Possible enhancement

[GoogleCodeExporter]

Hi there!

I have made a change in "a_next" which provides 30% faster throughput of Forth 
high level words.  But there is something I do not understand, I need to insert 
a "nop", instead of the instruction I have removed?

Code before.

a_next
a_debugonoff
if_never jmpret a_dum_ret, # a_dum ' when debug is loaded this address will be 
patched

rdword tregone,IP ' the next word
test tregone, fMask wz
if_z add IP, #2 ' if the one of the hi bits is not set, it is an assembler 
word, inc IP
if_z jmp tregone
rdword tregone, IP ' otherwise it is a forth word
mov tregfive, IP
add tregfive, #2
mov IP, tregone
call #a_rsPush
jmp #a_next

Code after.

a_next
a_debugonoff
if_never jmpret a_dum_ret, # a_dum ' when debug is loaded this address will be 
patched

rdword tregone,IP ' the next word
test tregone, fMask wz
if_z add IP, #2 ' if the one of the hi bits is not set, it is an assembler 
word, inc IP
if_z jmp tregone
'fjernet instr !PBP 'rdword tregone, IP ' otherwise it is a forth word
mov tregfive, IP
add tregfive, #2
mov IP, tregone
call #a_rsPush
jmp #a_next
a__maskin
and stTOS, ina wz
muxnz stTOS, fLongMask
jmp # a_next
'nop, der ikke betyder noget, men skal være der, for at det virker
'fordi jeg har fjernet rdword tregone, IP !PBP
nop
''''''''''''''''''''''''''''''''''''''''''!PBP

: tt sc cnt COG@ words cnt COG@ swap - . ;
I did run both programs, which ran "words" and got 56ms. faster drive. 

Original issue reported on code.google.com by prof.bra...@gmail.com on 2 Jun 2013 at 6:26

[GoogleCodeExporter]

Posted on parallax forum by

forum user: frida
Location     Denmark
Posts     3 

Original comment by prof.bra...@gmail.com on 2 Jun 2013 at 6:27

[GoogleCodeExporter]

marco says:

Looks like the dictionary has been generated for the before code. As it is just 
a dump it doesn't get updated automatically for your after changes. That said, 
the nop - if used - should appear before a__maskin but an updated dictionary 
would be the right solution.

Original comment by prof.bra...@gmail.com on 2 Jun 2013 at 3:34

[GoogleCodeExporter]

Thank you, you are right. I have looked at the code and it makes sense. I will 
incorporate into PropForth6. 

The noop is necessary in your code because changes in the assembler kernel 
requires that the PropForth image needs to be updated to reflect the changes.

This involves a rebuild from StartKernel forward. Mostly automated, but a 
little tricky. Will try to improve this process in 6.

Original comment by salsa...@gmail.com on 2 Jun 2013 at 3:56

[GoogleCodeExporter]

While you're at it, get rid of fMask too. I don't know how important kernel 
size is but its single use can easily be rewritten with either a cmp wc or a 
testn wz.

Original comment by marko.lu...@kyi.biglobe.ne.jp on 3 Jun 2013 at 12:07

[GoogleCodeExporter]

[deleted comment]

[GoogleCodeExporter]

[deleted comment]

[GoogleCodeExporter]

Further loop improvement. Currently IP is only incremented when it's not a 
FORTH word. Subsequently tregfive is loaded with IP and then incremented then 
IP reused. IOW IP can always be incremented.

a_next
a_debugonoff        
        if_never        jmpret a_dum_ret, # a_dum

                        rdword  tregone,IP
                        add     IP, #2
                        testn   tregone, #$1FF    wz
        if_z            jmp     tregone

                        mov     tregfive, IP
                        mov     IP, tregone       
                        call    #a_rsPush
                        jmp     #a_next

Original comment by marko.lu...@kyi.biglobe.ne.jp on 3 Jun 2013 at 12:52

[GoogleCodeExporter]

Agreed, will update 6, and make sure regression tests pass. (do not see why 
they would not)

Thank you.

Original comment by salsa...@gmail.com on 3 Jun 2013 at 2:53

[GoogleCodeExporter]

If I understand it correctly, the IP must not be updated if it is a Forth word, 
but it is the content of IP, which is updated.

frida

Original comment by PoulB...@gmail.com on 4 Jun 2013 at 9:04

[GoogleCodeExporter]

Yes, the operation is as follows. 
The IP is pointing to the next word to be executed. After the word is fetched, 
the IP is incremented to point to the next word. If it is an assembler word, we 
simply jump to the address. If it is a forth word, we push the IP onto the 
return stack, and start fetching the contents of the word.

Original comment by salsa...@gmail.com on 5 Jun 2013 at 12:05