bzero optimization rarely does

[Paul Koning]

>>>>> "Roger" == Roger Sayle <roger@eyesopen.com> writes:

 Roger> I've tracked down the source of the problem.  MOVE_BY_PIECES_P
 Roger> and thereby MOVE_RATIO is being inappropriately used to
 Roger> determine whether to call clear_by_pieces.  On targets that
 Roger> have movstr?i instructions MOVE_RATIO is set very low, so that
 Roger> individual moves are discouraged in favour of a hardware block
 Roger> move instruction.  Alas on targets, like MIPS, that have a
 Roger> movstr?i pattern but no clrstr?i, the use of clear_by_pieces
 Roger> is penalized in the hopes of using an efficient instruction
 Roger> that doesn't exist.  Similarly, targets with clrstr?i but no
 Roger> movstr?i will generate long sequences of move instructions
 Roger> rather than a hardware clear instruction.

Great.  That all make sense; I was slowly approaching the same
conclusion but without being able to propose a fix for it.

MIPS doesn't have a block-move instruction, but it does have a way to
do bulk moves that's more efficient than the load<mode>, store<mode>
you get without the movstr pattern.  The same would seem to be true
for clear, so I suspect that adding a clrstr<xx> pattern to mips.md
would be an additional useful thing to do.  But at least, in the
meantime, what you're doing should help a lot.

     paul