Re: Stack vs stackless operation

Paul Rubin <no.email@nospam.invalid> writes:

anton@mips.complang.tuwien.ac.at (Anton Ertl) writes:

Results (on Zen4):
gforth-fast (development): ...

>
It's interesting how little difference there is with gforth-fast.  Could
you also do gforth-itc?

gforth-itc (development):
      :=:        exchange          ex         ex-locals      exchange2
 7_527_256_553 5_224_615_325  6_825_283_178  9_238_357_501  7_036_128_309 c.
13_127_503_990 9_326_561_471 12_927_054_153 16_927_820_825 12_027_146_677 i.

For comparison: gforth-fast (development):
      :=:        exchange         ex        ex-locals      exchange2
  814_881_277   879_389_133   928_825_521   875_574_895   808_543_975 cyc.
3_908_874_164 3_708_891_336 4_508_966_770 4_209_778_557 3_708_865_505 inst.

exchange2 is a big win with VFX, suggesting its
optimizer could do better with some of the other versions.

On VFX exchange2 takes the same speed and the same number of
instructions as :=:.  EX is slower because VFX does not analyse the
return stack, unlike the data stack.  EX-LOCALS is slow because VFX's
locals implementation is not particularly good.

To see what a better analysis can do, let's look at lxf:

      :=:           ex         ex-locals     exchange2
  502_740_029   502_189_567   502_134_842   502_043_217 cycles
1_701_663_782 1_701_657_866 1_701_677_273 1_701_684_186 instructions

The cycles and instructions are worse (except for ex-locals) than with
VFX, but that's due to inlining (which VFX does and lxf does not).

E.g., here's lxf's code for EX-LOCALS:

 869204C  804FCE2  23  88C8000   5 normal  EX-LOCALS

 804FCE2 8B4500                 mov     eax , [ebp]
 804FCE5 8B00                   mov     eax , [eax]
 804FCE7 8BCB                   mov     ecx , ebx
 804FCE9 8B09                   mov     ecx , [ecx]
 804FCEB 8B5500                 mov     edx , [ebp]
 804FCEE 890A                   mov     [edx] , ecx
 804FCF0 8903                   mov     [ebx] , eax
 804FCF2 8B5D04                 mov     ebx , [ebp+4h]
 804FCF5 8D6D08                 lea     ebp , [ebp+8h]
 804FCF8 C3                     ret     near

It's the same code as lxf produces for :=:.

The code lxf produces for EX and EXCHANGE2 is:

 804FCF9 8BC3                   mov     eax , ebx
 804FCFB 8B00                   mov     eax , [eax]
 804FCFD 8B4D00                 mov     ecx , [ebp]
 804FD00 8B09                   mov     ecx , [ecx]
 804FD02 890B                   mov     [ebx] , ecx
 804FD04 8B5D00                 mov     ebx , [ebp]
 804FD07 8903                   mov     [ebx] , eax
 804FD09 8B5D04                 mov     ebx , [ebp+4h]
 804FD0C 8D6D08                 lea     ebp , [ebp+8h]
 804FD0F C3                     ret     near

- anton
--
M. Anton Ertl  http://www.complang.tuwien.ac.at/anton/home.html
comp.lang.forth FAQs: http://www.complang.tuwien.ac.at/forth/faq/toc.html
     New standard: https://forth-standard.org/
EuroForth 2023 proceedings: http://www.euroforth.org/ef23/papers/
EuroForth 2024 proceedings: http://www.euroforth.org/ef24/papers/