GP-2980 Analysis for unrolled switch guards

Ghidra/Features/Decompiler/src/decompile/cpp/block.cc

@@ -2616,6 +2616,70 @@ bool BlockBasic::noInterveningStatement(PcodeOp *first,int4 path,PcodeOp *last)
   return false;
 }
 
+/// If there exists a CPUI_MULTIEQUAL PcodeOp in the given basic block that takes this exact list of Varnodes
+/// as its inputs, return that PcodeOp. Otherwise return null.
+/// \param varArray is the exact list of Varnodes
+/// \param bl is the basic block
+/// \return the MULTIEQUAL or null
+PcodeOp *BlockBasic::findMultiequal(const vector<Varnode *> &varArray)
+
+{
+  Varnode *vn = varArray[0];
+  PcodeOp *op;
+  list<PcodeOp *>::const_iterator iter = vn->beginDescend();
+  for(;;) {
+    op = *iter;
+    if (op->code() == CPUI_MULTIEQUAL && op->getParent() == this)
+      break;
+    ++iter;
+    if (iter == vn->endDescend())
+      return (PcodeOp *)0;
+  }
+  for(int4 i=0;i<op->numInput();++i) {
+    if (op->getIn(i) != varArray[i])
+      return (PcodeOp *)0;
+  }
+  return op;
+}
+
+/// Each Varnode must be defined by a PcodeOp with the same OpCode.  The Varnode, within the array, is replaced
+/// with the input Varnode in the indicated slot.
+/// \param varArray is the given array of Varnodes
+/// \param slot is the indicated slot
+/// \return \true if all the Varnodes are defined in the same way
+bool BlockBasic::liftVerifyUnroll(vector<Varnode *> &varArray,int4 slot)
+
+{
+  OpCode opc;
+  Varnode *cvn;
+  Varnode *vn = varArray[0];
+  if (!vn->isWritten()) return false;
+  PcodeOp *op = vn->getDef();
+  opc = op->code();
+  if (op->numInput() == 2) {
+    cvn = op->getIn(1-slot);
+    if (!cvn->isConstant()) return false;
+  }
+  else
+    cvn = (Varnode *)0;
+  varArray[0] = op->getIn(slot);
+  for(int4 i=1;i<varArray.size();++i) {
+    vn = varArray[i];
+    if (!vn->isWritten()) return false;
+    op = vn->getDef();
+    if (op->code() != opc) return false;
+
+    if (cvn != (Varnode *)0) {
+      Varnode *cvn2 = op->getIn(1-slot);
+      if (!cvn2->isConstant()) return false;
+      if (cvn->getSize() != cvn2->getSize()) return false;
+      if (cvn->getOffset() != cvn2->getOffset()) return false;
+    }
+    varArray[i] = op->getIn(slot);
+  }
+  return true;
+}
+
 void BlockCopy::printHeader(ostream &s) const
 
 {

Ghidra/Features/Decompiler/src/decompile/cpp/block.hh

@@ -413,6 +413,8 @@ public:
   list<PcodeOp *>::const_iterator endOp(void) const { return op.end(); }	///< Return an iterator to the end of the PcodeOps
   bool emptyOp(void) const { return op.empty(); }		///< Return \b true if \b block contains no operations
   static bool noInterveningStatement(PcodeOp *first,int4 path,PcodeOp *last);
+  PcodeOp *findMultiequal(const vector<Varnode *> &varArray);		///< Find MULTIEQUAL with given inputs
+  static bool liftVerifyUnroll(vector<Varnode *> &varArray,int4 slot);	///< Verify given Varnodes are defined with same PcodeOp
 };
 
 /// \brief This class is used to mirror the BlockBasic objects in the fixed control-flow graph for a function

Ghidra/Features/Decompiler/src/decompile/cpp/jumptable.cc

@@ -502,6 +502,46 @@ uintb JumpBasic::backup2Switch(Funcdata *fd,uintb output,Varnode *outvn,Varnode
   return output;
 }
 
+/// If the Varnode has a restricted range due to masking via INT_AND, the maximum value of this range is returned.
+/// Otherwise, 0 is returned, indicating that the Varnode can take all possible values.
+/// \param vn is the given Varnode
+/// \return the maximum value or 0
+uintb JumpBasic::getMaxValue(Varnode *vn)
+
+{
+  uintb maxValue = 0;		// 0 indicates maximum possible value
+  if (!vn->isWritten())
+    return maxValue;
+  PcodeOp *op = vn->getDef();
+  if (op->code() == CPUI_INT_AND) {
+    Varnode *constvn = op->getIn(1);
+    if (constvn->isConstant()) {
+      maxValue = coveringmask( constvn->getOffset() );
+      maxValue = (maxValue + 1) & calc_mask(vn->getSize());
+    }
+  }
+  else if (op->code() == CPUI_MULTIEQUAL) {	// Its possible the AND is duplicated across multiple blocks
+    int4 i;
+    for(i=0;i<op->numInput();++i) {
+      Varnode *subvn = op->getIn(i);
+      if (!subvn->isWritten()) break;
+      PcodeOp *andOp = subvn->getDef();
+      if (andOp->code() != CPUI_INT_AND) break;
+      Varnode *constvn = andOp->getIn(1);
+      if (!constvn->isConstant()) break;
+      if (maxValue < constvn->getOffset())
+	maxValue = constvn->getOffset();
+    }
+    if (i == op->numInput()) {
+      maxValue = coveringmask( maxValue );
+      maxValue = (maxValue + 1) & calc_mask(vn->getSize());
+    }
+    else
+      maxValue = 0;
+  }
+  return maxValue;
+}
+
 /// \brief Calculate the initial set of Varnodes that might be switch variables
 ///
 /// Paths that terminate at the given PcodeOp are calculated and organized
@@ -566,7 +606,8 @@ static bool matching_constants(Varnode *vn1,Varnode *vn2)
 /// \param path is the specific branch to take from the CBRANCH to reach the switch
 /// \param rng is the range of values causing the switch path to be taken
 /// \param v is the Varnode holding the value controlling the CBRANCH
-GuardRecord::GuardRecord(PcodeOp *bOp,PcodeOp *rOp,int4 path,const CircleRange &rng,Varnode *v)
+/// \param unroll is \b true if the guard is duplicated across multiple blocks
+GuardRecord::GuardRecord(PcodeOp *bOp,PcodeOp *rOp,int4 path,const CircleRange &rng,Varnode *v,bool unr)
 
 {
   cbranch = bOp;
@@ -575,6 +616,7 @@ GuardRecord::GuardRecord(PcodeOp *bOp,PcodeOp *rOp,int4 path,const CircleRange &
   range = rng;
   vn = v;
   baseVn = quasiCopy(v,bitsPreserved);		// Look for varnode whose bits are copied
+  unrolled = unr;
 }
 
 /// \brief Determine if \b this guard applies to the given Varnode
@@ -1020,7 +1062,12 @@ void JumpBasic::analyzeGuards(BlockBasic *bl,int4 pathout)
     else {
       pathout = -1;		// Make sure not to use pathout next time around
       for(;;) {
-	if (bl->sizeIn() != 1) return; // Assume only 1 path to switch
+	if (bl->sizeIn() != 1) {
+	  if (bl->sizeIn() > 1)
+	    checkUnrolledGuard(bl, maxpullback, usenzmask);
+	  return;
+	}
+	// Only 1 flow path to the switch
 	prevbl = (BlockBasic *)bl->getIn(0);
 	if (prevbl->sizeOut() != 1) break; // Is it possible to deviate from switch path in this block
 	bl = prevbl;		// If not, back up to next block
@@ -1077,17 +1124,7 @@ void JumpBasic::calcRange(Varnode *vn,CircleRange &rng) const
   else if (vn->isWritten() && vn->getDef()->isBoolOutput())
     rng = CircleRange(0,2,1,1);	// Only 0 or 1 possible
   else {			// Should we go ahead and use nzmask in all cases?
-    uintb maxValue = 0;		// Every possible value
-    if (vn->isWritten()) {
-      PcodeOp *andop = vn->getDef();
-      if (andop->code() == CPUI_INT_AND) {
-	Varnode *constvn = andop->getIn(1);
-	if (constvn->isConstant()) {
-	  maxValue = coveringmask( constvn->getOffset() );
-	  maxValue = (maxValue + 1) & calc_mask(vn->getSize());
-	}
-      }
-    }
+    uintb maxValue = getMaxValue(vn);
     stride = getStride(vn);
     rng = CircleRange(0,maxValue,vn->getSize(),stride);
   }
@@ -1203,8 +1240,9 @@ void JumpBasic::markFoldableGuards(void)
   int4 bitsPreserved;
   Varnode *baseVn = GuardRecord::quasiCopy(vn, bitsPreserved);
   for(int4 i=0;i<selectguards.size();++i) {
-    if (selectguards[i].valueMatch(vn,baseVn,bitsPreserved)==0) {
-      selectguards[i].clear();		// Indicate this is not a true guard
+    GuardRecord &guardRecord(selectguards[i]);
+    if (guardRecord.valueMatch(vn,baseVn,bitsPreserved)==0 || guardRecord.isUnrolled()) {
+      guardRecord.clear();		// Indicate this guard was not used or should not be folded
     }
   }
 }
@@ -1243,6 +1281,75 @@ bool JumpBasic::flowsOnlyToModel(Varnode *vn,PcodeOp *trailOp)
   return true;
 }
 
+/// All CBRANCHs in addition to flowing to the given block, must also flow to another common block,
+/// and each boolean value must select between the given block and the common block in the same way.
+/// If this flow exists, \b true is returned and the boolean Varnode inputs to each CBRANCH are passed back.
+/// \param varArray will hold the input Varnodes being passed back
+/// \param bl is the given block
+/// \return \b true if the common CBRANCH flow exists across all incoming blocks
+bool JumpBasic::checkCommonCbranch(vector<Varnode *> &varArray,BlockBasic *bl)
+
+{
+  BlockBasic *curBlock = (BlockBasic *)bl->getIn(0);
+  PcodeOp *op  = curBlock->lastOp();
+  if (op == (PcodeOp *)0 || op->code() != CPUI_CBRANCH)
+    return false;
+  int4 outslot = bl->getInRevIndex(0);
+  bool isOpFlip = op->isBooleanFlip();
+  varArray.push_back(op->getIn(1));	// Pass back boolean input to CBRANCH
+  for(int4 i=1;i<bl->sizeIn();++i) {
+    curBlock = (BlockBasic *)bl->getIn(i);
+    op = curBlock->lastOp();
+    if (op == (PcodeOp *)0 || op->code() != CPUI_CBRANCH)
+      return false;				// All blocks must end with CBRANCH
+    if (op->isBooleanFlip() != isOpFlip)
+      return false;
+    if (outslot != bl->getInRevIndex(i))
+      return false;				// Boolean value must have some meaning
+    varArray.push_back(op->getIn(1));		// Pass back boolean input to CBRANCH
+  }
+  return true;
+}
+
+/// \brief Check for a guard that has been unrolled across multiple blocks
+///
+/// A guard calculation can be duplicated across multiple blocks that all branch to the basic block
+/// performing the final BRANCHIND.  In this case, the switch variable is also duplicated across multiple Varnodes
+/// that are all inputs to a MULTIEQUAL whose output is used for the final BRANCHIND calculation.  This method
+/// looks for this situation and creates a GuardRecord associated with this MULTIEQUAL output.
+/// \param bl is the basic block on the path to the switch with multiple incoming flows
+/// \param maxpullback is the maximum number of times to pull back from the guard CBRANCH to the putative switch variable
+/// \param usenzmask is \b true if the NZMASK should be used as part of the pull-back operation
+void JumpBasic::checkUnrolledGuard(BlockBasic *bl,int4 maxpullback,bool usenzmask)
+
+{
+  vector<Varnode *> varArray;
+  if (!checkCommonCbranch(varArray,bl))
+    return;
+  int4 indpath = bl->getInRevIndex(0);
+  bool toswitchval = (indpath == 1);
+  PcodeOp *cbranch = ((BlockBasic *)bl->getIn(0))->lastOp();
+  if (cbranch->isBooleanFlip())
+    toswitchval = !toswitchval;
+  CircleRange rng(toswitchval);
+  int4 indpathstore = bl->getIn(0)->getFlipPath() ? 1-indpath : indpath;
+  PcodeOp *readOp = cbranch;
+  for(int4 j=0;j<maxpullback;++j) {
+    PcodeOp *multiOp = bl->findMultiequal(varArray);
+    if (multiOp != (PcodeOp *)0) {
+      selectguards.push_back(GuardRecord(cbranch,readOp,indpathstore,rng,multiOp->getOut(),true));
+    }
+    Varnode *markup;		// Throw away markup information
+    Varnode *vn = varArray[0];
+    if (!vn->isWritten()) break;
+    PcodeOp *readOp = vn->getDef();
+    vn = rng.pullBack(readOp,&markup,usenzmask);
+    if (vn == (Varnode *)0) break;
+    if (rng.isEmpty()) break;
+    if (!BlockBasic::liftVerifyUnroll(varArray, readOp->getSlot(vn))) break;
+  }
+}
+
 bool JumpBasic::foldInOneGuard(Funcdata *fd,GuardRecord &guard,JumpTable *jump)
 
 {

Ghidra/Features/Decompiler/src/decompile/cpp/jumptable.hh

@@ -141,13 +141,15 @@ class JumpTable;
 class GuardRecord {
   PcodeOp *cbranch;		///< PcodeOp CBRANCH the branches around the switch
   PcodeOp *readOp;		///< The immediate PcodeOp causing the restriction
-  int4 indpath;			///< Specific CBRANCH path going to the switch
-  CircleRange range;		///< Range of values causing the CBRANCH to take the path to the switch
   Varnode *vn;			///< The Varnode being restricted
   Varnode *baseVn;		///< Value being (quasi)copied to the Varnode
+  int4 indpath;			///< Specific CBRANCH path going to the switch
   int4 bitsPreserved;		///< Number of bits copied (all other bits are zero)
+  CircleRange range;		///< Range of values causing the CBRANCH to take the path to the switch
+  bool unrolled;		///< \b true if guarding CBRANCH is duplicated across multiple blocks
 public:
-  GuardRecord(PcodeOp *bOp,PcodeOp *rOp,int4 path,const CircleRange &rng,Varnode *v);	///< Constructor
+  GuardRecord(PcodeOp *bOp,PcodeOp *rOp,int4 path,const CircleRange &rng,Varnode *v,bool unr=false);	///< Constructor
+  bool isUnrolled(void) const { return unrolled; }	///< Is \b this guard duplicated across multiple blocks
   PcodeOp *getBranch(void) const { return cbranch; }	///< Get the CBRANCH associated with \b this guard
   PcodeOp *getReadOp(void) const { return readOp; }	///< Get the PcodeOp immediately causing the restriction
   int4 getPath(void) const { return indpath; }		///< Get the specific path index going towards the switch
@@ -364,6 +366,7 @@ protected:
   static bool ispoint(Varnode *vn);	///< Is it possible for the given Varnode to be a switch variable?
   static int4 getStride(Varnode *vn);	///< Get the step/stride associated with the Varnode
   static uintb backup2Switch(Funcdata *fd,uintb output,Varnode *outvn,Varnode *invn);
+  static uintb getMaxValue(Varnode *vn);	///< Get maximum value associated with the given Varnode
   void findDeterminingVarnodes(PcodeOp *op,int4 slot);
   void analyzeGuards(BlockBasic *bl,int4 pathout);
   void calcRange(Varnode *vn,CircleRange &rng) const;
@@ -372,6 +375,8 @@ protected:
   void markFoldableGuards();
   void markModel(bool val);		///< Mark (or unmark) all PcodeOps involved in the model
   bool flowsOnlyToModel(Varnode *vn,PcodeOp *trailOp);	///< Check if the given Varnode flows to anything other than \b this model
+  bool checkCommonCbranch(vector<Varnode *> &varArray,BlockBasic *bl);	///< Check that all incoming blocks end with a CBRANCH
+  void checkUnrolledGuard(BlockBasic *bl,int4 maxpullback,bool usenzmask);
 
   /// \brief Eliminate the given guard to \b this switch
   ///