[llvm-bugs] [Bug 37428] [x86] Vector shift left generates sub-optimal code for shift by "select" from two constants or loop-invariant values

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=37428

Sanjay Patel <spatel+llvm at rotateright.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |FIXED
 Fixed By Commit(s)|                            |r363511
             Status|NEW                         |RESOLVED

--- Comment #13 from Sanjay Patel <spatel+llvm at rotateright.com> ---
I think we've solved all of the vector shift bugs in these examples with:
https://reviews.llvm.org/rL363511
...so I'm going to mark this as fixed. Feel free to reopen if I missed
anything.

AVX1 is still sub-optimal even after:
https://reviews.llvm.org/rL363508

I'll investigate that as a follow-up, but I see 2 potential symptoms depending
on the example:
1. We unrolled too much and ran out of registers, so we're spilling.
2. We failed to expand/hoist all of the vector shifts when fully unrolled.

Root cause may be cost model/unrolling. Ie, I don't see these problems if I use
"-fno-unroll-loops" (and perf improves with that setting).

-- 
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/20190617/c1298053/attachment.html>