[polly] scop detection: small cleanups

Tobias Grosser tobias at grosser.es
Fri Jun 14 18:41:10 PDT 2013


On 06/14/2013 01:32 PM, Sebastian Pop wrote:
> Tobias Grosser wrote:
>> On 06/12/2013 03:32 PM, Sebastian Pop wrote:
>>> Hi Tobi,
>>>
>>> I have 3 more small cleanups for compile time. I already sent an email about
>>> 4/4, so I'm not attaching it again here:
>>>
>>> [PATCH 1/4] scop detection: only execute print stmts under DEBUG
>>> [PATCH 2/4] scop detection: check flag before expensive call
>>> [PATCH 3/4] scop detection: inline and remove isValidBasicBlock
>>> [PATCH 4/4] scop detection: detect regions bottom up
>>
>> 2 and 3 are OK to commit.
>
> Committed r184000
> Committed r184001
>
>> The first patch causes minor troubles with:
>>
>> polly-opt test/Dependences/do_pluto_matmult.ll -view-scops
>>
>> Without the patch, the dotty graph gives the information that some
>> scops are not detected due to aliasing. With the patch the
>> information
>> is only printed if -debug is added. The other messages are still
>> printed without -debug.
>>
>> I understand that we want to move the formatting of the output
>> message out of the hot path and I fully agree with this. However,
>> just moving some of the formatting into DEBUG() causes inconsistent
>> behavior. We should either have all error messages showing or none
>> of them. A fast
>> solution would be to move this under a flag that is enabled if -debug
>> is set or if -polly-show or -polly-show-only is set.
>
> What about flipping -debug on when one of the flags -polly-show[-only] is used?

This will cause a lot of onrelated debug prints on the console. Also, I 
am unsure if it works in release mode.

Tobi




More information about the llvm-commits mailing list