Currently we use the `embedBufferInModule` function to store binary
strings containing device offloading data inside the host object to
create a fatbinary. In the case of LTO, we need to extract this object
from the LLVM-IR. This patch adds a metadata node for the embedded
objects containing the embedded pointers and the sections they were
stored at. This should create a cleaner interface for identifying these
values.
In the future it may be worthwhile to also encode an `ID` in the
metadata corresponding to the object's special section type if relevant.
This would allow us to extract the data from an object file and LLVM-IR
using the same ID.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D129033
We use LLD to perform AMDGPU linking. This linker accepts some arguments
through the `-plugin-opt` facilities. These options match what `Clang`
will output when given the same input.
Reviewed By: yaxunl
Differential Revision: https://reviews.llvm.org/D128923
Summary:
This patch adds some new sanity checks to make sure that the sizes of
the offsets are within the bounds of the file or what is expected by the
binary. This also improves the error handling of the version structure
to be built into the binary itself so we can change it easier.
The target features are necessary for correctly compiling most programs
in LTO mode. Currently, these are derived in clang at link time and
passed as an arguemnt to the linker wrapper. This is problematic because
it requires knowing the required toolchain at link time, which should
not be necessry. Instead, these features should be embedded into the
offloading binary so we can unify them in the linker wrapper for LTO.
This also required changing the offload packager to interpret multiple
arguments as concatenation with a comma. This is so we can still use the
`,` separator for the argument list.
Depends on D127246
Reviewed By: tra
Differential Revision: https://reviews.llvm.org/D127686
Summary:
Currently we use temporary files to write the intermediate results to.
However, these are stored as regular strings and we do a few unnecessary
copies and conversions of them. This patch simply replaces these strings
with a reference to the filename stored in the list of temporary files.
The temporary files will stay alive during the whole linking phase and
have stable pointers, so we should be able to cheaply pass references to
them rather than copying them every time.
Summary:
A recent patch added some new code paths to the linker wrapper. Older
compilers seem to have problems with returning errors wrapped in
an Excepted type without explicitly moving them. This caused failures in
some of the buildbots. This patch fixes that.
The linker wrapper currently eagerly extracts all identified offloading
binaries to a file. This isn't ideal because we will soon open these
files again to examine their symbols for LTO and other things.
Additionally, we may not use every extracted file in the case of static
libraries. This would be very noisy in the case of static libraries that
may contain code for several targets not participating in the current
link.
Recent changes allow us to treat an Offloading binary as a standard
binary class. So that allows us to use an OwningBinary to model the
file. Now we keep it in memory and only write it once we know which
files will be participating in the final link job. This also reworks a
lot of the structure around how we handle this by removing the old
DeviceFile class.
The main benefit from this is that the following doesn't output 32+ files and
instead will only output a single temp file for the linked module.
```
$ clang input.c -fopenmp --offload-arch=sm_70 -foffload-lto -save-temps
```
Reviewed By: JonChesterfield
Differential Revision: https://reviews.llvm.org/D127246
Summary:
The linker wrapper supports embedding bitcode images instead of linked
device images to facilitate JIT in the device runtime. However, we were
incorrectly passing in the file twice when this option was set. This
patch makes sure we only use the intermediate result of the LTO pass and
don't add the final output to the full job.
In the future we will want to add both of these andle handle that
accoridngly to allow the runtime to either use the AoT compiled version
or JIT compile the bitcode version if availible.
We use the clang-linker-wrapper to perform device linking of embedded
offloading object files. This is done by generating those jobs inside of
the linker-wrapper itself. This patch adds an argument in Clang and the
linker-wrapper that allows users to forward input to the device linking
phase. This can either be done for every device linker, or for a
specific target triple. We use the `-Xoffload-linker <arg>` and the
`-Xoffload-linker-<triple> <arg>` syntax to accomplish this.
Reviewed By: markdewing, tra
Differential Revision: https://reviews.llvm.org/D126226
Summary:
Static libraries need to be handled differently from regular inpout
files, namely they are loaded lazily. Previously we used a flag to
indicate a file camm from a static library. This patch simplifies this
by simply keeping a different array that contains the static libraries
so we don't need to parse them out again.
Summary:
The linker wrapper previously had functionality to strip the sections
manually. We don't use this at all because this is much better done by
the linker via the `SHF_EXCLUDE` flag. This patch simply removes the
support for thi sfeature to simplify the code.
Summary:
We use embedded binaries to extract offloading device code from the host
fatbinary. This uses a binary format whose necessary alignment is
eight bytes. The alignment is included within the ELF section type so
the data extracted from the ELF should always be aligned at that amount.
However, if this file was extraqcted from a static archive, it was being
sent as an offset in the archive file which did not have the same
alignment guaruntees as the ELF file. This was causing errors in the
UB-sanitizer build as it would occasionally try to access a misaligned
address. To fix this, I simply copy the memory directly to a new buffer
which is guarnteed to have worst-case alignment of 16 in the case that
it's not properly aligned.
This patch adds the necessary code generation to create the wrapper code
that registers all the globals in CUDA. We create the necessary
functions and iterate through the list of
`__start_cuda_offloading_entries` to find which globals must be
registered. This is very similar to the code generation done currently
in Clang for non-rdc builds, but here we are registering a fully linked
fatbinary and finding the globals via the above sections.
With this we should be able to fully support basic RDC / LTO building of CUDA
code.
It's also worth noting that this does not include the necessary PTX to JIT the
image, so to use this support the offloading architecture must match the
system's architecture.
Depends on D123810
Reviewed By: tra
Differential Revision: https://reviews.llvm.org/D123812
This patch adds the initial support for wrapping CUDA images. This
requires changing some of the logic for how we bundle images. We now
need to copy the image for all kinds that are active for the
architecture. Then we need to run a separate wrapping job if the Kind is
Cuda. For cuda wrapping we need to use the `fatbinary` program from the
CUDA SDK to bundle all the binaries together. This is then passed to a
new function to perfom the actual module code generation that will be
implemented in a later patch.
Depends on D120273 D123471
Reviewed By: tra
Differential Revision: https://reviews.llvm.org/D123810
Currently we handle static libraries like any other object in the
linker wrapper. However, this does not preserve the sematnics that
dictate static libraries should be lazily loaded as the symbols are
needed. This allows us to ignore linking in architectures that are not
used by the main application being compiled. This patch adds the basic
support for detecting if a file came from a static library, and only
including it in the link job if it's used by other object files.
This patch only adds the basic support, to be more correct we should
check the symbols and only inclue the library if the link job contains
symbols that are needed. Ideally we could just put this on the linker
itself, but nvlink doesn't seem to support `.a` files.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D125092
Summary:
A previous patch merged the command execution and printing into a helper
function. The old printing code wasn't removed causing each to be
printed twice.
After basic support for embedding and handling CUDA files was added to
the new driver, we should be able to call CUDA functions from OpenMP
code. This patch makes the necessary changes to successfuly link in CUDA
programs that were compiled using the new driver. With this patch it
should be possible to compile device-only CUDA code (no kernels) and
call it from OpenMP as follows:
```
$ clang++ cuda.cu -fopenmp-new-driver -offload-arch=sm_70 -c
$ clang++ openmp.cpp cuda.o -fopenmp-new-driver -fopenmp -fopenmp-targets=nvptx64 -Xopenmp-target=nvptx64 -march=sm_70
```
Currently this requires using a host variant to suppress the generation
of a CPU-side fallback call.
Depends on D120272
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D120273
Summary:
A previous patch updated the path searching in the linker wrapper. I
made an error and caused `lld`, which is necessary to link AMDGPU
images, to not be found on some systems. This patch fixes this by
correctly searching that linker-wrapper's binary path first again.
When we do LTO we consider ourselves to have whole program visibility if
every single input file we have contains LLVM bitcode. If we have whole
program visibliity then we can create a single image and utilize CUDA's
non-RDC mode by not passing `-c` to `ptxas` and ignoring the `nvlink`
job. This should be faster for some situations and also saves us the
time executing `nvlink`.
Reviewed By: tra
Differential Revision: https://reviews.llvm.org/D124292
The linker wrapper is used to perform linking and wrapping of embedded
device object files. Currently its internals are not able to be tested
easily. This patch adds the `--dry-run` and `--print-wrapped-module`
options to investigate the link jobs that will be run along with the
wrapped code that will be created to register the binaries.
Reviewed By: JonChesterfield
Differential Revision: https://reviews.llvm.org/D124039
The previous patch introduced the offloading binary format so we can
store some metada along with the binary image. This patch introduces
using this inside the linker wrapper and Clang instead of the previous
method that embedded the metadata in the section name.
Differential Revision: https://reviews.llvm.org/D122683
Summary:
The changes in D122987 ensures that the offloading sections always have
the SHF_EXCLUDE flag. This means that we do not need to manually strip
these sections for ELF or COFF targets.
Summary:
A previous patch added the option to use the default pipeline when
perfomring LTO rather than the regular LTO pipeline. This greatly
improved performance regressions we were observing with the LTO
pipeline. However, this should not be used if the user explicitly
disables optimizations as the default pipeline expects some
optimizatoins to be perfomed.
Summary:
Currently there is no option to configure the number of thin-backend
threads to use when performing thin-lto on the device, but we should
default to use all the threads rather than just one. In the future we
should use the same arguments that gold / lld use and parse it here.
This patch adds a configuration option to simply use the default pass
pipeline in favor of the LTO-specific one. We observed some severe
performance penalties when uding device-side LTO for OpenMP offloading
applications caused by the LTO-pass pipeline. This is primarily because
OpenMP uses an LLVM bitcode library to implement a GPU runtime library.
In a standard compilation we link this bitcode library into each source
file and optimize it with the default pipeline. When performing LTO we
link it late with all the files, but the bitcode library never has the
regular optimization pipeline applied to it so we miss a few
optimizations just using the LTO pipeline to optimize it.
I'm not committed to this solution, but it's the easiest method to solve
this performance regression when using LTO without changing the
optimizatin pipeline for other users.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D122133
This patch adds the offload kind to the embedded section name in
preparation for offloading to different kinda like CUDA or HIP.
Depends on D120288
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D120271
This patch implements a DenseMap info struct for the device file type.
This is used to help grouping device files that have the same triple and
architecture. Because of this the filename, which will always be unique
for each file, is not used.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D120288
Summary:
This patch correctly handles the `--sysroot=` option when passed to the
linker wrapper. This allows users to correctly find libraries that may
contain offloading code if using this option.
Summary;
This path adds printing support for the linker wrapper. When the user
passes `-v` it will not print the commands used by the linker wrapper to
indicate to the user what is happening during the linking.
Summary:
This patch removes the error we recieve when attempting to extract
offloading sections. We shouldn't consider this a failure because
extracting bitcode isn't necessarily required.
Summary:
We were not previously saving strings when saving symbol names during
LTO symbol resolution. This caused a crash inside the dense set when
some of the strings would rarely be moved internally by the object file
class.
This patch adds support for linking CPU offloading applications in the
linker wrapper. We generate the necessary linking job using the host
linker's path and library arguments. This may not be true for more
complex offloading schemes, but this is sufficient for now.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D119613
Summary:
This patch changes the ClangLinkerWrapper to use the executable path
when searching for the lld binary. Previously we relied on the program
name. Also not finding 'llvm-strip' is not considered an error anymore
because it is an optional optimization.
Add the build directory to the search path for llvm-strip instead
of solely relying on the PATH environment variable setting.
Reviewed By: jhuber6
Differential Revision: https://reviews.llvm.org/D118965
The linker wrapper tool uses the 'nvlink' and 'ptxas' binaries to link
and assemble device files. Previously we searched for this using the
binaries in the user's path. This didn't work in cases where the user
passed in a specific Cuda path to Clang. This patch changes the linker
wrapper to accept an argument for the Cuda path we can get from Clang.
This should fix#53573.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D118944
Summary:
This patch removes the system call to the `clang-offload-wrapper` tool
by replicating its functionality in a new file. This improves
performance and makes the future wrapping functionality easier to
change.
Differential Revision: https://reviews.llvm.org/D118198
Summary:
This patch replaces the system call to the `llc` binary with a library
call to the target machine interface. This should be faster than
relying on an external system call to compile the final wrapper binary.
Differential Revision: https://reviews.llvm.org/D118197
Summary:
This parses the executable name out of the linker arguments so we can
use it to give more informative temporary file names and so we don't
accidentally use it for device linking.
Summary:
This patch implements the `-save-temps` flag for the linker wrapper.
This allows the user to inspect the intermeditary outpout that the
linker wrapper creates.
Summary:
Various changes to the linker wrapper, and the bitcode embedding is not
done after the optimizations have run rather than after linking is done.
This saves time when doing JIT.
This patch improves the symbol resolution done for LTO with offloading
applications. The symbol resolution done here allows the LTO backend to
internalize more functions. The symbol resoltion done is a simplified
view that does not take into account various options like `--wrap` or
`--dyanimic-list` and always assumes we are creating a shared object.
The actual target may be an executable, but semantically it is used as a
shared object because certain objects need to be visible outside of the
executable when they are read by the OpenMP plugin.
Depends on D117246
Differential Revision: https://reviews.llvm.org/D118155
This patch adds support for linking AMDGPU images using the LLD binary.
AMDGPU files are always bitcode images and will always use the LTO
backend. Additionally we now pass the default architecture found with
the `amdgpu-arch` tool to the argument list.
Depends on D117156
Differential Revision: https://reviews.llvm.org/D117246
