Do not compact blocks around loops

[kunalspathak]

We would sometimes compact two blocks where 2nd block participates needs loop alignment. If the 1st block (in which 2nd block was merged) was part of an outer loop, then we would falsely think that we have seen a loop block before the top block and disable aligning that loop. That introduces regression in some benchmarks.

The fix would be to not compact such blocks. Also while working, I noticed that we also compact blocks that are part of different loops and that might make bbNatLoopNum information outdated. fgUpdateFlowGraph() (which calls the block compaction code) is called multiple times from various phases and it will be good if we have the bbNatLoopNum information as accurate as possible. Hence, I also added a check that will prohibit us from compacting such blocks that are part of different loops.

From this PR, we get back the alignment in the benchmark.

Relevant discussion: #71646 (comment)

Fixes: #71646

[ghost]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

Issue Details

---

We would sometimes compact two blocks where 2nd block participates needs loop alignment. If the 1st block (in which 2nd block was merged) was part of an outer loop, then we would falsely think that we have seen a loop block before the top block and disable aligning that loop. That introduces regression in some benchmarks.

The fix would be to not compact such blocks. Also while working, I noticed that we also compact blocks that are part of different loops and that might make bbNatLoopNum information outdated. fgUpdateFlowGraph() (which calls the block compaction code) is called multiple times from various phases and it will be good if we have the bbNatLoopNum information as accurate as possible. Hence, I also added a check that will prohibit us from compacting such blocks that are part of different loops.

From this PR, we get back the alignment in the benchmark.

Relevant discussion: #71646 (comment)

Fixes: #71646

Author:	kunalspathak
Assignees:	-
Labels:	area-CodeGen-coreclr
Milestone:	-

[kunalspathak]

@dotnet/jit-contrib

[BruceForstall]

Why can't:

if ((block->bbNatLoopNum != BasicBlock::NOT_IN_LOOP) && (bNext->bbNatLoopNum != BasicBlock::NOT_IN_LOOP) &&
        (block->bbNatLoopNum != bNext->bbNatLoopNum))

be:

if (block->bbNatLoopNum != bNext->bbNatLoopNum)

?

[EgorBo]

Improvements on Linux-arm64: dotnet/perf-autofiling-issues#6774