From b66bb2659eef7765f59ade741c56be440b502c6a Mon Sep 17 00:00:00 2001
From: iceman1001 <iceman@iuse.se>
Date: Tue, 15 Sep 2026 18:49:36 +0200
Subject: [PATCH] armsrc: copy words in memcpy, not just bytes

The device ships its own string.c because there is no libc linked, and
memcpy was a byte at a time loop: six instructions per byte, 17 cycles
per byte measured, 221us for a 624 byte frame. Every copy on the device
paid that.

Take words when source and destination share their offset within a word,
which is the only case ARM7TDMI can do at all, and unroll the byte tail
four ways since at -Os the loop bookkeeping otherwise costs more than the
copy. 624 bytes aligned goes 221us -> 34us, misaligned 221 -> 125.

192 bytes, was 24. The bootrom does not link string.c.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
---
 armsrc/string.c | 60 +++++++++++++++++++++++++++++++++++++++++++++----
 1 file changed, 56 insertions(+), 4 deletions(-)

diff --git a/armsrc/string.c b/armsrc/string.c
index b7e549ff1..d3c040e3e 100644
--- a/armsrc/string.c
+++ b/armsrc/string.c
@@ -17,13 +17,65 @@
 //-----------------------------------------------------------------------------
 #include "string.h"
 
+// The word view of the buffers has to be declared as aliasing, the tree builds
+// with -fstrict-aliasing.
+typedef uint32_t __attribute__((may_alias)) aliasing_u32;
+
 void *memcpy(void *dest, const void *src, int len) {
     uint8_t *d = dest;
     const uint8_t *s = src;
-    while ((len--) > 0) {
-        *d = *s;
-        d++;
-        s++;
+
+    if (len <= 0) {
+        return dest;
+    }
+
+    // ARM7TDMI cannot do misaligned word access, so words are only usable when
+    // both sides sit at the same offset within a word
+    if ((((uintptr_t)d ^ (uintptr_t)s) & 3) == 0) {
+
+        while ((((uintptr_t)d & 3) != 0) && (len > 0)) {
+            *d++ = *s++;
+            len--;
+        }
+
+        aliasing_u32 *dw = (aliasing_u32 *)d;
+        const aliasing_u32 *sw = (const aliasing_u32 *)s;
+
+        while (len >= 16) {
+            dw[0] = sw[0];
+            dw[1] = sw[1];
+            dw[2] = sw[2];
+            dw[3] = sw[3];
+            dw += 4;
+            sw += 4;
+            len -= 16;
+        }
+
+        while (len >= 4) {
+            *dw++ = *sw++;
+            len -= 4;
+        }
+
+        d = (uint8_t *)dw;
+        s = (const uint8_t *)sw;
+    }
+
+    // Byte tail, and the whole copy when the two sides are misaligned against
+    // each other. Unrolled because at -Os the compiler keeps a counter and
+    // indexes off it, so loop overhead otherwise costs more than the copy.
+    while (len >= 4) {
+        d[0] = s[0];
+        d[1] = s[1];
+        d[2] = s[2];
+        d[3] = s[3];
+        d += 4;
+        s += 4;
+        len -= 4;
+    }
+
+    while (len > 0) {
+        *d++ = *s++;
+        len--;
     }
     return dest;
 }