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

Tweak cost model to match what lowering actually does.

llvm-svn: 303013

3 files changed, 12 insertions, 9 deletions

diff --git a/llvm/lib/Target/X86/X86TargetTransformInfo.cpp b/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
index ba6462a0768..2952925ee41 100644
--- a/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
+++ b/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
@@ -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[] = {
diff --git a/llvm/test/Analysis/CostModel/X86/vshift-ashr-cost.ll b/llvm/test/Analysis/CostModel/X86/vshift-ashr-cost.ll
index d57e0a2824a..dd1003fe574 100644
--- a/llvm/test/Analysis/CostModel/X86/vshift-ashr-cost.ll
+++ b/llvm/test/Analysis/CostModel/X86/vshift-ashr-cost.ll
@@ -628,7 +628,7 @@ define <16 x i8> @splatconstant_shift_v16i8(<16 x i8> %a) {
 ; 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_v32i8(<32 x i8> %a) {
 ; 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_v64i8(<64 x i8> %a) {
 ; 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
diff --git a/llvm/test/Analysis/CostModel/X86/vshift-shl-cost.ll b/llvm/test/Analysis/CostModel/X86/vshift-shl-cost.ll
index e2f1a359b3e..99ffced8d7b 100644
--- a/llvm/test/Analysis/CostModel/X86/vshift-shl-cost.ll
+++ b/llvm/test/Analysis/CostModel/X86/vshift-shl-cost.ll
@@ -638,7 +638,7 @@ define <16 x i8> @splatconstant_shift_v16i8(<16 x i8> %a) {
 ; 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_v32i8(<32 x i8> %a) {
 ; 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_v64i8(<64 x i8> %a) {
 ; 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