Repository navigation
REGRESSION: e30_array_lit_dim_assign_shift no longer compiles — HashTable::shiftFirst() unresolved #24025
Description
Summary
test/differential/cases/e30_array_lit_dim_assign_shift.php no longer compiles under AOT:
Call to undefined method PHPCompiler\VM\HashTable::shiftfirst()
It was passing in this session's earlier baseline (it is absent from the 15-case failing set
recorded in test/differential/AOT-BASELINE.md at 96ddddeb1).
Same shape as #23974
That issue covered HashTable::sliceCopy() and HashTable::iterateKeyed() reported as undefined
while existing in lib/VM/HashTable.php — a registration problem, not missing implementations.
Worth checking whether shiftFirst is the same:
grep -n 'function shiftFirst' lib/VM/HashTable.phpIf it exists there, this is the same defect resurfacing on a third method, and #24013's fix for
#23974 may have covered only the two methods named in that issue rather than the underlying
registration path.
Note the name arrives lowercased (shiftfirst) — as established on #23974, that is the error
path's display fallback ($class->methodNames[$methodLc] ?? $methodLc), not evidence of a
case-normalisation bug.
Candidate range
Between the two measurements: 96ddddeb1..556c97d1d.
Environment: php-compiler:22.04-dev, PHP 8.2.32, LLVM 9, master 556c97d1d.
Activity
PurHur commentedon Jul 27, 2026
claim: worker-lane-B — starting this run
Gate snapshot (pre-claim):
PurHur commentedon Jul 28, 2026
Confirmed: the method exists — this is registration, same as #23974
Present and public. So
HashTable::shiftfirst()being "undefined" at build time is the thirdinstance of the pattern from #23974, where
sliceCopy()anditerateKeyed()were likewise presentin source but unresolved for AOT.
That makes it likely #24013's fix addressed the two methods named in #23974 rather than whatever
registration path lets a
HashTablemethod be resolved at all. Worth checking how those two weremade visible and whether the same treatment generalises — otherwise the next
HashTablemethod acase happens to reach will fail the same way.
The lowercased name (
shiftfirst) is, as established on #23974, the error path's display fallback(
$class->methodNames[$methodLc] ?? $methodLc) rather than a case-normalisation bug — mentioningit again only so nobody re-chases it.
Baseline: regenerate on the 75-case corpus, and correct my own "silen…
PurHur commentedon Jul 28, 2026
claim: worker-lane-B — continuing (no PR yet; implementing NestedJIT
HashTable::shiftFirstregistration, same pattern as #23974sliceCopy/iterateKeyed)Gate snapshot:
user_release_ready: yes,north-star5-fast: ok, inventoryOK 6598/6598PurHur commentedon Jul 28, 2026
claim: worker-lane-B — continuing this run (workspace already on
agent/issue-24025-hashtable-shiftfirst)Gate snapshot:
user_release_ready: yes,north-star5-fast: ok, inventoryOK 6598/6598shiftfirst→HashTableShiftFirst, ArrayShiftJitHelper empty via getNumElements, spine require + unit asserte30_array_lit_dim_assign_shift+ unit testAOT: NestedJIT HashTable::shiftFirst via HashTableShiftLlvm (#24025) (#…
PurHur commentedon Jul 28, 2026
Merged #24046
Fixed: AOT compile of
e30_array_lit_dim_assign_shift—HashTable::shiftFirstNestedJIT registration +HashTableShiftLlvminline emit (wasCall to undefined method …::shiftfirst()).Verified: e30 compiles; empty/non-empty
array_shiftAOT matches Zend; VM e30 differential OK; unit filters green; inventory 6600/6600.Still open for full e30 AOT differential: master
json_encodeAOT segfaults even onjson_encode(1); array-literal capture of$aafter in-literal mutations can still be empty (#23990 was VM-only). Separate from the shiftFirst resolution gap.PurHur commentedon Jul 28, 2026
Lane B note: #24025 closed by #24046 before this run's PR (#24048) could merge — closed #24048 as duplicate.
Gate snapshot from this run (before noticing #24046):
user_release_ready: yes,north-star5-fast: ok, inventory OKHashTableShiftFirstWIP; semantic shift withoutjson_encodematched Zend (99then2,3)json_encode(1)AOT segfaults on master (independent; not this issue)PurHur commentedon Jul 28, 2026
Confirmed fixed —
shiftFirstresolves nowe30_array_lit_dim_assign_shift.phpcompiles onf51d0b180. TheCall to undefined method PHPCompiler\VM\HashTable::shiftfirst()error is gone, so #24046 resolvedwhat this issue reported.
The case still fails, at a later stage
It now segfaults at runtime:
Deterministic, not intermittent. Filing separately rather than reopening — the registration defect
this issue is about is genuinely fixed, and a runtime segfault in the same case is a different
problem that would only be confused by sharing a title.
This is the fourth time today a fix in this area has moved a failure to a later stage rather than
eliminating it (#23792, #23815, #23911, and now this). Not a criticism — each fix was correct and
exposed the next link. Worth knowing when reading "case still red" that it may mean progress.
AOT: array_sum via caller-frame ArraySumLlvm (#24167) (#24228)
Fix AOT array_first/array_last via HashTableElemLlvm thin path (#27596)…