D26333: [AVX-512] Add unmasked version of shift by immediate and shift by single element in XMM.

[Demikhovsky, Elena via llvm-commits]

Do you want to replace all masked intrinsics with  "unmasked + select" or only this subset?

-  Elena


-----Original Message-----
From: Craig Topper [mailto:craig.topper at gmail.com] 
Sent: Thursday, November 10, 2016 23:59
To: craig.topper at gmail.com; llvm-dev at redking.me.uk; Demikhovsky, Elena <elena.demikhovsky at intel.com>; Rackover, Zvi <zvi.rackover at intel.com>
Cc: llvm-commits at lists.llvm.org
Subject: [PATCH] D26333: [AVX-512] Add unmasked version of shift by immediate and shift by single element in XMM.

craig.topper added a comment.

We could certainly teach InstCombine to handle the masking for the existing intrinsics. I was just going for consistency and not spreading understanding of masking IR to another place.

One concern i have about masking in general is that for a lot of legacy instructions we have have unmasked builtins and I've been wrapping them with selects in IR in the frontend. So the middle end optimizers see the selects and can maybe optimize them through constant folding and the like.. But for 512-bit intrinsics we don't have the selects in IR and instead we have instrinsics that don't expose the same ability to the middle end optimizers.


https://reviews.llvm.org/D26333



---------------------------------------------------------------------
Intel Israel (74) Limited

This e-mail and any attachments may contain confidential material for
the sole use of the intended recipient(s). Any review or distribution
by others is strictly prohibited. If you are not the intended
recipient, please contact the sender and delete all copies.