From: alexandre.rames@arm.com Date: Wed, 25 Jun 2014 15:11:00 +0000 (+0000) Subject: ARM: Use the shifter operand to merge in previous shift instructions. X-Git-Tag: upstream/4.7.83~8535 X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=a7408e512488b24f266a1e2e10b98bd0b46a225b;p=platform%2Fupstream%2Fv8.git ARM: Use the shifter operand to merge in previous shift instructions. When possible, we transform sequences of code of the form lsl r1, r2, #imm add r0, r5, r1 into add r0, r5, r2 LSL #imm This is an adaptation of r21161. R=ulan@chromium.org Review URL: https://codereview.chromium.org/312173002 git-svn-id: https://v8.googlecode.com/svn/branches/bleeding_edge@22017 ce2b1a6d-e550-0410-aec6-3dcde31c8c00 --- diff --git a/src/arm/assembler-arm.cc b/src/arm/assembler-arm.cc index 814bd8b..3a4eced 100644 --- a/src/arm/assembler-arm.cc +++ b/src/arm/assembler-arm.cc @@ -281,6 +281,7 @@ Operand::Operand(Handle handle) { Operand::Operand(Register rm, ShiftOp shift_op, int shift_imm) { ASSERT(is_uint5(shift_imm)); + ASSERT(shift_op != NO_SHIFT); rm_ = rm; rs_ = no_reg; @@ -301,7 +302,7 @@ Operand::Operand(Register rm, ShiftOp shift_op, int shift_imm) { Operand::Operand(Register rm, ShiftOp shift_op, Register rs) { - ASSERT(shift_op != RRX); + ASSERT((shift_op != RRX) && (shift_op != NO_SHIFT)); rm_ = rm; rs_ = no_reg; shift_op_ = shift_op; @@ -957,16 +958,16 @@ void Assembler::next(Label* L) { // If this returns true then you have to use the rotate_imm and immed_8 // that it returns, because it may have already changed the instruction // to match them! -static bool fits_shifter(uint32_t imm32, - uint32_t* rotate_imm, - uint32_t* immed_8, - Instr* instr) { +bool fits_shifter(uint32_t imm32, + uint32_t* rotate_imm, + uint32_t* immed_8, + Instr* instr) { // imm32 must be unsigned. for (int rot = 0; rot < 16; rot++) { uint32_t imm8 = (imm32 << 2*rot) | (imm32 >> (32 - 2*rot)); if ((imm8 <= 0xff)) { - *rotate_imm = rot; - *immed_8 = imm8; + if (rotate_imm != NULL) *rotate_imm = rot; + if (immed_8 != NULL) *immed_8 = imm8; return true; } } @@ -982,7 +983,8 @@ static bool fits_shifter(uint32_t imm32, if (imm32 < 0x10000) { *instr ^= kMovwLeaveCCFlip; *instr |= EncodeMovwImmediate(imm32); - *rotate_imm = *immed_8 = 0; // Not used for movw. + if (rotate_imm != NULL) *rotate_imm = 0; // Not used for movw. + if (immed_8 != NULL) *immed_8 = 0; // Not used for movw. return true; } } diff --git a/src/arm/assembler-arm.h b/src/arm/assembler-arm.h index 3d91628..f2a4b75 100644 --- a/src/arm/assembler-arm.h +++ b/src/arm/assembler-arm.h @@ -1598,6 +1598,10 @@ class Assembler : public AssemblerBase { }; +bool fits_shifter(uint32_t imm32, uint32_t* rotate_imm, + uint32_t* immed_8, Instr* instr); + + class EnsureSpace BASE_EMBEDDED { public: explicit EnsureSpace(Assembler* assembler) { diff --git a/src/arm/constants-arm.h b/src/arm/constants-arm.h index 6a2d82f..620d02d 100644 --- a/src/arm/constants-arm.h +++ b/src/arm/constants-arm.h @@ -236,6 +236,7 @@ enum ShiftOp { // as an argument, and will never actually be encoded. The Assembler will // detect it and emit the correct ROR shift operand with shift_imm == 0. RRX = -1, + NO_SHIFT = -2, kNumberOfShifts = 4 }; diff --git a/src/arm/lithium-arm.cc b/src/arm/lithium-arm.cc index 0392997..d1cc422 100644 --- a/src/arm/lithium-arm.cc +++ b/src/arm/lithium-arm.cc @@ -676,11 +676,127 @@ LInstruction* LChunkBuilder::DoDeoptimize(HDeoptimize* instr) { } +HBitwiseBinaryOperation* LChunkBuilder::CanTransformToShiftedOp(HValue* val, + HValue** left) { + if (!val->representation().IsInteger32()) return NULL; + if (!(val->IsBitwise() || val->IsAdd() || val->IsSub())) return NULL; + + HBinaryOperation* hinstr = HBinaryOperation::cast(val); + HValue* hleft = hinstr->left(); + HValue* hright = hinstr->right(); + ASSERT(hleft->representation().Equals(hinstr->representation())); + ASSERT(hright->representation().Equals(hinstr->representation())); + + if ((hright->IsConstant() && + LikelyFitsImmField(hinstr, HConstant::cast(hright)->Integer32Value())) || + (hinstr->IsCommutative() && hleft->IsConstant() && + LikelyFitsImmField(hinstr, HConstant::cast(hleft)->Integer32Value()))) { + // The constant operand will likely fit in the immediate field. We are + // better off with + // mov r1, r2 LSL #imm + // add r0, r1, #imm2 + // than with + // mov r5, #imm2 + // add r0, r5, r2 LSL #imm + return NULL; + } + + HBitwiseBinaryOperation* shift = NULL; + // TODO(aleram): We will miss situations where a shift operation is used by + // different instructions both as a left and right operands. + if (hright->IsBitwiseBinaryShift() && + HBitwiseBinaryOperation::cast(hright)->right()->IsConstant()) { + shift = HBitwiseBinaryOperation::cast(hright); + if (left != NULL) { + *left = hleft; + } + } else if (hinstr->IsCommutative() && + hleft->IsBitwiseBinaryShift() && + HBitwiseBinaryOperation::cast(hleft)->right()->IsConstant()) { + shift = HBitwiseBinaryOperation::cast(hleft); + if (left != NULL) { + *left = hright; + } + } else { + return NULL; + } + + if ((JSShiftAmountFromHConstant(shift->right()) == 0) && shift->IsShr()) { + // Logical shifts right by zero can deoptimize. + return NULL; + } + + return shift; +} + + +bool LChunkBuilder::ShiftCanBeOptimizedAway(HBitwiseBinaryOperation* shift) { + if (!shift->representation().IsInteger32()) { + return false; + } + for (HUseIterator it(shift->uses()); !it.Done(); it.Advance()) { + if (shift != CanTransformToShiftedOp(it.value())) { + return false; + } + } + return true; +} + + +LInstruction* LChunkBuilder::TryDoOpWithShiftedRightOperand( + HBinaryOperation* instr) { + HValue* left; + HBitwiseBinaryOperation* shift = CanTransformToShiftedOp(instr, &left); + + if ((shift != NULL) && ShiftCanBeOptimizedAway(shift)) { + return DoShiftedBinaryOp(instr, left, shift); + } + return NULL; +} + + +LInstruction* LChunkBuilder::DoShiftedBinaryOp( + HBinaryOperation* hinstr, HValue* hleft, HBitwiseBinaryOperation* hshift) { + ASSERT(hshift->IsBitwiseBinaryShift()); + ASSERT(!hshift->IsShr() || (JSShiftAmountFromHConstant(hshift->right()) > 0)); + + LTemplateResultInstruction<1>* res; + LOperand* left = UseRegisterAtStart(hleft); + LOperand* right = UseRegisterAtStart(hshift->left()); + LOperand* shift_amount = UseConstant(hshift->right()); + ShiftOp shift_op; + switch (hshift->opcode()) { + case HValue::kShl: shift_op = LSL; break; + case HValue::kShr: shift_op = LSR; break; + case HValue::kSar: shift_op = ASR; break; + default: UNREACHABLE(); shift_op = NO_SHIFT; + } + + if (hinstr->IsBitwise()) { + res = new(zone()) LBitI(left, right, shift_op, shift_amount); + } else if (hinstr->IsAdd()) { + res = new(zone()) LAddI(left, right, shift_op, shift_amount); + } else { + ASSERT(hinstr->IsSub()); + res = new(zone()) LSubI(left, right, shift_op, shift_amount); + } + if (hinstr->CheckFlag(HValue::kCanOverflow)) { + AssignEnvironment(res); + } + return DefineAsRegister(res); +} + + LInstruction* LChunkBuilder::DoShift(Token::Value op, HBitwiseBinaryOperation* instr) { if (instr->representation().IsSmiOrInteger32()) { ASSERT(instr->left()->representation().Equals(instr->representation())); ASSERT(instr->right()->representation().Equals(instr->representation())); + + if (ShiftCanBeOptimizedAway(instr)) { + return NULL; + } + LOperand* left = UseRegisterAtStart(instr->left()); HValue* right_value = instr->right(); @@ -690,7 +806,7 @@ LInstruction* LChunkBuilder::DoShift(Token::Value op, if (right_value->IsConstant()) { HConstant* constant = HConstant::cast(right_value); right = chunk_->DefineConstantOperand(constant); - constant_value = constant->Integer32Value() & 0x1f; + constant_value = JSShiftAmountFromHConstant(constant); // Left shifts can deoptimize if we shift by > 0 and the result cannot be // truncated to smi. if (instr->representation().IsSmi() && constant_value > 0) { @@ -1250,6 +1366,11 @@ LInstruction* LChunkBuilder::DoBitwise(HBitwise* instr) { ASSERT(instr->right()->representation().Equals(instr->representation())); ASSERT(instr->CheckFlag(HValue::kTruncatingToInt32)); + LInstruction* shifted_operation = TryDoOpWithShiftedRightOperand(instr); + if (shifted_operation != NULL) { + return shifted_operation; + } + LOperand* left = UseRegisterAtStart(instr->BetterLeftOperand()); LOperand* right = UseOrConstantAtStart(instr->BetterRightOperand()); return DefineAsRegister(new(zone()) LBitI(left, right)); @@ -1533,6 +1654,11 @@ LInstruction* LChunkBuilder::DoSub(HSub* instr) { ASSERT(instr->left()->representation().Equals(instr->representation())); ASSERT(instr->right()->representation().Equals(instr->representation())); + LInstruction* shifted_operation = TryDoOpWithShiftedRightOperand(instr); + if (shifted_operation != NULL) { + return shifted_operation; + } + if (instr->left()->IsConstant()) { // If lhs is constant, do reverse subtraction instead. return DoRSub(instr); @@ -1600,6 +1726,12 @@ LInstruction* LChunkBuilder::DoAdd(HAdd* instr) { if (instr->representation().IsSmiOrInteger32()) { ASSERT(instr->left()->representation().Equals(instr->representation())); ASSERT(instr->right()->representation().Equals(instr->representation())); + + LInstruction* shifted_operation = TryDoOpWithShiftedRightOperand(instr); + if (shifted_operation != NULL) { + return shifted_operation; + } + LOperand* left = UseRegisterAtStart(instr->BetterLeftOperand()); LOperand* right = UseOrConstantAtStart(instr->BetterRightOperand()); LAddI* add = new(zone()) LAddI(left, right); diff --git a/src/arm/lithium-arm.h b/src/arm/lithium-arm.h index 3d7f72b..344185e 100644 --- a/src/arm/lithium-arm.h +++ b/src/arm/lithium-arm.h @@ -1239,7 +1239,14 @@ class LBoundsCheck V8_FINAL : public LTemplateInstruction<0, 2, 0> { class LBitI V8_FINAL : public LTemplateInstruction<1, 2, 0> { public: - LBitI(LOperand* left, LOperand* right) { + LBitI(LOperand* left, LOperand* right) + : shift_(NO_SHIFT), shift_amount_(0) { + inputs_[0] = left; + inputs_[1] = right; + } + + LBitI(LOperand* left, LOperand* right, ShiftOp shift, LOperand* shift_amount) + : shift_(shift), shift_amount_(shift_amount) { inputs_[0] = left; inputs_[1] = right; } @@ -1247,10 +1254,17 @@ class LBitI V8_FINAL : public LTemplateInstruction<1, 2, 0> { LOperand* left() { return inputs_[0]; } LOperand* right() { return inputs_[1]; } + ShiftOp shift() const { return shift_; } + LOperand* shift_amount() const { return shift_amount_; } + Token::Value op() const { return hydrogen()->op(); } DECLARE_CONCRETE_INSTRUCTION(BitI, "bit-i") DECLARE_HYDROGEN_ACCESSOR(Bitwise) + + protected: + ShiftOp shift_; + LOperand* shift_amount_; }; @@ -1277,16 +1291,30 @@ class LShiftI V8_FINAL : public LTemplateInstruction<1, 2, 0> { class LSubI V8_FINAL : public LTemplateInstruction<1, 2, 0> { public: - LSubI(LOperand* left, LOperand* right) { + LSubI(LOperand* left, LOperand* right) + : shift_(NO_SHIFT), shift_amount_(0) { + inputs_[0] = left; + inputs_[1] = right; + } + + LSubI(LOperand* left, LOperand* right, ShiftOp shift, LOperand* shift_amount) + : shift_(shift), shift_amount_(shift_amount) { inputs_[0] = left; inputs_[1] = right; } + ShiftOp shift() const { return shift_; } + LOperand* shift_amount() const { return shift_amount_; } + LOperand* left() { return inputs_[0]; } LOperand* right() { return inputs_[1]; } DECLARE_CONCRETE_INSTRUCTION(SubI, "sub-i") DECLARE_HYDROGEN_ACCESSOR(Sub) + + protected: + ShiftOp shift_; + LOperand* shift_amount_; }; @@ -1455,7 +1483,14 @@ class LSeqStringSetChar V8_FINAL : public LTemplateInstruction<1, 4, 0> { class LAddI V8_FINAL : public LTemplateInstruction<1, 2, 0> { public: - LAddI(LOperand* left, LOperand* right) { + LAddI(LOperand* left, LOperand* right) + : shift_(NO_SHIFT), shift_amount_(0) { + inputs_[0] = left; + inputs_[1] = right; + } + + LAddI(LOperand* left, LOperand* right, ShiftOp shift, LOperand* shift_amount) + : shift_(shift), shift_amount_(shift_amount) { inputs_[0] = left; inputs_[1] = right; } @@ -1463,8 +1498,15 @@ class LAddI V8_FINAL : public LTemplateInstruction<1, 2, 0> { LOperand* left() { return inputs_[0]; } LOperand* right() { return inputs_[1]; } + ShiftOp shift() const { return shift_; } + LOperand* shift_amount() const { return shift_amount_; } + DECLARE_CONCRETE_INSTRUCTION(AddI, "add-i") DECLARE_HYDROGEN_ACCESSOR(Add) + + protected: + ShiftOp shift_; + LOperand* shift_amount_; }; @@ -2855,6 +2897,49 @@ class LChunkBuilder V8_FINAL : public LChunkBuilderBase { void AddInstruction(LInstruction* instr, HInstruction* current); void DoBasicBlock(HBasicBlock* block, HBasicBlock* next_block); + + int JSShiftAmountFromHConstant(HValue* constant) { + return HConstant::cast(constant)->Integer32Value() & 0x1f; + } + bool LikelyFitsImmField(HInstruction* instr, int imm) { + Instr instr_bits; + // All arithmetic and logical operations accept the same range of + // immediates. In some cases though, the operation itself can be changed to + // get a wider effective range of immediates. + if (instr->IsAdd() || instr->IsSub()) { + // ADD and SUB can be exchanged with a negate immediate. + instr_bits = ADD; + } else if (HBitwise::cast(instr)->op() == Token::BIT_AND) { + ASSERT(instr->IsBitwise()); + // AND and BIC can be exchanged with an inverted immediate. + instr_bits = AND; + } else { + ASSERT(instr->IsBitwise()); + // Use ORR for all other operations, since fits_shifter() can't adapt ORR. + instr_bits = ORR; + } + return fits_shifter(imm, NULL, NULL, &instr_bits); + } + + // Indicates if a sequence of the form + // mov r1, r2 LSL #imm + // add r0, r5, r1 + // can be replaced with: + // add r0, r5, r2 LSL #imm + // If this is not possible, the function returns NULL. Otherwise it returns a + // pointer to the shift instruction that would be optimized away. + HBitwiseBinaryOperation* CanTransformToShiftedOp(HValue* val, + HValue** left = NULL); + // Checks if all uses of the shift operation can optimize it away. + bool ShiftCanBeOptimizedAway(HBitwiseBinaryOperation* shift); + // Attempts to merge the binary operation and a previous shift operation into + // a single operation. Returns the merged instruction on success, and NULL + // otherwise. + LInstruction* TryDoOpWithShiftedRightOperand(HBinaryOperation* op); + LInstruction* DoShiftedBinaryOp(HBinaryOperation* instr, + HValue* left, + HBitwiseBinaryOperation* shift); + LInstruction* DoShift(Token::Value op, HBitwiseBinaryOperation* instr); LInstruction* DoArithmeticD(Token::Value op, HArithmeticBinaryOperation* instr); diff --git a/src/arm/lithium-codegen-arm.cc b/src/arm/lithium-codegen-arm.cc index 14c34ca..8a43146 100644 --- a/src/arm/lithium-codegen-arm.cc +++ b/src/arm/lithium-codegen-arm.cc @@ -480,6 +480,19 @@ bool LCodeGen::IsInteger32(LConstantOperand* op) const { } +template +Operand LCodeGen::ToShiftedRightOperand(LOperand* right, LI* shift_info) { + if (shift_info->shift() == NO_SHIFT) { + return ToOperand(right); + } else { + return Operand( + ToRegister(right), + shift_info->shift(), + JSShiftAmountFromLConstant(shift_info->shift_amount())); + } +} + + bool LCodeGen::IsSmi(LConstantOperand* op) const { return chunk_->LookupLiteralRepresentation(op).IsSmi(); } @@ -1712,11 +1725,13 @@ void LCodeGen::DoBitI(LBitI* instr) { Register result = ToRegister(instr->result()); Operand right(no_reg); + ASSERT(right_op->IsRegister() || (instr->shift() == NO_SHIFT)); + if (right_op->IsStackSlot()) { right = Operand(EmitLoadRegister(right_op, ip)); } else { ASSERT(right_op->IsRegister() || right_op->IsConstantOperand()); - right = ToOperand(right_op); + right = ToShiftedRightOperand(right_op, instr); } switch (instr->op()) { @@ -1773,9 +1788,7 @@ void LCodeGen::DoShiftI(LShiftI* instr) { break; } } else { - // Mask the right_op operand. - int value = ToInteger32(LConstantOperand::cast(right_op)); - uint8_t shift_count = static_cast(value & 0x1F); + int shift_count = JSShiftAmountFromLConstant(right_op); switch (instr->op()) { case Token::ROR: if (shift_count != 0) { @@ -1835,12 +1848,15 @@ void LCodeGen::DoSubI(LSubI* instr) { bool can_overflow = instr->hydrogen()->CheckFlag(HValue::kCanOverflow); SBit set_cond = can_overflow ? SetCC : LeaveCC; + ASSERT(right->IsRegister() || (instr->shift() == NO_SHIFT)); + if (right->IsStackSlot()) { Register right_reg = EmitLoadRegister(right, ip); __ sub(ToRegister(result), ToRegister(left), Operand(right_reg), set_cond); } else { ASSERT(right->IsRegister() || right->IsConstantOperand()); - __ sub(ToRegister(result), ToRegister(left), ToOperand(right), set_cond); + __ sub(ToRegister(result), ToRegister(left), + ToShiftedRightOperand(right, instr), set_cond); } if (can_overflow) { @@ -2029,12 +2045,15 @@ void LCodeGen::DoAddI(LAddI* instr) { bool can_overflow = instr->hydrogen()->CheckFlag(HValue::kCanOverflow); SBit set_cond = can_overflow ? SetCC : LeaveCC; + ASSERT(right->IsRegister() || (instr->shift() == NO_SHIFT)); + if (right->IsStackSlot()) { Register right_reg = EmitLoadRegister(right, ip); __ add(ToRegister(result), ToRegister(left), Operand(right_reg), set_cond); } else { ASSERT(right->IsRegister() || right->IsConstantOperand()); - __ add(ToRegister(result), ToRegister(left), ToOperand(right), set_cond); + __ add(ToRegister(result), ToRegister(left), + ToShiftedRightOperand(right, instr), set_cond); } if (can_overflow) { diff --git a/src/arm/lithium-codegen-arm.h b/src/arm/lithium-codegen-arm.h index b20b3f2..aa98c5b 100644 --- a/src/arm/lithium-codegen-arm.h +++ b/src/arm/lithium-codegen-arm.h @@ -86,6 +86,13 @@ class LCodeGen: public LCodeGenBase { // Returns a MemOperand pointing to the high word of a DoubleStackSlot. MemOperand ToHighMemOperand(LOperand* op) const; + template + Operand ToShiftedRightOperand(LOperand* right, LI* shift_info); + + int JSShiftAmountFromLConstant(LOperand* constant) { + return ToInteger32(LConstantOperand::cast(constant)) & 0x1f; + } + bool IsInteger32(LConstantOperand* op) const; bool IsSmi(LConstantOperand* op) const; Handle ToHandle(LConstantOperand* op) const;