[llvm] [AMDGPU] Look through bitcast when skipping musttail returns in UnifyDivergentExitNodes (PR #201021)

[Matt Arsenault via llvm-commits]

================
@@ -102,3 +103,29 @@ bb.0:
 bb.1:
   ret i32 %load
 }
+
+define ptr @musttail_call_with_bitcast(ptr %p) {
+; CHECK-LABEL: define ptr @musttail_call_with_bitcast(
+; CHECK-SAME: ptr [[P:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[LOAD:%.*]] = load i1, ptr [[P]], align 1
+; CHECK-NEXT:    br i1 [[LOAD]], label %[[BB_0:.*]], label %[[BB_1:.*]]
+; CHECK:       [[BB_0]]:
+; CHECK-NEXT:    [[RET:%.*]] = musttail call ptr @bar_ptr(ptr [[P]])
+; CHECK-NEXT:    [[RETC:%.*]] = bitcast ptr [[RET]] to ptr
+; CHECK-NEXT:    ret ptr [[RETC]]
+; CHECK:       [[BB_1]]:
+; CHECK-NEXT:    ret ptr null
+;
+entry:
+  %load = load i1, ptr %p, align 1
+  br i1 %load, label %bb.0, label %bb.1
+
+bb.0:
+  %ret = musttail call ptr @bar_ptr(ptr %p)
+  %retc = bitcast ptr %ret to ptr
----------------
arsenm wrote:

I'd prefer examples that aren't true no-ops, like bitcast from int to float 

https://github.com/llvm/llvm-project/pull/201021