add short load uop tests

Makefile

@@ -1,8 +1,11 @@
 CFLAGS=-lm -g -Wall -O1
 
-build/main: src/main.c src/repetition_tester.c build/multi_nop_loop.o build/write_loops.o build/load_uop.o build/store_uop.o src/rprof.h
+build/main: src/main.c src/repetition_tester.c build/multi_nop_loop.o build/write_loops.o build/load_uop.o build/store_uop.o build/short_load_uop.o src/rprof.h
 	mkdir -p build
-	gcc -o build/main src/main.c build/multi_nop_loop.o build/write_loops.o build/load_uop.o build/store_uop.o $(CFLAGS)
+	gcc -o build/main src/main.c build/multi_nop_loop.o build/write_loops.o build/load_uop.o build/store_uop.o build/short_load_uop.o $(CFLAGS)
+
+build/short_load_uop.o: src/short_load_uop.asm
+	nasm -g -f elf64 -o build/short_load_uop.o src/short_load_uop.asm
 
 build/load_uop.o: src/load_uop.asm
 	nasm -g -f elf64 -o build/load_uop.o src/load_uop.asm

src/main.c

@@ -16,6 +16,7 @@
 #include "main_write_loop.c"
 #include "main_load_uop.c"
 #include "main_store_uop.c"
+#include "main_short_load_uop.c"
 
 int main(int argc, char **argv) {
 	if (argc < 2) {
@@ -33,6 +34,8 @@ int main(int argc, char **argv) {
 		return main_test_load_uop();
 	} else if (!strncmp(test_name, "store_uop", sizeof("store_uop"))) {
 		return main_test_store_uop();
+	} else if (!strncmp(test_name, "short_load_uop", sizeof("short_load_uop"))) {
+		return main_test_short_load_uop();
 	} 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"))) {

src/main_short_load_uop.c

@@ -0,0 +1,41 @@
+#include "repetition_tester.c"
+#include "short_load_uop.h"
+
+int main_test_short_load_uop() {
+	typedef void (*test_cb)(uint8_t *buffer, uint64_t byte_count);
+	struct testcase {
+		char *name;
+		test_cb cb;
+	};
+
+	struct testcase cases[] = {
+		{ .name = "mov_load_1x2()", .cb = mov_load_1x2 },
+		{ .name = "mov_load_2x2()", .cb = mov_load_2x2 },
+		{ .name = "mov_load_4x2()", .cb = mov_load_4x2 },
+		{ .name = "mov_load_8x2()", .cb = mov_load_8x2 },
+	};
+
+	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 * 1024;
+	uint8_t buffer[byte_count];
+
+	for (int i = 0; i < ARRAY_LEN(cases); i++) {
+		struct testcase *testcase = &cases[i];
+		repetitor_clear(&repetitor);
+		while (repetitor_repeat(&repetitor, 2)) {
+			repetitor_start(&repetitor);
+			repetitor_measure_start(&repetitor);
+			testcase->cb(buffer, byte_count);
+			repetitor_measure_stop(&repetitor, byte_count);
+			repetitor_stop(&repetitor);
+		}
+		repetitor_print_results_label(&repetitor, testcase->name);
+	}
+
+	return 0;
+}
+
+

src/short_load_uop.asm

@@ -0,0 +1,54 @@
+global mov_load_1x2
+global mov_load_2x2
+global mov_load_4x2
+global mov_load_8x2
+
+section .text
+
+; rsi - byte_count
+; rdi - buffer
+mov_load_1x2:
+	mov rcx, rsi
+	align 64
+.loop:
+	mov al, [rdi]
+	mov al, [rdi]
+	sub rcx, 2
+	jnle .loop
+	ret
+
+; rsi - byte_count
+; rdi - buffer
+mov_load_2x2:
+	mov rcx, rsi
+	align 64
+.loop:
+	mov ax, [rdi]
+	mov ax, [rdi]
+	sub rcx, 2
+	jnle .loop
+	ret
+
+; rsi - byte_count
+; rdi - buffer
+mov_load_4x2:
+	mov rcx, rsi
+	align 64
+.loop:
+	mov eax, [rdi]
+	mov eax, [rdi]
+	sub rcx, 2
+	jnle .loop
+	ret
+
+; rsi - byte_count
+; rdi - buffer
+mov_load_8x2:
+	mov rcx, rsi
+	align 64
+.loop:
+	mov rax, [rdi]
+	mov rax, [rdi]
+	sub rcx, 2
+	jnle .loop
+	ret

src/short_load_uop.h

@@ -0,0 +1,6 @@
+#include <stdint.h>
+
+void mov_load_1x2(uint8_t *buffer, uint64_t byte_count);
+void mov_load_2x2(uint8_t *buffer, uint64_t byte_count);
+void mov_load_4x2(uint8_t *buffer, uint64_t byte_count);
+void mov_load_8x2(uint8_t *buffer, uint64_t byte_count);