[AVX] vector shifts are scalarized

[llvmbot]

Bugzilla Link	10581
Resolution	FIXED
Resolved on	Aug 08, 2011 16:31
Version	trunk
OS	All
Reporter	LLVM Bugzilla Contributor
CC	@bcardosolopes

Extended Description

Given:

define <8 x i32> @​foo___ff(<8 x i32> %a, <8 x i32> %b) nounwind readnone {
%s = shl <8 x i32> %a, <i32 2, i32 2, i32 2, i32 2, i32 2, i32 2, i32 2, i32 2>
ret <8 x i32> %s
}

llc -mattr=+avx deconstructs the vector into elements, shifts each one individually, and reconstructs the vector. It'd be nice if it just used the corresponding AVX vector shift instruction for this.

[bcardosolopes]

Hi,

Fixed in r137067