> Date: Thu, 10 May 2018 14:03:10 +0200 > From: Jakub Jelinek <ja...@redhat.com>
> On Thu, May 10, 2018 at 01:51:29PM +0200, Marc Glisse wrote: > > > > There are probably more > > > > complicated transformations this disables. > > > > > > I'm providing an example from *real* code where the > > > transformation is bad (admittedly just for the div-div case). > > Isn't the case of the posted real-world testcase that many targets have > instructions that can do division and modulo at the same time? Maybe they can, but they don't. What matters is what the tested targets actually do. I guess it's worth repeating: for the posted case, all the tested targets (aarch64-unknown-linux-gnu, powerpc64-unknown-linux-gnu, x86_64-pc-linux-gnu and mipsisa32r2el-linux-gnu) implement the div+mod in the test-case (by constant) using the high-part of a widened multiplication and shift. Actually, add ARM to that. (The only one I found that didn't was sparc.) > So perhaps don't punt just because of !single_use (), but punt if > !single_use () and one of the other uses is a modulo with the same constant > as the second division, or perhaps similarly if the first division is > !accompanied by modulo with the same constant too? Maybe, but I just saw no need to limit the case, and no way to test "other uses" anyway. brgds, H-P
