plain.cycle_accurate_model, None, "the target's own answer stands");
3254
3255        let bad = parse_args(&args(&["-Zcycle-accurate-model=maybe", "a.c"]))
3256            .expect_err("it takes yes or no");
3257        assert!(bad.message.contains("yes or no"), "{}", bad.message);
3258    }
3259
3260    #[test]
3261    fn a_bare_dash_o_means_o1_the_way_gcc_reads_it() {
3262        let (opts, _) = compile(&["-O", "a.c"]);
3263        assert_eq!(opts.opt_level, OptLevel::O1);
3264    }
3265
3266    #[test]
3267    fn dash_x_applies_to_later_inputs_only_and_none_stops_it() {
3268        let (_, plan) = compile(&["a.o", "-x", "c", "b.txt", "-x", "none", "c.o"]);
3269        assert_eq!(plan.jobs[0].kind, InputKind::LinkerInput);
3270        assert_eq!(plan.jobs[1].kind, InputKind::C);
3271        assert_eq!(plan.jobs[2].kind, InputKind::LinkerInput);
3272    }
3273
3274    #[test]
3275    fn dash_j_reaches_the_scheduler_and_defaults_to_the_machine() {
3276        let (_, _, jobs) = match parse_args(&args(&["-j4", "a.c"])).unwrap() {
3277            Action::Compile { opts, plan, jobs, .. } => (opts, plan, jobs),
3278            other => panic!("expected a compilation, got {other:?}"),
3279        };
3280        assert_eq!(jobs.count(), 4);
3281
3282        let default = match parse_args(&args(&["a.c"])).unwrap() {
3283            Action::Compile { jobs, .. } => jobs,
3284            other => panic!("expected a compilation, got {other:?}"),
3285        };
3286        assert_eq!(default, Jobs::available());
3287        assert!(parse_args(&args(&["-j0", "a.c"])).is_err());
3288    }
3289
3290    #[test]
3291    fn triple_hash_prints_the_plan_and_runs_nothing() {
3292        let a = parse_args(&args(&["-###", "-c", "a.c"])).unwrap();
3293        let Action::PrintPlan { plan, .. } = a else { panic!("expected a plan dump") };
3294        assert!(plan.render().contains("a.c: preprocess, compile, assemble -> a.o"));
3295    }
3296
3297    #[test]
3298    fn the_flag_that_keeps_the_intermediate_files_has_three_spellings_and_two_meanings() {
3299        // The bare one is `=obj` and not `=cwd`. gcc's manual says the opposite and gcc 16 does
3300        // this, and following the compiler is what makes a build that reads either of them find
3301        // the files where they are.
3302        assert_eq!(compile(&["-c", "-save-temps", "a.c"]).0.save_temps, SaveTemps::Object);
3303        assert_eq!(compile(&["-c", "-save-temps=obj", "a.c"]).0.save_temps, SaveTemps::Object);
3304        assert_eq!(compile(&["-c", "-save-temps=cwd", "a.c"]).0.save_temps, SaveTemps::Cwd);
3305        assert_eq!(compile(&["-c", "a.c"]).0.save_temps, SaveTemps::No);
3306        // The last one on the line decides, the way it does for every other flag with an
3307        // argument, and a keyword that is neither is fatal rather than ignored: a run that kept
3308        // nothing and said nothing looks exactly like one where the files were not produced.
3309        let (opts, _) = compile(&["-c", "-save-temps", "-save-temps=cwd", "a.c"]);
3310        assert_eq!(opts.save_temps, SaveTemps::Cwd);
3311        let e = parse_args(&args(&["-c", "-save-temps=nowhere", "a.c"])).unwrap_err();
3312        assert!(e.message.contains("accepted: cwd, obj"), "{}", e.message);
3313    }
3314
3315    #[test]
3316    fn the_flag_that_times_each_step_reaches_the_options_and_changes_nothing_else() {
3317        let (opts, plan) = compile(&["-c", "-time", "a.c"]);
3318        let (plain, without) = compile(&["-c", "a.c"]);
3319        assert!(opts.time);
3320        assert!(!plain.time);
3321        // Against the same line without the flag rather than against a spelling of the object's
3322        // name, since what the object is called is the host's business and this is not about that.
3323        assert_eq!(plan.jobs[0].output, without.jobs[0].output);
3324    }
3325
3326    #[test]
3327    fn dash_x_names_what_it_accepts_when_it_does_not_know_a_language() {
3328        let e = parse_args(&args(&["-x", "fortran", "a.c"])).unwrap_err();
3329        assert!(e.message.contains("assembler-with-cpp"), "{}", e.message);
3330    }
3331
3332    /// What `--fetch` says for a target this release pins nothing for, which is every target except
3333    /// the three windows-gnu ones today.
3334    #[test]
3335    fn a_fetch_of_a_target_nothing_is_pinned_for_says_so_rather_than_reaching_the_network() {
3336        let e = parse_args(&args(&["--fetch", "x86_64-linux-musl"])).unwrap_err();
3337        assert!(e.message.contains("pins no sysroot for x86_64-linux-musl"), "{}", e.message);
3338        // And what it does pin, because a release with some rows in the table and a release with
3339        // none are two situations and the second sentence is what tells them apart.
3340        assert!(e.message.contains("x86_64-windows-gnu"), "{}", e.message);
3341        // The joined spelling is the same flag.
3342        let joined = parse_args(&args(&["--fetch=x86_64-linux-musl"])).unwrap_err();
3343        assert_eq!(joined, e);
3344    }
3345
3346    /// The two targets a release will never pin, which is a different answer from the one above.
3347    ///
3348    /// Section 13.4. A person who reads "this release pins no sysroot yet" waits for a release that
3349    /// does, and no release of this compiler can ship either of these, so the message names the
3350    /// licence that decides it and what to do instead.
3351    #[test]
3352    fn a_fetch_of_a_target_behind_a_licence_wall_says_so_rather_than_saying_not_yet() {
3353        let e = parse_args(&args(&["--fetch", "aarch64-macos"])).unwrap_err();
3354        assert!(e.message.contains("Xcode licence"), "{}", e.message);
3355        assert!(e.message.contains("there never will be"), "{}", e.message);
3356        assert!(!e.message.contains("tamnd/rucc-cross"), "{}", e.message);
3357
3358        let e = parse_args(&args(&["--fetch", "x86_64-windows-msvc"])).unwrap_err();
3359        assert!(e.message.contains("redistributed"), "{}", e.message);
3360        // The way out of this one is a target rather than a download, and it is the default already.
3361        assert!(e.message.contains("mingw-w64"), "{}", e.message);
3362        // And the mingw-w64 target next to it is ours to ship and published, so the same flag has
3363        // something to get rather than a licence to explain.
3364        let action = parse_args(&args(&["--fetch", "x86_64-windows-gnu"])).expect("it is pinned");
3365        let Action::Fetch { what, .. } = action else { panic!("{action:?}") };
3366        assert_eq!(what.tuple, "x86_64-windows-gnu");
3367    }
3368
3369    /// An Apple target on a machine with no SDK, which is section 8.6's other host.
3370    ///
3371    /// Not run on a mac, where the SDK this is about is installed and the compile is the ordinary one
3372    /// that uses it. What the reason says is asserted in `rucc_sysroot::wall` and where it is printed
3373    /// is asserted in `rucc-pp`, so what is left here is that the driver works it out and leaves it
3374    /// where the preprocessor will find it, and that neither way past the wall leaves one behind.
3375    #[test]
3376    fn an_apple_target_with_no_sdk_anywhere_carries_the_licence_rather_than_a_missing_directory() {
3377        if cfg!(target_os = "macos") || std::env::var_os("SDKROOT").is_some() {
3378            return;
3379        }
3380        let (opts, _) = compile(&["--target=aarch64-macos", "-c", "a.c"]);
3381        let why = opts.search.missing_system().expect("the wall is the reason there are none");
3382        assert!(why.contains("aarch64-macos needs a macOS SDK"), "{why}");
3383        assert!(why.contains("Xcode licence"), "{why}");
3384        assert!(why.contains("-isysroot"), "{why}");
3385
3386        // A program that includes none of the library needs none of the SDK, which is what section
3387        // 8.6 means by being able to target the platform without one, so there is nothing to explain.
3388        let (opts, _) = compile(&["--target=aarch64-macos", "-nostdinc", "-c", "a.c"]);
3389        assert_eq!(opts.search.missing_system(), None);
3390        // And naming a path is the other way through, whether or not the path is there: a mistyped
3391        // directory is a mistake to report on its own terms rather than a licence to explain.
3392        let (opts, _) = compile(&["--target=aarch64-macos", "-isysroot", "/opt/sdk", "-c", "a.c"]);
3393        assert_eq!(opts.search.missing_system(), None);
3394    }
3395
3396    /// The same wall on the compile side of an MSVC target, where the way past it is a tuple.
3397    ///
3398    /// Not run on Windows, for the same reason the one above is not run on a mac: the wall stands in
3399    /// front of an SDK this machine does not have, and a Windows machine is the kind that does. The
3400    /// driver asks `vswhere` where Visual Studio is and takes the newest kit under it, so on a box
3401    /// with the build tools installed there are headers, no wall and nothing here to be about.
3402    /// `INCLUDE` is the other way a machine has one and is the other half of the guard, since a
3403    /// person can set that anywhere while Visual Studio is only found on the platform it runs on.
3404    #[test]
3405    fn an_msvc_target_with_no_sdk_named_says_which_environment_needs_nothing_installed() {
3406        if cfg!(target_os = "windows") || std::env::var_os("INCLUDE").is_some() {
3407            return;
3408        }
3409        let (opts, _) = compile(&["--target=x86_64-windows-msvc", "-c", "a.c"]);
3410        let why = opts.search.missing_system().expect("the wall is the reason there are none");
3411        assert!(why.contains("the Windows SDK and its universal CRT"), "{why}");
3412        assert!(why.contains("mingw-w64"), "{why}");
3413        // And the mingw-w64 target has its headers from us, so nothing is missing to explain.
3414        let (opts, _) = compile(&["--target=x86_64-windows-gnu", "-c", "a.c"]);
3415        assert_eq!(opts.search.missing_system(), None);
3416    }
3417
3418    #[test]
3419    fn a_fetch_with_no_target_and_a_fetch_of_a_tuple_that_is_not_one_both_say_which() {
3420        let e = parse_args(&args(&["--fetch"])).unwrap_err();
3421        assert!(e.message.contains("--fetch requires"), "{}", e.message);
3422        let e = parse_args(&args(&["--fetch", "sparc64-solaris-gnu"])).unwrap_err();
3423        assert!(e.message.contains("--fetch sparc64-solaris-gnu"), "{}", e.message);
3424        assert!(e.message.contains("no sysroot to get"), "{}", e.message);
3425    }
3426
3427    /// Both flags on one line ask for opposite things, in either order.
3428    #[test]
3429    fn a_fetch_and_offline_together_is_a_refusal_whichever_way_round_they_are_written() {
3430        for line in [
3431            vec!["--offline", "--fetch", "x86_64-linux-musl"],
3432            vec!["--fetch", "x86_64-linux-musl", "--offline"],
3433        ] {
3434            let e = parse_args(&args(&line)).unwrap_err();
3435            assert!(e.message.contains("two opposite things"), "{}", e.message);
3436        }
3437    }
3438
3439    #[test]
3440    fn a_fetch_does_not_compile_anything_and_says_so_when_it_is_handed_a_file() {
3441        let e = parse_args(&args(&["--fetch", "x86_64-linux-musl", "a.c"])).unwrap_err();
3442        assert!(e.message.contains("compiles nothing"), "{}", e.message);
3443        assert!(e.message.contains("a.c"), "{}", e.message);
3444    }
3445
3446    /// `--offline` on its own is accepted and changes nothing, because an ordinary compile
3447    /// downloads nothing with or without it. A build that passes it everywhere is the case this is
3448    /// for, and it must not lose the compilation it was passed beside.
3449    #[test]
3450    fn offline_on_a_compilation_is_the_same_compilation() {
3451        let (opts, plan) = compile(&["-c", "--offline", "a.c"]);
3452        let (plain, without) = compile(&["-c", "a.c"]);
3453        assert_eq!(opts.target, plain.target);
3454        assert_eq!(plan.jobs.len(), without.jobs.len());
3455        assert_eq!(plan.jobs[0].output, without.jobs[0].output);
3456    }
3457
3458    #[test]
3459    fn an_unknown_flag_is_an_error_rather_than_a_shrug() {
3460        let e = parse_args(&args(&["-fno-such-thing", "a.c"])).unwrap_err();
3461        assert!(e.message.contains("unknown option"), "{}", e.message);
3462    }
3463
3464    /// `-fpermissive` and the flag that turns it back off, which a build writes beside it when
3465    /// one directory needs the older rules and the rest of the tree does not.
3466    #[test]
3467    fn permissive_reads_in_both_directions_and_the_last_one_wins() {
3468        let (opts, _) = compile(&["-c", "a.c"]);
3469        assert!(!opts.permissive, "off unless it is asked for");
3470
3471        let (opts, _) = compile(&["-c", "-fpermissive", "a.c"]);
3472        assert!(opts.permissive);
3473
3474        let (opts, _) = compile(&["-c", "-fpermissive", "-fno-permissive", "a.c"]);
3475        assert!(!opts.permissive);
3476    }
3477
3478    #[test]
3479    fn asking_for_nested_functions_is_told_why_it_is_not_coming() {
3480        let e = parse_args(&args(&["-fnested-functions", "a.c"])).unwrap_err();
3481        assert!(e.message.contains("trampoline"), "{}", e.message);
3482        assert!(parse_args(&args(&["-fno-nested-functions", "a.c"])).is_ok());
3483    }
3484
3485    #[test]
3486    fn the_flag_every_configure_script_writes_is_taken() {
3487        // All four spellings, because a build writes whichever one its macros picked and a
3488        // compiler that takes three of them is a compiler that fails on the fourth.
3489        for flag in ["-fPIC", "-fpic", "-fPIE", "-fpie"] {
3490            let (opts, _) = compile(&["-c", flag, "a.c"]);
3491            assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3492        }
3493    }
3494
3495    #[test]
3496    fn a_table_is_written_unless_the_build_says_nothing_will_walk_it() {
3497        let (opts, _) = compile(&["-c", "a.c"]);
3498        assert!(opts.unwinds(), "the default is off");
3499        let (opts, _) = compile(&["-c", "-fno-asynchronous-unwind-tables", "a.c"]);
3500        assert!(!opts.unwinds(), "the build was not taken at its word");
3501        let (opts, _) = compile(&[
3502            "-c",
3503            "-fno-asynchronous-unwind-tables",
3504            "-fasynchronous-unwind-tables",
3505            "a.c",
3506        ]);
3507        assert!(opts.unwinds(), "the last flag did not win");
3508        // The weaker request, which the same table answers, so a line that asks for a table and
3509        // against an asynchronous one gets one. That is gcc's arrangement and it turns up when a
3510        // build turns the asynchronous one off globally and a directory asks for a table back.
3511        let (opts, _) =
3512            compile(&["-c", "-fno-asynchronous-unwind-tables", "-funwind-tables", "a.c"]);
3513        assert!(opts.unwinds(), "the weaker request was dropped");
3514        let (opts, _) = compile(&["-c", "-fno-unwind-tables", "a.c"]);
3515        assert!(opts.unwinds(), "the weaker negative turned off the stronger request");
3516        let (opts, _) =
3517            compile(&["-c", "-fno-unwind-tables", "-fno-asynchronous-unwind-tables", "a.c"]);
3518        assert!(!opts.unwinds(), "both were turned off and one stayed on");
3519    }
3520
3521    #[test]
3522    fn the_flags_that_describe_what_this_compiler_already_does_are_taken() {
3523        // Every one of these is on a real build line somewhere and every one of them was an
3524        // unknown option. What they have in common is that the answer rucc gives is the answer
3525        // they ask for, so there is nothing to implement and nothing to refuse.
3526        for flag in [
3527            "-fno-common",
3528            "-fstrict-aliasing",
3529            "-fno-strict-aliasing",
3530            "-fdelete-null-pointer-checks",
3531            "-fno-delete-null-pointer-checks",
3532            "-frounding-math",
3533            "-fno-rounding-math",
3534            "-fexcess-precision=standard",
3535            "-fexcess-precision=fast",
3536            "-fexcess-precision=16",
3537            "-pipe",
3538            "-fdiagnostics-color",
3539            "-fno-diagnostics-color",
3540            "-fdiagnostics-color=always",
3541            "-fdiagnostics-color=never",
3542            "-fdiagnostics-color=auto",
3543        ] {
3544            let (opts, _) = compile(&["-c", flag, "a.c"]);
3545            assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3546        }
3547    }
3548
3549    #[test]
3550    fn whether_an_exception_is_looked_at_is_kept_and_defaults_to_gccs_answer() {
3551        let (opts, _) = compile(&["-c", "a.c"]);
3552        assert!(opts.trapping_math, "the default was not gcc's");
3553        let (opts, _) = compile(&["-c", "-fno-trapping-math", "a.c"]);
3554        assert!(!opts.trapping_math);
3555        let (opts, _) = compile(&["-c", "-ftrapping-math", "a.c"]);
3556        assert!(opts.trapping_math, "spelling out the default turned it off");
3557        // The last one written wins, which is how a build line that inherits a flag from one
3558        // place and overrides it in another is read.
3559        let (opts, _) = compile(&["-c", "-fno-trapping-math", "-ftrapping-math", "a.c"]);
3560        assert!(opts.trapping_math);
3561    }
3562
3563    /// The flags a torture program writes on its own `dg-options` line, which is where most of
3564    /// these come from: a program reduced from a miscompilation names the pass that miscompiled
3565    /// it. Eighteen programs in the suite stopped on the driver before anything read them, and
3566    /// tamnd/rucc#1019 is the list.
3567    #[test]
3568    fn the_flags_that_name_a_pass_of_gccs_own_are_taken_and_dropped() {
3569        for flag in [
3570            "-fno-tree-ccp",
3571            "-fno-tree-dominator-opts",
3572            "-fno-tree-vrp",
3573            "-fno-tree-bit-ccp",
3574            "-fno-tree-coalesce-vars",
3575            "-ftree-vectorize",
3576            "-ftree-loop-distribution",
3577            "-fipa-pta",
3578            "-fmodulo-sched",
3579            "-fno-vect-cost-model",
3580            "-fvect-cost-model=unlimited",
3581            "-fsimd-cost-model=cheap",
3582            "-fexpensive-optimizations",
3583            "-fno-early-inlining",
3584            "-fno-inline",
3585            "-finline-functions",
3586            "-foptimize-strlen",
3587            "-fno-ira-share-spill-slots",
3588        ] {
3589            let (opts, _) = compile(&["-c", flag, "a.c"]);
3590            assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3591            assert!(opts.passes.is_empty(), "{flag} named a pass of gcc's and not one of ours");
3592        }
3593    }
3594
3595    /// The two namespaces are taken whole, so a name neither this test nor gcc 16 has heard of
3596    /// goes the same way as the ones above rather than stopping a build on the day gcc adds it.
3597    #[test]
3598    fn a_pass_name_in_either_family_is_taken_whether_or_not_it_is_one_gcc_has() {
3599        for flag in ["-ftree-no-such-pass", "-fno-ipa-no-such-pass"] {
3600            let (opts, _) = compile(&["-c", flag, "a.c"]);
3601            assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3602        }
3603    }
3604
3605    /// A pass this compiler has keeps its flag, since the arms that read the registry are above
3606    /// the family arms. `dce` is the one both compilers have a name for, and `execute/pr97421-2.c`
3607    /// is the program that writes it.
3608    #[test]
3609    fn a_pass_name_this_compiler_has_is_still_read_as_a_pass() {
3610        let (opts, _) = compile(&["-c", "-fno-dce", "a.c"]);
3611        assert_eq!(opts.passes, vec![("dce".to_owned(), false)]);
3612    }
3613
3614    /// gcc's name for the unroller reaches the unroller, in both directions. libtommath puts
3615    /// `-funroll-loops` in `CFLAGS` unconditionally, and before this it was an unknown option and
3616    /// the build stopped on its first file.
3617    #[test]
3618    fn the_gcc_spelling_of_the_unroller_turns_the_unroller_on_and_off() {
3619        let (opts, _) = compile(&["-c", "-funroll-loops", "a.c"]);
3620        assert_eq!(opts.passes, vec![("unroll".to_owned(), true)]);
3621        let (opts, _) = compile(&["-c", "-fno-unroll-loops", "a.c"]);
3622        assert_eq!(opts.passes, vec![("unroll".to_owned(), false)]);
3623    }
3624
3625    /// The three transformations that are a module at a time are named by a flag as well, even
3626    /// though none of them is a `rucc_opt::Pass` and so none is reached by the generic arms.
3627    ///
3628    /// A bisection over a miscompilation turns one thing off at a time, and a transformation with
3629    /// no spelling of its own cannot be the one turned off.
3630    #[test]
3631    fn the_transformations_that_are_not_passes_are_still_named_by_a_flag() {
3632        let (opts, _) = compile(&["-c", "-fno-ipa-cp", "-fipa-sra", "-fno-libcall", "a.c"]);
3633        assert_eq!(
3634            opts.passes,
3635            vec![
3636                (rucc_opt::ipcp::NAME.to_owned(), false),
3637                (rucc_opt::ipasra::NAME.to_owned(), true),
3638                (rucc_opt::libcall::NAME.to_owned(), false),
3639            ]
3640        );
3641        let (opts, _) = compile(&["-c", "-flibcall", "a.c"]);
3642        assert_eq!(opts.passes, vec![(rucc_opt::libcall::NAME.to_owned(), true)]);
3643    }
3644
3645    /// Where a function starts is a question this compiler answers, so the flag that asks about it
3646    /// is answered rather than dropped. femtolisp's Makefile writes the bare form on every compile
3647    /// of the project, and before this it was an unknown option and the build stopped on its first
3648    /// file. The numbers are gcc 16's, read off `-S` on x86-64: nothing and the bare form both
3649    /// give `.p2align 4`, `=32` gives 5, `=3` gives 2, and the negative form gives `.align 8`.
3650    #[test]
3651    fn the_alignment_of_a_function_is_a_request_this_compiler_can_answer() {
3652        let (opts, _) = compile(&["-c", "-falign-functions", "a.c"]);
3653        assert_eq!(opts.align_functions, None, "the bare form asks for the default");
3654
3655        let (opts, _) = compile(&["-c", "-falign-functions=32", "a.c"]);
3656        assert_eq!(opts.align_functions, Some(32));
3657
3658        let (opts, _) = compile(&["-c", "-falign-functions=3", "a.c"]);
3659        assert_eq!(opts.align_functions, Some(4), "rounded up rather than refused");
3660
3661        let (opts, _) = compile(&["-c", "-falign-functions=32:8", "a.c"]);
3662        assert_eq!(opts.align_functions, Some(32), "the boundary is the answerable half");
3663
3664        for flag in ["-falign-functions=0", "-falign-functions=1"] {
3665            let (opts, _) = compile(&["-c", flag, "a.c"]);
3666            assert_eq!(opts.align_functions, None, "{flag} means the default");
3667        }
3668
3669        let (opts, _) = compile(&["-c", "-fno-align-functions", "a.c"]);
3670        assert_eq!(opts.align_functions, Some(8), "the smallest boundary the target has");
3671
3672        // The last one on the line wins, which is how gcc reads a repeated flag.
3673        let (opts, _) = compile(&["-c", "-falign-functions=32", "-falign-functions", "a.c"]);
3674        assert_eq!(opts.align_functions, None);
3675
3676        let e = parse_args(&args(&["-c", "-falign-functions=big", "a.c"])).unwrap_err();
3677        assert!(e.message.contains("number of bytes"), "{}", e.message);
3678    }
3679
3680    /// The other three of the family are about padding inside a body, which nothing here writes,
3681    /// so they are taken and say nothing. Every spelling of each, since a build writes whichever
3682    /// one its author typed.
3683    #[test]
3684    fn the_alignment_flags_about_the_inside_of_a_body_are_taken_and_say_nothing() {
3685        for flag in [
3686            "-falign-labels",
3687            "-falign-loops",
3688            "-falign-jumps",
3689            "-falign-loops=16",
3690            "-falign-labels=32",
3691            "-fno-align-loops",
3692            "-fno-align-labels",
3693            "-fno-align-jumps",
3694        ] {
3695            let (opts, _) = compile(&["-c", flag, "a.c"]);
3696            assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3697            assert_eq!(opts.align_functions, None, "{flag} is not about where a function starts");
3698        }
3699    }
3700
3701    /// The encoding of the source is not a question about speed, so the one name that describes
3702    /// what the preprocessor does is taken and every other name is refused.
3703    #[test]
3704    fn the_input_charset_is_taken_when_it_names_the_one_that_is_read() {
3705        for flag in ["-finput-charset=utf-8", "-finput-charset=UTF-8", "-finput-charset=utf8"] {
3706            let (opts, _) = compile(&["-c", flag, "a.c"]);
3707            assert_eq!(opts.emit, EmitKind::Object, "{flag}");
3708        }
3709
3710        let e = parse_args(&args(&["-c", "-finput-charset=latin1", "a.c"])).unwrap_err();
3711        assert!(e.message.contains("latin1"), "{}", e.message);
3712        assert!(e.message.contains("UTF-8"), "what is read is worth saying: {}", e.message);
3713    }
3714
3715    /// The other half of the same rule. Each of these changes what the program does rather than
3716    /// how fast it does it, so each is refused with the reason, and the negative of each is what
3717    /// happens anyway and is taken.
3718    #[test]
3719    fn the_three_that_change_the_answer_are_refused_and_their_negatives_are_taken() {
3720        for (flag, word) in [
3721            ("-ffast-math", "__FAST_MATH__"),
3722            ("-fnon-call-exceptions", "landing pad"),
3723            ("-finstrument-functions", "__cyg_profile_func_enter"),
3724        ] {
3725            let e = parse_args(&args(&["-c", flag, "a.c"])).unwrap_err();
3726            assert!(e.message.contains(word), "{flag}: {}", e.message);
3727            assert!(!e.message.contains("unknown option"), "{flag} deserves a reason");
3728
3729            let off = format!("-fno-{}", flag.trim_start_matches("-f"));
3730            let (opts, _) = compile(&["-c", &off, "a.c"]);
3731            assert_eq!(opts.emit, EmitKind::Object, "{off}");
3732        }
3733    }
3734
3735    #[test]
3736    fn asking_the_linker_to_merge_tentative_definitions_is_told_why_it_is_not_coming() {
3737        // The one of that family that is a request rather than a description, and it is a real
3738        // difference: two files each writing `int g;` link under it and do not without it.
3739        let e = parse_args(&args(&["-fcommon", "a.c"])).unwrap_err();
3740        assert!(e.message.contains(".bss"), "{}", e.message);
3741        assert!(e.message.contains("extern"), "the way out is worth saying: {}", e.message);
3742    }
3743
3744    #[test]
3745    fn asking_for_position_dependent_code_is_told_why_it_is_not_coming() {
3746        for flag in ["-fno-pic", "-fno-pie"] {
3747            let e = parse_args(&args(&[flag, "a.c"])).unwrap_err();
3748            assert!(e.message.contains("global offset table"), "{flag}: {}", e.message);
3749            // The one it may have meant, since the two are a letter apart and one of them is
3750            // about linking and is taken.
3751            assert!(e.message.contains("-no-pie"), "{flag}: {}", e.message);
3752        }
3753    }
3754
3755    #[test]
3756    fn an_unsupported_target_names_itself() {
3757        let e = parse_args(&args(&["--target=sparc64-linux-gnu", "a.c"])).unwrap_err();
3758        assert!(e.message.contains("sparc64"), "{}", e.message);
3759    }
3760
3761    #[test]
3762    fn no_inputs_is_an_error_but_print_config_needs_none() {
3763        assert!(parse_args(&args(&[])).is_err());
3764        assert!(matches!(parse_args(&args(&["--print-config"])), Ok(Action::PrintConfig(_))));
3765    }
3766
3767    #[test]
3768    fn print_config_reports_the_target_it_was_given_not_the_host() {
3769        let a = parse_args(&args(&["--print-config", "--target=riscv64-linux-musl"])).unwrap();
3770        let Action::PrintConfig(opts) = a else { panic!("expected a configuration dump") };
3771        let text = print_config(&opts);
3772        assert!(text.contains("target: riscv64-unknown-linux-musl"), "{text}");
3773        assert!(text.contains("char-signed: false"), "{text}");
3774        assert!(text.contains("object-format: elf"), "{text}");
3775        assert!(text.contains("va-list: void-pointer"), "{text}");
3776        // RISC-V has a register file and this compiler has not written it down yet, and the
3777        // dump says which of those two it is rather than leaving the line out.
3778        assert!(text.contains("registers: none"), "{text}");
3779        assert!(text.contains("timing-model: none"), "{text}");
3780    }
3781
3782    /// The model the schedule was chosen with, which is a receipt anybody comparing two runs of a
3783    /// benchmark needs: two numbers that disagree are usually two models and not two compilers.
3784    #[test]
3785    fn print_config_names_the_model_the_schedule_was_chosen_with() {
3786        let opts = Options::new("x86_64-unknown-linux-gnu".parse().unwrap());
3787        let text = print_config(&opts);
3788        let line = text.lines().find(|l| l.starts_with("timing-model:")).expect("the model");
3789        assert!(line.contains("Skylake"), "{line}");
3790        assert!(line.contains("published"), "a sentence saying where it came from: {line}");
3791    }
3792
3793    #[test]
3794    fn print_config_has_one_key_per_line_and_a_fixed_order() {
3795        let opts = Options::new("x86_64-unknown-linux-gnu".parse().unwrap());
3796        let text = print_config(&opts);
3797        let keys: Vec<&str> =
3798            text.lines().map(|l| l.split(':').next().unwrap_or_default()).collect();
3799        assert_eq!(keys[0], "version");
3800        assert_eq!(keys[1], "target");
3801        assert_eq!(keys.len(), 26);
3802        assert!(text.ends_with('\n'));
3803    }
3804
3805    #[test]
3806    fn the_safety_tier_is_read_off_the_command_line_and_a_wrong_one_is_refused() {
3807        let (opts, _) = compile(&["a.c"]);
3808        assert_eq!(opts.safety, rucc_session::Safety::Off);
3809
3810        for (flag, tier) in [
3811            ("-fsafety=detect", rucc_session::Safety::Detect),
3812            ("-fsafety=enforce", rucc_session::Safety::Enforce),
3813            ("-fsafety=kernel", rucc_session::Safety::Kernel),
3814            ("-fsafety=off", rucc_session::Safety::Off),
3815        ] {
3816            let (opts, _) = compile(&[flag, "a.c"]);
3817            assert_eq!(opts.safety, tier, "{flag}");
3818        }
3819
3820        // The last one wins, the way every other repeated flag on this command line does.
3821        let (opts, _) = compile(&["-fsafety=enforce", "-fsafety=off", "a.c"]);
3822        assert_eq!(opts.safety, rucc_session::Safety::Off);
3823
3824        // A misspelled tier is refused rather than ignored. Silently compiling without the
3825        // monitor a build asked for is the one failure mode this feature cannot have.
3826        let e = parse_args(&args(&["-fsafety=on", "a.c"])).unwrap_err();
3827        assert!(e.message.contains("is not a safety tier"), "{}", e.message);
3828        assert!(parse_args(&args(&["-fsafety", "a.c"])).is_err());
3829    }
3830
3831    #[test]
3832    fn the_padding_mode_is_read_off_the_command_line_and_a_wrong_one_is_refused() {
3833        // The default is the one section 9.3 of document 09 gives library code, which is that
3834        // padding does not participate, so a record filled a member at a time is not reported.
3835        let (opts, _) = compile(&["a.c"]);
3836        assert_eq!(opts.padding, rucc_session::Padding::Ignored);
3837
3838        let (opts, _) = compile(&["-fsafety=detect", "-fsafety-init=padding", "a.c"]);
3839        assert_eq!(opts.padding, rucc_session::Padding::Tracked);
3840
3841        let (opts, _) = compile(&["-fsafety-init=padding", "-fsafety-init=nopadding", "a.c"]);
3842        assert_eq!(opts.padding, rucc_session::Padding::Ignored);
3843
3844        // The tier is still a tier. A flag whose name starts the same way must not be eaten by
3845        // the one above it, which is the thing worth pinning about a pair of names like these.
3846        let (opts, _) = compile(&["-fsafety-init=padding", "a.c"]);
3847        assert_eq!(opts.safety, rucc_session::Safety::Off);
3848
3849        let e = parse_args(&args(&["-fsafety-init=some", "a.c"])).unwrap_err();
3850        assert!(e.message.contains("is not a padding mode"), "{}", e.message);
3851    }
3852
3853    #[test]
3854    fn whether_a_write_has_to_stay_inside_its_member_is_read_off_the_command_line() {
3855        // Off by default, because a store to allocated storage sets its effective type and C 6.5
3856        // lets a program reuse a buffer as something else. Row S4 is a build opting out of that.
3857        let (opts, _) = compile(&["a.c"]);
3858        assert_eq!(opts.subobject, rucc_session::Subobject::Off);
3859
3860        let (opts, _) = compile(&["-fsafety=detect", "-fsafety-subobject", "a.c"]);
3861        assert_eq!(opts.subobject, rucc_session::Subobject::Members);
3862
3863        let (opts, _) = compile(&["-fsafety-subobject", "-fno-safety-subobject", "a.c"]);
3864        assert_eq!(opts.subobject, rucc_session::Subobject::Off);
3865
3866        // It takes no value. The form that would take one is the strict reading of section 9.4,
3867        // which is not written yet, so say so rather than accept a spelling that does nothing.
3868        let e = parse_args(&args(&["-fsafety-subobject=strict", "a.c"])).unwrap_err();
3869        assert!(e.message.contains("tamnd/rucc#967"), "{}", e.message);
3870    }
3871
3872    #[test]
3873    fn whether_two_restrict_pointers_may_meet_is_read_off_the_command_line() {
3874        // Off by default, because the record a block keeps is the union of what each pointer
3875        // reached, so two pointers striding through one array without landing on the same byte are
3876        // reported and by the letter of the standard those are different objects. Row Y8 is a build
3877        // deciding it would rather know.
3878        let (opts, _) = compile(&["a.c"]);
3879        assert_eq!(opts.promise, rucc_session::Promise::Off);
3880
3881        let (opts, _) = compile(&["-fsafety=detect", "-fsafety-restrict", "a.c"]);
3882        assert_eq!(opts.promise, rucc_session::Promise::Blocks);
3883
3884        let (opts, _) = compile(&["-fsafety-restrict", "-fno-safety-restrict", "a.c"]);
3885        assert_eq!(opts.promise, rucc_session::Promise::Off);
3886
3887        // The tier is still a tier, which is the thing worth pinning about a pair of names where
3888        // one is the front of the other.
3889        let (opts, _) = compile(&["-fsafety-restrict", "a.c"]);
3890        assert_eq!(opts.safety, rucc_session::Safety::Off);
3891
3892        let e = parse_args(&args(&["-fsafety-restrict=blocks", "a.c"])).unwrap_err();
3893        assert!(e.message.contains("takes no value"), "{}", e.message);
3894    }
3895
3896    #[test]
3897    fn safety_races_takes_a_mode_and_defaults_to_watching_nothing() {
3898        // Three modes rather than a bare flag, because section 9.5 gives two answers that record
3899        // the same thing and report different classes, so a flag with no value could not say which
3900        // was wanted. Off by default for the reason on `rucc_session::Races`, which is not a cost
3901        // argument: this is the one plane where an edge nobody interposed costs a false report.
3902        let (opts, _) = compile(&["a.c"]);
3903        assert_eq!(opts.races, rucc_session::Races::Off);
3904
3905        let (opts, _) = compile(&["-fsafety-races=metadata", "a.c"]);
3906        assert_eq!(opts.races, rucc_session::Races::Metadata);
3907
3908        let (opts, _) = compile(&["-fsafety-races=pointer", "a.c"]);
3909        assert_eq!(opts.races, rucc_session::Races::Pointer);
3910
3911        // Last one wins, as it does for every other mode flag here.
3912        let (opts, _) = compile(&["-fsafety-races=pointer", "-fno-safety-races", "a.c"]);
3913        assert_eq!(opts.races, rucc_session::Races::Off);
3914
3915        let e = parse_args(&args(&["-fsafety-races=all", "a.c"])).unwrap_err();
3916        assert!(e.message.contains("off, metadata or pointer"), "{}", e.message);
3917    }
3918
3919    #[test]
3920    fn print_pipeline_answers_with_the_passes_the_level_asked_for() {
3921        let a = parse_args(&args(&["--print-pipeline", "-O2"])).unwrap();
3922        let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3923        let text = print_pipeline(&opts);
3924        assert!(text.starts_with("level: -O2\n"), "{text}");
3925        assert!(text.contains("fold"), "{text}");
3926
3927        let a = parse_args(&args(&["--print-pipeline"])).unwrap();
3928        let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3929        // Two passes run at `-O0` and neither is an optimization. The first moves what
3930        // `__builtin_expect` said onto the branch and takes the instruction away, so that nothing
3931        // past the optimizer has to know the instruction exists. The second removes code nothing
3932        // reaches. See issue 359.
3933        assert!(print_pipeline(&opts).contains("1: expect,"), "{}", print_pipeline(&opts));
3934        assert!(print_pipeline(&opts).contains("2: simplify-cfg,"), "{}", print_pipeline(&opts));
3935
3936        let a = parse_args(&args(&["--print-pipeline", "-fno-simplify-cfg"])).unwrap();
3937        let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3938        // The second turns off and the first does not, because nothing below the optimizer lowers
3939        // what it removes, so `-fno-expect` is a compile that stops rather than one that runs.
3940        let text = print_pipeline(&opts);
3941        assert!(text.contains("1: expect,"), "{text}");
3942        assert!(!text.contains("simplify-cfg"), "{text}");
3943    }
3944
3945    #[test]
3946    fn print_pipeline_takes_the_toggles_into_account() {
3947        let a = parse_args(&args(&["--print-pipeline", "-O2", "-fno-fold"])).unwrap();
3948        let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3949        let text = print_pipeline(&opts);
3950        // The one that was named is gone and the rest of the level is not, which is the whole
3951        // of what a toggle promises.
3952        assert!(!text.contains("fold"), "{text}");
3953        assert!(text.contains("dce"), "{text}");
3954
3955        // Every pass the compiler has, named off. Built from the registry rather than written
3956        // out, so a pass added later is turned off here too and this keeps testing the thing it
3957        // is about, which is that the toggles can empty a level down to the passes that are not
3958        // optional. Those are named, because a listing that is all of them is a level nobody
3959        // emptied and the assertion would pass while saying nothing.
3960        let mut off = vec!["--print-pipeline".to_owned(), "-O2".to_owned()];
3961        off.extend(rucc_opt::PASSES.iter().map(|p| format!("-fno-{}", p.name())));
3962        let spelled: Vec<&str> = off.iter().map(String::as_str).collect();
3963        let a = parse_args(&args(&spelled)).unwrap();
3964        let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3965        let text = print_pipeline(&opts);
3966        let left: Vec<&str> =
3967            rucc_opt::PASSES.iter().filter(|p| p.required()).map(|p| p.name()).collect();
3968        assert_eq!(left, vec!["expect"], "{text}");
3969        for (at, name) in left.iter().enumerate() {
3970            assert!(text.contains(&format!("{}: {name},", at + 1)), "{text}");
3971        }
3972        assert!(!text.contains("dce"), "{text}");
3973    }
3974
3975    #[test]
3976    fn print_pipeline_says_when_a_budget_will_stop_the_run_short() {
3977        let a = parse_args(&args(&["--print-pipeline", "-O2"])).unwrap();
3978        let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3979        assert!(!print_pipeline(&opts).contains("global fuel"));
3980
3981        let a = parse_args(&args(&["--print-pipeline", "-O2", "-fpass-fuel-global=4"])).unwrap();
3982        let Action::PrintPipeline(opts) = a else { panic!("expected a pipeline dump") };
3983        let text = print_pipeline(&opts);
3984        // Because the listing is the answer to what this compilation will do, and a run that
3985        // stops after four rewrites is not doing what the level says it does.
3986        assert!(text.contains("global fuel: 4"), "{text}");
3987    }
3988
3989    /// A pass is turned on and off by its own name, and the order the flags were given in is
3990    /// kept, because the last spelling of a name is the one that decides.
3991    #[test]
3992    fn a_pass_is_named_by_dash_f_and_unnamed_by_dash_f_no() {
3993        let (opts, _) = compile(&["-c", "-O0", "-ffold", "-fno-fold", "-ffold", "a.c"]);
3994        assert_eq!(
3995            opts.passes,
3996            [("fold".to_owned(), true), ("fold".to_owned(), false), ("fold".to_owned(), true)]
3997        );
3998
3999        let e = parse_args(&args(&["-fno-such-pass", "a.c"])).unwrap_err();
4000        assert!(e.message.contains("unknown option"), "{}", e.message);
4001    }
4002
4003    #[test]
4004    fn pass_fuel_names_a_pass_and_a_count_and_refuses_anything_else() {
4005        let (opts, _) = compile(&["-c", "-O2", "-fpass-fuel=fold=3", "a.c"]);
4006        assert_eq!(opts.pass_fuel, [("fold".to_owned(), 3)]);
4007
4008        let e = parse_args(&args(&["-fpass-fuel=fold", "a.c"])).unwrap_err();
4009        assert!(e.message.contains("<pass>=<count>"), "{}", e.message);
4010        let e = parse_args(&args(&["-fpass-fuel=nosuch=3", "a.c"])).unwrap_err();
4011        assert!(e.message.contains("--print-pipeline"), "{}", e.message);
4012        let e = parse_args(&args(&["-fpass-fuel=fold=lots", "a.c"])).unwrap_err();
4013        assert!(e.message.contains("not a number"), "{}", e.message);
4014    }
4015
4016    #[test]
4017    fn global_pass_fuel_is_a_count_on_its_own_and_defaults_to_no_limit() {
4018        let (opts, _) = compile(&["-c", "-O2", "a.c"]);
4019        assert_eq!(opts.pass_fuel_global, None);
4020
4021        let (opts, _) = compile(&["-c", "-O2", "-fpass-fuel-global=12", "a.c"]);
4022        assert_eq!(opts.pass_fuel_global, Some(12));
4023        // And it is not the per pass flag with a longer name, so neither spelling swallows the
4024        // other.
4025        assert!(opts.pass_fuel.is_empty());
4026
4027        let e = parse_args(&args(&["-fpass-fuel-global=lots", "a.c"])).unwrap_err();
4028        assert!(e.message.contains("not a number"), "{}", e.message);
4029    }
4030
4031    #[test]
4032    fn a_gate_names_a_pass_and_optionally_the_functions_it_covers() {
4033        let (opts, _) = compile(&["-c", "-O2", "-fdisable-fold", "-fenable-fold=2-4,main", "a.c"]);
4034        assert_eq!(
4035            opts.pass_gates,
4036            [(false, "fold".to_owned()), (true, "fold=2-4,main".to_owned())],
4037            "the order is what decides, so it has to survive the parse"
4038        );
4039
4040        let e = parse_args(&args(&["-fdisable-nosuch", "a.c"])).unwrap_err();
4041        assert!(e.message.contains("--print-pipeline"), "{}", e.message);
4042        let e = parse_args(&args(&["-fenable-fold=9-2", "a.c"])).unwrap_err();
4043        assert!(e.message.contains("ends before it starts"), "{}", e.message);
4044        let e = parse_args(&args(&["-fdisable-fold=", "a.c"])).unwrap_err();
4045        assert!(e.message.contains("is empty"), "{}", e.message);
4046    }
4047
4048    #[test]
4049    fn the_pipeline_listing_says_which_passes_a_gate_touched() {
4050        let (opts, _) = compile(&["-c", "-O2", "-fdisable-fold=main", "a.c"]);
4051        let text = print_pipeline(&opts);
4052        assert!(text.contains("fold, "), "{text}");
4053        assert!(text.contains("[off for main]"), "{text}");
4054    }
4055
4056    /// The spelling is checked while the arguments are read, because a dump that names a pass
4057    /// this compiler does not have is a typo, and a typo found after the compilation has run is
4058    /// found too late to be any use.
4059    #[test]
4060    fn a_dump_is_checked_when_it_is_asked_for_rather_than_when_it_is_taken() {
4061        let (opts, _) = compile(&["-c", "-O2", "-fdump-ir=all", "-fdump-ir=after-fold", "a.c"]);
4062        assert_eq!(opts.dump_ir, ["all", "after-fold"]);
4063
4064        let e = parse_args(&args(&["-fdump-ir=after-nosuch", "a.c"])).unwrap_err();
4065        assert!(e.message.contains("nosuch"), "{}", e.message);
4066        assert!(parse_args(&args(&["-fdump-ir=sideways-fold", "a.c"])).is_err());
4067    }
4068
4069    /// Every spelling `-fopt-info` takes, and the one it does not.
4070    ///
4071    /// The keywords are checked here for the same reason a dump's pass name is: a person who
4072    /// misspelled one gets no output, and no output is also what a compilation where nothing
4073    /// happened looks like. Telling those two apart is the entire reason to reach for this flag.
4074    #[test]
4075    fn opt_info_takes_kinds_and_a_file_and_refuses_a_kind_it_does_not_have() {
4076        let (opts, _) = compile(&["-c", "-O2", "-fopt-info", "a.c"]);
4077        assert_eq!(opts.opt_info, [""], "a bare flag asks for the rewrites");
4078        assert_eq!(opts.opt_info_file, None, "and goes to standard error");
4079
4080        let (opts, _) = compile(&["-c", "-O2", "-fopt-info-missed-note", "a.c"]);
4081        assert_eq!(opts.opt_info, ["missed-note"]);
4082
4083        // Two flags add up rather than the second replacing the first, and the file is the last
4084        // one that named a file, which is how GCC treats both.
4085        let (opts, _) =
4086            compile(&["-c", "-O2", "-fopt-info-missed=one.txt", "-fopt-info-all=two.txt", "a.c"]);
4087        assert_eq!(opts.opt_info, ["missed", "all"]);
4088        assert_eq!(opts.opt_info_file.as_deref(), Some("two.txt"));
4089
4090        let e = parse_args(&args(&["-fopt-info-vectorized", "a.c"])).unwrap_err();
4091        assert!(e.message.contains("vectorized"), "{}", e.message);
4092        assert!(e.message.contains("`missed`"), "{}", e.message);
4093        let e = parse_args(&args(&["-fopt-info-missed=", "a.c"])).unwrap_err();
4094        assert!(e.message.contains("no file"), "{}", e.message);
4095    }
4096
4097    #[test]
4098    fn verify_each_is_unstable_and_off_unless_it_was_asked_for() {
4099        let (opts, _) = compile(&["-c", "-Zverify-each", "a.c"]);
4100        assert!(opts.verify_each);
4101        assert!(!USAGE.contains("verify-each"), "an unstable option stays out of the usage text");
4102    }
4103
4104    #[test]
4105    fn dash_o_needs_an_argument() {
4106        let e = parse_args(&args(&["a.c", "-o"])).unwrap_err();
4107        assert_eq!(e.message, "-o requires an argument");
4108    }
4109
4110    #[test]
4111    fn dash_d_and_dash_u_are_read_joined_or_separated_and_keep_their_order() {
4112        let (opts, _) = compile(&["-DFOO=1", "-D", "BAR", "-UBAZ", "-U", "QUX", "a.c"]);
4113        assert_eq!(opts.defines, ["FOO=1", "BAR"]);
4114        assert_eq!(opts.undefines, ["BAZ", "QUX"]);
4115    }
4116
4117    #[test]
4118    fn the_include_flags_land_on_the_chain_each_one_names() {
4119        // A sysroot with nothing under it, so that the library's own directories are the
4120        // same on every machine this test runs on, which is none of them.
4121        let (opts, _) = compile(&[
4122            "-Ii",
4123            "-iquote",
4124            "q",
4125            "-isystem",
4126            "sys",
4127            "-idirafter",
4128            "after",
4129            "--sysroot=/nowhere-at-all",
4130            "a.c",
4131        ]);
4132        let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
4133        // The compiler's own headers sit after every `-isystem` and before `-idirafter`,
4134        // which is where GCC puts its own: a directory the user named outranks ours.
4135        assert_eq!(dirs, ["q", "i", "sys", runtime::DIR, "after"]);
4136        assert!(!opts.search.dirs()[1].is_system);
4137        assert!(opts.search.dirs()[2].is_system);
4138    }
4139
4140    #[test]
4141    fn the_librarys_headers_come_after_the_compilers_own_and_go_away_with_them() {
4142        // Which machine this runs on decides what is on the path, so the test is about the
4143        // order rather than about the names: ours is on it, the library's follow it, and
4144        // `-nostdinc` is the one flag that takes both halves of the pair off at once.
4145        let (opts, _) = compile(&["a.c"]);
4146        let dirs = opts.search.dirs();
4147        let ours = dirs.iter().position(|d| d.path.to_str() == Some(runtime::DIR));
4148        assert_eq!(ours, Some(0), "{dirs:?}");
4149        assert!(dirs[1..].iter().all(|d| d.is_system), "{dirs:?}");
4150        let (bare, _) = compile(&["-nostdinc", "a.c"]);
4151        assert!(bare.search.dirs().is_empty(), "{:?}", bare.search.dirs());
4152    }
4153
4154    #[test]
4155    fn a_sysroot_moves_the_librarys_directories_and_nothing_else() {
4156        let (opts, _) = compile(&["-isystem", "sys", "--sysroot=/nowhere-at-all", "a.c"]);
4157        let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
4158        assert_eq!(dirs, ["sys", runtime::DIR]);
4159    }
4160
4161    #[test]
4162    fn a_cross_compile_reads_the_targets_own_headers_rather_than_the_ones_next_door() {
4163        // The target is not the machine this test runs on wherever it runs, so the answer is the
4164        // same on all of them: the libc's two include directories for that target, the kernel's
4165        // two, and nothing from here. A header read from here is the quiet failure of section 8.5, a
4166        // program that builds on the build machine and is wrong everywhere else.
4167        let (opts, _) = compile(&["--target=riscv64-linux-musl", "-c", "a.c"]);
4168        let dirs: Vec<&std::path::Path> =
4169            opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
4170        let root = cache::dir().join("sysroots").join("riscv64-linux-musl");
4171        let kernel = cache::dir().join("kernel-headers");
4172        assert_eq!(dirs.len(), 5, "{dirs:?}");
4173        assert_eq!(dirs[0], std::path::Path::new(runtime::DIR));
4174        assert_eq!(dirs[1], root.join("include").join("riscv64"));
4175        assert_eq!(dirs[2], root.join("include").join("generic"));
4176        // The kernel's, which are beside the sysroots rather than inside one, because every target
4177        // that shares an architecture reads the same files.
4178        assert_eq!(dirs[3], kernel.join("riscv"));
4179        assert_eq!(dirs[4], kernel.join("generic"));
4180    }
4181
4182    #[test]
4183    fn a_cross_compile_to_something_that_is_not_linux_reads_no_kernel_headers() {
4184        // The other side of the same answer. Windows has its own system headers and no `linux/` at
4185        // all, so the list is the libc's own and the question never arises, which is the `None` that
4186        // `link::cross_kernel` returns rather than a directory nothing would be found in.
4187        //
4188        // The libc's own is one directory rather than two here, because mingw-w64 publishes a single
4189        // header tree for every architecture and `Sysroot::splits_by_arch` says so.
4190        let (opts, _) = compile(&["--target=x86_64-pc-windows-gnu", "-c", "a.c"]);
4191        let dirs: Vec<&std::path::Path> =
4192            opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
4193        assert_eq!(dirs.len(), 2, "{dirs:?}");
4194        assert!(!dirs.iter().any(|dir| dir.ends_with("kernel-headers")), "{dirs:?}");
4195    }
4196
4197    #[test]
4198    fn the_glibc_version_macro_goes_with_the_bundled_tree_and_with_nothing_else() {
4199        // One tree serves every glibc release, so the release is what the target supplies, and the
4200        // condition is the same one that chose the directories. A host glibc and a tree somebody
4201        // named both define `__GLIBC_MINOR__` in their own `features.h`, and two definitions with
4202        // different values is a warning on every compilation of every file.
4203        //
4204        // The architecture is chosen against this machine's rather than written down, because the
4205        // bundled tree is only in effect for a target that is not this machine. The first version of
4206        // this test said x86_64-linux-gnu, which is a cross compile on a mac and this machine on a
4207        // Linux runner, so it passed here and failed there.
4208        let gnu = format!("--target={}-linux-gnu", cross_arch());
4209        let (bundled, _) = compile(&[&gnu, "-c", "a.c"]);
4210        assert_eq!(bundled.glibc_minor, Some(44));
4211        let pin = format!("{gnu}.2.28");
4212        let (pinned, _) = compile(&[&pin, "-c", "a.c"]);
4213        assert_eq!(pinned.glibc_minor, Some(28));
4214
4215        let (named, _) = compile(&[&gnu, "--sysroot=/nowhere-at-all", "-c", "a.c"]);
4216        assert_eq!(named.glibc_minor, None);
4217        let (none, _) = compile(&[&gnu, "-nostdinc", "-c", "a.c"]);
4218        assert_eq!(none.glibc_minor, None);
4219        let musl = format!("--target={}-linux-musl", cross_arch());
4220        let (musl, _) = compile(&[&musl, "-c", "a.c"]);
4221        assert_eq!(musl.glibc_minor, None);
4222
4223        // And this machine's own target gets nothing, whatever this machine is, because its headers
4224        // come from the machine and its own `features.h` defines the macro. On a glibc Linux box
4225        // that is the case this test had backwards; on a mac it is true for the other reason, which
4226        // is that Darwin is not a glibc target at all.
4227        if let Some(host) = Triple::host() {
4228            let native = format!("--target={}", host.tuple());
4229            let (native, _) = compile(&[&native, "-c", "a.c"]);
4230            assert_eq!(native.glibc_minor, None);
4231        }
4232    }
4233
4234    #[test]
4235    fn a_pinned_release_on_this_machines_own_target_reads_the_bundled_tree() {
4236        // The end to end half of the answer in `link::cross_for`. A release named for this machine's
4237        // own target is a cross compile, so the headers are the bundled tree's and the macro says
4238        // what was asked for rather than what this machine has.
4239        //
4240        // Only on a glibc box, because a release is a glibc release: a mac has no `__GLIBC_MINOR__`
4241        // to get wrong and nothing to pin. That makes this a test the Linux runners carry, which is
4242        // where the case lives.
4243        let Some(host) = Triple::host() else { return };
4244        if host.env != rucc_target::Env::Gnu {
4245            return;
4246        }
4247        let pin = format!("--target={}.2.28", host.tuple());
4248        let (opts, _) = compile(&[&pin, "-c", "a.c"]);
4249        assert_eq!(opts.glibc_minor, Some(28));
4250        let root = cache::dir().join("sysroots").join(format!("{}.2.28", host.tuple()));
4251        let dirs: Vec<&std::path::Path> =
4252            opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
4253        assert!(dirs.iter().any(|dir| dir.starts_with(&root)), "{dirs:?}");
4254        // And nothing of this machine's, which is the failure this was: a program compiled against
4255        // 2.44 declarations and told it was 2.28.
4256        assert!(!dirs.iter().any(|dir| *dir == std::path::Path::new("/usr/include")), "{dirs:?}");
4257    }
4258
4259    /// An architecture that is not this machine's, out of the three the driver has targets for.
4260    ///
4261    /// A test about the bundled sysroot has to name a target that is not the host, because a target
4262    /// that is the host reads the host's own headers and libraries. Asking which machine this is
4263    /// beats picking a row and hoping, and it is two lines.
4264    fn cross_arch() -> &'static str {
4265        match Triple::host().map(|host| host.arch) {
4266            Some(rucc_target::Arch::X86_64) => "aarch64",
4267            _ => "x86_64",
4268        }
4269    }
4270
4271    #[test]
4272    fn a_glibc_newer_than_the_bundled_tree_is_refused_by_name() {
4273        // Both versions in the message, because the two things a person can do about it are pin a
4274        // release the tree has and name a sysroot that has the one they asked for, and neither is a
4275        // choice they can make without knowing which release the tree is.
4276        //
4277        // Not this machine's architecture, for the reason the test above gives: the refusal is about
4278        // the bundled tree, and the bundled tree is not what a target that is this machine reads.
4279        let target = format!("--target={}-linux-gnu.2.99", cross_arch());
4280        let message = refused(&[&target, "-c", "a.c"]);
4281        assert!(message.contains("asked for glibc 2.99"), "{message}");
4282        assert!(message.contains("bundled headers are glibc 2.44"), "{message}");
4283        assert!(message.contains("--sysroot"), "{message}");
4284    }
4285
4286    #[test]
4287    fn a_sysroot_the_user_named_is_still_what_a_cross_compile_reads() {
4288        // The tree somebody assembled beats the one we would build, on the headers as on the
4289        // libraries. It is empty here, which is why the list comes out short: the directories under
4290        // it are checked for rather than assumed, and a tree that is not there offers nothing.
4291        let (opts, _) =
4292            compile(&["--target=riscv64-linux-musl", "--sysroot=/nowhere-at-all", "-c", "a.c"]);
4293        let dirs: Vec<&std::path::Path> =
4294            opts.search.dirs().iter().map(|d| d.path.as_path()).collect();
4295        assert_eq!(dirs, [std::path::Path::new(runtime::DIR)]);
4296    }
4297
4298    #[test]
4299    fn dash_i_dash_moves_the_bracket_directories_into_the_quoted_chain() {
4300        let (opts, _) =
4301            compile(&["-Iinc1", "-iquote", "inc2", "-I-", "-Iinc3", "-nostdinc", "a.c"]);
4302        let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
4303        assert_eq!(dirs, ["inc1", "inc2", "inc3"]);
4304        // An angled include sees only what came after the flag.
4305        assert_eq!(opts.search.start(IncludeForm::Angled), 2);
4306        assert!(!opts.search.searches_current_dir());
4307    }
4308
4309    #[test]
4310    fn the_prefix_flags_stick_what_iprefix_said_on_the_front_of_what_follows_it() {
4311        let (opts, _) = compile(&[
4312            "-iprefix",
4313            "/tools/",
4314            "-iwithprefix",
4315            "late",
4316            "-iwithprefixbefore",
4317            "early",
4318            "-iprefix",
4319            "/other/",
4320            "-iwithprefix",
4321            "last",
4322            "-nostdinc",
4323            "a.c",
4324        ]);
4325        let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
4326        // `-iwithprefixbefore` is an `-I` and the other two are `-isystem`, which is where GCC
4327        // puts them rather than where its manual says it does.
4328        assert_eq!(dirs, ["/tools/early", "/tools/late", "/other/last"]);
4329        assert!(!opts.search.dirs()[0].is_system);
4330        assert!(opts.search.dirs()[1].is_system);
4331    }
4332
4333    #[test]
4334    fn the_files_named_on_the_command_line_keep_their_order_and_which_flag_named_them() {
4335        let (opts, _) =
4336            compile(&["-include", "one.h", "-imacros", "two.h", "-include", "3.h", "a.c"]);
4337        let names: Vec<&str> = opts.preincludes.iter().map(|p| p.name.as_str()).collect();
4338        assert_eq!(names, ["one.h", "two.h", "3.h"]);
4339        assert_eq!(opts.preincludes.iter().filter(|p| p.macros_only).count(), 1);
4340    }
4341
4342    #[test]
4343    fn nostdinc_takes_the_compilers_own_headers_off_the_path() {
4344        let (opts, _) = compile(&["-Ii", "-nostdinc", "a.c"]);
4345        let dirs: Vec<&str> = opts.search.dirs().iter().filter_map(|d| d.path.to_str()).collect();
4346        assert_eq!(dirs, ["i"]);
4347    }
4348
4349    #[test]
4350    fn the_dialect_flags_set_the_language_and_the_extensions_separately() {
4351        let (opts, _) = compile(&["-std=gnu11", "a.c"]);
4352        assert_eq!(opts.std, Std::C11);
4353        assert!(opts.gnu_extensions);
4354
4355        let (opts, _) = compile(&["-std=iso9899:1999", "a.c"]);
4356        assert_eq!(opts.std, Std::C99);
4357        assert!(!opts.gnu_extensions);
4358
4359        let (opts, _) = compile(&["-ansi", "a.c"]);
4360        assert_eq!(opts.std, Std::C89);
4361        assert!(!opts.gnu_extensions);
4362
4363        let e = parse_args(&args(&["-std=c94jr", "a.c"])).unwrap_err();
4364        assert!(e.message.contains("unknown dialect"), "{}", e.message);
4365    }
4366
4367    #[test]
4368    fn the_dump_letters_are_a_family_and_everything_else_beginning_with_d_is_not() {
4369        let (opts, _) = compile(&["-dM", "a.c"]);
4370        assert!(opts.dumps.macros);
4371
4372        // Packed, the way GCC takes them, and a letter in the family we have not written yet
4373        // is accepted and does nothing rather than failing a build.
4374        let (opts, _) = compile(&["-dDM", "a.c"]);
4375        assert!(opts.dumps.macros);
4376        let (opts, _) = compile(&["-dD", "a.c"]);
4377        assert!(!opts.dumps.macros);
4378
4379        let (opts, _) = compile(&["a.c"]);
4380        assert!(!opts.dumps.any());
4381
4382        // `-dumpversion` is a different flag that happens to start the same way, and it is read
4383        // as itself rather than as a dump of nothing.
4384        assert_eq!(printed(&["-dumpversion", "a.c"]), VERSION);
4385    }
4386
4387    #[test]
4388    fn the_gcc_version_claimed_is_a_flag_and_the_short_spellings_are_the_ones_people_write() {
4389        let (opts, _) = compile(&["a.c"]);
4390        assert_eq!(
4391            opts.gnuc,
4392            GnucVersion { major: 16, minor: 0, patch: 0 },
4393            "the release this compiler is written against, and the earliest one of that series"
4394        );
4395
4396        let (opts, _) = compile(&["-fgnuc-version=15.1.0", "a.c"]);
4397        assert_eq!(opts.gnuc, GnucVersion { major: 15, minor: 1, patch: 0 });
4398
4399        // A missing component is zero. `gcc -dumpversion` says `15` on a release with no
4400        // patchlevel and a harness that pastes that back has to be understood.
4401        let (opts, _) = compile(&["-fgnuc-version=15", "a.c"]);
4402        assert_eq!(opts.gnuc, GnucVersion { major: 15, minor: 0, patch: 0 });
4403
4404        let (opts, _) = compile(&["-fgnuc-version=13.2", "a.c"]);
4405        assert_eq!(opts.gnuc, GnucVersion { major: 13, minor: 2, patch: 0 });
4406
4407        let e = parse_args(&args(&["-fgnuc-version=15.x", "a.c"])).unwrap_err();
4408        assert!(e.message.contains("minor that is not a number"), "{}", e.message);
4409
4410        let e = parse_args(&args(&["-fgnuc-version=1.2.3.4", "a.c"])).unwrap_err();
4411        assert!(e.message.contains("more than three"), "{}", e.message);
4412    }
4413
4414    #[test]
4415    fn pedantic_has_two_spellings_and_is_not_the_same_knob_as_the_dialect() {
4416        let (opts, _) = compile(&["-std=c17", "-pedantic", "a.c"]);
4417        assert!(opts.pedantic);
4418        assert_eq!(opts.std, Std::C17);
4419
4420        // The `-W` family's name for it, which is what a build that groups its warning flags
4421        // tends to write.
4422        let (opts, _) = compile(&["-Wpedantic", "a.c"]);
4423        assert!(opts.pedantic);
4424
4425        let (opts, _) = compile(&["-std=c17", "a.c"]);
4426        assert!(!opts.pedantic, "a dialect on its own does not diagnose an extension");
4427    }
4428
4429    #[test]
4430    fn dash_p_and_dash_ffreestanding_reach_the_options() {
4431        let (opts, _) = compile(&["-E", "-P", "-ffreestanding", "a.c"]);
4432        assert!(!opts.line_markers);
4433        assert!(!opts.hosted);
4434        assert_eq!(opts.emit, EmitKind::Preprocessed);
4435    }
4436
4437    /// The two ways a build says it means its own function by a name the C library also has.
4438    ///
4439    /// `-fno-builtin` is all of them and `-fno-builtin-<name>` is one, and the second is what a
4440    /// build writes when it means its own `memcpy` and the library's everything else. The name is
4441    /// kept as it was written and not checked against anything, because a program is allowed to
4442    /// mean something by a name this compiler has never heard of.
4443    #[test]
4444    fn the_builtin_flags_are_read_in_both_directions_and_one_name_at_a_time() {
4445        let (opts, _) = compile(&["-c", "a.c"]);
4446        assert!(opts.builtins, "a library name means the library function by default");
4447        assert!(opts.no_builtin.is_empty());
4448
4449        let (opts, _) = compile(&["-c", "-fno-builtin", "a.c"]);
4450        assert!(!opts.builtins);
4451
4452        let (opts, _) = compile(&["-c", "-fno-builtin", "-fbuiltin", "a.c"]);
4453        assert!(opts.builtins, "the last mention decides");
4454
4455        let (opts, _) = compile(&["-c", "-fno-builtin-memcpy", "-fno-builtin-nonesuch", "a.c"]);
4456        assert!(opts.builtins, "one name is not the family");
4457        assert_eq!(opts.no_builtin, vec!["memcpy".to_owned(), "nonesuch".to_owned()]);
4458    }
4459
4460    /// `-fvisibility=`, which is on every cmake project that cares about which names it exports
4461    /// and which was refused as an unknown option until now.
4462    ///
4463    /// Four spellings and three answers. `internal` is hidden plus a promise about never taking
4464    /// the address across a component boundary, and nothing derives anything from that promise
4465    /// here, so it comes out as the weaker of the two rather than as a refusal that stops a build
4466    /// over a distinction this compiler does not make.
4467    #[test]
4468    fn visibility_takes_the_four_spellings_gcc_takes_and_refuses_the_rest() {
4469        let (opts, _) = compile(&["-c", "a.c"]);
4470        assert_eq!(opts.visibility, Visibility::Default, "exported unless something says not");
4471
4472        for (written, wanted) in [
4473            ("default", Visibility::Default),
4474            ("hidden", Visibility::Hidden),
4475            ("internal", Visibility::Hidden),
4476            ("protected", Visibility::Protected),
4477        ] {
4478            let (opts, _) = compile(&["-c", &format!("-fvisibility={written}"), "a.c"]);
4479            assert_eq!(opts.visibility, wanted, "{written}");
4480        }
4481
4482        // The last mention decides, which is what every other flag of this shape does and what a
4483        // build that turns something off for one directory relies on.
4484        let (opts, _) = compile(&["-c", "-fvisibility=hidden", "-fvisibility=default", "a.c"]);
4485        assert_eq!(opts.visibility, Visibility::Default, "the last mention decides");
4486
4487        // A spelling gcc does not take is refused rather than read as the default, because a
4488        // build that meant hidden and got exported is a library with the wrong interface and
4489        // nothing said about it anywhere.
4490        let failed = parse_args(&args(&["-fvisibility=none", "a.c"])).expect_err("refused");
4491        assert!(failed.to_string().contains("is not a visibility"), "{failed}");
4492    }
4493
4494    /// `-ffp-contract=`, which is the one flag in the floating point group that is kept rather than
4495    /// described, and the values are gcc 16's three.
4496    #[test]
4497    fn how_far_a_multiply_and_an_addition_may_be_fused_is_asked_for() {
4498        let (opts, _) = compile(&["-c", "a.c"]);
4499        assert_eq!(opts.fp_contract, Contract::Off, "a licence nobody granted is not assumed");
4500
4501        for (written, wanted) in
4502            [("off", Contract::Off), ("on", Contract::On), ("fast", Contract::Fast)]
4503        {
4504            let (opts, _) = compile(&["-c", &format!("-ffp-contract={written}"), "a.c"]);
4505            assert_eq!(opts.fp_contract, wanted, "{written}");
4506        }
4507
4508        let (opts, _) = compile(&["-c", "-ffp-contract=fast", "-ffp-contract=off", "a.c"]);
4509        assert_eq!(opts.fp_contract, Contract::Off, "the last mention decides");
4510
4511        // Refused rather than read as one of the three, because a build that asked for no fusing
4512        // and was given the default would be one whose numbers change and whose command line says
4513        // they should not. gcc refuses the same spellings and names the same three in its message.
4514        for bad in ["-ffp-contract=none", "-ffp-contract=", "-ffp-contract=Fast"] {
4515            let failed = parse_args(&args(&[bad, "a.c"])).expect_err("refused");
4516            assert!(failed.to_string().contains("is not a contraction"), "{bad}: {failed}");
4517        }
4518
4519        // And the other one that takes a value, which is taken and kept nowhere: every operation
4520        // here is computed in the type it was written in, so `standard` is what happens and the
4521        // other two are permission to do something this does not do.
4522        let failed = parse_args(&args(&["-fexcess-precision=long", "a.c"])).expect_err("refused");
4523        assert!(failed.to_string().contains("is not an excess precision"), "{failed}");
4524    }
4525
4526    /// The four prefix mapping flags, which are what a distribution passes to get the same bytes
4527    /// out of `/build/pkg-1.2` and out of `/home/someone/pkg-1.2`. Three lists rather than one
4528    /// because gcc has three, and `-ffile-prefix-map=` is the three of them at once.
4529    #[test]
4530    fn a_prefix_mapping_flag_goes_on_the_list_its_spelling_names() {
4531        let (opts, _) = compile(&["-c", "a.c"]);
4532        assert!(opts.prefix_map.macros.is_empty(), "nothing is rewritten unless it is asked for");
4533        assert!(opts.prefix_map.debug.is_empty(), "nor here");
4534        assert!(opts.prefix_map.profile.is_empty(), "nor here");
4535
4536        let (opts, _) = compile(&["-c", "-fmacro-prefix-map=/build=.", "a.c"]);
4537        assert_eq!(opts.prefix_map.macros.apply("/build/a.c"), "./a.c", "the one it names");
4538        assert!(opts.prefix_map.debug.is_empty(), "and not the two it does not");
4539
4540        let (opts, _) = compile(&["-c", "-fdebug-prefix-map=/build=.", "a.c"]);
4541        assert_eq!(opts.prefix_map.debug.apply("/build/a.c"), "./a.c", "the one it names");
4542        assert!(opts.prefix_map.macros.is_empty(), "and not the two it does not");
4543
4544        let (opts, _) = compile(&["-c", "-fprofile-prefix-map=/build=.", "a.c"]);
4545        assert_eq!(opts.prefix_map.profile.apply("/build/a.c"), "./a.c", "the one it names");
4546        assert!(opts.prefix_map.macros.is_empty(), "and not the two it does not");
4547
4548        let (opts, _) = compile(&["-c", "-ffile-prefix-map=/build=.", "a.c"]);
4549        for list in [&opts.prefix_map.macros, &opts.prefix_map.debug, &opts.prefix_map.profile] {
4550            assert_eq!(list.apply("/build/a.c"), "./a.c", "all three at once");
4551        }
4552
4553        // Every mention is kept and the last one that matches wins, unlike the flags above whose
4554        // last mention replaces the earlier ones. A build writes one of these per source root and
4555        // expects all of them to be in force, which is the whole point of a list.
4556        let (opts, _) =
4557            compile(&["-c", "-ffile-prefix-map=/a=one", "-ffile-prefix-map=/b=two", "a.c"]);
4558        assert_eq!(opts.prefix_map.macros.apply("/a/x.c"), "one/x.c", "the earlier one still acts");
4559        assert_eq!(opts.prefix_map.macros.apply("/b/x.c"), "two/x.c", "and so does the later one");
4560
4561        // An argument with no `=` is refused rather than ignored, because a build whose paths were
4562        // meant to be rewritten and were not is one that ships the build directory's name and says
4563        // nothing about it. gcc refuses the same thing.
4564        for bad in ["-fmacro-prefix-map=nope", "-ffile-prefix-map=", "-fdebug-prefix-map=/build"] {
4565            let failed = parse_args(&args(&[bad, "a.c"])).expect_err("refused");
4566            assert!(failed.to_string().contains("is not a rewrite for"), "{bad}: {failed}");
4567        }
4568    }
4569
4570    /// `-ffunction-sections` and `-fdata-sections`, which are what make `--gc-sections` able to
4571    /// drop anything: a linker can leave out a section nothing reaches and cannot leave out half of
4572    /// one. A kernel and an embedded image are both linked that way.
4573    ///
4574    /// Two flags rather than one because gcc has two, and a build that asks for one of them and not
4575    /// the other is a build that measured something: splitting the code is nearly free at link time
4576    /// and splitting the data can defeat the linker's ordering of what is next to what.
4577    #[test]
4578    fn a_section_per_function_and_a_section_per_variable_are_asked_for_one_at_a_time() {
4579        let (opts, _) = compile(&["-c", "a.c"]);
4580        assert!(!opts.function_sections, "one text section unless something says otherwise");
4581        assert!(!opts.data_sections);
4582
4583        let (opts, _) = compile(&["-c", "-ffunction-sections", "a.c"]);
4584        assert!(opts.function_sections);
4585        assert!(!opts.data_sections, "one flag is not the other");
4586
4587        let (opts, _) = compile(&["-c", "-fdata-sections", "a.c"]);
4588        assert!(opts.data_sections);
4589        assert!(!opts.function_sections);
4590
4591        // Both directions taken, and the off one is what happens anyway rather than a refusal,
4592        // since a build that writes it is asking for the default.
4593        let (opts, _) = compile(&[
4594            "-c",
4595            "-ffunction-sections",
4596            "-fno-function-sections",
4597            "-fdata-sections",
4598            "-fno-data-sections",
4599            "a.c",
4600        ]);
4601        assert!(!opts.function_sections, "the last mention decides");
4602        assert!(!opts.data_sections, "the last mention decides");
4603    }
4604
4605    /// `-fgnu89-inline`, which is off by default and is not implied by anything on the command
4606    /// line, since the dialect asks for GNU's reading further in rather than through this.
4607    #[test]
4608    fn gnu89_inline_is_off_until_it_is_asked_for_and_the_last_mention_decides() {
4609        let (opts, _) = compile(&["-c", "a.c"]);
4610        assert!(!opts.gnu89_inline, "C's reading of inline by default");
4611
4612        let (opts, _) = compile(&["-c", "-fgnu89-inline", "a.c"]);
4613        assert!(opts.gnu89_inline);
4614
4615        let (opts, _) = compile(&["-c", "-fgnu89-inline", "-fno-gnu89-inline", "a.c"]);
4616        assert!(!opts.gnu89_inline, "the last mention decides");
4617
4618        // The C89 dialects are under GNU's reading whether this was written or not, so the flag
4619        // stays off there and the dialect is what the checker and the macro set both ask. That is
4620        // also why `-std=c89 -fno-gnu89-inline` needs no diagnostic: it asks for the reading the
4621        // dialect already has. gcc refuses that command line, which is measured in the issue.
4622        let (opts, _) = compile(&["-c", "-std=c89", "a.c"]);
4623        assert!(!opts.gnu89_inline);
4624    }
4625
4626    /// Both spellings of both frame flags, since a build that wants one usually writes the
4627    /// other beside it for the one file that has to be compiled the ordinary way.
4628    #[test]
4629    fn the_two_frame_flags_are_read_in_both_directions() {
4630        let (opts, _) = compile(&["-c", "a.c"]);
4631        assert!(!opts.frame_pointer, "gcc omits it above -O0 and so does this");
4632        assert!(opts.red_zone, "the psABI has one and nothing said not to use it");
4633
4634        let (opts, _) = compile(&["-c", "-fno-omit-frame-pointer", "-mno-red-zone", "a.c"]);
4635        assert!(opts.frame_pointer);
4636        assert!(!opts.red_zone);
4637
4638        let (opts, _) = compile(&[
4639            "-c",
4640            "-fno-omit-frame-pointer",
4641            "-fomit-frame-pointer",
4642            "-mno-red-zone",
4643            "-mred-zone",
4644            "a.c",
4645        ]);
4646        assert!(!opts.frame_pointer, "the last one wins, as it does in gcc");
4647        assert!(opts.red_zone);
4648    }
4649
4650    /// Four flags rather than one with an argument, which is how gcc spells them, and the negative
4651    /// spelled three ways because a build that turns one off writes whichever it turned on.
4652    #[test]
4653    fn the_stack_protector_is_four_flags_and_the_last_one_wins() {
4654        let (opts, _) = compile(&["-c", "a.c"]);
4655        assert_eq!(opts.protector, Protector::None, "gcc protects nothing unless it was asked");
4656
4657        for (flag, want) in [
4658            ("-fstack-protector", Protector::Buffers),
4659            ("-fstack-protector-strong", Protector::Strong),
4660            ("-fstack-protector-all", Protector::All),
4661        ] {
4662            let (opts, _) = compile(&["-c", flag, "a.c"]);
4663            assert_eq!(opts.protector, want, "{flag}");
4664        }
4665
4666        // What a package build does: the strong one in the global flags and one directory that
4667        // cannot have a protector turning it off on the line after.
4668        for off in ["-fno-stack-protector", "-fno-stack-protector-strong"] {
4669            let (opts, _) = compile(&["-c", "-fstack-protector-strong", off, "a.c"]);
4670            assert_eq!(opts.protector, Protector::None, "{off}");
4671        }
4672        let (opts, _) = compile(&["-c", "-fno-stack-protector", "-fstack-protector-all", "a.c"]);
4673        assert_eq!(opts.protector, Protector::All, "the last one wins either way round");
4674    }
4675
4676    /// A switch rather than a level, because how a frame is taken is one question and which
4677    /// functions get a canary is another, and gcc spells it that way for the same reason.
4678    #[test]
4679    fn taking_a_frame_a_page_at_a_time_is_off_until_it_is_asked_for() {
4680        let (opts, _) = compile(&["-c", "a.c"]);
4681        assert!(!opts.stack_clash, "gcc takes a frame in one subtraction unless it was asked");
4682
4683        let (opts, _) = compile(&["-c", "-fstack-clash-protection", "a.c"]);
4684        assert!(opts.stack_clash);
4685
4686        // The same shape a package build uses for the protector: on in the global flags and off
4687        // for the one directory that cannot have it.
4688        let (opts, _) =
4689            compile(&["-c", "-fstack-clash-protection", "-fno-stack-clash-protection", "a.c"]);
4690        assert!(!opts.stack_clash);
4691        let (opts, _) =
4692            compile(&["-c", "-fno-stack-clash-protection", "-fstack-clash-protection", "a.c"]);
4693        assert!(opts.stack_clash, "the last one wins either way round");
4694
4695        // The two are independent, since one is about the frame and the other about the function.
4696        let (opts, _) =
4697            compile(&["-c", "-fstack-clash-protection", "-fstack-protector-strong", "a.c"]);
4698        assert!(opts.stack_clash);
4699        assert_eq!(opts.protector, Protector::Strong);
4700    }
4701
4702    /// One flag with an argument rather than a family of spellings, because what it asks about is
4703    /// which of the two edges of a control flow transfer is checked and the two are not separate
4704    /// questions to the hardware.
4705    #[test]
4706    fn which_control_flow_edges_are_checked_is_asked_for_by_name() {
4707        let (opts, _) = compile(&["-c", "a.c"]);
4708        assert_eq!(opts.control, Control::None, "gcc's default on the targets this compiler has");
4709
4710        for (arg, want) in [
4711            ("-fcf-protection", Control::Full),
4712            ("-fcf-protection=full", Control::Full),
4713            ("-fcf-protection=branch", Control::Branch),
4714            ("-fcf-protection=return", Control::Return),
4715            ("-fcf-protection=none", Control::None),
4716            ("-fcf-protection=check", Control::Check),
4717        ] {
4718            let (opts, _) = compile(&["-c", arg, "a.c"]);
4719            assert_eq!(opts.control, want, "{arg}");
4720        }
4721
4722        // The shape a package build uses: on in the global flags and off for the one directory
4723        // that cannot have it, whichever of the two spellings of off it reaches for.
4724        let (opts, _) = compile(&["-c", "-fcf-protection=full", "-fno-cf-protection", "a.c"]);
4725        assert_eq!(opts.control, Control::None);
4726        let (opts, _) = compile(&["-c", "-fno-cf-protection", "-fcf-protection=branch", "a.c"]);
4727        assert_eq!(opts.control, Control::Branch, "the last one wins either way round");
4728    }
4729
4730    /// The profiler is asked for by two spellings, and where its hook goes by two more.
4731    ///
4732    /// The two halves are separate on purpose. `-mfentry` on its own says where a call would go and
4733    /// asks for no call, which is what gcc does with it, and a build system that sets it globally
4734    /// and asks for the profile per directory needs that to be true rather than an error.
4735    ///
4736    /// The link is asserted alongside, because the flag changes it too and a build that compiled
4737    /// with it and linked without it is a program that calls the hook everywhere and never writes a
4738    /// profile.
4739    #[test]
4740    fn the_profiler_and_where_its_hook_goes_are_two_separate_questions() {
4741        let (opts, _) = compile(&["-c", "a.c"]);
4742        assert!(!opts.profile);
4743        assert_eq!(opts.hook, Hook::Platform, "neither was named, so the target decides");
4744
4745        for arg in ["-pg", "-p"] {
4746            let (opts, _) = compile(&["-c", arg, "a.c"]);
4747            assert!(opts.profile, "{arg}");
4748            let (link, _) = linking(&[arg, "a.c"]);
4749            assert!(link.profile, "{arg} changes the link as well");
4750        }
4751
4752        for (arg, want) in [("-mfentry", Hook::Early), ("-mno-fentry", Hook::Late)] {
4753            let (opts, _) = compile(&["-c", arg, "a.c"]);
4754            assert_eq!(opts.hook, want, "{arg}");
4755            assert!(!opts.profile, "{arg} asks for no call of its own");
4756        }
4757
4758        let (opts, _) = compile(&["-c", "-mfentry", "-mno-fentry", "-pg", "a.c"]);
4759        assert_eq!(opts.hook, Hook::Late, "the last one wins");
4760        assert!(opts.profile);
4761    }
4762
4763    /// How much room a patcher is promised, which is one number or two.
4764    ///
4765    /// A command line that did not ask is asserted alongside, because the flag has to be written to
4766    /// mean anything and a build that reserved room nobody asked for would grow every function in
4767    /// it for nothing.
4768    #[test]
4769    fn the_room_a_patcher_is_promised_is_a_number_of_bytes_and_where_they_go() {
4770        let (opts, _) = compile(&["-c", "a.c"]);
4771        assert_eq!(opts.patchable, Patchable::default());
4772        assert!(!opts.patchable.any(), "nothing is reserved unless it was asked for");
4773
4774        let (opts, _) = compile(&["-c", "-fpatchable-function-entry=16", "a.c"]);
4775        assert_eq!(opts.patchable, Patchable { total: 16, before: 0 });
4776
4777        let (opts, _) = compile(&["-c", "-fpatchable-function-entry=5,3", "a.c"]);
4778        assert_eq!(opts.patchable, Patchable { total: 5, before: 3 });
4779        assert_eq!(opts.patchable.after(), 2);
4780
4781        // The last one wins, which is what every other flag of this shape does and what a build
4782        // that adds one to a command line it did not write is relying on.
4783        let (opts, _) = compile(&[
4784            "-c",
4785            "-fpatchable-function-entry=5,3",
4786            "-fpatchable-function-entry=2",
4787            "a.c",
4788        ]);
4789        assert_eq!(opts.patchable, Patchable { total: 2, before: 0 });
4790    }
4791
4792    /// And a request nothing could satisfy is refused rather than rounded into one that can be.
4793    #[test]
4794    fn room_in_front_of_the_label_that_is_more_than_the_room_asked_for_is_refused() {
4795        for arg in ["-fpatchable-function-entry=1,2", "-fpatchable-function-entry=x"] {
4796            let e = parse_args(&args(&["-c", arg, "a.c"])).unwrap_err();
4797            assert!(e.message.contains("is not an amount of room to reserve"), "{}", e.message);
4798        }
4799    }
4800
4801    /// What wraps rather than being undefined, which is two questions and three flags.
4802    ///
4803    /// The older flag is the pair of the newer two, which is gcc's own reading of it, so a build
4804    /// that writes `-fno-strict-overflow` gets both and a build that writes one of the others gets
4805    /// only what it asked for.
4806    #[test]
4807    fn what_overflows_rather_than_being_undefined_is_asked_for_two_ways() {
4808        let (opts, _) = compile(&["-c", "a.c"]);
4809        assert_eq!(opts.wrapping, Wrapping::NONE, "nothing wraps unless it was asked for");
4810
4811        let (opts, _) = compile(&["-c", "-fwrapv", "a.c"]);
4812        assert_eq!(opts.wrapping, Wrapping { signed: true, pointer: false, trap: false });
4813
4814        let (opts, _) = compile(&["-c", "-fwrapv-pointer", "a.c"]);
4815        assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: true, trap: false });
4816
4817        let (opts, _) = compile(&["-c", "-fno-strict-overflow", "a.c"]);
4818        assert_eq!(opts.wrapping, Wrapping::ALL);
4819
4820        // And the last one wins, in both directions. A build that turns one of these on globally
4821        // and off for one directory is relying on that, and so is one that writes the pair and
4822        // then takes half of it back.
4823        let (opts, _) = compile(&["-c", "-fwrapv", "-fno-wrapv", "a.c"]);
4824        assert_eq!(opts.wrapping, Wrapping::NONE);
4825
4826        let (opts, _) = compile(&["-c", "-fno-strict-overflow", "-fstrict-overflow", "a.c"]);
4827        assert_eq!(opts.wrapping, Wrapping::NONE);
4828
4829        let (opts, _) = compile(&["-c", "-fno-strict-overflow", "-fno-wrapv-pointer", "a.c"]);
4830        assert_eq!(opts.wrapping, Wrapping { signed: true, pointer: false, trap: false });
4831    }
4832
4833    /// And the other answer to the signed question cannot be held at the same time as the first.
4834    ///
4835    /// A program cannot both wrap and stop, so writing both is writing a contradiction, and gcc
4836    /// resolves it by letting the last one win rather than by reporting anything. That was measured
4837    /// against gcc 16 rather than read out of the manual, which says nothing about it: `-ftrapv
4838    /// -fwrapv` emits no checked calls and `-fwrapv -ftrapv` emits them.
4839    #[test]
4840    fn a_signed_overflow_that_stops_is_the_other_answer_and_not_a_third_one() {
4841        let (opts, _) = compile(&["-c", "-ftrapv", "a.c"]);
4842        assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: false, trap: true });
4843
4844        let (opts, _) = compile(&["-c", "-fwrapv", "-ftrapv", "a.c"]);
4845        assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: false, trap: true });
4846
4847        let (opts, _) = compile(&["-c", "-ftrapv", "-fwrapv", "a.c"]);
4848        assert_eq!(opts.wrapping, Wrapping { signed: true, pointer: false, trap: false });
4849
4850        let (opts, _) = compile(&["-c", "-ftrapv", "-fno-strict-overflow", "a.c"]);
4851        assert_eq!(opts.wrapping, Wrapping::ALL);
4852
4853        let (opts, _) = compile(&["-c", "-ftrapv", "-fno-trapv", "a.c"]);
4854        assert_eq!(opts.wrapping, Wrapping::NONE);
4855
4856        // And the flag that says what may be assumed says nothing about what happens, so it leaves
4857        // this alone where it takes the wrapping away. gcc does the same.
4858        let (opts, _) = compile(&["-c", "-ftrapv", "-fstrict-overflow", "a.c"]);
4859        assert_eq!(opts.wrapping, Wrapping { signed: false, pointer: false, trap: true });
4860    }
4861
4862    /// What a plain `char` is, which is four spellings of two answers and nothing by default.
4863    ///
4864    /// Nothing is the target's own answer and has to stay distinct from both of the others, since
4865    /// the same command line means a signed `char` on x86-64 and an unsigned one on Linux's arm64.
4866    /// The negative spellings are the other flag rather than a way of asking for the default, which
4867    /// was measured against gcc 16: `-fno-signed-char` defines `__CHAR_UNSIGNED__` and
4868    /// `-fno-unsigned-char` does not.
4869    #[test]
4870    fn the_signedness_of_a_plain_char_is_asked_for_in_four_ways() {
4871        let (opts, _) = compile(&["-c", "a.c"]);
4872        assert_eq!(opts.char_signed, None);
4873
4874        for flag in ["-fsigned-char", "-fno-unsigned-char"] {
4875            let (opts, _) = compile(&["-c", flag, "a.c"]);
4876            assert_eq!(opts.char_signed, Some(true), "{flag}");
4877        }
4878
4879        for flag in ["-funsigned-char", "-fno-signed-char"] {
4880            let (opts, _) = compile(&["-c", flag, "a.c"]);
4881            assert_eq!(opts.char_signed, Some(false), "{flag}");
4882        }
4883
4884        // And the last one wins, which is what a build that sets one globally and the other for a
4885        // directory relies on.
4886        let (opts, _) = compile(&["-c", "-funsigned-char", "-fsigned-char", "a.c"]);
4887        assert_eq!(opts.char_signed, Some(true));
4888
4889        // And what is asked for reaches the target, because that is what every other part of the
4890        // compiler asks. The triple is one whose own answer is the opposite, so a session that
4891        // ignored the flag would still read as signed here.
4892        let (opts, _) =
4893            compile(&["-c", "--target=aarch64-unknown-linux-gnu", "-fsigned-char", "a.c"]);
4894        assert!(Session::new(*opts).target.char_is_signed);
4895        let (opts, _) = compile(&["-c", "--target=aarch64-unknown-linux-gnu", "a.c"]);
4896        assert!(!Session::new(*opts).target.char_is_signed);
4897    }
4898
4899    /// And the size of an enumeration, which is one question with two spellings.
4900    #[test]
4901    fn the_smallest_enumeration_is_asked_for_and_taken_back() {
4902        let (opts, _) = compile(&["-c", "a.c"]);
4903        assert!(!opts.short_enums);
4904
4905        let (opts, _) = compile(&["-c", "-fshort-enums", "a.c"]);
4906        assert!(opts.short_enums);
4907
4908        let (opts, _) = compile(&["-c", "-fshort-enums", "-fno-short-enums", "a.c"]);
4909        assert!(!opts.short_enums);
4910
4911        let (opts, _) = compile(&["-c", "-fno-short-enums", "-fshort-enums", "a.c"]);
4912        assert!(opts.short_enums);
4913    }
4914
4915    /// And Microsoft's reading of an anonymous member, which the target answers where the command
4916    /// line said nothing. gcc's mingw build has it on and its Linux build has it off, so a header
4917    /// that closes a nameless union with a macro that expands to nothing is read the way the
4918    /// compiler that platform ships would read it.
4919    #[test]
4920    fn the_microsoft_reading_of_a_member_follows_the_target_until_it_is_asked_for() {
4921        // Named rather than left to the host, since the answer this asks for is the one a target
4922        // that is not Windows gives and on a Windows machine the host is not one of those.
4923        let (opts, _) = compile(&[LINUX, "-c", "a.c"]);
4924        assert!(!Session::new(*opts).ms_extensions());
4925
4926        let (opts, _) = compile(&["-c", "--target=x86_64-pc-windows-gnu", "a.c"]);
4927        assert!(Session::new(*opts).ms_extensions());
4928
4929        let (opts, _) = compile(&["-c", "-fms-extensions", "a.c"]);
4930        assert!(Session::new(*opts).ms_extensions());
4931
4932        let (opts, _) =
4933            compile(&["-c", "--target=x86_64-pc-windows-gnu", "-fno-ms-extensions", "a.c"]);
4934        assert!(!Session::new(*opts).ms_extensions());
4935    }
4936
4937    /// And a value nothing means is refused rather than taken for the nearest thing it looks like.
4938    ///
4939    /// `-fcf-protection=all` is the spelling somebody writes from memory, and a compiler that read
4940    /// it as `full` would be guessing, while one that let it fall through to the optimizer's `-f`
4941    /// family would report it as an unknown pass. Neither is the news the build wants.
4942    #[test]
4943    fn a_control_flow_protection_nothing_means_is_refused() {
4944        let e = parse_args(&args(&["-c", "-fcf-protection=all", "a.c"])).unwrap_err();
4945        assert!(e.message.contains("is not a control flow protection"), "{}", e.message);
4946        assert!(e.message.contains("full, branch, return, none or check"), "{}", e.message);
4947    }
4948
4949    #[test]
4950    fn the_link_flags_are_collected_apart_from_the_compilation() {
4951        let (link, _) = linking(&[
4952            "-static",
4953            "-nostartfiles",
4954            "-rdynamic",
4955            "-s",
4956            "-fuse-ld=mold",
4957            "-L/opt/lib",
4958            "-B",
4959            "/opt/tools",
4960            "a.c",
4961        ]);
4962        assert!(link.is_static);
4963        assert!(link.no_startfiles);
4964        assert!(link.export_dynamic);
4965        assert!(link.strip);
4966        assert_eq!(link.use_ld.as_deref(), Some("mold"));
4967        assert_eq!(link.search, vec![PathBuf::from("/opt/lib")]);
4968        assert_eq!(link.prefixes, vec![PathBuf::from("/opt/tools")]);
4969    }
4970
4971    #[test]
4972    fn a_comma_in_dash_wl_separates_two_arguments() {
4973        // The target is written down because the name of the object is derived from it, and `a.o`
4974        // on a Linux host is `a.obj` on a Windows one. What is under test is the splitting of the
4975        // argument, which has nothing to do with either.
4976        let (_, plan) = linking(&[LINUX, "-Wl,-rpath,/opt/lib", "-Xlinker", "--as-needed", "a.c"]);
4977        let link = plan.link.expect("expected a link step");
4978        assert_eq!(
4979            link.inputs,
4980            vec![
4981                link::Item::Linker("-rpath".into()),
4982                link::Item::Linker("/opt/lib".into()),
4983                link::Item::Linker("--as-needed".into()),
4984                link::Item::File("a.o".into()),
4985            ]
4986        );
4987    }
4988
4989    #[test]
4990    fn a_word_for_the_linker_keeps_its_place_among_the_files_too() {
4991        // What libtool writes around a set of convenience archives, and what #1279 was. Both words
4992        // are about the files between them, so the pair collected out of the line and appended to
4993        // the end is two options that bracket nothing and an archive that went in empty.
4994        let (_, plan) = linking(&[
4995            "--target=x86_64-unknown-linux-gnu",
4996            "a.c",
4997            "-Wl,--whole-archive",
4998            "libaesni.a",
4999            "-Wl,--no-whole-archive",
5000            "-lm",
5001        ]);
5002        let link = plan.link.expect("expected a link step");
5003        assert_eq!(
5004            link.inputs,
5005            vec![
5006                link::Item::File("a.o".into()),
5007                link::Item::Linker("--whole-archive".into()),
5008                link::Item::File("libaesni.a".into()),
5009                link::Item::Linker("--no-whole-archive".into()),
5010                link::Item::Library("m".into()),
5011            ]
5012        );
5013        // And it is not a job, because there is nothing to compile in a word for the linker.
5014        assert_eq!(plan.jobs.len(), 2);
5015    }
5016
5017    #[test]
5018    fn a_word_for_the_linker_on_a_dash_c_line_is_dropped_without_a_word() {
5019        // GCC says nothing about one either. `-Wl,` on a compile line is what a build system
5020        // writes when one variable holds the flags for both, and a note here would be a note on
5021        // every compile of every autotools project.
5022        let (_, plan) = linking(&["-c", "-Wl,--as-needed", "a.c"]);
5023        assert!(plan.link.is_none());
5024        assert!(plan.notes.is_empty(), "{:?}", plan.notes);
5025        assert_eq!(plan.jobs.len(), 1);
5026    }
5027
5028    #[test]
5029    fn a_library_keeps_its_place_between_the_objects() {
5030        // Link order is semantic: `-lm` written between two files resolves for the one before
5031        // it and not for the one after, so a library cannot be collected into a list of its own.
5032        // The target is named because the suffix of an object is the target's and this asserts
5033        // on the names: the same command line on a Windows host plans two `.obj` files.
5034        let (_, plan) = linking(&["--target=x86_64-unknown-linux-gnu", "a.c", "-lm", "b.c"]);
5035        let link = plan.link.expect("expected a link step");
5036        assert_eq!(
5037            link.inputs,
5038            vec![
5039                link::Item::File("a.o".into()),
5040                link::Item::Library("m".into()),
5041                link::Item::File("b.o".into()),
5042            ]
5043        );
5044        // And it is not a job, because there is nothing to compile in a library.
5045        assert_eq!(plan.jobs.len(), 2);
5046    }
5047
5048    #[test]
5049    fn a_library_on_a_dash_c_line_is_a_note_rather_than_an_error() {
5050        let (_, plan) = linking(&["-c", "-lm", "a.c"]);
5051        assert!(plan.link.is_none());
5052        assert!(plan.notes.iter().any(|n| n.contains("-lm")), "{:?}", plan.notes);
5053    }
5054
5055    #[test]
5056    fn the_sysroot_reaches_the_linker_as_well_as_the_headers() {
5057        let (link, _) = linking(&["--sysroot=/opt/root", "a.c"]);
5058        assert_eq!(link.sysroot, Some(PathBuf::from("/opt/root")));
5059    }
5060
5061    fn printed(s: &[&str]) -> String {
5062        match parse_args(&args(s)).expect("expected an answer") {
5063            Action::Print(line) => line,
5064            other => panic!("expected an answer, got {other:?}"),
5065        }
5066    }
5067
5068    fn refused(s: &[&str]) -> String {
5069        parse_args(&args(s)).expect_err("expected a refusal").message
5070    }
5071
5072    #[test]
5073    fn a_warning_flag_this_compiler_has_not_heard_of_is_taken_rather_than_refused() {
5074        // The rule in section 4.1, and the reason for it is autoconf: a configure script finds
5075        // out whether a warning flag exists by passing it and looking at the exit status, so a
5076        // compiler that refuses one it does not know fails a script written for a newer GCC.
5077        let (opts, _) = compile(&["-Wall", "-Wextra", "-Wno-format-truncation", "-c", "a.c"]);
5078        assert!(!opts.warnings_are_errors);
5079        assert!(opts.warnings);
5080        // The two spellings that do mean something are still read.
5081        let (opts, _) = compile(&["-Werror", "-c", "a.c"]);
5082        assert!(opts.warnings_are_errors);
5083        let (opts, _) = compile(&["-w", "-c", "a.c"]);
5084        assert!(!opts.warnings);
5085        // Off without being asked, the way gcc has it off, and both spellings are read.
5086        let (opts, _) = compile(&["-c", "a.c"]);
5087        assert!(!opts.system_header_warnings);
5088        let (opts, _) = compile(&["-Wsystem-headers", "-c", "a.c"]);
5089        assert!(opts.system_header_warnings);
5090        let (opts, _) = compile(&["-Wsystem-headers", "-Wno-system-headers", "-c", "a.c"]);
5091        assert!(!opts.system_header_warnings);
5092        let (opts, _) = compile(&["-pedantic-errors", "-c", "a.c"]);
5093        assert!(opts.pedantic && opts.warnings_are_errors);
5094    }
5095
5096    #[test]
5097    fn an_argument_for_a_separate_tool_is_refused_rather_than_dropped() {
5098        // Every one of these says something about the output, so the wrong answer is silence.
5099        assert!(refused(&["-Wa,--noexecstack", "-c", "a.c"]).contains("separate assembler"));
5100        assert!(refused(&["-Wp,-DX", "-c", "a.c"]).contains("separate assembler"));
5101        assert!(refused(&["-specs=/x", "a.c"]).contains("-specs= is not supported"));
5102        assert!(refused(&["-mcmodel=kernel", "-c", "a.c"]).contains("small code model"));
5103        assert!(refused(&["-gdwarf-4", "-c", "a.c"]).contains("DWARF 5"));
5104        assert!(refused(&["-Ofast", "-c", "a.c"]).contains("fast math"));
5105        // The word size the target does not have, which is a target this compiler was not asked
5106        // for rather than a flag it does not know.
5107        let no32 = refused(&["--target=x86_64-unknown-linux-gnu", "-m32", "-c", "a.c"]);
5108        assert!(no32.contains("32 bit target"), "{no32}");
5109    }
5110
5111    /// `-gz` and the two spellings of the split, which are the two questions about the shape of
5112    /// the debug output rather than about how much of it there is.
5113    ///
5114    /// Both answers here are about what happens when there is debug information to shape, and
5115    /// there is none yet, so what is being asserted is that the flags are read and remembered
5116    /// rather than that anything changed in the output. That is the whole of what taking them
5117    /// claims, and it is worth a test because the day `rucc-debug` writes a section this is where
5118    /// it comes to find out what the command line said.
5119    #[test]
5120    fn the_shape_of_the_debug_output_is_recorded_even_where_there_is_none_of_it() {
5121        let (opts, _) = compile(&["-c", "a.c"]);
5122        assert_eq!(opts.compress, Compress::None, "uncompressed unless somebody asks");
5123
5124        // Bare `-gz` is `-gz=zlib`, measured against gcc 16 rather than read out of the manual,
5125        // which describes the flag without ever saying which algorithm it picks.
5126        assert_eq!(compile(&["-gz", "-c", "a.c"]).0.compress, Compress::Zlib);
5127        for (spelling, want) in [
5128            ("none", Compress::None),
5129            ("zlib", Compress::Zlib),
5130            ("zlib-gnu", Compress::ZlibGnu),
5131            ("zstd", Compress::Zstd),
5132        ] {
5133            let (opts, _) = compile(&[&format!("-gz={spelling}"), "-c", "a.c"]);
5134            assert_eq!(opts.compress, want, "{spelling}");
5135        }
5136
5137        // A value nothing here has heard of is refused rather than rounded to the nearest one,
5138        // because a build that asked for `zstd` and quietly got `zlib` would ship a file its
5139        // reader may not understand and would have no way of finding out.
5140        for bad in ["-gz=gzip", "-gz="] {
5141            let failed = refused(&[bad, "-c", "a.c"]);
5142            assert!(failed.contains("is not a way to compress"), "{bad}: {failed}");
5143        }
5144
5145        // The split is refused in the direction that would have written a file and taken in the
5146        // direction that describes what happens. A build system that names the `.dwo` as an
5147        // output has to hear about it now rather than at the point the file is missing.
5148        let (opts, _) = compile(&["-gno-split-dwarf", "-g", "-c", "a.c"]);
5149        assert!(opts.debug_info, "the negative spelling says nothing about how much");
5150        let failed = refused(&["-gsplit-dwarf", "-c", "a.c"]);
5151        assert!(failed.contains(".dwo"), "the refusal names the file it would have written");
5152    }
5153
5154    /// The `-flto` family, which is the whole of an optimization this compiler does not do.
5155    ///
5156    /// Taken rather than refused because ignoring it gives a correct program that is slower than
5157    /// it could have been, which is section 4.1's hint about speed. The values are still held to
5158    /// gcc's, so a command line written for clang is told rather than quietly taken.
5159    #[test]
5160    fn the_link_time_family_is_read_and_checked_and_nothing_is_done_about_it() {
5161        let (opts, _) = compile(&["-c", "a.c"]);
5162        assert!(!opts.lto.requested, "nothing asks unless the command line does");
5163
5164        let (opts, _) = compile(&["-flto", "-c", "a.c"]);
5165        assert!(opts.lto.requested);
5166        assert_eq!(opts.lto.jobs, LtoJobs::One, "bare -flto is one process, the way gcc reads it");
5167
5168        // The last of the two directions wins, the same as every other pair of `-f` spellings.
5169        assert!(!compile(&["-flto", "-fno-lto", "-c", "a.c"]).0.lto.requested);
5170        assert!(compile(&["-fno-lto", "-flto", "-c", "a.c"]).0.lto.requested);
5171
5172        // A count is a count, and asking for one implies asking for the optimization.
5173        for (spelling, want) in [
5174            ("auto", LtoJobs::Auto),
5175            ("jobserver", LtoJobs::Jobserver),
5176            ("1", LtoJobs::One),
5177            ("8", LtoJobs::Count(8)),
5178        ] {
5179            let (opts, _) = compile(&[&format!("-flto={spelling}"), "-c", "a.c"]);
5180            assert_eq!(opts.lto.jobs, want, "{spelling}");
5181            assert!(opts.lto.requested, "{spelling} asks for it too");
5182        }
5183
5184        // gcc refuses a zero rather than reading it as `-fno-lto`, and `thin` is clang's spelling
5185        // of a question gcc answers with `-flto-partition=`, so somebody who wrote it meant a
5186        // different compiler and gets told so here rather than getting a serial link.
5187        for bad in ["-flto=0", "-flto=thin", "-flto=full", "-flto=-1"] {
5188            let failed = refused(&[bad, "-c", "a.c"]);
5189            assert!(failed.contains("link time jobs"), "{bad}: {failed}");
5190        }
5191
5192        // How the program is cut up before the work is spread over it.
5193        assert_eq!(compile(&["-c", "a.c"]).0.lto.partition, Partition::Balanced, "gcc's default");
5194        for (spelling, want) in [
5195            ("balanced", Partition::Balanced),
5196            ("1to1", Partition::OneToOne),
5197            ("one", Partition::One),
5198            ("max", Partition::Max),
5199            ("none", Partition::None),
5200        ] {
5201            let (opts, _) = compile(&[&format!("-flto-partition={spelling}"), "-c", "a.c"]);
5202            assert_eq!(opts.lto.partition, want, "{spelling}");
5203        }
5204        assert!(refused(&["-flto-partition=big", "-c", "a.c"]).contains("partitioning model"));
5205
5206        // And how hard the bytecode is compressed on its way into the object, which is zstd's
5207        // range of levels and is the range gcc checks an argument against.
5208        assert_eq!(compile(&["-c", "a.c"]).0.lto.compression, None, "whatever it does by default");
5209        assert_eq!(compile(&["-flto-compression-level=0", "-c", "a.c"]).0.lto.compression, Some(0));
5210        let (opts, _) = compile(&["-flto-compression-level=19", "-c", "a.c"]);
5211        assert_eq!(opts.lto.compression, Some(19));
5212        for bad in ["-flto-compression-level=20", "-flto-compression-level=-1"] {
5213            let failed = refused(&[bad, "-c", "a.c"]);
5214            assert!(failed.contains("compression level"), "{bad}: {failed}");
5215        }
5216
5217        // The two pairs that describe an arrangement rather than ask for one. Every object here
5218        // holds its machine code, so the fat spelling is what already happens and the other is a
5219        // smaller file rather than a different program, and the plugin pair is about a tool the
5220        // design in `spec/09-optimizer.md` never loads.
5221        for taken in [
5222            "-ffat-lto-objects",
5223            "-fno-fat-lto-objects",
5224            "-fuse-linker-plugin",
5225            "-fno-use-linker-plugin",
5226        ] {
5227            let (opts, _) = compile(&[taken, "-c", "a.c"]);
5228            assert!(!opts.lto.requested, "{taken} says nothing about whether to do it");
5229        }
5230    }
5231
5232    /// The profile family, which is the only one here that splits down the middle.
5233    ///
5234    /// Reading a profile is taken and writing one is refused, and the line between them is the one
5235    /// section 4.1 draws: ignoring a request to read the counts gives a correct program that is
5236    /// slower than it could have been, and ignoring a request to write them means a file the build
5237    /// declared as an output never appears.
5238    #[test]
5239    fn reading_a_profile_is_taken_and_writing_one_is_refused() {
5240        let (opts, _) = compile(&["-c", "a.c"]);
5241        assert!(!opts.profile_data.requested, "nothing asks unless the command line does");
5242        assert_eq!(opts.profile_data.path, None);
5243
5244        let (opts, _) = compile(&["-fprofile-use", "-c", "a.c"]);
5245        assert!(opts.profile_data.requested);
5246        assert_eq!(opts.profile_data.path, None, "beside the object, the way gcc looks");
5247
5248        let (opts, _) = compile(&["-fprofile-use=/counts", "-c", "a.c"]);
5249        assert!(opts.profile_data.requested, "naming a path asks for it too");
5250        assert_eq!(opts.profile_data.path.as_deref(), Some("/counts"));
5251
5252        // The last of the two directions wins, the same as every other pair of `-f` spellings.
5253        assert!(
5254            !compile(&["-fprofile-use", "-fno-profile-use", "-c", "a.c"]).0.profile_data.requested
5255        );
5256        assert!(
5257            compile(&["-fno-profile-use", "-fprofile-use", "-c", "a.c"]).0.profile_data.requested
5258        );
5259
5260        // The rest of the reading half, which is where the files are and three answers about what
5261        // to make of what is in them.
5262        let (opts, _) = compile(&[
5263            "-fprofile-dir=/build/profiles",
5264            "-fprofile-abs-path",
5265            "-fprofile-correction",
5266            "-fprofile-partial-training",
5267            "-c",
5268            "a.c",
5269        ]);
5270        assert_eq!(opts.profile_data.dir.as_deref(), Some("/build/profiles"));
5271        assert!(opts.profile_data.absolute);
5272        assert!(opts.profile_data.correction);
5273        assert!(opts.profile_data.partial_training);
5274
5275        // Writing one, which is refused by name. The first four instrument the program and the
5276        // last writes a file beside the object, and a build that got neither and no message would
5277        // go on to optimize against counts that were never gathered.
5278        for writing in [
5279            "-fprofile-generate",
5280            "-fprofile-generate=/build/profiles",
5281            "-fprofile-arcs",
5282            "--coverage",
5283            "-fcondition-coverage",
5284            "-fpath-coverage",
5285        ] {
5286            let failed = refused(&[writing, "-c", "a.c"]);
5287            assert!(failed.contains("instrument"), "{writing}: {failed}");
5288        }
5289        assert!(refused(&["-ftest-coverage", "-c", "a.c"]).contains(".gcno"), "it names the file");
5290
5291        // The negative spellings of the refused half are what already happens, so they are taken.
5292        for taken in ["-fno-profile-generate", "-fno-profile-arcs", "-fno-test-coverage"] {
5293            let (opts, _) = compile(&[taken, "-c", "a.c"]);
5294            assert!(!opts.profile_data.requested, "{taken} asks for nothing");
5295        }
5296
5297        // And the flags that describe the instrumentation that is refused above, which are checked
5298        // and dropped. Checked because a typo is worth finding here rather than on the day the
5299        // instrumentation lands.
5300        for taken in [
5301            "-fprofile-update=single",
5302            "-fprofile-update=atomic",
5303            "-fprofile-update=prefer-atomic",
5304            "-fprofile-reproducible=serial",
5305            "-fprofile-reproducible=parallel-runs",
5306            "-fprofile-reproducible=multithreaded",
5307            "-fprofile-values",
5308            "-fno-profile-values",
5309            "-fprofile-info-section",
5310            "-fprofile-filter-files=a.c",
5311            "-fprofile-exclude-files=b.c",
5312            "-fprofile-note=a.gcno",
5313        ] {
5314            let (opts, _) = compile(&[taken, "-c", "a.c"]);
5315            assert!(!opts.profile_data.requested, "{taken} says nothing about reading one");
5316        }
5317        assert!(refused(&["-fprofile-update=none", "-c", "a.c"]).contains("update method"));
5318        assert!(refused(&["-fprofile-reproducible=any", "-c", "a.c"]).contains("reproducibility"));
5319    }
5320
5321    /// The sanitizers, which are refused by name and are the one family refused for a reason that
5322    /// is not about the bytes.
5323    ///
5324    /// A sanitizer is a promise that the program is watched while it runs, so a build that asked
5325    /// for one and was quietly given a program with no checks in it gets a test suite that passes
5326    /// for the wrong reason rather than a slower program.
5327    #[test]
5328    fn a_sanitizer_that_is_still_asked_for_at_the_end_of_the_line_is_refused_by_name() {
5329        for asked in ["address", "undefined", "thread", "kernel-address", "leak", "memory"] {
5330            let failed = refused(&[&format!("-fsanitize={asked}"), "-c", "a.c"]);
5331            assert!(failed.contains(asked), "the refusal names what was asked for: {failed}");
5332            assert!(failed.contains("-fsafety=detect"), "and the nearest thing: {failed}");
5333        }
5334
5335        // A list is every name in it, and the first one still standing is the one named.
5336        let failed = refused(&["-fsanitize=address,undefined", "-c", "a.c"]);
5337        assert!(failed.contains("address"), "{failed}");
5338
5339        // A name that is not one, which is worth its own message: somebody who wrote `-fsanitize`
5340        // with a typo in it has a different problem from somebody who wrote a real one.
5341        for bad in ["-fsanitize=bogus", "-fsanitize=address,bogus", "-fno-sanitize=bogus"] {
5342            let failed = refused(&[bad, "-c", "a.c"]);
5343            assert!(failed.contains("is not a sanitizer"), "{bad}: {failed}");
5344        }
5345
5346        // gcc takes `all` only in the negative, and so does this.
5347        assert!(refused(&["-fsanitize=all", "-c", "a.c"]).contains("only `-fno-sanitize=all`"));
5348
5349        // Asking and then taking it back is asking for nothing, which is why the answer waits for
5350        // the end of the line. A build whose shared flags turn a check on and whose rule for one
5351        // file turns it off again compiles that file here.
5352        for pair in [
5353            ["-fsanitize=address", "-fno-sanitize=address"],
5354            ["-fsanitize=address,undefined", "-fno-sanitize=all"],
5355            ["-fsanitize=undefined", "-fno-sanitize=undefined"],
5356        ] {
5357            let (opts, _) = compile(&[pair[0], pair[1], "-c", "a.c"]);
5358            assert_eq!(opts.safety, rucc_session::Safety::Off, "{pair:?} asked for nothing");
5359        }
5360        // And the other order still asks, because the last word is the one that counts.
5361        assert!(!refused(&["-fno-sanitize=address", "-fsanitize=address", "-c", "a.c"]).is_empty());
5362
5363        // What a check does when it fires is an answer about checks that are refused, so there is
5364        // nothing left for it to change and it is taken.
5365        for taken in [
5366            "-fsanitize-recover=undefined",
5367            "-fno-sanitize-recover=all",
5368            "-fsanitize-trap=undefined",
5369            "-fno-sanitize-trap=all",
5370            "-fsanitize-undefined-trap-on-error",
5371            "-fsanitize-address-use-after-scope",
5372            "-fno-sanitize-address-use-after-scope",
5373            "-fsanitize-sections=.data",
5374        ] {
5375            let (opts, _) = compile(&[taken, "-c", "a.c"]);
5376            assert_eq!(opts.safety, rucc_session::Safety::Off, "{taken} asks for no checking");
5377        }
5378        assert!(refused(&["-fsanitize-recover=bogus", "-c", "a.c"]).contains("is not a sanitizer"));
5379
5380        // Coverage instrumentation is refused rather than dropped, because a fuzzer with no
5381        // feedback runs blind and never says so.
5382        let failed = refused(&["-fsanitize-coverage=trace-pc", "-c", "a.c"]);
5383        assert!(failed.contains("feedback"), "{failed}");
5384        let failed = refused(&["-fsanitize-coverage=trace-pc-guard", "-c", "a.c"]);
5385        assert!(failed.contains("trace-pc or trace-cmp"), "gcc takes two of them: {failed}");
5386    }
5387
5388    #[test]
5389    fn the_levels_gcc_spells_differently_are_the_levels_they_mean() {
5390        assert_eq!(compile(&["-O", "-c", "a.c"]).0.opt_level, OptLevel::O1);
5391        assert_eq!(compile(&["-Og", "-c", "a.c"]).0.opt_level, OptLevel::O1);
5392        assert_eq!(compile(&["-O2", "-c", "a.c"]).0.opt_level, OptLevel::O2);
5393    }
5394
5395    #[test]
5396    fn the_machine_flags_that_name_what_we_already_do_are_taken_and_the_rest_are_not() {
5397        let line = ["--target=x86_64-unknown-linux-gnu", "-m64", "-march=x86-64-v3"];
5398        let (opts, _) =
5399            compile(&[&line[..], &["-mtune=native", "-mabi=sysv", "-c", "a.c"]].concat());
5400        assert_eq!(opts.target.to_string(), "x86_64-unknown-linux-gnu");
5401        let wrong = refused(&["--target=x86_64-unknown-linux-gnu", "-mabi=ms", "-c", "a.c"]);
5402        assert!(wrong.contains("sysv convention"), "{wrong}");
5403    }
5404
5405    #[test]
5406    fn the_thread_flag_is_a_macro_and_a_library_and_the_library_goes_last() {
5407        let (opts, plan) = compile(&["-pthread", "-c", "a.c"]);
5408        assert!(opts.defines.iter().any(|d| d == "_REENTRANT"));
5409        // After the input, because a static link takes what it needs from a library when it
5410        // reaches it and not afterwards.
5411        let names: Vec<&str> = plan.jobs.iter().map(|j| j.input.as_str()).collect();
5412        assert_eq!(names, vec!["a.c"]);
5413    }
5414
5415    #[test]
5416    fn the_questions_a_build_system_asks_before_it_compiles_anything() {
5417        let target = "--target=x86_64-unknown-linux-gnu";
5418        assert_eq!(printed(&[target, "-dumpmachine"]), "x86_64-unknown-linux-gnu");
5419        assert_eq!(printed(&[target, "-dumpversion"]), VERSION);
5420        assert_eq!(printed(&[target, "-dumpfullversion"]), VERSION);
5421        assert_eq!(printed(&[target, "-print-multiarch"]), "x86_64-linux-gnu");
5422        // A name nothing holds comes back unchanged, which is GCC's rule and is what makes the
5423        // answer safe to paste into a link line whether or not the file is there.
5424        assert_eq!(printed(&[target, "-print-file-name=no-such-library.a"]), "no-such-library.a");
5425        assert_eq!(printed(&[target, "-print-prog-name=ld"]), "ld");
5426        let dirs = printed(&[target, "-print-search-dirs"]);
5427        assert!(dirs.starts_with("install: "), "{dirs}");
5428        assert!(dirs.contains("\nlibraries: ="), "{dirs}");
5429    }
5430
5431    #[test]
5432    fn the_sysroot_in_effect_is_the_one_the_command_line_named_or_the_one_for_the_target() {
5433        // A tree the user named is the answer whatever the target is, because it is the answer to
5434        // every other question too.
5435        assert_eq!(printed(&["--sysroot=/opt/cross", "-print-sysroot"]), "/opt/cross");
5436
5437        // A target that is no machine this suite runs on is read under the cache, and the answer is
5438        // the root rather than one of the directories under it, since what asks is looking for a
5439        // file of its own.
5440        let root = cache::dir().join("sysroots").join("riscv64-linux-musl");
5441        assert_eq!(
5442            printed(&["--target=riscv64-linux-musl", "-print-sysroot"]),
5443            root.display().to_string()
5444        );
5445
5446        // And a compile for this machine has no sysroot, which is the empty line GCC prints when it
5447        // was configured without one rather than a `/` that would be a claim about the filesystem.
5448        let host = Triple::host().expect("a host this compiler knows");
5449        assert_eq!(printed(&[&format!("--target={host}"), "-print-sysroot"]), "");
5450    }
5451
5452    #[test]
5453    fn the_provenance_of_a_sysroot_is_the_manifest_it_carries() {
5454        // Section 13.5 wants seven things per input and wants them machine readable, and the manifest
5455        // is the record that already has them, so the flag prints that rather than a second format.
5456        let manifest = "rucc sysroot manifest 3\n\
5457                        target\tx86_64-linux-musl\n\
5458                        kernel\t6.12\n\
5459                        include/generic/stdio.h\tmusl-1.2.5\t\
5460                        https://musl.libc.org/releases/musl-1.2.5.tar.gz\t\
5461                        0000000000000000000000000000000000000000000000000000000000000000\tmit\t\
5462                        bundled\n\
5463                        lib/libc.so\tmusl-1.2.5\t\
5464                        https://musl.libc.org/releases/musl-1.2.5.tar.gz\t\
5465                        1111111111111111111111111111111111111111111111111111111111111111\tmit\t\
5466                        generated\n";
5467        let tree = TempTree::new("provenance", &[("manifest", manifest)]);
5468        let sysroot = format!("--sysroot={}", tree.0.display());
5469        // The kernel line of tamnd/rucc#934 is in the answer without anything here naming it, because
5470        // the flag parses the record and renders it again rather than picking fields out of it. That
5471        // is the reason it prints a manifest and not a format of its own.
5472        //
5473        // The answer is the file without its last newline, because whatever prints it adds one. The
5474        // file is what somebody diffs the output against, so the two have to be the same bytes.
5475        assert_eq!(printed(&[&sysroot, "-print-sysroot-provenance"]) + "\n", manifest);
5476
5477        // A tree with no manifest in it is a tree somebody assembled themselves, and nothing here
5478        // knows where any of it came from. Saying nothing is the only honest answer, and a reader can
5479        // tell it from a manifest with no inputs because that one still has its two header lines.
5480        let bare = TempTree::new("provenance-bare", &[]);
5481        assert_eq!(
5482            printed(&[&format!("--sysroot={}", bare.0.display()), "-print-sysroot-provenance"]),
5483            ""
5484        );
5485
5486        // And a compile for this machine has no sysroot at all, which is the same empty answer
5487        // `-print-sysroot` gives for it.
5488        let host = Triple::host().expect("a host this compiler knows");
5489        assert_eq!(printed(&[&format!("--target={host}"), "-print-sysroot-provenance"]), "");
5490
5491        // And the other spelling, which section 13.5 is the document that writes.
5492        assert_eq!(printed(&[&sysroot, "--print-sysroot-provenance"]) + "\n", manifest);
5493
5494        // tamnd/rucc#1021. The digest of the same tree is the sha256 of that record, so it is one
5495        // line where the provenance is a few hundred, and it is checkable with `sha256sum` because
5496        // the bytes it is over are the bytes of the file. The number here is that hash of the
5497        // fixture above, computed by `sha256sum` rather than by this compiler.
5498        assert_eq!(
5499            printed(&[&sysroot, "-print-sysroot-digest"]),
5500            "d705ae6ebeafeb7fda4bd57cecc7882bf49784b17015664a09cfae25a1b2000a"
5501        );
5502        assert_eq!(
5503            printed(&[&sysroot, "--print-sysroot-digest"]),
5504            printed(&[&sysroot, "-print-sysroot-digest"])
5505        );
5506
5507        // And the two empty answers are empty here too, because a digest of nothing would read as a
5508        // claim about a sysroot rather than as the absence of one.
5509        assert_eq!(
5510            printed(&[&format!("--sysroot={}", bare.0.display()), "-print-sysroot-digest"]),
5511            ""
5512        );
5513        assert_eq!(printed(&[&format!("--target={host}"), "-print-sysroot-digest"]), "");
5514    }
5515
5516    #[test]
5517    fn a_manifest_this_build_cannot_read_is_refused_rather_than_printed() {
5518        // Passing a file we could not parse to whoever asked would make their parser the one that
5519        // finds the problem, and the three uses section 13.5 gives for this are all somebody else
5520        // parsing it.
5521        let tree = TempTree::new(
5522            "provenance-bad",
5523            &[("manifest", "rucc sysroot manifest 3\ntarget\tx86_64-linux-musl\nlib/libc.a\n")],
5524        );
5525        let message =
5526            refused(&[&format!("--sysroot={}", tree.0.display()), "-print-sysroot-provenance"]);
5527        assert!(message.contains("manifest"), "{message}");
5528        assert!(message.contains("1 fields where an input has six"), "{message}");
5529
5530        // The digest is refused for the same file and for a stronger reason: a hash of bytes this
5531        // build cannot read would be a number that names a record nobody can act on.
5532        let digest =
5533            refused(&[&format!("--sysroot={}", tree.0.display()), "-print-sysroot-digest"]);
5534        assert_eq!(digest, message);
5535    }
5536
5537    #[test]
5538    fn the_two_dependency_flags_that_stop_after_the_rule_stop_after_the_rule() {
5539        let (opts, _) = compile(&["-M", "a.c"]);
5540        assert!(opts.deps.emit && opts.deps.instead_of_compiling);
5541        assert!(opts.deps.system_headers, "plain -M lists them");
5542        assert_eq!(opts.emit, EmitKind::Preprocessed);
5543
5544        // Even where a later flag asked for something else, because the family is a mode and
5545        // the mode is what the run is for.
5546        let (opts, _) = compile(&["-M", "-c", "a.c"]);
5547        assert_eq!(opts.emit, EmitKind::Preprocessed);
5548
5549        let (opts, _) = compile(&["-MM", "a.c"]);
5550        assert!(!opts.deps.system_headers);
5551    }
5552
5553    #[test]
5554    fn the_two_that_end_in_d_leave_the_compilation_alone() {
5555        let (opts, _) = compile(&["-MD", "-c", "a.c"]);
5556        assert!(opts.deps.emit && !opts.deps.instead_of_compiling);
5557        assert!(opts.deps.system_headers);
5558        assert_eq!(opts.emit, EmitKind::Object);
5559
5560        let (opts, _) = compile(&["-MMD", "-c", "a.c"]);
5561        assert!(opts.deps.emit && !opts.deps.instead_of_compiling);
5562        assert!(!opts.deps.system_headers);
5563    }
5564
5565    #[test]
5566    fn nothing_puts_the_system_headers_back_once_a_flag_has_taken_them_out() {
5567        // GCC's rule, and not an oversight in it. The flag asking for fewer of them is read as
5568        // the answer, because the other one never asked the question.
5569        let (opts, _) = compile(&["-MM", "-M", "a.c"]);
5570        assert!(!opts.deps.system_headers);
5571        let (opts, _) = compile(&["-MD", "-MMD", "-c", "a.c"]);
5572        assert!(!opts.deps.system_headers);
5573        let (opts, _) = compile(&["-MMD", "-MD", "-c", "a.c"]);
5574        assert!(!opts.deps.system_headers);
5575    }
5576
5577    #[test]
5578    fn a_target_arrives_escaped_from_one_flag_and_untouched_from_the_other() {
5579        let (opts, _) = compile(&["-MM", "-MT", "a b.o", "-MQ", "a b.o", "a.c"]);
5580        assert_eq!(opts.deps.targets, vec!["a b.o".to_owned(), "a\\ b.o".to_owned()]);
5581    }
5582
5583    #[test]
5584    fn the_rest_of_the_family_is_a_file_and_a_switch() {
5585        let (opts, _) = compile(&["-MM", "-MF", "dep.d", "-MP", "a.c"]);
5586        assert_eq!(opts.deps.file.as_deref(), Some("dep.d"));
5587        assert!(opts.deps.phony);
5588
5589        for flag in ["-MF", "-MT", "-MQ"] {
5590            let e = parse_args(&args(&[flag])).unwrap_err();
5591            assert!(e.message.contains("requires an argument"), "{}", e.message);
5592        }
5593    }
5594
5595    /// A directory of sources for one test, removed when the test is done with it.
5596    struct TempTree(PathBuf);
5597
5598    impl Drop for TempTree {
5599        fn drop(&mut self) {
5600            let _ = std::fs::remove_dir_all(&self.0);
5601        }
5602    }
5603
5604    impl TempTree {
5605        fn new(name: &str, files: &[(&str, &str)]) -> TempTree {
5606            let dir = std::env::temp_dir().join(format!("rucc-deps-{}-{name}", std::process::id()));
5607            let _ = std::fs::remove_dir_all(&dir);
5608            std::fs::create_dir_all(&dir).expect("temporary directory should be writable");
5609            for (path, text) in files {
5610                let at = dir.join(path);
5611                if let Some(parent) = at.parent() {
5612                    std::fs::create_dir_all(parent).expect("creating a subdirectory should work");
5613                }
5614                std::fs::write(&at, text).expect("writing a temporary file should work");
5615            }
5616            TempTree(dir)
5617        }
5618
5619        fn path(&self, name: &str) -> String {
5620            self.0.join(name).to_string_lossy().into_owned()
5621        }
5622    }
5623
5624    #[test]
5625    fn the_rule_names_what_the_includes_found_and_names_each_of_them_once() {
5626        // End to end, because the list comes from the preprocessor and the format comes from
5627        // somewhere else, and a test of either half on its own would pass with the two of them
5628        // wired up backwards.
5629        let tree = TempTree::new(
5630            "found",
5631            &[
5632                ("a.c", "#include \"one.h\"\n#include \"two.h\"\nint main(void) { return X; }\n"),
5633                ("one.h", "#define X 0\n"),
5634                ("two.h", "#include \"one.h\"\n"),
5635            ],
5636        );
5637        let out = tree.path("dep.d");
5638        let code = run(&args(&["-MM", "-MF", &out, "-o", &tree.path("a.i"), &tree.path("a.c")]));
5639        assert_eq!(code, 0);
5640
5641        let text = std::fs::read_to_string(&out).expect("the rule should have been written");
5642        let names: Vec<&str> = text.split_whitespace().collect();
5643        // The target, the source, and each header once however many times it was reached.
5644        assert_eq!(names.first(), Some(&"a.o:"), "{text}");
5645        assert_eq!(names.iter().filter(|n| n.ends_with("one.h")).count(), 1, "{text}");
5646        assert_eq!(names.iter().filter(|n| n.ends_with("two.h")).count(), 1, "{text}");
5647        // And the `-o` went to the file the rule replaced, which is left empty rather than
5648        // absent because a makefile that named it as a target will look for it.
5649        assert_eq!(std::fs::read(tree.path("a.i")).expect("the output should exist"), b"");
5650    }
5651
5652    #[test]
5653    fn a_header_that_is_only_reached_under_a_guard_is_still_a_dependency() {
5654        // The multiple-include optimization means the second reach never opens the file. It is
5655        // still a file this translation unit was built from, so it is still in the rule.
5656        let tree = TempTree::new(
5657            "guarded",
5658            &[
5659                ("a.c", "#include \"g.h\"\n#include \"g.h\"\nint main(void) { return 0; }\n"),
5660                ("g.h", "#ifndef G\n#define G\n#endif\n"),
5661            ],
5662        );
5663        let out = tree.path("dep.d");
5664        let code = run(&args(&["-MM", "-MF", &out, "-o", &tree.path("a.i"), &tree.path("a.c")]));
5665        assert_eq!(code, 0);
5666        let text = std::fs::read_to_string(&out).expect("the rule should have been written");
5667        assert_eq!(text.split_whitespace().filter(|n| n.ends_with("g.h")).count(), 1, "{text}");
5668    }
5669
5670    #[test]
5671    fn every_imacros_file_is_read_before_every_include_file_whatever_order_they_were_written() {
5672        // Measured against GCC rather than read: the two flags the other way round produce the
5673        // same output byte for byte, so the command line order between the two families does not
5674        // decide anything and the order within one does. The `-include` file here can only see
5675        // the definition if the `-imacros` file that was written after it ran first.
5676        let tree = TempTree::new(
5677            "preinclude",
5678            &[
5679                ("a.c", "int main(void) { return 0; }\n"),
5680                ("i.h", "#ifdef FROM_MACROS\nint saw_it;\n#else\nint missed_it;\n#endif\n"),
5681                ("m.h", "#define FROM_MACROS 1\nint macros_text;\n"),
5682            ],
5683        );
5684        let out = tree.path("a.i");
5685        let code = run(&args(&[
5686            "-E",
5687            "-include",
5688            &tree.path("i.h"),
5689            "-imacros",
5690            &tree.path("m.h"),
5691            "-o",
5692            &out,
5693            &tree.path("a.c"),
5694        ]));
5695        assert_eq!(code, 0);
5696        let text = std::fs::read_to_string(&out).expect("the output should have been written");
5697        assert!(text.contains("saw_it"), "{text}");
5698        // And the text of the `-imacros` file is thrown away, which is the whole difference
5699        // between the two flags.
5700        assert!(!text.contains("macros_text"), "{text}");
5701    }
5702
5703    #[test]
5704    fn a_file_the_command_line_named_is_a_prerequisite_the_same_as_one_a_directive_named() {
5705        let tree = TempTree::new(
5706            "preinclude-deps",
5707            &[
5708                ("a.c", "int main(void) { return 0; }\n"),
5709                ("i.h", "int from_include;\n"),
5710                ("m.h", "#define M 1\n"),
5711            ],
5712        );
5713        let out = tree.path("dep.d");
5714        let code = run(&args(&[
5715            "-MM",
5716            "-MF",
5717            &out,
5718            "-include",
5719            &tree.path("i.h"),
5720            "-imacros",
5721            &tree.path("m.h"),
5722            "-o",
5723            &tree.path("a.i"),
5724            &tree.path("a.c"),
5725        ]));
5726        assert_eq!(code, 0);
5727        let text = std::fs::read_to_string(&out).expect("the rule should have been written");
5728        assert!(text.contains("i.h"), "{text}");
5729        assert!(text.contains("m.h"), "{text}");
5730    }
5731
5732    #[test]
5733    fn a_command_line_include_that_is_nowhere_on_the_path_is_an_error_and_not_a_warning() {
5734        // Including the directory of the source file, which is not on the path for these: the
5735        // command line was not written there, so a name in it is relative to where the compiler
5736        // was run rather than to where the source sits.
5737        let tree = TempTree::new(
5738            "preinclude-missing",
5739            &[("sub/a.c", "int main(void) { return 0; }\n"), ("sub/beside.h", "int x;\n")],
5740        );
5741        let code = run(&args(&["-E", "-include", "beside.h", "-o", "-", &tree.path("sub/a.c")]));
5742        assert_eq!(code, 1);
5743    }
5744
5745    #[test]
5746    fn a_command_line_that_links_names_the_executable_and_not_the_object_it_went_through() {
5747        // The object a link goes through is in a temporary directory and is gone before `make`
5748        // reads any of this, so the rule that named it would be a rule for a file that is never
5749        // there. The target and the file are both the `-o`, which is the executable.
5750        let (opts, plan) = compile(&["-MD", "sub/a.c", "-o", "prog"]);
5751        assert_eq!(plan.output.as_deref(), Some("prog"));
5752        assert_eq!(deps::default_target("sub/a.c", deps_target_output(&opts, &plan)), "prog");
5753        assert_eq!(
5754            deps::default_file(&opts.deps, "sub/a.c", plan.output.as_deref()).as_deref(),
5755            Some("prog.d")
5756        );
5757    }
5758
5759    #[test]
5760    fn the_plan_keeps_the_output_name_because_the_rule_is_written_from_it() {
5761        let (_, plan) = compile(&["-MMD", "-c", "sub/a.c", "-o", "obj/x.o"]);
5762        assert_eq!(plan.output.as_deref(), Some("obj/x.o"));
5763        let (_, plan) = compile(&["-MMD", "-c", "sub/a.c"]);
5764        assert_eq!(plan.output, None);
5765    }
5766
5767    #[test]
5768    fn usage_fits_on_a_screen() {
5769        // Not a style preference. A help text that scrolls is one nobody reads, and this is
5770        // the cheapest way to keep it honest as flags accumulate. The number goes up only when
5771        // a family of flags arrives that has nowhere to share a line, which the two pass gates
5772        // were and which the two fuel flags and `-fsafety=` now are, and it goes up by exactly
5773        // the lines that family took. The four it went up by last are the flags a build system
5774        // passes without being asked to: how much to say, what machine to generate for, threads,
5775        // and the questions `configure` asks before it compiles anything. The one it went up by
5776        // last is the second line of `--emit`, whose kinds are a family that has now outgrown
5777        // one line and has nowhere else to go. The two it went up by last are the dependency
5778        // family, which is eight flags that share nothing with anything above them. The one it
5779        // went up by last is the four spellings of position independent code, which every
5780        // configure script writes and which could only have shared the link line, and that line
5781        // is already four characters short of the limit. The two it went up by last are the rest
5782        // of the include family, which is six more flags that change where a header is looked for
5783        // and two that name a header outright. The one it went up by last is the pair that keeps
5784        // the intermediate files and times the steps, which belong next to the two flags above
5785        // them that are also about watching a compilation rather than changing one. The two it
5786        // went up by last are the section flags and the visibility flag, which are what a build
5787        // that cares about the size of what it ships and about which names it exports writes, and
5788        // the second of them was already taken and only missing from here. The one it went up by
5789        // last is the stack protector, which is four spellings of one question and which every
5790        // distribution puts on every command line it issues, so a build that reads this list
5791        // looking for it and does not find it has to go and read the specification instead. The one
5792        // it went up by last is the profiler, which is two spellings of the request and two of
5793        // where the call goes, and which is about watching a program run rather than about what is
5794        // generated, so it shares its subject with nothing above it. The one it went up by last is
5795        // the room a function opens with for something to be written over it later, which takes an
5796        // argument of its own shape and is what a kernel build asks for, so it fits beside the
5797        // profiler and nothing else. The one it went up by last is what overflows rather than being
5798        // undefined, which is three spellings of two questions and which a kernel build and a great
5799        // deal of code written before the standard settled both pass. The one it went up by last is
5800        // the other answer to the first of those questions, which could not share the line because
5801        // what it asks for is the opposite of what the flags on that line ask for. The one it went
5802        // up by last is the split of the line that lists what this compiler does anyway into that
5803        // and what it assumes anyway, which are two different claims that were sharing a line until
5804        // the second of them got a second flag and the line stopped fitting. The one it went up by
5805        // last is the three flags that change the ABI rather than the code, which have to be given
5806        // to every file in a program or none of them and which therefore belong somewhere a person
5807        // reading this list will see them. The one it went up by last is the floating point group,
5808        // which is two lines rather than one because the first of them is a choice this compiler
5809        // records and the rest are claims about what it does anyway, and putting a real setting on
5810        // the same line as three flags that change nothing would be misleading about both. The one
5811        // it went up by last is the flag that says a write has to stay inside the member it names,
5812        // which is a setting rather than a claim and so cannot share the line above it, that being
5813        // the one that picks a tier. The two it went up by last are the prefix mapping family,
5814        // which is four flags whose whole job is to keep a build's output the same from two
5815        // different directories, and which a person chasing a reproducible build comes here
5816        // looking for by name. The one it went up by last is how the debug sections are compressed
5817        // and whether they go in a file of their own, which are two questions about the shape of
5818        // the debug output, where the line above them is about how much of it there is. The one it
5819        // went up by last is the `restrict` contract, which is a setting for the same reason the
5820        // flag that keeps a write inside its member is and which is the check a person who has been
5821        // bitten by a vectorizer comes here looking for. The one it went up by last is link time
5822        // optimization, which is a whole optimization rather than a flag and which says so on its
5823        // own line, because a build that passes it and reads this looking for what it got is
5824        // asking a question no other line here answers. The one it went up by last is the sysroot,
5825        // which is the question somebody asks when a cross build read a file nobody expected, and
5826        // which has no room on the line above it because the answers there are a path each and this
5827        // one is the root all of them are under. The one it went up by last is what is inside that
5828        // root and where each of it came from, which is a question about a whole tree rather than
5829        // about a path and which is long enough on its own that it could not have shared a line with
5830        // anything. The one it went up by last is the profile family, which splits down the middle
5831        // where no other family here does, so the line has to name the half that is taken and the
5832        // half that is refused or it would be read as taking both. The one it went up by last is
5833        // the sanitizers, which are what somebody reaching for a checked build writes first and
5834        // which belong beside the tier that is the nearest thing here to what they asked for. The
5835        // one it went up by last is the digest of that record, which is the same tree as one number
5836        // and could not share the line above it because that line prints a few hundred lines and
5837        // this one prints sixty four characters, and a reader who wants the short answer is looking
5838        // for it by name rather than reading the long one. The one it went up by last is the
5839        // sysroot fetch, which is the only command here that gets something from somewhere else and
5840        // is therefore the one a person wants to have read before they run it rather than after.
5841        // And the flag beside it that forbids every download, which earns its line by being what a
5842        // build in a sealed environment passes and by meaning something even though an ordinary
5843        // compile downloads nothing either way.
5844        assert!(USAGE.lines().count() < 72, "usage text has grown past one screen");
5845    }
5846}