#include <stdio.h>

unsigned int 
rdtsc()
{
	unsigned int eax, edx;
	asm volatile ("rdtsc"
		      :"=a"(eax), "=d"(edx));

	return eax;
}

#define N 9

#define A16(P)					\
	P P P P					\
	P P P P					\
	P P P P					\
	P P P P					\

#define GEN(name, A)						\
int time_##name[N];						\
int								\
name(int a, int b, int n, int x)				\
{								\
	int ret;						\
	unsigned int begin, end;				\
	int i;							\
								\
	begin = rdtsc();					\
	for (i=n-1; i>=0; i--) {				\
		asm volatile (A16(A)				\
			      :::"eax", "ebx", "ecx", "edx");	\
	}							\
	asm("xor %%eax, %%eax":::"eax");			\
	asm("xor %%ebx, %%ebx":::"ebx");			\
	asm("xor %%ecx, %%ecx":::"ecx");			\
	asm("xor %%edx, %%edx":::"edx");			\
	end = rdtsc();						\
								\
	asm volatile ("emms");					\
	time_##name[x] = end-begin;				\
								\
	return ret;						\
}

GEN(add, "add %%eax, %%ecx\n\t")
GEN(mul, "imul %%ecx, %%edx\n\t")
GEN(paraadd, "add %%eax, %%ebx\n\tadd %%ecx, %%edx\n\t");
GEN(paramul, "imul %%eax, %%ebx\n\timul %%ecx, %%edx\n\t")
GEN(depadd, "add %%eax, %%ebx\n\tadd %%ecx, %%ebx\n\t");
GEN(empty, "#\n\t")
GEN(nop, "xchg %%eax, %%eax\n\t")
GEN(removedep, "add %%eax, %%ecx\n\txor %%ecx, %%ecx\n\t")
GEN(addrgen, "leal 4(%%eax,%%ecx), %%eax\n\t")
GEN(paraaddrgen, "leal 4(%%eax,%%ebx), %%eax\n\tleal 4(%%ecx,%%edx), %%ecx\n\t")
GEN(pmadd, "paddd %%mm1, %%mm0\n\t")
GEN(para_pmadd, "paddd %%mm4, %%mm5\n\tpaddd %%mm2, %%mm3\n\tpaddd %%mm1, %%mm0\n\t")
GEN(para_pmaddw, "paddw %%mm4, %%mm5\n\tpaddw %%mm2, %%mm3\n\tpaddw %%mm1, %%mm0\n\t")
GEN(para_pmaddw8,"paddw %%mm0, %%mm0\n\t"
    "paddw %%mm1, %%mm1\n\t"
    "paddw %%mm2, %%mm2\n\t"
    "paddw %%mm3, %%mm3\n\t"
    "paddw %%mm4, %%mm4\n\t"
    "paddw %%mm5, %%mm5\n\t"
    "paddw %%mm6, %%mm6\n\t"
    "paddw %%mm7, %%mm7\n\t");

GEN(incpc, "jmp 1f\n1:\n\t")
GEN(nopcall, "call hoge\n\t")
GEN(indirect, "movl hoge_adr, %%ecx\n\tmovl %%ecx, %%eax\n\tcall *%%eax\n\t")

int time_load[N];
int
load(int a, int b, int n, int x)
{
	int ret;
	unsigned int begin, end;
	int i;

	int stack, stack2;

	begin = rdtsc();
	for (i=0; i<n; i++) {
		asm volatile (A16("mov (%%eax), %%ecx\n\t"
				  "add %%ecx, %%edx\n\t")
			      :
			      :"a"(&stack2), "S"(&stack2)
			      :"ecx", "edx");
	}
	end = rdtsc();

	time_load[x] = end-begin;

	return ret;
}

asm("hoge:ret");

extern int hoge;
int *hoge_adr =  &hoge;

int alias_flag;

int time_para_loadstore[N];
int
para_loadstore(int a, int b, int n, int x)
{
	int ret;
	unsigned int begin, end;
	int i;
	int *p;

	int stack, stack2;

	if (alias_flag) {
		p = &stack2;
	} else {
		p = &stack;
	}

	begin = rdtsc();
	for (i=0; i<n; i++) {
		asm volatile (A16("add %%ecx, (%%eax)\n\t"
				  "add %%edx, (%%esi)\n\t")
			      :
			      :"a"(p), "S"(&stack2)
			      :"ecx", "edx");
	}
	end = rdtsc();

	time_para_loadstore[x] = end-begin;

	return ret;
}


void
run(const char *name, int*a, int (*f)(int,int,int,int))
{
	int i, j;

	puts(name);
	for (j=0, i=1; j<N; j++, i*=7) {
		f(100, 4, i, j);
	}

	for (j=0, i=1; j<N; j++, i*=7) {
		printf("%d:%f\n",
		       i, 
		       a[j]/(i*16.0));
	}
}


int
main()
{
	unsigned int t, u;
	t = rdtsc();
	u = rdtsc();
	printf("%d\n", u-t);

	run("empty", time_empty, empty);
	run("nop", time_nop, nop);
	run("add", time_add, add);
	run("mul", time_mul, mul);
	run("paraadd", time_paraadd, paraadd);
	run("paramul", time_paramul, paramul);
	run("depadd", time_depadd, depadd);
	run("removedep", time_removedep, removedep);
	run("addrgen", time_addrgen, addrgen);
	run("paraaddrgen", time_paraaddrgen, paraaddrgen);
	run("incpc", time_incpc, incpc);
	run("nopcall", time_nopcall, nopcall);
	run("indirect", time_indirect, indirect);
	run("load", time_load, load);
	run("pmadd", time_pmadd, pmadd);
	run("para_pmadd", time_para_pmadd, para_pmadd);
	run("para_pmaddw", time_para_pmaddw, para_pmaddw);
	run("para_pmaddw8", time_para_pmaddw8, para_pmaddw8);
	alias_flag=0;
	run("paraloadstore", time_para_loadstore, para_loadstore);
	alias_flag=1;
	run("paraloadstore", time_para_loadstore, para_loadstore);
}