rucc_codegen/finish.rs
1//! The prologue, the epilogue, and the moves the allocator asked for.
2//!
3//! Design: `spec/10-backend.md` sections 10.4 and 10.7.
4//!
5//! [`crate::frame`] works out what a function's stack looks like and writes nothing. This is what
6//! writes it. Three things are still missing from a function the allocator has finished with, and
7//! all three of them are instructions no lowering rule chose:
8//!
9//! ```text
10//! the prologue takes the frame the layout worked out, and puts away the registers a call
11//! leaves alone that this function writes anyway
12//! the moves every spill, every reload and every copy the allocator handed back as an
13//! edit, in the place it said and in the order it said
14//! the epilogue gives the frame back and puts the registers back, at the end of every block
15//! the function returns from
16//! ```
17//!
18//! There is a fourth thing and it is not an instruction but a number. The lowering wrote an
19//! instruction for every `alloca` that computes the address of the memory it asked for, and could
20//! not write how far into the frame that memory is, because when it ran there was no frame. So
21//! the displacement of each of those is filled in here, out of the same [`Frame`] everything else
22//! here reads, and off the same stack pointer every other offset in it is from.
23//!
24//! The loads that read the arguments the caller passed on the stack are waiting on the same number
25//! and on one more. Those bytes are the caller's rather than this function's, and a frame that had
26//! to force its own alignment cannot say how far away the caller's stack pointer was, so it reaches
27//! back through the frame pointer instead. Which register a load reads through is therefore settled
28//! here too, and it is the only base register in a finished function that was not settled by
29//! whoever wrote the instruction.
30//!
31//! After this the function is one an encoder can read: every register is physical, every offset
32//! into the frame is a constant, and the stack pointer is where the convention says it should be
33//! at every instruction that could look.
34//!
35//! # Why the moves go in first
36//!
37//! Every offset the frame reports is from the stack pointer as it stands in the body of the
38//! function. A spill written before the prologue exists would be written in front of the
39//! instruction it belongs to and behind nothing, which is where the prologue then goes, so the
40//! prologue ends up in front of it and the offsets stay true. Writing them the other way round
41//! would put the first reload above the instruction that takes the frame, and it would read from
42//! an address that is one frame out.
43//!
44//! # Where a return is
45//!
46//! A block that goes nowhere is a block the function leaves from. Mostly that is a return, and
47//! the other kind is a block ending in `unreachable`, which is a point the front end says control
48//! does not arrive at and which the lowering writes no instruction for. Both want the same thing
49//! here. A return wants the epilogue because that is what a return is once the frame is known,
50//! and an unreachable block wants it because the alternative is a function whose last instruction
51//! falls into whatever the assembler put after it, which is worse than an epilogue nothing runs.
52//! So the epilogue goes at the end of every block with an empty successor list, and there may be
53//! several, because nothing here insists a function has one exit.
54//!
55//! # What is target-specific here
56//!
57//! The names, and only the names. Which instruction pushes a register and which one moves the
58//! stack pointer is [`rucc_target::FrameInsts`], which the target says and this reads, so what
59//! is written below is the shape of a prologue rather than any particular machine's. That is
60//! `spec/10-backend.md` section 10.8 as it applies to the one pass that would otherwise be full
61//! of `x64.` by hand.
62
63use rucc_base::Interner;
64use rucc_mir::{Block, BlockCall, CfiOp, Func, Inst, Mem, Opcode, Operand, Reg};
65use rucc_regalloc::Allocation;
66use rucc_regalloc::assign::Place;
67use rucc_regalloc::rewrite::{At, Edit};
68use rucc_target::{BranchInsts, CallRegs, FrameInsts, Guard, PhysReg, Probe, RegClass};
69
70use crate::frame::Frame;
71use crate::lower::Stack;
72
73/// What the stack protector's check needs beyond the frame, in a function that has one.
74///
75/// Three things that come from three places, which is why they arrive together rather than being
76/// looked up here. Where the word the canary is copied from lives is a fact about the runtime the
77/// code is linked against. What a branch on a register is is a fact about the machine. And the two
78/// registers are neither: they are the ones the allocator was told to hold back, which is a
79/// decision about the allocator, and they are free at a return for exactly that reason.
80#[derive(Debug, Clone, Copy)]
81pub struct Protect<'a> {
82 /// Where the word the canary is a copy of lives, and what to call when the copy has changed.
83 pub guard: &'a Guard,
84 /// What a branch on a register is, which is what the check ends its block with.
85 pub branch: &'a BranchInsts,
86 /// The two registers the check may use, which are two the allocator never handed out.
87 pub scratch: [PhysReg; 2],
88}
89
90/// What a prologue that takes its frame a page at a time needs beyond the frame.
91///
92/// What `-fstack-clash-protection` asks for, and the same three kinds of thing [`Protect`] is:
93/// one fact about the platform, one about the machine, and two registers that are neither. See
94/// [`rucc_target::Probe`] for what the sequence is defending against.
95#[derive(Debug, Clone, Copy)]
96pub struct Probing<'a> {
97 /// What touches a page and how far apart the pages are.
98 pub probe: &'a Probe,
99 /// What a branch on a register is, which is what the loop under a large frame ends with.
100 pub branch: &'a BranchInsts,
101 /// The two registers the sequence may use, which are two the allocator never handed out.
102 pub scratch: [PhysReg; 2],
103}
104
105/// What the convention this function is compiled for says a frame is.
106///
107/// Five answers to the one question, which is why they travel together: where it puts things,
108/// which instructions build one, whether this function's carries a protector, whether it is taken
109/// a page at a time, and whether the function opens with a landing pad. The last three are the
110/// only ones about this function rather than about every function on the target, and they are here
111/// because what they need is the other two and nothing else.
112#[derive(Debug, Clone, Copy)]
113pub struct Convention<'a> {
114 /// Where the convention puts things.
115 pub regs: &'a CallRegs,
116 /// The instructions a prologue, an epilogue, a spill and a reload are made of on it.
117 pub insts: &'a FrameInsts,
118 /// What this function's stack protector needs, or `None` in a function with none.
119 pub protect: Option<Protect<'a>>,
120 /// What this function's probing prologue needs, or `None` when the frame is taken in one
121 /// subtraction, which is what a command line that did not ask asks for.
122 pub probe: Option<Probing<'a>>,
123 /// What says an indirect branch may arrive at the top of this function, or `None` when the
124 /// command line did not ask for one and on a target that has no such instruction.
125 ///
126 /// See [`rucc_target::FrameInsts::landing`]. A name rather than a flag because the flag has
127 /// already been read against the target by the time this is built, and because a prologue that
128 /// has the name has everything it needs.
129 pub landing: Option<&'static str>,
130}
131
132impl<'a> Convention<'a> {
133 /// That convention, for a function with no stack protector, no probing and no landing pad,
134 /// which is most of them.
135 #[must_use]
136 pub fn new(regs: &'a CallRegs, insts: &'a FrameInsts) -> Self {
137 Self { regs, insts, protect: None, probe: None, landing: None }
138 }
139}
140
141/// Writes the moves, the prologue and the epilogue into a function the allocator has finished
142/// with.
143///
144/// # Panics
145///
146/// Panics on a function with no blocks in it, on a frame whose slots or locals the allocation and
147/// the lowering do not match, and on a move of a class the target did not say how to move. All of
148/// them are the caller handing it a frame and a function that were not worked out from each other.
149pub fn finish(
150 func: &mut Func,
151 allocation: &Allocation,
152 frame: &Frame,
153 stack: &Stack,
154 convention: Convention<'_>,
155 names: &mut Interner,
156) {
157 let Convention { regs: conv, insts, protect, probe, landing } = convention;
158 let entry = func.entry().expect("a function with a block in it");
159 let returns: Vec<Block> = func.blocks().filter(|&block| func[block].succs.is_empty()).collect();
160
161 // Before anything is written, because these are instructions the lowering already put in the
162 // function and every one of them is somewhere the prologue is about to go in front of, which
163 // is what makes an offset from the stack pointer the right thing to write into them.
164 for &(inst, local) in &stack.addresses {
165 let at = frame.local(local).expect("a local the frame was worked out from");
166 let mem = func[inst].mem.expect("the address of a local is an address");
167 func[mem].disp = at;
168 }
169
170 // The same, one area further up, and through the frame pointer when that is what reaches it.
171 // These are in the entry block ahead of everything, so the prologue still goes in front of
172 // them, which is what makes both registers hold what these offsets are counted from.
173 let incoming = frame.incoming();
174 for &(inst, up) in &stack.arguments {
175 let mem = func[inst].mem.expect("an argument read out of memory is read from an address");
176 func[mem].disp = incoming.at + offset(up);
177 if incoming.through_frame_pointer {
178 // The base register is an operand of the instruction and the addressing mode holds
179 // where in the operand vector it is, so the register is changed there and not here.
180 let at = func[mem].base.expect("an address the lowering wrote a base register into");
181 let operands = func[inst].operands;
182 func[operands][usize::from(at)].reg = Reg::physical(conv.frame_pointer);
183 }
184 }
185
186 let mut writer = Writer { func, conv, insts, names, ahead: None };
187
188 let mut cursors: Vec<(At, Inst)> = Vec::new();
189 for edit in &allocation.edits {
190 let inst = writer.mov(edit, frame);
191 writer.put(&mut cursors, edit.at, inst);
192 }
193
194 let prologue = writer.prologue(frame, protect, probe, landing);
195 for &inst in prologue.iter().rev() {
196 writer.func.prepend_inst(entry, inst);
197 }
198 for block in returns {
199 // The check goes in front of the epilogue and takes the return with it. What is left in
200 // the block the function used to return from is the check, and the block the epilogue then
201 // goes in is the arm the canary was unchanged on.
202 let block = match protect {
203 Some(protect) => writer.check(block, frame, protect),
204 None => block,
205 };
206 let epilogue = writer.epilogue(frame);
207 for inst in epilogue {
208 writer.func.append_inst(block, inst);
209 }
210 }
211
212 // Last of everything, because the blocks a probing prologue made have to come in front of the
213 // block the function used to begin with and the ones the protector's check makes are made
214 // after that. Nothing has been laid out yet: `crate::layout` runs after this and puts every
215 // block in its own order, and all this decides is which block the function is entered at.
216 if let Some(ahead) = writer.ahead {
217 let rest: Vec<Block> =
218 writer.func.blocks().filter(|block| !ahead.contains(block)).collect();
219 let order: Vec<Block> = ahead.into_iter().chain(rest).collect();
220 writer.func.set_block_order(&order);
221 }
222}
223
224/// How many pages a probing prologue touches one after another before it writes a loop instead.
225///
226/// Three, which is what gcc unrolls to. The loop is four instructions however many pages it walks
227/// and a page written out is two, so three is the last size at which the straight line is no
228/// longer than the loop, and the straight line has no branch in it and needs no register.
229const UNROLLED: u32 = 3;
230
231/// One function having its frame written into it.
232struct Writer<'a> {
233 func: &'a mut Func,
234 conv: &'a CallRegs,
235 insts: &'a FrameInsts,
236 names: &'a mut Interner,
237 /// The blocks a probing prologue made, which go in front of the one the function began with.
238 ///
239 /// Empty in every function whose frame is taken in one subtraction, which is every function
240 /// on a command line that did not ask for the stack to be touched a page at a time and most
241 /// of them on one that did. See [`Writer::pages`].
242 ahead: Option<[Block; 2]>,
243}
244
245impl Writer<'_> {
246 /// The instructions the prologue is, in the order they run.
247 ///
248 /// The order is the one the epilogue undoes and it is not free. The frame pointer is saved
249 /// before anything else, so that it points at a fixed place whatever else happens. The
250 /// registers are pushed before the alignment is forced, so that the epilogue can find them
251 /// again from the frame pointer, since after the alignment is forced nothing else can. And the
252 /// vector registers are stored last, because until the frame has been taken there is nowhere
253 /// to store them.
254 ///
255 /// The landing pad is in front of all of it, because the address it makes reachable is the
256 /// address of the function and the address of the function is where the first instruction is.
257 /// It has to be written here rather than after the fact, since a probing prologue moves the
258 /// instructions written so far into a block of its own and the pad has to move with them.
259 fn prologue(
260 &mut self,
261 frame: &Frame,
262 protect: Option<Protect<'_>>,
263 probe: Option<Probing<'_>>,
264 landing: Option<&'static str>,
265 ) -> Vec<Inst> {
266 let sp = self.conv.stack_pointer;
267 let fp = self.conv.frame_pointer;
268 let int = self.conv.int_class;
269 let sse = self.conv.sse_class;
270 let word = offset(self.conv.word);
271 let mut out = Vec::new();
272 let landing = landing.map(|name| {
273 let opcode = self.opcode(name);
274 let inst = self.func.build_loose(opcode).finish();
275 out.push(inst);
276 inst
277 });
278 // How far the stack pointer is below the canonical frame address, and whether the address
279 // is still counted from the stack pointer at all. It starts at the return address the
280 // call itself pushed, which is the rule the CIE already states, so the first row here is
281 // the first thing this function does on top of that.
282 let mut below = offset(self.conv.return_address);
283 let mut from_sp = true;
284 if frame.frame_pointer() {
285 let inst = self.push(fp);
286 out.push(inst);
287 below += word;
288 self.row(inst, CfiOp::DefCfaOffset(below));
289 self.saved(inst, int, fp, -below);
290 let mov = self.opcode(self.insts.moves(int).expect("a move").mov);
291 let inst = self.two(mov, fp, sp);
292 out.push(inst);
293 let number = self.dwarf(int, fp);
294 self.row(inst, CfiOp::DefCfaRegister(number));
295 from_sp = false;
296 }
297 for ® in frame.saved_int() {
298 let inst = self.push(reg);
299 out.push(inst);
300 below += word;
301 if from_sp {
302 self.row(inst, CfiOp::DefCfaOffset(below));
303 }
304 self.saved(inst, int, reg, -below);
305 }
306 if let Some(to) = frame.realign() {
307 // Nothing is written for this and nothing can be. After it the stack pointer is a
308 // rounded-down version of where it was rather than a fixed distance from it, which is
309 // exactly what a rule cannot say. It is also why a frame that realigns is a frame
310 // with a frame pointer: by here the address is already counted from that instead.
311 assert!(!from_sp, "a frame that forces its own alignment has a frame pointer");
312 let and = self.opcode(self.insts.align);
313 out.push(self.arith(and, -i64::from(to)));
314 }
315 if frame.size() > 0 {
316 self.take(&mut out, frame.size(), &mut below, from_sp, probe);
317 }
318 for save in frame.saved_sse() {
319 let inst = self.store(sse, save.reg, save.at);
320 out.push(inst);
321 // Where it went is an offset from the stack pointer in the body, and the address is
322 // `below` above that, so the two make one constant. Unless the frame realigned, in
323 // which case there is no such constant and the rule is left out rather than guessed;
324 // the one convention that realigns and the one that preserves a vector register are
325 // not the same convention, so nothing reaches this today.
326 if frame.realign().is_none() {
327 self.saved(inst, sse, save.reg, save.at - below);
328 }
329 }
330 // Last of everything, because it writes into the frame and there is no frame to write into
331 // until the stack pointer has moved. Nothing is described for either instruction: they
332 // write a slot rather than save a register, and no unwinder wants to put a canary back.
333 if let Some(protect) = protect {
334 let at = frame.canary().expect("a protected function has a slot for its canary");
335 let [into, _] = protect.scratch;
336 out.push(self.read_guard(into, protect.guard));
337 out.push(self.store(self.conv.int_class, into, at));
338 }
339 // The rules the body runs under, kept so that each epilogue can put them back rather than
340 // leaving the next block reading whatever the last one ended on. See `epilogue`.
341 //
342 // Nothing is kept in a function whose whole prologue is the landing pad. The pad moves no
343 // register and takes no frame, so there is no rule to put back, and remembering anyway
344 // would give a function that needs no unwind rows a pair of them that cancel out.
345 if let Some(&last) = out.last() {
346 if Some(last) != landing {
347 self.row(last, CfiOp::RememberState);
348 }
349 }
350 out
351 }
352
353 /// Takes the frame, which is one subtraction unless the command line asked for the stack to be
354 /// touched a page at a time.
355 ///
356 /// `below` is how far the canonical frame address is above the stack pointer, and it comes
357 /// back as what it is once the frame has been taken.
358 fn take(
359 &mut self,
360 out: &mut Vec<Inst>,
361 size: u32,
362 below: &mut i32,
363 from_sp: bool,
364 probe: Option<Probing<'_>>,
365 ) {
366 let Some(probing) = probe.filter(|probing| size > probing.probe.interval) else {
367 let inst = self.sub(size);
368 out.push(inst);
369 *below += offset(size);
370 if from_sp {
371 self.row(inst, CfiOp::DefCfaOffset(*below));
372 }
373 return;
374 };
375 // Every step but the last is a whole page and is followed by a touch, and the last is
376 // whatever is left over, which is between one byte and one whole page. So the stack
377 // pointer never moves further than a page without something being written where it landed,
378 // and the unmapped page an operating system leaves below a stack cannot be stepped over.
379 //
380 // That is why the count is worked out from one less than the size. A frame that is an
381 // exact number of pages gets one fewer touch than it has pages, and the step left over is
382 // a whole page, which is a step that lands on the next page boundary rather than past it.
383 // gcc touches that last page as well, so this is one instruction shorter on a frame whose
384 // size is a multiple of the page and the same everywhere else.
385 let interval = probing.probe.interval;
386 let pages = (size - 1) / interval;
387 let rest = size - pages * interval;
388 let mut walked = false;
389 if pages <= UNROLLED {
390 for _ in 0..pages {
391 let inst = self.sub(interval);
392 out.push(inst);
393 *below += offset(interval);
394 if from_sp {
395 self.row(inst, CfiOp::DefCfaOffset(*below));
396 }
397 let touch = self.touch(probing.probe);
398 out.push(touch);
399 }
400 } else {
401 self.pages(out, pages, below, from_sp, probing);
402 walked = from_sp;
403 }
404 let inst = self.sub(rest);
405 out.push(inst);
406 *below += offset(rest);
407 if from_sp {
408 // A loop leaves the address counted from the register the stack pointer was compared
409 // against, since that is the one thing in it that holds still. This is where it goes
410 // back to being counted from the stack pointer, and it is written behind this
411 // instruction rather than behind the branch because a row is written behind an
412 // instruction and the branch is not one that survives [`crate::layout`].
413 let op = if walked {
414 let number = self.dwarf(self.conv.int_class, self.conv.stack_pointer);
415 CfiOp::DefCfa { reg: number, offset: *below }
416 } else {
417 CfiOp::DefCfaOffset(*below)
418 };
419 self.row(inst, op);
420 }
421 }
422
423 /// The loop that takes a frame too large for the touches to be written one after another.
424 ///
425 /// Three blocks, and the first two are new and go in front of the one the function began with:
426 ///
427 /// ```text
428 /// what the function is entered at everything the prologue did before this, and then the
429 /// address the stack pointer is walking down to
430 /// the loop one page, the touch, and the question of whether the
431 /// stack pointer has got there yet
432 /// what the function began with the rest of the prologue, and then the body
433 /// ```
434 ///
435 /// The instructions the prologue has written so far move into the first of them, because a
436 /// block is entered at the top and they have to run before the loop does. Nothing is laid out
437 /// here: which block comes first in memory is [`crate::layout`]'s answer, and all this decides
438 /// is which one the function is entered at.
439 fn pages(
440 &mut self,
441 out: &mut Vec<Inst>,
442 pages: u32,
443 below: &mut i32,
444 from_sp: bool,
445 probing: Probing<'_>,
446 ) {
447 let class = self.conv.int_class;
448 let sp = self.conv.stack_pointer;
449 let all = offset(pages * probing.probe.interval);
450 let [limit, byte] = probing.scratch;
451
452 let head = self.func.create_block();
453 for &inst in out.iter() {
454 self.func.append_inst(head, inst);
455 }
456 out.clear();
457 // Where the stack pointer is walking down to, worked out before it starts moving. A loop
458 // that counted down instead would need somewhere to keep the count, and this is somewhere
459 // to keep it that the comparison can read without arithmetic.
460 let lea = self.opcode(self.insts.lea);
461 let inst = self.address(lea, limit, sp, -all);
462 self.func.append_inst(head, inst);
463 if from_sp {
464 // The address is counted from that register for as long as the loop runs, and it has
465 // to be: the stack pointer moves once an iteration, so no fixed distance from it is
466 // true twice, and this register was written so that one distance is.
467 let number = self.dwarf(class, limit);
468 self.row(inst, CfiOp::DefCfa { reg: number, offset: *below + all });
469 }
470
471 let body = self.func.create_block();
472 *self.func.succs_mut(head) = vec![BlockCall::to(body)];
473 let inst = self.sub(probing.probe.interval);
474 self.func.append_inst(body, inst);
475 let touch = self.touch(probing.probe);
476 self.func.append_inst(body, touch);
477 let differ = self.opcode(self.insts.differ);
478 let inst = self
479 .func
480 .build_loose(differ)
481 .def(Reg::physical(byte), class)
482 .uses(Reg::physical(sp), class)
483 .uses(Reg::physical(limit), class)
484 .finish();
485 self.func.append_inst(body, inst);
486 let cond = Opcode::new(
487 self.names.intern(&format!("{}{}", probing.branch.prefix, probing.branch.cond)),
488 );
489 let inst = self.func.build_loose(cond).uses(Reg::physical(byte), class).finish();
490 self.func.append_inst(body, inst);
491 // The first arm is the one taken when the condition held, and the condition is that the
492 // stack pointer and the address it is walking down to still differ, so the first arm is
493 // another page.
494 let began = self.func.entry().expect("a function with a block in it");
495 *self.func.succs_mut(body) = vec![BlockCall::to(body), BlockCall::to(began)];
496 *below += all;
497 self.ahead = Some([head, body]);
498 }
499
500 /// Writes the page the stack pointer is on without changing what is there.
501 fn touch(&mut self, probe: &Probe) -> Inst {
502 let opcode = self.opcode(probe.inst);
503 let base = Operand::read(Reg::physical(self.conv.stack_pointer), self.conv.int_class);
504 self.func.build_loose(opcode).imm(0).mem(Mem::at(base)).finish()
505 }
506
507 /// Takes that many bytes off the stack pointer.
508 fn sub(&mut self, bytes: u32) -> Inst {
509 let sub = self.opcode(self.insts.sub);
510 self.arith(sub, i64::from(bytes))
511 }
512
513 /// The stack protector's check, written at the end of a block the function returns from.
514 ///
515 /// Gives back the block the epilogue goes in, which is a new one: the check has to be the last
516 /// thing the old block does, and what follows it is one of two arms rather than the return.
517 ///
518 /// ```text
519 /// block that returned reload the slot, read the word again, compare, branch
520 /// the arm it changed on call the function that does not come back, and nothing after
521 /// the arm it did not the epilogue, which the caller writes into what this gives back
522 /// ```
523 ///
524 /// The two registers are the ones the allocator was told to hold back, so nothing here has to
525 /// ask what is live: a scratch register holds nothing at the end of a block, because the only
526 /// thing that writes one is a move the rewriter put in and every one of those is read by the
527 /// instruction it was put in front of.
528 fn check(&mut self, block: Block, frame: &Frame, protect: Protect<'_>) -> Block {
529 let class = self.conv.int_class;
530 let at = frame.canary().expect("a protected function has a slot for its canary");
531 let [ours, theirs] = protect.scratch;
532
533 let inst = self.load(class, ours, at);
534 self.func.append_inst(block, inst);
535 let inst = self.read_guard(theirs, protect.guard);
536 self.func.append_inst(block, inst);
537 let differ = self.opcode(self.insts.differ);
538 let inst = self
539 .func
540 .build_loose(differ)
541 .def(Reg::physical(theirs), class)
542 .uses(Reg::physical(ours), class)
543 .uses(Reg::physical(theirs), class)
544 .finish();
545 self.func.append_inst(block, inst);
546
547 let failed = self.func.create_block();
548 let ok = self.func.create_block();
549 let cond = Opcode::new(
550 self.names.intern(&format!("{}{}", protect.branch.prefix, protect.branch.cond)),
551 );
552 let inst = self.func.build_loose(cond).uses(Reg::physical(theirs), class).finish();
553 self.func.append_inst(block, inst);
554 // The first arm is the one taken when the condition held, and the condition is that the
555 // two words differ, so the first arm is the one the canary was overwritten on.
556 *self.func.succs_mut(block) = vec![BlockCall::to(failed), BlockCall::to(ok)];
557
558 let call = self.opcode(self.insts.call);
559 let symbol = self.names.intern(protect.guard.fail);
560 self.func.build(failed, call).symbol(symbol).finish();
561 ok
562 }
563
564 /// Reads the word the canary is a copy of into a register.
565 ///
566 /// The address is a constant and names no register at all, because where the block a thread
567 /// has to itself begins is something only the machine knows and the segment register is what
568 /// holds it.
569 fn read_guard(&mut self, into: PhysReg, guard: &Guard) -> Inst {
570 let class = self.conv.int_class;
571 let load = self.opcode(self.insts.moves(class).expect("a class to load").load);
572 self.func
573 .build_loose(load)
574 .def(Reg::physical(into), class)
575 .mem(Mem::in_segment(guard.segment, guard.at))
576 .finish()
577 }
578
579 /// The instructions the epilogue is, in the order they run.
580 ///
581 /// The vector registers are read back while the stack pointer is still where the body left it,
582 /// because that is what their offsets are from. Then the stack pointer goes back to the last
583 /// register the prologue pushed, which is arithmetic when the prologue knew how far it had
584 /// moved and a read of the frame pointer when it did not.
585 fn epilogue(&mut self, frame: &Frame) -> Vec<Inst> {
586 let sp = self.conv.stack_pointer;
587 let fp = self.conv.frame_pointer;
588 let int = self.conv.int_class;
589 let sse = self.conv.sse_class;
590 let word = self.conv.word;
591 let described = !self.func.cfi.is_empty();
592 let mut out = Vec::new();
593 // Where the body left things, which is where every epilogue starts from.
594 let mut below = offset(self.conv.return_address)
595 + offset(word) * self.pushes(frame)
596 + offset(frame.size());
597 let from_sp = !frame.frame_pointer();
598 for save in frame.saved_sse() {
599 let inst = self.load(sse, save.reg, save.at);
600 out.push(inst);
601 if frame.realign().is_none() {
602 self.restored(inst, sse, save.reg);
603 }
604 }
605 let pushed = u32::try_from(frame.saved_int().len()).expect("a frame");
606 if frame.frame_pointer() {
607 // No row for either of these. The address is counted from the frame pointer here and
608 // this is what moves the stack pointer rather than the frame pointer, so the rule that
609 // was true before it is still true after it.
610 if pushed == 0 {
611 let mov = self.opcode(self.insts.moves(int).expect("a move").mov);
612 out.push(self.two(mov, sp, fp));
613 } else {
614 let lea = self.opcode(self.insts.lea);
615 let back = -offset(word * pushed);
616 out.push(self.address(lea, sp, fp, back));
617 }
618 } else if frame.size() > 0 {
619 let add = self.opcode(self.insts.add);
620 let inst = self.arith(add, i64::from(frame.size()));
621 out.push(inst);
622 below -= offset(frame.size());
623 self.row(inst, CfiOp::DefCfaOffset(below));
624 }
625 for ® in frame.saved_int().iter().rev() {
626 let inst = self.pop(reg);
627 out.push(inst);
628 self.restored(inst, int, reg);
629 below -= offset(word);
630 if from_sp {
631 self.row(inst, CfiOp::DefCfaOffset(below));
632 }
633 }
634 if frame.frame_pointer() {
635 let inst = self.pop(fp);
636 out.push(inst);
637 self.restored(inst, int, fp);
638 // The frame pointer holds the caller's value again, so the address goes back to being
639 // counted from the stack pointer, which by now is at the return address.
640 let number = self.dwarf(int, sp);
641 self.row(inst, CfiOp::DefCfa { reg: number, offset: offset(self.conv.return_address) });
642 }
643 let ret = self.opcode(self.insts.ret);
644 let inst = self.func.build_loose(ret).finish();
645 out.push(inst);
646 // These take effect at the address just past the return, which is where the next block
647 // begins, and the next block is body again. Popping the body's rules and pushing them
648 // straight back leaves the stack one deep however many blocks the function returns from,
649 // which is what makes one remembering in the prologue enough for all of them.
650 if described {
651 self.row(inst, CfiOp::RestoreState);
652 self.row(inst, CfiOp::RememberState);
653 }
654 out
655 }
656
657 /// How many general purpose registers the prologue put on the stack, the frame pointer
658 /// included.
659 fn pushes(&self, frame: &Frame) -> i32 {
660 let saved = i32::try_from(frame.saved_int().len()).expect("a frame");
661 saved + i32::from(frame.frame_pointer())
662 }
663
664 /// One row of the unwind table, taking effect after that instruction.
665 fn row(&mut self, inst: Inst, op: CfiOp) {
666 self.func.cfi.push((inst, op));
667 }
668
669 /// A row saying the caller's copy of that register is that far from the canonical frame
670 /// address, which is below it and so is negative.
671 fn saved(&mut self, inst: Inst, class: RegClass, reg: PhysReg, from_cfa: i32) {
672 let number = self.dwarf(class, reg);
673 self.row(inst, CfiOp::Offset { reg: number, offset: from_cfa });
674 }
675
676 /// A row saying that register holds what the caller left in it again.
677 fn restored(&mut self, inst: Inst, class: RegClass, reg: PhysReg) {
678 let number = self.dwarf(class, reg);
679 self.row(inst, CfiOp::Restore(number));
680 }
681
682 /// What an unwind table calls that register.
683 fn dwarf(&self, class: RegClass, reg: PhysReg) -> u16 {
684 self.conv.dwarf(class, reg).expect("a register a frame saves is one the table can name")
685 }
686
687 /// One edit as the instruction that makes it true.
688 fn mov(&mut self, edit: &Edit, frame: &Frame) -> Inst {
689 let moves = self.insts.moves(edit.class).expect("a class the target says how to move");
690 match (edit.mov.to, edit.mov.from) {
691 (Place::Reg(to), Place::Reg(from)) => {
692 let mov = self.opcode(moves.mov);
693 self.func
694 .build_loose(mov)
695 .def(Reg::physical(to), edit.class)
696 .uses(Reg::physical(from), edit.class)
697 .finish()
698 }
699 (Place::Reg(to), Place::Slot(slot)) => {
700 let at = self.slot(frame, slot);
701 self.load(edit.class, to, at)
702 }
703 (Place::Slot(slot), Place::Reg(from)) => {
704 let at = self.slot(frame, slot);
705 self.store(edit.class, from, at)
706 }
707 // The allocator expands this into two moves through a register of its own, because a
708 // machine that could do it in one is not a machine any of this is written for.
709 (Place::Slot(_), Place::Slot(_)) => {
710 unreachable!("a move from one stack slot straight into another")
711 }
712 }
713 }
714
715 /// Puts an instruction where an edit says it goes, after whatever earlier edits went there.
716 ///
717 /// The edits at one place are in the order they have to be made in, so each one goes behind
718 /// the last, and the first of them is what the place itself means.
719 fn put(&mut self, cursors: &mut Vec<(At, Inst)>, at: At, inst: Inst) {
720 if let Some(cursor) = cursors.iter_mut().find(|(place, _)| *place == at) {
721 self.func.insert_after(cursor.1, inst);
722 cursor.1 = inst;
723 return;
724 }
725 match at {
726 At::Before(before) => self.func.insert_before(before, inst),
727 At::After(after) => self.func.insert_after(after, inst),
728 At::StartOf(block) => self.func.prepend_inst(block, inst),
729 // Behind everything in the block. A block the allocator puts an edge's moves at the
730 // end of is one with a single edge out of it, and an edge like that is not an
731 // instruction here: [`crate::layout`] writes the jump it becomes after this has run.
732 // So the last instruction is an ordinary one, which may still be waiting on moves of
733 // its own that have to be made before the edge's are.
734 At::EndOf(block) => self.func.append_inst(block, inst),
735 }
736 cursors.push((at, inst));
737 }
738
739 /// Where a spill slot is, from the stack pointer in the body of the function.
740 fn slot(&self, frame: &Frame, slot: u32) -> i32 {
741 frame.slot(slot).expect("a slot the frame was worked out from")
742 }
743
744 /// Reads a register out of the frame.
745 fn load(&mut self, class: RegClass, reg: PhysReg, at: i32) -> Inst {
746 let load = self.opcode(self.insts.moves(class).expect("a class to load").load);
747 let base = Operand::read(Reg::physical(self.conv.stack_pointer), self.conv.int_class);
748 self.func
749 .build_loose(load)
750 .def(Reg::physical(reg), class)
751 .mem(Mem::at(base).plus(at))
752 .finish()
753 }
754
755 /// Writes a register into the frame.
756 fn store(&mut self, class: RegClass, reg: PhysReg, at: i32) -> Inst {
757 let store = self.opcode(self.insts.moves(class).expect("a class to store").store);
758 let base = Operand::read(Reg::physical(self.conv.stack_pointer), self.conv.int_class);
759 self.func
760 .build_loose(store)
761 .uses(Reg::physical(reg), class)
762 .mem(Mem::at(base).plus(at))
763 .finish()
764 }
765
766 /// Puts a general purpose register on the stack.
767 fn push(&mut self, reg: PhysReg) -> Inst {
768 let push = self.opcode(self.insts.push);
769 self.func.build_loose(push).uses(Reg::physical(reg), self.conv.int_class).finish()
770 }
771
772 /// Takes a general purpose register back off the stack.
773 fn pop(&mut self, reg: PhysReg) -> Inst {
774 let pop = self.opcode(self.insts.pop);
775 self.func.build_loose(pop).def(Reg::physical(reg), self.conv.int_class).finish()
776 }
777
778 /// One general purpose register written with another.
779 fn two(&mut self, opcode: Opcode, to: PhysReg, from: PhysReg) -> Inst {
780 let class = self.conv.int_class;
781 self.func
782 .build_loose(opcode)
783 .def(Reg::physical(to), class)
784 .uses(Reg::physical(from), class)
785 .finish()
786 }
787
788 /// Two-address arithmetic on the stack pointer, which reads it and writes it back.
789 fn arith(&mut self, opcode: Opcode, value: i64) -> Inst {
790 let class = self.conv.int_class;
791 let sp = Reg::physical(self.conv.stack_pointer);
792 self.func.build_loose(opcode).def(sp, class).uses(sp, class).imm(value).finish()
793 }
794
795 /// One register written with an address rather than with what is at it.
796 fn address(&mut self, opcode: Opcode, to: PhysReg, base: PhysReg, disp: i32) -> Inst {
797 let class = self.conv.int_class;
798 let base = Operand::read(Reg::physical(base), class);
799 self.func
800 .build_loose(opcode)
801 .def(Reg::physical(to), class)
802 .mem(Mem::at(base).plus(disp))
803 .finish()
804 }
805
806 /// The opcode of that name, in the machine IR's spelling, which is the target's prefix and
807 /// then the name the target gave.
808 fn opcode(&mut self, name: &str) -> Opcode {
809 Opcode::new(self.names.intern(&format!("{}{name}", self.insts.prefix)))
810 }
811}
812
813/// A distance in a frame, as the signed number every offset is.
814fn offset(bytes: u32) -> i32 {
815 i32::try_from(bytes).expect("a frame under two gigabytes")
816}
817
818#[cfg(test)]
819mod tests {
820 use rucc_base::Interner;
821 use rucc_mir::{BlockCall, print_func};
822 use rucc_regalloc::assign::Env;
823 use rucc_target::x86_64::{BRANCH, FRAME, GPR, PROBE, R10, R11, REGS, SYSV, WIN64, XMM, xmm};
824
825 use super::*;
826 use crate::frame::{Layout, Local};
827
828 /// An environment offering that many of the convention's registers, with everything after
829 /// them held back as scratch.
830 fn env(conv: &CallRegs, count: usize) -> Env {
831 Env::new().with(GPR, &conv.int_order[..count], &conv.int_order[count..])
832 }
833
834 /// A function of that many values, every one written before any is read, allocated with that
835 /// many registers to hand out. The same shape the frame layout's own tests are written
836 /// against, so that a frame here is one that has already been checked there.
837 fn pressure(conv: &CallRegs, values: usize, count: usize) -> (Func, Allocation, Interner) {
838 let mut names = Interner::new();
839 let mut func = Func::new(names.intern("f"));
840 let opcode = Opcode::new(names.intern("x64.nop"));
841 let block = func.create_block();
842 let regs: Vec<Reg> = (0..values).map(|_| func.new_vreg(GPR)).collect();
843 for ® in ®s {
844 func.build(block, opcode).def(reg, GPR).finish();
845 }
846 for ® in ®s {
847 func.build(block, opcode).uses(reg, GPR).finish();
848 }
849 let allocation = rucc_regalloc::run(&mut func, &env(conv, count), "test");
850 (func, allocation, names)
851 }
852
853 /// The function with its frame written into it, as the lines a dump would show.
854 fn written(
855 func: &mut Func,
856 allocation: &Allocation,
857 layout: &Layout<'_>,
858 names: &mut Interner,
859 ) -> Vec<String> {
860 with_protector(func, allocation, layout, None, names)
861 }
862
863 /// The same, for a function the caller has decided is protected or is not.
864 fn with_protector(
865 func: &mut Func,
866 allocation: &Allocation,
867 layout: &Layout<'_>,
868 protect: Option<Protect<'_>>,
869 names: &mut Interner,
870 ) -> Vec<String> {
871 let convention = Convention { protect, ..Convention::new(layout.conv, &FRAME) };
872 under(func, allocation, layout, convention, names)
873 }
874
875 /// The same, for a function whose frame the caller has decided is taken a page at a time.
876 fn with_probing(
877 func: &mut Func,
878 allocation: &Allocation,
879 layout: &Layout<'_>,
880 probe: Option<Probing<'_>>,
881 names: &mut Interner,
882 ) -> Vec<String> {
883 let convention = Convention { probe, ..Convention::new(layout.conv, &FRAME) };
884 under(func, allocation, layout, convention, names)
885 }
886
887 /// The function with its frame written into it under that convention.
888 fn under(
889 func: &mut Func,
890 allocation: &Allocation,
891 layout: &Layout<'_>,
892 convention: Convention<'_>,
893 names: &mut Interner,
894 ) -> Vec<String> {
895 let frame = Frame::of(func, allocation, layout);
896 finish(func, allocation, &frame, &Stack::default(), convention, names);
897 print_func(func, names, ®S)
898 .lines()
899 .filter(|line| !line.is_empty())
900 .map(|line| line.trim().to_string())
901 .collect()
902 }
903
904 /// Just the lines the frame put in, which is every line that is not the function it was
905 /// given and not the shape of the dump around it.
906 fn added(lines: &[String]) -> Vec<&str> {
907 lines
908 .iter()
909 .map(String::as_str)
910 .filter(|line| !line.contains("x64.nop"))
911 .filter(|line| !line.starts_with("mfunc") && !line.starts_with("block") && *line != "}")
912 .collect()
913 }
914
915 #[test]
916 fn a_function_that_needs_no_frame_is_given_a_return_and_nothing_else() {
917 let (mut func, allocation, mut names) = pressure(&SYSV, 2, 4);
918 let lines = written(&mut func, &allocation, &Layout::new(&SYSV, REGS), &mut names);
919
920 // Two values and four registers, so nothing is spilled, nothing is saved and the stack
921 // pointer never moves. A prologue of nothing is the right prologue for that.
922 assert_eq!(added(&lines), ["x64.ret"]);
923 }
924
925 #[test]
926 fn a_spill_is_a_store_and_a_reload_is_a_load() {
927 let (mut func, allocation, mut names) = pressure(&SYSV, 4, 2);
928 let lines = written(&mut func, &allocation, &Layout::new(&SYSV, REGS), &mut names);
929
930 // Two registers for four values, so two of them go to the stack. The store goes behind the
931 // instruction that wrote the value and the load in front of the one that wants it, both at
932 // the offsets the frame gave, which are below the stack pointer because a small leaf
933 // function is entitled to the red zone.
934 assert_eq!(
935 lines,
936 [
937 "mfunc @f {",
938 "block0:",
939 "$rax = x64.nop",
940 "$rcx = x64.nop",
941 "$rdx = x64.nop",
942 "x64.mov_mr_64 $rdx, [$rsp - 16]",
943 "$rdx = x64.nop",
944 "x64.mov_mr_64 $rdx, [$rsp - 8]",
945 "x64.nop $rax",
946 "x64.nop $rcx",
947 "$rdx = x64.mov_rm_64 [$rsp - 16]",
948 "x64.nop $rdx",
949 "$rdx = x64.mov_rm_64 [$rsp - 8]",
950 "x64.nop $rdx",
951 "x64.ret",
952 "}",
953 ]
954 );
955 }
956
957 #[test]
958 fn the_frame_the_prologue_takes_is_the_frame_the_epilogue_gives_back() {
959 let (mut func, allocation, mut names) = pressure(&SYSV, 4, 2);
960 let base = Layout::new(&SYSV, REGS);
961 let layout = Layout { red_zone: false, ..base };
962 let lines = written(&mut func, &allocation, &layout, &mut names);
963
964 // The same function told it may not use the red zone takes sixteen bytes instead, and
965 // every offset moves above the stack pointer to match.
966 assert_eq!(
967 added(&lines),
968 [
969 "$rsp = x64.sub_ri_64 $rsp, 16",
970 "x64.mov_mr_64 $rdx, [$rsp]",
971 "x64.mov_mr_64 $rdx, [$rsp + 8]",
972 "$rdx = x64.mov_rm_64 [$rsp]",
973 "$rdx = x64.mov_rm_64 [$rsp + 8]",
974 "$rsp = x64.add_ri_64 $rsp, 16",
975 "x64.ret",
976 ]
977 );
978 }
979
980 #[test]
981 fn the_registers_the_prologue_pushes_come_back_in_the_opposite_order() {
982 let (mut func, allocation, mut names) = pressure(&SYSV, 13, 13);
983 let lines = written(&mut func, &allocation, &Layout::new(&SYSV, REGS), &mut names);
984
985 // Four registers a call leaves alone, pushed in the convention's order and popped in the
986 // other one, which is the only order that gets each of them its own value back.
987 assert_eq!(
988 added(&lines),
989 [
990 "x64.push_64 $rbx",
991 "x64.push_64 $r12",
992 "x64.push_64 $r13",
993 "x64.push_64 $r14",
994 "$r14 = x64.pop_64",
995 "$r13 = x64.pop_64",
996 "$r12 = x64.pop_64",
997 "$rbx = x64.pop_64",
998 "x64.ret",
999 ]
1000 );
1001 }
1002
1003 #[test]
1004 fn a_function_that_keeps_a_frame_pointer_sets_it_up_and_leaves_by_it() {
1005 let (mut func, allocation, mut names) = pressure(&SYSV, 4, 2);
1006 let base = Layout::new(&SYSV, REGS);
1007 let layout = Layout { frame_pointer: true, red_zone: false, ..base };
1008 let lines = written(&mut func, &allocation, &layout, &mut names);
1009
1010 // The frame pointer is saved before anything else and points at where it was saved, so the
1011 // epilogue reaches the stack pointer through it rather than by counting the frame back.
1012 assert_eq!(
1013 added(&lines),
1014 [
1015 "x64.push_64 $rbp",
1016 "$rbp = x64.mov_rr_64 $rsp",
1017 "$rsp = x64.sub_ri_64 $rsp, 16",
1018 "x64.mov_mr_64 $rdx, [$rsp]",
1019 "x64.mov_mr_64 $rdx, [$rsp + 8]",
1020 "$rdx = x64.mov_rm_64 [$rsp]",
1021 "$rdx = x64.mov_rm_64 [$rsp + 8]",
1022 "$rsp = x64.mov_rr_64 $rbp",
1023 "$rbp = x64.pop_64",
1024 "x64.ret",
1025 ]
1026 );
1027 }
1028
1029 #[test]
1030 fn a_realigned_frame_forces_the_alignment_after_it_has_pushed_what_it_saves() {
1031 let (mut func, allocation, mut names) = pressure(&SYSV, 13, 13);
1032 let locals = [Local { size: 64, align: 32 }];
1033 let base = Layout::new(&SYSV, REGS);
1034 let layout = Layout { locals: &locals, ..base };
1035 let lines = written(&mut func, &allocation, &layout, &mut names);
1036
1037 // Forcing the alignment throws away how far the stack pointer had moved, so the registers
1038 // are pushed before it happens and the epilogue counts back from the frame pointer to find
1039 // them. The frame pointer is required here whatever the flags said.
1040 assert_eq!(
1041 added(&lines),
1042 [
1043 "x64.push_64 $rbp",
1044 "$rbp = x64.mov_rr_64 $rsp",
1045 "x64.push_64 $rbx",
1046 "x64.push_64 $r12",
1047 "x64.push_64 $r13",
1048 "x64.push_64 $r14",
1049 "$rsp = x64.and_ri_64 $rsp, -32",
1050 "$rsp = x64.sub_ri_64 $rsp, 64",
1051 "$rsp = x64.lea_64 [$rbp - 32]",
1052 "$r14 = x64.pop_64",
1053 "$r13 = x64.pop_64",
1054 "$r12 = x64.pop_64",
1055 "$rbx = x64.pop_64",
1056 "$rbp = x64.pop_64",
1057 "x64.ret",
1058 ]
1059 );
1060 }
1061
1062 #[test]
1063 fn every_block_the_function_returns_from_gets_an_epilogue() {
1064 let mut names = Interner::new();
1065 let mut func = Func::new(names.intern("f"));
1066 let opcode = Opcode::new(names.intern("x64.nop"));
1067 let head = func.create_block();
1068 let left = func.create_block();
1069 let right = func.create_block();
1070 func.build(head, opcode).finish();
1071 *func.succs_mut(head) = vec![BlockCall::to(left), BlockCall::to(right)];
1072 func.build(left, opcode).finish();
1073 func.build(right, opcode).finish();
1074 let allocation = rucc_regalloc::run(&mut func, &env(&SYSV, 4), "test");
1075 let base = Layout::new(&SYSV, REGS);
1076 let layout = Layout { leaf: false, ..base };
1077 let lines = written(&mut func, &allocation, &layout, &mut names);
1078
1079 // Both ways out get the frame given back, and the block that goes somewhere gets nothing,
1080 // because a block with an edge out of it is not a block anything returns from.
1081 assert_eq!(
1082 lines,
1083 [
1084 "mfunc @f {",
1085 "block0:",
1086 "$rsp = x64.sub_ri_64 $rsp, 8",
1087 "x64.nop block1, block2",
1088 "block1:",
1089 "x64.nop",
1090 "$rsp = x64.add_ri_64 $rsp, 8",
1091 "x64.ret",
1092 "block2:",
1093 "x64.nop",
1094 "$rsp = x64.add_ri_64 $rsp, 8",
1095 "x64.ret",
1096 "}",
1097 ]
1098 );
1099 }
1100
1101 #[test]
1102 fn a_protected_function_writes_the_canary_last_and_checks_it_before_it_returns() {
1103 let (mut func, allocation, mut names) = pressure(&SYSV, 4, 2);
1104 let base = Layout::new(&SYSV, REGS);
1105 let layout = Layout { leaf: false, protect: true, ..base };
1106 let guard = SYSV.guard.as_ref().expect("this convention has somewhere to keep the word");
1107 // The two the real pipeline holds back, which are held back in the environment above too:
1108 // it hands out the first two of the convention's order and keeps everything after them.
1109 let protect = Protect { guard, branch: &BRANCH, scratch: [R10, R11] };
1110 let lines = with_protector(&mut func, &allocation, &layout, Some(protect), &mut names);
1111
1112 // The read of the word and the store into the slot come after the stack pointer has moved,
1113 // because there is no slot to store into until it has. The check is the last thing the
1114 // block that returned does and the epilogue is on the arm the canary was unchanged on, so
1115 // a function whose canary changed never gives its frame back and never returns.
1116 assert_eq!(
1117 added(&lines),
1118 [
1119 "$rsp = x64.sub_ri_64 $rsp, 24",
1120 "$r10 = x64.mov_rm_64 [fs:40]",
1121 "x64.mov_mr_64 $r10, [$rsp + 16]",
1122 "x64.mov_mr_64 $rdx, [$rsp]",
1123 "x64.mov_mr_64 $rdx, [$rsp + 8]",
1124 "$rdx = x64.mov_rm_64 [$rsp]",
1125 "$rdx = x64.mov_rm_64 [$rsp + 8]",
1126 "$r10 = x64.mov_rm_64 [$rsp + 16]",
1127 "$r11 = x64.mov_rm_64 [fs:40]",
1128 "$r11 = x64.cmp_set_ne_64 $r10, $r11",
1129 "x64.br_cond_8 $r11, block1, block2",
1130 "x64.call @__stack_chk_fail",
1131 "$rsp = x64.add_ri_64 $rsp, 24",
1132 "x64.ret",
1133 ]
1134 );
1135 }
1136
1137 #[test]
1138 fn a_frame_that_fits_in_one_page_is_taken_in_one_subtraction_even_when_pages_are_touched() {
1139 let (mut func, allocation, mut names) = pressure(&SYSV, 2, 4);
1140 let locals = [Local { size: 4088, align: 16 }];
1141 let base = Layout::new(&SYSV, REGS);
1142 let layout = Layout { leaf: false, locals: &locals, ..base };
1143 let probing = Probing { probe: &PROBE, branch: &BRANCH, scratch: [R10, R11] };
1144 let lines = with_probing(&mut func, &allocation, &layout, Some(probing), &mut names);
1145
1146 // A frame of one page cannot step over the page below it, because the far end of it is the
1147 // near end of that page and anything written there is written to a page that is there. So
1148 // the flag costs such a function nothing, which is most functions.
1149 assert_eq!(
1150 added(&lines),
1151 ["$rsp = x64.sub_ri_64 $rsp, 4088", "$rsp = x64.add_ri_64 $rsp, 4088", "x64.ret",]
1152 );
1153 }
1154
1155 #[test]
1156 fn a_probing_prologue_touches_every_page_of_a_frame_a_few_pages_deep() {
1157 let (mut func, allocation, mut names) = pressure(&SYSV, 2, 4);
1158 let locals = [Local { size: 9000, align: 16 }];
1159 let base = Layout::new(&SYSV, REGS);
1160 let layout = Layout { leaf: false, locals: &locals, ..base };
1161 let probing = Probing { probe: &PROBE, branch: &BRANCH, scratch: [R10, R11] };
1162 let lines = with_probing(&mut func, &allocation, &layout, Some(probing), &mut names);
1163
1164 // A page of the stack pointer's own, then the touch that says the page is there, and only
1165 // then the next one, which is the whole of the defence: nothing here ever moves the stack
1166 // pointer further than one page without writing where it landed. The last subtraction is
1167 // the remainder and is smaller than a page, so it needs no touch of its own, and it exists
1168 // in every frame because the count of pages is taken off one less than the size.
1169 assert_eq!(
1170 added(&lines),
1171 [
1172 "$rsp = x64.sub_ri_64 $rsp, 4096",
1173 "x64.or_mi_8 [$rsp], 0",
1174 "$rsp = x64.sub_ri_64 $rsp, 4096",
1175 "x64.or_mi_8 [$rsp], 0",
1176 "$rsp = x64.sub_ri_64 $rsp, 808",
1177 "$rsp = x64.add_ri_64 $rsp, 9000",
1178 "x64.ret",
1179 ]
1180 );
1181 }
1182
1183 #[test]
1184 fn a_probing_prologue_deeper_than_that_walks_the_pages_in_a_loop() {
1185 let (mut func, allocation, mut names) = pressure(&SYSV, 2, 4);
1186 let locals = [Local { size: 100_000, align: 16 }];
1187 let base = Layout::new(&SYSV, REGS);
1188 let layout = Layout { leaf: false, locals: &locals, ..base };
1189 let probing = Probing { probe: &PROBE, branch: &BRANCH, scratch: [R10, R11] };
1190 let lines = with_probing(&mut func, &allocation, &layout, Some(probing), &mut names);
1191
1192 // Twenty-four pages, which is more than a straight line is worth, so the prologue works out
1193 // where it is going first and then walks there. The whole listing rather than the added
1194 // lines, because what matters as much as the instructions is that the two blocks the walk
1195 // is made of come in front of the block the function began with: the body the allocator
1196 // filled is block2 here and it was block0 before this ran.
1197 assert_eq!(
1198 lines,
1199 [
1200 "mfunc @f {",
1201 "block0:",
1202 "$r10 = x64.lea_64 [$rsp - 98304], block1",
1203 "block1:",
1204 "$rsp = x64.sub_ri_64 $rsp, 4096",
1205 "x64.or_mi_8 [$rsp], 0",
1206 "$r11 = x64.cmp_set_ne_64 $rsp, $r10",
1207 "x64.br_cond_8 $r11, block1, block2",
1208 "block2:",
1209 "$rsp = x64.sub_ri_64 $rsp, 1704",
1210 "$rax = x64.nop",
1211 "$rcx = x64.nop",
1212 "x64.nop $rax",
1213 "x64.nop $rcx",
1214 "$rsp = x64.add_ri_64 $rsp, 100008",
1215 "x64.ret",
1216 "}",
1217 ]
1218 );
1219 }
1220
1221 #[test]
1222 fn a_vector_register_a_windows_call_preserves_is_stored_and_read_back() {
1223 let mut names = Interner::new();
1224 let mut func = Func::new(names.intern("f"));
1225 let opcode = Opcode::new(names.intern("x64.nop"));
1226 let block = func.create_block();
1227 // An instruction that writes one of the vector registers Windows preserves, which is what
1228 // a rule for something that has to use it produces.
1229 func.build(block, opcode).operand(Operand::write(Reg::physical(xmm(6)), XMM)).finish();
1230 let allocation = rucc_regalloc::run(&mut func, &env(&WIN64, 4), "test");
1231 let lines = written(&mut func, &allocation, &Layout::new(&WIN64, REGS), &mut names);
1232
1233 // No machine here pushes a vector register, so it is stored into the frame rather than
1234 // pushed, and the frame has to be taken before there is anywhere to put it.
1235 assert_eq!(
1236 added(&lines),
1237 [
1238 "$rsp = x64.sub_ri_64 $rsp, 24",
1239 "x64.movaps_mr $xmm6, [$rsp]",
1240 "$xmm6 = x64.movaps_rm [$rsp]",
1241 "$rsp = x64.add_ri_64 $rsp, 24",
1242 "x64.ret",
1243 ]
1244 );
1245 }
1246}