What’s new in PyPy 4.0.0¶
Improve the performance of
dict.update() and a bunch of methods from
sets, by reusing the hash value stored in one dict when inspecting
or changing another dict with that key.
A major refactoring of the
ResOperations that kills Box. Also rewrote
unrolling to enable future enhancements. Should improve warmup time
by 20% or so.
Optimize common sequences of operations like
int_lt/cond_call in the JIT backends
Fix for missing headers in OpenBSD, already applied in downstream ports
Remove a source of non-incremental-ness in the GC: now
external_malloc() no longer runs
gc_step_until() any more. If there
is a currently-running major collection, we do only so many steps
before returning. This number of steps depends on the size of the
allocated object. It is controlled by tracking the general progress
of these major collection steps and the size of old objects that
keep adding up between them.
Issue #2141: fix a crash on Windows and OS/X and ARM when running at least 20 threads.
Add support for ndarray.ctypes property.
Share guard resume data between consecutive guards that have only pure operations and guards in between.
Fix performance regression on operations mixing numpy scalars and Python floats, cf. issue #2148.
__stdcall in CFFI.
Refactorings of annotation and rtyping of function calls.
Allow creation of fortran-ordered ndarrays
Remove some remnants of the old
ffi.new_handle() returns handles that work more like CPython’s: they
remain valid as long as the target exists (unlike the previous
version, where handles become invalid before the
__del__ is called).
allow automatic casting in ufuncs (and
frompypyfunc) to cast the
arguments to the allowed function type declarations, fixes various
failures in linalg CFFI functions
A new optimization pass to use emit vectorized loops
The PowerPC JIT backend is merged.
Improve the performance of simple trace functions by lazily calling
locals2fast only if
f_locals is actually accessed.