add assembly loops tests

Makefile

@@ -1,11 +1,14 @@
 CFLAGS=-lm -g -Wall -O1
 
-build/main: src/main.c src/repetition_tester.c build/nop_loop.o src/rprof.h
+build/main: src/main.c src/repetition_tester.c build/multi_nop_loop.o build/write_loops.o src/rprof.h
 	mkdir -p build
-	gcc -o build/main src/main.c build/nop_loop.o $(CFLAGS)
+	gcc -o build/main src/main.c build/multi_nop_loop.o build/write_loops.o $(CFLAGS)
 
-build/nop_loop.o: src/nop_loop.asm
-	nasm -g -f elf64 -o build/nop_loop.o src/nop_loop.asm
+build/write_loops.o: src/write_loops.asm
+	nasm -g -f elf64 -o build/write_loops.o src/write_loops.asm
+
+build/multi_nop_loop.o: src/multi_nop_loop.asm
+	nasm -g -f elf64 -o build/multi_nop_loop.o src/multi_nop_loop.asm
 
 run: ./build/main
 	./build/main $(TEST_NAME)

src/main.c

@@ -13,6 +13,7 @@
 #include "main_write_all_bytes.c"
 #include "main_malloc_read.c"
 #include "main_write_backward.c"
+#include "main_write_loop.c"
 
 int main(int argc, char **argv) {
 	if (argc < 2) {
@@ -26,6 +27,8 @@ int main(int argc, char **argv) {
 		return main_test_write_bytes_asm();
 	} else if (!strncmp(test_name, "write_all_bytes", sizeof("write_bytes"))) {
 		return main_test_write_all_bytes();
+	} else if (!strncmp(test_name, "write_loop", sizeof("write_loop"))) {
+		return main_test_write_loop();
 	} else if (!strncmp(test_name, "write_backward", sizeof("write_backward"))) {
 		return main_test_write_backward();
 	} else if (!strncmp(test_name, "read_file", sizeof("read_file"))) {

src/main_write_bytes_asm.c

@@ -1,5 +1,5 @@
 #include "repetition_tester.c"
-#include "nop_loop.h"
+#include "multi_nop_loop.h"
 
 int main_test_write_bytes_asm() {
 	typedef void (*write_bytes_cb)(uint8_t *buffer, uint64_t byte_count);

src/main_write_loop.c

@@ -0,0 +1,58 @@
+#include "repetition_tester.c"
+#include "write_loops.h"
+
+int main_test_write_loop() {
+	struct repetitor repetitor = {};
+	repetitor_init(&repetitor);
+	printf("CPU Frequency: %ldHz (~%.2fGHz)\n", repetitor.cpu_freq, (float)repetitor.cpu_freq/(1000*1000*1000));
+
+	uint64_t byte_count = 4096 * 512;
+	uint8_t *buffer = malloc(sizeof(uint8_t) * byte_count);
+
+	while (repeat_cycle_times(&repetitor, 3)) {
+		repetitor_clear(&repetitor);
+		while (repetitor_repeat(&repetitor, 2)) {
+			repetitor_start(&repetitor);
+			repetitor_measure_start(&repetitor);
+			mov_all_bytes_asm(buffer, byte_count);
+			repetitor_measure_stop(&repetitor, byte_count);
+			repetitor_stop(&repetitor);
+		}
+		repetitor_print_results_label(&repetitor, "mov_all_bytes");
+
+		repetitor_clear(&repetitor);
+		while (repetitor_repeat(&repetitor, 2)) {
+			repetitor_start(&repetitor);
+			repetitor_measure_start(&repetitor);
+			nop_all_bytes_asm(byte_count);
+			repetitor_measure_stop(&repetitor, byte_count);
+			repetitor_stop(&repetitor);
+		}
+		repetitor_print_results_label(&repetitor, "nop_all_bytes");
+
+		repetitor_clear(&repetitor);
+		while (repetitor_repeat(&repetitor, 2)) {
+			repetitor_start(&repetitor);
+			repetitor_measure_start(&repetitor);
+			cmp_all_bytes_asm(byte_count);
+			repetitor_measure_stop(&repetitor, byte_count);
+			repetitor_stop(&repetitor);
+		}
+		repetitor_print_results_label(&repetitor, "cmp_all_bytes");
+
+		repetitor_clear(&repetitor);
+		while (repetitor_repeat(&repetitor, 2)) {
+			repetitor_start(&repetitor);
+			repetitor_measure_start(&repetitor);
+			dec_all_bytes_asm(byte_count);
+			repetitor_measure_stop(&repetitor, byte_count);
+			repetitor_stop(&repetitor);
+		}
+		repetitor_print_results_label(&repetitor, "dec_all_bytes");
+
+	}
+
+	return 0;
+
+	return 0;
+}

src/write_loops.asm

@@ -0,0 +1,43 @@
+global mov_all_bytes_asm
+global nop_all_bytes_asm
+global cmp_all_bytes_asm
+global dec_all_bytes_asm
+
+section .text
+
+; rsi - byte_count
+; rdi - buffer
+mov_all_bytes_asm:
+	xor rax, rax
+.loop:
+	mov [rdi + rax], al
+	inc rax
+	cmp rax, rsi
+	jb .loop
+	ret
+
+; rdi - byte_count
+nop_all_bytes_asm:
+	xor rax, rax
+.loop:
+	db 0x0f, 0x1f, 0x00
+	inc rax
+	cmp rax, rdi
+	jb .loop
+	ret
+
+; rdi - byte_count
+cmp_all_bytes_asm:
+	xor rax, rax
+.loop:
+	inc rax
+	cmp rax, rdi
+	jb .loop
+	ret
+
+; rdi - byte_count
+dec_all_bytes_asm:
+.loop:
+	dec rdi
+	jnz .loop
+	ret

src/write_loops.h

@@ -0,0 +1,6 @@
+#include <stdint.h>
+
+void mov_all_bytes_asm(uint8_t *buffer, uint64_t byte_count);
+void nop_all_bytes_asm(uint64_t byte_count);
+void cmp_all_bytes_asm(uint64_t byte_count);
+void dec_all_bytes_asm(uint64_t byte_count);