[llvm] r303013 - [X86][XOP] XOP's general v16i8 shifts will be used instead of v8i16 shift + mask.

[Simon Pilgrim via llvm-commits]

Author: rksimon
Date: Sun May 14 12:59:46 2017
New Revision: 303013

URL: http://llvm.org/viewvc/llvm-project?rev=303013&view=rev
Log:
[X86][XOP] XOP's general v16i8 shifts will be used instead of v8i16 shift + mask.

Tweak cost model to match what lowering actually does.

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-shl-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=303013&r1=303012&r2=303013&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86TargetTransformInfo.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86TargetTransformInfo.cpp Sun May 14 12:59:46 2017
@@ -273,9 +273,12 @@ int X86TTIImpl::getArithmeticInstrCost(
     if (ISD == ISD::SDIV && LT.second == MVT::v4i32 && ST->hasSSE41())
       return LT.first * 15;
 
-    if (const auto *Entry = CostTableLookup(SSE2UniformConstCostTable, ISD,
-                                            LT.second))
-      return LT.first * Entry->Cost;
+    // XOP has faster vXi8 shifts.
+    if ((ISD != ISD::SHL && ISD != ISD::SRL && ISD != ISD::SRA) ||
+        !ST->hasXOP())
+      if (const auto *Entry =
+              CostTableLookup(SSE2UniformConstCostTable, ISD, LT.second))
+        return LT.first * Entry->Cost;
   }
 
   static const CostTblEntry AVX2UniformCostTable[] = {

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=303013&r1=303012&r2=303013&view=diff
==============================================================================
--- llvm/trunk/test/Analysis/CostModel/X86/vshift-ashr-cost.ll (original)
+++ llvm/trunk/test/Analysis/CostModel/X86/vshift-ashr-cost.ll Sun May 14 12:59:46 2017
@@ -628,7 +628,7 @@ define <16 x i8> @splatconstant_shift_v1
 ; AVX: Found an estimated cost of 4 for instruction:   %shift
 ; AVX2: Found an estimated cost of 4 for instruction:   %shift
 ; AVX512: Found an estimated cost of 4 for instruction:   %shift
-; XOP: Found an estimated cost of 4 for instruction:   %shift
+; XOP: Found an estimated cost of 2 for instruction:   %shift
   %shift = ashr <16 x i8> %a, <i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3>
   ret <16 x i8> %shift
 }
@@ -640,7 +640,7 @@ define <32 x i8> @splatconstant_shift_v3
 ; AVX: Found an estimated cost of 10 for instruction:   %shift
 ; AVX2: Found an estimated cost of 4 for instruction:   %shift
 ; AVX512: Found an estimated cost of 4 for instruction:   %shift
-; XOPAVX: Found an estimated cost of 10 for instruction:   %shift
+; XOPAVX: Found an estimated cost of 6 for instruction:   %shift
 ; XOPAVX2: Found an estimated cost of 4 for instruction:   %shift
   %shift = ashr <32 x i8> %a, <i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3>
   ret <32 x i8> %shift
@@ -654,7 +654,7 @@ define <64 x i8> @splatconstant_shift_v6
 ; AVX2: Found an estimated cost of 8 for instruction:   %shift
 ; AVX512F: Found an estimated cost of 8 for instruction:   %shift
 ; AVX512BW: Found an estimated cost of 4 for instruction:   %shift
-; XOPAVX: Found an estimated cost of 20 for instruction:   %shift
+; XOPAVX: Found an estimated cost of 12 for instruction:   %shift
 ; XOPAVX2: Found an estimated cost of 8 for instruction:   %shift
   %shift = ashr <64 x i8> %a, <i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3>
   ret <64 x i8> %shift

Modified: llvm/trunk/test/Analysis/CostModel/X86/vshift-shl-cost.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Analysis/CostModel/X86/vshift-shl-cost.ll?rev=303013&r1=303012&r2=303013&view=diff
==============================================================================
--- llvm/trunk/test/Analysis/CostModel/X86/vshift-shl-cost.ll (original)
+++ llvm/trunk/test/Analysis/CostModel/X86/vshift-shl-cost.ll Sun May 14 12:59:46 2017
@@ -638,7 +638,7 @@ define <16 x i8> @splatconstant_shift_v1
 ; 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 2 for instruction:   %shift
+; XOP: Found an estimated cost of 1 for instruction:   %shift
   %shift = shl <16 x i8> %a, <i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3>
   ret <16 x i8> %shift
 }
@@ -650,7 +650,7 @@ define <32 x i8> @splatconstant_shift_v3
 ; AVX: Found an estimated cost of 6 for instruction:   %shift
 ; AVX2: Found an estimated cost of 2 for instruction:   %shift
 ; AVX512: Found an estimated cost of 2 for instruction:   %shift
-; XOPAVX: Found an estimated cost of 6 for instruction:   %shift
+; XOPAVX: Found an estimated cost of 4 for instruction:   %shift
 ; XOPAVX2: Found an estimated cost of 2 for instruction:   %shift
   %shift = shl <32 x i8> %a, <i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3>
   ret <32 x i8> %shift
@@ -664,7 +664,7 @@ define <64 x i8> @splatconstant_shift_v6
 ; AVX2: Found an estimated cost of 4 for instruction:   %shift
 ; AVX512F: Found an estimated cost of 4 for instruction:   %shift
 ; AVX512BW: Found an estimated cost of 2 for instruction:   %shift
-; XOPAVX: Found an estimated cost of 12 for instruction:   %shift
+; XOPAVX: Found an estimated cost of 8 for instruction:   %shift
 ; XOPAVX2: Found an estimated cost of 4 for instruction:   %shift
   %shift = shl <64 x i8> %a, <i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3>
   ret <64 x i8> %shift