SLP Vectorizer too aggressive with simple stores

[MatzeB]

Bugzilla Link	31116
Version	trunk
OS	All
CC	@hfinkel,@RKSimon

Extended Description

In a situation like the following (modeled after what happens in a C++ constructor with two base classes), the SLP vectorizer seems too aggressive:

struct C {
  char *BaseA;
  char *BaseB;
};
extern char VTab[];
void initC(C *c) {
  c->BaseA = VTab + 48;
  c->BaseB = VTab + 16;
}

The SLP Vectorizer transforms:

  %BaseA = getelementptr inbounds %struct.C, %struct.C* %c, i64 0, i32 0
  store i8* getelementptr inbounds ([0 x i8], [0 x i8]* @VTab, i64 0, i64 48), i8** %BaseA, align 8, !tbaa !2
  %BaseB = getelementptr inbounds %struct.C, %struct.C* %c, i64 0, i32 1
  store i8* getelementptr inbounds ([0 x i8], [0 x i8]* @VTab, i64 0, i64 16), i8** %BaseB, align 8, !tbaa !7

into:

  %BaseA = getelementptr inbounds %struct.C, %struct.C* %c, i64 0, i32 0
  %BaseB = getelementptr inbounds %struct.C, %struct.C* %c, i64 0, i32 1
  %0 = bitcast i8** %BaseA to <2 x i8*>*
  store <2 x i8*> <i8* getelementptr inbounds ([0 x i8], [0 x i8]* @VTab, i64 0, i64 48), i8* getelementptr inbounds ([0 x i8], [0 x i8]* @VTab, i64 0, i64 16)>, <2 x i8*>* %0, align 8, !tbaa !2
  ret void

which results in this code on X86:

 	movq	_VTab@GOTPCREL(%rip), %rax
   leaq	16(%rax), %rcx
   movd	%rcx, %xmm0
   addq	$48, %rax
   movd	%rax, %xmm1
   punpcklqdq	%xmm0, %xmm1    ## xmm1 = xmm1[0],xmm0[0]
   movdqu	%xmm1, (%rdi)

while it can be this if the SLP Vectorizer doesn't do anything:

	movq	_VTab@GOTPCREL(%rip), %rax
	leaq	48(%rax), %rcx
	movq	%rcx, (%rdi)
	addq	$16, %rax
	movq	%rax, 8(%rdi)

[alexey-bataev]

Cannot be reproduced in trunk, fixed