Reland "Fix the calculation of jitinfo's usesBarrier in finalizer."

Calculate the number of barriers when translating the vISA barrier
instructions. For named barriers, derive the number from barrier ID
instead of the number of send barrier instructions. For the legacy
barrier, set it to 1 when there's a barrier instruction.

Also when propagating the barrier information from callee to caller,
change the caller's usesBarrier to callee's value if callee's value is
bigger than caller's.

Author: petechou

visa/BuildCISAIRImpl.cpp

@@ -1219,7 +1219,9 @@ static void retrieveBarrierInfoFromCallee(VISAKernelImpl* entry, std::set<VISAKe
         // path? For zebin path, the barrier information of an indirect
         // (external) function will be provided in the corresponding .ze_info
         // field.
-        entry->getIRBuilder()->getJitInfo()->usesBarrier |= callee->getIRBuilder()->getJitInfo()->usesBarrier;
+        entry->getIRBuilder()->getJitInfo()->usesBarrier =
+            std::max(entry->getIRBuilder()->getJitInfo()->usesBarrier,
+                     callee->getIRBuilder()->getJitInfo()->usesBarrier);
     }
 }
 

visa/Optimizer.cpp

@@ -8212,11 +8212,10 @@ bool Optimizer::foldPseudoAndOr(G4_BB* bb, INST_LIST_ITER& ii)
         }
     }
 
-    // perform simple stat collection (e.g., numBarriers, numSends)
+    // perform simple stat collection (e.g., numSends)
     // IR is not modified
     void Optimizer::collectStats()
     {
-        builder.getJitInfo()->usesBarrier = 0;
         uint32_t numSends = 0;
         for (auto bb : fg)
         {
@@ -8225,14 +8224,6 @@ bool Optimizer::foldPseudoAndOr(G4_BB* bb, INST_LIST_ITER& ii)
                 if (inst->isSend())
                 {
                     numSends++;
-                    if (inst->asSendInst()->getMsgDesc()->isBarrier())
-                    {
-                        // Propagate information about barriers presence back to IGC. It's safer to
-                        // depend on vISA statistics as IGC is not able to detect barriers if they are
-                        // used as a part of Inline vISA code.
-                        // This information is used by legacy CMRT as well as OpenCL/L0 runtime.
-                        builder.getJitInfo()->usesBarrier += 1;
-                    }
                 }
             }
         }

visa/VisaToG4/TranslateSendSync.cpp

@@ -277,7 +277,7 @@ static void checkNamedBarrierSrc(G4_Operand* src, bool isBarrierId)
         if (isBarrierId)
         {
             uint32_t val = (uint32_t)src->asImm()->getInt();
-            assert(val < 32 && "illegal named barrier id");
+            assert(val < FINALIZER_INFO::kMaxNamedBarriers && "illegal named barrier id");
         }
     }
     else if (src->isSrcRegRegion())
@@ -291,12 +291,36 @@ static void checkNamedBarrierSrc(G4_Operand* src, bool isBarrierId)
     }
 }
 
+static void updateNamedBarriersInJitInfo(FINALIZER_INFO* jitinfo, G4_Operand* barrierId)
+{
+    if (!jitinfo)
+        return;
+
+    if (barrierId->isImm())
+    {
+        // Update the number of named barriers to barrier id + 1 as the id is a
+        // 0-based number.
+        unsigned id = (unsigned)barrierId->asImm()->getInt();
+        jitinfo->usesBarrier = std::max(jitinfo->usesBarrier, id + 1);
+    }
+    else
+    {
+        // In order to be safe, set the number of named barriers to the max
+        // value allowed if the barrier id is unknown to vISA. We probably
+        // won't see this in typical cases as the barrier id provided by users
+        // like IGC should be an immediate value.
+        jitinfo->usesBarrier = FINALIZER_INFO::kMaxNamedBarriers;
+    }
+}
+
 int IR_Builder::translateVISANamedBarrierWait(G4_Operand* barrierId)
 {
     TIME_SCOPE(VISA_BUILDER_IR_CONSTRUCTION);
 
     checkNamedBarrierSrc(barrierId, true);
 
+    updateNamedBarriersInJitInfo(getJitInfo(), barrierId);
+
     G4_Operand* barSrc = barrierId;
     if (barrierId->isSrcRegRegion()) {
         // sync can take only flag src
@@ -319,6 +343,8 @@ int IR_Builder::translateVISANamedBarrierSignal(G4_Operand* barrierId, G4_Operan
     checkNamedBarrierSrc(barrierId, true);
     checkNamedBarrierSrc(threadCount, false);
 
+    updateNamedBarriersInJitInfo(getJitInfo(), barrierId);
+
     if (threadCount->isImm())
     {
         int numThreads = (int)threadCount->asImm()->getInt();
@@ -424,9 +450,21 @@ int IR_Builder::translateVISAWaitInst(G4_Operand* mask)
     return VISA_SUCCESS;
 }
 
+static void updateBarrierInJitInfo(FINALIZER_INFO* jitinfo)
+{
+    if (!jitinfo)
+        return;
+
+    // For the legacy barrier, update usesBarrier to 1 when there's a
+    // barrier.
+    if (jitinfo->usesBarrier == 0)
+        jitinfo->usesBarrier = 1;
+}
 
 void IR_Builder::generateBarrierSend()
 {
+    updateBarrierInJitInfo(getJitInfo());
+
     if (hasUnifiedBarrier())
     {
         generateSingleBarrier();
@@ -478,6 +516,8 @@ void IR_Builder::generateBarrierSend()
 
 void IR_Builder::generateBarrierWait()
 {
+    updateBarrierInJitInfo(getJitInfo());
+
     G4_Operand* waitSrc = nullptr;
     if (!hasUnifiedBarrier()) {
 

visa/include/JitterDataStruct.h

@@ -39,8 +39,17 @@ typedef struct _FINALIZER_INFO{
     unsigned numFlagSpillStore;
     unsigned numFlagSpillLoad;
 
-    // whether kernel uses a barrier
-    unsigned usesBarrier;
+    // Propagate information about barriers presence back to IGC. It's safer to
+    // depend on vISA statistics as IGC is not able to detect barriers if they
+    // are used as a part of Inline vISA code.
+    // This information is used by legacy CMRT as well as OpenCL/L0 runtime.
+    //
+    // The number of named barriers; it also indicates whether kernel uses a
+    // barrier.
+    // TODO: Maybe consider to use a bitset to track the barrier IDs used, and
+    // the number of bits set can be used to represent the number of barriers.
+    unsigned usesBarrier = 0;
+    static constexpr unsigned kMaxNamedBarriers = 32;
 
     unsigned BBNum;
     VISA_BB_INFO* BBInfo;