[PATCH] D28810: [X86][SSE] Use MOVMSK for all_of/any_of reduction patterns

[Andrea Di Biagio via Phabricator via llvm-commits]

andreadb added inline comments.


================
Comment at: test/CodeGen/X86/vector-compare-all_of.ll:739-746
+; SSE-NEXT:    pmovmskb %xmm0, %eax
+; SSE-NEXT:    cmpl $65535, %eax # imm = 0xFFFF
+; SSE-NEXT:    movb $-1, %al
+; SSE-NEXT:    je .LBB14_2
+; SSE-NEXT:  # BB#1:
+; SSE-NEXT:    xorl %eax, %eax
+; SSE-NEXT:  .LBB14_2:
----------------
RKSimon wrote:
> andreadb wrote:
> > I was expecting to see the usual `cmpl+movw+cmov` pattern here.
> > Is it because you would end up with a conditional move of i8 quantities?
> Hmm, it might be possible to lower this as:
> ```
>         pmovmskb %xmm0, %edi
>         xor %eax, %eax
>         cmpl $65535, %edi
>         sete %al
>         negl %eax
>         retq
> ```
I was about to suggest cmp+set+sra, but neg is even better :-).



Repository:
  rL LLVM

https://reviews.llvm.org/D28810