Code alignment testing

haversine_02/src/repetition_testing/main.cpp

@@ -15,6 +15,14 @@ extern "C" void nop_1x3_all_bytes_asm(u64 size);
 extern "C" void nop_1x9_all_bytes_asm(u64 size);
 extern "C" void inc_all_bytes_asm(u64 size);
 extern "C" void dec_all_bytes_asm(u64 size);
+extern "C" void align64_loop(u64 size);
+extern "C" void align1_loop(u64 size);
+extern "C" void align15_loop(u64 size);
+extern "C" void align31_loop(u64 size);
+extern "C" void align63_loop(u64 size);
+extern "C" void align75_loop(u64 size);
+extern "C" void align90_loop(u64 size);
+extern "C" void align112_loop(u64 size);
 
 void test_fread(reptester *tester, alloc_type type);
 void test_read(reptester *tester, alloc_type type);
@@ -25,6 +33,14 @@ void test_write_nop_1x3_all_bytes_asm(reptester *tester, alloc_type type);
 void test_write_nop_1x9_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);
+void test_align64_loop(reptester *tester, alloc_type type);
+void test_align1_loop(reptester *tester, alloc_type type);
+void test_align15_loop(reptester *tester, alloc_type type);
+void test_align31_loop(reptester *tester, alloc_type type);
+void test_align63_loop(reptester *tester, alloc_type type);
+void test_align75_loop(reptester *tester, alloc_type type);
+void test_align90_loop(reptester *tester, alloc_type type);
+void test_align112_loop(reptester *tester, alloc_type type);
 u64 get_file_length(FILE *fp);
 
 int main(int argc, char *argv[]) {
@@ -80,18 +96,26 @@ int main(int argc, char *argv[]) {
       // {{"WRITE", "WRITE WITH MALLOC"}, test_write},
       // {{"WRITE MOV ASM", "WRITE MOV ASM WITH MALLOC"},
       //  test_write_mov_all_bytes_asm},
-      {{"WRITE NOP ASM", "WRITE NOP ASM WITH MALLOC"},
+      // {{"WRITE NOP ASM", "WRITE NOP ASM WITH MALLOC"},
-       test_write_nop_all_bytes_asm},
+      //  test_write_nop_all_bytes_asm},
-      {{"WRITE NOP 1x3 ASM", "WRITE NOP 1x3 ASM WITH MALLOC"},
+      // {{"WRITE NOP 1x3 ASM", "WRITE NOP 1x3 ASM WITH MALLOC"},
-       test_write_nop_1x3_all_bytes_asm},
+      //  test_write_nop_1x3_all_bytes_asm},
-      {{"WRITE NOP 1x9 ASM", "WRITE NOP 1x9 ASM WITH MALLOC"},
+      // {{"WRITE NOP 1x9 ASM", "WRITE NOP 1x9 ASM WITH MALLOC"},
-       test_write_nop_1x9_all_bytes_asm},
+      //  test_write_nop_1x9_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},
+      {{"ALIGN 64", "ALIGN 64 WITH MALLOC"}, test_align64_loop},
+      {{"ALIGN 1", "ALIGN 1 WITH MALLOC"}, test_align1_loop},
+      {{"ALIGN 15", "ALIGN 15 WITH MALLOC"}, test_align15_loop},
+      {{"ALIGN 31", "ALIGN 31 WITH MALLOC"}, test_align31_loop},
+      {{"ALIGN 63", "ALIGN 63 WITH MALLOC"}, test_align63_loop},
+      {{"ALIGN 75", "ALIGN 75 WITH MALLOC"}, test_align75_loop},
+      {{"ALIGN 90", "ALIGN 90 WITH MALLOC"}, test_align90_loop},
+      {{"ALIGN 112", "ALIGN 112 WITH MALLOC"}, test_align112_loop},
   };
 
   tester.params.read_size = get_file_length(fp);
@@ -361,6 +385,208 @@ void test_write_dec_all_bytes_asm(reptester *tester, alloc_type type) {
   handle_free(tester, type);
 }
 
+void test_align64_loop(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;
+
+  align64_loop(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_align1_loop(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;
+
+  align1_loop(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_align15_loop(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;
+
+  align15_loop(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_align31_loop(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;
+
+  align31_loop(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_align63_loop(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;
+
+  align63_loop(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_align75_loop(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;
+
+  align75_loop(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_align90_loop(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;
+
+  align90_loop(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_align112_loop(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;
+
+  align112_loop(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);
+}
+
 u64 get_file_length(FILE *fp) {
   if (!fp) {
     return 0;

haversine_02/src/repetition_testing/reptest_functions.asm

@@ -4,6 +4,14 @@ global nop_1x3_all_bytes_asm
 global nop_1x9_all_bytes_asm
 global inc_all_bytes_asm
 global dec_all_bytes_asm
+global align64_loop
+global align1_loop
+global align15_loop
+global align31_loop
+global align63_loop
+global align75_loop
+global align90_loop
+global align112_loop
 
 mov_all_bytes_asm:
 	xor rax, rax
@@ -75,3 +83,110 @@ dec_all_bytes_asm:
 		jnz .loop
 
 	ret
+
+align64_loop:
+	xor rax, rax
+
+	align 64
+	.loop:
+		inc rax
+		cmp rdi, rax
+		jne .loop
+
+	ret
+
+align1_loop:
+	xor rax, rax
+
+	align 64
+	nop
+	.loop:
+		inc rax
+		cmp rdi, rax
+		jne .loop
+
+	ret
+
+align15_loop:
+	xor rax, rax
+
+	align 64
+	%rep 15
+	nop
+	%endrep
+	.loop:
+		inc rax
+		cmp rdi, rax
+		jne .loop
+
+	ret
+
+align31_loop:
+	xor rax, rax
+
+	align 64
+	%rep 31
+	nop
+	%endrep
+	.loop:
+		inc rax
+		cmp rdi, rax
+		jne .loop
+
+	ret
+
+align63_loop:
+	xor rax, rax
+
+	align 64
+	%rep 63
+	nop
+	%endrep
+	.loop:
+		inc rax
+		cmp rdi, rax
+		jne .loop
+
+	ret
+
+align75_loop:
+	xor rax, rax
+
+	align 64
+	%rep 75
+	nop
+	%endrep
+	.loop:
+		inc rax
+		cmp rdi, rax
+		jne .loop
+
+	ret
+
+align90_loop:
+	xor rax, rax
+
+	align 64
+	%rep 90
+	nop
+	%endrep
+	.loop:
+		inc rax
+		cmp rdi, rax
+		jne .loop
+
+	ret
+
+align112_loop:
+	xor rax, rax
+
+	align 64
+	%rep 112
+	nop
+	%endrep
+	.loop:
+		inc rax
+		cmp rdi, rax
+		jne .loop
+
+	ret