improve inferability of base #22019
improve inferability of base #22019
Conversation
|
Is this really safe to backport to 0.6? It seems awfully large and potentially dangerous for that. It would have to be fixing some really bad issue in order to warrant that backport. |
|
Its somewhat large, but mostly mechanical. I probably wouldn't bother if we already had a release, but I think it's useful to keep master closer during the RC period. There's a few bug fix commits in here too (1, 3, and 7) that should be backported regardless. |
|
sysimg also seems to be about a megabyte smaller too. |
|
As always, can we have just the bug fixes in their own PRs please? Those can and should be merged as soon as they pass CI. The rest is less clear. |
|
This is the bugfix PR (feature PR is #21933). It's just not quite as essential that we backport the correctness fixes as the assertion fixes. |
|
What bugs are being fixed? Where are the tests? |
base/abstractarray.jl
Outdated
| _cshp(tail(dims), (out..., next), tail(shape), tail(nshape)) | ||
| @inline function _cshp(ndim::Int, dims, shape, nshape) | ||
| a = shape[1] | ||
| b = nshape[1] |
|
The fixes here are to ensure that inference will always eventually terminate when any of these methods enter its queue. Tests are PR #21933 which addresses that bug more broadly. |
would segfault when reaching the jl_datatype_nfields call
TODO: this is helping to avoid a type-system bug mis-computing sparams during intersection, but that can already cause significant problems elsewhere too
…to be non-terminating this same issue applied to several other similar functions similarly, need to avoid nesting the use of functional code (like map) too deeply inside the array code to avoid the appearance of indeterminate recursion in inference
vtjnash
force-pushed
the
jn/inferrable-functions
branch
from
736fa71 to
4287761
Compare
|
@nanosoldier |
|
Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. cc @jrevels |
There is exactly one line of tests changed in https://github.com/JuliaLang/julia/pull/21933/files. I'm against backporting this without any tests. |
|
This broke |
|
cf0c6a1 fixes an issue for me and is trivial to backport, would like to see that in a point release. |
|
can you submit a unit test that it fixes? |
Avoid growing an input parameter during recursion. In general, permitting that would require that inference be potentially non-terminating (as proving that it will converge requires solving the halting problem). However, the methods in base using this pattern only actually requires the simpler pattern of tail-recursing on some element and concatenating the output.
(This is the learnings of #21933 without the policy change.)