[llvm] [Attributor][AMDGPU] Enable AAIndirectCallInfo for AMDAttributorPass (PR #100952)

[Shilei Tian via llvm-commits]

================
@@ -231,7 +231,19 @@ define amdgpu_kernel void @indirect_calls_none_agpr(i1 %cond) {
 ; CHECK-LABEL: define amdgpu_kernel void @indirect_calls_none_agpr(
 ; CHECK-SAME: i1 [[COND:%.*]]) #[[ATTR0]] {
 ; CHECK-NEXT:    [[FPTR:%.*]] = select i1 [[COND]], ptr @empty, ptr @also_empty
-; CHECK-NEXT:    call void [[FPTR]]()
+; CHECK-NEXT:    [[TMP1:%.*]] = icmp eq ptr [[FPTR]], @also_empty
+; CHECK-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP3:%.*]]
+; CHECK:       2:
+; CHECK-NEXT:    call void @also_empty()
+; CHECK-NEXT:    br label [[TMP6:%.*]]
+; CHECK:       3:
+; CHECK-NEXT:    br i1 true, label [[TMP4:%.*]], label [[TMP5:%.*]]
+; CHECK:       4:
+; CHECK-NEXT:    call void @empty()
+; CHECK-NEXT:    br label [[TMP6]]
+; CHECK:       5:
+; CHECK-NEXT:    unreachable
+; CHECK:       6:
----------------
shiltian wrote:

It is not an accident. `AAIndirectCallInfo` will do both analysis and manifest. This change is expected (for now) because it replaces an indirect function call with two direct function calls under condition. This is assumed to have better performance. The reason it is for now is, in https://github.com/llvm/llvm-project/pull/100954 we add uniform check, and this change doesn't conform with it.

https://github.com/llvm/llvm-project/pull/100952