Horrible x86 codegen for _mm_loadl_epi64

[efriedma-quic]

Bugzilla Link	2108
Resolution	FIXED
Resolved on	Apr 10, 2008 00:38
Version	unspecified
OS	Linux
CC	@asl

Extended Description

Testcase:
#include <xmmintrin.h>
__m128i doload64(unsigned long long x) { return _mm_loadl_epi64(&x);}

Generated il:
define <2 x i64> @​doload64(i64 %x) nounwind {
entry:
%tmp717 = bitcast i64 %x to double
%tmp8 = insertelement <2 x double> undef, double %tmp717, i32 0
%tmp9 = insertelement <2 x double> %tmp8, double 0.000000e+00, i32 1
%tmp11 = bitcast <2 x double> %tmp9 to <2 x i64>
ret <2 x i64> %tmp11
}

On x86, the this codegens to the following:

doload64:
subl $12, %esp
movl 20(%esp), %eax
movl %eax, 4(%esp)
movl 16(%esp), %eax
movl %eax, (%esp)
movsd (%esp), %xmm0
addl $12, %esp
ret

which is 6 instructions longer than it needs to be. The code generator should really be smart enough to load from the original stack slot.

[lattner]

Yep, we generate poor code for this. In x86-32 we should do a load from the stack slot. In x86-64 we get:

_doload64:
Leh_func_begin1:
Llabel1:
movd %rdi, %xmm1
xorps %xmm0, %xmm0
movsd %xmm1, %xmm0
ret

which could be a single "movd rdi -> xmm0".

This sounds like generic vector improvements.

-Chris

[lattner]

The x86-64 side of this is fixed:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20080407/060904.html

We now emit:

_doload64:
movd %rdi, %xmm0
ret

[lattner]

On the 32-bit side, amusingly gcc generates almost identical code:

_doload64:
subl $28, %esp
movl 32(%esp), %eax
movl %eax, 8(%esp)
movl 36(%esp), %eax
movl %eax, 12(%esp)
movq 8(%esp), %xmm0
addl $28, %esp
ret

GCC does use more stack space though for mysterious reasons.

[lattner]

This x86-32 case is somewhat tricky. By the time dagcombine1 runs, the argument load has already been lowered into 2 x i32 loads. dagcombine would have to notice that the two i32's are fusable and the whole thing can be trned into an f64 load by changing the shuffle.

[lattner]

Here is a reduced testcase:

define fastcc double @​doload64(i64 %x) nounwind {
entry:
%tmp717 = bitcast i64 %x to double ; [#uses=1]
ret double %tmp717
}

compiles to:

_doload64:
subl $12, %esp
movl 20(%esp), %eax
movl %eax, 4(%esp)
movl 16(%esp), %eax
movl %eax, (%esp)
movsd (%esp), %xmm0
addl $12, %esp
ret

I moved this to the X86 readme for the x86-32 piece. The x86-64 piece is fixed:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20080407/060905.html

-Chris