| OLD | NEW |
| 1 //===- subzero/src/IceTargetLoweringX8632.cpp - x86-32 lowering -----------===// | 1 //===- subzero/src/IceTargetLoweringX8632.cpp - x86-32 lowering -----------===// |
| 2 // | 2 // |
| 3 // The Subzero Code Generator | 3 // The Subzero Code Generator |
| 4 // | 4 // |
| 5 // This file is distributed under the University of Illinois Open Source | 5 // This file is distributed under the University of Illinois Open Source |
| 6 // License. See LICENSE.TXT for details. | 6 // License. See LICENSE.TXT for details. |
| 7 // | 7 // |
| 8 //===----------------------------------------------------------------------===// | 8 //===----------------------------------------------------------------------===// |
| 9 // | 9 // |
| 10 // This file implements the TargetLoweringX8632 class, which | 10 // This file implements the TargetLoweringX8632 class, which |
| (...skipping 494 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 505 if (!hasFramePointer()) | 505 if (!hasFramePointer()) |
| 506 Offset += getStackAdjustment(); | 506 Offset += getStackAdjustment(); |
| 507 if (Offset) { | 507 if (Offset) { |
| 508 if (Offset > 0) | 508 if (Offset > 0) |
| 509 Str << "+"; | 509 Str << "+"; |
| 510 Str << Offset; | 510 Str << Offset; |
| 511 } | 511 } |
| 512 Str << "]"; | 512 Str << "]"; |
| 513 } | 513 } |
| 514 | 514 |
| 515 x86::Address TargetX8632::stackVarToAsmOperand(const Variable *Var) const { |
| 516 assert(!Var->hasReg()); |
| 517 int32_t Offset = Var->getStackOffset(); |
| 518 if (!hasFramePointer()) |
| 519 Offset += getStackAdjustment(); |
| 520 return x86::Address(x86::Register(getFrameOrStackReg()), Offset); |
| 521 } |
| 522 |
| 515 void TargetX8632::lowerArguments() { | 523 void TargetX8632::lowerArguments() { |
| 516 VarList &Args = Func->getArgs(); | 524 VarList &Args = Func->getArgs(); |
| 517 // The first four arguments of vector type, regardless of their | 525 // The first four arguments of vector type, regardless of their |
| 518 // position relative to the other arguments in the argument list, are | 526 // position relative to the other arguments in the argument list, are |
| 519 // passed in registers xmm0 - xmm3. | 527 // passed in registers xmm0 - xmm3. |
| 520 unsigned NumXmmArgs = 0; | 528 unsigned NumXmmArgs = 0; |
| 521 | 529 |
| 522 Context.init(Func->getEntryNode()); | 530 Context.init(Func->getEntryNode()); |
| 523 Context.setInsertPoint(Context.getCur()); | 531 Context.setInsertPoint(Context.getCur()); |
| 524 | 532 |
| (...skipping 68 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 593 } | 601 } |
| 594 if (isVectorType(Ty)) { | 602 if (isVectorType(Ty)) { |
| 595 InArgsSizeBytes = applyStackAlignment(InArgsSizeBytes); | 603 InArgsSizeBytes = applyStackAlignment(InArgsSizeBytes); |
| 596 } | 604 } |
| 597 Arg->setStackOffset(BasicFrameOffset + InArgsSizeBytes); | 605 Arg->setStackOffset(BasicFrameOffset + InArgsSizeBytes); |
| 598 InArgsSizeBytes += typeWidthInBytesOnStack(Ty); | 606 InArgsSizeBytes += typeWidthInBytesOnStack(Ty); |
| 599 if (Arg->hasReg()) { | 607 if (Arg->hasReg()) { |
| 600 assert(Ty != IceType_i64); | 608 assert(Ty != IceType_i64); |
| 601 OperandX8632Mem *Mem = OperandX8632Mem::create( | 609 OperandX8632Mem *Mem = OperandX8632Mem::create( |
| 602 Func, Ty, FramePtr, | 610 Func, Ty, FramePtr, |
| 603 Ctx->getConstantInt(IceType_i32, Arg->getStackOffset())); | 611 Ctx->getConstantInt32(IceType_i32, Arg->getStackOffset())); |
| 604 if (isVectorType(Arg->getType())) { | 612 if (isVectorType(Arg->getType())) { |
| 605 _movp(Arg, Mem); | 613 _movp(Arg, Mem); |
| 606 } else { | 614 } else { |
| 607 _mov(Arg, Mem); | 615 _mov(Arg, Mem); |
| 608 } | 616 } |
| 609 } | 617 } |
| 610 } | 618 } |
| 611 | 619 |
| 612 Type TargetX8632::stackSlotType() { return IceType_i32; } | 620 Type TargetX8632::stackSlotType() { return IceType_i32; } |
| 613 | 621 |
| (...skipping 178 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 792 // Align esp if necessary. | 800 // Align esp if necessary. |
| 793 if (NeedsStackAlignment) { | 801 if (NeedsStackAlignment) { |
| 794 uint32_t StackOffset = X86_RET_IP_SIZE_BYTES + PreservedRegsSizeBytes; | 802 uint32_t StackOffset = X86_RET_IP_SIZE_BYTES + PreservedRegsSizeBytes; |
| 795 uint32_t StackSize = applyStackAlignment(StackOffset + SpillAreaSizeBytes); | 803 uint32_t StackSize = applyStackAlignment(StackOffset + SpillAreaSizeBytes); |
| 796 SpillAreaSizeBytes = StackSize - StackOffset; | 804 SpillAreaSizeBytes = StackSize - StackOffset; |
| 797 } | 805 } |
| 798 | 806 |
| 799 // Generate "sub esp, SpillAreaSizeBytes" | 807 // Generate "sub esp, SpillAreaSizeBytes" |
| 800 if (SpillAreaSizeBytes) | 808 if (SpillAreaSizeBytes) |
| 801 _sub(getPhysicalRegister(Reg_esp), | 809 _sub(getPhysicalRegister(Reg_esp), |
| 802 Ctx->getConstantInt(IceType_i32, SpillAreaSizeBytes)); | 810 Ctx->getConstantInt32(IceType_i32, SpillAreaSizeBytes)); |
| 803 | 811 |
| 804 resetStackAdjustment(); | 812 resetStackAdjustment(); |
| 805 | 813 |
| 806 // Fill in stack offsets for stack args, and copy args into registers | 814 // Fill in stack offsets for stack args, and copy args into registers |
| 807 // for those that were register-allocated. Args are pushed right to | 815 // for those that were register-allocated. Args are pushed right to |
| 808 // left, so Arg[0] is closest to the stack/frame pointer. | 816 // left, so Arg[0] is closest to the stack/frame pointer. |
| 809 Variable *FramePtr = getPhysicalRegister(getFrameOrStackReg()); | 817 Variable *FramePtr = getPhysicalRegister(getFrameOrStackReg()); |
| 810 size_t BasicFrameOffset = PreservedRegsSizeBytes + X86_RET_IP_SIZE_BYTES; | 818 size_t BasicFrameOffset = PreservedRegsSizeBytes + X86_RET_IP_SIZE_BYTES; |
| 811 if (!IsEbpBasedFrame) | 819 if (!IsEbpBasedFrame) |
| 812 BasicFrameOffset += SpillAreaSizeBytes; | 820 BasicFrameOffset += SpillAreaSizeBytes; |
| (...skipping 95 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 908 Context.setInsertPoint(InsertPoint); | 916 Context.setInsertPoint(InsertPoint); |
| 909 | 917 |
| 910 Variable *esp = getPhysicalRegister(Reg_esp); | 918 Variable *esp = getPhysicalRegister(Reg_esp); |
| 911 if (IsEbpBasedFrame) { | 919 if (IsEbpBasedFrame) { |
| 912 Variable *ebp = getPhysicalRegister(Reg_ebp); | 920 Variable *ebp = getPhysicalRegister(Reg_ebp); |
| 913 _mov(esp, ebp); | 921 _mov(esp, ebp); |
| 914 _pop(ebp); | 922 _pop(ebp); |
| 915 } else { | 923 } else { |
| 916 // add esp, SpillAreaSizeBytes | 924 // add esp, SpillAreaSizeBytes |
| 917 if (SpillAreaSizeBytes) | 925 if (SpillAreaSizeBytes) |
| 918 _add(esp, Ctx->getConstantInt(IceType_i32, SpillAreaSizeBytes)); | 926 _add(esp, Ctx->getConstantInt32(IceType_i32, SpillAreaSizeBytes)); |
| 919 } | 927 } |
| 920 | 928 |
| 921 // Add pop instructions for preserved registers. | 929 // Add pop instructions for preserved registers. |
| 922 llvm::SmallBitVector CalleeSaves = | 930 llvm::SmallBitVector CalleeSaves = |
| 923 getRegisterSet(RegSet_CalleeSave, RegSet_None); | 931 getRegisterSet(RegSet_CalleeSave, RegSet_None); |
| 924 for (SizeT i = 0; i < CalleeSaves.size(); ++i) { | 932 for (SizeT i = 0; i < CalleeSaves.size(); ++i) { |
| 925 SizeT j = CalleeSaves.size() - i - 1; | 933 SizeT j = CalleeSaves.size() - i - 1; |
| 926 if (j == Reg_ebp && IsEbpBasedFrame) | 934 if (j == Reg_ebp && IsEbpBasedFrame) |
| 927 continue; | 935 continue; |
| 928 if (CalleeSaves[j] && RegsUsed[j]) { | 936 if (CalleeSaves[j] && RegsUsed[j]) { |
| (...skipping 97 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 1026 } | 1034 } |
| 1027 | 1035 |
| 1028 Operand *TargetX8632::loOperand(Operand *Operand) { | 1036 Operand *TargetX8632::loOperand(Operand *Operand) { |
| 1029 assert(Operand->getType() == IceType_i64); | 1037 assert(Operand->getType() == IceType_i64); |
| 1030 if (Operand->getType() != IceType_i64) | 1038 if (Operand->getType() != IceType_i64) |
| 1031 return Operand; | 1039 return Operand; |
| 1032 if (Variable *Var = llvm::dyn_cast<Variable>(Operand)) { | 1040 if (Variable *Var = llvm::dyn_cast<Variable>(Operand)) { |
| 1033 split64(Var); | 1041 split64(Var); |
| 1034 return Var->getLo(); | 1042 return Var->getLo(); |
| 1035 } | 1043 } |
| 1036 if (ConstantInteger *Const = llvm::dyn_cast<ConstantInteger>(Operand)) { | 1044 if (ConstantInteger64 *Const = llvm::dyn_cast<ConstantInteger64>(Operand)) { |
| 1037 uint64_t Mask = (1ull << 32) - 1; | 1045 return Ctx->getConstantInt32(IceType_i32, |
| 1038 return Ctx->getConstantInt(IceType_i32, Const->getValue() & Mask); | 1046 static_cast<uint32_t>(Const->getValue())); |
| 1039 } | 1047 } |
| 1040 if (OperandX8632Mem *Mem = llvm::dyn_cast<OperandX8632Mem>(Operand)) { | 1048 if (OperandX8632Mem *Mem = llvm::dyn_cast<OperandX8632Mem>(Operand)) { |
| 1041 return OperandX8632Mem::create(Func, IceType_i32, Mem->getBase(), | 1049 return OperandX8632Mem::create(Func, IceType_i32, Mem->getBase(), |
| 1042 Mem->getOffset(), Mem->getIndex(), | 1050 Mem->getOffset(), Mem->getIndex(), |
| 1043 Mem->getShift(), Mem->getSegmentRegister()); | 1051 Mem->getShift(), Mem->getSegmentRegister()); |
| 1044 } | 1052 } |
| 1045 llvm_unreachable("Unsupported operand type"); | 1053 llvm_unreachable("Unsupported operand type"); |
| 1046 return NULL; | 1054 return NULL; |
| 1047 } | 1055 } |
| 1048 | 1056 |
| 1049 Operand *TargetX8632::hiOperand(Operand *Operand) { | 1057 Operand *TargetX8632::hiOperand(Operand *Operand) { |
| 1050 assert(Operand->getType() == IceType_i64); | 1058 assert(Operand->getType() == IceType_i64); |
| 1051 if (Operand->getType() != IceType_i64) | 1059 if (Operand->getType() != IceType_i64) |
| 1052 return Operand; | 1060 return Operand; |
| 1053 if (Variable *Var = llvm::dyn_cast<Variable>(Operand)) { | 1061 if (Variable *Var = llvm::dyn_cast<Variable>(Operand)) { |
| 1054 split64(Var); | 1062 split64(Var); |
| 1055 return Var->getHi(); | 1063 return Var->getHi(); |
| 1056 } | 1064 } |
| 1057 if (ConstantInteger *Const = llvm::dyn_cast<ConstantInteger>(Operand)) { | 1065 if (ConstantInteger64 *Const = llvm::dyn_cast<ConstantInteger64>(Operand)) { |
| 1058 return Ctx->getConstantInt(IceType_i32, Const->getValue() >> 32); | 1066 return Ctx->getConstantInt32( |
| 1067 IceType_i32, static_cast<uint32_t>(Const->getValue() >> 32)); |
| 1059 } | 1068 } |
| 1060 if (OperandX8632Mem *Mem = llvm::dyn_cast<OperandX8632Mem>(Operand)) { | 1069 if (OperandX8632Mem *Mem = llvm::dyn_cast<OperandX8632Mem>(Operand)) { |
| 1061 Constant *Offset = Mem->getOffset(); | 1070 Constant *Offset = Mem->getOffset(); |
| 1062 if (Offset == NULL) | 1071 if (Offset == NULL) |
| 1063 Offset = Ctx->getConstantInt(IceType_i32, 4); | 1072 Offset = Ctx->getConstantInt32(IceType_i32, 4); |
| 1064 else if (ConstantInteger *IntOffset = | 1073 else if (ConstantInteger32 *IntOffset = |
| 1065 llvm::dyn_cast<ConstantInteger>(Offset)) { | 1074 llvm::dyn_cast<ConstantInteger32>(Offset)) { |
| 1066 Offset = Ctx->getConstantInt(IceType_i32, 4 + IntOffset->getValue()); | 1075 Offset = Ctx->getConstantInt32(IceType_i32, 4 + IntOffset->getValue()); |
| 1067 } else if (ConstantRelocatable *SymOffset = | 1076 } else if (ConstantRelocatable *SymOffset = |
| 1068 llvm::dyn_cast<ConstantRelocatable>(Offset)) { | 1077 llvm::dyn_cast<ConstantRelocatable>(Offset)) { |
| 1069 Offset = Ctx->getConstantSym(IceType_i32, 4 + SymOffset->getOffset(), | 1078 Offset = Ctx->getConstantSym(IceType_i32, 4 + SymOffset->getOffset(), |
| 1070 SymOffset->getName()); | 1079 SymOffset->getName()); |
| 1071 } | 1080 } |
| 1072 return OperandX8632Mem::create(Func, IceType_i32, Mem->getBase(), Offset, | 1081 return OperandX8632Mem::create(Func, IceType_i32, Mem->getBase(), Offset, |
| 1073 Mem->getIndex(), Mem->getShift(), | 1082 Mem->getIndex(), Mem->getShift(), |
| 1074 Mem->getSegmentRegister()); | 1083 Mem->getSegmentRegister()); |
| 1075 } | 1084 } |
| 1076 llvm_unreachable("Unsupported operand type"); | 1085 llvm_unreachable("Unsupported operand type"); |
| (...skipping 47 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 1124 // For default align=0, set it to the real value 1, to avoid any | 1133 // For default align=0, set it to the real value 1, to avoid any |
| 1125 // bit-manipulation problems below. | 1134 // bit-manipulation problems below. |
| 1126 AlignmentParam = std::max(AlignmentParam, 1u); | 1135 AlignmentParam = std::max(AlignmentParam, 1u); |
| 1127 | 1136 |
| 1128 // LLVM enforces power of 2 alignment. | 1137 // LLVM enforces power of 2 alignment. |
| 1129 assert((AlignmentParam & (AlignmentParam - 1)) == 0); | 1138 assert((AlignmentParam & (AlignmentParam - 1)) == 0); |
| 1130 assert((X86_STACK_ALIGNMENT_BYTES & (X86_STACK_ALIGNMENT_BYTES - 1)) == 0); | 1139 assert((X86_STACK_ALIGNMENT_BYTES & (X86_STACK_ALIGNMENT_BYTES - 1)) == 0); |
| 1131 | 1140 |
| 1132 uint32_t Alignment = std::max(AlignmentParam, X86_STACK_ALIGNMENT_BYTES); | 1141 uint32_t Alignment = std::max(AlignmentParam, X86_STACK_ALIGNMENT_BYTES); |
| 1133 if (Alignment > X86_STACK_ALIGNMENT_BYTES) { | 1142 if (Alignment > X86_STACK_ALIGNMENT_BYTES) { |
| 1134 _and(esp, Ctx->getConstantInt(IceType_i32, -Alignment)); | 1143 _and(esp, Ctx->getConstantInt32(IceType_i32, -Alignment)); |
| 1135 } | 1144 } |
| 1136 if (ConstantInteger *ConstantTotalSize = | 1145 if (ConstantInteger32 *ConstantTotalSize = |
| 1137 llvm::dyn_cast<ConstantInteger>(TotalSize)) { | 1146 llvm::dyn_cast<ConstantInteger32>(TotalSize)) { |
| 1138 uint32_t Value = ConstantTotalSize->getValue(); | 1147 uint32_t Value = ConstantTotalSize->getValue(); |
| 1139 Value = applyAlignment(Value, Alignment); | 1148 Value = applyAlignment(Value, Alignment); |
| 1140 _sub(esp, Ctx->getConstantInt(IceType_i32, Value)); | 1149 _sub(esp, Ctx->getConstantInt32(IceType_i32, Value)); |
| 1141 } else { | 1150 } else { |
| 1142 // Non-constant sizes need to be adjusted to the next highest | 1151 // Non-constant sizes need to be adjusted to the next highest |
| 1143 // multiple of the required alignment at runtime. | 1152 // multiple of the required alignment at runtime. |
| 1144 Variable *T = makeReg(IceType_i32); | 1153 Variable *T = makeReg(IceType_i32); |
| 1145 _mov(T, TotalSize); | 1154 _mov(T, TotalSize); |
| 1146 _add(T, Ctx->getConstantInt(IceType_i32, Alignment - 1)); | 1155 _add(T, Ctx->getConstantInt32(IceType_i32, Alignment - 1)); |
| 1147 _and(T, Ctx->getConstantInt(IceType_i32, -Alignment)); | 1156 _and(T, Ctx->getConstantInt32(IceType_i32, -Alignment)); |
| 1148 _sub(esp, T); | 1157 _sub(esp, T); |
| 1149 } | 1158 } |
| 1150 _mov(Dest, esp); | 1159 _mov(Dest, esp); |
| 1151 } | 1160 } |
| 1152 | 1161 |
| 1153 void TargetX8632::lowerArithmetic(const InstArithmetic *Inst) { | 1162 void TargetX8632::lowerArithmetic(const InstArithmetic *Inst) { |
| 1154 Variable *Dest = Inst->getDest(); | 1163 Variable *Dest = Inst->getDest(); |
| 1155 Operand *Src0 = legalize(Inst->getSrc(0)); | 1164 Operand *Src0 = legalize(Inst->getSrc(0)); |
| 1156 Operand *Src1 = legalize(Inst->getSrc(1)); | 1165 Operand *Src1 = legalize(Inst->getSrc(1)); |
| 1157 if (Dest->getType() == IceType_i64) { | 1166 if (Dest->getType() == IceType_i64) { |
| (...skipping 89 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 1247 // t2 = shl t2, t1 | 1256 // t2 = shl t2, t1 |
| 1248 // test t1, 0x20 | 1257 // test t1, 0x20 |
| 1249 // je L1 | 1258 // je L1 |
| 1250 // use(t3) | 1259 // use(t3) |
| 1251 // t3 = t2 | 1260 // t3 = t2 |
| 1252 // t2 = 0 | 1261 // t2 = 0 |
| 1253 // L1: | 1262 // L1: |
| 1254 // a.lo = t2 | 1263 // a.lo = t2 |
| 1255 // a.hi = t3 | 1264 // a.hi = t3 |
| 1256 Variable *T_1 = NULL, *T_2 = NULL, *T_3 = NULL; | 1265 Variable *T_1 = NULL, *T_2 = NULL, *T_3 = NULL; |
| 1257 Constant *BitTest = Ctx->getConstantInt(IceType_i32, 0x20); | 1266 Constant *BitTest = Ctx->getConstantInt32(IceType_i32, 0x20); |
| 1258 Constant *Zero = Ctx->getConstantZero(IceType_i32); | 1267 Constant *Zero = Ctx->getConstantZero(IceType_i32); |
| 1259 InstX8632Label *Label = InstX8632Label::create(Func, this); | 1268 InstX8632Label *Label = InstX8632Label::create(Func, this); |
| 1260 _mov(T_1, Src1Lo, Reg_ecx); | 1269 _mov(T_1, Src1Lo, Reg_ecx); |
| 1261 _mov(T_2, Src0Lo); | 1270 _mov(T_2, Src0Lo); |
| 1262 _mov(T_3, Src0Hi); | 1271 _mov(T_3, Src0Hi); |
| 1263 _shld(T_3, T_2, T_1); | 1272 _shld(T_3, T_2, T_1); |
| 1264 _shl(T_2, T_1); | 1273 _shl(T_2, T_1); |
| 1265 _test(T_1, BitTest); | 1274 _test(T_1, BitTest); |
| 1266 _br(InstX8632Br::Br_e, Label); | 1275 _br(InstX8632Br::Br_e, Label); |
| 1267 // Because of the intra-block control flow, we need to fake a use | 1276 // Because of the intra-block control flow, we need to fake a use |
| (...skipping 15 matching lines...) Expand all Loading... |
| 1283 // t3 = shr t3, t1 | 1292 // t3 = shr t3, t1 |
| 1284 // test t1, 0x20 | 1293 // test t1, 0x20 |
| 1285 // je L1 | 1294 // je L1 |
| 1286 // use(t2) | 1295 // use(t2) |
| 1287 // t2 = t3 | 1296 // t2 = t3 |
| 1288 // t3 = 0 | 1297 // t3 = 0 |
| 1289 // L1: | 1298 // L1: |
| 1290 // a.lo = t2 | 1299 // a.lo = t2 |
| 1291 // a.hi = t3 | 1300 // a.hi = t3 |
| 1292 Variable *T_1 = NULL, *T_2 = NULL, *T_3 = NULL; | 1301 Variable *T_1 = NULL, *T_2 = NULL, *T_3 = NULL; |
| 1293 Constant *BitTest = Ctx->getConstantInt(IceType_i32, 0x20); | 1302 Constant *BitTest = Ctx->getConstantInt32(IceType_i32, 0x20); |
| 1294 Constant *Zero = Ctx->getConstantZero(IceType_i32); | 1303 Constant *Zero = Ctx->getConstantZero(IceType_i32); |
| 1295 InstX8632Label *Label = InstX8632Label::create(Func, this); | 1304 InstX8632Label *Label = InstX8632Label::create(Func, this); |
| 1296 _mov(T_1, Src1Lo, Reg_ecx); | 1305 _mov(T_1, Src1Lo, Reg_ecx); |
| 1297 _mov(T_2, Src0Lo); | 1306 _mov(T_2, Src0Lo); |
| 1298 _mov(T_3, Src0Hi); | 1307 _mov(T_3, Src0Hi); |
| 1299 _shrd(T_2, T_3, T_1); | 1308 _shrd(T_2, T_3, T_1); |
| 1300 _shr(T_3, T_1); | 1309 _shr(T_3, T_1); |
| 1301 _test(T_1, BitTest); | 1310 _test(T_1, BitTest); |
| 1302 _br(InstX8632Br::Br_e, Label); | 1311 _br(InstX8632Br::Br_e, Label); |
| 1303 // Because of the intra-block control flow, we need to fake a use | 1312 // Because of the intra-block control flow, we need to fake a use |
| (...skipping 15 matching lines...) Expand all Loading... |
| 1319 // t3 = sar t3, t1 | 1328 // t3 = sar t3, t1 |
| 1320 // test t1, 0x20 | 1329 // test t1, 0x20 |
| 1321 // je L1 | 1330 // je L1 |
| 1322 // use(t2) | 1331 // use(t2) |
| 1323 // t2 = t3 | 1332 // t2 = t3 |
| 1324 // t3 = sar t3, 0x1f | 1333 // t3 = sar t3, 0x1f |
| 1325 // L1: | 1334 // L1: |
| 1326 // a.lo = t2 | 1335 // a.lo = t2 |
| 1327 // a.hi = t3 | 1336 // a.hi = t3 |
| 1328 Variable *T_1 = NULL, *T_2 = NULL, *T_3 = NULL; | 1337 Variable *T_1 = NULL, *T_2 = NULL, *T_3 = NULL; |
| 1329 Constant *BitTest = Ctx->getConstantInt(IceType_i32, 0x20); | 1338 Constant *BitTest = Ctx->getConstantInt32(IceType_i32, 0x20); |
| 1330 Constant *SignExtend = Ctx->getConstantInt(IceType_i32, 0x1f); | 1339 Constant *SignExtend = Ctx->getConstantInt32(IceType_i32, 0x1f); |
| 1331 InstX8632Label *Label = InstX8632Label::create(Func, this); | 1340 InstX8632Label *Label = InstX8632Label::create(Func, this); |
| 1332 _mov(T_1, Src1Lo, Reg_ecx); | 1341 _mov(T_1, Src1Lo, Reg_ecx); |
| 1333 _mov(T_2, Src0Lo); | 1342 _mov(T_2, Src0Lo); |
| 1334 _mov(T_3, Src0Hi); | 1343 _mov(T_3, Src0Hi); |
| 1335 _shrd(T_2, T_3, T_1); | 1344 _shrd(T_2, T_3, T_1); |
| 1336 _sar(T_3, T_1); | 1345 _sar(T_3, T_1); |
| 1337 _test(T_1, BitTest); | 1346 _test(T_1, BitTest); |
| 1338 _br(InstX8632Br::Br_e, Label); | 1347 _br(InstX8632Br::Br_e, Label); |
| 1339 // Because of the intra-block control flow, we need to fake a use | 1348 // Because of the intra-block control flow, we need to fake a use |
| 1340 // of T_3 to prevent its earlier definition from being dead-code | 1349 // of T_3 to prevent its earlier definition from being dead-code |
| (...skipping 100 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 1441 // # T2 = {Src0[1] * Src1[1], Src0[3] * Src1[3]} | 1450 // # T2 = {Src0[1] * Src1[1], Src0[3] * Src1[3]} |
| 1442 // pmuludq T2, T3 | 1451 // pmuludq T2, T3 |
| 1443 // # T1 = {lo(T1[0]), lo(T1[2]), lo(T2[0]), lo(T2[2])} | 1452 // # T1 = {lo(T1[0]), lo(T1[2]), lo(T2[0]), lo(T2[2])} |
| 1444 // shufps T1, T2, {0,2,0,2} | 1453 // shufps T1, T2, {0,2,0,2} |
| 1445 // pshufd T4, T1, {0,2,1,3} | 1454 // pshufd T4, T1, {0,2,1,3} |
| 1446 // movups Dest, T4 | 1455 // movups Dest, T4 |
| 1447 | 1456 |
| 1448 // Mask that directs pshufd to create a vector with entries | 1457 // Mask that directs pshufd to create a vector with entries |
| 1449 // Src[1, 0, 3, 0] | 1458 // Src[1, 0, 3, 0] |
| 1450 const unsigned Constant1030 = 0x31; | 1459 const unsigned Constant1030 = 0x31; |
| 1451 Constant *Mask1030 = Ctx->getConstantInt(IceType_i8, Constant1030); | 1460 Constant *Mask1030 = Ctx->getConstantInt32(IceType_i8, Constant1030); |
| 1452 // Mask that directs shufps to create a vector with entries | 1461 // Mask that directs shufps to create a vector with entries |
| 1453 // Dest[0, 2], Src[0, 2] | 1462 // Dest[0, 2], Src[0, 2] |
| 1454 const unsigned Mask0202 = 0x88; | 1463 const unsigned Mask0202 = 0x88; |
| 1455 // Mask that directs pshufd to create a vector with entries | 1464 // Mask that directs pshufd to create a vector with entries |
| 1456 // Src[0, 2, 1, 3] | 1465 // Src[0, 2, 1, 3] |
| 1457 const unsigned Mask0213 = 0xd8; | 1466 const unsigned Mask0213 = 0xd8; |
| 1458 Variable *T1 = makeReg(IceType_v4i32); | 1467 Variable *T1 = makeReg(IceType_v4i32); |
| 1459 Variable *T2 = makeReg(IceType_v4i32); | 1468 Variable *T2 = makeReg(IceType_v4i32); |
| 1460 Variable *T3 = makeReg(IceType_v4i32); | 1469 Variable *T3 = makeReg(IceType_v4i32); |
| 1461 Variable *T4 = makeReg(IceType_v4i32); | 1470 Variable *T4 = makeReg(IceType_v4i32); |
| 1462 _movp(T1, Src0); | 1471 _movp(T1, Src0); |
| 1463 _pshufd(T2, Src0, Mask1030); | 1472 _pshufd(T2, Src0, Mask1030); |
| 1464 _pshufd(T3, Src1, Mask1030); | 1473 _pshufd(T3, Src1, Mask1030); |
| 1465 _pmuludq(T1, Src1); | 1474 _pmuludq(T1, Src1); |
| 1466 _pmuludq(T2, T3); | 1475 _pmuludq(T2, T3); |
| 1467 _shufps(T1, T2, Ctx->getConstantInt(IceType_i8, Mask0202)); | 1476 _shufps(T1, T2, Ctx->getConstantInt32(IceType_i8, Mask0202)); |
| 1468 _pshufd(T4, T1, Ctx->getConstantInt(IceType_i8, Mask0213)); | 1477 _pshufd(T4, T1, Ctx->getConstantInt32(IceType_i8, Mask0213)); |
| 1469 _movp(Dest, T4); | 1478 _movp(Dest, T4); |
| 1470 } else { | 1479 } else { |
| 1471 assert(Dest->getType() == IceType_v16i8); | 1480 assert(Dest->getType() == IceType_v16i8); |
| 1472 scalarizeArithmetic(Inst->getOp(), Dest, Src0, Src1); | 1481 scalarizeArithmetic(Inst->getOp(), Dest, Src0, Src1); |
| 1473 } | 1482 } |
| 1474 } break; | 1483 } break; |
| 1475 case InstArithmetic::Shl: | 1484 case InstArithmetic::Shl: |
| 1476 case InstArithmetic::Lshr: | 1485 case InstArithmetic::Lshr: |
| 1477 case InstArithmetic::Ashr: | 1486 case InstArithmetic::Ashr: |
| 1478 case InstArithmetic::Udiv: | 1487 case InstArithmetic::Udiv: |
| (...skipping 272 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 1751 assert(Ty == IceType_i32 || Ty == IceType_f32 || Ty == IceType_i64 || | 1760 assert(Ty == IceType_i32 || Ty == IceType_f32 || Ty == IceType_i64 || |
| 1752 Ty == IceType_f64 || isVectorType(Ty)); | 1761 Ty == IceType_f64 || isVectorType(Ty)); |
| 1753 if (isVectorType(Ty) && XmmArgs.size() < X86_MAX_XMM_ARGS) { | 1762 if (isVectorType(Ty) && XmmArgs.size() < X86_MAX_XMM_ARGS) { |
| 1754 XmmArgs.push_back(Arg); | 1763 XmmArgs.push_back(Arg); |
| 1755 } else { | 1764 } else { |
| 1756 StackArgs.push_back(Arg); | 1765 StackArgs.push_back(Arg); |
| 1757 if (isVectorType(Arg->getType())) { | 1766 if (isVectorType(Arg->getType())) { |
| 1758 ParameterAreaSizeBytes = applyStackAlignment(ParameterAreaSizeBytes); | 1767 ParameterAreaSizeBytes = applyStackAlignment(ParameterAreaSizeBytes); |
| 1759 } | 1768 } |
| 1760 Variable *esp = Func->getTarget()->getPhysicalRegister(Reg_esp); | 1769 Variable *esp = Func->getTarget()->getPhysicalRegister(Reg_esp); |
| 1761 Constant *Loc = Ctx->getConstantInt(IceType_i32, ParameterAreaSizeBytes); | 1770 Constant *Loc = |
| 1771 Ctx->getConstantInt32(IceType_i32, ParameterAreaSizeBytes); |
| 1762 StackArgLocations.push_back(OperandX8632Mem::create(Func, Ty, esp, Loc)); | 1772 StackArgLocations.push_back(OperandX8632Mem::create(Func, Ty, esp, Loc)); |
| 1763 ParameterAreaSizeBytes += typeWidthInBytesOnStack(Arg->getType()); | 1773 ParameterAreaSizeBytes += typeWidthInBytesOnStack(Arg->getType()); |
| 1764 } | 1774 } |
| 1765 } | 1775 } |
| 1766 | 1776 |
| 1767 // Adjust the parameter area so that the stack is aligned. It is | 1777 // Adjust the parameter area so that the stack is aligned. It is |
| 1768 // assumed that the stack is already aligned at the start of the | 1778 // assumed that the stack is already aligned at the start of the |
| 1769 // calling sequence. | 1779 // calling sequence. |
| 1770 ParameterAreaSizeBytes = applyStackAlignment(ParameterAreaSizeBytes); | 1780 ParameterAreaSizeBytes = applyStackAlignment(ParameterAreaSizeBytes); |
| 1771 | 1781 |
| (...skipping 76 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 1848 Operand *CallTarget = legalize(Instr->getCallTarget(), Legal_All); | 1858 Operand *CallTarget = legalize(Instr->getCallTarget(), Legal_All); |
| 1849 Inst *NewCall = InstX8632Call::create(Func, ReturnReg, CallTarget); | 1859 Inst *NewCall = InstX8632Call::create(Func, ReturnReg, CallTarget); |
| 1850 Context.insert(NewCall); | 1860 Context.insert(NewCall); |
| 1851 if (ReturnRegHi) | 1861 if (ReturnRegHi) |
| 1852 Context.insert(InstFakeDef::create(Func, ReturnRegHi)); | 1862 Context.insert(InstFakeDef::create(Func, ReturnRegHi)); |
| 1853 | 1863 |
| 1854 // Add the appropriate offset to esp. The call instruction takes care | 1864 // Add the appropriate offset to esp. The call instruction takes care |
| 1855 // of resetting the stack offset during emission. | 1865 // of resetting the stack offset during emission. |
| 1856 if (ParameterAreaSizeBytes) { | 1866 if (ParameterAreaSizeBytes) { |
| 1857 Variable *esp = Func->getTarget()->getPhysicalRegister(Reg_esp); | 1867 Variable *esp = Func->getTarget()->getPhysicalRegister(Reg_esp); |
| 1858 _add(esp, Ctx->getConstantInt(IceType_i32, ParameterAreaSizeBytes)); | 1868 _add(esp, Ctx->getConstantInt32(IceType_i32, ParameterAreaSizeBytes)); |
| 1859 } | 1869 } |
| 1860 | 1870 |
| 1861 // Insert a register-kill pseudo instruction. | 1871 // Insert a register-kill pseudo instruction. |
| 1862 VarList KilledRegs; | 1872 VarList KilledRegs; |
| 1863 for (SizeT i = 0; i < ScratchRegs.size(); ++i) { | 1873 for (SizeT i = 0; i < ScratchRegs.size(); ++i) { |
| 1864 if (ScratchRegs[i]) | 1874 if (ScratchRegs[i]) |
| 1865 KilledRegs.push_back(Func->getTarget()->getPhysicalRegister(i)); | 1875 KilledRegs.push_back(Func->getTarget()->getPhysicalRegister(i)); |
| 1866 } | 1876 } |
| 1867 Context.insert(InstFakeKill::create(Func, KilledRegs, NewCall)); | 1877 Context.insert(InstFakeKill::create(Func, KilledRegs, NewCall)); |
| 1868 | 1878 |
| (...skipping 64 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 1933 Variable *T = makeReg(DestTy); | 1943 Variable *T = makeReg(DestTy); |
| 1934 _movp(T, Src0RM); | 1944 _movp(T, Src0RM); |
| 1935 _pand(T, OneMask); | 1945 _pand(T, OneMask); |
| 1936 Variable *Zeros = makeVectorOfZeros(Dest->getType()); | 1946 Variable *Zeros = makeVectorOfZeros(Dest->getType()); |
| 1937 _pcmpgt(T, Zeros); | 1947 _pcmpgt(T, Zeros); |
| 1938 _movp(Dest, T); | 1948 _movp(Dest, T); |
| 1939 } else { | 1949 } else { |
| 1940 // width = width(elty) - 1; dest = (src << width) >> width | 1950 // width = width(elty) - 1; dest = (src << width) >> width |
| 1941 SizeT ShiftAmount = | 1951 SizeT ShiftAmount = |
| 1942 X86_CHAR_BIT * typeWidthInBytes(typeElementType(DestTy)) - 1; | 1952 X86_CHAR_BIT * typeWidthInBytes(typeElementType(DestTy)) - 1; |
| 1943 Constant *ShiftConstant = Ctx->getConstantInt(IceType_i8, ShiftAmount); | 1953 Constant *ShiftConstant = |
| 1954 Ctx->getConstantInt32(IceType_i8, ShiftAmount); |
| 1944 Variable *T = makeReg(DestTy); | 1955 Variable *T = makeReg(DestTy); |
| 1945 _movp(T, Src0RM); | 1956 _movp(T, Src0RM); |
| 1946 _psll(T, ShiftConstant); | 1957 _psll(T, ShiftConstant); |
| 1947 _psra(T, ShiftConstant); | 1958 _psra(T, ShiftConstant); |
| 1948 _movp(Dest, T); | 1959 _movp(Dest, T); |
| 1949 } | 1960 } |
| 1950 } else if (Dest->getType() == IceType_i64) { | 1961 } else if (Dest->getType() == IceType_i64) { |
| 1951 // t1=movsx src; t2=t1; t2=sar t2, 31; dst.lo=t1; dst.hi=t2 | 1962 // t1=movsx src; t2=t1; t2=sar t2, 31; dst.lo=t1; dst.hi=t2 |
| 1952 Constant *Shift = Ctx->getConstantInt(IceType_i32, 31); | 1963 Constant *Shift = Ctx->getConstantInt32(IceType_i32, 31); |
| 1953 Variable *DestLo = llvm::cast<Variable>(loOperand(Dest)); | 1964 Variable *DestLo = llvm::cast<Variable>(loOperand(Dest)); |
| 1954 Variable *DestHi = llvm::cast<Variable>(hiOperand(Dest)); | 1965 Variable *DestHi = llvm::cast<Variable>(hiOperand(Dest)); |
| 1955 Variable *T_Lo = makeReg(DestLo->getType()); | 1966 Variable *T_Lo = makeReg(DestLo->getType()); |
| 1956 if (Src0RM->getType() == IceType_i32) { | 1967 if (Src0RM->getType() == IceType_i32) { |
| 1957 _mov(T_Lo, Src0RM); | 1968 _mov(T_Lo, Src0RM); |
| 1958 } else if (Src0RM->getType() == IceType_i1) { | 1969 } else if (Src0RM->getType() == IceType_i1) { |
| 1959 _movzx(T_Lo, Src0RM); | 1970 _movzx(T_Lo, Src0RM); |
| 1960 _shl(T_Lo, Shift); | 1971 _shl(T_Lo, Shift); |
| 1961 _sar(T_Lo, Shift); | 1972 _sar(T_Lo, Shift); |
| 1962 } else { | 1973 } else { |
| 1963 _movsx(T_Lo, Src0RM); | 1974 _movsx(T_Lo, Src0RM); |
| 1964 } | 1975 } |
| 1965 _mov(DestLo, T_Lo); | 1976 _mov(DestLo, T_Lo); |
| 1966 Variable *T_Hi = NULL; | 1977 Variable *T_Hi = NULL; |
| 1967 _mov(T_Hi, T_Lo); | 1978 _mov(T_Hi, T_Lo); |
| 1968 if (Src0RM->getType() != IceType_i1) | 1979 if (Src0RM->getType() != IceType_i1) |
| 1969 // For i1, the sar instruction is already done above. | 1980 // For i1, the sar instruction is already done above. |
| 1970 _sar(T_Hi, Shift); | 1981 _sar(T_Hi, Shift); |
| 1971 _mov(DestHi, T_Hi); | 1982 _mov(DestHi, T_Hi); |
| 1972 } else if (Src0RM->getType() == IceType_i1) { | 1983 } else if (Src0RM->getType() == IceType_i1) { |
| 1973 // t1 = src | 1984 // t1 = src |
| 1974 // shl t1, dst_bitwidth - 1 | 1985 // shl t1, dst_bitwidth - 1 |
| 1975 // sar t1, dst_bitwidth - 1 | 1986 // sar t1, dst_bitwidth - 1 |
| 1976 // dst = t1 | 1987 // dst = t1 |
| 1977 size_t DestBits = X86_CHAR_BIT * typeWidthInBytes(Dest->getType()); | 1988 size_t DestBits = X86_CHAR_BIT * typeWidthInBytes(Dest->getType()); |
| 1978 Constant *ShiftAmount = Ctx->getConstantInt(IceType_i32, DestBits - 1); | 1989 Constant *ShiftAmount = Ctx->getConstantInt32(IceType_i32, DestBits - 1); |
| 1979 Variable *T = makeReg(Dest->getType()); | 1990 Variable *T = makeReg(Dest->getType()); |
| 1980 if (typeWidthInBytes(Dest->getType()) <= | 1991 if (typeWidthInBytes(Dest->getType()) <= |
| 1981 typeWidthInBytes(Src0RM->getType())) { | 1992 typeWidthInBytes(Src0RM->getType())) { |
| 1982 _mov(T, Src0RM); | 1993 _mov(T, Src0RM); |
| 1983 } else { | 1994 } else { |
| 1984 // Widen the source using movsx or movzx. (It doesn't matter | 1995 // Widen the source using movsx or movzx. (It doesn't matter |
| 1985 // which one, since the following shl/sar overwrite the bits.) | 1996 // which one, since the following shl/sar overwrite the bits.) |
| 1986 _movzx(T, Src0RM); | 1997 _movzx(T, Src0RM); |
| 1987 } | 1998 } |
| 1988 _shl(T, ShiftAmount); | 1999 _shl(T, ShiftAmount); |
| (...skipping 22 matching lines...) Expand all Loading... |
| 2011 Constant *Zero = Ctx->getConstantZero(IceType_i32); | 2022 Constant *Zero = Ctx->getConstantZero(IceType_i32); |
| 2012 Variable *DestLo = llvm::cast<Variable>(loOperand(Dest)); | 2023 Variable *DestLo = llvm::cast<Variable>(loOperand(Dest)); |
| 2013 Variable *DestHi = llvm::cast<Variable>(hiOperand(Dest)); | 2024 Variable *DestHi = llvm::cast<Variable>(hiOperand(Dest)); |
| 2014 Variable *Tmp = makeReg(DestLo->getType()); | 2025 Variable *Tmp = makeReg(DestLo->getType()); |
| 2015 if (Src0RM->getType() == IceType_i32) { | 2026 if (Src0RM->getType() == IceType_i32) { |
| 2016 _mov(Tmp, Src0RM); | 2027 _mov(Tmp, Src0RM); |
| 2017 } else { | 2028 } else { |
| 2018 _movzx(Tmp, Src0RM); | 2029 _movzx(Tmp, Src0RM); |
| 2019 } | 2030 } |
| 2020 if (Src0RM->getType() == IceType_i1) { | 2031 if (Src0RM->getType() == IceType_i1) { |
| 2021 Constant *One = Ctx->getConstantInt(IceType_i32, 1); | 2032 Constant *One = Ctx->getConstantInt32(IceType_i32, 1); |
| 2022 _and(Tmp, One); | 2033 _and(Tmp, One); |
| 2023 } | 2034 } |
| 2024 _mov(DestLo, Tmp); | 2035 _mov(DestLo, Tmp); |
| 2025 _mov(DestHi, Zero); | 2036 _mov(DestHi, Zero); |
| 2026 } else if (Src0RM->getType() == IceType_i1) { | 2037 } else if (Src0RM->getType() == IceType_i1) { |
| 2027 // t = Src0RM; t &= 1; Dest = t | 2038 // t = Src0RM; t &= 1; Dest = t |
| 2028 Constant *One = Ctx->getConstantInt(IceType_i32, 1); | 2039 Constant *One = Ctx->getConstantInt32(IceType_i32, 1); |
| 2029 Variable *T = makeReg(IceType_i32); | 2040 Variable *T = makeReg(IceType_i32); |
| 2030 _movzx(T, Src0RM); | 2041 _movzx(T, Src0RM); |
| 2031 _and(T, One); | 2042 _and(T, One); |
| 2032 _mov(Dest, T); | 2043 _mov(Dest, T); |
| 2033 } else { | 2044 } else { |
| 2034 // t1 = movzx src; dst = t1 | 2045 // t1 = movzx src; dst = t1 |
| 2035 Variable *T = makeReg(Dest->getType()); | 2046 Variable *T = makeReg(Dest->getType()); |
| 2036 _movzx(T, Src0RM); | 2047 _movzx(T, Src0RM); |
| 2037 _mov(Dest, T); | 2048 _mov(Dest, T); |
| 2038 } | 2049 } |
| (...skipping 11 matching lines...) Expand all Loading... |
| 2050 _movp(Dest, T); | 2061 _movp(Dest, T); |
| 2051 } else { | 2062 } else { |
| 2052 Operand *Src0 = Inst->getSrc(0); | 2063 Operand *Src0 = Inst->getSrc(0); |
| 2053 if (Src0->getType() == IceType_i64) | 2064 if (Src0->getType() == IceType_i64) |
| 2054 Src0 = loOperand(Src0); | 2065 Src0 = loOperand(Src0); |
| 2055 Operand *Src0RM = legalize(Src0, Legal_Reg | Legal_Mem); | 2066 Operand *Src0RM = legalize(Src0, Legal_Reg | Legal_Mem); |
| 2056 // t1 = trunc Src0RM; Dest = t1 | 2067 // t1 = trunc Src0RM; Dest = t1 |
| 2057 Variable *T = NULL; | 2068 Variable *T = NULL; |
| 2058 _mov(T, Src0RM); | 2069 _mov(T, Src0RM); |
| 2059 if (Dest->getType() == IceType_i1) | 2070 if (Dest->getType() == IceType_i1) |
| 2060 _and(T, Ctx->getConstantInt(IceType_i1, 1)); | 2071 _and(T, Ctx->getConstantInt32(IceType_i1, 1)); |
| 2061 _mov(Dest, T); | 2072 _mov(Dest, T); |
| 2062 } | 2073 } |
| 2063 break; | 2074 break; |
| 2064 } | 2075 } |
| 2065 case InstCast::Fptrunc: | 2076 case InstCast::Fptrunc: |
| 2066 case InstCast::Fpext: { | 2077 case InstCast::Fpext: { |
| 2067 Operand *Src0RM = legalize(Inst->getSrc(0), Legal_Reg | Legal_Mem); | 2078 Operand *Src0RM = legalize(Inst->getSrc(0), Legal_Reg | Legal_Mem); |
| 2068 // t1 = cvt Src0RM; Dest = t1 | 2079 // t1 = cvt Src0RM; Dest = t1 |
| 2069 Variable *T = makeReg(Dest->getType()); | 2080 Variable *T = makeReg(Dest->getType()); |
| 2070 _cvt(T, Src0RM); | 2081 _cvt(T, Src0RM); |
| (...skipping 24 matching lines...) Expand all Loading... |
| 2095 Call->addArg(Inst->getSrc(0)); | 2106 Call->addArg(Inst->getSrc(0)); |
| 2096 lowerCall(Call); | 2107 lowerCall(Call); |
| 2097 } else { | 2108 } else { |
| 2098 Operand *Src0RM = legalize(Inst->getSrc(0), Legal_Reg | Legal_Mem); | 2109 Operand *Src0RM = legalize(Inst->getSrc(0), Legal_Reg | Legal_Mem); |
| 2099 // t1.i32 = cvt Src0RM; t2.dest_type = t1; Dest = t2.dest_type | 2110 // t1.i32 = cvt Src0RM; t2.dest_type = t1; Dest = t2.dest_type |
| 2100 Variable *T_1 = makeReg(IceType_i32); | 2111 Variable *T_1 = makeReg(IceType_i32); |
| 2101 Variable *T_2 = makeReg(Dest->getType()); | 2112 Variable *T_2 = makeReg(Dest->getType()); |
| 2102 _cvtt(T_1, Src0RM); | 2113 _cvtt(T_1, Src0RM); |
| 2103 _mov(T_2, T_1); // T_1 and T_2 may have different integer types | 2114 _mov(T_2, T_1); // T_1 and T_2 may have different integer types |
| 2104 if (Dest->getType() == IceType_i1) | 2115 if (Dest->getType() == IceType_i1) |
| 2105 _and(T_2, Ctx->getConstantInt(IceType_i1, 1)); | 2116 _and(T_2, Ctx->getConstantInt32(IceType_i1, 1)); |
| 2106 _mov(Dest, T_2); | 2117 _mov(Dest, T_2); |
| 2107 T_2->setPreferredRegister(T_1, true); | 2118 T_2->setPreferredRegister(T_1, true); |
| 2108 } | 2119 } |
| 2109 break; | 2120 break; |
| 2110 case InstCast::Fptoui: | 2121 case InstCast::Fptoui: |
| 2111 if (isVectorType(Dest->getType())) { | 2122 if (isVectorType(Dest->getType())) { |
| 2112 assert(Dest->getType() == IceType_v4i32 && | 2123 assert(Dest->getType() == IceType_v4i32 && |
| 2113 Inst->getSrc(0)->getType() == IceType_v4f32); | 2124 Inst->getSrc(0)->getType() == IceType_v4f32); |
| 2114 const SizeT MaxSrcs = 1; | 2125 const SizeT MaxSrcs = 1; |
| 2115 InstCall *Call = makeHelperCall("Sz_fptoui_v4f32", Dest, MaxSrcs); | 2126 InstCall *Call = makeHelperCall("Sz_fptoui_v4f32", Dest, MaxSrcs); |
| (...skipping 16 matching lines...) Expand all Loading... |
| 2132 lowerCall(Call); | 2143 lowerCall(Call); |
| 2133 return; | 2144 return; |
| 2134 } else { | 2145 } else { |
| 2135 Operand *Src0RM = legalize(Inst->getSrc(0), Legal_Reg | Legal_Mem); | 2146 Operand *Src0RM = legalize(Inst->getSrc(0), Legal_Reg | Legal_Mem); |
| 2136 // t1.i32 = cvt Src0RM; t2.dest_type = t1; Dest = t2.dest_type | 2147 // t1.i32 = cvt Src0RM; t2.dest_type = t1; Dest = t2.dest_type |
| 2137 Variable *T_1 = makeReg(IceType_i32); | 2148 Variable *T_1 = makeReg(IceType_i32); |
| 2138 Variable *T_2 = makeReg(Dest->getType()); | 2149 Variable *T_2 = makeReg(Dest->getType()); |
| 2139 _cvtt(T_1, Src0RM); | 2150 _cvtt(T_1, Src0RM); |
| 2140 _mov(T_2, T_1); // T_1 and T_2 may have different integer types | 2151 _mov(T_2, T_1); // T_1 and T_2 may have different integer types |
| 2141 if (Dest->getType() == IceType_i1) | 2152 if (Dest->getType() == IceType_i1) |
| 2142 _and(T_2, Ctx->getConstantInt(IceType_i1, 1)); | 2153 _and(T_2, Ctx->getConstantInt32(IceType_i1, 1)); |
| 2143 _mov(Dest, T_2); | 2154 _mov(Dest, T_2); |
| 2144 T_2->setPreferredRegister(T_1, true); | 2155 T_2->setPreferredRegister(T_1, true); |
| 2145 } | 2156 } |
| 2146 break; | 2157 break; |
| 2147 case InstCast::Sitofp: | 2158 case InstCast::Sitofp: |
| 2148 if (isVectorType(Dest->getType())) { | 2159 if (isVectorType(Dest->getType())) { |
| 2149 assert(Dest->getType() == IceType_v4f32 && | 2160 assert(Dest->getType() == IceType_v4f32 && |
| 2150 Inst->getSrc(0)->getType() == IceType_v4i32); | 2161 Inst->getSrc(0)->getType() == IceType_v4i32); |
| 2151 Operand *Src0RM = legalize(Inst->getSrc(0), Legal_Reg | Legal_Mem); | 2162 Operand *Src0RM = legalize(Inst->getSrc(0), Legal_Reg | Legal_Mem); |
| 2152 Variable *T = makeReg(Dest->getType()); | 2163 Variable *T = makeReg(Dest->getType()); |
| (...skipping 190 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 2343 _movp(Dest, legalizeToVar(Src0)); | 2354 _movp(Dest, legalizeToVar(Src0)); |
| 2344 } break; | 2355 } break; |
| 2345 } | 2356 } |
| 2346 break; | 2357 break; |
| 2347 } | 2358 } |
| 2348 } | 2359 } |
| 2349 } | 2360 } |
| 2350 | 2361 |
| 2351 void TargetX8632::lowerExtractElement(const InstExtractElement *Inst) { | 2362 void TargetX8632::lowerExtractElement(const InstExtractElement *Inst) { |
| 2352 Operand *SourceVectNotLegalized = Inst->getSrc(0); | 2363 Operand *SourceVectNotLegalized = Inst->getSrc(0); |
| 2353 ConstantInteger *ElementIndex = | 2364 ConstantInteger32 *ElementIndex = |
| 2354 llvm::dyn_cast<ConstantInteger>(Inst->getSrc(1)); | 2365 llvm::dyn_cast<ConstantInteger32>(Inst->getSrc(1)); |
| 2355 // Only constant indices are allowed in PNaCl IR. | 2366 // Only constant indices are allowed in PNaCl IR. |
| 2356 assert(ElementIndex); | 2367 assert(ElementIndex); |
| 2357 | 2368 |
| 2358 unsigned Index = ElementIndex->getValue(); | 2369 unsigned Index = ElementIndex->getValue(); |
| 2359 Type Ty = SourceVectNotLegalized->getType(); | 2370 Type Ty = SourceVectNotLegalized->getType(); |
| 2360 Type ElementTy = typeElementType(Ty); | 2371 Type ElementTy = typeElementType(Ty); |
| 2361 Type InVectorElementTy = getInVectorElementType(Ty); | 2372 Type InVectorElementTy = getInVectorElementType(Ty); |
| 2362 Variable *ExtractedElementR = makeReg(InVectorElementTy); | 2373 Variable *ExtractedElementR = makeReg(InVectorElementTy); |
| 2363 | 2374 |
| 2364 // TODO(wala): Determine the best lowering sequences for each type. | 2375 // TODO(wala): Determine the best lowering sequences for each type. |
| 2365 bool CanUsePextr = | 2376 bool CanUsePextr = |
| 2366 Ty == IceType_v8i16 || Ty == IceType_v8i1 || InstructionSet >= SSE4_1; | 2377 Ty == IceType_v8i16 || Ty == IceType_v8i1 || InstructionSet >= SSE4_1; |
| 2367 if (CanUsePextr && Ty != IceType_v4f32) { | 2378 if (CanUsePextr && Ty != IceType_v4f32) { |
| 2368 // Use pextrb, pextrw, or pextrd. | 2379 // Use pextrb, pextrw, or pextrd. |
| 2369 Constant *Mask = Ctx->getConstantInt(IceType_i8, Index); | 2380 Constant *Mask = Ctx->getConstantInt32(IceType_i8, Index); |
| 2370 Variable *SourceVectR = legalizeToVar(SourceVectNotLegalized); | 2381 Variable *SourceVectR = legalizeToVar(SourceVectNotLegalized); |
| 2371 _pextr(ExtractedElementR, SourceVectR, Mask); | 2382 _pextr(ExtractedElementR, SourceVectR, Mask); |
| 2372 } else if (Ty == IceType_v4i32 || Ty == IceType_v4f32 || Ty == IceType_v4i1) { | 2383 } else if (Ty == IceType_v4i32 || Ty == IceType_v4f32 || Ty == IceType_v4i1) { |
| 2373 // Use pshufd and movd/movss. | 2384 // Use pshufd and movd/movss. |
| 2374 Variable *T = NULL; | 2385 Variable *T = NULL; |
| 2375 if (Index) { | 2386 if (Index) { |
| 2376 // The shuffle only needs to occur if the element to be extracted | 2387 // The shuffle only needs to occur if the element to be extracted |
| 2377 // is not at the lowest index. | 2388 // is not at the lowest index. |
| 2378 Constant *Mask = Ctx->getConstantInt(IceType_i8, Index); | 2389 Constant *Mask = Ctx->getConstantInt32(IceType_i8, Index); |
| 2379 T = makeReg(Ty); | 2390 T = makeReg(Ty); |
| 2380 _pshufd(T, legalize(SourceVectNotLegalized, Legal_Reg | Legal_Mem), Mask); | 2391 _pshufd(T, legalize(SourceVectNotLegalized, Legal_Reg | Legal_Mem), Mask); |
| 2381 } else { | 2392 } else { |
| 2382 T = legalizeToVar(SourceVectNotLegalized); | 2393 T = legalizeToVar(SourceVectNotLegalized); |
| 2383 } | 2394 } |
| 2384 | 2395 |
| 2385 if (InVectorElementTy == IceType_i32) { | 2396 if (InVectorElementTy == IceType_i32) { |
| 2386 _movd(ExtractedElementR, T); | 2397 _movd(ExtractedElementR, T); |
| 2387 } else { // Ty == Icetype_f32 | 2398 } else { // Ty == Icetype_f32 |
| 2388 // TODO(wala): _movss is only used here because _mov does not | 2399 // TODO(wala): _movss is only used here because _mov does not |
| (...skipping 118 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 2507 bool HasC1 = (TableFcmp[Index].C1 != InstX8632Br::Br_None); | 2518 bool HasC1 = (TableFcmp[Index].C1 != InstX8632Br::Br_None); |
| 2508 bool HasC2 = (TableFcmp[Index].C2 != InstX8632Br::Br_None); | 2519 bool HasC2 = (TableFcmp[Index].C2 != InstX8632Br::Br_None); |
| 2509 if (HasC1) { | 2520 if (HasC1) { |
| 2510 Src0 = legalize(Src0); | 2521 Src0 = legalize(Src0); |
| 2511 Operand *Src1RM = legalize(Src1, Legal_Reg | Legal_Mem); | 2522 Operand *Src1RM = legalize(Src1, Legal_Reg | Legal_Mem); |
| 2512 Variable *T = NULL; | 2523 Variable *T = NULL; |
| 2513 _mov(T, Src0); | 2524 _mov(T, Src0); |
| 2514 _ucomiss(T, Src1RM); | 2525 _ucomiss(T, Src1RM); |
| 2515 } | 2526 } |
| 2516 Constant *Default = | 2527 Constant *Default = |
| 2517 Ctx->getConstantInt(IceType_i32, TableFcmp[Index].Default); | 2528 Ctx->getConstantInt32(IceType_i32, TableFcmp[Index].Default); |
| 2518 _mov(Dest, Default); | 2529 _mov(Dest, Default); |
| 2519 if (HasC1) { | 2530 if (HasC1) { |
| 2520 InstX8632Label *Label = InstX8632Label::create(Func, this); | 2531 InstX8632Label *Label = InstX8632Label::create(Func, this); |
| 2521 _br(TableFcmp[Index].C1, Label); | 2532 _br(TableFcmp[Index].C1, Label); |
| 2522 if (HasC2) { | 2533 if (HasC2) { |
| 2523 _br(TableFcmp[Index].C2, Label); | 2534 _br(TableFcmp[Index].C2, Label); |
| 2524 } | 2535 } |
| 2525 Context.insert(InstFakeUse::create(Func, Dest)); | 2536 Context.insert(InstFakeUse::create(Func, Dest)); |
| 2526 Constant *NonDefault = | 2537 Constant *NonDefault = |
| 2527 Ctx->getConstantInt(IceType_i32, !TableFcmp[Index].Default); | 2538 Ctx->getConstantInt32(IceType_i32, !TableFcmp[Index].Default); |
| 2528 _mov(Dest, NonDefault); | 2539 _mov(Dest, NonDefault); |
| 2529 Context.insert(Label); | 2540 Context.insert(Label); |
| 2530 } | 2541 } |
| 2531 } | 2542 } |
| 2532 | 2543 |
| 2533 void TargetX8632::lowerIcmp(const InstIcmp *Inst) { | 2544 void TargetX8632::lowerIcmp(const InstIcmp *Inst) { |
| 2534 Operand *Src0 = legalize(Inst->getSrc(0)); | 2545 Operand *Src0 = legalize(Inst->getSrc(0)); |
| 2535 Operand *Src1 = legalize(Inst->getSrc(1)); | 2546 Operand *Src1 = legalize(Inst->getSrc(1)); |
| 2536 Variable *Dest = Inst->getDest(); | 2547 Variable *Dest = Inst->getDest(); |
| 2537 | 2548 |
| (...skipping 120 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 2658 _br(getIcmp32Mapping(Inst->getCondition()), NextBr->getTargetTrue(), | 2669 _br(getIcmp32Mapping(Inst->getCondition()), NextBr->getTargetTrue(), |
| 2659 NextBr->getTargetFalse()); | 2670 NextBr->getTargetFalse()); |
| 2660 // Skip over the following branch instruction. | 2671 // Skip over the following branch instruction. |
| 2661 Context.advanceNext(); | 2672 Context.advanceNext(); |
| 2662 return; | 2673 return; |
| 2663 } | 2674 } |
| 2664 } | 2675 } |
| 2665 | 2676 |
| 2666 // a=icmp cond, b, c ==> cmp b,c; a=1; br cond,L1; FakeUse(a); a=0; L1: | 2677 // a=icmp cond, b, c ==> cmp b,c; a=1; br cond,L1; FakeUse(a); a=0; L1: |
| 2667 Constant *Zero = Ctx->getConstantZero(IceType_i32); | 2678 Constant *Zero = Ctx->getConstantZero(IceType_i32); |
| 2668 Constant *One = Ctx->getConstantInt(IceType_i32, 1); | 2679 Constant *One = Ctx->getConstantInt32(IceType_i32, 1); |
| 2669 if (Src0->getType() == IceType_i64) { | 2680 if (Src0->getType() == IceType_i64) { |
| 2670 InstIcmp::ICond Condition = Inst->getCondition(); | 2681 InstIcmp::ICond Condition = Inst->getCondition(); |
| 2671 size_t Index = static_cast<size_t>(Condition); | 2682 size_t Index = static_cast<size_t>(Condition); |
| 2672 assert(Index < TableIcmp64Size); | 2683 assert(Index < TableIcmp64Size); |
| 2673 Operand *Src0LoRM = legalize(loOperand(Src0), Legal_Reg | Legal_Mem); | 2684 Operand *Src0LoRM = legalize(loOperand(Src0), Legal_Reg | Legal_Mem); |
| 2674 Operand *Src0HiRM = legalize(hiOperand(Src0), Legal_Reg | Legal_Mem); | 2685 Operand *Src0HiRM = legalize(hiOperand(Src0), Legal_Reg | Legal_Mem); |
| 2675 Operand *Src1LoRI = legalize(loOperand(Src1), Legal_Reg | Legal_Imm); | 2686 Operand *Src1LoRI = legalize(loOperand(Src1), Legal_Reg | Legal_Imm); |
| 2676 Operand *Src1HiRI = legalize(hiOperand(Src1), Legal_Reg | Legal_Imm); | 2687 Operand *Src1HiRI = legalize(hiOperand(Src1), Legal_Reg | Legal_Imm); |
| 2677 if (Condition == InstIcmp::Eq || Condition == InstIcmp::Ne) { | 2688 if (Condition == InstIcmp::Eq || Condition == InstIcmp::Ne) { |
| 2678 InstX8632Label *Label = InstX8632Label::create(Func, this); | 2689 InstX8632Label *Label = InstX8632Label::create(Func, this); |
| (...skipping 30 matching lines...) Expand all Loading... |
| 2709 _mov(Dest, One); | 2720 _mov(Dest, One); |
| 2710 _br(getIcmp32Mapping(Inst->getCondition()), Label); | 2721 _br(getIcmp32Mapping(Inst->getCondition()), Label); |
| 2711 Context.insert(InstFakeUse::create(Func, Dest)); | 2722 Context.insert(InstFakeUse::create(Func, Dest)); |
| 2712 _mov(Dest, Zero); | 2723 _mov(Dest, Zero); |
| 2713 Context.insert(Label); | 2724 Context.insert(Label); |
| 2714 } | 2725 } |
| 2715 | 2726 |
| 2716 void TargetX8632::lowerInsertElement(const InstInsertElement *Inst) { | 2727 void TargetX8632::lowerInsertElement(const InstInsertElement *Inst) { |
| 2717 Operand *SourceVectNotLegalized = Inst->getSrc(0); | 2728 Operand *SourceVectNotLegalized = Inst->getSrc(0); |
| 2718 Operand *ElementToInsertNotLegalized = Inst->getSrc(1); | 2729 Operand *ElementToInsertNotLegalized = Inst->getSrc(1); |
| 2719 ConstantInteger *ElementIndex = | 2730 ConstantInteger32 *ElementIndex = |
| 2720 llvm::dyn_cast<ConstantInteger>(Inst->getSrc(2)); | 2731 llvm::dyn_cast<ConstantInteger32>(Inst->getSrc(2)); |
| 2721 // Only constant indices are allowed in PNaCl IR. | 2732 // Only constant indices are allowed in PNaCl IR. |
| 2722 assert(ElementIndex); | 2733 assert(ElementIndex); |
| 2723 unsigned Index = ElementIndex->getValue(); | 2734 unsigned Index = ElementIndex->getValue(); |
| 2724 assert(Index < typeNumElements(SourceVectNotLegalized->getType())); | 2735 assert(Index < typeNumElements(SourceVectNotLegalized->getType())); |
| 2725 | 2736 |
| 2726 Type Ty = SourceVectNotLegalized->getType(); | 2737 Type Ty = SourceVectNotLegalized->getType(); |
| 2727 Type ElementTy = typeElementType(Ty); | 2738 Type ElementTy = typeElementType(Ty); |
| 2728 Type InVectorElementTy = getInVectorElementType(Ty); | 2739 Type InVectorElementTy = getInVectorElementType(Ty); |
| 2729 | 2740 |
| 2730 if (ElementTy == IceType_i1) { | 2741 if (ElementTy == IceType_i1) { |
| 2731 // Expand the element to the appropriate size for it to be inserted | 2742 // Expand the element to the appropriate size for it to be inserted |
| 2732 // in the vector. | 2743 // in the vector. |
| 2733 Variable *Expanded = | 2744 Variable *Expanded = |
| 2734 Func->makeVariable(InVectorElementTy, Context.getNode()); | 2745 Func->makeVariable(InVectorElementTy, Context.getNode()); |
| 2735 InstCast *Cast = InstCast::create(Func, InstCast::Zext, Expanded, | 2746 InstCast *Cast = InstCast::create(Func, InstCast::Zext, Expanded, |
| 2736 ElementToInsertNotLegalized); | 2747 ElementToInsertNotLegalized); |
| 2737 lowerCast(Cast); | 2748 lowerCast(Cast); |
| 2738 ElementToInsertNotLegalized = Expanded; | 2749 ElementToInsertNotLegalized = Expanded; |
| 2739 } | 2750 } |
| 2740 | 2751 |
| 2741 if (Ty == IceType_v8i16 || Ty == IceType_v8i1 || InstructionSet >= SSE4_1) { | 2752 if (Ty == IceType_v8i16 || Ty == IceType_v8i1 || InstructionSet >= SSE4_1) { |
| 2742 // Use insertps, pinsrb, pinsrw, or pinsrd. | 2753 // Use insertps, pinsrb, pinsrw, or pinsrd. |
| 2743 Operand *ElementRM = | 2754 Operand *ElementRM = |
| 2744 legalize(ElementToInsertNotLegalized, Legal_Reg | Legal_Mem); | 2755 legalize(ElementToInsertNotLegalized, Legal_Reg | Legal_Mem); |
| 2745 Operand *SourceVectRM = | 2756 Operand *SourceVectRM = |
| 2746 legalize(SourceVectNotLegalized, Legal_Reg | Legal_Mem); | 2757 legalize(SourceVectNotLegalized, Legal_Reg | Legal_Mem); |
| 2747 Variable *T = makeReg(Ty); | 2758 Variable *T = makeReg(Ty); |
| 2748 _movp(T, SourceVectRM); | 2759 _movp(T, SourceVectRM); |
| 2749 if (Ty == IceType_v4f32) | 2760 if (Ty == IceType_v4f32) |
| 2750 _insertps(T, ElementRM, Ctx->getConstantInt(IceType_i8, Index << 4)); | 2761 _insertps(T, ElementRM, Ctx->getConstantInt32(IceType_i8, Index << 4)); |
| 2751 else | 2762 else |
| 2752 _pinsr(T, ElementRM, Ctx->getConstantInt(IceType_i8, Index)); | 2763 _pinsr(T, ElementRM, Ctx->getConstantInt32(IceType_i8, Index)); |
| 2753 _movp(Inst->getDest(), T); | 2764 _movp(Inst->getDest(), T); |
| 2754 } else if (Ty == IceType_v4i32 || Ty == IceType_v4f32 || Ty == IceType_v4i1) { | 2765 } else if (Ty == IceType_v4i32 || Ty == IceType_v4f32 || Ty == IceType_v4i1) { |
| 2755 // Use shufps or movss. | 2766 // Use shufps or movss. |
| 2756 Variable *ElementR = NULL; | 2767 Variable *ElementR = NULL; |
| 2757 Operand *SourceVectRM = | 2768 Operand *SourceVectRM = |
| 2758 legalize(SourceVectNotLegalized, Legal_Reg | Legal_Mem); | 2769 legalize(SourceVectNotLegalized, Legal_Reg | Legal_Mem); |
| 2759 | 2770 |
| 2760 if (InVectorElementTy == IceType_f32) { | 2771 if (InVectorElementTy == IceType_f32) { |
| 2761 // ElementR will be in an XMM register since it is floating point. | 2772 // ElementR will be in an XMM register since it is floating point. |
| 2762 ElementR = legalizeToVar(ElementToInsertNotLegalized); | 2773 ElementR = legalizeToVar(ElementToInsertNotLegalized); |
| (...skipping 30 matching lines...) Expand all Loading... |
| 2793 // ElementR := ElementR[0, 0] T[0, 3] | 2804 // ElementR := ElementR[0, 0] T[0, 3] |
| 2794 // T := T[0, 1] ElementR[0, 3] | 2805 // T := T[0, 1] ElementR[0, 3] |
| 2795 // | 2806 // |
| 2796 // insertelement into index 3 (result is stored in T): | 2807 // insertelement into index 3 (result is stored in T): |
| 2797 // T := SourceVectRM | 2808 // T := SourceVectRM |
| 2798 // ElementR := ElementR[0, 0] T[0, 2] | 2809 // ElementR := ElementR[0, 0] T[0, 2] |
| 2799 // T := T[0, 1] ElementR[3, 0] | 2810 // T := T[0, 1] ElementR[3, 0] |
| 2800 const unsigned char Mask1[3] = {0, 192, 128}; | 2811 const unsigned char Mask1[3] = {0, 192, 128}; |
| 2801 const unsigned char Mask2[3] = {227, 196, 52}; | 2812 const unsigned char Mask2[3] = {227, 196, 52}; |
| 2802 | 2813 |
| 2803 Constant *Mask1Constant = Ctx->getConstantInt(IceType_i8, Mask1[Index - 1]); | 2814 Constant *Mask1Constant = |
| 2804 Constant *Mask2Constant = Ctx->getConstantInt(IceType_i8, Mask2[Index - 1]); | 2815 Ctx->getConstantInt32(IceType_i8, Mask1[Index - 1]); |
| 2816 Constant *Mask2Constant = |
| 2817 Ctx->getConstantInt32(IceType_i8, Mask2[Index - 1]); |
| 2805 | 2818 |
| 2806 if (Index == 1) { | 2819 if (Index == 1) { |
| 2807 _shufps(ElementR, SourceVectRM, Mask1Constant); | 2820 _shufps(ElementR, SourceVectRM, Mask1Constant); |
| 2808 _shufps(ElementR, SourceVectRM, Mask2Constant); | 2821 _shufps(ElementR, SourceVectRM, Mask2Constant); |
| 2809 _movp(Inst->getDest(), ElementR); | 2822 _movp(Inst->getDest(), ElementR); |
| 2810 } else { | 2823 } else { |
| 2811 Variable *T = makeReg(Ty); | 2824 Variable *T = makeReg(Ty); |
| 2812 _movp(T, SourceVectRM); | 2825 _movp(T, SourceVectRM); |
| 2813 _shufps(ElementR, T, Mask1Constant); | 2826 _shufps(ElementR, T, Mask1Constant); |
| 2814 _shufps(T, ElementR, Mask2Constant); | 2827 _shufps(T, ElementR, Mask2Constant); |
| (...skipping 19 matching lines...) Expand all Loading... |
| 2834 Variable *T = makeReg(Ty); | 2847 Variable *T = makeReg(Ty); |
| 2835 _movp(T, Slot); | 2848 _movp(T, Slot); |
| 2836 _movp(Inst->getDest(), T); | 2849 _movp(Inst->getDest(), T); |
| 2837 } | 2850 } |
| 2838 } | 2851 } |
| 2839 | 2852 |
| 2840 void TargetX8632::lowerIntrinsicCall(const InstIntrinsicCall *Instr) { | 2853 void TargetX8632::lowerIntrinsicCall(const InstIntrinsicCall *Instr) { |
| 2841 switch (Instr->getIntrinsicInfo().ID) { | 2854 switch (Instr->getIntrinsicInfo().ID) { |
| 2842 case Intrinsics::AtomicCmpxchg: { | 2855 case Intrinsics::AtomicCmpxchg: { |
| 2843 if (!Intrinsics::VerifyMemoryOrder( | 2856 if (!Intrinsics::VerifyMemoryOrder( |
| 2844 llvm::cast<ConstantInteger>(Instr->getArg(3))->getValue())) { | 2857 llvm::cast<ConstantInteger32>(Instr->getArg(3))->getValue())) { |
| 2845 Func->setError("Unexpected memory ordering (success) for AtomicCmpxchg"); | 2858 Func->setError("Unexpected memory ordering (success) for AtomicCmpxchg"); |
| 2846 return; | 2859 return; |
| 2847 } | 2860 } |
| 2848 if (!Intrinsics::VerifyMemoryOrder( | 2861 if (!Intrinsics::VerifyMemoryOrder( |
| 2849 llvm::cast<ConstantInteger>(Instr->getArg(4))->getValue())) { | 2862 llvm::cast<ConstantInteger32>(Instr->getArg(4))->getValue())) { |
| 2850 Func->setError("Unexpected memory ordering (failure) for AtomicCmpxchg"); | 2863 Func->setError("Unexpected memory ordering (failure) for AtomicCmpxchg"); |
| 2851 return; | 2864 return; |
| 2852 } | 2865 } |
| 2853 Variable *DestPrev = Instr->getDest(); | 2866 Variable *DestPrev = Instr->getDest(); |
| 2854 Operand *PtrToMem = Instr->getArg(0); | 2867 Operand *PtrToMem = Instr->getArg(0); |
| 2855 Operand *Expected = Instr->getArg(1); | 2868 Operand *Expected = Instr->getArg(1); |
| 2856 Operand *Desired = Instr->getArg(2); | 2869 Operand *Desired = Instr->getArg(2); |
| 2857 if (tryOptimizedCmpxchgCmpBr(DestPrev, PtrToMem, Expected, Desired)) | 2870 if (tryOptimizedCmpxchgCmpBr(DestPrev, PtrToMem, Expected, Desired)) |
| 2858 return; | 2871 return; |
| 2859 lowerAtomicCmpxchg(DestPrev, PtrToMem, Expected, Desired); | 2872 lowerAtomicCmpxchg(DestPrev, PtrToMem, Expected, Desired); |
| 2860 return; | 2873 return; |
| 2861 } | 2874 } |
| 2862 case Intrinsics::AtomicFence: | 2875 case Intrinsics::AtomicFence: |
| 2863 if (!Intrinsics::VerifyMemoryOrder( | 2876 if (!Intrinsics::VerifyMemoryOrder( |
| 2864 llvm::cast<ConstantInteger>(Instr->getArg(0))->getValue())) { | 2877 llvm::cast<ConstantInteger32>(Instr->getArg(0))->getValue())) { |
| 2865 Func->setError("Unexpected memory ordering for AtomicFence"); | 2878 Func->setError("Unexpected memory ordering for AtomicFence"); |
| 2866 return; | 2879 return; |
| 2867 } | 2880 } |
| 2868 _mfence(); | 2881 _mfence(); |
| 2869 return; | 2882 return; |
| 2870 case Intrinsics::AtomicFenceAll: | 2883 case Intrinsics::AtomicFenceAll: |
| 2871 // NOTE: FenceAll should prevent and load/store from being moved | 2884 // NOTE: FenceAll should prevent and load/store from being moved |
| 2872 // across the fence (both atomic and non-atomic). The InstX8632Mfence | 2885 // across the fence (both atomic and non-atomic). The InstX8632Mfence |
| 2873 // instruction is currently marked coarsely as "HasSideEffects". | 2886 // instruction is currently marked coarsely as "HasSideEffects". |
| 2874 _mfence(); | 2887 _mfence(); |
| 2875 return; | 2888 return; |
| 2876 case Intrinsics::AtomicIsLockFree: { | 2889 case Intrinsics::AtomicIsLockFree: { |
| 2877 // X86 is always lock free for 8/16/32/64 bit accesses. | 2890 // X86 is always lock free for 8/16/32/64 bit accesses. |
| 2878 // TODO(jvoung): Since the result is constant when given a constant | 2891 // TODO(jvoung): Since the result is constant when given a constant |
| 2879 // byte size, this opens up DCE opportunities. | 2892 // byte size, this opens up DCE opportunities. |
| 2880 Operand *ByteSize = Instr->getArg(0); | 2893 Operand *ByteSize = Instr->getArg(0); |
| 2881 Variable *Dest = Instr->getDest(); | 2894 Variable *Dest = Instr->getDest(); |
| 2882 if (ConstantInteger *CI = llvm::dyn_cast<ConstantInteger>(ByteSize)) { | 2895 if (ConstantInteger32 *CI = llvm::dyn_cast<ConstantInteger32>(ByteSize)) { |
| 2883 Constant *Result; | 2896 Constant *Result; |
| 2884 switch (CI->getValue()) { | 2897 switch (CI->getValue()) { |
| 2885 default: | 2898 default: |
| 2886 // Some x86-64 processors support the cmpxchg16b intruction, which | 2899 // Some x86-64 processors support the cmpxchg16b intruction, which |
| 2887 // can make 16-byte operations lock free (when used with the LOCK | 2900 // can make 16-byte operations lock free (when used with the LOCK |
| 2888 // prefix). However, that's not supported in 32-bit mode, so just | 2901 // prefix). However, that's not supported in 32-bit mode, so just |
| 2889 // return 0 even for large sizes. | 2902 // return 0 even for large sizes. |
| 2890 Result = Ctx->getConstantZero(IceType_i32); | 2903 Result = Ctx->getConstantZero(IceType_i32); |
| 2891 break; | 2904 break; |
| 2892 case 1: | 2905 case 1: |
| 2893 case 2: | 2906 case 2: |
| 2894 case 4: | 2907 case 4: |
| 2895 case 8: | 2908 case 8: |
| 2896 Result = Ctx->getConstantInt(IceType_i32, 1); | 2909 Result = Ctx->getConstantInt32(IceType_i32, 1); |
| 2897 break; | 2910 break; |
| 2898 } | 2911 } |
| 2899 _mov(Dest, Result); | 2912 _mov(Dest, Result); |
| 2900 return; | 2913 return; |
| 2901 } | 2914 } |
| 2902 // The PNaCl ABI requires the byte size to be a compile-time constant. | 2915 // The PNaCl ABI requires the byte size to be a compile-time constant. |
| 2903 Func->setError("AtomicIsLockFree byte size should be compile-time const"); | 2916 Func->setError("AtomicIsLockFree byte size should be compile-time const"); |
| 2904 return; | 2917 return; |
| 2905 } | 2918 } |
| 2906 case Intrinsics::AtomicLoad: { | 2919 case Intrinsics::AtomicLoad: { |
| 2907 // We require the memory address to be naturally aligned. | 2920 // We require the memory address to be naturally aligned. |
| 2908 // Given that is the case, then normal loads are atomic. | 2921 // Given that is the case, then normal loads are atomic. |
| 2909 if (!Intrinsics::VerifyMemoryOrder( | 2922 if (!Intrinsics::VerifyMemoryOrder( |
| 2910 llvm::cast<ConstantInteger>(Instr->getArg(1))->getValue())) { | 2923 llvm::cast<ConstantInteger32>(Instr->getArg(1))->getValue())) { |
| 2911 Func->setError("Unexpected memory ordering for AtomicLoad"); | 2924 Func->setError("Unexpected memory ordering for AtomicLoad"); |
| 2912 return; | 2925 return; |
| 2913 } | 2926 } |
| 2914 Variable *Dest = Instr->getDest(); | 2927 Variable *Dest = Instr->getDest(); |
| 2915 if (Dest->getType() == IceType_i64) { | 2928 if (Dest->getType() == IceType_i64) { |
| 2916 // Follow what GCC does and use a movq instead of what lowerLoad() | 2929 // Follow what GCC does and use a movq instead of what lowerLoad() |
| 2917 // normally does (split the load into two). | 2930 // normally does (split the load into two). |
| 2918 // Thus, this skips load/arithmetic op folding. Load/arithmetic folding | 2931 // Thus, this skips load/arithmetic op folding. Load/arithmetic folding |
| 2919 // can't happen anyway, since this is x86-32 and integer arithmetic only | 2932 // can't happen anyway, since this is x86-32 and integer arithmetic only |
| 2920 // happens on 32-bit quantities. | 2933 // happens on 32-bit quantities. |
| (...skipping 12 matching lines...) Expand all Loading... |
| 2933 lowerLoad(Load); | 2946 lowerLoad(Load); |
| 2934 // Make sure the atomic load isn't elided when unused, by adding a FakeUse. | 2947 // Make sure the atomic load isn't elided when unused, by adding a FakeUse. |
| 2935 // Since lowerLoad may fuse the load w/ an arithmetic instruction, | 2948 // Since lowerLoad may fuse the load w/ an arithmetic instruction, |
| 2936 // insert the FakeUse on the last-inserted instruction's dest. | 2949 // insert the FakeUse on the last-inserted instruction's dest. |
| 2937 Context.insert( | 2950 Context.insert( |
| 2938 InstFakeUse::create(Func, Context.getLastInserted()->getDest())); | 2951 InstFakeUse::create(Func, Context.getLastInserted()->getDest())); |
| 2939 return; | 2952 return; |
| 2940 } | 2953 } |
| 2941 case Intrinsics::AtomicRMW: | 2954 case Intrinsics::AtomicRMW: |
| 2942 if (!Intrinsics::VerifyMemoryOrder( | 2955 if (!Intrinsics::VerifyMemoryOrder( |
| 2943 llvm::cast<ConstantInteger>(Instr->getArg(3))->getValue())) { | 2956 llvm::cast<ConstantInteger32>(Instr->getArg(3))->getValue())) { |
| 2944 Func->setError("Unexpected memory ordering for AtomicRMW"); | 2957 Func->setError("Unexpected memory ordering for AtomicRMW"); |
| 2945 return; | 2958 return; |
| 2946 } | 2959 } |
| 2947 lowerAtomicRMW(Instr->getDest(), | 2960 lowerAtomicRMW(Instr->getDest(), |
| 2948 static_cast<uint32_t>(llvm::cast<ConstantInteger>( | 2961 static_cast<uint32_t>(llvm::cast<ConstantInteger32>( |
| 2949 Instr->getArg(0))->getValue()), | 2962 Instr->getArg(0))->getValue()), |
| 2950 Instr->getArg(1), Instr->getArg(2)); | 2963 Instr->getArg(1), Instr->getArg(2)); |
| 2951 return; | 2964 return; |
| 2952 case Intrinsics::AtomicStore: { | 2965 case Intrinsics::AtomicStore: { |
| 2953 if (!Intrinsics::VerifyMemoryOrder( | 2966 if (!Intrinsics::VerifyMemoryOrder( |
| 2954 llvm::cast<ConstantInteger>(Instr->getArg(2))->getValue())) { | 2967 llvm::cast<ConstantInteger32>(Instr->getArg(2))->getValue())) { |
| 2955 Func->setError("Unexpected memory ordering for AtomicStore"); | 2968 Func->setError("Unexpected memory ordering for AtomicStore"); |
| 2956 return; | 2969 return; |
| 2957 } | 2970 } |
| 2958 // We require the memory address to be naturally aligned. | 2971 // We require the memory address to be naturally aligned. |
| 2959 // Given that is the case, then normal stores are atomic. | 2972 // Given that is the case, then normal stores are atomic. |
| 2960 // Add a fence after the store to make it visible. | 2973 // Add a fence after the store to make it visible. |
| 2961 Operand *Value = Instr->getArg(0); | 2974 Operand *Value = Instr->getArg(0); |
| 2962 Operand *Ptr = Instr->getArg(1); | 2975 Operand *Ptr = Instr->getArg(1); |
| 2963 if (Value->getType() == IceType_i64) { | 2976 if (Value->getType() == IceType_i64) { |
| 2964 // Use a movq instead of what lowerStore() normally does | 2977 // Use a movq instead of what lowerStore() normally does |
| (...skipping 27 matching lines...) Expand all Loading... |
| 2992 _bswap(T_Hi); | 3005 _bswap(T_Hi); |
| 2993 _mov(DestLo, T_Hi); | 3006 _mov(DestLo, T_Hi); |
| 2994 _mov(DestHi, T_Lo); | 3007 _mov(DestHi, T_Lo); |
| 2995 } else if (Val->getType() == IceType_i32) { | 3008 } else if (Val->getType() == IceType_i32) { |
| 2996 Variable *T = legalizeToVar(Val); | 3009 Variable *T = legalizeToVar(Val); |
| 2997 _bswap(T); | 3010 _bswap(T); |
| 2998 _mov(Dest, T); | 3011 _mov(Dest, T); |
| 2999 } else { | 3012 } else { |
| 3000 assert(Val->getType() == IceType_i16); | 3013 assert(Val->getType() == IceType_i16); |
| 3001 Val = legalize(Val); | 3014 Val = legalize(Val); |
| 3002 Constant *Eight = Ctx->getConstantInt(IceType_i16, 8); | 3015 Constant *Eight = Ctx->getConstantInt32(IceType_i16, 8); |
| 3003 Variable *T = NULL; | 3016 Variable *T = NULL; |
| 3004 _mov(T, Val); | 3017 _mov(T, Val); |
| 3005 _rol(T, Eight); | 3018 _rol(T, Eight); |
| 3006 _mov(Dest, T); | 3019 _mov(Dest, T); |
| 3007 } | 3020 } |
| 3008 return; | 3021 return; |
| 3009 } | 3022 } |
| 3010 case Intrinsics::Ctpop: { | 3023 case Intrinsics::Ctpop: { |
| 3011 Variable *Dest = Instr->getDest(); | 3024 Variable *Dest = Instr->getDest(); |
| 3012 Operand *Val = Instr->getArg(0); | 3025 Operand *Val = Instr->getArg(0); |
| (...skipping 462 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 3475 // bit position conversion, and the speculation is reversed. | 3488 // bit position conversion, and the speculation is reversed. |
| 3476 assert(Ty == IceType_i32 || Ty == IceType_i64); | 3489 assert(Ty == IceType_i32 || Ty == IceType_i64); |
| 3477 Variable *T = makeReg(IceType_i32); | 3490 Variable *T = makeReg(IceType_i32); |
| 3478 Operand *FirstValRM = legalize(FirstVal, Legal_Mem | Legal_Reg); | 3491 Operand *FirstValRM = legalize(FirstVal, Legal_Mem | Legal_Reg); |
| 3479 if (Cttz) { | 3492 if (Cttz) { |
| 3480 _bsf(T, FirstValRM); | 3493 _bsf(T, FirstValRM); |
| 3481 } else { | 3494 } else { |
| 3482 _bsr(T, FirstValRM); | 3495 _bsr(T, FirstValRM); |
| 3483 } | 3496 } |
| 3484 Variable *T_Dest = makeReg(IceType_i32); | 3497 Variable *T_Dest = makeReg(IceType_i32); |
| 3485 Constant *ThirtyTwo = Ctx->getConstantInt(IceType_i32, 32); | 3498 Constant *ThirtyTwo = Ctx->getConstantInt32(IceType_i32, 32); |
| 3486 Constant *ThirtyOne = Ctx->getConstantInt(IceType_i32, 31); | 3499 Constant *ThirtyOne = Ctx->getConstantInt32(IceType_i32, 31); |
| 3487 if (Cttz) { | 3500 if (Cttz) { |
| 3488 _mov(T_Dest, ThirtyTwo); | 3501 _mov(T_Dest, ThirtyTwo); |
| 3489 } else { | 3502 } else { |
| 3490 Constant *SixtyThree = Ctx->getConstantInt(IceType_i32, 63); | 3503 Constant *SixtyThree = Ctx->getConstantInt32(IceType_i32, 63); |
| 3491 _mov(T_Dest, SixtyThree); | 3504 _mov(T_Dest, SixtyThree); |
| 3492 } | 3505 } |
| 3493 _cmov(T_Dest, T, InstX8632::Br_ne); | 3506 _cmov(T_Dest, T, InstX8632::Br_ne); |
| 3494 if (!Cttz) { | 3507 if (!Cttz) { |
| 3495 _xor(T_Dest, ThirtyOne); | 3508 _xor(T_Dest, ThirtyOne); |
| 3496 } | 3509 } |
| 3497 if (Ty == IceType_i32) { | 3510 if (Ty == IceType_i32) { |
| 3498 _mov(Dest, T_Dest); | 3511 _mov(Dest, T_Dest); |
| 3499 return; | 3512 return; |
| 3500 } | 3513 } |
| (...skipping 108 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 3609 if (Index == NULL) | 3622 if (Index == NULL) |
| 3610 return false; | 3623 return false; |
| 3611 const Inst *IndexInst = Index->getDefinition(); | 3624 const Inst *IndexInst = Index->getDefinition(); |
| 3612 if (IndexInst == NULL) | 3625 if (IndexInst == NULL) |
| 3613 return false; | 3626 return false; |
| 3614 if (IndexInst->getSrcSize() < 2) | 3627 if (IndexInst->getSrcSize() < 2) |
| 3615 return false; | 3628 return false; |
| 3616 if (const InstArithmetic *ArithInst = | 3629 if (const InstArithmetic *ArithInst = |
| 3617 llvm::dyn_cast<InstArithmetic>(IndexInst)) { | 3630 llvm::dyn_cast<InstArithmetic>(IndexInst)) { |
| 3618 if (Variable *Var = llvm::dyn_cast<Variable>(ArithInst->getSrc(0))) { | 3631 if (Variable *Var = llvm::dyn_cast<Variable>(ArithInst->getSrc(0))) { |
| 3619 if (ConstantInteger *Const = | 3632 if (ConstantInteger32 *Const = |
| 3620 llvm::dyn_cast<ConstantInteger>(ArithInst->getSrc(1))) { | 3633 llvm::dyn_cast<ConstantInteger32>(ArithInst->getSrc(1))) { |
| 3621 if (ArithInst->getOp() == InstArithmetic::Mul && | 3634 if (ArithInst->getOp() == InstArithmetic::Mul && |
| 3622 !Var->getIsMultidef() && Const->getType() == IceType_i32) { | 3635 !Var->getIsMultidef() && Const->getType() == IceType_i32) { |
| 3623 uint64_t Mult = Const->getValue(); | 3636 uint64_t Mult = Const->getValue(); |
| 3624 uint32_t LogMult; | 3637 uint32_t LogMult; |
| 3625 switch (Mult) { | 3638 switch (Mult) { |
| 3626 case 1: | 3639 case 1: |
| 3627 LogMult = 0; | 3640 LogMult = 0; |
| 3628 break; | 3641 break; |
| 3629 case 2: | 3642 case 2: |
| 3630 LogMult = 1; | 3643 LogMult = 1; |
| (...skipping 30 matching lines...) Expand all Loading... |
| 3661 const Inst *BaseInst = Base->getDefinition(); | 3674 const Inst *BaseInst = Base->getDefinition(); |
| 3662 if (BaseInst == NULL) | 3675 if (BaseInst == NULL) |
| 3663 return false; | 3676 return false; |
| 3664 if (const InstArithmetic *ArithInst = | 3677 if (const InstArithmetic *ArithInst = |
| 3665 llvm::dyn_cast<const InstArithmetic>(BaseInst)) { | 3678 llvm::dyn_cast<const InstArithmetic>(BaseInst)) { |
| 3666 if (ArithInst->getOp() != InstArithmetic::Add && | 3679 if (ArithInst->getOp() != InstArithmetic::Add && |
| 3667 ArithInst->getOp() != InstArithmetic::Sub) | 3680 ArithInst->getOp() != InstArithmetic::Sub) |
| 3668 return false; | 3681 return false; |
| 3669 bool IsAdd = ArithInst->getOp() == InstArithmetic::Add; | 3682 bool IsAdd = ArithInst->getOp() == InstArithmetic::Add; |
| 3670 Variable *Var = NULL; | 3683 Variable *Var = NULL; |
| 3671 ConstantInteger *Const = NULL; | 3684 ConstantInteger32 *Const = NULL; |
| 3672 if (Variable *VariableOperand = | 3685 if (Variable *VariableOperand = |
| 3673 llvm::dyn_cast<Variable>(ArithInst->getSrc(0))) { | 3686 llvm::dyn_cast<Variable>(ArithInst->getSrc(0))) { |
| 3674 Var = VariableOperand; | 3687 Var = VariableOperand; |
| 3675 Const = llvm::dyn_cast<ConstantInteger>(ArithInst->getSrc(1)); | 3688 Const = llvm::dyn_cast<ConstantInteger32>(ArithInst->getSrc(1)); |
| 3676 } else if (IsAdd) { | 3689 } else if (IsAdd) { |
| 3677 Const = llvm::dyn_cast<ConstantInteger>(ArithInst->getSrc(0)); | 3690 Const = llvm::dyn_cast<ConstantInteger32>(ArithInst->getSrc(0)); |
| 3678 Var = llvm::dyn_cast<Variable>(ArithInst->getSrc(1)); | 3691 Var = llvm::dyn_cast<Variable>(ArithInst->getSrc(1)); |
| 3679 } | 3692 } |
| 3680 if (Var == NULL || Const == NULL || Var->getIsMultidef()) | 3693 if (Var == NULL || Const == NULL || Var->getIsMultidef()) |
| 3681 return false; | 3694 return false; |
| 3682 Base = Var; | 3695 Base = Var; |
| 3683 Offset += IsAdd ? Const->getValue() : -Const->getValue(); | 3696 Offset += IsAdd ? Const->getValue() : -Const->getValue(); |
| 3684 Reason = BaseInst; | 3697 Reason = BaseInst; |
| 3685 return true; | 3698 return true; |
| 3686 } | 3699 } |
| 3687 return false; | 3700 return false; |
| (...skipping 117 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 3805 // Vanilla ICE load instructions should not use the segment registers, | 3818 // Vanilla ICE load instructions should not use the segment registers, |
| 3806 // and computeAddressOpt only works at the level of Variables and Constants, | 3819 // and computeAddressOpt only works at the level of Variables and Constants, |
| 3807 // not other OperandX8632Mem, so there should be no mention of segment | 3820 // not other OperandX8632Mem, so there should be no mention of segment |
| 3808 // registers there either. | 3821 // registers there either. |
| 3809 const OperandX8632Mem::SegmentRegisters SegmentReg = | 3822 const OperandX8632Mem::SegmentRegisters SegmentReg = |
| 3810 OperandX8632Mem::DefaultSegment; | 3823 OperandX8632Mem::DefaultSegment; |
| 3811 Variable *Base = llvm::dyn_cast<Variable>(Addr); | 3824 Variable *Base = llvm::dyn_cast<Variable>(Addr); |
| 3812 computeAddressOpt(Func, Inst, Base, Index, Shift, Offset); | 3825 computeAddressOpt(Func, Inst, Base, Index, Shift, Offset); |
| 3813 if (Base && Addr != Base) { | 3826 if (Base && Addr != Base) { |
| 3814 Inst->setDeleted(); | 3827 Inst->setDeleted(); |
| 3815 Constant *OffsetOp = Ctx->getConstantInt(IceType_i32, Offset); | 3828 Constant *OffsetOp = Ctx->getConstantInt32(IceType_i32, Offset); |
| 3816 Addr = OperandX8632Mem::create(Func, Dest->getType(), Base, OffsetOp, Index, | 3829 Addr = OperandX8632Mem::create(Func, Dest->getType(), Base, OffsetOp, Index, |
| 3817 Shift, SegmentReg); | 3830 Shift, SegmentReg); |
| 3818 Context.insert(InstLoad::create(Func, Dest, Addr)); | 3831 Context.insert(InstLoad::create(Func, Dest, Addr)); |
| 3819 } | 3832 } |
| 3820 } | 3833 } |
| 3821 | 3834 |
| 3822 void TargetX8632::randomlyInsertNop(float Probability) { | 3835 void TargetX8632::randomlyInsertNop(float Probability) { |
| 3823 RandomNumberGeneratorWrapper RNG(Ctx->getRNG()); | 3836 RandomNumberGeneratorWrapper RNG(Ctx->getRNG()); |
| 3824 if (RNG.getTrueWithProbability(Probability)) { | 3837 if (RNG.getTrueWithProbability(Probability)) { |
| 3825 _nop(RNG.next(X86_NUM_NOP_VARIANTS)); | 3838 _nop(RNG.next(X86_NUM_NOP_VARIANTS)); |
| (...skipping 46 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 3872 if (InstructionSet >= SSE4_1) { | 3885 if (InstructionSet >= SSE4_1) { |
| 3873 // TODO(wala): If the condition operand is a constant, use blendps | 3886 // TODO(wala): If the condition operand is a constant, use blendps |
| 3874 // or pblendw. | 3887 // or pblendw. |
| 3875 // | 3888 // |
| 3876 // Use blendvps or pblendvb to implement select. | 3889 // Use blendvps or pblendvb to implement select. |
| 3877 if (SrcTy == IceType_v4i1 || SrcTy == IceType_v4i32 || | 3890 if (SrcTy == IceType_v4i1 || SrcTy == IceType_v4i32 || |
| 3878 SrcTy == IceType_v4f32) { | 3891 SrcTy == IceType_v4f32) { |
| 3879 Operand *ConditionRM = legalize(Condition, Legal_Reg | Legal_Mem); | 3892 Operand *ConditionRM = legalize(Condition, Legal_Reg | Legal_Mem); |
| 3880 Variable *xmm0 = makeReg(IceType_v4i32, Reg_xmm0); | 3893 Variable *xmm0 = makeReg(IceType_v4i32, Reg_xmm0); |
| 3881 _movp(xmm0, ConditionRM); | 3894 _movp(xmm0, ConditionRM); |
| 3882 _psll(xmm0, Ctx->getConstantInt(IceType_i8, 31)); | 3895 _psll(xmm0, Ctx->getConstantInt32(IceType_i8, 31)); |
| 3883 _movp(T, SrcFRM); | 3896 _movp(T, SrcFRM); |
| 3884 _blendvps(T, SrcTRM, xmm0); | 3897 _blendvps(T, SrcTRM, xmm0); |
| 3885 _movp(Dest, T); | 3898 _movp(Dest, T); |
| 3886 } else { | 3899 } else { |
| 3887 assert(typeNumElements(SrcTy) == 8 || typeNumElements(SrcTy) == 16); | 3900 assert(typeNumElements(SrcTy) == 8 || typeNumElements(SrcTy) == 16); |
| 3888 Type SignExtTy = Condition->getType() == IceType_v8i1 ? IceType_v8i16 | 3901 Type SignExtTy = Condition->getType() == IceType_v8i1 ? IceType_v8i16 |
| 3889 : IceType_v16i8; | 3902 : IceType_v16i8; |
| 3890 Variable *xmm0 = makeReg(SignExtTy, Reg_xmm0); | 3903 Variable *xmm0 = makeReg(SignExtTy, Reg_xmm0); |
| 3891 lowerCast(InstCast::create(Func, InstCast::Sext, xmm0, Condition)); | 3904 lowerCast(InstCast::create(Func, InstCast::Sext, xmm0, Condition)); |
| 3892 _movp(T, SrcFRM); | 3905 _movp(T, SrcFRM); |
| (...skipping 94 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 3987 Variable *Base = llvm::dyn_cast<Variable>(Addr); | 4000 Variable *Base = llvm::dyn_cast<Variable>(Addr); |
| 3988 // Vanilla ICE store instructions should not use the segment registers, | 4001 // Vanilla ICE store instructions should not use the segment registers, |
| 3989 // and computeAddressOpt only works at the level of Variables and Constants, | 4002 // and computeAddressOpt only works at the level of Variables and Constants, |
| 3990 // not other OperandX8632Mem, so there should be no mention of segment | 4003 // not other OperandX8632Mem, so there should be no mention of segment |
| 3991 // registers there either. | 4004 // registers there either. |
| 3992 const OperandX8632Mem::SegmentRegisters SegmentReg = | 4005 const OperandX8632Mem::SegmentRegisters SegmentReg = |
| 3993 OperandX8632Mem::DefaultSegment; | 4006 OperandX8632Mem::DefaultSegment; |
| 3994 computeAddressOpt(Func, Inst, Base, Index, Shift, Offset); | 4007 computeAddressOpt(Func, Inst, Base, Index, Shift, Offset); |
| 3995 if (Base && Addr != Base) { | 4008 if (Base && Addr != Base) { |
| 3996 Inst->setDeleted(); | 4009 Inst->setDeleted(); |
| 3997 Constant *OffsetOp = Ctx->getConstantInt(IceType_i32, Offset); | 4010 Constant *OffsetOp = Ctx->getConstantInt32(IceType_i32, Offset); |
| 3998 Addr = OperandX8632Mem::create(Func, Data->getType(), Base, OffsetOp, Index, | 4011 Addr = OperandX8632Mem::create(Func, Data->getType(), Base, OffsetOp, Index, |
| 3999 Shift, SegmentReg); | 4012 Shift, SegmentReg); |
| 4000 Context.insert(InstStore::create(Func, Data, Addr)); | 4013 Context.insert(InstStore::create(Func, Data, Addr)); |
| 4001 } | 4014 } |
| 4002 } | 4015 } |
| 4003 | 4016 |
| 4004 void TargetX8632::lowerSwitch(const InstSwitch *Inst) { | 4017 void TargetX8632::lowerSwitch(const InstSwitch *Inst) { |
| 4005 // This implements the most naive possible lowering. | 4018 // This implements the most naive possible lowering. |
| 4006 // cmp a,val[0]; jeq label[0]; cmp a,val[1]; jeq label[1]; ... jmp default | 4019 // cmp a,val[0]; jeq label[0]; cmp a,val[1]; jeq label[1]; ... jmp default |
| 4007 Operand *Src0 = Inst->getComparison(); | 4020 Operand *Src0 = Inst->getComparison(); |
| 4008 SizeT NumCases = Inst->getNumCases(); | 4021 SizeT NumCases = Inst->getNumCases(); |
| 4009 // OK, we'll be slightly less naive by forcing Src into a physical | 4022 // OK, we'll be slightly less naive by forcing Src into a physical |
| 4010 // register if there are 2 or more uses. | 4023 // register if there are 2 or more uses. |
| 4011 if (NumCases >= 2) | 4024 if (NumCases >= 2) |
| 4012 Src0 = legalizeToVar(Src0, true); | 4025 Src0 = legalizeToVar(Src0, true); |
| 4013 else | 4026 else |
| 4014 Src0 = legalize(Src0, Legal_Reg | Legal_Mem, true); | 4027 Src0 = legalize(Src0, Legal_Reg | Legal_Mem, true); |
| 4015 for (SizeT I = 0; I < NumCases; ++I) { | 4028 for (SizeT I = 0; I < NumCases; ++I) { |
| 4016 // TODO(stichnot): Correct lowering for IceType_i64. | 4029 // TODO(stichnot): Correct lowering for IceType_i64. |
| 4017 Constant *Value = Ctx->getConstantInt(IceType_i32, Inst->getValue(I)); | 4030 Constant *Value = Ctx->getConstantInt32(IceType_i32, Inst->getValue(I)); |
| 4018 _cmp(Src0, Value); | 4031 _cmp(Src0, Value); |
| 4019 _br(InstX8632Br::Br_e, Inst->getLabel(I)); | 4032 _br(InstX8632Br::Br_e, Inst->getLabel(I)); |
| 4020 } | 4033 } |
| 4021 | 4034 |
| 4022 _br(Inst->getLabelDefault()); | 4035 _br(Inst->getLabelDefault()); |
| 4023 } | 4036 } |
| 4024 | 4037 |
| 4025 void TargetX8632::scalarizeArithmetic(InstArithmetic::OpKind Kind, | 4038 void TargetX8632::scalarizeArithmetic(InstArithmetic::OpKind Kind, |
| 4026 Variable *Dest, Operand *Src0, | 4039 Variable *Dest, Operand *Src0, |
| 4027 Operand *Src1) { | 4040 Operand *Src1) { |
| 4028 assert(isVectorType(Dest->getType())); | 4041 assert(isVectorType(Dest->getType())); |
| 4029 Type Ty = Dest->getType(); | 4042 Type Ty = Dest->getType(); |
| 4030 Type ElementTy = typeElementType(Ty); | 4043 Type ElementTy = typeElementType(Ty); |
| 4031 SizeT NumElements = typeNumElements(Ty); | 4044 SizeT NumElements = typeNumElements(Ty); |
| 4032 | 4045 |
| 4033 Operand *T = Ctx->getConstantUndef(Ty); | 4046 Operand *T = Ctx->getConstantUndef(Ty); |
| 4034 for (SizeT I = 0; I < NumElements; ++I) { | 4047 for (SizeT I = 0; I < NumElements; ++I) { |
| 4035 Constant *Index = Ctx->getConstantInt(IceType_i32, I); | 4048 Constant *Index = Ctx->getConstantInt32(IceType_i32, I); |
| 4036 | 4049 |
| 4037 // Extract the next two inputs. | 4050 // Extract the next two inputs. |
| 4038 Variable *Op0 = Func->makeVariable(ElementTy, Context.getNode()); | 4051 Variable *Op0 = Func->makeVariable(ElementTy, Context.getNode()); |
| 4039 lowerExtractElement(InstExtractElement::create(Func, Op0, Src0, Index)); | 4052 lowerExtractElement(InstExtractElement::create(Func, Op0, Src0, Index)); |
| 4040 Variable *Op1 = Func->makeVariable(ElementTy, Context.getNode()); | 4053 Variable *Op1 = Func->makeVariable(ElementTy, Context.getNode()); |
| 4041 lowerExtractElement(InstExtractElement::create(Func, Op1, Src1, Index)); | 4054 lowerExtractElement(InstExtractElement::create(Func, Op1, Src1, Index)); |
| 4042 | 4055 |
| 4043 // Perform the arithmetic as a scalar operation. | 4056 // Perform the arithmetic as a scalar operation. |
| 4044 Variable *Res = Func->makeVariable(ElementTy, Context.getNode()); | 4057 Variable *Res = Func->makeVariable(ElementTy, Context.getNode()); |
| 4045 lowerArithmetic(InstArithmetic::create(Func, Kind, Res, Op0, Op1)); | 4058 lowerArithmetic(InstArithmetic::create(Func, Kind, Res, Op0, Op1)); |
| (...skipping 68 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 4114 _psub(Dest, MinusOne); | 4127 _psub(Dest, MinusOne); |
| 4115 return Dest; | 4128 return Dest; |
| 4116 } | 4129 } |
| 4117 | 4130 |
| 4118 Variable *TargetX8632::makeVectorOfHighOrderBits(Type Ty, int32_t RegNum) { | 4131 Variable *TargetX8632::makeVectorOfHighOrderBits(Type Ty, int32_t RegNum) { |
| 4119 assert(Ty == IceType_v4i32 || Ty == IceType_v4f32 || Ty == IceType_v8i16 || | 4132 assert(Ty == IceType_v4i32 || Ty == IceType_v4f32 || Ty == IceType_v8i16 || |
| 4120 Ty == IceType_v16i8); | 4133 Ty == IceType_v16i8); |
| 4121 if (Ty == IceType_v4f32 || Ty == IceType_v4i32 || Ty == IceType_v8i16) { | 4134 if (Ty == IceType_v4f32 || Ty == IceType_v4i32 || Ty == IceType_v8i16) { |
| 4122 Variable *Reg = makeVectorOfOnes(Ty, RegNum); | 4135 Variable *Reg = makeVectorOfOnes(Ty, RegNum); |
| 4123 SizeT Shift = typeWidthInBytes(typeElementType(Ty)) * X86_CHAR_BIT - 1; | 4136 SizeT Shift = typeWidthInBytes(typeElementType(Ty)) * X86_CHAR_BIT - 1; |
| 4124 _psll(Reg, Ctx->getConstantInt(IceType_i8, Shift)); | 4137 _psll(Reg, Ctx->getConstantInt32(IceType_i8, Shift)); |
| 4125 return Reg; | 4138 return Reg; |
| 4126 } else { | 4139 } else { |
| 4127 // SSE has no left shift operation for vectors of 8 bit integers. | 4140 // SSE has no left shift operation for vectors of 8 bit integers. |
| 4128 const uint32_t HIGH_ORDER_BITS_MASK = 0x80808080; | 4141 const uint32_t HIGH_ORDER_BITS_MASK = 0x80808080; |
| 4129 Constant *ConstantMask = | 4142 Constant *ConstantMask = |
| 4130 Ctx->getConstantInt(IceType_i32, HIGH_ORDER_BITS_MASK); | 4143 Ctx->getConstantInt32(IceType_i32, HIGH_ORDER_BITS_MASK); |
| 4131 Variable *Reg = makeReg(Ty, RegNum); | 4144 Variable *Reg = makeReg(Ty, RegNum); |
| 4132 _movd(Reg, legalize(ConstantMask, Legal_Reg | Legal_Mem)); | 4145 _movd(Reg, legalize(ConstantMask, Legal_Reg | Legal_Mem)); |
| 4133 _pshufd(Reg, Reg, Ctx->getConstantZero(IceType_i8)); | 4146 _pshufd(Reg, Reg, Ctx->getConstantZero(IceType_i8)); |
| 4134 return Reg; | 4147 return Reg; |
| 4135 } | 4148 } |
| 4136 } | 4149 } |
| 4137 | 4150 |
| 4138 OperandX8632Mem *TargetX8632::getMemoryOperandForStackSlot(Type Ty, | 4151 OperandX8632Mem *TargetX8632::getMemoryOperandForStackSlot(Type Ty, |
| 4139 Variable *Slot, | 4152 Variable *Slot, |
| 4140 uint32_t Offset) { | 4153 uint32_t Offset) { |
| 4141 // Ensure that Loc is a stack slot. | 4154 // Ensure that Loc is a stack slot. |
| 4142 assert(Slot->getWeight() == RegWeight::Zero); | 4155 assert(Slot->getWeight() == RegWeight::Zero); |
| 4143 assert(Slot->getRegNum() == Variable::NoRegister); | 4156 assert(Slot->getRegNum() == Variable::NoRegister); |
| 4144 // Compute the location of Loc in memory. | 4157 // Compute the location of Loc in memory. |
| 4145 // TODO(wala,stichnot): lea should not be required. The address of | 4158 // TODO(wala,stichnot): lea should not be required. The address of |
| 4146 // the stack slot is known at compile time (although not until after | 4159 // the stack slot is known at compile time (although not until after |
| 4147 // addProlog()). | 4160 // addProlog()). |
| 4148 const Type PointerType = IceType_i32; | 4161 const Type PointerType = IceType_i32; |
| 4149 Variable *Loc = makeReg(PointerType); | 4162 Variable *Loc = makeReg(PointerType); |
| 4150 _lea(Loc, Slot); | 4163 _lea(Loc, Slot); |
| 4151 Constant *ConstantOffset = Ctx->getConstantInt(IceType_i32, Offset); | 4164 Constant *ConstantOffset = Ctx->getConstantInt32(IceType_i32, Offset); |
| 4152 return OperandX8632Mem::create(Func, Ty, Loc, ConstantOffset); | 4165 return OperandX8632Mem::create(Func, Ty, Loc, ConstantOffset); |
| 4153 } | 4166 } |
| 4154 | 4167 |
| 4155 // Helper for legalize() to emit the right code to lower an operand to a | 4168 // Helper for legalize() to emit the right code to lower an operand to a |
| 4156 // register of the appropriate type. | 4169 // register of the appropriate type. |
| 4157 Variable *TargetX8632::copyToReg(Operand *Src, int32_t RegNum) { | 4170 Variable *TargetX8632::copyToReg(Operand *Src, int32_t RegNum) { |
| 4158 Type Ty = Src->getType(); | 4171 Type Ty = Src->getType(); |
| 4159 Variable *Reg = makeReg(Ty, RegNum); | 4172 Variable *Reg = makeReg(Ty, RegNum); |
| 4160 if (isVectorType(Ty)) { | 4173 if (isVectorType(Ty)) { |
| 4161 _movp(Reg, Src); | 4174 _movp(Reg, Src); |
| (...skipping 109 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 4271 OperandX8632Mem *TargetX8632::FormMemoryOperand(Operand *Operand, Type Ty) { | 4284 OperandX8632Mem *TargetX8632::FormMemoryOperand(Operand *Operand, Type Ty) { |
| 4272 OperandX8632Mem *Mem = llvm::dyn_cast<OperandX8632Mem>(Operand); | 4285 OperandX8632Mem *Mem = llvm::dyn_cast<OperandX8632Mem>(Operand); |
| 4273 // It may be the case that address mode optimization already creates | 4286 // It may be the case that address mode optimization already creates |
| 4274 // an OperandX8632Mem, so in that case it wouldn't need another level | 4287 // an OperandX8632Mem, so in that case it wouldn't need another level |
| 4275 // of transformation. | 4288 // of transformation. |
| 4276 if (!Mem) { | 4289 if (!Mem) { |
| 4277 Variable *Base = llvm::dyn_cast<Variable>(Operand); | 4290 Variable *Base = llvm::dyn_cast<Variable>(Operand); |
| 4278 Constant *Offset = llvm::dyn_cast<Constant>(Operand); | 4291 Constant *Offset = llvm::dyn_cast<Constant>(Operand); |
| 4279 assert(Base || Offset); | 4292 assert(Base || Offset); |
| 4280 if (Offset) { | 4293 if (Offset) { |
| 4281 assert(llvm::isa<ConstantInteger>(Offset) || | 4294 assert(llvm::isa<ConstantInteger32>(Offset) || |
| 4282 llvm::isa<ConstantRelocatable>(Offset)); | 4295 llvm::isa<ConstantRelocatable>(Offset)); |
| 4283 } | 4296 } |
| 4284 Mem = OperandX8632Mem::create(Func, Ty, Base, Offset); | 4297 Mem = OperandX8632Mem::create(Func, Ty, Base, Offset); |
| 4285 } | 4298 } |
| 4286 return llvm::cast<OperandX8632Mem>(legalize(Mem)); | 4299 return llvm::cast<OperandX8632Mem>(legalize(Mem)); |
| 4287 } | 4300 } |
| 4288 | 4301 |
| 4289 Variable *TargetX8632::makeReg(Type Type, int32_t RegNum) { | 4302 Variable *TargetX8632::makeReg(Type Type, int32_t RegNum) { |
| 4290 // There aren't any 64-bit integer registers for x86-32. | 4303 // There aren't any 64-bit integer registers for x86-32. |
| 4291 assert(Type != IceType_i64); | 4304 assert(Type != IceType_i64); |
| (...skipping 103 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 4395 if (WhiteList[RegNum]) | 4408 if (WhiteList[RegNum]) |
| 4396 FreedRegisters[RegNum] = true; | 4409 FreedRegisters[RegNum] = true; |
| 4397 } | 4410 } |
| 4398 } | 4411 } |
| 4399 } | 4412 } |
| 4400 } | 4413 } |
| 4401 AvailableRegisters |= FreedRegisters; | 4414 AvailableRegisters |= FreedRegisters; |
| 4402 } | 4415 } |
| 4403 } | 4416 } |
| 4404 | 4417 |
| 4405 template <> void ConstantInteger::emit(GlobalContext *Ctx) const { | 4418 template <> void ConstantInteger32::emit(GlobalContext *Ctx) const { |
| 4406 Ostream &Str = Ctx->getStrEmit(); | 4419 Ostream &Str = Ctx->getStrEmit(); |
| 4407 Str << (int64_t) getValue(); | 4420 Str << (int32_t)getValue(); |
| 4421 } |
| 4422 |
| 4423 template <> void ConstantInteger64::emit(GlobalContext *) const { |
| 4424 llvm_unreachable("Not expecting to emit 64-bit integers"); |
| 4408 } | 4425 } |
| 4409 | 4426 |
| 4410 template <> void ConstantFloat::emit(GlobalContext *Ctx) const { | 4427 template <> void ConstantFloat::emit(GlobalContext *Ctx) const { |
| 4411 Ostream &Str = Ctx->getStrEmit(); | 4428 Ostream &Str = Ctx->getStrEmit(); |
| 4412 // It would be better to prefix with ".L$" instead of "L$", but | 4429 // It would be better to prefix with ".L$" instead of "L$", but |
| 4413 // llvm-mc doesn't parse "dword ptr [.L$foo]". | 4430 // llvm-mc doesn't parse "dword ptr [.L$foo]". |
| 4414 Str << "dword ptr [L$" << IceType_f32 << "$" << getPoolEntryID() << "]"; | 4431 Str << "dword ptr [L$" << IceType_f32 << "$" << getPoolEntryID() << "]"; |
| 4415 } | 4432 } |
| 4416 | 4433 |
| 4417 template <> void ConstantDouble::emit(GlobalContext *Ctx) const { | 4434 template <> void ConstantDouble::emit(GlobalContext *Ctx) const { |
| (...skipping 95 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 4513 Str << "\t.align\t" << Align << "\n"; | 4530 Str << "\t.align\t" << Align << "\n"; |
| 4514 Str << MangledName << ":\n"; | 4531 Str << MangledName << ":\n"; |
| 4515 for (SizeT i = 0; i < Size; ++i) { | 4532 for (SizeT i = 0; i < Size; ++i) { |
| 4516 Str << "\t.byte\t" << (((unsigned)Data[i]) & 0xff) << "\n"; | 4533 Str << "\t.byte\t" << (((unsigned)Data[i]) & 0xff) << "\n"; |
| 4517 } | 4534 } |
| 4518 Str << "\t.size\t" << MangledName << ", " << Size << "\n"; | 4535 Str << "\t.size\t" << MangledName << ", " << Size << "\n"; |
| 4519 } | 4536 } |
| 4520 } | 4537 } |
| 4521 | 4538 |
| 4522 } // end of namespace Ice | 4539 } // end of namespace Ice |
| OLD | NEW |