Re: A bug in IA32 assembler

[Richard Henderson <rth at redhat dot com>]

On Fri, Jun 29, 2001 at 12:15:56AM -0700, H . J . Lu wrote:
> This one:
> 
>         asm("movaps (%0),%%xmm0"                                //SSE
>         :
>         :"g"(A));

This one is wrong because A is an array, which decomposes to a pointer,
which has the constant value A.2.  And you've told the compiler that
immediates are acceptable.  So it prints A.2 as an immediate.

It's also wrong in that you're _not_ telling the compiler that you are
reading from memory.  So it'll be happy to move stores across this asm,
or even remove them as dead.

>         asm("movaps %%xmm0,(%0)":"=g"(C));                        //SSE

Likewise, you're not telling the compiler that you are storing to memory.

The easiest way to write things correctly is like so.



r~


#include <stdio.h>

typedef struct
{
  float x[4] __attribute__((aligned(16)));
} V4SF;

static void dump (const char *name, const V4SF *v)
{
	int i;

	printf("float %s=(", name);
	for(i = 0; i < 4; i++)
		printf("%f,", v->x[i]);
	printf("\b)\n");
}

int main()
{
	static V4SF A = {1.,1.,1.,1.};
	static V4SF B = {0.,1.,2.,3.};
	static V4SF C;

	dump("A", &A);
	dump("B", &B);

	asm("movaps %0,%%xmm0" : :"m"(A));
	asm("movaps %0,%%xmm1" : :"m"(B));

	asm("addps %xmm1,%xmm0");
//      asm("mulps %xmm1,%xmm0");

	asm("movaps %%xmm0,%0" : "=m"(C));

	dump("C", &C);

	return 0;
}