[llvm-branch-commits] [llvm] [SimplifyCFG][profcheck] Profile propagation for `indirectbr` (PR #161747)

[Alan Zhao via llvm-branch-commits]

================
@@ -8,15 +8,19 @@ declare void @A()
 declare void @B(i32)
 declare void @C()
 
-define void @indbrtest0(ptr %P, ptr %Q) {
+;.
+; CHECK: @anchor = constant [13 x ptr] [ptr blockaddress(@indbrtest3, %L1), ptr blockaddress(@indbrtest3, %L2), ptr inttoptr (i32 1 to ptr), ptr blockaddress(@indbrtest4, %L1), ptr inttoptr (i32 1 to ptr), ptr inttoptr (i32 1 to ptr), ptr inttoptr (i32 1 to ptr), ptr inttoptr (i32 1 to ptr), ptr inttoptr (i32 1 to ptr), ptr inttoptr (i32 1 to ptr), ptr inttoptr (i32 1 to ptr), ptr inttoptr (i32 1 to ptr), ptr inttoptr (i32 1 to ptr)]
+; CHECK: @xblkx.bbs = internal unnamed_addr constant [9 x ptr] [ptr blockaddress(@indbrtest7, %xlab4x), ptr blockaddress(@indbrtest7, %xlab4x), ptr blockaddress(@indbrtest7, %v2j), ptr blockaddress(@indbrtest7, %xlab4x), ptr blockaddress(@indbrtest7, %xlab4x), ptr blockaddress(@indbrtest7, %xlab4x), ptr blockaddress(@indbrtest7, %xlab4x), ptr blockaddress(@indbrtest7, %xlab4x), ptr blockaddress(@indbrtest7, %v2j)]
+;.
+define void @indbrtest0(ptr %P, ptr %Q) !prof !0 {
 ; CHECK-LABEL: @indbrtest0(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    store ptr blockaddress(@indbrtest0, [[BB0:%.*]]), ptr [[P:%.*]], align 8
 ; CHECK-NEXT:    store ptr blockaddress(@indbrtest0, [[BB1:%.*]]), ptr [[P]], align 8
 ; CHECK-NEXT:    store ptr blockaddress(@indbrtest0, [[BB2:%.*]]), ptr [[P]], align 8
 ; CHECK-NEXT:    call void @foo()
 ; CHECK-NEXT:    [[T:%.*]] = load ptr, ptr [[Q:%.*]], align 8
-; CHECK-NEXT:    indirectbr ptr [[T]], [label [[BB0]], label [[BB1]], label %BB2]
+; CHECK-NEXT:    indirectbr ptr [[T]], [label [[BB0]], label [[BB1]], label %BB2], !prof [[PROF1:![0-9]+]]
----------------
alanzhao1 wrote:

Shouldn't the branch weights be summed grouped by destination instead of dropped here?

If I understand correctly, the purpose of this transformation is to remove redundant labels in indirectbr statements. As an example,

```
indirectbr ptr %foo [label %L1, label %L2, label %L1]
```

is simplified to

```
indirectbr ptr %foo [label %L1, label %L2]
```

Now assume in the original statement the weights are 1, 1, and 1. This implies that `%L1` is twice as likely to be branched to compared to `%L2`. The simplification removes redundant labels, but since nothing else gets changed, shouldn't the branch weights be 2 and 1 instead of 1 and 1 as the instruction is identical?

https://github.com/llvm/llvm-project/pull/161747