[llvm] [AMDGPU] Implement readcyclecounter for GFX12 (PR #76965)

[Matt Arsenault via llvm-commits]

================
@@ -4827,6 +4827,48 @@ MachineBasicBlock *SITargetLowering::EmitInstrWithCustomInserter(
     MI.eraseFromParent();
     return BB;
   }
+  case AMDGPU::GET_SHADERCYCLESHILO: {
+    assert(MF->getSubtarget<GCNSubtarget>().hasShaderCyclesHiLoRegisters());
+    MachineRegisterInfo &MRI = MF->getRegInfo();
+    const DebugLoc &DL = MI.getDebugLoc();
+    // The algorithm is:
+    //
+    // hi1 = getreg(SHADER_CYCLES_HI)
+    // lo1 = getreg(SHADER_CYCLES_LO)
+    // hi2 = getreg(SHADER_CYCLES_HI)
+    //
+    // If hi1 == hi2 then there was no overflow and the result is hi2:lo1.
+    // Otherwise there was overflow and the result is hi2:0. In both cases the
+    // result should represent the actual time at some point during the sequence
+    // of three getregs.
----------------
arsenm wrote:

This is pretty ugly

https://github.com/llvm/llvm-project/pull/76965