>
> Hmm, I didn't know the IR could represent this! ;-)
It was already used but I would prefer to avoid it.
For example, when copying a structure:
struct s {
char name[5];
} s, d;
...
d = s;
will linearize into a single 40-bit load + store.
In this case, it's quite OK because it directly translate to
a memcpy().
> Is the 'lowering' code already present? Maybe next patch.
I had several versions, all more ugly than the others. It's why
I ended with this 'OK, keep things simple for now'.
Also, there is several ways of doing this and I'm not convinced
of which one should be used. Worse, the case:
struct {
a:10;
f:14;
};
should probably not be handled like the case:
struct {
a:5;
f:30;
z:5;
};
since the problems are different (the first one is just a question
of not doing an out-of-bound access, while for the second case we
have a field not wider than 4-bytes but which can't be accessed
in less than 5 bytes).
-- Luc
