For those interested in how the native backend for Gambit is progressing, here are some links to the paper that was presented yesterday at the Scheme workshop, and the slides of the talk. Laurent Huberdeau did a great job!
http://www.iro.umontreal.ca/~gambit/SW2018-talk.pdf
https://www.brinckerhoff.org/scheme2018/papers/Huberdeau_Feeley.pdf
Marc
Afficher les réponses par date
Hallo,
On 29/09/18 21:29, Marc Feeley wrote:
For those interested in how the native backend for Gambit is progressing, here are some links to the paper that was presented yesterday at the Scheme workshop, and the slides of the talk. Laurent Huberdeau did a great job!
This is very interesting! In which branch is this being developed? How can we try it out?
Cheers,
On Oct 1, 2018, at 4:24 AM, Alex Silva asandroq@gmail.com wrote:
Hallo,
On 29/09/18 21:29, Marc Feeley wrote:
For those interested in how the native backend for Gambit is progressing, here are some links to the paper that was presented yesterday at the Scheme workshop, and the slides of the talk. Laurent Huberdeau did a great job!
This is very interesting! In which branch is this being developed? How can we try it out?
Here are the steps (on a x86-64 platform):
% git checkout cpu % ./configure --enable-single-host CC="gcc -D___SUPPORT_LOWLEVEL_EXEC" % make % cp bin/gambcomp-C bin/gambcomp-x86-64 % cat > fib.scm (declare (standard-bindings) (block) (not safe))
(define (fib n) (if (fx< n 2) n (fx+ (fib (fx- n 1)) (fib (fx- n 2)))))
(pp (time (fib 40))) ^D % gsc/gsc -:=. -target x86-64 fib.scm % gsi/gsi fib (time (fib 40)) 289 ms real time 288 ms cpu time (288 user, 0 system) no collections no bytes allocated 1 minor fault no major faults 102334155 % gsc/gsc -:=. fib.scm % gsi/gsi fib (time (fib 40)) 315 ms real time 316 ms cpu time (316 user, 0 system) no collections no bytes allocated 1 minor fault no major faults 102334155
Here is a list of the primitives that are inlined by the native backend. If you use other primitives, they will be actual calls to the runtime library compiled with the C backend (i.e. slow). Remember that this is work in progress and that some things are not yet optimized and some things are unimplemented or broken.
##not ##identity
##fixnum? ##pair? ##null? ##char? ##special? ##mem-allocated?
##fx+ ##fx+? ##fx- ##fx-? ##fx< ##fx<= ##fx> ##fx>= ##fx=
##cons ##car ##cdr ##set-car! ##set-cdr!
##vector-ref ##vector-set! ##vector-length
So at this point the native backend is more a “proof of concept” and there is still lots of work to be done before this is a practical alternative to the C backend.
Marc
On 10/01/2018 11:42 AM, Marc Feeley wrote:
Here is a list of the primitives that are inlined by the native backend. If you use other primitives, they will be actual calls to the runtime library compiled with the C backend (i.e. slow). Remember that this is work in progress and that some things are not yet optimized and some things are unimplemented or broken.
##not ##identity
##fixnum? ##pair? ##null? ##char? ##special? ##mem-allocated?
##fx+ ##fx+? ##fx- ##fx-? ##fx< ##fx<= ##fx> ##fx>= ##fx=
##cons ##car ##cdr ##set-car! ##set-cdr!
##vector-ref ##vector-set! ##vector-length
So at this point the native backend is more a “proof of concept” and there is still lots of work to be done before this is a practical alternative to the C backend.
And here's roughly how many "primitives" there are in a backend:
heine:~/programs/gambit/gambit/gsc> grep '("##' _prims.scm | wc 459 3230 34354
Many of the remaining vector primitives (##u8vector-ref, ...) can be modeled on the currently implemented primitives, but there are many others that will require some thought. And perhaps many can remain accessed in the existing C versions through the new "bridge" put together by Marc and Laurent.
Once complex routines (beyond ack, fib, tak, ...) are tested, issues about register allocation, etc. (all the things that are big issues in any compiler) will become more important.
Brad