Opt dcmp

src/crt/dcmp.src

@@ -4,23 +4,71 @@
 
 	public	__dcmp
 
-; int _dcmp_c(f64_cmp_arg const *__restrict const arg)
 __dcmp:
-	; f64_cmp_arg*_ret_i24
-	push	bc, de, hl
-	or	a, a
-	sbc	hl, hl
-	add	hl, sp
-	push	iy, af
-	push	hl	; f64_cmp_arg*
-	call	__dcmp_c
-	pop	af
-	pop	af, iy
-	; Set the comparison flags
-	add	hl, de
-	or	a, a
-	sbc	hl, de
-	pop	hl, de, bc
+	push	iy
+	ld	iy, 0
+	add	iy, sp
+	push	bc
+	push	hl
+	ld	hl, (iy + 12)	; y.hi
+	rlc	b
+	sla	h	; Carry = signbit y
+	bit	0, b	; signbit x
+	jr	z, .positve_x
+	ccf
+.positve_x:
+	res	0, b
+	; H and B now have their signbits cleared
+	jr	c, .different_sign
+	; same_sign
+	sbc.s	hl, bc	; y.hi - x.hi
+	jr	nz, .not_equal
+	ld	hl, (iy + 9)	; y.md
+	sbc	hl, de	; y.md - x.md
+	jr	nz, .not_equal
+	ld	bc, (iy - 6)	; x.lo
+	ld	hl, (iy + 6)	; y.lo
+	sbc	hl, bc	; y.lo - x.lo
+	jr	z, .equal
+.not_equal:
+	ld	l, a	; preserve A
+	rra	; bit 7 of A = Carry
+	cpl
+.ret_flags:
+	xor	a, (iy - 2)	; bit 7 of B = sign of x
+	rla
+	sbc	a, a	; A is 0 or -1
+	ccf	; carry is set when A is 0, so A will be 1 or -1 in the end
+	adc	a, e
+	sub	a, e
+	ld	a, l	; restore A
+.equal:
+.finish_zero:
+	pop	hl
+	pop	bc
+	pop	iy
 	ret
 
-	extern	__dcmp_c
+.different_sign:
+; test for zero
+	; x[hi-md-lo]
+	; y[hi-md-lo]
+	add.s	hl, bc
+	adc	hl, de
+	jr	nz, .non_zero	; path taken by normal, inf, and NaN
+	jr	c, .non_zero
+	; both x and y are subnormal or zero
+	; x[00-00-lo]
+	; y[00-md-lo]
+	; HL is zero, Carry is reset
+	ld	hl, (iy + 6)	; y.lo
+	ld	bc, (iy + 9)	; y.md
+	add	hl, bc
+	ld	bc, (iy - 6)	; x.lo
+	adc	hl, bc
+	jr	nz, .non_zero
+	jr	nc, .finish_zero	; Z, NC, and P
+.non_zero:
+	ld	l, a	; preserve A
+	xor	a, a
+	jr	.ret_flags

src/crt/dtoll.src

@@ -13,28 +13,31 @@ __dtoull:
 	jr	z, .zero_or_one
 	bit	7, b
 	push	af
-	res	7, b
+;	res	7, b	; not needed since this bit is shifted out later on
 	push	hl
 	; -((Float64_mant_bits + Float64_bias) << 4)
-	ld	hl, $FFC010 ; -16368 ; -$3FF0
+;	ld	hl, $FFC010 ; -16368 ; -$3FF0
 
-	; clears the exponent field without touching the mantissa
-	; sets the LSB of the exponent since x is normalized
-	ld	a, c
-	or	a, l	; or a, $10
-	and	a, $1F
-
-	add	hl, bc
+	; extract the exponent
+	ld	h, b
+	ld	l, c
 	; HL <<= 4
 	add	hl, hl
 	add	hl, hl
 	add	hl, hl
 	add	hl, hl
+;	inc	hl
 
+	; clears the exponent field without touching the mantissa
+	; sets the LSB of the exponent since x is normalized
+	ld	a, c
+	or	a, $10
+	and	a, $1F
 	ld	c, a
 	ld	b, 0
-	ld	a, h
-	sub	a, 52 + 1	; float64 mantissa bits
+
+	ld	a, h	; expon
+	sub	a, 52 + 1 - 1	; float64 mantissa bits + CPL trick - inc HL
 	jr	c, .shift_right
 	; shift_left
 	; expon >= 52 or [52, 63]

src/crt/ltod.src

@@ -48,25 +48,24 @@ __ltod:
 	private	__ltod.hijack
 __ltod.hijack:
 	call	__lctlz
-	cp	a, 32
+	inc.s	bc	; clear UBC
+	ld	b, a	; <<= 8
+	xor	a, $20	; turns 32 into zero and clears carry flag
 	jr	z, .zero
 	; clears the MSB since the float will be normalized
 	; x <<= clz_result + 1; /* shift by 32 is UB */
 if 0
 	; calculate the exponent
 	push	hl
 	; 1023 + 31 = 1054 = 0x41E
-	inc.s	bc
-	or	a, a
-	ld	b, a	; <<= 8
 	ld	hl, $041E00
 	ld	c, l	; ld c, 0
 	sbc	hl, bc
 	ld	l, e	; (expon16 << (16 + 24)) | (mant48)
 	ex	de, hl
 	pop	hl
 
-	ld	b, a
+;	ld	b, a
 	inc	b
 	ld	a, e
 .loop32:	; shift by 32 is not UB here!
@@ -78,17 +77,14 @@ else
 	; calculate the exponent
 	push	hl
 	; 1023 + 31 = 1054 = 0x41E
-	inc.s	bc
-	or	a, a
-	ld	b, a	; <<= 8
 	ld	hl, $041E00
 	ld	c, l	; ld c, 0
 	sbc	hl, bc
 	ld	l, e	; (expon16 << (16 + 24)) | (mant48)
 	ex	de, hl
 
+	ld	l, b
 	pop	bc
-	ld	l, a
 	ld	a, e
 	call	__lshl
 	push	bc
@@ -112,7 +108,7 @@ end if
 	ret	; negative
 
 .zero:
-	sbc	hl, hl
+	; E:UHL and A are zero
 	ex	de, hl
 	sbc	hl, hl
 	ld	b, e