* Negate syscall error return value in sysdep.c.
* Change vfork flag constant to correctly pass signal number for SIGCHLD.

* Seems to correctly place ld.so.1's relocations into .data.rel.ro
  and not .rodata.
* .rodata is no longer marked with the W flag in readelf -S output
  for ld.so.1.
* I don't fully understand this commit.

* Prevents (or at least so I think) register %4 from being reused.
* Fixes problem in which busybox attempts this snippet of code:
  1:   lr.w    v1, 0(a1)
        bne    v1, a2, 1f
       sc.w    a1, v0, 0(a1)
       bnez    a1, 1b
  1:
  If sc.w fails, a1 receives 1, and the lr.w is attempted again.

* Update linux-headers for changed register usage.
* Fixes "can't fork" problems in Busybox.

This also fixes the erroneous emission of a call instead of a sibcall.

use JAL instead of AUIPC/LD/JALR

this way, they can be used in shared libraries

constants >32b were emitting SLL instead of SLLI.

ra     -> x1
s0-s11 -> x2-x13
sp     -> x14
tp     -> x15
v0-v1  -> x16-x17
a0-a13 -> x18-x31

* Tweak Makefile parameters and correctly inhibit libc.
* Disable bootstrap on both stages' builds of riscv-linux-gcc.
* Adjust the #define for ABI_64 from 2 to 3, because glibc assumes
  that _RISCV_SIM_ABI64 = 3 as per sysdeps/riscv/sgidefs.h.
* Comment out the store word instructions in pipe.S, as they're no
  longer needed.
* Fix sysdeps/unix/sysv/linux/riscv/bits/fcntl.h, mman.h, signum.h,
  and kernel_stat.h to match the ones provided in Linux.
* Remove sysdeps/unix/sysv/linux/riscv/xstatconv.c, rv64/fxstatat64.c,
  and bits/stat.h, to fall back on the ones already provided for
  all architectures.
* Negate the value created from the INTERNAL_SYSCALL_ERRNO macro
  to convert negative error values from the kernel.

the dynamic linker's such a clusterfuck; it's probably not even worth
trying to improve.

riscv-gcc is now capable of making the relevant optimization.

We now make use of xori (for consts between 0x7ffff800 and 0x7fffffff) and
make use of two registers to accelerate the hardest cases (especially for
superscalars); the worst case is now 6 insns.

the frame pointer will point just above RA and FP.