1 files changed, 94 insertions, 4 deletions
diff --git a/lib/libc/string/memcpy.c b/lib/libc/string/memcpy.c
index f87d95cd..83a29170 100644
--- a/lib/libc/string/memcpy.c
+++ b/lib/libc/string/memcpy.c
@@ -1,5 +1,6 @@
 #include <errno.h>  // for EINVAL, ERANGE
 #include <stddef.h> // for NULL, errno_t
+#include <stdint.h>
 #include <string.h> // for rsize_t, memcpy, size_t, memcpy_s
 #include <sys/cdefs.h>
 
@@ -7,12 +8,101 @@ __weak void *memcpy(void *restrict s1, const void *restrict s2, size_t n);
 
 void *memcpy(void *restrict s1, const void *restrict s2, size_t n)
 {
-	unsigned char *dest = (unsigned char *)s1;
-	const unsigned char *src = (const unsigned char *)s2;
+	unsigned char *d = (unsigned char *)s1;
+	const unsigned char *s = (const unsigned char *)s2;
 
-	while (n--) {
-		*dest++ = *src++;
+	if (n == 0)
+		return s1;
+
+	if (n < 16) {
+		if (n >= 8) {
+			*(uint64_t *)d = *(const uint64_t *)s;
+			d += 8;
+			s += 8;
+			n -= 8;
+		}
+		if (n >= 4) {
+			*(uint32_t *)d = *(const uint32_t *)s;
+			d += 4;
+			s += 4;
+			n -= 4;
+		}
+		if (n >= 2) {
+			*(uint16_t *)d = *(const uint16_t *)s;
+			d += 2;
+			s += 2;
+			n -= 2;
+		}
+		if (n)
+			*d = *s;
+		return s1;
+	}
+
+	uintptr_t align_mask = 15;
+	size_t misalign = (uintptr_t)d & align_mask;
+	if (misalign) {
+		size_t to_align = 16 - misalign;
+		if (to_align > n)
+			to_align = n;
+		for (size_t i = 0; i < to_align; ++i)
+			*d++ = *s++;
+		n -= to_align;
+	}
+
+	while (n >= 128) {
+#pragma unroll
+		for (int i = 0; i < 8; ++i) {
+			((i128 *)d)[i] = ((const i128 *)s)[i];
+		}
+		d += 128;
+		s += 128;
+		n -= 128;
+	}
+	while (n >= 64) {
+#pragma unroll
+		for (int i = 0; i < 4; ++i) {
+			((i128 *)d)[i] = ((const i128 *)s)[i];
+		}
+		d += 64;
+		s += 64;
+		n -= 64;
+	}
+	while (n >= 32) {
+#pragma unroll
+		for (int i = 0; i < 2; ++i) {
+			((i128 *)d)[i] = ((const i128 *)s)[i];
+		}
+		d += 32;
+		s += 32;
+		n -= 32;
+	}
+	while (n >= 16) {
+		*(i128 *)d = *(const i128 *)s;
+		d += 16;
+		s += 16;
+		n -= 16;
+	}
+
+	if (n >= 8) {
+		*(uint64_t *)d = *(const uint64_t *)s;
+		d += 8;
+		s += 8;
+		n -= 8;
+	}
+	if (n >= 4) {
+		*(uint32_t *)d = *(const uint32_t *)s;
+		d += 4;
+		s += 4;
+		n -= 4;
+	}
+	if (n >= 2) {
+		*(uint16_t *)d = *(const uint16_t *)s;
+		d += 2;
+		s += 2;
+		n -= 2;
 	}
+	if (n)
+		*d = *s;
 
 	return s1;
 }