Add NEON

src/deluge/dsp/memmove.c

@@ -1,10 +1,11 @@
 #include "mem_functions.h"
+#include <arm_neon.h>
 #include <stdbool.h>
 #include <stddef.h>
 #include <stdint.h>
 #include <stdlib.h>
 
-void* memmove(void* dst, const void* src, size_t len) {
+void* my_memmove(void* dst, const void* src, size_t len) {
 	ptrdiff_t result;
 	asm("sub %0, %1, %2" : "=r"(result) : "r"(dst), "r"(src));
 	if (abs(result) >= (ptrdiff_t)len) {
@@ -30,9 +31,51 @@ void* memmove(void* dst, const void* src, size_t len) {
 	// copy backwards by word...
 	const uint32_t* s_32 = (const uint32_t*)s_16;
 	uint32_t* d_32 = (uint32_t*)d_16;
-	while ((intptr_t)d_32 > (intptr_t)dst) {
+	if (len % 8) {
 		*--d_32 = *--s_32;
 	}
 
+	// doubleword
+	s = (const char*)s_32;
+	d = (char*)d_32;
+	if (len % 16) {
+		s -= 8;
+		d -= 8;
+		vst1_u8(d, vld1_u8(s));
+	}
+
+	// quadword
+	if (len % 32) {
+		s -= 16;
+		d -= 16;
+		vst1q_u8(d, vld1q_u8(s));
+	}
+
+	// quadword x2
+	if (len % 64) {
+		s -= 32;
+		d -= 32;
+		vst1q_u8_x2(d, vld1q_u8_x2(s));
+	}
+
+	// quadword x4
+	if (len % 128) {
+		s -= 64;
+		d -= 64;
+		vst1q_u8_x4(d, vld1q_u8_x4(s));
+	}
+
+	// quadword x8
+	while ((intptr_t)d > (intptr_t)dst) {
+		s -= 64;
+		uint8x16x4_t ld1 = vld1q_u8_x4(s);
+		s -= 64;
+		uint8x16x4_t ld2 = vld1q_u8_x4(s);
+		d -= 64;
+		vst1q_u8_x4(d, ld1);
+		d -= 64;
+		vst1q_u8_x4(d, ld2);
+	}
+
 	return dst;
 }