Darren New escreveu:
> Contrast with something that generates machine code natively, which 
> might output code that your C compiler would never output, for example.
> 
> If your front-end outputs C, you're never going to out-perform the C 
> code it generates. If your C compiler can't take advantage of the GPU to 
> do math or can't hook into interrupt vectors directly or ..., then your 
> language isn't going to do better.

It is very difficult to write compilers that generate optimized native 
code as good as the ones already provided by mature C compilers that 
have been around for ages.

Larceny is the only free Scheme implementation that tries to generate 
native code directly (or .NET bytecodes) and is written by one of the 
top Scheme implementers around.  It generates native code that is a good 
deal faster than JIT-compiled like in mzscheme or ikarus, but still much 
slower ones by Gambit or Stalin, both of which generate C code.

BTW, if you look at C code so produced, it hardly looks like C at all. 
Like:

___DEF_OLD_SYM_GLO(___S_pcase9,___G_pcase9)
___DEF_OLD_SYM_GLO(___S_peek_2d_char,___G_peek_2d_char)
___DEF_OLD_SYM_GLO(___S_peek_2d_eof_3f_,___G_peek_2d_eof_3f_)
___DEF_OLD_SYM_GLO(___S_peer,___G_peer)
___DEF_OLD_SYM_GLO(___S_permanent_2d_close,___G_permanent_2d_close)
___DEF_OLD_SYM_GLO(___S_permissions,___G_permissions)

from Gambit.  Truly insane.  But the good thing is that code is now 
portable to different 32-bit platforms as long as there's a C compiler. 
  Would it be worth the pain generating assebly not as optimized as a C 
compiler?...

-- 
a game sig: http://tinyurl.com/d3rxz9

Post a reply to this message