[llvm-dev] LLVM - getAnalysisUsage()

Mehdi Amini via llvm-dev llvm-dev at lists.llvm.org
Sat Jan 23 11:07:06 PST 2016


> On Jan 22, 2016, at 6:33 PM, Syed Rafiul Hussain <rafiul1906 at gmail.com> wrote:
> 
> Sorry for my typo:
> 
> The actual outputs are following:
> 
>  ModulePass Manager
>    Module pass example
>      Unnamed pass: implement Pass::getPassName()
>    FunctionPass Manager
>      Module Verifier
>    Bitcode Writer
> Pass Arguments:  -functionscalls -domtree -dominancefrontier
>  FunctionPass Manager
>    Function Call Analysis
>    Dominator Tree Construction
>    Dominance Frontier Construction


The issue is not clear to me on this output?


> 
> Yes, I got your point that getAnalysis() is not scheduling the
> execution. However, I do not want a function-pass to be executed more
> than once for a single function. Do you have any suggestions to do
> that?

It has to be marked as preserved by all the other subsequent passes.
There is no other way that I know of.

— 
Mehdi



> 
> Thanks.
> 
> On Fri, Jan 22, 2016 at 4:43 PM, Mehdi Amini <mehdi.amini at apple.com> wrote:
>> 
>>> On Jan 22, 2016, at 12:49 PM, Syed Rafiul Hussain <rafiul1906 at gmail.com> wrote:
>>> 
>>> I have added -debug-pass=Structure, and found the following:
>>> 
>>> ModulePass Manager
>>>   X Analysis
>>>     Unnamed pass: implement Pass::getPassName()
>>>   FunctionPass Manager
>>>     Module Verifier
>>>   Bitcode Writer
>>> Pass Arguments:  -x -y -z
>>> FunctionPass Manager
>>>   X Analysis
>>>   Y Construction
>>>   Z Construction
>> 
>> This is suspicious to me, I think you try to anonymize the names but reused "X Analysis” two times (it shows up as a module pass the first time, and in a functionpass manager the second times).
>> 
>>> 
>>> Even for getAnalysis<Y>(*F) and getAnalysis<Z>(&F), all the passes X,
>>> Y, Z are executed again. In total there are three places where
>>> getAnalysis() directive has been used for X, Y, Z. However, for each
>>> getAnalysis() call, all three passes are scheduled/executed.
>> 
>> Again the “getAnalysis()” has nothing to do with scheduling execution. All is decided with the getRequired calls.
>> 
>>>> Mehdi
>> 
>> 
>>> 
>>> 
>>> 
>>> 
>>> On Fri, Jan 22, 2016 at 12:24 PM, Mehdi Amini <mehdi.amini at apple.com> wrote:
>>>> 
>>>>> On Jan 21, 2016, at 9:05 PM, Syed Rafiul Hussain via llvm-dev <llvm-dev at lists.llvm.org> wrote:
>>>>> 
>>>>> Hi,
>>>>> 
>>>>> I am using llvm-3.8 for my project. Following is my getAnalysisUsage() method:
>>>>> 
>>>>> virtual void getAnalysisUsage(AnalysisUsage &AU) const override
>>>>> {
>>>>> AU.setPreservesAll();
>>>>> AU.addRequired<X>();
>>>>> AU.addRequired<Y>();
>>>>> AU.addRequired<Z>();
>>>>> }
>>>>> 
>>>>> Now, if I call getAnalysis<X>(*F), instead of invoking just the X
>>>>> pass, all the passes, i.e., X, Y and Z are being invoked. Could anyone
>>>>> help me in this regard?
>>>> 
>>>> This is expected. With the current pass manager, the scheduling is done upfront. The “addRequired” directive tells the pass manager that before even starting to run your pass, it needs to schedule X, Y, and Z.
>>>> They will run even without the call to  getAnalysis<X>(*F). The new pass manager (not complete yet but almost) will change that.
>>>> 
>>>> When running your pass with opt, you can add —debug-pass=Structure to see what the scheduling is.
>>>> 
>>>> I hope I didn’t misunderstand your question.
>>>> 
>>>>>>>> Mehdi
>>>> 
>>> 
>>> 
>>> 
>>> --
>>> Rafi
>> 
> 
> 
> 
> -- 
> Rafi



More information about the llvm-dev mailing list