[PATCH] D68154: [Dominators][CodeGen] Fix MachineDominatorTree preservation in PHIElimination

[Jakub Kuderski via Phabricator via llvm-commits]

kuhar planned changes to this revision.
kuhar added a comment.

Thanks for the comments, @arsenm and @hiraditya.

I tried not declaring MDT as a preserved analysis at all, but I got issues with OrcJit unittests:

  [2019-09-29 15:31:37.806980254] 0x54de520     Made Modification 'Simple Register Coalescing' on Function 'bar'...
  [2019-09-29 15:31:37.807045279] 0x54de520      Freeing Pass 'Simple Register Coalescing' on Function 'bar'...
  [2019-09-29 15:31:37.807075980] 0x54de520      Freeing Pass 'MachineDominator Tree Construction' on Function 'bar'...
  [2019-09-29 15:31:37.807107079] 0x54de520     Executing Pass 'Rename Disconnected Subregister Components' on Function 'bar'...
  [2019-09-29 15:31:37.807177117] 0x54de520      Freeing Pass 'Rename Disconnected Subregister Components' on Function 'bar'...
  [2019-09-29 15:31:37.807204793] 0x54de520     Executing Pass 'Machine Instruction Scheduler' on Function 'bar'...
  OrcJITTests: /home/kuba/llvm/llvm-project/llvm/include/llvm/PassAnalysisSupport.h:235: AnalysisType &llvm::Pass::getAnalysisID(llvm::AnalysisID) const [AnalysisType = llvm::MachineDominatorTree]: Assertion `ResultPass && "getAnalysis*() called on an analysis that was not " "'required' by pass!"' failed.

Not sure why this is the case, as I marked MDT as a required analysis for MachineScheduler. Do you have some tips here? I'm a newbie when it comes to MIR and Machine passes.

When I preserve the analysis in PHIElimination everything works.


Repository:
  rL LLVM

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D68154/new/

https://reviews.llvm.org/D68154