On Sandybridge loading unaligned 256bits using two XMM loads (vmovups and vinsertf128) is faster than using a single vmovups instruction.
llvm-svn: 172868
diff --git a/llvm/test/CodeGen/X86/sandybridge-loads.ll b/llvm/test/CodeGen/X86/sandybridge-loads.ll
new file mode 100644
index 0000000..d85c32e
--- /dev/null
+++ b/llvm/test/CodeGen/X86/sandybridge-loads.ll
@@ -0,0 +1,21 @@
+; RUN: llc -march=x86-64 -mtriple=x86_64-apple-darwin -mcpu=corei7-avx -o - < %s | FileCheck %s
+
+;CHECK: wideloads
+;CHECK: vmovaps
+;CHECK: vinsertf128
+;CHECK: vmovups
+;CHECK-NOT: vinsertf128
+;CHECK: ret
+
+define void @wideloads(<8 x float>* %a, <8 x float>* %b, <8 x float>* %c) nounwind uwtable noinline ssp {
+ %v0 = load <8 x float>* %a, align 16 ; <---- unaligned!
+ %v1 = load <8 x float>* %b, align 32 ; <---- aligned!
+ %m0 = fcmp olt <8 x float> %v1, %v0
+ %v2 = load <8 x float>* %c, align 16
+ %m1 = fcmp olt <8 x float> %v2, %v0
+ %mand = and <8 x i1> %m1, %m0
+ %r = zext <8 x i1> %mand to <8 x i32>
+ store <8 x i32> %r, <8 x i32>* undef, align 16
+ ret void
+}
+
diff --git a/llvm/test/CodeGen/X86/v8i1-masks.ll b/llvm/test/CodeGen/X86/v8i1-masks.ll
index abb4b39..ea231af 100644
--- a/llvm/test/CodeGen/X86/v8i1-masks.ll
+++ b/llvm/test/CodeGen/X86/v8i1-masks.ll
@@ -1,7 +1,7 @@
; RUN: llc -march=x86-64 -mtriple=x86_64-apple-darwin -mcpu=corei7-avx -o - < %s | FileCheck %s
;CHECK: and_masks
-;CHECK: vmovups
+;CHECK: vmovaps
;CHECK: vcmpltp
;CHECK: vcmpltp
;CHECK: vandps