Optimize the sequence blend(sign_extend(x)) to blend(shl(x)) since SSE blend instructions only look at the highest bit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@147426 91177308-0d34-0410-b5e6-96231b3b80d8

4 files changed, 31 insertions, 10 deletions

diff --git a/test/CodeGen/X86/2011-12-28-vselecti8.ll b/test/CodeGen/X86/2011-12-28-vselecti8.ll
index dbc122a..fc1b83b 100644
--- a/test/CodeGen/X86/2011-12-28-vselecti8.ll
+++ b/test/CodeGen/X86/2011-12-28-vselecti8.ll
@@ -5,8 +5,10 @@ target triple = "x86_64-apple-darwin11.2.0"
 
 ; CHECK: @foo8
 ; CHECK: psll
-; CHECK: psraw
-; CHECK: pblendvb
+; CHECK-NOT: sra
+; CHECK: pandn
+; CHECK: pand
+; CHECK: or
 ; CHECK: ret
 define void @foo8(float* nocapture %RET) nounwind {
 allocas:
diff --git a/test/CodeGen/X86/sext-blend.ll b/test/CodeGen/X86/sext-blend.ll
new file mode 100644
index 0000000..b1f9573
--- /dev/null
+++ b/test/CodeGen/X86/sext-blend.ll
@@ -0,0 +1,15 @@
+; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=corei7 -promote-elements -mattr=+sse41 | FileCheck %s
+
+; CHECK: foo
+define <4 x double> @foo(<4 x double> %x, <4 x double> %y) {
+  ; CHECK: cmpnlepd
+  ; CHECK: psllq
+  ; CHECK-NEXT: blendvpd
+  ; CHECK: psllq
+  ; CHECK-NEXT: blendvpd
+  ; CHECK: ret
+  %min_is_x = fcmp ult <4 x double> %x, %y
+  %min = select <4 x i1> %min_is_x, <4 x double> %x, <4 x double> %y
+  ret <4 x double> %min
+}
+
diff --git a/test/CodeGen/X86/sse2-blend.ll b/test/CodeGen/X86/sse2-blend.ll
index 2f4317b..c6602d3 100644
--- a/test/CodeGen/X86/sse2-blend.ll
+++ b/test/CodeGen/X86/sse2-blend.ll
@@ -28,10 +28,10 @@ define void@vsel_i32(<4 x i32>* %v1, <4 x i32>* %v2) {
 
 ; Without forcing instructions, fall back to the preferred PS domain.
 ; CHECK: vsel_i64
-; CHECK: xorps
-; CHECK: andps
-; CHECK: andnps
-; CHECK: orps
+; CHECK: pxor
+; CHECK: and
+; CHECK: andn
+; CHECK: or
 ; CHECK: ret
 
 define void@vsel_i64(<4 x i64>* %v1, <4 x i64>* %v2) {
@@ -44,10 +44,10 @@ define void@vsel_i64(<4 x i64>* %v1, <4 x i64>* %v2) {
 
 ; Without forcing instructions, fall back to the preferred PS domain.
 ; CHECK: vsel_double
-; CHECK: xorps
-; CHECK: andps
-; CHECK: andnps
-; CHECK: orps
+; CHECK: xor
+; CHECK: and
+; CHECK: andn
+; CHECK: or
 ; CHECK: ret
 
 define void@vsel_double(<4 x double>* %v1, <4 x double>* %v2) {
diff --git a/test/CodeGen/X86/sse41-blend.ll b/test/CodeGen/X86/sse41-blend.ll
index 78604a0..0a71dd0 100644
--- a/test/CodeGen/X86/sse41-blend.ll
+++ b/test/CodeGen/X86/sse41-blend.ll
@@ -36,6 +36,7 @@ define <4 x i32> @vsel_i32(<4 x i32> %v1, <4 x i32> %v2) {
 
 
 ;CHECK: vsel_double
+;CHECK-NOT: sra
 ;CHECK: blendvpd
 ;CHECK: ret
 define <4 x double> @vsel_double(<4 x double> %v1, <4 x double> %v2) {
@@ -54,6 +55,7 @@ define <4 x i64> @vsel_i64(<4 x i64> %v1, <4 x i64> %v2) {
 
 
 ;CHECK: vsel_i8
+;CHECK-NOT: sra
 ;CHECK: pblendvb
 ;CHECK: ret
 define <16 x i8> @vsel_i8(<16 x i8> %v1, <16 x i8> %v2) {
@@ -65,6 +67,7 @@ define <16 x i8> @vsel_i8(<16 x i8> %v1, <16 x i8> %v2) {
 ; CHECK: A
 define <2 x double> @A(<2 x double> %x, <2 x double> %y) {
   ; CHECK: cmplepd
+  ; CHECK-NOT: sra
   ; CHECK: blendvpd
   %max_is_x = fcmp oge <2 x double> %x, %y
   %max = select <2 x i1> %max_is_x, <2 x double> %x, <2 x double> %y
@@ -74,6 +77,7 @@ define <2 x double> @A(<2 x double> %x, <2 x double> %y) {
 ; CHECK: B
 define <2 x double> @B(<2 x double> %x, <2 x double> %y) {
   ; CHECK: cmpnlepd
+  ; CHECK-NOT: sra
   ; CHECK: blendvpd
   %min_is_x = fcmp ult <2 x double> %x, %y
   %min = select <2 x i1> %min_is_x, <2 x double> %x, <2 x double> %y