Skip to main content

rucc_codegen/
split.rs

1//! Splitting critical edges, so that every edge that carries values has somewhere to put them.
2//!
3//! Design: `spec/10-backend.md` section 10.4.
4//!
5//! An edge carries values when the block it goes to takes parameters, and giving a parameter its
6//! value is a move. The move has to happen on the edge and not before it or after it, because
7//! before it is a block that goes somewhere else too and after it is a block that is arrived at
8//! from somewhere else too, and in either case the move would run on a path it was not written
9//! for. An edge out of a block with one successor can put its moves at the end of that block,
10//! since every path through it takes the edge. An edge into a block with one predecessor can put
11//! them at the start of that block, for the same reason the other way round. An edge that is
12//! neither, which is what a critical edge is, has neither place, and the allocator says so:
13//! `rucc_regalloc` asserts that it never sees one.
14//!
15//! So one is turned into two. A block with nothing in it goes on the edge, the arguments move on
16//! to the second half, and both halves are now uncritical: the first goes to a block with one
17//! predecessor and the second leaves a block with one successor. Which of the two the moves end
18//! up in is the allocator's answer and not this one's, and either is correct.
19//!
20//! # What it leaves behind
21//!
22//! An empty block, which is a jump to the next thing unless the layout puts it where it falls
23//! through. That is a cost, and it is why an edge with nothing to carry is left alone: there are
24//! no moves to find a place for, so splitting it would buy a jump and nothing else.
25//!
26//! # The other edge with nowhere to put a move
27//!
28//! A computed `goto` leaves its block through a register, and the moves an edge out of it carries
29//! would have to be written somewhere the jump has already gone past. So there is a second pass
30//! here, [`indirect`], which takes the values off those edges and puts them in a block of their
31//! own in front of each label. It runs first, and what it leaves behind is edges the splitting
32//! below then has nothing to do about.
33//!
34//! [`pads`] is here for the same reason and not for a reason of its own: the blocks those labels
35//! begin at are addresses an indirect branch arrives at, and a machine that checks the forward edge
36//! wants a landing pad at every one of them. Which block an address names is settled by the pass
37//! above, so the pad is written after it and not where the prologue's own pad is written.
38
39use std::collections::HashMap;
40
41use rucc_base::Interner;
42use rucc_mir as mir;
43use rucc_target::{BranchInsts, FrameInsts, RegClass};
44
45/// Splits every critical edge that carries values, and gives back how many it split.
46///
47/// Run after lowering and before allocation. Running it twice is running it once, because the
48/// blocks it adds have one successor each and are never the source of a critical edge.
49pub fn critical(func: &mut mir::Func) -> usize {
50    let preds = preds(func);
51    let blocks: Vec<mir::Block> = func.blocks().collect();
52    let mut split = 0;
53    for block in blocks {
54        if func[block].succs.len() < 2 {
55            continue;
56        }
57        for index in 0..func[block].succs.len() {
58            let call = func[block].succs[index].clone();
59            if call.args.is_empty() || preds[call.block.index()] < 2 {
60                continue;
61            }
62            // The new block is at the end of the layout, which is where a block that is a jump
63            // and nothing else does the least harm before the layout pass has an opinion.
64            //
65            // It runs exactly as often as the edge it sits on is taken, and both halves of that
66            // edge are now that edge, which is why the weight is copied onto all three rather
67            // than left at what a block nobody told anything runs. A block on a cold edge that
68            // claimed to run once per call would be one the layout put in the middle of the hot
69            // path.
70            let weight = call.weight;
71            let half = func.create_block();
72            func.set_weight(half, weight);
73            *func.succs_mut(half) = vec![call];
74            func.succs_mut(block)[index] = mir::BlockCall::to(half).taken(weight);
75            split += 1;
76        }
77    }
78    split
79}
80
81/// Takes the values off every edge out of a computed `goto`, and gives back how many blocks it
82/// made to hold them.
83///
84/// Run after lowering and before [`critical`], which then sees edges with nothing on them and
85/// leaves them alone. Running it twice is running it once, for the reason the splitting above is:
86/// the blocks it adds end in a jump rather than in a branch through a register.
87///
88/// # What is wrong with the edge it takes the values off
89///
90/// Every other edge in the function is out of a block whose last instruction the layout writes, so
91/// an edge that is the only way out of its block can put its moves at the end of that block and
92/// they land in front of the jump. A block that leaves through a register already ends in the jump
93/// when the allocator runs, because where it goes is a value and a value is something selection
94/// reads rather than something the layout knows. Moves at the end of that block would be written
95/// after the jump, where nothing runs them, and moves in front of it would be written across the
96/// register the jump reads, which the allocator believes is dead from the jump onwards and is free
97/// to hand to one of the moves.
98///
99/// So the moves go somewhere else. Each label an indirect branch reaches gets a block in front of
100/// it that carries the values, the branch goes to that block with nothing on the edge, and the
101/// address the `&&label` produces is the address of that block rather than of the label's own. The
102/// new block is arrived at one way and leaves one way, so its own edge has both of the places the
103/// splitting above talks about and the allocator is content.
104///
105/// # One label, one address, and two branches that disagree
106///
107/// A label has one address, so two computed `goto`s that reach it both arrive at whatever block
108/// that address names, and the values they carry are not the same values. One block in front of
109/// the label cannot move two different sets of registers.
110///
111/// So they are made to agree first. Each parameter of the label gets a register of its own, every
112/// branch writes that register in front of its jump, and the block in front of the label carries
113/// those registers and nothing else. That is what gcc does about the same problem, which it calls
114/// coalescing across an abnormal edge, done here rather than while the values are still the
115/// optimizer's.
116///
117/// Writing them in front of the jump is safe, which is not obvious, since a branch that goes five
118/// ways writes the registers of one of those ways on the path to all five. What makes it safe is
119/// that nothing reads those registers except the block in front of the label, and the only way to
120/// reach that block is an edge out of a branch, which writes them on the way. So a value written
121/// here and not used is a value overwritten before anything looks, whichever way the jump went.
122///
123/// # One register for one value, and not one for every place it is given to
124///
125/// That safety is also what makes the cost of it worth watching. A branch writes the registers of
126/// every label it can reach, so a register for every parameter of every label is a whole table's
127/// worth of moves in front of every jump in the function, and a dispatch table is a branch that
128/// reaches hundreds of labels. An interpreter hands each of them whatever its loop had in hand at
129/// the jump, which is the same few values over and over, so a register for each place one of them
130/// lands means those values written a hundred times over before every instruction the interpreter
131/// runs. That is not a small constant. It is what makes an interpreter built this way ten times
132/// slower than the same interpreter built with a `switch` instead of the computed `goto`.
133///
134/// So the register belongs to the value rather than to the place. Two parameters are given one
135/// register when they are drawn from the same class and every branch in the function gives them
136/// the same register, which is exactly when one register can stand for both, and a branch writes
137/// each register it has to write once however many labels asked for it. A dispatch table where
138/// every label wants the instruction pointer writes the instruction pointer once. A label
139/// something else reaches, or a label given something no other label is given, keeps a register of
140/// its own, and a branch that gives one label nothing shares nothing with it, since a register
141/// that branch never wrote is not one the label can be given.
142///
143/// # Panics
144///
145/// Panics on a class of register the machine named no move for, which is a function carrying a
146/// value of a kind the target never said how to copy, and on a branch that has lost the terminator
147/// it was found by, which nothing between the finding and the use of it can do. Both are a target
148/// description or a function that was built wrongly, and both are worth finding here rather than as
149/// a value that arrives somewhere it was never written.
150pub fn indirect(
151    func: &mut mir::Func,
152    branch: &BranchInsts,
153    frame: &FrameInsts,
154    names: &mut Interner,
155) -> usize {
156    let jump = mir::Opcode::new(names.intern(&format!("{}{}", branch.prefix, branch.indirect)));
157    let branches: Vec<mir::Block> = func
158        .blocks()
159        .filter(|&block| func.terminator(block).is_some_and(|last| func[last].opcode == jump))
160        .collect();
161    // Nothing at all in almost every function, and the walk at the bottom is over every instruction
162    // in it, so the answer is arrived at here rather than paid for everywhere.
163    if branches.is_empty() {
164        return 0;
165    }
166    // In the order the branches name them rather than in whatever order a hash gives, so that two
167    // runs of the compiler over one program write the same blocks.
168    let mut targets: Vec<mir::Block> = Vec::new();
169    for &block in &branches {
170        for call in &func[block].succs {
171            if !call.args.is_empty() && !targets.contains(&call.block) {
172                targets.push(call.block);
173            }
174        }
175    }
176
177    // One register per thing a branch has to give, rather than one per place it is given to. The
178    // key is what every branch gives that parameter, so two parameters given the same register by
179    // the same branches are given it in one register and a branch writes that register once.
180    let mut homes: HashMap<Given, mir::Reg> = HashMap::new();
181    // What each branch writes in front of its jump, in the order it was first asked for, and never
182    // the same register twice. Two parameters that share a register are given it by the one move.
183    let mut writes: Vec<Vec<(mir::Reg, mir::Reg, RegClass)>> = vec![Vec::new(); branches.len()];
184    let mut entries: HashMap<mir::Block, mir::Block> = HashMap::new();
185
186    for target in targets {
187        let params = func[target].params.clone();
188        let given = given(func, &branches, target);
189        let mut carried: Vec<mir::Reg> = Vec::new();
190        for (index, param) in params.iter().enumerate() {
191            let key: Given = (
192                param.class,
193                given.iter().map(|edges| edges.iter().map(|args| args[index]).collect()).collect(),
194            );
195            let home = *homes.entry(key).or_insert_with(|| func.new_vreg(param.class));
196            carried.push(home);
197            for (branch, edges) in given.iter().enumerate() {
198                for args in edges {
199                    if !writes[branch].iter().any(|&(written, _, _)| written == home) {
200                        writes[branch].push((home, args[index], param.class));
201                    }
202                }
203            }
204        }
205        let entry = func.create_block();
206        let mut total = mir::Weight::NEVER;
207        for &block in &branches {
208            for index in 0..func[block].succs.len() {
209                if func[block].succs[index].block != target {
210                    continue;
211                }
212                // The block in front of the label runs as often as every branch that reaches it,
213                // which is the same sum the weight of a block with that many edges into it would
214                // be.
215                let weight = func[block].succs[index].weight;
216                total = mir::Weight::parts(total.raw().saturating_add(weight.raw()));
217                func.succs_mut(block)[index] = mir::BlockCall::to(entry).taken(weight);
218            }
219        }
220        func.set_weight(entry, total);
221        *func.succs_mut(entry) = vec![mir::BlockCall::with(target, carried).taken(total)];
222        entries.insert(target, entry);
223    }
224
225    // And the moves themselves, once every label has asked for what it wants, since what one label
226    // asks for is what another may already have asked the same branch for.
227    for (branch, moves) in branches.iter().zip(&writes) {
228        let last = func.terminator(*branch).expect("a block that ends in a jump");
229        for &(home, arg, class) in moves {
230            let name = frame.moves(class).expect("a class this machine can move").mov;
231            let opcode = mir::Opcode::new(names.intern(&format!("{}{name}", frame.prefix)));
232            let inst = func.build_loose(opcode).def(home, class).uses(arg, class).finish();
233            func.insert_before(last, inst);
234        }
235    }
236
237    // And the addresses, which is the half of this that is not about edges. Every `&&label` in the
238    // function names a block, and a label with a block in front of it now begins at that block, so
239    // an address left pointing at the label's own block would be a jump past the moves.
240    let mut addresses: Vec<mir::MemRef> = Vec::new();
241    for block in func.blocks() {
242        for inst in func.insts(block) {
243            if let Some(mem) = func[inst].mem {
244                addresses.push(mem);
245            }
246        }
247    }
248    for mem in addresses {
249        if let Some(named) = func[mem].block {
250            if let Some(&entry) = entries.get(&named) {
251                func[mem].block = Some(entry);
252            }
253        }
254    }
255    // And the names, for the same reason. A block an image points at is one a `goto *p` arrives at,
256    // so a name left on the label's own block would be an address in a table that skips the moves,
257    // which is the one way into the block that would not have made them.
258    for (block, _) in &mut func.labels {
259        if let Some(&entry) = entries.get(block) {
260            *block = entry;
261        }
262    }
263    entries.len()
264}
265
266/// Puts a landing pad at the front of every block whose address is taken, and gives back how many
267/// it wrote.
268///
269/// Run after [`indirect`], because the block an address names is not settled until that has moved
270/// the addresses on to the blocks it made, and only when the command line asked for the forward
271/// edge to be checked. Nothing is written otherwise, which is why the name comes in as an option
272/// and why a target with no such instruction is a target this does nothing on.
273///
274/// The pad a prologue opens with is written elsewhere, in `crate::finish`, because the address it
275/// makes reachable is the address of the function rather than a place inside it. These are the
276/// other addresses an indirect branch may arrive at, and a machine that checks the forward edge
277/// faults on one that has no pad, so a computed `goto` compiled without this would be a program
278/// that ran everywhere except on the hardware the flag was turned on for.
279pub fn pads(
280    func: &mut mir::Func,
281    frame: &FrameInsts,
282    landing: Option<&'static str>,
283    names: &mut Interner,
284) -> usize {
285    let Some(name) = landing else { return 0 };
286    let opcode = mir::Opcode::new(names.intern(&format!("{}{name}", frame.prefix)));
287    let mut addressed: Vec<mir::Block> = Vec::new();
288    for block in func.blocks() {
289        for inst in func.insts(block) {
290            if let Some(mem) = func[inst].mem {
291                if let Some(named) = func[mem].block {
292                    if !addressed.contains(&named) {
293                        addressed.push(named);
294                    }
295                }
296            }
297        }
298    }
299    for &block in &addressed {
300        let inst = func.build_loose(opcode).finish();
301        func.prepend_inst(block, inst);
302    }
303    addressed.len()
304}
305
306/// What decides whether two parameters can be given their value in one register: the class the
307/// parameter is drawn from, and the register every branch in the function gives it, in the order
308/// the branches are in and with one entry per edge inside that. A branch that does not reach the
309/// label gives nothing, which is a length of zero and is as much a part of the answer as a
310/// register is, since sharing with a parameter a branch never gives anything to would be reading a
311/// register that branch never wrote.
312type Given = (RegClass, Vec<Vec<mir::Reg>>);
313
314/// What each branch gives that label, edge by edge.
315///
316/// One entry per branch and in the branches' own order, since a label two branches reach and a
317/// label one branch reaches twice are not given the same thing. A branch is allowed to reach one
318/// label twice, which a table with the same label in two of its cells is, so what a branch gives
319/// is a list of what it gives rather than one set of registers.
320fn given(func: &mir::Func, branches: &[mir::Block], target: mir::Block) -> Vec<Vec<Vec<mir::Reg>>> {
321    branches
322        .iter()
323        .map(|&block| {
324            func[block]
325                .succs
326                .iter()
327                .filter(|call| call.block == target)
328                .map(|call| call.args.clone())
329                .collect()
330        })
331        .collect()
332}
333
334/// How many edges arrive at each block, counted by index rather than in layout order so that a
335/// block added while splitting can be looked up in the same table.
336fn preds(func: &mir::Func) -> Vec<usize> {
337    let mut counts = vec![0; func.block_count()];
338    for block in func.blocks() {
339        for call in &func[block].succs {
340            counts[call.block.index()] += 1;
341        }
342    }
343    counts
344}
345
346#[cfg(test)]
347mod tests {
348    use rucc_base::Interner;
349    use rucc_target::x86_64::{BRANCH, FRAME, GPR, REGS};
350
351    use super::*;
352
353    /// A diamond: one block that goes two ways and one block both ways arrive at, with as many
354    /// parameters on the block they arrive at as the test asks for.
355    fn diamond(params: usize) -> (Interner, mir::Func, [mir::Block; 4]) {
356        let mut names = Interner::new();
357        let mut func = mir::Func::new(names.intern("f"));
358        let head = func.create_block();
359        let left = func.create_block();
360        let right = func.create_block();
361        let join = func.create_block();
362        // The values arrive in the head, so that they have somewhere to be defined and the
363        // printer has a name for them. Nothing here runs an allocator, which is the one thing
364        // that would object to a first block with parameters.
365        let args: Vec<mir::Reg> = (0..params).map(|_| func.append_param(head, GPR)).collect();
366        for _ in 0..params {
367            func.append_param(join, GPR);
368        }
369        *func.succs_mut(head) = vec![mir::BlockCall::to(left), mir::BlockCall::to(right)];
370        *func.succs_mut(left) = vec![mir::BlockCall::with(join, args.clone())];
371        *func.succs_mut(right) = vec![mir::BlockCall::with(join, args)];
372        (names, func, [head, left, right, join])
373    }
374
375    /// Where each block goes, which is the whole of what this changes.
376    fn edges(func: &mir::Func) -> Vec<Vec<usize>> {
377        func.blocks()
378            .map(|block| func[block].succs.iter().map(|call| call.block.index()).collect())
379            .collect()
380    }
381
382    #[test]
383    fn an_edge_that_is_the_only_way_out_is_left_alone() {
384        let (_, mut func, _) = diamond(1);
385        // The two edges into the join carry a value each and neither is critical, because the
386        // block each leaves goes nowhere else.
387        assert_eq!(critical(&mut func), 0);
388        assert_eq!(edges(&func), vec![vec![1, 2], vec![3], vec![3], vec![]]);
389    }
390
391    #[test]
392    fn a_critical_edge_carrying_a_value_is_split_in_two() {
393        let (_, mut func, [head, _, _, join]) = diamond(1);
394        // Now the head goes straight to the join as well, so both of its arms are critical: it
395        // has two ways out and the join has three ways in.
396        let arg = func.append_param(head, GPR);
397        func.succs_mut(head).push(mir::BlockCall::with(join, vec![arg]));
398        func.succs_mut(head).swap(1, 2);
399
400        assert_eq!(critical(&mut func), 1);
401        assert_eq!(
402            edges(&func),
403            // The head's second arm is the new block and the new block goes to the join. The
404            // other two arms are untouched, because each goes to a block with one way in.
405            vec![vec![1, 4, 2], vec![3], vec![3], vec![], vec![3]]
406        );
407    }
408
409    #[test]
410    fn a_critical_edge_carrying_nothing_is_left_alone() {
411        let (_, mut func, [head, _, _, join]) = diamond(0);
412        func.succs_mut(head).push(mir::BlockCall::to(join));
413
414        // Critical and not split, because there is no move to find a place for and a block that
415        // is a jump and nothing else is worth more than nothing.
416        assert_eq!(critical(&mut func), 0);
417    }
418
419    #[test]
420    fn the_arguments_move_on_to_the_half_that_arrives() {
421        let (names, mut func, [head, _, _, join]) = diamond(1);
422        let arg = func.append_param(head, GPR);
423        func.succs_mut(head).push(mir::BlockCall::with(join, vec![arg]));
424
425        assert_eq!(critical(&mut func), 1);
426        // What the first half carries is nothing, since the block it goes to asks for nothing,
427        // and what the second half carries is what the whole edge used to.
428        let half = func.blocks().last().expect("the block the split added");
429        assert_eq!(func[head].succs[2].args, Vec::new());
430        assert_eq!(func[half].succs[0].args, vec![arg]);
431        assert_eq!(
432            mir::print_func(&func, &names, &REGS),
433            "mfunc @f {\nblock0(%0:gpr, %1:gpr):\n    block1, block2, block4\n\n\
434             block1:\n    block3(%0)\n\nblock2:\n    block3(%0)\n\n\
435             block3(%2:gpr):\n\nblock4:\n    block3(%1)\n}\n"
436        );
437    }
438
439    #[test]
440    fn splitting_twice_is_splitting_once() {
441        let (_, mut func, [head, _, _, join]) = diamond(1);
442        let arg = func.append_param(head, GPR);
443        func.succs_mut(head).push(mir::BlockCall::with(join, vec![arg]));
444
445        assert_eq!(critical(&mut func), 1);
446        assert_eq!(critical(&mut func), 0);
447    }
448
449    /// A function with one label whose address is taken and as many blocks leaving through that
450    /// address as the test asks for, each carrying as many values to the label as it asks for.
451    fn computed(branches: usize, params: usize) -> (Interner, mir::Func) {
452        let mut names = Interner::new();
453        let mut func = mir::Func::new(names.intern("f"));
454        let head = func.create_block();
455        let label = func.create_block();
456        for _ in 0..params {
457            func.append_param(label, GPR);
458        }
459        let lea = mir::Opcode::new(names.intern("x64.lea_64"));
460        let jump = mir::Opcode::new(names.intern("x64.jmp_reg"));
461        for _ in 0..branches {
462            // Every branch works the address out for itself, which is what a program that takes
463            // the address of a label twice looks like once the values are in registers.
464            let args: Vec<mir::Reg> = (0..params).map(|_| func.append_param(head, GPR)).collect();
465            let address = func.new_vreg(GPR);
466            let at = if branches == 1 { head } else { func.create_block() };
467            func.build(at, lea).def(address, GPR).mem(mir::Mem::block(label)).finish();
468            func.build(at, jump).operand(mir::Operand::read(address, GPR)).finish();
469            *func.succs_mut(at) = vec![mir::BlockCall::with(label, args)];
470        }
471        (names, func)
472    }
473
474    /// Which block each address in the function names, in the order the instructions are in.
475    fn addressed(func: &mir::Func) -> Vec<usize> {
476        func.blocks()
477            .flat_map(|block| func.insts(block).collect::<Vec<_>>())
478            .filter_map(|inst| func[inst].mem)
479            .filter_map(|mem| func[mem].block)
480            .map(mir::Block::index)
481            .collect()
482    }
483
484    #[test]
485    fn the_values_a_computed_goto_carries_move_into_a_block_in_front_of_the_label() {
486        let (mut names, mut func) = computed(1, 1);
487        assert_eq!(indirect(&mut func, &BRANCH, &FRAME, &mut names), 1);
488
489        // The branch goes to the new block carrying nothing, and the new block carries the value
490        // the branch used to. The address the `lea` works out is the new block's as well, since
491        // arriving at the label without going through the new block is arriving without the value.
492        assert_eq!(edges(&func), vec![vec![2], vec![], vec![1]]);
493        assert_eq!(func[mir::Block::new(0)].succs[0].args, Vec::new());
494        assert_eq!(addressed(&func), vec![2]);
495    }
496
497    #[test]
498    fn two_computed_gotos_that_reach_one_label_are_made_to_agree() {
499        let (mut names, mut func) = computed(2, 1);
500        assert_eq!(indirect(&mut func, &BRANCH, &FRAME, &mut names), 1);
501
502        // One block in front of the label and not two, because the label has one address and both
503        // branches arrive at it. What makes that sound is the move each branch writes in front of
504        // its own jump, which puts its value in the register that block carries.
505        assert_eq!(edges(&func), vec![vec![], vec![], vec![4], vec![4], vec![1]]);
506        let text = mir::print_func(&func, &names, &REGS);
507        assert_eq!(text.matches("x64.mov_rr_64").count(), 2, "{text}");
508        // In front of the jump rather than behind it, since nothing behind a jump runs.
509        for line in text.lines().collect::<Vec<_>>().windows(2) {
510            if line[1].contains("x64.jmp_reg") {
511                assert!(line[0].contains("x64.mov_rr_64"), "{text}");
512            }
513        }
514        assert_eq!(addressed(&func), vec![4, 4]);
515    }
516
517    /// A function with one computed `goto` that reaches as many labels as the test asks for, each
518    /// given the one value the branch has in hand, which is the shape of a dispatch table.
519    fn table(labels: usize) -> (Interner, mir::Func, Vec<mir::Block>) {
520        let mut names = Interner::new();
521        let mut func = mir::Func::new(names.intern("f"));
522        let head = func.create_block();
523        let arg = func.append_param(head, GPR);
524        let lea = mir::Opcode::new(names.intern("x64.lea_64"));
525        let jump = mir::Opcode::new(names.intern("x64.jmp_reg"));
526        let mut targets = Vec::new();
527        for _ in 0..labels {
528            let label = func.create_block();
529            func.append_param(label, GPR);
530            targets.push(label);
531            func.succs_mut(head).push(mir::BlockCall::with(label, vec![arg]));
532        }
533        let address = func.new_vreg(GPR);
534        func.build(head, lea).def(address, GPR).mem(mir::Mem::block(targets[0])).finish();
535        func.build(head, jump).operand(mir::Operand::read(address, GPR)).finish();
536        (names, func, targets)
537    }
538
539    #[test]
540    fn labels_a_branch_gives_the_same_value_are_given_it_in_one_register() {
541        let (mut names, mut func, _) = table(8);
542        assert_eq!(indirect(&mut func, &BRANCH, &FRAME, &mut names), 8);
543
544        // One move in front of the jump and not eight, because the eight labels are given the one
545        // value and it is now in the one register. Eight blocks were still made, since each label
546        // needs the block that moves that register on to its own parameter.
547        let text = mir::print_func(&func, &names, &REGS);
548        assert_eq!(text.matches("x64.mov_rr_64").count(), 1, "{text}");
549    }
550
551    #[test]
552    fn a_label_given_something_else_keeps_a_register_of_its_own() {
553        let (mut names, mut func, targets) = table(8);
554        let head = mir::Block::new(0);
555        let other = func.append_param(head, GPR);
556        let last = func[head].succs.len() - 1;
557        func.succs_mut(head)[last] = mir::BlockCall::with(targets[7], vec![other]);
558
559        assert_eq!(indirect(&mut func, &BRANCH, &FRAME, &mut names), 8);
560        // Two moves: one register for the seven labels given the same value, and one for the label
561        // given the other. Sharing is about what a label is given and not about how many there are.
562        let text = mir::print_func(&func, &names, &REGS);
563        assert_eq!(text.matches("x64.mov_rr_64").count(), 2, "{text}");
564    }
565
566    #[test]
567    fn labels_that_take_different_numbers_of_values_still_share_the_ones_they_agree_on() {
568        let (mut names, mut func, targets) = table(8);
569        let head = mir::Block::new(0);
570        let arg = func[head].params[0].reg;
571        let other = func.append_param(head, GPR);
572        // The last label takes a second value, which is what an interpreter looks like: each of
573        // its labels uses what it needs and no two of them need quite the same list.
574        func.append_param(targets[7], GPR);
575        let last = func[head].succs.len() - 1;
576        func.succs_mut(head)[last] = mir::BlockCall::with(targets[7], vec![arg, other]);
577
578        assert_eq!(indirect(&mut func, &BRANCH, &FRAME, &mut names), 8);
579        // Two moves, not nine. The first parameter of the long label is given what the other seven
580        // are given, so it takes the same register, and only the value nothing else is given needs
581        // one of its own.
582        let text = mir::print_func(&func, &names, &REGS);
583        assert_eq!(text.matches("x64.mov_rr_64").count(), 2, "{text}");
584    }
585
586    #[test]
587    fn an_edge_out_of_a_computed_goto_that_carries_nothing_is_left_alone() {
588        let (mut names, mut func) = computed(1, 0);
589
590        // No values to carry, so no block to carry them, and the address stays the label's own.
591        assert_eq!(indirect(&mut func, &BRANCH, &FRAME, &mut names), 0);
592        assert_eq!(addressed(&func), vec![1]);
593    }
594
595    #[test]
596    fn a_function_with_no_computed_goto_in_it_is_left_alone() {
597        let (mut names, mut func, _) = diamond(1);
598        assert_eq!(indirect(&mut func, &BRANCH, &FRAME, &mut names), 0);
599        assert_eq!(edges(&func), vec![vec![1, 2], vec![3], vec![3], vec![]]);
600    }
601
602    #[test]
603    fn what_it_leaves_is_nothing_for_the_splitting_below_to_do() {
604        let (mut names, mut func) = computed(2, 1);
605        indirect(&mut func, &BRANCH, &FRAME, &mut names);
606        // The edges out of the branches carry nothing now, and the edges out of the blocks it
607        // added are the only way out of those blocks, so neither kind is critical.
608        assert_eq!(critical(&mut func), 0);
609    }
610
611    /// The first instruction of each block, by opcode, and an empty string for a block with
612    /// nothing in it.
613    fn opens(func: &mir::Func, names: &Interner) -> Vec<String> {
614        func.blocks()
615            .map(|block| match func.insts(block).next() {
616                Some(inst) => names.resolve(func[inst].opcode.name()).to_owned(),
617                None => String::new(),
618            })
619            .collect()
620    }
621
622    #[test]
623    fn the_block_a_label_begins_at_gets_a_landing_pad_when_the_forward_edge_is_checked() {
624        let (mut names, mut func) = computed(2, 1);
625        indirect(&mut func, &BRANCH, &FRAME, &mut names);
626
627        // One pad, at the block in front of the label, because that is the block both addresses
628        // name once the values have been moved on to it. The label's own block is arrived at by an
629        // ordinary edge from there and wants nothing.
630        assert_eq!(pads(&mut func, &FRAME, FRAME.landing, &mut names), 1);
631        assert_eq!(opens(&func, &names), ["", "", "x64.lea_64", "x64.lea_64", "x64.endbr64"]);
632    }
633
634    #[test]
635    fn a_label_with_no_block_in_front_of_it_gets_the_pad_itself() {
636        let (mut names, mut func) = computed(1, 0);
637        indirect(&mut func, &BRANCH, &FRAME, &mut names);
638
639        // Nothing was moved on to anything, so the address still names the label and the pad goes
640        // where the address goes.
641        assert_eq!(pads(&mut func, &FRAME, FRAME.landing, &mut names), 1);
642        assert_eq!(opens(&func, &names), ["x64.lea_64", "x64.endbr64"]);
643    }
644
645    #[test]
646    fn nothing_is_written_when_the_forward_edge_is_not_checked() {
647        let (mut names, mut func) = computed(1, 0);
648        assert_eq!(pads(&mut func, &FRAME, None, &mut names), 0);
649        assert_eq!(opens(&func, &names), ["x64.lea_64", ""]);
650    }
651}