Re: [PATCH] PR 10339: Remove unsafe memcmp optimizations

[Falk Hueffner <falk dot hueffner at student dot uni-tuebingen dot de>]

"Kaveh R. Ghazi" <ghazi at caip dot rutgers dot edu> writes:

> PS: Do we have a compare_by_pieces for small comparisons analogous
> to the move_by_pieces, clear_by_pieces and store_by_pieces?  This
> would be a good optimization for platforms without cmpstr.

We don't, it seems. It would be a good idea though, IMHO. On big
endian, it's obvious, on Alpha for example, you can do something like

    diff = x0 ^ x1;
    if (diff == 0) return 0;
    eq = cmpbge(0, diff);
    x1_gt_x0 = cmpbge(x1, x0) & ~eq;
    return (x1_gt_x0 & (eq + 1)) ? -1 : 1;

I guess other little endian platforms might have similar tricks.
Additionally, about 90% of all memcmps seem to occur in a boolean
context, where they could be simplified even more.

-- 
	Falk