[llvm] [TailDuplicator] Add a limit on the size of predecessors (PR #78582)

[Quentin Dian via llvm-commits]

DianQK wrote:

I made some progress.

There are usually only two instructions that can be duplicated. Indirect branches increase this limit to 20.

https://github.com/llvm/llvm-project/blob/4b2381a5f05dcd576e50615ec1d9661fbae3282c/llvm/lib/CodeGen/TailDuplicator.cpp#L591-L602

I understand from the comments that this is to improve the accuracy of branch prediction. I want to know that if it is appropriate with numerous of indirect branches. So I did some experimenting on https://github.com/DianQK/llvm-tail-dup-indirect-succ-size.

I'm using the result of `instructions:u` and `branch-misses:u` that don't call `f1` as a baseline. The arguments of `dup` are `-tail-dup-indirect-size=20`, and the arguments of `nodup` are `-tail-dup-indirect-size=2`. I think the replication of this scenario may not make sense.

One of my results is as follows:

```
==================
Iterations: 1
dup: instructions: 418, branches: 32, branch-misses: 64
nodup: instructions: 280, branches: 53, branch-misses: 1
==================
Iterations: 2
dup: instructions: 840, branches: 63, branch-misses: 60
nodup: instructions: 565, branches: 106, branch-misses: 30
==================
Iterations: 3
dup: instructions: 1260, branches: 93, branch-misses: 46
nodup: instructions: 857, branches: 159, branch-misses: 62
==================
Iterations: 4
dup: instructions: 1681, branches: 123, branch-misses: 48
nodup: instructions: 1135, branches: 212, branch-misses: 74
==================
Iterations: 5
dup: instructions: 2102, branches: 152, branch-misses: 46
nodup: instructions: 1434, branches: 265, branch-misses: 67
==================
Iterations: 6
dup: instructions: 2536, branches: 183, branch-misses: 55
nodup: instructions: 1705, branches: 318, branch-misses: 82
==================
Iterations: 7
dup: instructions: 2944, branches: 212, branch-misses: 60
nodup: instructions: 1990, branches: 371, branch-misses: 77
==================
Iterations: 8
dup: instructions: 3371, branches: 243, branch-misses: 45
nodup: instructions: 2279, branches: 424, branch-misses: 57
==================
Iterations: 9
dup: instructions: 3793, branches: 273, branch-misses: 48
nodup: instructions: 2565, branches: 477, branch-misses: 56
==================
Iterations: 10
dup: instructions: 4232, branches: 309, branch-misses: 61
nodup: instructions: 2870, branches: 536, branch-misses: 49
==================
Iterations: 100
dup: instructions: 42138, branches: 3015, branch-misses: 54
nodup: instructions: 28540, branches: 5312, branch-misses: 55
==================
Iterations: 1000
dup: instructions: 421057, branches: 30020, branch-misses: 42
nodup: instructions: 285055, branches: 53018, branch-misses: 84
==================
Iterations: 10000
dup: instructions: 4210078, branches: 300027, branch-misses: 51
nodup: instructions: 2850075, branches: 530024, branch-misses: 15
==================
Iterations: 100000
dup: instructions: 42100100, branches: 3000035, branch-misses: 60
nodup: instructions: 28500098, branches: 5300032, branch-misses: 52
==================
Iterations: 1000000
dup: instructions: 421000146, branches: 30000068, branch-misses: 157
nodup: instructions: 285000149, branches: 53000061, branch-misses: 117
==================
Iterations: 10000000
dup: instructions: 4210000439, branches: 300000341, branch-misses: 749
nodup: instructions: 2850000384, branches: 530000291, branch-misses: 621
==================
Iterations: 100000000
dup: instructions: 42100003182, branches: 3000003075, branch-misses: 6277
nodup: instructions: 28500002671, branches: 5300002564, branch-misses: 5186
```

If this is the right route, I'll continue to figure out the other two problems.

- Why does execute more instructions? (Related to increased PHI?)
- Finding a suitable successor size?


https://github.com/llvm/llvm-project/pull/78582