On Mon, Jul 29, 2019 at 11:52 PM Segher Boessenkool <seg...@kernel.crashing.org> wrote: > > On Mon, Jul 29, 2019 at 01:32:46PM -0700, Nathan Chancellor wrote: > > For the record: > > > > https://godbolt.org/z/z57VU7 > > > > This seems consistent with what Michael found so I don't think a revert > > is entirely unreasonable. > > Try this: > > https://godbolt.org/z/6_ZfVi > > This matters in non-trivial loops, for example. But all current cases > where such non-trivial loops are done with cache block instructions are > actually written in real assembler already, using two registers. > Because performance matters. Not that I recommend writing code as > critical as memset in C with inline asm :-)
Upon a second look, I think the issue is that the "Z" is an input argument when it should be an output. clang decides that it can make a copy of the input and pass that into the inline asm. This is not the most efficient way, but it seems entirely correct according to the constraints. Changing it to an output "=Z" constraint seems to make it work: https://godbolt.org/z/FwEqHf Clang still doesn't use the optimum form, but it passes the correct pointer. Arnd
