The problem is we create the call to __kmpc_kernel_parallel in the
openmp-opt pass but while we optimize the code, the call is not there
yet. Thus, we assume we never reach it from __kmpc_target_deinit. That
allows us to remove the store in there (`ParallelRegionFn = nullptr`),
which leads to bad results later on.
This is a shortstop solution until we come up with something better.
Fixes https://github.com/llvm/llvm-project/issues/57064
We will add some simple implementation of libc functions starting from
this patch, and the first one is `memcmp`, which is reported in #56929. Note that
`malloc` and `free` are not included in this patch because of the use of
`declare variant`. In the near future we will implement the two functions w/o
using any vendor provided function.
This fixes#56929.
Reviewed By: jhuber6
Differential Revision: https://reviews.llvm.org/D131182
The behaviour of this patch is not great, but it has some side-effects
that are required for OpenMPOpt to work. The problem is that when we use
`-mlink-builtin-bitcode` we only import used symbols from the runtime.
Then OpenMPOpt will insert calls to symbols that were not previously
included. This patch removed this implicit behaviour as these functions
were kept alive by the `noinline` simply because it kept calls to them
in the module. This caused regression in some tests that relied on some
OpenMPOpt passes without using LTO. Reverting for the LLVM15 release but
will try to fix it more correctly on main.
This reverts commit d61d72dae6.
Fixes#56752
A previous patch removed the need to set the auxiliary architecture as
it was no longer needed for the clang invocation after moving to using
the clang frontend. However, this had a second use of preventing
unsupported host architectures from building the device runtime. This
caused failures when trying to build on 32-bit hosts for example.
Fixes#56699
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D130509
We previously used the `noinline` attributes to specify some defintions
which should be kept alive in the runtime. These were then stripped
immediately in the OpenMPOpt module pass. However, Since the changes in
D130298, we not explicitly state which functions will have external
visiblity in the bitcode library. Additionally the OpenMPOpt module pass
should run before the inliner pass, so this shouldn't make a difference
in whether or not the functions will be alive for the initial pass of
OpenMPOpt. This should simplify the interface, and additionally save
time spend on scanning funciton names for noinline.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D130368
Multiple calls to `omp_get_wtime` could be optimized out due to the function
is mistakenly marked as `readnone`. This patch fixes the issue, and also add the
support to run optimization on `libomptarget` tests.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D130179
Multiple calls to `omp_get_wtime` could be optimized out due to the function
is mistakenly marked as `readnone`. This patch fixes the issue, and also add the
support to run optimization on `libomptarget` tests.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D130179
This patch adds the use of the `-internalize-public-api-file` option in
the internalization pass to internalize any definition that isn't
explicitly needed for the interface. This will allow us to perform more
optimizations on the file that normally would not have been possible
with functions internal to the library not being internal.
Depends on D130293
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D130298
Currently the bitcode library is build using the clang front-end
manually. This was originally done because we did not support device
only compilation. Now we support device only compilation, at least for a
single offloading toolchain, so we can instead use clang directly rather
than using the front-end. This saves us needing to define things like
`aux_triple`.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D130293
We can help optimizations by making sure we use the team state whenever
it is clear there is no thread state. To this end we introduce a new
state flag (`state::HasThreadState`) and explicit control for the
`state::ValueRAII` helpers, including a dedicated "assert equal".
Differential Revision: https://reviews.llvm.org/D130113
Our conditional writes in the runtime look like this:
```
if (active)
*ptr = value;
```
In the RAII we need to assign `ptr` which comes from a lookup call.
If a thread that is not the main thread calls lookup with the intention
to write the pointer, we'll create a new thread state. As such, we need
to avoid calling lookup for inactive threads. We used to use `nullptr`
as their `ptr` value but that can cause pessimistic reasoning. We now
use `undef` instead.
Differential Revision: https://reviews.llvm.org/D130114
We used to inline the `lookup` calls such that the runtime had "known"
access offsets when it was shipped. With the new static library build it
doesn't as the lookup is an indirection we cannot look through. This
should help us optimize the code better until we can do LTO for the
runtime again.
Differential Revision: https://reviews.llvm.org/D130111
We previously had some logic that stopped us from building the device runtime if
there were no NVPTX architectures provided. This is incorrect because we could
have AMDGPU libraries. Even if the lists are empty we should be able to attempt
to build these and get dummy output. THis wilil make it much easier for our
tooling which expects certain libraries. If the user wishes to disable the
library entirely they should use `-DLIBOMPTARGET_BUILD_DEVICERTL_BCLIB=OFF"
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D130266
The device runtime uses the address space attribute to control the
placement of important constants on the GPU. The changes made in D126061
caused these to start emitting errors as they were not applied to the
type. This patch fixes the issues to make the warnings go away.
Reviewed By: ye-luo
Differential Revision: https://reviews.llvm.org/D129896
This patch implements omp_get_device_num() in the host and the device.
It uses the already existing getDeviceNum in the device config for the device.
And in the host it uses the omp_get_num_devices().
Two simple tests added
Differential Revision: https://reviews.llvm.org/D128347
Summary:
This patch removes a duplicated exit from the OpenMP data envrionment.
We already have an RAII method that guards this environment so it is
unnecessary.
Make libomptarget.device.a built when using -DLLVM_ENABLE_PROJECTS=openmp
Use add_custom_command.
Reviewed By: jhuber6
Differential Revision: https://reviews.llvm.org/D128130
Old LLVM installation may expose its internal omptarget CMake target when being used by find_package(LLVM) and caused issues in the CMake of libomptarget that is being built. Trap the issue early.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D128129
When we build the libomptarget device runtime library targeting bitcode,
we need special care to make sure that certain functions are not
optimized out. This is because we manually internalize and optimize
these definitions, ignoring their standard linkage semantics. When we
build with the static library, we can maintain these semantics and we do
not need these to be kept-alive. Furthermore, if they are kept-alive it
prevents them from being removed during LTO. This prevents us from
completely internalizing `IsSPMDMode` and removing several other
functions. This patch removes these for the static library target by
using a macro definition to enable them.
Reviewed By: JonChesterfield
Differential Revision: https://reviews.llvm.org/D126701
Summary:
We usually used the `OMP_LIKELY` and `OMP_UNLIKELY` macros to add branch
prediction intrinsics to help the optimizer ignore unlikely loops. This
wasn't applied to this one loop so add that in.
Summary:
This patch adds the `leaf` attribute to the `vprintf` declaration in the
OpenMP runtime. This attribute allows us to determine that the `vprintf`
function will not call any functions within the translation unit,
allowing us to deduce `norecurse` attributes on the caller.
The OpenMP device offloading library is a bitcode library and thus only
expect to build and linked with the same version of clang that was used
to create it. This somewhat copmlicates the building process as we
require the Clang that was just built to be used to create the library.
This is either done with a two-step build, where OpenMP is built with
the Clang that was just installed, or through the
`-DLLLVM_ENABLE_RUNTIMES=openmp` option. This has always been the case,
but recent changes have caused this to make it difficult to build the
rest of OpenMP. This patchs adds a check to not build the OpenMP device
runtime if the current compiler is not Clang with the same version as
the LLVM installation. This should allow users to build OpenMP as a
project using any compiler without it erroring out due to the bitcode
library, but if users require it they will need to use the above methods
to compile it.
Reviewed By: jdoerfert, tianshilei1992, ye-luo
Differential Revision: https://reviews.llvm.org/D125698
Summary:
This patch allows users to compile the static library without CUDA
installed on the system. This requires the new flag `--cuda-feature` to
indicate that we need `+ptx61` in order to compile the runtime.
This patch adds the necessary CMake configuration to build a static
library version of the device runtime, `libomptarget.devicertl.a`.
Various improvements in how we handle static libraries and generating
offloading code should allow us to treat the device library as a regular
project without needing to invoke the clang front-end directly. Here we
generate a job for each offloading architecture supported. Each
offloading architecture will be embedded into the static library and
used as-needed by the host.
This library will primarily be used to replace the bitcode library when
performing LTO. Currently, we need to manually pass in the bitcode
library which requires foreknowledge of the offloading architecture.
This approach lets us handle that in the linker wrapper instead.
Furthermore this should improve our interface to the device runtime. We
can now build it fully under a release build and have all the expected
entry points, as well as supporting debug builds.
Depends on D125265 D125256 D125260 D125314 D125563
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D125315
This patche attemps to address the current warnings in the OpenMP
offloading device runtime. Previously we did not see these because we
compiled the runtime without the standard warning flags enabled.
However, these warnings are used when we now build the static library
version of this runtime. This became extremely noisy when coupled with
the fact the we compile each file roughly 32 times when all the
architectures are considered. So it would be ideal to not have all these
warnings show up when building.
Most of these errors were simply implicit switch-case fallthroughs,
which can be addressed using C++17's fallthrough attribute. Additionally
there was a volatile variable that was being casted away. This is most
likely safe to remove because we cast it away before its even used and
didn't seem to affect anything in testing.
Depends on D125260
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D125339
Currently the OpenMP offloading device runtime is only expected to be
compiled for the specific architecture it's targeting. This is
problematic if we want to make compiling the device runtime more general
via the standar `clang` driver rather than invoking the clang front-end
directly. This patch addresses this by primarily changing the declare
type to `nohost` so the host will not contain any of this code.
Additionally we forward declare the functions that are defined via
variants, otherwise these would cause problems on the host.
Reviewed By: jdoerfert, tianshilei1992
Differential Revision: https://reviews.llvm.org/D125260
This patch adds the `llvm_omp_target_dynamic_shared_alloc` function to
the `omp.h` header file so users can access it by default. Also changed
the name to keep it consistent with the other target allocators. Added
some documentation so users know how to use it. Didn't add the interface
for Fortran since there's no way to test it right now.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D123246
In a clean build directory, `check-openmp` or `check-libomptarget` will fail because of missing device RTL .bc files. Ensure that the new targets new custom targets `omptarget.devicertl.nvptx` and `omptarget.devicertl.amdgpu` (corresponding to the plugin rtl targets `omptarget.rtl.cuda`, respectively `omptarget.rlt.amdgpu` ) are dependencies of the regression tests.
Reviewed By: JonChesterfield
Differential Revision: https://reviews.llvm.org/D123177
Inline assembly is scary but we need to support it for the OpenMP GPU
device runtime. The new assumption expresses the fact that it may not
have call semantics, that is, it will not call another function but
simply perform an operation or side-effect. This is important for
reachability in the presence of inline assembly.
Differential Revision: https://reviews.llvm.org/D109986
The unroll pragma did not properly work as the loop bound was not known
when we optimize the runtime and we then added a "unroll disable"
metadata which prevented unrolling later when the bounds were known.
For now we manually unroll to make sure up to 16 elements are handled
nicely. This helps optimizations to look through the argument passing.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D109164
`LIBOMPTARGET_LLVM_INCLUDE_DIRS` is currently checked and included for
multiple times redundantly. This patch is simply a clean up.
Reviewed By: jhuber6
Differential Revision: https://reviews.llvm.org/D121055
Libomptarget uses some shared variables to track certain internal stated
in the runtime. This causes problems when we have code that contains no
OpenMP kernels. These variables are normally initialized upon kernel
entry, but if there are no kernels we will see no initialization.
Currently we load the runtime into each source file when not running in
LTO mode, so these variables will be erroneously considered undefined or
dead and removed, causing miscompiles. This patch temporarily works
around the most obvious case, but others still exhibit this problem. We
will need to fix this more soundly later.
Fixes#54208.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D121007
The runtime uses thread state values to indicate when we use an ICV or
are in nested parallelism. This is done for OpenMP correctness, but it
not needed in the majority of cases. The new flag added is
`-fopenmp-assume-no-thread-state`.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D120106
The `IsSPMD` global can only be read by threads other than the main
thread *after* initialization is complete. To allow usage of
`mapping::getBlockSize` before initialization is done, we can pass the
`IsSPMD` state explicitly. This is similar to other APIs that take
`IsSPMD` explicitly to avoid such a race, e.g.,
`mapping::isInitialThreadInLevel0(IsSPMD)`
Fixes https://github.com/llvm/llvm-project/issues/53857
Currently whenever we compile the device runtime we get the following
'Mapping.cpp:32:32: warning: inline function '_OMP::impl::getGridValue'
is not defined [-Wundefined-inline]' warning. This can be silenced by
removing the constexpr attribute for this function. Doing this doesn't
change the generated bitcode at all but prevents the screen from getting
filled with warnings whenver we build the runtime.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D119747
This patch manually adds the runtime include files to the list of
dependencies when we build the bitcode runtime library. Previously if
only the header was changed we would not recompile the source files.
The solution used here isn't optimal because every source file not has a
dependency on each header file regardless of if it was actually used by
that file.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D119254
This patch replaces the ValueRAII pointer with a default 'nullptr'
value. Previously this was initialized as a reference to an existing
variable. The use of this variable caused overhead as the compiler could
not look through the uses and determine that it was unused if 'Active'
was not set. Because of this accesses to the variable would be left in
the runtime once compiled.
Fixes#53641
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D119187
This patch completely removes the old OpenMP device runtime. Previously,
the old runtime had the prefix `libomptarget-new-` and the old runtime
was simply called `libomptarget-`. This patch makes the formerly new
runtime the only runtime available. The entire project has been deleted,
and all references to the `libomptarget-new` runtime has been replaced
with `libomptarget-`.
Reviewed By: JonChesterfield
Differential Revision: https://reviews.llvm.org/D118934
Due to num_threads (probably also other reasons) we cannot assume
explicit barriers are always executed by all threads in an aligned
fashion. We can optimize them if that property can be proven but
that is different.
Johannes Doerfert