[llvm] [AMDGPU] Avoid setting GLC for image atomics when result is unused (PR #150742)

[Matt Arsenault via llvm-commits]

================
@@ -2123,8 +2123,10 @@ bool AMDGPUInstructionSelector::selectImageIntrinsic(
   assert((!IsTexFail || DMaskLanes >= 1) && "should have legalized this");
 
   unsigned CPol = MI.getOperand(ArgOffset + Intr->CachePolicyIndex).getImm();
-  if (BaseOpcode->Atomic)
-    CPol |= AMDGPU::CPol::GLC; // TODO no-return optimization
+  // Keep GLC only when the atomic's result is actually used.
+  if (BaseOpcode->Atomic && VDataOut && !MRI->use_empty(VDataOut))
----------------
arsenm wrote:

use_nodbg_empty 

https://github.com/llvm/llvm-project/pull/150742