| Issue |
123201
|
| Summary |
[LV][EVL] Support interleaved accesses for EVL tail folding.
|
| Labels |
vectorizers
|
| Assignees |
Mel-Chen
|
| Reporter |
Mel-Chen
|
The motivation for this issue is to provide better support for RVV unit-strided segment load/store.
The following scenarios need to be supported:
* Interleaved load (vp.load + interleave)
* Interleaved load with tail gaps (Requires scalar epilogue to run the last iteration)
* Fully interleaved store (deinterleave + vp.store)
* Interleaved store with gaps (This can not emit unit-strided segment store. We can only emit a wide masked store for that)
Due to the high complexity of `VPInterleaveRecipe::execute()`, creating a new recipe or converting it into `VPWidenIntrinsicRecipe` does not seem like a wise approach.
A tentative approach I have in mind is to first split `VPInterleaveRecipe` into `VPWidenLoad + VPDeinterleave` and `VPInterleave + VPWidenStore`. During the EVL lowering phase, we would only need to transform `VPWidenLoad/VPWidenStore` into `VPWidenLoadEVL/VPWidenStoreEVL`.
For now, the focus will be on supporting factor 2 (`interleave2/deinterleave2`) as the initial target, with support for factors 3 to 8 planned after test results are stable.
Related IAP support: #120490 .
![]()
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
