Add assembly loops

haversine_02/compile

@@ -4,6 +4,12 @@ CC=clang
 CXX=clang++
 CFLAGS="-Wall -Wextra -I$(realpath ./include) "
 
+ASM=nasm
+ASM_FLAGS="-f elf64 "
+
+AR=ar
+AR_FLAGS="rcs"
+
 # PARSE ARGUMENTS
 # From this StackOverflow answer https://stackoverflow.com/a/14203146
 while [[ $# > 0 ]];do
@@ -71,6 +77,18 @@ MEMTESTOUT=memtest
 (set -x ; $CC $CFLAGS $MEMTESTSRC -o $MEMTESTOUT)
 echo
 
+# REPTEST ASSEMBLY
+ASM_BUILD_DIR=reptest_build
+ASM_SRC="./src/repetition_testing/reptest_functions.asm"
+ASM_OBJ="./$ASM_BUILD_DIR/funcs.o"
+ASM_LIB="./$ASM_BUILD_DIR/libfuncs.a"
+
+mkdir $ASM_BUILD_DIR
+
+(set -x ; $ASM $ASM_FLAGS $ASM_SRC -o $ASM_OBJ)
+(set -x ; $AR $AR_FLAGS $ASM_LIB $ASM_OBJ)
+echo
+
 if [[ $BASIC_PROFILING == true ]] || [[ $FULL_PROFILING == true ]]; then
 	if [[ $FULL_PROFILING == true ]]; then
 		JSONFLAGS+="-DFULL_PROFILING"
@@ -95,7 +113,7 @@ if [[ $BASIC_PROFILING == true ]] || [[ $FULL_PROFILING == true ]]; then
 	cd ../
 
 	# REPETITION TESTING
-	REPTESTSRC="./src/repetition_testing/*.cpp ./$PROF_BUILD_DIR/*.o"
+	REPTESTSRC="./src/repetition_testing/*.cpp ./$PROF_BUILD_DIR/*.o $ASM_LIB"
 	REPTESTOUT=reptest
 
 	(set -x ; $CXX $CFLAGS $REPTESTFLAGS $REPTESTSRC -o $REPTESTOUT)
@@ -114,4 +132,4 @@ cd ../
 echo
 
 # CLEAR BUILD FILES
-rm -rvf $JSON_BUILD_DIR $PROF_BUILD_DIR
+rm -rvf $JSON_BUILD_DIR $PROF_BUILD_DIR $ASM_BUILD_DIR

haversine_02/src/repetition_testing/main.cpp

@@ -9,9 +9,18 @@
 
 #define ARR_LEN(ARR) sizeof(ARR) / sizeof(*ARR)
 
+extern "C" void mov_all_bytes_asm(char *buffer, u64 size);
+extern "C" void nop_all_bytes_asm(u64 size);
+extern "C" void inc_all_bytes_asm(u64 size);
+extern "C" void dec_all_bytes_asm(u64 size);
+
 void test_fread(reptester *tester, alloc_type type);
 void test_read(reptester *tester, alloc_type type);
 void test_write(reptester *tester, alloc_type type);
+void test_write_mov_all_bytes_asm(reptester *tester, alloc_type type);
+void test_write_nop_all_bytes_asm(reptester *tester, alloc_type type);
+void test_write_inc_all_bytes_asm(reptester *tester, alloc_type type);
+void test_write_dec_all_bytes_asm(reptester *tester, alloc_type type);
 u64 get_file_length(FILE *fp);
 
 int main(int argc, char *argv[]) {
@@ -65,8 +74,16 @@ int main(int argc, char *argv[]) {
 
   func_data funcs[] = {
       {{"WRITE", "WRITE WITH MALLOC"}, test_write},
-      {{"READ", "READ WITH MALLOC"}, test_read},
+      {{"WRITE MOV ASM", "WRITE MOV ASM WITH MALLOC"},
-      {{"FREAD", "FREAD WITH MALLOC"}, test_fread},
+       test_write_mov_all_bytes_asm},
+      {{"WRITE NOP ASM", "WRITE NOP ASM WITH MALLOC"},
+       test_write_nop_all_bytes_asm},
+      {{"WRITE INC ASM", "WRITE INC ASM WITH MALLOC"},
+       test_write_inc_all_bytes_asm},
+      {{"WRITE DEC ASM", "WRITE DEC ASM WITH MALLOC"},
+       test_write_dec_all_bytes_asm},
+      // {{"READ", "READ WITH MALLOC"}, test_read},
+      // {{"FREAD", "FREAD WITH MALLOC"}, test_fread},
   };
 
   tester.params.read_size = get_file_length(fp);
@@ -155,6 +172,12 @@ void test_read(reptester *tester, alloc_type type) {
   fclose(fp);
 }
 
+void write_to_all_bytes(char *buffer, u64 size) {
+  for (u64 i = 0; i < size; ++i) {
+    buffer[i] = i;
+  }
+}
+
 void test_write(reptester *tester, alloc_type type) {
   u64 start = read_cpu_timer();
   u64 fault_count_start = page_fault_count();
@@ -163,10 +186,108 @@ void test_write(reptester *tester, alloc_type type) {
 
   u64 total_size = tester->params.read_size * tester->params.read_count;
 
-  for (u64 i = 0; i < total_size; ++i) {
+  write_to_all_bytes(tester->params.buffer, total_size);
-    tester->params.buffer[i] = '0';
+
+  u64 fault_count_end = page_fault_count();
+  u64 end = read_cpu_timer();
+
+  u64 read_time = end - start;
+  u64 page_faults = fault_count_end - fault_count_start;
+
+  tester->results = {
+      total_size,
+      read_time,
+      page_faults,
+  };
+
+  handle_free(tester, type);
 }
 
+void test_write_mov_all_bytes_asm(reptester *tester, alloc_type type) {
+  u64 start = read_cpu_timer();
+  u64 fault_count_start = page_fault_count();
+
+  handle_alloc(tester, type);
+
+  u64 total_size = tester->params.read_size * tester->params.read_count;
+
+  mov_all_bytes_asm(tester->params.buffer, total_size);
+
+  u64 fault_count_end = page_fault_count();
+  u64 end = read_cpu_timer();
+
+  u64 read_time = end - start;
+  u64 page_faults = fault_count_end - fault_count_start;
+
+  tester->results = {
+      total_size,
+      read_time,
+      page_faults,
+  };
+
+  handle_free(tester, type);
+}
+
+void test_write_nop_all_bytes_asm(reptester *tester, alloc_type type) {
+  u64 start = read_cpu_timer();
+  u64 fault_count_start = page_fault_count();
+
+  handle_alloc(tester, type);
+
+  u64 total_size = tester->params.read_size * tester->params.read_count;
+
+  nop_all_bytes_asm(total_size);
+
+  u64 fault_count_end = page_fault_count();
+  u64 end = read_cpu_timer();
+
+  u64 read_time = end - start;
+  u64 page_faults = fault_count_end - fault_count_start;
+
+  tester->results = {
+      total_size,
+      read_time,
+      page_faults,
+  };
+
+  handle_free(tester, type);
+}
+
+void test_write_inc_all_bytes_asm(reptester *tester, alloc_type type) {
+  u64 start = read_cpu_timer();
+  u64 fault_count_start = page_fault_count();
+
+  handle_alloc(tester, type);
+
+  u64 total_size = tester->params.read_size * tester->params.read_count;
+
+  inc_all_bytes_asm(total_size);
+
+  u64 fault_count_end = page_fault_count();
+  u64 end = read_cpu_timer();
+
+  u64 read_time = end - start;
+  u64 page_faults = fault_count_end - fault_count_start;
+
+  tester->results = {
+      total_size,
+      read_time,
+      page_faults,
+  };
+
+  handle_free(tester, type);
+}
+
+void test_write_dec_all_bytes_asm(reptester *tester, alloc_type type) {
+  u64 start = read_cpu_timer();
+  u64 fault_count_start = page_fault_count();
+
+  handle_alloc(tester, type);
+
+  u64 total_size = tester->params.read_size * tester->params.read_count;
+
+  dec_all_bytes_asm(total_size);
+
   u64 fault_count_end = page_fault_count();
   u64 end = read_cpu_timer();
 

haversine_02/src/repetition_testing/reptest_functions.asm

@@ -0,0 +1,43 @@
+global mov_all_bytes_asm
+global nop_all_bytes_asm
+global inc_all_bytes_asm
+global dec_all_bytes_asm
+
+mov_all_bytes_asm:
+	xor rax, rax
+
+	.loop:
+		mov BYTE [rdi + rax * 1], al
+		inc rax
+		cmp rsi, rax
+		jne .loop
+
+	ret
+
+nop_all_bytes_asm:
+	xor rax, rax
+
+	.loop:
+		db 0x0f, 0x1f, 0x00
+		inc rax
+		cmp rdi, rax
+		jne .loop
+
+	ret
+
+inc_all_bytes_asm:
+	xor rax, rax
+
+	.loop:
+		inc rax
+		cmp rdi, rax
+		jne .loop
+
+	ret
+
+dec_all_bytes_asm:
+	.loop:
+		dec rdi
+		jnz .loop
+
+	ret