diff options
-rw-r--r-- | compiler/optimizing/code_generator_mips64.cc | 667 | ||||
-rw-r--r-- | compiler/optimizing/code_generator_mips64.h | 10 | ||||
-rw-r--r-- | compiler/optimizing/instruction_simplifier.cc | 7 | ||||
-rw-r--r-- | compiler/utils/mips64/assembler_mips64.cc | 114 | ||||
-rw-r--r-- | compiler/utils/mips64/assembler_mips64.h | 26 | ||||
-rw-r--r-- | compiler/utils/mips64/assembler_mips64_test.cc | 152 |
6 files changed, 729 insertions, 247 deletions
diff --git a/compiler/optimizing/code_generator_mips64.cc b/compiler/optimizing/code_generator_mips64.cc index fb45ef938d..1e428a06e1 100644 --- a/compiler/optimizing/code_generator_mips64.cc +++ b/compiler/optimizing/code_generator_mips64.cc @@ -1752,11 +1752,7 @@ void InstructionCodeGeneratorMIPS64::VisitClinitCheck(HClinitCheck* check) { void LocationsBuilderMIPS64::VisitCompare(HCompare* compare) { Primitive::Type in_type = compare->InputAt(0)->GetType(); - LocationSummary::CallKind call_kind = Primitive::IsFloatingPointType(in_type) - ? LocationSummary::kCall - : LocationSummary::kNoCall; - - LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(compare, call_kind); + LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(compare); switch (in_type) { case Primitive::kPrimLong: @@ -1766,13 +1762,11 @@ void LocationsBuilderMIPS64::VisitCompare(HCompare* compare) { break; case Primitive::kPrimFloat: - case Primitive::kPrimDouble: { - InvokeRuntimeCallingConvention calling_convention; - locations->SetInAt(0, Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(0))); - locations->SetInAt(1, Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(1))); - locations->SetOut(calling_convention.GetReturnLocation(Primitive::kPrimInt)); + case Primitive::kPrimDouble: + locations->SetInAt(0, Location::RequiresFpuRegister()); + locations->SetInAt(1, Location::RequiresFpuRegister()); + locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); break; - } default: LOG(FATAL) << "Unexpected type for compare operation " << in_type; @@ -1781,14 +1775,15 @@ void LocationsBuilderMIPS64::VisitCompare(HCompare* compare) { void InstructionCodeGeneratorMIPS64::VisitCompare(HCompare* instruction) { LocationSummary* locations = instruction->GetLocations(); + GpuRegister res = locations->Out().AsRegister<GpuRegister>(); Primitive::Type in_type = instruction->InputAt(0)->GetType(); + bool gt_bias = instruction->IsGtBias(); // 0 if: left == right // 1 if: left > right // -1 if: left < right switch (in_type) { case Primitive::kPrimLong: { - GpuRegister dst = locations->Out().AsRegister<GpuRegister>(); GpuRegister lhs = locations->InAt(0).AsRegister<GpuRegister>(); Location rhs_location = locations->InAt(1); bool use_imm = rhs_location.IsConstant(); @@ -1803,35 +1798,52 @@ void InstructionCodeGeneratorMIPS64::VisitCompare(HCompare* instruction) { rhs = rhs_location.AsRegister<GpuRegister>(); } __ Slt(TMP, lhs, rhs); - __ Slt(dst, rhs, lhs); - __ Subu(dst, dst, TMP); + __ Slt(res, rhs, lhs); + __ Subu(res, res, TMP); break; } - case Primitive::kPrimFloat: - case Primitive::kPrimDouble: { - int32_t entry_point_offset; - if (in_type == Primitive::kPrimFloat) { - entry_point_offset = instruction->IsGtBias() ? QUICK_ENTRY_POINT(pCmpgFloat) - : QUICK_ENTRY_POINT(pCmplFloat); + case Primitive::kPrimFloat: { + FpuRegister lhs = locations->InAt(0).AsFpuRegister<FpuRegister>(); + FpuRegister rhs = locations->InAt(1).AsFpuRegister<FpuRegister>(); + Mips64Label done; + __ CmpEqS(FTMP, lhs, rhs); + __ LoadConst32(res, 0); + __ Bc1nez(FTMP, &done); + if (gt_bias) { + __ CmpLtS(FTMP, lhs, rhs); + __ LoadConst32(res, -1); + __ Bc1nez(FTMP, &done); + __ LoadConst32(res, 1); } else { - entry_point_offset = instruction->IsGtBias() ? QUICK_ENTRY_POINT(pCmpgDouble) - : QUICK_ENTRY_POINT(pCmplDouble); + __ CmpLtS(FTMP, rhs, lhs); + __ LoadConst32(res, 1); + __ Bc1nez(FTMP, &done); + __ LoadConst32(res, -1); } - codegen_->InvokeRuntime(entry_point_offset, instruction, instruction->GetDexPc(), nullptr); - if (in_type == Primitive::kPrimFloat) { - if (instruction->IsGtBias()) { - CheckEntrypointTypes<kQuickCmpgFloat, int32_t, float, float>(); - } else { - CheckEntrypointTypes<kQuickCmplFloat, int32_t, float, float>(); - } + __ Bind(&done); + break; + } + + case Primitive::kPrimDouble: { + FpuRegister lhs = locations->InAt(0).AsFpuRegister<FpuRegister>(); + FpuRegister rhs = locations->InAt(1).AsFpuRegister<FpuRegister>(); + Mips64Label done; + __ CmpEqD(FTMP, lhs, rhs); + __ LoadConst32(res, 0); + __ Bc1nez(FTMP, &done); + if (gt_bias) { + __ CmpLtD(FTMP, lhs, rhs); + __ LoadConst32(res, -1); + __ Bc1nez(FTMP, &done); + __ LoadConst32(res, 1); } else { - if (instruction->IsGtBias()) { - CheckEntrypointTypes<kQuickCmpgDouble, int32_t, double, double>(); - } else { - CheckEntrypointTypes<kQuickCmplDouble, int32_t, double, double>(); - } + __ CmpLtD(FTMP, rhs, lhs); + __ LoadConst32(res, 1); + __ Bc1nez(FTMP, &done); + __ LoadConst32(res, -1); } + __ Bind(&done); break; } @@ -1842,8 +1854,19 @@ void InstructionCodeGeneratorMIPS64::VisitCompare(HCompare* instruction) { void LocationsBuilderMIPS64::VisitCondition(HCondition* instruction) { LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction); - locations->SetInAt(0, Location::RequiresRegister()); - locations->SetInAt(1, Location::RegisterOrConstant(instruction->InputAt(1))); + switch (instruction->InputAt(0)->GetType()) { + default: + case Primitive::kPrimLong: + locations->SetInAt(0, Location::RequiresRegister()); + locations->SetInAt(1, Location::RegisterOrConstant(instruction->InputAt(1))); + break; + + case Primitive::kPrimFloat: + case Primitive::kPrimDouble: + locations->SetInAt(0, Location::RequiresFpuRegister()); + locations->SetInAt(1, Location::RequiresFpuRegister()); + break; + } if (instruction->NeedsMaterialization()) { locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); } @@ -1854,129 +1877,42 @@ void InstructionCodeGeneratorMIPS64::VisitCondition(HCondition* instruction) { return; } - // TODO: generalize to long - DCHECK_NE(instruction->InputAt(0)->GetType(), Primitive::kPrimLong); - + Primitive::Type type = instruction->InputAt(0)->GetType(); LocationSummary* locations = instruction->GetLocations(); - GpuRegister dst = locations->Out().AsRegister<GpuRegister>(); - GpuRegister lhs = locations->InAt(0).AsRegister<GpuRegister>(); - Location rhs_location = locations->InAt(1); - - GpuRegister rhs_reg = ZERO; - int64_t rhs_imm = 0; - bool use_imm = rhs_location.IsConstant(); - if (use_imm) { - rhs_imm = CodeGenerator::GetInt32ValueOf(rhs_location.GetConstant()); - } else { - rhs_reg = rhs_location.AsRegister<GpuRegister>(); - } - - IfCondition if_cond = instruction->GetCondition(); + Mips64Label true_label; - switch (if_cond) { - case kCondEQ: - case kCondNE: - if (use_imm && IsUint<16>(rhs_imm)) { - __ Xori(dst, lhs, rhs_imm); - } else { - if (use_imm) { - rhs_reg = TMP; - __ LoadConst32(rhs_reg, rhs_imm); - } - __ Xor(dst, lhs, rhs_reg); - } - if (if_cond == kCondEQ) { - __ Sltiu(dst, dst, 1); - } else { - __ Sltu(dst, ZERO, dst); - } - break; + switch (type) { + default: + // Integer case. + GenerateIntLongCompare(instruction->GetCondition(), /* is64bit */ false, locations); + return; + case Primitive::kPrimLong: + GenerateIntLongCompare(instruction->GetCondition(), /* is64bit */ true, locations); + return; - case kCondLT: - case kCondGE: - if (use_imm && IsInt<16>(rhs_imm)) { - __ Slti(dst, lhs, rhs_imm); - } else { - if (use_imm) { - rhs_reg = TMP; - __ LoadConst32(rhs_reg, rhs_imm); - } - __ Slt(dst, lhs, rhs_reg); - } - if (if_cond == kCondGE) { - // Simulate lhs >= rhs via !(lhs < rhs) since there's - // only the slt instruction but no sge. - __ Xori(dst, dst, 1); - } + case Primitive::kPrimFloat: + case Primitive::kPrimDouble: + // TODO: don't use branches. + GenerateFpCompareAndBranch(instruction->GetCondition(), + instruction->IsGtBias(), + type, + locations, + &true_label); break; + } - case kCondLE: - case kCondGT: - if (use_imm && IsInt<16>(rhs_imm + 1)) { - // Simulate lhs <= rhs via lhs < rhs + 1. - __ Slti(dst, lhs, rhs_imm + 1); - if (if_cond == kCondGT) { - // Simulate lhs > rhs via !(lhs <= rhs) since there's - // only the slti instruction but no sgti. - __ Xori(dst, dst, 1); - } - } else { - if (use_imm) { - rhs_reg = TMP; - __ LoadConst32(rhs_reg, rhs_imm); - } - __ Slt(dst, rhs_reg, lhs); - if (if_cond == kCondLE) { - // Simulate lhs <= rhs via !(rhs < lhs) since there's - // only the slt instruction but no sle. - __ Xori(dst, dst, 1); - } - } - break; + // Convert the branches into the result. + Mips64Label done; - case kCondB: - case kCondAE: - if (use_imm && 0 <= rhs_imm && rhs_imm <= 0x7fff) { - __ Sltiu(dst, lhs, rhs_imm); - } else { - if (use_imm) { - rhs_reg = TMP; - __ LoadConst32(rhs_reg, rhs_imm); - } - __ Sltu(dst, lhs, rhs_reg); - } - if (if_cond == kCondAE) { - // Simulate lhs >= rhs via !(lhs < rhs) since there's - // only the sltu instruction but no sgeu. - __ Xori(dst, dst, 1); - } - break; + // False case: result = 0. + __ LoadConst32(dst, 0); + __ Bc(&done); - case kCondBE: - case kCondA: - if (use_imm && 0 <= rhs_imm && rhs_imm <= 0x7ffe) { - // Simulate lhs <= rhs via lhs < rhs + 1. - __ Sltiu(dst, lhs, rhs_imm + 1); - if (if_cond == kCondA) { - // Simulate lhs > rhs via !(lhs <= rhs) since there's - // only the sltiu instruction but no sgtiu. - __ Xori(dst, dst, 1); - } - } else { - if (use_imm) { - rhs_reg = TMP; - __ LoadConst32(rhs_reg, rhs_imm); - } - __ Sltu(dst, rhs_reg, lhs); - if (if_cond == kCondBE) { - // Simulate lhs <= rhs via !(rhs < lhs) since there's - // only the sltu instruction but no sleu. - __ Xori(dst, dst, 1); - } - } - break; - } + // True case: result = 1. + __ Bind(&true_label); + __ LoadConst32(dst, 1); + __ Bind(&done); } void InstructionCodeGeneratorMIPS64::DivRemOneOrMinusOne(HBinaryOperation* instruction) { @@ -2375,6 +2311,329 @@ void InstructionCodeGeneratorMIPS64::VisitTryBoundary(HTryBoundary* try_boundary } } +void InstructionCodeGeneratorMIPS64::GenerateIntLongCompare(IfCondition cond, + bool is64bit, + LocationSummary* locations) { + GpuRegister dst = locations->Out().AsRegister<GpuRegister>(); + GpuRegister lhs = locations->InAt(0).AsRegister<GpuRegister>(); + Location rhs_location = locations->InAt(1); + GpuRegister rhs_reg = ZERO; + int64_t rhs_imm = 0; + bool use_imm = rhs_location.IsConstant(); + if (use_imm) { + if (is64bit) { + rhs_imm = CodeGenerator::GetInt64ValueOf(rhs_location.GetConstant()); + } else { + rhs_imm = CodeGenerator::GetInt32ValueOf(rhs_location.GetConstant()); + } + } else { + rhs_reg = rhs_location.AsRegister<GpuRegister>(); + } + int64_t rhs_imm_plus_one = rhs_imm + UINT64_C(1); + + switch (cond) { + case kCondEQ: + case kCondNE: + if (use_imm && IsUint<16>(rhs_imm)) { + __ Xori(dst, lhs, rhs_imm); + } else { + if (use_imm) { + rhs_reg = TMP; + __ LoadConst64(rhs_reg, rhs_imm); + } + __ Xor(dst, lhs, rhs_reg); + } + if (cond == kCondEQ) { + __ Sltiu(dst, dst, 1); + } else { + __ Sltu(dst, ZERO, dst); + } + break; + + case kCondLT: + case kCondGE: + if (use_imm && IsInt<16>(rhs_imm)) { + __ Slti(dst, lhs, rhs_imm); + } else { + if (use_imm) { + rhs_reg = TMP; + __ LoadConst64(rhs_reg, rhs_imm); + } + __ Slt(dst, lhs, rhs_reg); + } + if (cond == kCondGE) { + // Simulate lhs >= rhs via !(lhs < rhs) since there's + // only the slt instruction but no sge. + __ Xori(dst, dst, 1); + } + break; + + case kCondLE: + case kCondGT: + if (use_imm && IsInt<16>(rhs_imm_plus_one)) { + // Simulate lhs <= rhs via lhs < rhs + 1. + __ Slti(dst, lhs, rhs_imm_plus_one); + if (cond == kCondGT) { + // Simulate lhs > rhs via !(lhs <= rhs) since there's + // only the slti instruction but no sgti. + __ Xori(dst, dst, 1); + } + } else { + if (use_imm) { + rhs_reg = TMP; + __ LoadConst64(rhs_reg, rhs_imm); + } + __ Slt(dst, rhs_reg, lhs); + if (cond == kCondLE) { + // Simulate lhs <= rhs via !(rhs < lhs) since there's + // only the slt instruction but no sle. + __ Xori(dst, dst, 1); + } + } + break; + + case kCondB: + case kCondAE: + if (use_imm && IsInt<16>(rhs_imm)) { + // Sltiu sign-extends its 16-bit immediate operand before + // the comparison and thus lets us compare directly with + // unsigned values in the ranges [0, 0x7fff] and + // [0x[ffffffff]ffff8000, 0x[ffffffff]ffffffff]. + __ Sltiu(dst, lhs, rhs_imm); + } else { + if (use_imm) { + rhs_reg = TMP; + __ LoadConst64(rhs_reg, rhs_imm); + } + __ Sltu(dst, lhs, rhs_reg); + } + if (cond == kCondAE) { + // Simulate lhs >= rhs via !(lhs < rhs) since there's + // only the sltu instruction but no sgeu. + __ Xori(dst, dst, 1); + } + break; + + case kCondBE: + case kCondA: + if (use_imm && (rhs_imm_plus_one != 0) && IsInt<16>(rhs_imm_plus_one)) { + // Simulate lhs <= rhs via lhs < rhs + 1. + // Note that this only works if rhs + 1 does not overflow + // to 0, hence the check above. + // Sltiu sign-extends its 16-bit immediate operand before + // the comparison and thus lets us compare directly with + // unsigned values in the ranges [0, 0x7fff] and + // [0x[ffffffff]ffff8000, 0x[ffffffff]ffffffff]. + __ Sltiu(dst, lhs, rhs_imm_plus_one); + if (cond == kCondA) { + // Simulate lhs > rhs via !(lhs <= rhs) since there's + // only the sltiu instruction but no sgtiu. + __ Xori(dst, dst, 1); + } + } else { + if (use_imm) { + rhs_reg = TMP; + __ LoadConst64(rhs_reg, rhs_imm); + } + __ Sltu(dst, rhs_reg, lhs); + if (cond == kCondBE) { + // Simulate lhs <= rhs via !(rhs < lhs) since there's + // only the sltu instruction but no sleu. + __ Xori(dst, dst, 1); + } + } + break; + } +} + +void InstructionCodeGeneratorMIPS64::GenerateIntLongCompareAndBranch(IfCondition cond, + bool is64bit, + LocationSummary* locations, + Mips64Label* label) { + GpuRegister lhs = locations->InAt(0).AsRegister<GpuRegister>(); + Location rhs_location = locations->InAt(1); + GpuRegister rhs_reg = ZERO; + int64_t rhs_imm = 0; + bool use_imm = rhs_location.IsConstant(); + if (use_imm) { + if (is64bit) { + rhs_imm = CodeGenerator::GetInt64ValueOf(rhs_location.GetConstant()); + } else { + rhs_imm = CodeGenerator::GetInt32ValueOf(rhs_location.GetConstant()); + } + } else { + rhs_reg = rhs_location.AsRegister<GpuRegister>(); + } + + if (use_imm && rhs_imm == 0) { + switch (cond) { + case kCondEQ: + case kCondBE: // <= 0 if zero + __ Beqzc(lhs, label); + break; + case kCondNE: + case kCondA: // > 0 if non-zero + __ Bnezc(lhs, label); + break; + case kCondLT: + __ Bltzc(lhs, label); + break; + case kCondGE: + __ Bgezc(lhs, label); + break; + case kCondLE: + __ Blezc(lhs, label); + break; + case kCondGT: + __ Bgtzc(lhs, label); + break; + case kCondB: // always false + break; + case kCondAE: // always true + __ Bc(label); + break; + } + } else { + if (use_imm) { + rhs_reg = TMP; + __ LoadConst64(rhs_reg, rhs_imm); + } + switch (cond) { + case kCondEQ: + __ Beqc(lhs, rhs_reg, label); + break; + case kCondNE: + __ Bnec(lhs, rhs_reg, label); + break; + case kCondLT: + __ Bltc(lhs, rhs_reg, label); + break; + case kCondGE: + __ Bgec(lhs, rhs_reg, label); + break; + case kCondLE: + __ Bgec(rhs_reg, lhs, label); + break; + case kCondGT: + __ Bltc(rhs_reg, lhs, label); + break; + case kCondB: + __ Bltuc(lhs, rhs_reg, label); + break; + case kCondAE: + __ Bgeuc(lhs, rhs_reg, label); + break; + case kCondBE: + __ Bgeuc(rhs_reg, lhs, label); + break; + case kCondA: + __ Bltuc(rhs_reg, lhs, label); + break; + } + } +} + +void InstructionCodeGeneratorMIPS64::GenerateFpCompareAndBranch(IfCondition cond, + bool gt_bias, + Primitive::Type type, + LocationSummary* locations, + Mips64Label* label) { + FpuRegister lhs = locations->InAt(0).AsFpuRegister<FpuRegister>(); + FpuRegister rhs = locations->InAt(1).AsFpuRegister<FpuRegister>(); + if (type == Primitive::kPrimFloat) { + switch (cond) { + case kCondEQ: + __ CmpEqS(FTMP, lhs, rhs); + __ Bc1nez(FTMP, label); + break; + case kCondNE: + __ CmpEqS(FTMP, lhs, rhs); + __ Bc1eqz(FTMP, label); + break; + case kCondLT: + if (gt_bias) { + __ CmpLtS(FTMP, lhs, rhs); + } else { + __ CmpUltS(FTMP, lhs, rhs); + } + __ Bc1nez(FTMP, label); + break; + case kCondLE: + if (gt_bias) { + __ CmpLeS(FTMP, lhs, rhs); + } else { + __ CmpUleS(FTMP, lhs, rhs); + } + __ Bc1nez(FTMP, label); + break; + case kCondGT: + if (gt_bias) { + __ CmpUltS(FTMP, rhs, lhs); + } else { + __ CmpLtS(FTMP, rhs, lhs); + } + __ Bc1nez(FTMP, label); + break; + case kCondGE: + if (gt_bias) { + __ CmpUleS(FTMP, rhs, lhs); + } else { + __ CmpLeS(FTMP, rhs, lhs); + } + __ Bc1nez(FTMP, label); + break; + default: + LOG(FATAL) << "Unexpected non-floating-point condition"; + } + } else { + DCHECK_EQ(type, Primitive::kPrimDouble); + switch (cond) { + case kCondEQ: + __ CmpEqD(FTMP, lhs, rhs); + __ Bc1nez(FTMP, label); + break; + case kCondNE: + __ CmpEqD(FTMP, lhs, rhs); + __ Bc1eqz(FTMP, label); + break; + case kCondLT: + if (gt_bias) { + __ CmpLtD(FTMP, lhs, rhs); + } else { + __ CmpUltD(FTMP, lhs, rhs); + } + __ Bc1nez(FTMP, label); + break; + case kCondLE: + if (gt_bias) { + __ CmpLeD(FTMP, lhs, rhs); + } else { + __ CmpUleD(FTMP, lhs, rhs); + } + __ Bc1nez(FTMP, label); + break; + case kCondGT: + if (gt_bias) { + __ CmpUltD(FTMP, rhs, lhs); + } else { + __ CmpLtD(FTMP, rhs, lhs); + } + __ Bc1nez(FTMP, label); + break; + case kCondGE: + if (gt_bias) { + __ CmpUleD(FTMP, rhs, lhs); + } else { + __ CmpLeD(FTMP, rhs, lhs); + } + __ Bc1nez(FTMP, label); + break; + default: + LOG(FATAL) << "Unexpected non-floating-point condition"; + } + } +} + void InstructionCodeGeneratorMIPS64::GenerateTestAndBranch(HInstruction* instruction, size_t condition_input_index, Mips64Label* true_target, @@ -2420,97 +2679,27 @@ void InstructionCodeGeneratorMIPS64::GenerateTestAndBranch(HInstruction* instruc // The condition instruction has not been materialized, use its inputs as // the comparison and its condition as the branch condition. HCondition* condition = cond->AsCondition(); + Primitive::Type type = condition->InputAt(0)->GetType(); + LocationSummary* locations = cond->GetLocations(); + IfCondition if_cond = condition->GetCondition(); + Mips64Label* branch_target = true_target; - GpuRegister lhs = condition->GetLocations()->InAt(0).AsRegister<GpuRegister>(); - Location rhs_location = condition->GetLocations()->InAt(1); - GpuRegister rhs_reg = ZERO; - int32_t rhs_imm = 0; - bool use_imm = rhs_location.IsConstant(); - if (use_imm) { - rhs_imm = CodeGenerator::GetInt32ValueOf(rhs_location.GetConstant()); - } else { - rhs_reg = rhs_location.AsRegister<GpuRegister>(); - } - - IfCondition if_cond; - Mips64Label* non_fallthrough_target; if (true_target == nullptr) { if_cond = condition->GetOppositeCondition(); - non_fallthrough_target = false_target; - } else { - if_cond = condition->GetCondition(); - non_fallthrough_target = true_target; - } - - if (use_imm && rhs_imm == 0) { - switch (if_cond) { - case kCondEQ: - __ Beqzc(lhs, non_fallthrough_target); - break; - case kCondNE: - __ Bnezc(lhs, non_fallthrough_target); - break; - case kCondLT: - __ Bltzc(lhs, non_fallthrough_target); - break; - case kCondGE: - __ Bgezc(lhs, non_fallthrough_target); - break; - case kCondLE: - __ Blezc(lhs, non_fallthrough_target); - break; - case kCondGT: - __ Bgtzc(lhs, non_fallthrough_target); - break; - case kCondB: - break; // always false - case kCondBE: - __ Beqzc(lhs, non_fallthrough_target); // <= 0 if zero - break; - case kCondA: - __ Bnezc(lhs, non_fallthrough_target); // > 0 if non-zero - break; - case kCondAE: - __ Bc(non_fallthrough_target); // always true - break; - } - } else { - if (use_imm) { - rhs_reg = TMP; - __ LoadConst32(rhs_reg, rhs_imm); - } - switch (if_cond) { - case kCondEQ: - __ Beqc(lhs, rhs_reg, non_fallthrough_target); - break; - case kCondNE: - __ Bnec(lhs, rhs_reg, non_fallthrough_target); - break; - case kCondLT: - __ Bltc(lhs, rhs_reg, non_fallthrough_target); - break; - case kCondGE: - __ Bgec(lhs, rhs_reg, non_fallthrough_target); - break; - case kCondLE: - __ Bgec(rhs_reg, lhs, non_fallthrough_target); - break; - case kCondGT: - __ Bltc(rhs_reg, lhs, non_fallthrough_target); - break; - case kCondB: - __ Bltuc(lhs, rhs_reg, non_fallthrough_target); - break; - case kCondAE: - __ Bgeuc(lhs, rhs_reg, non_fallthrough_target); - break; - case kCondBE: - __ Bgeuc(rhs_reg, lhs, non_fallthrough_target); - break; - case kCondA: - __ Bltuc(rhs_reg, lhs, non_fallthrough_target); - break; - } + branch_target = false_target; + } + + switch (type) { + default: + GenerateIntLongCompareAndBranch(if_cond, /* is64bit */ false, locations, branch_target); + break; + case Primitive::kPrimLong: + GenerateIntLongCompareAndBranch(if_cond, /* is64bit */ true, locations, branch_target); + break; + case Primitive::kPrimFloat: + case Primitive::kPrimDouble: + GenerateFpCompareAndBranch(if_cond, condition->IsGtBias(), type, locations, branch_target); + break; } } diff --git a/compiler/optimizing/code_generator_mips64.h b/compiler/optimizing/code_generator_mips64.h index 85e3a4a3ce..1593cec2a6 100644 --- a/compiler/optimizing/code_generator_mips64.h +++ b/compiler/optimizing/code_generator_mips64.h @@ -237,6 +237,16 @@ class InstructionCodeGeneratorMIPS64 : public HGraphVisitor { void DivRemByPowerOfTwo(HBinaryOperation* instruction); void GenerateDivRemWithAnyConstant(HBinaryOperation* instruction); void GenerateDivRemIntegral(HBinaryOperation* instruction); + void GenerateIntLongCompare(IfCondition cond, bool is64bit, LocationSummary* locations); + void GenerateIntLongCompareAndBranch(IfCondition cond, + bool is64bit, + LocationSummary* locations, + Mips64Label* label); + void GenerateFpCompareAndBranch(IfCondition cond, + bool gt_bias, + Primitive::Type type, + LocationSummary* locations, + Mips64Label* label); void HandleGoto(HInstruction* got, HBasicBlock* successor); Mips64Assembler* const assembler_; diff --git a/compiler/optimizing/instruction_simplifier.cc b/compiler/optimizing/instruction_simplifier.cc index 67097deaeb..c504ded54c 100644 --- a/compiler/optimizing/instruction_simplifier.cc +++ b/compiler/optimizing/instruction_simplifier.cc @@ -777,13 +777,6 @@ void InstructionSimplifierVisitor::VisitLessThanOrEqual(HLessThanOrEqual* condit void InstructionSimplifierVisitor::VisitCondition(HCondition* condition) { // Try to fold an HCompare into this HCondition. - // This simplification is currently supported on x86, x86_64, ARM and ARM64. - // TODO: Implement it for MIPS64. - InstructionSet instruction_set = GetGraph()->GetInstructionSet(); - if (instruction_set == kMips64) { - return; - } - HInstruction* left = condition->GetLeft(); HInstruction* right = condition->GetRight(); // We can only replace an HCondition which compares a Compare to 0. diff --git a/compiler/utils/mips64/assembler_mips64.cc b/compiler/utils/mips64/assembler_mips64.cc index 107d5bb572..cfd8421e93 100644 --- a/compiler/utils/mips64/assembler_mips64.cc +++ b/compiler/utils/mips64/assembler_mips64.cc @@ -616,6 +616,14 @@ void Mips64Assembler::Bnezc(GpuRegister rs, uint32_t imm21) { EmitI21(0x3E, rs, imm21); } +void Mips64Assembler::Bc1eqz(FpuRegister ft, uint16_t imm16) { + EmitFI(0x11, 0x9, ft, imm16); +} + +void Mips64Assembler::Bc1nez(FpuRegister ft, uint16_t imm16) { + EmitFI(0x11, 0xD, ft, imm16); +} + void Mips64Assembler::EmitBcondc(BranchCondition cond, GpuRegister rs, GpuRegister rt, @@ -669,6 +677,14 @@ void Mips64Assembler::EmitBcondc(BranchCondition cond, case kCondGEU: Bgeuc(rs, rt, imm16_21); break; + case kCondF: + CHECK_EQ(rt, ZERO); + Bc1eqz(static_cast<FpuRegister>(rs), imm16_21); + break; + case kCondT: + CHECK_EQ(rt, ZERO); + Bc1nez(static_cast<FpuRegister>(rs), imm16_21); + break; case kUncond: LOG(FATAL) << "Unexpected branch condition " << cond; UNREACHABLE(); @@ -827,6 +843,86 @@ void Mips64Assembler::MaxD(FpuRegister fd, FpuRegister fs, FpuRegister ft) { EmitFR(0x11, 0x11, ft, fs, fd, 0x1e); } +void Mips64Assembler::CmpUnS(FpuRegister fd, FpuRegister fs, FpuRegister ft) { + EmitFR(0x11, 0x14, ft, fs, fd, 0x01); +} + +void Mips64Assembler::CmpEqS(FpuRegister fd, FpuRegister fs, FpuRegister ft) { + EmitFR(0x11, 0x14, ft, fs, fd, 0x02); +} + +void Mips64Assembler::CmpUeqS(FpuRegister fd, FpuRegister fs, FpuRegister ft) { + EmitFR(0x11, 0x14, ft, fs, fd, 0x03); +} + +void Mips64Assembler::CmpLtS(FpuRegister fd, FpuRegister fs, FpuRegister ft) { + EmitFR(0x11, 0x14, ft, fs, fd, 0x04); +} + +void Mips64Assembler::CmpUltS(FpuRegister fd, FpuRegister fs, FpuRegister ft) { + EmitFR(0x11, 0x14, ft, fs, fd, 0x05); +} + +void Mips64Assembler::CmpLeS(FpuRegister fd, FpuRegister fs, FpuRegister ft) { + EmitFR(0x11, 0x14, ft, fs, fd, 0x06); +} + +void Mips64Assembler::CmpUleS(FpuRegister fd, FpuRegister fs, FpuRegister ft) { + EmitFR(0x11, 0x14, ft, fs, fd, 0x07); +} + +void Mips64Assembler::CmpOrS(FpuRegister fd, FpuRegister fs, FpuRegister ft) { + EmitFR(0x11, 0x14, ft, fs, fd, 0x11); +} + +void Mips64Assembler::CmpUneS(FpuRegister fd, FpuRegister fs, FpuRegister ft) { + EmitFR(0x11, 0x14, ft, fs, fd, 0x12); +} + +void Mips64Assembler::CmpNeS(FpuRegister fd, FpuRegister fs, FpuRegister ft) { + EmitFR(0x11, 0x14, ft, fs, fd, 0x13); +} + +void Mips64Assembler::CmpUnD(FpuRegister fd, FpuRegister fs, FpuRegister ft) { + EmitFR(0x11, 0x15, ft, fs, fd, 0x01); +} + +void Mips64Assembler::CmpEqD(FpuRegister fd, FpuRegister fs, FpuRegister ft) { + EmitFR(0x11, 0x15, ft, fs, fd, 0x02); +} + +void Mips64Assembler::CmpUeqD(FpuRegister fd, FpuRegister fs, FpuRegister ft) { + EmitFR(0x11, 0x15, ft, fs, fd, 0x03); +} + +void Mips64Assembler::CmpLtD(FpuRegister fd, FpuRegister fs, FpuRegister ft) { + EmitFR(0x11, 0x15, ft, fs, fd, 0x04); +} + +void Mips64Assembler::CmpUltD(FpuRegister fd, FpuRegister fs, FpuRegister ft) { + EmitFR(0x11, 0x15, ft, fs, fd, 0x05); +} + +void Mips64Assembler::CmpLeD(FpuRegister fd, FpuRegister fs, FpuRegister ft) { + EmitFR(0x11, 0x15, ft, fs, fd, 0x06); +} + +void Mips64Assembler::CmpUleD(FpuRegister fd, FpuRegister fs, FpuRegister ft) { + EmitFR(0x11, 0x15, ft, fs, fd, 0x07); +} + +void Mips64Assembler::CmpOrD(FpuRegister fd, FpuRegister fs, FpuRegister ft) { + EmitFR(0x11, 0x15, ft, fs, fd, 0x11); +} + +void Mips64Assembler::CmpUneD(FpuRegister fd, FpuRegister fs, FpuRegister ft) { + EmitFR(0x11, 0x15, ft, fs, fd, 0x12); +} + +void Mips64Assembler::CmpNeD(FpuRegister fd, FpuRegister fs, FpuRegister ft) { + EmitFR(0x11, 0x15, ft, fs, fd, 0x13); +} + void Mips64Assembler::Cvtsw(FpuRegister fd, FpuRegister fs) { EmitFR(0x11, 0x14, static_cast<FpuRegister>(0), fs, fd, 0x20); } @@ -1134,6 +1230,10 @@ Mips64Assembler::Branch::Branch(uint32_t location, CHECK_NE(lhs_reg, ZERO); CHECK_EQ(rhs_reg, ZERO); break; + case kCondF: + case kCondT: + CHECK_EQ(rhs_reg, ZERO); + break; case kUncond: UNREACHABLE(); } @@ -1188,6 +1288,10 @@ Mips64Assembler::BranchCondition Mips64Assembler::Branch::OppositeCondition( return kCondGEU; case kCondGEU: return kCondLTU; + case kCondF: + return kCondT; + case kCondT: + return kCondF; case kUncond: LOG(FATAL) << "Unexpected branch condition " << cond; } @@ -1567,7 +1671,7 @@ void Mips64Assembler::EmitBranch(Mips64Assembler::Branch* branch) { case Branch::kCondBranch: CHECK_EQ(overwrite_location_, branch->GetOffsetLocation()); EmitBcondc(condition, lhs, rhs, offset); - Nop(); // TODO: improve by filling the forbidden slot. + Nop(); // TODO: improve by filling the forbidden/delay slot. break; case Branch::kCall: CHECK_EQ(overwrite_location_, branch->GetOffsetLocation()); @@ -1657,6 +1761,14 @@ void Mips64Assembler::Bnezc(GpuRegister rs, Mips64Label* label) { Bcond(label, kCondNEZ, rs); } +void Mips64Assembler::Bc1eqz(FpuRegister ft, Mips64Label* label) { + Bcond(label, kCondF, static_cast<GpuRegister>(ft), ZERO); +} + +void Mips64Assembler::Bc1nez(FpuRegister ft, Mips64Label* label) { + Bcond(label, kCondT, static_cast<GpuRegister>(ft), ZERO); +} + void Mips64Assembler::LoadFromOffset(LoadOperandType type, GpuRegister reg, GpuRegister base, int32_t offset) { if (!IsInt<16>(offset)) { diff --git a/compiler/utils/mips64/assembler_mips64.h b/compiler/utils/mips64/assembler_mips64.h index 57fc19a6e9..883f013f87 100644 --- a/compiler/utils/mips64/assembler_mips64.h +++ b/compiler/utils/mips64/assembler_mips64.h @@ -227,6 +227,8 @@ class Mips64Assembler FINAL : public Assembler { void Bnec(GpuRegister rs, GpuRegister rt, uint16_t imm16); void Beqzc(GpuRegister rs, uint32_t imm21); void Bnezc(GpuRegister rs, uint32_t imm21); + void Bc1eqz(FpuRegister ft, uint16_t imm16); + void Bc1nez(FpuRegister ft, uint16_t imm16); void AddS(FpuRegister fd, FpuRegister fs, FpuRegister ft); void SubS(FpuRegister fd, FpuRegister fs, FpuRegister ft); @@ -266,6 +268,26 @@ class Mips64Assembler FINAL : public Assembler { void MinD(FpuRegister fd, FpuRegister fs, FpuRegister ft); void MaxS(FpuRegister fd, FpuRegister fs, FpuRegister ft); void MaxD(FpuRegister fd, FpuRegister fs, FpuRegister ft); + void CmpUnS(FpuRegister fd, FpuRegister fs, FpuRegister ft); + void CmpEqS(FpuRegister fd, FpuRegister fs, FpuRegister ft); + void CmpUeqS(FpuRegister fd, FpuRegister fs, FpuRegister ft); + void CmpLtS(FpuRegister fd, FpuRegister fs, FpuRegister ft); + void CmpUltS(FpuRegister fd, FpuRegister fs, FpuRegister ft); + void CmpLeS(FpuRegister fd, FpuRegister fs, FpuRegister ft); + void CmpUleS(FpuRegister fd, FpuRegister fs, FpuRegister ft); + void CmpOrS(FpuRegister fd, FpuRegister fs, FpuRegister ft); + void CmpUneS(FpuRegister fd, FpuRegister fs, FpuRegister ft); + void CmpNeS(FpuRegister fd, FpuRegister fs, FpuRegister ft); + void CmpUnD(FpuRegister fd, FpuRegister fs, FpuRegister ft); + void CmpEqD(FpuRegister fd, FpuRegister fs, FpuRegister ft); + void CmpUeqD(FpuRegister fd, FpuRegister fs, FpuRegister ft); + void CmpLtD(FpuRegister fd, FpuRegister fs, FpuRegister ft); + void CmpUltD(FpuRegister fd, FpuRegister fs, FpuRegister ft); + void CmpLeD(FpuRegister fd, FpuRegister fs, FpuRegister ft); + void CmpUleD(FpuRegister fd, FpuRegister fs, FpuRegister ft); + void CmpOrD(FpuRegister fd, FpuRegister fs, FpuRegister ft); + void CmpUneD(FpuRegister fd, FpuRegister fs, FpuRegister ft); + void CmpNeD(FpuRegister fd, FpuRegister fs, FpuRegister ft); void Cvtsw(FpuRegister fd, FpuRegister fs); void Cvtdw(FpuRegister fd, FpuRegister fs); @@ -317,6 +339,8 @@ class Mips64Assembler FINAL : public Assembler { void Bnec(GpuRegister rs, GpuRegister rt, Mips64Label* label); void Beqzc(GpuRegister rs, Mips64Label* label); void Bnezc(GpuRegister rs, Mips64Label* label); + void Bc1eqz(FpuRegister ft, Mips64Label* label); + void Bc1nez(FpuRegister ft, Mips64Label* label); void EmitLoad(ManagedRegister m_dst, GpuRegister src_register, int32_t src_offset, size_t size); void LoadFromOffset(LoadOperandType type, GpuRegister reg, GpuRegister base, int32_t offset); @@ -474,6 +498,8 @@ class Mips64Assembler FINAL : public Assembler { kCondNEZ, kCondLTU, kCondGEU, + kCondF, // Floating-point predicate false. + kCondT, // Floating-point predicate true. kUncond, }; friend std::ostream& operator<<(std::ostream& os, const BranchCondition& rhs); diff --git a/compiler/utils/mips64/assembler_mips64_test.cc b/compiler/utils/mips64/assembler_mips64_test.cc index 29a5a88316..bac4375b35 100644 --- a/compiler/utils/mips64/assembler_mips64_test.cc +++ b/compiler/utils/mips64/assembler_mips64_test.cc @@ -403,6 +403,106 @@ TEST_F(AssemblerMIPS64Test, MaxD) { DriverStr(RepeatFFF(&mips64::Mips64Assembler::MaxD, "max.d ${reg1}, ${reg2}, ${reg3}"), "max.d"); } +TEST_F(AssemblerMIPS64Test, CmpUnS) { + DriverStr(RepeatFFF(&mips64::Mips64Assembler::CmpUnS, "cmp.un.s ${reg1}, ${reg2}, ${reg3}"), + "cmp.un.s"); +} + +TEST_F(AssemblerMIPS64Test, CmpEqS) { + DriverStr(RepeatFFF(&mips64::Mips64Assembler::CmpEqS, "cmp.eq.s ${reg1}, ${reg2}, ${reg3}"), + "cmp.eq.s"); +} + +TEST_F(AssemblerMIPS64Test, CmpUeqS) { + DriverStr(RepeatFFF(&mips64::Mips64Assembler::CmpUeqS, "cmp.ueq.s ${reg1}, ${reg2}, ${reg3}"), + "cmp.ueq.s"); +} + +TEST_F(AssemblerMIPS64Test, CmpLtS) { + DriverStr(RepeatFFF(&mips64::Mips64Assembler::CmpLtS, "cmp.lt.s ${reg1}, ${reg2}, ${reg3}"), + "cmp.lt.s"); +} + +TEST_F(AssemblerMIPS64Test, CmpUltS) { + DriverStr(RepeatFFF(&mips64::Mips64Assembler::CmpUltS, "cmp.ult.s ${reg1}, ${reg2}, ${reg3}"), + "cmp.ult.s"); +} + +TEST_F(AssemblerMIPS64Test, CmpLeS) { + DriverStr(RepeatFFF(&mips64::Mips64Assembler::CmpLeS, "cmp.le.s ${reg1}, ${reg2}, ${reg3}"), + "cmp.le.s"); +} + +TEST_F(AssemblerMIPS64Test, CmpUleS) { + DriverStr(RepeatFFF(&mips64::Mips64Assembler::CmpUleS, "cmp.ule.s ${reg1}, ${reg2}, ${reg3}"), + "cmp.ule.s"); +} + +TEST_F(AssemblerMIPS64Test, CmpOrS) { + DriverStr(RepeatFFF(&mips64::Mips64Assembler::CmpOrS, "cmp.or.s ${reg1}, ${reg2}, ${reg3}"), + "cmp.or.s"); +} + +TEST_F(AssemblerMIPS64Test, CmpUneS) { + DriverStr(RepeatFFF(&mips64::Mips64Assembler::CmpUneS, "cmp.une.s ${reg1}, ${reg2}, ${reg3}"), + "cmp.une.s"); +} + +TEST_F(AssemblerMIPS64Test, CmpNeS) { + DriverStr(RepeatFFF(&mips64::Mips64Assembler::CmpNeS, "cmp.ne.s ${reg1}, ${reg2}, ${reg3}"), + "cmp.ne.s"); +} + +TEST_F(AssemblerMIPS64Test, CmpUnD) { + DriverStr(RepeatFFF(&mips64::Mips64Assembler::CmpUnD, "cmp.un.d ${reg1}, ${reg2}, ${reg3}"), + "cmp.un.d"); +} + +TEST_F(AssemblerMIPS64Test, CmpEqD) { + DriverStr(RepeatFFF(&mips64::Mips64Assembler::CmpEqD, "cmp.eq.d ${reg1}, ${reg2}, ${reg3}"), + "cmp.eq.d"); +} + +TEST_F(AssemblerMIPS64Test, CmpUeqD) { + DriverStr(RepeatFFF(&mips64::Mips64Assembler::CmpUeqD, "cmp.ueq.d ${reg1}, ${reg2}, ${reg3}"), + "cmp.ueq.d"); +} + +TEST_F(AssemblerMIPS64Test, CmpLtD) { + DriverStr(RepeatFFF(&mips64::Mips64Assembler::CmpLtD, "cmp.lt.d ${reg1}, ${reg2}, ${reg3}"), + "cmp.lt.d"); +} + +TEST_F(AssemblerMIPS64Test, CmpUltD) { + DriverStr(RepeatFFF(&mips64::Mips64Assembler::CmpUltD, "cmp.ult.d ${reg1}, ${reg2}, ${reg3}"), + "cmp.ult.d"); +} + +TEST_F(AssemblerMIPS64Test, CmpLeD) { + DriverStr(RepeatFFF(&mips64::Mips64Assembler::CmpLeD, "cmp.le.d ${reg1}, ${reg2}, ${reg3}"), + "cmp.le.d"); +} + +TEST_F(AssemblerMIPS64Test, CmpUleD) { + DriverStr(RepeatFFF(&mips64::Mips64Assembler::CmpUleD, "cmp.ule.d ${reg1}, ${reg2}, ${reg3}"), + "cmp.ule.d"); +} + +TEST_F(AssemblerMIPS64Test, CmpOrD) { + DriverStr(RepeatFFF(&mips64::Mips64Assembler::CmpOrD, "cmp.or.d ${reg1}, ${reg2}, ${reg3}"), + "cmp.or.d"); +} + +TEST_F(AssemblerMIPS64Test, CmpUneD) { + DriverStr(RepeatFFF(&mips64::Mips64Assembler::CmpUneD, "cmp.une.d ${reg1}, ${reg2}, ${reg3}"), + "cmp.une.d"); +} + +TEST_F(AssemblerMIPS64Test, CmpNeD) { + DriverStr(RepeatFFF(&mips64::Mips64Assembler::CmpNeD, "cmp.ne.d ${reg1}, ${reg2}, ${reg3}"), + "cmp.ne.d"); +} + TEST_F(AssemblerMIPS64Test, CvtDL) { DriverStr(RepeatFF(&mips64::Mips64Assembler::Cvtdl, "cvt.d.l ${reg1}, ${reg2}"), "cvt.d.l"); } @@ -591,6 +691,58 @@ TEST_F(AssemblerMIPS64Test, Bgeuc) { BranchCondTwoRegsHelper(&mips64::Mips64Assembler::Bgeuc, "Bgeuc"); } +TEST_F(AssemblerMIPS64Test, Bc1eqz) { + mips64::Mips64Label label; + __ Bc1eqz(mips64::F0, &label); + constexpr size_t kAdduCount1 = 63; + for (size_t i = 0; i != kAdduCount1; ++i) { + __ Addu(mips64::ZERO, mips64::ZERO, mips64::ZERO); + } + __ Bind(&label); + constexpr size_t kAdduCount2 = 64; + for (size_t i = 0; i != kAdduCount2; ++i) { + __ Addu(mips64::ZERO, mips64::ZERO, mips64::ZERO); + } + __ Bc1eqz(mips64::F31, &label); + + std::string expected = + ".set noreorder\n" + "bc1eqz $f0, 1f\n" + "nop\n" + + RepeatInsn(kAdduCount1, "addu $zero, $zero, $zero\n") + + "1:\n" + + RepeatInsn(kAdduCount2, "addu $zero, $zero, $zero\n") + + "bc1eqz $f31, 1b\n" + "nop\n"; + DriverStr(expected, "Bc1eqz"); +} + +TEST_F(AssemblerMIPS64Test, Bc1nez) { + mips64::Mips64Label label; + __ Bc1nez(mips64::F0, &label); + constexpr size_t kAdduCount1 = 63; + for (size_t i = 0; i != kAdduCount1; ++i) { + __ Addu(mips64::ZERO, mips64::ZERO, mips64::ZERO); + } + __ Bind(&label); + constexpr size_t kAdduCount2 = 64; + for (size_t i = 0; i != kAdduCount2; ++i) { + __ Addu(mips64::ZERO, mips64::ZERO, mips64::ZERO); + } + __ Bc1nez(mips64::F31, &label); + + std::string expected = + ".set noreorder\n" + "bc1nez $f0, 1f\n" + "nop\n" + + RepeatInsn(kAdduCount1, "addu $zero, $zero, $zero\n") + + "1:\n" + + RepeatInsn(kAdduCount2, "addu $zero, $zero, $zero\n") + + "bc1nez $f31, 1b\n" + "nop\n"; + DriverStr(expected, "Bc1nez"); +} + TEST_F(AssemblerMIPS64Test, LongBeqc) { mips64::Mips64Label label; __ Beqc(mips64::A0, mips64::A1, &label); |