[llvm-bugs] [Bug 26084] [AArch64] Missed CCMP opportunity

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=26084

Eli Friedman <efriedma at quicinc.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|RESOLVED                    |REOPENED
                 CC|                            |efriedma at quicinc.com,
                   |                            |florian_hahn at apple.com,
                   |                            |sjoerd.meijer at arm.com
         Resolution|FIXED                       |---

--- Comment #4 from Eli Friedman <efriedma at quicinc.com> ---
The patch in question was reverted, and apparently nobody has looked into this
since.

I briefly tried rebasing the patch.  It basically works, but the "custom"
markings have unintended consequences, and as noted in the revert, the
DAGCombine changes are suspect.  (One of those consequences is suppressing
add->or conversion after legalization; filed bug 46151 and bug 46152.)

I don't think marking integer OR for custom lowering is an appropriate
solution; there are too many potential weird consequences in terms of
suppressing target-independent DAGCombines.  Probably should be a DAGCombine
itself.  (Not sure if it would run before or after legalization.  Before,
checking profitability is hard.  After, we have to write some new code to dig
into target-specific compare nodes.)

-- 
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/20200601/26d90a45/attachment.html>