[llvm] r291366 - [CostModel][X86] Fix AVX2 v16i16 shift 'splat' costs.

[Simon Pilgrim via llvm-commits]

Author: rksimon
Date: Sat Jan  7 16:08:09 2017
New Revision: 291366

URL: http://llvm.org/viewvc/llvm-project?rev=291366&view=rev
Log:
[CostModel][X86] Fix AVX2 v16i16 shift 'splat' costs.

Modified:
    llvm/trunk/lib/Target/X86/X86TargetTransformInfo.cpp
    llvm/trunk/test/Analysis/CostModel/X86/vshift-ashr-cost.ll
    llvm/trunk/test/Analysis/CostModel/X86/vshift-lshr-cost.ll

Modified: llvm/trunk/lib/Target/X86/X86TargetTransformInfo.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86TargetTransformInfo.cpp?rev=291366&r1=291365&r2=291366&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86TargetTransformInfo.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86TargetTransformInfo.cpp Sat Jan  7 16:08:09 2017
@@ -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.

Modified: llvm/trunk/test/Analysis/CostModel/X86/vshift-ashr-cost.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Analysis/CostModel/X86/vshift-ashr-cost.ll?rev=291366&r1=291365&r2=291366&view=diff
==============================================================================
--- llvm/trunk/test/Analysis/CostModel/X86/vshift-ashr-cost.ll (original)
+++ llvm/trunk/test/Analysis/CostModel/X86/vshift-ashr-cost.ll Sat Jan  7 16:08:09 2017
@@ -578,9 +578,10 @@ define <16 x i16> @splatconstant_shift_v
 ; 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_v
 ; 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
 }

Modified: llvm/trunk/test/Analysis/CostModel/X86/vshift-lshr-cost.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Analysis/CostModel/X86/vshift-lshr-cost.ll?rev=291366&r1=291365&r2=291366&view=diff
==============================================================================
--- llvm/trunk/test/Analysis/CostModel/X86/vshift-lshr-cost.ll (original)
+++ llvm/trunk/test/Analysis/CostModel/X86/vshift-lshr-cost.ll Sat Jan  7 16:08:09 2017
@@ -589,9 +589,10 @@ define <16 x i16> @splatconstant_shift_v
 ; 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_v
 ; 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
 }