clear_user.cc

//
// Compile with:
//
//   g++ clear_user.cc -std=c++11 -isystem <path to googlebenchmark>/include \
//       -L<path to googlebenchmark>/build/src -lbenchmark -lpthread -o clear_user
//

#include <benchmark/benchmark.h>

#define BUF_SIZE (64*1024)
static char buf[BUF_SIZE];

void __clear_mov_imm(void *addr)
{
        long __d0;
        asm volatile("movq $0,(%[dst])\n"
		: [dst] "=&D" (__d0)
		: "[dst]"(addr));
}

void __clear_mov_reg(void *addr)
{
        long __d0;
        asm volatile("movq %[zero],(%[dst])\n"
		: [dst] "=&D" (__d0)
		: "[dst]"(addr), [zero] "r" (0UL));
}

static void BM_mov_reg(benchmark::State& state) {
	for (auto _ : state)
		__clear_mov_reg(buf);
}
BENCHMARK(BM_mov_reg);

static void BM_mov_imm(benchmark::State& state) {
	for (auto _ : state)
		__clear_mov_imm(buf);
}
BENCHMARK(BM_mov_imm);

static void BM_mov_reg_inline(benchmark::State& state) {
	for (auto _ : state) {
	        long __d0;
        	asm volatile("movq %[zero],(%[dst])\n"
			: [dst] "=&D" (__d0)
			: "[dst]"(buf), [zero] "r" (0UL));
}
}
BENCHMARK(BM_mov_reg_inline);

static void BM_mov_imm_inline(benchmark::State& state) {
	for (auto _ : state) {
	        long __d0;
	        asm volatile("movq $0,(%[dst])\n"
			: [dst] "=&D" (__d0)
			: "[dst]"(buf));
	}
}
BENCHMARK(BM_mov_imm_inline);

BENCHMARK_MAIN();