ssair: fix phi edge cleanup for current block

[KristofferC]

🤖: When removing an edge during compaction, phi pruning treated only blocks with to < active_bb as already compacted. If the edge was removed into the current block, its phi nodes had already been rewritten with renamed predecessors, but the cleanup path still searched using the old predecessor ID in the pre-compact IR. This left a live block with a phi edge from a dead predecessor and broke IR verification. Treat to == active_bb as already compacted so phi edges are removed from the result stream with renamed predecessor IDs.

AI fix (take 2) of #60660. Putting up for consideration.

The test only reproduces on 1.13 but figured I might as well add it in for potential backport.

[Keno]

Fix looks plausible, but I'll need to take a deeper look. It'd be nice to get an ir-level test also in addition to the regression test.

[Keno]

So I think the identification of the issue is correct, but i don't think the fix quite works, because the statement range in result_bbs will not be correct for the active_bb. Instead the already compacted range for the current bb should be computed as StmtRange(first(bb.stmts), compact.result_idx-1) as it would be when finishing the bb (can pull this out into a helper). Thus if to == active_bb, this needs to be special cased. We are also making the assumption that all phi block statements have already been processed here. This is true, but should be documented and perhaps encoded in the name of the function, e.g. kill_edge_terminator! (only the version that has the active_bb arg of course). Plus needs an IR level test as mentioned.

(If you paste this into your claude session it'll probably give you correct code).

[KristofferC]

updated