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
|
/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 4 -*-
* vim: set ts=8 sts=4 et sw=4 tw=99:
* This Source Code Form is subject to the terms of the Mozilla Public
* License, v. 2.0. If a copy of the MPL was not distributed with this
* file, You can obtain one at http://mozilla.org/MPL/2.0/. */
#include "jit/MoveResolver.h"
#include "mozilla/Attributes.h"
#include "jit/MacroAssembler.h"
#include "jit/RegisterSets.h"
using namespace js;
using namespace js::jit;
MoveOperand::MoveOperand(MacroAssembler& masm, const ABIArg& arg)
{
switch (arg.kind()) {
case ABIArg::GPR:
kind_ = REG;
code_ = arg.gpr().code();
break;
#ifdef JS_CODEGEN_REGISTER_PAIR
case ABIArg::GPR_PAIR:
kind_ = REG_PAIR;
code_ = arg.evenGpr().code();
MOZ_ASSERT(code_ % 2 == 0);
MOZ_ASSERT(code_ + 1 == arg.oddGpr().code());
break;
#endif
case ABIArg::FPU:
kind_ = FLOAT_REG;
code_ = arg.fpu().code();
break;
case ABIArg::Stack:
kind_ = MEMORY;
code_ = masm.getStackPointer().code();
disp_ = arg.offsetFromArgBase();
break;
}
}
MoveResolver::MoveResolver()
: numCycles_(0), curCycles_(0)
{
}
void
MoveResolver::resetState()
{
numCycles_ = 0;
curCycles_ = 0;
}
bool
MoveResolver::addMove(const MoveOperand& from, const MoveOperand& to, MoveOp::Type type)
{
// Assert that we're not doing no-op moves.
MOZ_ASSERT(!(from == to));
PendingMove* pm = movePool_.allocate();
if (!pm)
return false;
new (pm) PendingMove(from, to, type);
pending_.pushBack(pm);
return true;
}
// Given move (A -> B), this function attempts to find any move (B -> *) in the
// pending move list, and returns the first one.
MoveResolver::PendingMove*
MoveResolver::findBlockingMove(const PendingMove* last)
{
for (PendingMoveIterator iter = pending_.begin(); iter != pending_.end(); iter++) {
PendingMove* other = *iter;
if (other->from().aliases(last->to())) {
// We now have pairs in the form (A -> X) (X -> y). The second pair
// blocks the move in the first pair, so return it.
return other;
}
}
// No blocking moves found.
return nullptr;
}
// Given move (A -> B), this function attempts to find any move (B -> *) in the
// move list iterator, and returns the first one.
// N.B. It is unclear if a single move can complete more than one cycle, so to be
// conservative, this function operates on iterators, so the caller can process all
// instructions that start a cycle.
MoveResolver::PendingMove*
MoveResolver::findCycledMove(PendingMoveIterator* iter, PendingMoveIterator end, const PendingMove* last)
{
for (; *iter != end; (*iter)++) {
PendingMove* other = **iter;
if (other->from().aliases(last->to())) {
// We now have pairs in the form (A -> X) (X -> y). The second pair
// blocks the move in the first pair, so return it.
(*iter)++;
return other;
}
}
// No blocking moves found.
return nullptr;
}
#ifdef JS_CODEGEN_ARM
static inline bool
MoveIsDouble(const MoveOperand& move)
{
if (!move.isFloatReg())
return false;
return move.floatReg().isDouble();
}
#endif
#ifdef JS_CODEGEN_ARM
static inline bool
MoveIsSingle(const MoveOperand& move)
{
if (!move.isFloatReg())
return false;
return move.floatReg().isSingle();
}
#endif
#ifdef JS_CODEGEN_ARM
bool
MoveResolver::isDoubleAliasedAsSingle(const MoveOperand& move)
{
if (!MoveIsDouble(move))
return false;
for (auto iter = pending_.begin(); iter != pending_.end(); ++iter) {
PendingMove* other = *iter;
if (other->from().aliases(move) && MoveIsSingle(other->from()))
return true;
if (other->to().aliases(move) && MoveIsSingle(other->to()))
return true;
}
return false;
}
#endif
#ifdef JS_CODEGEN_ARM
static MoveOperand
SplitIntoLowerHalf(const MoveOperand& move)
{
if (MoveIsDouble(move)) {
FloatRegister lowerSingle = move.floatReg().asSingle();
return MoveOperand(lowerSingle);
}
MOZ_ASSERT(move.isMemoryOrEffectiveAddress());
return move;
}
#endif
#ifdef JS_CODEGEN_ARM
static MoveOperand
SplitIntoUpperHalf(const MoveOperand& move)
{
if (MoveIsDouble(move)) {
FloatRegister lowerSingle = move.floatReg().asSingle();
FloatRegister upperSingle = VFPRegister(lowerSingle.code() + 1, VFPRegister::Single);
return MoveOperand(upperSingle);
}
MOZ_ASSERT(move.isMemoryOrEffectiveAddress());
return MoveOperand(move.base(), move.disp() + sizeof(float));
}
#endif
bool
MoveResolver::resolve()
{
resetState();
orderedMoves_.clear();
#ifdef JS_CODEGEN_ARM
// Some of ARM's double registers alias two of its single registers,
// but the algorithm below assumes that every register can participate
// in at most one cycle. To satisfy the algorithm, any double registers
// that may conflict are split into their single-register halves.
//
// This logic is only applicable because ARM only uses registers d0-d15,
// all of which alias s0-s31. Double registers d16-d31 are unused.
// Therefore there is never a double move that cannot be split.
// If this changes in the future, the algorithm will have to be fixed.
for (auto iter = pending_.begin(); iter != pending_.end(); ++iter) {
PendingMove* pm = *iter;
if (isDoubleAliasedAsSingle(pm->from()) || isDoubleAliasedAsSingle(pm->to())) {
PendingMove* lower = movePool_.allocate();
if (!lower)
return false;
// Insert the new node before the current position to not affect iteration.
MoveOperand fromLower = SplitIntoLowerHalf(pm->from());
MoveOperand toLower = SplitIntoLowerHalf(pm->to());
new (lower) PendingMove(fromLower, toLower, MoveOp::FLOAT32);
pending_.insertBefore(pm, lower);
// Overwrite pm in place for the upper move. Iteration proceeds as normal.
MoveOperand fromUpper = SplitIntoUpperHalf(pm->from());
MoveOperand toUpper = SplitIntoUpperHalf(pm->to());
pm->overwrite(fromUpper, toUpper, MoveOp::FLOAT32);
}
}
#endif
InlineList<PendingMove> stack;
// This is a depth-first-search without recursion, which tries to find
// cycles in a list of moves.
//
// Algorithm.
//
// S = Traversal stack.
// P = Pending move list.
// O = Ordered list of moves.
//
// As long as there are pending moves in P:
// Let |root| be any pending move removed from P
// Add |root| to the traversal stack.
// As long as S is not empty:
// Let |L| be the most recent move added to S.
//
// Find any pending move M whose source is L's destination, thus
// preventing L's move until M has completed.
//
// If a move M was found,
// Remove M from the pending list.
// If M's destination is |root|,
// Annotate M and |root| as cycles.
// Add M to S.
// do not Add M to O, since M may have other conflictors in P that have not yet been processed.
// Otherwise,
// Add M to S.
// Otherwise,
// Remove L from S.
// Add L to O.
//
while (!pending_.empty()) {
PendingMove* pm = pending_.popBack();
// Add this pending move to the cycle detection stack.
stack.pushBack(pm);
while (!stack.empty()) {
PendingMove* blocking = findBlockingMove(stack.peekBack());
if (blocking) {
PendingMoveIterator stackiter = stack.begin();
PendingMove* cycled = findCycledMove(&stackiter, stack.end(), blocking);
if (cycled) {
// Find the cycle's start.
// We annotate cycles at each move in the cycle, and
// assert that we do not find two cycles in one move chain
// traversal (which would indicate two moves to the same
// destination).
// Since there can be more than one cycle, find them all.
do {
cycled->setCycleEnd(curCycles_);
cycled = findCycledMove(&stackiter, stack.end(), blocking);
} while (cycled);
blocking->setCycleBegin(pm->type(), curCycles_);
curCycles_++;
pending_.remove(blocking);
stack.pushBack(blocking);
} else {
// This is a new link in the move chain, so keep
// searching for a cycle.
pending_.remove(blocking);
stack.pushBack(blocking);
}
} else {
// Otherwise, pop the last move on the search stack because it's
// complete and not participating in a cycle. The resulting
// move can safely be added to the ordered move list.
PendingMove* done = stack.popBack();
if (!addOrderedMove(*done))
return false;
movePool_.free(done);
}
}
// If the current queue is empty, it is certain that there are
// all previous cycles cannot conflict with future cycles,
// so re-set the counter of pending cycles, while keeping a high-water mark.
if (numCycles_ < curCycles_)
numCycles_ = curCycles_;
curCycles_ = 0;
}
return true;
}
bool
MoveResolver::addOrderedMove(const MoveOp& move)
{
// Sometimes the register allocator generates move groups where multiple
// moves have the same source. Try to optimize these cases when the source
// is in memory and the target of one of the moves is in a register.
MOZ_ASSERT(!move.from().aliases(move.to()));
if (!move.from().isMemory() || move.isCycleBegin() || move.isCycleEnd())
return orderedMoves_.append(move);
// Look for an earlier move with the same source, where no intervening move
// touches either the source or destination of the new move.
for (int i = orderedMoves_.length() - 1; i >= 0; i--) {
const MoveOp& existing = orderedMoves_[i];
if (existing.from() == move.from() &&
!existing.to().aliases(move.to()) &&
existing.type() == move.type() &&
!existing.isCycleBegin() &&
!existing.isCycleEnd())
{
MoveOp* after = orderedMoves_.begin() + i + 1;
if (existing.to().isGeneralReg() || existing.to().isFloatReg()) {
MoveOp nmove(existing.to(), move.to(), move.type());
return orderedMoves_.insert(after, nmove);
} else if (move.to().isGeneralReg() || move.to().isFloatReg()) {
MoveOp nmove(move.to(), existing.to(), move.type());
orderedMoves_[i] = move;
return orderedMoves_.insert(after, nmove);
}
}
if (existing.aliases(move))
break;
}
return orderedMoves_.append(move);
}
void
MoveResolver::reorderMove(size_t from, size_t to)
{
MOZ_ASSERT(from != to);
MoveOp op = orderedMoves_[from];
if (from < to) {
for (size_t i = from; i < to; i++)
orderedMoves_[i] = orderedMoves_[i + 1];
} else {
for (size_t i = from; i > to; i--)
orderedMoves_[i] = orderedMoves_[i - 1];
}
orderedMoves_[to] = op;
}
void
MoveResolver::sortMemoryToMemoryMoves()
{
// Try to reorder memory->memory moves so that they are executed right
// before a move that clobbers some register. This will allow the move
// emitter to use that clobbered register as a scratch register for the
// memory->memory move, if necessary.
for (size_t i = 0; i < orderedMoves_.length(); i++) {
const MoveOp& base = orderedMoves_[i];
if (!base.from().isMemory() || !base.to().isMemory())
continue;
if (base.type() != MoveOp::GENERAL && base.type() != MoveOp::INT32)
continue;
// Look for an earlier move clobbering a register.
bool found = false;
for (int j = i - 1; j >= 0; j--) {
const MoveOp& previous = orderedMoves_[j];
if (previous.aliases(base) || previous.isCycleBegin() || previous.isCycleEnd())
break;
if (previous.to().isGeneralReg()) {
reorderMove(i, j);
found = true;
break;
}
}
if (found)
continue;
// Look for a later move clobbering a register.
if (i + 1 < orderedMoves_.length()) {
bool found = false, skippedRegisterUse = false;
for (size_t j = i + 1; j < orderedMoves_.length(); j++) {
const MoveOp& later = orderedMoves_[j];
if (later.aliases(base) || later.isCycleBegin() || later.isCycleEnd())
break;
if (later.to().isGeneralReg()) {
if (skippedRegisterUse) {
reorderMove(i, j);
found = true;
} else {
// There is no move that uses a register between the
// original memory->memory move and this move that
// clobbers a register. The move should already be able
// to use a scratch register, so don't shift anything
// around.
}
break;
}
if (later.from().isGeneralReg())
skippedRegisterUse = true;
}
if (found) {
// Redo the search for memory->memory moves at the current
// index, so we don't skip the move just shifted back.
i--;
}
}
}
}
|