JuliaLang / julia / 275
78%

DEFAULT BRANCH: master
Ran
Jobs 0
Files 0
Run time –
Badge
Embed ▾
README BADGES
x

If you need to use a raster PNG badge, change the '.svg' to '.png' in the link

Markdown

Textile

RDoc

HTML

Rst
275

push

buildkite

web-flow 
support: range containment is (start, end], not [start, end) (#62703)

🤖 #62316 re-encoded the Eytzinger image-range tree to textbook half-open
[start, end) containment, reasoning that images have padding around them
so boundary queries would not occur. That assumption fails for
heap-image blobs smaller than GC_PERM_POOL_LIMIT, which are carved from
the same perm bump pool as runtime permobjs: a zero-size singleton's
value pointer is one-past its own tag word — exactly the pool's next
bump address — so a small blob loaded right after such a singleton is
registered starting at the singleton's address. Under start-inclusive
containment the tree then claims the heap singleton is in-image while
its header bit correctly says otherwise, tripping the jl_object_in_image
consistency assertion (deterministically reproducible via ArDCA's
dependency stack; observed as Static.StaticInt in PkgEval, fixes #62521)
— and silently corrupting serialization decisions on non-assert builds.
The mirror case is equally real: a trailing zero-size singleton's value
pointer equals its blob's end, which start-inclusive/end-exclusive
containment reports out-of-image.

Object pointers can never equal their blob's base (a blob begins with
its first object's tag word) and can equal its end, so the correct
convention is start-exclusive, end-inclusive — the one staticdata.c
itself asserts throughout (base < obj && obj <= base + size). Query with
addr + 1 instead of addr + 3: the strict-predecessor search then answers
(start, end], and at abutting ranges the shared address belongs to the
earlier range, matching the trailing-singleton case.

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Source Files on build 275
Detailed source file information is not available for this build.