It's checked in a cmake function that builds up a list called
`config_options`. But the check was appending to a list called
`compile_options`, so the intended `-DLIBC_MATH=whatever` didn't end up
on the actual compile command lines.
In IEEE 754 and C standards, when calling `frexp` with Inf/Nan inputs,
the exponent result is unspecified. In this case, FreeBSD libc and musl
just passthrough `exp`, while glibc, FreeBSD libm set exp = 0, and MSVC
set exp = -1.
By default, LLVM libc will passthrough `exp` just as FreeBSD libc and
musl, but we also allow users to explicitly choose the return exp value
in this case for compatibility with other libc.
Notice that, gcc did generate passthrough `exp` for `frexp(NaN/Inf,
exp)`: https://godbolt.org/z/sM8fEej4E
Refer: 7.3.1 from [ISO
SPEC](https://www.open-std.org/jtc1/sc22/wg14/www/docs/n3220.pdf)
I have added complex variants of F16 and F128 in libc doc but have
omitted support for them since we will have to first investigate how
their support matrix for clang and gcc looks like, and then add header
guards for them accordingly. Planning to add them in follow up PRs once
this gets landed.
Currently the nan* functions use nullptr dereferencing to crash with
SIGSEGV if the input is nullptr. Both `nan(nullptr)` and `nullptr`
dereferencing are undefined behaviors according to the C standard.
Employing `nullptr` dereference in the `nan` function implementation is
ok if users only linked against the pre-built library, but it might be
completely removed by the compilers' optimizations if it is built from
source together with the users' code.
See for instance: https://godbolt.org/z/fd8KcM9bx
This PR uses volatile load to prevent the undefined behavior if libc is
built without sanitizers, and leave the current undefined behavior if
libc is built with sanitizers, so that the undefined behavior can be
caught for users' codes.
The 32-bit Arm builds of libc define time_t to be `__INTPTR_TYPE__`,
i.e. a 32-bit integer. This is commented in the commit introducing it
(75398f28ebdb600) as being for compatibility with glibc. But in the near
future not even every AArch32 build of glibc will have a 32-bit time_t:
Debian is planning that their next release (trixie) will have switched
to 64-bit. And non-Linux builds of this libc (e.g. baremetal) have no
reason to need glibc compatibility in the first place – and every reason
_not_ to want to start using a 32-bit time_t in 2024 or later.
So I've replaced the `#ifdef` in `llvm-libc-types/time_t.h` with two
versions of the header file, chosen in `CMakeLists.txt` via a new
configuration option. This involved adding an extra parameter to the
cmake `add_header` function to specify different names for the header
file in the source and destination directories.
Summary:
The GPU build needs to be able to inline stuff in LTO. Builtin
transformations cause problems on the functions that the optimizer does
heavy libcall recognition on. Previously we moved to using
`-fno-builtin-*` to allow us to only disable the problematic ones.
However, this still didn't allow inlining because each function had the
attribute that told the inliner not to inlining a nobuiltin function
into a non-nobuiltin function
This patch fixes that by only applying these attributes to the
entrypoints that define them. That is enough to prevent recursive calls
within the definitoins themselves.
When crosscompiling, we need to search for the linux kernel headers in the sysroot but since #97486 the linux kernel headers were always searched in /usr/include.
This patch fixes this behaviour by prepending a '=' to where we search for the kernel headers. As per the gcc/clang's documentation a '=' before the path is replaced by the sysroot.
This patch also includes a fix for rv32, that fails to compile due to a missing definition of CLOCK_REALTIME after this change.
Summary:
This patch removed the `-ffreestanding` and `-fno-builtin` flags from
the publically installed version of the GPU library. The presense of
`-fno-builtin` caused issues that prevented all inlining in the GPU C
library, see
https://discourse.llvm.org/t/rfc-libc-ffreestanding-fno-builtin/75883.
Previously, I attempted to fix this by loosening the restriction that
`"no-builtins"` functions cannot be inlined into functions without that
attribute. However, that opened up a lot of extra issues that stalled
that approach.
This patch instead removes that and instead passes `-fno-builtin-<xyz>`
for the few calls that are known to be problematic. I believe this works
in general as the GPU backends do not emit any libcalls and the
implementations of most of these simply reduce to builtins right now.
This is a very useful patch as we can now actually inline calls.
Summary:
If the user does not have the selected backend enabled, we should still
be able to build the LLVM-IR an ddistribute it. This patch makes logic
to suppress tests if the backend can't build it, as well as removing a
flag for the building that's only present int he NVPTX backend.
Summary:
This warning is intended to indicate if `-march=native` returns
different values in a single compilation sense. As it stands we don't
care and it absolutely spams the test output if you run it on a machine
with more than one GPU like any cluster machine. Disable these warnings
everywhere we compile.
Summary:
This patch silences two warnings that may occur during the building of
GPU tests. These are not informative or helpful and just make the test
output longer.
Summary:
This is a massive patch because it reworks the entire build and
everything that depends on it. This is not split up because various bots
would fail otherwise. I will attempt to describe the necessary changes
here.
This patch completely reworks how the GPU build is built and targeted.
Previously, we used a standard runtimes build and handled both NVPTX and
AMDGPU in a single build via multi-targeting. This added a lot of
divergence in the build system and prevented us from doing various
things like building for the CPU / GPU at the same time, or exporting
the startup libraries or running tests without a full rebuild.
The new appraoch is to handle the GPU builds as strict cross-compiling
runtimes. The first step required
https://github.com/llvm/llvm-project/pull/81557 to allow the `LIBC`
target to build for the GPU without touching the other targets. This
means that the GPU uses all the same handling as the other builds in
`libc`.
The new expected way to build the GPU libc is with
`LLVM_LIBC_RUNTIME_TARGETS=amdgcn-amd-amdhsa;nvptx64-nvidia-cuda`.
The second step was reworking how we generated the embedded GPU library
by moving it into the library install step. Where we previously had one
`libcgpu.a` we now have `libcgpu-amdgpu.a` and `libcgpu-nvptx.a`. This
patch includes the necessary clang / OpenMP changes to make that not
break the bots when this lands.
We unfortunately still require that the NVPTX target has an `internal`
target for tests. This is because the NVPTX target needs to do LTO for
the provided version (The offloading toolchain can handle it) but cannot
use it for the native toolchain which is used for making tests.
This approach is vastly superior in every way, allowing us to treat the
GPU as a standard cross-compiling target. We can now install the GPU
utilities to do things like use the offload tests and other fun things.
Some certain utilities need to be built with
`--target=${LLVM_HOST_TRIPLE}` as well. I think this is a fine
workaround as we
will always assume that the GPU `libc` is a cross-build with a
functioning host.
Depends on https://github.com/llvm/llvm-project/pull/81557
