GENERAL PROTECTION FAULT - Why?

[hutch--]

You guys worry me with your register usage, if you properly comply with the Intel ABI (preserve EBX ESI EDI EBP and ESP and trash EAX ECX and EDX) you will never get into trouble. Do it wrong and you will keep getting mysterious crashes. (Voice crying in the wilderness etc ....)

Dave, the main reason why you make a deviant branching mechanism is for speed yet you are using an antique and really SSSLLLOOOOOOOWWWWWWWWW "lodsd" instruction that will kill any speed advantage with a dispatcher. This is DOS brain era yet Intel in manuals from early Pentiums upwards to current i3/5/7 hardware recommend against using LODS, MOVS etc ... without the REP(E) prefix.

[dedndave]

in the grander scheme of things, i doubt the difference between
LODSD | JMP EAX
and
ADD ESI,4 | JMP [ESI]
is nearly as big as you think   :P

if you put any kind of code in the little module, it will likely overshadow the difference
but, if you like, i can whip out Michael's timing macro and we can test it   

[jj2007]

in the grander scheme of things, ...

It's definitely a bit shorter ;-)

Intel(R) Celeron(R) M CPU        420  @ 1.60GHz (SSE3)
loop overhead is approx. 188/100 cycles

?    cycles for 100 * lodsd & jmp
?    cycles for 100 * add esi & jmp

30      bytes for lodsd & jmp
46      bytes for add esi & jmp

[dedndave]

that's not the test we were talking about
there has to be some "signifigant" code in each module
nonetheless, here are my results...

Code Select Expand

Intel(R) Pentium(R) 4 CPU 3.00GHz (SSE3)
+19 of 20 tests valid, loop overhead is approx. 243/100 cycles

28529   cycles for 100 * lodsd & jmp
19006   cycles for 100 * add esi & jmp

28495   cycles for 100 * lodsd & jmp
18970   cycles for 100 * add esi & jmp

28514   cycles for 100 * lodsd & jmp
18981   cycles for 100 * add esi & jmp

i am going to call that 29 vs 19 cycles
if you put 500 cycles of code in each section, it becomes rather insignifigant

[japheth]

It's definitely a bit shorter ;-)

But it's also cheating, because "ADD ESI,4 | JMP [ESI]" is not even vaguely equivalent to "LODSD | JMP EAX". Perhaps you meant "ADD ESI,4 | JMP [ESI-4]"?

[jj2007]

It's definitely a bit shorter ;-)

But it's also cheating, because "ADD ESI,4 | JMP [ESI]" is not even vaguely equivalent to "LODSD | JMP EAX". Perhaps you meant "ADD ESI,4 | JMP [ESI-4]"?

It is, it is, just read the posts before writing. Don't be scared, it's plain assembler, not MasmBasic :icon_mrgreen:

And surprisingly, on my cheap Intel CPU it's even faster...

Intel(R) Celeron(R) M CPU        420  @ 1.60GHz (SSE3)
2033    cycles for 100 * lodsd & jmp
2132    cycles for 100 * add esi & jmp

On AMD instead, it's 12 vs 9 cycles for the five jumps:

AMD Athlon(tm) Dual Core Processor 4450B (SSE3)
12744   cycles for 100 * lodsd & jmp
9169    cycles for 100 * add esi & jmp

[japheth]

It is, it is, ...

It is what? It is ... cheating? - or: It is ... equivalent? Try to express yourself clearly!

... just read the posts before writing.

Ok, I finally did - but it didn't help.

Don't be scared, it's plain assembler, not MasmBasic :icon_mrgreen:

Ok .. but it still looks awe.......some.

[jj2007]

"ADD ESI,4 | JMP [ESI]" is not even vaguely equivalent to "LODSD | JMP EAX"

It is, it is, ...

It is what? It is ... cheating? - or: It is ... equivalent? Try to express yourself clearly!

I am so sorry! What I meant is: "ADD ESI,4 | JMP [ESI]" is not even vaguely fully equivalent to "LODSD | JMP EAX" in the context of the ten lines of plain assembly code posted above by Dave. Apologies for not expressing myself clearly - I'm not a native English speaker, y'know ;-)

[dedndave]

the difference is that EAX gets destroyed   :P

[frktons]

the difference is that EAX gets destroyed   :P

How beautiful. Let's destroy something togheter  :P

Intel(R) Core(TM)2 CPU          6600  @ 2.40GHz (SSE4)
loop overhead is approx. 187/100 cycles

3153    cycles for 100 * lodsd & jmp
3000    cycles for 100 * add esi & jmp

3151    cycles for 100 * lodsd & jmp
3000    cycles for 100 * add esi & jmp

3149    cycles for 100 * lodsd & jmp
3003    cycles for 100 * add esi & jmp

30      bytes for lodsd & jmp
46      bytes for add esi & jmp


--- ok ---

It's about 5% difference. Not that much I suppose.