SetMaxRegForThreadDispatch was hardcoded for up to 128 GRFs.

SetMaxRegForThreadDispatch was hardcoded for up to 128 GRFs, so errors
could occur on future platforms if larger sets of GRFs allowed.

Author: mtargows

IGC/Compiler/CISACodeGen/CISABuilder.cpp

@@ -5414,7 +5414,7 @@ namespace IGC
             context->type == ShaderType::GEOMETRY_SHADER || context->type == ShaderType::HULL_SHADER)
         {
             unsigned maxReg = m_program->GetMaxRegForThreadDispatch();
-            V(vKernel->AddKernelAttribute("MaxRegThreadDispatch", 1, &maxReg));
+            V(vKernel->AddKernelAttribute("MaxRegThreadDispatch", sizeof(maxReg), &maxReg));
         }
     }
 

IGC/Compiler/CISACodeGen/CoalescingEngine.cpp

@@ -251,8 +251,8 @@ namespace IGC
             SetCurrentPart(tupleGeneratingInstruction, numPart);
             const uint numOperands = GetNumPayloadElements(tupleGeneratingInstruction);
             bool isAnyNodeAnchored = false, isAnyNodeCoalescable = false;
-            SmallPtrSet<CCTuple*, 8> touchedTuplesSet;
-            SmallVector<CCTuple*, 8> touchedTuples;
+            SmallPtrSet<CCTuple*, MaxTouchedTuples> touchedTuplesSet;
+            SmallVector<CCTuple*, MaxTouchedTuples> touchedTuples;
 
             //Step: Prepare.
             PrepareTuple(
@@ -371,7 +371,7 @@ namespace IGC
                     //provided that elements in dominatorsForDisplacement are displaced, and other nodes are attached.
                     //If interferes is true, then no element will be attached to the ccTuple.
                     if (!interferes) {
-                        SmallPtrSet<Value*, 8> touchedValuesSet;
+                        SmallPtrSet<Value*, MaxTouchedTuples> touchedValuesSet;
                         for (uint i = 0; i < numOperands; i++) {
                             Value* val = GetPayloadElementToValueMapping(tupleGeneratingInstruction, i);
 
@@ -575,8 +575,8 @@ namespace IGC
     void CoalescingEngine::PrepareTuple(
         const uint numOperands,
         Instruction* tupleGeneratingInstruction,
-        SmallPtrSet<CCTuple*, 8> & touchedTuplesSet,
-        SmallVector<CCTuple*, 8> & touchedTuples,
+        SmallPtrSet<CCTuple*, MaxTouchedTuples> & touchedTuplesSet,
+        SmallVector<CCTuple*, MaxTouchedTuples> & touchedTuples,
         bool& isAnyNodeAnchored,
         bool& isAnyNodeCoalescable)
     {
@@ -843,7 +843,7 @@ namespace IGC
         CCTuple* ccTuple,
         ElementFunctor* functor)
     {
-        SmallPtrSet<Value*, 8> touchedValuesSet;
+        SmallPtrSet<Value*, MaxTouchedTuples> touchedValuesSet;
 
         for (uint i = 0; i < numOperands; i++) {
             functor->SetIndex(i);
@@ -1001,7 +1001,7 @@ namespace IGC
         CCTuple* ccTuple,
         ProcessInterferencesElementFunctor* interferencesFunctor)
     {
-        SmallPtrSet<Value*, 8> touchedValuesSet;
+        SmallPtrSet<Value*, MaxTouchedTuples> touchedValuesSet;
         GatherWeightElementFunctor gatherFunctor;
         ProcessElements(numOperands, tupleInst, offsetDiff, ccTuple, &gatherFunctor);
         bool forceEviction =
@@ -1071,7 +1071,7 @@ namespace IGC
 
         if (ccTuple)
         {
-            SmallPtrSet<Value*, 8> touchedValuesSet;
+            SmallPtrSet<Value*, MaxTouchedTuples> touchedValuesSet;
 
             //index = 0;
             payloadCovered = true;
@@ -1179,7 +1179,7 @@ namespace IGC
         }
 
         if (payloadCovered) {
-            SmallPtrSet<Value*, 8> touchedValuesSet;
+            SmallPtrSet<Value*, MaxTouchedTuples> touchedValuesSet;
 
             for (uint index = 0; index < numOperands; index++)
             {

IGC/Compiler/CISACodeGen/CoalescingEngine.hpp

@@ -34,7 +34,8 @@ namespace IGC {
     class CoalescingEngine : public llvm::FunctionPass, public llvm::InstVisitor<CoalescingEngine>
     {
         //TODO: this is fixed for now, but once we have pressure heuristic, could be relaxed
-        static const int MaxTupleSize = 12;
+        static const int MaxTouchedTuples = 8;
+        static const int MaxTupleSize = MaxTouchedTuples + 4;
 
     public:
         static char ID; // Pass identification, replacement for typeid
@@ -327,7 +328,7 @@ namespace IGC {
             else {
                 uint numUsers = 0;
                 {
-                    llvm::SmallPtrSet<llvm::User*, 8> touchedUsers;
+                    llvm::SmallPtrSet<llvm::User*, MaxTouchedTuples> touchedUsers;
                     for (llvm::Value::user_iterator i = val->user_begin(), e = val->user_end(); i != e; ++i) {
                         llvm::User* user = *i;
                         if (llvm::isa<llvm::Instruction>(user)) {
@@ -388,8 +389,8 @@ namespace IGC {
         void PrepareTuple(
             const uint numOperands,
             llvm::Instruction* tupleGeneratingInstruction,
-            llvm::SmallPtrSet<CCTuple*, 8> & touchedTuplesSet,
-            llvm::SmallVector<CCTuple*, 8> & touchedTuples,
+            llvm::SmallPtrSet<CCTuple*, MaxTouchedTuples> & touchedTuplesSet,
+            llvm::SmallVector<CCTuple*, MaxTouchedTuples> & touchedTuples,
             bool& isAnyNodeAnchored,
             bool& isAnyNodeCoalescable);
 
@@ -712,7 +713,7 @@ namespace IGC {
             CCTuple* m_ccTuple;
             CoalescingEngine* m_CE;
             int m_index;
-            llvm::SmallPtrSet<llvm::Value*, 8> m_valuesForIsolation;
+            llvm::SmallPtrSet<llvm::Value*, MaxTouchedTuples> m_valuesForIsolation;
 
         public:
             ProcessInterferencesElementFunctor(
@@ -732,7 +733,7 @@ namespace IGC {
 
             }
 
-            llvm::SmallPtrSet<llvm::Value*, 8> & GetComputedValuesForIsolation()
+            llvm::SmallPtrSet<llvm::Value*, MaxTouchedTuples> & GetComputedValuesForIsolation()
             {
                 return m_valuesForIsolation;
             }