#170525
Previously, lookup tables were defined in a header so that they can be
constexpr, but this lead to each translation unit having it's own copy,
bloating the library size. Looking at current usages of wctype
functions, they are always used at runtime, so there's no benefit of
having them constexpr, so I'm moving them to a cpp file, so that there's
only a single copy in the library.
[#172040](https://github.com/llvm/llvm-project/issues/172040)
This patch implements the scripts for generating the lookup tables and
associated utils for wctype classification functions. Not all Unicode
properties are covered as not all need a lookup table, the rest will be
hardcoded. The size of the generated tables is 47,8KB.
Summary:
These were originally intended to represent the functions that are
present on the GPU as to be provided by the LLVM libc implementation.
The original plan was that LLVM libc would report which functions were
supported and then the offload interface would mark those as supported.
The problem is that these wrapper headers are very difficult to make
work given the various libc extensions everyone does so they were
extremely fragile.
OpenMP already declares all functions used inside of a target region as
implicitly host / device, while these headers weren't even used for CUDA
/ HIP yet anyway. The only things we need to define right now are the
stdio FILE types. If we want to make this work for CUDA we'd need to
define these manually, but we're a ways off and that's way easier
because they do proper overloading.
This makes the sorting behavior more uniform: functions and
macros are always sorted (separately), not only when merging.
This changes the sort order used for functions and other things
sorted by their symbol names. Symbols are sorted alphabetically
without regard to leading underscores, and then for identifiers
that differ only in the number of leading underscores, the fewer
underscores the earlier in the sort order. For the functions
declared in a generated header, adjacent names with and without
underscores will be grouped together without blank lines.
This is implemented by factoring the name field, equality, and
sorting support out of the various entity classes into a new
common superclass (hdrgen.Symbol).
This uncovered YAML's requirement to quote the string "NULL" to
avoid pyyaml parsing it as None (equivalent to Javascript null)
rather than a string.
This adds a few new features to hdrgen, all meant to facilitate
using it with inputs and outputs that are outside the llvm-libc
source tree.
The new `extra_standards` field is a dictionary to augment the
set of names that can be used in `standards` lists. The keys are
the identifiers used in YAML ("stdc") and the values are the
pretty names generated in the header comments ("Standard C").
This lets a libc project that's leveraging the llvm-libc sources
along with its own code define new APIs outside the formal and de
facto standards that llvm-libc draws its supported APIs from.
The new `license_text` field is a list of lines of license text
that replaces the standard LLVM license text used at the top of
each generated header. This lets other projects use hdrgen with
their own inputs to produce generated headers that are not tied
to the LLVM project.
Finally, for any function attributes that are not in a canonical
list known to be provided by __llvm-libc-common.h, an include
will be generated for "llvm-libc-macros/{attribute name}.h",
expecting that file to define the "attribute" name as a macro.
All this can be used immediately by builds that drive hdrgen and
build libc code outside the LLVM CMake build. Future changes
could add CMake plumbing to facilitate augmenting the LLVM CMake
build of libc with outside sources via overlays and cache files.
The JSON output support in hdrgen had a bug that tripped when
used with headers that use special-case headers like <stdint.h>
to supply some times, as well as llvm-libc-types/*.h headers.
This PR adds the following basic math functions for BFloat16 type along
with the tests:
- nearbyintbf16
- rintbf16
- lrintbf16
- llrintbf16
- lroundbf16
- llroundbf16
---------
Signed-off-by: Krishna Pandey <kpandey81930@gmail.com>
This PR adds the following basic math functions for BFloat16 type along
with the tests:
- bf16fma
- bf16fmaf
- bf16fmal
- bf16fmaf128
---------
Signed-off-by: Krishna Pandey <kpandey81930@gmail.com>
This PR adds implements following basic math functions for BFloat16 type
along with the tests:
- bf16add
- bf16addf
- bf16addl
- bf16addf128
- bf16sub
- bf16subf
- bf16subl
- bf16subf128
---------
Signed-off-by: Krishna Pandey <kpandey81930@gmail.com>
Reverts llvm/llvm-project#146226
The MPFR test uses `mpfr_asinpi` which requires MPFR 4.2.0 or later, but
the Buildbots are running an older version of MPFR.
See https://lab.llvm.org/buildbot/#/builders/104/builds/27743 for
example.
I said I was going to revert the PR until we have a workaround for older
versions of MPFR, but then I forgot and I just disabled the entrypoints
which doesn't fix the Buildbot builds.
The function is implemented using the following Taylor series that's
generated using [python-sympy](https://www.sympy.org/en/index.html), and
it is very accurate for |x| $$\in [0, 0.5]$$ and has been verified using
Geogebra. The range reduction is used for the rest range (0.5, 1].
$$
\frac{\arcsin(x)}{\pi} \approx
\begin{aligned}[t]
& 0.318309886183791x \\
& + 0.0530516476972984x^3 \\
& + 0.0238732414637843x^5 \\
& + 0.0142102627760621x^7 \\
& + 0.00967087327815336x^9 \\
& + 0.00712127941391293x^{11} \\
& + 0.00552355646848375x^{13} \\
& + 0.00444514782463692x^{15} \\
& + 0.00367705242846804x^{17} \\
& + 0.00310721681820837x^{19} + O(x^{21})
\end{aligned}
$$
## Geogebra graph
Closes#132210
This PR enables support for BFloat16 type in LLVM libc along with
support for testing BFloat16 functions via MPFR.
---------
Signed-off-by: krishna2803 <kpandey81930@gmail.com>
Signed-off-by: Krishna Pandey <kpandey81930@gmail.com>
Co-authored-by: OverMighty <its.overmighty@gmail.com>
Summary:
There were a few issues with the first one, leading to some errors and
warnings. Most importantly, this was building on MSVC which isn't
supported.
Summary:
These tools `amdhsa-loader` and `nvptx-loader` are used to execute unit
tests directly on the GPU. We use this for `libc` and `libcxx` unit
tests as well as general GPU experimentation. It looks like this.
```console
> clang++ main.cpp --target=amdgcn-amd-amdhsa -mcpu=native -flto -lc ./lib/amdgcn-amd-amdhsa/crt1.o
> llvm-gpu-loader a.out
Hello World!
```
Currently these are a part of the `libc` project, but this creates
issues as `libc` itself depends on them to run tests. Right now we get
around this by force-including the `libc` project prior to running the
runtimes build so that this dependency can be built first. We should
instead just make this a simple LLVM tool so it's always available.
This has the effect of installing these by default now instead of just
when `libc` was enabled, but they should be relatively small. Right now
this only supports a 'static' configuration. That is, we locate the CUDA
and HSA dependencies at LLVM compile time. In the future we should be
able to provide this by default using `dlopen` and some API.
I don't know if it's required to reformat all of these names since they
used the `libc` naming convention so I just left it for now.
This fleshes out the <link.h> a little more, including the
`struct dl_phdr_info` type and declaring the dl_iterate_phdr
function. There is only a no-op implementation without tests, as
for the existing dlfcn functions.
Fixes incorrect logic that went unnoticed until the function was tested
with output and input types that have the same underlying floating-point
format.
Initial UEFI OS target support after the headers. This just defines
enough that stuff might try and compile. Test with:
```
$ cmake -S llvm -B build -G Ninja -DLLVM_RUNTIME_TARGETS=x86_64-unknown-uefi-llvm -DRUNTIMES_x86_64-unknown-uefi-llvm_LLVM_ENABLE_RUNTIMES=libc -DRUNTIMES_x86_64-unknown-uefi-llvm_LLVM_LIBC_FULL_BUILD=true -DCMAKE_C_COMPILER_WORKS=true -DCMAKE_CXX_COMPILER_WORKS=true -DLLVM_ENABLE_PROJECTS="clang;lld" -DCMAKE_BUILD_TYPE=Debug -DLLVM_ENABLE_LIBCXX=true -DLLVM_HOST_TRIPLE=aarch64-unknown-linux-gnu -DLLVM_DEFAULT_TARGET_TRIPLE=x86_64-unknown-uefi-llvm -DCMAKE_INSTALL_LIBDIR=build/target/lib
$ ninja -C build
```
Summary:
This patch moves the RPC server handling to be a header only utility
stored in the `shared/` directory. This is intended to be shared within
LLVM for the loaders and `offload/` handling.
Generally, this makes it easier to share code without weird
cross-project binaries being plucked out of the build system. It also
allows us to soon move the loader interface out of the `libc` project so
that we don't need to bootstrap those and can build them in LLVM.
Summary:
This is done in preparation for making this header only so we can use it
without an object library. This will be a transient interface that
internal LLVM stuff with use. External people can use it too if they go
through the LLVM libc interface for shared stuff but there's no backward
compatibility guarantees.
Patch just cleans it up prior to the move.
Summary:
Currently we dispatch the writing mode off of a runtime enum passed in
by the constructor. This causes very unfortunate codegen for the GPU
targets where we get worst-case codegen because of the unused function
pointer for `sprintf`. Instead, this patch moves all of this to a
template so it can be masked out. This results in no dynamic stack and
uses 60 VGPRs instead of 117. It also compiles about 5x as fast.
This involved a little bit of yak shaving because one of the new tests
depends on MPC, and we didn't have targets for it yet, so I ended up
needing to add a similar setup to what we have for MPFR.
`cpp::is_complex_type_same<T1, T2>` is a function, so we need
parentheses in order to call it. Otherwise the expression is treated
like a function pointer which is always true in this boolean context.
Move the hdrgen code under a subdirectory to treat it as a Python
module.
This mimics the structure used by llvm/utils/lit and
llvm/utils/mlgo-utils and simplifies integration of hdrgen to the build
system which rely on Python modules. In addition to that, it clarifies
which imports are coming from the hdrgen-specific helpers (e.g. "from
type import ..." becomes "from hdrgen.type import ...".
Leave the entrypoints (top-level main.py and yaml_to_classes.py) as-is:
they can keep being referred by the CMake build system w/o any changes.
This adds a feature to hdrgen to emit JSON summaries of header
files for build system integration. For now the summaries have
only the basic information about each header that is relevant for
build and testing purposes: the standards and includes lists.
Macros starting with alphabetic characters such as "LLVM" are in
the application name space and cannot be defined or used by a
conforming implementation's headers. This fixes the headers that
are entirely generated, and the __llvm-libc-common.h header to
use a conforming macro name for the header guard. That is, it
starts with "_LLVM_LIBC_" instead of "LLVM_LIBC_", as identifiers
starting with an underscore followed by a capital letter are in
the name space reserved for the implementation.
The remaining headers either will be fixed implicitly by removal
of their custom template files, or will need to be fixed by hand.
