[llvm] Add IRBuilder::CreateFMA (PR #131112)
Frederik Harwath via llvm-commits
llvm-commits at lists.llvm.org
Thu Mar 13 07:02:40 PDT 2025
https://github.com/frederik-h updated https://github.com/llvm/llvm-project/pull/131112
>From f1ce369163e842e4f44172bfdc9ad53d48d1f0b4 Mon Sep 17 00:00:00 2001
From: Frederik Harwath <fharwath at amd.com>
Date: Thu, 13 Mar 2025 05:11:54 -0400
Subject: [PATCH 1/5] Add IRBuilder::CreateFMA
- Adjust some existing CreateIntrinsic uses appropriately
- Adjust IRBuilderTest.cpp and remove duplicate test case
---
llvm/include/llvm/IR/IRBuilder.h | 7 +++++++
llvm/lib/IR/AutoUpgrade.cpp | 4 ++--
llvm/unittests/IR/IRBuilderTest.cpp | 12 ++----------
3 files changed, 11 insertions(+), 12 deletions(-)
diff --git a/llvm/include/llvm/IR/IRBuilder.h b/llvm/include/llvm/IR/IRBuilder.h
index 67e357c600d3b..13665dcd992a8 100644
--- a/llvm/include/llvm/IR/IRBuilder.h
+++ b/llvm/include/llvm/IR/IRBuilder.h
@@ -1065,6 +1065,13 @@ class IRBuilderBase {
{Src, Exp}, FMFSource, Name);
}
+ /// Create call to the fma intrinsic.
+ Value *CreateFMA(Value *Factor1, Value *Factor2, Value *Summand,
+ FMFSource FMFSource = {}, const Twine &Name = "") {
+ return CreateIntrinsic(Intrinsic::fma, {Factor1->getType()},
+ {Factor1, Factor2, Summand}, FMFSource, Name);
+ }
+
/// Create a call to the arithmetic_fence intrinsic.
CallInst *CreateArithmeticFence(Value *Val, Type *DstType,
const Twine &Name = "") {
diff --git a/llvm/lib/IR/AutoUpgrade.cpp b/llvm/lib/IR/AutoUpgrade.cpp
index cb4ecc60aa473..ce3b2c90a41a1 100644
--- a/llvm/lib/IR/AutoUpgrade.cpp
+++ b/llvm/lib/IR/AutoUpgrade.cpp
@@ -3755,7 +3755,7 @@ static Value *upgradeX86IntrinsicCall(StringRef Name, CallBase *CI, Function *F,
IID = Intrinsic::x86_avx512_vfmadd_f32;
Rep = Builder.CreateIntrinsic(IID, {}, Ops);
} else {
- Rep = Builder.CreateIntrinsic(Intrinsic::fma, A->getType(), {A, B, C});
+ Rep = Builder.CreateFMA(A, B, C);
}
Value *PassThru = IsMaskZ ? Constant::getNullValue(Rep->getType())
@@ -3808,7 +3808,7 @@ static Value *upgradeX86IntrinsicCall(StringRef Name, CallBase *CI, Function *F,
Rep = Builder.CreateIntrinsic(IID, {}, {A, B, C, CI->getArgOperand(4)});
} else {
- Rep = Builder.CreateIntrinsic(Intrinsic::fma, A->getType(), {A, B, C});
+ Rep = Builder.CreateFMA(A, B, C);
}
Value *PassThru = IsMaskZ ? llvm::Constant::getNullValue(CI->getType())
diff --git a/llvm/unittests/IR/IRBuilderTest.cpp b/llvm/unittests/IR/IRBuilderTest.cpp
index 3a55d88f03d49..2170524d6eb3c 100644
--- a/llvm/unittests/IR/IRBuilderTest.cpp
+++ b/llvm/unittests/IR/IRBuilderTest.cpp
@@ -109,21 +109,13 @@ TEST_F(IRBuilderTest, Intrinsics) {
EXPECT_TRUE(II->hasNoInfs());
EXPECT_FALSE(II->hasNoNaNs());
- Result = Builder.CreateIntrinsic(Intrinsic::fma, {V->getType()}, {V, V, V});
+ Result = Builder.CreateFMA(V, V, V);
II = cast<IntrinsicInst>(Result);
EXPECT_EQ(II->getIntrinsicID(), Intrinsic::fma);
EXPECT_FALSE(II->hasNoInfs());
EXPECT_FALSE(II->hasNoNaNs());
- Result =
- Builder.CreateIntrinsic(Intrinsic::fma, {V->getType()}, {V, V, V}, I);
- II = cast<IntrinsicInst>(Result);
- EXPECT_EQ(II->getIntrinsicID(), Intrinsic::fma);
- EXPECT_TRUE(II->hasNoInfs());
- EXPECT_FALSE(II->hasNoNaNs());
-
- Result =
- Builder.CreateIntrinsic(Intrinsic::fma, {V->getType()}, {V, V, V}, I);
+ Result = Builder.CreateFMA(V, V, V, I);
II = cast<IntrinsicInst>(Result);
EXPECT_EQ(II->getIntrinsicID(), Intrinsic::fma);
EXPECT_TRUE(II->hasNoInfs());
>From 1191eea7e92e4549c80385876c116de6a2b7975d Mon Sep 17 00:00:00 2001
From: Frederik Harwath <fharwath at amd.com>
Date: Thu, 13 Mar 2025 06:34:59 -0400
Subject: [PATCH 2/5] CreateFMA: Add assert to document that
experimental.constrained.fma is not supported
---
llvm/include/llvm/IR/IRBuilder.h | 1 +
1 file changed, 1 insertion(+)
diff --git a/llvm/include/llvm/IR/IRBuilder.h b/llvm/include/llvm/IR/IRBuilder.h
index 13665dcd992a8..389c7882502af 100644
--- a/llvm/include/llvm/IR/IRBuilder.h
+++ b/llvm/include/llvm/IR/IRBuilder.h
@@ -1068,6 +1068,7 @@ class IRBuilderBase {
/// Create call to the fma intrinsic.
Value *CreateFMA(Value *Factor1, Value *Factor2, Value *Summand,
FMFSource FMFSource = {}, const Twine &Name = "") {
+ assert(!IsFPConstrained && "TODO: Support experimental_constrained_fma");
return CreateIntrinsic(Intrinsic::fma, {Factor1->getType()},
{Factor1, Factor2, Summand}, FMFSource, Name);
}
>From 22f87c11f3a77b5a5274fbeeab1b324c942dfa82 Mon Sep 17 00:00:00 2001
From: Frederik Harwath <fharwath at amd.com>
Date: Thu, 13 Mar 2025 09:49:50 -0400
Subject: [PATCH 3/5] Add IRBuilder::CreateConstraintedFPIntrinsic
This is derived from CreateConstrainedFPBinOp but supports arbitrary
intrinsics.
---
llvm/include/llvm/IR/IRBuilder.h | 11 +++++++++++
llvm/lib/IR/IRBuilder.cpp | 20 ++++++++++++++++++++
llvm/unittests/IR/IRBuilderTest.cpp | 9 +++++++++
3 files changed, 40 insertions(+)
diff --git a/llvm/include/llvm/IR/IRBuilder.h b/llvm/include/llvm/IR/IRBuilder.h
index 389c7882502af..a9e1cb9bb5235 100644
--- a/llvm/include/llvm/IR/IRBuilder.h
+++ b/llvm/include/llvm/IR/IRBuilder.h
@@ -1731,6 +1731,17 @@ class IRBuilderBase {
return Accum;
}
+ /// This function is like @ref CreateIntrinsic for constrained fp
+ /// intrinsics. It sets the rounding mode and exception behavior of
+ /// the created intrinsic call according to \p Rounding and \p
+ /// Except and it sets \p FPMathTag as the 'fpmath' metadata, using
+ /// defaults if a value equals nullopt/null.
+ CallInst *CreateConstrainedFPIntrinsic(
+ Intrinsic::ID ID, ArrayRef<Type *> Types, ArrayRef<Value *> Args,
+ FMFSource FMFSource, const Twine &Name, MDNode *FPMathTag = nullptr,
+ std::optional<RoundingMode> Rounding = std::nullopt,
+ std::optional<fp::ExceptionBehavior> Except = std::nullopt);
+
CallInst *CreateConstrainedFPBinOp(
Intrinsic::ID ID, Value *L, Value *R, FMFSource FMFSource = {},
const Twine &Name = "", MDNode *FPMathTag = nullptr,
diff --git a/llvm/lib/IR/IRBuilder.cpp b/llvm/lib/IR/IRBuilder.cpp
index 134459265cecb..421b617a5fb7e 100644
--- a/llvm/lib/IR/IRBuilder.cpp
+++ b/llvm/lib/IR/IRBuilder.cpp
@@ -950,6 +950,26 @@ CallInst *IRBuilderBase::CreateConstrainedFPBinOp(
return C;
}
+CallInst *IRBuilderBase::CreateConstrainedFPIntrinsic(
+ Intrinsic::ID ID, ArrayRef<Type *> Types, ArrayRef<Value *> Args,
+ FMFSource FMFSource, const Twine &Name, MDNode *FPMathTag,
+ std::optional<RoundingMode> Rounding,
+ std::optional<fp::ExceptionBehavior> Except) {
+ Value *RoundingV = getConstrainedFPRounding(Rounding);
+ Value *ExceptV = getConstrainedFPExcept(Except);
+
+ FastMathFlags UseFMF = FMFSource.get(FMF);
+
+ llvm::SmallVector<Value *, 5> ExtArgs(Args);
+ ExtArgs.push_back(RoundingV);
+ ExtArgs.push_back(ExceptV);
+
+ CallInst *C = CreateIntrinsic(ID, Types, ExtArgs, nullptr, Name);
+ setConstrainedFPCallAttr(C);
+ setFPAttrs(C, FPMathTag, UseFMF);
+ return C;
+}
+
CallInst *IRBuilderBase::CreateConstrainedFPUnroundedBinOp(
Intrinsic::ID ID, Value *L, Value *R, FMFSource FMFSource,
const Twine &Name, MDNode *FPMathTag,
diff --git a/llvm/unittests/IR/IRBuilderTest.cpp b/llvm/unittests/IR/IRBuilderTest.cpp
index 2170524d6eb3c..321fbd358bce5 100644
--- a/llvm/unittests/IR/IRBuilderTest.cpp
+++ b/llvm/unittests/IR/IRBuilderTest.cpp
@@ -390,6 +390,15 @@ TEST_F(IRBuilderTest, ConstrainedFP) {
EXPECT_EQ(fp::ebMayTrap, CII->getExceptionBehavior());
EXPECT_EQ(RoundingMode::TowardNegative, CII->getRoundingMode());
+ // Same as previous test for CreateConstrainedFPIntrinsic
+ Call = Builder.CreateConstrainedFPIntrinsic(
+ Intrinsic::experimental_constrained_fadd, {V->getType()}, {V, V}, nullptr,
+ "", nullptr, RoundingMode::TowardNegative, fp::ebMayTrap);
+ CII = cast<ConstrainedFPIntrinsic>(Call);
+ EXPECT_EQ(CII->getIntrinsicID(), Intrinsic::experimental_constrained_fadd);
+ EXPECT_EQ(fp::ebMayTrap, CII->getExceptionBehavior());
+ EXPECT_EQ(RoundingMode::TowardNegative, CII->getRoundingMode());
+
Builder.CreateRetVoid();
EXPECT_FALSE(verifyModule(*M));
}
>From c08871d3c91135f3b245618a24add2b99954d6e1 Mon Sep 17 00:00:00 2001
From: Frederik Harwath <fharwath at amd.com>
Date: Thu, 13 Mar 2025 09:52:44 -0400
Subject: [PATCH 4/5] Add experimental_constrained_fma support to
IRBuilder::CreateFMA
---
llvm/include/llvm/IR/IRBuilder.h | 7 ++++++-
llvm/unittests/IR/IRBuilderTest.cpp | 14 ++++++++++++++
2 files changed, 20 insertions(+), 1 deletion(-)
diff --git a/llvm/include/llvm/IR/IRBuilder.h b/llvm/include/llvm/IR/IRBuilder.h
index a9e1cb9bb5235..750a99cc50dd7 100644
--- a/llvm/include/llvm/IR/IRBuilder.h
+++ b/llvm/include/llvm/IR/IRBuilder.h
@@ -1068,7 +1068,12 @@ class IRBuilderBase {
/// Create call to the fma intrinsic.
Value *CreateFMA(Value *Factor1, Value *Factor2, Value *Summand,
FMFSource FMFSource = {}, const Twine &Name = "") {
- assert(!IsFPConstrained && "TODO: Support experimental_constrained_fma");
+ if (IsFPConstrained) {
+ return CreateConstrainedFPIntrinsic(
+ Intrinsic::experimental_constrained_fma, {Factor1->getType()},
+ {Factor1, Factor2, Summand}, FMFSource, Name);
+ }
+
return CreateIntrinsic(Intrinsic::fma, {Factor1->getType()},
{Factor1, Factor2, Summand}, FMFSource, Name);
}
diff --git a/llvm/unittests/IR/IRBuilderTest.cpp b/llvm/unittests/IR/IRBuilderTest.cpp
index 321fbd358bce5..95534ea2d6de6 100644
--- a/llvm/unittests/IR/IRBuilderTest.cpp
+++ b/llvm/unittests/IR/IRBuilderTest.cpp
@@ -121,6 +121,15 @@ TEST_F(IRBuilderTest, Intrinsics) {
EXPECT_TRUE(II->hasNoInfs());
EXPECT_FALSE(II->hasNoNaNs());
+ FastMathFlags SavedFMF = Builder.getFastMathFlags();
+ Builder.setFastMathFlags(FastMathFlags::getFast());
+ Result = Builder.CreateFMA(V, V, V);
+ II = cast<IntrinsicInst>(Result);
+ EXPECT_EQ(II->getIntrinsicID(), Intrinsic::fma);
+ EXPECT_TRUE(II->hasNoInfs());
+ EXPECT_TRUE(II->hasNoNaNs());
+ Builder.setFastMathFlags(SavedFMF);
+
Result = Builder.CreateUnaryIntrinsic(Intrinsic::roundeven, V);
II = cast<IntrinsicInst>(Result);
EXPECT_EQ(II->getIntrinsicID(), Intrinsic::roundeven);
@@ -299,6 +308,11 @@ TEST_F(IRBuilderTest, ConstrainedFP) {
II = cast<IntrinsicInst>(V);
EXPECT_EQ(II->getIntrinsicID(), Intrinsic::experimental_constrained_frem);
+ V = Builder.CreateFMA(V, V, V);
+ ASSERT_TRUE(isa<IntrinsicInst>(V));
+ II = cast<IntrinsicInst>(V);
+ EXPECT_EQ(II->getIntrinsicID(), Intrinsic::experimental_constrained_fma);
+
VInt = Builder.CreateFPToUI(VDouble, Builder.getInt32Ty());
ASSERT_TRUE(isa<IntrinsicInst>(VInt));
II = cast<IntrinsicInst>(VInt);
>From 208b91739e9ec6b92e42f83e4409d0d6ebf4072c Mon Sep 17 00:00:00 2001
From: Frederik Harwath <fharwath at amd.com>
Date: Thu, 13 Mar 2025 10:00:19 -0400
Subject: [PATCH 5/5] IRBuilderTest: Pass FastMathFlags as FMFSource instead of
using the Builder
---
llvm/unittests/IR/IRBuilderTest.cpp | 5 +----
1 file changed, 1 insertion(+), 4 deletions(-)
diff --git a/llvm/unittests/IR/IRBuilderTest.cpp b/llvm/unittests/IR/IRBuilderTest.cpp
index 95534ea2d6de6..e9e9d7b11a36c 100644
--- a/llvm/unittests/IR/IRBuilderTest.cpp
+++ b/llvm/unittests/IR/IRBuilderTest.cpp
@@ -121,14 +121,11 @@ TEST_F(IRBuilderTest, Intrinsics) {
EXPECT_TRUE(II->hasNoInfs());
EXPECT_FALSE(II->hasNoNaNs());
- FastMathFlags SavedFMF = Builder.getFastMathFlags();
- Builder.setFastMathFlags(FastMathFlags::getFast());
- Result = Builder.CreateFMA(V, V, V);
+ Result = Builder.CreateFMA(V, V, V, FastMathFlags::getFast());
II = cast<IntrinsicInst>(Result);
EXPECT_EQ(II->getIntrinsicID(), Intrinsic::fma);
EXPECT_TRUE(II->hasNoInfs());
EXPECT_TRUE(II->hasNoNaNs());
- Builder.setFastMathFlags(SavedFMF);
Result = Builder.CreateUnaryIntrinsic(Intrinsic::roundeven, V);
II = cast<IntrinsicInst>(Result);
More information about the llvm-commits
mailing list