[llvm] [AMDGPU] Filter candidates of LiveRegOptimizer for profitable cases (PR #124624)

[Jeffrey Byrnes via llvm-commits]

================
@@ -1446,3 +1446,25 @@ void GCNTTIImpl::collectKernelLaunchBounds(
   LB.push_back({"amdgpu-waves-per-eu[0]", WavesPerEU.first});
   LB.push_back({"amdgpu-waves-per-eu[1]", WavesPerEU.second});
 }
+
+/// Check if operation is legal.
+/// TODO: If we had IR<->SDag mapping, we could use TLI->isOperationLegal
+bool GCNTTIImpl::isOpLegal(Instruction *I) {
----------------
jrbyrnes wrote:

Can you move this to AMDGPULateCodeGenPrepare. Don't whitelist shuffles, and should only whitelist v_dot / mfma / wmma (instead of all intrinsics).

https://github.com/llvm/llvm-project/pull/124624