Added 48 bit Routines

[fundudeone]

Added the following 48 bit crt routines:

slland: 348 cycles
sllor: 348 cycles
sllxor: 381 cycles
sllnot: 158 cycles
sllcmpu: 210 cycles
sllcmpzero: 45 cycles
sllshl: varies. It's quite long so it could be rewritten with a loop like shr routines if need be.
sllshru: varies
sllshrs: varies
sllmulu: 1592 cycles

[runer112]

Before I really get into review, just checking, does the calling convention for these routines match what ez80-clang emits?

[fundudeone]

Before I really get into review, just checking, does the calling convention for these routines match what ez80-clang emits?

Yes, they should work with the Z80_LibCall calling convention (once everything's fixed at least)

Edit: sllcmpu and sllcmpzero both use Z80_LibCall_F actually

[mateoconlechuga]

I don't like the fact that some CRT routines are disabling interrupts... but we can't really use them anyway so maybe I'm less conflcted 🙃

[runer112]

I only reviewed a small portion of the changes so far. Before I continue, we really ought to get @jacobly0's support to (a) confirm the signature expected for each of these functions and (b) produce a version of ez80-clang that actually emits calls to dedicated 48-bit math functions so they can be tested.

[runer112]

This optimization seems valid? -10 bytes, -10F cycles.

Suggested change

	; Push 2/3 done values
	push iy
	ld iy, 0
	add iy, sp
	push de
	push hl
	push bc
	; Deal with Upper Byte
	ld a, (iy-1) ; a = u[de] (modified)
	and a, (iy+2) ; a &= u[iy]
	ld (iy-1), a ; u[de] (modified) = a
	ld a, (iy-4) ; a = u[hl] (modified)
	and a, (iy-7) ; a &= u[bc]
	ld (iy-4), a ; u[hl] (modified) = a
	; Load values back into registers and clean up stack
	pop bc
	pop hl
	pop de
	pop iy
	; Push 2/3 done values
	push de
	push iy
	push hl
	sbc hl, hl
	add hl, sp ; uhl = &l
	push bc
	; Deal with Upper Byte
	ld de, 3
	dec hl
	ld a, (hl) ; a = bcu
	add hl, de
	and a, (hl) ; a &= hlu
	ld (hl), a ; hlu = a
	add hl, de
	ld a, (hl) ; a = iyu
	add hl, de
	and a, (hl) ; a &= deu
	ld (hl), a ; deu = a
	; Load values back into registers and clean up stack
	pop bc
	pop hl
	pop iy
	pop de

If valid, the same pattern can be applied to sllor and sllxor, of course.

[fundudeone]

Sorry this took me so long to get back to. This makes sense and works with everything I've thrown at it, so go ahead with or and xor.

[calc84maniac]

Just FYI, support has now been added to the compiler for these routines, and they use an i48 naming scheme rather than sll. I've added stub implementations for these in this file which ideally should make the routines "work" (suboptimally) by wrapping the 64-bit routines, and those can be replaced with full implementations via this PR.

My observations about the libcall calling convention (hopefully accurate, haven't confirmed but should be testable now) are:
Routines that take one 48-bit parameter: input in UDEUHL
Routines that take two 48-bit parameters: inputs in UDEUHL, UIYUBC
Routines that take one 48-bit parameter and one 8-bit parameter: inputs in UDEUHL, C
Routines that return one 48-bit result: output in UDEUHL
Routines that return two 48-bit results (just __i48dvrmu): output in UDEUHL, UIYUBC
Routines that return one 8-bit result: output in A
__i48cmpu outputs C and Z flags
__i48cmpzero outputs S and Z flags (just fixed the 32-bit implementation based on this information)
__i48cmps outputs S and V flags (does it also need the Z flag? probably not, the 32-bit and 64-bit implementations seem to indicate no, and the LLVM implementation seems to reorder comparisons to avoid explicit equality checks)