[PATCH] D38671: [X86][AVX512] lowering shuffle i/f intrinsic - llvm part

[Craig Topper via Phabricator via llvm-commits]

craig.topper added a comment.

For the 64-bit case its probably still better to do the conversion from combineSelect similar to what we do for the bitcasted case. Its somewhat unusual to walk the users in DAG combine.

Though it might make sense to turn all X86ISD::VPERM2X128 that we can into SHUF128 regardless of it being masked. Assuming they have the same latency and port bindings. This would open up the XMM16-XMM31 register space for them.


https://reviews.llvm.org/D38671