implemented realloc as well as a bunch of fixes and some small changes like changing the order of arguments for !, making alloc drop etc.

main.py

@@ -2290,18 +2290,18 @@ class Assembler:
             "%define RSTK_BYTES 65536",
             "%define PRINT_BUF_BYTES 128",
             "global _start",
-            "global argc",
+            "global sys_argc",
-            "global argv",
+            "global sys_argv",
             "section .data",
-            "argc: dq 0",
+            "sys_argc: dq 0",
-            "argv: dq 0",
+            "sys_argv: dq 0",
             "section .text",
             "_start:",
             "    ; Linux x86-64 startup: argc/argv from stack",
             "    mov rdi, [rsp]",         # argc
             "    lea rsi, [rsp+8]",      # argv
-            "    mov [rel argc], rdi",
+            "    mov [rel sys_argc], rdi",
-            "    mov [rel argv], rsi",
+            "    mov [rel sys_argv], rsi",
             "    ; initialize data/return stack pointers",
             "    lea r12, [rel dstack_top]",
             "    mov r15, r12",
@@ -2480,6 +2480,7 @@ def macro_with(ctx: MacroContext) -> Optional[List[Op]]:
     for name in reversed(names):
         helper = helper_for[name]
         emitted.append(helper)
+        emitted.append("swap")
         emitted.append("!")
 
     i = 0
@@ -2491,6 +2492,7 @@ def macro_with(ctx: MacroContext) -> Optional[List[Op]]:
             next_tok = body[i + 1] if i + 1 < len(body) else None
             if next_tok is not None and next_tok.lexeme == "!":
                 emitted.append(helper)
+                emitted.append("swap")
                 emitted.append("!")
                 i += 2
                 continue
@@ -3210,7 +3212,7 @@ def macro_struct_begin(ctx: MacroContext) -> Optional[List[Op]]:
             generated,
             name_token,
             f"{struct_name}.{field.name}!",
-            [offset_word, "+", "!"],
+            ["swap", offset_word, "+", "swap", "!"],
         )
 
     parser.tokens[parser.pos:parser.pos] = generated

nob.sl

@@ -31,25 +31,20 @@ word sh
         0
         c!
 
+        mem
         "/bin/sh" drop
-        0
-        mem +
         !
+        mem 8 +
         "-c" drop
-        8
-        mem +
         !
+        mem 16 +
         0 rpick
-        16
-        mem +
         !
+        mem 24 +
         0
-        24
-        mem +
         !
+        mem 32 +
         0
-        32
-        mem +
         !
 
         syscall.fork

stdlib/core.sl

@@ -11,8 +11,8 @@
 
 # : argc ( -- n )
 :asm argc {
-	extern argc
+	extern sys_argc
-	mov rax, [rel argc]
+	mov rax, [rel sys_argc]
 	sub r12, 8
 	mov [r12], rax
 	ret
@@ -21,8 +21,8 @@
 
 # : argv ( -- ptr )
 :asm argv {
-	extern argv
+	extern sys_argv
-	mov rax, [rel argv]
+	mov rax, [rel sys_argv]
 	sub r12, 8
 	mov [r12], rax
 	ret
@@ -31,9 +31,9 @@
 
 # : argv@ ( n -- ptr )
 :asm argv@ {
-	extern argv
+	extern sys_argv
 	mov rbx, [r12]      ; n
-	mov rax, [rel argv]
+	mov rax, [rel sys_argv]
 	mov rax, [rax + rbx*8]
 	mov [r12], rax
 	ret
@@ -54,6 +54,7 @@
 	mov rax, [r12]         ; get address from stack
 	add r12, 8             ; pop address
 	mov rbx, [r12]         ; get byte value
+	add r12, 8
 	mov [rbx], al          ; store byte at address
 	ret
 }
@@ -310,11 +311,11 @@
 
 # : ! ( x addr -- )
 :asm ! {
-	mov rax, [r12]         ; get address
+	mov rax, [r12]         ; get value (TOS)
-	add r12, 8             ; pop address
-	mov rbx, [r12]         ; get value
-	mov [rax], rbx         ; store value at address
 	add r12, 8             ; pop value
+	mov rbx, [r12]         ; get addr (NOS)
+	add r12, 8			   ; pop addr
+	mov [rbx], rax         ; store value at address
 }
 ;
 

stdlib/mem.sl

@@ -8,6 +8,7 @@ word alloc
     -1     # fd
     0      # offset
     mmap
+    swap drop
 end
 
 word free
@@ -21,7 +22,6 @@ word memcpy #(dst_addr src_addr len -- dst_addr len)
     dup c@
     3 pick swap
     c!
-    drop
     swap
     for
         1 + dup
@@ -33,10 +33,30 @@ word memcpy #(dst_addr src_addr len -- dst_addr len)
         dup
         rot
         c!
-        drop
         swap
     end
     swap
     nip
     r> dup -rot - swap
+end
+
+word memset #( value len addr -- )
+    swap
+    0 swap for
+        -rot swap 2 pick + 2dup swap ! 1 + -rot swap
+    end
+    2drop drop
+end
+
+word memdump #( len addr -- addr )
+    for
+        dup c@ puti cr 1 +
+    end
+end
+
+word realloc #( addr old_len new_len -- new_addr )
+    2 pick swap alloc
+    rot rot swap
+    memcpy
+    swap -rot free
 end

stdlib/utils.sl

@@ -99,7 +99,7 @@ word tostr
     dup
     count_digits
     2dup >r alloc
-    nip swap rot swap
+    swap rot swap
     for
         dup 10 % swap 10 /
     end
@@ -114,7 +114,6 @@ word tostr
         2 + pick
         3 pick rot +
         swap 48 + swap 1 - swap c!
-        drop
         swap
         1 +
         swap

tests/alloc.sl

@@ -3,7 +3,7 @@ import ../stdlib/io.sl
 import ../stdlib/mem.sl
 
 word test-mem-alloc
-    4096 alloc dup 1337 swap !   # allocate 4096 bytes, store 1337 at start
+    4096 alloc dup 1337 !   # allocate 4096 bytes, store 1337 at start
     dup @ puti cr                # print value at start
     4096 free                    # free the memory
 end
@@ -16,9 +16,9 @@ end
 word main
     32 alloc           # allocate 32 bytes (enough for a Point struct)
     dup
-    dup 111 swap Point.x!
+    dup 111 Point.x!     # store 111 at offset 0 (Point.x)
-    dup 222 swap Point.y!
+    dup 222 Point.y! # store 222 at offset 8 (Point.y)
-    dup Point.x@ puti cr
+    dup Point.x@ puti cr      # print x
-    Point.y@ puti cr
+    dup Point.y@ puti cr  # print y
     32 free            # free the memory
 end

tests/integration_core.sl

@@ -66,7 +66,7 @@ end
 
 word test-store
     mem-slot dup
-    123 swap !
+    123 !
     @ puti cr
 end
 
@@ -80,7 +80,7 @@ word test-mmap
     0      # offset
     mmap
     dup
-    1337 swap !
+    1337 !
     dup
     @ puti cr
     4096 munmap drop
@@ -132,10 +132,10 @@ end
 
 word test-struct
     mem-slot
-    dup 111 swap Point.x!
+    dup 111 !       # Point.x = 111
-    dup 222 swap Point.y!
+    dup 8 + 222 !   # Point.y = 222
-    dup Point.x@ puti cr
+    dup @ puti cr        # print Point.x
-    Point.y@ puti cr
+    dup 8 + @ puti cr    # print Point.y
     Point.size puti cr
 end
 

tests/mem.expected

@@ -1,2 +1,9 @@
 5
-6
+6
+aaaaaaaaaa
+111
+222
+111
+222
+333
+444

tests/mem.sl

@@ -1,9 +1,39 @@
 import ../stdlib/stdlib.sl
-import ../stdlib/io.sl
 
 word main
-    mem 5 swap !
+    mem 5 !
-    mem 8 + 6 swap !
+    mem 8 + 6 !
     mem @ puti cr
     mem 8 + @ puti cr
+
+    10 alloc
+    dup
+    10 97 memset
+    10 2dup puts
+    free
+
+
+    # Test realloc: allocate 16 bytes, fill with data, then grow to 32 bytes
+    16 alloc
+    dup 111 !       # Store 111 at offset 0
+    dup 8 + 222 !   # Store 222 at offset 8
+
+    # Realloc to 32 bytes
+    16 32 realloc  # ( addr old_len=16 new_len=32 ) -> ( new_addr )
+
+    # Verify old data is preserved
+    dup @ puti cr         # Should print 111
+    dup 8 + @ puti cr     # Should print 222
+
+    # Write new data to the expanded region
+    dup 16 + 333 !        # Store 333 at offset 16
+    dup 24 + 444 !        # Store 444 at offset 24
+
+    # Print all values to verify
+    dup @ puti cr         # 111
+    dup 8 + @ puti cr     # 222
+    dup 16 + @ puti cr    # 333
+    dup 24 + @ puti cr    # 444
+
+    32 free
 end