summaryrefslogtreecommitdiffstats
path: root/js/src/jit/LoopUnroller.cpp
blob: 5481d6978042b8056effd77635190e270478f1fb (plain)
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
/* -*- 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/LoopUnroller.h"

#include "jit/MIRGraph.h"

using namespace js;
using namespace js::jit;

using mozilla::ArrayLength;

namespace {

struct LoopUnroller
{
    typedef HashMap<MDefinition*, MDefinition*,
                    PointerHasher<MDefinition*, 2>, SystemAllocPolicy> DefinitionMap;

    explicit LoopUnroller(MIRGraph& graph)
      : graph(graph), alloc(graph.alloc()),
        header(nullptr), backedge(nullptr),
        unrolledHeader(nullptr), unrolledBackedge(nullptr),
        oldPreheader(nullptr), newPreheader(nullptr)
    {}

    MIRGraph& graph;
    TempAllocator& alloc;

    // Header and body of the original loop.
    MBasicBlock* header;
    MBasicBlock* backedge;

    // Header and body of the unrolled loop.
    MBasicBlock* unrolledHeader;
    MBasicBlock* unrolledBackedge;

    // Old and new preheaders. The old preheader starts out associated with the
    // original loop, but becomes the preheader of the new loop. The new
    // preheader will be given to the original loop.
    MBasicBlock* oldPreheader;
    MBasicBlock* newPreheader;

    // Map terms in the original loop to terms in the current unrolled iteration.
    DefinitionMap unrolledDefinitions;

    MDefinition* getReplacementDefinition(MDefinition* def);
    MResumePoint* makeReplacementResumePoint(MBasicBlock* block, MResumePoint* rp);
    bool makeReplacementInstruction(MInstruction* ins);

    void go(LoopIterationBound* bound);
};

} // namespace

MDefinition*
LoopUnroller::getReplacementDefinition(MDefinition* def)
{
    if (def->block()->id() < header->id()) {
        // The definition is loop invariant.
        return def;
    }

    DefinitionMap::Ptr p = unrolledDefinitions.lookup(def);
    if (!p) {
        // After phi analysis (TypeAnalyzer::replaceRedundantPhi) the resume
        // point at the start of a block can contain definitions from within
        // the block itself.
        MOZ_ASSERT(def->isConstant());

        MConstant* constant = MConstant::Copy(alloc, def->toConstant());
        oldPreheader->insertBefore(*oldPreheader->begin(), constant);
        return constant;
    }

    return p->value();
}

bool
LoopUnroller::makeReplacementInstruction(MInstruction* ins)
{
    MDefinitionVector inputs(alloc);
    for (size_t i = 0; i < ins->numOperands(); i++) {
        MDefinition* old = ins->getOperand(i);
        MDefinition* replacement = getReplacementDefinition(old);
        if (!inputs.append(replacement))
            return false;
    }

    MInstruction* clone = ins->clone(alloc, inputs);

    unrolledBackedge->add(clone);

    if (!unrolledDefinitions.putNew(ins, clone))
        return false;

    if (MResumePoint* old = ins->resumePoint()) {
        MResumePoint* rp = makeReplacementResumePoint(unrolledBackedge, old);
        clone->setResumePoint(rp);
    }

    return true;
}

MResumePoint*
LoopUnroller::makeReplacementResumePoint(MBasicBlock* block, MResumePoint* rp)
{
    MDefinitionVector inputs(alloc);
    for (size_t i = 0; i < rp->numOperands(); i++) {
        MDefinition* old = rp->getOperand(i);
        MDefinition* replacement = old->isUnused() ? old : getReplacementDefinition(old);
        if (!inputs.append(replacement))
            return nullptr;
    }

    MResumePoint* clone = MResumePoint::New(alloc, block, rp, inputs);
    if (!clone)
        return nullptr;

    return clone;
}

void
LoopUnroller::go(LoopIterationBound* bound)
{
    // For now we always unroll loops the same number of times.
    static const size_t UnrollCount = 10;

    JitSpew(JitSpew_Unrolling, "Attempting to unroll loop");

    header = bound->header;

    // UCE might have determined this isn't actually a loop.
    if (!header->isLoopHeader())
        return;

    backedge = header->backedge();
    oldPreheader = header->loopPredecessor();

    MOZ_ASSERT(oldPreheader->numSuccessors() == 1);

    // Only unroll loops with two blocks: an initial one ending with the
    // bound's test, and the body ending with the backedge.
    MTest* test = bound->test;
    if (header->lastIns() != test)
        return;
    if (test->ifTrue() == backedge) {
        if (test->ifFalse()->id() <= backedge->id())
            return;
    } else if (test->ifFalse() == backedge) {
        if (test->ifTrue()->id() <= backedge->id())
            return;
    } else {
        return;
    }
    if (backedge->numPredecessors() != 1 || backedge->numSuccessors() != 1)
        return;
    MOZ_ASSERT(backedge->phisEmpty());

    MBasicBlock* bodyBlocks[] = { header, backedge };

    // All instructions in the header and body must be clonable.
    for (size_t i = 0; i < ArrayLength(bodyBlocks); i++) {
        MBasicBlock* block = bodyBlocks[i];
        for (MInstructionIterator iter(block->begin()); iter != block->end(); iter++) {
            MInstruction* ins = *iter;
            if (ins->canClone())
                continue;
            if (ins->isTest() || ins->isGoto() || ins->isInterruptCheck())
                continue;
#ifdef JS_JITSPEW
            JitSpew(JitSpew_Unrolling, "Aborting: can't clone instruction %s", ins->opName());
#endif
            return;
        }
    }

    // Compute the linear inequality we will use for exiting the unrolled loop:
    //
    // iterationBound - iterationCount - UnrollCount >= 0
    //
    LinearSum remainingIterationsInequality(bound->boundSum);
    if (!remainingIterationsInequality.add(bound->currentSum, -1))
        return;
    if (!remainingIterationsInequality.add(-int32_t(UnrollCount)))
        return;

    // Terms in the inequality need to be either loop invariant or phis from
    // the original header.
    for (size_t i = 0; i < remainingIterationsInequality.numTerms(); i++) {
        MDefinition* def = remainingIterationsInequality.term(i).term;
        if (def->isDiscarded())
            return;
        if (def->block()->id() < header->id())
            continue;
        if (def->block() == header && def->isPhi())
            continue;
        return;
    }

    // OK, we've checked everything, now unroll the loop.

    JitSpew(JitSpew_Unrolling, "Unrolling loop");

    // The old preheader will go before the unrolled loop, and the old loop
    // will need a new empty preheader.
    const CompileInfo& info = oldPreheader->info();
    if (header->trackedPc()) {
        unrolledHeader =
            MBasicBlock::New(graph, nullptr, info,
                             oldPreheader, header->trackedSite(), MBasicBlock::LOOP_HEADER);
        unrolledBackedge =
            MBasicBlock::New(graph, nullptr, info,
                             unrolledHeader, backedge->trackedSite(), MBasicBlock::NORMAL);
        newPreheader =
            MBasicBlock::New(graph, nullptr, info,
                             unrolledHeader, oldPreheader->trackedSite(), MBasicBlock::NORMAL);
    } else {
        unrolledHeader = MBasicBlock::New(graph, info, oldPreheader, MBasicBlock::LOOP_HEADER);
        unrolledBackedge = MBasicBlock::New(graph, info, unrolledHeader, MBasicBlock::NORMAL);
        newPreheader = MBasicBlock::New(graph, info, unrolledHeader, MBasicBlock::NORMAL);
    }

    unrolledHeader->discardAllResumePoints();
    unrolledBackedge->discardAllResumePoints();
    newPreheader->discardAllResumePoints();

    // Insert new blocks at their RPO position, and update block ids.
    graph.insertBlockAfter(oldPreheader, unrolledHeader);
    graph.insertBlockAfter(unrolledHeader, unrolledBackedge);
    graph.insertBlockAfter(unrolledBackedge, newPreheader);
    graph.renumberBlocksAfter(oldPreheader);

    // We don't tolerate allocation failure after this point.
    // TODO: This is a bit drastic, is it possible to improve this?
    AutoEnterOOMUnsafeRegion oomUnsafe;

    if (!unrolledDefinitions.init())
        oomUnsafe.crash("LoopUnroller::go");

    // Add phis to the unrolled loop header which correspond to the phis in the
    // original loop header.
    MOZ_ASSERT(header->getPredecessor(0) == oldPreheader);
    for (MPhiIterator iter(header->phisBegin()); iter != header->phisEnd(); iter++) {
        MPhi* old = *iter;
        MOZ_ASSERT(old->numOperands() == 2);
        MPhi* phi = MPhi::New(alloc);
        phi->setResultType(old->type());
        phi->setResultTypeSet(old->resultTypeSet());
        phi->setRange(old->range());

        unrolledHeader->addPhi(phi);

        if (!phi->reserveLength(2))
            oomUnsafe.crash("LoopUnroller::go");

        // Set the first input for the phi for now. We'll set the second after
        // finishing the unroll.
        phi->addInput(old->getOperand(0));

        // The old phi will now take the value produced by the unrolled loop.
        old->replaceOperand(0, phi);

        if (!unrolledDefinitions.putNew(old, phi))
            oomUnsafe.crash("LoopUnroller::go");
    }

    // The loop condition can bail out on e.g. integer overflow, so make a
    // resume point based on the initial resume point of the original header.
    MResumePoint* headerResumePoint = header->entryResumePoint();
    if (headerResumePoint) {
        MResumePoint* rp = makeReplacementResumePoint(unrolledHeader, headerResumePoint);
        if (!rp)
            oomUnsafe.crash("LoopUnroller::makeReplacementResumePoint");
        unrolledHeader->setEntryResumePoint(rp);

        // Perform an interrupt check at the start of the unrolled loop.
        unrolledHeader->add(MInterruptCheck::New(alloc));
    }

    // Generate code for the test in the unrolled loop.
    for (size_t i = 0; i < remainingIterationsInequality.numTerms(); i++) {
        MDefinition* def = remainingIterationsInequality.term(i).term;
        MDefinition* replacement = getReplacementDefinition(def);
        remainingIterationsInequality.replaceTerm(i, replacement);
    }
    MCompare* compare = ConvertLinearInequality(alloc, unrolledHeader, remainingIterationsInequality);
    MTest* unrolledTest = MTest::New(alloc, compare, unrolledBackedge, newPreheader);
    unrolledHeader->end(unrolledTest);

    // Make an entry resume point for the unrolled body. The unrolled header
    // does not have side effects on stack values, even if the original loop
    // header does, so use the same resume point as for the unrolled header.
    if (headerResumePoint) {
        MResumePoint* rp = makeReplacementResumePoint(unrolledBackedge, headerResumePoint);
        if (!rp)
            oomUnsafe.crash("LoopUnroller::makeReplacementResumePoint");
        unrolledBackedge->setEntryResumePoint(rp);
    }

    // Make an entry resume point for the new preheader. There are no
    // instructions which use this but some other stuff wants one to be here.
    if (headerResumePoint) {
        MResumePoint* rp = makeReplacementResumePoint(newPreheader, headerResumePoint);
        if (!rp)
            oomUnsafe.crash("LoopUnroller::makeReplacementResumePoint");
        newPreheader->setEntryResumePoint(rp);
    }

    // Generate the unrolled code.
    MOZ_ASSERT(UnrollCount > 1);
    size_t unrollIndex = 0;
    while (true) {
        // Clone the contents of the original loop into the unrolled loop body.
        for (size_t i = 0; i < ArrayLength(bodyBlocks); i++) {
            MBasicBlock* block = bodyBlocks[i];
            for (MInstructionIterator iter(block->begin()); iter != block->end(); iter++) {
                MInstruction* ins = *iter;
                if (ins->canClone()) {
                    if (!makeReplacementInstruction(*iter))
                        oomUnsafe.crash("LoopUnroller::makeReplacementDefinition");
                } else {
                    // Control instructions are handled separately.
                    MOZ_ASSERT(ins->isTest() || ins->isGoto() || ins->isInterruptCheck());
                }
            }
        }

        // Compute the value of each loop header phi after the execution of
        // this unrolled iteration.
        MDefinitionVector phiValues(alloc);
        MOZ_ASSERT(header->getPredecessor(1) == backedge);
        for (MPhiIterator iter(header->phisBegin()); iter != header->phisEnd(); iter++) {
            MPhi* old = *iter;
            MDefinition* oldInput = old->getOperand(1);
            if (!phiValues.append(getReplacementDefinition(oldInput)))
                oomUnsafe.crash("LoopUnroller::go");
        }

        unrolledDefinitions.clear();

        if (unrollIndex == UnrollCount - 1) {
            // We're at the end of the last unrolled iteration, set the
            // backedge input for the unrolled loop phis.
            size_t phiIndex = 0;
            for (MPhiIterator iter(unrolledHeader->phisBegin()); iter != unrolledHeader->phisEnd(); iter++) {
                MPhi* phi = *iter;
                phi->addInput(phiValues[phiIndex++]);
            }
            MOZ_ASSERT(phiIndex == phiValues.length());
            break;
        }

        // Update the map for the phis in the next iteration.
        size_t phiIndex = 0;
        for (MPhiIterator iter(header->phisBegin()); iter != header->phisEnd(); iter++) {
            MPhi* old = *iter;
            if (!unrolledDefinitions.putNew(old, phiValues[phiIndex++]))
                oomUnsafe.crash("LoopUnroller::go");
        }
        MOZ_ASSERT(phiIndex == phiValues.length());

        unrollIndex++;
    }

    MGoto* backedgeJump = MGoto::New(alloc, unrolledHeader);
    unrolledBackedge->end(backedgeJump);

    // Place the old preheader before the unrolled loop.
    MOZ_ASSERT(oldPreheader->lastIns()->isGoto());
    oldPreheader->discardLastIns();
    oldPreheader->end(MGoto::New(alloc, unrolledHeader));

    // Place the new preheader before the original loop.
    newPreheader->end(MGoto::New(alloc, header));

    // Cleanup the MIR graph.
    if (!unrolledHeader->addPredecessorWithoutPhis(unrolledBackedge))
        oomUnsafe.crash("LoopUnroller::go");
    header->replacePredecessor(oldPreheader, newPreheader);
    oldPreheader->setSuccessorWithPhis(unrolledHeader, 0);
    newPreheader->setSuccessorWithPhis(header, 0);
    unrolledBackedge->setSuccessorWithPhis(unrolledHeader, 1);
}

bool
jit::UnrollLoops(MIRGraph& graph, const LoopIterationBoundVector& bounds)
{
    if (bounds.empty())
        return true;

    for (size_t i = 0; i < bounds.length(); i++) {
        LoopUnroller unroller(graph);
        unroller.go(bounds[i]);
    }

    // The MIR graph is valid, but now has several new blocks. Update or
    // recompute previous analysis information for the remaining optimization
    // passes.
    ClearDominatorTree(graph);
    if (!BuildDominatorTree(graph))
        return false;

    return true;
}