[SPIRV] Added support for the constrained arithmetic(Fmuladd) intrinsic

[SubashBoopathi]

Added SPIR-V support for constrained arithmetic intrinsic fmuladd, lowered as a sequence of OpFMul and OpFAdd with roundingmode, consistent with the SPIR-V translator.

[github-actions]

🐧 Linux x64 Test Results

• 186879 tests passed
• 4916 tests skipped

✅ The build succeeded and all tests passed.

[llvmbot]

@llvm/pr-subscribers-backend-spir-v

Author: Subash B (SubashBoopathi)

Changes

Added SPIR-V support for constrained arithmetic intrinsic fmuladd, lowered as a sequence of OpFMul and OpFAdd with roundingmode, consistent with the SPIR-V translator.

---

Full diff: https://github.com/llvm/llvm-project/pull/170270.diff

2 Files Affected:

• (modified) llvm/lib/Target/SPIRV/SPIRVPrepareFunctions.cpp (+22)
• (added) llvm/test/CodeGen/SPIRV/llvm-intrinsics/constrained-fmuladd.ll (+64)

diff --git a/llvm/lib/Target/SPIRV/SPIRVPrepareFunctions.cpp b/llvm/lib/Target/SPIRV/SPIRVPrepareFunctions.cpp
index be88f334d2171..01670d8cb4c6b 100644
--- a/llvm/lib/Target/SPIRV/SPIRVPrepareFunctions.cpp
+++ b/llvm/lib/Target/SPIRV/SPIRVPrepareFunctions.cpp
@@ -393,6 +393,24 @@ static bool toSpvLifetimeIntrinsic(IntrinsicInst *II, Intrinsic::ID NewID) {
   return true;
 }
 
+static void
+lowerConstrainedFmuladd(IntrinsicInst *II,
+                        SmallVector<Instruction *> &EraseFromParent) {
+  auto *FPI = cast<ConstrainedFPIntrinsic>(II);
+  Value *A = FPI->getArgOperand(0);
+  Value *Mul = FPI->getArgOperand(1);
+  Value *Add = FPI->getArgOperand(2);
+  IRBuilder<> Builder(II->getParent());
+  Builder.SetInsertPoint(II);
+  std::optional<RoundingMode> Rounding = FPI->getRoundingMode();
+  Value *Product = Builder.CreateFMul(A, Mul, II->getName() + ".mul");
+  Value *Result = Builder.CreateConstrainedFPBinOp(
+      Intrinsic::experimental_constrained_fadd, Product, Add, {},
+      II->getName() + ".add", nullptr, Rounding);
+  II->replaceAllUsesWith(Result);
+  EraseFromParent.push_back(II);
+}
+
 // Substitutes calls to LLVM intrinsics with either calls to SPIR-V intrinsics
 // or calls to proper generated functions. Returns True if F was modified.
 bool SPIRVPrepareFunctions::substituteIntrinsicCalls(Function *F) {
@@ -446,6 +464,10 @@ bool SPIRVPrepareFunctions::substituteIntrinsicCalls(Function *F) {
         lowerPtrAnnotation(II);
         Changed = true;
         break;
+      case Intrinsic::experimental_constrained_fmuladd:
+        lowerConstrainedFmuladd(II, EraseFromParent);
+        Changed = true;
+        break;
       case Intrinsic::experimental_constrained_fcmp:
       case Intrinsic::experimental_constrained_fcmps:
         lowerConstrainedFPCmpIntrinsic(dyn_cast<ConstrainedFPCmpIntrinsic>(II),
diff --git a/llvm/test/CodeGen/SPIRV/llvm-intrinsics/constrained-fmuladd.ll b/llvm/test/CodeGen/SPIRV/llvm-intrinsics/constrained-fmuladd.ll
new file mode 100644
index 0000000000000..340f2d78fc21b
--- /dev/null
+++ b/llvm/test/CodeGen/SPIRV/llvm-intrinsics/constrained-fmuladd.ll
@@ -0,0 +1,64 @@
+; RUN: llc -mtriple=spirv64-unknown-unknown %s -o - | FileCheck %s
+; RUN: %if spirv-tools %{ llc -mtriple=spirv64-unknown-unknown %s -o - -filetype=obj | spirv-val %}
+
+; CHECK-DAG: OpDecorate %[[#]] FPRoundingMode RTE
+; CHECK-DAG: OpDecorate %[[#]] FPRoundingMode RTZ
+; CHECK-DAG: OpDecorate %[[#]] FPRoundingMode RTP
+; CHECK-DAG: OpDecorate %[[#]] FPRoundingMode RTN
+; CHECK-DAG: OpDecorate %[[#]] FPRoundingMode RTE
+
+; CHECK: OpFMul %[[#]] %[[#]] %[[#]]
+; CHECK: OpFAdd %[[#]] %[[#]] %[[#]]
+define spir_kernel void @test_f32(float %a) {
+entry:
+  %r = tail call float @llvm.experimental.constrained.fmuladd.f32(
+              float %a, float %a, float %a,
+              metadata !"round.tonearest", metadata !"fpexcept.strict")
+  ret void
+}
+
+; CHECK: OpFMul %[[#]] %[[#]] %[[#]] 
+; CHECK: OpFAdd %[[#]] %[[#]] %[[#]]
+define spir_kernel void @test_f64(double %a) {
+entry:
+  %r = tail call double @llvm.experimental.constrained.fmuladd.f64(
+              double %a, double %a, double %a,
+              metadata !"round.towardzero", metadata !"fpexcept.strict")
+  ret void
+}
+
+; CHECK: OpFMul %[[#]] %[[#]] %[[#]]
+; CHECK: OpFAdd %[[#]] %[[#]] %[[#]]
+define spir_kernel void @test_v2f32(<2 x float> %a) {
+entry:
+  %r = tail call <2 x float> @llvm.experimental.constrained.fmuladd.v2f32(
+              <2 x float> %a, <2 x float> %a, <2 x float> %a,
+              metadata !"round.upward", metadata !"fpexcept.strict")
+  ret void
+}
+
+; CHECK: OpFMul %[[#]] %[[#]] %[[#]]
+; CHECK: OpFAdd %[[#]] %[[#]] %[[#]]
+define spir_kernel void @test_v4f32(<4 x float> %a) {
+entry:
+  %r = tail call <4 x float> @llvm.experimental.constrained.fmuladd.v4f32(
+              <4 x float> %a, <4 x float> %a, <4 x float> %a,
+              metadata !"round.downward", metadata !"fpexcept.strict")
+  ret void
+}
+
+; CHECK: OpFMul %[[#]] %[[#]] %[[#]]
+; CHECK: OpFAdd %[[#]] %[[#]] %[[#]]
+define spir_kernel void @test_v2f64(<2 x double> %a) {
+entry:
+  %r = tail call <2 x double> @llvm.experimental.constrained.fmuladd.v2f64(
+              <2 x double> %a, <2 x double> %a, <2 x double> %a,
+              metadata !"round.tonearest", metadata !"fpexcept.strict")
+  ret void
+}
+
+declare float  @llvm.experimental.constrained.fmuladd.f32(float, float, float, metadata, metadata)
+declare double @llvm.experimental.constrained.fmuladd.f64(double, double, double, metadata, metadata)
+declare <2 x float> @llvm.experimental.constrained.fmuladd.v2f32(<2 x float>, <2 x float>, <2 x float>, metadata, metadata)
+declare <4 x float> @llvm.experimental.constrained.fmuladd.v4f32(<4 x float>, <4 x float>, <4 x float>, metadata, metadata)
+declare <2 x double> @llvm.experimental.constrained.fmuladd.v2f64(<2 x double>, <2 x double>, <2 x double>, metadata, metadata)