[llvm] AMDGPU: VALU data fast-forwarding needs no s_delay (PR #205481)

[Jay Foad via llvm-commits]

================
@@ -346,6 +352,205 @@ class AMDGPUInsertDelayAlu {
     return (Imm & 0x780) ? nullptr : DelayAlu;
   }
 
+  static bool isFastForwardProducer(const MachineInstr &MI,
+                                    const MachineOperand &MO, Register VccReg,
+                                    Register ExecReg) {
+    if (!MO.isReg() || !MO.isDef())
+      return false;
+
+    if (!SIInstrInfo::isVALU(MI, /*AllowLDSDMA=*/false))
+      return false;
+
+    const TargetRegisterInfo *TRI =
+        MI.getMF()->getSubtarget().getRegisterInfo();
+    Register Reg = MO.getReg();
+    if (TRI->isSubRegisterEq(Reg, VccReg)) {
----------------
jayfoad wrote:

I think you might be able to simplify the code from here to the end of the function. All of these opcodes define exactly one SReg (either an SGPR or EXEC or VCC) so you should be able to handle them all uniformly with something like:
```
  if (isSGPR()) { // assuming isSGPR includes EXEC and VCC? If not, there should be another predicate that does include them
    switch (Opcode) {
    case ...:
      return true;
    default:
      return isCompare();
    }
  }
```


https://github.com/llvm/llvm-project/pull/205481