Merge pull request #559 from boriel/feature/optimize_mul8_and_mul16_for_zxnext

Feature/optimize mul8 and mul16 for zxnext

Author: boriel

src/arch/zx48k/library-asm/_mul32.asm

@@ -28,10 +28,12 @@ __MUL32_64START:
     ; sinclair zx81 rom for the mantissa multiply
 
 __LMUL:
-    and     a               ; reset carry flag
-    sbc     hl,hl           ; result bits 32..47 = 0
+    xor     a               ; reset carry flag
+    ld      h, a            ; result bits 32..47 = 0
+    ld      l, a
     exx
-    sbc     hl,hl           ; result bits 48..63 = 0
+    ld      h, a            ; result bits 48..63 = 0
+    ld      l, a
     exx
     ld      a,b             ; mpr is b'c'ac
     ld      b,33            ; initialize loop counter

src/arch/zxnext/library-asm/_mul32.asm

@@ -28,10 +28,12 @@ __MUL32_64START:
     ; sinclair zx81 rom for the mantissa multiply
 
 __LMUL:
-    and     a               ; reset carry flag
-    sbc     hl,hl           ; result bits 32..47 = 0
+    xor     a               ; reset carry flag
+    ld      h, a            ; result bits 32..47 = 0
+    ld      l, a
     exx
-    sbc     hl,hl           ; result bits 48..63 = 0
+    ld      h, a            ; result bits 48..63 = 0
+    ld      l, a
     exx
     ld      a,b             ; mpr is b'c'ac
     ld      b,33            ; initialize loop counter

src/arch/zxnext/library-asm/heapinit.asm

@@ -65,7 +65,7 @@
 ; An init directive is useful for initialization routines.
 ; They will be added automatically if needed.
 
-#init "core.__MEM_INIT"
+#init ".core.__MEM_INIT"
 
 
 ; ---------------------------------------------------------------------

src/arch/zxnext/library-asm/mul16.asm

@@ -5,32 +5,30 @@ __MUL16:	; Mutiplies HL with the last value stored into de stack
 
     PROC
 
-    LOCAL __MUL16LOOP
-    LOCAL __MUL16NOADD
-
     ex de, hl
     pop hl		; Return address
     ex (sp), hl ; CALLEE caller convention
 
 __MUL16_FAST:
-    ld b, 16
-    ld a, h
-    ld c, l
-    ld hl, 0
-
-__MUL16LOOP:
-    add hl, hl  ; hl << 1
-    sla c
-    rla         ; a,c << 1
-    jp nc, __MUL16NOADD
-    add hl, de
-
-__MUL16NOADD:
-    djnz __MUL16LOOP
+    ld a,d                      ; a = xh
+    ld d,h                      ; d = yh
+    ld h,a                      ; h = xh
+    ld c,e                      ; c = xl
+    ld b,l                      ; b = yl
+    mul d,e                     ; yh * yl
+    ex de,hl
+    mul d,e                     ; xh * yl
+    add hl,de                   ; add cross products
+    ld e,c
+    ld d,b
+    mul d,e                     ; yl * xl
+    ld a,l                      ; cross products lsb
+    add a,d                     ; add to msb final
+    ld h,a
+    ld l,e                      ; hl = final
 
     ret	; Result in hl (16 lower bits)
 
     ENDP
 
     pop namespace
-

src/arch/zxnext/library-asm/mul8.asm

@@ -2,54 +2,18 @@
 
 __MUL8:		; Performs 8bit x 8bit multiplication
     PROC
-
-    ;LOCAL __MUL8A
-    LOCAL __MUL8LOOP
-    LOCAL __MUL8B
     ; 1st operand (byte) in A, 2nd operand into the stack (AF)
     pop hl	; return address
     ex (sp), hl ; CALLE convention
 
-;;__MUL8_FAST: ; __FASTCALL__ entry
-;;	ld e, a
-;;	ld d, 0
-;;	ld l, d
-;;
-;;	sla h
-;;	jr nc, __MUL8A
-;;	ld l, e
-;;
-;;__MUL8A:
-;;
-;;	ld b, 7
-;;__MUL8LOOP:
-;;	add hl, hl
-;;	jr nc, __MUL8B
-;;
-;;	add hl, de
-;;
-;;__MUL8B:
-;;	djnz __MUL8LOOP
-;;
-;;	ld a, l ; result = A and HL  (Truncate to lower 8 bits)
-
 __MUL8_FAST: ; __FASTCALL__ entry, a = a * h (8 bit mul) and Carry
-
-    ld b, 8
-    ld l, a
-    xor a
-
-__MUL8LOOP:
-    add a, a ; a *= 2
-    sla l
-    jp nc, __MUL8B
-    add a, h
-
-__MUL8B:
-    djnz __MUL8LOOP
-
-    ret		; result = HL
+	; zx next 			
+	ld e,a 			; 4t 	
+	ld d,h 			; 4t 
+	mul d,e 		; 8t 
+	ld a,e			; 4			; 20t
+    ret		; result = DE & A
+	
     ENDP
 
     pop namespace
-

src/arch/zxnext/library-asm/print.asm

@@ -21,7 +21,7 @@
 ; Putting a comment starting with @INIT <address>
 ; will make the compiler to add a CALL to <address>
 ; It is useful for initialization routines.
-#init core.__PRINT_INIT
+#init .core.__PRINT_INIT
 
     push namespace core
 

src/arch/zxnext/library-asm/read_restore.asm

@@ -68,7 +68,7 @@ read_restart:
     jr nz, cont
     ;; Signals out of data
 
-    ld hl, __DATA__0
+    ld hl, .DATA.__DATA__0
     ld (__DATA_ADDR), hl
     jr read_restart  ; Start again
 cont:
@@ -339,7 +339,7 @@ __09_decode_float:
     ret
 
 __DATA_ADDR:  ;; Stores current DATA ptr
-    dw __DATA__0
+    dw .DATA.__DATA__0
     ENDP
 
 #undef _str

tests/functional/fact.asm

@@ -226,10 +226,12 @@ __MUL32_64START:
 	    ; this routine was with tiny differences in the
 	    ; sinclair zx81 rom for the mantissa multiply
 __LMUL:
-	    and     a               ; reset carry flag
-	    sbc     hl,hl           ; result bits 32..47 = 0
+	    xor     a               ; reset carry flag
+	    ld      h, a            ; result bits 32..47 = 0
+	    ld      l, a
 	    exx
-	    sbc     hl,hl           ; result bits 48..63 = 0
+	    ld      h, a            ; result bits 48..63 = 0
+	    ld      l, a
 	    exx
 	    ld      a,b             ; mpr is b'c'ac
 	    ld      b,33            ; initialize loop counter

tests/functional/ltee9.asm

@@ -222,10 +222,12 @@ __MUL32_64START:
 	    ; this routine was with tiny differences in the
 	    ; sinclair zx81 rom for the mantissa multiply
 __LMUL:
-	    and     a               ; reset carry flag
-	    sbc     hl,hl           ; result bits 32..47 = 0
+	    xor     a               ; reset carry flag
+	    ld      h, a            ; result bits 32..47 = 0
+	    ld      l, a
 	    exx
-	    sbc     hl,hl           ; result bits 48..63 = 0
+	    ld      h, a            ; result bits 48..63 = 0
+	    ld      l, a
 	    exx
 	    ld      a,b             ; mpr is b'c'ac
 	    ld      b,33            ; initialize loop counter

tests/functional/modf16c.asm

@@ -331,10 +331,12 @@ __MUL32_64START:
 	    ; this routine was with tiny differences in the
 	    ; sinclair zx81 rom for the mantissa multiply
 __LMUL:
-	    and     a               ; reset carry flag
-	    sbc     hl,hl           ; result bits 32..47 = 0
+	    xor     a               ; reset carry flag
+	    ld      h, a            ; result bits 32..47 = 0
+	    ld      l, a
 	    exx
-	    sbc     hl,hl           ; result bits 48..63 = 0
+	    ld      h, a            ; result bits 48..63 = 0
+	    ld      l, a
 	    exx
 	    ld      a,b             ; mpr is b'c'ac
 	    ld      b,33            ; initialize loop counter