Extend the <target>.optimized-compiler-builtins bootstrap option to accept a
path to a prebuilt compiler-rt builtins library, and update compiler-builtins
to enable optimized builtins without building compiler-rt builtins.
LLVM does not currently emit these, but it is being discussed as an
option on platforms where `f32` is not hardware supported. Glibc/libgcc
also has the comparison functions [1] already.
The generic implementations for addition, subtraction, and
multiplication work for f16 without any complications, as do
comparisons, so add them here.
[1]: https://sourceware.org/git/?p=glibc.git;a=commit;h=6ec6c77867af4ddfec7323e0ac6ede89effca852
Currently we run the `rustc` from the `RUSTC` environment variable to
figure out whether or not to enable `f16` and `f128`, based on the
`target_has_reliable_{f16,f128}` config. However, this does not know
about the codegen backend used, and the backend isn't trivial to check
in a build script (usually it gets set via `RUSTFLAGS`).
It turns out we don't actually need to run `rustc` here: Cargo
unconditionally emits all config from the relevant compiler as
`CARGO_CFG_*` variables, regardless of whether or not they are known
options. Switch to checking these for setting config rather than
invoking `rustc`.
As an added advantage, this will work with target.json files without any
special handling.
Fixes: ed17b95715dd ("Use the compiler to determine whether or not to enable `f16` and `f128`")
Add dynamic support for aarch64 LSE atomic ops on linux targets
when optimized-compiler-builtins is not enabled.
A hook, __enable_rust_lse, is provided for the runtime to enable
them if available. A future patch will use this to enable them
if available.
The resulting asm should exactly match that of LLVM's compiler-rt
builtins, though the symbol naming for the support function and
global does not.
Assembly-related configuration was added in 1621c6dbf9eb ("Use
`specialized-div-rem` 1.0.0 for division algorithms") to account for
Cranelift not yet supporting assembly. This hasn't been relevant for a
while, so we no longer need to gate `asm!` behind this configuration.
Thus, remove `cfg(not(feature = "no-asm"))` in places where there is no
generic fallback.
There are other cases, however, where setting the `no-asm` configuration
enables testing of generic version of builtins when there are platform-
specific implementations available; these cases are left unchanged. This
could be improved in the future by exposing both versions for testing
rather than using a configuration and running the entire testsuite
twice.
This is the compiler-builtins portion of
https://github.com/rust-lang/rust/pull/144471.
Emit `x86_no_sse` in the compiler-builtins (and builtins-test) build
script, and use it to simplify `all(target_arch = "x86",
not(target_fefature = "sse))` configuration.
This alias was added in 9897bfb8a ("Fix memset arguments for MSP430
target"), which predates `core::ffi`. Now that it exists we can just use
`core::ffi::c_int`.
Disable docs for `compiler-builtins` and `sysroot`
Bootstrap already had a manual doc filter for the `sysroot` crate, but
other library crates keep themselves out of the public docs by setting
`[lib] doc = false` in their manifest. This seems like a better solution
to hide `compiler-builtins` docs, and removes the `sysroot` hack too.
Fixesrust-lang/rust#143215 (after backport)
```@rustbot``` label beta-nominated
Bootstrap already had a manual doc filter for the `sysroot` crate, but
other library crates keep themselves out of the public docs by setting
`[lib] doc = false` in their manifest. This seems like a better solution
to hide `compiler-builtins` docs, and removes the `sysroot` hack too.
The `rustc` probe done in our build scripts needs to pass `--target` to
get the correct configuration, which usually comes from the `TARGET`
environment variable. However, for targets specified via a `target.json`
file, `TARGET` gets set to the file name without an extension or path.
`rustc` will check a search path to attempt to locate the file, but this
is likely to fail since the directory where Cargo invokes build scripts
(and hence where those scripts invoke `rustc`) might not have any
relation to the JSON spec file.
Resolve this for now by leaving `f16` and `f128` disabled if the `rustc`
command fails. Result of the discussion at CARGO-14208 may eventually
provide a better solution.
A CI test is also added since custom JSON files are an edge case that
could fail in other ways. I verified this fails without the fix here.
The JSON file is the output for `thumbv7em-none-eabi`, just renamed so
`rustc` doesn't identify it.
Currently we whether or not to build and test `f16` and `f128` support
mostly based on the target triple. This isn't always accurate, however,
since support also varies by backend and the backend version.
Since recently, `rustc` is aware of this with the unstable config option
`target_has_reliable_{f16,f128}`, which better represents when the types
are actually expected to be available and usable. Switch our
compiler-builtins and libm configuration to use this by probing `rustc`
for the target's settings.
A few small `cfg` fixes are needed with this.
Out-of-tree testing is broken with the most recent update from
rust-lang/rust because it makes `compiler-builtins` depend on `core` by
path, which isn't usually available. In order to enable testing outside
of rust-lang/rust, add a new crate `builtins-shim` that uses the same
source as `compiler-builtins` but drops the `core` dependency. This has
replaced `compiler-builtins` as the workspace member and entrypoint for
tests.
Use a consistent ordering for top-level manifest keys, and remove those
that are now redundant (`homapage` isn't supposed to be the same as
`repository`, and `documentation` automatically points to docs.rs now).
The `build.rs` entrypoint returns early for some targets, so emscripten
and OpenBSD were not getting check-cfg set. Emit these earlier to avoid
the `unexpected_cfgs` lint.
On the ILP32 `x86_64-unknown-linux-gnux32` target, `usize` is 32 bits so
there is a sub-register alignment warning. Specify the 64-bit `r`
registers, which matches the current default as well as the size of the
other operands in the routines.
There are a few places that violate this lint, which showed up in
rust-lang/rust CI (the relevent module is gated behind
`kernel_user_helpers` which is only set for `armv4t`, `armv5te`, and
`arm-linux-androideabi`; none of these are tested in compiler-builtins
CI). Add new `unsafe { /* ... */ }` blocks where needed to address this.
Some blocks should get a more thorough review of their preconditions, so
their safety comments are left as `FIXME`s.
Many of `std`'s dependency have a dependency on the crates.io
`compiler-builtins` when used with the feature
`rustc-std-workspace-core`. Use a Cargo patch to select the in-tree
version instead.
`compiler-builtins` is also added as a dependency of
`rustc-std-workspace-core` so these crates can remove their crates.io
dependency in the future.
This was introduced before `#[panic_handler]` was stable, but should no
longer be needed. Additionally, we only need it for
`builtins-test-intrinsics`, not as a dependency of `compiler-builtins`.
compiler-builtins has a symlink to the `libm` source directory so the
two crates can share files but still act as two separate crates. This
causes problems with some sysroot-related tooling, however, since
directory symlinks seem to not be supported.
The reason this was a symlink in the first place is that there isn't an
easy for Cargo to publish two crates that share source (building works
fine but publishing rejects `include`d files from parent directories, as
well as nested package roots). However, after the switch to a subtree,
we no longer need to publish compiler-builtins; this means that we can
eliminate the link and just use `#[path]`.
Similarly, the LICENSE file was symlinked so it could live in the
repository root but be included in the package. This is also removed as
it caused problems with the dist job (error from bootstrap's
`tarball.rs`, "generated a symlink in a tarball").
If we need to publish compiler-builtins again for any reason, it would
be easy to revert these changes in a preprocess step.
These are now provided by `compiler-builtins`, so there is no need to
also build the C versions. This was detected by checking for duplicate
symbols and not excluding weak symbols (like CI currently does).
As seen at [1], LLVM uses `long long` on LLP64 (to get a 64-bit integer
matching pointer size) and `long` on everything else, with exceptions
for AArch64 and AVR. Our current logic always uses an `i32`. This
happens to work because LLVM uses 32-bit instructions to check the
output on x86-64, but the GCC checks the full 64-bit register so garbage
in the upper half leads to incorrect results.
Update our return type to be `isize`, with exceptions for AArch64 and
AVR.
Fixes: https://github.com/rust-lang/compiler-builtins/issues/919
[1]: 0cf3c437c1/compiler-rt/lib/builtins/fp_compare_impl.inc (L11-L27)
