projects
/
oota-llvm.git
/ blobdiff
commit
grep
author
committer
pickaxe
?
search:
re
summary
|
shortlog
|
log
|
commit
|
commitdiff
|
tree
raw
|
inline
| side by side
Fix merges of non-zero vector stores
[oota-llvm.git]
/
test
/
CodeGen
/
X86
/
sse-align-12.ll
diff --git
a/test/CodeGen/X86/sse-align-12.ll
b/test/CodeGen/X86/sse-align-12.ll
index 2351fd6fa77bd65ac98e31b4b51ff5247f12eb44..9441cc0002fbe86382aa9e75947058d35e0f0948 100644
(file)
--- a/
test/CodeGen/X86/sse-align-12.ll
+++ b/
test/CodeGen/X86/sse-align-12.ll
@@
-1,10
+1,12
@@
-; RUN: llc < %s -march=x86-64 -mcpu=nehalem | FileCheck %s
+; RUN: llc < %s -m
triple=x86_64-unknown-unknown -m
arch=x86-64 -mcpu=nehalem | FileCheck %s
-; CHECK-LABEL: a:
-; CHECK: movdqu
-; CHECK: pshufd
define <4 x float> @a(<4 x float>* %y) nounwind {
define <4 x float> @a(<4 x float>* %y) nounwind {
- %x = load <4 x float>* %y, align 4
+; CHECK-LABEL: a:
+; CHECK: # BB#0:
+; CHECK-NEXT: movups (%rdi), %xmm0
+; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,2,1,0]
+; CHECK-NEXT: retq
+ %x = load <4 x float>, <4 x float>* %y, align 4
%a = extractelement <4 x float> %x, i32 0
%b = extractelement <4 x float> %x, i32 1
%c = extractelement <4 x float> %x, i32 2
%a = extractelement <4 x float> %x, i32 0
%b = extractelement <4 x float> %x, i32 1
%c = extractelement <4 x float> %x, i32 2
@@
-16,11
+18,13
@@
define <4 x float> @a(<4 x float>* %y) nounwind {
ret <4 x float> %s
}
ret <4 x float> %s
}
-; CHECK-LABEL: b:
-; CHECK: movups
-; CHECK: unpckhps
define <4 x float> @b(<4 x float>* %y, <4 x float> %z) nounwind {
define <4 x float> @b(<4 x float>* %y, <4 x float> %z) nounwind {
- %x = load <4 x float>* %y, align 4
+; CHECK-LABEL: b:
+; CHECK: # BB#0:
+; CHECK-NEXT: movups (%rdi), %xmm1
+; CHECK-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; CHECK-NEXT: retq
+ %x = load <4 x float>, <4 x float>* %y, align 4
%a = extractelement <4 x float> %x, i32 2
%b = extractelement <4 x float> %x, i32 3
%c = extractelement <4 x float> %z, i32 2
%a = extractelement <4 x float> %x, i32 2
%b = extractelement <4 x float> %x, i32 3
%c = extractelement <4 x float> %z, i32 2
@@
-32,11
+36,13
@@
define <4 x float> @b(<4 x float>* %y, <4 x float> %z) nounwind {
ret <4 x float> %s
}
ret <4 x float> %s
}
-; CHECK-LABEL: c:
-; CHECK: movupd
-; CHECK: shufpd
define <2 x double> @c(<2 x double>* %y) nounwind {
define <2 x double> @c(<2 x double>* %y) nounwind {
- %x = load <2 x double>* %y, align 8
+; CHECK-LABEL: c:
+; CHECK: # BB#0:
+; CHECK-NEXT: movupd (%rdi), %xmm0
+; CHECK-NEXT: shufpd {{.*#+}} xmm0 = xmm0[1,0]
+; CHECK-NEXT: retq
+ %x = load <2 x double>, <2 x double>* %y, align 8
%a = extractelement <2 x double> %x, i32 0
%c = extractelement <2 x double> %x, i32 1
%p = insertelement <2 x double> undef, double %c, i32 0
%a = extractelement <2 x double> %x, i32 0
%c = extractelement <2 x double> %x, i32 1
%p = insertelement <2 x double> undef, double %c, i32 0
@@
-44,11
+50,13
@@
define <2 x double> @c(<2 x double>* %y) nounwind {
ret <2 x double> %r
}
ret <2 x double> %r
}
-; CHECK-LABEL: d:
-; CHECK: movupd
-; CHECK: unpckhpd
define <2 x double> @d(<2 x double>* %y, <2 x double> %z) nounwind {
define <2 x double> @d(<2 x double>* %y, <2 x double> %z) nounwind {
- %x = load <2 x double>* %y, align 8
+; CHECK-LABEL: d:
+; CHECK: # BB#0:
+; CHECK-NEXT: movupd (%rdi), %xmm1
+; CHECK-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
+; CHECK-NEXT: retq
+ %x = load <2 x double>, <2 x double>* %y, align 8
%a = extractelement <2 x double> %x, i32 1
%c = extractelement <2 x double> %z, i32 1
%p = insertelement <2 x double> undef, double %c, i32 0
%a = extractelement <2 x double> %x, i32 1
%c = extractelement <2 x double> %z, i32 1
%p = insertelement <2 x double> undef, double %c, i32 0