[llvm-bugs] [Bug 11210] [AVX] opportunity for better code with masked stores

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=11210

Simon Pilgrim <llvm-dev at redking.me.uk> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |FIXED
                 CC|                            |craig.topper at gmail.com
 Fixed By Commit(s)|                            |3170d5484265,e2dee9af8db6
          Component|Scalar Optimizations        |Backend: X86
             Status|NEW                         |RESOLVED

--- Comment #15 from Simon Pilgrim <llvm-dev at redking.me.uk> ---
Tentatively resolving this - test cases were added at rGe2dee9af8db6 to show
@kparzysz's DSE fixes in action.

I played with auto upgrading the intrinsics but the non-optimized codegen was
really awful - I can revisit it if people think its worth it.

Instead, I've just extended the instcombine handling at rG3170d5484265 - where
before we just folded constant masks to generic masked loads/stores, we now do
it for any sign-extended bool vector. This already worked nicely for
BLENDV->SELECT combines (which have a similar sign-bit selection mask
mechanism) so should be fine for masked loads/stores as well.

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20200913/239b29eb/attachment.html>