[llvm] [AMDGPU] Add support for GFX12 expert scheduling mode 2 (PR #170319)

[Stephen Thomas via llvm-commits]

================
@@ -1778,6 +1895,39 @@ bool WaitcntGeneratorGFX12Plus::applyPreexistingWaitcnt(
     }
   }
 
+  if (WaitcntDepctrInstr) {
+    // Get the encoded Depctr immediate and override the VA_VDST and VM_VSRC
+    // subfields with the new required values.
+    unsigned Enc =
+        TII->getNamedOperand(*WaitcntDepctrInstr, AMDGPU::OpName::simm16)
+            ->getImm();
+    Enc = AMDGPU::DepCtr::encodeFieldVmVsrc(Enc, Wait.VmVsrc);
+    Enc = AMDGPU::DepCtr::encodeFieldVaVdst(Enc, Wait.VaVdst);
+
+    ScoreBrackets.applyWaitcnt(VA_VDST, Wait.VaVdst);
+    ScoreBrackets.applyWaitcnt(VM_VSRC, Wait.VmVsrc);
+    Wait.VaVdst = ~0u;
+    Wait.VmVsrc = ~0u;
+
+    // If that new encoded Depctr immediate would actually still wait
+    // for anything, update the instruction's operand. Otherwise it can
+    // just be deleted.
+    if (Enc != 0xffff) {
----------------
stepthomas wrote:

See above

https://github.com/llvm/llvm-project/pull/170319