max.asm

包含几个高效的矢量运算的数学函数

	.include "ccall.asm"
	
	.def _ti_max
	.text
_ti_max:
	pre_ccall 2,AR_X,AR_N
	setc SXM
	
	sar AR_X,*+,AR_X ; xmax=stack[0]=&x[max]=&x[0];
	lacc *,AR_STACK
	sacl *+,AR_STACK ; stack[1] = x[max]
	
; Compute N%7 to make N divisible by 8 in the unrolled loop	
	sar AR_N,*			; stack[2]=N         
	lacl *+				; acc=N
	and #07h			; acc&=0111b - Unroll 8 iterations
	sacl *-				; stack[3] = N % 0111b
	lacc *,13			; acch=N>>3
	sach *+				; stack[2]=N>>3
	lar AR_N,*-			; AR_N = stack[3] = N%7
	mar *-,AR_N
	banz NEXT0,AR_X		; AR_N--, Skip loop if AR_N = 0
	b NEXT1

	; SP = &stack[1]
	
	; Local variables
	; &xmax - stack[0]
	; *xmax - stack[1]

; ARP=AR_X
NEXT0:                     
; Is x[AR_X++] > x[max] ?
	lacc *+,AR_STACK			; ACC = x[AR_X++];
	sub *,AR_N			; ACC -= x[max];
	cc NEWMAX2,gt		; if(x[AR_X]>x[max]) newmax
	; while(N--)
	banz NEXT0,AR_X

NEXT1:
	mar *,AR_STACK
	mar *+
	lar AR_N,*-,AR_N 		; AR_N = stack[2]=N>>3 
	banz NEXT2,AR_X		; Go to 8x loop if AR_X > 0
	b DONE

NEXT2:  
	.loop 7
	lacc *+,AR_STACK			; ACC = x[AR_X++];
	sub *,AR_X			; ACC -= x[max];
	cc NEWMAX1,gt		; if(x[AR_X]>x[max]) newmax
	.endloop
	lacc *+,AR_STACK			; ACC = x[AR_X++];
	sub *,AR_N			; ACC -= x[max];
	cc NEWMAX2,gt		; if(x[AR_X]>x[max]) newmax
	; while(N--)
	banz NEXT2,AR_X

DONE:
	mar *,AR_STACK
	mar *-				; SP=0
	lacl *				; acc=stack[0]=&x[max]
	post_ccall 2


NEWMAX .macro AR
	mar *,AR_X
	mar *-,AR_STACK			; i--
	sbrk #1				; SP--
	sar	AR_X,*+,AR_X		; stack[0]=&x[i] (address of new max)
	lacc *+,AR_STACK			; acc=x[i]
	sacl *,AR			; stack[1]=x[i]
	ret
 .endm

; We need one NEWMAX that returns with ARP=AR_X and
; one with ARP=AR_N
NEWMAX1:    
	NEWMAX AR_X

NEWMAX2:
	NEWMAX AR_N