[X86][AVX2] Fix costs for v4i64 ashr by splat

llvm-svn: 303022

diff --git a/llvm/lib/Target/X86/X86TargetTransformInfo.cpp b/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
index 4ea66d2..4a5c7e3 100644 (file)
--- a/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
+++ b/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
@@ -471,6 +471,11 @@ int X86TTIImpl::getArithmeticInstrCost(
   if (ST->hasSSE2() &&
       ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
        (Op2Info == TargetTransformInfo::OK_UniformValue))) {
+
+    // Handle AVX2 uniform v4i64 ISD::SRA, it's not worth a table.
+    if (ISD == ISD::SRA && LT.second == MVT::v4i64 && ST->hasAVX2())
+      return LT.first * 4; // 2*psrad + shuffle.
+
     if (const auto *Entry =
             CostTableLookup(SSE2UniformShiftCostTable, ISD, LT.second))
       return LT.first * Entry->Cost;

diff --git a/llvm/test/Analysis/CostModel/X86/vshift-ashr-cost.ll b/llvm/test/Analysis/CostModel/X86/vshift-ashr-cost.ll
index c2e2960..9b3b58a 100644 (file)
--- a/llvm/test/Analysis/CostModel/X86/vshift-ashr-cost.ll
+++ b/llvm/test/Analysis/CostModel/X86/vshift-ashr-cost.ll
@@ -192,7 +192,7 @@ define <4 x i64> @splatvar_shift_v4i64(<4 x i64> %a, i64 %b) {
 ; SSE2: Found an estimated cost of 8 for instruction:   %shift
 ; SSE41: Found an estimated cost of 8 for instruction:   %shift
 ; AVX: Found an estimated cost of 10 for instruction:   %shift
-; AVX2: 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 1 for instruction:   %shift
 ; XOP: Found an estimated cost of 6 for instruction:   %shift
   %insert = insertelement <4 x i64> undef, i64 %b, i32 0
@@ -206,7 +206,7 @@ define <8 x i64> @splatvar_shift_v8i64(<8 x i64> %a, i64 %b) {
 ; SSE2: Found an estimated cost of 16 for instruction:   %shift
 ; SSE41: Found an estimated cost of 16 for instruction:   %shift
 ; AVX: Found an estimated cost of 20 for instruction:   %shift
-; AVX2: Found an estimated cost of 20 for instruction:   %shift
+; AVX2: Found an estimated cost of 8 for instruction:   %shift
 ; AVX512: Found an estimated cost of 1 for instruction:   %shift
 ; XOP: Found an estimated cost of 12 for instruction:   %shift
   %insert = insertelement <8 x i64> undef, i64 %b, i32 0