[AtomicExpandPass] Allow for pointer types in insertRMWCmpXchgLoop()

[arichardson]

Summary

In our CHERI target we support RMW operations on capability types which
are implemented as i8 addrspace(200)*. In certain case we have to fall
back to the cmpxchg expansion and were hitting an assertion here since
getPrimitiveSizeInBits() returns zero for pointer types. Fix this by
querying the DataLayout instead.

[arichardson]

ping?

[arichardson]

ping?

[arichardson]

ping?

[jyknight]

Sorry for never looking at this before now. :(

However, now that atomicrmw instructions have an alignment attribute, this change is superseded by D97223, which fixes the issue, instead, by passing through the atomicrmw alignment, instead of coming up with a new value.

Thanks!

[arichardson]

No longer required.

llvm/lib/CodeGen/AtomicExpandPass.cpp

Value *AtomicExpand::insertRMWCmpXchgLoop(
 
   // The split call above "helpfully" added a branch at the end of BB (to the
   // wrong place), but we want a load. It's easiest to just remove
   // the branch entirely.
   std::prev(BB->end())->eraseFromParent();
   Builder.SetInsertPoint(BB);
   LoadInst *InitLoaded = Builder.CreateLoad(ResultTy, Addr);
   // Atomics require at least natural alignment.
-  InitLoaded->setAlignment(Align(ResultTy->getPrimitiveSizeInBits() / 8));
+  auto TypeSize = F->getParent()->getDataLayout().getTypeSizeInBits(ResultTy);
+  InitLoaded->setAlignment(Align(TypeSize.getFixedSize() / 8));
   Builder.CreateBr(LoopBB);
 
   // Start the main loop block now that we've taken care of the preliminaries.
   Builder.SetInsertPoint(LoopBB);
   PHINode *Loaded = Builder.CreatePHI(ResultTy, 2, "loaded");
   Loaded->addIncoming(InitLoaded, BB);
 
   Value *NewVal = PerformOp(Builder, Loaded);