https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92080
Bug ID: 92080
Summary: Missed CSE of _mm512_set1_epi8(c) with
_mm256_set1_epi8(c)
Product: gcc
Version: 10.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: peter at cordes dot ca
Target Milestone: ---
Target: x86_64-*-*, i?86-*-*
As a workaround for PR 82887 some code (e.g. a memset) uses
__m512i zmm = _mm512_set1_epi8((char)c);
__m256i ymm = _mm256_set1_epi8((char)c);
instead of
ymm = _mm512_castsi512_si256(zmm);
(found in the persistent-memory library
https://github.com/pmem/pmdk/blob/a6031710f7c102c6b8b6b19dc9708a3b7d43e87b/src/libpmem/x86_64/memset/memset_nt_avx512f.h#L193
)
Obviously we'd like to CSE that instead of actually broadcasting twice. MVCE:
#include <immintrin.h>
__m512i sinkz;
__m256i sinky;
void foo(char c) {
sinkz = _mm512_set1_epi8(c);
sinky = _mm256_set1_epi8(c);
}
https://godbolt.org/z/CeXhi8 g++ (Compiler-Explorer-Build) 10.0.0 20191012
# g++ -O3 -march=skylake-avx512 (AVX512BW + AVX512VL are the relevant ones)
foo(char):
vpbroadcastb %edi, %zmm0
vmovdqa64 %zmm0, sinkz(%rip)
vpbroadcastb %edi, %ymm0 # wasted insn
vmovdqa64 %ymm0, sinky(%rip) # wasted EVEX prefix
vzeroupper
ret
Without AVX512VL it wastes even more instructions (vmovd + AVX2 vpbroadcastb
xmm,ymm), even though AVX512BW vpbroadcastb zmm does set the YMM register.
(There are no CPUs with AVX512BW but not AVX512VL; if people compile that way
it's their own fault. But this might be relevant for set1_epi32() on KNL).
Clang finds this optimization, and uses a shorter vmovdqa for the YMM store
saving another 2 bytes of code size:
vpbroadcastb %edi, %zmm0
vmovdqa64 %zmm0, sinkz(%rip)
vmovdqa %ymm0, sinky(%rip)
vzeroupper
ret
