Commits

Author Commit Message Labels Comments Date
Hakan Ardo
In case of conflicts, prioritize among the potential short ops that can produce one specific box instead of duplicating recursivly every op that uses that box as an argument. The priority order is: - ops found in the original trace - synthetic ops (setfields converted to getfields) - inputargs - potential ops that was never promoted to short_boxes This makes the effect of the optimizations less random and should always remove loop invariant ops. Non loop invar…
Branches
jit-duplicated_short_boxes
Armin Rigo
merge heads
Armin Rigo
Rename this file. (For now it's always skipped.)
Alex Gaynor
merged upstream
Alex Gaynor
Fix a crash (escaped ValueError) in various division methods on floats with an inf LHS.
Armin Rigo
Better order for this. Corner-casish and not important.
Armin Rigo
Fix a corner case.
Armin Rigo
Don't use shutil and don't remove all files within __pycache__. Instead only remove .pyc/.pyo files from these __pycache__ directories, and only kill the directory if it is empty afterwards (common case).
Armin Rigo
Rewrite py.cleanup to also remove __pycache__ directories, created by CPython 3.x.
Armin Rigo
Reorganization. Probably more jit-friendly.
Armin Rigo
Rename these two methods to their more precise meaning.
Armin Rigo
Split the class WeakrefLifeline in a parent class without __del__ and a subclass with __del__, that is used if we create weakrefs with callbacks only.
Armin Rigo
Test and fix for issue852.
Armin Rigo
Add the type for c_bool here.
Armin Rigo
Test and fix for issue856.
Armin Rigo
A passing test.
Armin Rigo
Fix for an obscure case in which prejit_optimizations() renames some variables in the jit_merge_point, after which the original op cannot be found in the graph any more.
Alex Gaynor
remove completely misnamed method after the previous commit.
Alex Gaynor
remove unnecsary use of weakrefs from our sqlite3. This is worth about 25% on the sqlite benchmark.
Alex Gaynor
allow inlining into more of the thread library, specifically for get_ident, which used to be rendered as a call_may_force, and is now properly elided. Seems to be worth about 5% on the sqlite benchmark.
Justin Peel
fix test_sin to deal with Float32's decreased precision
Branches
unsigned-dtypes
Justin Peel
(alex gaynor) add a proper Float32 dtype and drop the fake Float96 dtype.
Branches
unsigned-dtypes
Justin Peel
force _trace_drag_out to always be inlined. The function is only called by 2-3 other functions. This change really helps with gc-intensive scripts like using a dict as a counter (can be a 10% speed-up).
Alex Gaynor
Fix a test I broke, by better optimizing things :)
Antonio Cuni
cache the _ffiargtype; this saves a dict lookup in hot loops
Armin Rigo
merge heads
Armin Rigo
Test and fix (baaa*a*a*ah, took too long): IR_QUASIIMMUTABLE fields must of course not be considered as fully immutable fields!
Maciej Fijalkowski
minor improvements
Carl Friedrich Bolz
add some helpful comments
Branches
improve-heap-caching-tracing
Carl Friedrich Bolz
add the knowledge that two boxes with that were the result of a "new" operation can never alias each other.
Branches
improve-heap-caching-tracing
  1. Prev
  2. Next