Improvements for xxHash performance, both on small keys as well as large slices.
* std.hash: make benchmarks consistent for xxhash
There is some odd behaviour in the timings for the XXHash benchmarks
introduced in 879f0b9, specifically the changes to the allocation in
benchmarkHash. The problem is somewhere in the difference between
9628243 and 9362d61 (these are commit that were force-pushed over but
can be found in PR #15917).
* std.hash: correctly calculate throughput in benchmark
* std.hash: add hashes per sec to small key output
* std.hash: add exact and small xxhash routines
* std.hash: add --small-only flag to benchmark
* std.hash.xxhash: extract stateless Accumulator type
* std.hash.xxhash: cleanup hash() and improve small key perf
* std.hash.xxhash: port xxhash64 changes to xxhash32
* std.hash: change benchmark --small-only flag to --include-array
Specifically this is to make sure llvm data layout generation doesn't
regress. The no emit bin is to allow testing targets that can't
currently be linked. The commented out targets are ones that fail in
the linker anyway when no emit bin is passed.
6e4fff6ba62ae3e61a948c98fa8fea7e35732cc0 regressed the direct CLI
interface of -femit-docs.
This commit fixes it and also hooks up to the progress bar.
This fixes a regression caused by https://github.com/ziglang/zig/pull/13993
As an optimization, the first call to `NtQueryDirectoryFile` would only ask for a single result and assume that if the result returned did not match the app_name exactly, then the unappended app_name did not exist. However, this relied on the assumption that the unappended app_name would always be returned first, but that only seems to be the case on NTFS. On FAT filesystems, the order of returned files can be different, which meant that it could assume the unappended file doesn't exist when it actually does.
This commit fixes that by fully iterating the wildcard matches via `NtQueryDirectoryFile` and taking note of any unappended/PATHEXT-appended filenames it finds. In practice, this strategy does not introduce a speed regression compared to the previous (buggy) implementation.
Benchmark 1 (10 runs): winpathbench-master.exe
measurement mean ± σ min … max outliers delta
wall_time 508ms ± 4.08ms 502ms … 517ms 1 (10%) 0%
peak_rss 3.62MB ± 2.76KB 3.62MB … 3.63MB 0 ( 0%) 0%
Benchmark 2 (10 runs): winpathbench-fat32-fix.exe
measurement mean ± σ min … max outliers delta
wall_time 500ms ± 21.4ms 480ms … 535ms 0 ( 0%) - 1.5% ± 2.8%
peak_rss 3.62MB ± 2.76KB 3.62MB … 3.63MB 0 ( 0%) - 0.0% ± 0.1%
---
Partially addresses #16374 (it fixes `zig build` on FAT32 when no `zig-cache` is present)
Since the same Key.Func data structure is used for coerced function
bodies as well as uncoerced function bodies, there is danger of them
being hashed and equality-checked as the same. When that happens, the
type of a function body value might be wrong, causing lots of problems.
In this instance, it causes an assertion failure.
This commit fixes it by introducing an `uncoerced_ty` field which
is different than `ty` in the case of `func_coerced` and is used to
differentiate when doing hashing and equality checking.
I have a new behavior test to cover this bug, but it revealed *another*
bug at the same time, so I will fix it in the next commit and and add
the new test therein.
These two tests can't be disambiguated at comptime, since the filesystem that the test is running on also matters for whether or not POSIX_SEMANTICS / IGNORE_READONLY_ATTRIBUTE can actually be used (since they are only supported on NTFS).
Using FileDispositionInformationEx (and therefore flags like FILE_DISPOSITION_POSIX_SEMANTICS and FILE_DISPOSITION_IGNORE_READONLY_ATTRIBUTE) is only supported on NTFS, so the comptime Windows version range check is not enough to determine whether or not the NtSetInformationFile will succeed.
This commit makes DeleteFile always try using FileDispositionInformationEx first, but if INVALID_PARAMETER is received (which is the status that occurs when the filesystem doesn't support FileDispositionInformationEx), then it will fallback and try calling NtSetInformationFile with FileDispositionInformation.
This keeps NTFS as fast as it was before, since it will do at most 1 NtSetInformationFile call, but on non-NTFS filesystems (e.g. FAT32), DeleteFile may need to do 2 NtSetInformationFile calls.
Closes#16497
There were two missing places.
Regressed in the #16318 branch.
Found from compiling Bun. Unfortunately we do not have a behavior test
reduction for this bug.
The logic incorrectly assumed that adhoc_inferred_error_set_type would
be part of the inferred_error_set InternPool.Key when it actually is
part of `simple_type`.
Regressed in the #16318 branch.
Found from compiling Bun. Unfortunately we do not have a behavior test
reduction for this bug.
Fixes regression introduced in
db33ee45b7261c9ec62a1087cfc9377bc4e7aa8f.
Found from compiling bun. Unfortunately we do not have a behavior test
reduction for this bug.
Apparently hard things:
* run autodoc on the std lib
* build the langref
If any zig distros want these documentation, they'll need to get them
from a stage3 `zig build` instead of stage2 (LLVM backend instead of C
backend).
* build.zig: introduce `-Dflat` option which makes the installation
match what we want to ship for our download tarballs. This allows
deleting a bunch of shell script logic from the CI.
- for example it puts the executable directly in prefix/zig rather
than prefix/bin/zig and it additionally includes prefix/LICENSE.
* build.zig: by default also install std lib documentation to doc/std/
- this can be disabled by `-Dno-autodocs` similar to how there is
already `-Dno-langref`.
* build.zig: add `std-docs` and `langref` steps which build and install
the std lib autodocs and langref to prefix/doc/std and
prefix/doc/langref.html, respectively.
* std.Build: implement proper handling of `-femit-docs` using the
LazyPath system. This is a breaking change.
- this is a partial implementation of #16351
* frontend: fixed the handling of Autodocs with regards to caching and
putting the artifacts in the proper location to integrate with the
build system.
- closes#15864
* CI: delete the logic for autodocs since it is now handled by build.zig
and is enabled by default.
- in the future we should strive to have nearly all the CI shell
script logic deleted in favor of `zig build` commands.
* CI: pass `-DZIG_NO_LIB=ON`/`-Dno-lib` except for the one command where
we want to actually generate the langref and autodocs. Generating the
langref takes 14 minutes right now (why?!) so we don't want to do that
more times than necessary.
* Autodoc: fixed use of a global variable. It works fine as a local
variable instead.
- note that in the future we will want to make Autodoc run
simultaneously using the job system, but for now the principle of
YAGNI dictates that we don't have an init()/deinit() API and instead
simply call the function that does the things.
* Autodoc: only do it when there are no compile errors
Both ways do the same thing but I think the compiler might have an
easier time optimizing `appendNTimes` because it does less
things/the path is shorter.
I have not done any benchmarking at runtime but have compared the
instruction count of both ways a little here: https://zig.godbolt.org/z/vr193W9oj
`b` (`appendNTimes`) is ~103 instructions while `a`
(`writer().writeByteNTimes`) is ~117 instructions.
And looking at the implementation of `writeByteNTimes`, it only seems to
buffer up 256 bytes before doing another `writeAll` which for
`std.ArrayList` probably means another allocation, whereas when directly
using `appendNTimes`, the entire exact additional capacity required is known from the start.
Either way, this would be more consistent anyway.
When a signed integer's bitsize is not 32 or 64, but the given
bitsize and wanted bitsize are either both represented by Wasm's i32
or i64, we must either sign extend or wrap the integer.
