Skip to content

Commit 156d2aa

Browse files
stomfaigstomfaig
committed
moving rounding functionality to helper
1 parent aab58a9 commit 156d2aa

File tree

1 file changed

+69
-86
lines changed
  • clang/lib/CodeGen/TargetBuiltins

1 file changed

+69
-86
lines changed

‎clang/lib/CodeGen/TargetBuiltins/X86.cpp‎

Lines changed: 69 additions & 86 deletions
Original file line numberDiff line numberDiff line change
@@ -75,6 +75,70 @@ static Value *getMaskVecValue(CodeGenFunction &CGF, Value *Mask,
7575
return MaskVec;
7676
}
7777

78+
static Value *emitX86Round(CodeGenFunction &CGF,
79+
Value *X,
80+
unsigned M) {
81+
unsigned RoundingMask = 0b11;
82+
unsigned UpdatePEBit = 0b100;
83+
unsigned UseMXCSRBit = 0b1000;
84+
85+
unsigned roundingMode = M & RoundingMask;
86+
bool updatePE = M & UpdatePEBit;
87+
bool useMXCSR = M & UseMXCSRBit;
88+
89+
Intrinsic::ID ID = Intrinsic::not_intrinsic;
90+
LLVMContext &Ctx = CGF.CGM.getLLVMContext();
91+
92+
if (useMXCSR) {
93+
ID = Intrinsic::experimental_constrained_nearbyint;
94+
95+
auto PE_metatadata = updatePE ? "fpexcept.strict" : "fpexcept.ignore";
96+
97+
Value *ExceptMode = MetadataAsValue::get(
98+
Ctx,
99+
MDString::get(Ctx, PE_metatadata)
100+
);
101+
102+
Value *RoundingMode = MetadataAsValue::get(
103+
Ctx,
104+
MDString::get(Ctx, "rounding.dynamic")
105+
);
106+
107+
Function *F = CGF.CGM.getIntrinsic(ID, X->getType());
108+
return CGF.Builder.CreateCall(F, {X, ExceptMode, RoundingMode});
109+
}
110+
111+
if (updatePE) {
112+
switch (roundingMode) {
113+
case 0b00: ID = Intrinsic::experimental_constrained_roundeven; break;
114+
case 0b01: ID = Intrinsic::experimental_constrained_floor; break;
115+
case 0b10: ID = Intrinsic::experimental_constrained_ceil; break;
116+
case 0b11: ID = Intrinsic::experimental_constrained_trunc; break;
117+
default: llvm_unreachable("Invalid rounding mode");
118+
}
119+
120+
Value *ExceptMode =MetadataAsValue::get(
121+
Ctx,
122+
MDString::get(Ctx, "fpexcept.strict")
123+
);
124+
125+
Function *F = CGF.CGM.getIntrinsic(ID, X->getType());
126+
return CGF.Builder.CreateCall(F, {X, ExceptMode});
127+
}
128+
129+
// Otherwise we can use the standard ops
130+
switch (roundingMode) {
131+
case 0b00: ID = Intrinsic::roundeven; break;
132+
case 0b01: ID = Intrinsic::floor; break;
133+
case 0b10: ID = Intrinsic::ceil; break;
134+
case 0b11: ID = Intrinsic::trunc; break;
135+
default: llvm_unreachable("Invalid rounding mode");
136+
}
137+
138+
Function *F = CGF.CGM.getIntrinsic(ID, X->getType());
139+
return CGF.Builder.CreateCall(F, {X});
140+
}
141+
78142
static Value *EmitX86MaskedStore(CodeGenFunction &CGF, ArrayRef<Value *> Ops,
79143
Align Alignment) {
80144
Value *Ptr = Ops[0];
@@ -843,100 +907,19 @@ Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID,
843907
case X86::BI__builtin_ia32_roundps:
844908
case X86::BI__builtin_ia32_roundpd:
845909
case X86::BI__builtin_ia32_roundps256:
846-
case X86::BI__builtin_ia32_roundpd256: {
910+
case X86::BI__builtin_ia32_roundpd256: {
847911
unsigned M = cast<ConstantInt>(Ops[1])->getZExtValue();
848-
unsigned roundingModeAndPE = M & 0b111;
849-
unsigned updatePE = M & 0b100;
850-
unsigned use_MXCSR = M & 0b1000;
851-
852-
Intrinsic::ID ID;
853-
854-
// Currently no ops for MXCSR bit set, so lower directly to SSE41 instruction
855-
if (use_MXCSR) {
856-
switch (BuiltinID) {
857-
case X86::BI__builtin_ia32_roundps: ID = Intrinsic::x86_sse41_round_ps; break;
858-
case X86::BI__builtin_ia32_roundpd: ID = Intrinsic::x86_sse41_round_pd; break;
859-
}
860-
return nullptr;
861-
} else {
862-
switch (roundingModeAndPE) {
863-
default: return nullptr;
864-
case 0b000: ID = Intrinsic::nearbyint; break;
865-
case 0b001: ID = Intrinsic::floor; break;
866-
case 0b010: ID = Intrinsic::ceil; break;
867-
case 0b011: ID = Intrinsic::trunc; break;
868-
case 0b100: ID = Intrinsic::experimental_constrained_floor; break; // TODO: replace with actual op
869-
case 0b101: ID = Intrinsic::experimental_constrained_floor; break;
870-
case 0b110: ID = Intrinsic::experimental_constrained_ceil; break;
871-
case 0b111: ID = Intrinsic::experimental_constrained_trunc; break;
872-
}
873-
}
874-
875-
Function *F = CGM.getIntrinsic(ID, Ops[0]->getType());
876-
877-
if (updatePE) {
878-
LLVMContext &Ctx = CGM.getLLVMContext();
879-
880-
Value *ExceptMode =MetadataAsValue::get(
881-
Ctx,
882-
MDString::get(Ctx, "fpexcept.strict")
883-
);
884-
885-
return Builder.CreateCall(F, {Ops[0], ExceptMode});
886-
}
887-
888-
return Builder.CreateCall(F, {Ops[0]});
912+
return emitX86Round(*this, Ops[0], M);
889913
}
890914
case X86::BI__builtin_ia32_roundss:
891915
case X86::BI__builtin_ia32_roundsd: {
892916
unsigned M = cast<ConstantInt>(Ops[2])->getZExtValue();
893-
unsigned roundingModeAndPE = M & 0b111;
894-
unsigned updatePE = M & 0b100;
895-
unsigned use_MXCSR = M & 0b1000;
896917

897-
Intrinsic::ID ID;
898-
899-
// Currently no ops for MXCSR bit set, so lower directly to SSE41 instruction
900-
if (use_MXCSR) {
901-
switch (BuiltinID) {
902-
case X86::BI__builtin_ia32_roundss: ID = Intrinsic::x86_sse41_round_ss; break;
903-
case X86::BI__builtin_ia32_roundsd: ID = Intrinsic::x86_sse41_round_sd; break;
904-
}
905-
return nullptr;
906-
} else {
907-
switch (roundingModeAndPE) {
908-
default: return nullptr;
909-
case 0b000: ID = Intrinsic::nearbyint; break;
910-
case 0b001: ID = Intrinsic::floor; break;
911-
case 0b010: ID = Intrinsic::ceil; break;
912-
case 0b011: ID = Intrinsic::trunc; break;
913-
case 0b100: ID = Intrinsic::experimental_constrained_floor; break; // TODO: replace with actual op
914-
case 0b101: ID = Intrinsic::experimental_constrained_floor; break;
915-
case 0b110: ID = Intrinsic::experimental_constrained_ceil; break;
916-
case 0b111: ID = Intrinsic::experimental_constrained_trunc; break;
917-
}
918-
}
919-
920918
Value *idx = Builder.getInt32(0);
921-
Value *b0 = Builder.CreateExtractElement(Ops[1], idx);
922-
Value *rounded0;
923-
924-
Function *F = CGM.getIntrinsic(ID, b0->getType());
925-
926-
if (updatePE) {
927-
LLVMContext &Ctx = CGM.getLLVMContext();
928-
929-
Value *ExceptMode =MetadataAsValue::get(
930-
Ctx,
931-
MDString::get(Ctx, "fpexcept.strict")
932-
);
933-
934-
rounded0 = Builder.CreateCall(F, {b0, ExceptMode});
935-
} else {
936-
rounded0 = Builder.CreateCall(F, {b0});
937-
}
919+
Value *ValAt0 = Builder.CreateExtractElement(Ops[1], idx);
920+
Value *RoundedAt0 = emitX86Round(*this, ValAt0, M);
938921

939-
return Builder.CreateInsertElement(Ops[0], rounded0, idx);
922+
return Builder.CreateInsertElement(Ops[0], RoundedAt0, idx);
940923
}
941924
case X86::BI__builtin_ia32_lzcnt_u16:
942925
case X86::BI__builtin_ia32_lzcnt_u32:

0 commit comments

Comments
 (0)