balaji.i...@gtri.gatech.edu wrote:
Hello Everyone,
I am creating a benchmark, where I have a following code:
x1 = a * b
x2 = x1 * a;
x3 = x1* x2;
x4 = x2 * x3;
x5 = x3 * x4;
x6 = x5 * x6;
. . .
. . .
x1000 = x999 * x998;
When I do this, and compile using -O3/-O2/-O1, the compiler freezes. Am I doing anything wrong?
GCC Version: 4.2.1
Platform: x86_64 (in macOS Snow Leopard).
I also tested this on a i386 (Red Hat 4.3.2-7) machine with gcc 4.3.2 and I had
the same response.
Please CC me in the response since I m not a subscribed member of this mailing list.
I have no real idea on this particular issue, but I did empirically observe that on some programs (for me, it was the
MELT generated C code on my MELT [1] branch eg before rev 154917) the register allocator (or some optimizing passes)
might have in practice an O(n^2) behavior, and might need a lot of memory
I would suggest:
* trying a very recent gcc (that you probably should build from its released source code), at least gcc 4.4.3
http://gcc.gnu.org/gcc-4.4/ or perhaps a gcc 4.5 snapshot.
* reducing the number of variables (to e.g. 300 or 500 instead of 1000).
* compiling first with -O1
* compiling with -ftime-report
* observe the cpu time and memory consumption of the compilation, eg by running
the top utility in a separate terminal.
So try first to reduce the number of variables (perhaps even to 200) and to
compile with
gcc-4.4 -O1 -ftime-report -v -c benchbalaji.c
I believe that recent GCC (ie 4.4 at least) had serious progress both on register allocation and on some optimization
passes, so they probably should work better on your [pathological] case.
Regards.
Note [1]: MELT is a lisp like dialect -translated into C- for coding GCC passes in plugins, and it used to generate huge
initialization routines. This is no more the case since a few weeks (because there is no more huge routines, only
several big ones), MELT can be reasonably compiled on a desktop even with optimization like -O2.
--
Basile STARYNKEVITCH http://starynkevitch.net/Basile/
email: basile<at>starynkevitch<dot>net mobile: +33 6 8501 2359
8, rue de la Faiencerie, 92340 Bourg La Reine, France
*** opinions {are only mines, sont seulement les miennes} ***
