Fix problems with TypeVars in intersection and type_match_ #11194
Conversation
timholy
changed the title
|
The second commit fixes #8652 |
|
...whoops, no it doesn't (or rather, it breaks other stuff). It made it through I'll discuss more in #8652. |
timholy
force-pushed
the
teh/tvar_intersect
branch
from
deb0fb2 to
2e26314
Compare
|
OK, this one should work (it passes locally). Interestingly, it also fixes #8045. |
|
Hmm, the PR (not the push) has now failed twice on the |
|
It is possible that there's a difference — the PR build merges the branch into master. You could also try rebasing to see if the push then fails, too. |
|
I just did an up-to-the-minute rebase, and it passed locally (but this was 64-bit, which also passed on Travis). |
|
The CI problems seem to have been fixed. Nice! |
|
Of my various open PRs, now that Jeff seems back (already!) I'm going to bump this particular one. It's short, I think it's one of the last barriers to merging static compilation, and Jeff seems likely to be the only suitable reviewer. |
timholy
force-pushed
the
teh/tvar_intersect
branch
from
57a47b3 to
f9d0d83
Compare
|
Hah, but I just discovered that a small addition also fixes #11367. |
|
Thanks @timholy, amazing work, as always! |
| if (jl_is_typevar(b)) { | ||
| for(i=0; i < penv->n; i+=2) { | ||
| if (penv->data[i] == b && !jl_is_typevar(penv->data[i+1])) { | ||
| if (jl_types_equal((jl_value_t*)a, penv->data[i+1])) { |
There was a problem hiding this comment.
This seems very odd to me, since a is a typevar. I would think jl_types_equal would always return false in this case.
There was a problem hiding this comment.
You're right, it probably can be deleted. I'll check it out.
|
Putting us all to shame --- just try to find a neuroscientist who knows as much about type systems! :) |
|
I'm just trying to keep up with @simonster in doing our clan proud. Besides, #11242 was type-system boot camp, in every respect except shaving my head (which is good, since there is so little to shave). |
…8652) In jl_args_morespecific(a,b), one of the important tests is to call type_match_(b,a), essentially reasoning that if b <: a then a cannot be more specific than b. The problem is that this test can fail for an independent reason, like invariance. This takes the drastic step of keeping invariance=false for the type_match_ check in jl_args_morespecific. Surprisingly, this doesn't appear to break anything.
timholy
force-pushed
the
teh/tvar_intersect
branch
from
f9d0d83 to
65cc7f3
Compare
|
Right as usual, Jeff. And thanks for the speedy review! |
Fix problems with TypeVars in intersection and type_match_
|
👏 |
Fixes #11136. Don't merge until @JeffBezanson gets enough free time to look this over.