[llvm-commits] [llvm] r51472 - in /llvm/trunk: lib/Transforms/Scalar/InstructionCombining.cpp test/Transforms/InstCombine/2008-05-22-FoldAddIVec.ll

[Chris Lattner]

On May 22, 2008, at 9:25 PM, Chris Lattner wrote:

>
> On May 22, 2008, at 9:14 PM, Nick Lewycky wrote:
>
>> Author: nicholas
>> Date: Thu May 22 23:14:51 2008
>> New Revision: 51472
>>
>> URL: http://llvm.org/viewvc/llvm-project?rev=51472&view=rev
>> Log:
>> Implement X + X for vectors.
>
> Is the code generator turning this back into x+x?  If not, this is a
> major pessimization.  What does:
>
> define <4 x i32> @f(<4 x i32> %i) nounwind {
>   %A = add <4 x i32> %i, %i
>   ret <4 x i32> %A
> }
>
> compile into?  Without your patch I get:
>
> _f:
> 	paddd	%xmm0, %xmm0
> 	ret

Please revert.  Your patch optimizes this into:

LCPI1_0:					##  <4 x i32>
	.long	2
	.long	2
	.long	2
	.long	2
	.text
	.align	4,0x90
	.globl	_f
_f:
	pshufd	$3, %xmm0, %xmm1
	movd	%xmm1, %eax
	imull	LCPI1_0+12, %eax
	movd	%eax, %xmm1
	pshufd	$1, %xmm0, %xmm2
	movd	%xmm2, %eax
	imull	LCPI1_0+4, %eax
	movd	%eax, %xmm2
	punpckldq	%xmm1, %xmm2
	movd	%xmm0, %eax
	imull	LCPI1_0, %eax
	movd	%eax, %xmm1
	movhlps	%xmm0, %xmm0
	movd	%xmm0, %eax
	imull	LCPI1_0+8, %eax
	movd	%eax, %xmm0
	punpckldq	%xmm0, %xmm1
	movaps	%xmm1, %xmm0
	punpckldq	%xmm2, %xmm0
	ret