[llvm] [AMDGPU] Fix cost of fast unsafe f32 fdiv (PR #68988)

[Jay Foad via llvm-commits]

https://github.com/jayfoad created https://github.com/llvm/llvm-project/pull/68988

None

>From bee9d8f04feff4d0c6c401de96d8a37cf560907b Mon Sep 17 00:00:00 2001
From: Jay Foad <jay.foad at amd.com>
Date: Fri, 13 Oct 2023 14:30:38 +0100
Subject: [PATCH] [AMDGPU] Fix cost of fast unsafe f32 fdiv

---
 .../AMDGPU/AMDGPUTargetTransformInfo.cpp      |  9 +++
 llvm/test/Analysis/CostModel/AMDGPU/fdiv.ll   | 62 +++++++++++++++++++
 2 files changed, 71 insertions(+)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
index 074c8626572b938..cb877a4695f1ece 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
@@ -650,6 +650,15 @@ InstructionCost GCNTTIImpl::getArithmeticInstrCost(
       return LT.first * Cost * NElts;
     }
 
+    if (SLT == MVT::f32 && ((CxtI && CxtI->hasApproxFunc()) ||
+                            TLI->getTargetMachine().Options.UnsafeFPMath)) {
+      // Fast unsafe fdiv lowering:
+      // f32 rcp
+      // f32 fmul
+      int Cost = getQuarterRateInstrCost(CostKind) + getFullRateInstrCost();
+      return LT.first * Cost * NElts;
+    }
+
     if (SLT == MVT::f32 || SLT == MVT::f16) {
       // 4 more v_cvt_* insts without f16 insts support
       int Cost = (SLT == MVT::f16 ? 14 : 10) * getFullRateInstrCost() +
diff --git a/llvm/test/Analysis/CostModel/AMDGPU/fdiv.ll b/llvm/test/Analysis/CostModel/AMDGPU/fdiv.ll
index 11ce416b7fd7954..2830bfcdaed2025 100644
--- a/llvm/test/Analysis/CostModel/AMDGPU/fdiv.ll
+++ b/llvm/test/Analysis/CostModel/AMDGPU/fdiv.ll
@@ -43,6 +43,37 @@ define amdgpu_kernel void @fdiv_f32_ieee() #0 {
   ret void
 }
 
+define amdgpu_kernel void @fdiv_f32_afn_ieee() #0 {
+; ALL-LABEL: 'fdiv_f32_afn_ieee'
+; ALL-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %f32 = fdiv afn float undef, undef
+; ALL-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %v2f32 = fdiv afn <2 x float> undef, undef
+; ALL-NEXT:  Cost Model: Found an estimated cost of 15 for instruction: %v3f32 = fdiv afn <3 x float> undef, undef
+; ALL-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %v4f32 = fdiv afn <4 x float> undef, undef
+; ALL-NEXT:  Cost Model: Found an estimated cost of 25 for instruction: %v5f32 = fdiv afn <5 x float> undef, undef
+; ALL-NEXT:  Cost Model: Found an estimated cost of 40 for instruction: %v8f32 = fdiv afn <8 x float> undef, undef
+; ALL-NEXT:  Cost Model: Found an estimated cost of 135 for instruction: %v9f32 = fdiv afn <9 x float> undef, undef
+; ALL-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: ret void
+;
+; ALL-SIZE-LABEL: 'fdiv_f32_afn_ieee'
+; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %f32 = fdiv afn float undef, undef
+; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %v2f32 = fdiv afn <2 x float> undef, undef
+; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 9 for instruction: %v3f32 = fdiv afn <3 x float> undef, undef
+; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %v4f32 = fdiv afn <4 x float> undef, undef
+; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 15 for instruction: %v5f32 = fdiv afn <5 x float> undef, undef
+; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %v8f32 = fdiv afn <8 x float> undef, undef
+; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 81 for instruction: %v9f32 = fdiv afn <9 x float> undef, undef
+; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret void
+;
+  %f32 = fdiv afn float undef, undef
+  %v2f32 = fdiv afn <2 x float> undef, undef
+  %v3f32 = fdiv afn <3 x float> undef, undef
+  %v4f32 = fdiv afn <4 x float> undef, undef
+  %v5f32 = fdiv afn <5 x float> undef, undef
+  %v8f32 = fdiv afn <8 x float> undef, undef
+  %v9f32 = fdiv afn <9 x float> undef, undef
+  ret void
+}
+
 define amdgpu_kernel void @fdiv_f32_ftzdaz() #1 {
 ; ALL-LABEL: 'fdiv_f32_ftzdaz'
 ; ALL-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %f32 = fdiv float undef, undef
@@ -74,6 +105,37 @@ define amdgpu_kernel void @fdiv_f32_ftzdaz() #1 {
   ret void
 }
 
+define amdgpu_kernel void @fdiv_f32_afn_ftzdaz() #1 {
+; ALL-LABEL: 'fdiv_f32_afn_ftzdaz'
+; ALL-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %f32 = fdiv afn float undef, undef
+; ALL-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %v2f32 = fdiv afn <2 x float> undef, undef
+; ALL-NEXT:  Cost Model: Found an estimated cost of 15 for instruction: %v3f32 = fdiv afn <3 x float> undef, undef
+; ALL-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %v4f32 = fdiv afn <4 x float> undef, undef
+; ALL-NEXT:  Cost Model: Found an estimated cost of 25 for instruction: %v5f32 = fdiv afn <5 x float> undef, undef
+; ALL-NEXT:  Cost Model: Found an estimated cost of 40 for instruction: %v8f32 = fdiv afn <8 x float> undef, undef
+; ALL-NEXT:  Cost Model: Found an estimated cost of 135 for instruction: %v9f32 = fdiv afn <9 x float> undef, undef
+; ALL-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: ret void
+;
+; ALL-SIZE-LABEL: 'fdiv_f32_afn_ftzdaz'
+; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %f32 = fdiv afn float undef, undef
+; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %v2f32 = fdiv afn <2 x float> undef, undef
+; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 9 for instruction: %v3f32 = fdiv afn <3 x float> undef, undef
+; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %v4f32 = fdiv afn <4 x float> undef, undef
+; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 15 for instruction: %v5f32 = fdiv afn <5 x float> undef, undef
+; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %v8f32 = fdiv afn <8 x float> undef, undef
+; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 81 for instruction: %v9f32 = fdiv afn <9 x float> undef, undef
+; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret void
+;
+  %f32 = fdiv afn float undef, undef
+  %v2f32 = fdiv afn <2 x float> undef, undef
+  %v3f32 = fdiv afn <3 x float> undef, undef
+  %v4f32 = fdiv afn <4 x float> undef, undef
+  %v5f32 = fdiv afn <5 x float> undef, undef
+  %v8f32 = fdiv afn <8 x float> undef, undef
+  %v9f32 = fdiv afn <9 x float> undef, undef
+  ret void
+}
+
 define amdgpu_kernel void @fdiv_f64() #0 {
 ; CIFASTF64-LABEL: 'fdiv_f64'
 ; CIFASTF64-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %f64 = fdiv double undef, undef