https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109695
--- Comment #23 from Aldy Hernandez <aldyh at gcc dot gnu.org> --- An update on the int_range_max memory bloat work. As Andrew mentioned, having int_range<25> solves the problem, but is just kicking the can down the road. I ran some stats on what we actually need on a bootstrap, and 99.7% of ranges fit in a 3 sub-range range, but we need more to represent switches, etc. There's no clear winner for choosing <N>, as the distribution for anything past <3> is rather random. What I did see was that at no point do we need more than 125 sub-ranges (on a set of .ii files from a boostrap). I've implemented various alternatives using a dynamic approach similar to what we do for auto_vec. I played with allocating 2x as much as needed, and allocating 10 or 20 more than needed, as well going from N to 255 in one go. All of it required some shuffling to make sure the penalty isn't much wrt virtuals, etc, but I think the dynamic approach is the way to go. The question is how much of a performance hit are we willing take in order to reduce the memory footprint. Memory to speed is a linear relationship here, so we just have to pick a number we're happy with. Here are some numbers for various sub-ranges (the sub-ranges grow automatically in union, intersect, invert, and assignment, which are the methods that grow in sub-ranges). trunk (wide_ints <255>) => 40912 bytes GCC 12 (trees <255>) => 4112 bytes auto_int_range<2> => 432 bytes (5.14% penalty for VRP) auto_int_range<3> => 592 bytes (4.01% penalty) auto_int_range<8> => 1392 bytes (2.68% penalty) auto_int_range<10> => 1712 bytes (2.14% penalty) As you can see, even at N=10, we're still 24X smaller than trunk and 2.4X smaller than GCC12 for a 2.14% performance drop. I'm tempted to just pick a number and tweak this later as we have ultimate flexibility here. Plus, we can also revert to a very small N, and have passes that care about switches use their own temporaries (auto_int_range<20> or such). Note that we got a 13.22% improvement for the wide_int+legacy work, so even the absolute worst case of a 5.14% penalty would have us sitting on a net 8.76% improvement over GCC12. Bike shedding welcome ;-)
