hare

commit 9c2d4ba4dafa4679127edcb26078a9ddb8695010
parent e00d4a9b82076fde6a6f032753518ee37715bcd4
Author: Autumn! <autumnull@posteo.net>
Date:   Thu,  7 Sep 2023 14:47:23 +0000

rt: copy by words in memmove

Signed-off-by: Autumn! <autumnull@posteo.net>

Diffstat:
M
rt/memmove.ha
 | 
63
++++++++++++++++++++++++++++++++++++++++++++++++++++++---------

1 file changed, 54 insertions(+), 9 deletions(-)
diff --git a/rt/memmove.ha b/rt/memmove.ha
@@ -1,19 +1,64 @@
 // License: MPL-2.0
 // (c) 2021 Drew DeVault <sir@cmpwn.com>
 
+def SZSZ = size(size);
+def MODSZ = size(size) - 1;
+
 export fn memmove(dest: *opaque, src: *const opaque, n: size) void = {
-	let d = dest: *[*]u8, s = src: *const [*]u8;
-	if (d: uintptr == s: uintptr) {
-		return;
-	};
+	// implementation adapted from musl libc
+
+	let d = memfunc_ptr { byte = dest: *[*]u8 };
+	let s = memfunc_ptr { byte = src: *[*]u8 };
 
-	if (d: uintptr < s: uintptr) {
-		for (let i = 0z; i < n; i += 1) {
-			d[i] = s[i];
+	if (d.uptr == s.uptr) return;
+	// equivalent to n <= abs(src - dest)
+	if (s.uptr - d.uptr - n <= -2*n) return memcpy(dest, src, n);
+
+	if (d.uptr < s.uptr) {
+		// if alignment matches, copy the majority as []size
+		if (s.uptr & MODSZ == d.uptr & MODSZ) {
+			// use u8 until we reach a word boundary
+			for (0 < d.uptr & MODSZ) {
+				if (n == 0) return else n -= 1;
+				d.byte[0] = s.byte[0];
+				d.uptr += 1;
+				s.uptr += 1;
+			};
+			for (SZSZ <= n; n -= SZSZ) {
+				d.sz[0] = s.sz[0];
+				// TODO: should be +=, blocked on compiler bug
+				d.uptr = d.uptr + SZSZ;
+				s.uptr = s.uptr + SZSZ;
+			};
+		};
+		for (0 < n; n -= 1) {
+			d.byte[0] = s.byte[0];
+			d.uptr += 1;
+			s.uptr += 1;
 		};
 	} else {
-		for (let i = 0z; i < n; i += 1) {
-			d[n - i - 1] = s[n - i - 1];
+		d.uptr = d.uptr + n;
+		s.uptr = s.uptr + n;
+		// if alignment matches, copy the majority as []size
+		if (s.uptr & MODSZ == d.uptr & MODSZ) {
+			// use u8 until we reach a word boundary
+			for (0 < d.uptr & MODSZ) {
+				if (n == 0) return else n -= 1;
+				d.uptr -= 1;
+				s.uptr -= 1;
+				d.byte[0] = s.byte[0];
+			};
+			for (SZSZ <= n; n -= SZSZ) {
+				// TODO: should be -=, blocked on compiler bug
+				d.uptr = d.uptr - SZSZ;
+				s.uptr = s.uptr - SZSZ;
+				d.sz[0] = s.sz[0];
+			};
+		};
+		for (0 < n; n -= 1) {
+			d.uptr -= 1;
+			s.uptr -= 1;
+			d.byte[0] = s.byte[0];
 		};
 	};
 };