[llvm] [AMDGPU] Remove s_delay_alu for VALU->SGPR->SALU (PR #127212)

[Nicolai Hähnle via llvm-commits]

================
@@ -371,12 +376,51 @@ class AMDGPUInsertDelayAlu : public MachineFunctionPass {
             for (MCRegUnit Unit : TRI->regunits(Op.getReg())) {
               auto It = State.find(Unit);
               if (It != State.end()) {
-                Delay.merge(It->second);
-                State.erase(Unit);
+                if (SII->isSALU(MI.getOpcode()) &&
+                    AMDGPU::isSGPR(Op.getReg(), TRI) &&
+                    It->second.VALUCycles > 0) {
+                  deletedCyclesNum = It->second.VALUCycles;
+                  State.erase(Unit);
+                  VALUSALUStall = true;
+                } else {
+                  Delay.merge(It->second);
+                  State.erase(Unit);
+                }
               }
             }
           }
         }
+        unsigned maxCycles = 0;
+        unsigned lastWrite = 0;
+        if (Type != OTHER) {
+          for (const auto &Op : MI.defs()) {
+            for (MCRegUnit Unit : TRI->regunits(Op.getReg())) {
+              if (AMDGPU::isSGPR(Op.getReg(), TRI)) {
+                maxCycles =
+                    (State.find(longestWait) == State.end())
+                        ? std::max(deletedCyclesNum, (unsigned)0)
+                        : std::max(State[longestWait].VALUCycles,
+                                   State[longestWait].SALUCycles);
+                lastWrite =
----------------
nhaehnle wrote:

Please still address the issue of repeated lookups in State.

https://github.com/llvm/llvm-project/pull/127212