[NFI]: XPUBackend refactoring to facilitate arch-specific implementations #5329
Conversation
etiotto
changed the title
XPUBackend is the base class. Arch-specific implementations derive from it and override the required methods. For example, they could override
It could be done, by defining a metaclass, but the inheritance should be used in most use-cases. |
AndreyPavlenko
force-pushed
the
AndreyPavlenko/backend-refactor
branch
from
2c0c89d to
055acd2
Compare
| return pm | ||
|
|
||
| @classmethod | ||
| def get_ttir_passes(cls, opt): |
There was a problem hiding this comment.
How about we separate the changes about get_*ir_passes in another PR? I am not sure I like it, but the other changes don't need to be blocked by it.
- Allow metaclasses injection: use XPUBackendMeta from the .meta submodule as the metaclass, if it exists. - Try to find an arch-specific implementation in the .arch.<name> submodule. - Create the list of passes in separate methods to allow subclasses to modify it.
AndreyPavlenko
force-pushed
the
AndreyPavlenko/backend-refactor
branch
from
055acd2 to
abbf291
Compare
| def make_ttir(mod, metadata, opt): | ||
| pm = ir.pass_manager(mod.context) | ||
| @classmethod | ||
| def create_pass_manager(cls, context): |
There was a problem hiding this comment.
what's the benefit of create_pass_manager? is it going to be different for other arch?
There was a problem hiding this comment.
In the initial version it also added passes to the pm - 65c9f8a#diff-5ec0c116ff4361ab1b5db32636583aaf2da599bda46558aab877dc5670a77a39R206 . I've removed this functionality, but kept the method to save on duplication of 2 lines of code :)
There was a problem hiding this comment.
IMO it is ok not to save 2 lines of code duplication in exchange for easy comparison with other backends.
The purpose of this refactoring is to simplify development for new architectures by subclassing the base class instead of adding conditional code blocks.
The main changes:
.arch.<name>submodule.