Refactoring push analysis pass, change PushInfo gather metadata

from. map to vector so that we can re-arrange the order of
gather constants if. required. Resubmit bc2991ae after fixing
functional regression.

Change-Id: I8bc935e9228f1bb3f8ba864438636d32cadab363

Author: pguo-igc

IGC/Compiler/CISACodeGen/CShader.cpp

@@ -409,7 +409,7 @@ void CShader::AllocateConstants(uint& offset)
 {
     m_ConstantBufferLength = 0;
     for (auto I = pushInfo.constants.begin(), E = pushInfo.constants.end(); I != E; I++) {
-        CVariable* var = GetSymbol(m_argListCache[I->second]);
+        CVariable* var = GetSymbol(m_argListCache[I->argIndex]);
         AllocateInput(var, offset + m_ConstantBufferLength);
         m_ConstantBufferLength += var->GetSize();
     }
@@ -453,7 +453,7 @@ void CShader::CreateGatherMap()
     gatherMap.reserve(pushInfo.constants.size());
     for(auto I = pushInfo.constants.begin(), E = pushInfo.constants.end();I!=E;I++)
     {
-		unsigned int address = (I->first.bufId * 256 * 4) + (I->first.eltId);
+		unsigned int address = (I->bufId * 256 * 4) + (I->eltId);
 		unsigned int cstOffset = address / 4;
 		unsigned int cstChannel = address % 4;
         if(cstOffset!=index)

IGC/Compiler/CISACodeGen/PushAnalysis.cpp

@@ -519,7 +519,6 @@ bool PushAnalysis::CanPushConstants()
     {
         return false;
     }
-
 
     switch(m_context->type)
     {
@@ -796,6 +795,40 @@ PushConstantMode PushAnalysis::GetPushConstantMode()
     return pushConstantMode;
 }
 
+// if multiple entries fall into cb[].xyzw, we need to insert in order
+static void insertGatherEntry(std::vector<ConstantAddress>& constants,
+    ConstantAddress& address)
+{
+    bool hasSameSlot = false;
+    bool inserted = false;
+
+    for (unsigned i = 0; i < constants.size(); i++)
+    {
+        if (constants[i].eltId / 4 == address.eltId / 4)
+        {
+            hasSameSlot = true;
+            if (constants[i].eltId > address.eltId)
+            {
+                constants.insert(constants.begin() + i, address);
+                inserted = true;
+                break;
+            }
+        }
+        else
+            if (hasSameSlot)
+            {
+                constants.insert(constants.begin() + i, address);
+                inserted = true;
+                break;
+            }
+    }
+
+    if (!inserted)
+    {
+        constants.push_back(address);
+    }
+}
+
 // process gather constants, update PushInfo.constants
 void PushAnalysis::processGather(Instruction* inst, uint bufId, uint eltId)
 {
@@ -838,7 +871,8 @@ void PushAnalysis::processGather(Instruction* inst, uint bufId, uint eltId)
         address.bufId = bufId;
         address.eltId = eltId + i;
 
-        auto it = pushInfo.constants.find(address);
+        auto it = std::find(pushInfo.constants.begin(),
+            pushInfo.constants.end(), address);
         if (it != pushInfo.constants.end() ||
             (pushInfo.constantReg.size() + pushInfo.constants.size() < m_pullConstantHeuristics->getPushConstantThreshold(m_pFunction) * 8))
         {
@@ -861,13 +895,14 @@ void PushAnalysis::processGather(Instruction* inst, uint bufId, uint eltId)
                 if (pTypeToPush != value->getType())
                     value = CastInst::CreateZExtOrBitCast(value, pTypeToPush, "", inst);
 
-                pushInfo.constants[address] = m_argIndex;
+                address.argIndex = m_argIndex;
+                insertGatherEntry(pushInfo.constants, address);
             }
             else
             {
-                assert((it->second <= m_argIndex) &&
+                assert((it->argIndex <= m_argIndex) &&
                     "Function arguments list and metadata are out of sync!");
-                value = m_argList[it->second];
+                value = m_argList[it->argIndex];
                 if (pTypeToPush != value->getType())
                     value = CastInst::CreateZExtOrBitCast(value, pTypeToPush, "", inst);
             }

IGC/common/MDFrameWork.cpp

@@ -65,6 +65,14 @@ namespace IGC
 
 		return false;
 	}
+
+    bool operator == (const ConstantAddress &a, const ConstantAddress &b)
+    {
+        if (a.bufId == b.bufId && a.eltId == b.eltId)
+            return true;
+
+        return false;
+    }
 }
 
 // non-autogen functions implementations below

IGC/common/MDFrameWork.h

@@ -143,14 +143,16 @@ namespace IGC
 		std::map<unsigned int, int> simplePushLoads;
 	};
 
-	struct ConstantAddress
-	{
-		unsigned int bufId = 0;
-		unsigned int eltId = 0;
+    struct ConstantAddress
+    {
+        unsigned int bufId = 0;
+        unsigned int eltId = 0;
         int size = 0;
-	};
+        int argIndex = 0;
+    };
 
-	bool operator < (const ConstantAddress &a, const ConstantAddress &b);
+    bool operator < (const ConstantAddress &a, const ConstantAddress &b);
+    bool operator == (const ConstantAddress &a, const ConstantAddress &b);
 
     struct StatelessPushInfo
     {
@@ -168,7 +170,7 @@ namespace IGC
         unsigned int inlineConstantBufferSlot = INVALID_CONSTANT_BUFFER_INVALID_ADDR; // slot of the inlined constant buffer
         unsigned int inlineConstantBufferOffset = INVALID_CONSTANT_BUFFER_INVALID_ADDR;    // offset of the inlined constant buffer
 
-		std::map<ConstantAddress, int> constants;
+		std::vector<ConstantAddress> constants;
 		std::map<unsigned int, SInputDesc> inputs;
 		std::map<unsigned int, int> constantReg;
 		std::array<SimplePushInfo, g_c_maxNumberOfBufferPushed> simplePushInfoArr;