[AMD] Fix uniform offset computation

[giuseros]

There was a bug in how we were splitting the uniform/non-uniform offset contribution for addptr.

Consider this IR (where U is a uniform value, e.g., , coming from a splat and NU is non-uniform, coming e.g., from a make_range).

%a = %U+%NU
%b = %a + %NU
%c = addptr %ptr, %b

It would have been rewritten to

%b = %NU+%NU
%basePtr = addptr %basePtrOld, %U
%c = addi %offset, %b

The main issue here is that %b's operand #0 has changed, i.e., the scalar contribution has been removed. This is fine if addptr is the only operation that uses %b. If any other operation uses %b, they need the "old" %b.

The solution is to accumulate both the uniform and non-uniform contributions in a separate IR and leave the original %b untouched. Possible duplications will be removed by the canonicalizer .

Doing things in this way, I also could generalize the pass to all expressions of the form (U+NU)*(U+NU).

I tried enabling this pass and running all the suite and it is working fine

[antiagainst]

Nice. This is much better than mutating directly; thanks! I just have a few small nits.

[antiagainst]

I addressed the small nits given I need to retrigger the CI after some network issues anyway.