// To compile: gcc -std=gnu99 -O2 -o performance performance.c

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

#define arrcpy(dest,src,size) for(size_t zz=0;zz<size;zz++)dest[zz]=src[zz]
#define dalloc(size) (test_t*)malloc(sizeof(test_t)*size)

typedef double test_t;
const size_t SIZE = 1E4;
const size_t ITERATIONS = 1E6;
const int MATH_TEST = 2;
const int MEMORY_TEST = 1;

inline void ptrcpy( test_t *src, test_t *dest, size_t size ) {
	test_t *srcptr = src+size, *destptr = dest+size;
	while ( srcptr > src ) {
		*(--destptr) = *(--srcptr);
	}
}

inline void ptrcpy2( test_t *src, test_t *dest, size_t size ) {
	test_t *srcptr = src, *destptr = dest;
	test_t *srcend = src+size;
	while ( srcptr < srcend ) {
		*(destptr++) = *(srcptr++);
	}
}


int main( int argc, char **argv ) {
	size_t i, j;
	test_t *src = dalloc( SIZE );
	test_t *dest = dalloc( SIZE );
	test_t *dest2 = dalloc( SIZE );
	time_t start_time, run_time;
	int tests = MATH_TEST | MEMORY_TEST;

	if ( argc > 1 && !strcmp( argv[1], "math" )) {
		tests = MATH_TEST;
	}
	if ( argc > 1 && !strcmp( argv[1], "mem" )) {
		tests = MEMORY_TEST;
	}

	if ( tests & MEMORY_TEST ) {
		printf( "Test data type size: %d bits\n", sizeof( test_t ) << 3 );
		printf( "Number of elements: %d\n", SIZE );
		printf( "Number of iterations: %d\n", ITERATIONS );

		printf( "Running memory read/write tests...\n" );
		time( &start_time );
		for( i=0; i < ITERATIONS; i++ ) {
			memcpy( dest, src, sizeof( test_t ) * SIZE );
			memcpy( dest2, src, sizeof( test_t ) * SIZE );
		}
		run_time = time( NULL ) - start_time;
		printf( "memcpy() run time %d seconds.\n", (int)run_time );

		time( &start_time );
		for( i=0; i < ITERATIONS; i++ ) {
			for( j=0; j < SIZE; j++ ) {
				dest[ j ] = src[ j ];
				dest2[ j ] = src[ j ];
			}
		}
		run_time = time( NULL ) - start_time;
		printf( "for() run time %d seconds.\n", (int)run_time );

		time( &start_time );
		for( i=0; i < ITERATIONS; i++ ) {
			for( j=0; j < SIZE; j++ ) {
				dest[ j ] = src[ j ];
			}
			for( j=0; j < SIZE; j++ ) {
				dest2[ j ] = src[ j ];
			}
		}
		run_time = time( NULL ) - start_time;
		printf( "dual for() run time %d seconds.\n", (int)run_time );

//		time( &start_time );
//		for( i=0; i < ITERATIONS; i++ ) {
//			arrcpy( dest, src, SIZE );
//			arrcpy( dest2, src, SIZE );
//		}
//		run_time = time( NULL ) - start_time;
//		printf( "arrcpy() run time %d seconds.\n", (int)run_time );

		time( &start_time );
		for( i=0; i < ITERATIONS; i++ ) {
			ptrcpy( dest, src, SIZE );
			ptrcpy( dest2, src, SIZE );
		}
		run_time = time( NULL ) - start_time;
		printf( "ptrcpy() run time %d seconds.\n", (int)run_time );


		time( &start_time );
		for( i=0; i < ITERATIONS; i++ ) {
			ptrcpy2( dest, src, SIZE );
			ptrcpy2( dest2, src, SIZE );
		}
		run_time = time( NULL ) - start_time;
		printf( "ptrcpy2() run time %d seconds.\n", (int)run_time );

	}

	free( src );
	free( dest );
	free( dest2 );

	// I'm not sure these are really giving an accurate picture. May need more work.
	if ( tests & MATH_TEST ) {
		printf( "Running math tests...\n" );
		time( &start_time );
		for( i=0; i < ITERATIONS * SIZE / 10; i++ ) {
			drand48( ) * drand48( );
		}
		run_time = time( NULL ) - start_time;
		printf( "Floating point multiplication test run time %d seconds.\n", (int)run_time );

		time( &start_time );
		for( i=0; i < ITERATIONS * SIZE / 10; i++ ) {
			rand( ) * rand( );
		}
		run_time = time( NULL ) - start_time;
		printf( "Integer multiplication test run time %d seconds.\n", (int)run_time );
	}

	return 0;
}