design
design copied to clipboard
WebAssembly
Bit reversing instructions
Future features mentioned a nice set for instruction set future improvements. For example bswap and bswap16 but missing bireverse/bitrev/rbit operations which I think quite important due to it could be useful for wide range of compression algorithms, Cooley–Tukey's FFT, FHT, dyadic rational numbers and etc.
Clang has builtins for that:
__builtin_bitreverse8
__builtin_bitreverse16
__builtin_bitreverse32
__builtin_bitreverse64
which quite useful due to implement bit reversing is quite challenging and some architectures like ARM64 could be lowered to single instruction RBIT.
So it would be interesting to hear your opinion on whether to add bitrev to post-MVP?
Proposed unary instructions:
i32.bitrev8
i32.bitrev16
i32.bitrev
i64.bitrev
Implementation details for runtimes
For x64 / ARMv7 / RISCV / MIPS / POWER and etc:
- i32.bitrev8: one read from 8-bit lookup table
- i32.bitrev16: two reads from 8-bit lookup table
- i32.bitrev: four reads from 8-bit lookup table
- i64.bitrev: use Knuth's method from Hacker's delight or 8 lookups if it faster on target architecture.
For more modern x64 architecture like Zen3 and Skylake bit reversing could be implemented via PDEP / PEXT instructions.
For ARM64:
- i32.bitrev8:
rbit w1, w0 lsr w0, w1, #24 - i32.bitrev16:
rbit w1, w0 lsr w0, w1, #16 - i32.bitrev:
rbit w0, w1 - i64.bitrev:
rbit x0, x1
I second this. SHR, SHR, AND, and BSWAP will be of help on the X86 implementation. Bit reversal is really useful, but really expensive in the absence of a competent primitive.
@MaxGraey Are you still there? Perhaps we should propose a discussion of this at a future community group meeting.
Related: register exchange instructions. Let's double check that these are already well supported.
Perhaps we should propose a discussion of this at a future community group meeting.
That's will be great.
@MaxGraey Cool, care to pick a date? Ideally one with little or nothing already slated for discussion.
https://github.com/WebAssembly/meetings/tree/master/main/2021
If you have the opportunity, could you take it upon yourself? In the near future, I will be too busy to participate in meeting calls and try to propose this to WG as post-MVP op codes.
