[LLVMdev] DominatorTree not updated properly after calling the llvm::SplitBlock.

[Paul Vario]

Hi Andy,

         As it turns out, it was my own mistake. I added a couple of BBs,
and forgot to perform "DT->addNewBlock" on them. Still, in the situation
where CFG gets modified, the behavior of the DominatorTree is
counter-intuitive to what LLVM newbies expected to find .. you would think
that DT keeps track of how CFG is being modified, and perform analysis
updates on-the-fly automatically. Thanks a lot for the help.

Best Regards,
Paul


On Fri, Feb 14, 2014 at 1:13 PM, Andrew Trick <atrick at apple.com> wrote:

>
> On Feb 14, 2014, at 9:50 AM, Paul Vario <paul.paul.mit at gmail.com> wrote:
>
> Hi Andrew,
>
>      Thanks a lot. But the function "DT->dominate(A,B)" decides the
> dominance relationship through comparing the DFS numbers, right? At
> least, in my example, when I check whether the newly split node (i.e., %
> *for.end.split*) DOMINATES the original node (I.e., %for.end), the answer
> is true, which is obviously wrong.
>
>
> That sounds like a bug. When DFSInfoValid==false, the DFS number does not
> affect the query. So I don't know why you're getting the wrong answer.
>
> I know you don't have a test case, but I suggest filling a bugzilla PR and
> attaching a dump of both your IR and Domtree before splitting and after
> splitting.
>
> -Andy
>
> Paul
>
>
> On Fri, Feb 14, 2014 at 1:59 AM, Andrew Trick <atrick at apple.com> wrote:
>
>>
>> On Feb 13, 2014, at 4:34 PM, Paul Vario <paul.paul.mit at gmail.com> wrote:
>>
>> Hi Fellows,
>>
>>       I am writing a pass that requires the DominatorTree pass. In the
>> runOnFunction of my own pass, I call llvm::SplitBlock to split a block
>> (%for.end) at it's first insert point. Then, when I dump my dominator tree
>> immediately after, I can see the newly split %for.end.split, but its
>> DFSNumIn and DFSNumOut are not properly updated in llvm::SplitBlock (i.e.,
>> still the initialized -1). At the same time, the DT->dump() results also
>> show that DFSInfoValid equals false! I could not figure out what I did
>> wrong, b/c it seems all the analysis updating should be handled by
>> llvm::SplitBlock. Any suggestions on how to approach this situation? Thanks
>> a bunch in advance.
>>
>>
>> Hi Paul,
>>
>> DFSNumbers is a cache within the DominatorTree analysis. It is lazily
>> computed after a certain number of queries. It's kind of hard to
>> incrementally update these numbers because we may have to shift them all.
>>
>> At any rate, invalid DFS numbers does not mean something is wrong with
>> the analysis.
>>
>> -Andy
>>
>>
>>
>> -------------------------  The actual dump of my dominator tree after
>> SplitBlock --------------------
>>
>> Inorder Dominator Tree: DFSNumbers invalid: 1 slow queries.
>>   [1] %entry.barrier {0,27}
>>     [2] %entry {1,26}
>>       [3] %for.cond {2,25}
>>         [4] %for.body {3,6}
>>           [5] %for.inc {4,5}
>>         [4] %for.end {7,24}
>>           *[5] %for.end.split {4294967295,4294967295}*
>>             [6] %if.then {8,15}
>>               [7] %if.then13 {9,10}
>>               [7] %if.end {11,12}
>>               [7] %if.then.if.end_crit_edge {13,14}
>>             [6] %if.end14 {16,21}
>>               [7] %exit.b4.barrier {17,20}
>>                 [8] %exit.barrier {18,19}
>>             [6] %for.end.if.end14_crit_edge {22,23}
>>
>>  _______________________________________________
>> LLVM Developers mailing list
>> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>>
>>
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20140215/f346b146/attachment.html>