Summary:
The 'new driver' sets up the lifetime of a registered liftime using
global constructors and destructors. Currently, this is put at priority
1 which isn't strictly conformant as it will conflict with system
utilities. We now use 101 as this is the loweest suggested for
non-system constructors and will still run before user constructors.
Secondly, there were issues with the CUDA runtime when destructed with a
global destructor. Because the global ones are in any order and
potentially run before other things we were hitting an edge case where
the OpenMP runtime was uninitialized *after* `_dl_fini` was called. This
would result in us erroring when we call into a destroyed `libcuda.so`
instance. using `atexit` is what CUDA / HIP use and it prevents this
from happening. Most everything uses `atexit` except system utilities
and because of the constructor priority it will be unregistered *after*
everything else but not after `_fl_fini`.
Summary:
The standard GPU compilation process embeds each intermediate object
file into the host file at the `.llvm.offloading` section so it can be
linked later. We also use a special section called something like
`omp_offloading_entries` to store all the globals that need to be
registered by the runtime. The linker-wrapper's job is to link the
embedded device code stored at this section and then emit code to
register the linked image and the kernels and globals in the offloading
entry section.
One downside to RDC linking is that it can become quite big for very
large projects that wish to make use of static linking. This patch
changes the support for relocatable linking via `-r` to support a kind
of "partial" RDC compilation for offloading languages.
This primarily requires manually editing the embedded data in the
output object file for the relocatable link. We need to rename the
output section to make it distinct from the input sections that will be
merged. We then delete the old embedded object code so it won't be
linked further. We then need to rename the old offloading section so
that it is private to the module. A runtime solution could also be done
to defer entries that don't belong to the given GPU executable, but this
is easier. Note that this does not work with COFF linking, only the ELF
method for handling offloading entries, that could be made to work
similarly.
Given this support, the following compilation path should produce two
distinct images for OpenMP offloading.
```
$ clang foo.c -fopenmp --offload-arch=native -c
$ clang foo.c -lomptarget.devicertl --offload-link -r -o merged.o
$ clang main.c merged.o -fopenmp --offload-arch=native
$ ./a.out
```
Or similarly for HIP to effectively perform non-RDC mode compilation for
a subset of files.
```
$ clang -x hip foo.c --offload-arch=native --offload-new-driver -fgpu-rdc -c
$ clang -x hip foo.c -lomptarget.devicertl --offload-link -r -o merged.o
$ clang -x hip main.c merged.o --offload-arch=native --offload-new-driver -fgpu-rdc
$ ./a.out
```
One question is whether or not this should be the default behavior of
`-r` when run through the linker-wrapper or a special option. Standard
`-r` behavior is still possible if used without invoking the
linker-wrapper and it guaranteed to be correct.
This patch moves `clang/tools/clang-linker-wrapper/OffloadWrapper.*` to
`llvm/Frontend/Offloading` allowing them to be re-utilized by other
projects.
Additionally, it makes minor modifications to the API to make it more
flexible.
Concretely:
- The `wrap*` methods now have additional arguments `EntryArray`,
`Suffix` and `EmitSurfacesAndTextures` to specify some additional options.
- The `EntryArray` is now constructed by the caller. This change is needed to
enable JIT compilation, as ORC doesn't fully support `__start_` and `__stop_`
symbols. Thus, to JIT the code, the `EntryArray` has to be constructed explicitly in the IR.
- The `Suffix` field is used when emitting the descriptor, registration
methods, etc, to make them more readable. It is empty by default.
- The `EmitSurfacesAndTextures` field controls whether to emit surface
and texture registration code, as those functions were removed from `CUDART`
in CUDA 12. It is true by default.
- The function `getOffloadingEntryInitializer` was added to help create
the `EntryArray`, as it returns the constant initializer and not a global
variable.
