Bug 1645986 solved the problem for most generated files by moving their
rules to the top-level, but we're going to add rules that will end up in
subdirectories, so we have to solve the same problem again, in the
subdirectories.
Differential Revision: https://phabricator.services.mozilla.com/D88389
It turns out setting CARGO_PROFILE_RELEASE_LTO has unwanted side
effects.
First it's not actually strictly equivalent to using `cargo rustc --
-Clto`. For instance, it apparently also enables cross-language LTO in
newer versions of cargo.
Second, it changes the rust computed hash for all the dependencies of
the crate being built with the variable set, which makes them diverge
from when the same dependencies are built through another crate in the
tree that is not LTOed. This effectively makes us build a _lot_ of
crates twice, many of which are not cacheable.
Since the original problem is that cargo >= 1.45 passes extra flags (`-C
embed-bitcode=no`) to rustc that are incompatible with `-Clto`, and while
it knows to adjust based on the `lto` setting in the build profile
(which CARGO_PROFILE_RELEASE_LTO overrides the default of), cargo
ignores flags passed via `cargo rustc -- ...` when making those
adjustments.
So, we need to override with `-C embed-bitcode=yes` on our own at the
same time we pass `-Clto`. But doing that through `cargo rustc -- ...`
is not enough because all the dependencies of the crate built with
`-Clto` need to be built with `-C embed-bitcode=yes`. So we need to
override with `RUSTFLAGS`, which will affect all the dependencies.
But we also need to do this consistently across all crates, not only the
dependencies of crates built with `-Clto`, otherwise we'd still end up
building crates twice (once with and once without the override).
Unfortunately, the `-C embed-bitcode=*` flag is also not supported in
versions older than 1.45, so we have to avoid adding it on older
versions.
We unfortunately support a large range of versions of rustc (albeit only
for tools/crashreporter), but we actually need to upgrade the smaller
supported version because rustc < 1.38 doesn't support our top-level
Cargo.lock. This makes the version check slightly less awful.
Differential Revision: https://phabricator.services.mozilla.com/D84652
The main reason we look at the complete symbols table is that before bug
1541792, we needed to look at that table for _NSModule symbols.
In bug 1516228, we also made everything llvm-objdump to limit the
differences cross-platform, but that's not necessary anymore per the
previous change.
llvm-objdump doesn't support getting only the dynamic symbols table, so
we go back to what we were using before bug 1516228, namely readelf,
while preserving a code path to use the complete symbols table for the
networking test on libgkrust.a, which doesn't have a dynamic symbols
table.
With this change, check_binary goes from 45s to 0.2s on my machine.
Differential Revision: https://phabricator.services.mozilla.com/D84305
Otherwise, crates that are shared between host and target are always
rebuilt because of the difference in rust flags.
Differential Revision: https://phabricator.services.mozilla.com/D83412
Doing this means that make will correctly understand that the underlying
program has changed and therefore it needs to trigger install rules.
Differential Revision: https://phabricator.services.mozilla.com/D83326
The previous logic would check the value of the `PYTHON3` environment variable, which is not set by the Firefox build system, so this build script tends to just use the `python3` at the tip of the `PATH`, regardless of which value you've configured for it. Instead, it should prefer to read from the `OBJDIR` if one is present.
Differential Revision: https://phabricator.services.mozilla.com/D82981
This makes it much simpler for the for the static JS services cache to store
and lookup interface IDs for components.
Differential Revision: https://phabricator.services.mozilla.com/D81416
This makes it much simpler for the for the static JS services cache to store
and lookup interface IDs for components.
Differential Revision: https://phabricator.services.mozilla.com/D81416
`ply`, [by design](https://github.com/dabeaz/ply/issues/79), does not produce reproducible table files; hence bug 1633156. (Note that this was *always* true, but only became a problem once we switched to Python 3, which has more unpredictable dict iteration order than Python 2.7, at least prior to [3.7](https://docs.python.org/3/whatsnew/3.7.html#summary-release-highlights).)
In any other circumstance I would consider submitting a patch to `ply` to fix this, but as of the [in-progress version 4.0 of the library](https://github.com/dabeaz/ply/blob/master/CHANGES), it doesn't even emit this cached data any more, and indeed the [latest version of the code](1fac9fed64/ply) doesn't even call `open()` at all except to do logging or to read the text data to be parsed from `stdin`. So if we were going to pin our future on `ply` and upgrade to later versions of the library in the future, we would have to live in a world where `ply` doesn't generate cached table files for us anyway.
Emitting the cached table files so later build steps can consume them is an "optimization", but it's not clear exactly how much actual value that optimization provides overall. Quoth the `CHANGES` file from that repository:
```
PLY no longer writes cached table files. Honestly, the use of
the cached files made more sense when I was developing PLY on
my 200Mhz PC in 2001. It's not as much as an issue now. For small
to medium sized grammars, PLY should be almost instantaneous.
```
In practice, I have found this to be true; namely, `./mach build pre-export export` takes just about as long on my machine after this patch as it did before, and in a try push I performed, there's no noticeable performance regression from applying this patch. In local testing I also found that generating the LALR tables in calls to `yacc()` takes about 0.01s on my machine generally, and we generate these tables a couple dozen times total over the course of the `export` tier now. This isn't *nothing*, but in my opinion it's also not nearly long enough where it would be a concern given how long `export` already takes.
That `CHANGES` file also stresses that if caching this data is important, we have the option of doing so via `pickle`. If and when we decide that re-enabling this optimization is valuable for us, we should take control of this process and perform the generation in such a way that we can guarantee reproducibility.
Differential Revision: https://phabricator.services.mozilla.com/D73484
For native sanitizer builds, we currently do not pass the linker flags
to cargo, as they were causing crashes in some build scripts. Without
this, however, the linker is unable to find libstdc++. Instead, do
tell cargo to use the linker wrapper, but omit the problematic flags
from MOZ_CARGO_WRAP_LDFLAGS.
Differential Revision: https://phabricator.services.mozilla.com/D70354
They all fail for the same reason: the sanitizer runtime in compiler-rt installs an interceptor for `getsockname` which then contains a call to the real implementation.
Differential Revision: https://phabricator.services.mozilla.com/D70889
--HG--
extra : moz-landing-system : lando
By default the Rust compiler will perform a limited kind of ThinLTO on each
crate. For local builds this additional optimization is not worth the
increase in compile time.
Differential Revision: https://phabricator.services.mozilla.com/D61821
--HG--
extra : moz-landing-system : lando
The current coreaudio-sys in gecko is a custom 0.2.2 version that used
to avoid the cross-compiling issue mentioned in bug 1569003. The issue
has been fixed in the coreaudio-sys 0.2.3, so we should follow the
upstream instead of using a custom version. As a result, the
coreaudio-sys would generate API bindings based on the MacOS SDK defined
in the build settings.
Differential Revision: https://phabricator.services.mozilla.com/D50531
--HG--
extra : moz-landing-system : lando
If the run task generates bad profile data, the merge step in the
profile-use task will fail. However, retrying the profile-use task
doesn't fix the problem, and there isn't a straightforward way to retry
the run task in this situation. Instead we can add a clang toolchain to
all the run tasks, and perform the merge there.
This means the output from the run task will always be a successfully
merged file called 'merged.profdata', and we no longer need to perform
the merge as part of the profile-use build as a GENERATED_FILES step.
Depends on D45262
Differential Revision: https://phabricator.services.mozilla.com/D45263
--HG--
extra : moz-landing-system : lando
We don't actually care that much about LTO'ing the rust parts of libxul
for gtests, and not LTO'ing it would save multiple minutes of build time
on automation.
Differential Revision: https://phabricator.services.mozilla.com/D42812
--HG--
extra : moz-landing-system : lando
With the addition of toolkit/library/build because of the rust
shenanigans, bug 1573314 and bug 1572046 don't do anything useful
anymore. We're going to do something better anyways.
Differential Revision: https://phabricator.services.mozilla.com/D42251
--HG--
extra : moz-landing-system : lando
When a directory, like toolkit/library, builds both a static and a
shared library, and another, like toolkit/library/gtest, depends on the
static part, it currently needs to wait for the shared library to be
finished building, preventing both libraries being built in parallel.
By separating shared libraries to a different target, we allow more
parallelism to the build.
Differential Revision: https://phabricator.services.mozilla.com/D41099
--HG--
extra : moz-landing-system : lando
We don't need to run binary checks on the instrumentation builds, only the final optimized build.
Differential Revision: https://phabricator.services.mozilla.com/D38382
--HG--
extra : moz-landing-system : lando
