diff --git a/src/aig/gia/giaIf.c b/src/aig/gia/giaIf.c
index e733bccad..682d7b171 100644
--- a/src/aig/gia/giaIf.c
+++ b/src/aig/gia/giaIf.c
@@ -1190,7 +1190,7 @@ int Gia_ManFromIfLogicCreateLutSpecialJ( Gia_Man_t * pNew, word * pRes, Vec_Int_
 {
     word Truth;
     int i, iObjLit1, iObjLit2, iObjLit3;
-    word z = If_CutPerformDeriveJ( NULL, (unsigned *)pRes, Vec_IntSize(vLeaves), Vec_IntSize(vLeaves), NULL, 1 );
+    word z = If_CutPerformDeriveJ( NULL, (unsigned *)pRes, Vec_IntSize(vLeaves), Vec_IntSize(vLeaves), NULL, 1, 0 );
     assert( z != 0 );
     if ( ((z >> 63) & 1) == 0 )
     {   
@@ -2135,7 +2135,7 @@ void Gia_ManConfigPrint2( unsigned char * pConfigData, int nLeaves )
   SeeAlso     []
 
 ***********************************************************************/
-void Gia_ManFromIfGetConfig2( Vec_Str_t * vConfigs2, If_Man_t * pIfMan, word * pTruth, int nLeaves )
+void Gia_ManFromIfGetConfig2( Vec_Str_t * vConfigs2, If_Man_t * pIfMan, word * pTruth, int nLeaves, int fDelay )
 {
     int i, CellId;
     int startPos = Vec_StrSize(vConfigs2);
@@ -2143,13 +2143,18 @@ void Gia_ManFromIfGetConfig2( Vec_Str_t * vConfigs2, If_Man_t * pIfMan, word * p
     // Determine cell type based on the number of leaves and configuration
     if ( nLeaves <= 4 ) // 7 bytes = 1 byte CellId + 4 bytes mapping + 2 bytes truth table
     {
+        word z = If_CutPerformDeriveJ( pIfMan, (unsigned *)pTruth, nLeaves, nLeaves, NULL, 1, fDelay );
+        int fHavePerm = (z != 0) && ((z & ABC_CONST(0x4000000000000000)) != 0);
         // Cell type 0: Simple LUT4
         CellId = 0;
         // Write CellId
         Vec_StrPush( vConfigs2, (char)CellId );
         // Write mapping
         for ( i = 0; i < nLeaves; i++ )
-            Vec_StrPush( vConfigs2, 2+i );
+        {
+            int v = fHavePerm ? (int)((z >> (2 * i)) & 3) : i;
+            Vec_StrPush( vConfigs2, 2 + v );
+        }
         for ( ; i < 4; i++ )
             Vec_StrPush( vConfigs2, 0 );
         // Write truth table (16 bits for LUT4)
@@ -2161,7 +2166,7 @@ void Gia_ManFromIfGetConfig2( Vec_Str_t * vConfigs2, If_Man_t * pIfMan, word * p
     }
     else // 12 bytes = 1 byte CellId + 7 bytes mapping + 4 bytes truth tables
     {
-        word z = If_CutPerformDeriveJ( pIfMan, (unsigned *)pTruth, nLeaves, nLeaves, NULL, 1 );
+        word z = If_CutPerformDeriveJ( pIfMan, (unsigned *)pTruth, nLeaves, nLeaves, NULL, 1, fDelay );
         //Gia_ManConfigPrint( 0, z, nLeaves );
         if ( ((z >> 63) & 1) == 0 )
         {
@@ -2551,7 +2556,16 @@ Gia_Man_t * Gia_ManFromIfLogic( If_Man_t * pIfMan )
                             Abc_TtFlip( pTruth, Abc_TtWordNum(pCutBest->nLeaves), k );
                     if ( Abc_LitIsCompl(pIfObj->iCopy) ^ pCutBest->fCompl )
                         Abc_TtNot( pTruth, Abc_TtWordNum(pCutBest->nLeaves) );
-                    Gia_ManFromIfGetConfig2( vConfigs2, pIfMan, pTruth, pCutBest->nLeaves );
+                    if ( pIfMan->pPars->fDelayOptCell )
+                    {
+                        pIfMan->nCutLeavesCur = pCutBest->nLeaves;
+                        If_CutForEachLeaf( pIfMan, pCutBest, pIfLeaf, k )
+                        {
+                            pIfMan->pCutLeavesCur[k] = pIfLeaf->Id;
+                            pIfMan->pCutLeafArrCur[k] = If_ObjCutBest(pIfLeaf)->Delay;
+                        }
+                    }
+                    Gia_ManFromIfGetConfig2( vConfigs2, pIfMan, pTruth, pCutBest->nLeaves, pIfMan->pPars->fDelayOptCell );
                 }
             }
             else
@@ -3337,4 +3351,3 @@ Gia_Man_t * Gia_ManDupUnhashMapping( Gia_Man_t * p )
 
 
 ABC_NAMESPACE_IMPL_END
-
diff --git a/src/aig/gia/giaSpeedup.c b/src/aig/gia/giaSpeedup.c
index 4212c8966..e5caba8fc 100644
--- a/src/aig/gia/giaSpeedup.c
+++ b/src/aig/gia/giaSpeedup.c
@@ -33,6 +33,54 @@ ABC_NAMESPACE_IMPL_START
 ///                     FUNCTION DEFINITIONS                         ///
 ////////////////////////////////////////////////////////////////////////
 
+static int Gia_ManConfig2GetBytePos( Gia_Man_t * p, int iObj )
+{
+    int iLut, bytePos = 0;
+    if ( p == NULL || p->vConfigs2 == NULL )
+        return -1;
+    Gia_ManForEachLut( p, iLut )
+    {
+        unsigned char CellId;
+        if ( bytePos >= Vec_StrSize(p->vConfigs2) )
+            return -1;
+        if ( iLut == iObj )
+            return bytePos;
+        CellId = (unsigned char)Vec_StrEntry( p->vConfigs2, bytePos );
+        if ( CellId == 0 )
+            bytePos += 7;
+        else if ( CellId == 1 )
+            bytePos += 12;
+        else if ( CellId == 2 )
+            bytePos += 14;
+        else
+            return -1;
+    }
+    return -1;
+}
+static int Gia_ManConfig2DerivePinDelays( Gia_Man_t * p, int iObj, If_LibCell_t * pCellLib, int * pPinDelay, int nLutSize )
+{
+    int bytePos, i, nPins;
+    unsigned char CellId;
+    if ( pCellLib == NULL || pPinDelay == NULL || nLutSize < 1 || nLutSize > 32 )
+        return 0;
+    for ( i = 0; i < nLutSize; i++ )
+        pPinDelay[i] = 1;
+    bytePos = Gia_ManConfig2GetBytePos( p, iObj );
+    if ( bytePos < 0 || bytePos >= Vec_StrSize(p->vConfigs2) )
+        return 0;
+    CellId = (unsigned char)Vec_StrEntry( p->vConfigs2, bytePos );
+    if ( CellId >= pCellLib->nCellNum )
+        return 0;
+    nPins = Abc_MinInt( pCellLib->nCellInputs[CellId], 9 );
+    for ( i = 0; i < nPins; i++ )
+    {
+        int v = (unsigned char)Vec_StrEntry( p->vConfigs2, bytePos + 1 + i );
+        if ( v >= 2 && v < 2 + nLutSize )
+            pPinDelay[v - 2] = pCellLib->pCellPinDelays[CellId][i];
+    }
+    return 1;
+}
+
 /**Function*************************************************************
 
   Synopsis    [Sorts the pins in the decreasing order of delays.]
@@ -113,7 +161,7 @@ float Gia_ObjComputeArrival( Gia_Man_t * p, int iObj, int fUseSorting )
     If_LibLut_t * pLutLib = (If_LibLut_t *)p->pLutLib;
     If_LibCell_t * pCellLib = (If_LibCell_t *)p->pCellLib;
     Gia_Obj_t * pObj = Gia_ManObj( p, iObj );
-    int k, iFanin, pPinPerm[32];
+    int k, iFanin, pPinPerm[32], pPinDelay[32];
     float pPinDelays[32];
     float tArrival, * pDelays;
     if ( Gia_ObjIsCi(pObj) )
@@ -132,15 +180,17 @@ float Gia_ObjComputeArrival( Gia_Man_t * p, int iObj, int fUseSorting )
     {
         // Handle cell library delays (use integer delays directly)
         int nLutSize = Gia_ObjLutSize(p, iObj);
+        int fHaveCfg2 = Gia_ManConfig2DerivePinDelays( p, iObj, pCellLib, pPinDelay, nLutSize );
         // Find matching cell (simple approach: use first cell with enough inputs)
         int cellId = -1;
         int i;
-        for ( i = 0; i < pCellLib->nCellNum; i++ )
-            if ( pCellLib->nCellInputs[i] >= nLutSize )
-            {
-                cellId = i;
-                break;
-            }
+        if ( !fHaveCfg2 )
+            for ( i = 0; i < pCellLib->nCellNum; i++ )
+                if ( pCellLib->nCellInputs[i] >= nLutSize )
+                {
+                    cellId = i;
+                    break;
+                }
         if ( cellId >= 0 )
         {
             // Use cell delays as integers from the library
@@ -151,6 +201,15 @@ float Gia_ObjComputeArrival( Gia_Man_t * p, int iObj, int fUseSorting )
                     tArrival = Gia_ObjTimeArrival(p, iFanin) + delay;
             }
         }
+        else if ( fHaveCfg2 )
+        {
+            Gia_LutForEachFanin( p, iObj, iFanin, k )
+            {
+                float delay = (float)pPinDelay[k];
+                if ( tArrival < Gia_ObjTimeArrival(p, iFanin) + delay )
+                    tArrival = Gia_ObjTimeArrival(p, iFanin) + delay;
+            }
+        }
         else
         {
             // Fall back to default delay if no matching cell
@@ -204,7 +263,7 @@ float Gia_ObjPropagateRequired( Gia_Man_t * p, int iObj, int fUseSorting )
 {
     If_LibLut_t * pLutLib = (If_LibLut_t *)p->pLutLib;
     If_LibCell_t * pCellLib = (If_LibCell_t *)p->pCellLib;
-    int k, iFanin, pPinPerm[32];
+    int k, iFanin, pPinPerm[32], pPinDelay[32];
     float pPinDelays[32];
     float tRequired = 0.0; // Suppress "might be used uninitialized"
     float * pDelays;
@@ -223,15 +282,17 @@ float Gia_ObjPropagateRequired( Gia_Man_t * p, int iObj, int fUseSorting )
     {
         // Handle cell library delays (use integer delays directly)
         int nLutSize = Gia_ObjLutSize(p, iObj);
+        int fHaveCfg2 = Gia_ManConfig2DerivePinDelays( p, iObj, pCellLib, pPinDelay, nLutSize );
         // Find matching cell (simple approach: use first cell with enough inputs)
         int cellId = -1;
         int i;
-        for ( i = 0; i < pCellLib->nCellNum; i++ )
-            if ( pCellLib->nCellInputs[i] >= nLutSize )
-            {
-                cellId = i;
-                break;
-            }
+        if ( !fHaveCfg2 )
+            for ( i = 0; i < pCellLib->nCellNum; i++ )
+                if ( pCellLib->nCellInputs[i] >= nLutSize )
+                {
+                    cellId = i;
+                    break;
+                }
         if ( cellId >= 0 )
         {
             // Use cell delays as integers from the library
@@ -243,6 +304,15 @@ float Gia_ObjPropagateRequired( Gia_Man_t * p, int iObj, int fUseSorting )
                     Gia_ObjSetTimeRequired( p, iFanin, tRequired );
             }
         }
+        else if ( fHaveCfg2 )
+        {
+            Gia_LutForEachFanin( p, iObj, iFanin, k )
+            {
+                tRequired = Gia_ObjTimeRequired( p, iObj ) - (float)pPinDelay[k];
+                if ( Gia_ObjTimeRequired(p, iFanin) > tRequired )
+                    Gia_ObjSetTimeRequired( p, iFanin, tRequired );
+            }
+        }
         else
         {
             // Fall back to default delay if no matching cell
@@ -977,4 +1047,3 @@ Gia_Man_t * Gia_ManSpeedup( Gia_Man_t * p, int Percentage, int Degree, int fVerb
 
 
 ABC_NAMESPACE_IMPL_END
-
diff --git a/src/base/abci/abc.c b/src/base/abci/abc.c
index 63847c88b..e42dfb8a4 100644
--- a/src/base/abci/abc.c
+++ b/src/base/abci/abc.c
@@ -22285,7 +22285,7 @@ int Abc_CommandIf( Abc_Frame_t * pAbc, int argc, char ** argv )
     If_ManSetDefaultPars( pPars );
     pPars->pLutLib = (If_LibLut_t *)Abc_FrameReadLibLut();
     Extra_UtilGetoptReset();
-    while ( ( c = Extra_UtilGetopt( argc, argv, "KCFAGRNTXYUZDEWSJqalepmrsdbgxyzuojiktncfvh" ) ) != EOF )
+    while ( ( c = Extra_UtilGetopt( argc, argv, "KCFAGRNTXYZUDEWSJqalepmrsdbgxyzuoiktncfvh" ) ) != EOF )
     {
         switch ( c )
         {
@@ -22535,9 +22535,9 @@ int Abc_CommandIf( Abc_Frame_t * pAbc, int argc, char ** argv )
         case 'o':
             pPars->fUseBuffs ^= 1;
             break;
-        case 'j':
-            pPars->fEnableCheck07 ^= 1;
-            break;
+        //case 'j':
+        //    pPars->fEnableCheck07 ^= 1;
+        //    break;
         case 'i':
             pPars->fUseCofVars ^= 1;
             break;
@@ -22867,7 +22867,7 @@ usage:
         sprintf(LutSize, "library" );
     else
         sprintf(LutSize, "%d", pPars->nLutSize );
-    Abc_Print( -2, "usage: if [-KCFAGRNTXYUZ num] [-DEW float] [-SJ str] [-qarlepmsdbgxyuojiktnczfvh]\n" );
+    Abc_Print( -2, "usage: if [-KCFAGRNTXYZMU num] [-DEW float] [-SJ str] [-qarlepmsdbgxyuoiktnczfvh]\n" );
     Abc_Print( -2, "\t           performs FPGA technology mapping of the network\n" );
     Abc_Print( -2, "\t-K num   : the number of LUT inputs (2 < num < %d) [default = %s]\n", IF_MAX_LUTSIZE+1, LutSize );
     Abc_Print( -2, "\t-C num   : the max number of priority cuts (0 < num < 2^12) [default = %d]\n", pPars->nCutsMax );
@@ -22901,7 +22901,7 @@ usage:
     Abc_Print( -2, "\t-y       : toggles delay optimization with recorded library [default = %s]\n", pPars->fUserRecLib? "yes": "no" );
     Abc_Print( -2, "\t-u       : toggles delay optimization with SAT-based library [default = %s]\n", pPars->fUserSesLib? "yes": "no" );
     Abc_Print( -2, "\t-o       : toggles using buffers to decouple combinational outputs [default = %s]\n", pPars->fUseBuffs? "yes": "no" );
-    Abc_Print( -2, "\t-j       : toggles enabling additional check [default = %s]\n", pPars->fEnableCheck07? "yes": "no" );
+    //Abc_Print( -2, "\t-j       : toggles enabling additional check [default = %s]\n", pPars->fEnableCheck07? "yes": "no" );
     Abc_Print( -2, "\t-i       : toggles using cofactoring variables [default = %s]\n", pPars->fUseCofVars? "yes": "no" );
     Abc_Print( -2, "\t-k       : toggles matching based on precomputed DSD manager [default = %s]\n", pPars->fUseDsdTune? "yes": "no" );
     Abc_Print( -2, "\t-t       : toggles optimizing average rather than maximum level [default = %s]\n", pPars->fDoAverage? "yes": "no" );
@@ -44067,8 +44067,9 @@ int Abc_CommandAbc9If( Abc_Frame_t * pAbc, int argc, char ** argv )
         Abc_FrameSetLibLut( If_LibLutSetSimple( 6 ) );
     }
     pPars->pLutLib = (If_LibLut_t *)Abc_FrameReadLibLut();
+    pPars->pCellLib = (If_LibCell_t *)Abc_FrameReadLibCell();
     Extra_UtilGetoptReset();
-    while ( ( c = Extra_UtilGetopt( argc, argv, "KCFAGRDEWSJTXYZqalepmrsdbgxyofuijkztncvwh" ) ) != EOF )
+    while ( ( c = Extra_UtilGetopt( argc, argv, "KCFAGRDEWSJTXYZMqalepmrsdbgxyofuijkztncvwh" ) ) != EOF )
     {
         switch ( c )
         {
@@ -44084,6 +44085,7 @@ int Abc_CommandAbc9If( Abc_Frame_t * pAbc, int argc, char ** argv )
                 goto usage;
             // if the LUT size is specified, disable library
             pPars->pLutLib = NULL;
+            pPars->pCellLib = NULL;
             break;
         case 'C':
             if ( globalUtilOptind >= argc )
@@ -44173,6 +44175,21 @@ int Abc_CommandAbc9If( Abc_Frame_t * pAbc, int argc, char ** argv )
             if ( pPars->nAndDelay < 0 )
                 goto usage;
             break;
+        case 'M':
+        {
+            int Value;
+            if ( globalUtilOptind >= argc )
+            {
+                Abc_Print( -1, "Command line switch \"-M\" should be followed by 0 or 1.\n" );
+                goto usage;
+            }
+            Value = atoi(argv[globalUtilOptind]);
+            globalUtilOptind++;
+            if ( Value < 0 || Value > 1 )
+                goto usage;
+            pPars->fDelayOptCell = Value;
+            break;
+        }
         case 'D':
             if ( globalUtilOptind >= argc )
             {
@@ -44349,6 +44366,7 @@ int Abc_CommandAbc9If( Abc_Frame_t * pAbc, int argc, char ** argv )
                     Abc_Print( 1, "Auto-detected K=%d from cell library (max inputs).\n", nMaxInputs );
                 // Disable LUT library since we're using K from cell library
                 pPars->pLutLib = NULL;
+                pPars->pCellLib = pCellLib;
             }
         }
     }
@@ -44537,6 +44555,11 @@ int Abc_CommandAbc9If( Abc_Frame_t * pAbc, int argc, char ** argv )
         pPars->nLutSize  = 4;
     }
 
+    if ( pPars->fEnableCheck07 )
+        pPars->nCutsMax = Abc_MaxInt( pPars->nCutsMax, 16 );
+    else
+        pPars->pCellLib = NULL;
+
     // enable truth table computation if cut minimization is selected
     if ( pPars->fCutMin || pPars->fDeriveLuts )
     {
@@ -44672,7 +44695,7 @@ usage:
         sprintf(LutSize, "library" );
     else
         sprintf(LutSize, "%d", pPars->nLutSize );
-    Abc_Print( -2, "usage: &if [-KCFAGRTXY num] [-DEW float] [-SJ str] [-qarlepmsdbgxyofuijkztnchvw]\n" );
+    Abc_Print( -2, "usage: &if [-KCFAGRTXYZM num] [-DEW float] [-SJ str] [-qarlepmsdbgxyofuijkztnchvw]\n" );
     Abc_Print( -2, "\t           performs FPGA technology mapping of the network\n" );
     Abc_Print( -2, "\t-K num   : the number of LUT inputs (2 < num < %d) [default = %s]\n", IF_MAX_LUTSIZE+1, LutSize );
     Abc_Print( -2, "\t-C num   : the max number of priority cuts (0 < num < 2^12) [default = %d]\n", pPars->nCutsMax );
@@ -44683,6 +44706,7 @@ usage:
     Abc_Print( -2, "\t-T num   : the type of LUT structures [default = any]\n", pPars->nStructType );
     Abc_Print( -2, "\t-X num   : delay of AND-gate in LUT library units [default = %d]\n", pPars->nAndDelay );
     Abc_Print( -2, "\t-Y num   : area of AND-gate in LUT library units [default = %d]\n", pPars->nAndArea );
+    Abc_Print( -2, "\t-M num   : enables delay-driven decomposition [default = %d]\n", pPars->fDelayOptCell );
     Abc_Print( -2, "\t-D float : sets the delay constraint for the mapping [default = %s]\n", Buffer );
     Abc_Print( -2, "\t-E float : sets epsilon used for tie-breaking [default = %f]\n", pPars->Epsilon );
     Abc_Print( -2, "\t-W float : sets wire delay between adjects LUTs [default = %f]\n", pPars->WireDelay );
@@ -48585,22 +48609,13 @@ int Abc_CommandAbc9Trace( Abc_Frame_t * pAbc, int argc, char ** argv )
     pAbc->pGia->pLutLib = fUseLutLib ? Abc_FrameReadLibLut() : NULL;
     pAbc->pGia->pCellLib = fUseCellLib ? Abc_FrameReadLibCell() : NULL;
 
-    if ( pFileName )
-    {
-        // Dump the delay trace to file
-        Gia_ManDelayTraceDump( pAbc->pGia, (char *)pFileName );
-    }
-    else
-    {
-        // Print the delay trace to console
-        Gia_ManDelayTraceLutPrint( pAbc->pGia, fVerbose );
-    }
+    Gia_ManDelayTraceDump( pAbc->pGia, (char *)pFileName );
     return 0;
 
 usage:
     Abc_Print( -2, "usage: &trace [-F file] [-lcvh]\n" );
     Abc_Print( -2, "\t           performs delay trace of LUT-mapped network\n" );
-    Abc_Print( -2, "\t-F file  : dump the critical path to a file [default = console output]\n" );
+    Abc_Print( -2, "\t-F file  : dump the critical path to a file [default = stdout]\n" );
     Abc_Print( -2, "\t-l       : toggle using LUT-library-delay model [default = %s]\n", fUseLutLib? "yes": "no" );
     Abc_Print( -2, "\t-c       : toggle using cell-library-delay model [default = %s]\n", fUseCellLib? "yes": "no" );
     Abc_Print( -2, "\t-v       : toggle printing optimization summary [default = %s]\n", fVerbose? "yes": "no" );
diff --git a/src/map/if/if.h b/src/map/if/if.h
index 3f9ddd101..bfc451c6c 100644
--- a/src/map/if/if.h
+++ b/src/map/if/if.h
@@ -126,6 +126,7 @@ struct If_Par_t_
     int                fCutMin;       // performs cut minimization by removing functionally reducdant variables
     int                fDelayOpt;     // special delay optimization
     int                fDelayOptLut;  // delay optimization for LUTs
+    int                fDelayOptCell; // delay optimization for cells
     int                fDsdBalance;   // special delay optimization
     int                fUserRecLib;   // use recorded library
     int                fUserSesLib;   // use SAT-based synthesis
@@ -175,6 +176,7 @@ struct If_Par_t_
     float              FinalDelay;    // final delay after mapping
     float              FinalArea;     // final area after mapping
     If_LibLut_t *      pLutLib;       // the LUT library
+    If_LibCell_t *     pCellLib;      // the cell library
     float *            pTimesArr;     // arrival times
     float *            pTimesReq;     // required times
     int (* pFuncCost)  (If_Man_t *, If_Cut_t *);  // procedure to compute the user's cost of a cut
@@ -299,6 +301,14 @@ struct If_Man_t_
     Vec_Int_t *        vVisited2;
     Vec_Int_t *        vCuts;
     Vec_Int_t *        vCutCosts;
+    // current cut context for user callbacks
+    If_Obj_t *         pCutObjCur;    // current object whose cut is being checked
+    If_Cut_t *         pCutCur;       // current cut being checked
+    int                nCutLeavesCur; // number of leaves in current cut
+    int                pCutLeavesCur[IF_MAX_LUTSIZE]; // current cut leaves (IDs)
+    float              pCutLeafArrCur[IF_MAX_LUTSIZE]; // current cut leaf arrivals
+    float              CutDelayCur;   // current cut delay returned by user callback
+    int                fCutDelayCurValid; // indicates CutDelayCur is valid
 
     // timing manager
     Tim_Man_t *        pManTim;
@@ -319,6 +329,7 @@ struct If_Cut_t_
     float              Edge;          // the edge flow
     float              Power;         // the power flow
     float              Delay;         // delay of the cut
+    word               Config;        // configuration string
     int                iCutFunc;      // TT ID of the cut
     int                uMaskFunc;     // polarity bitmask
     unsigned           uSign;         // cut signature
@@ -564,7 +575,8 @@ extern float           If_CutPowerDerefed( If_Man_t * p, If_Cut_t * pCut, If_Obj
 extern float           If_CutPowerRefed( If_Man_t * p, If_Cut_t * pCut, If_Obj_t * pRoot );
 /*=== ifDec.c =============================================================*/
 extern word            If_CutPerformDerive07( If_Man_t * p, unsigned * pTruth, int nVars, int nLeaves, char * pStr );
-extern word            If_CutPerformDeriveJ( If_Man_t * p, unsigned * pTruth, int nVars, int nLeaves, char * pStr, int fDerive );
+extern word            If_CutPerformDeriveJ( If_Man_t * p, unsigned * pTruth, int nVars, int nLeaves, char * pStr, int fDerive, int fDelay );
+extern void            If_CutComputeIntrinsicJ( If_Man_t * p, word Config, int nLeaves, int * pIntrinsicDelays );
 extern int             If_CutPerformCheck07( If_Man_t * p, unsigned * pTruth, int nVars, int nLeaves, char * pStr );
 extern int             If_CutPerformCheck08( If_Man_t * p, unsigned * pTruth, int nVars, int nLeaves, char * pStr );
 extern int             If_CutPerformCheck10( If_Man_t * p, unsigned * pTruth, int nVars, int nLeaves, char * pStr );
diff --git a/src/map/if/ifCore.c b/src/map/if/ifCore.c
index 58649b720..975a57306 100644
--- a/src/map/if/ifCore.c
+++ b/src/map/if/ifCore.c
@@ -207,4 +207,3 @@ int If_ManPerformMappingComb( If_Man_t * p )
 
 
 ABC_NAMESPACE_IMPL_END
-
diff --git a/src/map/if/ifDecJ.c b/src/map/if/ifDecJ.c
index 018ad4969..d9676727b 100644
--- a/src/map/if/ifDecJ.c
+++ b/src/map/if/ifDecJ.c
@@ -35,10 +35,13 @@ int If_CutPerformCheckJ( If_Man_t * p, unsigned * pTruth, int nVars, int nLeaves
 {
     return 1;
 }
-word If_CutPerformDeriveJ( If_Man_t * p, unsigned * pTruth, int nVars, int nLeaves, char * pStr, int fDerive )
+word If_CutPerformDeriveJ( If_Man_t * p, unsigned * pTruth, int nVars, int nLeaves, char * pStr, int fDerive, int fDelay )
 {
     return 0;
 }
+void If_CutComputeIntrinsicJ( If_Man_t * p, word Config, int nLeaves, int * pIntrinsicDelays )
+{
+}
 void If_PermUnpack( unsigned Value, int Pla2Var[9] )
 {
 }
@@ -52,4 +55,3 @@ void Gia_ManDelayTraceDump( Gia_Man_t * p, char * pFileName )
 
 
 ABC_NAMESPACE_IMPL_END
-
diff --git a/src/map/if/ifMan.c b/src/map/if/ifMan.c
index 1a79014da..ebf650db4 100644
--- a/src/map/if/ifMan.c
+++ b/src/map/if/ifMan.c
@@ -719,6 +719,7 @@ void If_ManSetupCutTriv( If_Man_t * p, If_Cut_t * pCut, int ObjId )
     pCut->uSign      = If_ObjCutSign( pCut->pLeaves[0] );
     pCut->iCutFunc   = p->pPars->fUseTtPerm ? 3 : (p->pPars->fTruth ? 2: -1);
     pCut->uMaskFunc  = 0;
+    pCut->Config     = 0;
     assert( pCut->pLeaves[0] < p->vObjs->nSize );
 }
 
diff --git a/src/map/if/ifMap.c b/src/map/if/ifMap.c
index ac62e26e4..b03201a6b 100644
--- a/src/map/if/ifMap.c
+++ b/src/map/if/ifMap.c
@@ -162,6 +162,7 @@ int * If_CutArrTimeProfile( If_Man_t * p, If_Cut_t * pCut )
 void If_ObjPerformMappingAnd( If_Man_t * p, If_Obj_t * pObj, int Mode, int fPreprocess, int fFirst )
 {
     If_Set_t * pCutSet;
+    If_Obj_t * pLeaf;
     If_Cut_t * pCut0, * pCut1, * pCut;
     If_Cut_t * pCut0R, * pCut1R;
     int fFunc0R, fFunc1R;
@@ -193,6 +194,31 @@ void If_ObjPerformMappingAnd( If_Man_t * p, If_Obj_t * pObj, int Mode, int fPrep
             pCut->Delay = If_CutSopBalanceEval( p, pCut, NULL );
         else if ( p->pPars->fDsdBalance )
             pCut->Delay = If_CutDsdBalanceEval( p, pCut, NULL );
+        else if ( p->pPars->fEnableCheck07 && p->pPars->fDelayOptCell )
+        {
+            int iLeaf, Intrinsic[IF_MAX_LUTSIZE];
+            float Delay = -IF_FLOAT_LARGE;
+            if ( pCut->nLeaves == 0 )
+            {
+                pCut->fUseless = 0;
+                pCut->Delay = 0.0;
+                goto IfMapBestCutDone;
+            }
+            assert( pCut->nLeaves == 1 || pCut->Config );
+            If_CutComputeIntrinsicJ( p, pCut->Config, pCut->nLeaves, Intrinsic );
+            If_CutForEachLeaf( p, pCut, pLeaf, iLeaf )
+            {
+                If_Cut_t * pBestCut = If_ObjCutBest( pLeaf );
+                assert( pBestCut != NULL );
+                assert( pBestCut->fUseless == 0 );
+                Delay = IF_MAX( Delay, If_ObjArrTime(pLeaf) + (float)Intrinsic[iLeaf] );
+            }
+            pCut->fUseless = (Delay > IF_FLOAT_LARGE/2);
+            pCut->Delay = Delay;
+IfMapBestCutDone:
+//            if ( pCut->nLeaves == 9 && !pCut->fUseless )
+//                Abc_Print( 1, "Delay-debug(Map9-best): obj=%d delay=%.2f\n", pObj->Id, pCut->Delay );
+        }
         else if ( p->pPars->fUserRecLib )
             pCut->Delay = If_CutDelayRecCost3( p, pCut, pObj ); 
         else if ( p->pPars->fUserSesLib )
@@ -324,12 +350,32 @@ void If_ObjPerformMappingAnd( If_Man_t * p, If_Obj_t * pObj, int Mode, int fPrep
             {
                 assert( p->pPars->fUseTtPerm == 0 );
                 assert( pCut->nLimit >= 4 && pCut->nLimit <= 16 );
+                pCut->Config = 0;
                 if ( p->pPars->fUseDsd )
                     pCut->fUseless = If_DsdManCheckDec( p->pIfDsdMan, If_CutDsdLit(p, pCut) );
                 else if ( p->pPars->pFuncCell2 )
                     pCut->fUseless = !p->pPars->pFuncCell2( p, (word *)If_CutTruthW(p, pCut), pCut->nLeaves, NULL, NULL );
                 else
+                {
+                    int iLeaf;
+                    // Expose current cut context for user callbacks.
+                    p->pCutObjCur = pObj;
+                    p->pCutCur = pCut;
+                    p->nCutLeavesCur = pCut->nLeaves;
+                    for ( iLeaf = 0; iLeaf < p->nCutLeavesCur; iLeaf++ )
+                    {
+                        If_Obj_t * pLeaf = If_CutLeaf( p, pCut, iLeaf );
+                        if ( p->pPars->fEnableCheck07 && p->pPars->fDelayOptCell && pCut->nLeaves > 1 )
+                        {
+                            If_Cut_t * pBestCut = If_ObjCutBest( pLeaf );
+                            assert( pBestCut != NULL );
+                            assert( pBestCut->fUseless == 0 );
+                        }
+                        p->pCutLeavesCur[iLeaf] = pLeaf->Id;
+                        p->pCutLeafArrCur[iLeaf] = If_ObjArrTime( pLeaf );
+                    }
                     pCut->fUseless = !p->pPars->pFuncCell( p, If_CutTruth(p, pCut), Abc_MaxInt(6, pCut->nLeaves), pCut->nLeaves, p->pPars->pLutStruct );
+                }
                 p->nCutsUselessAll += pCut->fUseless;
                 p->nCutsUseless[pCut->nLeaves] += pCut->fUseless;
                 p->nCutsCountAll++;
@@ -427,6 +473,28 @@ void If_ObjPerformMappingAnd( If_Man_t * p, If_Obj_t * pObj, int Mode, int fPrep
             pCut->Delay = If_CutSopBalanceEval( p, pCut, NULL );
         else if ( p->pPars->fDsdBalance )
             pCut->Delay = If_CutDsdBalanceEval( p, pCut, NULL );
+        else if ( p->pPars->fEnableCheck07 && p->pPars->fDelayOptCell )
+        {
+            if ( pCut->nLeaves == 0 )
+            {
+                pCut->Delay = 0.0;
+                pCut->fUseless = 0;
+                goto IfMapCutEvalDone;
+            }
+            if ( pCut->nLeaves == 1 )
+            {
+                pLeaf = If_ManObj( p, pCut->pLeaves[0] );
+                pCut->Delay = If_ObjArrTime( pLeaf );
+                pCut->fUseless = 0;
+                goto IfMapCutEvalDone;
+            }
+            assert( pCut->fUseless || p->fCutDelayCurValid );
+            assert( pCut->fUseless || pCut->Config != 0 );
+            pCut->Delay = pCut->fUseless ? IF_FLOAT_LARGE : p->CutDelayCur;
+IfMapCutEvalDone:
+//            if ( pCut->nLeaves == 9 && !pCut->fUseless )
+//                Abc_Print( 1, "Delay-debug(Map9): obj=%d delay=%.2f\n", pObj->Id, pCut->Delay );
+        }
         else if ( p->pPars->fUserRecLib )
             pCut->Delay = If_CutDelayRecCost3( p, pCut, pObj );
         else if ( p->pPars->fUserLutDec )
@@ -502,7 +570,7 @@ void If_ObjPerformMappingAnd( If_Man_t * p, If_Obj_t * pObj, int Mode, int fPrep
     if ( Mode && pObj->nRefs > 0 )
         If_CutAreaRef( p, If_ObjCutBest(pObj) );
     if ( If_ObjCutBest(pObj)->fUseless )
-        Abc_Print( 1, "The best cut is useless.\n" );
+        Abc_Print( 1, "The best cut is useless.  Please increase the number of cuts used by the mapper, for example: \"&if -C 32\"\n" );
     // call the user specified function for each cut
     if ( p->pPars->pFuncUser )
         If_ObjForEachCut( pObj, pCut, i )
@@ -702,4 +770,3 @@ int If_ManPerformMappingRound( If_Man_t * p, int nCutsUsed, int Mode, int fPrepr
 
 
 ABC_NAMESPACE_IMPL_END
-
diff --git a/src/map/if/ifTime.c b/src/map/if/ifTime.c
index 216e4ce5c..2b3e21a5a 100644
--- a/src/map/if/ifTime.c
+++ b/src/map/if/ifTime.c
@@ -131,6 +131,20 @@ float If_CutDelay( If_Man_t * p, If_Obj_t * pObj, If_Cut_t * pCut )
     }
     else
     {
+        if ( p->pPars->fEnableCheck07 && p->pPars->fDelayOptCell && p->pPars->pCellLib )
+        {
+            int Intrinsic[IF_MAX_LUTSIZE];
+            if ( pCut->nLeaves == 0 )
+                return 0.0;
+            assert( pCut->nLeaves == 1 || pCut->Config != 0 );
+            If_CutComputeIntrinsicJ( p, pCut->Config, pCut->nLeaves, Intrinsic );
+            If_CutForEachLeaf( p, pCut, pLeaf, i )
+            {
+                DelayCur = If_ObjCutBest(pLeaf)->Delay + (float)Intrinsic[i];
+                Delay = IF_MAX( Delay, DelayCur );
+            }
+            return Delay;
+        }
         if ( pCut->fUser )
         {
             assert( !p->pPars->fLiftLeaves );
@@ -219,6 +233,17 @@ void If_CutPropagateRequired( If_Man_t * p, If_Obj_t * pObj, If_Cut_t * pCut, fl
     }
     else
     {
+        if ( p->pPars->fEnableCheck07 && p->pPars->fDelayOptCell && p->pPars->pCellLib )
+        {
+            int Intrinsic[IF_MAX_LUTSIZE];
+            if ( pCut->nLeaves == 0 )
+                return;
+            assert( (pObj && pObj->Id == 0) || pCut->nLeaves == 1 || pCut->Config != 0 );
+            If_CutComputeIntrinsicJ( p, pCut->Config, pCut->nLeaves, Intrinsic );
+            If_CutForEachLeaf( p, pCut, pLeaf, i )
+                pLeaf->Required = IF_MIN( pLeaf->Required, ObjRequired - (float)Intrinsic[i] );
+            return;
+        }
         if ( pCut->fUser )
         {
             char Perm[IF_MAX_FUNC_LUTSIZE], * pPerm = Perm;
@@ -528,4 +553,3 @@ void If_ManComputeRequired( If_Man_t * p )
 
 
 ABC_NAMESPACE_IMPL_END
-