0.5: Type too large error #15464
Comments
|
This appears to be fixed (at least in the development branch). |
|
Would be great to get a reduced test case if at all possible, so we can make sure it stays fixed. |
|
The original bug was only triggered when running all tests in a row: running even just the line in question wasn't sufficient. So I'm not sure a reduced test case is realistic. I didn't check whether the bug is indeed fixed on the branch above. But new deprecations would make that time consuming. |
|
The test failure was fixed by e529058 But the bug is a tvar construction error, which results in roughly trying to compute the following intersection: julia> type RealBasis end; type IntervalDomain end; type SumSpace{a,b,c,d} end;
julia> a=Tuple{typeof(==), Type{TypeVar(:_, SumSpace)}, Type{TypeVar(:_, SumSpace{TypeVar(:_, TypeVar(:SV, Any)), RealBasis, TypeVar(:DD, IntervalDomain, true), 1})}}
ERROR: TypeError: typeassert: expected Type{T}, got TypeVar
julia> b=Tuple{typeof(==), Type{TypeVar(:T, true)}, Type{TypeVar(:T, true)}}
julia> typeintersect(a,b)
ERROR: type too largeAFAICT, there's only one place in the code where this type could have been constructed, requiring the following patch to fix this issue: diff --git a/src/jltypes.c b/src/jltypes.c
index 95460b2..33c3f69 100644
--- a/src/jltypes.c
+++ b/src/jltypes.c
@@ -796,7 +796,7 @@ static jl_value_t *approxify_type(jl_datatype_t *dt, jl_svec_t *pp)
JL_GC_PUSH1(&p);
for(i=0; i < l; i++) {
jl_value_t *el = jl_svecref(dt->parameters, i);
- if (jl_has_typevars_from(el, pp))
+ if (!jl_is_typevar(el) && jl_has_typevars_from(el, pp))
jl_svecset(p, i, jl_new_typevar(underscore_sym, jl_bottom_type, el));
else
jl_svecset(p, i, el);@JeffBezanson @carnaval do you know how to trigger this directly to make a test? |
|
bump. @JeffBezanson can you write a test for this? |
|
bump again. |
StefanKarpinski
added
help wanted
|
@dlfivefifty Worth trying this again now that #18457 is merged. Will be very interesting to see what happens. The |
|
@JeffBezanson the branch that triggered the bug won't work in 0.6, so I won't be able to test it, unfortunately. |
|
Seems to be fixed for practical purposes, and current subtype and intersection issues are tracked elsewhere. |
I'm trying to get ApproxFun working in 0.5, but the tests fail due to a "type too large" error, see error dump below. This only occurs when I run all the tests at once: if I run just the line that throughs the error:
I don't get the error message, though the compile time seems very long compared to 0.4.
You can reproduce it with:
My
versioninfoisHere is the error dump:
The text was updated successfully, but these errors were encountered: