[llvm-commits] [llvm] r166971 - in /llvm/trunk: lib/Target/X86/X86ISelLowering.cpp test/CodeGen/X86/vec_shuffle-30.ll
Jakub Staszak
kubastaszak at gmail.com
Mon Oct 29 14:56:35 PDT 2012
Author: kuba
Date: Mon Oct 29 16:56:35 2012
New Revision: 166971
URL: http://llvm.org/viewvc/llvm-project?rev=166971&view=rev
Log:
Allow to fold vector load if there is more than one bitcast, so in the case:
%0 = load <8 x i16>* %dest
%1 = shufflevector <8 x i16> %0, <8 x i16> %in,
<8 x i32> < i32 0, i32 1, i32 2, i32 3, i32 13, i32 undef, i32 14, i32 14>
store <8 x i16> %1, <8 x i16>* %dest
We get:
vmovlpd (%eax), %xmm0, %xmm0
instead of:
vmovaps (%eax), %xmm1
vmovsd %xmm1, %xmm0, %xmm0
No extra test-case is added. I just fixed the existing one
(also it uses FileCheck now).
Modified:
llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
llvm/trunk/test/CodeGen/X86/vec_shuffle-30.ll
Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=166971&r1=166970&r2=166971&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Mon Oct 29 16:56:35 2012
@@ -6438,17 +6438,17 @@
}
static bool MayFoldVectorLoad(SDValue V) {
- if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
+ while (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
V = V.getOperand(0);
+
if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
V = V.getOperand(0);
if (V.hasOneUse() && V.getOpcode() == ISD::BUILD_VECTOR &&
V.getNumOperands() == 2 && V.getOperand(1).getOpcode() == ISD::UNDEF)
// BUILD_VECTOR (load), undef
V = V.getOperand(0);
- if (MayFoldLoad(V))
- return true;
- return false;
+
+ return MayFoldLoad(V);
}
// FIXME: the version above should always be used. Since there's
Modified: llvm/trunk/test/CodeGen/X86/vec_shuffle-30.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/vec_shuffle-30.ll?rev=166971&r1=166970&r2=166971&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/vec_shuffle-30.ll (original)
+++ llvm/trunk/test/CodeGen/X86/vec_shuffle-30.ll Mon Oct 29 16:56:35 2012
@@ -1,21 +1,25 @@
-; RUN: llc < %s -march=x86 -mattr=sse41 -o %t
-; RUN: grep pshufhw %t | grep -- -95 | count 1
-; RUN: grep shufps %t | count 1
-; RUN: not grep pslldq %t
+; RUN: llc < %s -march=x86 -mattr=sse41 | FileCheck %s
+; CHECK: test
; Test case when creating pshufhw, we incorrectly set the higher order bit
; for an undef,
define void @test(<8 x i16>* %dest, <8 x i16> %in) nounwind {
entry:
+; CHECK-NOT: vmovaps
+; CHECK: vmovlpd
+; CHECK: vpshufhw $-95
%0 = load <8 x i16>* %dest
%1 = shufflevector <8 x i16> %0, <8 x i16> %in, <8 x i32> < i32 0, i32 1, i32 2, i32 3, i32 13, i32 undef, i32 14, i32 14>
store <8 x i16> %1, <8 x i16>* %dest
ret void
-}
+}
+; CHECK: test2
; A test case where we shouldn't generate a punpckldq but a pshufd and a pslldq
define void @test2(<4 x i32>* %dest, <4 x i32> %in) nounwind {
entry:
+; CHECK-NOT: pslldq
+; CHECK: shufps
%0 = shufflevector <4 x i32> %in, <4 x i32> <i32 0, i32 0, i32 0, i32 0>, <4 x i32> < i32 undef, i32 5, i32 undef, i32 2>
store <4 x i32> %0, <4 x i32>* %dest
ret void
More information about the llvm-commits
mailing list