#28 OnBC: Fix codegen for functions returning structs

[Steve Little]

As per Philippe Guillot's email:

"the generated code for functions that return a
structure is not alway correct. Code like:

typedef union {
UInt32 u;
struct {UInt16 hi;UInt16 lo} hh
} memoryword;

memoryword foo() { blablabla }
memoryword bar() { ... return foo(); }

compiles but does not run properly due to incorrectly
generated code.
(one has to write ... m=foo(); return m; }"

[Steve Little]

Logged In: YES
user_id=583634

Slight change of description, plus full details from Philippe:

"Contrarily to what I said, it does not occur when the return
value is another function result, but it occurs when the return
value is an array element:

This is the source code:
#pragma segment 2
typedef union {
UInt32 u;
Int32 i;
struct { UInt16 hi; UInt16 lo; } hh;
struct { UInt8 b3; UInt8 b2; UInt8 b1; UInt8 b0; } qqqq;
} MWord;

MWord* Mem; // a global variable
MWord brol(int n) {
return Mem[n];
}

And this is the generated code (I added comments) :

; function Brol
0000 link a6,#0
0004 move.l -4(a5),a0 ; get global Mem address in a0
0008 move 12(a6),d0 ; get n value in d0 (parameter)
000C lsl 2,d0 ; multiply by 4 (because size of MWord
is 4)
000E add d0,d0 ; ????!!!! this is the bug.
0010 lea 0(a0,d0),a0 ; get address of Mem[n] in a0
0014 move.l 8(a6),a1 ; get de destination address in a1
0018 move.l (a0)+,(a1)+ ; transfert the data
001A unlk a6 ; epilog
001C rts

I have observed this bug only in a function return value.
Otherwize, the code that generate the address of similar
expression (such as Mem[n].hh.hi=5;) seems correct."

[Steve Little]

Logged In: YES
user_id=583634

A bit more info: (from Philippe)

" - The same bug occurs if MWord is a 4 bytes structure
- if MWord is 6 bytes then there are TWO instructions muls
#6,d0 (in place of lsl 2,d0 above)"

[John Wilund]

Logged In: YES
user_id=574706

I sencond! Please make this a voting matter.