Add support for struct > 4B returned via registers

On ARM with hardfloat calling convention, structure containing 4 fields
or less of the same float type are returned via float registers. This
means that a structure can be returned in up to 4 double registers in a
structure is composed of 4 doubles. This commit adds support for return
of structures in several registers.

1 files changed, 30 insertions, 18 deletions

diff --git a/arm-gen.c b/arm-gen.c
index 92b080d..6d0acd8 100644
--- a/arm-gen.c
+++ b/arm-gen.c
@@ -746,24 +746,6 @@ static void gcall_or_jmp(int is_jmp)
   }
 }
 
-/* Return 1 if this function returns via an sret pointer, 0 otherwise */
-ST_FUNC int gfunc_sret(CType *vt, CType *ret, int *ret_align) {
-#ifdef TCC_ARM_EABI
-    int size, align;
-    size = type_size(vt, &align);
-    if (size > 4) {
-        return 1;
-    } else {
-        *ret_align = 4;
-        ret->ref = NULL;
-        ret->t = VT_INT;
-        return 0;
-    }
-#else
-    return 1;
-#endif
-}
-
 #ifdef TCC_ARM_HARDFLOAT
 /* Return whether a structure is an homogeneous float aggregate or not.
    The answer is true if all the elements of the structure are of the same
@@ -831,6 +813,33 @@ int assign_vfpreg(struct avail_regs *avregs, int align, int size)
 }
 #endif
 
+/* Return the number of registers needed to return the struct, or 0 if
+   returning via struct pointer. */
+ST_FUNC int gfunc_sret(CType *vt, CType *ret, int *ret_align) {
+#ifdef TCC_ARM_EABI
+    int size, align;
+    size = type_size(vt, &align);
+#ifdef TCC_ARM_HARDFLOAT
+    if (is_float(vt->t) || is_hgen_float_aggr(vt)) {
+        *ret_align = 8;
+        ret->ref = NULL;
+        ret->t = VT_DOUBLE;
+        return (size + 7) >> 3;
+    } else
+#endif
+    if (size > 4) {
+        return 0;
+    } else {
+        *ret_align = 4;
+        ret->ref = NULL;
+        ret->t = VT_INT;
+        return 1;
+    }
+#else
+    return 0;
+#endif
+}
+
 /* Parameters are classified according to how they are copied to their final
    destination for the function call. Because the copying is performed class
    after class according to the order in the union below, it is important that
@@ -1198,6 +1207,9 @@ void gfunc_prolog(CType *func_type)
   n = nf = 0;
   variadic = (func_type->ref->c == FUNC_ELLIPSIS);
   if((func_vt.t & VT_BTYPE) == VT_STRUCT
+#ifdef TCC_ARM_HARDFLOAT
+     && (variadic || !is_hgen_float_aggr(&func_vt))
+#endif
      && type_size(&func_vt,&align) > 4)
   {
     n++;