SplatNodes within ArrayNodes (e.g. [*1..2, 3]) need to be special
cased in the compiler because they use a combination of concatarray
and newarray instructions to treat each sequence of splat or non-splat
elements as independent arrays which get concatenated. This commit
implements those cases.
This entirely changes how it is tested. Rather than to use counters
we now record the timeline of events with associated threads which
makes it much easier to assert that certains events are only preceded
by a specific event, and makes it much easier to debug unexpected
timelines.
Co-Authored-By: Étienne Barrié <etienne.barrie@gmail.com>
Co-Authored-By: JP Camara <jp@jpcamara.com>
Co-Authored-By: John Hawthorn <john@hawthorn.email>
There have been some sproradically flaky tests related to GC compaction,
which fail with:
1) Failure:
TestGCCompact#test_moving_hashes_down_size_pools [/test/ruby/test_gc_compact.rb:442]:
Expected 499 to be >= 500.
What's happening here, is that, _sometimes_, depending on very unlucky
combinations of machine things, one of the expected-to-be-moved hashes
might be found on the machine stack during GC, and thus pinned.
One factor which seems to make this _more_ likely is that GCC 11 on
Ubuntu 22.04 seems to want to allocate 440 bytes of stack space for
`gc_start`, which is much more than it actually uses on the common code
path. The result is that there are some 50-odd VALUE-sized cells "live"
on the stack which may well contain valid heap pointers from previous
function calls, and will need to be pinned.
This is, of course, totally normal and expected; Ruby's GC is
conservative and if there is the possibility that a VALUE might be live
on the machine stack, it can't be moved. However, it does make these
tests flaky.
This commit "fixes" the tests by performing the work in a fiber; the
fiber goes out of scope and should be collected by the call to
verify_compaction_references, so there should be no references to the
to-be-moved objects floating around on the machine stack.
Fixes [#20021]
* Before this it was compiled but not used, because TruffleRuby has
a stringio.rb in stdlib and .rb has precedence over .so.
In fact that extension never worked on TruffleRuby,
because rb_io_extract_modeenc() has never been defined on TruffleRuby.
* So this just skip compiling the extension since compilation of it now fails:
https://github.com/ruby/openssl/issues/699https://github.com/ruby/stringio/commit/d791b63df6
This changes the CompactIndexClient to store etags received from the
compact index in separate files rather than relying on the MD5 checksum
of the file as the etag.
Smoothes the upgrade from md5 etags to opaque by generating them when no
etag file exists. This should reduce the initial impact of changing the
caching behavior by reducing cache misses when the MD5 etag is the same.
Eventually, the MD5 behavior should be retired and the etag should be
considered completely opaque with no assumption that MD5 would match.
```
==> memprof.after.txt <==
Total allocated: 1.13 MB (2352 objects)
Total retained: 10.08 kB (78 objects)
==> memprof.before.txt <==
Total allocated: 46.27 MB (38439 objects)
Total retained: 9.94 kB (75 objects)
```
Yes, we were allocating 45MB of arrays in `dependencies_installed?`,
it was accidentally cubic.
https://github.com/rubygems/rubygems/commit/13ab874388
- Unless `sizeof(BDIGIT) == 4`, (8-byte integer not available), the
size to be loaded was wrong.
- Since `BDIGIT`s are dumped as raw binary, the loaded byte order was
inverted unless little-endian.
