void MBasicBlock::inheritPhis(MBasicBlock* header) { MResumePoint* headerRp = header->entryResumePoint(); size_t stackDepth = headerRp->stackDepth(); for (size_t slot = 0; slot < stackDepth; slot++) { MDefinition* exitDef = getSlot(slot); MDefinition* loopDef = headerRp->getOperand(slot); if (loopDef->block() != header) { MOZ_ASSERT(loopDef->block()->id() < header->id()); MOZ_ASSERT(loopDef == exitDef); continue; } // Phis are allocated by NewPendingLoopHeader. MPhi* phi = loopDef->toPhi(); MOZ_ASSERT(phi->numOperands() == 2); // The entry definition is always the leftmost input to the phi. MDefinition* entryDef = phi->getOperand(0); if (entryDef != exitDef) continue; // If the entryDef is the same as exitDef, then we must propagate the // phi down to this successor. This chance was missed as part of // setBackedge() because exits are not captured in resume points. setSlot(slot, phi); } }
bool MBasicBlock::inheritPhisFromBackedge(MBasicBlock* backedge, bool* hadTypeChange) { // We must be a pending loop header MOZ_ASSERT(kind_ == PENDING_LOOP_HEADER); size_t stackDepth = entryResumePoint()->stackDepth(); for (size_t slot = 0; slot < stackDepth; slot++) { // Get the value stack-slot of the back edge. MDefinition* exitDef = backedge->getSlot(slot); // Get the value of the loop header. MDefinition* loopDef = entryResumePoint()->getOperand(slot); if (loopDef->block() != this) { // If we are finishing a pending loop header, then we need to ensure // that all operands are phis. This is usualy the case, except for // object/arrays build with generators, in which case we share the // same allocations across all blocks. MOZ_ASSERT(loopDef->block()->id() < id()); MOZ_ASSERT(loopDef == exitDef); continue; } // Phis are allocated by NewPendingLoopHeader. MPhi* entryDef = loopDef->toPhi(); MOZ_ASSERT(entryDef->block() == this); if (entryDef == exitDef) { // If the exit def is the same as the entry def, make a redundant // phi. Since loop headers have exactly two incoming edges, we // know that that's just the first input. // // Note that we eliminate later rather than now, to avoid any // weirdness around pending continue edges which might still hold // onto phis. exitDef = entryDef->getOperand(0); } bool typeChange = false; if (!entryDef->addInputSlow(exitDef)) return false; if (!entryDef->checkForTypeChange(exitDef, &typeChange)) return false; *hadTypeChange |= typeChange; setSlot(slot, entryDef); } return true; }
// If an equivalent and dominating value already exists in the set, return it. // Otherwise insert |def| into the set and return it. MDefinition* ValueNumberer::leader(MDefinition* def) { // If the value isn't suitable for eliminating, don't bother hashing it. The // convention is that congruentTo returns false for node kinds that wish to // opt out of redundance elimination. // TODO: It'd be nice to clean up that convention (bug 1031406). if (!def->isEffectful() && def->congruentTo(def)) { // Look for a match. VisibleValues::AddPtr p = values_.findLeaderForAdd(def); if (p) { MDefinition* rep = *p; if (!rep->isDiscarded() && rep->block()->dominates(def->block())) { // We found a dominating congruent value. return rep; } // The congruent value doesn't dominate. It never will again in this // dominator tree, so overwrite it. values_.overwrite(p, def); } else { // No match. Add a new entry. if (!values_.add(p, def)) return nullptr; } #ifdef DEBUG JitSpew(JitSpew_GVN, " Recording %s%u", def->opName(), def->id()); #endif } return def; }
MDefinition * ValueNumberer::simplify(MDefinition *def, bool useValueNumbers) { if (def->isEffectful()) return def; MDefinition *ins = def->foldsTo(useValueNumbers); if (ins == def || !ins->updateForFolding(def)) return def; // ensure this instruction has a VN if (!ins->valueNumberData()) ins->setValueNumberData(new ValueNumberData); if (!ins->block()) { // In this case, we made a new def by constant folding, for // example, we replaced add(#3,#4) with a new const(#7) node. // We will only fold a phi into one of its operands. JS_ASSERT(!def->isPhi()); def->block()->insertAfter(def->toInstruction(), ins->toInstruction()); ins->setValueNumber(lookupValue(ins)); } JS_ASSERT(ins->id() != 0); def->replaceAllUsesWith(ins); IonSpew(IonSpew_GVN, "Folding %d to be %d", def->id(), ins->id()); return ins; }
void TypeAnalyzer::adjustPhiInputs(MPhi *phi) { MIRType phiType = phi->type(); if (phiType == MIRType_Double) { // Convert int32 operands to double. for (size_t i = 0; i < phi->numOperands(); i++) { MDefinition *in = phi->getOperand(i); if (in->type() == MIRType_Int32) { MToDouble *toDouble = MToDouble::New(in); in->block()->insertBefore(in->block()->lastIns(), toDouble); phi->replaceOperand(i, toDouble); } else { JS_ASSERT(in->type() == MIRType_Double); } } return; } if (phiType != MIRType_Value) return; // Box every typed input. for (size_t i = 0; i < phi->numOperands(); i++) { MDefinition *in = phi->getOperand(i); if (in->type() == MIRType_Value) continue; if (in->isUnbox()) { // The input is being explicitly unboxed, so sneak past and grab // the original box. phi->replaceOperand(i, in->toUnbox()->input()); } else { MBox *box = MBox::New(in); in->block()->insertBefore(in->block()->lastIns(), box); phi->replaceOperand(i, box); } } }
bool MBasicBlock::addPredecessorPopN(TempAllocator& alloc, MBasicBlock* pred, uint32_t popped) { MOZ_ASSERT(pred); MOZ_ASSERT(predecessors_.length() > 0); // Predecessors must be finished, and at the correct stack depth. MOZ_ASSERT(pred->hasLastIns()); MOZ_ASSERT(pred->stackPosition_ == stackPosition_ + popped); for (uint32_t i = 0, e = stackPosition_; i < e; ++i) { MDefinition* mine = getSlot(i); MDefinition* other = pred->getSlot(i); if (mine != other) { // If the current instruction is a phi, and it was created in this // basic block, then we have already placed this phi and should // instead append to its operands. if (mine->isPhi() && mine->block() == this) { MOZ_ASSERT(predecessors_.length()); if (!mine->toPhi()->addInputSlow(other)) return false; } else { // Otherwise, create a new phi node. MPhi* phi; if (mine->type() == other->type()) phi = MPhi::New(alloc.fallible(), mine->type()); else phi = MPhi::New(alloc.fallible()); if (!phi) return false; addPhi(phi); // Prime the phi for each predecessor, so input(x) comes from // predecessor(x). if (!phi->reserveLength(predecessors_.length() + 1)) return false; for (size_t j = 0, numPreds = predecessors_.length(); j < numPreds; ++j) { MOZ_ASSERT(predecessors_[j]->getSlot(i) == mine); phi->addInput(mine); } phi->addInput(other); setSlot(i, phi); if (entryResumePoint()) entryResumePoint()->replaceOperand(i, phi); } } } return predecessors_.append(pred); }
static void UnboxSimdPhi(const JitCompartment* jitCompartment, MIRGraph& graph, MPhi* phi, SimdType unboxType) { TempAllocator& alloc = graph.alloc(); // Unbox and replace all operands. for (size_t i = 0, e = phi->numOperands(); i < e; i++) { MDefinition* op = phi->getOperand(i); MSimdUnbox* unbox = MSimdUnbox::New(alloc, op, unboxType); op->block()->insertAtEnd(unbox); phi->replaceOperand(i, unbox); } // Change the MIRType of the Phi. MIRType mirType = SimdTypeToMIRType(unboxType); phi->setResultType(mirType); MBasicBlock* phiBlock = phi->block(); MInstruction* atRecover = phiBlock->safeInsertTop(nullptr, MBasicBlock::IgnoreRecover); MInstruction* at = phiBlock->safeInsertTop(atRecover); // Note, we capture the uses-list now, as new instructions are not visited. MUseIterator i(phi->usesBegin()), e(phi->usesEnd()); // Add a MSimdBox, and replace all the Phi uses with it. JSObject* templateObject = jitCompartment->maybeGetSimdTemplateObjectFor(unboxType); InlineTypedObject* inlineTypedObject = &templateObject->as<InlineTypedObject>(); MSimdBox* recoverBox = MSimdBox::New(alloc, nullptr, phi, inlineTypedObject, unboxType, gc::DefaultHeap); recoverBox->setRecoveredOnBailout(); phiBlock->insertBefore(atRecover, recoverBox); MSimdBox* box = nullptr; while (i != e) { MUse* use = *i++; MNode* ins = use->consumer(); if ((ins->isDefinition() && ins->toDefinition()->isRecoveredOnBailout()) || (ins->isResumePoint() && ins->toResumePoint()->isRecoverableOperand(use))) { use->replaceProducer(recoverBox); continue; } if (!box) { box = MSimdBox::New(alloc, nullptr, phi, inlineTypedObject, unboxType, gc::DefaultHeap); phiBlock->insertBefore(at, box); } use->replaceProducer(box); } }
bool MBasicBlock::addPredecessorPopN(MBasicBlock *pred, uint32_t popped) { JS_ASSERT(pred); JS_ASSERT(predecessors_.length() > 0); // Predecessors must be finished, and at the correct stack depth. JS_ASSERT(pred->lastIns_); JS_ASSERT(pred->stackPosition_ == stackPosition_ + popped); for (uint32_t i = 0; i < stackPosition_; i++) { MDefinition *mine = getSlot(i); MDefinition *other = pred->getSlot(i); if (mine != other) { // If the current instruction is a phi, and it was created in this // basic block, then we have already placed this phi and should // instead append to its operands. if (mine->isPhi() && mine->block() == this) { JS_ASSERT(predecessors_.length()); if (!mine->toPhi()->addInputSlow(other)) return false; } else { // Otherwise, create a new phi node. MPhi *phi = MPhi::New(i); addPhi(phi); // Prime the phi for each predecessor, so input(x) comes from // predecessor(x). if (!phi->initLength(predecessors_.length() + 1)) return false; for (size_t j = 0; j < predecessors_.length(); j++) { JS_ASSERT(predecessors_[j]->getSlot(i) == mine); phi->setOperand(j, mine); } phi->setOperand(predecessors_.length(), other); setSlot(i, phi); if (entryResumePoint()) entryResumePoint()->replaceOperand(i, phi); } } } return predecessors_.append(pred); }
// Visit |def|. bool ValueNumberer::visitDefinition(MDefinition* def) { // Nop does not fit in any of the previous optimization, as its only purpose // is to reduce the register pressure by keeping additional resume // point. Still, there is no need consecutive list of MNop instructions, and // this will slow down every other iteration on the Graph. if (def->isNop()) { MNop* nop = def->toNop(); MBasicBlock* block = nop->block(); // We look backward to know if we can remove the previous Nop, we do not // look forward as we would not benefit from the folding made by GVN. MInstructionReverseIterator iter = ++block->rbegin(nop); // This nop is at the beginning of the basic block, just replace the // resume point of the basic block by the one from the resume point. if (iter == block->rend()) { JitSpew(JitSpew_GVN, " Removing Nop%u", nop->id()); nop->moveResumePointAsEntry(); block->discard(nop); return true; } // The previous instruction is also a Nop, no need to keep it anymore. MInstruction* prev = *iter; if (prev->isNop()) { JitSpew(JitSpew_GVN, " Removing Nop%u", prev->id()); block->discard(prev); return true; } return true; } // If this instruction has a dependency() into an unreachable block, we'll // need to update AliasAnalysis. MInstruction* dep = def->dependency(); if (dep != nullptr && (dep->isDiscarded() || dep->block()->isDead())) { JitSpew(JitSpew_GVN, " AliasAnalysis invalidated"); if (updateAliasAnalysis_ && !dependenciesBroken_) { // TODO: Recomputing alias-analysis could theoretically expose more // GVN opportunities. JitSpew(JitSpew_GVN, " Will recompute!"); dependenciesBroken_ = true; } // Temporarily clear its dependency, to protect foldsTo, which may // wish to use the dependency to do store-to-load forwarding. def->setDependency(def->toInstruction()); } else { dep = nullptr; } // Look for a simplified form of |def|. MDefinition* sim = simplified(def); if (sim != def) { if (sim == nullptr) return false; // If |sim| doesn't belong to a block, insert it next to |def|. if (sim->block() == nullptr) def->block()->insertAfter(def->toInstruction(), sim->toInstruction()); #ifdef DEBUG JitSpew(JitSpew_GVN, " Folded %s%u to %s%u", def->opName(), def->id(), sim->opName(), sim->id()); #endif MOZ_ASSERT(!sim->isDiscarded()); ReplaceAllUsesWith(def, sim); // The node's foldsTo said |def| can be replaced by |rep|. If |def| is a // guard, then either |rep| is also a guard, or a guard isn't actually // needed, so we can clear |def|'s guard flag and let it be discarded. def->setNotGuardUnchecked(); if (DeadIfUnused(def)) { if (!discardDefsRecursively(def)) return false; // If that ended up discarding |sim|, then we're done here. if (sim->isDiscarded()) return true; } // Otherwise, procede to optimize with |sim| in place of |def|. def = sim; } // Now that foldsTo is done, re-enable the original dependency. Even though // it may be pointing into a discarded block, it's still valid for the // purposes of detecting congruent loads. if (dep != nullptr) def->setDependency(dep); // Look for a dominating def which makes |def| redundant. MDefinition* rep = leader(def); if (rep != def) { if (rep == nullptr) return false; if (rep->updateForReplacement(def)) { #ifdef DEBUG JitSpew(JitSpew_GVN, " Replacing %s%u with %s%u", def->opName(), def->id(), rep->opName(), rep->id()); #endif ReplaceAllUsesWith(def, rep); // The node's congruentTo said |def| is congruent to |rep|, and it's // dominated by |rep|. If |def| is a guard, it's covered by |rep|, // so we can clear |def|'s guard flag and let it be discarded. def->setNotGuardUnchecked(); if (DeadIfUnused(def)) { // discardDef should not add anything to the deadDefs, as the // redundant operation should have the same input operands. mozilla::DebugOnly<bool> r = discardDef(def); MOZ_ASSERT(r, "discardDef shouldn't have tried to add anything to the worklist, " "so it shouldn't have failed"); MOZ_ASSERT(deadDefs_.empty(), "discardDef shouldn't have added anything to the worklist"); } def = rep; } } return true; }
void RangeAnalysis::analyzeLoopPhi(MBasicBlock *header, LoopIterationBound *loopBound, MPhi *phi) { // Given a bound on the number of backedges taken, compute an upper and // lower bound for a phi node that may change by a constant amount each // iteration. Unlike for the case when computing the iteration bound // itself, the phi does not need to change the same amount every iteration, // but is required to change at most N and be either nondecreasing or // nonincreasing. if (phi->numOperands() != 2) return; MBasicBlock *preLoop = header->loopPredecessor(); JS_ASSERT(!preLoop->isMarked() && preLoop->successorWithPhis() == header); MBasicBlock *backedge = header->backedge(); JS_ASSERT(backedge->isMarked() && backedge->successorWithPhis() == header); MDefinition *initial = phi->getOperand(preLoop->positionInPhiSuccessor()); if (initial->block()->isMarked()) return; SimpleLinearSum modified = ExtractLinearSum(phi->getOperand(backedge->positionInPhiSuccessor())); if (modified.term != phi || modified.constant == 0) return; if (!phi->range()) phi->setRange(new Range()); LinearSum initialSum; if (!initialSum.add(initial, 1)) return; // The phi may change by N each iteration, and is either nondecreasing or // nonincreasing. initial(phi) is either a lower or upper bound for the // phi, and initial(phi) + loopBound * N is either an upper or lower bound, // at all points within the loop, provided that loopBound >= 0. // // We are more interested, however, in the bound for phi at points // dominated by the loop bound's test; if the test dominates e.g. a bounds // check we want to hoist from the loop, using the value of the phi at the // head of the loop for this will usually be too imprecise to hoist the // check. These points will execute only if the backedge executes at least // one more time (as the test passed and the test dominates the backedge), // so we know both that loopBound >= 1 and that the phi's value has changed // at most loopBound - 1 times. Thus, another upper or lower bound for the // phi is initial(phi) + (loopBound - 1) * N, without requiring us to // ensure that loopBound >= 0. LinearSum limitSum(loopBound->sum); if (!limitSum.multiply(modified.constant) || !limitSum.add(initialSum)) return; int32_t negativeConstant; if (!SafeSub(0, modified.constant, &negativeConstant) || !limitSum.add(negativeConstant)) return; if (modified.constant > 0) { phi->range()->setSymbolicLower(new SymbolicBound(NULL, initialSum)); phi->range()->setSymbolicUpper(new SymbolicBound(loopBound, limitSum)); } else { phi->range()->setSymbolicUpper(new SymbolicBound(NULL, initialSum)); phi->range()->setSymbolicLower(new SymbolicBound(loopBound, limitSum)); } IonSpew(IonSpew_Range, "added symbolic range on %d", phi->id()); SpewRange(phi); }
LoopIterationBound * RangeAnalysis::analyzeLoopIterationCount(MBasicBlock *header, MTest *test, BranchDirection direction) { SimpleLinearSum lhs(NULL, 0); MDefinition *rhs; bool lessEqual; if (!ExtractLinearInequality(test, direction, &lhs, &rhs, &lessEqual)) return NULL; // Ensure the rhs is a loop invariant term. if (rhs && rhs->block()->isMarked()) { if (lhs.term && lhs.term->block()->isMarked()) return NULL; MDefinition *temp = lhs.term; lhs.term = rhs; rhs = temp; if (!SafeSub(0, lhs.constant, &lhs.constant)) return NULL; lessEqual = !lessEqual; } JS_ASSERT_IF(rhs, !rhs->block()->isMarked()); // Ensure the lhs is a phi node from the start of the loop body. if (!lhs.term || !lhs.term->isPhi() || lhs.term->block() != header) return NULL; // Check that the value of the lhs changes by a constant amount with each // loop iteration. This requires that the lhs be written in every loop // iteration with a value that is a constant difference from its value at // the start of the iteration. if (lhs.term->toPhi()->numOperands() != 2) return NULL; // The first operand of the phi should be the lhs' value at the start of // the first executed iteration, and not a value written which could // replace the second operand below during the middle of execution. MDefinition *lhsInitial = lhs.term->toPhi()->getOperand(0); if (lhsInitial->block()->isMarked()) return NULL; // The second operand of the phi should be a value written by an add/sub // in every loop iteration, i.e. in a block which dominates the backedge. MDefinition *lhsWrite = lhs.term->toPhi()->getOperand(1); if (lhsWrite->isBeta()) lhsWrite = lhsWrite->getOperand(0); if (!lhsWrite->isAdd() && !lhsWrite->isSub()) return NULL; if (!lhsWrite->block()->isMarked()) return NULL; MBasicBlock *bb = header->backedge(); for (; bb != lhsWrite->block() && bb != header; bb = bb->immediateDominator()) {} if (bb != lhsWrite->block()) return NULL; SimpleLinearSum lhsModified = ExtractLinearSum(lhsWrite); // Check that the value of the lhs at the backedge is of the form // 'old(lhs) + N'. We can be sure that old(lhs) is the value at the start // of the iteration, and not that written to lhs in a previous iteration, // as such a previous value could not appear directly in the addition: // it could not be stored in lhs as the lhs add/sub executes in every // iteration, and if it were stored in another variable its use here would // be as an operand to a phi node for that variable. if (lhsModified.term != lhs.term) return NULL; LinearSum bound; if (lhsModified.constant == 1 && !lessEqual) { // The value of lhs is 'initial(lhs) + iterCount' and this will end // execution of the loop if 'lhs + lhsN >= rhs'. Thus, an upper bound // on the number of backedges executed is: // // initial(lhs) + iterCount + lhsN == rhs // iterCount == rhsN - initial(lhs) - lhsN if (rhs) { if (!bound.add(rhs, 1)) return NULL; } if (!bound.add(lhsInitial, -1)) return NULL; int32_t lhsConstant; if (!SafeSub(0, lhs.constant, &lhsConstant)) return NULL; if (!bound.add(lhsConstant)) return NULL; } else if (lhsModified.constant == -1 && lessEqual) { // The value of lhs is 'initial(lhs) - iterCount'. Similar to the above // case, an upper bound on the number of backedges executed is: // // initial(lhs) - iterCount + lhsN == rhs // iterCount == initial(lhs) - rhs + lhsN if (!bound.add(lhsInitial, 1)) return NULL; if (rhs) { if (!bound.add(rhs, -1)) return NULL; } if (!bound.add(lhs.constant)) return NULL; } else { return NULL; } return new LoopIterationBound(header, test, bound); }
// Visit |def|. bool ValueNumberer::visitDefinition(MDefinition *def) { // If this instruction has a dependency() into an unreachable block, we'll // need to update AliasAnalysis. MDefinition *dep = def->dependency(); if (dep != nullptr && (dep->isDiscarded() || dep->block()->isDead())) { JitSpew(JitSpew_GVN, " AliasAnalysis invalidated"); if (updateAliasAnalysis_ && !dependenciesBroken_) { // TODO: Recomputing alias-analysis could theoretically expose more // GVN opportunities. JitSpew(JitSpew_GVN, " Will recompute!"); dependenciesBroken_ = true; } // Temporarily clear its dependency, to protect foldsTo, which may // wish to use the dependency to do store-to-load forwarding. def->setDependency(def->toInstruction()); } else { dep = nullptr; } // Look for a simplified form of |def|. MDefinition *sim = simplified(def); if (sim != def) { if (sim == nullptr) return false; // If |sim| doesn't belong to a block, insert it next to |def|. if (sim->block() == nullptr) def->block()->insertAfter(def->toInstruction(), sim->toInstruction()); #ifdef DEBUG JitSpew(JitSpew_GVN, " Folded %s%u to %s%u", def->opName(), def->id(), sim->opName(), sim->id()); #endif ReplaceAllUsesWith(def, sim); // The node's foldsTo said |def| can be replaced by |rep|. If |def| is a // guard, then either |rep| is also a guard, or a guard isn't actually // needed, so we can clear |def|'s guard flag and let it be discarded. def->setNotGuardUnchecked(); if (DeadIfUnused(def)) { if (!discardDefsRecursively(def)) return false; } def = sim; } // Now that foldsTo is done, re-enable the original dependency. Even though // it may be pointing into a discarded block, it's still valid for the // purposes of detecting congruent loads. if (dep != nullptr) def->setDependency(dep); // Look for a dominating def which makes |def| redundant. MDefinition *rep = leader(def); if (rep != def) { if (rep == nullptr) return false; if (rep->updateForReplacement(def)) { #ifdef DEBUG JitSpew(JitSpew_GVN, " Replacing %s%u with %s%u", def->opName(), def->id(), rep->opName(), rep->id()); #endif ReplaceAllUsesWith(def, rep); // The node's congruentTo said |def| is congruent to |rep|, and it's // dominated by |rep|. If |def| is a guard, it's covered by |rep|, // so we can clear |def|'s guard flag and let it be discarded. def->setNotGuardUnchecked(); if (DeadIfUnused(def)) { // discardDef should not add anything to the deadDefs, as the // redundant operation should have the same input operands. mozilla::DebugOnly<bool> r = discardDef(def); MOZ_ASSERT(r, "discardDef shouldn't have tried to add anything to the worklist, " "so it shouldn't have failed"); MOZ_ASSERT(deadDefs_.empty(), "discardDef shouldn't have added anything to the worklist"); } def = rep; } } return true; }
bool ValueNumberer::eliminateRedundancies() { // A definition is 'redundant' iff it is dominated by another definition // with the same value number. // // So, we traverse the dominator tree in pre-order, maintaining a hashmap // from value numbers to instructions. // // For each definition d with value number v, we look up v in the hashmap. // // If there is a definition d' in the hashmap, and the current traversal // index is within that instruction's dominated range, then we eliminate d, // replacing all uses of d with uses of d'. // // If there is no valid definition in the hashtable (the current definition // is not in dominated scope), then we insert the current instruction, // since it is the most dominant instruction with the given value number. InstructionMap defs; if (!defs.init()) return false; IonSpew(IonSpew_GVN, "Eliminating redundant instructions"); // Stack for pre-order CFG traversal. Vector<MBasicBlock *, 1, IonAllocPolicy> worklist; // The index of the current block in the CFG traversal. size_t index = 0; // Add all self-dominating blocks to the worklist. // This includes all roots. Order does not matter. for (MBasicBlockIterator i(graph_.begin()); i != graph_.end(); i++) { MBasicBlock *block = *i; if (block->immediateDominator() == block) { if (!worklist.append(block)) return false; } } // Starting from each self-dominating block, traverse the CFG in pre-order. while (!worklist.empty()) { MBasicBlock *block = worklist.popCopy(); IonSpew(IonSpew_GVN, "Looking at block %d", block->id()); // Add all immediate dominators to the front of the worklist. for (size_t i = 0; i < block->numImmediatelyDominatedBlocks(); i++) { if (!worklist.append(block->getImmediatelyDominatedBlock(i))) return false; } // For each instruction, attempt to look up a dominating definition. for (MDefinitionIterator iter(block); iter; ) { MDefinition *ins = simplify(*iter, true); // Instruction was replaced, and all uses have already been fixed. if (ins != *iter) { iter = block->discardDefAt(iter); continue; } // Instruction has side-effects and cannot be folded. if (!ins->isMovable() || ins->isEffectful()) { iter++; continue; } MDefinition *dom = findDominatingDef(defs, ins, index); if (!dom) return false; // Insertion failed. if (dom == ins || !dom->updateForReplacement(ins)) { iter++; continue; } IonSpew(IonSpew_GVN, "instruction %d is dominated by instruction %d (from block %d)", ins->id(), dom->id(), dom->block()->id()); ins->replaceAllUsesWith(dom); JS_ASSERT(!ins->hasUses()); JS_ASSERT(ins->block() == block); JS_ASSERT(!ins->isEffectful()); JS_ASSERT(ins->isMovable()); iter = ins->block()->discardDefAt(iter); } index++; } JS_ASSERT(index == graph_.numBlocks()); return true; }
MBasicBlock* MBasicBlock::NewSplitEdge(MIRGraph& graph, MBasicBlock* pred, size_t predEdgeIdx, MBasicBlock* succ) { MBasicBlock* split = nullptr; if (!succ->pc()) { // The predecessor does not have a PC, this is a Wasm compilation. split = MBasicBlock::New(graph, succ->info(), pred, SPLIT_EDGE); if (!split) return nullptr; } else { // The predecessor has a PC, this is an IonBuilder compilation. MResumePoint* succEntry = succ->entryResumePoint(); BytecodeSite* site = new(graph.alloc()) BytecodeSite(succ->trackedTree(), succEntry->pc()); split = new(graph.alloc()) MBasicBlock(graph, succ->info(), site, SPLIT_EDGE); if (!split->init()) return nullptr; // A split edge is used to simplify the graph to avoid having a // predecessor with multiple successors as well as a successor with // multiple predecessors. As instructions can be moved in this // split-edge block, we need to give this block a resume point. To do // so, we copy the entry resume points of the successor and filter the // phis to keep inputs from the current edge. // Propagate the caller resume point from the inherited block. split->callerResumePoint_ = succ->callerResumePoint(); // Split-edge are created after the interpreter stack emulation. Thus, // there is no need for creating slots. split->stackPosition_ = succEntry->stackDepth(); // Create a resume point using our initial stack position. MResumePoint* splitEntry = new(graph.alloc()) MResumePoint(split, succEntry->pc(), MResumePoint::ResumeAt); if (!splitEntry->init(graph.alloc())) return nullptr; split->entryResumePoint_ = splitEntry; // The target entry resume point might have phi operands, keep the // operands of the phi coming from our edge. size_t succEdgeIdx = succ->indexForPredecessor(pred); for (size_t i = 0, e = splitEntry->numOperands(); i < e; i++) { MDefinition* def = succEntry->getOperand(i); // This early in the pipeline, we have no recover instructions in // any entry resume point. MOZ_ASSERT_IF(def->block() == succ, def->isPhi()); if (def->block() == succ) def = def->toPhi()->getOperand(succEdgeIdx); splitEntry->initOperand(i, def); } // This is done in the New variant for wasm, so we cannot keep this // line below, where the rest of the graph is modified. if (!split->predecessors_.append(pred)) return nullptr; } split->setLoopDepth(succ->loopDepth()); // Insert the split edge block in-between. split->end(MGoto::New(graph.alloc(), succ)); graph.insertBlockAfter(pred, split); pred->replaceSuccessor(predEdgeIdx, split); succ->replacePredecessor(pred, split); return split; }
bool Sink(MIRGenerator* mir, MIRGraph& graph) { TempAllocator& alloc = graph.alloc(); bool sinkEnabled = mir->optimizationInfo().sinkEnabled(); for (PostorderIterator block = graph.poBegin(); block != graph.poEnd(); block++) { if (mir->shouldCancel("Sink")) return false; for (MInstructionReverseIterator iter = block->rbegin(); iter != block->rend(); ) { MInstruction* ins = *iter++; // Only instructions which can be recovered on bailout can be moved // into the bailout paths. if (ins->isGuard() || ins->isGuardRangeBailouts() || ins->isRecoveredOnBailout() || !ins->canRecoverOnBailout()) { continue; } // Compute a common dominator for all uses of the current // instruction. bool hasLiveUses = false; bool hasUses = false; MBasicBlock* usesDominator = nullptr; for (MUseIterator i(ins->usesBegin()), e(ins->usesEnd()); i != e; i++) { hasUses = true; MNode* consumerNode = (*i)->consumer(); if (consumerNode->isResumePoint()) continue; MDefinition* consumer = consumerNode->toDefinition(); if (consumer->isRecoveredOnBailout()) continue; hasLiveUses = true; // If the instruction is a Phi, then we should dominate the // predecessor from which the value is coming from. MBasicBlock* consumerBlock = consumer->block(); if (consumer->isPhi()) consumerBlock = consumerBlock->getPredecessor(consumer->indexOf(*i)); usesDominator = CommonDominator(usesDominator, consumerBlock); if (usesDominator == *block) break; } // Leave this instruction for DCE. if (!hasUses) continue; // We have no uses, so sink this instruction in all the bailout // paths. if (!hasLiveUses) { MOZ_ASSERT(!usesDominator); ins->setRecoveredOnBailout(); JitSpewDef(JitSpew_Sink, " No live uses, recover the instruction on bailout\n", ins); continue; } // This guard is temporarly moved here as the above code deals with // Dead Code elimination, which got moved into this Sink phase, as // the Dead Code elimination used to move instructions with no-live // uses to the bailout path. if (!sinkEnabled) continue; // To move an effectful instruction, we would have to verify that the // side-effect is not observed. In the mean time, we just inhibit // this optimization on effectful instructions. if (ins->isEffectful()) continue; // If all the uses are under a loop, we might not want to work // against LICM by moving everything back into the loop, but if the // loop is it-self inside an if, then we still want to move the // computation under this if statement. while (block->loopDepth() < usesDominator->loopDepth()) { MOZ_ASSERT(usesDominator != usesDominator->immediateDominator()); usesDominator = usesDominator->immediateDominator(); } // Only move instructions if there is a branch between the dominator // of the uses and the original instruction. This prevent moving the // computation of the arguments into an inline function if there is // no major win. MBasicBlock* lastJoin = usesDominator; while (*block != lastJoin && lastJoin->numPredecessors() == 1) { MOZ_ASSERT(lastJoin != lastJoin->immediateDominator()); MBasicBlock* next = lastJoin->immediateDominator(); if (next->numSuccessors() > 1) break; lastJoin = next; } if (*block == lastJoin) continue; // Skip to the next instruction if we cannot find a common dominator // for all the uses of this instruction, or if the common dominator // correspond to the block of the current instruction. if (!usesDominator || usesDominator == *block) continue; // Only instruction which can be recovered on bailout and which are // sinkable can be moved into blocks which are below while filling // the resume points with a clone which is recovered on bailout. // If the instruction has live uses and if it is clonable, then we // can clone the instruction for all non-dominated uses and move the // instruction into the block which is dominating all live uses. if (!ins->canClone()) continue; // If the block is a split-edge block, which is created for folding // test conditions, then the block has no resume point and has // multiple predecessors. In such case, we cannot safely move // bailing instruction to these blocks as we have no way to bailout. if (!usesDominator->entryResumePoint() && usesDominator->numPredecessors() != 1) continue; JitSpewDef(JitSpew_Sink, " Can Clone & Recover, sink instruction\n", ins); JitSpew(JitSpew_Sink, " into Block %u", usesDominator->id()); // Copy the arguments and clone the instruction. MDefinitionVector operands(alloc); for (size_t i = 0, end = ins->numOperands(); i < end; i++) { if (!operands.append(ins->getOperand(i))) return false; } MInstruction* clone = ins->clone(alloc, operands); ins->block()->insertBefore(ins, clone); clone->setRecoveredOnBailout(); // We should not update the producer of the entry resume point, as // it cannot refer to any instruction within the basic block excepts // for Phi nodes. MResumePoint* entry = usesDominator->entryResumePoint(); // Replace the instruction by its clone in all the resume points / // recovered-on-bailout instructions which are not in blocks which // are dominated by the usesDominator block. for (MUseIterator i(ins->usesBegin()), e(ins->usesEnd()); i != e; ) { MUse* use = *i++; MNode* consumer = use->consumer(); // If the consumer is a Phi, then we look for the index of the // use to find the corresponding predecessor block, which is // then used as the consumer block. MBasicBlock* consumerBlock = consumer->block(); if (consumer->isDefinition() && consumer->toDefinition()->isPhi()) { consumerBlock = consumerBlock->getPredecessor( consumer->toDefinition()->toPhi()->indexOf(use)); } // Keep the current instruction for all dominated uses, except // for the entry resume point of the block in which the // instruction would be moved into. if (usesDominator->dominates(consumerBlock) && (!consumer->isResumePoint() || consumer->toResumePoint() != entry)) { continue; } use->replaceProducer(clone); } // As we move this instruction in a different block, we should // verify that we do not carry over a resume point which would refer // to an outdated state of the control flow. if (ins->resumePoint()) ins->clearResumePoint(); // Now, that all uses which are not dominated by usesDominator are // using the cloned instruction, we can safely move the instruction // into the usesDominator block. MInstruction* at = usesDominator->safeInsertTop(nullptr, MBasicBlock::IgnoreRecover); block->moveBefore(at, ins); } } return true; }
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(); JS_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; JS_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 DEBUG 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->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. CompileInfo &info = oldPreheader->info(); if (header->trackedSite().pc()) { 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::NewAsmJS(graph, info, oldPreheader, MBasicBlock::LOOP_HEADER); unrolledBackedge = MBasicBlock::NewAsmJS(graph, info, unrolledHeader, MBasicBlock::NORMAL); newPreheader = MBasicBlock::NewAsmJS(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); if (!unrolledDefinitions.init()) CrashAtUnhandlableOOM("LoopUnroller::go"); // Add phis to the unrolled loop header which correspond to the phis in the // original loop header. JS_ASSERT(header->getPredecessor(0) == oldPreheader); for (MPhiIterator iter(header->phisBegin()); iter != header->phisEnd(); iter++) { MPhi *old = *iter; JS_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)) CrashAtUnhandlableOOM("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)) CrashAtUnhandlableOOM("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); 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); 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); newPreheader->setEntryResumePoint(rp); } // Generate the unrolled code. JS_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()) { makeReplacementInstruction(*iter); } else { // Control instructions are handled separately. JS_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); JS_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))) CrashAtUnhandlableOOM("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++]); } JS_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++])) CrashAtUnhandlableOOM("LoopUnroller::go"); } JS_ASSERT(phiIndex == phiValues.length()); unrollIndex++; } MGoto *backedgeJump = MGoto::New(alloc, unrolledHeader); unrolledBackedge->end(backedgeJump); // Place the old preheader before the unrolled loop. JS_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)) CrashAtUnhandlableOOM("LoopUnroller::go"); header->replacePredecessor(oldPreheader, newPreheader); oldPreheader->setSuccessorWithPhis(unrolledHeader, 0); newPreheader->setSuccessorWithPhis(header, 0); unrolledBackedge->setSuccessorWithPhis(unrolledHeader, 1); }
// Visit |def|. bool ValueNumberer::visitDefinition(MDefinition* def) { // Nop does not fit in any of the previous optimization, as its only purpose // is to reduce the register pressure by keeping additional resume // point. Still, there is no need consecutive list of MNop instructions, and // this will slow down every other iteration on the Graph. if (def->isNop()) { MNop* nop = def->toNop(); MBasicBlock* block = nop->block(); // We look backward to know if we can remove the previous Nop, we do not // look forward as we would not benefit from the folding made by GVN. MInstructionReverseIterator iter = ++block->rbegin(nop); // This nop is at the beginning of the basic block, just replace the // resume point of the basic block by the one from the resume point. if (iter == block->rend()) { JitSpew(JitSpew_GVN, " Removing Nop%u", nop->id()); nop->moveResumePointAsEntry(); block->discard(nop); return true; } // The previous instruction is also a Nop, no need to keep it anymore. MInstruction* prev = *iter; if (prev->isNop()) { JitSpew(JitSpew_GVN, " Removing Nop%u", prev->id()); block->discard(prev); return true; } // The Nop is introduced to capture the result and make sure the operands // are not live anymore when there are no further uses. Though when // all operands are still needed the Nop doesn't decrease the liveness // and can get removed. MResumePoint* rp = nop->resumePoint(); if (rp && rp->numOperands() > 0 && rp->getOperand(rp->numOperands() - 1) == prev && !nop->block()->lastIns()->isThrow() && !prev->isAssertRecoveredOnBailout()) { size_t numOperandsLive = 0; for (size_t j = 0; j < prev->numOperands(); j++) { for (size_t i = 0; i < rp->numOperands(); i++) { if (prev->getOperand(j) == rp->getOperand(i)) { numOperandsLive++; break; } } } if (numOperandsLive == prev->numOperands()) { JitSpew(JitSpew_GVN, " Removing Nop%u", nop->id()); block->discard(nop); } } return true; } // Skip optimizations on instructions which are recovered on bailout, to // avoid mixing instructions which are recovered on bailouts with // instructions which are not. if (def->isRecoveredOnBailout()) return true; // If this instruction has a dependency() into an unreachable block, we'll // need to update AliasAnalysis. MDefinition* dep = def->dependency(); if (dep != nullptr && (dep->isDiscarded() || dep->block()->isDead())) { JitSpew(JitSpew_GVN, " AliasAnalysis invalidated"); if (updateAliasAnalysis_ && !dependenciesBroken_) { // TODO: Recomputing alias-analysis could theoretically expose more // GVN opportunities. JitSpew(JitSpew_GVN, " Will recompute!"); dependenciesBroken_ = true; } // Temporarily clear its dependency, to protect foldsTo, which may // wish to use the dependency to do store-to-load forwarding. def->setDependency(def->toInstruction()); } else { dep = nullptr; } // Look for a simplified form of |def|. MDefinition* sim = simplified(def); if (sim != def) { if (sim == nullptr) return false; bool isNewInstruction = sim->block() == nullptr; // If |sim| doesn't belong to a block, insert it next to |def|. if (isNewInstruction) def->block()->insertAfter(def->toInstruction(), sim->toInstruction()); #ifdef JS_JITSPEW JitSpew(JitSpew_GVN, " Folded %s%u to %s%u", def->opName(), def->id(), sim->opName(), sim->id()); #endif MOZ_ASSERT(!sim->isDiscarded()); ReplaceAllUsesWith(def, sim); // The node's foldsTo said |def| can be replaced by |rep|. If |def| is a // guard, then either |rep| is also a guard, or a guard isn't actually // needed, so we can clear |def|'s guard flag and let it be discarded. def->setNotGuardUnchecked(); if (def->isGuardRangeBailouts()) sim->setGuardRangeBailoutsUnchecked(); if (DeadIfUnused(def)) { if (!discardDefsRecursively(def)) return false; // If that ended up discarding |sim|, then we're done here. if (sim->isDiscarded()) return true; } if (!rerun_ && def->isPhi() && !sim->isPhi()) { rerun_ = true; JitSpew(JitSpew_GVN, " Replacing phi%u may have enabled cascading optimisations; " "will re-run", def->id()); } // Otherwise, procede to optimize with |sim| in place of |def|. def = sim; // If the simplified instruction was already part of the graph, then we // probably already visited and optimized this instruction. if (!isNewInstruction) return true; } // Now that foldsTo is done, re-enable the original dependency. Even though // it may be pointing into a discarded block, it's still valid for the // purposes of detecting congruent loads. if (dep != nullptr) def->setDependency(dep); // Look for a dominating def which makes |def| redundant. MDefinition* rep = leader(def); if (rep != def) { if (rep == nullptr) return false; if (rep->updateForReplacement(def)) { #ifdef JS_JITSPEW JitSpew(JitSpew_GVN, " Replacing %s%u with %s%u", def->opName(), def->id(), rep->opName(), rep->id()); #endif ReplaceAllUsesWith(def, rep); // The node's congruentTo said |def| is congruent to |rep|, and it's // dominated by |rep|. If |def| is a guard, it's covered by |rep|, // so we can clear |def|'s guard flag and let it be discarded. def->setNotGuardUnchecked(); if (DeadIfUnused(def)) { // discardDef should not add anything to the deadDefs, as the // redundant operation should have the same input operands. mozilla::DebugOnly<bool> r = discardDef(def); MOZ_ASSERT(r, "discardDef shouldn't have tried to add anything to the worklist, " "so it shouldn't have failed"); MOZ_ASSERT(deadDefs_.empty(), "discardDef shouldn't have added anything to the worklist"); } def = rep; } } return true; }