AMDGPU: Fix worklist management in simplifyDemandedVectorEltsIntrinsic

Fixes bot sanitizer error, but it does leave behind a dead instruction
if there is a bundle for some reason.

Author: arsenm

llvm/lib/Target/AMDGPU/AMDGPUInstCombineIntrinsic.cpp

@@ -1599,11 +1599,10 @@ Value *GCNTTIImpl::simplifyAMDGCNLaneIntrinsicDemanded(
   // TODO: Preserve callsite attributes?
   CallInst *NewCall = IC.Builder.CreateCall(Remangled, {Extract}, OpBundles);
 
-  Value *Result = IC.Builder.CreateInsertElement(PoisonValue::get(II.getType()),
-                                                 NewCall, FirstElt);
-  IC.replaceInstUsesWith(II, Result);
-  IC.eraseInstFromFunction(II);
-  return Result;
+  // FIXME: If the call has a convergence bundle, we end up leaving the dead
+  // call behind.
+  return IC.Builder.CreateInsertElement(PoisonValue::get(II.getType()), NewCall,
+                                        FirstElt);
 }
 
 std::optional<Value *> GCNTTIImpl::simplifyDemandedVectorEltsIntrinsic(

llvm/test/Transforms/InstCombine/AMDGPU/simplify-demanded-vector-elts-lane-intrinsics.ll

@@ -349,6 +349,7 @@ define i32 @extract_elt0_v2i32_readfirstlane_convergencetoken(<2 x i32> %src) co
 ; CHECK-LABEL: define i32 @extract_elt0_v2i32_readfirstlane_convergencetoken(
 ; CHECK-SAME: <2 x i32> [[SRC:%.*]]) #[[ATTR1:[0-9]+]] {
 ; CHECK-NEXT:    [[T:%.*]] = call token @llvm.experimental.convergence.entry()
+; CHECK-NEXT:    [[VEC:%.*]] = call <2 x i32> @llvm.amdgcn.readfirstlane.v2i32(<2 x i32> [[SRC]]) [ "convergencectrl"(token [[T]]) ]
 ; CHECK-NEXT:    [[TMP1:%.*]] = extractelement <2 x i32> [[SRC]], i64 0
 ; CHECK-NEXT:    [[ELT:%.*]] = call i32 @llvm.amdgcn.readfirstlane.i32(i32 [[TMP1]]) [ "convergencectrl"(token [[T]]) ]
 ; CHECK-NEXT:    ret i32 [[ELT]]