`-fgpu-rdc` mode allows device functions call device functions in
different TU. However, currently all device objects have to be linked
together since only one fat binary is supported. This is time consuming
for AMDGPU backend since it only supports LTO.
There are use cases that objects can be divided into groups in which
device functions are self-contained but host functions are not. It is
desirable to link/optimize/codegen the device code and generate a fatbin
for each group, whereas partially link the host code with `ld -r` or
generate a static library by using the `--emit-static-lib` option of
clang. This avoids linking all device code together, therefore decreases
the linking time for `-fgpu-rdc`.
Previously, clang emits an external symbol `__hip_fatbin` for all
objects for `-fgpu-rdc`. With this patch, clang emits an unique external
symbol `__hip_fatbin_{cuid}` for the fat binary for each object. When a
group of objects are linked together to generate a fatbin, the symbols
are merged by alias and point to the same fat binary. Each group has its
own fat binary. One executable or shared library can have multiple fat
binaries. Device linking is done for undefined fab binary symbols only
to avoid repeated linking. `__hip_gpubin_handle` is also uniquefied and
merged to avoid repeated registering. Symbol `__hip_cuid_{cuid}` is
introduced to facilitate debugging and tooling.
Fixes: https://github.com/llvm/llvm-project/issues/77018
This adds -no-opaque-pointers to clang tests whose output will
change when opaque pointers are enabled by default. This is
intended to be part of the migration approach described in
https://discourse.llvm.org/t/enabling-opaque-pointers-by-default/61322/9.
The patch has been produced by replacing %clang_cc1 with
%clang_cc1 -no-opaque-pointers for tests that fail with opaque
pointers enabled. Worth noting that this doesn't cover all tests,
there's a remaining ~40 tests not using %clang_cc1 that will need
a followup change.
Differential Revision: https://reviews.llvm.org/D123115
- ``externally_initialized`` variables would be initialized or modified
elsewhere. Particularly, CUDA or HIP may have host code to initialize
or modify ``externally_initialized`` device variables, which may not
be explicitly referenced on the device side but may still be used
through the host side interfaces. Not preserving them triggers the
elimination of them in the GlobalDCE and breaks the user code.
Reviewed By: yaxunl
Differential Revision: https://reviews.llvm.org/D105135
variables emitted on both host and device side with different addresses
when ODR-used by host function should not cause device side counter-part
to be force emitted.
This fixes the regression caused by https://reviews.llvm.org/D102237
Reviewed by: Artem Belevich
Differential Revision: https://reviews.llvm.org/D102801
Currently clang does not emit device template variables
instantiated only in host functions, however, nvcc is
able to do that:
https://godbolt.org/z/fneEfferY
This patch fixes this issue by refactoring and extending
the existing mechanism for emitting static device
var ODR-used by host only. Basically clang records
device variables ODR-used by host code and force
them to be emitted in device compilation. The existing
mechanism makes sure these device variables ODR-used
by host code are added to llvm.compiler-used, therefore
they are guaranteed not to be deleted.
It also fixes non-ODR-use of static device variable by host code
causing static device variable to be emitted and registered,
which should not.
Reviewed by: Artem Belevich
Differential Revision: https://reviews.llvm.org/D102237
Add device variables to llvm.compiler.used if they are
ODR-used by either host or device functions.
This is necessary to prevent them from being
eliminated by whole-program optimization
where the compiler has no way to know a device
variable is used by some host code.
Reviewed by: Artem Belevich
Differential Revision: https://reviews.llvm.org/D98814
