Add ByteAddressBuffer support

[spencer-lunarg]

attempt 2 of #236 (closes #234)

When ByteAddressBuffer are collapsed they start to look like

OpDecorateString %var UserTypeGOOGLE "byteaddressbuffer"
%var = OpVariable %22 StorageBuffer

%a = OpLoad %type %input
%b = OpIAdd %type %a %uint_16
%c = OpShiftRightLogical %type %b uint_2
%_ = OpAccessChain %e %var 0 %c

using the UserType we can detect which variable is used, from here we can build up an offset by working out the math

// offset + 16
%b = OpIAdd %type %a %uint_16
// offset >> 2
%c = OpShiftRightLogical %type %b uint_2
// offset == 4

[danginsburg]

Should this check be above this? Otherwise it will return if you have 8 operations when it could have succeeded (if you have exactly 8). I also don't understand why there is such a low cap on this, it seems rather inflexible. Would it be better to scan for how many first then allocate an array with a cap of like the size of the spir-v or something?

The problem is that SPIR-V reflection will completely fail in this case, not just that you won't get the offsets.

[spencer-lunarg]

so before it failed because I returned an error, not I return the same as "could not find" it it hits the cap

I guess there is no reason for a hard limit, it just I have yet to find a case where it is more then 2 here

Would it be better to scan for how many first

So the other issue is I do need "a cap" when doing this search anyways, there is no "correct" way for HLSL/Slang to flatten the ByteAddressBuffer so in the future, if some strange pattern arises, I still need some arbitrary number to stop at incase I get in some strange situation and don't want to crash

[danginsburg]

I've tested the latest PR with our full shader corpus and can confirm it's working properly and producing proper ByteAddressBuffer offsets. Looks good to merge to me.

[chaoticbob]

LGTM