Dear Philipp,
> thanks for your help with Z80 multiplication.
Thank you for your answer!
> study, but it has been shown, that small integer values are far more
> common than bigger ones. AFAIR values 0,1,2 combined are about as likely
> to occur as all other values combined in an 8 bit integer.
I'm sure you are right, because "char" is often used to store "boolean"
values or for small cycles. But the real question is: are those values
used as operands in multiplications?
If you use the byteXbyte product as a building block for some more complex
mathematical function, you usually have "random" bytes to handle. If you
use it to access elements of a bi-dimensional array A[x+ROWLEN*y], then
the result must be small... and so must be the two operands.
A far faster code for small operands is:
ld hl, #0
ld d, h
or a
jp 3
1:
add hl,de
2:
rl e
rl d
3:
rra
jp C, 1
jp NZ, 2
It assumes the two operands are in A and E; put the result in HL... It is
very fast when A<32, but code size is too big to be inlined...
>> Drawbacks:
>> - overwrites the accumulator A
> For builtin code generation this is bad, since A is used to backup B
> (when B is in use).
If both DE and B are used, wrap into push bc/pop pc, (saving ld a,b/ld b,a).
If only B is, save it to D or E (same as current).
If only DE is in use, do nothing (save push/pop).
> You might want to have a look at the 16 bit multiplication in mul.s or
There is a faster way for 16x16->16, but it would be slower for
16x8->16... I have to think about it a little bit longer... anyway I can
suggest some very small corrections:
---------8<---------8<---------8<---------8<---------8<---------8<-----
--- sdcc/device/lib/z80/mul.s 2009-01-05 11:20:47.000000000 +0100
+++ ./mul.s 2009-03-30 15:30:08.000000000 +0200
@@ -29,11 +29,11 @@
;; DE = multiplier
;;
;; Exit conditions
- ;; DE = less significant word of product
+ ;; HL = less significant word of product
;;
;; Register used: AF,BC,DE,HL
__mul16::
- ld hl,#0
+ ld l,#0
ld a,b
; ld c,c
ld b,#16
@@ -41,7 +41,7 @@
;; Optimise for the case when this side has 8 bits of data or
;; less. This is often the case with support address calls.
or a
- jr NZ,1$
+ jr NZ,3$
ld b,#8
ld a,c
@@ -49,6 +49,7 @@
;; Taken from z88dk, which originally borrowed from the
;; Spectrum rom.
add hl,hl
+3$:
rl c
rla ;DLE 27/11/98
jr NC,2$
---------8<---------8<---------8<---------8<---------8<---------8<-----
> I wonder how these would compare at device/lib/_mulllong.c
This is 32x32->32, isn't it? It should be nice to write an ad-hoc function
for this, if one can use BC,DE,HL,BC',DE',HL',IX,IY,A,A'... there as a lot
of register-space to work on... Maybe also 32x32->64 is possible.
Regards,
Marco
--
http://bodrato.it/
------------------------------------------------------------------------------
_______________________________________________
Sdcc-user mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/sdcc-user
