[LLVMdev] [Polly] Move Polly's execution later
Star Tan
tanmx_star at yeah.net
Thu Oct 3 11:13:19 PDT 2013
At 2013-10-03 19:17:51,"Tobias Grosser" <tobias at grosser.es> wrote:>On 09/30/2013 04:11 AM, Star Tan wrote:
>> At 2013-09-25 18:03:18,"Tobias Grosser" <tobias at grosser.es> wrote:>
>>> I think this is too early, as most of the canonicalization is not yet
>>> done. We probably don't need to investigate this bug immediately, but
>>> it would be nice if we could make it reproducible without your changes
>>> to polly. For this please run the command with -debug-pass=Arguments
>>> and replace the -O3 with the actual set of commands that is needed to
>>> trigger the bug. If you can reduce the set of commands using bugpoint,
>>> that would be even better.
>>>
>>
>> I am trying to figure out why "moving Polly before loop rotate pass would
>> lead to incorrect code generation.
>>
>> With the help of "-debug-pass=Arguments", I have reproduced the
>> bug without modify LLVM and Polly. For the attached testcase, it will produce
>> incorrect code if we run a very simple command:
>> $ opt -load LLVMPolly.so -basicaa -mem2reg -polly-codegen -loop-simplify -print-module foo.preopt.ll
>>
>> The output LLVM IR code would contain a exit basic block that is incorrectly
>> marked as "unreachable".
>>
>> I have updated the bug17323 (http://llvm.org/bugs/show_bug.cgi?id=17323)
>> for the further debugging.
>
>The problem is that at some point we failt to properly update the
>LoopInfo. I updated the bug report with more details.
>
Thanks a lot! I will try to find out why LoopInfo is not updated properly.
Star Tan
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20131004/8137fac9/attachment.html>
More information about the llvm-dev
mailing list