[gambit-list] (declare (mostly-fixnum)) vs (declare (mostly-fixnum-flonum))

[Bradley Lucier]

Marc:

I compiled an image processing application with mostly-fixnum and  
mostly-fixnum-flonum and here is what I found:

[lindv2:Arrays/srfi3/gsc-test] lucier% ll
total 107960
-rw-r--r--   1 lucier  lucier  4103052 Mar 21 15:51 all.c.fixnum
-rw-r--r--   1 lucier  lucier  5489576 Mar 22 21:15 all.c.fixnum-flonum
-rw-r--r--   1 lucier  lucier  6556015 Mar 22 20:28 all.i.fixnum
-rw-r--r--   1 lucier  lucier  8689958 Mar 22 21:15 all.i.fixnum-flonum
-rw-r--r--   1 lucier  lucier  1811056 Mar 22 20:45 all.o.fixnum
-rw-r--r--   1 lucier  lucier  2492112 Mar 22 21:45 all.o.fixnum-flonum
-rwxr-xr-x   1 lucier  lucier  1449209 Mar 22 20:46 all.o1.fixnum*
-rwxr-xr-x   1 lucier  lucier  2029916 Mar 22 21:45 all.o1.fixnum- 
flonum*
-rw-r--r--   1 lucier  lucier  5280685 Mar 22 20:45 all.s.fixnum
-rw-r--r--   1 lucier  lucier  7388908 Mar 22 21:45 all.s.fixnum-flonum
-rw-r--r--   1 lucier  lucier      558 Mar 22 21:15 all.scm
-rw-r--r--   1 lucier  lucier  1631997 Mar 22 22:01 fixnum-expansion.scm
-rw-r--r--   1 lucier  lucier  2151785 Mar 22 21:49 fixnum-flonum- 
expansion.scm
-rwxr--r--   1 lucier  lucier      371 Mar 21 15:43 gaussian-random.scm*
-rwxr--r--   1 lucier  lucier    88397 Mar 21 15:45 generic-arrays.scm*
-rwxr--r--   1 lucier  lucier    25985 Mar 21 15:46 multi-calculus.scm*
-rw-r--r--   1 lucier  lucier     6683 Mar 21 15:48 pgm.scm
-rwxr--r--   1 lucier  lucier     9778 Mar 21 15:43 pixel-method.scm*
-rwxr--r--   1 lucier  lucier     7113 Mar 21 15:46 point.scm*

You can see the size of the .scm files, the .c files, the .i files  
after macro expansion, the .s files, etc.

On a 2GHz Mac G5, the mostly-fixnum version took 2.4GB and 17 minutes  
to compile and the mostly-fixnum-flonum version took 3.4GB and 30  
minutes to compile.  This was with gcc-4.1.0 and not the compiler  
that apple ships (since the compiler that apple ships is 32-bit and  
can't allocate over 2.0 GB ;-).

The mostly-fixnum-flonum code is not as big as one might expect, for  
the following reasons:

1.  If you have (+ i 1), then because 1 is not a flonum, it doesn't  
generate the tests or the code for flonum.+.

2.  Comparisons are relatively simple, e.g., (< x y) just adds "check  
whether x and y are flonums and compare them" to the code, no boxing,  
etc.

So I don't see mostly-flonum-fixnum as that much of an added burden  
over mostly-fixnum.

Brad