diff options
Diffstat (limited to 'compiler/optimizing/nodes.h')
| -rw-r--r-- | compiler/optimizing/nodes.h | 115 |
1 files changed, 108 insertions, 7 deletions
diff --git a/compiler/optimizing/nodes.h b/compiler/optimizing/nodes.h index fb0c889792..52a02c2285 100644 --- a/compiler/optimizing/nodes.h +++ b/compiler/optimizing/nodes.h @@ -400,6 +400,12 @@ class HGraph : public ArenaObject<kArenaAllocGraph> { // put deoptimization instructions, etc. void TransformLoopHeaderForBCE(HBasicBlock* header); + // Adds a new loop directly after the loop with the given header and exit. + // Returns the new preheader. + HBasicBlock* TransformLoopForVectorization(HBasicBlock* header, + HBasicBlock* body, + HBasicBlock* exit); + // Removes `block` from the graph. Assumes `block` has been disconnected from // other blocks and has no instructions or phis. void DeleteDeadEmptyBlock(HBasicBlock* block); @@ -1363,6 +1369,25 @@ class HLoopInformationOutwardIterator : public ValueObject { M(TypeConversion, Instruction) \ M(UShr, BinaryOperation) \ M(Xor, BinaryOperation) \ + M(VecReplicateScalar, VecUnaryOperation) \ + M(VecSetScalars, VecUnaryOperation) \ + M(VecSumReduce, VecUnaryOperation) \ + M(VecCnv, VecUnaryOperation) \ + M(VecNeg, VecUnaryOperation) \ + M(VecNot, VecUnaryOperation) \ + M(VecAdd, VecBinaryOperation) \ + M(VecSub, VecBinaryOperation) \ + M(VecMul, VecBinaryOperation) \ + M(VecDiv, VecBinaryOperation) \ + M(VecAnd, VecBinaryOperation) \ + M(VecAndNot, VecBinaryOperation) \ + M(VecOr, VecBinaryOperation) \ + M(VecXor, VecBinaryOperation) \ + M(VecShl, VecBinaryOperation) \ + M(VecShr, VecBinaryOperation) \ + M(VecUShr, VecBinaryOperation) \ + M(VecLoad, VecMemoryOperation) \ + M(VecStore, VecMemoryOperation) \ /* * Instructions, shared across several (not all) architectures. @@ -1424,7 +1449,11 @@ class HLoopInformationOutwardIterator : public ValueObject { M(Constant, Instruction) \ M(UnaryOperation, Instruction) \ M(BinaryOperation, Instruction) \ - M(Invoke, Instruction) + M(Invoke, Instruction) \ + M(VecOperation, Instruction) \ + M(VecUnaryOperation, VecOperation) \ + M(VecBinaryOperation, VecOperation) \ + M(VecMemoryOperation, VecOperation) #define FOR_EACH_INSTRUCTION(M) \ FOR_EACH_CONCRETE_INSTRUCTION(M) \ @@ -2081,6 +2110,7 @@ class HInstruction : public ArenaObject<kArenaAllocInstruction> { void SetLocations(LocationSummary* locations) { locations_ = locations; } void ReplaceWith(HInstruction* instruction); + void ReplaceUsesDominatedBy(HInstruction* dominator, HInstruction* replacement); void ReplaceInput(HInstruction* replacement, size_t index); // This is almost the same as doing `ReplaceWith()`. But in this helper, the @@ -2944,28 +2974,97 @@ class HTryBoundary FINAL : public HTemplateInstruction<0> { }; // Deoptimize to interpreter, upon checking a condition. -class HDeoptimize FINAL : public HTemplateInstruction<1> { +class HDeoptimize FINAL : public HVariableInputSizeInstruction { public: + enum class Kind { + kBCE, + kInline, + kLast = kInline + }; + + // Use this constructor when the `HDeoptimize` acts as a barrier, where no code can move + // across. + HDeoptimize(ArenaAllocator* arena, HInstruction* cond, Kind kind, uint32_t dex_pc) + : HVariableInputSizeInstruction( + SideEffects::All(), + dex_pc, + arena, + /* number_of_inputs */ 1, + kArenaAllocMisc) { + SetPackedFlag<kFieldCanBeMoved>(false); + SetPackedField<DeoptimizeKindField>(kind); + SetRawInputAt(0, cond); + } + + // Use this constructor when the `HDeoptimize` guards an instruction, and any user + // that relies on the deoptimization to pass should have its input be the `HDeoptimize` + // instead of `guard`. // We set CanTriggerGC to prevent any intermediate address to be live // at the point of the `HDeoptimize`. - HDeoptimize(HInstruction* cond, uint32_t dex_pc) - : HTemplateInstruction(SideEffects::CanTriggerGC(), dex_pc) { + HDeoptimize(ArenaAllocator* arena, + HInstruction* cond, + HInstruction* guard, + Kind kind, + uint32_t dex_pc) + : HVariableInputSizeInstruction( + SideEffects::CanTriggerGC(), + dex_pc, + arena, + /* number_of_inputs */ 2, + kArenaAllocMisc) { + SetPackedFlag<kFieldCanBeMoved>(true); + SetPackedField<DeoptimizeKindField>(kind); SetRawInputAt(0, cond); + SetRawInputAt(1, guard); } - bool CanBeMoved() const OVERRIDE { return true; } - bool InstructionDataEquals(const HInstruction* other ATTRIBUTE_UNUSED) const OVERRIDE { - return true; + bool CanBeMoved() const OVERRIDE { return GetPackedFlag<kFieldCanBeMoved>(); } + + bool InstructionDataEquals(const HInstruction* other) const OVERRIDE { + return (other->CanBeMoved() == CanBeMoved()) && (other->AsDeoptimize()->GetKind() == GetKind()); } + bool NeedsEnvironment() const OVERRIDE { return true; } + bool CanThrow() const OVERRIDE { return true; } + Kind GetKind() const { return GetPackedField<DeoptimizeKindField>(); } + + Primitive::Type GetType() const OVERRIDE { + return GuardsAnInput() ? GuardedInput()->GetType() : Primitive::kPrimVoid; + } + + bool GuardsAnInput() const { + return InputCount() == 2; + } + + HInstruction* GuardedInput() const { + DCHECK(GuardsAnInput()); + return InputAt(1); + } + + void RemoveGuard() { + RemoveInputAt(1); + } + DECLARE_INSTRUCTION(Deoptimize); private: + static constexpr size_t kFieldCanBeMoved = kNumberOfGenericPackedBits; + static constexpr size_t kFieldDeoptimizeKind = kNumberOfGenericPackedBits + 1; + static constexpr size_t kFieldDeoptimizeKindSize = + MinimumBitsToStore(static_cast<size_t>(Kind::kLast)); + static constexpr size_t kNumberOfDeoptimizePackedBits = + kFieldDeoptimizeKind + kFieldDeoptimizeKindSize; + static_assert(kNumberOfDeoptimizePackedBits <= kMaxNumberOfPackedBits, + "Too many packed fields."); + using DeoptimizeKindField = BitField<Kind, kFieldDeoptimizeKind, kFieldDeoptimizeKindSize>; + DISALLOW_COPY_AND_ASSIGN(HDeoptimize); }; +std::ostream& operator<<(std::ostream& os, const HDeoptimize::Kind& rhs); + // Represents a should_deoptimize flag. Currently used for CHA-based devirtualization. // The compiled code checks this flag value in a guard before devirtualized call and // if it's true, starts to do deoptimization. @@ -6619,6 +6718,8 @@ class HParallelMove FINAL : public HTemplateInstruction<0> { } // namespace art +#include "nodes_vector.h" + #if defined(ART_ENABLE_CODEGEN_arm) || defined(ART_ENABLE_CODEGEN_arm64) #include "nodes_shared.h" #endif |