From a64353ce71f68b0386011d42060059aac5c8a8e3 Mon Sep 17 00:00:00 2001
From: grischka <grischka>
Date: Mon, 23 Sep 2019 17:45:39 +0200
Subject: [PATCH] tccgen.c: generic char/short promotion for function return
 values

---
 arm64-gen.c     |  6 +-----
 i386-gen.c      | 28 ----------------------------
 tccgen.c        |  6 ++++++
 tests/tcctest.c |  4 ++++
 x86_64-gen.c    | 35 +----------------------------------
 5 files changed, 12 insertions(+), 67 deletions(-)

diff --git a/arm64-gen.c b/arm64-gen.c
index da54b3d7..5ea47db4 100644
--- a/arm64-gen.c
+++ b/arm64-gen.c
@@ -959,11 +959,7 @@ ST_FUNC void gfunc_call(int nb_args)
     {
         int rt = return_type->t;
         int bt = rt & VT_BTYPE;
-        if (bt == VT_BYTE || bt == VT_SHORT)
-            // Promote small integers:
-            o(0x13001c00 | (bt == VT_SHORT) << 13 |
-              (uint32_t)!!(rt & VT_UNSIGNED) << 30); // [su]xt[bh] w0,w0
-        else if (bt == VT_STRUCT && !(a[0] & 1)) {
+        if (bt == VT_STRUCT && !(a[0] & 1)) {
             // A struct was returned in registers, so write it out:
             gv(RC_R(8));
             --vtop;
diff --git a/i386-gen.c b/i386-gen.c
index 47568b6e..8b81c66f 100644
--- a/i386-gen.c
+++ b/i386-gen.c
@@ -357,34 +357,6 @@ static void gcall_or_jmp(int is_jmp)
         o(0xff); /* call/jmp *r */
         o(0xd0 + r + (is_jmp << 4));
     }
-    if (!is_jmp) {
-        int rt;
-        /* extend the return value to the whole register if necessary
-           visual studio and gcc do not always set the whole eax register
-           when assigning the return value of a function  */
-        rt = vtop->type.ref->type.t;
-        switch (rt & VT_BTYPE) {
-            case VT_BYTE:
-            case VT_BOOL:
-                if (rt & VT_UNSIGNED) {
-                    o(0xc0b60f); /* movzx %al, %eax */
-                }
-                else {
-                    o(0xc0be0f); /* movsx %al, %eax */
-                }
-                break;
-            case VT_SHORT:
-                if (rt & VT_UNSIGNED) {
-                    o(0xc0b70f); /* movzx %ax, %eax */
-                }
-                else {
-                    o(0xc0bf0f); /* movsx %ax, %eax */
-                }
-                break;
-            default:
-                break;
-        }
-    }
 }
 
 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
diff --git a/tccgen.c b/tccgen.c
index 43c142b2..0475b3c5 100644
--- a/tccgen.c
+++ b/tccgen.c
@@ -5629,6 +5629,12 @@ special_math_val:
                     }
                     vset(&s->type, VT_LOCAL | VT_LVAL, addr);
                 }
+
+                /* Promote char/short return values. This is matters only
+                   for calling function that were not compiled by TCC */
+                t = s->type.t & VT_BTYPE;
+                if (t == VT_BYTE || t == VT_SHORT || t == VT_BOOL)
+                    vtop->r |= BFVAL(VT_MUSTCAST, 1);
             }
             if (s->f.func_noreturn)
                 CODE_OFF();
diff --git a/tests/tcctest.c b/tests/tcctest.c
index 0c28d316..3333e878 100644
--- a/tests/tcctest.c
+++ b/tests/tcctest.c
@@ -3316,6 +3316,9 @@ void override_func2 (void)
 extern int bug_table[] __attribute__((section("__bug_table")));
 char * get_asm_string (void)
 {
+#ifdef __i386__
+  char *str = "(not tested)";
+#else
   extern int some_symbol;
   asm volatile (".globl some_symbol\n"
 		"jmp .+6\n"
@@ -3330,6 +3333,7 @@ char * get_asm_string (void)
 		"2:\t.long 1b - 2b, %c0 - 2b\n"
 		".popsection\n" : : "i" ("A string"));
   char * str = ((char*)bug_table) + bug_table[1];
+#endif
   return str;
 }
 
diff --git a/x86_64-gen.c b/x86_64-gen.c
index 5d42f195..fd2d6936 100644
--- a/x86_64-gen.c
+++ b/x86_64-gen.c
@@ -902,23 +902,6 @@ void gfunc_call(int nb_args)
 #endif
 
     }
-
-    /* other compilers don't clear the upper bits when returning char/short */
-    bt = vtop->type.ref->type.t & (VT_BTYPE | VT_UNSIGNED);
-    if (bt == (VT_BYTE | VT_UNSIGNED) || (bt & VT_TYPE) == VT_BOOL)
-        o(0xc0b60f);  /* movzbl %al, %eax */
-    else if (bt == VT_BYTE)
-        o(0xc0be0f); /* movsbl %al, %eax */
-    else if (bt == VT_SHORT)
-        o(0x98); /* cwtl */
-    else if (bt == (VT_SHORT | VT_UNSIGNED))
-        o(0xc0b70f);  /* movzbl %al, %eax */
-#if 0 /* handled in gen_cast() */
-    else if (bt == VT_INT)
-        o(0x9848); /* cltq */
-    else if (bt == (VT_INT | VT_UNSIGNED))
-        o(0xc089); /* mov %eax,%eax */
-#endif
     vtop--;
 }
 
@@ -1278,7 +1261,7 @@ void gfunc_call(int nb_args)
 {
     X86_64_Mode mode;
     CType type;
-    int size, align, r, args_size, stack_adjust, i, reg_count, bt;
+    int size, align, r, args_size, stack_adjust, i, reg_count;
     int nb_reg_args = 0;
     int nb_sse_args = 0;
     int sse_reg, gen_reg;
@@ -1466,25 +1449,9 @@ void gfunc_call(int nb_args)
     gcall_or_jmp(0);
     if (args_size)
         gadd_sp(args_size);
-    /* other compilers don't clear the upper bits when returning char/short,
-       TCC does so for convenience.  When we'd stay purely within TCC compiled
-       code we wouldn't need this, but for compatibility we have to extend.
-       Ideally TCC wouldn't extend at return statements to not do double
-       extensions, or would understand sub-int types during expression
-       evaluation.  */
-    bt = vtop->type.ref->type.t & (VT_BTYPE | VT_UNSIGNED);
-    if (bt == (VT_BYTE | VT_UNSIGNED) || (bt & VT_TYPE) == VT_BOOL)
-        o(0xc0b60f);  /* movzbl %al, %eax */
-    else if (bt == VT_BYTE)
-        o(0xc0be0f); /* movsbl %al, %eax */
-    else if (bt == VT_SHORT)
-        o(0x98); /* cwtl */
-    else if (bt == (VT_SHORT | VT_UNSIGNED))
-        o(0xc0b70f);  /* movzwl %al, %eax */
     vtop--;
 }
 
-
 #define FUNC_PROLOG_SIZE 11
 
 static void push_arg_reg(int i) {