[PATCH] D12781: PGO IR-level instrumentation infrastructure

[Rong Xu via llvm-commits]

On Tue, Nov 24, 2015 at 4:46 PM, Xinliang David Li <davidxl at google.com>
wrote:

>
>
> On Tue, Nov 24, 2015 at 4:20 PM, Sean Silva <chisophugis at gmail.com> wrote:
>
>>
>>
>> On Tue, Nov 24, 2015 at 4:07 PM, Rong Xu <xur at google.com> wrote:
>>
>>> Hi Sean,
>>>
>>> Sorry about this. I revert the change a moment ago.
>>>
>>> As for the test. Here is the way I want to fix:
>>> In profile-generate tests, I will test (1) number of counters inserted
>>> in each function, and (2) If there is a instrumentation in a specific bb. I
>>> will skip the count index.
>>>
>>> In profile-use, I will invoke a new pass branch-prob that computes the
>>> branch probability.  I will check the branch probability rather the branch
>>> weight metadata.
>>>
>>> Let me know if you think this is reasonable, or you have a better way to
>>> do the test.
>>>
>>
>> You can use FileCheck variable captures (
>> http://llvm.org/docs/CommandGuide/FileCheck.html#filecheck-variables) to
>> match up the indexes (see the example where it captures a register
>> assignment and checks that a later instruction uses that register).
>> Together with good use of CHECK lines, you should be able to capture
>> exactly what you need and make the test very readable.
>>
>> The profile-use pass only adds metadata, so we should be checking the
>> metadata. We should not need to invoke branch-prob to test the behavior.
>>
>> Also, Hal's point is worth investigating: why is this pass not
>> deterministic across platforms? LLVM code should have deterministic output
>> independent of the host platform.
>>
>
> One thing I noticed is that std::stable_sort should be probably be used
> instead of std::sort.
>

It seems all the test failures are caused by std::sort.
I used darwin to verify. Once using stable_sort, the instrumentation order
became
identical in two platforms and all the test turns green.

-Rong


>
> Another thing is that the dumping traverses a map indexed by address --
> this can result in different order on different host -- but the test cases
> do not depend on that (debug dump).
>
> david
>
>
>
>>
>> -- Sean Silva
>>
>>
>>>
>>> Thanks,
>>>
>>> -Rong
>>>
>>>
>>>
>>> On Tue, Nov 24, 2015 at 3:16 PM, Sean Silva <chisophugis at gmail.com>
>>> wrote:
>>>
>>>> Hi Rong,
>>>>
>>>> Typically, when there have been multiple active reviewers, it is common
>>>> courtesy to wait for all of them to LGTM the patch.
>>>>
>>>> I had further comments on the testing here (which the bots seem to have
>>>> caught you on anyway), so I recommend reverting this patch for the moment
>>>> and reopening the review.
>>>>
>>>> -- Sean Silva
>>>>
>>>>
>>>> On Tue, Nov 24, 2015 at 1:34 PM, Rong Xu <xur at google.com> wrote:
>>>>
>>>>> This revision was automatically updated to reflect the committed
>>>>> changes.
>>>>> Closed by commit rL254021: [PGO] MST based PGO instrumentation
>>>>> infrastructure (authored by xur).
>>>>>
>>>>> Changed prior to commit:
>>>>>   http://reviews.llvm.org/D12781?vs=41062&id=41079#toc
>>>>>
>>>>> Repository:
>>>>>   rL LLVM
>>>>>
>>>>> http://reviews.llvm.org/D12781
>>>>>
>>>>> Files:
>>>>>   llvm/trunk/include/llvm/IR/DiagnosticInfo.h
>>>>>   llvm/trunk/include/llvm/InitializePasses.h
>>>>>   llvm/trunk/include/llvm/LinkAllPasses.h
>>>>>   llvm/trunk/include/llvm/Transforms/Instrumentation.h
>>>>>   llvm/trunk/lib/IR/DiagnosticInfo.cpp
>>>>>   llvm/trunk/lib/Transforms/IPO/LLVMBuild.txt
>>>>>   llvm/trunk/lib/Transforms/Instrumentation/CFGMST.h
>>>>>   llvm/trunk/lib/Transforms/Instrumentation/CMakeLists.txt
>>>>>   llvm/trunk/lib/Transforms/Instrumentation/Instrumentation.cpp
>>>>>   llvm/trunk/lib/Transforms/Instrumentation/LLVMBuild.txt
>>>>>   llvm/trunk/lib/Transforms/Instrumentation/PGOInstrumentation.cpp
>>>>>   llvm/trunk/test/Transforms/PGOProfile/Inputs/branch1.proftext
>>>>>   llvm/trunk/test/Transforms/PGOProfile/Inputs/branch2.proftext
>>>>>   llvm/trunk/test/Transforms/PGOProfile/Inputs/criticaledge.proftext
>>>>>   llvm/trunk/test/Transforms/PGOProfile/Inputs/landingpad.proftext
>>>>>   llvm/trunk/test/Transforms/PGOProfile/Inputs/loop1.proftext
>>>>>   llvm/trunk/test/Transforms/PGOProfile/Inputs/loop2.proftext
>>>>>   llvm/trunk/test/Transforms/PGOProfile/Inputs/loop3.proftext
>>>>>   llvm/trunk/test/Transforms/PGOProfile/Inputs/single_bb.proftext
>>>>>   llvm/trunk/test/Transforms/PGOProfile/Inputs/switch.proftext
>>>>>   llvm/trunk/test/Transforms/PGOProfile/branch1_gen.ll
>>>>>   llvm/trunk/test/Transforms/PGOProfile/branch1_use.ll
>>>>>   llvm/trunk/test/Transforms/PGOProfile/branch2_gen.ll
>>>>>   llvm/trunk/test/Transforms/PGOProfile/branch2_use.ll
>>>>>   llvm/trunk/test/Transforms/PGOProfile/checksum_mismatch.ll
>>>>>   llvm/trunk/test/Transforms/PGOProfile/criticaledge_gen.ll
>>>>>   llvm/trunk/test/Transforms/PGOProfile/criticaledge_use.ll
>>>>>   llvm/trunk/test/Transforms/PGOProfile/landingpad_gen.ll
>>>>>   llvm/trunk/test/Transforms/PGOProfile/landingpad_use.ll
>>>>>   llvm/trunk/test/Transforms/PGOProfile/loop1_gen.ll
>>>>>   llvm/trunk/test/Transforms/PGOProfile/loop1_use.ll
>>>>>   llvm/trunk/test/Transforms/PGOProfile/loop2_gen.ll
>>>>>   llvm/trunk/test/Transforms/PGOProfile/loop2_use.ll
>>>>>   llvm/trunk/test/Transforms/PGOProfile/loop3_gen.ll
>>>>>   llvm/trunk/test/Transforms/PGOProfile/loop3_use.ll
>>>>>   llvm/trunk/test/Transforms/PGOProfile/noprofile_use.ll
>>>>>   llvm/trunk/test/Transforms/PGOProfile/single_bb_gen.ll
>>>>>   llvm/trunk/test/Transforms/PGOProfile/switch_gen.ll
>>>>>   llvm/trunk/test/Transforms/PGOProfile/switch_use.ll
>>>>>
>>>>>
>>>>
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20151124/616040fa/attachment.html>