[X86][SSE] Account for cost of extract/insert of v32i8 vector shifts

llvm-svn: 303012

diff --git a/llvm/lib/Target/X86/X86TargetTransformInfo.cpp b/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
index bfbd33e..ba6462a 100644 (file)
--- a/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
+++ b/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
@@ -251,9 +251,9 @@ int X86TTIImpl::getArithmeticInstrCost(
     { ISD::SRL,  MVT::v16i8,   2 }, // psrlw + pand.
     { ISD::SRA,  MVT::v16i8,   4 }, // psrlw, pand, pxor, psubb.
 
-    { ISD::SHL,  MVT::v32i8,   4 }, // 2*(psllw + pand).
-    { ISD::SRL,  MVT::v32i8,   4 }, // 2*(psrlw + pand).
-    { ISD::SRA,  MVT::v32i8,   8 }, // 2*(psrlw, pand, pxor, psubb).
+    { ISD::SHL,  MVT::v32i8, 4+2 }, // 2*(psllw + pand) + split.
+    { ISD::SRL,  MVT::v32i8, 4+2 }, // 2*(psrlw + pand) + split.
+    { ISD::SRA,  MVT::v32i8, 8+2 }, // 2*(psrlw, pand, pxor, psubb) + split.
 
     { ISD::SDIV, MVT::v16i16, 12 }, // pmulhw sequence
     { ISD::SDIV, MVT::v8i16,   6 }, // pmulhw sequence

diff --git a/llvm/test/Analysis/CostModel/X86/vshift-ashr-cost.ll b/llvm/test/Analysis/CostModel/X86/vshift-ashr-cost.ll
index f2c6922..d57e0a2 100644 (file)
--- a/llvm/test/Analysis/CostModel/X86/vshift-ashr-cost.ll
+++ b/llvm/test/Analysis/CostModel/X86/vshift-ashr-cost.ll
@@ -637,10 +637,10 @@ define <32 x i8> @splatconstant_shift_v32i8(<32 x i8> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'splatconstant_shift_v32i8':
 ; 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 8 for instruction:   %shift
+; 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 8 for instruction:   %shift
+; XOPAVX: Found an estimated cost of 10 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
@@ -650,11 +650,11 @@ define <64 x i8> @splatconstant_shift_v64i8(<64 x i8> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'splatconstant_shift_v64i8':
 ; 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 16 for instruction:   %shift
+; AVX: Found an estimated cost of 20 for instruction:   %shift
 ; 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 16 for instruction:   %shift
+; XOPAVX: Found an estimated cost of 20 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-lshr-cost.ll b/llvm/test/Analysis/CostModel/X86/vshift-lshr-cost.ll
index ab402d1..c19a696 100644 (file)
--- a/llvm/test/Analysis/CostModel/X86/vshift-lshr-cost.ll
+++ b/llvm/test/Analysis/CostModel/X86/vshift-lshr-cost.ll
@@ -644,10 +644,10 @@ define <32 x i8> @splatconstant_shift_v32i8(<32 x i8> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'splatconstant_shift_v32i8':
 ; 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
+; 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 4 for instruction:   %shift
+; XOPAVX: Found an estimated cost of 6 for instruction:   %shift
 ; XOPAVX2: Found an estimated cost of 2 for instruction:   %shift
   %shift = lshr <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
@@ -657,11 +657,11 @@ define <64 x i8> @splatconstant_shift_v64i8(<64 x i8> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'splatconstant_shift_v64i8':
 ; 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 8 for instruction:   %shift
+; AVX: Found an estimated cost of 12 for instruction:   %shift
 ; 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 8 for instruction:   %shift
+; XOPAVX: Found an estimated cost of 12 for instruction:   %shift
 ; XOPAVX2: Found an estimated cost of 4 for instruction:   %shift
   %shift = lshr <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 6ec524a..e2f1a35 100644 (file)
--- a/llvm/test/Analysis/CostModel/X86/vshift-shl-cost.ll
+++ b/llvm/test/Analysis/CostModel/X86/vshift-shl-cost.ll
@@ -647,10 +647,10 @@ define <32 x i8> @splatconstant_shift_v32i8(<32 x i8> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'splatconstant_shift_v32i8':
 ; 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
+; 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 4 for instruction:   %shift
+; XOPAVX: Found an estimated cost of 6 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
@@ -660,11 +660,11 @@ define <64 x i8> @splatconstant_shift_v64i8(<64 x i8> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'splatconstant_shift_v64i8':
 ; 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 8 for instruction:   %shift
+; AVX: Found an estimated cost of 12 for instruction:   %shift
 ; 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 8 for instruction:   %shift
+; XOPAVX: Found an estimated cost of 12 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