[PATCH] D73444: AMDGPU/GlobalISel: Legalize TFE image result loads

[Nicolai Hähnle via Phabricator via llvm-commits]

nhaehnle added a comment.

Correct, TFE does not apply to store or atomic. Stores or atomics to unmapped partially-resident images will simply be ignored by the hardware (and atomic return values are undefined).

Do all cases still work in terms of how the intended semantics for the (super ugly) unpacked f16 case are represented? It seems a bit fragile.



================
Comment at: llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp:2779-2780
 
-  // Must be an image load.
+  if (BaseOpcode->Atomic) // No d16 atomics, or TFE?
+    return true;
+
----------------
This doesn't need to be moved.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D73444/new/

https://reviews.llvm.org/D73444