[AMDGPU][CostModel] Add fneg cost estimation

[dfukalov]

Summary

The estimation uses AMDGPUTargetLowering::isFNegFree()

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp

553

If you're putting in in target code, might as well just return 0

[dfukalov]

Added comment, fixed typo

llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp

553

There are different cases when fneg is not free. E.g. neg modifier can make instruction bigger. Or even explicit xor will be generated.
I tried to make cost model less independent from actual transforms. If one changes
isFNegFree() behavior then cost model possible should be updated.

[rampitec]

LGTM

llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp

if (SLT == MVT::f32 || SLT == MVT::f16) {
       if (!HasFP32Denormals) {
         // FP mode switches.
         Cost += 2 * getFullRateInstrCost();
       }
 
       return LT.first * NElts * Cost;
     }
     break;
+  case ISD::FNEG:
+    // Use the backend' estimation. If fneg is not free each element will cost
+    // one additional instruction.
+    return TLI->isFNegFree(SLT) ? 0 : NElts;
   default:
     break;
   }
 
   return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Opd1Info,
                                        Opd2Info,
                                        Opd1PropInfo, Opd2PropInfo);
 }

llvm/test/Analysis/CostModel/AMDGPU/fneg.ll

+; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa < %s | FileCheck %s
+; RUN: opt -cost-model -cost-kind=code-size -analyze -mtriple=amdgcn-unknown-amdhsa < %s | FileCheck %s
+
+; CHECK-LABEL: 'fneg_f32'
+; CHECK: estimated cost of 0 for instruction:   %fneg = fneg float
+define amdgpu_kernel void @fneg_f32(float addrspace(1)* %out, float addrspace(1)* %vaddr) {
+  %vec = load float, float addrspace(1)* %vaddr
+  %fadd = fadd float %vec, undef
+  %fneg = fneg float %fadd
+  store float %fneg, float addrspace(1)* %out
+  ret void
+}
+
+; CHECK-LABEL: 'fneg_v2f32'
+; CHECK: estimated cost of 0 for instruction:   %fneg = fneg <2 x float>
+define amdgpu_kernel void @fneg_v2f32(<2 x float> addrspace(1)* %out, <2 x float> addrspace(1)* %vaddr) {
+  %vec = load <2 x float>, <2 x float> addrspace(1)* %vaddr
+  %fadd = fadd <2 x float> %vec, undef
+  %fneg = fneg <2 x float> %fadd
+  store <2 x float> %fneg, <2 x float> addrspace(1)* %out
+  ret void
+}
+
+; CHECK-LABEL: 'fneg_v3f32'
+; CHECK: estimated cost of 0 for instruction:   %fneg = fneg <3 x float>
+define amdgpu_kernel void @fneg_v3f32(<3 x float> addrspace(1)* %out, <3 x float> addrspace(1)* %vaddr) {
+  %vec = load <3 x float>, <3 x float> addrspace(1)* %vaddr
+  %fadd = fadd <3 x float> %vec, undef
+  %fneg = fneg <3 x float> %fadd
+  store <3 x float> %fneg, <3 x float> addrspace(1)* %out
+  ret void
+}
+
+; CHECK-LABEL: 'fneg_v5f32'
+; CHECK: estimated cost of 0 for instruction:   %fneg = fneg <5 x float>
+define amdgpu_kernel void @fneg_v5f32(<5 x float> addrspace(1)* %out, <5 x float> addrspace(1)* %vaddr) {
+  %vec = load <5 x float>, <5 x float> addrspace(1)* %vaddr
+  %fadd = fadd <5 x float> %vec, undef
+  %fneg = fneg <5 x float> %fadd
+  store <5 x float> %fneg, <5 x float> addrspace(1)* %out
+  ret void
+}
+
+; CHECK-LABEL: 'fneg_f64'
+; CHECK: estimated cost of 0 for instruction:   %fneg = fneg double
+define amdgpu_kernel void @fneg_f64(double addrspace(1)* %out, double addrspace(1)* %vaddr) {
+  %vec = load double, double addrspace(1)* %vaddr
+  %fadd = fadd double %vec, undef
+  %fneg = fneg double %fadd
+  store double %fneg, double addrspace(1)* %out
+  ret void
+}
+
+; CHECK-LABEL: 'fneg_v2f64'
+; CHECK: estimated cost of 0 for instruction:   %fneg = fneg <2 x double>
+define amdgpu_kernel void @fneg_v2f64(<2 x double> addrspace(1)* %out, <2 x double> addrspace(1)* %vaddr) {
+  %vec = load <2 x double>, <2 x double> addrspace(1)* %vaddr
+  %fadd = fadd <2 x double> %vec, undef
+  %fneg = fneg <2 x double> %fadd
+  store <2 x double> %fneg, <2 x double> addrspace(1)* %out
+  ret void
+}
+
+; CHECK-LABEL: 'fneg_v3f64'
+; CHECK: estimated cost of 0 for instruction:   %fneg = fneg <3 x double>
+define amdgpu_kernel void @fneg_v3f64(<3 x double> addrspace(1)* %out, <3 x double> addrspace(1)* %vaddr) {
+  %vec = load <3 x double>, <3 x double> addrspace(1)* %vaddr
+  %fadd = fadd <3 x double> %vec, undef
+  %fneg = fneg <3 x double> %fadd
+  store <3 x double> %fneg, <3 x double> addrspace(1)* %out
+  ret void
+}
+
+; CHECK-LABEL: 'fneg_f16'
+; CHECK: estimated cost of 0 for instruction:   %fneg = fneg half
+define amdgpu_kernel void @fneg_f16(half addrspace(1)* %out, half addrspace(1)* %vaddr) {
+  %vec = load half, half addrspace(1)* %vaddr
+  %fadd = fadd half %vec, undef
+  %fneg = fneg half %fadd
+  store half %fneg, half addrspace(1)* %out
+  ret void
+}
+
+; CHECK-LABEL: 'fneg_v2f16'
+; CHECK: estimated cost of 0 for instruction:   %fneg = fneg <2 x half>
+define amdgpu_kernel void @fneg_v2f16(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %vaddr) {
+  %vec = load <2 x half>, <2 x half> addrspace(1)* %vaddr
+  %fadd = fadd <2 x half> %vec, undef
+  %fneg = fneg <2 x half> %fadd
+  store <2 x half> %fneg, <2 x half> addrspace(1)* %out
+  ret void
+}
+
+; CHECK-LABEL: 'fneg_v3f16'
+; CHECK: estimated cost of 0 for instruction:   %fneg = fneg <3 x half>
+define amdgpu_kernel void @fneg_v3f16(<3 x half> addrspace(1)* %out, <3 x half> addrspace(1)* %vaddr) {
+  %vec = load <3 x half>, <3 x half> addrspace(1)* %vaddr
+  %fadd = fadd <3 x half> %vec, undef
+  %fneg = fneg <3 x half> %fadd
+  store <3 x half> %fneg, <3 x half> addrspace(1)* %out
+  ret void
+}