tccgen: factor out gfunc_return

Also:
- on windows i386 and x86-64, structures of size <= 8 are
  NOT returned in registers if size is not one of 1,2,4,8.
- cleanup: put all tv-push/pop/swap/rot into one place

arm64-gen.c

@@ -1193,9 +1193,9 @@ ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret,
     return 0;
 }
 
-ST_FUNC void greturn(void)
+ST_FUNC void gfunc_return(CType *func_type)
 {
-    CType *t = &func_vt;
+    CType *t = func_type;
     unsigned long a;
 
     arm64_pcs(0, &t, &a);
@@ -1203,8 +1203,8 @@ ST_FUNC void greturn(void)
     case -1:
         break;
     case 0:
-        if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
+        if ((func_type->t & VT_BTYPE) == VT_STRUCT) {
-            int align, size = type_size(&func_vt, &align);
+            int align, size = type_size(func_type, &align);
             gaddrof();
             gv(RC_R(0));
             arm64_ldrs(0, size);
@@ -1213,7 +1213,7 @@ ST_FUNC void greturn(void)
             gv(RC_IRET);
         break;
     case 1: {
-        CType type = func_vt;
+        CType type = *func_type;
         mk_pointer(&type);
         vset(&type, VT_LOCAL | VT_LVAL, func_vc);
         indir();
@@ -1222,7 +1222,7 @@ ST_FUNC void greturn(void)
         break;
     }
     case 16:
-        if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
+        if ((func_type->t & VT_BTYPE) == VT_STRUCT) {
           uint32_t j, sz, n = arm64_hfa(&vtop->type, &sz);
           gaddrof();
           gv(RC_R(0));
@@ -1237,6 +1237,7 @@ ST_FUNC void greturn(void)
     default:
       assert(0);
     }
+    vtop--;
 }
 
 ST_FUNC void gfunc_epilog(void)

i386-gen.c

@@ -396,21 +396,21 @@ ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align, int
 {
 #ifdef TCC_TARGET_PE
     int size, align;
-
     *ret_align = 1; // Never have to re-align return values for x86
     *regsize = 4;
     size = type_size(vt, &align);
-    if (size > 8) {
+    if (size > 8 || (size & (size - 1)))
         return 0;
-    } else if (size > 4) {
+    if (size == 8)
-        ret->ref = NULL;
         ret->t = VT_LLONG;
-        return 1;
+    else if (size == 4)
-    } else {
-        ret->ref = NULL;
         ret->t = VT_INT;
-        return 1;
+    else if (size == 2)
-    }
+        ret->t = VT_SHORT;
+    else
+        ret->t = VT_BYTE;
+    ret->ref = NULL;
+    return 1;
 #else
     *ret_align = 1; // Never have to re-align return values for x86
     return 0;
@@ -547,7 +547,8 @@ ST_FUNC void gfunc_prolog(CType *func_type)
     func_var = (sym->c == FUNC_ELLIPSIS);
 #ifdef TCC_TARGET_PE
     size = type_size(&func_vt,&align);
-    if (((func_vt.t & VT_BTYPE) == VT_STRUCT) && (size > 8)) {
+    if (((func_vt.t & VT_BTYPE) == VT_STRUCT)
+        && (size > 8 || (size & (size - 1)))) {
 #else
     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 #endif

tcc.h

@@ -1273,7 +1273,6 @@ ST_FUNC void check_vstack(void);
 ST_INLN int is_float(int t);
 ST_FUNC int ieee_finite(double d);
 ST_FUNC void test_lvalue(void);
-ST_FUNC void swap(int *p, int *q);
 ST_FUNC void vpushi(int v);
 ST_FUNC Sym *external_global_sym(int v, CType *type, int r);
 ST_FUNC void vset(CType *type, int r, long v);
@@ -1519,7 +1518,7 @@ ST_FUNC void gen_cvt_itof1(int t);
 #ifdef TCC_TARGET_ARM64
 ST_FUNC void gen_cvt_sxtw(void);
 ST_FUNC void gen_opl(int op);
-ST_FUNC void greturn(void);
+ST_FUNC void gfunc_return(CType *func_type);
 ST_FUNC void gen_va_start(void);
 ST_FUNC void gen_va_arg(CType *t);
 ST_FUNC void gen_clear_cache(void);

tccgen.c

@@ -559,14 +559,6 @@ static void apply_visibility(Sym *sym, CType *type)
 
 /* ------------------------------------------------------------------------- */
 
-ST_FUNC void swap(int *p, int *q)
-{
-    int t;
-    t = *p;
-    *p = *q;
-    *q = t;
-}
-
 static void vsetc(CType *type, int r, CValue *vc)
 {
     int v;
@@ -583,12 +575,15 @@ static void vsetc(CType *type, int r, CValue *vc)
        as their value might still be used for real.  All values
        we push under nocode_wanted will eventually be popped
        again, so that the VT_CMP/VT_JMP value will be in vtop
-       when code is unsuppressed again.  */
+       when code is unsuppressed again.
+
+       Same logic below in vswap(); */
     if (vtop >= vstack && !nocode_wanted) {
         v = vtop->r & VT_VALMASK;
         if (v == VT_CMP || (v & ~1) == VT_JMP)
             gv(RC_INT);
     }
+
     vtop++;
     vtop->type = *type;
     vtop->r = r;
@@ -597,6 +592,38 @@ static void vsetc(CType *type, int r, CValue *vc)
     vtop->sym = NULL;
 }
 
+ST_FUNC void vswap(void)
+{
+    SValue tmp;
+    /* cannot vswap cpu flags. See comment at vsetc() above */
+    if (vtop >= vstack && !nocode_wanted) {
+        int v = vtop->r & VT_VALMASK;
+        if (v == VT_CMP || (v & ~1) == VT_JMP)
+            gv(RC_INT);
+    }
+    tmp = vtop[0];
+    vtop[0] = vtop[-1];
+    vtop[-1] = tmp;
+}
+
+/* pop stack value */
+ST_FUNC void vpop(void)
+{
+    int v;
+    v = vtop->r & VT_VALMASK;
+#if defined(TCC_TARGET_I386) || defined(TCC_TARGET_X86_64)
+    /* for x86, we need to pop the FP stack */
+    if (v == TREG_ST0) {
+        o(0xd8dd); /* fstp %st(0) */
+    } else
+#endif
+    if (v == VT_JMP || v == VT_JMPI) {
+        /* need to put correct jump if && or || without test */
+        gsym(vtop->c.i);
+    }
+    vtop--;
+}
+
 /* push constant of type "type" with useless value */
 ST_FUNC void vpush(CType *type)
 {
@@ -637,6 +664,71 @@ static inline void vpushll(long long v)
     vpush64(VT_LLONG, v);
 }
 
+ST_FUNC void vset(CType *type, int r, long v)
+{
+    CValue cval;
+
+    cval.i = v;
+    vsetc(type, r, &cval);
+}
+
+static void vseti(int r, int v)
+{
+    CType type;
+    type.t = VT_INT;
+    type.ref = 0;
+    vset(&type, r, v);
+}
+
+ST_FUNC void vpushv(SValue *v)
+{
+    if (vtop >= vstack + (VSTACK_SIZE - 1))
+        tcc_error("memory full (vstack)");
+    vtop++;
+    *vtop = *v;
+}
+
+static void vdup(void)
+{
+    vpushv(vtop);
+}
+
+/* rotate n first stack elements to the bottom
+   I1 ... In -> I2 ... In I1 [top is right]
+*/
+ST_FUNC void vrotb(int n)
+{
+    int i;
+    SValue tmp;
+
+    tmp = vtop[-n + 1];
+    for(i=-n+1;i!=0;i++)
+        vtop[i] = vtop[i+1];
+    vtop[0] = tmp;
+}
+
+/* rotate the n elements before entry e towards the top
+   I1 ... In ... -> In I1 ... I(n-1) ... [top is right]
+ */
+ST_FUNC void vrote(SValue *e, int n)
+{
+    int i;
+    SValue tmp;
+
+    tmp = *e;
+    for(i = 0;i < n - 1; i++)
+        e[-i] = e[-i - 1];
+    e[-n + 1] = tmp;
+}
+
+/* rotate n first stack elements to the top
+   I1 ... In -> In I1 ... I(n-1)  [top is right]
+ */
+ST_FUNC void vrott(int n)
+{
+    vrote(vtop, n);
+}
+
 /* push a symbol value of TYPE */
 static inline void vpushsym(CType *type, Sym *sym)
 {
@@ -715,56 +807,6 @@ ST_FUNC void vpush_global_sym(CType *type, int v)
     vpushsym(type, external_global_sym(v, type, 0));
 }
 
-ST_FUNC void vset(CType *type, int r, long v)
-{
-    CValue cval;
-
-    cval.i = v;
-    vsetc(type, r, &cval);
-}
-
-static void vseti(int r, int v)
-{
-    CType type;
-    type.t = VT_INT;
-    type.ref = 0;
-    vset(&type, r, v);
-}
-
-ST_FUNC void vswap(void)
-{
-    SValue tmp;
-    /* cannot let cpu flags if other instruction are generated. Also
-       avoid leaving VT_JMP anywhere except on the top of the stack
-       because it would complicate the code generator. */
-    if (vtop >= vstack) {
-        int v = vtop->r & VT_VALMASK;
-        if (v == VT_CMP || (v & ~1) == VT_JMP)
-            gv(RC_INT);
-    }
-    tmp = vtop[0];
-    vtop[0] = vtop[-1];
-    vtop[-1] = tmp;
-
-/* XXX: +2% overall speed possible with optimized memswap
- *
- *  memswap(&vtop[0], &vtop[1], sizeof *vtop);
- */
-}
-
-ST_FUNC void vpushv(SValue *v)
-{
-    if (vtop >= vstack + (VSTACK_SIZE - 1))
-        tcc_error("memory full (vstack)");
-    vtop++;
-    *vtop = *v;
-}
-
-static void vdup(void)
-{
-    vpushv(vtop);
-}
-
 /* save registers up to (vtop - n) stack entry */
 ST_FUNC void save_regs(int n)
 {
@@ -1297,60 +1339,6 @@ static void lbuild(int t)
 }
 #endif
 
-/* rotate n first stack elements to the bottom 
-   I1 ... In -> I2 ... In I1 [top is right]
-*/
-ST_FUNC void vrotb(int n)
-{
-    int i;
-    SValue tmp;
-
-    tmp = vtop[-n + 1];
-    for(i=-n+1;i!=0;i++)
-        vtop[i] = vtop[i+1];
-    vtop[0] = tmp;
-}
-
-/* rotate the n elements before entry e towards the top
-   I1 ... In ... -> In I1 ... I(n-1) ... [top is right]
- */
-ST_FUNC void vrote(SValue *e, int n)
-{
-    int i;
-    SValue tmp;
-
-    tmp = *e;
-    for(i = 0;i < n - 1; i++)
-        e[-i] = e[-i - 1];
-    e[-n + 1] = tmp;
-}
-
-/* rotate n first stack elements to the top
-   I1 ... In -> In I1 ... I(n-1)  [top is right]
- */
-ST_FUNC void vrott(int n)
-{
-    vrote(vtop, n);
-}
-
-/* pop stack value */
-ST_FUNC void vpop(void)
-{
-    int v;
-    v = vtop->r & VT_VALMASK;
-#if defined(TCC_TARGET_I386) || defined(TCC_TARGET_X86_64)
-    /* for x86, we need to pop the FP stack */
-    if (v == TREG_ST0) {
-        o(0xd8dd); /* fstp %st(0) */
-    } else
-#endif
-    if (v == VT_JMP || v == VT_JMPI) {
-        /* need to put correct jump if && or || without test */
-        gsym(vtop->c.i);
-    }
-    vtop--;
-}
-
 /* convert stack entry to register and duplicate its value in another
    register */
 static void gv_dup(void)
@@ -2025,7 +2013,7 @@ redo:
             /* Put pointer as first operand */
             if (bt2 == VT_PTR) {
                 vswap();
-                swap(&t1, &t2);
+                t = t1, t1 = t2, t2 = t;
             }
 #if PTR_SIZE == 4
             if ((vtop[0].type.t & VT_BTYPE) == VT_LLONG)
@@ -4668,11 +4656,11 @@ ST_FUNC void unary(void)
 	       subtract(-0, x).  */
 	    vpush(&vtop->type);
 	    if (t == VT_FLOAT)
-	        vtop->c.f = -0.0f;
+	        vtop->c.f = -1.0 * 0.0;
 	    else if (t == VT_DOUBLE)
-	        vtop->c.d = -0.0;
+	        vtop->c.d = -1.0 * 0.0;
 	    else
-	        vtop->c.ld = -0.0;
+	        vtop->c.ld = -1.0 * 0.0;
 	} else
 	    vpushi(0);
 	vswap();
@@ -5460,6 +5448,72 @@ static void label_or_decl(int l)
     decl(l);
 }
 
+#ifndef TCC_TARGET_ARM64
+static void gfunc_return(CType *func_type)
+{
+    if ((func_type->t & VT_BTYPE) == VT_STRUCT) {
+        CType type, ret_type;
+        int ret_align, ret_nregs, regsize;
+        ret_nregs = gfunc_sret(func_type, func_var, &ret_type,
+                               &ret_align, &regsize);
+        if (0 == ret_nregs) {
+            /* if returning structure, must copy it to implicit
+               first pointer arg location */
+            type = *func_type;
+            mk_pointer(&type);
+            vset(&type, VT_LOCAL | VT_LVAL, func_vc);
+            indir();
+            vswap();
+            /* copy structure value to pointer */
+            vstore();
+        } else {
+            /* returning structure packed into registers */
+            int r, size, addr, align;
+            size = type_size(func_type,&align);
+            if ((vtop->r != (VT_LOCAL | VT_LVAL) ||
+                 (vtop->c.i & (ret_align-1)))
+                && (align & (ret_align-1))) {
+                loc = (loc - size) & -ret_align;
+                addr = loc;
+                type = *func_type;
+                vset(&type, VT_LOCAL | VT_LVAL, addr);
+                vswap();
+                vstore();
+                vpop();
+                vset(&ret_type, VT_LOCAL | VT_LVAL, addr);
+            }
+            vtop->type = ret_type;
+            if (is_float(ret_type.t))
+                r = rc_fret(ret_type.t);
+            else
+                r = RC_IRET;
+
+            if (ret_nregs == 1)
+                gv(r);
+            else {
+                for (;;) {
+                    vdup();
+                    gv(r);
+                    vpop();
+                    if (--ret_nregs == 0)
+                      break;
+                    /* We assume that when a structure is returned in multiple
+                       registers, their classes are consecutive values of the
+                       suite s(n) = 2^n */
+                    r <<= 1;
+                    vtop->c.i += regsize;
+                }
+            }
+        }
+    } else if (is_float(func_type->t)) {
+        gv(rc_fret(func_type->t));
+    } else {
+        gv(RC_IRET);
+    }
+    vtop--; /* NOT vpop() because on x86 it would flush the fp stack */
+}
+#endif
+
 static int case_cmp(const void *pa, const void *pb)
 {
     int64_t a = (*(struct case_t**) pa)->v1;
@@ -5655,71 +5709,7 @@ static void block(int *bsym, int *csym, int is_expr)
         if (tok != ';') {
             gexpr();
             gen_assign_cast(&func_vt);
-#ifdef TCC_TARGET_ARM64
+            gfunc_return(&func_vt);
-            // Perhaps it would be better to use this for all backends:
-            greturn();
-#else
-            if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
-                CType type, ret_type;
-                int ret_align, ret_nregs, regsize;
-                ret_nregs = gfunc_sret(&func_vt, func_var, &ret_type,
-                                       &ret_align, &regsize);
-                if (0 == ret_nregs) {
-                    /* if returning structure, must copy it to implicit
-                       first pointer arg location */
-                    type = func_vt;
-                    mk_pointer(&type);
-                    vset(&type, VT_LOCAL | VT_LVAL, func_vc);
-                    indir();
-                    vswap();
-                    /* copy structure value to pointer */
-                    vstore();
-                } else {
-                    /* returning structure packed into registers */
-                    int r, size, addr, align;
-                    size = type_size(&func_vt,&align);
-                    if ((vtop->r != (VT_LOCAL | VT_LVAL) ||
-                         (vtop->c.i & (ret_align-1)))
-                        && (align & (ret_align-1))) {
-                        loc = (loc - size) & -ret_align;
-                        addr = loc;
-                        type = func_vt;
-                        vset(&type, VT_LOCAL | VT_LVAL, addr);
-                        vswap();
-                        vstore();
-                        vpop();
-                        vset(&ret_type, VT_LOCAL | VT_LVAL, addr);
-                    }
-                    vtop->type = ret_type;
-                    if (is_float(ret_type.t))
-                        r = rc_fret(ret_type.t);
-                    else
-                        r = RC_IRET;
-
-                    if (ret_nregs == 1)
-                        gv(r);
-                    else {
-                        for (;;) {
-                            vdup();
-                            gv(r);
-                            vpop();
-                            if (--ret_nregs == 0)
-                              break;
-                            /* We assume that when a structure is returned in multiple
-                               registers, their classes are consecutive values of the
-                               suite s(n) = 2^n */
-                            r <<= 1;
-                            vtop->c.i += regsize;
-                        }
-                    }
-                }
-            } else if (is_float(func_vt.t)) {
-                gv(rc_fret(func_vt.t));
-            } else {
-                gv(RC_IRET);
-            }
-#endif
-            vtop--; /* NOT vpop() because on x86 it would flush the fp stack */
         }
         skip(';');
         /* jump unless last stmt in top-level block */

x86_64-gen.c

@@ -752,25 +752,21 @@ void gen_offs_sp(int b, int r, int d)
 ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align, int *regsize)
 {
     int size, align;
-    *regsize = 8;
     *ret_align = 1; // Never have to re-align return values for x86-64
+    *regsize = 8;
     size = type_size(vt, &align);
-    ret->ref = NULL;
+    if (size > 8 || (size & (size - 1)))
-    if (size > 8) {
         return 0;
-    } else if (size > 4) {
+    if (size == 8)
         ret->t = VT_LLONG;
-        return 1;
+    else if (size == 4)
-    } else if (size > 2) {
         ret->t = VT_INT;
-        return 1;
+    else if (size == 2)
-    } else if (size > 1) {
         ret->t = VT_SHORT;
-        return 1;
+    else
-    } else {
         ret->t = VT_BYTE;
-        return 1;
+    ret->ref = NULL;
-    }
+    return 1;
 }
 
 static int is_sse_float(int t) {