Make shared_objects a StringArrayHashMap so that deduping does not
need to happen in flush. That deduping code also was using an O(N^2)
algorithm, which is not allowed in this codebase. There is another
violation of this rule in resolveSymbols but this commit does not
address it.
This required reworking shared object parsing, breaking it into
independent components so that we could access soname earlier.
Shared object parsing had a few problems that I noticed and fixed in
this commit:
* Many instances of incorrect use of align(1).
* `shnum * @sizeOf(elf.Elf64_Shdr)` can overflow based on user data.
* `@divExact` can cause illegal behavior based on user data.
* Strange versyms logic that wasn't present in mold nor lld. The logic
was not commented and there is no git blame information in ziglang/zig
nor kubkon/zld. I changed it to match mold and lld instead.
* Use of ArrayList for slices of memory that are never resized.
* finding DT_VERDEFNUM in a different loop than finding DT_SONAME.
Ultimately I think we should follow mold's lead and ignore this
integer, relying on null termination instead.
* Doing logic based on VER_FLG_BASE rather than ignoring it like mold
and LLD do. No comment explaining why the behavior is different.
* Mutating the original ELF symbols rather than only storing the mangled
name on the new Symbol struct.
I noticed something that I didn't try to address in this commit: Symbol
stores a lot of redundant information that is already present in the ELF
symbols. I suspect that the codebase could benefit from reworking Symbol
to not store redundant information.
Additionally:
* Add some type safety to std.elf.
* Eliminate 1-3 file system reads for determining the kind of input
files, by taking advantage of file name extension and handling error
codes properly.
* Move more error handling methods to link.Diags and make them
infallible and thread-safe
* Make the data dependencies obvious in the parameters of
parseSharedObject. It's now clear that the first two steps (Header and
Parsed) can be done during the main Compilation pipeline, rather than
waiting for flush().
Embrace the Path abstraction, doing more operations based on directory
handles rather than absolute file paths. Most of the diff noise here
comes from this one.
Fix sorting of crtbegin/crtend atoms. Previously it would look at all
path components for those strings.
Make the C runtime path detection partially a pure function, and move
some logic to glibc.zig where it belongs.
This commit changes how `std.io.poll` is implemented on Windows. The new
implementation unfortunately incurs a little extra system call overhead,
but fixes several bugs in the old implementation:
* The `lpNumberOfBytesRead` parameter of `ReadFile` was used with
overlapped I/O. This is explicitly disallowed by the documentation, as
the value written to this pointer is "potentially erroneous"; instead,
`GetOverlappedResult` must always be used, even if the operation
immediately returns. Documentation states that `lpNumberOfBytesRead`
cannot be passed as null on Windows 7, so for compatibility, the
parameter is passed as a pointer to a dummy global.
* If the initial `ReadFile` returned data, and the next read returned
`BROKEN_PIPE`, the received data was silently ignored in the sense
that `pollWindows` did not `return`, instead waiting for data to come
in on another file (or for all files to close).
* The asynchronous `ReadFile` calls which were left pending between
calls to `pollWindows` pointed to a potentially unstable buffer, since
the user of `poll` may use part of the `LinearFifo` API which rotate
its ring buffer. This race condition was causing CI failures in some
uses of the compiler server protocol.
These issues are all resolved. Now, `pollWindows` will queue an initial
read to a small (128-byte) stable buffer per file. When this read is
completed, reads directly into the FIFO's writable slice are performed
until one is left pending, at which point that read is cancelled (with a
check to see if it was completed between the `ReadFile` and `CancelIo`
calls) and the next read into the small stable buffer is queued. These
small buffer reads are the ones left pending between `pollWindows`
calls, avoiding the race condition described above.
Related: #21565
