[PATCH] D54276: [SelectionDAG][AArch64][X86] Move legalization of vector MULHS/MULHU from LegalizeDAG to LegalizeVectorOps

[Craig Topper via Phabricator via llvm-commits]

craig.topper added inline comments.


================
Comment at: test/CodeGen/X86/combine-udiv.ll:726
+; XOP-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; XOP-NEXT:    vpperm {{.*#+}} xmm1 = xmm1[1,3,5,7,9,11,13,15],xmm2[1,3,5,7,9,11,13,15]
 ; XOP-NEXT:    movl $249, %eax
----------------
RKSimon wrote:
> It's a little odd that this hasn't been simplified to a VPSHUFB (or at least removed the xmm2 zero register - VPPERM can generates its own zero elements) - at that point this wouldn't be XOP specific at all......
It looks like a packus was turned in vpperm before some zero_extend_vector_inreg and mul constant folding kicked in later in the same dag combine round. So at the time the VPPERM was created it wasn't known that it was zero. This all occurred on the DAG combine between vector op legalization and DAG legalization. This vpperm itself didn't get revisited after the input became 0 due to a bitcast. But the last DAG combine after legalize DAG should have revisited it. Should we have been able to recombine the VPPERM to a VPPERM with zero controls?


Repository:
  rL LLVM

https://reviews.llvm.org/D54276