| Issue |
178792
|
| Summary |
[X86] Missed fold: `vgf2p8affineqb` + 8-bit shifts
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
WalterKruger
|
Performing a 8-bit shift on `vgf2p8affineqb` can be folded by together by shifting its matrix by 8 times the shift amount in the opposite direction. The position of each matrix byte determines a corresponding bit in the output, in reverse order, so shifting the matrix is equivalent. (Further reading: https://wunkolo.github.io/post/2020/11/gf2p8affineqb-int8-shifting/)
This applies to logical, rotational, and arithmetic shifts in either direction (but with the arithmetic shifting-in the entire least significant byte rather then LS bit). Only seem beneficial when the matrix and shift amount are constant.
```asm
bitRevSHL1_clang:
gf2p8affineqb xmm0, xmmword ptr [rip + .LCPI0_0], 0
paddb xmm0, xmm0
ret
```
```asm
bitRevSHL1_tgt:
gf2p8affineqb xmm0, xmmword ptr [rip + .LCPI0_0], 0
ret
```
https://godbolt.org/z/rd7frehEd
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
