|
|
@@ -241,85 +241,12 @@ unit cpupara;
|
|
|
|
|
|
|
|
|
function taarch64paramanager.ret_in_param(def:tdef;pd:tabstractprocdef):boolean;
|
|
|
- var
|
|
|
- i: longint;
|
|
|
- sym: tsym;
|
|
|
- fpufield: boolean;
|
|
|
begin
|
|
|
if handle_common_ret_in_param(def,pd,result) then
|
|
|
exit;
|
|
|
- case def.typ of
|
|
|
- recorddef:
|
|
|
- begin
|
|
|
- result:=def.size>4;
|
|
|
- if not result and
|
|
|
- (target_info.abi in [abi_default,abi_armeb]) then
|
|
|
- begin
|
|
|
- { in case of the old ARM abi (APCS), a struct is returned in
|
|
|
- a register only if it is simple. And what is a (non-)simple
|
|
|
- struct:
|
|
|
-
|
|
|
- "A non-simple type is any non-floating-point type of size
|
|
|
- greater than one word (including structures containing only
|
|
|
- floating-point fields), and certain single-word structured
|
|
|
- types."
|
|
|
- (-- ARM APCS documentation)
|
|
|
-
|
|
|
- So only floating point types or more than one word ->
|
|
|
- definitely non-simple (more than one word is already
|
|
|
- checked above). This includes unions/variant records with
|
|
|
- overlaid floating point and integer fields.
|
|
|
-
|
|
|
- Smaller than one word struct types are simple if they are
|
|
|
- "integer-like", and:
|
|
|
-
|
|
|
- "A structure is termed integer-like if its size is less than
|
|
|
- or equal to one word, and the offset of each of its
|
|
|
- addressable subfields is zero."
|
|
|
- (-- ARM APCS documentation)
|
|
|
-
|
|
|
- An "addressable subfield" is a field of which you can take
|
|
|
- the address, which in practive means any non-bitfield.
|
|
|
- In Pascal, there is no way to express the difference that
|
|
|
- you can have in C between "char" and "int :8". In this
|
|
|
- context, we use the fake distinction that a type defined
|
|
|
- inside the record itself (such as "a: 0..255;") indicates
|
|
|
- a bitpacked field while a field using a different type
|
|
|
- (such as "a: byte;") is not.
|
|
|
- }
|
|
|
- for i:=0 to trecorddef(def).symtable.SymList.count-1 do
|
|
|
- begin
|
|
|
- sym:=tsym(trecorddef(def).symtable.SymList[i]);
|
|
|
- if sym.typ<>fieldvarsym then
|
|
|
- continue;
|
|
|
- { bitfield -> ignore }
|
|
|
- if (trecordsymtable(trecorddef(def).symtable).usefieldalignment=bit_alignment) and
|
|
|
- (tfieldvarsym(sym).vardef.typ in [orddef,enumdef]) and
|
|
|
- (tfieldvarsym(sym).vardef.owner.defowner=def) then
|
|
|
- continue;
|
|
|
- { all other fields must be at offset zero }
|
|
|
- if tfieldvarsym(sym).fieldoffset<>0 then
|
|
|
- begin
|
|
|
- result:=true;
|
|
|
- exit;
|
|
|
- end;
|
|
|
- { floating point field -> also by reference }
|
|
|
- if tfieldvarsym(sym).vardef.typ=floatdef then
|
|
|
- begin
|
|
|
- result:=true;
|
|
|
- exit;
|
|
|
- end;
|
|
|
- end;
|
|
|
- end;
|
|
|
- end;
|
|
|
- procvardef:
|
|
|
- if not tprocvardef(def).is_addressonly then
|
|
|
- result:=true
|
|
|
- else
|
|
|
- result:=false
|
|
|
- else
|
|
|
- result:=inherited ret_in_param(def,pd);
|
|
|
- end;
|
|
|
+ { ABI: if the parameter would be passed in registers, it is returned
|
|
|
+ in those registers; otherwise, it's returned by reference }
|
|
|
+ result:=push_addr_param(vs_value,def,pd.proccalloption);
|
|
|
end;
|
|
|
|
|
|
|
Jonas Maebe