On Jul  6, 2004, Richard Henderson <rth@redhat.com> wrote:

> (1) Rename the existing rtl movstr patterns to movmem.

Can do.  Patch is attached.  Will test i686-pc-linux-gnu native
overnight.  Ok to install if it passes?

> (2) Create new movstr patterns that mirror how cmpstr patterns operate.
> (3) Expand strcpy/strpcpy like we do strcmp elsewhere in builtins.c.

At first I'd got the impression h8sx didn't have an actual stpcpy
instruction but, as it turns out, Richard Sandiford actually
introduced a pattern that uses movsd and a conditional branch for
that.

On second thought, I realized I'm not sure how to best approach
introducing movstr.  I suppose on most machines that can use these
patterns, we'd have some register modified in such a way that at least
movstr is pretty close to the requirements of stpcpy(), and
implementing strcpy() in terms of that would be quite trivial (just
refrain from using the modified register, and use the original one).

However, even if it's close to stpcpy(), some arches will have the
final register point one-past the NUL terminator, whereas others might
have it point to the NUL terminator itself.  I suppose we should
probably model movstr exactly after stpcpy(), even if that takes an
additional increment/decrement instruction to get the right value.
This will make the pattern useful for stpcpy and strcpy, and the
additional instruction will likely be optimized away if unneeded, or
combined with other insns, unless it has the needed effect anyway.

So, should movstr be modeled exactly after stpcpy()?  With a single
output operand for end-of-dest, or another for end-of-src?  I don't
see that end-of-src could be used with the code we have today, but it
might if we could somehow tell GCC that the register that scanned src
got modified similarly to the one that scanned dest.  Does it sound
like it might be useful to add an output operand for end-of-src to the
standard pattern?

E.g.:

(define_expand "movstrsi"
  ;; intuitive notion of the implementation:
  [(set (match_operand:BLK 0 "" "")  ;; dest
        (match_operand:BLK 1 "" "")) ;; src
   ;; wouldn't it be so nice if :P was a valid mode? :-)
   (clobber (match_operand:P 2 "" "")) ;; end-of-dest
   (clobber (match_operand:P 3 "" "")) ;; end-of-src
   ] ...)

Should we perhaps make operands 3, and maybe even 2, optional?