[AArch64] Change the cost of fma and fmuladd to match fmul. (#152963)

As fmul and fmadd are so similar, their performance characteristics tend
to be the same on most platforms, at least in terms of reciprocal
throughputs. Processors capable of performing a given number of fmul per
cycle can usually perform the same number of fma, with the extra add
being relatively simple on top. This patch makes the scores of the two
operations the same, which brings the throughput cost of a fma/fmuladd
to 2, and the latency to 3, which are the defaults for fmul.

Note that we might also want to change the throughput cost of a fmul to
1, as most processors have ample bandwidth for them, but they should
still stay in-line with one another.

Author: davemgreen

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp

@@ -666,6 +666,16 @@ AArch64TTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
       return LT.first;
     break;
   }
+  case Intrinsic::fma:
+  case Intrinsic::fmuladd: {
+    // Given a fma or fmuladd, cost it the same as a fmul instruction which are
+    // usually the same for costs. TODO: Add fp16 and bf16 expansion costs.
+    Type *EltTy = RetTy->getScalarType();
+    if (EltTy->isFloatTy() || EltTy->isDoubleTy() ||
+        (EltTy->isHalfTy() && ST->hasFullFP16()))
+      return getArithmeticInstrCost(Instruction::FMul, RetTy, CostKind);
+    break;
+  }
   case Intrinsic::stepvector: {
     InstructionCost Cost = 1; // Cost of the `index' instruction
     auto LT = getTypeLegalizationCost(RetTy);

llvm/test/Analysis/CostModel/AArch64/arith-fp.ll

@@ -663,13 +663,13 @@ define void @fcopysign_fp16() {
 
 define void @fma() {
 ; CHECK-LABEL: 'fma'
-; CHECK-NEXT:  Cost Model: Found costs of 1 for: %F32 = call float @llvm.fma.f32(float undef, float undef, float undef)
-; CHECK-NEXT:  Cost Model: Found costs of 1 for: %V2F32 = call <2 x float> @llvm.fma.v2f32(<2 x float> undef, <2 x float> undef, <2 x float> undef)
-; CHECK-NEXT:  Cost Model: Found costs of 1 for: %V4F32 = call <4 x float> @llvm.fma.v4f32(<4 x float> undef, <4 x float> undef, <4 x float> undef)
-; CHECK-NEXT:  Cost Model: Found costs of 4 for: %V8F32 = call <8 x float> @llvm.fma.v8f32(<8 x float> undef, <8 x float> undef, <8 x float> undef)
-; CHECK-NEXT:  Cost Model: Found costs of 1 for: %F64 = call double @llvm.fma.f64(double undef, double undef, double undef)
-; CHECK-NEXT:  Cost Model: Found costs of 1 for: %V2F64 = call <2 x double> @llvm.fma.v2f64(<2 x double> undef, <2 x double> undef, <2 x double> undef)
-; CHECK-NEXT:  Cost Model: Found costs of 4 for: %V4F64 = call <4 x double> @llvm.fma.v4f64(<4 x double> undef, <4 x double> undef, <4 x double> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %F32 = call float @llvm.fma.f32(float undef, float undef, float undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %V2F32 = call <2 x float> @llvm.fma.v2f32(<2 x float> undef, <2 x float> undef, <2 x float> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %V4F32 = call <4 x float> @llvm.fma.v4f32(<4 x float> undef, <4 x float> undef, <4 x float> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:4 CodeSize:1 Lat:3 SizeLat:1 for: %V8F32 = call <8 x float> @llvm.fma.v8f32(<8 x float> undef, <8 x float> undef, <8 x float> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %F64 = call double @llvm.fma.f64(double undef, double undef, double undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %V2F64 = call <2 x double> @llvm.fma.v2f64(<2 x double> undef, <2 x double> undef, <2 x double> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:4 CodeSize:1 Lat:3 SizeLat:1 for: %V4F64 = call <4 x double> @llvm.fma.v4f64(<4 x double> undef, <4 x double> undef, <4 x double> undef)
 ; CHECK-NEXT:  Cost Model: Found costs of RThru:10 CodeSize:1 Lat:10 SizeLat:10 for: %F128 = call fp128 @llvm.fma.f128(fp128 undef, fp128 undef, fp128 undef)
 ; CHECK-NEXT:  Cost Model: Found costs of RThru:20 CodeSize:2 Lat:20 SizeLat:20 for: %V2F128 = call <2 x fp128> @llvm.fma.v2f128(<2 x fp128> undef, <2 x fp128> undef, <2 x fp128> undef)
 ; CHECK-NEXT:  Cost Model: Found costs of RThru:0 CodeSize:1 Lat:1 SizeLat:1 for: ret void
@@ -698,10 +698,10 @@ define void @fma_fp16() {
 ; CHECK-BASE-NEXT:  Cost Model: Found costs of RThru:0 CodeSize:1 Lat:1 SizeLat:1 for: ret void
 ;
 ; CHECK-FP16-LABEL: 'fma_fp16'
-; CHECK-FP16-NEXT:  Cost Model: Found costs of 1 for: %F16 = call half @llvm.fma.f16(half undef, half undef, half undef)
-; CHECK-FP16-NEXT:  Cost Model: Found costs of 1 for: %V4F16 = call <4 x half> @llvm.fma.v4f16(<4 x half> undef, <4 x half> undef, <4 x half> undef)
-; CHECK-FP16-NEXT:  Cost Model: Found costs of 1 for: %V8F16 = call <8 x half> @llvm.fma.v8f16(<8 x half> undef, <8 x half> undef, <8 x half> undef)
-; CHECK-FP16-NEXT:  Cost Model: Found costs of 4 for: %V16F16 = call <16 x half> @llvm.fma.v16f16(<16 x half> undef, <16 x half> undef, <16 x half> undef)
+; CHECK-FP16-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %F16 = call half @llvm.fma.f16(half undef, half undef, half undef)
+; CHECK-FP16-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %V4F16 = call <4 x half> @llvm.fma.v4f16(<4 x half> undef, <4 x half> undef, <4 x half> undef)
+; CHECK-FP16-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %V8F16 = call <8 x half> @llvm.fma.v8f16(<8 x half> undef, <8 x half> undef, <8 x half> undef)
+; CHECK-FP16-NEXT:  Cost Model: Found costs of RThru:4 CodeSize:1 Lat:3 SizeLat:1 for: %V16F16 = call <16 x half> @llvm.fma.v16f16(<16 x half> undef, <16 x half> undef, <16 x half> undef)
 ; CHECK-FP16-NEXT:  Cost Model: Found costs of RThru:0 CodeSize:1 Lat:1 SizeLat:1 for: ret void
 ;
   %F16 = call half @llvm.fma.f16(half undef, half undef, half undef)
@@ -713,13 +713,13 @@ define void @fma_fp16() {
 
 define void @fmuladd() {
 ; CHECK-LABEL: 'fmuladd'
-; CHECK-NEXT:  Cost Model: Found costs of 1 for: %F32 = call float @llvm.fmuladd.f32(float undef, float undef, float undef)
-; CHECK-NEXT:  Cost Model: Found costs of 1 for: %V2F32 = call <2 x float> @llvm.fmuladd.v2f32(<2 x float> undef, <2 x float> undef, <2 x float> undef)
-; CHECK-NEXT:  Cost Model: Found costs of 1 for: %V4F32 = call <4 x float> @llvm.fmuladd.v4f32(<4 x float> undef, <4 x float> undef, <4 x float> undef)
-; CHECK-NEXT:  Cost Model: Found costs of 4 for: %V8F32 = call <8 x float> @llvm.fmuladd.v8f32(<8 x float> undef, <8 x float> undef, <8 x float> undef)
-; CHECK-NEXT:  Cost Model: Found costs of 1 for: %F64 = call double @llvm.fmuladd.f64(double undef, double undef, double undef)
-; CHECK-NEXT:  Cost Model: Found costs of 1 for: %V2F64 = call <2 x double> @llvm.fmuladd.v2f64(<2 x double> undef, <2 x double> undef, <2 x double> undef)
-; CHECK-NEXT:  Cost Model: Found costs of 4 for: %V4F64 = call <4 x double> @llvm.fmuladd.v4f64(<4 x double> undef, <4 x double> undef, <4 x double> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %F32 = call float @llvm.fmuladd.f32(float undef, float undef, float undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %V2F32 = call <2 x float> @llvm.fmuladd.v2f32(<2 x float> undef, <2 x float> undef, <2 x float> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %V4F32 = call <4 x float> @llvm.fmuladd.v4f32(<4 x float> undef, <4 x float> undef, <4 x float> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:4 CodeSize:1 Lat:3 SizeLat:1 for: %V8F32 = call <8 x float> @llvm.fmuladd.v8f32(<8 x float> undef, <8 x float> undef, <8 x float> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %F64 = call double @llvm.fmuladd.f64(double undef, double undef, double undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %V2F64 = call <2 x double> @llvm.fmuladd.v2f64(<2 x double> undef, <2 x double> undef, <2 x double> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:4 CodeSize:1 Lat:3 SizeLat:1 for: %V4F64 = call <4 x double> @llvm.fmuladd.v4f64(<4 x double> undef, <4 x double> undef, <4 x double> undef)
 ; CHECK-NEXT:  Cost Model: Found costs of RThru:8 CodeSize:2 Lat:6 SizeLat:2 for: %F128 = call fp128 @llvm.fmuladd.f128(fp128 undef, fp128 undef, fp128 undef)
 ; CHECK-NEXT:  Cost Model: Found costs of RThru:16 CodeSize:2 Lat:6 SizeLat:2 for: %V2F128 = call <2 x fp128> @llvm.fmuladd.v2f128(<2 x fp128> undef, <2 x fp128> undef, <2 x fp128> undef)
 ; CHECK-NEXT:  Cost Model: Found costs of RThru:0 CodeSize:1 Lat:1 SizeLat:1 for: ret void
@@ -748,10 +748,10 @@ define void @fmuladd_fp16() {
 ; CHECK-BASE-NEXT:  Cost Model: Found costs of RThru:0 CodeSize:1 Lat:1 SizeLat:1 for: ret void
 ;
 ; CHECK-FP16-LABEL: 'fmuladd_fp16'
-; CHECK-FP16-NEXT:  Cost Model: Found costs of 1 for: %F16 = call half @llvm.fmuladd.f16(half undef, half undef, half undef)
-; CHECK-FP16-NEXT:  Cost Model: Found costs of 1 for: %V4F16 = call <4 x half> @llvm.fmuladd.v4f16(<4 x half> undef, <4 x half> undef, <4 x half> undef)
-; CHECK-FP16-NEXT:  Cost Model: Found costs of 1 for: %V8F16 = call <8 x half> @llvm.fmuladd.v8f16(<8 x half> undef, <8 x half> undef, <8 x half> undef)
-; CHECK-FP16-NEXT:  Cost Model: Found costs of 4 for: %V16F16 = call <16 x half> @llvm.fmuladd.v16f16(<16 x half> undef, <16 x half> undef, <16 x half> undef)
+; CHECK-FP16-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %F16 = call half @llvm.fmuladd.f16(half undef, half undef, half undef)
+; CHECK-FP16-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %V4F16 = call <4 x half> @llvm.fmuladd.v4f16(<4 x half> undef, <4 x half> undef, <4 x half> undef)
+; CHECK-FP16-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %V8F16 = call <8 x half> @llvm.fmuladd.v8f16(<8 x half> undef, <8 x half> undef, <8 x half> undef)
+; CHECK-FP16-NEXT:  Cost Model: Found costs of RThru:4 CodeSize:1 Lat:3 SizeLat:1 for: %V16F16 = call <16 x half> @llvm.fmuladd.v16f16(<16 x half> undef, <16 x half> undef, <16 x half> undef)
 ; CHECK-FP16-NEXT:  Cost Model: Found costs of RThru:0 CodeSize:1 Lat:1 SizeLat:1 for: ret void
 ;
   %F16 = call half @llvm.fmuladd.f32(half undef, half undef, half undef)

llvm/test/Analysis/CostModel/AArch64/sve-arith-fp.ll

@@ -167,14 +167,14 @@ define void @frem() {
 
 define void @fma() {
 ; CHECK-LABEL: 'fma'
-; CHECK-NEXT:  Cost Model: Found costs of 2 for: %V4F16 = call <vscale x 4 x half> @llvm.fma.nxv4f16(<vscale x 4 x half> undef, <vscale x 4 x half> undef, <vscale x 4 x half> undef)
-; CHECK-NEXT:  Cost Model: Found costs of 2 for: %V8F16 = call <vscale x 8 x half> @llvm.fma.nxv8f16(<vscale x 8 x half> undef, <vscale x 8 x half> undef, <vscale x 8 x half> undef)
-; CHECK-NEXT:  Cost Model: Found costs of 4 for: %V16F16 = call <vscale x 16 x half> @llvm.fma.nxv16f16(<vscale x 16 x half> undef, <vscale x 16 x half> undef, <vscale x 16 x half> undef)
-; CHECK-NEXT:  Cost Model: Found costs of 2 for: %V2F32 = call <vscale x 2 x float> @llvm.fma.nxv2f32(<vscale x 2 x float> undef, <vscale x 2 x float> undef, <vscale x 2 x float> undef)
-; CHECK-NEXT:  Cost Model: Found costs of 2 for: %V4F32 = call <vscale x 4 x float> @llvm.fma.nxv4f32(<vscale x 4 x float> undef, <vscale x 4 x float> undef, <vscale x 4 x float> undef)
-; CHECK-NEXT:  Cost Model: Found costs of 4 for: %V8F32 = call <vscale x 8 x float> @llvm.fma.nxv8f32(<vscale x 8 x float> undef, <vscale x 8 x float> undef, <vscale x 8 x float> undef)
-; CHECK-NEXT:  Cost Model: Found costs of 2 for: %V2F64 = call <vscale x 2 x double> @llvm.fma.nxv2f64(<vscale x 2 x double> undef, <vscale x 2 x double> undef, <vscale x 2 x double> undef)
-; CHECK-NEXT:  Cost Model: Found costs of 4 for: %V4F64 = call <vscale x 4 x double> @llvm.fma.nxv4f64(<vscale x 4 x double> undef, <vscale x 4 x double> undef, <vscale x 4 x double> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %V4F16 = call <vscale x 4 x half> @llvm.fma.nxv4f16(<vscale x 4 x half> undef, <vscale x 4 x half> undef, <vscale x 4 x half> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %V8F16 = call <vscale x 8 x half> @llvm.fma.nxv8f16(<vscale x 8 x half> undef, <vscale x 8 x half> undef, <vscale x 8 x half> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:4 CodeSize:1 Lat:3 SizeLat:1 for: %V16F16 = call <vscale x 16 x half> @llvm.fma.nxv16f16(<vscale x 16 x half> undef, <vscale x 16 x half> undef, <vscale x 16 x half> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %V2F32 = call <vscale x 2 x float> @llvm.fma.nxv2f32(<vscale x 2 x float> undef, <vscale x 2 x float> undef, <vscale x 2 x float> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %V4F32 = call <vscale x 4 x float> @llvm.fma.nxv4f32(<vscale x 4 x float> undef, <vscale x 4 x float> undef, <vscale x 4 x float> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:4 CodeSize:1 Lat:3 SizeLat:1 for: %V8F32 = call <vscale x 8 x float> @llvm.fma.nxv8f32(<vscale x 8 x float> undef, <vscale x 8 x float> undef, <vscale x 8 x float> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %V2F64 = call <vscale x 2 x double> @llvm.fma.nxv2f64(<vscale x 2 x double> undef, <vscale x 2 x double> undef, <vscale x 2 x double> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:4 CodeSize:1 Lat:3 SizeLat:1 for: %V4F64 = call <vscale x 4 x double> @llvm.fma.nxv4f64(<vscale x 4 x double> undef, <vscale x 4 x double> undef, <vscale x 4 x double> undef)
 ; CHECK-NEXT:  Cost Model: Found costs of RThru:0 CodeSize:1 Lat:1 SizeLat:1 for: ret void
 ;
   %V4F16 = call <vscale x 4 x half> @llvm.fma.v4f16(<vscale x 4 x half> undef, <vscale x 4 x half> undef, <vscale x 4 x half> undef)
@@ -193,14 +193,14 @@ define void @fma() {
 
 define void @fmuladd() {
 ; CHECK-LABEL: 'fmuladd'
-; CHECK-NEXT:  Cost Model: Found costs of 2 for: %V4F16 = call <vscale x 4 x half> @llvm.fmuladd.nxv4f16(<vscale x 4 x half> undef, <vscale x 4 x half> undef, <vscale x 4 x half> undef)
-; CHECK-NEXT:  Cost Model: Found costs of 2 for: %V8F16 = call <vscale x 8 x half> @llvm.fmuladd.nxv8f16(<vscale x 8 x half> undef, <vscale x 8 x half> undef, <vscale x 8 x half> undef)
-; CHECK-NEXT:  Cost Model: Found costs of 4 for: %V16F16 = call <vscale x 16 x half> @llvm.fmuladd.nxv16f16(<vscale x 16 x half> undef, <vscale x 16 x half> undef, <vscale x 16 x half> undef)
-; CHECK-NEXT:  Cost Model: Found costs of 2 for: %V2F32 = call <vscale x 2 x float> @llvm.fmuladd.nxv2f32(<vscale x 2 x float> undef, <vscale x 2 x float> undef, <vscale x 2 x float> undef)
-; CHECK-NEXT:  Cost Model: Found costs of 2 for: %V4F32 = call <vscale x 4 x float> @llvm.fmuladd.nxv4f32(<vscale x 4 x float> undef, <vscale x 4 x float> undef, <vscale x 4 x float> undef)
-; CHECK-NEXT:  Cost Model: Found costs of 4 for: %V8F32 = call <vscale x 8 x float> @llvm.fmuladd.nxv8f32(<vscale x 8 x float> undef, <vscale x 8 x float> undef, <vscale x 8 x float> undef)
-; CHECK-NEXT:  Cost Model: Found costs of 2 for: %V2F64 = call <vscale x 2 x double> @llvm.fmuladd.nxv2f64(<vscale x 2 x double> undef, <vscale x 2 x double> undef, <vscale x 2 x double> undef)
-; CHECK-NEXT:  Cost Model: Found costs of 4 for: %V4F64 = call <vscale x 4 x double> @llvm.fmuladd.nxv4f64(<vscale x 4 x double> undef, <vscale x 4 x double> undef, <vscale x 4 x double> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %V4F16 = call <vscale x 4 x half> @llvm.fmuladd.nxv4f16(<vscale x 4 x half> undef, <vscale x 4 x half> undef, <vscale x 4 x half> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %V8F16 = call <vscale x 8 x half> @llvm.fmuladd.nxv8f16(<vscale x 8 x half> undef, <vscale x 8 x half> undef, <vscale x 8 x half> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:4 CodeSize:1 Lat:3 SizeLat:1 for: %V16F16 = call <vscale x 16 x half> @llvm.fmuladd.nxv16f16(<vscale x 16 x half> undef, <vscale x 16 x half> undef, <vscale x 16 x half> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %V2F32 = call <vscale x 2 x float> @llvm.fmuladd.nxv2f32(<vscale x 2 x float> undef, <vscale x 2 x float> undef, <vscale x 2 x float> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %V4F32 = call <vscale x 4 x float> @llvm.fmuladd.nxv4f32(<vscale x 4 x float> undef, <vscale x 4 x float> undef, <vscale x 4 x float> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:4 CodeSize:1 Lat:3 SizeLat:1 for: %V8F32 = call <vscale x 8 x float> @llvm.fmuladd.nxv8f32(<vscale x 8 x float> undef, <vscale x 8 x float> undef, <vscale x 8 x float> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:2 CodeSize:1 Lat:3 SizeLat:1 for: %V2F64 = call <vscale x 2 x double> @llvm.fmuladd.nxv2f64(<vscale x 2 x double> undef, <vscale x 2 x double> undef, <vscale x 2 x double> undef)
+; CHECK-NEXT:  Cost Model: Found costs of RThru:4 CodeSize:1 Lat:3 SizeLat:1 for: %V4F64 = call <vscale x 4 x double> @llvm.fmuladd.nxv4f64(<vscale x 4 x double> undef, <vscale x 4 x double> undef, <vscale x 4 x double> undef)
 ; CHECK-NEXT:  Cost Model: Found costs of RThru:0 CodeSize:1 Lat:1 SizeLat:1 for: ret void
 ;
   %V4F16 = call <vscale x 4 x half> @llvm.fmuladd.v4f16(<vscale x 4 x half> undef, <vscale x 4 x half> undef, <vscale x 4 x half> undef)