Comment reproduced here:
If we're linking libc, some naughty applications may have
registered atexit handlers which we really do not want to
run in the fork child. I caught LLVM doing this and it
caused a deadlock instead of doing an exit syscall. In
the words of Avril Lavigne, "Why'd you have to go and
make things so complicated?"
When we get a cache hit for a stage1 compilation, we need to know about
some of the flags such as have_winmain or have_dllmain to know which
subsystem to infer during linking.
To do this, we append a hex-encoded byte into the intentionally-dangling
symlink which contains the cache hash digest rather than a filename. The
hex-encoded byte contains the flags we need to infer the subsystem
during linking.
Before:
gimli-hash: 120 MiB/s
gimli-aead: 130 MiB/s
After:
gimli-hash: 195 MiB/s
gimli-aead: 208 MiB/s
Also fixes in-place decryption by the way.
If the input & output buffers were the same, decryption used to fail.
Return on decryption error in the benchmark to detect similar issues
in future AEADs even in non release-fast mode.
stage1 was unable to parse ranges whose starting point was written in
binary/octal as the first dot in '...' was incorrectly interpreted as
decimal point.
stage2 forgot to reset the literal type to IntegerLiteral when it
discovered the dot was not a decimal point.
I've only stumbled across this bug because zig fmt keeps formatting the
ranges without any space around the ...
Previously if a LinearFifo was empty and discard was called
an unsigned overflow would occur. However it is safe to perform
this overflow as a bitwise & operation with 0xFFFFFFFFFFFFFF is a noop
* std.log: still print error messages in ReleaseSmall builds.
- when start code gets an error code from main, it uses std.log.err
to report the error. this resulted in a test failure because
ReleaseSmall wasn't printing `error: TheErrorCode` when an error
was returned from main. But that seems like it should keep working.
So I changed the std.log defaults. I plan to follow this up with a
proposal to change the names of and reduce the quantity of the
log levels.
* warning emitted when using -femit-h when using stage1 backend; fatal
log message when using -femit-h with self-hosted backend (because the
feature is not yet available)
* fix double `test-cli` build steps in zig's build.zig
* update docgen to use new CLI
* translate-c uses `-x c` and generates a temporary basename with a
`.h` extension. Otherwise clang reports an error.
* --show-builtin implies -fno-emit-bin
* restore the compile error for using an extern "c" function without
putting -lc on the build line. we have to know about the libc
dependency up front.
* Fix ReleaseFast and ReleaseSmall getting swapped when passing the
value to the stage1 backend.
* correct the zig0 CLI usage text.
* update test harness code to the new CLI.
This is helpful on systems with multiple LLVM installations.
For example, OpenBSD ships with LLVM in the base system, but without
support for extra targets.
A full LLVM version can be installed using the ports system, but even
when it is, `cmake` is not going to pick it up unless `/usr/local` is
explicitly added to the `cmake` search paths.
Having the full `llvm-config` path printed on error is helpful to
understand that the detected LLVM version was not the expected one.
These CLI options are now forwarded to the stage1 backend.
We're not going to support the -mllvm CLI option any longer. As a
compromise, we unconditionally tell LLVM to output intel x86 syntax when
using -femit-asm.
Simplify stage1 logic; it no longer has the concept of an output
directory. --output-dir is no longer a valid CLI option. cmake uses
the `-femit-bin=[path]` option.
Note the changes to test/cli.zig. This breaks the CLI API that Godbolt
is using so we're going to want to open a PR to help them upgrade to the
new CLI for the upcoming Zig 0.7.0 release.
This cleans up how the CLI parses and handles -E, -S, and -c.
Compilation explicitly acknowledges when it is being used to do C
preprocessing.
-S is properly translated to -fno-emit-bin -femit-asm but Compilation
does not yet handle -femit-asm.
There is not yet a mechanism for skipping the linking step when there is
only a single object file, and so to make this work we have to do a file
copy in link.flush() to copy the file from zig-cache into the output
directory.
These APIs allow one to write code that is agnostic of whether it is
using an ArrayHashMap or a HashMap, which can be valuable. Specify
intent precisely: if you only need the count of the items, it makes
sense to have a function for that.
