ARM: Add support for moving predicated false dest operands from sources.

14 files changed, 315 insertions, 3 deletions

diff --git a/src/arch/arm/isa/insts/misc.isa b/src/arch/arm/isa/insts/misc.isa
index a1cf895f8..68320d3c5 100644
--- a/src/arch/arm/isa/insts/misc.isa
+++ b/src/arch/arm/isa/insts/misc.isa
@@ -467,8 +467,7 @@ let {{
     exec_output += PredOpExecute.subst(usada8Iop)
 
     bkptCode = 'return new PrefetchAbort(PC, ArmFault::DebugEvent);\n'
-    bkptIop = InstObjParams("bkpt", "BkptInst", "ArmStaticInst",
-            bkptCode)
+    bkptIop = InstObjParams("bkpt", "BkptInst", "PredOp", bkptCode)
     header_output += BasicDeclare.subst(bkptIop)
     decoder_output += BasicConstructor.subst(bkptIop)
     exec_output += BasicExecute.subst(bkptIop)
diff --git a/src/arch/arm/isa/templates/basic.isa b/src/arch/arm/isa/templates/basic.isa
index 843f90840..0728b66e3 100644
--- a/src/arch/arm/isa/templates/basic.isa
+++ b/src/arch/arm/isa/templates/basic.isa
@@ -52,6 +52,11 @@ def template BasicConstructor {{
         inline %(class_name)s::%(class_name)s(ExtMachInst machInst)  : %(base_class)s("%(mnemonic)s", machInst, %(op_class)s)
         {
                 %(constructor)s;
+                if (!(condCode == COND_AL || condCode == COND_UC)) {
+                    for (int x = 0; x < _numDestRegs; x++) {
+                        _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+                    }
+                }
         }
 }};
 
diff --git a/src/arch/arm/isa/templates/branch.isa b/src/arch/arm/isa/templates/branch.isa
index a666c28fa..d1f581f51 100644
--- a/src/arch/arm/isa/templates/branch.isa
+++ b/src/arch/arm/isa/templates/branch.isa
@@ -53,6 +53,11 @@ def template BranchImmConstructor {{
         : %(base_class)s("%(mnemonic)s", machInst, %(op_class)s, _imm)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -75,6 +80,11 @@ def template BranchImmCondConstructor {{
                          _imm, _condCode)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -94,6 +104,11 @@ def template BranchRegConstructor {{
         : %(base_class)s("%(mnemonic)s", machInst, %(op_class)s, _op1)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -116,6 +131,11 @@ def template BranchRegCondConstructor {{
                          _op1, _condCode)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -152,6 +172,11 @@ def template BranchRegRegConstructor {{
         : %(base_class)s("%(mnemonic)s", machInst, %(op_class)s, _op1, _op2)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -173,5 +198,10 @@ def template BranchImmRegConstructor {{
         : %(base_class)s("%(mnemonic)s", machInst, %(op_class)s, _imm, _op1)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
diff --git a/src/arch/arm/isa/templates/macromem.isa b/src/arch/arm/isa/templates/macromem.isa
index c7ebfcd06..7620fb871 100644
--- a/src/arch/arm/isa/templates/macromem.isa
+++ b/src/arch/arm/isa/templates/macromem.isa
@@ -69,6 +69,11 @@ def template MicroMemConstructor {{
                          _ura, _urb, _up, _imm)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -89,6 +94,11 @@ def template MicroNeonMemDeclare {{
         {
             memAccessFlags |= extraMemFlags;
             %(constructor)s;
+            if (!(condCode == COND_AL || condCode == COND_UC)) {
+                for (int x = 0; x < _numDestRegs; x++) {
+                    _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+                }
+            }
         }
 
         %(BasicExecDeclare)s
@@ -121,6 +131,11 @@ def template MicroIntConstructor {{
                          _ura, _urb, _urc)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -158,6 +173,11 @@ def template MicroNeonMixDeclare {{
                            _dest, _op1, _step)
         {
             %(constructor)s;
+            if (!(condCode == COND_AL || condCode == COND_UC)) {
+                for (int x = 0; x < _numDestRegs; x++) {
+                    _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+                }
+            }
         }
 
         %(BasicExecDeclare)s
@@ -208,6 +228,11 @@ def template MicroNeonMixLaneDeclare {{
                            _dest, _op1, _step, _lane)
         {
             %(constructor)s;
+            if (!(condCode == COND_AL || condCode == COND_UC)) {
+                for (int x = 0; x < _numDestRegs; x++) {
+                    _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+                }
+            }
         }
 
         %(BasicExecDeclare)s
@@ -236,6 +261,11 @@ def template MicroIntMovConstructor {{
                          _ura, _urb)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -264,6 +294,11 @@ def template MicroIntImmConstructor {{
                          _ura, _urb, _imm)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -286,6 +321,11 @@ def template MicroIntRegConstructor {{
                          _ura, _urb, _urc, _shiftAmt, _shiftType)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -317,6 +357,11 @@ def template MacroMemConstructor {{
                      index, up, user, writeback, load, reglist)
 {
     %(constructor)s;
+    if (!(condCode == COND_AL || condCode == COND_UC)) {
+        for (int x = 0; x < _numDestRegs; x++) {
+            _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+        }
+    }
 }
 
 }};
@@ -341,6 +386,11 @@ def template VMemMultConstructor {{
                      rn, vd, regs, inc, size, align, rm)
 {
     %(constructor)s;
+    if (!(condCode == COND_AL || condCode == COND_UC)) {
+        for (int x = 0; x < _numDestRegs; x++) {
+            _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+        }
+    }
 }
 }};
 
@@ -364,6 +414,11 @@ def template VMemSingleConstructor {{
                      rn, vd, regs, inc, size, align, rm, lane)
 {
     %(constructor)s;
+    if (!(condCode == COND_AL || condCode == COND_UC)) {
+        for (int x = 0; x < _numDestRegs; x++) {
+            _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+        }
+    }
 }
 }};
 
@@ -390,6 +445,11 @@ def template MacroVFPMemConstructor {{
                      vd, single, up, writeback, load, offset)
 {
     %(constructor)s;
+    if (!(condCode == COND_AL || condCode == COND_UC)) {
+        for (int x = 0; x < _numDestRegs; x++) {
+            _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+        }
+    }
 }
 
 }};
diff --git a/src/arch/arm/isa/templates/mem.isa b/src/arch/arm/isa/templates/mem.isa
index ab4f1b01b..66384331b 100644
--- a/src/arch/arm/isa/templates/mem.isa
+++ b/src/arch/arm/isa/templates/mem.isa
@@ -913,6 +913,11 @@ def template RfeConstructor {{
                  (IntRegIndex)_base, (AddrMode)_mode, _wb)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
 #if %(use_uops)d
         assert(numMicroops >= 2);
         uops = new StaticInstPtr[numMicroops];
@@ -931,6 +936,11 @@ def template SrsConstructor {{
                  (OperatingMode)_regMode, (AddrMode)_mode, _wb)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
 #if %(use_uops)d
         assert(numMicroops >= 2);
         uops = new StaticInstPtr[numMicroops];
@@ -949,6 +959,11 @@ def template SwapConstructor {{
                  (IntRegIndex)_dest, (IntRegIndex)_op1, (IntRegIndex)_base)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -961,6 +976,11 @@ def template LoadStoreDImmConstructor {{
                  (IntRegIndex)_base, _add, _imm)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
 #if %(use_uops)d
         assert(numMicroops >= 2);
         uops = new StaticInstPtr[numMicroops];
@@ -982,6 +1002,11 @@ def template StoreExDImmConstructor {{
                  (IntRegIndex)_base, _add, _imm)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
 #if %(use_uops)d
         assert(numMicroops >= 2);
         uops = new StaticInstPtr[numMicroops];
@@ -1001,6 +1026,11 @@ def template LoadStoreImmConstructor {{
                  (IntRegIndex)_dest, (IntRegIndex)_base, _add, _imm)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
 #if %(use_uops)d
         assert(numMicroops >= 2);
         uops = new StaticInstPtr[numMicroops];
@@ -1021,6 +1051,11 @@ def template StoreExImmConstructor {{
                  (IntRegIndex)_base, _add, _imm)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
 #if %(use_uops)d
         assert(numMicroops >= 2);
         uops = new StaticInstPtr[numMicroops];
@@ -1044,6 +1079,11 @@ def template StoreDRegConstructor {{
                  (IntRegIndex)_index)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
 #if %(use_uops)d
         assert(numMicroops >= 2);
         uops = new StaticInstPtr[numMicroops];
@@ -1066,6 +1106,11 @@ def template StoreRegConstructor {{
                  (IntRegIndex)_index)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
 #if %(use_uops)d
         assert(numMicroops >= 2);
         uops = new StaticInstPtr[numMicroops];
@@ -1089,6 +1134,11 @@ def template LoadDRegConstructor {{
                  (IntRegIndex)_index)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
 #if %(use_uops)d
         assert(numMicroops >= 2);
         uops = new StaticInstPtr[numMicroops];
@@ -1123,6 +1173,11 @@ def template LoadRegConstructor {{
                  (IntRegIndex)_index)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
 #if %(use_uops)d
         assert(numMicroops >= 2);
         uops = new StaticInstPtr[numMicroops];
@@ -1164,6 +1219,11 @@ def template LoadImmConstructor {{
                  (IntRegIndex)_dest, (IntRegIndex)_base, _add, _imm)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
 #if %(use_uops)d
         assert(numMicroops >= 2);
         uops = new StaticInstPtr[numMicroops];
diff --git a/src/arch/arm/isa/templates/misc.isa b/src/arch/arm/isa/templates/misc.isa
index 915dea9b0..0347869f8 100644
--- a/src/arch/arm/isa/templates/misc.isa
+++ b/src/arch/arm/isa/templates/misc.isa
@@ -54,6 +54,11 @@ def template MrsConstructor {{
         : %(base_class)s("%(mnemonic)s", machInst, %(op_class)s, _dest)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -75,6 +80,11 @@ def template MsrRegConstructor {{
         : %(base_class)s("%(mnemonic)s", machInst, %(op_class)s, _op1, mask)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -96,6 +106,11 @@ def template MsrImmConstructor {{
         : %(base_class)s("%(mnemonic)s", machInst, %(op_class)s, imm, mask)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -115,6 +130,11 @@ def template ImmOpConstructor {{
         : %(base_class)s("%(mnemonic)s", machInst, %(op_class)s, _imm)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -135,6 +155,11 @@ def template RegImmOpConstructor {{
         : %(base_class)s("%(mnemonic)s", machInst, %(op_class)s, _dest, _imm)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -156,6 +181,11 @@ def template RegRegOpConstructor {{
         : %(base_class)s("%(mnemonic)s", machInst, %(op_class)s, _dest, _op1)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -182,6 +212,11 @@ def template RegRegRegImmOpConstructor {{
                          _dest, _op1, _op2, _imm)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -208,6 +243,11 @@ def template RegRegRegRegOpConstructor {{
                          _dest, _op1, _op2, _op3)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -232,6 +272,11 @@ def template RegRegRegOpConstructor {{
                          _dest, _op1, _op2)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -257,6 +302,11 @@ def template RegRegImmOpConstructor {{
                          _dest, _op1, _imm)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -283,6 +333,11 @@ def template RegRegImmImmOpConstructor {{
                          _dest, _op1, _imm1, _imm2)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -307,6 +362,11 @@ def template RegImmRegOpConstructor {{
                          _dest, _imm, _op1)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -334,6 +394,11 @@ def template RegImmRegShiftOpConstructor {{
                          _dest, _imm, _op1, _shiftAmt, _shiftType)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
diff --git a/src/arch/arm/isa/templates/mult.isa b/src/arch/arm/isa/templates/mult.isa
index d5001fb5f..dd4847fd4 100644
--- a/src/arch/arm/isa/templates/mult.isa
+++ b/src/arch/arm/isa/templates/mult.isa
@@ -57,6 +57,11 @@ def template Mult3Constructor {{
                          _reg0, _reg1, _reg2)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -82,5 +87,10 @@ def template Mult4Constructor {{
                          _reg0, _reg1, _reg2, _reg3)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
diff --git a/src/arch/arm/isa/templates/neon.isa b/src/arch/arm/isa/templates/neon.isa
index 0e592c6e4..02c2bb30d 100644
--- a/src/arch/arm/isa/templates/neon.isa
+++ b/src/arch/arm/isa/templates/neon.isa
@@ -59,6 +59,11 @@ class %(class_name)s : public %(base_class)s
                          _dest, _op1, _op2)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 
     %(BasicExecDeclare)s
@@ -80,6 +85,11 @@ class %(class_name)s : public %(base_class)s
                          _dest, _op1, _op2, _imm)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 
     %(BasicExecDeclare)s
@@ -100,6 +110,11 @@ class %(class_name)s : public %(base_class)s
                          _dest, _op1, _imm)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 
     %(BasicExecDeclare)s
@@ -118,6 +133,11 @@ class %(class_name)s : public %(base_class)s
         : %(base_class)s("%(mnemonic)s", machInst, %(op_class)s, _dest, _imm)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 
     %(BasicExecDeclare)s
@@ -138,6 +158,11 @@ class %(class_name)s : public %(base_class)s
                          _dest, _op1)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 
     %(BasicExecDeclare)s
diff --git a/src/arch/arm/isa/templates/pred.isa b/src/arch/arm/isa/templates/pred.isa
index b5bdbc40e..c9e7b1803 100644
--- a/src/arch/arm/isa/templates/pred.isa
+++ b/src/arch/arm/isa/templates/pred.isa
@@ -71,6 +71,11 @@ def template DataImmConstructor {{
                          _dest, _op1, _imm, _rotC)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -97,6 +102,11 @@ def template DataRegConstructor {{
                          _dest, _op1, _op2, _shiftAmt, _shiftType)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -123,6 +133,11 @@ def template DataRegRegConstructor {{
                          _dest, _op1, _op2, _shift, _shiftType)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
diff --git a/src/arch/arm/isa/templates/vfp.isa b/src/arch/arm/isa/templates/vfp.isa
index 8ccfedd0d..8888dc0ae 100644
--- a/src/arch/arm/isa/templates/vfp.isa
+++ b/src/arch/arm/isa/templates/vfp.isa
@@ -84,6 +84,11 @@ def template FpRegRegOpConstructor {{
                 _dest, _op1, mode)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -105,6 +110,11 @@ def template FpRegImmOpConstructor {{
                 _dest, _imm, mode)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -130,6 +140,11 @@ def template FpRegRegImmOpConstructor {{
                          _dest, _op1, _imm, mode)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
 
@@ -155,5 +170,10 @@ def template FpRegRegRegOpConstructor {{
                          _dest, _op1, _op2, mode)
     {
         %(constructor)s;
+        if (!(condCode == COND_AL || condCode == COND_UC)) {
+            for (int x = 0; x < _numDestRegs; x++) {
+                _srcRegIdx[_numSrcRegs++] = _destRegIdx[x];
+            }
+        }
     }
 }};
diff --git a/src/arch/arm/registers.hh b/src/arch/arm/registers.hh
index 61e467e24..30289ff27 100644
--- a/src/arch/arm/registers.hh
+++ b/src/arch/arm/registers.hh
@@ -49,7 +49,11 @@
 
 namespace ArmISA {
 
-using ArmISAInst::MaxInstSrcRegs;
+
+// For a predicated instruction, we need all the
+// destination registers to also be sources
+const int MaxInstSrcRegs = ArmISAInst::MaxInstDestRegs +
+    ArmISAInst::MaxInstSrcRegs;
 using ArmISAInst::MaxInstDestRegs;
 
 typedef uint16_t  RegIndex;
diff --git a/src/cpu/o3/dyn_inst.hh b/src/cpu/o3/dyn_inst.hh
index b62068111..487c284e6 100644
--- a/src/cpu/o3/dyn_inst.hh
+++ b/src/cpu/o3/dyn_inst.hh
@@ -181,6 +181,18 @@ class BaseO3DynInst : public BaseDynInst<Impl>
         this->thread->inSyscall = in_syscall;
     }
 
+    void forwardOldRegs()
+    {
+
+        for (int idx = 0; idx < this->numDestRegs(); idx++) {
+            PhysRegIndex prev_phys_reg = this->prevDestRegIdx(idx);
+            TheISA::RegIndex original_dest_reg = this->staticInst->destRegIdx(idx);
+            if (original_dest_reg <  TheISA::FP_Base_DepTag)
+                this->setIntRegOperand(this->staticInst.get(), idx, this->cpu->readIntReg(prev_phys_reg));
+            else if (original_dest_reg < TheISA::Ctrl_Base_DepTag)
+                this->setFloatRegOperandBits(this->staticInst.get(), idx, this->cpu->readFloatRegBits(prev_phys_reg));
+        }
+    }
 #if FULL_SYSTEM
     /** Calls hardware return from error interrupt. */
     Fault hwrei();
diff --git a/src/cpu/o3/iew_impl.hh b/src/cpu/o3/iew_impl.hh
index 2acb941e0..e1af20852 100644
--- a/src/cpu/o3/iew_impl.hh
+++ b/src/cpu/o3/iew_impl.hh
@@ -1261,6 +1261,8 @@ DefaultIEW<Impl>::executeInsts()
             // will be replaced and we will lose it.
             if (inst->getFault() == NoFault) {
                 inst->execute();
+                if (inst->readPredicate() == false)
+                    inst->forwardOldRegs();
             }
 
             inst->setExecuted();
diff --git a/src/cpu/o3/lsq_unit_impl.hh b/src/cpu/o3/lsq_unit_impl.hh
index 64d674666..7be5e4e5b 100644
--- a/src/cpu/o3/lsq_unit_impl.hh
+++ b/src/cpu/o3/lsq_unit_impl.hh
@@ -458,6 +458,8 @@ LSQUnit<Impl>::executeLoad(DynInstPtr &inst)
         // realizes there is activity.
         // Mark it as executed unless it is an uncached load that
         // needs to hit the head of commit.
+        if (inst->readPredicate() == false)
+            inst->forwardOldRegs();
         DPRINTF(LSQUnit, "Load [sn:%lli] not executed from %s\n",
                 inst->seqNum,
                 (load_fault != NoFault ? "fault" : "predication"));
@@ -530,6 +532,9 @@ LSQUnit<Impl>::executeStore(DynInstPtr &store_inst)
 
     Fault store_fault = store_inst->initiateAcc();
 
+    if (store_inst->readPredicate() == false)
+        store_inst->forwardOldRegs();
+
     if (storeQueue[store_idx].size == 0) {
         DPRINTF(LSQUnit,"Fault on Store PC %s, [sn:%lli], Size = 0\n",
                 store_inst->pcState(), store_inst->seqNum);