[PATCH] D38563: [MachineBlockPlacement] Make sure PreferredLoopExit is cleared everytime new loop is processed

Xin Tong via llvm-commits llvm-commits at lists.llvm.org
Thu Oct 5 16:03:35 PDT 2017


Hi @gyiu

Sorry about the build break. I have reverted that part of the code
in r315032. The patch which this assert comes with is a fix for an existing
bug in MBP. I will add a test case when I do more investigation, hopefully
from what you provided to me.

Thanks
-Xin

On Fri, Oct 6, 2017 at 7:57 AM, Graham Yiu via Phabricator <
reviews at reviews.llvm.org> wrote:

> gyiu added a comment.
>
> @trentxintong Hello, this change is actually causing the assert to trigger
> in 15 SPEC2006 benchmarks with '-O3 -mcpu=pwr8 -m64 -fprofile-use
> -flto=thin'.  Failures are during the link step.
>
>   lib/CodeGen/MachineBlockPlacement.cpp:2246: void {anonymous}::
> MachineBlockPlacement::buildLoopChains(const llvm::MachineLoop&):
> Assertion `L.isLoopExiting(PreferredLoopExit) && "not an exiting block of
> current loop"' failed.
>
> Benchmarks: 400.perlbench 401.bzip2 403.gcc 433.milc 445.gobmk 447.dealII
> 450.soplex 453.povray 456.hmmer 458.sjeng 464.h264ref 471.omnetpp 473.astar
> 482.sphinx3 483.xalancbmk
>
>
> Repository:
>   rL LLVM
>
> https://reviews.llvm.org/D38563
>
>
>
>


-- 
Software Engineer - Compiler Toolchain
Employee of Facebook Inc.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20171006/0810a8b9/attachment.html>


More information about the llvm-commits mailing list