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

[Reem Elkhouly via llvm-commits]

================
@@ -121,6 +122,18 @@ class AMDGPUInsertDelayAlu {
     // until it completes.
     uint8_t SALUCycles = 0;
 
+    // Types of data fast-forwarding for producer/consumer operands.
----------------
amd-relkhoul wrote:

I found that using `MCInstrDesc`to identify operands that are part of the instruction definition, rather than those added for liveness tracking, helps eliminate the false positives for implicit operands in the fast-forward and delay-insertion cases.
With a change in the logic to check the operand before checking the opcode, the differentiation of FFKinds proved to be unnecessary, as you had suggested.

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