Fix gc_verify_internal_consistency error for pattern_matching in ripper #7627
Conversation
`gc_verify_internal_consistency` reports "found internal inconsistency" for "test_pattern_matching.rb". http://ci.rvm.jp/results/trunk-gc-asserts@ruby-sp2-docker/4501173 Ruby's parser manages objects by two different ways. 1. For parser * markable node holds objects * call `RB_OBJ_WRITTEN` with `p->ast` as parent * `mark_ast_value` marks objects 2. For ripper * unmarkable node, NODE_RIPPER/NODE_CDECL, holds objects * call `rb_ast_add_mark_object`. This function calls `rb_hash_aset` then `RB_OBJ_WRITTEN` is called with `mark_hash` as parent * `mark_hash` marks objects However in current pattern_matching implementation * markable node holds objects * call `rb_ast_add_mark_object` This commit fix it to be #2. This was inconsistency however always `mark_hash` is made young by `rb_ast_add_mark_object` call then objects are not collected.
There was a problem hiding this comment.
This was inconsistency however always mark_hash is made young by rb_ast_add_mark_object call then objects are not collected.
Do I understand correctly that it was a write barrier timing problem?rb_ast_add_mark_object can yield to the GC due to allocation, and previously, when it did yield, the GC observed new out edges from imemo_ast added without write barriers.
|
I don't think it's a write barrier timing problem. |
|
Ah, so previously, the references were added to |
|
Previously, the references were added to both (1) |
|
Thank you for patiently helping me understand! |
gc_verify_internal_consistencyreports "found internal inconsistency" for "test_pattern_matching.rb".http://ci.rvm.jp/results/trunk-gc-asserts@ruby-sp2-docker/4501173
Ruby's parser manages objects by two different ways.
RB_OBJ_WRITTENwithp->astas parentmark_ast_valuemarks objectsrb_ast_add_mark_object. This function callsrb_hash_asetthenRB_OBJ_WRITTENis called withmark_hashas parentmark_hashmarks objectsHowever in current pattern_matching implementation
rb_ast_add_mark_objectThis commit fix it to be #2.
This was inconsistency however always
mark_hashis made young byrb_ast_add_mark_objectcall then objects are not collected.