Don't lower splat vector load to relative to the esp if the

stack may be misaligned.

Update test accordingly.

Patch by Evan Cheng!


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@94291 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 4 insertions, 37 deletions

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index e3bd694ca1..11e07dfec3 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -3390,17 +3390,10 @@ X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
     MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
     if (DAG.InferPtrAlignment(Ptr) < 16) {
       if (MFI->isFixedObjectIndex(FI)) {
-        // Can't change the alignment. Reference stack + offset explicitly
-        // if stack pointer is at least 16-byte aligned.
-        unsigned StackAlign = Subtarget->getStackAlignment();
-        if (StackAlign < 16)
-          return SDValue();
-        Offset = MFI->getObjectOffset(FI) + Offset;
-        SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
-                                              getPointerTy());
-        Ptr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
-                          DAG.getConstant(Offset & ~15, getPointerTy()));
-        Offset %= 16;
+        // Can't change the alignment. FIXME: It's possible to compute
+        // the exact stack offset and reference FI + adjust offset instead.
+        // If someone *really* cares about this. That's the way to implement it.
+        return SDValue();
       } else {
         MFI->setObjectAlignment(FI, 16);
       }
diff --git a/test/CodeGen/X86/splat-scalar-load.ll b/test/CodeGen/X86/splat-scalar-load.ll
index 32d3ab661d..2b13029896 100644
--- a/test/CodeGen/X86/splat-scalar-load.ll
+++ b/test/CodeGen/X86/splat-scalar-load.ll
@@ -1,21 +1,6 @@
 ; RUN: llc < %s -mtriple=i386-apple-darwin -mattr=+sse2 | FileCheck %s
 ; rdar://7434544
 
-define <2 x i64> @t1() nounwind ssp {
-entry:
-; CHECK: t1:
-; CHECK: pshufd	$0, (%esp), %xmm0
-  %array = alloca [8 x float], align 16
-  %arrayidx = getelementptr inbounds [8 x float]* %array, i32 0, i32 0
-  %tmp2 = load float* %arrayidx
-  %vecinit = insertelement <4 x float> undef, float %tmp2, i32 0
-  %vecinit5 = insertelement <4 x float> %vecinit, float %tmp2, i32 1
-  %vecinit7 = insertelement <4 x float> %vecinit5, float %tmp2, i32 2
-  %vecinit9 = insertelement <4 x float> %vecinit7, float %tmp2, i32 3
-  %0 = bitcast <4 x float> %vecinit9 to <2 x i64>
-  ret <2 x i64> %0
-}
-
 define <2 x i64> @t2() nounwind ssp {
 entry:
 ; CHECK: t2:
@@ -30,14 +15,3 @@ entry:
   %0 = bitcast <4 x float> %vecinit9 to <2 x i64>
   ret <2 x i64> %0
 }
-
-define <4 x float> @t3(float %tmp1, float %tmp2, float %tmp3) nounwind readnone ssp {
-entry:
-; CHECK: t3:
-; CHECK: pshufd	$-86, (%esp), %xmm0
-  %0 = insertelement <4 x float> undef, float %tmp3, i32 0
-  %1 = insertelement <4 x float> %0, float %tmp3, i32 1
-  %2 = insertelement <4 x float> %1, float %tmp3, i32 2
-  %3 = insertelement <4 x float> %2, float %tmp3, i32 3
-  ret <4 x float> %3
-}