SSE2 generation bug with 4.1.2 and -O3

[Prakash Punnoor <prakash at punnoor dot de>]

Hi,

this is my simple test proggie:


#include <stdio.h>

#include <emmintrin.h>

union __u128i {
	unsigned int u[4];
	__m128i v;
};

static const union __u128i mask = {{0xffffffffU, 0x00000000U, 0xffffffffU, 
0x00000000U}};


static void test1()
{
	union __u128i input = {{1, 2, 3, 4}};
	int i;
	
	input.v = _mm_and_si128(input.v, mask.v);
	
	for(i=0;i<4;i++)
		printf("%u\n", input.u[i]);
	
}

static void test2()
{
	void *input;
	int i;
	
	input = _mm_malloc(4*sizeof(int), 16);
	
	for(i=0;i<4;i++)
		((unsigned int*)input)[i] = i+1;
	
	*(__m128i*)input = _mm_and_si128(*(__m128i*)input, mask.v);
	
	for(i=0;i<4;i++)
		printf("%u\n",((unsigned int*)input)[i]);
	
	_mm_free(input);
}

int main(){
	test1();
	printf("\n");
	test2();
	
	return 0;
}

test1 is always ok, but test2 fails with -O3 (produces output 1 0 0 0 for 
test2), but -O2 makes it behave.


I  am on x86_64. -m32 has interesting effects:

-O3 -Wall -m32 -msse2 -march=i486 is ok (output 1 0 3 0, as expected)
-O3 -Wall -m32 -msse2 -march=i586 is broken (output 0 2 3 4)
-O3 -Wall -m32 -msse2 -march=athlon is broken (output 1 0 0 0)

Is this the aliasing issues I read about with 4.1 compiler? Is there an easy 
work-around? In fact I need to pass data as aligned buffer, so I need test2 
to behave correctly. :-(

Do you need anything else?

Cheers,
-- 
(Â=                 =Â)
//\ Prakash Punnoor /\\
V_/                 \_V

Attachment: pgp00000.pgp
Description: PGP signature