Improve the Neon vector permute intrinsics.
The Neon vector permute intrinsics namely
vtrn
vtbl
vtbx
vrev32/64/16
vzip
vuzp
vext
can be implemented using the generic __builtin_shuffle extension mechanism. This will also benefit and make sure that all tests in this area work properly.
This is something that came up only in 4.7 and should have a reasonably high priority as it can improve our vectorization numbers in quite a few cases and end up improving
quite a few cases in the intrinsics as we now have a way of telling the vectorizer and the generic optimizers that this really isn't a builtin_call but just a vector permute operation :)
Blueprint information
- Status:
- Complete
- Approver:
- Michael Hope
- Priority:
- Medium
- Drafter:
- Ramana Radhakrishnan
- Direction:
- Approved
- Assignee:
- Ramana Radhakrishnan
- Definition:
- Approved
- Series goal:
- Accepted for 4.7
- Implementation:
-
Implemented
- Milestone target:
- None
- Started by
- Ramana Radhakrishnan
- Completed by
- Matthew Gretton-Dann
Related branches
Related bugs
Sprints
Whiteboard
[2013-05-22] This has been completed and is in FSF GCC 4.8
Work Items
Work items:
* Improve vrev64/32/16 , vzip, vuzp, vtrn using __builtin_shuffle Commit upstream: DONE
* Implement support for __builtin_shuffle in the C++ frontend : DONE
* Implement support for constexpr and __builtin_shuffle in C++ frontend : DONE
* Improve support for vdup intrinsics in the backend : INPROGRESS
* Improve neon intrinsics tests : INPROGRESS
* Improve support for vld3 intrinsics in the backend- Do not stack values to and from the stack gratuitously. Done by tweaking costs to allow lower-subreg only to lower values you want : INPROGRESS
* Validate improvements obtained : TODO
* Backport to Linaro 4.7: TODO
