bcd2int_asm.c

#include <stdio.h>
#include <inttypes.h>
#include <x86intrin.h>
 
int bcd2int(int bcd)
{
	static __v2di output = {0,0};
	output[0]=bcd;
	asm volatile(
		“\tmovq\t%0, %%xmm0;\n”		// use 128 bit SSE xmm registers
		“\tmovq\t%%xmm0, %%xmm1;\n”	// make copy “04030201”
		“\tmovq\t$0x000f000f000f000f, %0;\n”
		“\tmovq\t%0, %%xmm2;\n”		// low nibble mask “000f”
		“\tshlq\t$8, %0;\n”
		“\tmovq\t%0, %%xmm3;\n”		// high nibble mask “0f00”
		“\tpand\t%%xmm2, %%xmm0;\n”	// only low “0001”
		“\tpand\t%%xmm3, %%xmm1;\n”	// only hi “0200”
		“\tpsrlw\t$4, %%xmm1;\n”		// “0020”
		“\tpaddw\t%%xmm1, %%xmm0;\n”	// “0020”+“0001“ => ”0021"
		“\tpackuswb\t%%xmm0, %%xmm0;\n”	// “0043” “0021” => “4321”
		“\tmovq\t%%xmm0, %2;\n”		//stored packed bcd
		“\tfbld\t%2;\n”				//convert pbcd->float
		“\tfistp\t%2;\n”			//float->int
		“\temms;\n”					//free FPU
		“\tmovq\t%2, %1;\n”
		:“=r” (output[0])
		:“0” (output[0]), “o” (output) // o instead m means try to keep memory reference short
	);
	return output[0];
}
 
int main()
{
	register uint64_t t;
	printf(“%d\n”, bcd2int(0x04050602));
	printf(“%d\n”, bcd2int(0x0));
	printf(“%d\n”, bcd2int(0x01020304));
	printf(“%d\n”, bcd2int(0x01010101));
	printf(“%d\n”, bcd2int(0x09090909));
	printf(“%d\n”, bcd2int(0x04030201));
 
	asm volatile(“cpuid;cpuid;cpuid”);
	t=__rdtsc()+72;
	for(register unsigned int i=1000000–1;i;i--) bcd2int(0x09080706);
	t=__rdtsc()-t;
	printf(“cycles: %ld\n”, t/1000000);
}