https://github.com/RKSimon updated
https://github.com/llvm/llvm-project/pull/157106
>From 4d6078f92740efa3f109b85492f43b8fadd39b03 Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-...@redking.me.uk>
Date: Fri, 5 Sep 2025 14:32:07 +0100
Subject: [PATCH 1/2] [Clang][bytecode] Add
interp__builtin_elementwise_triop_fp to handle general 3-operand floating
point intrinsics
Refactor interp__builtin_elementwise_fma into something similar to
interp__builtin_elementwise_triop with a callback function argument to allow
reuse with other intrinsics
This will allow reuse with some upcoming x86 intrinsics
---
clang/lib/AST/ByteCode/InterpBuiltin.cpp | 24 ++++++++++++++++--------
1 file changed, 16 insertions(+), 8 deletions(-)
diff --git a/clang/lib/AST/ByteCode/InterpBuiltin.cpp
b/clang/lib/AST/ByteCode/InterpBuiltin.cpp
index 713895b191c88..c0937ecb683bf 100644
--- a/clang/lib/AST/ByteCode/InterpBuiltin.cpp
+++ b/clang/lib/AST/ByteCode/InterpBuiltin.cpp
@@ -2736,12 +2736,14 @@ static bool interp__builtin_ia32_pmul(InterpState &S,
CodePtr OpPC,
return true;
}
-static bool interp__builtin_elementwise_fma(InterpState &S, CodePtr OpPC,
- const CallExpr *Call) {
+static bool interp__builtin_elementwise_triop_fp(
+ InterpState &S, CodePtr OpPC, const CallExpr *Call,
+ llvm::function_ref<APFloat(const APFloat &, const APFloat &,
+ const APFloat &, const FPOptions &)>
+ Fn) {
assert(Call->getNumArgs() == 3);
FPOptions FPO = Call->getFPFeaturesInEffect(S.Ctx.getLangOpts());
- llvm::RoundingMode RM = getRoundingMode(FPO);
const QualType Arg1Type = Call->getArg(0)->getType();
const QualType Arg2Type = Call->getArg(1)->getType();
const QualType Arg3Type = Call->getArg(2)->getType();
@@ -2756,8 +2758,7 @@ static bool interp__builtin_elementwise_fma(InterpState
&S, CodePtr OpPC,
const Floating &Z = S.Stk.pop<Floating>();
const Floating &Y = S.Stk.pop<Floating>();
const Floating &X = S.Stk.pop<Floating>();
- APFloat F = X.getAPFloat();
- F.fusedMultiplyAdd(Y.getAPFloat(), Z.getAPFloat(), RM);
+ APFloat F = Fn(X.getAPFloat(), Y.getAPFloat(), Z.getAPFloat(), FPO);
Floating Result = S.allocFloat(X.getSemantics());
Result.copy(F);
S.Stk.push<Floating>(Result);
@@ -2788,8 +2789,8 @@ static bool interp__builtin_elementwise_fma(InterpState
&S, CodePtr OpPC,
APFloat X = VX.elem<T>(I).getAPFloat();
APFloat Y = VY.elem<T>(I).getAPFloat();
APFloat Z = VZ.elem<T>(I).getAPFloat();
- (void)X.fusedMultiplyAdd(Y, Z, RM);
- Dst.elem<Floating>(I) = Floating(X);
+ APFloat F = Fn(X, Y, Z, FPO);
+ Dst.elem<Floating>(I) = Floating(F);
}
Dst.initializeAllElements();
return true;
@@ -3410,7 +3411,14 @@ bool InterpretBuiltin(InterpState &S, CodePtr OpPC,
const CallExpr *Call,
return interp__builtin_ia32_pmul(S, OpPC, Call, BuiltinID);
case Builtin::BI__builtin_elementwise_fma:
- return interp__builtin_elementwise_fma(S, OpPC, Call);
+ return interp__builtin_elementwise_triop_fp(
+ S, OpPC, Call,
+ [](const APFloat &X, const APFloat &Y, const APFloat &Z,
+ const FPOptions &FPO) {
+ APFloat F = X;
+ F.fusedMultiplyAdd(Y, Z, getRoundingMode(FPO));
+ return F;
+ });
case X86::BI__builtin_ia32_selectb_128:
case X86::BI__builtin_ia32_selectb_256:
>From d970c40289896ad388d08eb46e7aebf423a63ded Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-...@redking.me.uk>
Date: Fri, 5 Sep 2025 15:15:27 +0100
Subject: [PATCH 2/2] interp__builtin_elementwise_triop_fp - just pass
RoundingMode not all FPOptions
---
clang/lib/AST/ByteCode/InterpBuiltin.cpp | 11 ++++++-----
1 file changed, 6 insertions(+), 5 deletions(-)
diff --git a/clang/lib/AST/ByteCode/InterpBuiltin.cpp
b/clang/lib/AST/ByteCode/InterpBuiltin.cpp
index c0937ecb683bf..65c1761792de4 100644
--- a/clang/lib/AST/ByteCode/InterpBuiltin.cpp
+++ b/clang/lib/AST/ByteCode/InterpBuiltin.cpp
@@ -2739,11 +2739,12 @@ static bool interp__builtin_ia32_pmul(InterpState &S,
CodePtr OpPC,
static bool interp__builtin_elementwise_triop_fp(
InterpState &S, CodePtr OpPC, const CallExpr *Call,
llvm::function_ref<APFloat(const APFloat &, const APFloat &,
- const APFloat &, const FPOptions &)>
+ const APFloat &, llvm::RoundingMode)>
Fn) {
assert(Call->getNumArgs() == 3);
FPOptions FPO = Call->getFPFeaturesInEffect(S.Ctx.getLangOpts());
+ llvm::RoundingMode RM = getRoundingMode(FPO);
const QualType Arg1Type = Call->getArg(0)->getType();
const QualType Arg2Type = Call->getArg(1)->getType();
const QualType Arg3Type = Call->getArg(2)->getType();
@@ -2758,7 +2759,7 @@ static bool interp__builtin_elementwise_triop_fp(
const Floating &Z = S.Stk.pop<Floating>();
const Floating &Y = S.Stk.pop<Floating>();
const Floating &X = S.Stk.pop<Floating>();
- APFloat F = Fn(X.getAPFloat(), Y.getAPFloat(), Z.getAPFloat(), FPO);
+ APFloat F = Fn(X.getAPFloat(), Y.getAPFloat(), Z.getAPFloat(), RM);
Floating Result = S.allocFloat(X.getSemantics());
Result.copy(F);
S.Stk.push<Floating>(Result);
@@ -2789,7 +2790,7 @@ static bool interp__builtin_elementwise_triop_fp(
APFloat X = VX.elem<T>(I).getAPFloat();
APFloat Y = VY.elem<T>(I).getAPFloat();
APFloat Z = VZ.elem<T>(I).getAPFloat();
- APFloat F = Fn(X, Y, Z, FPO);
+ APFloat F = Fn(X, Y, Z, RM);
Dst.elem<Floating>(I) = Floating(F);
}
Dst.initializeAllElements();
@@ -3414,9 +3415,9 @@ bool InterpretBuiltin(InterpState &S, CodePtr OpPC, const
CallExpr *Call,
return interp__builtin_elementwise_triop_fp(
S, OpPC, Call,
[](const APFloat &X, const APFloat &Y, const APFloat &Z,
- const FPOptions &FPO) {
+ llvm::RoundingMode RM) {
APFloat F = X;
- F.fusedMultiplyAdd(Y, Z, getRoundingMode(FPO));
+ F.fusedMultiplyAdd(Y, Z, RM);
return F;
});
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
