* Scope Kamek bl-patch LR-continuation detection to genuine skip-return targets
Fix crash from Kamek skip-return hooks (Item Rain crash) (#182) added every
Kamek BranchLink patch target to lrContinuationCallTargets unconditionally,
with no filter analogous to the RetroWfcHookSetsLinkRegister check already
used for RetroWFC hooks. Since bl is the ordinary PowerPC call instruction,
this made the codegen treat effectively every patched call in the mod as a
potential skip-return hook, forcing conservative handling (full register
reload, disabled resident-call fast paths, local LR-continuation dispatch
tables) onto thousands of calls that just return normally.
For Retro Rewind this inflated total translated mod size by +42%
(1,414,327 -> 2,005,284 lines), concentrated in ~10 unrelated overlay
functions that happened to call a patched target, and was enough to make
one aggregate build shard pathologically slow to compile (hangs Linux CI).
Instead, only mark a bl target as LR-continuation-aware if a lightweight
discovery-only decode of its own body actually finds evidence of
skip-return behavior via DiscoverLrRelativeIndirectJumpOffsets. Falls back
to the conservative (old) behavior if a target can't be statically
analyzed, so no skip-return case is silently missed.
Verified against the real Retro Rewind mod: total mod size returns to
1,416,350 lines (+0.14% vs. pre-fix, down from +42%), all 6 genuinely new
continuation functions from the original fix are preserved, zero
functions lost, and all 609 existing translator tests still pass.
* Distinguish exhausted from truncated LR-relative offset search
CodeRabbit flagged that TargetExhibitsLrSkipReturn (added in ad2d4e7) treated
an empty DiscoverLrRelativeIndirectJumpOffsets result as a verified "this
target never skip-returns," but the analysis silently drops any path state
once more than MaxStatesPerInstruction (16) distinct states reach one
instruction - a bctr/return on a dropped state can never contribute its
offset, so an empty result could be an incomplete search rather than a real
negative. Treating every capped case as "skip-return possible" outright was
rejected as too broad a fallback given how conservative/expensive that path
already is.
Instead: raise MaxStatesPerInstruction 16 -> 512 (an arbitrary conservative
bound to begin with, not something correctness depended on) so genuinely
branchy functions have far more headroom to reach an exhaustive answer, and
give DiscoverLrRelativeIndirectJumpOffsets an optional onStateCapExceeded
callback that fires exactly when a state is dropped. TargetExhibitsLrSkipReturn
now only falls back to the conservative "treat as skip-return" answer when
the search both found nothing and the cap was actually hit during that run -
not whenever the cap merely exists - so a target is trusted as clean once the
search genuinely exhausts it.
Verified: all 609 translator tests pass, and a full translate-mod run against
the real Retro Rewind mod produces byte-for-byte identical output to the
prior fix (same 4,065 functions, 1,416,350 total lines) - confirming the
16-state cap was never actually the limiting factor in practice and this
change is a pure safety-net closure, not a behavior change for this mod.
* Add LR continuation regression tests
* Refine LR continuation hook analysis
---------
Co-authored-by: patchzyy <64382339+patchzyy@users.noreply.github.com>
* fix: Kamek LR-continuation hook discovery and dispatch
* test: cover branching Kamek LR continuations
* review fix
* another review fix
fix: get the new tests to pass
test: expose LR restore and loop continuation regressions
* Update translator/src/Translator.Core/Mods/ContinuationPlanner.cs
Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com>
* test: cover continuation regressions from the new path-sensitive planner
* Update translator/src/Translator.Core/Mods/ContinuationPlanner.cs
Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com>
* test: cover continuation regressions from the new path-sensitive planner
* fix: preserve LR continuation analysis across large handlers and clobbers
* fix: track LR-relative r1 across update-form stack stores
* Harden LR-relative continuation test coverage
* Fix LR/SP continuation state tracking
---------
Co-authored-by: patchzyy <64382339+patchzyy@users.noreply.github.com>
Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com>
