[X86] Make X86TTIImpl::getCastInstrCost properly handle the case where AVX512 is enabled, but 512-bit vectors aren't legal.

Unlike most cost model functions this code makes a lot of table lookups without using the results from getTypeLegalizationCost. This means 512-bit vectors can be looked up even when the type isn't legal.

This patch adds a check around the two tables that contain 512-bit types to make sure that neither of the types would be split by type legalization. Meaning 512 bit types are illegal. I wanted to write this in a somewhat generic way that uses type legalization query hooks. But if prefered, I can switch to just using is512BitVector and the subtarget feature.

Differential Revision: https://reviews.llvm.org/D54984

llvm-svn: 347786

2 files changed, 40 insertions, 55 deletions

diff --git a/llvm/lib/Target/X86/X86TargetTransformInfo.cpp b/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
index a77d8e04451..9ead6a614fc 100644
--- a/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
+++ b/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
@@ -1570,49 +1570,51 @@ int X86TTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
   if (!SrcTy.isSimple() || !DstTy.isSimple())
     return BaseT::getCastInstrCost(Opcode, Dst, Src);
 
-  if (ST->hasBWI())
-    if (const auto *Entry = ConvertCostTableLookup(AVX512BWConversionTbl, ISD,
-                                                   DstTy.getSimpleVT(),
-                                                   SrcTy.getSimpleVT()))
-      return Entry->Cost;
+  MVT SimpleSrcTy = SrcTy.getSimpleVT();
+  MVT SimpleDstTy = DstTy.getSimpleVT();
+
+  // Make sure that neither type is going to be split before using the
+  // AVX512 tables. This handles -mprefer-vector-width=256
+  // with -min-legal-vector-width<=256
+  if (TLI->getTypeAction(SimpleSrcTy) != TargetLowering::TypeSplitVector &&
+      TLI->getTypeAction(SimpleDstTy) != TargetLowering::TypeSplitVector) {
+    if (ST->hasBWI())
+      if (const auto *Entry = ConvertCostTableLookup(AVX512BWConversionTbl, ISD,
+                                                     SimpleDstTy, SimpleSrcTy))
+        return Entry->Cost;
 
-  if (ST->hasDQI())
-    if (const auto *Entry = ConvertCostTableLookup(AVX512DQConversionTbl, ISD,
-                                                   DstTy.getSimpleVT(),
-                                                   SrcTy.getSimpleVT()))
-      return Entry->Cost;
+    if (ST->hasDQI())
+      if (const auto *Entry = ConvertCostTableLookup(AVX512DQConversionTbl, ISD,
+                                                     SimpleDstTy, SimpleSrcTy))
+        return Entry->Cost;
 
-  if (ST->hasAVX512())
-    if (const auto *Entry = ConvertCostTableLookup(AVX512FConversionTbl, ISD,
-                                                   DstTy.getSimpleVT(),
-                                                   SrcTy.getSimpleVT()))
-      return Entry->Cost;
+    if (ST->hasAVX512())
+      if (const auto *Entry = ConvertCostTableLookup(AVX512FConversionTbl, ISD,
+                                                     SimpleDstTy, SimpleSrcTy))
+        return Entry->Cost;
+  }
 
   if (ST->hasAVX2()) {
     if (const auto *Entry = ConvertCostTableLookup(AVX2ConversionTbl, ISD,
-                                                   DstTy.getSimpleVT(),
-                                                   SrcTy.getSimpleVT()))
+                                                   SimpleDstTy, SimpleSrcTy))
       return Entry->Cost;
   }
 
   if (ST->hasAVX()) {
     if (const auto *Entry = ConvertCostTableLookup(AVXConversionTbl, ISD,
-                                                   DstTy.getSimpleVT(),
-                                                   SrcTy.getSimpleVT()))
+                                                   SimpleDstTy, SimpleSrcTy))
       return Entry->Cost;
   }
 
   if (ST->hasSSE41()) {
     if (const auto *Entry = ConvertCostTableLookup(SSE41ConversionTbl, ISD,
-                                                   DstTy.getSimpleVT(),
-                                                   SrcTy.getSimpleVT()))
+                                                   SimpleDstTy, SimpleSrcTy))
       return Entry->Cost;
   }
 
   if (ST->hasSSE2()) {
     if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD,
-                                                   DstTy.getSimpleVT(),
-                                                   SrcTy.getSimpleVT()))
+                                                   SimpleDstTy, SimpleSrcTy))
       return Entry->Cost;
   }
 
diff --git a/llvm/test/Analysis/CostModel/X86/min-legal-vector-width.ll b/llvm/test/Analysis/CostModel/X86/min-legal-vector-width.ll
index 3cd1b335f87..028ed87a989 100644
--- a/llvm/test/Analysis/CostModel/X86/min-legal-vector-width.ll
+++ b/llvm/test/Analysis/CostModel/X86/min-legal-vector-width.ll
@@ -4,21 +4,13 @@
 ; RUN: opt < %s  -cost-model -analyze -mtriple=x86_64-apple-macosx10.8.0 -mattr=+avx512vl,+avx512bw,+avx512dq,-prefer-256-bit | FileCheck %s --check-prefixes=CHECK,VEC512
 
 define void @zext256() "min-legal-vector-width"="256" {
-; AVX-LABEL: 'zext256'
-; AVX-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %A = zext <8 x i16> undef to <8 x i64>
-; AVX-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %B = zext <8 x i32> undef to <8 x i64>
-; AVX-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %C = zext <16 x i8> undef to <16 x i32>
-; AVX-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %D = zext <16 x i16> undef to <16 x i32>
-; AVX-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %E = zext <32 x i8> undef to <32 x i16>
-; AVX-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret void
-;
-; SKX256-LABEL: 'zext256'
-; SKX256-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %A = zext <8 x i16> undef to <8 x i64>
-; SKX256-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %B = zext <8 x i32> undef to <8 x i64>
-; SKX256-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %C = zext <16 x i8> undef to <16 x i32>
-; SKX256-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %D = zext <16 x i16> undef to <16 x i32>
-; SKX256-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %E = zext <32 x i8> undef to <32 x i16>
-; SKX256-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret void
+; VEC256-LABEL: 'zext256'
+; VEC256-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %A = zext <8 x i16> undef to <8 x i64>
+; VEC256-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %B = zext <8 x i32> undef to <8 x i64>
+; VEC256-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %C = zext <16 x i8> undef to <16 x i32>
+; VEC256-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %D = zext <16 x i16> undef to <16 x i32>
+; VEC256-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %E = zext <32 x i8> undef to <32 x i16>
+; VEC256-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret void
 ;
 ; VEC512-LABEL: 'zext256'
 ; VEC512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %A = zext <8 x i16> undef to <8 x i64>
@@ -70,23 +62,14 @@ define void @zext512() "min-legal-vector-width"="512" {
 }
 
 define void @sext256() "min-legal-vector-width"="256" {
-; AVX-LABEL: 'sext256'
-; AVX-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %A = sext <8 x i8> undef to <8 x i64>
-; AVX-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %B = sext <8 x i16> undef to <8 x i64>
-; AVX-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %C = sext <8 x i32> undef to <8 x i64>
-; AVX-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %D = sext <16 x i8> undef to <16 x i32>
-; AVX-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %E = sext <16 x i16> undef to <16 x i32>
-; AVX-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %F = sext <32 x i8> undef to <32 x i16>
-; AVX-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret void
-;
-; SKX256-LABEL: 'sext256'
-; SKX256-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %A = sext <8 x i8> undef to <8 x i64>
-; SKX256-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %B = sext <8 x i16> undef to <8 x i64>
-; SKX256-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %C = sext <8 x i32> undef to <8 x i64>
-; SKX256-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %D = sext <16 x i8> undef to <16 x i32>
-; SKX256-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %E = sext <16 x i16> undef to <16 x i32>
-; SKX256-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %F = sext <32 x i8> undef to <32 x i16>
-; SKX256-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret void
+; VEC256-LABEL: 'sext256'
+; VEC256-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %A = sext <8 x i8> undef to <8 x i64>
+; VEC256-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %B = sext <8 x i16> undef to <8 x i64>
+; VEC256-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %C = sext <8 x i32> undef to <8 x i64>
+; VEC256-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %D = sext <16 x i8> undef to <16 x i32>
+; VEC256-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %E = sext <16 x i16> undef to <16 x i32>
+; VEC256-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %F = sext <32 x i8> undef to <32 x i16>
+; VEC256-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret void
 ;
 ; VEC512-LABEL: 'sext256'
 ; VEC512-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %A = sext <8 x i8> undef to <8 x i64>