https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101096
--- Comment #2 from Hongtao.liu <crazylht at gmail dot com> ---
For foo, after supporting downconvert instruction, below it's difference for
codegen.
@@ -6,15 +6,17 @@
foo:
.LFB0:
.cfi_startproc
- movl $255, %eax
- vpbroadcastw %eax, %xmm0
- vpand 16(%rsi), %xmm0, %xmm2
- vpand (%rsi), %xmm0, %xmm1
- vpackuswb %xmm2, %xmm1, %xmm1
- vpand (%rdi), %xmm0, %xmm2
- vpand 16(%rdi), %xmm0, %xmm0
- vpackuswb %xmm0, %xmm2, %xmm0
- vpaddb %xmm0, %xmm1, %xmm0
+ vmovdqu16 (%rsi), %xmm1
+ vmovdqu16 16(%rsi), %xmm0
+ vmovdqu16 16(%rdi), %xmm2
+ vpmovwb %xmm0, %xmm0
+ vpmovwb %xmm1, %xmm1
+ vpunpcklqdq %xmm0, %xmm1, %xmm1
+ vmovdqu16 (%rdi), %xmm0
+ vpmovwb %xmm2, %xmm2
+ vpmovwb %xmm0, %xmm0
+ vpunpcklqdq %xmm2, %xmm0, %xmm0
+ vpaddb %xmm1, %xmm0, %xmm0
vmovdqu8 %xmm0, (%rdx)
If GCC vectorizer support different vector length(then we don't need to down
convert and pack), vpmovwb may be better, but if not, the instructions number
seems more or less, but vpmovwb is more expensive than vpand.
