AMDGPU: Restrict the number of instructions to scan for getPointerDependencyFrom

[cfang]

Summary

This is for getPointerDependencyFrom in AMDGPUAnnotateUniformValues.
With the default scan limit of 100 instruction per block, we can reduce the compile time
of a special kernel from longer than 2 hours to 7 minutes.

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUAnnotateUniformValues.cpp

109

I don't see how this changes anything. This is what the default null case does?

unsigned DefaultLimit = getDefaultBlockScanLimit();
if (!Limit)
  Limit = &DefaultLimit;

[cfang]

llvm/lib/Target/AMDGPU/AMDGPUAnnotateUniformValues.cpp

109

Ooh, the change is even not correct due to the weird design of the Limit usage:

--*Limit;
if (!*Limit)
  return MemDepResult::getUnknown();

This will usually causes trouble if you don't use default nullptr. You will have to set the Limit before each call.

To say that, the use at DeadStoreElimination.cpp is also not correct.

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUAnnotateUniformValues.cpp

109

Seems like you should fix that instead of working around it here, or make it a mandatory argument?

[cfang]

llvm/lib/Target/AMDGPU/AMDGPUAnnotateUniformValues.cpp

109

We should fix DeadStoreElimination case anyway.

And it seems DeadStoreElimination is the only case using explicit argument. However I am using
an argument is still useful which allows the user to specify the limit instead of using the default.
We may just pass by value with a negative meaning default.

[cfang]

https://reviews.llvm.org/D84873 addresses the same issue. So this is no longer needed.

llvm/lib/Target/AMDGPU/AMDGPUAnnotateUniformValues.cpp

       L = P;
       P = P->getParentLoop();
     } while (P);
     Checklist.insert(L->block_begin(), L->block_end());
     Start = L->getHeader();
   }
 
   DFS(Start, Checklist);
+  // Maximum number of instructions to search in a block (default 100).
+  auto Limit = MDR->getDefaultBlockScanLimit();
   for (auto &BB : Checklist) {
     BasicBlock::iterator StartIt = (!L && (BB == Load->getParent())) ?
       BasicBlock::iterator(Load) : BB->end();
     auto Q = MDR->getPointerDependencyFrom(MemoryLocation(Ptr), true,
-                                           StartIt, BB, Load);
+                                           StartIt, BB, Load, &Limit);
     if (Q.isClobber() || Q.isUnknown())
       return true;
   }
   return false;
 }
 
 void AMDGPUAnnotateUniformValues::visitBranchInst(BranchInst &I) {
   if (DA->isUniform(&I))