[CIR][CodeGen] Updates GlobalViewAttr's indices computation for the union type

[gitoleg]

This PR introduces a new attribute for the explicit byte offset for GlobalViewAttr. It's a proposal, so feel free to reject it once it's not good from your point of view.

The problem is (as usually) with globals and unions: looks like we can not really use indexes in the GlobalView to address an array of unions. For example, the next program prints 4 now, but it should be 42:

typedef struct {
    long s0;
    int  s1;
} S;

typedef union {
   int  f0;
   S f1;
} U;

static U g1[3] = {{42},{42},{42}};
int* g2 = &g1[1].f1.s1;

int main() {       
   (*g2) = 4;
   printf("%d\n", g1[1].f0);    
   return0;
}

The problem is that we compute wrong indices in CIRGenBuilder::computeGlobalViewIndicesFromFlatOffset. Maybe it can be even fixed in this case, but I have a feeling that the fix would be a bit fragile.

So instead of trying to support indexes for the array of unions I suggest to use the offset explicitly.

From the implementation point of view there are some changes in CIRGenBuilder - but nothing really new is in there - I just did not want to introduce copy-pasta for the isOffsetInUnion function that is pretty the same as former computeGlobalViewIndicesFromFlatOffset.

[gitoleg]

I think I have a better solution and just update the index computation - so once you are good with it - I can rename the PR.
The only downside of this approach and it could be a chance that indices will be computed wrong:

typedef struct {
   double x;
   int y;
} S1;

typedef struct {
     long x;
     int y;
} S2; 

typedef union {
   S1 s1;
   S2 s2;
} U;

static U g1[2] = {{42},{42}};
int* g2 = &g1[1].s1.x;
int* g3 = &g1[1].s2.x;

Both of g2 and g3 will have the same indexes. But the point is - both of them will have the same offset on LLVM IR side, which is more important since it's correct and even with the original LLVM with CIR involved.

[bcardosolopes]

This is much better, apologies for the delay! LGTM pending some nits

[bcardosolopes]

Can you add some comment here so it's easy to understand why unions are different?

[bcardosolopes]

Similar here!

[gitoleg]

done!