improving code length

**memcpy** · 03-04-2012

This is my x86-specific memcmp() that works much faster than default (or even the C rendition of the same concept), the only problem is that it occupies a lot of lines.

Code:

int asm_memcmp_block (const void *s1, const void *s2, int len)
{
	register int retval;
	
	asm volatile (
		      "xorq	%%rbx, %%rbx\n"


		      "cmp_8:"
		      "cmpq	$8, %%rax\n"
		      "jl	cmp_4\n"
		      "movq	(%%rsi, %%rbx, 0x8), %%rcx\n"
		      "movq	(%%rdi, %%rbx, 0x8), %%rdx\n"
		      "cmpq	%%rcx, %%rdx\n"
		      "jne	_diff\n"
		      "addq	$8, %%rbx\n"
		      "subq	$8, %%rax\n"
		      "jmp	cmp_8\n"
		      
		      "cmp_4:"
		      "cmpq	$4, %%rax\n"
		      "jl	cmp_2\n"
		      "movl	(%%rsi, %%rbx, 0x4), %%ecx\n"
		      "movl	(%%rdi, %%rbx, 0x4), %%edx\n"
		      "cmpl	%%ecx, %%edx\n"
		      "jne	_diff\n"
		      "addq	$4, %%rbx\n"
		      "subq	$4, %%rax\n"
		      "jmp	cmp_4\n"
		      
		      "cmp_2:"
		      "cmpq	$2, %%rax\n"
		      "jl	cmp_1\n"
		      "movw	(%%rsi, %%rbx, 0x2), %%cx\n"
		      "movw	(%%rdi, %%rbx, 0x2), %%dx\n"
		      "cmpw	%%cx, %%dx\n"
		      "jne	_diff\n"
		      "addq	$2, %%rbx\n"
		      "subq	$2, %%rax\n"
		      "jmp	cmp_2\n"
		      
		      "cmp_1:"
		      "cmpq	$1, %%rax\n"
		      "jl	_same\n"
		      "movb	(%%rsi, %%rbx, 0x1), %%cl\n"
		      "movb	(%%rdi, %%rbx, 0x1), %%dl\n"
		      "cmp	%%cl, %%dl\n"
		      "jne	_diff\n"
		      "incq	%%rbx\n"
		      "decq	%%rax\n"
		      "jmp	cmp_1\n"
		      
		      "_same:"
		      "xorq	%%rcx, %%rcx\n"
		      "jmp	_out\n"


		      "_diff:"
		      "movq	$1, %%rcx\n"
		      "jmp	_out\n"
		      
		      "_out:"
		      
		      : "=c" (retval)
		      : "S" (s1), "D" (s2), "a" (len)
		      : "rbx", "rcx", "rdx"
		      );
	
	return retval;
}

Does anyone have any suggestions as to shorten and/or improve the code? Perhaps using a macro for each value?

**oogabooga** · 03-05-2012

Two ideas.

1. Could you compare the blocks backwards, starting at the end, and thereby not use the b register? (But that might cause alignment problems and wouldn't work if you wanted to have it act more like memcmp and say whether the first different byte is greater or less.)

2. You could presumably get rid of cmp_4 and cmp_2 without any real loss, since if you do that cmp_1 will only need to compare 7 bytes in the worst case. If you don't want to do that, then cmp_1 can get rid of it's loopback since it's only comparing one byte.

EDIT: Now that I think about it, cmp_4 and cmp_2 will never loop either.

**memcpy** · 03-05-2012

Thanks, that helped. Also, in addition to those changes, is this code standard C:

Code:

int asm_memcmp_block (const void *s1, const void *s2, int len)
{
    register int retval;    
    
#define MEMCMP_BLK(x,y) "cmp_"#x":"                            \
                    "cmpq    $"#x", %%rax\n"                   \
                    "jl      cmp_"#y"\n"                       \
                    "movq    (%%rdi, %%rax, "#x"), %%rbx\n"    \
                    "cmpq    (%%rsi, %%rax, "#x"), %%rbx\n"    \
                    "jne     _diff\n"                          \
                    "subq    $"#x", %%rax\n"                   \  
                    "jmp     cmp_"#x"\n"                      


    asm volatile (
              MEMCMP_BLK ( 8, 1 )
              
              MEMCMP_BLK ( 1, 0 )


              "cmp_0:"
              "xorq    %%rcx, %%rcx\n"
              "jmp    _out\n"


              "_diff:"
              "movq    $1, %%rcx\n"
              "jmp    _out\n"
              
              "_out:"
              
              : "=c" (retval)
              : "S" (s1), "D" (s2), "a" (len)
              : "rbx"
              );
    
    return retval;
}

**iMalc** · 03-06-2012

The moment you start using assembly, you're into compiler specific extensions, so no there's nothing standard about it.

**oogabooga** · 03-06-2012

You could also maybe look into the cmps instruction.

**adeyblue** · 03-06-2012

memcmp also returns a value less than 0 when the data dictates

Thread: improving code length

Thread Tools

Search Thread

Display

improving code length

Similar Threads

help improving code

Improving Code Interface

Improving my code

Code improving

Improving code