Bind in-place update operators (e.g. +=) in Python #8683
Conversation
alexreinking
added
release_notes
alexreinking
changed the title
|
Converting to draft as my attempted workaround for the double-updates issue broke something else. Python |
|
Looks like upstream LLVM broke... something. Seeing the same errors in #8684 |
|
During the dev meeting, we discussed an alternative implementation strategy where |
alexreinking
force-pushed
the
python/implicit-pure-def
branch
from
d54ff7d to
f0f7baf
Compare
|
I attempted the alternate implementation and it seems to work, perhaps even more cleanly than before! |
|
MSVC doesn't like a template... diagnosing |
alexreinking
force-pushed
the
python/implicit-pure-def
branch
from
85a8132 to
d2be2ee
Compare
src/Func.h
Outdated
| } | ||
|
|
||
| /** What are the arguments to the function? */ | ||
| const std::vector<Expr> &get_args() const { |
There was a problem hiding this comment.
Two thoughts:
- Our getters don't normally have
get_ - This should be explicit about not expanding implicit vars. Here's a test case:
f[_] += g[x, y]I think it probably works because_will be unexpanded on both sides for the check.
So maybe rename the member to args_, the method to args(), and add a note to the comment that any implicit vars aren't expanded? Or maybe it would be better to rename the method to unexpanded_args, or raw_args or something like that. There should be something to make sure callers stop and think about implicit vars for a second.
There was a problem hiding this comment.
Yeah, the get_ prefix was to avoid renaming the member. If I use the raw_args name, I should kill two birds with one stone. I'll add the test case, too.
This enables inference of units (and types) for pure definitions in reduction-y update stages.
Fixes #8661