I've added an additional linker relaxation that relaxes two instruction
pc-relative sequences to one instruction gp relative sequences when
possible.  This sequence now optimizes the initial gp generation to

  mv gp, gp

which is obviously bogus.  The fix is to disable relaxations when
setting up gp, preventing the linker from relaxing away this setup code.

https://github.com/riscv/riscv-elf-psabi-doc/pull/15

When we upstream glibc+linux, we should revisit this value.

This patch uses INLINE_SYSCALL_ERROR_RETURN_VALUE to avoid reading
and writing errno directly on architectures for which factoring out
the position-independent code for loading the address of errno
improves performance and/or code size in the common case.

	* sysdeps/unix/sysv/linux/generic/epoll_create.c (epoll_create):
	Use INLINE_SYSCALL_ERROR_RETURN_VALUE.
	* sysdeps/unix/sysv/linux/generic/futimesat.c (futimesat): Likewise.
	* sysdeps/unix/sysv/linux/generic/sysctl.c (sysctl): Likewise.
	* sysdeps/unix/sysv/linux/generic/ustat.c (ustat): Likewise.
	* sysdeps/unix/sysv/linux/generic/wordsize-32/fcntl.c (do_fcntl):
	Likewise.
	* sysdeps/unix/sysv/linux/generic/wordsize-32/getdents.c (__getdents):
	Likewise.
	* sysdeps/unix/sysv/linux/generic/wordsize-32/mmap.c (__mmap):
	Likewise.
	* sysdeps/unix/sysv/linux/generic/wordsize-32/overflow.h
	(lseek_overflow, stat_overflow, statfs_overflow): Likewise.
	* sysdeps/unix/sysv/linux/generic/wordsize-32/sendfile.c (sendfile):
	Likewise.
	* sysdeps/unix/sysv/linux/generic/xmknod.c (__xmknod): Likewise.
	* sysdeps/unix/sysv/linux/generic/setegid.c (setegid): Likewise.
	* sysdeps/unix/sysv/linux/generic/seteuid.c (seteuid): Likewise.
	* sysdeps/unix/sysv/linux/generic/wordsize-64/fxstat.c (__fxstat):
	Likewise.
	* sysdeps/unix/sysv/linux/generic/wordsize-64/fxstatat.c (__fxstatat):
	Likewise.
	* sysdeps/unix/sysv/linux/generic/wordsize-64/lxstat.c (__lxstat):
	Likewise.
	* sysdeps/unix/sysv/linux/generic/wordsize-64/xstat.c (__xstat):
	Likewise.

This factors out the setting of setting errno and returning -1, reducing
code size.

Forgot to add the thread pointer, derp.

This makes the common case of no error faster by avoiding clobbering
most of the caller-saved registers.

This resolves the ABI bug mentioned in #6.

Happily, falling back to the built-in versions in the no-atomic case
does the right thing (invoking sysriscv).

Since that's the only version we support, there's no reason to
support older kernel ABIs.

This deliberately breaks ABI compatibility, since we are changing
the calling convention etc.