The result of a function with noundef return attribute must be well defined

[reames]

Summary

This does for readability of returns within said function as what we do for the caller side when reasoning about what might be poison.

Couldn't figure out a small test. I'm working on this from SCEV, but that's blocked by other things as well.

[nikic]

If nothing else you could add a unit test, see programUndefinedIfPoison/isGuaranteedNotToBePoison tests in ValueTrackingTest.

llvm/lib/Analysis/ValueTracking.cpp

5423

This should guard against void returns, in which case getOperand(0) would assert.

[nikic]

llvm/lib/Analysis/ValueTracking.cpp

5423

Hm, I guess this is not strictly necessary thanks to https://github.com/llvm/llvm-project/blob/7d541eb4d49aaaab6a51a3568b9214fd8691e2d3/llvm/lib/IR/Attributes.cpp#L1840. Having noundef implies that the return type cannot be void.

[reames]

Added D111186 as a dependency. Will rebase over updated tests once I can exercise this from SCEV.

[reames]

Add tests - key detail is that removing udiv no-longer has any effect as noimplicit defs serves as a immediate UB root.

[nikic]

LG

llvm/lib/Analysis/ValueTracking.cpp

case Instruction::Invoke: {
         Operands.insert(CB->getCalledOperand());
       for (unsigned i = 0; i < CB->arg_size(); ++i) {
         if (CB->paramHasAttr(i, Attribute::NoUndef) ||
             CB->paramHasAttr(i, Attribute::Dereferenceable))
           Operands.insert(CB->getArgOperand(i));
       }
       break;
     }
-
+    case Instruction::Ret:
+      if (I->getFunction()->hasRetAttribute(Attribute::NoUndef))
+        Operands.insert(I->getOperand(0));
+      break;
     default:
       break;
   }
 }
 
 void llvm::getGuaranteedNonPoisonOps(const Instruction *I,
                                      SmallPtrSetImpl<const Value *> &Operands) {
   getGuaranteedWellDefinedOps(I, Operands);