Re: Mode change for bswap pattern expansion

[Richard Sandiford <rdsandiford@googlemail.com>]

Paulo Matos <pmatos@broadcom.com> writes:
> On a vector processor we can do a bswapsi with two instructions, by first rotating half-words (16 bits) by 8 and then rotating full words by 16. 
> However, this means expanding:
> (set (match_operand:SI 0 "register_operand" "")
>      (bswap:SI (match_operand:SI 1 "register_operand" "")))
>
> to:
> (set (match_dup:V2HI 0)
>      (rotate:V2HI (match_dup:V2HI 1)
>                   (const_int 8)))
> (set (match_dup:SI 0)
>      (rotate:SI (match_dup:SI 0)
>                 (const_int 16)))
>
> This is obviously not correct, because match_dup cannot set the mode. The point I am trying to make is that I can't find a good way to deal with the mode changes. I don't think GCC is too happy if I change the modes of the same operand from one instruction to the other right? The only other way is to emit paradoxical subregs. So something along these lines:
> (set (subreg:V2HI (match_dup 0) 0)
>      (rotate:V2HI (subreg:V2HI (match_dup 1) 0)
>                   (const_int 8)))
> (set (match_dup 0)
>      (rotate:SI (match_dup 0)
>                 (const_int 16)))

It's usually better not to hard-code the subregs in the pattern.
Instead you could use C code to create the subregs, e.g.:

  [(set (match_dup 3)
        (rotate:V2HI (match_dup 2)
                     (const_int 8)))
   (set (match_dup 0)
        (rotate:SI (match_dup 4)
                   (const_int 16)))]
  ""
{
  operands[2] = gen_lowpart (V2HImode, operands[1]);
  operands[3] = gen_reg_rtx (V2HImode);
  operands[4] = gen_lowpart (SImode, operands[3]);
}

so that any hard regs are correctly handled.  Or it might be easier to code
it using emit_insn (gen_* (...))s instead.

BTW, paradoxical subregs are where the outer mode is strictly larger
than the inner mode.

MIPS uses essentially the same sequence, except that it has a special
instruction to do the first rotate (WSBH), rather than it being an instance
of a general vector rotate.  For MIPS we just model it as an unspec SImode
operation.  Maybe that would be easier here too.

Thanks,
Richard