merge my index fix and lisa's fix

--HG--
extra : convert_revision : 5f2c7d46c96fa061bbfb66edf188d405ca600020

10 files changed, 367 insertions, 219 deletions

diff --git a/src/arch/sparc/floatregfile.cc b/src/arch/sparc/floatregfile.cc
index 585782ddb..e1b5ea7c8 100644
--- a/src/arch/sparc/floatregfile.cc
+++ b/src/arch/sparc/floatregfile.cc
@@ -69,22 +69,25 @@ FloatReg FloatRegFile::readReg(int floatReg, int width)
     switch(width)
     {
       case SingleWidth:
-        float32_t result32;
+        uint32_t result32;
+        float32_t fresult32;
         memcpy(&result32, regSpace + 4 * floatReg, sizeof(result32));
-        result = htog(result32);
-        DPRINTF(Sparc, "Read FP32 register %d = 0x%x\n", floatReg, result);
+        result32 = htog(result32);
+        memcpy(&fresult32, &result32, sizeof(result32));
+        result = fresult32;
+        DPRINTF(Sparc, "Read FP32 register %d = [%f]0x%x\n", floatReg, result, result32);
         break;
       case DoubleWidth:
-        float64_t result64;
+        uint64_t result64;
+        float64_t fresult64;
         memcpy(&result64, regSpace + 4 * floatReg, sizeof(result64));
-        result = htog(result64);
-        DPRINTF(Sparc, "Read FP64 register %d = 0x%x\n", floatReg, result);
+        result64 = htog(result64);
+        memcpy(&fresult64, &result64, sizeof(result64));
+        result = fresult64;
+        DPRINTF(Sparc, "Read FP64 register %d = [%f]0x%x\n", floatReg, result, result64);
         break;
       case QuadWidth:
-        float128_t result128;
-        memcpy(&result128, regSpace + 4 * floatReg, sizeof(result128));
-        result = htog(result128);
-        DPRINTF(Sparc, "Read FP128 register %d = 0x%x\n", floatReg, result);
+        panic("Quad width FP not implemented.");
         break;
       default:
         panic("Attempted to read a %d bit floating point register!", width);
@@ -113,10 +116,7 @@ FloatRegBits FloatRegFile::readRegBits(int floatReg, int width)
         DPRINTF(Sparc, "Read FP64 bits register %d = 0x%x\n", floatReg, result);
         break;
       case QuadWidth:
-        uint64_t result128;
-        memcpy(&result128, regSpace + 4 * floatReg, sizeof(result128));
-        result = htog(result128);
-        DPRINTF(Sparc, "Read FP128 bits register %d = 0x%x\n", floatReg, result);
+        panic("Quad width FP not implemented.");
         break;
       default:
         panic("Attempted to read a %d bit floating point register!", width);
@@ -132,15 +132,21 @@ Fault FloatRegFile::setReg(int floatReg, const FloatReg &val, int width)
 
     uint32_t result32;
     uint64_t result64;
+    float32_t fresult32;
+    float64_t fresult64;
     switch(width)
     {
       case SingleWidth:
-        result32 = gtoh((uint32_t)val);
+        fresult32 = val;
+        memcpy(&result32, &fresult32, sizeof(result32));
+        result32 = gtoh(result32);
         memcpy(regSpace + 4 * floatReg, &result32, sizeof(result32));
         DPRINTF(Sparc, "Write FP64 register %d = 0x%x\n", floatReg, result32);
         break;
       case DoubleWidth:
-        result64 = gtoh((uint64_t)val);
+        fresult64 = val;
+        memcpy(&result64, &fresult64, sizeof(result64));
+        result64 = gtoh(result64);
         memcpy(regSpace + 4 * floatReg, &result64, sizeof(result64));
         DPRINTF(Sparc, "Write FP64 register %d = 0x%x\n", floatReg, result64);
         break;
diff --git a/src/arch/sparc/interrupts.hh b/src/arch/sparc/interrupts.hh
index 3af4f6342..dc3b235fe 100644
--- a/src/arch/sparc/interrupts.hh
+++ b/src/arch/sparc/interrupts.hh
@@ -46,164 +46,167 @@ enum interrupts_t {
     num_interrupt_types
 };
 
-    class Interrupts
-    {
+class Interrupts
+{
 
-      private:
+  private:
 
-        bool interrupts[num_interrupt_types];
-        int numPosted;
+    bool interrupts[num_interrupt_types];
+    int numPosted;
 
-      public:
-        Interrupts()
-        {
-            for (int i = 0; i < num_interrupt_types; ++i) {
-                interrupts[i] = false;
-            }
-            numPosted = 0;
+  public:
+    Interrupts()
+    {
+        for (int i = 0; i < num_interrupt_types; ++i) {
+            interrupts[i] = false;
         }
+        numPosted = 0;
+    }
 
-        void post(int int_type)
-        {
-            if (int_type < 0 || int_type >= num_interrupt_types)
-                panic("posting unknown interrupt!\n");
-
-            if (interrupts[int_type] == false) {
-                interrupts[int_type] = true;
-                ++numPosted;
-            }
+    void post(int int_type)
+    {
+        if (int_type < 0 || int_type >= num_interrupt_types)
+            panic("posting unknown interrupt!\n");
+        if (interrupts[int_type] == false) {
+            interrupts[int_type] = true;
+            ++numPosted;
         }
+    }
 
-        void post(int int_num, int index)
-        {
-
-        }
+    void post(int int_num, int index)
+    {
 
-        void clear(int int_num, int index)
-        {
+    }
 
-        }
+    void clear(int int_num, int index)
+    {
 
-        void clear_all()
-        {
+    }
 
-        }
+    void clear_all()
+    {
 
-        bool check_interrupts(ThreadContext * tc) const
-        {
-            if (numPosted)
-                return true;
-            else
-                return false;
-        }
+    }
 
-        Fault getInterrupt(ThreadContext * tc)
-        {
-            int hpstate = tc->readMiscReg(MISCREG_HPSTATE);
-            int pstate = tc->readMiscReg(MISCREG_PSTATE);
-            bool ie = pstate & PSTATE::ie;
-
-            // THESE ARE IN ORDER OF PRIORITY
-            // since there are early returns, and the highest
-            // priority interrupts should get serviced,
-            // it is v. important that new interrupts are inserted
-            // in the right order of processing
-            if (hpstate & HPSTATE::hpriv) {
-                if (ie) {
-                    if (interrupts[hstick_match]) {
-                        if (tc->readMiscReg(MISCREG_HINTP) & 1) {
-                            interrupts[hstick_match] = false;
-                            --numPosted;
-                            return new HstickMatch;
-                        }
-                    }
-                    if (interrupts[interrupt_vector]) {
-                        interrupts[interrupt_vector] = false;
-                        --numPosted;
-                        //HAVEN'T IMPLed THIS YET
-                        return NoFault;
-                    }
-                } else {
-                    if (interrupts[hstick_match]) {
-                        return NoFault;
-                    }
+    bool check_interrupts(ThreadContext * tc) const
+    {
+        if (numPosted)
+            return true;
+        else
+            return false;
+    }
 
-                }
-            } else {
-                if (interrupts[trap_level_zero]) {
-                    if ((pstate & HPSTATE::tlz) && (tc->readMiscReg(MISCREG_TL) == 0)) {
-                        interrupts[trap_level_zero] = false;
-                        --numPosted;
-                        return new TrapLevelZero;
-                    }
-                }
+    Fault getInterrupt(ThreadContext * tc)
+    {
+        int hpstate = tc->readMiscReg(MISCREG_HPSTATE);
+        int pstate = tc->readMiscReg(MISCREG_PSTATE);
+        bool ie = pstate & PSTATE::ie;
+
+        // THESE ARE IN ORDER OF PRIORITY
+        // since there are early returns, and the highest
+        // priority interrupts should get serviced,
+        // it is v. important that new interrupts are inserted
+        // in the right order of processing
+        if (hpstate & HPSTATE::hpriv) {
+            if (ie) {
                 if (interrupts[hstick_match]) {
                     if (tc->readMiscReg(MISCREG_HINTP) & 1) {
                         interrupts[hstick_match] = false;
                         --numPosted;
                         return new HstickMatch;
-                        }
-                }
-                if (ie) {
-                    if (interrupts[cpu_mondo]) {
-                        interrupts[cpu_mondo] = false;
-                        --numPosted;
-                        return new CpuMondo;
                     }
-                    if (interrupts[dev_mondo]) {
-                        interrupts[dev_mondo] = false;
-                        --numPosted;
-                        return new DevMondo;
+                }
+                if (interrupts[interrupt_vector]) {
+                    interrupts[interrupt_vector] = false;
+                    --numPosted;
+                    //HAVEN'T IMPLed THIS YET
+                    return NoFault;
+                }
+            } else {
+                if (interrupts[hstick_match]) {
+                    return NoFault;
+                }
+
+            }
+        } else {
+            if (interrupts[trap_level_zero]) {
+                if ((pstate & HPSTATE::tlz) && (tc->readMiscReg(MISCREG_TL) == 0)) {
+                    interrupts[trap_level_zero] = false;
+                    --numPosted;
+                    return new TrapLevelZero;
+                }
+            }
+            if (interrupts[hstick_match]) {
+                if (tc->readMiscReg(MISCREG_HINTP) & 1) {
+                    interrupts[hstick_match] = false;
+                    --numPosted;
+                    return new HstickMatch;
                     }
-                    if (interrupts[soft_interrupt]) {
-                        int il = InterruptLevel(tc->readMiscReg(MISCREG_SOFTINT));
-                        // it seems that interrupt vectors are right in
-                        // the middle of interrupt levels with regard to
-                        // priority, so have to check
-                        if ((il < 6) &&
-                            interrupts[interrupt_vector]) {
-                                // may require more details here since there
-                                // may be lots of interrupts embedded in an
-                                // platform interrupt vector
-                                interrupts[interrupt_vector] = false;
+            }
+            if (ie) {
+                if (interrupts[cpu_mondo]) {
+                    interrupts[cpu_mondo] = false;
+                    --numPosted;
+                    return new CpuMondo;
+                }
+                if (interrupts[dev_mondo]) {
+                    interrupts[dev_mondo] = false;
+                    --numPosted;
+                    return new DevMondo;
+                }
+                if (interrupts[soft_interrupt]) {
+                    int il = InterruptLevel(tc->readMiscReg(MISCREG_SOFTINT));
+                    // it seems that interrupt vectors are right in
+                    // the middle of interrupt levels with regard to
+                    // priority, so have to check
+                    if ((il < 6) &&
+                        interrupts[interrupt_vector]) {
+                            // may require more details here since there
+                            // may be lots of interrupts embedded in an
+                            // platform interrupt vector
+                            interrupts[interrupt_vector] = false;
+                            --numPosted;
+                            //HAVEN'T IMPLed YET
+                            return NoFault;
+                    } else {
+                        if (il > tc->readMiscReg(MISCREG_PIL)) {
+                            uint64_t si = tc->readMiscReg(MISCREG_SOFTINT);
+                            uint64_t more = si & ~(1 << (il + 1));
+                            if (!InterruptLevel(more)) {
+                                interrupts[soft_interrupt] = false;
                                 --numPosted;
-                                //HAVEN'T IMPLed YET
-                                return NoFault;
-                        } else {
-                            if (il > tc->readMiscReg(MISCREG_PIL)) {
-                                uint64_t si = tc->readMiscReg(MISCREG_SOFTINT);
-                                uint64_t more = si & ~(1 << (il + 1));
-                                if (!InterruptLevel(more)) {
-                                    interrupts[soft_interrupt] = false;
-                                    --numPosted;
-                                }
-                                return new InterruptLevelN(il);
                             }
+                            return new InterruptLevelN(il);
                         }
                     }
-                    if (interrupts[resumable_error]) {
-                        interrupts[resumable_error] = false;
-                        --numPosted;
-                        return new ResumableError;
-                    }
+                }
+                if (interrupts[resumable_error]) {
+                    interrupts[resumable_error] = false;
+                    --numPosted;
+                    return new ResumableError;
                 }
             }
-            return NoFault;
         }
+        return NoFault;
+    }
 
-        void updateIntrInfo(ThreadContext * tc)
-        {
+    void updateIntrInfo(ThreadContext * tc)
+    {
 
-        }
+    }
 
-        void serialize(std::ostream &os)
-        {
-        }
+    void serialize(std::ostream &os)
+    {
+        SERIALIZE_ARRAY(interrupts,num_interrupt_types);
+        SERIALIZE_SCALAR(numPosted);
+    }
 
-        void unserialize(Checkpoint *cp, const std::string &section)
-        {
-        }
-    };
-}
+    void unserialize(Checkpoint *cp, const std::string &section)
+    {
+        UNSERIALIZE_ARRAY(interrupts,num_interrupt_types);
+        UNSERIALIZE_SCALAR(numPosted);
+    }
+};
+} // namespace SPARC_ISA
 
 #endif // __ARCH_SPARC_INTERRUPT_HH__
diff --git a/src/arch/sparc/isa/decoder.isa b/src/arch/sparc/isa/decoder.isa
index 81443fecb..fb606c7cc 100644
--- a/src/arch/sparc/isa/decoder.isa
+++ b/src/arch/sparc/isa/decoder.isa
@@ -479,10 +479,10 @@ decode OP default Unknown::unknown()
                 0x11: PrivCheck::rdpic({{Rd = Pic;}}, {{Pcr<0:>}});
                 //0x12 should cause an illegal instruction exception
                 0x13: NoPriv::rdgsr({{
-                    if(Fprs<2:> == 0 || Pstate<4:> == 0)
-                        Rd = Gsr;
-                    else
-                        fault = new FpDisabled;
+                       fault = checkFpEnableFault(xc);
+                       if (fault)
+                            return fault;
+                       Rd = Gsr;
                 }});
                 //0x14-0x15 should cause an illegal instruction exception
                 0x16: Priv::rdsoftint({{Rd = Softint;}});
@@ -718,7 +718,7 @@ decode OP default Unknown::unknown()
                 0x1F: HPriv::wrhprhstick_cmpr({{HstickCmpr = Rs1 ^ Rs2_or_imm13;}});
             }
             0x34: decode OPF{
-                format BasicOperate{
+                format FpBasic{
                     0x01: fmovs({{
                         Frds.uw = Frs2s.uw;
                         //fsr.ftt = fsr.cexc = 0
@@ -765,7 +765,7 @@ decode OP default Unknown::unknown()
                     0x42: faddd({{Frd.df = Frs1.df + Frs2.df;}});
                     0x43: FpUnimpl::faddq();
                     0x45: fsubs({{Frds.sf = Frs1s.sf - Frs2s.sf;}});
-                    0x46: fsubd({{Frd.df = Frs1.df - Frs2.df;}});
+                    0x46: fsubd({{Frd.df = Frs1.df - Frs2.df; }});
                     0x47: FpUnimpl::fsubq();
                     0x49: fmuls({{Frds.sf = Frs1s.sf * Frs2s.sf;}});
                     0x4A: fmuld({{Frd.df = Frs1.df * Frs2.df;}});
@@ -776,26 +776,26 @@ decode OP default Unknown::unknown()
                     0x69: fsmuld({{Frd.df = Frs1s.sf * Frs2s.sf;}});
                     0x6E: FpUnimpl::fdmulq();
                     0x81: fstox({{
-                            Frd.df = (double)static_cast<int64_t>(Frs2s.sf);
+                            Frd.sdw = static_cast<int64_t>(Frs2s.sf);
                     }});
                     0x82: fdtox({{
-                            Frd.df = (double)static_cast<int64_t>(Frs2.df);
+                            Frd.sdw = static_cast<int64_t>(Frs2.df);
                     }});
                     0x83: FpUnimpl::fqtox();
                     0x84: fxtos({{
-                            Frds.sf = static_cast<float>((int64_t)Frs2.df);
+                            Frds.sf = static_cast<float>(Frs2.sdw);
                     }});
                     0x88: fxtod({{
-                            Frd.df = static_cast<double>((int64_t)Frs2.df);
+                            Frd.df = static_cast<double>(Frs2.sdw);
                     }});
                     0x8C: FpUnimpl::fxtoq();
                     0xC4: fitos({{
-                            Frds.sf = static_cast<float>((int32_t)Frs2s.sf);
+                            Frds.sf = static_cast<float>(Frs2s.sw);
                     }});
                     0xC6: fdtos({{Frds.sf = Frs2.df;}});
                     0xC7: FpUnimpl::fqtos();
                     0xC8: fitod({{
-                            Frd.df = static_cast<double>((int32_t)Frs2s.sf);
+                            Frd.df = static_cast<double>(Frs2s.sw);
                     }});
                     0xC9: fstod({{Frd.df = Frs2s.sf;}});
                     0xCB: FpUnimpl::fqtod();
@@ -803,17 +803,23 @@ decode OP default Unknown::unknown()
                     0xCD: FpUnimpl::fstoq();
                     0xCE: FpUnimpl::fdtoq();
                     0xD1: fstoi({{
-                            Frds.sf = (float)static_cast<int32_t>(Frs2s.sf);
+                            Frds.sw = static_cast<int32_t>(Frs2s.sf);
+                            float t = Frds.sw;
+                            if (t != Frs2s.sf)
+                               Fsr = insertBits(Fsr, 4,0, 0x01);
                     }});
                     0xD2: fdtoi({{
-                            Frds.sf = (float)static_cast<int32_t>(Frs2.df);
+                            Frds.sw = static_cast<int32_t>(Frs2.df);
+                            double t = Frds.sw;
+                            if (t != Frs2.df)
+                               Fsr = insertBits(Fsr, 4,0, 0x01);
                     }});
                     0xD3: FpUnimpl::fqtoi();
                     default: FailUnimpl::fpop1();
                 }
             }
             0x35: decode OPF{
-                format BasicOperate{
+                format FpBasic{
                     0x51: fcmps({{
                           uint8_t fcc;
                           if(isnan(Frs1s) || isnan(Frs2s))
@@ -831,11 +837,11 @@ decode OP default Unknown::unknown()
                     }});
                     0x52: fcmpd({{
                           uint8_t fcc;
-                          if(isnan(Frs1s) || isnan(Frs2s))
+                          if(isnan(Frs1) || isnan(Frs2))
                               fcc = 3;
-                          else if(Frs1s < Frs2s)
+                          else if(Frs1 < Frs2)
                               fcc = 1;
-                          else if(Frs1s > Frs2s)
+                          else if(Frs1 > Frs2)
                               fcc = 2;
                           else
                               fcc = 0;
@@ -860,11 +866,11 @@ decode OP default Unknown::unknown()
                     }});
                     0x56: fcmped({{
                           uint8_t fcc = 0;
-                          if(isnan(Frs1s) || isnan(Frs2s))
+                          if(isnan(Frs1) || isnan(Frs2))
                               fault = new FpExceptionIEEE754;
-                          if(Frs1s < Frs2s)
+                          if(Frs1 < Frs2)
                               fcc = 1;
-                          else if(Frs1s > Frs2s)
+                          else if(Frs1 > Frs2)
                               fcc = 2;
                           uint8_t firstbit = 10;
                           if(FCMPCC)
@@ -960,24 +966,24 @@ decode OP default Unknown::unknown()
                 0x55: FailUnimpl::fpsub16s();
                 0x56: FailUnimpl::fpsub32();
                 0x57: FailUnimpl::fpsub32s();
-                0x60: BasicOperate::fzero({{Frd.df = 0;}});
-                0x61: BasicOperate::fzeros({{Frds.sf = 0;}});
+                0x60: FpBasic::fzero({{Frd.df = 0;}});
+                0x61: FpBasic::fzeros({{Frds.sf = 0;}});
                 0x62: FailUnimpl::fnor();
                 0x63: FailUnimpl::fnors();
                 0x64: FailUnimpl::fandnot2();
                 0x65: FailUnimpl::fandnot2s();
-                0x66: BasicOperate::fnot2({{
+                0x66: FpBasic::fnot2({{
                         Frd.df = (double)(~((uint64_t)Frs2.df));
                 }});
-                0x67: BasicOperate::fnot2s({{
+                0x67: FpBasic::fnot2s({{
                         Frds.sf = (float)(~((uint32_t)Frs2s.sf));
                 }});
                 0x68: FailUnimpl::fandnot1();
                 0x69: FailUnimpl::fandnot1s();
-                0x6A: BasicOperate::fnot1({{
+                0x6A: FpBasic::fnot1({{
                         Frd.df = (double)(~((uint64_t)Frs1.df));
                 }});
-                0x6B: BasicOperate::fnot1s({{
+                0x6B: FpBasic::fnot1s({{
                         Frds.sf = (float)(~((uint32_t)Frs1s.sf));
                 }});
                 0x6C: FailUnimpl::fxor();
@@ -988,18 +994,18 @@ decode OP default Unknown::unknown()
                 0x71: FailUnimpl::fands();
                 0x72: FailUnimpl::fxnor();
                 0x73: FailUnimpl::fxnors();
-                0x74: BasicOperate::fsrc1({{Frd.udw = Frs1.udw;}});
-                0x75: BasicOperate::fsrc1s({{Frds.uw = Frs1s.uw;}});
+                0x74: FpBasic::fsrc1({{Frd.udw = Frs1.udw;}});
+                0x75: FpBasic::fsrc1s({{Frds.uw = Frs1s.uw;}});
                 0x76: FailUnimpl::fornot2();
                 0x77: FailUnimpl::fornot2s();
-                0x78: BasicOperate::fsrc2({{Frd.udw = Frs2.udw;}});
-                0x79: BasicOperate::fsrc2s({{Frds.uw = Frs2s.uw;}});
+                0x78: FpBasic::fsrc2({{Frd.udw = Frs2.udw;}});
+                0x79: FpBasic::fsrc2s({{Frds.uw = Frs2s.uw;}});
                 0x7A: FailUnimpl::fornot1();
                 0x7B: FailUnimpl::fornot1s();
                 0x7C: FailUnimpl::for();
                 0x7D: FailUnimpl::fors();
-                0x7E: BasicOperate::fone({{Frd.udw = std::numeric_limits<uint64_t>::max();}});
-                0x7F: BasicOperate::fones({{Frds.uw = std::numeric_limits<uint32_t>::max();}});
+                0x7E: FpBasic::fone({{Frd.udw = std::numeric_limits<uint64_t>::max();}});
+                0x7F: FpBasic::fones({{Frds.uw = std::numeric_limits<uint32_t>::max();}});
                 0x80: Trap::shutdown({{fault = new IllegalInstruction;}});
                 0x81: FailUnimpl::siam();
             }
@@ -1236,16 +1242,32 @@ decode OP default Unknown::unknown()
                Rd.uw = uReg0;}}, {{EXT_ASI}});
         format Trap {
             0x20: Load::ldf({{Frds.uw = Mem.uw;}});
-            0x21: decode X {
-                0x0: Load::ldfsr({{Fsr = Mem.uw | Fsr<63:32>;}});
-                0x1: Load::ldxfsr({{Fsr = Mem.udw;}});
+            0x21: decode RD {
+                0x0: Load::ldfsr({{fault = checkFpEnableFault(xc);
+                                     if (fault)
+                                         return fault;
+                                   Fsr = Mem.uw | Fsr<63:32>;}});
+                0x1: Load::ldxfsr({{fault = checkFpEnableFault(xc);
+                                     if (fault)
+                                         return fault;
+                                    Fsr = Mem.udw;}});
+                default: FailUnimpl::ldfsrOther();
             }
             0x22: ldqf({{fault = new FpDisabled;}});
             0x23: Load::lddf({{Frd.udw = Mem.udw;}});
             0x24: Store::stf({{Mem.uw = Frds.uw;}});
-            0x25: decode X {
-                0x0: Store::stfsr({{Mem.uw = Fsr<31:0>;}});
-                0x1: Store::stxfsr({{Mem.udw = Fsr;}});
+            0x25: decode RD {
+                0x0: Store::stfsr({{fault = checkFpEnableFault(xc);
+                                     if (fault)
+                                         return fault;
+                                    Mem.uw = Fsr<31:0>;
+                                    Fsr = insertBits(Fsr,16,14,0);}});
+                0x1: Store::stxfsr({{fault = checkFpEnableFault(xc);
+                                     if (fault)
+                                         return fault;
+                                     Mem.udw = Fsr;
+                                     Fsr = insertBits(Fsr,16,14,0);}});
+                default: FailUnimpl::stfsrOther();
             }
             0x26: stqf({{fault = new FpDisabled;}});
             0x27: Store::stdf({{Mem.udw = Frd.udw;}});
diff --git a/src/arch/sparc/isa/formats/basic.isa b/src/arch/sparc/isa/formats/basic.isa
index e8762a205..017f43780 100644
--- a/src/arch/sparc/isa/formats/basic.isa
+++ b/src/arch/sparc/isa/formats/basic.isa
@@ -103,3 +103,46 @@ def format BasicOperate(code, *flags) {{
         decode_block = BasicDecode.subst(iop)
         exec_output = BasicExecute.subst(iop)
 }};
+
+def format FpBasic(code, *flags) {{
+        fp_code = """
+    Fsr |= bits(Fsr,4,0) << 5;
+    Fsr = insertBits(Fsr,4,0,0);
+#if defined(__sun) || defined (__OpenBSD__)
+    fp_rnd newrnd = FP_RN;
+    switch (Fsr<31:30>) {
+      case 0: newrnd = FP_RN; break;
+      case 1: newrnd = FP_RZ; break;
+      case 2: newrnd = FP_RP; break;
+      case 3: newrnd = FP_RM; break;
+    }
+    fp_rnd oldrnd = fpsetround(newrnd);
+#else
+    int newrnd = FE_TONEAREST;
+    switch (Fsr<31:30>) {
+      case 0: newrnd = FE_TONEAREST; break;
+      case 1: newrnd = FE_TOWARDZERO; break;
+      case 2: newrnd = FE_UPWARD; break;
+      case 3: newrnd = FE_DOWNWARD; break;
+    }
+    int oldrnd = fegetround();
+    fesetround(newrnd);
+#endif
+"""
+
+        fp_code += code
+
+
+        fp_code += """
+#if defined(__sun) || defined (__OpenBSD__)
+    fpsetround(oldrnd);
+#else
+    fesetround(oldrnd);
+#endif
+"""
+        iop = InstObjParams(name, Name, 'SparcStaticInst', fp_code, flags)
+        header_output = BasicDeclare.subst(iop)
+        decoder_output = BasicConstructor.subst(iop)
+        decode_block = BasicDecode.subst(iop)
+        exec_output = BasicExecute.subst(iop)
+}};
diff --git a/src/arch/sparc/isa/includes.isa b/src/arch/sparc/isa/includes.isa
index a6dca9bf1..d2ef67154 100644
--- a/src/arch/sparc/isa/includes.isa
+++ b/src/arch/sparc/isa/includes.isa
@@ -53,7 +53,7 @@ output decoder {{
 #include "cpu/thread_context.hh"  // for Jump::branchTarget()
 #include "mem/packet.hh"
 
-#if defined(linux)
+#if defined(linux) || defined(__APPLE__)
 #include <fenv.h>
 #endif
 #include <algorithm>
@@ -62,9 +62,14 @@ using namespace SparcISA;
 }};
 
 output exec {{
-#if defined(linux)
+#if defined(linux) || defined(__APPLE__)
 #include <fenv.h>
 #endif
+
+#if defined(__sun) || defined (__OpenBSD__)
+#include <ieeefp.h>
+#endif
+
 #include <limits>
 
 #include <cmath>
diff --git a/src/arch/sparc/miscregfile.cc b/src/arch/sparc/miscregfile.cc
index 0fe3e96b2..8b612e8b4 100644
--- a/src/arch/sparc/miscregfile.cc
+++ b/src/arch/sparc/miscregfile.cc
@@ -232,6 +232,7 @@ MiscReg MiscRegFile::readReg(int miscReg)
 
         /** Floating Point Status Register */
       case MISCREG_FSR:
+        DPRINTF(Sparc, "FSR read as: %#x\n", fsr);
         return fsr;
 
       case MISCREG_MMU_P_CONTEXT:
@@ -337,10 +338,6 @@ MiscReg MiscRegFile::readRegWithEffect(int miscReg, ThreadContext * tc)
       case MISCREG_PCR:
       case MISCREG_PIC:
         panic("Performance Instrumentation not impl\n");
-        /** Floating Point Status Register */
-      case MISCREG_FSR:
-        warn("Reading FSR Floating Point not implemented\n");
-        break;
       case MISCREG_SOFTINT_CLR:
       case MISCREG_SOFTINT_SET:
         panic("Can read from softint clr/set\n");
@@ -488,6 +485,7 @@ void MiscRegFile::setReg(int miscReg, const MiscReg &val)
         /** Floating Point Status Register */
       case MISCREG_FSR:
         fsr = val;
+        DPRINTF(Sparc, "FSR written with: %#x\n", fsr);
         break;
 
       case MISCREG_MMU_P_CONTEXT:
diff --git a/src/arch/sparc/tlb.cc b/src/arch/sparc/tlb.cc
index ebc8c0e7a..293f667d6 100644
--- a/src/arch/sparc/tlb.cc
+++ b/src/arch/sparc/tlb.cc
@@ -668,8 +668,6 @@ DTB::translate(RequestPtr &req, ThreadContext *tc, bool write)
     if (!implicit && asi != ASI_P && asi != ASI_S) {
         if (AsiIsLittle(asi))
             panic("Little Endian ASIs not supported\n");
-        if (AsiIsBlock(asi))
-            panic("Block ASIs not supported\n");
         if (AsiIsNoFault(asi))
             panic("No Fault ASIs not supported\n");
 
@@ -688,7 +686,7 @@ DTB::translate(RequestPtr &req, ThreadContext *tc, bool write)
             goto handleSparcErrorRegAccess;
 
         if (!AsiIsReal(asi) && !AsiIsNucleus(asi) && !AsiIsAsIfUser(asi) &&
-                !AsiIsTwin(asi))
+                !AsiIsTwin(asi) && !AsiIsBlock(asi))
             panic("Accessing ASI %#X. Should we?\n", asi);
     }
 
diff --git a/src/cpu/exetrace.cc b/src/cpu/exetrace.cc
index e34ae3731..5108d7338 100644
--- a/src/cpu/exetrace.cc
+++ b/src/cpu/exetrace.cc
@@ -411,8 +411,14 @@ Trace::InstRecord::dump(ostream &outs)
                     if(shared_data->y !=
                             thread->readIntReg(NumIntArchRegs + 1))
                         diffY = true;
-                    if(shared_data->fsr != thread->readMiscReg(MISCREG_FSR))
+                    if(shared_data->fsr != thread->readMiscReg(MISCREG_FSR)) {
                         diffFsr = true;
+                        if (mbits(shared_data->fsr, 63,10) ==
+                                mbits(thread->readMiscReg(MISCREG_FSR), 63,10)) {
+                            thread->setMiscReg(MISCREG_FSR, shared_data->fsr);
+                            diffFsr = false;
+                        }
+                    }
                     //if(shared_data->ccr != thread->readMiscReg(MISCREG_CCR))
                     if(shared_data->ccr !=
                             thread->readIntReg(NumIntArchRegs + 2))
@@ -450,16 +456,13 @@ Trace::InstRecord::dump(ostream &outs)
                                 diffTlb = true;
                     }
 
-                    if ((diffPC || diffCC || diffInst || diffIntRegs ||
+                    if (diffPC || diffCC || diffInst || diffIntRegs ||
                          diffFpRegs || diffTpc || diffTnpc || diffTstate ||
                          diffTt || diffHpstate || diffHtstate || diffHtba ||
                          diffPstate || diffY || diffCcr || diffTl || diffFsr ||
                          diffGl || diffAsi || diffPil || diffCwp || diffCansave ||
                          diffCanrestore || diffOtherwin || diffCleanwin || diffTlb)
-                        && !((staticInst->machInst & 0xC1F80000) == 0x81D00000)
-                        && !(((staticInst->machInst & 0xC0000000) == 0xC0000000)
-                            && shared_data->tl == thread->readMiscReg(MISCREG_TL) + 1)
-                       ) {
+                       {
 
                         outs << "Differences found between M5 and Legion:";
                         if (diffPC)
@@ -639,7 +642,7 @@ Trace::InstRecord::dump(ostream &outs)
                                 char label[8];
                                 sprintf(label, "%%f%d", x);
                                 printRegPair(outs, label,
-                                 thread->readFloatRegBits(x,FloatRegFile::DoubleWidth),
+                                 thread->readFloatRegBits(x*2,FloatRegFile::DoubleWidth),
                                  shared_data->fpregs[x]);
                             }
                         }
@@ -667,7 +670,7 @@ Trace::InstRecord::dump(ostream &outs)
                         }
 
                         diffcount++;
-                        if (diffcount > 2)
+                        if (diffcount > 3)
                             fatal("Differences found between Legion and M5\n");
                     } else
                         diffcount = 0;
diff --git a/src/dev/sparc/mm_disk.cc b/src/dev/sparc/mm_disk.cc
index 018415f6c..b8cabd0cf 100644
--- a/src/dev/sparc/mm_disk.cc
+++ b/src/dev/sparc/mm_disk.cc
@@ -47,7 +47,7 @@
 MmDisk::MmDisk(Params *p)
     : BasicPioDevice(p), image(p->image), curSector((uint64_t)-1), dirty(false)
 {
-    std::memset(&bytes, 0, SectorSize);
+    std::memset(&diskData, 0, SectorSize);
     pioSize = image->size() * SectorSize;
 }
 
@@ -57,9 +57,9 @@ MmDisk::read(PacketPtr pkt)
     Addr accessAddr;
     off_t sector;
     off_t bytes_read;
-    uint16_t *d16;
-    uint32_t *d32;
-    uint64_t *d64;
+    uint16_t d16;
+    uint32_t d32;
+    uint64_t d64;
 
     assert(pkt->result == Packet::Unknown);
     assert(pkt->getAddr() >= pioAddr && pkt->getAddr() < pioAddr + pioSize);
@@ -68,26 +68,34 @@ MmDisk::read(PacketPtr pkt)
     sector = accessAddr / SectorSize;
 
     if (sector != curSector) {
-        if (dirty)
-            bytes_read = image->write(bytes, curSector);
-        bytes_read = image->read(bytes,  sector);
+        if (dirty) {
+            bytes_read = image->write(diskData, curSector);
+            assert(bytes_read == SectorSize);
+        }
+        bytes_read = image->read(diskData,  sector);
+        assert(bytes_read == SectorSize);
         curSector = sector;
     }
     switch (pkt->getSize()) {
       case sizeof(uint8_t):
-        pkt->set(bytes[accessAddr % SectorSize]);
+        pkt->set(diskData[accessAddr % SectorSize]);
+        DPRINTF(IdeDisk, "reading byte %#x value= %#x\n", accessAddr, diskData[accessAddr %
+                SectorSize]);
         break;
       case sizeof(uint16_t):
-        d16 = (uint16_t*)bytes + (accessAddr % SectorSize)/2;
-        pkt->set(htobe(*d16));
+        memcpy(&d16, diskData + (accessAddr % SectorSize), 2);
+        pkt->set(htobe(d32));
+        DPRINTF(IdeDisk, "reading word %#x value= %#x\n", accessAddr, d16);
         break;
       case sizeof(uint32_t):
-        d32 = (uint32_t*)bytes + (accessAddr % SectorSize)/4;
-        pkt->set(htobe(*d32));
+        memcpy(&d32, diskData + (accessAddr % SectorSize), 4);
+        pkt->set(htobe(d32));
+        DPRINTF(IdeDisk, "reading dword %#x value= %#x\n", accessAddr, d32);
         break;
       case sizeof(uint64_t):
-        d64 = (uint64_t*)bytes + (accessAddr % SectorSize)/8;
-        pkt->set(htobe(*d64));
+        memcpy(&d64, diskData + (accessAddr % SectorSize), 8);
+        pkt->set(htobe(d64));
+        DPRINTF(IdeDisk, "reading qword %#x value= %#x\n", accessAddr, d64);
         break;
       default:
         panic("Invalid access size\n");
@@ -100,11 +108,74 @@ MmDisk::read(PacketPtr pkt)
 Tick
 MmDisk::write(PacketPtr pkt)
 {
-   panic("need to implement\n");
-   M5_DUMMY_RETURN
+    Addr accessAddr;
+    off_t sector;
+    off_t bytes_read;
+    uint16_t d16;
+    uint32_t d32;
+    uint64_t d64;
+
+    assert(pkt->result == Packet::Unknown);
+    assert(pkt->getAddr() >= pioAddr && pkt->getAddr() < pioAddr + pioSize);
+    accessAddr = pkt->getAddr() - pioAddr;
+
+    sector = accessAddr / SectorSize;
+
+    if (sector != curSector) {
+        if (dirty) {
+            bytes_read = image->write(diskData, curSector);
+            assert(bytes_read == SectorSize);
+        }
+        bytes_read = image->read(diskData,  sector);
+        assert(bytes_read == SectorSize);
+        curSector = sector;
+    }
+    dirty = true;
+
+    switch (pkt->getSize()) {
+      case sizeof(uint8_t):
+        diskData[accessAddr % SectorSize] = htobe(pkt->get<uint8_t>());
+        DPRINTF(IdeDisk, "writing byte %#x value= %#x\n", accessAddr, diskData[accessAddr %
+                SectorSize]);
+        break;
+      case sizeof(uint16_t):
+        d16 = htobe(pkt->get<uint16_t>());
+        memcpy(diskData + (accessAddr % SectorSize), &d16, 2);
+        DPRINTF(IdeDisk, "writing word %#x value= %#x\n", accessAddr, d16);
+        break;
+      case sizeof(uint32_t):
+        d32 = htobe(pkt->get<uint32_t>());
+        memcpy(diskData + (accessAddr % SectorSize), &d32, 4);
+        DPRINTF(IdeDisk, "writing dword %#x value= %#x\n", accessAddr, d32);
+        break;
+      case sizeof(uint64_t):
+        d64 = htobe(pkt->get<uint64_t>());
+        memcpy(diskData + (accessAddr % SectorSize), &d64, 8);
+        DPRINTF(IdeDisk, "writing qword %#x value= %#x\n", accessAddr, d64);
+        break;
+      default:
+        panic("Invalid access size\n");
+    }
+
+    pkt->result = Packet::Success;
+    return pioDelay;
+}
+
+void
+MmDisk::serialize(std::ostream &os)
+{
+    // just write any dirty changes to the cow layer it will take care of
+    // serialization
+    int bytes_read;
+    if (dirty) {
+        bytes_read = image->write(diskData, curSector);
+        assert(bytes_read == SectorSize);
+    }
 }
 
 
+
+
 BEGIN_DECLARE_SIM_OBJECT_PARAMS(MmDisk)
     Param<Addr> pio_addr;
     Param<Tick> pio_latency;
diff --git a/src/dev/sparc/mm_disk.hh b/src/dev/sparc/mm_disk.hh
index 0a4626067..30028d2b6 100644
--- a/src/dev/sparc/mm_disk.hh
+++ b/src/dev/sparc/mm_disk.hh
@@ -46,10 +46,7 @@ class MmDisk : public BasicPioDevice
     DiskImage *image;
     off_t curSector;
     bool dirty;
-    union {
-        uint8_t bytes[SectorSize];
-        uint32_t words[SectorSize/4];
-    };
+    uint8_t diskData[SectorSize];
 
   public:
     struct Params : public BasicPioDevice::Params
@@ -64,6 +61,8 @@ class MmDisk : public BasicPioDevice
 
     virtual Tick read(PacketPtr pkt);
     virtual Tick write(PacketPtr pkt);
+
+    virtual void serialize(std::ostream &os);
 };
 
 #endif //__DEV_SPARC_MM_DISK_HH__