1. Brocki2301
  2. pypy

Commits

Show all
Author Commit Message Date Builds
Armin Rigo
shadowstack support in CALL_RELEASE_GIL. Much easier than asmgcc. Still a bit hard to test correctly at this level...
Branches
jitypes2
Armin Rigo
Double this test, once with asmgcc and once with shadowstack.
Branches
jitypes2
Antonio Cuni
hg merge default
Branches
jitypes2
Antonio Cuni
(antocuni, arigo): merge the invalidate-virtualrefs branch; this makes virtualrefs *less* powerful, because they can no longer be forced after the call to virtual_ref_finish (unless they were already forced before), but at the same time it lets the JIT to optimize them better; in particular, we can avoid to allocate the frames after a call whose effectinfo is none
Antonio Cuni
close about-to-be-merged branch
Antonio Cuni
revert 75c1ede14db1, tests pass
Antonio Cuni
temporarily run only these tests, trying to understand why they failed
Armin Rigo
Speed up a little bit str.split('char'). Unsure why CPython is so much faster at doing this on a big string (benchmark: a 60MB string with 3 million times the separator character).
Benjamin Peterson
detect future features after source is in utf-8 This should fix #732.
Armin Rigo
Test and (bah) fix.
Armin Rigo
merge heads
Armin Rigo
Call _preprocess_libary_dirs() from here too.
Alex Gaynor
Mark a function in the tests as dont_look_inside so it works with the new stricter requirements. Also remove it from a test that marked every function as don't look inside to stree the JIT.
Alex Gaynor
Show the function that can't be access direct.
Alex Gaynor
Merged upstream.
Alex Gaynor
We shipped 1.5 already, this file is basically completely out of date.
Benjamin Peterson
reject graphs with access_directly that we no longer want to look inside
Benjamin Peterson
some rarely used frame methods which we don't have to look inside
Benjamin Peterson
BUILD_SET is @unroll_safe
Armin Rigo
merge heads
Armin Rigo
Found where the #includes come from. Remove them; they belong to debug_print.c. Give up writing a test, given that on x86 and x86_64 we use a custom version of READ_TIMESTAMP anyway.
Armin Rigo
A missing #include means that CLOCK_THREAD_CPUTIME_ID is not defined, even on Linux. Bad fijal, no cookie. But it's true that it's hard to test: both versions give numbers that are reasonable. I'll try to come up with a test anyway...
Benjamin Peterson
fix compiling an empty module
Armin Rigo
A quick hack as a workaround for the current (hard) issue: introduce "stack-critical code", enabled with a pair of macros from rlib.rstack. When running such a "stack-critical" piece of code, StackOverflows are never raised. Used to protect key pieces of the reconstruction of data from a guard failure, where it is essential that vrefinfo.continue_tracing() is called on all virtualrefs.
Armin Rigo
hg merge default
Armin Rigo
Kill old unused function.
Armin Rigo
merge heads
Armin Rigo
When compiling with "make lldebug", increase the number of RPython traceback entries recorded.
Armin Rigo
merge heads
Armin Rigo
Fix the comment.
  1. Prev
  2. Next