[llvm] [SampleFDO] Support enabling sample loader pass in O0 mode (PR #113985)

[Lei Wang via llvm-commits]

================
@@ -103,3 +104,10 @@ declare i32 @printf(ptr, ...)
 ; MERGE: !{!"branch_weights", i32 10}
 ; MERGE: name: "sub"
 ; MERGE-NEXT: {!"function_entry_count", i64 3}
+
+; FLATTEN: name: "sum"
+; FLATTEN-NEXT: {!"function_entry_count", i64 35}
+; FLATTEN: !{!"branch_weights", i32 13, i32 23}
+; FLATTEN: !{!"branch_weights", i32 12}
----------------
wlei-llvm wrote:

OK, after switch using flattened profile, it triggered a test failure, I took a deep look at this, the reason might explain the difference using flattened profile vs merging not inlined profile on-the-fly(merged profile):
Original profile (two sum are in different context)
```

...
 sum:46
  2:10 sub:10
...
 sum:11:22
  2: sub:2
   1: 2
```

The merged profile:
```
sum:46
 2:10 sub:10
 ...
 2: sub:2
  1: 2
```
For a flattened profile
```
sum:46
 2:12 sub:12
```
Then for the line 2, it gives different block count 12 vs 10, because the inlinee profile isn't counted for the block count, which I think it should be, so this seems using flattened profile makes more sense(though the difference is not big)



https://github.com/llvm/llvm-project/pull/113985