From b5cc68d6e2afe6d249ad31c83bf64382eade7b49 Mon Sep 17 00:00:00 2001
From: Kugel Fuhr <98353208+kugelfuhr@users.noreply.github.com>
Date: Sun, 1 Sep 2024 12:41:25 +0200
Subject: [PATCH] Do not save any register variables when entering main(). Do
 not restore the C stack when leaving main(). Both are unnecessary and just
 bloat the executable.

---
 src/cc65/codegen.c  | 45 +++++++++++++++++++++++++--------------------
 src/cc65/codegen.h  |  2 +-
 src/cc65/function.c |  8 +++++---
 src/cc65/locals.c   | 44 ++++++++++++++++++++++++++++++++++----------
 4 files changed, 65 insertions(+), 34 deletions(-)

diff --git a/src/cc65/codegen.c b/src/cc65/codegen.c
index 69dcc1c6c..45823fab6 100644
--- a/src/cc65/codegen.c
+++ b/src/cc65/codegen.c
@@ -507,34 +507,39 @@ void g_enter (unsigned flags, unsigned argsize)
 
 
 
-void g_leave (void)
+void g_leave (int IsMainFunc)
 /* Function epilogue */
 {
-    /* How many bytes of locals do we have to drop? */
-    unsigned ToDrop = (unsigned) -StackPtr;
+    /* In the main function nothing has to be dropped because the program
+    ** is terminated anyway.
+    */
+    if (!IsMainFunc) {
+        /* How many bytes of locals do we have to drop? */
+        unsigned ToDrop = (unsigned) -StackPtr;
 
-    /* If we didn't have a variable argument list, don't call leave */
-    if (funcargs >= 0) {
+        /* If we didn't have a variable argument list, don't call leave */
+        if (funcargs >= 0) {
 
-        /* Drop stackframe if needed */
-        g_drop (ToDrop + funcargs);
+            /* Drop stackframe if needed */
+            g_drop (ToDrop + funcargs);
 
-    } else if (StackPtr != 0) {
+        } else if (StackPtr != 0) {
+
+            /* We've a stack frame to drop */
+            if (ToDrop > 255) {
+                g_drop (ToDrop);            /* Inlines the code */
+                AddCodeLine ("jsr leave");
+            } else {
+                AddCodeLine ("ldy #$%02X", ToDrop);
+                AddCodeLine ("jsr leavey");
+            }
 
-        /* We've a stack frame to drop */
-        if (ToDrop > 255) {
-            g_drop (ToDrop);            /* Inlines the code */
-            AddCodeLine ("jsr leave");
         } else {
-            AddCodeLine ("ldy #$%02X", ToDrop);
-            AddCodeLine ("jsr leavey");
+
+            /* Nothing to drop */
+            AddCodeLine ("jsr leave");
+
         }
-
-    } else {
-
-        /* Nothing to drop */
-        AddCodeLine ("jsr leave");
-
     }
 
     /* Add the final rts */
diff --git a/src/cc65/codegen.h b/src/cc65/codegen.h
index 8e04b45e4..b95df5cfb 100644
--- a/src/cc65/codegen.h
+++ b/src/cc65/codegen.h
@@ -247,7 +247,7 @@ void g_scale (unsigned flags, long val);
 void g_enter (unsigned flags, unsigned argsize);
 /* Function prologue */
 
-void g_leave (void);
+void g_leave (int IsMainFunc);
 /* Function epilogue */
 
 
diff --git a/src/cc65/function.c b/src/cc65/function.c
index a4b860251..d5cab3993 100644
--- a/src/cc65/function.c
+++ b/src/cc65/function.c
@@ -646,11 +646,13 @@ void NewFunc (SymEntry* Func, FuncDesc* D)
     /* Output the function exit code label */
     g_defcodelabel (F_GetRetLab (CurrentFunc));
 
-    /* Restore the register variables */
-    F_RestoreRegVars (CurrentFunc);
+    /* Restore the register variables (not necessary for main function) */
+    if (!F_IsMainFunc (CurrentFunc)) {
+        F_RestoreRegVars (CurrentFunc);
+    }
 
     /* Generate the exit code */
-    g_leave ();
+    g_leave (F_IsMainFunc (CurrentFunc));
 
     /* Emit references to imports/exports */
     EmitExternals ();
diff --git a/src/cc65/locals.c b/src/cc65/locals.c
index 34d762324..c4d0aa25b 100644
--- a/src/cc65/locals.c
+++ b/src/cc65/locals.c
@@ -111,22 +111,31 @@ static void ParseRegisterDecl (Declarator* Decl, int Reg)
     /* Get the size of the variable */
     unsigned Size = SizeOf (Decl->Type);
 
-    /* Save the current contents of the register variable on stack */
-    F_AllocLocalSpace (CurrentFunc);
-    g_save_regvars (Reg, Size);
-
-    /* Add the symbol to the symbol table. We do that now, because for register
-    ** variables the current stack pointer is implicitly used as location for
-    ** the save area.
-    */
-    Sym = AddLocalSym (Decl->Ident, Decl->Type, Decl->StorageClass, Reg);
-
     /* Check for an optional initialization */
     if (CurTok.Tok == TOK_ASSIGN) {
 
         /* Skip the '=' */
         NextToken ();
 
+        /* If the register variable is initialized, the initialization code may
+        ** access other already declared variables. This means that we have to
+        ** allocate them now.
+        */
+        F_AllocLocalSpace (CurrentFunc);
+
+        /* Save the current contents of the register variable on stack. This is
+        ** not necessary for the main function.
+        */
+        if (!F_IsMainFunc (CurrentFunc)) {
+            g_save_regvars (Reg, Size);
+        }
+
+        /* Add the symbol to the symbol table. We do that now, because for
+        ** register variables the current stack pointer is implicitly used
+        ** as location for the save area (unused in case of main()).
+        */
+        Sym = AddLocalSym (Decl->Ident, Decl->Type, Decl->StorageClass, Reg);
+
         /* Special handling for compound types */
         if (IsCompound) {
 
@@ -173,6 +182,21 @@ static void ParseRegisterDecl (Declarator* Decl, int Reg)
 
         /* Mark the variable as referenced */
         Sym->Flags |= SC_REF;
+    } else {
+
+        /* Save the current contents of the register variable on stack. This is
+        ** not necessary for the main function.
+        */
+        if (!F_IsMainFunc (CurrentFunc)) {
+            F_AllocLocalSpace (CurrentFunc);
+            g_save_regvars (Reg, Size);
+        }
+
+        /* Add the symbol to the symbol table. We do that now, because for
+        ** register variables the current stack pointer is implicitly used
+        ** as location for the save area (unused in case of main()).
+        */
+        Sym = AddLocalSym (Decl->Ident, Decl->Type, Decl->StorageClass, Reg);
     }
 
     /* Cannot allocate a variable of unknown size */