Skip to content

Compiler no longer optimizing some horizontal add/sub patterns after r310260 #34072

@dyung

Description

@dyung
Bugzilla Link 34724
Version trunk
OS Linux
CC @alexey-bataev,@efriedma-quic,@hfinkel,@RKSimon,@rotateright,@ZviRackover

Extended Description

Prior to r310260, the compiler was optimizing some add/sub + shuffle patterns to horizontal add/sub instructions, but no longer seems to do so.

Consider this example:

__attribute__((noinline))
__m128 add_ps_001(__m128 a, __m128 b) {
  __m128 r = (__m128){ a[0] + a[1], a[2] + a[3], b[0] + b[1], b[2] + b[3] };
  return __builtin_shufflevector(r, a, -1, 1, 2, 3);
}

When compiled with the options "-S -O2 -march=bdver2" using a compiler prior to r310260, the compiler would generate the following assembly:

vmovaps (%rcx), %xmm0
vhaddps (%rdx), %xmm0, %xmm0
retq

After r310260, the compiler is now generating the following code for the same function:

vmovaps (%rdx), %xmm0
vmovddup        8(%rcx), %xmm1  # xmm1 = mem[0,0]
vinsertps       $28, %xmm0, %xmm1, %xmm2 # xmm2 = xmm1[0],xmm0[0],zero,zero
vinsertps       $76, %xmm1, %xmm0, %xmm1 # xmm1 = xmm1[1],xmm0[1],zero,zero
vaddps  %xmm1, %xmm2, %xmm1
vpermilpd       $1, %xmm0, %xmm2 # xmm2 = xmm0[1,0]
vpermilps       $231, %xmm0, %xmm0 # xmm0 = xmm0[3,1,2,3]
vaddss  %xmm0, %xmm2, %xmm0
vpermilps       $208, %xmm1, %xmm1 # xmm1 = xmm1[0,0,1,3]
vinsertps       $48, %xmm0, %xmm1, %xmm0 # xmm0 = xmm1[0,1,2],xmm0[0]
retq

Metadata

Metadata

Assignees

Labels

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions