On 7/11/18, U.Mutlu <um@xxxxxxxxxxx> wrote: > I think these wide int routines make it possible to create 64bit targets > on 32bit hosts, ie. crosscompiling. See a later reply by Martin Sebor: they also make it possible to create 65bit to 128bit targets on 64bit hosts, and I suspect we could create beyond 128bit targets. > Regarding make: it should work when touch'ing a source file. > I just tried it out with the ../gcc_src/gcc/expmed.c: yes, it behaves as > expected. > > Maybe you did the build (configure and make) via a script, > but now calling make not from the same script: in this case > the environment can be different. One should avoid that; > better is doing the make in the same script as the configure, > in this case just disabling the configure step beforehand. I just extracted GCC from the tar file, ran contrib prerequisites, and then set up a standard make file, and just ran make, which worked before, so I've no idea why it no longer works! > OTOH, maybe just your "hook" is causing some side effects leading to that > unexpected behavior? :-) I don't think so, but w/o seeing the modification > one can only guess. > If it's not much code, then you can post your hook here, I then can try > to compile it and tell you about the outcome. That's most helpful; what I'm really interested in is some rough benchmarks/timings, so I'll think about the best way to do that and then, as you suggest, post the code here. It may be a day or two until I manage to do that. What I'm suspecting is that using "wide int" even just once might be slower than completely avoiding it. I posted something on the main gcc mailing list on Making GNU GCC choose_multiplier in expmed.c significantly faster https://gcc.gnu.org/ml/gcc/2018-07/msg00159.html but I haven't had any "bites" yet!