[llvm] [AMDGPU] Extend zero initialization of return values for TFE (PR #85759)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> Yes, I thought you might say that :) I'll take another look. Any suggestions? Image support is done in AMDGPUInstructionSelector.cpp:selectImageIntrinsic - so not appropriate for the buffer instructions - but it seems like overkill to implement a selectBufferLoadTfe when selectImpl does a good job already (but maybe that's the best approach??)

The handling of all of these intrinsics is split between legalizeImageIntrinsic / legalizeBuffer* and the select* functions. Anything changing the register types used should be done in the legalizer 


https://github.com/llvm/llvm-project/pull/85759