On Tue, Jun 21, 2016 at 11:26:19AM -0400, Chris Metcalf wrote: > On 6/21/2016 10:14 AM, Peter Zijlstra wrote: > >On Tue, Jun 21, 2016 at 04:04:08PM +0200, Peter Zijlstra wrote: > >>>I'm not sure who builds the toolchains, but tilepro is in upstream > >>>gcc/binutils/etc > >>>so should be easy enough to include. There's also a cross-toolchain for > >>>x64 I put > >>>up a while ago [1] that you could grab if you wanted to try it. > >>I usually build my own set -- and just did. But tilepro was not > >>included. Lemme go do so. > >binutils-2_26-branch builds for tilepro-linux > >gcc-6-branch does _not_ build for tilepro-linux > > I figured I would take a stab at diagnosing this myself, and it looks like the > toolchain build assumes that the kernel headers have already been installed > and therefore we have <asm/unistd.h> available. This actually seems like > a reasonable prerequisite for building the toolchain. I'm guessing you > don't? Should you? Or perhaps the compiler shouldn't make that assumption?
I can build: ls -la /opt/cross/bin/*-gcc | wc -l 27 compilers without installing kernel headers. > This has been true since gcc 4.x when tilepro support was first added. > > In any case if you replace the #include <asm/unistd.h> with > > #define __NR_FAST_cmpxchg -1 > #define __NR_FAST_atomic_update -2 > #define __NR_FAST_cmpxchg64 -3 > > that should also probably fix it, though I haven't tested it. It probably > wouldn't be crazy to just put those #defines directly in tilepro's atomic.h, > since it's not like those fast system call numbers will ever change. OK, I'll go test that right after I've got the kid in bed .. I'll let you know.
