[cfe-dev] One little test, comparing clang and gcc

[Giangiacomo Mariotti]

On Thu, Oct 29, 2009 at 8:17 PM, Daniel Dunbar <daniel at zuster.org> wrote:
> I don't know the answers to these questions without digging in to your
> example -- there is a lot of information here which I find hard to
> read. I think you are asking about the time difference between
> clang_02 and gcc_02, this is probably due to some missed optimization
> which of course we always would like to fix. However, you'll have to
> actually dig in to the example to find out what is going on. If you
> can identify a small test case where clang clearly isn't optimizing
> something it should, please file a bugzilla for it
> (http://llvm.org/bugs).
>
Let's make the example code much shorter:

#include <stdio.h>
#include <time.h>
#include <stdlib.h>

#define MAXN 100000
int x[MAXN];
int startn = 5000;
int n;

#define T(s, n)\
do {\
	typeof(s) s__ = s;\
	typeof(n) n__ = n;\
	printf("%s (n=%d)\n", s__, n__);\
} while (0)

#define TRIALS 5

#define M(op)	do {						\
	printf(" %-26s", #op);				\
	k = 0;								\
	timesum = 0;						\
	for (ex = 0; ex < TRIALS; ex++) {	\
		start = clock();				\
		for (i = 1; i <= n; i++) {		\
			fi = (float) i;				\
			for (j = 1; j <= n; j++) {	\
				op;						\
			}							\
		}								\
		t = clock()-start;				\
		printf("%8d", t);				\
		timesum += t;					\
	}									\
	nans = 1e9 * timesum / ((double)	\
		n*n * TRIALS * CLOCKS_PER_SEC);	\
	printf("%8.0f\n", nans); } while (0)

int main(void)
{
	int i, j, k;
	float fi;
	int t, ex, timesum, start;
	double nans;

	for (i = 0; i < n; i++)
		x[i] = rand();
	n = startn;
	printf("C Time Cost Model, n=%d\n", n);

	M({});

	return 0;
}

Here, if we look at the instruction that is timed and we ask ourself,
what is that instruction doing? The answer is simply: absolutely
nothing.

Now these are the results for the code compiled as explained on the
previous mail:

->time_tests_001_gcc_O2
C Time Cost Model, n=5000
 {}                               0       0       0       0       0       0

->time_tests_001_clang_O2
C Time Cost Model, n=5000
 {}                           10000   20000   20000   20000   20000       1

So the point is, at optimization level -O2, gcc compiles out code that
does nothing, while clang optimizes it, but it still leaves the code
there.
(Notice that I haven't inspected the object file, so I may be wrong. I
wanted to post this here, before doing that)

Giangiacomo Mariotti