rucc_codegen/abi.rs
1//! Where a function's arguments already are when it starts running, and where a call puts its own.
2//!
3//! Design: `spec/12-abi-and-runtime.md`.
4//!
5//! This is the one part of the calling convention that is not a lowering rule, and it is worth
6//! saying why, because everything else in this crate is. A rule matches a term and rewrites it,
7//! and which register the third argument arrives in is not a fact about any term: it depends on
8//! the argument's position and on the classification of every argument before it. A pattern has
9//! nowhere to put that. So the arguments are built here, by hand, out of what the convention
10//! says, the same way [`crate::finish`] builds a prologue.
11//!
12//! The classification itself is not here either. `rucc-lower` has already run it by the time a
13//! function reaches this crate, which is why the parameters read here are nearly all plain
14//! scalars: an aggregate has been split into the pieces it travels in, and a return through memory
15//! is an ordinary pointer parameter in front of the rest. What is left for this is the step after
16//! classification, from how a value travels to which register it is actually in, which is
17//! [`rucc_target::Places`].
18//!
19//! The one parameter that is not a scalar is an aggregate the classification put in the argument
20//! area whole, which is [`rucc_ir::Abi::ByVal`]. The IR calls it a pointer, because a pointer is
21//! what an instruction reading it has to have, and the convention says the bytes travel and the
22//! pointer does not. So this is the one place that reads what the classification said rather than
23//! only the type, on both sides of the call, and the two sides are the two halves of one copy.
24//!
25//! # What it writes
26//!
27//! One `x64.arg_val_*` per parameter that arrived in a register, at the top of the entry block,
28//! each defining a fresh register constrained to the one the argument arrived in. They encode to
29//! nothing. The point of them is that a parameter has to be defined somewhere for the allocator to
30//! have anything to move, and the entry block cannot define it as a block parameter: there is no
31//! edge into the entry block for the move to go on, which is what `rucc_regalloc::rewrite` asserts.
32//!
33//! What the allocator does with them is the whole of the argument sequence. A parameter that is
34//! read where it arrived costs nothing, and one that is not gets a copy, which is the same
35//! bargain the return already makes and is decided by the same code.
36//!
37//! A parameter whose bytes travelled is the exception to all of that. Its bytes are already in
38//! this function, at a place in the caller's argument area the same walk gives, so nothing is
39//! brought in at all: what the parameter is is where they are, and that is one `lea`. It waits on
40//! the frame the way the loads below it do, and for the same reason.
41//!
42//! A parameter past the last register arrived in the caller's memory rather than in a register, so
43//! it is a load and not a pseudo, and it is a real instruction that encodes to real bytes. How far
44//! up the caller's argument area it is is a number [`rucc_target::Places`] answers here, but where
45//! that area is from inside this function is a distance into a frame, and no frame exists until
46//! after allocation. So the load is written with nothing in its displacement, which of the two
47//! registers it reads through is left to be settled too, and both are filled in by [`crate::finish`]
48//! out of [`crate::frame::Frame::incoming`]. That is the same bargain an `alloca` already makes,
49//! for the same reason and in the same two places.
50//!
51//! # A call
52//!
53//! The same reasoning the other way round, and one instruction rather than several. `x64.call`
54//! and `x64.call_reg` are the only opcodes in the description whose operand vector is empty
55//! there, because nothing about a call's operands is the same from one call to the next, so they
56//! are built here: one read per argument constrained to the register the convention passes it in,
57//! one definition for the value that comes back constrained to the register it comes back in, and
58//! one definition per register the convention does not preserve.
59//!
60//! A call through an address has one operand more, which is the address, and it is the one
61//! operand of a call that is a fact about the instruction rather than about the signature. It
62//! goes in front of the arguments, because the assembler has to find it and an index into a
63//! vector whose length depends on the convention is not a way of finding anything.
64//!
65//! Those last ones are the clobbers, and they are the whole of what the allocator has to know
66//! about a call besides where the values go. Each is a definition of the physical register itself
67//! rather than of a value, since there is no value: it says the register is written here, which
68//! is exactly what stops the allocator from leaving something in one across the call. A register
69//! an argument or the result already names is not repeated, because naming it once already blocks
70//! it for the length of the instruction, which is all a clobber does.
71//!
72//! An argument past the last register the convention has for it is a store into the outgoing area
73//! rather than an operand of the call, written in front of the call in the same block. Where that
74//! area is does not have to wait for the frame the way the incoming one does, because the outgoing
75//! area is at the bottom of the frame and the bottom of the frame is where the stack pointer is:
76//! that is the whole reason the frame puts it there, since it is where the callee will look. So the
77//! offset [`rucc_target::Places`] gives back is the offset the store is written with.
78//!
79//! An object passed by value in memory is the same thing again and a copy rather than a store. The
80//! caller owes the callee a copy it is free to write to, which is what makes a C call by value
81//! different from passing a pointer the callee must not keep, and the argument area is where the
82//! convention says that copy goes. So the bytes are read out of the object and written into the
83//! area a word at a time, in front of the call, with the words chosen by the same function that
84//! chooses them for a `memcpy`. An object with more words than that unrolls to is copied by a call
85//! to the runtime's `memcpy` instead, written in front of the outer call in the same block.
86//!
87//! That is one call built in the middle of building another, which sounds worse than it is.
88//! Nothing of the outer call is in a physical register when the copy is written: every argument
89//! that travels in one is still a virtual register, and the register it has to end up in is a
90//! constraint on an operand of the call instruction, which is not built until every copy in front
91//! of it has been. So what the inner call destroys is what any call destroys, and the allocator
92//! keeps the outer call's values out of those registers the same way it does across a call the
93//! program wrote. The arguments already stored into the outgoing area are safe for a plainer
94//! reason: the inner call's own frame is below the stack pointer and that area is above it.
95//!
96//! The call still reports how many bytes it needed, because the frame reserves as many as the
97//! widest call in the function asked for and cannot know that until every call has been seen.
98
99use rucc_base::{Interner, Symbol};
100use rucc_diag::Span;
101use rucc_ir::{Abi, Drains, Param, Type};
102use rucc_mir as mir;
103use rucc_target::{CallRegs, Constraint, PhysReg, Places, RegClass, Variadic, Where};
104
105use crate::capability;
106use crate::varargs::Area;
107
108pub mod aarch64;
109
110/// Why a parameter could not be brought in.
111#[derive(Debug, Clone, Copy, PartialEq, Eq)]
112pub enum Missing {
113 /// It travels on the x87 stack, which is a `long double` and nothing else. That stack is a
114 /// third register file, it is not one the allocator has, and no instruction in the
115 /// description touches it.
116 OnX87,
117 /// It is a width no pseudo covers, which is anything a machine register does not hold.
118 Width,
119 /// It is a value that comes back in more registers than the convention returns in. A structure
120 /// of at most sixteen bytes comes back in up to two, which is as many as SysV has, and a
121 /// convention with fewer of them returns such a structure through a hidden pointer instead. So
122 /// this is what a signature the classification did not produce would get.
123 NoRoom,
124 /// It is an object whose bytes travel in the argument area and there are more of them than
125 /// any count of them can be written down as. A copy too long to unroll is a call to the
126 /// runtime and not a refusal, so what is left here is an object of two gigabytes or more,
127 /// which is a size the immediate holding the byte count has nowhere to put and a structure no
128 /// program passes.
129 TooBig,
130}
131
132impl Missing {
133 /// What it says when a function could not be compiled because of it.
134 ///
135 /// Worded so that it reads the same about a value arriving and a value being passed, since
136 /// the two are the same fact seen from the two ends of one call.
137 #[must_use]
138 pub fn why(self) -> &'static str {
139 match self {
140 Missing::OnX87 => "is on the x87 stack",
141 Missing::Width => "is a width no argument register holds",
142 Missing::NoRoom => "takes more registers than this convention has for it",
143 Missing::TooBig => "is more bytes than a count of them can be written down as",
144 }
145 }
146}
147
148/// Which register file a value of that type travels in.
149///
150/// The whole of what the two files mean to this module. A float is in the vector one and
151/// everything else is in the general purpose one, which is what both of this machine's conventions
152/// say. A `long double` is in neither and what its register holds is an address, which is a general
153/// purpose value like every other address, so it answers with the other file rather than with the
154/// one its type would suggest.
155fn class_of(ty: Type, conv: &CallRegs) -> RegClass {
156 if ty.is_float() && !on_the_stack(ty) { conv.sse_class } else { conv.int_class }
157}
158
159/// How many bytes of the argument area a value in the vector file takes.
160///
161/// Its own width, lanes included, which is what [`Places::float`] wants and is a number that only
162/// matters to a value the registers ran out before. Everything the machine computes in is a word
163/// or narrower and takes a word either way. The `_Float128` is the one that is not: two words, and
164/// aligned to two words, which is the difference between the argument behind it being placed after
165/// it and being placed on top of half of it.
166pub(crate) fn float_bytes(ty: Type) -> u32 {
167 ty.bits().div_ceil(8).saturating_mul(ty.lanes())
168}
169
170/// Spends the registers an object in the argument area said nothing after it may have.
171pub(crate) fn drain(places: &mut Places<'_>, drains: Drains) {
172 match drains {
173 Drains::Nothing => {}
174 Drains::Integers => places.drain_integers(),
175 Drains::Floats => places.drain_floats(),
176 }
177}
178
179/// How many bytes of the argument area a value in the general purpose file takes.
180///
181/// Its own width, which only a convention that packs the argument area reads, since everywhere
182/// else anything this narrow takes a word. A pointer has no width of its own in the IR, so it is
183/// the word.
184pub(crate) fn int_bytes(ty: Type, conv: &CallRegs) -> u32 {
185 if ty.is_ptr() { conv.word } else { ty.bits().saturating_mul(ty.lanes()).div_ceil(8).max(1) }
186}
187
188/// Whether a value of that type travels as bytes in the argument area because of what it is.
189///
190/// One type does, and it is the `long double`. SysV classifies it X87 and X87UP, which is the
191/// classification that means memory, so it goes where a structure the classification put in memory
192/// goes and what the two ends pass is the address of the bytes. That is not a decision about
193/// registers running out: a `long double` travels in the argument area when it is the only argument
194/// there is.
195///
196/// Sixteen bytes aligned to sixteen, which is what the psABI says the type takes and is the same
197/// number [`crate::lower`] gives one in the frame, so a value being passed and a value being worked
198/// on are the same shape of object in two places.
199#[must_use]
200pub fn on_the_stack(ty: Type) -> bool {
201 ty.is_float() && ty.is_scalar() && ty.bits() == 80
202}
203
204/// Whether what a function gives back is left on the x87 stack, which is one `long double` in
205/// `st(0)` or a `_Complex long double` with its real half there and its imaginary half in `st(1)`.
206///
207/// Those two are the whole of what SysV returns on that stack, and a structure holding a `long
208/// double` is not one of them, since the lowering sends that back through memory.
209#[must_use]
210pub fn back_on_x87(returns: &[Type]) -> bool {
211 matches!(returns.len(), 1 | 2) && returns.iter().all(|&ty| on_the_stack(ty))
212}
213
214/// How much room one takes in the argument area, as a size and an alignment.
215pub(crate) const X87_AREA: (u32, u32) = (16, 16);
216
217/// Why a value of that type cannot travel at all, or nothing if it can.
218///
219/// The width question and the file question in one place, so that the two ends of a call give the
220/// same answer about the same type, and so that a `return` this cannot make says the same thing
221/// about a type as the call that would have received it.
222///
223/// A `long double` is not one of them any more when it is an argument, since an argument of that
224/// type travels as bytes and [`on_the_stack`] is what says so before this is asked. What is left
225/// here is the value that comes back, because coming back is the one direction where it is not
226/// bytes: it arrives in `st(0)`, which is a register file this cannot name.
227#[must_use]
228pub fn refuses(ty: Type, insts: &Insts) -> Option<Missing> {
229 if (insts.arg)(ty).is_some() {
230 return None;
231 }
232 // A `long double` is the one type here that is in neither of the two files. Saying so is worth
233 // more than calling it a width, because eighty bits is a width this machine computes in and
234 // the file it computes in is what actually stands in the way.
235 if on_the_stack(ty) {
236 return Some(Missing::OnX87);
237 }
238 Some(Missing::Width)
239}
240
241/// What a function's parameters came to.
242#[derive(Debug, Default, Clone, PartialEq, Eq)]
243pub struct Arrived {
244 /// The register each parameter is in, in the order the parameters were given, so the caller can
245 /// bind each IR parameter to the one at its position.
246 pub regs: Vec<mir::Reg>,
247 /// The loads that read a parameter out of the caller's argument area, and how far up that area
248 /// each of them reads.
249 ///
250 /// Empty for almost every function, because almost every function has few enough parameters to
251 /// have been handed all of them in registers. The distance is from the bottom of the caller's
252 /// argument area, which is somewhere [`crate::finish`] works out and this cannot.
253 pub stack: Vec<(mir::Inst, u32)>,
254 /// How many general purpose argument registers the parameters took, and how many vector ones.
255 ///
256 /// Nothing about an ordinary function needs this. A variadic one does: the first argument its
257 /// signature does not name is the one after the last it does, so where each of the two walks
258 /// stopped is where `va_start` has to say the next argument begins.
259 pub took: (usize, usize),
260 /// How far up the caller's argument area the first argument the signature does not name is.
261 ///
262 /// However much of that area the named parameters took, on a convention that counts the two
263 /// register files apart, because there the area holds only the arguments no register was left
264 /// for. On one that counts them as one run it is not the same number: that area begins with the
265 /// shadow space the caller reserved, every argument owns one word of it whether it also arrived
266 /// in a register or not, and the named ones own the first few of those words. So it is where
267 /// the walk over the registers stopped, which is a position rather than a size, until the named
268 /// parameters have used every register and the two agree again.
269 pub beyond: u32,
270 /// The argument registers left over for the arguments the signature does not name, as the
271 /// register each was bound into and how far up the save area its slot is.
272 ///
273 /// Empty unless a save area was asked for. The ones a named parameter took are not here,
274 /// because their slots are behind where `va_start` sets the two offsets and nothing ever reads
275 /// them, so writing them would be fourteen stores where six are wanted.
276 pub spare: Vec<(mir::Reg, RegClass, u32)>,
277 /// The parameters that arrived in an argument register, as the position of each and how far up
278 /// the save area the slot of its register is.
279 ///
280 /// Empty unless a save area was asked for, and read only by a function that saves every
281 /// argument register rather than the spare ones, which is one holding `__builtin_apply_args`.
282 /// Together with [`Arrived::spare`] it is every register an argument can arrive in.
283 pub named: Vec<(usize, u32)>,
284}
285
286/// Binds a function's parameters to where the convention says they arrive.
287///
288/// # Errors
289///
290/// The first parameter this cannot bring in, and why. A function with one is reported rather
291/// than compiled, because the alternative is a function that reads an argument from wherever the
292/// last one happened to leave a register.
293pub fn entry(
294 out: &mut mir::Func,
295 block: mir::Block,
296 params: &[Param],
297 conv: &CallRegs,
298 insts: &Insts,
299 names: &mut Interner,
300 save: Option<Area>,
301) -> Result<Arrived, (usize, Missing)> {
302 // Where everything is, worked out before anything is written, both so that a parameter this
303 // cannot bring in stops the function before half of one is built and so that the two loops
304 // below can be two loops. Asking for the place of a parameter that cannot be brought in is
305 // still done, because every place after it depends on it and a reader stepping through this
306 // should see the same numbers a working version would.
307 let mut places = Places::new(conv);
308 let mut where_from = Vec::with_capacity(params.len());
309 for (index, &Param { ty, abi }) in params.iter().enumerate() {
310 // A structure the classification put in the argument area arrived as bytes, and the
311 // parameter the IR sees is a pointer to them. So there is nothing to bring in: the bytes
312 // are already in this function's frame, and what the pointer holds is where they are.
313 if let Abi::ByVal { size, align, drains } = abi {
314 let size = u32::try_from(size).map_err(|_| (index, Missing::TooBig))?;
315 where_from.push((ty, places.object(size, align), abi));
316 drain(&mut places, drains);
317 continue;
318 }
319 // And an eighty bit float, which arrives the same way for the same reason and is told
320 // apart only by the classification having said nothing about it: the front end passes it
321 // as a value of its own type, and it is this that knows the type is one that travels as
322 // bytes. What arrives is the address of those bytes, which is what an object in the
323 // argument area always hands over.
324 if on_the_stack(ty) {
325 let (size, align) = X87_AREA;
326 where_from.push((
327 ty,
328 places.on_stack(size, align),
329 Abi::ByVal { size: size.into(), align, drains: Drains::Nothing },
330 ));
331 continue;
332 }
333 let at = match (abi, conv.sret) {
334 // The address a result goes back through, on a convention with a register of its own
335 // for it, which takes no position from the arguments after it.
336 (Abi::Sret { .. }, Some(sret)) => Where::Reg(sret),
337 _ if ty.is_float() => places.float(float_bytes(ty)),
338 _ => places.integer(int_bytes(ty, conv)),
339 };
340 if let Some(missing) = refuses(ty, insts) {
341 return Err((index, missing));
342 }
343 where_from.push((ty, at, abi));
344 }
345 let took = (places.integers(), places.floats());
346 // Where the first argument the signature does not name is, which is the two sentences on
347 // [`Arrived::beyond`] written out. The run of words is contiguous from the bottom of the area on
348 // a convention that homes its register arguments, so a position multiplied by a word is the
349 // answer there until the positions run out and the named parameters start taking room of their
350 // own, at which point what they took is the answer again.
351 let reached = took.0 + took.1;
352 let beyond = match conv.shared_positions && reached < conv.int_args.len() {
353 true => conv.word * u32::try_from(reached).unwrap_or(0),
354 false => places.size(),
355 };
356 let mut arrived =
357 Arrived { regs: Vec::with_capacity(params.len()), took, beyond, ..Arrived::default() };
358
359 // Every pseudo first and everything else after, which is not a preference. A pseudo says a
360 // register holds an argument and defines nothing before it, so as far as the allocator can see
361 // the register was dead until then and is free to be used as a scratch. That is true of a
362 // register no pseudo has named yet, and it stops being true the moment one does. Anything that
363 // needs a scratch has to come after all of them, and a load out of the caller's stack needs one
364 // for the value it loads.
365 for (index, &(ty, at, _)) in where_from.iter().enumerate() {
366 let class = class_of(ty, conv);
367 let reg = out.new_vreg(class);
368 if class == conv.sse_class {
369 out.set_width(reg, float_bytes(ty));
370 }
371 arrived.regs.push(reg);
372 let Where::Reg(arrived_in) = at else { continue };
373 let head = (insts.arg)(ty).ok_or((index, Missing::Width))?;
374 let opcode = mir::Opcode::new(names.intern(head));
375 let operand = mir::Operand::write(reg, class).with(Constraint::Fixed(arrived_in));
376 out.build(block, opcode).operand(operand).finish();
377 if let Some(slot) = save.and_then(|area| slot_of(conv, area, arrived_in)) {
378 arrived.named.push((index, slot));
379 }
380 }
381 if let Some(area) = save {
382 arrived.spare = spare(out, block, conv, insts, names, area, arrived.took);
383 }
384
385 // The stack pointer is written down as the register to read through because it is the one that
386 // reaches the caller's stack in almost every function, and a realigned frame is the exception
387 // that [`crate::finish`] rewrites. Putting something here rather than nothing keeps the
388 // instruction printable and verifiable in between.
389 for (index, &(ty, at, abi)) in where_from.iter().enumerate() {
390 let Where::Stack(up) = at else { continue };
391 let class = class_of(ty, conv);
392 // The bytes of an object that travelled as bytes are read by whatever reads the parameter,
393 // and what the parameter is is their address, so this takes the address rather than a
394 // value out of it. Everything else about it is the load's, including the two fields
395 // [`crate::finish`] fills in, because where the caller's argument area is is the same
396 // question for both.
397 let name = match abi {
398 Abi::ByVal { .. } => insts.lea,
399 _ => (insts.load)(ty).ok_or((index, Missing::Width))?,
400 };
401 let opcode = mir::Opcode::new(names.intern(name));
402 let sp = mir::Operand::read(mir::Reg::physical(conv.stack_pointer), conv.int_class);
403 let made =
404 out.build(block, opcode).def(arrived.regs[index], class).mem(mir::Mem::at(sp)).finish();
405 arrived.stack.push((made, up));
406 }
407 Ok(arrived)
408}
409
410/// How far up a save area the slot of an argument register is, or nothing for a register the area
411/// has no slot for, which is the one a convention passes the address of a result in.
412fn slot_of(conv: &CallRegs, area: Area, reg: PhysReg) -> Option<u32> {
413 let files = [(false, conv.int_args), (true, conv.sse_args)];
414 files.into_iter().find_map(|(float, file)| {
415 let at = u32::try_from(file.iter().position(|&it| it == reg)?).ok()?;
416 (at < area.holds(float)).then(|| area.starts_at(float) + at * area.stride(float))
417 })
418}
419
420/// Binds the argument registers no parameter the signature names took, which are the ones the
421/// arguments it does not name arrived in.
422///
423/// One pseudo each and nothing else, for the reason the loop above them gives: what these do is say
424/// the register holds something, and the stores that put it in the save area are written by
425/// [`crate::lower`] once it has an address to store to, which is after every pseudo in the block.
426///
427/// Where each file's walk stopped is the file's own count on a convention that keeps two, and the
428/// sum of both on one that counts the files as a single run of positions, since there an argument
429/// of either kind steps the one counter. A file the area holds no slots of is skipped entirely,
430/// which is the vector file on the second kind: a variadic float travels in the general purpose
431/// register at its position as well, so the copy the walk reads is already the one being spilled.
432fn spare(
433 out: &mut mir::Func,
434 block: mir::Block,
435 conv: &CallRegs,
436 insts: &Insts,
437 names: &mut Interner,
438 area: Area,
439 took: (usize, usize),
440) -> Vec<(mir::Reg, RegClass, u32)> {
441 let word = Type::int(64);
442 let double = Type::float(rucc_ir::Float::F64);
443 let reached = |own: usize| if conv.shared_positions { took.0 + took.1 } else { own };
444 let files = [
445 (conv.int_args, reached(took.0), word, false),
446 (conv.sse_args, reached(took.1), double, true),
447 ];
448 let mut spare = Vec::new();
449 for (regs, taken, ty, float) in files {
450 let held = usize::try_from(area.holds(float)).unwrap_or(0);
451 let Some(head) = (insts.arg)(ty) else { continue };
452 let class = class_of(ty, conv);
453 for (index, &arrived_in) in regs.iter().enumerate().take(held).skip(taken) {
454 let reg = out.new_vreg(class);
455 let opcode = mir::Opcode::new(names.intern(head));
456 let operand = mir::Operand::write(reg, class).with(Constraint::Fixed(arrived_in));
457 out.build(block, opcode).operand(operand).finish();
458 let at = area.starts_at(float) + area.stride(float) * u32::try_from(index).unwrap_or(0);
459 spare.push((reg, class, at));
460 }
461 }
462 spare
463}
464
465/// The instructions this file writes, for one machine.
466///
467/// Every one of them is written here by hand rather than by a rule, for the reasons the module
468/// comment gives, so this file is where a machine's names for them have to come from. Each answer
469/// is the whole name, prefix and all, because each is interned as it is.
470///
471/// The four functions answer for the same set of types, and a type one of them has no name for is
472/// a type none of them should have: an argument that can arrive in a register but not be read off
473/// the stack is a function turned away for where its sixth argument happened to land.
474#[derive(Debug)]
475pub struct Insts {
476 /// The pseudo a parameter of that type arrives in a register as. See [`head_of`].
477 pub arg: fn(Type) -> Option<&'static str>,
478 /// The load that reads one of that type out of the argument area. See [`load_of`].
479 pub load: fn(Type) -> Option<&'static str>,
480 /// The store that writes one of that type into the argument area. See [`store_of`].
481 pub store: fn(Type) -> Option<&'static str>,
482 /// The pseudo a value of that type leaves in the register at that place. See [`ret_of`].
483 pub ret: fn(Type, usize) -> Option<&'static str>,
484 /// The call of a name. See [`CALL`].
485 pub call: &'static str,
486 /// The call of an address in a register. See [`CALL_REG`].
487 pub call_reg: &'static str,
488 /// The instruction that puts an address in a register without reading what is at it.
489 pub lea: &'static str,
490 /// The instruction that writes a small constant into a general purpose register, which is
491 /// what a byte count and a SysV vector count are.
492 pub small: &'static str,
493 /// The instruction that extends a value of that type the way the attribute asks, for the one
494 /// side of a call that owes the other the bits above it, and [`None`] where nothing does.
495 pub extend: fn(Type, Abi) -> Option<&'static str>,
496}
497
498/// The x86-64 ones.
499pub static X86_64: Insts = Insts {
500 arg: head_of,
501 load: load_of,
502 store: store_of,
503 ret: ret_of,
504 call: CALL,
505 call_reg: CALL_REG,
506 lea: "x64.lea_64",
507 small: "x64.mov_ri_32",
508 extend: |_, _| None,
509};
510
511/// What the instruction that calls a name is called.
512///
513/// Here rather than in a rule for the same reason the arguments are: a rule pattern sees one term
514/// and a call's operands are whatever the signature made them, so no pattern could name them.
515pub const CALL: &str = "x64.call";
516
517/// What the instruction that calls an address in a register is called.
518///
519/// A different instruction rather than the same one with a different operand, which is what the
520/// machine says too: one carries the distance to somewhere in the program and takes a relocation,
521/// and the other carries the register the address is in and takes none. Sharing an opcode would
522/// mean an instruction whose bytes depend on whether a field beside it happens to be set.
523pub const CALL_REG: &str = "x64.call_reg";
524
525/// One value a call passes.
526#[derive(Debug, Clone, Copy, PartialEq, Eq)]
527pub struct Passing {
528 /// The type it travels as, which for an object travelling as bytes is the pointer's rather
529 /// than the object's, because the pointer is what the machine IR has.
530 pub ty: Type,
531 /// The register holding it, or holding its address when the bytes are what travel.
532 pub reg: mir::Reg,
533 /// What the classification asked of it. The one thing read here is whether the object behind
534 /// the pointer is the argument, since everything else it can say is about a value that is
535 /// already in a register in the form it travels in.
536 pub abi: Abi,
537}
538
539/// What one call came to.
540#[derive(Debug, Clone, PartialEq, Eq)]
541pub struct Made {
542 /// The registers the value came back in, in the order the signature returns them, which is
543 /// empty for a call that gives nothing back and holds two for a structure that comes back in a
544 /// pair. Which register each of them is is the classification's answer and is worked out here
545 /// rather than in a table, for the reason the second half of [`Calling::returns`] gives.
546 pub results: Vec<mir::Reg>,
547 /// How many bytes below the stack pointer this call needs for the arguments it passes there.
548 ///
549 /// Not always zero for a call that passes everything in registers: a Windows caller reserves
550 /// thirty two bytes for the callee to spill its register arguments into whether it uses them
551 /// or not, and that reservation is this.
552 pub outgoing: u32,
553}
554
555/// Which of a call's values could not be passed, and why.
556#[derive(Debug, Clone, Copy, PartialEq, Eq)]
557pub struct Refused {
558 /// Its position among the arguments, or `None` for the value that comes back.
559 pub argument: Option<usize>,
560 /// What is wrong with where it travels.
561 pub missing: Missing,
562}
563
564/// What a call goes to.
565///
566/// The whole of the difference between the two calls. Everything else about them, which is what
567/// they pass and what comes back and which registers they destroy, is the signature's answer and
568/// is the same answer either way.
569#[derive(Debug, Clone, Copy, PartialEq, Eq)]
570pub enum Callee {
571 /// A name, which the linker resolves.
572 Named(Symbol),
573 /// An address in a register, which nothing resolves because there is nothing to resolve: the
574 /// value is not known until the program runs.
575 ///
576 /// The register is unconstrained, and it has to be, because every register the convention
577 /// does not preserve is one this instruction writes and every register an argument travels in
578 /// is spoken for. What is left is the registers the callee has to put back, which is where
579 /// the allocator will put the address, and it is the right answer for the same reason it is
580 /// the only one.
581 Through(mir::Reg),
582}
583
584/// One call, as everything about it that is not the function it is being built into.
585#[derive(Debug, Clone, Copy)]
586pub struct Calling<'a> {
587 /// What it calls.
588 pub callee: Callee,
589 /// What it passes, in the order the signature holds them, which is the order the convention
590 /// places them in.
591 pub args: &'a [Passing],
592 /// What comes back, which is empty for a call that gives nothing back, one type for a value,
593 /// and two for a structure small enough to come back in a pair of registers.
594 ///
595 /// A pair is placed here rather than named by a rule for the reason the arguments are: which
596 /// register each half goes in depends on the halves before it, since the two files are walked
597 /// separately, and a pattern over a term cannot see them.
598 pub returns: &'a [Type],
599 /// Whether the callee takes arguments beyond the ones its signature names, which is what says
600 /// whether it reads the count of vector registers the call passed arguments in.
601 pub variadic: bool,
602 /// How many of the arguments the signature does name, so that the ones past it can be told
603 /// apart from the ones before it.
604 ///
605 /// A convention that passes a variadic float in both register files needs that, because which
606 /// arguments get the second copy is exactly the ones the callee has no prototype for. Every
607 /// other convention treats the two the same and never asks.
608 pub named: usize,
609 /// Where the call was written, which every instruction built for it is filed under.
610 ///
611 /// A call is one of the few places in the machine IR where a run of instructions comes from no
612 /// term in the IR at all: the stores into the outgoing area, the copies a structure passed by
613 /// value is made of and the call itself are the convention's answer rather than anything a rule
614 /// matched. So there is nothing for them to inherit a span from, and without this the bytes of
615 /// an entire call statement are covered by whichever row came before them, which is usually the
616 /// line above. gcc names the line the call is written on over all of it.
617 ///
618 /// [`Span::DUMMY`] in a call this crate builds for itself, which is the copy into the argument
619 /// area that the runtime does, since that one is under whatever the call it belongs to is under.
620 pub at: Span,
621}
622
623/// Builds one call: what it passes, what comes back, and what it destroys.
624///
625/// # Errors
626///
627/// The first value this cannot pass, and why, before anything is written. A call with one is
628/// reported rather than compiled, because the alternative is a call that leaves an argument
629/// wherever the last one happened to put a register.
630///
631/// # Panics
632///
633/// If a call passes two gigabytes of arguments on the stack, which is a distance no offset in a
634/// frame can hold and a call no program makes.
635pub fn call(
636 out: &mut mir::Func,
637 block: mir::Block,
638 made: &Calling<'_>,
639 conv: &CallRegs,
640 insts: &Insts,
641 names: &mut Interner,
642) -> Result<Made, Refused> {
643 let &Calling { callee, args, returns, variadic, named, at: span } = made;
644 // Where everything goes, worked out before anything is built, so that a call this cannot make
645 // leaves no half of one behind.
646 let mut places = Places::new(conv);
647 let mut passed = Vec::with_capacity(args.len());
648 // The ones with no register left for them, as the store each of them becomes and how far up
649 // the outgoing area it writes. Almost always empty.
650 let mut on_stack = Vec::new();
651 // How many of them went in vector registers, which is what a SysV variadic callee is told.
652 let mut vectors = 0u32;
653 // The ones whose bytes travel rather than their address, as the register that address is in,
654 // how far up the outgoing area they go and which words the copy is made of. Almost always
655 // empty too, and never at the same time as a register: an object in the argument area is in
656 // the argument area whatever is left of the register files.
657 let mut as_bytes = Vec::new();
658 // The arguments beyond the ones the signature names that travel in a vector register and have
659 // to travel in a general purpose one at the same time, as the register each is in and the
660 // general purpose register its copy belongs in. Empty on every convention but the one that says
661 // so, and on that one this is what makes `printf("%f", x)` read the right register.
662 let mut in_both = Vec::new();
663 // Which argument position the next one that gets a register is at, which is only the same as
664 // the index when nothing ahead of it went to memory. A convention that counts the two register
665 // files as one run is what needs it: the general purpose register a variadic float's second
666 // copy goes in is the one at that position.
667 let mut position = 0usize;
668 // The ones the callee reads as a whole 32 bit register when they are narrower than that, as
669 // where in `passed` each is and the instruction that makes its upper bits what the callee
670 // expects. Only an argument in a register, since one in memory is read at its own width.
671 let mut widen = Vec::new();
672 for (index, &Passing { ty, reg, abi }) in args.iter().enumerate() {
673 let refused = |missing| Refused { argument: Some(index), missing };
674 // An eighty bit float is bytes in the argument area whatever the classification said, for
675 // the reason [`on_the_stack`] gives, and the register holding it holds their address. So it
676 // joins the objects below rather than being a case of its own, and the copy it becomes is
677 // the copy any other sixteen byte object gets.
678 let abi = match abi {
679 _ if on_the_stack(ty) => {
680 let (size, align) = X87_AREA;
681 Abi::ByVal { size: size.into(), align, drains: Drains::Nothing }
682 }
683 abi => abi,
684 };
685 if let Abi::ByVal { size, align, drains } = abi {
686 let size = u32::try_from(size).map_err(|_| refused(Missing::TooBig))?;
687 // One past the `...` is never packed, so it is words wherever it is.
688 let at = if variadic && index >= named {
689 places.on_stack(size, align)
690 } else {
691 places.object(size, align)
692 };
693 drain(&mut places, drains);
694 let Where::Stack(up) = at else {
695 unreachable!("an object in the argument area is in the argument area")
696 };
697 // Nothing here refuses a plan it did not get, because a copy the words give up on is
698 // a call to the runtime below. What the byte count has to fit in is the immediate the
699 // call passes it as, and an object that large is what is left of the old refusal.
700 let plan = crate::expand::plan(u64::from(size), align, conv.word);
701 let count = i32::try_from(size).map_err(|_| refused(Missing::TooBig))?;
702 as_bytes.push((reg, up, count, plan));
703 continue;
704 }
705 // The address a result comes back through goes in its own register where the convention
706 // has one, and is not an argument position, the same as on the other side in [`entry`].
707 if let (Abi::Sret { .. }, Some(sret)) = (abi, conv.sret) {
708 passed.push((reg, sret, class_of(ty, conv)));
709 continue;
710 }
711 // Apple's AArch64 puts every argument past the named ones in memory, a word or more each,
712 // whatever registers are left.
713 let unnamed = variadic && index >= named && conv.abi.variadic == Variadic::AlwaysMemory;
714 let at = if unnamed {
715 let bytes = int_bytes(ty, conv);
716 places.on_stack(bytes, bytes)
717 } else if ty.is_float() {
718 places.float(float_bytes(ty))
719 } else {
720 places.integer(int_bytes(ty, conv))
721 };
722 if let Some(missing) = refuses(ty, insts) {
723 return Err(refused(missing));
724 }
725 let class = class_of(ty, conv);
726 match at {
727 Where::Reg(at) => {
728 if let Some(name) = (insts.extend)(ty, abi) {
729 widen.push((passed.len(), names.intern(name)));
730 }
731 // Only a register counts, because the count is of registers. An argument that went
732 // to memory is one the callee reads from memory whatever this says.
733 if class == conv.sse_class {
734 vectors += 1;
735 // Windows passes a float the callee has no prototype for in the vector register
736 // and in the general purpose register at the same position, both at once,
737 // because the callee has no way to know which file to look in and its walk over
738 // the arguments reads the second one. An argument the signature does name needs
739 // no second copy, since the callee's parameter says where it is.
740 let both = conv.shared_positions && variadic && index >= named;
741 if let Some(&also) = conv.int_args.get(position).filter(|_| both) {
742 in_both.push((reg, also));
743 }
744 }
745 passed.push((reg, at, class));
746 position += 1;
747 }
748 Where::Stack(up) => {
749 let store = (insts.store)(ty).ok_or(refused(Missing::Width))?;
750 on_stack.push((reg, class, names.intern(store), up));
751 }
752 }
753 }
754 // A `long double` comes back in `st(0)`, which is not a register in either file and not one
755 // this call can be said to write. So nothing is placed for it and nothing is constrained, and
756 // the call gives back no register at all: what takes the value off that stack is the `fstp`
757 // [`crate::lower`] writes straight after the call, which is the same shape every other use of
758 // the x87 stack is written in. A complex one is the same with its imaginary half in `st(1)`,
759 // and a second `fstp` takes that one off.
760 let comes_back =
761 if back_on_x87(returns) { Vec::new() } else { places_back(returns, conv, insts)? };
762
763 // A variadic callee on SysV reads how many vector registers the call passed arguments in and
764 // skips saving them when the answer is none, which is what makes `printf` with no floating
765 // point argument cheap. It is an obligation rather than an optimization: leaving whatever was
766 // in the register there makes the callee save a register file it was not given, and a count
767 // that is too low makes it read an argument out of a register nothing put one in.
768 let counted = if variadic { conv.vector_count } else { None };
769
770 for (at, name) in widen {
771 let (reg, _, class) = passed[at];
772 let wide = out.new_vreg(class);
773 out.build(block, mir::Opcode::new(name))
774 .at(span)
775 .def(wide, class)
776 .uses(reg, class)
777 .finish();
778 passed[at].0 = wide;
779 }
780
781 // The arguments that go to memory go there now, in front of the call and after everything this
782 // could have refused, so that a call it cannot make leaves no store behind either. The offset
783 // is written straight in rather than left for [`crate::finish`]: the outgoing area is at the
784 // bottom of the frame because that is where the callee looks for it, and the bottom of the
785 // frame is where the stack pointer already is.
786 for (reg, class, store, up) in on_stack {
787 let sp = mir::Operand::read(mir::Reg::physical(conv.stack_pointer), conv.int_class);
788 let up = i32::try_from(up).expect("an argument area under two gigabytes");
789 let build = out.build(block, mir::Opcode::new(store)).at(span);
790 build.uses(reg, class).mem(mir::Mem::at(sp).plus(up)).finish();
791 }
792
793 // How many bytes the copies below needed for calls of their own, which is nothing on a
794 // convention that passes three pointers in registers and thirty two bytes on the one that
795 // reserves a place for them anyway. The frame has to hear about it, since a call built here is
796 // still a call this function makes.
797 let mut nested = 0u32;
798
799 // And the objects whose bytes go there, as a load and a store for each word of each of them.
800 // This is the copy the caller owes a callee that takes a structure by value: the callee is
801 // free to write to what it was handed, so what it was handed cannot be the caller's own copy,
802 // and the argument area is where the convention says the caller's copy goes. The words are the
803 // same words `crate::expand` would have chosen for a `memcpy` of the same block, because they
804 // are chosen by the same function.
805 for (from, up, count, plan) in as_bytes {
806 let up = i32::try_from(up).expect("an argument area under two gigabytes");
807 let Some(plan) = plan else {
808 let what = Copying { from, up, count, span };
809 nested = nested.max(by_runtime(out, block, conv, insts, names, what));
810 continue;
811 };
812 for (at, width) in plan {
813 let ty = Type::int(width * 8);
814 let at = i32::try_from(at).expect("an object under two gigabytes");
815 let word = out.new_vreg(conv.int_class);
816 let load = names.intern(
817 (insts.load)(ty).ok_or(Refused { argument: None, missing: Missing::Width })?,
818 );
819 let there = mir::Operand::read(from, conv.int_class);
820 let build = out.build(block, mir::Opcode::new(load)).at(span);
821 build.def(word, conv.int_class).mem(mir::Mem::at(there).plus(at)).finish();
822 let store = names.intern(
823 (insts.store)(ty).ok_or(Refused { argument: None, missing: Missing::Width })?,
824 );
825 let sp = mir::Operand::read(mir::Reg::physical(conv.stack_pointer), conv.int_class);
826 let build = out.build(block, mir::Opcode::new(store)).at(span);
827 build.uses(word, conv.int_class).mem(mir::Mem::at(sp).plus(up + at)).finish();
828 }
829 }
830
831 // And the second copy of each float the callee has no prototype for, which is one `movq` out of
832 // the vector register it is already in. What the callee reads out of the general purpose
833 // register is the sixty four bits and not a value of any type, so the bits are what move, and
834 // the copy joins the arguments rather than being a thing of its own: it is passed in a register
835 // the convention names, which is what every other argument here is.
836 for (from, into) in in_both {
837 let word = out.new_vreg(conv.int_class);
838 let movq = mir::Opcode::new(names.intern("x64.movq_from_xmm"));
839 out.build(block, movq)
840 .at(span)
841 .def(word, conv.int_class)
842 .uses(from, conv.sse_class)
843 .finish();
844 passed.push((word, into, conv.int_class));
845 }
846
847 // The definitions first and the reads after, which is the order every operand vector in the
848 // machine IR is in and the order `rucc_mir::defs` counts.
849 let mut operands = Vec::with_capacity(args.len() + conv.int_order.len() + 2);
850 let results: Vec<mir::Reg> = comes_back
851 .iter()
852 .map(|&(at, class)| {
853 let reg = out.new_vreg(class);
854 operands.push(mir::Operand::write(reg, class).with(Constraint::Fixed(at)));
855 reg
856 })
857 .collect();
858 // One list per file, because a physical register is a number and the class is what says which
859 // file it is a number in. One list would have `xmm0` blocking `rax`.
860 let spoken_for = |class: RegClass| -> Vec<PhysReg> {
861 comes_back
862 .iter()
863 .filter(|&&(_, at)| at == class)
864 .map(|&(reg, _)| reg)
865 .chain(counted.filter(|_| class == conv.int_class))
866 .chain(passed.iter().filter(|&&(_, _, at)| at == class).map(|&(_, reg, _)| reg))
867 .collect()
868 };
869 let named = spoken_for(conv.int_class);
870 for ® in conv.int_order {
871 if !conv.preserves_int(reg) && !named.contains(®) {
872 operands.push(mir::Operand::write(mir::Reg::physical(reg), conv.int_class));
873 }
874 }
875 let named = spoken_for(conv.sse_class);
876 for ® in conv.sse_order {
877 if named.contains(®) {
878 continue;
879 }
880 // One the convention keeps only the bottom of is written above that, which leaves a value
881 // narrow enough to fit where it is and takes the register from anything wider.
882 let clobber = mir::Operand::write(mir::Reg::physical(reg), conv.sse_class);
883 match conv.sse_kept {
884 _ if !conv.preserves_sse(reg) => operands.push(clobber),
885 Some(kept) => operands.push(clobber.with(Constraint::Above(kept))),
886 None => {}
887 }
888 }
889 // The address in front of the arguments, because a call through one is written with the
890 // register it goes through and nothing in the operand vector is at a place a table could name.
891 // First read is a place that does not depend on the signature, which is what
892 // [`rucc_target::x86_64::Arg::Through`] is written against.
893 if let Callee::Through(reg) = callee {
894 operands.push(mir::Operand::read(reg, conv.int_class));
895 }
896 for (reg, at, class) in passed {
897 operands.push(mir::Operand::read(reg, class).with(Constraint::Fixed(at)));
898 }
899 if let Some(at) = counted {
900 let count = out.new_vreg(conv.int_class);
901 let zero = mir::Opcode::new(names.intern(insts.small));
902 out.build(block, zero).at(span).def(count, conv.int_class).imm(i64::from(vectors)).finish();
903 operands.push(mir::Operand::read(count, conv.int_class).with(Constraint::Fixed(at)));
904 }
905
906 let opcode = mir::Opcode::new(names.intern(match callee {
907 Callee::Named(_) => insts.call,
908 Callee::Through(_) => insts.call_reg,
909 }));
910 let mut build = out.build(block, opcode).at(span);
911 if let Callee::Named(symbol) = callee {
912 build = build.symbol(symbol);
913 }
914 for operand in operands {
915 build = build.operand(operand);
916 }
917 build.finish();
918 Ok(Made { results, outgoing: places.size().max(nested) })
919}
920
921/// One object whose bytes go into the argument area by a call to the runtime.
922#[derive(Debug, Clone, Copy)]
923struct Copying {
924 /// The register its address is in.
925 from: mir::Reg,
926 /// How far up the outgoing area its copy goes.
927 up: i32,
928 /// How many bytes it is.
929 count: i32,
930 /// Where the call it is an argument of was written.
931 span: Span,
932}
933
934/// One object with more words than a copy into the argument area unrolls to, copied there by a
935/// call to the runtime, and how many bytes that call itself needed below the stack pointer.
936///
937/// Which routine it is is [`crate::capability`]'s answer and not a name written here, because a
938/// call standing in for an operation the machine has no instruction for is what that table is a
939/// list of, and a copy too large to unroll is already on it: this is the same row `crate::expand`
940/// reads for a `memcpy` in the IR of the same size.
941fn by_runtime(
942 out: &mut mir::Func,
943 block: mir::Block,
944 conv: &CallRegs,
945 insts: &Insts,
946 names: &mut Interner,
947 what: Copying,
948) -> u32 {
949 let Copying { from, up, count, span } = what;
950 let routine = capability::libcall(rucc_ir::Opcode::Memcpy, "big")
951 .expect("the runtime copies a block too large to unroll");
952
953 // Where the copy goes, which is a distance up the outgoing area and so is the stack pointer
954 // plus that distance. A `lea` rather than an add, because the stack pointer is not this
955 // function's to move and what the call wants is the address in a register of the allocator's
956 // choosing.
957 let sp = mir::Operand::read(mir::Reg::physical(conv.stack_pointer), conv.int_class);
958 let into = out.new_vreg(conv.int_class);
959 let lea = mir::Opcode::new(names.intern(insts.lea));
960 out.build(block, lea)
961 .at(span)
962 .def(into, conv.int_class)
963 .mem(mir::Mem::at(sp).plus(up))
964 .finish();
965
966 // And how many bytes, which C takes as a `size_t` and this has as a number. A thirty two bit
967 // move carries it, because writing the low half of a general purpose register clears the high
968 // half, so the sixty four bit count it becomes is the count as long as the count fits in the
969 // immediate, which is what the caller checked before anything was built.
970 let bytes = out.new_vreg(conv.int_class);
971 let mov = mir::Opcode::new(names.intern(insts.small));
972 out.build(block, mov).at(span).def(bytes, conv.int_class).imm(i64::from(count)).finish();
973
974 let args = [
975 Passing { ty: Type::PTR, reg: into, abi: Abi::Plain },
976 Passing { ty: Type::PTR, reg: from, abi: Abi::Plain },
977 Passing { ty: Type::int(conv.word * 8), reg: bytes, abi: Abi::Plain },
978 ];
979 let made = Calling {
980 callee: Callee::Named(names.intern(routine)),
981 args: &args,
982 returns: &[],
983 variadic: false,
984 named: args.len(),
985 at: span,
986 };
987 // Three pointer sized arguments and nothing coming back is a call every convention here has
988 // registers for, so the only way this could refuse is a convention with fewer than three
989 // argument registers, and there is no such convention.
990 call(out, block, &made, conv, insts, names)
991 .expect("the runtime's copy passes three words and takes nothing back")
992 .outgoing
993}
994
995/// Which register each value comes back in, walked the way the arguments are.
996///
997/// The two files are counted separately, because a structure of a `double` and a `long` comes back
998/// with the `double` in the first vector register and the `long` in the first integer one, and a
999/// single count would put the second half one place further along a list it is not on.
1000///
1001/// # Errors
1002///
1003/// The first value that cannot come back at all, and why, so that a call this cannot make leaves
1004/// nothing behind. Nothing here reports which value it was, because the caller has one answer for
1005/// all of them: the value that comes back is not an argument and has no position among them.
1006fn places_back(
1007 returns: &[Type],
1008 conv: &CallRegs,
1009 insts: &Insts,
1010) -> Result<Vec<(PhysReg, RegClass)>, Refused> {
1011 let refused = |missing| Refused { argument: None, missing };
1012 let mut back = Vec::with_capacity(returns.len());
1013 let (mut ints, mut sses) = (0usize, 0usize);
1014 for &ty in returns {
1015 if let Some(missing) = refuses(ty, insts) {
1016 return Err(refused(missing));
1017 }
1018 let class = class_of(ty, conv);
1019 let (file, at) = if class == conv.sse_class {
1020 (conv.sse_returns, &mut sses)
1021 } else {
1022 (conv.int_returns, &mut ints)
1023 };
1024 let reg = *file.get(*at).ok_or_else(|| refused(Missing::NoRoom))?;
1025 *at += 1;
1026 back.push((reg, class));
1027 }
1028 Ok(back)
1029}
1030
1031/// Which of the four widths a value travels at, where a truth value travels as the byte it lives
1032/// in.
1033///
1034/// [`crate::term::slot`] is the question the rule set asks and it answers nothing for one bit,
1035/// because there is no register of that width and so no instruction written at it. A convention
1036/// asks a different question. It has nothing narrower than a byte to put an argument in either,
1037/// and what it says about the one type that is a bit is that the byte holding it is the argument,
1038/// with the seven bits above unspecified. So the two lists differ by exactly this entry.
1039///
1040/// It is written here and not in [`crate::term::slot`] because moving it there would tell the rule
1041/// set that one bit is a byte, and then every byte rule in the file would match a term that is not
1042/// one. What the convention needs is narrower: a name for the register an argument arrives in, and
1043/// that name says a width because a listing is easier to read when it does.
1044fn place(ty: Type) -> Option<usize> {
1045 if crate::term::is_bit(ty) { Some(0) } else { crate::term::slot(ty) }
1046}
1047
1048/// What the pseudo for an argument of that type is called.
1049///
1050/// The width is in the name for the same reason it is in every other opcode here: it is what the
1051/// instruction is about. Nothing encodes it, so nothing depends on it being right, but a listing
1052/// that says an argument arrived and does not say how much of it did is a listing worth less.
1053///
1054/// Which widths there are is `place` above, and not a list of its own, because it has to be the
1055/// same list the three below use. An argument brought in at a width nothing downstream has a name
1056/// for is a register nothing could then read, and a width the others cover that this refuses is a
1057/// function turned away for no reason. Asking one question in one place is what keeps the four
1058/// answers from drifting, and an address is what they used to disagree about.
1059#[must_use]
1060pub fn head_of(ty: Type) -> Option<&'static str> {
1061 // The format that fills a whole vector register, which travels in one of them: the psABI
1062 // classifies it SSE and SSEUP, and those two eightbytes are the one register the pair names
1063 // rather than two registers.
1064 if crate::term::is_quad(ty) {
1065 return Some("x64.arg_val_f128");
1066 }
1067 // The half, which arrives in the low sixteen bits of a vector register the same way a `float`
1068 // arrives in the low thirty two. It is a name of its own rather than the `f32` one for the
1069 // reason every other width here has a name of its own: what arrived is two bytes, and a
1070 // listing that said four would be saying something the convention does not.
1071 if crate::term::is_half(ty) {
1072 return Some("x64.arg_val_f16");
1073 }
1074 if let Some(at) = crate::term::float_slot(ty) {
1075 return Some(["x64.arg_val_f32", "x64.arg_val_f64"][at]);
1076 }
1077 let names = ["x64.arg_val_8", "x64.arg_val_16", "x64.arg_val_32", "x64.arg_val_64"];
1078 Some(names[place(ty)?])
1079}
1080
1081/// What the instruction that reads an argument of that type out of memory is called.
1082///
1083/// Keyed off the same two questions [`head_of`] asks and answering for the same set of types, so
1084/// that a parameter this compiler can bring in from a register is one it can bring in from the
1085/// caller's stack as well. A width one of them covered and the other did not would be a function
1086/// turned away for where its sixth argument happened to land.
1087///
1088/// Reading a narrow argument at its own width and not at a word is deliberate. The caller wrote a
1089/// whole word, but what it put in the part above the value is not something the convention says, so
1090/// the bits this reads are exactly the bits that mean anything. That is the same thing an argument
1091/// arriving in a register gets: `x64.arg_val_8` says the low byte of that register is the argument
1092/// and says nothing at all about the rest of it.
1093#[must_use]
1094pub fn load_of(ty: Type) -> Option<&'static str> {
1095 // Sixteen bytes, which is the whole register and is also the whole value, so the instruction
1096 // a spill uses and the instruction an argument uses are the same one here. They are two
1097 // different instructions at the two narrower formats because there the value is part of the
1098 // register, and at this format there is no part of it to leave behind.
1099 if crate::term::is_quad(ty) {
1100 return Some("x64.movaps_rm");
1101 }
1102 // Two bytes out of the argument area and into the low lane of a vector register, which is one
1103 // instruction and is the same one a rule writes for a program's own read of a `_Float16`.
1104 if crate::term::is_half(ty) {
1105 return Some("x64.pinsrw_rm");
1106 }
1107 if let Some(at) = crate::term::float_slot(ty) {
1108 return Some(["x64.movss_rm", "x64.movsd_rm"][at]);
1109 }
1110 let names = ["x64.mov_rm_8", "x64.mov_rm_16", "x64.mov_rm_32", "x64.mov_rm_64"];
1111 Some(names[place(ty)?])
1112}
1113
1114/// What the instruction that writes an argument of that type into memory is called.
1115///
1116/// The mirror of [`load_of`], keyed off the same two questions and answering for the same set of
1117/// types, so that the two ends of one call agree about what travels. A type a callee can read out
1118/// of the argument area and a caller cannot write into it would be a call turned away for a reason
1119/// the function it calls does not have.
1120///
1121/// Writing a narrow argument at its own width leaves whatever was already in the rest of the word.
1122/// That is allowed, and it is what [`load_of`] is written against: the convention does not say what
1123/// is above the value, so the callee reads only the bits that mean anything and neither end has to
1124/// agree about the rest.
1125#[must_use]
1126pub fn store_of(ty: Type) -> Option<&'static str> {
1127 if crate::term::is_quad(ty) {
1128 return Some("x64.movaps_mr");
1129 }
1130 // The half goes out four bytes wide, which is the one place here where the instruction is not
1131 // the width of the value. SSE2 has no store of sixteen bits out of a vector register: the form
1132 // of `pextrw` that writes memory arrived with SSE4.1 and is above this target's baseline, so
1133 // the choices are a four byte store or a pair of instructions through a general purpose
1134 // register, and a pair is not something one name can be.
1135 //
1136 // Four bytes is safe here and would not be everywhere. An argument on the stack sits in an
1137 // eightbyte of its own, the paragraph above says the convention promises nothing about the
1138 // part of it above the value, and [`load_of`] reads back exactly the two bytes that mean
1139 // anything. What lands in the two bytes beside them is whatever was in the register, which is
1140 // no more than any other narrow argument leaves behind.
1141 if crate::term::is_half(ty) {
1142 return Some("x64.movss_mr");
1143 }
1144 if let Some(at) = crate::term::float_slot(ty) {
1145 return Some(["x64.movss_mr", "x64.movsd_mr"][at]);
1146 }
1147 let names = ["x64.mov_mr_8", "x64.mov_mr_16", "x64.mov_mr_32", "x64.mov_mr_64"];
1148 Some(names[place(ty)?])
1149}
1150
1151/// What the instruction that leaves a returned value in its register is called, for the value at
1152/// that place in its own register file.
1153///
1154/// Keyed off the same two questions [`head_of`] asks, so a type this can give back is a type it can
1155/// take in. The place is the one a call counted to when it laid the return out, which is per file
1156/// rather than over the whole list: a structure of a `double` and a `long` gives both of them back
1157/// at place zero.
1158///
1159/// The register itself is not here. It is in the operand table in `rucc_target::x86_64`, which is
1160/// where the first one has always been, and the two names below are how a value says which of the
1161/// two it is. A convention with more than two registers to come back in would need more names, and
1162/// there is none, which is what the `None` at the end is about.
1163#[must_use]
1164pub fn ret_of(ty: Type, at: usize) -> Option<&'static str> {
1165 if crate::term::is_quad(ty) {
1166 return Some(*["x64.ret_val_f128", "x64.ret_val2_f128"].get(at)?);
1167 }
1168 if crate::term::is_half(ty) {
1169 return Some(*["x64.ret_val_f16", "x64.ret_val2_f16"].get(at)?);
1170 }
1171 if let Some(width) = crate::term::float_slot(ty) {
1172 let names =
1173 [["x64.ret_val_f32", "x64.ret_val_f64"], ["x64.ret_val2_f32", "x64.ret_val2_f64"]];
1174 return Some(names.get(at)?[width]);
1175 }
1176 let names = [
1177 ["x64.ret_val_8", "x64.ret_val_16", "x64.ret_val_32", "x64.ret_val_64"],
1178 ["x64.ret_val2_8", "x64.ret_val2_16", "x64.ret_val2_32", "x64.ret_val2_64"],
1179 ];
1180 Some(names.get(at)?[place(ty)?])
1181}
1182
1183#[cfg(test)]
1184mod tests {
1185 use rucc_target::Convention;
1186 use rucc_target::x86_64::{REGS, SYSV, WIN64};
1187
1188 use super::*;
1189
1190 /// Those types as parameters that travel as the values they are, which is every one of them
1191 /// that is not a structure the classification put in the argument area.
1192 fn plain(params: &[Type]) -> Vec<Param> {
1193 params.iter().copied().map(Param::new).collect()
1194 }
1195
1196 /// The parameters of a function under a convention, as machine IR text.
1197 fn bind(params: &[Type], conv: &CallRegs) -> String {
1198 let mut names = Interner::new();
1199 let mut out = mir::Func::new(names.intern("f"));
1200 let block = out.create_block();
1201 entry(&mut out, block, &plain(params), conv, &X86_64, &mut names, None)
1202 .expect("every parameter arrives");
1203 mir::print_func(&out, &names, ®S)
1204 }
1205
1206 #[test]
1207 fn the_first_arguments_arrive_where_the_convention_puts_them() {
1208 let i32 = Type::int(32);
1209 assert_eq!(
1210 bind(&[i32, i32, Type::int(64)], &SYSV),
1211 "mfunc @f {\nblock0:\n %0:gpr($rdi) = x64.arg_val_32\n \
1212 %1:gpr($rsi) = x64.arg_val_32\n %2:gpr($rdx) = x64.arg_val_64\n}\n"
1213 );
1214 }
1215
1216 #[test]
1217 fn the_other_convention_puts_the_same_arguments_somewhere_else() {
1218 // The first argument is in `rcx` here and in `rdi` above, which is the difference that
1219 // makes a SysV binary calling a Windows one read the wrong value rather than fail.
1220 let i64 = Type::int(64);
1221 assert_eq!(
1222 bind(&[i64, i64], &WIN64),
1223 "mfunc @f {\nblock0:\n %0:gpr($rcx) = x64.arg_val_64\n \
1224 %1:gpr($rdx) = x64.arg_val_64\n}\n"
1225 );
1226 }
1227
1228 /// The parameters of a function under a convention, and what each of the ones that arrived in
1229 /// memory is waiting on.
1230 fn arrive(params: &[Type], conv: &CallRegs) -> (String, Vec<u32>) {
1231 let mut names = Interner::new();
1232 let mut out = mir::Func::new(names.intern("f"));
1233 let block = out.create_block();
1234 let arrived = entry(&mut out, block, &plain(params), conv, &X86_64, &mut names, None)
1235 .expect("every parameter");
1236 let up = arrived.stack.iter().map(|&(_, up)| up).collect();
1237 (mir::print_func(&out, &names, ®S), up)
1238 }
1239
1240 #[test]
1241 fn an_argument_past_the_last_register_is_read_out_of_the_caller_s_stack() {
1242 let (text, up) = arrive(&[Type::int(64); 7], &SYSV);
1243
1244 // Six of them got registers and the seventh did not, so the seventh is a load rather than
1245 // a pseudo. It reads through the stack pointer with nothing in its displacement, because
1246 // where the caller's argument area is from in here is a distance into a frame that does
1247 // not exist yet, and it is at the bottom of that area because it is the first one in it.
1248 assert_eq!(up, [0]);
1249 assert!(text.contains("%6:gpr = x64.mov_rm_64 [$rsp]"), "{text}");
1250 assert_eq!(text.matches("x64.arg_val_64").count(), 6, "{text}");
1251 }
1252
1253 #[test]
1254 fn the_other_convention_runs_out_of_registers_three_arguments_earlier() {
1255 let (text, up) = arrive(&[Type::int(64); 7], &WIN64);
1256
1257 // Windows passes four integers in registers and reserves thirty two bytes below the call
1258 // whether they are used or not, so the fifth argument is not at the bottom of the argument
1259 // area but above the shadow space, and the three after it follow it a word at a time.
1260 assert_eq!(up, [32, 40, 48]);
1261 assert_eq!(text.matches("x64.arg_val_64").count(), 4, "{text}");
1262 assert!(text.contains("%4:gpr = x64.mov_rm_64 [$rsp]"), "{text}");
1263 }
1264
1265 /// The parameters of a function under a convention, with one of them a structure whose bytes
1266 /// travel, and what each of the ones that arrived in memory is waiting on.
1267 fn arrive_with(params: &[Param], conv: &CallRegs) -> (String, Vec<u32>) {
1268 let mut names = Interner::new();
1269 let mut out = mir::Func::new(names.intern("f"));
1270 let block = out.create_block();
1271 let arrived = entry(&mut out, block, params, conv, &X86_64, &mut names, None)
1272 .expect("every parameter");
1273 let up = arrived.stack.iter().map(|&(_, up)| up).collect();
1274 (mir::print_func(&out, &names, ®S), up)
1275 }
1276
1277 #[test]
1278 fn a_structure_that_arrived_as_bytes_is_an_address_and_not_a_load() {
1279 let byval =
1280 Param::with_abi(Type::PTR, Abi::ByVal { size: 32, align: 8, drains: Drains::Nothing });
1281 let (text, up) =
1282 arrive_with(&[Param::new(Type::int(32)), byval, Param::new(Type::int(32))], &SYSV);
1283
1284 // `int f(int a, struct Big b, int c)`. The bytes of `b` are already in this function, at
1285 // the bottom of the caller's argument area, so nothing is read out of them here: what the
1286 // parameter is is where they are, which is one address. The two integers still travel in
1287 // registers, because an object in the argument area takes no register and the arguments
1288 // behind it do not shift along.
1289 assert_eq!(up, [0]);
1290 assert_eq!(text.matches("x64.arg_val_32").count(), 2, "{text}");
1291 assert!(text.contains("%2:gpr = x64.lea_64 [$rsp]"), "{text}");
1292 assert!(!text.contains("mov_rm"), "nothing is read out of the bytes: {text}");
1293 }
1294
1295 #[test]
1296 fn the_argument_behind_a_structure_that_travelled_as_bytes_is_above_all_of_them() {
1297 let byval =
1298 Param::with_abi(Type::PTR, Abi::ByVal { size: 24, align: 16, drains: Drains::Nothing });
1299 let params: Vec<Param> = (0..7).map(|_| Param::new(Type::int(64))).collect();
1300 let (_, up) = arrive_with(&[¶ms[..], &[byval], ¶ms[..1]].concat(), &SYSV);
1301
1302 // Six integers take the six registers, the seventh is at the bottom of the argument area,
1303 // and the structure is above it at the alignment its type asks for rather than at a word.
1304 // The one behind the structure is above all twenty four of its bytes, rounded up to a
1305 // whole number of words, because the area is a run of words.
1306 assert_eq!(up, [0, 16, 40]);
1307 }
1308
1309 /// A parameter narrower than a word is read at its own width rather than at a word, and one in
1310 /// the other register file is read with the other file's instruction. Both are the same list
1311 /// [`head_of`] answers from, which is what stops a function being turned away for the width of
1312 /// its seventh argument alone.
1313 #[test]
1314 fn what_a_stack_argument_is_read_with_is_its_own_width_and_its_own_file() {
1315 let f32 = Type::float(rucc_ir::Float::F32);
1316 let params = [Type::int(64), Type::int(64), Type::int(64), Type::int(64), Type::int(8)];
1317 let (text, up) = arrive(¶ms, &WIN64);
1318 assert_eq!(up, [32]);
1319 assert!(text.contains("x64.mov_rm_8 [$rsp]"), "{text}");
1320
1321 let floats = [f32; 5];
1322 let (text, up) = arrive(&floats, &WIN64);
1323 assert_eq!(up, [32]);
1324 assert!(text.contains("%4:xmm = x64.movss_rm [$rsp]"), "{text}");
1325 }
1326
1327 /// Every type a parameter can arrive in a register at is one it can be read from memory at.
1328 /// The two lists are keyed off the same two questions so that they cannot drift, and this is
1329 /// what says so: a width one covered and the other did not would be a function turned away for
1330 /// where its arguments happened to land rather than for anything about it.
1331 #[test]
1332 fn the_two_lists_of_widths_answer_for_the_same_types() {
1333 let types = [
1334 Type::int(1),
1335 Type::int(8),
1336 Type::int(16),
1337 Type::int(32),
1338 Type::int(64),
1339 Type::int(128),
1340 Type::PTR,
1341 Type::float(rucc_ir::Float::F32),
1342 Type::float(rucc_ir::Float::F64),
1343 Type::float(rucc_ir::Float::F80),
1344 Type::float(rucc_ir::Float::F128),
1345 ];
1346 for ty in types {
1347 assert_eq!(head_of(ty).is_some(), load_of(ty).is_some(), "{ty:?}");
1348 assert_eq!(head_of(ty).is_some(), store_of(ty).is_some(), "{ty:?}");
1349 assert_eq!(head_of(ty).is_some(), ret_of(ty, 0).is_some(), "{ty:?}");
1350 }
1351 }
1352
1353 /// A hundred and twenty eight bit float arrives in a vector register like the two narrower
1354 /// formats, and it takes one of them rather than two: the psABI classifies it SSE and SSEUP,
1355 /// and what that pair names is the one register both eightbytes are in.
1356 ///
1357 /// The second float here is what says so. If the quad had taken two vector registers the
1358 /// `double` after it would be in `xmm2`.
1359 #[test]
1360 fn a_quad_float_arrives_in_one_vector_register_and_not_in_two() {
1361 let quad = Type::float(rucc_ir::Float::F128);
1362 let f64 = Type::float(rucc_ir::Float::F64);
1363 assert_eq!(
1364 bind(&[quad, f64], &SYSV),
1365 "mfunc @f {\nblock0:\n %0:xmm($xmm0) = x64.arg_val_f128\n \
1366 %1:xmm($xmm1) = x64.arg_val_f64\n}\n"
1367 );
1368 }
1369
1370 /// And it is not the width that was refused before there was an instruction to move it with,
1371 /// which is the one thing about this type that used to turn a whole function away.
1372 #[test]
1373 fn a_quad_float_is_no_longer_a_width_nothing_can_carry() {
1374 assert_eq!(refuses(Type::float(rucc_ir::Float::F128), &X86_64), None);
1375 assert_eq!(refuses(Type::float(rucc_ir::Float::F80), &X86_64), Some(Missing::OnX87));
1376 assert_eq!(refuses(Type::int(128), &X86_64), Some(Missing::Width));
1377 }
1378
1379 /// A float arrives in the other file, and the two files are counted apart on SysV: the
1380 /// integer here is the first integer argument and the float is the first float one, so they
1381 /// are in `rdi` and `xmm0` rather than in the first and second of anything.
1382 #[test]
1383 fn a_float_arrives_in_a_vector_register_and_is_counted_apart_from_the_integers() {
1384 let f32 = Type::float(rucc_ir::Float::F32);
1385 let f64 = Type::float(rucc_ir::Float::F64);
1386 assert_eq!(
1387 bind(&[Type::int(32), f64, f32], &SYSV),
1388 "mfunc @f {\nblock0:\n %0:gpr($rdi) = x64.arg_val_32\n \
1389 %1:xmm($xmm0) = x64.arg_val_f64\n %2:xmm($xmm1) = x64.arg_val_f32\n}\n"
1390 );
1391 }
1392
1393 /// Windows counts the two files together, so the same three arguments land in different
1394 /// registers: the float is the second argument and takes the second vector register rather
1395 /// than the first, which is the difference that makes a mismatched call read the wrong value.
1396 #[test]
1397 fn the_other_convention_counts_the_two_files_as_one_run_of_positions() {
1398 let f64 = Type::float(rucc_ir::Float::F64);
1399 assert_eq!(
1400 bind(&[Type::int(32), f64, Type::int(64)], &WIN64),
1401 "mfunc @f {\nblock0:\n %0:gpr($rcx) = x64.arg_val_32\n \
1402 %1:xmm($xmm1) = x64.arg_val_f64\n %2:gpr($r8) = x64.arg_val_64\n}\n"
1403 );
1404 }
1405
1406 /// A `long double` is in neither file and travels in the argument area, which is what SysV's
1407 /// X87 classification comes to. So it arrives the way a structure the classification put in
1408 /// memory arrives, as the address of its bytes in a general purpose register, and it does that
1409 /// while the vector file is untouched: this one is in the argument area because of what it is
1410 /// rather than because the registers ran out.
1411 #[test]
1412 fn a_long_double_arrives_as_the_address_of_its_bytes_in_the_argument_area() {
1413 let params = [Type::int(32), Type::float(rucc_ir::Float::F80)];
1414 assert_eq!(
1415 bind(¶ms, &SYSV),
1416 "mfunc @f {\nblock0:\n %0:gpr($rdi) = x64.arg_val_32\n \
1417 %1:gpr = x64.lea_64 [$rsp]\n}\n"
1418 );
1419 }
1420
1421 /// It still cannot come back beside another value, and what it is turned away for says which
1422 /// file is in the way rather than calling eighty bits a width no register holds. A pair comes
1423 /// back in a pair of registers and there is no pair with the x87 stack in it.
1424 #[test]
1425 fn a_long_double_in_a_pair_is_reported_as_the_x87_stack_it_travels_on() {
1426 let returns = [Type::float(rucc_ir::Float::F80), Type::int(64)];
1427 assert_eq!(
1428 make(&[], &returns, false, &SYSV).2,
1429 Err(Refused { argument: None, missing: Missing::OnX87 })
1430 );
1431 }
1432
1433 /// One call to `g`, with a register for each argument arriving in the block that makes it.
1434 ///
1435 /// A variadic call here names none of its arguments, which is the shape that asks the most of a
1436 /// convention. [`made_naming`] is for the tests that care where the line between the named ones
1437 /// and the rest actually falls.
1438 fn make(
1439 args: &[Type],
1440 returns: &[Type],
1441 variadic: bool,
1442 conv: &CallRegs,
1443 ) -> (Interner, mir::Func, Result<Made, Refused>) {
1444 let named = if variadic { 0 } else { args.len() };
1445 made_naming(args, returns, named, variadic, conv)
1446 }
1447
1448 /// The same, for a callee whose signature names that many of the arguments.
1449 fn made_naming(
1450 args: &[Type],
1451 returns: &[Type],
1452 named: usize,
1453 variadic: bool,
1454 conv: &CallRegs,
1455 ) -> (Interner, mir::Func, Result<Made, Refused>) {
1456 let mut names = Interner::new();
1457 let mut out = mir::Func::new(names.intern("f"));
1458 let block = out.create_block();
1459 let passed: Vec<Passing> = args
1460 .iter()
1461 .map(|&ty| Passing {
1462 ty,
1463 reg: out.append_param(block, class_of(ty, conv)),
1464 abi: Abi::Plain,
1465 })
1466 .collect();
1467 let callee = Callee::Named(names.intern("g"));
1468 let what = Calling { callee, args: &passed, returns, variadic, named, at: Span::DUMMY };
1469 let made = call(&mut out, block, &what, conv, &X86_64, &mut names);
1470 (names, out, made)
1471 }
1472
1473 /// What the call in that function reads and writes, by register name, in the order the
1474 /// operands are in.
1475 fn operands(func: &mir::Func) -> (Vec<String>, Vec<String>) {
1476 let block = func.entry().expect("a function with a block in it");
1477 let call = func.terminator(block).expect("the call is the last thing in the block");
1478 let name = |operand: &mir::Operand| match (operand.reg.phys(), operand.constraint) {
1479 (Some(reg), _) | (None, Constraint::Fixed(reg)) => {
1480 REGS.name(operand.class, reg).expect("a register the file describes").to_string()
1481 }
1482 _ => format!("{:?}", operand.reg),
1483 };
1484 let mut written = Vec::new();
1485 let mut read = Vec::new();
1486 for operand in &func[func[call].operands] {
1487 let into = if operand.role == mir::Role::Use { &mut read } else { &mut written };
1488 into.push(name(operand));
1489 }
1490 (written, read)
1491 }
1492
1493 #[test]
1494 fn a_call_passes_its_arguments_where_the_convention_puts_them() {
1495 let i32 = Type::int(32);
1496 let (_, func, made) = make(&[i32, i32, i32], &[], false, &SYSV);
1497 assert_eq!(made.expect("three integers all fit in registers").results, []);
1498 assert_eq!(operands(&func).1, ["rdi", "rsi", "rdx"]);
1499 }
1500
1501 #[test]
1502 fn the_other_convention_passes_the_same_arguments_somewhere_else() {
1503 let i64 = Type::int(64);
1504 let (_, func, made) = make(&[i64, i64], &[], false, &WIN64);
1505 // Thirty two bytes of stack for a call that passes nothing on the stack, which is what
1506 // Windows asks a caller to leave the callee whether the callee uses it or not.
1507 assert_eq!(made.expect("two integers fit in registers").outgoing, 32);
1508 assert_eq!(operands(&func).1, ["rcx", "rdx"]);
1509 }
1510
1511 #[test]
1512 fn what_a_call_gives_back_comes_out_of_the_register_the_convention_returns_in() {
1513 let (names, func, made) = make(&[], &[Type::int(32)], false, &SYSV);
1514 let made = made.expect("an integer comes back");
1515 let [result] = made.results[..] else { panic!("one register") };
1516 // The first thing written is the result, and it is the only thing written that is a value
1517 // rather than a register the callee destroyed.
1518 assert_eq!(operands(&func).0.first().map(String::as_str), Some("rax"));
1519 assert_eq!(func.class_of(result), Some(SYSV.int_class));
1520 assert!(mir::print_func(&func, &names, ®S).contains("x64.call"));
1521 }
1522
1523 #[test]
1524 fn every_register_the_callee_may_destroy_is_written_by_the_call() {
1525 let (_, func, _) = make(&[Type::int(64)], &[Type::int(64)], false, &SYSV);
1526 let (written, read) = operands(&func);
1527 // The callee saved registers are not here, because a value in one of those survives a
1528 // call and that is the whole difference between the two halves of the convention.
1529 for saved in ["rbx", "rbp", "r12", "r13", "r14", "r15"] {
1530 assert!(!written.contains(&saved.to_string()), "{saved} survives a call");
1531 }
1532 // Every other integer register is, once. The two named ones are named by the result and
1533 // by the argument instead, and naming one twice would be blocking it twice.
1534 for destroyed in ["rcx", "rdx", "rsi", "r8", "r9", "r10", "r11"] {
1535 let count = written.iter().filter(|name| *name == destroyed).count();
1536 assert_eq!(count, 1, "{destroyed} is destroyed by a call and is written {count} times");
1537 }
1538 assert_eq!(written.iter().filter(|name| *name == "rax").count(), 1);
1539 assert_eq!(read, ["rdi"]);
1540 // The vector registers are all destroyed on SysV, and they are in the other class.
1541 assert!(written.contains(&"xmm0".to_string()));
1542 }
1543
1544 #[test]
1545 fn a_variadic_call_says_how_many_vector_registers_it_passed_arguments_in() {
1546 let (names, func, made) = make(&[Type::int(64)], &[], true, &SYSV);
1547 made.expect("an integer argument to a variadic callee");
1548 let (_, read) = operands(&func);
1549 // Zero of them here, and `al` is where a SysV callee looks for it. Leaving whatever was in
1550 // the register there would make a callee that saves its vector registers save ones it was
1551 // never given.
1552 assert_eq!(read, ["rdi", "rax"]);
1553 assert_eq!(
1554 mir::print_func(&func, &names, ®S).lines().nth(2),
1555 Some(" %1:gpr = x64.mov_ri_32 0")
1556 );
1557
1558 // Two of them here, which is the number that decides how much of the register save area a
1559 // callee like `printf` fills in. A count of zero with a float in `xmm0` would be a callee
1560 // reading its first `%f` out of a register nothing wrote.
1561 let f64 = Type::float(rucc_ir::Float::F64);
1562 let (names, func, made) = make(&[Type::int(64), f64, f64], &[], true, &SYSV);
1563 made.expect("one integer and two floats all fit in registers");
1564 assert_eq!(operands(&func).1, ["rdi", "xmm0", "xmm1", "rax"]);
1565 assert!(mir::print_func(&func, &names, ®S).contains("x64.mov_ri_32 2"));
1566 }
1567
1568 /// Windows passes a float the callee has no prototype for in both files at once, because the
1569 /// callee has no way to know which file to look in and its walk over the arguments reads the
1570 /// general purpose one. This is the whole of what `printf("%f", x)` needs from the caller.
1571 #[test]
1572 fn a_float_a_variadic_callee_has_no_prototype_for_travels_in_both_files_on_windows() {
1573 let f64 = Type::float(rucc_ir::Float::F64);
1574 let (names, func, made) = made_naming(&[Type::int(32), f64], &[], 1, true, &WIN64);
1575 made.expect("an integer and a float both fit in registers");
1576
1577 // The float is the second argument, so its position is one and both of its registers are
1578 // the second of their file. `rdx` holds the bits and nothing converts them, which is what
1579 // the `movq` is: the callee reads bits out of it and not a value of any type.
1580 assert_eq!(operands(&func).1, ["rcx", "xmm1", "rdx"]);
1581 let text = mir::print_func(&func, &names, ®S);
1582 assert!(text.contains("x64.movq_from_xmm %1"), "{text}");
1583 }
1584
1585 /// An argument the signature does name needs no second copy, since the callee's parameter says
1586 /// where it is, and neither does one on a convention that keeps the two files apart.
1587 #[test]
1588 fn an_argument_the_signature_names_travels_in_one_file() {
1589 let f64 = Type::float(rucc_ir::Float::F64);
1590 let (names, func, made) = made_naming(&[Type::int(32), f64], &[], 2, true, &WIN64);
1591 made.expect("both are named");
1592 assert_eq!(operands(&func).1, ["rcx", "xmm1"]);
1593 assert!(!mir::print_func(&func, &names, ®S).contains("movq_from_xmm"));
1594
1595 let (names, func, made) = made_naming(&[Type::int(32), f64], &[], 1, true, &SYSV);
1596 made.expect("an integer and a float");
1597 assert!(!mir::print_func(&func, &names, ®S).contains("movq_from_xmm"));
1598 }
1599
1600 /// A float past the position the registers run out at is in the argument area and nowhere else,
1601 /// which is where the second copy stops being a thing there is room for. The callee reads it
1602 /// out of memory whichever file it would have been in.
1603 #[test]
1604 fn a_float_the_registers_ran_out_before_gets_no_second_copy() {
1605 let f64 = Type::float(rucc_ir::Float::F64);
1606 let (names, func, made) = made_naming(&[f64; 6], &[], 0, true, &WIN64);
1607 made.expect("four in registers and two in memory");
1608 let text = mir::print_func(&func, &names, ®S);
1609 assert_eq!(text.matches("movq_from_xmm").count(), 4, "{text}");
1610 assert!(text.contains("x64.movsd_mr %4, [$rsp + 32]"), "{text}");
1611 }
1612
1613 #[test]
1614 fn a_call_through_an_address_reads_it_in_front_of_the_arguments() {
1615 let i32 = Type::int(32);
1616 let mut names = Interner::new();
1617 let mut out = mir::Func::new(names.intern("f"));
1618 let block = out.create_block();
1619 let address = out.append_param(block, SYSV.int_class);
1620 let reg = out.append_param(block, SYSV.int_class);
1621 let passed = vec![Passing { ty: i32, reg, abi: Abi::Plain }];
1622 let what = Calling {
1623 callee: Callee::Through(address),
1624 args: &passed,
1625 returns: &[i32],
1626 variadic: false,
1627 named: passed.len(),
1628 at: Span::DUMMY,
1629 };
1630 call(&mut out, block, &what, &SYSV, &X86_64, &mut names)
1631 .expect("one integer fits in a register");
1632
1633 // The address is the first thing read and the arguments follow it, which is the order the
1634 // assembler counts on, and it is in no particular register because every register a call
1635 // could insist on is one the call has already spoken for.
1636 let text = mir::print_func(&out, &names, ®S);
1637 assert!(text.contains("= x64.call_reg %0, %1($rdi)\n"), "{text}");
1638 assert!(!text.contains("@g"), "a call through an address names nobody: {text}");
1639 }
1640
1641 #[test]
1642 fn a_call_with_no_register_left_writes_the_argument_into_the_outgoing_area() {
1643 let i64 = Type::int(64);
1644 let (names, func, made) = make(&[i64; 7], &[], false, &SYSV);
1645 let made = made.expect("the seventh goes to memory");
1646
1647 // At the stack pointer, because the outgoing area is at the bottom of the frame, and in
1648 // front of the call rather than as an operand of it.
1649 let text = mir::print_func(&func, &names, ®S);
1650 assert!(text.contains("x64.mov_mr_64 %6, [$rsp]\n"), "{text}");
1651 let store = text.find("x64.mov_mr_64").expect("the store");
1652 assert!(store < text.find("x64.call").expect("the call"), "{text}");
1653 // One word of it, which is what the frame has to reserve for this call.
1654 assert_eq!(made.outgoing, 8);
1655 }
1656
1657 /// One call to `g`, passing that many words, then an object of that size and alignment by
1658 /// value, then one more integer, which is `int g(long.., struct Big, int)` after the
1659 /// classification.
1660 fn pass_bytes(
1661 before: usize,
1662 size: u64,
1663 align: u32,
1664 conv: &CallRegs,
1665 ) -> (Interner, mir::Func, Result<Made, Refused>) {
1666 let mut names = Interner::new();
1667 let mut out = mir::Func::new(names.intern("f"));
1668 let block = out.create_block();
1669 let mut args: Vec<Passing> = (0..before)
1670 .map(|_| Passing {
1671 ty: Type::int(64),
1672 reg: out.append_param(block, conv.int_class),
1673 abi: Abi::Plain,
1674 })
1675 .collect();
1676 args.push(Passing {
1677 ty: Type::PTR,
1678 reg: out.append_param(block, conv.int_class),
1679 abi: Abi::ByVal { size, align, drains: Drains::Nothing },
1680 });
1681 args.push(Passing {
1682 ty: Type::int(32),
1683 reg: out.append_param(block, conv.int_class),
1684 abi: Abi::Plain,
1685 });
1686 let callee = Callee::Named(names.intern("g"));
1687 let what = Calling {
1688 callee,
1689 args: &args,
1690 returns: &[],
1691 variadic: false,
1692 named: args.len(),
1693 at: Span::DUMMY,
1694 };
1695 let made = call(&mut out, block, &what, conv, &X86_64, &mut names);
1696 (names, out, made)
1697 }
1698
1699 #[test]
1700 fn a_structure_passed_by_value_in_memory_is_copied_into_the_outgoing_area() {
1701 let (names, func, made) = pass_bytes(1, 24, 8, &SYSV);
1702 let made = made.expect("an object of three words is copied a word at a time");
1703
1704 // The bytes travel and the address does not, so the copy is a load and a store for each
1705 // word of it, in front of the call, and the callee's copy is at the bottom of the outgoing
1706 // area. The caller owes it this copy: the callee is free to write to what it was handed,
1707 // so what it was handed cannot be the object itself.
1708 let text = mir::print_func(&func, &names, ®S);
1709 assert!(text.contains("x64.mov_mr_64 %3, [$rsp]\n"), "{text}");
1710 assert!(text.contains("x64.mov_mr_64 %4, [$rsp + 8]\n"), "{text}");
1711 assert!(text.contains("x64.mov_mr_64 %5, [$rsp + 16]\n"), "{text}");
1712 assert_eq!(text.matches("x64.mov_rm_64").count(), 3, "{text}");
1713 assert!(text.find("x64.mov_mr_64") < text.find("x64.call"), "{text}");
1714 assert_eq!(made.outgoing, 24);
1715 }
1716
1717 #[test]
1718 fn the_integers_beside_it_still_travel_in_registers() {
1719 let (_, func, _) = pass_bytes(1, 24, 8, &SYSV);
1720
1721 // An object in the argument area takes no argument register, so the integer behind it is
1722 // in the second one and not the third. Counting it as a register is the mistake that would
1723 // shift every argument after it along by one.
1724 let (clobbered, read) = operands(&func);
1725 assert_eq!(read, ["rdi", "rsi"]);
1726 assert!(clobbered.contains(&"rdx".to_owned()), "the third is free: {clobbered:?}");
1727 }
1728
1729 #[test]
1730 fn an_object_wanting_more_alignment_than_a_word_gets_it() {
1731 let (names, func, made) = pass_bytes(7, 24, 16, &SYSV);
1732 let made = made.expect("an object of three words");
1733
1734 // Six of the integers took the registers and the seventh is at the bottom of the area, so
1735 // the object cannot start where it left off: sixteen byte alignment moves it up to the
1736 // next multiple of sixteen and leaves a word of nothing behind it. The integer after it is
1737 // above all three of its words.
1738 let text = mir::print_func(&func, &names, ®S);
1739 assert!(text.contains("x64.mov_mr_64 %9, [$rsp + 16]\n"), "{text}");
1740 assert!(text.contains("x64.mov_mr_32 %8, [$rsp + 40]\n"), "{text}");
1741 assert_eq!(made.outgoing, 48);
1742 }
1743
1744 #[test]
1745 fn an_object_too_large_to_copy_a_word_at_a_time_is_copied_by_the_runtime() {
1746 let (names, func, made) = pass_bytes(1, 4096, 8, &SYSV);
1747 let made = made.expect("an object the runtime copies");
1748
1749 // Five hundred and twelve words is past what unrolling is worth, so the copy is the call
1750 // the same size of `memcpy` in the IR becomes: the address in the outgoing area, the
1751 // address of the object, and the count, and then the call the object was an argument of.
1752 let text = mir::print_func(&func, &names, ®S);
1753 assert!(text.contains("x64.lea_64 [$rsp]"), "{text}");
1754 assert!(text.contains("x64.mov_ri_32 4096"), "{text}");
1755 assert_eq!(text.matches("x64.call").count(), 2, "the copy and the call: {text}");
1756 assert!(text.find("@memcpy") < text.find("@g"), "the copy comes first: {text}");
1757
1758 // And the argument area is the object, because a call passing three words in registers
1759 // needs none of its own.
1760 assert_eq!(made.outgoing, 4096);
1761 }
1762
1763 #[test]
1764 fn an_object_too_large_to_count_the_bytes_of_is_reported_rather_than_passed() {
1765 let (_, _, made) = pass_bytes(1, 1 << 31, 8, &SYSV);
1766
1767 // Two gigabytes is more than the immediate the byte count travels in holds, and a count
1768 // that does not fit is the whole of what is left to refuse. No program passes a structure
1769 // that size by value, and one that tried would rather hear about it than be handed a copy
1770 // of the low part of it.
1771 assert_eq!(made, Err(Refused { argument: Some(1), missing: Missing::TooBig }));
1772 assert_eq!(
1773 Missing::TooBig.why(),
1774 "is more bytes than a count of them can be written down as"
1775 );
1776 }
1777
1778 /// An `ms_abi` function on Linux takes its arguments the Windows way: four in registers from
1779 /// `rcx`, and the fifth above the 32 bytes of shadow space.
1780 #[test]
1781 fn an_ms_abi_function_on_linux_takes_its_arguments_the_windows_way() {
1782 let ms = SYSV.under(Convention::Ms).expect("Linux has the Windows convention too");
1783 let (text, up) = arrive(&[Type::int(64); 7], ms);
1784 assert_eq!(up, [32, 40, 48]);
1785 assert!(text.contains("%0:gpr($rcx) = x64.arg_val_64"), "{text}");
1786 assert_eq!(text.matches("x64.arg_val_64").count(), 4, "{text}");
1787 }
1788
1789 /// A `sysv_abi` function on Windows takes six in registers from `rdi`, and the seventh at the
1790 /// bottom of the argument area with no shadow space under it.
1791 #[test]
1792 fn a_sysv_abi_function_on_windows_takes_its_arguments_the_system_v_way() {
1793 let sysv = WIN64.under(Convention::Sysv).expect("Windows has the System V convention too");
1794 let (text, up) = arrive(&[Type::int(64); 7], sysv);
1795 assert_eq!(up, [0]);
1796 assert!(text.contains("%0:gpr($rdi) = x64.arg_val_64"), "{text}");
1797 assert_eq!(text.matches("x64.arg_val_64").count(), 6, "{text}");
1798 }
1799
1800 /// A call to an `ms_abi` function from Linux leaves the callee its shadow space, and does not
1801 /// count `rsi`, `rdi` or the upper vector registers as destroyed, since the callee keeps them.
1802 #[test]
1803 fn a_call_to_an_ms_abi_function_keeps_what_the_callee_keeps() {
1804 let ms = SYSV.under(Convention::Ms).expect("Linux has the Windows convention too");
1805 let i64 = Type::int(64);
1806 let (_, func, made) = make(&[i64, i64], &[i64], false, ms);
1807 assert_eq!(made.expect("two integers fit in registers").outgoing, 32);
1808 let (written, read) = operands(&func);
1809 assert_eq!(read, ["rcx", "rdx"]);
1810 for kept in ["rsi", "rdi", "xmm6", "xmm15"] {
1811 assert!(!written.contains(&kept.to_string()), "{kept} survives the call");
1812 }
1813 assert!(written.contains(&"xmm5".to_string()), "{written:?}");
1814 }
1815
1816 /// A call to a `sysv_abi` function from Windows is the other way round: no shadow space, the
1817 /// arguments from `rdi`, and `rsi`, `rdi` and every vector register destroyed, which is what
1818 /// makes the calling Windows function save the ones it owes its own caller.
1819 #[test]
1820 fn a_call_to_a_sysv_abi_function_destroys_what_windows_would_keep() {
1821 let sysv = WIN64.under(Convention::Sysv).expect("Windows has the System V convention too");
1822 let i64 = Type::int(64);
1823 let (_, func, made) = make(&[i64, i64], &[i64], false, sysv);
1824 assert_eq!(made.expect("two integers fit in registers").outgoing, 0);
1825 let (written, read) = operands(&func);
1826 assert_eq!(read, ["rdi", "rsi"]);
1827 for destroyed in ["xmm6", "xmm15"] {
1828 assert!(written.contains(&destroyed.to_string()), "{destroyed}: {written:?}");
1829 }
1830 // `rsi` is named by an argument and so is not repeated as a clobber, and `rdi` likewise.
1831 assert!(!written.contains(&"rbx".to_string()), "{written:?}");
1832 }
1833
1834 /// The other convention runs out three arguments earlier and starts its argument area above the
1835 /// shadow space it also has to reserve, and both of those are what `Places` already said.
1836 #[test]
1837 fn where_the_outgoing_area_starts_is_the_convention_s_answer() {
1838 let i64 = Type::int(64);
1839 let (names, func, made) = make(&[i64; 7], &[], false, &WIN64);
1840 assert_eq!(made.expect("the last three go to memory").outgoing, 56);
1841
1842 // Thirty two bytes of shadow space first, which the caller writes nothing into and the
1843 // callee owns, and the fifth argument above it.
1844 let text = mir::print_func(&func, &names, ®S);
1845 assert!(text.contains("x64.mov_mr_64 %4, [$rsp + 32]\n"), "{text}");
1846 assert!(text.contains("x64.mov_mr_64 %5, [$rsp + 40]\n"), "{text}");
1847 assert!(text.contains("x64.mov_mr_64 %6, [$rsp + 48]\n"), "{text}");
1848 }
1849
1850 /// What a stack argument is written with is its own width and its own register file, matching
1851 /// what the callee reads it back with.
1852 #[test]
1853 fn a_narrow_or_floating_argument_keeps_its_own_store() {
1854 let i64 = Type::int(64);
1855 let narrow = [i64, i64, i64, i64, i64, i64, Type::int(8)];
1856 let (names, func, made) = make(&narrow, &[], false, &SYSV);
1857 made.expect("the seventh goes to memory");
1858 let text = mir::print_func(&func, &names, ®S);
1859 assert!(text.contains("x64.mov_mr_8 %6, [$rsp]\n"), "{text}");
1860
1861 let f32 = Type::float(rucc_ir::Float::F32);
1862 let (names, func, made) = make(&[f32; 9], &[], false, &SYSV);
1863 made.expect("the ninth goes to memory");
1864 let text = mir::print_func(&func, &names, ®S);
1865 assert!(text.contains("x64.movss_mr %8, [$rsp]\n"), "{text}");
1866 }
1867
1868 /// The count a SysV variadic callee reads is a count of registers, so an argument that went to
1869 /// memory instead is not in it.
1870 #[test]
1871 fn an_argument_in_memory_is_not_counted_as_a_vector_register() {
1872 let f64 = Type::float(rucc_ir::Float::F64);
1873 let (names, func, made) = make(&[f64; 9], &[], true, &SYSV);
1874 made.expect("the ninth goes to memory");
1875 let text = mir::print_func(&func, &names, ®S);
1876 assert!(text.contains("x64.mov_ri_32 8\n"), "eight registers, not nine: {text}");
1877 }
1878
1879 /// The two lists of widths answer for the same set of types, so that a value the callee can
1880 /// read out of the argument area is one the caller can write into it.
1881 #[test]
1882 fn what_can_be_read_can_be_written() {
1883 let types = [
1884 Type::int(1),
1885 Type::int(8),
1886 Type::int(16),
1887 Type::int(32),
1888 Type::int(64),
1889 Type::int(128),
1890 Type::PTR,
1891 Type::float(rucc_ir::Float::F32),
1892 Type::float(rucc_ir::Float::F64),
1893 Type::float(rucc_ir::Float::F80),
1894 ];
1895 for ty in types {
1896 assert_eq!(load_of(ty).is_some(), store_of(ty).is_some(), "{ty:?}");
1897 }
1898 }
1899
1900 /// A float travels in the other file at both ends of a call, and the register it comes back in
1901 /// is the first of that file rather than the first of the other one.
1902 #[test]
1903 fn a_call_passes_and_returns_a_float_in_a_vector_register() {
1904 let f64 = Type::float(rucc_ir::Float::F64);
1905 let (_, func, made) = make(&[Type::int(32), f64], &[f64], false, &SYSV);
1906 let result = made.expect("an integer and a float both fit in registers");
1907 let (written, read) = operands(&func);
1908 assert_eq!(read, ["rdi", "xmm0"]);
1909 assert_eq!(written.first().map(String::as_str), Some("xmm0"));
1910 assert_eq!(func.class_of(result.results[0]), Some(SYSV.sse_class));
1911 // Written once, because the register the result comes back in is already blocked by being
1912 // named and a clobber that repeated it would be blocking it twice. `rax` is a clobber here
1913 // rather than the result, which is the same register number in the other file and is the
1914 // whole reason the two lists are counted apart.
1915 assert_eq!(written.iter().filter(|name| *name == "xmm0").count(), 1);
1916 assert!(written.contains(&"rax".to_string()));
1917 }
1918
1919 #[test]
1920 fn a_call_at_a_width_no_register_holds_is_reported_on_either_side() {
1921 let i128 = Type::int(128);
1922 assert_eq!(
1923 make(&[i128], &[], false, &SYSV).2,
1924 Err(Refused { argument: Some(0), missing: Missing::Width })
1925 );
1926 assert_eq!(
1927 make(&[], &[i128], false, &SYSV).2,
1928 Err(Refused { argument: None, missing: Missing::Width })
1929 );
1930 }
1931
1932 #[test]
1933 fn an_argument_wider_than_a_register_has_no_name() {
1934 assert_eq!(head_of(Type::int(128)), None);
1935 assert_eq!(head_of(Type::int(8)), Some("x64.arg_val_8"));
1936 assert_eq!(head_of(Type::int(64)), Some("x64.arg_val_64"));
1937 }
1938
1939 /// An address arrives in a general purpose register like any other integer of its width, and
1940 /// used to be turned away here as a width no register holds, which is what issue 274 is.
1941 /// `int g(char *s)` is the smallest program that was.
1942 #[test]
1943 fn an_address_arrives_in_a_register_like_the_integer_it_is() {
1944 assert_eq!(head_of(Type::PTR), Some("x64.arg_val_64"));
1945 assert_eq!(
1946 bind(&[Type::PTR], &SYSV),
1947 "mfunc @f {\nblock0:\n %0:gpr($rdi) = x64.arg_val_64\n}\n"
1948 );
1949 // And it travels the same way at a call, on both sides of one.
1950 assert!(make(&[Type::PTR], &[Type::PTR], false, &SYSV).2.is_ok());
1951 }
1952}