Fixed __ldvrmu calling convention stability and fixed C ldiv under PREFER_OS_CRT. Created __ldivu_lremu_common to allow for custom calling conventions to be used for implementing 32-bit division/remainder.

Author: ZERICO2005

src/crt/ldivu.src

@@ -10,14 +10,15 @@
 
 .else
 
+; REMEMBER to keep ldivu.src, lremu.src, and ldiv.src in sync with this choice
 .if 1
 
 __ldivu:
 ; I: EUHL=dividend, AUBC=divisor
 ; O: euhl=EUHL/AUBC
 	push	bc
 
-	call	__ldvrmu
+	call	__ldivu_lremu_common
 
 	ld	a, b
 	pop	bc
@@ -26,6 +27,8 @@ __ldivu:
 	ei
 	ret
 
+	.extern	__ldivu_lremu_common
+
 .else
 
 __ldivu:
@@ -41,8 +44,9 @@ __ldivu:
 	pop	hl
 
 	ret
-.endif
 
 	.extern	__ldvrmu
 
 .endif
+
+.endif

src/crt/ldivu_lremu_common.src

@@ -0,0 +1,86 @@
+	.assume	adl=1
+
+	.section	.text
+
+	.global	__ldivu_lremu_common
+	.type	__ldivu_lremu_common, @function
+
+;;; struct u32div_t {
+;;;     uint32_t rem;
+;;;     uint32_t quot;
+;;; };
+;;; u32div_t _ldivu_lremu_common(uint32_t dividend, uint32_t divisor) {
+
+__ldivu_lremu_common:
+; Similar to __ldvrmu but faster and with a different calling convention.
+; Note: Uses shadow registers and disables interrupts.
+; I: EUHL=dividend, AUBC=divisor
+; O: a[uhl']=EUHL%AUBC, bcu=0, b=A, c=?, euhl=EUHL/AUBC, eubc'=AUBC, zf=!IEF2
+
+;;;     u32div_t result;
+;;;     result.quot = dividend;
+					; euhl : result.quot
+
+	push	bc
+
+	ld	c, a			; c = A
+	ld	a, i			; a = I
+					; pf = IEF2
+	di
+
+	ld	a, c			; a = A
+	exx
+	pop	bc
+	ld	e, a			; eubc' : divisor
+
+	push	af
+
+;;;     result.rem = 0;
+	xor	a, a
+	sbc	hl, hl			; auhl' : result.rem
+
+;;;     int i = 32;
+	exx
+	ld	b, 32			; b : i
+
+;;;     do {
+.L.c.loop:
+
+;;;         bool dividendBit = result.quot >> 31;
+;;;         result.quot <<= 1;
+	add	hl, hl
+	rl	e
+;;;         result.rem = (result.rem << 1) + dividendBit;
+	exx
+	adc	hl, hl
+	adc	a, a
+
+;;;         bool quotBit = result.rem >= divisor;
+;;;         result.rem -= divisor;
+	sbc	hl, bc
+	sbc	a, e
+
+;;;         if (!quotBit) {
+	jr	nc, .L.c.restore_skip
+;;;             result.rem += divisor;
+	add	hl, bc
+	adc	a, e
+;;;         }
+.L.c.restore_skip:
+
+;;;         if (quotBit) {
+	exx
+	jr	c, .L.c.skip
+;;;             result.quot++;
+	inc	l
+;;;         }
+.L.c.skip:
+
+;;;     } while (--i != 0);
+	djnz	.L.c.loop
+
+;;;     return result;
+	pop	bc
+	bit	2, c
+	ret
+;;; }

src/crt/ldvrmu.src

@@ -1,11 +1,7 @@
 	.assume	adl=1
 
 	.section	.text
-;;; struct u32div_t {
-;;;     uint32_t rem;
-;;;     uint32_t quot;
-;;; };
-;;; u32div_t _ldvrmu(uint32_t dividend, uint32_t divisor) {
+
 	.global	__ldvrmu
 	.type	__ldvrmu, @function
 
@@ -15,83 +11,6 @@
 
 .else
 
-; REMEMBER to also modify ldivu.src, lremu.src, and ldiv.src if you change this
-.if 1
-
-__ldvrmu:
-; I: EUHL=dividend, AUBC=divisor
-; O: a[uhl']=EUHL%AUBC, bcu=0, b=A, c=?, euhl=EUHL/AUBC, eubc'=AUBC, zf=!IEF2
-
-;;;     u32div_t result;
-;;;     result.quot = dividend;
-					; euhl : result.quot
-
-	push	bc
-
-	ld	c, a			; c = A
-	ld	a, i			; a = I
-					; pf = IEF2
-	di
-
-	ld	a, c			; a = A
-	exx
-	pop	bc
-	ld	e, a			; eubc' : divisor
-
-	push	af
-
-;;;     result.rem = 0;
-	xor	a, a
-	sbc	hl, hl			; auhl' : result.rem
-
-;;;     int i = 32;
-	exx
-	ld	b, 32			; b : i
-
-;;;     do {
-.L.c.loop:
-
-;;;         bool dividendBit = result.quot >> 31;
-;;;         result.quot <<= 1;
-	add	hl, hl
-	rl	e
-;;;         result.rem = (result.rem << 1) + dividendBit;
-	exx
-	adc	hl, hl
-	adc	a, a
-
-;;;         bool quotBit = result.rem >= divisor;
-;;;         result.rem -= divisor;
-	sbc	hl, bc
-	sbc	a, e
-
-;;;         if (!quotBit) {
-	jr	nc, .L.c.restore_skip
-;;;             result.rem += divisor;
-	add	hl, bc
-	adc	a, e
-;;;         }
-.L.c.restore_skip:
-
-;;;         if (quotBit) {
-	exx
-	jr	c, .L.c.skip
-;;;             result.quot++;
-	inc	l
-;;;         }
-.L.c.skip:
-
-;;;     } while (--i != 0);
-	djnz	.L.c.loop
-
-;;;     return result;
-	pop	bc
-	bit	2, c
-	ret
-;;; }
-
-.else
-
 __ldvrmu:
 ; I: EUHL=dividend, AUBC=divisor
 ; O: auhl=EUHL%AUBC, euix=EUHL/AUBC, iyh=A, iyl=0
@@ -128,5 +47,3 @@ __ldvrmu:
 	ret
 
 .endif
-
-.endif

src/crt/lremu.src

@@ -10,12 +10,13 @@
 
 .else
 
+; REMEMBER to keep ldivu.src, lremu.src, and ldiv.src in sync with this choice
 .if 1
 
 __lremu:
 ; I: EUHL=dividend, AUBC=divisor
 ; O: euhl=EUHL%AUBC
-	call	__ldvrmu
+	call	__ldivu_lremu_common
 	ld	e, a
 	push	de
 	exx
@@ -26,6 +27,8 @@ __lremu:
 	ei
 	ret
 
+	.extern	__ldivu_lremu_common
+
 .else
 
 __lremu:
@@ -41,8 +44,9 @@ __lremu:
 	pop	iy
 	pop	ix
 	ret
-.endif
 
 	.extern	__ldvrmu
 
 .endif
+
+.endif

src/libc/ldiv.src

@@ -4,6 +4,7 @@
 	.global	_ldiv
 	.type	_ldiv, @function
 
+; REMEMBER to keep ldivu.src, lremu.src, and ldiv.src in sync with this choice
 .if 1
 
 _ldiv:
@@ -27,7 +28,7 @@ _ldiv:
 	bit	7, e
 	call	__ldivs_lrems_common
 
-	call	__ldvrmu
+	call	__ldivu_lremu_common
 
 	exx
 	ld	e, a
@@ -53,6 +54,8 @@ _ldiv:
 .L.ei_skip:
 	jp	(hl)
 
+	.extern	__ldivu_lremu_common
+
 .else
 
 _ldiv:
@@ -111,8 +114,9 @@ _ldiv:
 	push	de
 	jp	(hl)
 
+	.extern	__ldvrmu
+
 .endif
 
 	.extern	__ldivs_lrems_common
-	.extern	__ldvrmu
 	.extern	__lneg