Re: PATCH: fine-tuning for can_store_by_pieces

[Richard Sandiford <richard at codesourcery dot com>]

Sandra Loosemore <sandra@codesourcery.com> writes:
> Richard Sandiford wrote:
>> What did you think about the other suggestion: moving the magic
>> "1 instruction" bound for optimize_size from builtins.c to SET_RATIO?
>
> Perhaps other maintainers can jump in and say something here, but my
> gut feeling is that it doesn't make sense to remove that.  Doing a
> one-byte store inline always has to be cheaper than a function call,
> and doing the optimization early makes more sense than relying on a
> target-specific expansion, because it might allow recognition of other
> optimization patterns along the way.  I found during my earlier
> testing that setting MOVE_RATIO too low to catch that case had the
> side-effect of causing one of the profile-guided optimization test
> cases to fail, for instance.

Hmm, I'm not sure I follow.  You seem to be implying that 1-byte stores
are always done "by pieces" when optimize_size, but I don't think that's
true.  I was referring the 1-instruction bound in code like this:

      if (host_integerp (len, 1)
--->	  && !(optimize_size && tree_low_cst (len, 1) > 1)
	  && can_store_by_pieces (tree_low_cst (len, 1),
				  builtin_memset_read_str, &c, dest_align))
	{
	  val_rtx = force_reg (TYPE_MODE (unsigned_char_type_node),
			       val_rtx);
	  store_by_pieces (dest_mem, tree_low_cst (len, 1),
			   builtin_memset_gen_str, val_rtx, dest_align, 0);
	}
      else if (!set_storage_via_setmem (dest_mem, len_rtx, val_rtx,
					dest_align, expected_align,
					expected_size))
	goto do_libcall;
       
This code still uses can_store_by_pieces for single-byte stores when
optimize_size (and can still fall back to setmem or libcalls for that
case if can_store_by_pieces returns false, although I agree that's an
odd thing to do for single-byte stores).  What I was objecting to was
that the target doesn't get any chance to say that _2-byte stores_ (or
bigger) are better implemented "by pieces" than via a setmem or libcall
pattern.

You referred to this limit yourself when I queried the MIPS
optimize_size value of SET_RATIO.  You said that the value only really
matters for 1-byte stores, and looking at the patch, I thought I could
see why.  All calls to can_store_by_pieces with a "true" argument seemed
to be guarded by a check like the above.  So the suggestion to move the
check was really following on from that.  As far as I could tell,
CLEAR_RATIO and CLEAR_BY_PIECES_P have no single-byte limit for
optimize_size, so I was thinking it would be better if SET_RATIO and
SET_BY_PIECES_P didn't either.

Richard