ARCompact has three-address instructions, but it has a number of shorter two-address instruction codes.
When optimizing or speed, I think it would generally be best to choose the three-address alternatives for input reloads, since this allows more reload inheritance. OTOH, when optimizing for size, the two-address alternatives are probably better. I would like to define a machine-specific contraint modifier that adjustes reject in find_reloads similarly as '?' and '!' do, but has the actual amount defined in target code. I.e. for the size/speed tradeoff explained above, I would make the reject adjustment zero when optimizing for size, something nonzero like - e.g. six - when optimizing for speed, and then use this constraint modifier to adjust the two-address alternatives (which come first so that they will dominate when there is an exact match). Would such a patch acceptable? What do you think about which characters to use for the first constraint modifier character; should I use a letter like ordinary constraints, or should I rather grab one of the currently unused 'punctuation' (i.e. non-alphanumeric, printable) characters?
