fixup and solve the dst/src overlap issue caused by spill

Author: bcheng0127

visa/GraphColor.cpp

@@ -9763,7 +9763,8 @@ int GlobalRA::coloringRegAlloc()
                     spillRegSize,
                     indrSpillRegSize,
                     enableSpillSpaceCompression,
-                    useScratchMsgForSpill);
+                    useScratchMsgForSpill,
+                    builder.avoidDstSrcOverlap());
 
                 bool success = spillGRF.insertSpillFillCode(&kernel, pointsToAnalysis);
                 nextSpillOffset = spillGRF.getNextOffset();
@@ -9887,7 +9888,6 @@ int GlobalRA::coloringRegAlloc()
     computePhyReg();
 
     stopTimer(TimerID::GRF_GLOBAL_RA);
-
     //
     // Report failure to allocate due to excessive register pressure.
     //

visa/LinearScanRA.cpp

@@ -1031,7 +1031,8 @@ int LinearScanRA::linearScanRA()
                 &l,
                 &spillLRs,
                 enableSpillSpaceCompression,
-                useScratchMsgForSpill);
+                useScratchMsgForSpill,
+                builder.avoidDstSrcOverlap());
 
             spillGRF.spillLiveRanges(&kernel);
             nextSpillOffset = spillGRF.getNextOffset();;

visa/SpillManagerGMRF.cpp

@@ -143,7 +143,8 @@ SpillManagerGRF::SpillManagerGRF(
     unsigned spillRegSize,
     unsigned indrSpillRegSize,
     bool enableSpillSpaceCompression,
-    bool useScratchMsg)
+    bool useScratchMsg,
+    bool avoidDstSrcOverlap)
     : gra(g)
     , builder_(g.kernel.fg.builder)
     , varIdCount_(varIdCount)
@@ -158,6 +159,7 @@ SpillManagerGRF::SpillManagerGRF(
     , spillIntf_(intf)
     , mem_(1024)
     , useScratchMsg_(useScratchMsg)
+    , avoidDstSrcOverlap_(avoidDstSrcOverlap)
 {
     const unsigned size = sizeof(unsigned) * varIdCount;
     spillRangeCount_ = (unsigned*)allocMem(size);
@@ -194,7 +196,8 @@ SpillManagerGRF::SpillManagerGRF(
     const LivenessAnalysis* lvInfo,
     LSLR_LIST* spilledLSLRs,
     bool enableSpillSpaceCompression,
-    bool useScratchMsg)
+    bool useScratchMsg,
+    bool avoidDstSrcOverlap)
     : gra(g)
     , builder_(g.kernel.fg.builder)
     , varIdCount_(varIdCount)
@@ -206,6 +209,7 @@ SpillManagerGRF::SpillManagerGRF(
     , failSafeSpill_(false)
     , mem_(1024)
     , useScratchMsg_(useScratchMsg)
+    , avoidDstSrcOverlap_(avoidDstSrcOverlap)
 {
     const unsigned size = sizeof(unsigned) * varIdCount;
     spillRangeCount_ = (unsigned*)allocMem(size);
@@ -2950,12 +2954,13 @@ void SpillManagerGRF::insertSpillRangeCode(
         initMWritePayload(
             spillRangeDcl, mRangeDcl, spilledRegion, execSize);
 
+
         {
             spillSendInst = createSpillSendInstr(
                 spillRangeDcl, mRangeDcl, spilledRegion, execSize, spillSendOption);
         }
 
-        if (failSafeSpill_)
+        if (failSafeSpill_ && !avoidDstSrcOverlap_)
         {
             spillRegOffset_ = spillRegStart_;
         }

visa/SpillManagerGMRF.h

@@ -89,7 +89,8 @@ class SpillManagerGRF
         unsigned                 spillRegSize,
         unsigned                 indrSpillRegSize,
         bool                     enableSpillSpaceCompression,
-        bool                     useScratchMsg
+        bool                     useScratchMsg,
+        bool                     avoidDstSrcOverlap
     );
 
     SpillManagerGRF(
@@ -99,7 +100,9 @@ class SpillManagerGRF
         const LivenessAnalysis* lvInfo,
         LSLR_LIST* spilledLSLRs,
         bool enableSpillSpaceCompression,
-        bool useScratchMsg);
+        bool useScratchMsg,
+        bool avoidDstSrcOverlap
+    );
 
     ~SpillManagerGRF() {}
 
@@ -638,7 +641,7 @@ class SpillManagerGRF
     int globalScratchOffset;
 
     bool useScratchMsg_;
-
+    bool avoidDstSrcOverlap_;
     // spilled declares that represent a scalar immediate (created due to encoding restrictions)
     // We rematerialize the immediate value instead of spill/fill them
     std::unordered_map<G4_Declare*, G4_Imm*> scalarImmSpill;