Pat Hickey (Jan 14 2020 at 18:44): Pat Hickey (Jan 14 2020 at 18:44):so, just before break my witx-test tool (a proptest generator for witx specs, plus a driver to use wasi-headers to generate a header and validate the header with clang) found some parts of the witx grammar that wasi-headers couldn't handle correctly
Pat Hickey (Jan 14 2020 at 18:45):most of it is easy to fix, the tough part to crack is that the representation of witx arrays in C is tricky/inconstient
Pat Hickey (Jan 14 2020 at 18:46):we currently don't have any way to deal with an array as a member of a struct, or a variant in a union
Pat Hickey (Jan 14 2020 at 18:49):i think the correct way to deal with arrays in either of those cases is to define a struct array_$typename { $typename *ptr; size_t len; };
Pat Hickey (Jan 14 2020 at 18:50):this is going to complicate the one-pass nature of the tool, but i dont think theres any other way to handle it correctly. does anyone have objections?
Pat Hickey (Jan 14 2020 at 18:53):along those same lines - right now the wasi-headers tool hard codes its input files and output files. im going to extend the main to take an optional input and output path - if both are blank it will do what it does now, which is use the latest snapshot and output to the right spot in the libc. but in general i think its a pretty useful tool to use elsewhere
Dan Gohman (Jan 14 2020 at 18:55):One option is just to return an error; WASI itself has no arrays inside of structs or unions
Pat Hickey (Jan 14 2020 at 18:56):id argue that it already does have arrays inside structs, $iovec and $ciovec
Pat Hickey (Jan 14 2020 at 18:56):which are a pair of pointer and len
Pat Hickey (Jan 14 2020 at 18:58):those should be arrays, i think, in order to get rid of pointers in witx. i dont want to figure out how to get rid of pointers today, but i want to eventaully :)
Pat Hickey (Jan 14 2020 at 18:58):(reason being that pointers arent an interface type, but arrays are)
Pat Hickey (Jan 14 2020 at 18:59):also, i just realized this issue would happen if you ever put a string inside a struct or union as well
Dan Gohman (Jan 14 2020 at 19:00):Yeah, that does make sense. I guess I'm just feeling a little conservative about supporting "arrays" which become pointers with not-entirely-obvious lifetimes outside of function parameters.
Pat Hickey (Jan 14 2020 at 19:00):hmm. thats a fair concern
Pat Hickey (Jan 14 2020 at 19:01):well, we can punt on it for now.
Pat Hickey (Jan 14 2020 at 19:01):ill make a PR that fixes some simpler stuff, and panics in the code generator in these unsupported cases
Dan Gohman (Jan 14 2020 at 19:05):Admittadly, the existing pointer type isn't exactly watertight either.
Dan Gohman (Jan 14 2020 at 19:05):But it feels different -- with interface types focusing on by-value semantics, if we have something that we describe as an "array", the value of the array should be passed by copy.
Dan Gohman (Jan 14 2020 at 19:08):Also, there are ongoing discussions at the interface types level about the possibilty of making memory slices a proper thing in interface types, and if we get that, that could change how all of these things work.
Pat Hickey (Jan 14 2020 at 19:08):i havent kept up to date on those discussions
Dan Gohman (Jan 14 2020 at 19:08):So my instinct here is to not get too far ahead of those discussions.
Pat Hickey (Jan 14 2020 at 19:08):ok, works for me
Last updated: Nov 22 2024 at 16:03 UTC