Peephole optimizations

src/libc/ceilf.src

@@ -15,28 +15,28 @@
 
 _ceil:
 _ceilf:
-	call	__frameset0
-	ld	bc, (ix+6)
-	ld	a, (ix+9)
+	ld	hl, 6
+	add	hl, sp
+	ld	a, (hl)
+	dec	hl
+	dec	hl
+	dec	hl
+	ld	bc, (hl)
 	call	__fneg
 	ld	l, a
-	ld	h, 0
 	push	hl
 	push	bc
 	call	_floor
 	pop	bc
-	pop	bc
-	push	hl
+	ex	(sp), hl
 	pop	bc
 	ld	a, e
 	call	__fneg
 	ld	e, a
 	push	bc
 	pop	hl
-	pop	ix
 	ret
 
-	.extern	__frameset0
 	.extern	__fneg
 	.extern	_floor
 

src/libc/cimagf.src

@@ -10,9 +10,11 @@
 ; float cimagf(float _Complex)
 _cimagf:
 _cimag:
-	ld	iy, 0
-	add	iy, sp
-	ld	sp, iy
-	ld	hl, (iy + 7)
-	ld	e, (iy + 10)
+	ld	hl, 10
+	add	hl, sp
+	ld	e, (hl)
+	dec	hl
+	dec	hl
+	dec	hl
+	ld	hl, (hl)
 	ret

src/libc/cimagl.src

@@ -9,7 +9,6 @@
 _cimagl:
 	ld	iy, 0
 	add	iy, sp
-	ld	sp, iy
 	ld	hl, (iy + 11)
 	ld	de, (iy + 14)
 	ld	bc, (iy + 17)

src/libc/ez80_builtin.src

@@ -285,17 +285,20 @@ ___ez80_bitreverse48:
 	.type	___ez80_rotateleft24, @function
 ___ez80_rotateleft24:
 	; unoptimized
-	ld	iy, 0
-	lea	de, iy + 0
-	add	iy, sp
-	ld	hl, (iy + 3)
-	ld	a, (iy + 6)
-	ld	c, 24
+	ld	hl, 6
+	add	hl, sp
+	ld	a, (hl)
+	dec	hl
+	dec	hl
+	dec	hl
+	ld	hl, (hl)
+	ld	de, 24
 ___ez80_rotateleft24.mod24:
-	sub	a, c
+	sub	a, e
 	jr	nc, ___ez80_rotateleft24.mod24
-	add	a, c
+	add	a, e
 	ret	z
+	ld	e, d		; ld de, 0
 	ld	b, a
 ___ez80_rotateleft24.loop:
 	add	hl, hl
@@ -312,18 +315,17 @@ ___ez80_rotateleft24.loop:
 ___ez80_rotateleft48:
 	; unoptimized
 	ld	iy, 0
-	lea	bc, iy + 0
+	lea	bc, iy + 48	; ld bc, 48
 	add	iy, sp
 	ld	hl, (iy + 3)
 	ld	de, (iy + 6)
 	ld	a, (iy + 9)
-	ld	c, 48
 ___ez80_rotateleft48.mod48:
 	sub	a, c
 	jr	nc, ___ez80_rotateleft48.mod48
 	add	a, c
 	ret	z
-	ld	c, 0
+	ld	c, b		; ld bc, 0
 ___ez80_rotateleft48.loop:
 	add	hl, hl
 	ex	de, hl
@@ -342,18 +344,21 @@ ___ez80_rotateleft48.loop:
 	.type	___ez80_rotateright24, @function
 ___ez80_rotateright24:
 	; unoptimized
-	ld	iy, 0
-	lea	de, iy + 0
-	add	iy, sp
-	ld	hl, (iy + 3)
-	ld	a, (iy + 6)
-	ld	c, 24
+	ld	hl, 6
+	add	hl, sp
+	ld	a, (hl)
+	dec	hl
+	dec	hl
+	dec	hl
+	ld	hl, (hl)
+	ld	de, 24
 ___ez80_rotateright24.mod24:
-	sub	a, c
+	sub	a, e
 	jr	nc, ___ez80_rotateright24.mod24
 	neg
-	cp	a, c
+	cp	a, e
 	ret	z
+	ld	e, d		; ld de, 0
 	ld	b, a
 ___ez80_rotateright24.loop:
 	add	hl, hl
@@ -370,19 +375,18 @@ ___ez80_rotateright24.loop:
 ___ez80_rotateright48:
 	; unoptimized
 	ld	iy, 0
-	lea	bc, iy + 0
+	lea	bc, iy + 48	; ld bc, 48
 	add	iy, sp
 	ld	hl, (iy + 3)
 	ld	de, (iy + 6)
 	ld	a, (iy + 9)
-	ld	c, 48
 ___ez80_rotateright48.mod48:
 	sub	a, c
 	jr	nc, ___ez80_rotateright48.mod48
 	neg
 	cp	a, c
 	ret	z
-	ld	c, 0
+	ld	c, b		; ld bc, 0
 ___ez80_rotateright48.loop:
 	add	hl, hl
 	ex	de, hl

src/libc/strrstr.src

@@ -17,17 +17,16 @@ _strrstr:
 	ld	bc, 0
 	cpir
 	sbc	hl, hl
-	dec	hl
-	sbc	hl, bc		; carry will be cleared
+	scf
+	sbc	hl, bc		; carry will be set
 	ex	(sp), hl	; strlen(needle)
 	push	hl		; needle
 
 	ex	de, hl
 	push	hl
-	ld	bc, 0
+	ld	bc, -1
 	cpir
 	sbc	hl, hl
-	dec	hl
 	sbc	hl, bc
 	ex	(sp), hl	; strlen(haystack)
 	push	hl		; haystack

src/libc/strstr.src

@@ -24,17 +24,16 @@ _strstr:
 	ld	bc, 0
 	cpir
 	sbc	hl, hl
-	dec	hl
-	sbc	hl, bc		; carry will be cleared
+	scf
+	sbc	hl, bc		; carry will be set
 	ex	(sp), hl	; strlen(needle)
 	push	hl		; needle
 
 	ex	de, hl
 	push	hl
-	ld	bc, 0
+	ld	bc, -1
 	cpir
 	sbc	hl, hl
-	dec	hl
 	sbc	hl, bc
 	ex	(sp), hl	; strlen(haystack)
 	push	hl		; haystack

src/libc/wcsnlen.src

@@ -27,17 +27,16 @@ _wcsnlen:
 	cp	a, c		; cp a, 0
 	ret	nz		; high byte was non-zero
 	dec	hl
-	; ret	; We can use the RET PO below since overflow won't be set
-.L.loop_lo:
-	ret	po
-.L.loop_hi:
+	ret
+
+.L.loop:
 	inc	hl
 .L.loop_start:
 	cpi
-	jr	nz, .L.loop_lo
 	ret	po
+	jr	nz, .L.loop
 	cp	a, (hl)
-	jr	nz, .L.loop_hi
+	jr	nz, .L.loop
 	ex	de, hl
 	pop	de		; reset SP
 	scf

src/srldrvce/srldrvce.asm

@@ -523,8 +523,7 @@ srl_GetCDCStandardDescriptors:
 	jq	nz,.noserial
 
 	ld	a,3
-	ld	hl,.device+deviceDescriptor.iSerialNumber
-	ld	(hl),a
+	ld	(.device+deviceDescriptor.iSerialNumber),a
 
 	ld	de,.stringserialnum + 2
 	ld	hl,ti.OP4
@@ -550,9 +549,7 @@ srl_GetCDCStandardDescriptors:
 	jq	c,.store
 	add	a,'A'-'9'-1
 .store:
-	ex	de,hl
-	ld	(hl),a
-	ex	de,hl
+	ld	(de),a
 	inc	de
 	inc	de
 	ret
@@ -615,15 +612,12 @@ srl_UsbEventCallback:
 	push	de
 	push	de
 	call	usb_FindDevice
-	pop	de
-	pop	de
-	pop	de
+	ld	sp, ix
 	ld	de, 0
 	push	de
 	push	hl
 	call	usb_GetDeviceEndpoint
-	pop	de
-	pop	de
+	ld	sp, ix
 	ld	de, 0
 	push	de
 	push	de
@@ -634,11 +628,7 @@ srl_UsbEventCallback:
 	push	hl
 	call	usb_ScheduleControlTransfer
 	ld	bc, 1
-	pop	hl
-	pop	hl
-	pop	hl
-	pop	hl
-	pop	hl
+	ld	sp, ix
 .BB0_5:
 	push	bc
 	pop	hl
@@ -668,7 +658,7 @@ get_device_type:
 	call	next_descriptor
 	inc	hl
 	inc	hl
-	ld	de,0
+	inc.s	de	; clear UDE
 	ld	e,(hl)	; total descriptor length
 	inc	hl
 	ld	d,(hl)
@@ -828,7 +818,7 @@ set_rate_cdc:
 	ld	(.linecoding),hl
 	ld	bc,0
 	push	bc	; transferred
-	ld	bc,50
+	ld	c,50	; ld bc, 50
 	push	bc	; num retries
 	ld	bc,.linecoding
 	push	bc	; data
@@ -984,8 +974,8 @@ ring_buf_push:
 	sbc	hl,hl					; check if bc is 0
 	adc	hl,bc
 	ret	z
+	ld	hl,(xring_buf_ctrl.data_end)
 	ex	de,hl					; hl = data
-	ld	de,(xring_buf_ctrl.data_end)
 	push	bc
 	ldir						; de = data_end + len
 	ld	(xring_buf_ctrl.data_end),de

src/usbdrvce/usbdrvce.asm

@@ -2952,8 +2952,8 @@ assert endpointDescriptor.wMaxPacketSize+2 = endpointDescriptor.bInterval
 	inc	hl
 	ld	b,(hl)
 	inc	b
-	djnz	.validInterval
-	jq	.invalidParam
+	dec	b
+	jq	z, .invalidParam
 .validInterval:
 	ex	de,hl
 	inc.s	hl