Commits

Show all
Author Commit Message Labels Comments Date
Lukas Diekmann
quicksort turned out to be not that good. instead we use a special TimSort implementation for ints
Branches
list-strategies
Lukas Diekmann
more test coverage
Branches
list-strategies
Carl Friedrich Bolz
that means we can now use the split in rlib.rstring
Branches
list-strategies
Carl Friedrich Bolz
add a space.newlist_str method that can be used to wrap a list of unwrapped strings. implement a fast path in the std object space.
Branches
list-strategies
Carl Friedrich Bolz
add a new space method listview_str which returns None except if the argument is a list using the string strategy. then return an unwrapped list of unwrapped strings. use this in str.join.
Branches
list-strategies
Carl Friedrich Bolz
clean up getitem variants a bit
Branches
list-strategies
Carl Friedrich Bolz
unroll getitem as well, if the list has a constant size.
Branches
list-strategies
Carl Friedrich Bolz
making an empty wrapped list takes a lot less allocations now.
Branches
list-strategies
Carl Friedrich Bolz
don't rely on the file encoding, which seems to be brittle in some situations.
Branches
list-strategies
Carl Friedrich Bolz
The ticker does not necessarily decrease by one. it depends on the trace length.
Branches
list-strategies
Carl Friedrich Bolz
for some reason some of the list fields are not cached across the loop. is this due to the opaque pointer restriction of the short preambles?
Branches
list-strategies
Carl Friedrich Bolz
slightly annoying: cast_opaque_ptr killed the tracing heap cache. I cannot just remove it, seems the short preamble stuff seems to need it. Thus I need to consider equivalences of boxes in the heapcache.
Branches
list-strategies
Carl Friedrich Bolz
merge default
Branches
list-strategies
Carl Friedrich Bolz
comment
Branches
list-strategies
Carl Friedrich Bolz
specialization always gives this functions different names, so just use "_" here.
Branches
list-strategies
Armin Rigo
Fix test.
Alex Gaynor
appears to fix the rlib test_jit test, might just be a phase of the moon thing :/
Alex Gaynor
simplify some code
Alex Gaynor
added numpy.e
Alex Gaynor
added numpy.arctanh.
Justin Peel
merge unsigned-dtypes branch
Justin Peel
merge in default
Branches
unsigned-dtypes
Hakan Ardo
Merging jit-optimizeopt-cleanups. It makes the the different Optimization stages of optimizeopt more stricly follow the model of emitting operations to the next stage instead of manipulating Optimizer.newoperations directly. This means fewer optimizations are prevented due to unwanted interactions between the different stages (for example test_remove_duplicate_pure_op_ovf_with_lazy_setfield). Operations produced by forcing boxes are emitted by the stage doing the forcing. Two new Optimization stages have been intorduced, OptPure and OptEarlyForce. OptEarlyForce forces things early in the chain and thus allows the following stages to work on the ops produced (moste notable the OptHeap and OptPure stages). Operations produced as an effect of forincing a lazy setfiled are still not seen …
Hakan Ardo
closing branch for merge
Branches
jit-optimizeopt-cleanups
Hakan Ardo
hg merge default
Branches
jit-optimizeopt-cleanups
Justin Peel
merge in default
Branches
unsigned-dtypes
Justin Peel
remove iter stuff that I hadn't intended to merge
Branches
unsigned-dtypes
Justin Peel
remove some commented code and skip a zjit test until we change llimpl
Branches
unsigned-dtypes
Armin Rigo
Merge gil-improvement: improve the balance of threads getting the GIL, notably when there is a mixture of CPU-intensive and I/O threads. Previously, any CPU-intensive thread would release and immediately re-acquire the GIL every 100 bytecodes; but most of the time on multi-CPU machines this didn't actually gave other threads a chance to run. And so when an I/O thread did any operation, chances were that they would loose control and not regain it for quite a long time. Now, the yield-after-N-bytecodes step is different from the re…
Armin Rigo
Tweaks.
Branches
gil-improvement
  1. Prev
  2. Next