-
Notifications
You must be signed in to change notification settings - Fork 13.3k
/
Copy pathsimplify.rs
420 lines (358 loc) · 15.1 KB
/
simplify.rs
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
//! A number of passes which remove various redundancies in the CFG.
//!
//! The `SimplifyCfg` pass gets rid of unnecessary blocks in the CFG, whereas the `SimplifyLocals`
//! gets rid of all the unnecessary local variable declarations.
//!
//! The `SimplifyLocals` pass is kinda expensive and therefore not very suitable to be run often.
//! Most of the passes should not care or be impacted in meaningful ways due to extra locals
//! either, so running the pass once, right before codegen, should suffice.
//!
//! On the other side of the spectrum, the `SimplifyCfg` pass is considerably cheap to run, thus
//! one should run it after every pass which may modify CFG in significant ways. This pass must
//! also be run before any analysis passes because it removes dead blocks, and some of these can be
//! ill-typed.
//!
//! The cause of this typing issue is typeck allowing most blocks whose end is not reachable have
//! an arbitrary return type, rather than having the usual () return type (as a note, typeck's
//! notion of reachability is in fact slightly weaker than MIR CFG reachability - see #31617). A
//! standard example of the situation is:
//!
//! ```rust
//! fn example() {
//! let _a: char = { return; };
//! }
//! ```
//!
//! Here the block (`{ return; }`) has the return type `char`, rather than `()`, but the MIR we
//! naively generate still contains the `_a = ()` write in the unreachable block "after" the
//! return.
use crate::transform::{MirPass, MirSource};
use rustc_index::bit_set::BitSet;
use rustc_index::vec::{Idx, IndexVec};
use rustc_middle::mir::visit::{MutVisitor, MutatingUseContext, PlaceContext, Visitor};
use rustc_middle::mir::*;
use rustc_middle::ty::{self, TyCtxt};
use std::borrow::Cow;
pub struct SimplifyCfg {
label: String,
}
impl SimplifyCfg {
pub fn new(label: &str) -> Self {
SimplifyCfg { label: format!("SimplifyCfg-{}", label) }
}
}
pub fn simplify_cfg(body: &mut BodyAndCache<'_>) {
CfgSimplifier::new(body).simplify();
remove_dead_blocks(body);
// FIXME: Should probably be moved into some kind of pass manager
body.basic_blocks_mut().raw.shrink_to_fit();
}
impl<'tcx> MirPass<'tcx> for SimplifyCfg {
fn name(&self) -> Cow<'_, str> {
Cow::Borrowed(&self.label)
}
fn run_pass(&self, _tcx: TyCtxt<'tcx>, _src: MirSource<'tcx>, body: &mut BodyAndCache<'tcx>) {
debug!("SimplifyCfg({:?}) - simplifying {:?}", self.label, body);
simplify_cfg(body);
}
}
pub struct CfgSimplifier<'a, 'tcx> {
basic_blocks: &'a mut IndexVec<BasicBlock, BasicBlockData<'tcx>>,
pred_count: IndexVec<BasicBlock, u32>,
}
impl<'a, 'tcx> CfgSimplifier<'a, 'tcx> {
pub fn new(body: &'a mut BodyAndCache<'tcx>) -> Self {
let mut pred_count = IndexVec::from_elem(0u32, body.basic_blocks());
// we can't use mir.predecessors() here because that counts
// dead blocks, which we don't want to.
pred_count[START_BLOCK] = 1;
for (_, data) in traversal::preorder(body) {
if let Some(ref term) = data.terminator {
for &tgt in term.successors() {
pred_count[tgt] += 1;
}
}
}
let basic_blocks = body.basic_blocks_mut();
CfgSimplifier { basic_blocks, pred_count }
}
pub fn simplify(mut self) {
self.strip_nops();
let mut start = START_BLOCK;
// Vec of the blocks that should be merged. We store the indices here, instead of the
// statements itself to avoid moving the (relatively) large statements twice.
// We do not push the statements directly into the target block (`bb`) as that is slower
// due to additional reallocations
let mut merged_blocks = Vec::new();
loop {
let mut changed = false;
self.collapse_goto_chain(&mut start, &mut changed);
for bb in self.basic_blocks.indices() {
if self.pred_count[bb] == 0 {
continue;
}
debug!("simplifying {:?}", bb);
let mut terminator =
self.basic_blocks[bb].terminator.take().expect("invalid terminator state");
for successor in terminator.successors_mut() {
self.collapse_goto_chain(successor, &mut changed);
}
let mut inner_changed = true;
merged_blocks.clear();
while inner_changed {
inner_changed = false;
inner_changed |= self.simplify_branch(&mut terminator);
inner_changed |= self.merge_successor(&mut merged_blocks, &mut terminator);
changed |= inner_changed;
}
let statements_to_merge =
merged_blocks.iter().map(|&i| self.basic_blocks[i].statements.len()).sum();
if statements_to_merge > 0 {
let mut statements = std::mem::take(&mut self.basic_blocks[bb].statements);
statements.reserve(statements_to_merge);
for &from in &merged_blocks {
statements.append(&mut self.basic_blocks[from].statements);
}
self.basic_blocks[bb].statements = statements;
}
self.basic_blocks[bb].terminator = Some(terminator);
changed |= inner_changed;
}
if !changed {
break;
}
}
if start != START_BLOCK {
debug_assert!(self.pred_count[START_BLOCK] == 0);
self.basic_blocks.swap(START_BLOCK, start);
self.pred_count.swap(START_BLOCK, start);
// pred_count == 1 if the start block has no predecessor _blocks_.
if self.pred_count[START_BLOCK] > 1 {
for (bb, data) in self.basic_blocks.iter_enumerated_mut() {
if self.pred_count[bb] == 0 {
continue;
}
for target in data.terminator_mut().successors_mut() {
if *target == start {
*target = START_BLOCK;
}
}
}
}
}
}
// Collapse a goto chain starting from `start`
fn collapse_goto_chain(&mut self, start: &mut BasicBlock, changed: &mut bool) {
let mut terminator = match self.basic_blocks[*start] {
BasicBlockData {
ref statements,
terminator:
ref mut terminator @ Some(Terminator { kind: TerminatorKind::Goto { .. }, .. }),
..
} if statements.is_empty() => terminator.take(),
// if `terminator` is None, this means we are in a loop. In that
// case, let all the loop collapse to its entry.
_ => return,
};
let target = match terminator {
Some(Terminator { kind: TerminatorKind::Goto { ref mut target }, .. }) => {
self.collapse_goto_chain(target, changed);
*target
}
_ => unreachable!(),
};
self.basic_blocks[*start].terminator = terminator;
debug!("collapsing goto chain from {:?} to {:?}", *start, target);
*changed |= *start != target;
if self.pred_count[*start] == 1 {
// This is the last reference to *start, so the pred-count to
// to target is moved into the current block.
self.pred_count[*start] = 0;
} else {
self.pred_count[target] += 1;
self.pred_count[*start] -= 1;
}
*start = target;
}
// merge a block with 1 `goto` predecessor to its parent
fn merge_successor(
&mut self,
merged_blocks: &mut Vec<BasicBlock>,
terminator: &mut Terminator<'tcx>,
) -> bool {
let target = match terminator.kind {
TerminatorKind::Goto { target } if self.pred_count[target] == 1 => target,
_ => return false,
};
debug!("merging block {:?} into {:?}", target, terminator);
*terminator = match self.basic_blocks[target].terminator.take() {
Some(terminator) => terminator,
None => {
// unreachable loop - this should not be possible, as we
// don't strand blocks, but handle it correctly.
return false;
}
};
merged_blocks.push(target);
self.pred_count[target] = 0;
true
}
// turn a branch with all successors identical to a goto
fn simplify_branch(&mut self, terminator: &mut Terminator<'tcx>) -> bool {
match terminator.kind {
TerminatorKind::SwitchInt { .. } => {}
_ => return false,
};
let first_succ = {
if let Some(&first_succ) = terminator.successors().next() {
if terminator.successors().all(|s| *s == first_succ) {
let count = terminator.successors().count();
self.pred_count[first_succ] -= (count - 1) as u32;
first_succ
} else {
return false;
}
} else {
return false;
}
};
debug!("simplifying branch {:?}", terminator);
terminator.kind = TerminatorKind::Goto { target: first_succ };
true
}
fn strip_nops(&mut self) {
for blk in self.basic_blocks.iter_mut() {
blk.statements
.retain(|stmt| if let StatementKind::Nop = stmt.kind { false } else { true })
}
}
}
pub fn remove_dead_blocks(body: &mut BodyAndCache<'_>) {
let mut seen = BitSet::new_empty(body.basic_blocks().len());
for (bb, _) in traversal::preorder(body) {
seen.insert(bb.index());
}
let basic_blocks = body.basic_blocks_mut();
let num_blocks = basic_blocks.len();
let mut replacements: Vec<_> = (0..num_blocks).map(BasicBlock::new).collect();
let mut used_blocks = 0;
for alive_index in seen.iter() {
replacements[alive_index] = BasicBlock::new(used_blocks);
if alive_index != used_blocks {
// Swap the next alive block data with the current available slot. Since
// alive_index is non-decreasing this is a valid operation.
basic_blocks.raw.swap(alive_index, used_blocks);
}
used_blocks += 1;
}
basic_blocks.raw.truncate(used_blocks);
for block in basic_blocks {
for target in block.terminator_mut().successors_mut() {
*target = replacements[target.index()];
}
}
}
pub struct SimplifyLocals;
impl<'tcx> MirPass<'tcx> for SimplifyLocals {
fn run_pass(&self, tcx: TyCtxt<'tcx>, source: MirSource<'tcx>, body: &mut BodyAndCache<'tcx>) {
trace!("running SimplifyLocals on {:?}", source);
let locals = {
let read_only_cache = read_only!(body);
let mut marker = DeclMarker { locals: BitSet::new_empty(body.local_decls.len()), body };
marker.visit_body(&read_only_cache);
// Return pointer and arguments are always live
marker.locals.insert(RETURN_PLACE);
for arg in body.args_iter() {
marker.locals.insert(arg);
}
marker.locals
};
let map = make_local_map(&mut body.local_decls, locals);
// Update references to all vars and tmps now
LocalUpdater { map, tcx }.visit_body(body);
body.local_decls.shrink_to_fit();
}
}
/// Construct the mapping while swapping out unused stuff out from the `vec`.
fn make_local_map<V>(
vec: &mut IndexVec<Local, V>,
mask: BitSet<Local>,
) -> IndexVec<Local, Option<Local>> {
let mut map: IndexVec<Local, Option<Local>> = IndexVec::from_elem(None, &*vec);
let mut used = Local::new(0);
for alive_index in mask.iter() {
map[alive_index] = Some(used);
if alive_index != used {
vec.swap(alive_index, used);
}
used.increment_by(1);
}
vec.truncate(used.index());
map
}
struct DeclMarker<'a, 'tcx> {
pub locals: BitSet<Local>,
pub body: &'a Body<'tcx>,
}
impl<'a, 'tcx> Visitor<'tcx> for DeclMarker<'a, 'tcx> {
fn visit_local(&mut self, local: &Local, ctx: PlaceContext, location: Location) {
// Ignore storage markers altogether, they get removed along with their otherwise unused
// decls.
// FIXME: Extend this to all non-uses.
if ctx.is_storage_marker() {
return;
}
// Ignore stores of constants because `ConstProp` and `CopyProp` can remove uses of many
// of these locals. However, if the local is still needed, then it will be referenced in
// another place and we'll mark it as being used there.
if ctx == PlaceContext::MutatingUse(MutatingUseContext::Store)
|| ctx == PlaceContext::MutatingUse(MutatingUseContext::Projection)
{
let block = &self.body.basic_blocks()[location.block];
if location.statement_index != block.statements.len() {
let stmt = &block.statements[location.statement_index];
if let StatementKind::Assign(box (dest, rvalue)) = &stmt.kind {
if !dest.is_indirect() && dest.local == *local {
if let Rvalue::Use(Operand::Constant(c)) = rvalue {
match c.literal.val {
// Keep assignments from unevaluated constants around, since the
// evaluation may report errors, even if the use of the constant
// is dead code.
ty::ConstKind::Unevaluated(..) => {}
_ => {
trace!("skipping store of const value {:?} to {:?}", c, dest);
return;
}
}
} else if let Rvalue::Discriminant(d) = rvalue {
trace!("skipping store of discriminant value {:?} to {:?}", d, dest);
return;
}
}
}
}
}
self.locals.insert(*local);
}
}
struct LocalUpdater<'tcx> {
map: IndexVec<Local, Option<Local>>,
tcx: TyCtxt<'tcx>,
}
impl<'tcx> MutVisitor<'tcx> for LocalUpdater<'tcx> {
fn tcx(&self) -> TyCtxt<'tcx> {
self.tcx
}
fn visit_basic_block_data(&mut self, block: BasicBlock, data: &mut BasicBlockData<'tcx>) {
// Remove unnecessary StorageLive and StorageDead annotations.
data.statements.retain(|stmt| match &stmt.kind {
StatementKind::StorageLive(l) | StatementKind::StorageDead(l) => self.map[*l].is_some(),
StatementKind::Assign(box (place, _)) => self.map[place.local].is_some(),
_ => true,
});
self.super_basic_block_data(block, data);
}
fn visit_local(&mut self, l: &mut Local, _: PlaceContext, _: Location) {
*l = self.map[*l].unwrap();
}
}