path-type allow non-subtype struct acc #10
Conversation
|
Can we make a test for this? Also, can you use #:when instead of nevermind? |
|
Okay, this initial fix is not right (I'll update the pull req in a sec if my next approach passes all the TR tests). Here are two files which together produce the error when you run main.rkt: Someone who is better at modules could probably turn this into a test right quick. I can probably figure out how to with some more tinkering... Basically, this code (which is a simpler version of the funkytown code) involves a module ("foo.rkt") which exports a macro ("in-foo") which references a struct ("foo") defined in that module also, however that struct ("foo") is not exported. When another module ("main.rkt") imports and uses that macro ("in-foo"), it is typechecking code which is doing type-tests and unsafe-struct accesses on a struct it doesn't even have a type for. This would normally error with a type checking error (user level) -- let aliasing got involved because it made some aliases to struct-path objects before the "foo is an unbound type" errors kicked in and instead it hit that internal error (which was annoying! Why didn't that "foo unbound" error get thrown sooner? Then this (probably) wouldn't have been an issue at all!). |
proper path-type mismatch error
|
So, because the internal error in path-type was getting called before the proper typechecking errors were being reported, I tried returning the "Error" type if path-type is unable to extract a type -- this worked, causing all the code which was triggering the internal error to instead give the proper user-level typechecking errors. All TR test cases are passing. |
path-type now allows normal internal errors
fixes odd corner case internal errors where an alias's path doesn't check as a subtype of the type of the variable.
For example error, download tar:
https://www.dropbox.com/s/m0lj8n82z2xwogr/funky-typed.tar?dl=0
untar
racket funky-typed/synth/synth.rkt
With this commit, the internal error this currently triggers disappears and instead typechecking errors with a type mismatch on Array (this file does some non-standard usage of some of the private internals of the math/array library that doesn't work out). So both break, but this commit ensures it is a user level error for this case (i.e. a readable typechecker error) instead of a misleading internal path-mismatch error.