Implement and test support for all RISC-V inline asm constraints

[asb]

As documented here.

[luismarques]

The present inline asm constraints implementation incompleteness seems to be contributing to a confusing situation. For instance, here's my implementation of a UART in terms of a Linux-compatible system call, in D (LDC LLVM inline assembly):

__asm("ecall", "{x10},{x11},{x12},{x17},~{x10}", 1, str.ptr, str.length, 64);

You pass the arguments in registers a0-a2, and specify the system call 64 in a7. a0 is overwritten with the result, which I don't use, so I just treat it as a clobber. The obvious improvement would be to use the ABI register names directly. Changing just one of the input regs to use the ABI name:

__asm("ecall", "{a0},{x11},{x12},{x17},~{x10}", 1, str.ptr, str.length, 64);
error: couldn't allocate input reg for constraint '{a0}'

So it seems that the ABI names aren't supported, right? But changing it in the clobber works:

__asm("ecall", "{x10},{x11},{x12},{x17},~{a0}", 1, str.ptr, str.length, 64);

Trying to use the result even with the canonical register names doesn't seem to work either, so maybe this is my fault, or LDC's?

int r = __asm!int("ecall", "={x10},{x11},{x12},{x17}", 1, str.ptr, str.length, 64);
Error: __asm constraint argument is invalid

[markww]

I'd like to express my interest in this bug being fixed, as it is currently causing Clang to fault when attempting to compile code from musl-riscv libc. The implementation of atomic helper functions in musl-riscv heavily uses these constraints, particularly J and A: https://github.com/lluixhi/musl-riscv/blob/master/arch/riscv64/atomic_a.h .

[luismarques]

All of the described problems now work properly. The issue can be closed.