3 files changed, 27 insertions, 21 deletions
diff --git a/CHANGELOG.txt b/CHANGELOG.txt
index d8b90a7..e4c5386 100644
--- a/CHANGELOG.txt
+++ b/CHANGELOG.txt
@@ -64,6 +64,10 @@
 * Add more run-time tests;
 * Rename likly and ulikly to lik and ulik;
 
+* Fix memcpy results on ARM;
+* Fix order of parameters in memcpy on ARM;
+* Add more assembly comments;
+
 # 0.0.2
 
 * Migrate to CMake;
diff --git a/zp/source/arm/mem/memcpy.s b/zp/source/arm/mem/memcpy.s
index b07c1f6..3a9d0b6 100644
--- a/zp/source/arm/mem/memcpy.s
+++ b/zp/source/arm/mem/memcpy.s
@@ -12,39 +12,41 @@
 .thumb_func
 
 zp_memcpy:
- @ zp_i02 tmp4;
- @ zp_i8 tmp1;
+ @ zp_i02 tmp4;
+ @ zp_i8 tmp1;
 
-.wrdcpy: @ wrdcpy:;
+.wrdcpy: @ wrdcpy:;
 @ Check if there are at least four bytes remaining:
- cmp r1,0x4
- blt .bytcpy @ if (num < 0x4u) goto bytcpy;
+ cmp r3,0x4
+ blt .bytcpy @ if (num < 0x4u) goto bytcpy;
 
 @ Copy one word:
- ldm r0!,{r3} @ tmp4 = *(zp_i02 *)in; /* We use ldm/stm with an exclamation mark after the source/destination as this version saves the incremented address into the register, meaning we don't have to icrement it ourselves. */
- stm r2!,{r3} @ *(zp_i02 *)out = tmp4;
+ ldm r2!,{r4} @ tmp4 = *(zp_i02 *)src; // We use ldm/stm with an exclamation mark after the source/destination as this version saves the incremented address into the register, meaning we don't have to icrement it ourselves.
+ stm r1!,{r4} @ *(zp_i02 *)dst = tmp4;
 
 @ Continue to the next word:
- subs r1,0x4 @ num -= 0x4u;
- b .wrdcpy @ goto wrdcpy;
+ subs r3,0x4 @ num -= 0x4u;
+ b .wrdcpy @ goto wrdcpy;
 
-.bytcpy: @ bytcpy:;
+.bytcpy: @ bytcpy:;
 @ Check if we have any bytes remaining:
- cmp r1,0x0
- beq .done @ if (num == 0x0u) goto done;
+ cmp r3,0x0
+ beq .done @ if (num == 0x0u) goto done;
 
 @ Copy one byte:
- ldrb r3,[r0] @ tmp1 = *(zp_i8 *)in;
- strb r3,[r2] @ *(zp_i8 *)out = tmp1;
+ ldrb r4,[r2] @ tmp1 = *(zp_i8 *)src;
+ strb r4,[r1] @ *(zp_i8 *)dst = tmp1;
 
 @ Continue to the next byte:
- adds r0,0x1 @ ++in;
- adds r2,0x1 @ ++out;
- subs r1,0x1 @ --num;
- b .bytcpy @ goto bytcpy;
+ adds r1,0x1 @ ++dst;
+ adds r2,0x1 @ ++src;
+ subs r3,0x1 @ --num;
+ b .bytcpy @ goto bytcpy;
 
-.done: @ done:;
+.done: @ done:;
 @ Return:
- bx lr @ return;
+ str r2,[r0] @ We place the structure members in the buffer given by the caller, the pointer to which is always passed in r0.
+ str r3,[r0,0x4]
+ bx lr @ return (zp_cpyres) {.dst = dst,.src = src};
 
 .endfunc
diff --git a/zp/source/arm64/mem/memcpy.s b/zp/source/arm64/mem/memcpy.s
index 6d5631e..91fbcf7 100644
--- a/zp/source/arm64/mem/memcpy.s
+++ b/zp/source/arm64/mem/memcpy.s
@@ -35,6 +35,6 @@ zp_cp:
 b .bytcpy // goto bytcpy;
 
 .done: // done:;
- ret // return;
+ ret // return; // The structure members are already in the appropriate registers.
 
 .endfunc