regsiter allocation support for NoMask WA

regsiter allocation support for NoMask WA

Author: jgu222

visa/GraphColor.cpp

@@ -9767,27 +9767,28 @@ int GlobalRA::coloringRegAlloc()
     }
 
 
-    //
-    // If the graph has stack calls, then add the caller-save/callee-save pseudo
-    // declares and code. This currently must be done after flag/addr RA due to
-    // the assumption about the location of the pseudo save/restore instructions
-    //
-    bool euWADone = false;
-    if (hasStackCall)
+    if (builder.hasFusedEUWA() && !builder.getIsPayload())
     {
-        if (builder.hasFusedEUWA() && !euWADone)
+        G4_BB* entryBB = (*kernel.fg.begin());
+        if (entryBB)
         {
-            G4_INST* euWAInst = builder.createEUWASpill(false);
-            G4_BB* entryBB = (*kernel.fg.begin());
             INST_LIST_ITER inst_it = entryBB->begin();
-            while ((*inst_it)->isLabel())
+            while (inst_it != entryBB->end() && (*inst_it)->isLabel())
             {
                 inst_it++;
             }
+            G4_INST* euWAInst = builder.createEUWASpill(false);
             entryBB->insertBefore(inst_it, euWAInst);
-            euWADone = true;
         }
+    }
 
+    //
+    // If the graph has stack calls, then add the caller-save/callee-save pseudo
+    // declares and code. This currently must be done after flag/addr RA due to
+    // the assumption about the location of the pseudo save/restore instructions
+    //
+    if (hasStackCall)
+    {
         addCallerSavePseudoCode();
 
         // Only GENX sub-graphs require callee-save code.
@@ -10194,19 +10195,6 @@ int GlobalRA::coloringRegAlloc()
                 bool success = spillGRF.insertSpillFillCode(&kernel, pointsToAnalysis);
                 nextSpillOffset = spillGRF.getNextOffset();
 
-                if (builder.hasFusedEUWA() && !euWADone)
-                {
-                    G4_INST * euWAInst = builder.createEUWASpill(false);
-                    G4_BB* entryBB = (*kernel.fg.begin());
-                    INST_LIST_ITER inst_it = entryBB->begin();
-                    while ((*inst_it)->isLabel())
-                    {
-                        inst_it++;
-                    }
-                    entryBB->insertBefore(inst_it, euWAInst);
-                    euWADone = true;
-                }
-
                 if (builder.hasScratchSurface() && !hasStackCall &&
                     (nextSpillOffset + globalScratchOffset) > SCRATCH_MSG_LIMIT)
                 {

visa/HWCaps.inc

@@ -785,7 +785,7 @@ SPDX-License-Identifier: MIT
     bool hasFusedEUWA() const
     {
         return ((getuint32Option(vISA_noMaskWA) & 0x3) > 0 ||
-                getOption(vISA_forceNoMaskWA));
+                getOption(vISA_forceNoMaskWA)) && !getOption(vISA_LinearScan) && hasFusedEU();
     }
 
     bool hasFusedEU() const

visa/LocalRA.cpp

@@ -2343,6 +2343,10 @@ void PhyRegsLocalRA::markPhyRegs(G4_Declare* topdcl)
     unsigned numwords = 0;
     unsigned int regnum = 0;
 
+    if (builder->hasFusedEUWA() && rvar->getPhyReg() == nullptr)
+    {
+        return;
+    }
     // Calculate number of physical registers required by this dcl
     if (numrows == 1)
     {