insertelement of undef generates horrible code (e.g. _mm_sll_epi16)

[isanbard]

Bugzilla Link	1286
Resolution	FIXED
Resolved on	Feb 22, 2010 12:48
Version	trunk
OS	MacOS X

Extended Description

This:

#include <xmmintrin.h>

void foo(__m128i *A, __m128i *B) {
*A = _mm_sll_epi16 (*A, *B);
}

generates this code:

_foo:
movl 8(%esp), %eax
movdqa (%eax), %xmm0
#IMPLICIT_DEF %eax
pinsrw $2, %eax, %xmm0
xorl %ecx, %ecx
pinsrw $3, %ecx, %xmm0
pinsrw $4, %eax, %xmm0
pinsrw $5, %ecx, %xmm0
pinsrw $6, %eax, %xmm0
pinsrw $7, %ecx, %xmm0
movl 4(%esp), %eax
movdqa (%eax), %xmm1
psllw %xmm0, %xmm1
movdqa %xmm1, (%eax)
ret

This is obviously bad! We should be as good GCC at least.

[isanbard]

assigned to @lattner

[lattner]

Fixed, patch here:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20070402/047039.html

Testcase here: Transforms/InstCombine/vec_insertelt.ll

We now compile this to:

_foo:
movl 4(%esp), %eax
movdqa (%eax), %xmm0
movl 8(%esp), %ecx
psllw (%ecx), %xmm0
movdqa %xmm0, (%eax)
ret

GCC manages:

_foo:
subl $12, %esp
movl 16(%esp), %edx
movl 20(%esp), %eax
movdqa (%eax), %xmm0
movdqa (%edx), %xmm1
psllw %xmm0, %xmm1
movdqa %xmm1, (%edx)
addl $12, %esp
ret

-Chris

[nunoplopes]

mentioned in issue llvm/llvm-bugzilla-archive#45481