[llvm] [AMDGPU] Fix for GFX11 VOPD interlock hazard (PR #220348)

[via llvm-commits]

hidekisaito wrote:

> > I'd prefer if this was implemented by adding a new mask table such as the following to enforce the src parity constraint. 'constexpr unsigned VOPD_GFX11_VGPR_BANK_MASKS[] = {1, 1, 1, 1}; '
> > That way we don't need yet another bool argument to getInvalidCompOperandIndex
> 
> We don't have access to ST in InstInfo::getInvalidCompOperandIndex(). So, passing one more argument (bool, mask, or ST) to it or adding a new member (ST?) to InstInfo look inevitable. Isn't that the case? If so, what's the least evil? Pass ST instead of bool?

Used VOPD_GFX11_VGPR_BANK_MASKS[], but kept the bool parameter (which was renamed to make it more obvious). That seemed more natural than passing ST or pointer to gfx11 mask.

https://github.com/llvm/llvm-project/pull/220348