[x86] adjust cost model values for minnum/maxnum with fast-math-flags

[spatel]

Summary

Without FMF, we lower these intrinsics into something like this:

vmaxsd	%xmm0, %xmm1, %xmm2
vcmpunordsd	%xmm0, %xmm0, %xmm0
vblendvpd	%xmm0, %xmm1, %xmm2, %xmm0

But if we can ignore NANs, the single min/max instruction is enough because there is no need to fix up the x86 logic that corresponds to something like X>Y ? X : Y.
We probably want to make other adjustments for FP intrinsics with FMF to account for specialized codegen (for example, FSQRT).

[pengfei]

LGTM.

llvm/test/Analysis/CostModel/X86/fmaxnum.ll

3–4

I think you can use --check-prefixes=AVX,AVX1 to reduce some duplicated checks. The same to fminnum.ll

[spatel]

llvm/test/Analysis/CostModel/X86/fmaxnum.ll

3–4

Thanks! I'll update the RUN lines on commit.

llvm/lib/Target/X86/X86TargetTransformInfo.cpp

if (ISD == ISD::BITREVERSE && ST->hasGFNI() && ST->hasSSSE3() &&
       // instructions. We also need an extract and an insert.
       if (!(MTy.is128BitVector() || (ST->hasAVX2() && MTy.is256BitVector()) ||
             (ST->hasBWI() && MTy.is512BitVector())))
         Cost = Cost * 2 + 2;
 
       return LT.first * Cost;
     }
 
+    auto adjustTableCost = [](const CostTblEntry &Entry, int LegalizationCost,
+                              FastMathFlags FMF) {
+      // If there are no NANs to deal with, then these are reduced to a
+      // single MIN** or MAX** instruction instead of the MIN/CMP/SELECT that we
+      // assume is used in the non-fast case.
+      if (Entry.ISD == ISD::FMAXNUM || Entry.ISD == ISD::FMINNUM) {
+        if (FMF.noNaNs())
+          return LegalizationCost * 1;
+      }
+      return LegalizationCost * (int)Entry.Cost;
+    };
+
     if (ST->useGLMDivSqrtCosts())
       if (const auto *Entry = CostTableLookup(GLMCostTbl, ISD, MTy))
-        return LT.first * Entry->Cost;
+        return adjustTableCost(*Entry, LT.first, ICA.getFlags());
 
     if (ST->isSLM())
       if (const auto *Entry = CostTableLookup(SLMCostTbl, ISD, MTy))
-        return LT.first * Entry->Cost;
+        return adjustTableCost(*Entry, LT.first, ICA.getFlags());
 
     if (ST->hasCDI())
       if (const auto *Entry = CostTableLookup(AVX512CDCostTbl, ISD, MTy))
-        return LT.first * Entry->Cost;
+        return adjustTableCost(*Entry, LT.first, ICA.getFlags());
 
     if (ST->hasBWI())
       if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy))
-        return LT.first * Entry->Cost;
+        return adjustTableCost(*Entry, LT.first, ICA.getFlags());
 
     if (ST->hasAVX512())
       if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
-        return LT.first * Entry->Cost;
+        return adjustTableCost(*Entry, LT.first, ICA.getFlags());
 
     if (ST->hasXOP())
       if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy))
-        return LT.first * Entry->Cost;
+        return adjustTableCost(*Entry, LT.first, ICA.getFlags());
 
     if (ST->hasAVX2())
       if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy))
-        return LT.first * Entry->Cost;
+        return adjustTableCost(*Entry, LT.first, ICA.getFlags());
 
     if (ST->hasAVX())
       if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy))
-        return LT.first * Entry->Cost;
+        return adjustTableCost(*Entry, LT.first, ICA.getFlags());
 
     if (ST->hasSSE42())
       if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy))
-        return LT.first * Entry->Cost;
+        return adjustTableCost(*Entry, LT.first, ICA.getFlags());
 
     if (ST->hasSSE41())
       if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy))
-        return LT.first * Entry->Cost;
+        return adjustTableCost(*Entry, LT.first, ICA.getFlags());
 
     if (ST->hasSSSE3())
       if (const auto *Entry = CostTableLookup(SSSE3CostTbl, ISD, MTy))
-        return LT.first * Entry->Cost;
+        return adjustTableCost(*Entry, LT.first, ICA.getFlags());
 
     if (ST->hasSSE2())
       if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy))
-        return LT.first * Entry->Cost;
+        return adjustTableCost(*Entry, LT.first, ICA.getFlags());
 
     if (ST->hasSSE1())
       if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy))
-        return LT.first * Entry->Cost;
+        return adjustTableCost(*Entry, LT.first, ICA.getFlags());
 
     if (ST->hasBMI()) {
       if (ST->is64Bit())
         if (const auto *Entry = CostTableLookup(BMI64CostTbl, ISD, MTy))
-          return LT.first * Entry->Cost;
+          return adjustTableCost(*Entry, LT.first, ICA.getFlags());
 
       if (const auto *Entry = CostTableLookup(BMI32CostTbl, ISD, MTy))
-        return LT.first * Entry->Cost;
+        return adjustTableCost(*Entry, LT.first, ICA.getFlags());
     }
 
     if (ST->hasLZCNT()) {
       if (ST->is64Bit())
         if (const auto *Entry = CostTableLookup(LZCNT64CostTbl, ISD, MTy))
-          return LT.first * Entry->Cost;
+          return adjustTableCost(*Entry, LT.first, ICA.getFlags());
 
       if (const auto *Entry = CostTableLookup(LZCNT32CostTbl, ISD, MTy))
-        return LT.first * Entry->Cost;
+        return adjustTableCost(*Entry, LT.first, ICA.getFlags());
     }
 
     if (ST->hasPOPCNT()) {
       if (ST->is64Bit())
         if (const auto *Entry = CostTableLookup(POPCNT64CostTbl, ISD, MTy))
-          return LT.first * Entry->Cost;
+          return adjustTableCost(*Entry, LT.first, ICA.getFlags());
 
       if (const auto *Entry = CostTableLookup(POPCNT32CostTbl, ISD, MTy))
-        return LT.first * Entry->Cost;
+        return adjustTableCost(*Entry, LT.first, ICA.getFlags());
     }
 
     // TODO - add BMI (TZCNT) scalar handling
 
     if (ST->is64Bit())
       if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy))
-        return LT.first * Entry->Cost;
+        return adjustTableCost(*Entry, LT.first, ICA.getFlags());
 
     if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy))
-      return LT.first * Entry->Cost;
+      return adjustTableCost(*Entry, LT.first, ICA.getFlags());
   }
 
   return BaseT::getIntrinsicInstrCost(ICA, CostKind);
 }
 
 int X86TTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
                                       TTI::TargetCostKind CostKind) {
   if (ICA.isTypeBasedOnly())

llvm/test/Analysis/CostModel/X86/fmaxnum.ll

 ; NOTE: Assertions have been autogenerated by utils/update_analyze_test_checks.py
 ; RUN: opt < %s -cost-model -analyze -mtriple=x86_64-- -mattr=+sse2 | FileCheck %s --check-prefixes=SSE
 ; RUN: opt < %s -cost-model -analyze -mtriple=x86_64-- -mattr=+sse4.2 | FileCheck %s --check-prefixes=SSE
-; RUN: opt < %s -cost-model -analyze -mtriple=x86_64-- -mattr=+avx | FileCheck %s --check-prefixes=AVX1
-; RUN: opt < %s -cost-model -analyze -mtriple=x86_64-- -mattr=+avx2 | FileCheck %s --check-prefixes=AVX2
+; RUN: opt < %s -cost-model -analyze -mtriple=x86_64-- -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX1
+; RUN: opt < %s -cost-model -analyze -mtriple=x86_64-- -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX2
 ; RUN: opt < %s -cost-model -analyze -mtriple=x86_64-- -mattr=+avx512f | FileCheck %s --check-prefixes=AVX512
 ; RUN: opt < %s -cost-model -analyze -mtriple=x86_64-- -mattr=+avx512f,+avx512bw | FileCheck %s --check-prefixes=AVX512
 ;
 ; RUN: opt < %s -cost-model -analyze -mtriple=x86_64-- -mcpu=slm | FileCheck %s --check-prefixes=SSE
 ; RUN: opt < %s -cost-model -analyze -mtriple=x86_64-- -mcpu=goldmont | FileCheck %s --check-prefixes=SSE
-; RUN: opt < %s -cost-model -analyze -mtriple=x86_64-- -mcpu=btver2 | FileCheck %s --check-prefixes=AVX1
+; RUN: opt < %s -cost-model -analyze -mtriple=x86_64-- -mcpu=btver2 | FileCheck %s --check-prefixes=AVX,AVX1
 
 define i32 @f32(i32 %arg) {
 ; SSE-LABEL: 'f32'
 ; SSE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %F32 = call float @llvm.maxnum.f32(float undef, float undef)
 ; SSE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V2F32 = call <2 x float> @llvm.maxnum.v2f32(<2 x float> undef, <2 x float> undef)
 ; SSE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4F32 = call <4 x float> @llvm.maxnum.v4f32(<4 x float> undef, <4 x float> undef)
 ; SSE-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8F32 = call <8 x float> @llvm.maxnum.v8f32(<8 x float> undef, <8 x float> undef)
 ; SSE-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16F32 = call <16 x float> @llvm.maxnum.v16f32(<16 x float> undef, <16 x float> undef)
;
   %V4f64 = call <4 x double> @llvm.maxnum.v4f64(<4 x double> undef, <4 x double> undef)
   %V8f64 = call <8 x double> @llvm.maxnum.v8f64(<8 x double> undef, <8 x double> undef)
   %V16f64 = call <16 x double> @llvm.maxnum.v16f64(<16 x double> undef, <16 x double> undef)
   ret i32 undef
 }
 
 define i32 @f32_nnan(i32 %arg) {
 ; SSE-LABEL: 'f32_nnan'
-; SSE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %F32 = call nnan float @llvm.maxnum.f32(float undef, float undef)
-; SSE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V2F32 = call nnan <2 x float> @llvm.maxnum.v2f32(<2 x float> undef, <2 x float> undef)
-; SSE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4F32 = call nnan <4 x float> @llvm.maxnum.v4f32(<4 x float> undef, <4 x float> undef)
-; SSE-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8F32 = call nnan <8 x float> @llvm.maxnum.v8f32(<8 x float> undef, <8 x float> undef)
-; SSE-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16F32 = call nnan <16 x float> @llvm.maxnum.v16f32(<16 x float> undef, <16 x float> undef)
+; SSE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %F32 = call nnan float @llvm.maxnum.f32(float undef, float undef)
+; SSE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2F32 = call nnan <2 x float> @llvm.maxnum.v2f32(<2 x float> undef, <2 x float> undef)
+; SSE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4F32 = call nnan <4 x float> @llvm.maxnum.v4f32(<4 x float> undef, <4 x float> undef)
+; SSE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8F32 = call nnan <8 x float> @llvm.maxnum.v8f32(<8 x float> undef, <8 x float> undef)
+; SSE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16F32 = call nnan <16 x float> @llvm.maxnum.v16f32(<16 x float> undef, <16 x float> undef)
 ; SSE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
 ;
-; AVX1-LABEL: 'f32_nnan'
-; AVX1-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %F32 = call nnan float @llvm.maxnum.f32(float undef, float undef)
-; AVX1-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V2F32 = call nnan <2 x float> @llvm.maxnum.v2f32(<2 x float> undef, <2 x float> undef)
-; AVX1-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4F32 = call nnan <4 x float> @llvm.maxnum.v4f32(<4 x float> undef, <4 x float> undef)
-; AVX1-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V8F32 = call nnan <8 x float> @llvm.maxnum.v8f32(<8 x float> undef, <8 x float> undef)
-; AVX1-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V16F32 = call nnan <16 x float> @llvm.maxnum.v16f32(<16 x float> undef, <16 x float> undef)
-; AVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
-;
-; AVX2-LABEL: 'f32_nnan'
-; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %F32 = call nnan float @llvm.maxnum.f32(float undef, float undef)
-; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V2F32 = call nnan <2 x float> @llvm.maxnum.v2f32(<2 x float> undef, <2 x float> undef)
-; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4F32 = call nnan <4 x float> @llvm.maxnum.v4f32(<4 x float> undef, <4 x float> undef)
-; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V8F32 = call nnan <8 x float> @llvm.maxnum.v8f32(<8 x float> undef, <8 x float> undef)
-; AVX2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V16F32 = call nnan <16 x float> @llvm.maxnum.v16f32(<16 x float> undef, <16 x float> undef)
-; AVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+; AVX-LABEL: 'f32_nnan'
+; AVX-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %F32 = call nnan float @llvm.maxnum.f32(float undef, float undef)
+; AVX-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2F32 = call nnan <2 x float> @llvm.maxnum.v2f32(<2 x float> undef, <2 x float> undef)
+; AVX-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4F32 = call nnan <4 x float> @llvm.maxnum.v4f32(<4 x float> undef, <4 x float> undef)
+; AVX-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8F32 = call nnan <8 x float> @llvm.maxnum.v8f32(<8 x float> undef, <8 x float> undef)
+; AVX-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16F32 = call nnan <16 x float> @llvm.maxnum.v16f32(<16 x float> undef, <16 x float> undef)
+; AVX-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
 ;
 ; AVX512-LABEL: 'f32_nnan'
-; AVX512-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %F32 = call nnan float @llvm.maxnum.f32(float undef, float undef)
-; AVX512-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V2F32 = call nnan <2 x float> @llvm.maxnum.v2f32(<2 x float> undef, <2 x float> undef)
-; AVX512-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4F32 = call nnan <4 x float> @llvm.maxnum.v4f32(<4 x float> undef, <4 x float> undef)
-; AVX512-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8F32 = call nnan <8 x float> @llvm.maxnum.v8f32(<8 x float> undef, <8 x float> undef)
-; AVX512-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16F32 = call nnan <16 x float> @llvm.maxnum.v16f32(<16 x float> undef, <16 x float> undef)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %F32 = call nnan float @llvm.maxnum.f32(float undef, float undef)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2F32 = call nnan <2 x float> @llvm.maxnum.v2f32(<2 x float> undef, <2 x float> undef)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4F32 = call nnan <4 x float> @llvm.maxnum.v4f32(<4 x float> undef, <4 x float> undef)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8F32 = call nnan <8 x float> @llvm.maxnum.v8f32(<8 x float> undef, <8 x float> undef)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16F32 = call nnan <16 x float> @llvm.maxnum.v16f32(<16 x float> undef, <16 x float> undef)
 ; AVX512-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
 ;
   %F32 = call nnan float @llvm.maxnum.f32(float undef, float undef)
   %V2F32 = call nnan <2 x float> @llvm.maxnum.v2f32(<2 x float> undef, <2 x float> undef)
   %V4F32 = call nnan <4 x float> @llvm.maxnum.v4f32(<4 x float> undef, <4 x float> undef)
   %V8F32 = call nnan <8 x float> @llvm.maxnum.v8f32(<8 x float> undef, <8 x float> undef)
   %V16F32 = call nnan <16 x float> @llvm.maxnum.v16f32(<16 x float> undef, <16 x float> undef)
   ret i32 undef
 }
 
 define i32 @f64_nnan(i32 %arg) {
 ; SSE-LABEL: 'f64_nnan'
-; SSE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %f64 = call nnan double @llvm.maxnum.f64(double undef, double undef)
-; SSE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V2f64 = call nnan <2 x double> @llvm.maxnum.v2f64(<2 x double> undef, <2 x double> undef)
-; SSE-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V4f64 = call nnan <4 x double> @llvm.maxnum.v4f64(<4 x double> undef, <4 x double> undef)
-; SSE-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V8f64 = call nnan <8 x double> @llvm.maxnum.v8f64(<8 x double> undef, <8 x double> undef)
-; SSE-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %V16f64 = call nnan <16 x double> @llvm.maxnum.v16f64(<16 x double> undef, <16 x double> undef)
+; SSE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %f64 = call nnan double @llvm.maxnum.f64(double undef, double undef)
+; SSE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2f64 = call nnan <2 x double> @llvm.maxnum.v2f64(<2 x double> undef, <2 x double> undef)
+; SSE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4f64 = call nnan <4 x double> @llvm.maxnum.v4f64(<4 x double> undef, <4 x double> undef)
+; SSE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8f64 = call nnan <8 x double> @llvm.maxnum.v8f64(<8 x double> undef, <8 x double> undef)
+; SSE-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16f64 = call nnan <16 x double> @llvm.maxnum.v16f64(<16 x double> undef, <16 x double> undef)
 ; SSE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
 ;
-; AVX1-LABEL: 'f64_nnan'
-; AVX1-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %f64 = call nnan double @llvm.maxnum.f64(double undef, double undef)
-; AVX1-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V2f64 = call nnan <2 x double> @llvm.maxnum.v2f64(<2 x double> undef, <2 x double> undef)
-; AVX1-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V4f64 = call nnan <4 x double> @llvm.maxnum.v4f64(<4 x double> undef, <4 x double> undef)
-; AVX1-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V8f64 = call nnan <8 x double> @llvm.maxnum.v8f64(<8 x double> undef, <8 x double> undef)
-; AVX1-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V16f64 = call nnan <16 x double> @llvm.maxnum.v16f64(<16 x double> undef, <16 x double> undef)
-; AVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
-;
-; AVX2-LABEL: 'f64_nnan'
-; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %f64 = call nnan double @llvm.maxnum.f64(double undef, double undef)
-; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V2f64 = call nnan <2 x double> @llvm.maxnum.v2f64(<2 x double> undef, <2 x double> undef)
-; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4f64 = call nnan <4 x double> @llvm.maxnum.v4f64(<4 x double> undef, <4 x double> undef)
-; AVX2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V8f64 = call nnan <8 x double> @llvm.maxnum.v8f64(<8 x double> undef, <8 x double> undef)
-; AVX2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V16f64 = call nnan <16 x double> @llvm.maxnum.v16f64(<16 x double> undef, <16 x double> undef)
-; AVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+; AVX-LABEL: 'f64_nnan'
+; AVX-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %f64 = call nnan double @llvm.maxnum.f64(double undef, double undef)
+; AVX-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2f64 = call nnan <2 x double> @llvm.maxnum.v2f64(<2 x double> undef, <2 x double> undef)
+; AVX-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4f64 = call nnan <4 x double> @llvm.maxnum.v4f64(<4 x double> undef, <4 x double> undef)
+; AVX-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8f64 = call nnan <8 x double> @llvm.maxnum.v8f64(<8 x double> undef, <8 x double> undef)
+; AVX-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16f64 = call nnan <16 x double> @llvm.maxnum.v16f64(<16 x double> undef, <16 x double> undef)
+; AVX-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
 ;
 ; AVX512-LABEL: 'f64_nnan'
-; AVX512-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %f64 = call nnan double @llvm.maxnum.f64(double undef, double undef)
-; AVX512-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V2f64 = call nnan <2 x double> @llvm.maxnum.v2f64(<2 x double> undef, <2 x double> undef)
-; AVX512-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4f64 = call nnan <4 x double> @llvm.maxnum.v4f64(<4 x double> undef, <4 x double> undef)
-; AVX512-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8f64 = call nnan <8 x double> @llvm.maxnum.v8f64(<8 x double> undef, <8 x double> undef)
-; AVX512-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16f64 = call nnan <16 x double> @llvm.maxnum.v16f64(<16 x double> undef, <16 x double> undef)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %f64 = call nnan double @llvm.maxnum.f64(double undef, double undef)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2f64 = call nnan <2 x double> @llvm.maxnum.v2f64(<2 x double> undef, <2 x double> undef)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4f64 = call nnan <4 x double> @llvm.maxnum.v4f64(<4 x double> undef, <4 x double> undef)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8f64 = call nnan <8 x double> @llvm.maxnum.v8f64(<8 x double> undef, <8 x double> undef)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16f64 = call nnan <16 x double> @llvm.maxnum.v16f64(<16 x double> undef, <16 x double> undef)
 ; AVX512-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
 ;
   %f64 = call nnan double @llvm.maxnum.f64(double undef, double undef)
   %V2f64 = call nnan <2 x double> @llvm.maxnum.v2f64(<2 x double> undef, <2 x double> undef)
   %V4f64 = call nnan <4 x double> @llvm.maxnum.v4f64(<4 x double> undef, <4 x double> undef)
   %V8f64 = call nnan <8 x double> @llvm.maxnum.v8f64(<8 x double> undef, <8 x double> undef)
   %V16f64 = call nnan <16 x double> @llvm.maxnum.v16f64(<16 x double> undef, <16 x double> undef)
   ret i32 undef

llvm/test/Analysis/CostModel/X86/fminnum.ll

 ; NOTE: Assertions have been autogenerated by utils/update_analyze_test_checks.py
 ; RUN: opt < %s -cost-model -analyze -mtriple=x86_64-- -mattr=+sse2 | FileCheck %s --check-prefixes=SSE
 ; RUN: opt < %s -cost-model -analyze -mtriple=x86_64-- -mattr=+sse4.2 | FileCheck %s --check-prefixes=SSE
-; RUN: opt < %s -cost-model -analyze -mtriple=x86_64-- -mattr=+avx | FileCheck %s --check-prefixes=AVX1
-; RUN: opt < %s -cost-model -analyze -mtriple=x86_64-- -mattr=+avx2 | FileCheck %s --check-prefixes=AVX2
+; RUN: opt < %s -cost-model -analyze -mtriple=x86_64-- -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX1
+; RUN: opt < %s -cost-model -analyze -mtriple=x86_64-- -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX2
 ; RUN: opt < %s -cost-model -analyze -mtriple=x86_64-- -mattr=+avx512f | FileCheck %s --check-prefixes=AVX512
 ; RUN: opt < %s -cost-model -analyze -mtriple=x86_64-- -mattr=+avx512f,+avx512bw | FileCheck %s --check-prefixes=AVX512
 ;
 ; RUN: opt < %s -cost-model -analyze -mtriple=x86_64-- -mcpu=slm | FileCheck %s --check-prefixes=SSE
 ; RUN: opt < %s -cost-model -analyze -mtriple=x86_64-- -mcpu=goldmont | FileCheck %s --check-prefixes=SSE
-; RUN: opt < %s -cost-model -analyze -mtriple=x86_64-- -mcpu=btver2 | FileCheck %s --check-prefixes=AVX1
+; RUN: opt < %s -cost-model -analyze -mtriple=x86_64-- -mcpu=btver2 | FileCheck %s --check-prefixes=AVX,AVX1
 
 define i32 @f32(i32 %arg) {
 ; SSE-LABEL: 'f32'
 ; SSE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %F32 = call float @llvm.minnum.f32(float undef, float undef)
 ; SSE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V2F32 = call <2 x float> @llvm.minnum.v2f32(<2 x float> undef, <2 x float> undef)
 ; SSE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4F32 = call <4 x float> @llvm.minnum.v4f32(<4 x float> undef, <4 x float> undef)
 ; SSE-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8F32 = call <8 x float> @llvm.minnum.v8f32(<8 x float> undef, <8 x float> undef)
 ; SSE-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16F32 = call <16 x float> @llvm.minnum.v16f32(<16 x float> undef, <16 x float> undef)
;
   %V4f64 = call <4 x double> @llvm.minnum.v4f64(<4 x double> undef, <4 x double> undef)
   %V8f64 = call <8 x double> @llvm.minnum.v8f64(<8 x double> undef, <8 x double> undef)
   %V16f64 = call <16 x double> @llvm.minnum.v16f64(<16 x double> undef, <16 x double> undef)
   ret i32 undef
 }
 
 define i32 @f32_nnan(i32 %arg) {
 ; SSE-LABEL: 'f32_nnan'
-; SSE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %F32 = call nnan float @llvm.minnum.f32(float undef, float undef)
-; SSE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V2F32 = call nnan <2 x float> @llvm.minnum.v2f32(<2 x float> undef, <2 x float> undef)
-; SSE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4F32 = call nnan <4 x float> @llvm.minnum.v4f32(<4 x float> undef, <4 x float> undef)
-; SSE-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8F32 = call nnan <8 x float> @llvm.minnum.v8f32(<8 x float> undef, <8 x float> undef)
-; SSE-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16F32 = call nnan <16 x float> @llvm.minnum.v16f32(<16 x float> undef, <16 x float> undef)
+; SSE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %F32 = call nnan float @llvm.minnum.f32(float undef, float undef)
+; SSE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2F32 = call nnan <2 x float> @llvm.minnum.v2f32(<2 x float> undef, <2 x float> undef)
+; SSE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4F32 = call nnan <4 x float> @llvm.minnum.v4f32(<4 x float> undef, <4 x float> undef)
+; SSE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8F32 = call nnan <8 x float> @llvm.minnum.v8f32(<8 x float> undef, <8 x float> undef)
+; SSE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16F32 = call nnan <16 x float> @llvm.minnum.v16f32(<16 x float> undef, <16 x float> undef)
 ; SSE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
 ;
-; AVX1-LABEL: 'f32_nnan'
-; AVX1-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %F32 = call nnan float @llvm.minnum.f32(float undef, float undef)
-; AVX1-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V2F32 = call nnan <2 x float> @llvm.minnum.v2f32(<2 x float> undef, <2 x float> undef)
-; AVX1-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4F32 = call nnan <4 x float> @llvm.minnum.v4f32(<4 x float> undef, <4 x float> undef)
-; AVX1-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V8F32 = call nnan <8 x float> @llvm.minnum.v8f32(<8 x float> undef, <8 x float> undef)
-; AVX1-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V16F32 = call nnan <16 x float> @llvm.minnum.v16f32(<16 x float> undef, <16 x float> undef)
-; AVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
-;
-; AVX2-LABEL: 'f32_nnan'
-; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %F32 = call nnan float @llvm.minnum.f32(float undef, float undef)
-; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V2F32 = call nnan <2 x float> @llvm.minnum.v2f32(<2 x float> undef, <2 x float> undef)
-; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4F32 = call nnan <4 x float> @llvm.minnum.v4f32(<4 x float> undef, <4 x float> undef)
-; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V8F32 = call nnan <8 x float> @llvm.minnum.v8f32(<8 x float> undef, <8 x float> undef)
-; AVX2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V16F32 = call nnan <16 x float> @llvm.minnum.v16f32(<16 x float> undef, <16 x float> undef)
-; AVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+; AVX-LABEL: 'f32_nnan'
+; AVX-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %F32 = call nnan float @llvm.minnum.f32(float undef, float undef)
+; AVX-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2F32 = call nnan <2 x float> @llvm.minnum.v2f32(<2 x float> undef, <2 x float> undef)
+; AVX-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4F32 = call nnan <4 x float> @llvm.minnum.v4f32(<4 x float> undef, <4 x float> undef)
+; AVX-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8F32 = call nnan <8 x float> @llvm.minnum.v8f32(<8 x float> undef, <8 x float> undef)
+; AVX-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16F32 = call nnan <16 x float> @llvm.minnum.v16f32(<16 x float> undef, <16 x float> undef)
+; AVX-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
 ;
 ; AVX512-LABEL: 'f32_nnan'
-; AVX512-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %F32 = call nnan float @llvm.minnum.f32(float undef, float undef)
-; AVX512-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V2F32 = call nnan <2 x float> @llvm.minnum.v2f32(<2 x float> undef, <2 x float> undef)
-; AVX512-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4F32 = call nnan <4 x float> @llvm.minnum.v4f32(<4 x float> undef, <4 x float> undef)
-; AVX512-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8F32 = call nnan <8 x float> @llvm.minnum.v8f32(<8 x float> undef, <8 x float> undef)
-; AVX512-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16F32 = call nnan <16 x float> @llvm.minnum.v16f32(<16 x float> undef, <16 x float> undef)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %F32 = call nnan float @llvm.minnum.f32(float undef, float undef)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2F32 = call nnan <2 x float> @llvm.minnum.v2f32(<2 x float> undef, <2 x float> undef)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4F32 = call nnan <4 x float> @llvm.minnum.v4f32(<4 x float> undef, <4 x float> undef)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8F32 = call nnan <8 x float> @llvm.minnum.v8f32(<8 x float> undef, <8 x float> undef)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16F32 = call nnan <16 x float> @llvm.minnum.v16f32(<16 x float> undef, <16 x float> undef)
 ; AVX512-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
 ;
   %F32 = call nnan float @llvm.minnum.f32(float undef, float undef)
   %V2F32 = call nnan <2 x float> @llvm.minnum.v2f32(<2 x float> undef, <2 x float> undef)
   %V4F32 = call nnan <4 x float> @llvm.minnum.v4f32(<4 x float> undef, <4 x float> undef)
   %V8F32 = call nnan <8 x float> @llvm.minnum.v8f32(<8 x float> undef, <8 x float> undef)
   %V16F32 = call nnan <16 x float> @llvm.minnum.v16f32(<16 x float> undef, <16 x float> undef)
   ret i32 undef
 }
 
 define i32 @f64_nnan(i32 %arg) {
 ; SSE-LABEL: 'f64_nnan'
-; SSE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %f64 = call nnan double @llvm.minnum.f64(double undef, double undef)
-; SSE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V2f64 = call nnan <2 x double> @llvm.minnum.v2f64(<2 x double> undef, <2 x double> undef)
-; SSE-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V4f64 = call nnan <4 x double> @llvm.minnum.v4f64(<4 x double> undef, <4 x double> undef)
-; SSE-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V8f64 = call nnan <8 x double> @llvm.minnum.v8f64(<8 x double> undef, <8 x double> undef)
-; SSE-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %V16f64 = call nnan <16 x double> @llvm.minnum.v16f64(<16 x double> undef, <16 x double> undef)
+; SSE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %f64 = call nnan double @llvm.minnum.f64(double undef, double undef)
+; SSE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2f64 = call nnan <2 x double> @llvm.minnum.v2f64(<2 x double> undef, <2 x double> undef)
+; SSE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4f64 = call nnan <4 x double> @llvm.minnum.v4f64(<4 x double> undef, <4 x double> undef)
+; SSE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8f64 = call nnan <8 x double> @llvm.minnum.v8f64(<8 x double> undef, <8 x double> undef)
+; SSE-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16f64 = call nnan <16 x double> @llvm.minnum.v16f64(<16 x double> undef, <16 x double> undef)
 ; SSE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
 ;
-; AVX1-LABEL: 'f64_nnan'
-; AVX1-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %f64 = call nnan double @llvm.minnum.f64(double undef, double undef)
-; AVX1-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V2f64 = call nnan <2 x double> @llvm.minnum.v2f64(<2 x double> undef, <2 x double> undef)
-; AVX1-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V4f64 = call nnan <4 x double> @llvm.minnum.v4f64(<4 x double> undef, <4 x double> undef)
-; AVX1-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V8f64 = call nnan <8 x double> @llvm.minnum.v8f64(<8 x double> undef, <8 x double> undef)
-; AVX1-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V16f64 = call nnan <16 x double> @llvm.minnum.v16f64(<16 x double> undef, <16 x double> undef)
-; AVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
-;
-; AVX2-LABEL: 'f64_nnan'
-; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %f64 = call nnan double @llvm.minnum.f64(double undef, double undef)
-; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V2f64 = call nnan <2 x double> @llvm.minnum.v2f64(<2 x double> undef, <2 x double> undef)
-; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4f64 = call nnan <4 x double> @llvm.minnum.v4f64(<4 x double> undef, <4 x double> undef)
-; AVX2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V8f64 = call nnan <8 x double> @llvm.minnum.v8f64(<8 x double> undef, <8 x double> undef)
-; AVX2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V16f64 = call nnan <16 x double> @llvm.minnum.v16f64(<16 x double> undef, <16 x double> undef)
-; AVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+; AVX-LABEL: 'f64_nnan'
+; AVX-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %f64 = call nnan double @llvm.minnum.f64(double undef, double undef)
+; AVX-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2f64 = call nnan <2 x double> @llvm.minnum.v2f64(<2 x double> undef, <2 x double> undef)
+; AVX-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4f64 = call nnan <4 x double> @llvm.minnum.v4f64(<4 x double> undef, <4 x double> undef)
+; AVX-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8f64 = call nnan <8 x double> @llvm.minnum.v8f64(<8 x double> undef, <8 x double> undef)
+; AVX-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16f64 = call nnan <16 x double> @llvm.minnum.v16f64(<16 x double> undef, <16 x double> undef)
+; AVX-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
 ;
 ; AVX512-LABEL: 'f64_nnan'
-; AVX512-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %f64 = call nnan double @llvm.minnum.f64(double undef, double undef)
-; AVX512-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V2f64 = call nnan <2 x double> @llvm.minnum.v2f64(<2 x double> undef, <2 x double> undef)
-; AVX512-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4f64 = call nnan <4 x double> @llvm.minnum.v4f64(<4 x double> undef, <4 x double> undef)
-; AVX512-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8f64 = call nnan <8 x double> @llvm.minnum.v8f64(<8 x double> undef, <8 x double> undef)
-; AVX512-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16f64 = call nnan <16 x double> @llvm.minnum.v16f64(<16 x double> undef, <16 x double> undef)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %f64 = call nnan double @llvm.minnum.f64(double undef, double undef)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2f64 = call nnan <2 x double> @llvm.minnum.v2f64(<2 x double> undef, <2 x double> undef)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4f64 = call nnan <4 x double> @llvm.minnum.v4f64(<4 x double> undef, <4 x double> undef)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8f64 = call nnan <8 x double> @llvm.minnum.v8f64(<8 x double> undef, <8 x double> undef)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16f64 = call nnan <16 x double> @llvm.minnum.v16f64(<16 x double> undef, <16 x double> undef)
 ; AVX512-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
 ;
   %f64 = call nnan double @llvm.minnum.f64(double undef, double undef)
   %V2f64 = call nnan <2 x double> @llvm.minnum.v2f64(<2 x double> undef, <2 x double> undef)
   %V4f64 = call nnan <4 x double> @llvm.minnum.v4f64(<4 x double> undef, <4 x double> undef)
   %V8f64 = call nnan <8 x double> @llvm.minnum.v8f64(<8 x double> undef, <8 x double> undef)
   %V16f64 = call nnan <16 x double> @llvm.minnum.v16f64(<16 x double> undef, <16 x double> undef)
   ret i32 undef