https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82731
Bug ID: 82731
Summary: _mm256_set_epi8(array[offset[0]], array[offset[1]],
...) byte gather makes slow code, trying to
zero-extend all the uint16_t offsets first and
spilling them.
Product: gcc
Version: 8.0
Status: UNCONFIRMED
Keywords: missed-optimization, ssemmx
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: peter at cordes dot ca
Target Milestone: ---
Target: x86_64-*-*, i?86-*-*
#include "immintrin.h"
#include "inttypes.h"
__m256i gather(char *array, uint16_t *offset) {
return _mm256_set_epi8(array[offset[0]], array[offset[1]], array[offset[2]],
array[offset[3]], array[offset[4]], array[offset[5]], array[offset[6]],
array[offset[7]],
array[offset[8]],array[offset[9]],array[offset[10]],array[offset[11]],
array[offset[12]], array[offset[13]], array[offset[14]], array[offset[15]],
array[offset[16]],array[offset[17]], array[offset[18]],
array[offset[19]], array[offset[20]], array[offset[21]], array[offset[22]],
array[offset[23]],
array[offset[24]],array[offset[25]],array[offset[26]], array[offset[27]],
array[offset[28]], array[offset[29]], array[offset[30]],array[offset[31]]);
}
https://stackoverflow.com/questions/46881656/avx2-byte-gather-with-uint16-indices-into-a-m256i
https://godbolt.org/g/LEVVwt
pushq %rbp
movq %rsp, %rbp
pushq %r15
pushq %r14
pushq %r13
pushq %r12
pushq %rbx
andq $-32, %rsp
subq $40, %rsp
movzwl 40(%rsi), %eax
... # more movzwl
movq %rax, 32(%rsp) # spill
movzwl 38(%rsi), %eax # and reuse
... # more movzwl
movzwl 46(%rsi), %r8d
movq %rax, 24(%rsp) # spill
movzwl 36(%rsi), %eax
movzwl 42(%rsi), %edx
movq %rax, 16(%rsp)
movzwl 34(%rsi), %eax
...
...
vpinsrb $1, (%rdi,%r9), %xmm6, %xmm6
vpinsrb $1, (%rdi,%rcx), %xmm5, %xmm5
movq 24(%rsp), %rcx # more reloading
vpunpcklwd %xmm6, %xmm3, %xmm3
movzbl (%rdi,%rcx), %edx # and using as a gather index
movq 8(%rsp), %rcx
vpunpcklwd %xmm5, %xmm1, %xmm1
vpunpckldq %xmm3, %xmm2, %xmm2
vmovd %edx, %xmm0
movzbl (%rdi,%rcx), %edx
vpinsrb $1, (%rdi,%rbx), %xmm0, %xmm0
I think gcc is missing the point of vpinsrb, and making too many separate dep
chains which it then has to shuffle together. It doesn't have such good
throughput on any CPUs that you need more than 2 or 3 dep chains to max out its
1 or 2 per clock throughput.
But the main point here is doing all the zero-extension of offset[0..31] before
doing *any* of the loads from array[], running out of registers and spilling.
See also discussion on that SO question about byte gathers and possibilities
for VPGATHERDD being maybe worth it on Skylake.
