8080 question

[dxforth]

I have a count in a double register.

Is there a way to decrement and test for 0 without using the A register?

[Peter Heitzer]

dxforth <dxf...@gmail.com> wrote:
>I have a count in a double register.

>Is there a way to decrement and test for 0 without using the A register?

No fast way.
Only the 8 bit register opcodes set the flags. So you cannot use
a 16 bit decrement but have to do the decrement with the high and low
registers.
Decrement low register.
If carry was set decrement the high register and continue.
If z was set decrement the high register and check if carry was set.
That means that the high register was zero before the decrement.
If no carry was set you have to increment the high register in order to
restore the value before the decrement.





--
Dipl.-Inform(FH) Peter Heitzer, peter....@rz.uni-regensburg.de

[Douglas Miller]

NOTE! 8-bit increment/decrement does NOT affect the carry. Only sign or zero may be detected after those. If you use zero as the trigger for a dual 8-bit decrement loop (i.e. accomplishing a 16-bit decrement using two 8-bit decrements) you need to take into account that zero is not the same as carry and you will have to adjust your starting counts accordingly.

[Peter Heitzer]

Douglas Miller <durga...@gmail.com> wrote:
>NOTE! 8-bit increment/decrement does NOT affect the carry. Only sign or zero may be detected after those. If you use zero as the trigger for a dual 8-bit decrement loop (i.e. accomplishing a 16-bit decrement using two 8-bit decrements) you need to take into account that zero is not the same as carry and you will have to adjust your starting counts accordingly.

You are right. My mistake. Using only the detection of zero a possible
solution is doing a 16 bit decrement and for the low and high bytes
an increment followed by a decrement. So the value of the register does
not change but only the flags are set. A test costs 36 cycles if
both registers have to be checked or at least 18 cycles if the low byte
is not zero after the 16 Bit decrement.

[Douglas Miller]

On Wednesday, August 2, 2023 at 6:48:32 AM UTC-5, Peter Heitzer wrote:
> You are right. My mistake. Using only the detection of zero a possible
> solution is doing a 16 bit decrement and for the low and high bytes
> an increment followed by a decrement. So the value of the register does
> not change but only the flags are set. A test costs 36 cycles if
> both registers have to be checked or at least 18 cycles if the low byte
> is not zero after the 16 Bit decrement.
> --
> Dipl.-Inform(FH) Peter Heitzer, peter....@rz.uni-regensburg.de

Typically, I just see that the starting count is adjusted to account for the difference. These cases just use the zero flag to detect the carry over, but the starting count makes less sense (possibly, depending on whether it was an arbitrary number or had specific meaning). I've also seen these sorts of loops be "inverted" (do increment instead of decrement), which means the starting count needs to be negative - but usually makes more sense in terms of absolute numbers.

[dxforth]

My initial thought was to save/restore A e.g.

push h
mov l,a
mov a,e
ora d
mov a,l
pop h

but hoping there was something smarter. I later discovered a logical error
in my code and after correcting it the problem went away...

[Peter Heitzer]

dxforth <dxf...@gmail.com> wrote:
>On 2/08/2023 9:48 pm, Peter Heitzer wrote:
>> Douglas Miller <durga...@gmail.com> wrote:
>>> NOTE! 8-bit increment/decrement does NOT affect the carry. Only sign or zero may be detected after those. If you use zero as the trigger for a dual 8-bit decrement loop (i.e. accomplishing a 16-bit decrement using two 8-bit decrements) you need to take into account that zero is not the same as carry and you will have to adjust your starting counts accordingly.
>>
>> You are right. My mistake. Using only the detection of zero a possible
>> solution is doing a 16 bit decrement and for the low and high bytes
>> an increment followed by a decrement. So the value of the register does
>> not change but only the flags are set. A test costs 36 cycles if
>> both registers have to be checked or at least 18 cycles if the low byte
>> is not zero after the 16 Bit decrement.

>My initial thought was to save/restore A e.g.

> push h
> mov l,a
> mov a,e
> ora d
> mov a,l
> pop h

That makes 38 cycles + 10 cycles for the jump which is more than my
suggested code but it uses less bytes and is better readable.

>but hoping there was something smarter. I later discovered a logical error
>in my code and after correcting it the problem went away...

[Russell Marks]

dxforth <dxf...@gmail.com> wrote:

[Re: testing if a register pair is zero on an 8080]

>
> My initial thought was to save/restore A e.g.
>
> push h
> mov l,a
> mov a,e
> ora d
> mov a,l
> pop h
>
> but hoping there was something smarter. I later discovered a logical error
> in my code and after correcting it the problem went away...

That makes it sound like you don't need the test any more, but I may
as well still say that self-modifying code would be slightly faster:

sta aop+1
mov a,e
ora d
aop: mvi a,0

-Rus.

[dxforth]

Only 2 extra instructions which makes it very clean. Must remember it
for next time!

Despite the 8080 having few registers I'm often surprised by how few
times that's been a problem.

[Mark Lougheed]

8080 or 8085? If 8085, there is the undocumented instruction "DSUB" (08h), where:
HL = HL - BC (Z, S, P, CY, AC and X5, V all flag receives influence)

LXI H some number
LXI B 1
DSUB
JZ (or JNZ) somewhere

https://electronicerror.blogspot.com/2007/08/undocumented-flags-and-instructions.html

[Peter Heitzer]

In 8080:
PUSH B
LXI B 0ffffh
DAD B
POP B
JC iszero

Needs 51 cycles and the counter to be tested must be in HL.

[Mark Ogden]

Depending on usage, one option is to pre increment the two bytes of the count as in
lxi b,count + 101h

then the test for 0 can be done with
dcr c
jnz not0
dcr b
jnz not0
; count was zero

or similar variants
e.g.
lxi b,count + 101h
loop:
do something
...
dcr c
jnz loop
dcr b
jnz loop
; all done

this can lead to quicker code as the dcr b is only done once every time the c register is zero
note if count is calculated, just do inr c, inr b

Mark

[Phillip Stevens]

On Wednesday, 2 August 2023, dxforth wrote:
> I have a count in a double register.
> Is there a way to decrement and test for 0 without using the A register?

If the target is 8085 then there is an undocumented flag and jump instruction available.

The flag is usually called K and I often use the JP NK instruction to emulate a z80 LDIR instruction. The K flag is set on 0 to 0xFFFF so the loop counter needs to be pre-decremented.

Usage xamples are here.
https://github.com/RC2014Z80/RC2014/blob/master/ROMs/CPM-IDE/acia85cf/cpm22preamble.asm#L34

Of course this is only going to work if your target is actually 8085.

[dxforth]

A pity Intel didn't seriously look at an extended instruction set.
Why Faggin left?

[Phillip Stevens]

A bit OT but since you asked. The enhancements Intel made to the 8085, and then decided not to support, made the 8085 almost perfect imho. Undocumented stack access using DE instructions is much faster than IX/IY in the z80, and the 16 bit rotations really accelerated math (given no hardware multiply).

Using these enhancements, and a “native” C compiler, the 8085 actually beats the z80 in some of our z88dk benchmarks.

So one excellent code table (256) of instructions is a pretty nice design.
https://feilipu.me/2021/09/27/8085-software/

> Why Faggin left?

Wasn’t there, but I guess it is easier to get rich working for yourself. 😊
Something he perhaps wouldn’t have achieved at Intel.

[dxforth]

On 9/08/2023 8:10 am, Phillip Stevens wrote:
> On Tuesday, 8 August 2023 at 22:54:05 UTC+10, dxforth wrote:
>> On 8/08/2023 2:11 pm, Phillip Stevens wrote:
>>> Of course this is only going to work if your target is actually 8085.
>
>> A pity Intel didn't seriously look at an extended instruction set.
>
> A bit OT but since you asked. The enhancements Intel made to the 8085, and then decided not to support, made the 8085 almost perfect imho. Undocumented stack access using DE instructions is much faster than IX/IY in the z80, and the 16 bit rotations really accelerated math (given no hardware multiply).
>
> Using these enhancements, and a “native” C compiler, the 8085 actually beats the z80 in some of our z88dk benchmarks.
>
> So one excellent code table (256) of instructions is a pretty nice design.
> https://feilipu.me/2021/09/27/8085-software/

I assumed any undoc instructions would be either side-effects or planned but
broken in some way. From what you say this doesn't appear to be the case here.
Sounds like it was an executive decision to leave them out. Unfortunately the
result is the same - software that uses undoc instructions have a certain odour
to them.

>> Why Faggin left?
>
> Wasn’t there, but I guess it is easier to get rich working for yourself. 😊
> Something he perhaps wouldn’t have achieved at Intel.

Though for creative folk leaving tends to be more about frustration. For Intel
execs it would have been about where best to invest their money.

[Phillip Stevens]

> > Using these enhancements, and a “native” C compiler, the 8085 actually beats the z80 in some of our z88dk benchmarks.
> >

> > So one excellent code table of (256) instructions is a pretty nice design.

> > https://feilipu.me/2021/09/27/8085-software/
> I assumed any undoc instructions would be either side-effects or planned but
> broken in some way. From what you say this doesn't appear to be the case here.
> Sounds like it was an executive decision to leave them out. Unfortunately the
> result is the same - software that uses undoc instructions have a certain odour
> to them.

I wouldn’t worry too much. Tundra Semiconductor licensed the 80c85 design and published them as “enhanced instructions” in their datasheet. Complete with their own mistakes and mis interpretation.

http://images.100y.com.tw/pdf_file/34-TUNDRA-CA80C85B.pdf

Ken Shirriff covers the reverse engineering in some details and corrects the error in the Tundra datasheet.
His whole series is great reading on the 8085.

http://www.righto.com/2013/02/looking-at-silicon-to-understanding.html?m=1

So safe to use with no hesitation or odour. :-)

[George Phillips]

If you're willing to alter the count you can do the 16 bit loop as two 8 bit tests which will be faster. For example, suppose your 16 bit loop count is in DE. This code will execute the loop DE times:

; transform DE to split count
dcx d
inr d
inr e
loop:
; ... processing here
dcr e
jnz loop
dcr d
jnz loop

To see how it works consider the two cases going in where E is zero and not zero. Or, equivalently, when DE is a exact multiple of 256 or not. Note that the value of E never changes. If E is not zero then 1 is added to D. The inner loop will do E iterations and then after that it will do 256 iterations D times (the original value of D). Check it with $0003 and $0103 to get the idea.

If E is zero then D is not altered either. The inner loop will always be 256 iterations and will be executed D times. Checking $0100 and $0000 will show the correctness. Like with a conventional 16 bit loop, $0000 means 65536 iterations.

[dxforth]

Intel gave Tundra license to reproduce the silicon. Publishing undoc instructions
that were never their own, not industry standard, and unbelievably - screwing it
up - might give some customers pause to think. Marketing gimmick gone wrong?

[dxforth]

Also ROMable

[Mark Ogden]

If a count of zero is valid, then the tests need to be moved to the start of the loop
; de = count
inr d ; or lxi d, count + 101h
inr e
endtest:

dcr e
jnz loop
dcr d

jz done
loop:
; processing here
jmp endtest