Now that `73884ae` is in some nightly release We can add ledf2vfp/leds2vfp
and so these two functions be aliased to aeabi_fcmple/aeabi_dcmple on soft-float targets.
1. Avoid undefined references as:
undefined reference to `__modsi3'
undefined reference to `__umodsi3'
2. We can't remove assembly implementations that are not in the list
Collection of VFP intrinsics
Nothing really exciting here, just a list of trivial VFP intrinsics.
First of all set `mfloat-abi=hard` not only for thumb targets, then add support for the following intrinsics:
```
__gesf2vfp
__gedf2vfp
__gtsf2vfp
__gtdf2vfp
__ltsf2vfp
__ltdf2vfp
__nesf2vfp
__nedf2vfp
__eqsf2vfp
__eqdf2vfp
__extendsfdf2vfp
```
Resulting implementation is really trivial thanks to native code generated by LLVM on hard-float targets
Be sure to do not mix hard-float and soft-float calls.
Disassembled code show exactly this.
So replace add with an explicit call to __addsf3/__adddf3
This seems the root cause of some sporadic failures.
I was able to trigger an issue extending f32::MAX or f32::MIN to a f64.
Issue was not triggered by `testcrate` mainly because f32::MAX/MIN are
not in the list of special values to generate.
This PR fix the issue and improve `testcrate` adding MAX/MIN/MIN_POSITIVE
in the list of special values.
Add `extend` module to implement conversion from a narrower to a wider
floating-point type.
This implementation is only intended to support *widening* operations.
Module to convert a *narrower* floating-point will be added in the future.
All tests are moved to a separate crate in this repository to enable features by
default. Additionally the test generation is moved to a seprate build script and
simplified to reduce the amount of boilerplate needed per test.
Overall this should still be testing everything, just in a different location!
Here using `"C"` the compiler will use `"aapcs"` or `"aapcs-vfp"`
depending on target configuration.
Of course this translates in a call to `__aeabi_fdiv` / `__aeabi_fmul`
on non-HF targets.
On `eabi` targets with +vfpv2/vfpv3 LLVM generate:
vmov s0, r1
vmov s2, r0
vdiv.f32 s0, s2, s0
vmov r0, s0
bx lr
On `eabihf` targets with +vfpv3-d16/d32/f32 +fp-only-sp LLVM generate:
vdiv.f32 s0, s0, s1
bx lr
That's exactly what We need for [div/mul][s/d]f3vfp.S
* I believe `__gtdf2` erroneously used `f32` instead of `f64`
* Most of these needed `#[arm_aeabi_alias]` to ensure they're correctly called
through the alias
* Some existing aliases were corrected with the right names
First of all aeabi_cdcmp and aeabi_cfcmp are never called LLVM.
Second are little-endian only, so build fail on big-endian targets.
Temporally workaround: exclude these files for big-endian targets.
Note that this changes semantics:
pub extern "C" fn __eqsf2(a: f32, b: f32) -> bool {
cmp(a, b).to_le_abi() != 0
}
is not the same as
pub extern "C" fn __eqsf2(a: f32, b: f32) -> i32 {
cmp(a, b).to_le_abi()
}
However, compiler-rt does the latter, so this is actually
an improvement.
E.g. take a look at the assembly output for:
pub fn comparesf2(a: f32, b: f32) -> bool { a > b }
pub fn comparedf2(a: f64, b: f64) -> bool { a > b }
which will include calls to __gtsf2 and __gtdf2.
Use the "volatile" option and the "memory" clobber on inline asm that does
things like return directly, to reduce the chances of compilers rearranging
the code.
This commit prepares the build script for a wasm32 target that doesn't use
Emcripten, notably forcing the `mem` feature to get activated and forcibly
ignoring the `c` feature, even if activated, for the wasm32 target.
Instead use the Cargo runner feature. The binfmt_misc approach requires running a privileged
container for setup. Not all docker setups support privileged containers so the test suite should be
more accessible with this change as no privileged container is needed.
Implement x86 chkstk in "rust"
cc #183
Basically the same as the x86_64 ones, except `__alloca` doesn't need to fix the parameter register. I've manually verified that the disassembly is the same, and that these work in a compiled rust program.
The second commit disables compiling probestack functions for `feature = mangled-names`. They aren't needed during testing because they aren't comparison tested and the unmangled versions are the ones that actually get used.
r? @alexcrichton
Update rustbuild commentary
r? @nikomatsakis
This is the reason that https://github.com/rust-lang/rust/pull/44509 doesn't work - the `rustbuild` feature _is_ actually used, it was just incorrectly documented here and I missed it.
It looks like the old `__rust_probestack` routine is incompatible with newer
linux kernels. My best guess for this is that the kernel's auto-growth logic is
failing to trigger, causing what looks like a legitimate segfault to get
delivered. My best guess for why *that's* happening is that the faulting address
is below `%rsp`, whereas previously all faulting stack addresses were above
`%rsp`. The probestack routine does not modify `%rsp` as it's probing the stack,
and presumably newer kernels are interpreting this as a legitimate violation.
This commit tweaks the probestack routine to instead update `%rsp` incrementally
as probing happens. The ABI of the function, however, requires that `%rsp`
isn't changed as part of the function so it's restored at the end to the
previous value.
Although compiler-rt presumably has a more optimized implementation written in
assembly, it appears buggy for whatever reason, causing #173.
For now let's see if integration into rust-lang/rust will work with the
Rust-defined implementation!
* Don't run `intrinsics` tests on thumb
* Disable `compiler_builtins` attribute on `feature = "gen-tests"`
* Disable mangling on `feature = "gen-tests"` instead of `cfg(test)`
