[llvm] llvm-reduce: Avoid removing convergent with convergence tokens (PR #132946)

[Sameer Sahasrabuddhe via llvm-commits]

================
@@ -44,9 +44,25 @@ namespace {
 class AttributeRemapper : public InstVisitor<AttributeRemapper> {
   Oracle &O;
   LLVMContext &Context;
+  bool HasControlledConvergence = true;
 
 public:
-  AttributeRemapper(Oracle &O, LLVMContext &C) : O(O), Context(C) {}
+  AttributeRemapper(Oracle &O, Module &M) : O(O), Context(M.getContext()) {
+
+    // Check if there are any convergence intrinsics used. We cannot remove the
----------------
ssahasra wrote:

Perhaps the last part of my comment about uses of the loop intrinsic was a distraction. My point is that the convergence control token can be viewed as an improvement over the convergent intrinsic, to the point that the intrinsic will become redundant if tokens are used everywhere. The token does not itself participate in control flow or data flow. In that sense, deleting the token on a convergent call while removing the convergent attribute is a valid interpretation of what this pass is doing. Even if we made a separate pass that removes token operands, that pass must remove the convergent attribute to keep the IR valid. Then we might as well do both right here.

https://github.com/llvm/llvm-project/pull/132946