Change schedule priority according to dep type

For barrier dep, shouldn't use latency cycle to calcuate priority, because barrier is order issue, not the latecy issue. Use occupancy steady.

Author: bcheng0127

IGC/ocloc_tests/optimizations/PromoteToPredicatedMemoryAccess/LoadSubDW.cl

@@ -13,8 +13,6 @@ SPDX-License-Identifier: MIT
 // RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'EnablePromoteToPredicatedMemoryAccess=1 VISAOptions=-asmToConsole'" 2>&1 | FileCheck %s --check-prefixes=CHECK-ASM
 
 // CHECK-ASM: kernel test_i8_0
-// CHECK-ASM:             cmp (32|M0)   (lt)[[F1:f[0-9\.]+]]
-// CHECK-ASM:   ([[F1]]) goto (32|M0)
 // CHECK-ASM: {{[_a-z0-9A-Z]+}}:
 // CHECK-ASM-DAG:         cmp (32|M0)   (le)[[F2:f[0-9\.]+]]   null<1>:d     r{{[0-9\.]+}}
 // CHECK-ASM-DAG:         cmp (32|M0)   (ge)[[F3:f[0-9\.]+]]   null<1>:d     r{{[0-9\.]+}}

visa/LocalScheduler/Dependencies_G4IR.cpp

@@ -189,6 +189,15 @@ DepType vISA::getDepScratchSend(G4_INST *curInst, G4_INST *liveInst) {
   return NODEP;
 }
 
+bool vISA::isNotLatencyBarrier (DepType type) {
+  if (type == CONTROL_FLOW_BARRIER ||
+      type == OPT_BARRIER ||
+      type == SEND_BARRIER ||
+      type == MSG_BARRIER)
+    return true;
+  return false;
+}
+
 DepType vISA::CheckBarrier(G4_INST *inst) {
   if (inst->isOptBarrier() || inst->isAtomicInst() || inst->opcode() == G4_madm) {
     return OPT_BARRIER;

visa/LocalScheduler/Dependencies_G4IR.h

@@ -38,6 +38,8 @@ DepType getDepSend(G4_INST *curInst, G4_INST *liveInst, bool BTIIsRestrict);
 
 DepType getDepScratchSend(G4_INST *curInst, G4_INST *liveInst);
 
+bool isNotLatencyBarrier(DepType type);
+
 DepType CheckBarrier(G4_INST *inst);
 
 DepType getDepForOpnd(Gen4_Operand_Number cur, Gen4_Operand_Number liv);

visa/LocalScheduler/LocalScheduler_G4IR.cpp

@@ -1887,9 +1887,21 @@ void DDD::collectRoots() {
 void DDD::setPriority(Node *pred, const Edge &edge) {
   // Calculate PRED's priority (pred->priority), based on SUCC's priority
   Node *succ = edge.getNode();
+  DepType type = edge.getType();
   vISA_ASSERT(succ->priority != Node::PRIORITY_UNINIT,
          "succ node has no priority?");
-  int newPriority = succ->priority + edge.getLatency();
+  int newPriority = succ->priority;
+  // Note that, node->isBarrier cannot be used here. Because there may be
+  // non-barrier dep.
+  bool isSend = false;
+  if (!pred->getInstructions()->empty()) {
+    isSend = pred->getInstructions()->front()->isSend();
+  }
+  if (isNotLatencyBarrier(type) &&
+      !isSend) // send has long latency, it's better be scheduled to the front.
+    newPriority += pred->getOccupancy();
+  else
+    newPriority += edge.getLatency();
   pred->priority =
       (newPriority > pred->priority) ? newPriority : pred->priority;
 }