Objmode with-block

[sklam]

(Based on #3017)

• can't assign anything yet fixed
• can't return anything yet fixed
• pass thru objects deferred to Pass thru pyobjects #3282

[codecov-io]

Codecov Report

Merging #3166 into master will increase coverage by 0.07%.
The diff coverage is 89%.

@@            Coverage Diff             @@
##           master    #3166      +/-   ##
==========================================
+ Coverage   81.09%   81.17%   +0.07%     
==========================================
  Files         386      388       +2     
  Lines       76285    76921     +636     
  Branches     8575     8631      +56     
==========================================
+ Hits        61867    62440     +573     
- Misses      13100    13155      +55     
- Partials     1318     1326       +8

[sklam]

Merged #3230 into this patch.

[sklam]

Feature is completed. Ready for early review. I will just be adding more tests for crazy cases; like looplifting calling into objmode and vice versa.

[ehsantn]

@sklam this works great! A couple of minor issues:

• Could you add developer docs for this feature's design? Also, more comments throughout the code could help.
• Using numba.withcontexts.objmode_context doesn't work since get_ctxmgr_obj expects a Global.
• If a variable's type is not defined by mistake, this throws a KeyError instead of proper error message.
• Can objmode_context be exported at the top level (i.e. numba.objmode_context)?

[ehsantn]

@DrTodd13 @fschlimb could you take a look at this feature as well (since it's a critical feature)?

[ehsantn]

@sklam with this feature in place, I'm wondering if it is now easier to just declare a python function as object mode inside all njit functions. The user just provides type inference and Numba handles boxing/unboxing automatically. For example:

@infer_object_mode(foo)
class FooInfer(AbstractTemplate):
    def generic(self, args, kws):
        return signature(types.float64, *args)

[sklam]

Notes on new changes:

• objmode_context renamed to objmode; exposed to top-level ns.
  no need to worry about the "objmode" target because the objmode context is marking a region for that target.
• numba/objmode.py is not numba/pylowering.py (which contains class PyLower) to avoid name collision.
• @stuartarchibald 's torture tests are included. Some are addressed/fixed and some are too hard to support; i.e. list and function types are not allowed to go into a objmode context. For the closure-using-objmode one, I have to mark it as expected-failure as there are no ways to handle it.
• new docs. most of the docs are autogen from docstring in help(numba.objmode).

[sklam]

@ehsantn , regarding #3166 (comment), it's definitely easier now. I think one can also just use @overload; e.g.:

@overload(mytoy)
def mytoy_ov(x, y):
    def impl():
        with objmode(out=...):
               out = mytoy(x, y)
        return out
    return impl

[ehsantn]

Interesting solution. Makes sense.

[seibert]

need to indent this line so that sphinx doesn't think the list has ended

[seibert]

same need for indent here

[seibert]

Current failures:

• Minor sphinx rendering warning in docstring that needs to be fixed
• Crash bug on 32-bit Linux during testing. Not sure about cause...