Shop OBEX P1 Docs P2 Docs Learn Events
Using ALTI for stacks — Parallax Forums

Using ALTI for stacks

SeairthSeairth Posts: 2,474
edited 2017-03-02 18:03 in Propeller 2
I'm was thinking about how to use ALTI effectively for stack manipulation (as was suggested in the "New Spin" thread). This is what I came up with (untested):
loop    SETD    sp, #in_val
        CALL    #push
        
        SETR    sp, #out_val
        CALL    #pop
        
        JMP     #loop

push    ALTI    sp, #%100_111           'substitute sp[17:9] into D, substitute sp[8:0] into S, then increment sp[8:0]
        WRLUT   0-0, 0-0        
        RET

pop     ALTI    sp, #%010               'increment sp[8:0]
        ALTI    sp, #%100_000_100       'substitute sp[17:9] into D, substitute sp[8:0] into S
        RDLUT   0-0, 0-0
        RET

sp      LONG    0
in_val  RES     1
out_val RES     1

If you wanted to avoid the call overhead, the de-normalized code would be:
loop    ALTI    sp, #%111               'substitute sp[17:9] into D, substitute sp[8:0] into S, then increment sp[8:0]
        WRLUT   in_val, 0-0        
        
        ALTI    sp, #%010               'increment sp[8:0]
        ALTI    sp, #%100               'substitute sp[27:19] into R, substitute sp[8:0] into S
        RDLUT   out_val, 0-0
        
        JMP     #loop
        RET

sp      LONG    0
in_val  RES     1
out_val RES     1

As noted, I don't think spacer instructions are needed, as you are not altering an instruction in the pipeline.

Does this look like what others were thinking? I'm, not crazy about the need for two ALTI instructions in the POP, but that doesn't seem too bad overall. Simple PUSH/POP instructions would certainly be nicer, though.

Edit: I just realized that the RDLUT probably requires SETR, not SETD. Code is updated to reflect that.
Edit: Removed unnecessary instructions from inlined version

Comments

  • SeairthSeairth Posts: 2,474
    edited 2017-03-02 18:04
    Here's a slightly less trivial example:
            MOV     in_val1, #6
            MOV     in_val2, #7
            
    loop    SETD    sp, #in_val1
            CALL    #push                   ' push first operand on LUT stack
            SETD    sp, #in_val2
            CALL    #push                   ' push second operand on LUT stack
            CALL    #_mul
            SETR    sp, #out_val
            CALL    #pop                    ' pop result off LUT stack
            CMP     out_val, #42 WZ
      IF_ E JMP     #0                      'halt
            JMP     #loop                   'maybe the answer will be different next time?
    
    _mul    SETR    sp, #temp2
            CALL    #pop                    ' pop second operand off LUT stack
            SETR    sp, #temp
            CALL    #pop                    ' pop first operand off LUT stack
            MUL     temp, temp2
            SETD    sp, #temp
            CALL    #push                   ' push result on LUT stack
            RET
           
    
    push    ALTI    sp, #%100_111           'substitute sp[17:9] into D, substitute sp[8:0] into S, then increment sp[8:0]
            WRLUT   0-0, 0-0        
            RET
    
    pop     ALTI    sp, #%010               'increment sp[8:0]
            ALTI    sp, #%100_000_100       'substitute sp[17:9] into D, substitute sp[8:0] into S
            RDLUT   0-0, 0-0
            RET
    
    sp      LONG    0
    in_val1 RES     1
    in_val2 RES     1
    out_val RES     1
    temp    RES     1
    temp2   RES     1
    

    It seems to me (totally unsubstantiated, mind you) that this will not perform much better than a hub stack using inlined PUSHx/POPx. Of course, you could also inline the push/pop operations in the above code to avoid the CALL/RET overhead, but at a cost of readability.
    loop    ALTI    sp, #%111               'substitute sp[8:0] into S, then increment sp[8:0]
            WRLUT   #6, 0-0                 'push first operand on LUT stack
            ALTI    sp, #%111               'substitute sp[8:0] into S, then increment sp[8:0]
            WRLUT   #7, 0-0                 'push second operand on LUT stack
            CALL    #_mul
            ALTI    sp, #%010               'increment sp[8:0]
            ALTI    sp, #%100               'substitute sp[8:0] into S
            RDLUT   out_val, 0-0            ' pop result off LUT stack
            CMP     out_val, #42 WZ
      IF_ E JMP     #0                      'halt
            JMP     #loop                   'maybe the answer will be different next time?
    
    _mul    ALTI    sp, #%010               'increment sp[8:0]
            ALTI    sp, #%100               'substitute sp[8:0] into S
            RDLUT   temp2, 0-0              'pop second operand off LUT stack
            ALTI    sp, #%010               'increment sp[8:0]
            ALTI    sp, #%100               'substitute sp[8:0] into S
            RDLUT   temp, 0-0               'pop first operand off LUT stack
            MUL     temp, temp2
            ALTI    sp, #%111               'substitute sp[8:0] into S, then increment sp[8:0]
            WRLUT   temp, 0-0               'push result on LUT stack
            RET
    
    sp      LONG    0
    out_val RES     1
    temp    RES     1
    temp2   RES     1
    
  • Actually, I just realized the inlined versions were doing unnecessary steps. Further, once inlined, it also made it possible to push immediate values on the stack.

    Question to Chip: what happens if you mix ALTI and AUG? For instance, in the above code, what if I wanted to push a large immediate value? Is this allowed?
            ALTI    sp, #%111               'substitute sp[8:0] into S, then increment sp[8:0]
            WRLUT   ##7, 0-0                'push large immediate value on LUT stack
    
  • A quick look on my debugger shows it steps over 3 instructions as expected but the result is incorrect.
  • ozpropdev wrote: »
    A quick look on my debugger shows it steps over 3 instructions as expected but the result is incorrect.

    Oh well! I kinda figured, but was hopeful. That's bound to bite someone in the future, though. The assembler should probably flag it as an error.
  • cgraceycgracey Posts: 14,133
    Seairth wrote: »
    Actually, I just realized the inlined versions were doing unnecessary steps. Further, once inlined, it also made it possible to push immediate values on the stack.

    Question to Chip: what happens if you mix ALTI and AUG? For instance, in the above code, what if I wanted to push a large immediate value? Is this allowed?
            ALTI    sp, #%111               'substitute sp[8:0] into S, then increment sp[8:0]
            WRLUT   ##7, 0-0                'push large immediate value on LUT stack
    

    The ALTxx instructions modify the next instruction, no matter what it is. The AUGS/AUGD instructions, however, will work around the ALTxx instructions. So, you'd do one or both AUGx instruction(s), then the ALTxx instruction, then the final instruction. That should work okay.
  • cgraceycgracey Posts: 14,133
    edited 2017-03-03 09:42
    Here's how I would do a LUT push/pop stack:
    ' Using LUT as a push/pop stack
    
    CON	stktop = $1FF
    
    DAT	org
    
    push_x	wrlut	x,ptr		'write x at lut[ptr]
    	incmod	ptr,#stktop	'increment/wrap ptr
    	ret
    
    pop_x	decmod	ptr,#stktop	'decrement/wrap ptr
    	rdlut	x,ptr		'read x at lut[ptr]
    	ret
    
    ptr	long	0
    
    x	res	1
    

    All those ALTxx instructions are only useful for indirect addressing of cog registers.
  • cgraceycgracey Posts: 14,133
    edited 2017-03-03 09:46
    Here's how I would do a cog register push/pop stack:
    ' Using cog registers as a push/pop stack
    
    CON	stksize = 100
    
    DAT	org
    
    push_x	altd	ptr,#stk	'set next d field
    	mov	0,x		'write x at stk[ptr]
    	incmod	ptr,#stksize-1	'increment/wrap ptr
    	ret
    
    pop_x	decmod	ptr,#stksize-1	'decrement/wrap ptr
    	alts	ptr,#stk	'set next s field
    	mov	x,0		'read x at stk[ptr]
    	ret
    
    ptr	long	0
    
    stk	res	stksize
    
    x	res	1
    
  • cgraceycgracey Posts: 14,133
    ALTI is best used in some kind of loop:
    ' Using REP+ALTI to automate register operations
    
    CON	stksize = 100
    
    DAT	org
    
    scroll	sets	x,#stk+1	's field of x points to stk[1]
    	setd	x,#stk+0	'd field of x points to stk[0]
    
    	rep	#2,#stksize-1	'ready to scroll registers
    	alti	x,#%111_111	'set next d and s fields, increment them in x
    	mov	0,0		'do 'mov stk[n+0],stk[n+1]', n = 0..stksize-2
    	ret
    
    stk	res	stksize
    
    x	res	1
    
  • cgracey wrote: »
    Seairth wrote: »
    what happens if you mix ALTI and AUG? For instance, in the above code, what if I wanted to push a large immediate value? Is this allowed?
            ALTI    sp, #%111               'substitute sp[8:0] into S, then increment sp[8:0]
            WRLUT   ##7, 0-0                'push large immediate value on LUT stack
    

    The ALTxx instructions modify the next instruction, no matter what it is. The AUGS/AUGD instructions, however, will work around the ALTxx instructions. So, you'd do one or both AUGx instruction(s), then the ALTxx instruction, then the final instruction. That should work okay.

    That's good to know. This would also imply that you can't use AUGx to modify ALTx, which seems reasonable.
  • cgraceycgracey Posts: 14,133
    Seairth wrote: »
    ...This would also imply that you can't use AUGx to modify ALTx, which seems reasonable.

    That's correct.
  • Ahh, yes. The venerable INCMOD/DECMOD! That is definitely a cleaner and more readable way to implement stacks.

    That does point out a minor issue with doing hub stacks with PTRx: they are not bounded. Of course, if you exceed your stack size, all bets are off anyhow. It's probably a toss-up whether it's worse to corrupt the bottom of your stack or non-stack memory.
  • Seairth wrote: »
    Ahh, yes. The venerable INCMOD/DECMOD! That is definitely a cleaner and more readable way to implement stacks.

    Yes it is. Didn't even consider that one.

    In any case, stacks aren't really an issue. We've got reasonable ways to do them.

    Nice!





Sign In or Register to comment.