[CostModel][X86] Fix AVX2 v16i16 shift 'splat' costs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@291366 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86TargetTransformInfo.cpp b/lib/Target/X86/X86TargetTransformInfo.cpp
index 829b47b7aa274fed3e2ec7c4fe8a3d8d35915ccd..30b205552570cf89c475a5239f9afd8bbf1a4cea 100644 (file)
--- a/lib/Target/X86/X86TargetTransformInfo.cpp
+++ b/lib/Target/X86/X86TargetTransformInfo.cpp
@@ -291,6 +291,20 @@ int X86TTIImpl::getArithmeticInstrCost(
       return LT.first * Entry->Cost;
   }
 
+  static const CostTblEntry AVX2UniformCostTable[] = {
+    // Uniform splats are cheaper for the following instructions.
+    { ISD::SRL,  MVT::v16i16, 1 }, // psrlw.
+    { ISD::SRA,  MVT::v16i16, 1 }, // psraw.
+  };
+
+  if (ST->hasAVX2() &&
+      ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
+       (Op2Info == TargetTransformInfo::OK_UniformValue))) {
+    if (const auto *Entry =
+            CostTableLookup(AVX2UniformCostTable, ISD, LT.second))
+      return LT.first * Entry->Cost;
+  }
+
   static const CostTblEntry XOPShiftCostTable[] = {
     // 128bit shifts take 1cy, but right shifts require negation beforehand.
     { ISD::SHL,     MVT::v16i8,    1 },
@@ -325,8 +339,7 @@ int X86TTIImpl::getArithmeticInstrCost(
     if (const auto *Entry = CostTableLookup(XOPShiftCostTable, ISD, LT.second))
       return LT.first * Entry->Cost;
 
-  static const CostTblEntry
-  SSE2UniformCostTable[] = {
+  static const CostTblEntry SSE2UniformCostTable[] = {
     // Uniform splats are cheaper for the following instructions.
     { ISD::SHL,  MVT::v16i8,  1 }, // psllw.
     { ISD::SHL,  MVT::v32i8,  2 }, // psllw.

diff --git a/test/Analysis/CostModel/X86/vshift-ashr-cost.ll b/test/Analysis/CostModel/X86/vshift-ashr-cost.ll
index ab1eb730109ba047db1f55b945a94d691813ba8b..e75b5dc3ddc0ee3320661bb1fdab4ef618cb5d91 100644 (file)
--- a/test/Analysis/CostModel/X86/vshift-ashr-cost.ll
+++ b/test/Analysis/CostModel/X86/vshift-ashr-cost.ll
@@ -578,9 +578,10 @@ define <16 x i16> @splatconstant_shift_v16i16(<16 x i16> %a) {
 ; SSE2: Found an estimated cost of 2 for instruction:   %shift
 ; SSE41: Found an estimated cost of 2 for instruction:   %shift
 ; AVX: Found an estimated cost of 2 for instruction:   %shift
-; AVX2: Found an estimated cost of 2 for instruction:   %shift
-; AVX512: Found an estimated cost of 2 for instruction:   %shift
-; XOP: Found an estimated cost of 4 for instruction:   %shift
+; AVX2: Found an estimated cost of 1 for instruction:   %shift
+; AVX512: Found an estimated cost of 1 for instruction:   %shift
+; XOPAVX: Found an estimated cost of 4 for instruction:   %shift
+; XOPAVX2: Found an estimated cost of 1 for instruction:   %shift
   %shift = ashr <16 x i16> %a, <i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3>
   ret <16 x i16> %shift
 }
@@ -590,10 +591,11 @@ define <32 x i16> @splatconstant_shift_v32i16(<32 x i16> %a) {
 ; SSE2: Found an estimated cost of 4 for instruction:   %shift
 ; SSE41: Found an estimated cost of 4 for instruction:   %shift
 ; AVX: Found an estimated cost of 4 for instruction:   %shift
-; AVX2: Found an estimated cost of 4 for instruction:   %shift
-; AVX512F: Found an estimated cost of 4 for instruction:   %shift
+; AVX2: Found an estimated cost of 2 for instruction:   %shift
+; AVX512F: Found an estimated cost of 2 for instruction:   %shift
 ; AVX512BW: Found an estimated cost of 1 for instruction:   %shift
-; XOP: Found an estimated cost of 8 for instruction:   %shift
+; XOPAVX: Found an estimated cost of 8 for instruction:   %shift
+; XOPAVX2: Found an estimated cost of 2 for instruction:   %shift
   %shift = ashr <32 x i16> %a, <i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3>
   ret <32 x i16> %shift
 }

diff --git a/test/Analysis/CostModel/X86/vshift-lshr-cost.ll b/test/Analysis/CostModel/X86/vshift-lshr-cost.ll
index fea727147ff600879945e376848677835b300dfe..c1c674875c989952cd42be148ff491e18eb93428 100644 (file)
--- a/test/Analysis/CostModel/X86/vshift-lshr-cost.ll
+++ b/test/Analysis/CostModel/X86/vshift-lshr-cost.ll
@@ -589,9 +589,10 @@ define <16 x i16> @splatconstant_shift_v16i16(<16 x i16> %a) {
 ; SSE2: Found an estimated cost of 2 for instruction:   %shift
 ; SSE41: Found an estimated cost of 2 for instruction:   %shift
 ; AVX: Found an estimated cost of 2 for instruction:   %shift
-; AVX2: Found an estimated cost of 2 for instruction:   %shift
-; AVX512: Found an estimated cost of 2 for instruction:   %shift
-; XOP: Found an estimated cost of 4 for instruction:   %shift
+; AVX2: Found an estimated cost of 1 for instruction:   %shift
+; AVX512: Found an estimated cost of 1 for instruction:   %shift
+; XOPAVX: Found an estimated cost of 4 for instruction:   %shift
+; XOPAVX2: Found an estimated cost of 1 for instruction:   %shift
   %shift = lshr <16 x i16> %a, <i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3>
   ret <16 x i16> %shift
 }
@@ -601,10 +602,11 @@ define <32 x i16> @splatconstant_shift_v32i16(<32 x i16> %a) {
 ; SSE2: Found an estimated cost of 4 for instruction:   %shift
 ; SSE41: Found an estimated cost of 4 for instruction:   %shift
 ; AVX: Found an estimated cost of 4 for instruction:   %shift
-; AVX2: Found an estimated cost of 4 for instruction:   %shift
-; AVX512F: Found an estimated cost of 4 for instruction:   %shift
+; AVX2: Found an estimated cost of 2 for instruction:   %shift
+; AVX512F: Found an estimated cost of 2 for instruction:   %shift
 ; AVX512BW: Found an estimated cost of 1 for instruction:   %shift
-; XOP: Found an estimated cost of 8 for instruction:   %shift
+; XOPAVX: Found an estimated cost of 8 for instruction:   %shift
+; XOPAVX2: Found an estimated cost of 2 for instruction:   %shift
   %shift = lshr <32 x i16> %a, <i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3>
   ret <32 x i16> %shift
 }