Unaligned loads should use the VMOVUPS opcode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@177130 91177308-0d34-0410-b5e6-96231b3b80d8

Author: nadavrot

lib/Target/X86/X86InstrSSE.td

@@ -1009,7 +1009,7 @@ let Predicates = [HasAVX] in {
             (VMOVUPSmr addr:$dst, (v4i32 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
   def : Pat<(store (v8i16 (extract_subvector
                            (v16i16 VR256:$src), (iPTR 0))), addr:$dst),
-            (VMOVAPSmr addr:$dst, (v8i16 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
+            (VMOVUPSmr addr:$dst, (v8i16 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
   def : Pat<(store (v16i8 (extract_subvector
                            (v32i8 VR256:$src), (iPTR 0))), addr:$dst),
             (VMOVUPSmr addr:$dst, (v16i8 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;

test/CodeGen/X86/avx-load-store.ll

@@ -55,7 +55,7 @@ define void @storev16i16(<16 x i16> %a) nounwind {
 
 ; CHECK: storev16i16_01
 ; CHECK: vextractf128
-; CHECK: vmovaps  %xmm
+; CHECK: vmovups  %xmm
 define void @storev16i16_01(<16 x i16> %a) nounwind {
   store <16 x i16> %a, <16 x i16>* undef, align 4
   unreachable

test/CodeGen/X86/vec_align_i256.ll

@@ -0,0 +1,11 @@
+; RUN: llc < %s -mcpu=corei7-avx | FileCheck %s 
+
+; Make sure that we are not generating a movaps because the vector is aligned to 1.
+;CHECK: @foo
+;CHECK: xor
+;CHECK-NEXT: vmovups
+;CHECK-NEXT: ret
+define void @foo() {
+  store <16 x i16> zeroinitializer, <16 x i16>* undef, align 1
+  ret void
+}