| Issue |
208502
|
| Summary |
[RISCV] Remove performReverseEVLCombine and performVP_STORECombine
|
| Labels |
backend:RISC-V
|
| Assignees |
|
| Reporter |
lukel97
|
These combines convert a reverse + load/store into a strided load/store with stride -1. However a `vlse.v` or `vsse.v` isn't always profitable, and depends on the target and types. We should remove these combines and instead do this as a cost driven transform in the loop vectorizer, see https://github.com/llvm/llvm-project/issues/204154
Below are the reciprocal throughput measurements on a bpi-f3 for each approach:
| EEW | LMUL | VL | A: vle+vrgather | B: vlse(-stride) | Speedup B/A | Highest throughput |
|----:|-----:|----:|----------------:|-----------------:|:-----------:|:-------|
| 8 | 1 | 32 | 5.0 | 32.0 | 0.16x | A |
| 8 | 2 | 64 | 16.0 | 64.0 | 0.25x | A |
| 8 | 4 | 128 | 64.0 | 128.0 | 0.50x | A |
| 8 | 8 | 256 | 263.0 | 256.0 | 1.03x | tie |
| 16 | 1 | 16 | 5.0 | 16.0 | 0.31x | A |
| 16 | 2 | 32 | 16.0 | 32.0 | 0.50x | A |
| 16 | 4 | 64 | 64.0 | 64.0 | 1.00x | tie |
| 16 | 8 | 128 | 263.0 | 128.0 | 2.05x | B |
| 32 | 1 | 8 | 5.0 | 8.0 | 0.63x | A |
| 32 | 2 | 16 | 16.0 | 16.0 | 1.00x | tie |
| 32 | 4 | 32 | 64.0 | 32.0 | 2.00x | B |
| 32 | 8 | 64 | 263.0 | 64.0 | 4.11x | B |
| 64 | 1 | 4 | 5.0 | 4.0 | 1.25x | B |
| 64 | 2 | 8 | 16.0 | 8.0 | 2.00x | B |
| 64 | 4 | 16 | 64.0 | 16.0 | 4.00x | B |
| 64 | 8 | 32 | 263.0 | 32.0 | 8.22x | B |
| EEW | LMUL | VL | A: vrgather+vse | B: vsse(-stride) | Speedup B/A | Highest profitable |
|----:|-----:|----:|----------------:|-----------------:|:-----------:|:-------|
| 8 | 1 | 32 | 5.0 | 32.0 | 0.16x | A |
| 8 | 2 | 64 | 16.0 | 64.0 | 0.25x | A |
| 8 | 4 | 128 | 64.0 | 128.0 | 0.50x | A |
| 8 | 8 | 256 | 256.0 | 256.0 | 1.00x | tie |
| 16 | 1 | 16 | 5.0 | 16.0 | 0.31x | A |
| 16 | 2 | 32 | 16.0 | 32.0 | 0.50x | A |
| 16 | 4 | 64 | 64.0 | 65.0 | 0.98x | tie |
| 16 | 8 | 128 | 256.0 | 130.3 | 1.97x | B |
| 32 | 1 | 8 | 5.0 | 8.0 | 0.63x | A |
| 32 | 2 | 16 | 16.0 | 16.0 | 1.00x | tie |
| 32 | 4 | 32 | 64.0 | 32.0 | 2.00x | B |
| 32 | 8 | 64 | 256.0 | 64.0 | 4.00x | B |
| 64 | 1 | 4 | 5.0 | 5.9 | 0.85x | A |
| 64 | 2 | 8 | 16.0 | 12.0 | 1.33x | B |
| 64 | 4 | 16 | 64.0 | 16.0 | 4.00x | B |
| 64 | 8 | 32 | 256.0 | 32.0 | 8.00x | B |
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
