Segher Boessenkool <seg...@kernel.crashing.org> writes: > On Tue, May 05, 2020 at 05:40:21PM +0200, Christophe Leroy wrote: >> >>+#define __put_user_asm_goto(x, addr, label, op) \ >> >>+ asm volatile goto( \ >> >>+ "1: " op "%U1%X1 %0,%1 # put_user\n" \ >> >>+ EX_TABLE(1b, %l2) \ >> >>+ : \ >> >>+ : "r" (x), "m<>" (*addr) \ >> > >> >The "m<>" here is breaking GCC 4.6.3, which we allegedly still support. >> > >> >Plain "m" works, how much does the "<>" affect code gen in practice? >> > >> >A quick diff here shows no difference from removing "<>". >> >> It was recommended by Segher, there has been some discussion about it on >> v1 of this patch, see >> https://patchwork.ozlabs.org/project/linuxppc-dev/patch/4fdc2aba6f5e51887d1cd0fee94be0989eada2cd.1586942312.git.christophe.le...@c-s.fr/ >> >> As far as I understood that's mandatory on recent gcc to get the >> pre-update form of the instruction. With older versions "m" was doing >> the same, but not anymore. > > Yes. How much that matters depends on the asm. On older CPUs (6xx/7xx, > say) the update form was just as fast as the non-update form. On newer > or bigger CPUs it is usually executed just the same as an add followed > by the memory access, so it just saves a bit of code size.
The update-forms are stdux, sthux etc. right? I don't see any change in the number of those with or without the constraint. That's using GCC 9.3.0. >> Should we ifdef the "m<>" or "m" based on GCC >> version ? > > That will be a lot of churn. Just make 4.8 minimum? As I said in my other mail that's not really up to us. We could mandate a higher minimum for powerpc, but I'd rather not. I think for now I'm inclined to just drop the "<>", and we can revisit in a release or two when hopefully GCC 4.8 has become the minimum. cheers