ASM Code Problems

mlpa · 09-02-2010, 08:44 AM

Hi everyone, I'm making a multimedia library to speedup vector and matrix computation.

But I'm getting a strange error, in this code

Code:

void memcpy_unrolled_sse2(void *dst, void *src, int size)
{
	int i = 0,
	    length = (size/16)*16,
	    length2 = (size/32)*32,
	    length4 = (size/64)*64,
	    length8 = (size/128)*128;

	char *a = (char *)src;
	char *b = (char *)dst;

	for(i = 0; i < length8; i+=128)
	{
		__asm__ volatile
                ( // instruction             comment
                "\n\t movdqa     %8,%%xmm0         \t#"
		"\n\t movdqa     %9,%%xmm1         \t#"
		"\n\t movdqa     %10,%%xmm2        \t#"
		"\n\t movdqa     %11,%%xmm3        \t#"
		"\n\t movdqa     %12,%%xmm4        \t#"
		"\n\t movdqa     %13,%%xmm5        \t#"
		"\n\t movdqa     %14,%%xmm6        \t#"
		"\n\t movdqa     %15,%%xmm7        \t#"
                "\n\t movdqa     %%xmm0,%0         \t#"
		"\n\t movdqa     %%xmm1,%1         \t#"
		"\n\t movdqa     %%xmm2,%2         \t#"
		"\n\t movdqa     %%xmm3,%3         \t#"
		"\n\t movdqa     %%xmm4,%4         \t#"
		"\n\t movdqa     %%xmm5,%5         \t#"
		"\n\t movdqa     %%xmm6,%6         \t#"
		"\n\t movdqa     %%xmm7,%7         \t#"
		: "=m" (b[i]),     // %0
		  "=m" (b[i+16]),  // %1
		  "=m" (b[i+32]),  // %2
		  "=m" (b[i+48]),  // %3
		  "=m" (b[i+64]),  // %4
		  "=m" (b[i+80]),  // %5
		  "=m" (b[i+96]),  // %6
		  "=m" (b[i+112])  // %7
		: "m"  (a[i]),     // %8
		  "m"  (a[i+16]),  // %9
		  "m"  (a[i+32]),  // %10
		  "m"  (a[i+48]),  // %11
		  "m"  (a[i+64]),  // %12
		  "m"  (a[i+80]),  // %13
		  "m"  (a[i+96]),  // %14
		  "m"  (a[i+112])  // %15
		);
	}

	for(i = length8; i < length4; i+=64)
	{
		__asm__ volatile
                ( // instruction             comment
                "\n\t movdqa     %4,%%xmm0         \t#"
		"\n\t movdqa     %5,%%xmm1         \t#"
		"\n\t movdqa     %6,%%xmm2         \t#"
		"\n\t movdqa     %7,%%xmm3         \t#"
	        "\n\t movdqa     %%xmm0,%0         \t#"
		"\n\t movdqa     %%xmm1,%1         \t#"
		"\n\t movdqa     %%xmm2,%2         \t#"
		"\n\t movdqa     %%xmm3,%3         \t#"
		: "=m" (b[i]),     // %0
		  "=m" (b[i+16]),  // %1
		  "=m" (b[i+32]),  // %2
		  "=m" (b[i+48])   // %3
		: "m"  (a[i]),     // %4
		  "m"  (a[i+16]),  // %5
		  "m"  (a[i+32]),  // %6
		  "m"  (a[i+48])   // %7
		);
	}
	
	for(i = length4; i < length2; i+=32)
	{
		__asm__ volatile
                ( // instruction             comment
                "\n\t movdqa     %2,%%xmm0         \t#"
		"\n\t movdqa     %3,%%xmm1         \t#"
                "\n\t movdqa     %%xmm0,%0         \t#"
		"\n\t movdqa     %%xmm1,%1         \t#"
		: "=m" (b[i]),     // %0
		  "=m" (b[i+16])   // %1
		: "m"  (a[i]),     // %2
		  "m"  (a[i+16])   // %3
		);
	}

	for(i = length2; i < length; i+=16)
	{
		__asm__ volatile
                ( // instruction             comment
                "\n\t movdqa     %1,%%xmm0         \t#"
                "\n\t movdqa     %%xmm0,%0         \t#"
		: "=m" (b[i])      // %0
		: "m"  (a[i])      // %1
		);
	}


	for(i = length; i < size; i++)
		b[i] = a[i];
}

When I compile without -O1,2,3 I get this error:

Code:

libMultimedia.cpp: In function 'void memcpy_unrolled_sse2(void*, void*, int)':
libMultimedia.cpp:5033: error: can't find a register in class 'GENERAL_REGS' while reloading 'asm'
libMultimedia.cpp:5033: error: 'asm' operand has impossible constraints
make: *** [libMultimedia.o] Error 1

Can some one help me?

grail · 09-03-2010, 12:59 AM

As we obviously don't have the more than 5000 lines of code where the error is, are you able to identify which line it is in the code you have presented?

Kenny_Strawn · 09-03-2010, 03:02 AM

It doesn't look like it's the full source code. It looks like a single function, possibly defined even before main() is declared. So we know the error is somewhere within that function.

mlpa · 09-03-2010, 03:31 AM

The error is in the first block of asm code the line with );
This project is hosted in google code here.

I searched google but all I can find is pieces of asm code with the same error, never a solution.

ntubski · 09-05-2010, 10:07 AM

When compiling without optimization (-O0) gcc ends up allocating one register per operand, but there are only 15 registers in total including the stack and frame pointers, thus the error. You can rewrite the assembly so only 2 registers are used:

Code:

            asm
                ( // instruction             comment
                 "\n\t movdqa 0x00(%1), %%xmm0           \t#"
                 "\n\t movdqa 0x10(%1), %%xmm1           \t#"
                 "\n\t movdqa 0x20(%1), %%xmm2           \t#"
                 "\n\t movdqa 0x30(%1), %%xmm3           \t#"
                 "\n\t movdqa 0x40(%1), %%xmm4           \t#"
                 "\n\t movdqa 0x50(%1), %%xmm5           \t#"
                 "\n\t movdqa 0x60(%1), %%xmm6           \t#"
                 "\n\t movdqa 0x70(%1), %%xmm7           \t#"

                 "\n\t movdqa %%xmm0,0x00(%0)         \t#"
                 "\n\t movdqa %%xmm1,0x10(%0)         \t#"
                 "\n\t movdqa %%xmm2,0x20(%0)         \t#"
                 "\n\t movdqa %%xmm3,0x30(%0)         \t#"
                 "\n\t movdqa %%xmm4,0x40(%0)         \t#"
                 "\n\t movdqa %%xmm5,0x50(%0)         \t#"
                 "\n\t movdqa %%xmm6,0x60(%0)         \t#"
                 "\n\t movdqa %%xmm7,0x70(%0)         \t#"
                 : "=r" (b) : "r" (a)
                 : "xmm0", "xmm1", "xmm2", "xmm3",
                   "xmm4", "xmm5", "xmm6", "xmm7"
		);