AMDGPU: Don't run redundant GlobalDCE

[arsenm]

Summary

The internalize pass is added early, and
GlobalOpt / GlobalDCE are already run later.

[rampitec]

The idea to run it was to speedup compilation. After internalize we have an opportunity to remove some functions and do not spend time optimizing them. As far as I understand GlobalDCE is really cheap, unlike passes running before its next invocation.

Did you make any compile time measurements to support it?

[arsenm]

In D71263#1777570, @rampitec wrote:

The idea to run it was to speedup compilation. After internalize we have an opportunity to remove some functions and do not spend time optimizing them. As far as I understand GlobalDCE is really cheap, unlike passes running before its next invocation.

Did you make any compile time measurements to support it?

I found this patch lying around from 2017, and don't remember why I was looking at this. There aren't many passes run between the internalize here and GlobalOpt:

Internalize Global Symbols
Dead Global Elimination
Interprocedural Sparse Conditional Constant Propagation
  FunctionPass Manager
    Dominator Tree Construction
Called Value Propagation
Deduce and propagate attributes
Global Variable Optimizer

[rampitec]

In D71263#1777573, @arsenm wrote:

In D71263#1777570, @rampitec wrote:

The idea to run it was to speedup compilation. After internalize we have an opportunity to remove some functions and do not spend time optimizing them. As far as I understand GlobalDCE is really cheap, unlike passes running before its next invocation.

Did you make any compile time measurements to support it?

I found this patch lying around from 2017, and don't remember why I was looking at this. There aren't many passes run between the internalize here and GlobalOpt:

Internalize Global Symbols
Dead Global Elimination
Interprocedural Sparse Conditional Constant Propagation
  FunctionPass Manager
    Dominator Tree Construction
Called Value Propagation
Deduce and propagate attributes
Global Variable Optimizer

AFAIR initially GlobalDCE was added to skip unused library functions after link. Anyway, do you think GlobalDCE adds any visible slowdown, which is more than DT construction?

llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

[Internalize, EarlyInline, AMDGPUAA, this](const PassManagerBuilder &,
         PM.add(createAMDGPUAAWrapperPass());
         PM.add(createAMDGPUExternalAAWrapperPass());
       }
       PM.add(createAMDGPUUnifyMetadataPass());
       PM.add(createAMDGPUPrintfRuntimeBinding());
       PM.add(createAMDGPUPropagateAttributesLatePass(this));
       if (Internalize) {
         PM.add(createInternalizePass(mustPreserveGV));
-        PM.add(createGlobalDCEPass());
       }
       if (EarlyInline)
         PM.add(createAMDGPUAlwaysInlinePass(false));
   });
 
   const auto &Opt = Options;
   Builder.addExtension(
     PassManagerBuilder::EP_EarlyAsPossible,