Equivalent of PEXT instruction on ARM

Question

Do ARMv7 and/or ARMv8 instruction set provide an instruction similar to PEXT on x86?

If not, what is the most efficient instructions combination to achieve the same behavior?

There is no such instruction and unfortunately `pext` is rather difficult to emulate in the general case. If you have a specific use case in mind, I might be able to suggest an efficient work around. — fuz, Nov 18 '21 at 17:09
Unfortunately [Anderson's "Bit Twiddling Hacks"](https://graphics.stanford.edu/~seander/bithacks.html) doesn't include this operation :-( — Nate Eldredge, Nov 18 '21 at 17:52
I suppose one could look at how a translator like Rosetta does it. — Nate Eldredge, Nov 18 '21 at 18:10
@NateEldredge Rosetta does not support this instruction as far as I know. — fuz, Nov 18 '21 at 18:42
`BFC`/`BFI` and `UBFX` come to mind, however not nearly as powerful as `PEXT`, you would need multiple of those and possibly some and-masking to emulate a `PEXT` with a *given known mask*. — Marco Bonelli, Nov 18 '21 at 18:43
Rosetta does indeed crash on `pext` instructions with SIGILL. But if you have any C compiler that emits the instruction for a particular input code (without explicitly using intrinsics), you could try and see what it generates for ARM? — Siguza, Nov 18 '21 at 19:00
There's a possible algorithm at https://stackoverflow.com/questions/21144237/standard-c11-code-equivalent-to-the-pext-haswell-instruction-and-likely-to-be — Nate Eldredge, Nov 19 '21 at 00:48
@Siguza: From a quick grep of gcc and clang source code, it's not clear to me that either compiler is capable of emitting `pext` under any circumstances, except from calling an intrinsic. — Nate Eldredge, Nov 19 '21 at 00:55

score 0 · Answer 1 · edited Jan 21 '22 at 18:07

This is my implementation

    // w14 = pext(val: w13, mask: w15)
    // Used Regs x12-x15
    // (w13 & w15 will BROKEN after the loop)
    mov     w14, wzr;
.loop:
    cbz     w15, .pext_end;   // may accelerate by prepare with (w13 & w15)
    clz     w12, w15
    lsl     w13, w13, w12
    lsl     w15, w15, w12
    extr    w14, w14, w13, #31
    bfc     w15, #31, #1
    b      .loop
.pext_end:

Equivalent of PEXT instruction on ARM

1 Answers1