This fixes (on my end) #3340, #3117, #2945, #2922, and #2762
A general refactor of debug info generation in order to fix issues and
increase stability.
What I believe is the root cause of a bunch of issues is that we use the
temporary metadata/forward declarations too much (/ hold onto them for
too long). It seems to cause problems with the reference counting inside
LLVM.
This PR reduces the use of these forward declarations to a minimum, it
creates it, fills in the fields, and resolves it, instead of waiting
until the end of generating code.
Some smaller issues I came across have also been solved.
We did some profiling for #3343 and this seems to be the biggest
problem. `realpath` is expensive, and we are locking here for no reason
that I can think of.
This improves the "check procedure bodies" timing (of the linked issue)
from 2.4s to .4s on my machine.
Trying to fix all linker warnings that macOS comes up with, when
building a dynamic library it currently emits `ld: warning: ignoring -e, not used for output type`
After #3316 we set a default minimum version, now this will warn if you
link with a library that is targeting later versions.
This might be a bit annoying, especially when the user hasn't actually
given Odin a minimum target.
So this PR makes these warnings only show when you explicitly give a
target version (afaik that is the only thing that -mmacosx-min-version
actually does for us because we don't use it to compile anything, just
to link).
Since commit b4fe9677a1 some core tests
segfault during build, upon investigation it is because some arg types
were of size 0 and you can't have a 0 sized int.
It also applies the earlier fix for parameters to the return types, this
fixes#3223
Thought I would put this PR up, but I am in no way an expert in abi so
feel free to close for a better fix if there is one.
I am able to run the entire core test suite with `-sanitize:address`
with these changes.
the required procs within wasm would compile to take native v128
arguments in, but the procs are supposed to take in i64's causing bad wasm
modules.
Fixes#3263
`path_to_fullpath` did different things on Windows&Unix, an attempt to
bring them closer together was made here.
This was prompted by the compiler completely ignoring
`foreign import "foo.dylib"` when `foo.dylib` does not exist (because
`path_to_fullpath` returns an empty string).
Causing just unresolved symbol errors, when on Windows it would pass
along the path to the linker and actually say it doesn't exist, which
is now also the case for Unix.
This also fixes some checker errors that relied on the Windows
behaviour, for example: `Error: File name, , cannot be as a library name
as it is not a valid identifier`.
Made `-no-crt` require either `-default-to-nil-allocator` or
`-default-to-panic-allocator` on Unix, the current default allocators
rely on libc and this was not obvious and would immediately bring up
unresolved symbol errors for the linked memory management functions, or
just link with libc anyways because it was foreign imported.
Added a suggestion to install `nasm` with the user's package manager
when assembling using `nasm` fails on Unix, I saw some confusion about
it in the Discord.
Ignore explicit foreign imports of libc. It is already linked in later
on in the linking process and would otherwise (at least on macOS) cause
linker warnings for duplicate libraries. This also makes it so when
using `-no-crt` and importing something that requires libc, linker
errors are given (like I would expect), instead of silently still
linking with libc because it was foreign imported.
Watching the sporadic CI failures it seems to come from these tests a
lot of the time, this PR cleans up and simplifies (while testing the
same things):
1. Lots of tests were using threads without a need for it
2. Tests had hardcoded `time.sleep` calls which is never a good idea
3. An unclear abstraction was implemented without a real need
4. They weren't being ran on non-windows
5. The `client_connects_to_open_but_not_accepting_port` was not doing
what you wanted to test for, the `tcp_server` proc was returning, and
then `dial` was called, which meant that the server already closed
and you got a refusal error. Now it correctly listens without
accepting, which even results in a different error because the kernel
buffer would have buffered the send
We already have 2 different folders for the different architectures, so
this does should not have any usage impact.
Main motivation for doing this is hombrew requiring non-universal
binaries in their installation, so this is a step to getting homebrew
and vendor to play nice.
Using `system:miniaudio` is suboptimal, we already provide the
`Makefile` that builds the `lib/miniaudio.a` and this works on MacOS.
This PR makes linking with that library the default.
Laytan Laats