Fix a case where vISA spill size is not reported correctly.

Spill code cleanup pass may create OOB spill/fill during coalescing,
and we need to adjust spill size accordingly.

Author: weiyu-chen

visa/GraphColor.cpp

@@ -10051,12 +10051,8 @@ int GlobalRA::coloringRegAlloc()
         return VISA_SPILL;
     }
 
-    // do not double count the spill mem offset
-    // Note that this includes both SLM and scratch space spills
-    uint32_t spillMemUsed = nextSpillOffset;
-    //
-    // Report spill memory usage information.
-    //
+    // this includes vISA's scratch space use only and does not include whatever IGC may use for private memory
+    uint32_t spillMemUsed = std::max(this->actualSpillSize, nextSpillOffset);
     if (spillMemUsed)
     {
         builder.criticalMsgStream() << "Spill memory used = " << spillMemUsed << " bytes for kernel " <<

visa/GraphColor.h

@@ -791,6 +791,10 @@ namespace vISA
         uint32_t numGRFSpill = 0;
         uint32_t numGRFFill = 0;
 
+        // kernel's spill size
+        // We need this because spill code cleanup may result in OOB fill/spill during coalescing.
+        uint32_t actualSpillSize = 0;
+
         void expandFillNonStackcall(uint32_t numRows, uint32_t offset, short rowOffset, G4_SrcRegRegion* header, G4_DstRegRegion* resultRgn, G4_BB* bb, INST_LIST_ITER& instIt);
         void expandSpillNonStackcall(uint32_t numRows, uint32_t offset, short rowOffset, G4_SrcRegRegion* header, G4_SrcRegRegion* payload, G4_BB* bb, INST_LIST_ITER& instIt);
         void expandFillStackcall(uint32_t numRows, uint32_t offset, short rowOffset, G4_SrcRegRegion* header, G4_DstRegRegion* resultRgn, G4_BB* bb, INST_LIST_ITER& instIt);

visa/SpillManagerGMRF.cpp

@@ -4831,6 +4831,8 @@ void GlobalRA::expandSpillIntrinsic(G4_BB* bb)
             auto header = inst->getSrc(0)->asSrcRegRegion();
             auto payload = inst->getSrc(1)->asSrcRegRegion();
             auto spillIt = instIt;
+            auto byteOff = inst->asSpillIntrinsic()->getOffsetInBytes() + numRows * numEltPerGRF<Type_UB>();
+            actualSpillSize = std::max(actualSpillSize, byteOff);
 
             auto rowOffset = payload->getRegOff();
             {
@@ -5005,6 +5007,8 @@ void GlobalRA::expandFillIntrinsic(G4_BB* bb)
             auto header = inst->getSrc(0)->asSrcRegRegion();
             auto resultRgn = inst->getDst();
             auto fillIt = instIt;
+            auto byteOff = inst->asFillIntrinsic()->getOffsetInBytes() + numRows * numEltPerGRF<Type_UB>();
+            actualSpillSize = std::max(actualSpillSize, byteOff);
 
             auto rowOffset = resultRgn->getRegOff();
             {