Fix the bug for SWSB dependence missing on undefined variable

cmp.eq  v5444, v5444
v5444 is undefined and will be assigned with any register.
As a result there may be RAW dependence in the SIMD control.
Current SWSB only track WAR WAW along the SIMD control.

Author: bcheng0127

visa/G4_Declare.h

@@ -93,6 +93,7 @@ class G4_Declare {
   uint16_t addrSpillFill : 1;
   uint16_t forceSpilled : 1;
   uint16_t exclusiveLoad : 1;
+  uint16_t isCmpUseOnly : 1;
 
   unsigned declId; // global decl id for this builder
 
@@ -330,6 +331,9 @@ class G4_Declare {
   void setPreDefinedVar(bool b) { PreDefinedVar = b; }
   bool isPreDefinedVar() const { return PreDefinedVar; }
 
+  void setIsCmpUseOnly(bool b) { isCmpUseOnly = b; }
+  bool getIsCmpUseOnly() const { return isCmpUseOnly; }
+
   unsigned getNumRegNeeded() const;
 
   void emit(std::ostream &output) const;

visa/G4_IR.cpp

@@ -5154,6 +5154,7 @@ G4_Declare::G4_Declare(const IR_Builder &builder, const char *n,
   addrSpillFill = false;
   forceSpilled = false;
   exclusiveLoad = false;
+  isCmpUseOnly = false;
   scopeID = 0;
 
   declId = (unsigned)dcllist.size();

visa/GraphColor.h

@@ -1405,6 +1405,7 @@ class GlobalRA {
   static const RAVarInfo defaultValues;
   std::vector<RAVarInfo> vars;
   std::vector<G4_Declare *> UndeclaredVars;
+  std::vector<G4_Declare *> UndefinedCmpVars;
 
   // fake declares for each GRF reg, used by HRA
   // note only GRFs that are used by LRA get a declare
@@ -1617,6 +1618,7 @@ class GlobalRA {
   void addSpillCodeInBB(G4_BB *bb) { BBsWithSpillCode.insert(bb); }
 
   void addUndefinedDcl(G4_Declare *dcl) { UndeclaredVars.push_back(dcl); }
+  void addUndefinedCmpDcl(G4_Declare *dcl) { UndefinedCmpVars.push_back(dcl); }
 
   bool isUndefinedDcl(const G4_Declare *dcl) const {
     return std::find(UndeclaredVars.begin(), UndeclaredVars.end(), dcl) !=
@@ -1689,6 +1691,7 @@ class GlobalRA {
       setBBId(dcl, UINT_MAX);
       resetLocalLR(dcl);
     }
+    UndefinedCmpVars.clear();
   }
 
   void clearLocalLiveRanges() {
@@ -2050,6 +2053,7 @@ class GlobalRA {
   void addCalleeSavePseudoCode();
   void addStoreRestoreToReturn();
   void storeCEInProlog();
+  void setUndefinedVarCmp();
   void markGraphBlockLocalVars();
   void verifyRA(LivenessAnalysis &liveAnalysis);
   void verifySpillFill();

visa/LocalRA.h

@@ -135,6 +135,7 @@ class LocalLiveRange {
 
   bool assigned;
   bool isSplit;
+  bool defined;
 
   IR_Builder &builder;
 
@@ -153,6 +154,7 @@ class LocalLiveRange {
     assigned = false;
     eot = false;
     isSplit = false;
+    defined = false;
 
     if (!builder.canWriteR0())
       addForbidden(0);
@@ -224,6 +226,9 @@ class LocalLiveRange {
   void markSplit() { isSplit = true; }
   bool getSplit() const { return isSplit; }
 
+  void markDefined() { defined = true; }
+  bool isDefined() const { return defined; }
+
   void addForbidden(unsigned int f) { forbiddenGRFs.insert(f); }
   std::unordered_set<unsigned int> &getForbidden() { return forbiddenGRFs; }
 };

visa/LocalScheduler/SWSB_G4IR.cpp

@@ -1043,6 +1043,8 @@ SBFootprint *G4_BB_SB::getFootprintForGRF(G4_Operand *opnd,
   GenPrecision precision = GenPrecision::INVALID;
   bool isFcvtByteType = false;
   bool isPrecision = false;
+  bool isCmpUseOnly = inst->opcode() == G4_cmp && opnd_num == Opnd_src0 &&
+                      opnd->getTopDcl()->getIsCmpUseOnly();
 
   if (inst->opcode() == G4_fcvt &&
       (IS_BTYPE(type) ||
@@ -1175,7 +1177,7 @@ SBFootprint *G4_BB_SB::getFootprintForGRF(G4_Operand *opnd,
                 SBFootprint(GRF_T, precision, LB, RB, inst, isFcvtByteType)
           : new (allocedMem)
                 SBFootprint(GRF_T, type, LB, RB, inst, isFcvtByteType);
-
+  footprint->isCmpUseOnly = isCmpUseOnly;
   return footprint;
 }
 
@@ -5600,6 +5602,9 @@ void G4_BB_SB::setSendOpndMayKilled(SBNODE_VECT &SBNodes, PointsToAnalysis &p,
           send_may_kill.src.set(globalID);
         } else if (dep == RAW) {
           send_may_kill.dst.set(globalID);
+          if (curFootprint->isCmpUseOnly) {
+            send_WAW_may_kill.set(globalID);
+          }
           // Exclusive WAW has no overlap
         } else if (dep == WAW && (isSend || !isDclExclusiveLoad(
                                                 nodeInfo.topDeclare, topDcl))) {
@@ -8001,9 +8006,20 @@ void SWSB::addGlobalDependence(unsigned globalSendNum,
           unsigned short internalOffset = 0;
           bool hasOverlap =
               curFootprint->hasOverlap(liveFootprint, internalOffset);
+          DepType dep = getDepForOpnd(liveOpnd, curOpnd);
 
+          // Following special dependence checking is to handle following case
+          // in SIMD control follow:
+          // cmp.eq  v5444, v5444
+          // where v5444 is undefined and can be assigned with any register. As
+          // a result there may be RAW dependence in the SIMD control follow.
+          // For this special case, we treat it as a WAW dependence by checking
+          // if it's uninitialized declare used in cmp
+          if (!afterWrite && liveOpnd == Opnd_dst &&
+              curFootprint->isCmpUseOnly) {
+            dep = getDepForOpnd(liveOpnd, Opnd_dst);
+          }
           // Find DEP type
-          DepType dep = getDepForOpnd(liveOpnd, curOpnd);
           if (!hasOverlap && dep == RAW) {
             hasOverlap =
                 sb_bb->hasExtraOverlap(liveInst, curInst, liveFootprint,

visa/LocalScheduler/SWSB_G4IR.h

@@ -141,7 +141,8 @@ struct SBFootprint {
   const unsigned short RightB;
   unsigned short offset = 0;
   bool isPrecision = false;
-  bool isFcvtByteType = false;;
+  bool isFcvtByteType = false;
+  bool isCmpUseOnly = false;
   G4_INST *inst;
 
   // FIXME: The choice of C-style linked list seems suspect given that there are

visa/RegAlloc.cpp

@@ -1923,6 +1923,23 @@ void GlobalRA::markBlockLocalVars() {
          it++) {
       G4_INST *inst = *it;
 
+      // Chjeck if there is undefine variable used in CMP instruction, which is
+      // used to detect the execution mask.
+      //     cmp.eq (M1, 16) P12 V0147(0,0)<0;1,0> V0147(0,0)<0;1,0>
+      if (inst->opcode() == G4_cmp) {
+        const bool isModEq =
+            inst->getCondMod() && inst->getCondMod()->getMod() == Mod_e;
+        const bool isNullDst = !inst->getDst() || inst->hasNULLDst();
+        const bool isSrc0SameAsSrc1 = inst->getSrc(0)->asSrcRegRegion() &&
+                                      inst->getSrc(1)->asSrcRegRegion() &&
+                                      *inst->getSrc(0)->asSrcRegRegion() ==
+                                          *inst->getSrc(1)->asSrcRegRegion();
+        if (isModEq && isNullDst && isSrc0SameAsSrc1) {
+          G4_Declare *topdcl = GetTopDclFromRegRegion(inst->getSrc(0));
+          addUndefinedCmpDcl(topdcl);
+        }
+      }
+
       // Track direct dst references.
 
       G4_DstRegRegion *dst = inst->getDst();
@@ -1945,6 +1962,7 @@ void GlobalRA::markBlockLocalVars() {
               lr->setFirstRef(inst, 0);
             }
             lr->recordRef(bb);
+            lr->markDefined();
             recordRef(topdcl);
           }
         }
@@ -2087,6 +2105,20 @@ bool GlobalRA::canSkipFDE() const {
   return !kernel.fg.getHasStackCalls() && kernel.getOption(vISA_skipFDE);
 }
 
+void GlobalRA::setUndefinedVarCmp() {
+  // Iterate over all dcls and remove those with 0
+  // ref count and not addressed. This is done only for
+  // GRF dcls.
+
+  // Propagate top dcl info to aliases
+  for (auto dcl : UndefinedCmpVars) {
+    LocalLiveRange *lr = getLocalLR(dcl);
+    if (!lr->isDefined()) {
+      dcl->setIsCmpUseOnly(true);
+    }
+  }
+}
+
 //
 // Mark block local (temporary) variables.
 //
@@ -2097,6 +2129,9 @@ void GlobalRA::markGraphBlockLocalVars() {
   // Create live ranges and record the reference info
   markBlockLocalVars();
 
+  // Set undefined variable used in cmp
+  setUndefinedVarCmp();
+
   VISA_DEBUG_VERBOSE({
     std::cout << "\t--LOCAL VARIABLES--\n";
     for (auto dcl : kernel.Declares) {