From 04c2d491c4bbb424a59273d4ebee62ddfe3379f9 Mon Sep 17 00:00:00 2001
From: Andrew Waterman <waterman@cs.berkeley.edu>
Date: Sun, 11 Aug 2013 19:10:51 -0700
Subject: [PATCH] Instructions are no longer member functions

---
 configure               |   2 +-
 configure.ac            |   2 +-
 riscv/decode.h          |  35 ++++----
 riscv/htif.cc           |   6 +-
 riscv/insn_template.cc  |   4 +-
 riscv/insns/amoadd_d.h  |   4 +-
 riscv/insns/amoadd_w.h  |   4 +-
 riscv/insns/amoand_d.h  |   4 +-
 riscv/insns/amoand_w.h  |   4 +-
 riscv/insns/amomax_d.h  |   4 +-
 riscv/insns/amomax_w.h  |   4 +-
 riscv/insns/amomaxu_d.h |   4 +-
 riscv/insns/amomaxu_w.h |   4 +-
 riscv/insns/amomin_d.h  |   4 +-
 riscv/insns/amomin_w.h  |   4 +-
 riscv/insns/amominu_d.h |   4 +-
 riscv/insns/amominu_w.h |   4 +-
 riscv/insns/amoor_d.h   |   4 +-
 riscv/insns/amoor_w.h   |   4 +-
 riscv/insns/amoswap_d.h |   4 +-
 riscv/insns/amoswap_w.h |   4 +-
 riscv/insns/break.h     |   2 +-
 riscv/insns/clearpcr.h  |   4 +-
 riscv/insns/eret.h      |   8 +-
 riscv/insns/fence_i.h   |   2 +-
 riscv/insns/fld.h       |   2 +-
 riscv/insns/flw.h       |   2 +-
 riscv/insns/frsr.h      |   2 +-
 riscv/insns/fsd.h       |   2 +-
 riscv/insns/fssr.h      |   4 +-
 riscv/insns/fsw.h       |   2 +-
 riscv/insns/lb.h        |   2 +-
 riscv/insns/lbu.h       |   2 +-
 riscv/insns/ld.h        |   2 +-
 riscv/insns/lh.h        |   2 +-
 riscv/insns/lhu.h       |   2 +-
 riscv/insns/lr_d.h      |   3 +-
 riscv/insns/lr_w.h      |   3 +-
 riscv/insns/lw.h        |   2 +-
 riscv/insns/lwu.h       |   2 +-
 riscv/insns/mfpcr.h     |   2 +-
 riscv/insns/mtpcr.h     |   4 +-
 riscv/insns/rdcycle.h   |   2 +-
 riscv/insns/rdinstret.h |   2 +-
 riscv/insns/rdtime.h    |   2 +-
 riscv/insns/sb.h        |   2 +-
 riscv/insns/sc_d.h      |   8 +-
 riscv/insns/sc_w.h      |   8 +-
 riscv/insns/sd.h        |   2 +-
 riscv/insns/setpcr.h    |   4 +-
 riscv/insns/sh.h        |   2 +-
 riscv/insns/slli.h      |   2 +-
 riscv/insns/srai.h      |   2 +-
 riscv/insns/srli.h      |   2 +-
 riscv/insns/sw.h        |   2 +-
 riscv/insns/syscall.h   |   2 +-
 riscv/interactive.cc    |  29 +++----
 riscv/mmu.cc            |  17 ++--
 riscv/mmu.h             |  27 +-----
 riscv/processor.cc      | 181 ++++++++++++++++++++++------------------
 riscv/processor.h       |  69 ++++++++-------
 riscv/sim.cc            |  18 ++--
 riscv/sim.h             |   2 +-
 riscv/trap.cc           |  17 ++--
 riscv/trap.h            |  75 +++++++++++------
 65 files changed, 337 insertions(+), 309 deletions(-)

diff --git a/configure b/configure
index 32f29c2..9340bfb 100755
--- a/configure
+++ b/configure
@@ -4004,7 +4004,7 @@ fi
 
 CFLAGS="-Wall -O2 -Wno-unused"
 
-CXXFLAGS="-Wall -O2 -std=c++0x -Wno-pmf-conversions"
+CXXFLAGS="-Wall -O2 -std=c++0x"
 
 
 #-------------------------------------------------------------------------
diff --git a/configure.ac b/configure.ac
index bb672e6..335100b 100644
--- a/configure.ac
+++ b/configure.ac
@@ -72,7 +72,7 @@ AC_HEADER_STDC
 #-------------------------------------------------------------------------
 
 AC_SUBST([CFLAGS],  ["-Wall -O2 -Wno-unused"])
-AC_SUBST([CXXFLAGS],["-Wall -O2 -std=c++0x -Wno-pmf-conversions"])
+AC_SUBST([CXXFLAGS],["-Wall -O2 -std=c++0x"])
 
 #-------------------------------------------------------------------------
 # MCPPBS subproject list
diff --git a/riscv/decode.h b/riscv/decode.h
index 5f65c27..1529927 100644
--- a/riscv/decode.h
+++ b/riscv/decode.h
@@ -168,14 +168,15 @@ private:
 };
 
 // helpful macros, etc
-#define RS1 XPR[insn.rtype.rs1]
-#define RS2 XPR[insn.rtype.rs2]
-#define RD XPR.write_port(insn.rtype.rd)
-#define RA XPR.write_port(1)
-#define FRS1 FPR[insn.ftype.rs1]
-#define FRS2 FPR[insn.ftype.rs2]
-#define FRS3 FPR[insn.ftype.rs3]
-#define FRD FPR.write_port(insn.ftype.rd)
+#define MMU (*p->get_mmu())
+#define RS1 p->get_state()->XPR[insn.rtype.rs1]
+#define RS2 p->get_state()->XPR[insn.rtype.rs2]
+#define RD p->get_state()->XPR.write_port(insn.rtype.rd)
+#define RA p->get_state()->XPR.write_port(1)
+#define FRS1 p->get_state()->FPR[insn.ftype.rs1]
+#define FRS2 p->get_state()->FPR[insn.ftype.rs2]
+#define FRS3 p->get_state()->FPR[insn.ftype.rs3]
+#define FRD p->get_state()->FPR.write_port(insn.ftype.rd)
 #define BIGIMM insn.ltype.bigimm
 #define SIMM insn.itype.imm12
 #define BIMM ((signed)insn.btype.immlo | (insn.btype.immhi << IMMLO_BITS))
@@ -187,23 +188,23 @@ private:
 #define ITYPE_EADDR sext_xprlen(RS1 + SIMM)
 #define BTYPE_EADDR sext_xprlen(RS1 + BIMM)
 #define RM ({ int rm = insn.ftype.rm; \
-              if(rm == 7) rm = (fsr & FSR_RD) >> FSR_RD_SHIFT; \
-              if(rm > 4) throw trap_illegal_instruction; \
+              if(rm == 7) rm = (p->get_state()->fsr & FSR_RD) >> FSR_RD_SHIFT; \
+              if(rm > 4) throw trap_illegal_instruction(); \
               rm; })
 
 #define xpr64 (xprlen == 64)
 
-#define require_supervisor if(unlikely(!(sr & SR_S))) throw trap_privileged_instruction
-#define require_xpr64 if(unlikely(!xpr64)) throw trap_illegal_instruction
-#define require_xpr32 if(unlikely(xpr64)) throw trap_illegal_instruction
+#define require_supervisor if(unlikely(!(p->get_state()->sr & SR_S))) throw trap_privileged_instruction()
+#define require_xpr64 if(unlikely(!xpr64)) throw trap_illegal_instruction()
+#define require_xpr32 if(unlikely(xpr64)) throw trap_illegal_instruction()
 #ifndef RISCV_ENABLE_FPU
-# define require_fp throw trap_illegal_instruction
+# define require_fp throw trap_illegal_instruction()
 #else
-# define require_fp if(unlikely(!(sr & SR_EF))) throw trap_fp_disabled
+# define require_fp if(unlikely(!(p->get_state()->sr & SR_EF))) throw trap_fp_disabled()
 #endif
 
 #define cmp_trunc(reg) (reg_t(reg) << (64-xprlen))
-#define set_fp_exceptions ({ set_fsr(fsr | \
+#define set_fp_exceptions ({ p->set_fsr(p->get_state()->fsr | \
                                (softfloat_exceptionFlags << FSR_AEXC_SHIFT)); \
                              softfloat_exceptionFlags = 0; })
 
@@ -220,7 +221,7 @@ private:
 
 #define set_pc(x) \
   do { if ((x) & 3 /* For now... */) \
-         throw trap_instruction_address_misaligned; \
+         throw trap_instruction_address_misaligned(); \
        npc = (x); \
      } while(0)
 
diff --git a/riscv/htif.cc b/riscv/htif.cc
index fc90375..679ef93 100644
--- a/riscv/htif.cc
+++ b/riscv/htif.cc
@@ -41,7 +41,7 @@ void htif_isasim_t::tick_once()
 
       uint64_t buf[hdr.data_size];
       for (size_t i = 0; i < hdr.data_size; i++)
-        buf[i] = sim->mmu->load_uint64((hdr.addr+i)*HTIF_DATA_ALIGN);
+        buf[i] = sim->debug_mmu->load_uint64((hdr.addr+i)*HTIF_DATA_ALIGN);
       send(buf, hdr.data_size * sizeof(buf[0]));
       break;
     }
@@ -49,7 +49,7 @@ void htif_isasim_t::tick_once()
     {
       const uint64_t* buf = (const uint64_t*)p.get_payload();
       for (size_t i = 0; i < hdr.data_size; i++)
-        sim->mmu->store_uint64((hdr.addr+i)*HTIF_DATA_ALIGN, buf[i]);
+        sim->debug_mmu->store_uint64((hdr.addr+i)*HTIF_DATA_ALIGN, buf[i]);
 
       packet_header_t ack(HTIF_CMD_ACK, seqno, 0, 0);
       send(&ack, sizeof(ack));
@@ -77,7 +77,7 @@ void htif_isasim_t::tick_once()
       send(&old_val, sizeof(old_val));
 
       if (regno == PCR_TOHOST)
-          sim->procs[coreid]->tohost = 0;
+          sim->procs[coreid]->state.tohost = 0;
 
       if (hdr.cmd == HTIF_CMD_WRITE_CONTROL_REG)
       {
diff --git a/riscv/insn_template.cc b/riscv/insn_template.cc
index 5bd986e..a901077 100644
--- a/riscv/insn_template.cc
+++ b/riscv/insn_template.cc
@@ -8,7 +8,7 @@
 #include "internals.h" // ditto
 #include <assert.h>
 
-reg_t processor_t::rv32_NAME(insn_t insn, reg_t pc)
+reg_t rv32_NAME(processor_t* p, insn_t insn, reg_t pc)
 {
   int xprlen = 32;
   reg_t npc = sext_xprlen(pc + insn_length(OPCODE));
@@ -16,7 +16,7 @@ reg_t processor_t::rv32_NAME(insn_t insn, reg_t pc)
   return npc;
 }
 
-reg_t processor_t::rv64_NAME(insn_t insn, reg_t pc)
+reg_t rv64_NAME(processor_t* p, insn_t insn, reg_t pc)
 {
   int xprlen = 64;
   reg_t npc = sext_xprlen(pc + insn_length(OPCODE));
diff --git a/riscv/insns/amoadd_d.h b/riscv/insns/amoadd_d.h
index b8450bf..bba975c 100644
--- a/riscv/insns/amoadd_d.h
+++ b/riscv/insns/amoadd_d.h
@@ -1,4 +1,4 @@
 require_xpr64;
-reg_t v = mmu.load_uint64(RS1);
-mmu.store_uint64(RS1, RS2 + v);
+reg_t v = MMU.load_uint64(RS1);
+MMU.store_uint64(RS1, RS2 + v);
 RD = v;
diff --git a/riscv/insns/amoadd_w.h b/riscv/insns/amoadd_w.h
index 033b3c8..07c9c9a 100644
--- a/riscv/insns/amoadd_w.h
+++ b/riscv/insns/amoadd_w.h
@@ -1,3 +1,3 @@
-reg_t v = mmu.load_int32(RS1);
-mmu.store_uint32(RS1, RS2 + v);
+reg_t v = MMU.load_int32(RS1);
+MMU.store_uint32(RS1, RS2 + v);
 RD = v;
diff --git a/riscv/insns/amoand_d.h b/riscv/insns/amoand_d.h
index 586eb7f..1bb3402 100644
--- a/riscv/insns/amoand_d.h
+++ b/riscv/insns/amoand_d.h
@@ -1,4 +1,4 @@
 require_xpr64;
-reg_t v = mmu.load_uint64(RS1);
-mmu.store_uint64(RS1, RS2 & v);
+reg_t v = MMU.load_uint64(RS1);
+MMU.store_uint64(RS1, RS2 & v);
 RD = v;
diff --git a/riscv/insns/amoand_w.h b/riscv/insns/amoand_w.h
index 18a9249..91866dc 100644
--- a/riscv/insns/amoand_w.h
+++ b/riscv/insns/amoand_w.h
@@ -1,3 +1,3 @@
-reg_t v = mmu.load_int32(RS1);
-mmu.store_uint32(RS1, RS2 & v);
+reg_t v = MMU.load_int32(RS1);
+MMU.store_uint32(RS1, RS2 & v);
 RD = v;
diff --git a/riscv/insns/amomax_d.h b/riscv/insns/amomax_d.h
index 1a0bc8a..dfd2b33 100644
--- a/riscv/insns/amomax_d.h
+++ b/riscv/insns/amomax_d.h
@@ -1,4 +1,4 @@
 require_xpr64;
-sreg_t v = mmu.load_int64(RS1);
-mmu.store_uint64(RS1, std::max(sreg_t(RS2),v));
+sreg_t v = MMU.load_int64(RS1);
+MMU.store_uint64(RS1, std::max(sreg_t(RS2),v));
 RD = v;
diff --git a/riscv/insns/amomax_w.h b/riscv/insns/amomax_w.h
index ff9c2da..1f68a8b 100644
--- a/riscv/insns/amomax_w.h
+++ b/riscv/insns/amomax_w.h
@@ -1,3 +1,3 @@
-int32_t v = mmu.load_int32(RS1);
-mmu.store_uint32(RS1, std::max(int32_t(RS2),v));
+int32_t v = MMU.load_int32(RS1);
+MMU.store_uint32(RS1, std::max(int32_t(RS2),v));
 RD = v;
diff --git a/riscv/insns/amomaxu_d.h b/riscv/insns/amomaxu_d.h
index ccfaf1d..8d50a0a 100644
--- a/riscv/insns/amomaxu_d.h
+++ b/riscv/insns/amomaxu_d.h
@@ -1,4 +1,4 @@
 require_xpr64;
-reg_t v = mmu.load_uint64(RS1);
-mmu.store_uint64(RS1, std::max(RS2,v));
+reg_t v = MMU.load_uint64(RS1);
+MMU.store_uint64(RS1, std::max(RS2,v));
 RD = v;
diff --git a/riscv/insns/amomaxu_w.h b/riscv/insns/amomaxu_w.h
index 075847d..d507e4f 100644
--- a/riscv/insns/amomaxu_w.h
+++ b/riscv/insns/amomaxu_w.h
@@ -1,3 +1,3 @@
-uint32_t v = mmu.load_int32(RS1);
-mmu.store_uint32(RS1, std::max(uint32_t(RS2),v));
+uint32_t v = MMU.load_int32(RS1);
+MMU.store_uint32(RS1, std::max(uint32_t(RS2),v));
 RD = (int32_t)v;
diff --git a/riscv/insns/amomin_d.h b/riscv/insns/amomin_d.h
index 4f3b6d6..a20ace8 100644
--- a/riscv/insns/amomin_d.h
+++ b/riscv/insns/amomin_d.h
@@ -1,4 +1,4 @@
 require_xpr64;
-sreg_t v = mmu.load_int64(RS1);
-mmu.store_uint64(RS1, std::min(sreg_t(RS2),v));
+sreg_t v = MMU.load_int64(RS1);
+MMU.store_uint64(RS1, std::min(sreg_t(RS2),v));
 RD = v;
diff --git a/riscv/insns/amomin_w.h b/riscv/insns/amomin_w.h
index 529ad50..d8f95af 100644
--- a/riscv/insns/amomin_w.h
+++ b/riscv/insns/amomin_w.h
@@ -1,3 +1,3 @@
-int32_t v = mmu.load_int32(RS1);
-mmu.store_uint32(RS1, std::min(int32_t(RS2),v));
+int32_t v = MMU.load_int32(RS1);
+MMU.store_uint32(RS1, std::min(int32_t(RS2),v));
 RD = v;
diff --git a/riscv/insns/amominu_d.h b/riscv/insns/amominu_d.h
index c09c51a..4f83c0f 100644
--- a/riscv/insns/amominu_d.h
+++ b/riscv/insns/amominu_d.h
@@ -1,4 +1,4 @@
 require_xpr64;
-reg_t v = mmu.load_uint64(RS1);
-mmu.store_uint64(RS1, std::min(RS2,v));
+reg_t v = MMU.load_uint64(RS1);
+MMU.store_uint64(RS1, std::min(RS2,v));
 RD = v;
diff --git a/riscv/insns/amominu_w.h b/riscv/insns/amominu_w.h
index d8d6377..a3a537a 100644
--- a/riscv/insns/amominu_w.h
+++ b/riscv/insns/amominu_w.h
@@ -1,3 +1,3 @@
-uint32_t v = mmu.load_int32(RS1);
-mmu.store_uint32(RS1, std::min(uint32_t(RS2),v));
+uint32_t v = MMU.load_int32(RS1);
+MMU.store_uint32(RS1, std::min(uint32_t(RS2),v));
 RD = (int32_t)v;
diff --git a/riscv/insns/amoor_d.h b/riscv/insns/amoor_d.h
index 76a4508..87b6f2a 100644
--- a/riscv/insns/amoor_d.h
+++ b/riscv/insns/amoor_d.h
@@ -1,4 +1,4 @@
 require_xpr64;
-reg_t v = mmu.load_uint64(RS1);
-mmu.store_uint64(RS1, RS2 | v);
+reg_t v = MMU.load_uint64(RS1);
+MMU.store_uint64(RS1, RS2 | v);
 RD = v;
diff --git a/riscv/insns/amoor_w.h b/riscv/insns/amoor_w.h
index 741fbef..0733fad 100644
--- a/riscv/insns/amoor_w.h
+++ b/riscv/insns/amoor_w.h
@@ -1,3 +1,3 @@
-reg_t v = mmu.load_int32(RS1);
-mmu.store_uint32(RS1, RS2 | v);
+reg_t v = MMU.load_int32(RS1);
+MMU.store_uint32(RS1, RS2 | v);
 RD = v;
diff --git a/riscv/insns/amoswap_d.h b/riscv/insns/amoswap_d.h
index 43e3538..3423b91 100644
--- a/riscv/insns/amoswap_d.h
+++ b/riscv/insns/amoswap_d.h
@@ -1,4 +1,4 @@
 require_xpr64;
-reg_t v = mmu.load_uint64(RS1);
-mmu.store_uint64(RS1, RS2);
+reg_t v = MMU.load_uint64(RS1);
+MMU.store_uint64(RS1, RS2);
 RD = v;
diff --git a/riscv/insns/amoswap_w.h b/riscv/insns/amoswap_w.h
index 30e6102..b888235 100644
--- a/riscv/insns/amoswap_w.h
+++ b/riscv/insns/amoswap_w.h
@@ -1,3 +1,3 @@
-reg_t v = mmu.load_int32(RS1);
-mmu.store_uint32(RS1, RS2);
+reg_t v = MMU.load_int32(RS1);
+MMU.store_uint32(RS1, RS2);
 RD = v;
diff --git a/riscv/insns/break.h b/riscv/insns/break.h
index 7fd3d66..c22776c 100644
--- a/riscv/insns/break.h
+++ b/riscv/insns/break.h
@@ -1 +1 @@
-throw trap_breakpoint;
+throw trap_breakpoint();
diff --git a/riscv/insns/clearpcr.h b/riscv/insns/clearpcr.h
index 7acf221..56c3584 100644
--- a/riscv/insns/clearpcr.h
+++ b/riscv/insns/clearpcr.h
@@ -1,4 +1,2 @@
 require_supervisor;
-reg_t temp = get_pcr(insn.rtype.rs1);
-set_pcr(insn.rtype.rs1, temp & ~SIMM);
-RD = temp;
+RD = p->set_pcr(insn.rtype.rs1, p->get_pcr(insn.rtype.rs1) & ~SIMM);
diff --git a/riscv/insns/eret.h b/riscv/insns/eret.h
index 9ea8baf..d4517ee 100644
--- a/riscv/insns/eret.h
+++ b/riscv/insns/eret.h
@@ -1,5 +1,5 @@
 require_supervisor;
-set_pcr(PCR_SR, ((sr & ~(SR_S | SR_EI)) |
-                ((sr & SR_PS) ? SR_S : 0)) |
-                ((sr & SR_PEI) ? SR_EI : 0));
-set_pc(epc);
+p->set_pcr(PCR_SR, ((p->get_state()->sr & ~(SR_S | SR_EI)) |
+                ((p->get_state()->sr & SR_PS) ? SR_S : 0)) |
+                ((p->get_state()->sr & SR_PEI) ? SR_EI : 0));
+set_pc(p->get_state()->epc);
diff --git a/riscv/insns/fence_i.h b/riscv/insns/fence_i.h
index a2dbffe..38dcaf3 100644
--- a/riscv/insns/fence_i.h
+++ b/riscv/insns/fence_i.h
@@ -1 +1 @@
-mmu.flush_icache();
+MMU.flush_icache();
diff --git a/riscv/insns/fld.h b/riscv/insns/fld.h
index bc8b9c7..2704a4d 100644
--- a/riscv/insns/fld.h
+++ b/riscv/insns/fld.h
@@ -1,2 +1,2 @@
 require_fp;
-FRD = mmu.load_int64(ITYPE_EADDR);
+FRD = MMU.load_int64(ITYPE_EADDR);
diff --git a/riscv/insns/flw.h b/riscv/insns/flw.h
index 74374b9..afab636 100644
--- a/riscv/insns/flw.h
+++ b/riscv/insns/flw.h
@@ -1,2 +1,2 @@
 require_fp;
-FRD = mmu.load_int32(ITYPE_EADDR);
+FRD = MMU.load_int32(ITYPE_EADDR);
diff --git a/riscv/insns/frsr.h b/riscv/insns/frsr.h
index 29debc4..ef121e3 100644
--- a/riscv/insns/frsr.h
+++ b/riscv/insns/frsr.h
@@ -1,2 +1,2 @@
 require_fp;
-RD = fsr;
+RD = p->get_fsr();
diff --git a/riscv/insns/fsd.h b/riscv/insns/fsd.h
index 95fbd26..0e1c38a 100644
--- a/riscv/insns/fsd.h
+++ b/riscv/insns/fsd.h
@@ -1,2 +1,2 @@
 require_fp;
-mmu.store_uint64(BTYPE_EADDR, FRS2);
+MMU.store_uint64(BTYPE_EADDR, FRS2);
diff --git a/riscv/insns/fssr.h b/riscv/insns/fssr.h
index cc6f9ea..a9acca6 100644
--- a/riscv/insns/fssr.h
+++ b/riscv/insns/fssr.h
@@ -1,4 +1,2 @@
 require_fp;
-uint32_t tmp = fsr;
-set_fsr(RS1);
-RD = tmp;
+RD = p->set_fsr(RS1);
diff --git a/riscv/insns/fsw.h b/riscv/insns/fsw.h
index 59d9066..c921123 100644
--- a/riscv/insns/fsw.h
+++ b/riscv/insns/fsw.h
@@ -1,2 +1,2 @@
 require_fp;
-mmu.store_uint32(BTYPE_EADDR, FRS2);
+MMU.store_uint32(BTYPE_EADDR, FRS2);
diff --git a/riscv/insns/lb.h b/riscv/insns/lb.h
index c88ee2d..56a5f32 100644
--- a/riscv/insns/lb.h
+++ b/riscv/insns/lb.h
@@ -1 +1 @@
-RD = mmu.load_int8(ITYPE_EADDR);
+RD = MMU.load_int8(ITYPE_EADDR);
diff --git a/riscv/insns/lbu.h b/riscv/insns/lbu.h
index bcef96b..66621c0 100644
--- a/riscv/insns/lbu.h
+++ b/riscv/insns/lbu.h
@@ -1 +1 @@
-RD = mmu.load_uint8(ITYPE_EADDR);
+RD = MMU.load_uint8(ITYPE_EADDR);
diff --git a/riscv/insns/ld.h b/riscv/insns/ld.h
index 2bbe5c3..f214294 100644
--- a/riscv/insns/ld.h
+++ b/riscv/insns/ld.h
@@ -1,2 +1,2 @@
 require_xpr64;
-RD = mmu.load_int64(ITYPE_EADDR);
+RD = MMU.load_int64(ITYPE_EADDR);
diff --git a/riscv/insns/lh.h b/riscv/insns/lh.h
index c04302c..fea2a8e 100644
--- a/riscv/insns/lh.h
+++ b/riscv/insns/lh.h
@@ -1 +1 @@
-RD = mmu.load_int16(ITYPE_EADDR);
+RD = MMU.load_int16(ITYPE_EADDR);
diff --git a/riscv/insns/lhu.h b/riscv/insns/lhu.h
index 99d0985..71c21be 100644
--- a/riscv/insns/lhu.h
+++ b/riscv/insns/lhu.h
@@ -1 +1 @@
-RD = mmu.load_uint16(ITYPE_EADDR);
+RD = MMU.load_uint16(ITYPE_EADDR);
diff --git a/riscv/insns/lr_d.h b/riscv/insns/lr_d.h
index 5c8eff1..3d2aace 100644
--- a/riscv/insns/lr_d.h
+++ b/riscv/insns/lr_d.h
@@ -1,2 +1,3 @@
 require_xpr64;
-RD = mmu.load_reserved_int64(RS1);
+p->get_state()->load_reservation = RS1;
+RD = MMU.load_int64(RS1);
diff --git a/riscv/insns/lr_w.h b/riscv/insns/lr_w.h
index 3ac4746..7ff48ea 100644
--- a/riscv/insns/lr_w.h
+++ b/riscv/insns/lr_w.h
@@ -1 +1,2 @@
-RD = mmu.load_reserved_int32(RS1);
+p->get_state()->load_reservation = RS1;
+RD = MMU.load_int32(RS1);
diff --git a/riscv/insns/lw.h b/riscv/insns/lw.h
index 639d0e7..77f735e 100644
--- a/riscv/insns/lw.h
+++ b/riscv/insns/lw.h
@@ -1 +1 @@
-RD = mmu.load_int32(ITYPE_EADDR);
+RD = MMU.load_int32(ITYPE_EADDR);
diff --git a/riscv/insns/lwu.h b/riscv/insns/lwu.h
index b3ebc0c..e731178 100644
--- a/riscv/insns/lwu.h
+++ b/riscv/insns/lwu.h
@@ -1,2 +1,2 @@
 require_xpr64;
-RD = mmu.load_uint32(ITYPE_EADDR);
+RD = MMU.load_uint32(ITYPE_EADDR);
diff --git a/riscv/insns/mfpcr.h b/riscv/insns/mfpcr.h
index f7aea9f..0f23426 100644
--- a/riscv/insns/mfpcr.h
+++ b/riscv/insns/mfpcr.h
@@ -1,2 +1,2 @@
 require_supervisor;
-RD = get_pcr(insn.rtype.rs1);
+RD = p->get_pcr(insn.rtype.rs1);
diff --git a/riscv/insns/mtpcr.h b/riscv/insns/mtpcr.h
index 5cd0134..770dfd5 100644
--- a/riscv/insns/mtpcr.h
+++ b/riscv/insns/mtpcr.h
@@ -1,4 +1,2 @@
 require_supervisor;
-reg_t val = get_pcr(insn.rtype.rs1);
-set_pcr(insn.rtype.rs1, RS2);
-RD = val;
+RD = p->set_pcr(insn.rtype.rs1, RS2);
diff --git a/riscv/insns/rdcycle.h b/riscv/insns/rdcycle.h
index 9b966a6..7ebe986 100644
--- a/riscv/insns/rdcycle.h
+++ b/riscv/insns/rdcycle.h
@@ -1 +1 @@
-RD = cycle;
+RD = sext_xprlen(p->get_state()->cycle);
diff --git a/riscv/insns/rdinstret.h b/riscv/insns/rdinstret.h
index 9b966a6..df56cb7 100644
--- a/riscv/insns/rdinstret.h
+++ b/riscv/insns/rdinstret.h
@@ -1 +1 @@
-RD = cycle;
+#include "insns/rdcycle.h"
diff --git a/riscv/insns/rdtime.h b/riscv/insns/rdtime.h
index 9b966a6..df56cb7 100644
--- a/riscv/insns/rdtime.h
+++ b/riscv/insns/rdtime.h
@@ -1 +1 @@
-RD = cycle;
+#include "insns/rdcycle.h"
diff --git a/riscv/insns/sb.h b/riscv/insns/sb.h
index 54176c0..db4d523 100644
--- a/riscv/insns/sb.h
+++ b/riscv/insns/sb.h
@@ -1 +1 @@
-mmu.store_uint8(BTYPE_EADDR, RS2);
+MMU.store_uint8(BTYPE_EADDR, RS2);
diff --git a/riscv/insns/sc_d.h b/riscv/insns/sc_d.h
index a29b9f7..9ad962c 100644
--- a/riscv/insns/sc_d.h
+++ b/riscv/insns/sc_d.h
@@ -1,2 +1,8 @@
 require_xpr64;
-RD = mmu.store_conditional_uint64(RS1, RS2);
+if (RS1 == p->get_state()->load_reservation)
+{
+  MMU.store_uint64(RS1, RS2);
+  RD = 0;
+}
+else
+  RD = 1;
diff --git a/riscv/insns/sc_w.h b/riscv/insns/sc_w.h
index caf7683..3ad79ac 100644
--- a/riscv/insns/sc_w.h
+++ b/riscv/insns/sc_w.h
@@ -1 +1,7 @@
-RD = mmu.store_conditional_uint32(RS1, RS2);
+if (RS1 == p->get_state()->load_reservation)
+{
+  MMU.store_uint32(RS1, RS2);
+  RD = 0;
+}
+else
+  RD = 1;
diff --git a/riscv/insns/sd.h b/riscv/insns/sd.h
index 558428e..24c0de9 100644
--- a/riscv/insns/sd.h
+++ b/riscv/insns/sd.h
@@ -1,2 +1,2 @@
 require_xpr64;
-mmu.store_uint64(BTYPE_EADDR, RS2);
+MMU.store_uint64(BTYPE_EADDR, RS2);
diff --git a/riscv/insns/setpcr.h b/riscv/insns/setpcr.h
index d645626..4a25d80 100644
--- a/riscv/insns/setpcr.h
+++ b/riscv/insns/setpcr.h
@@ -1,4 +1,2 @@
 require_supervisor;
-reg_t temp = get_pcr(insn.rtype.rs1);
-set_pcr(insn.rtype.rs1, temp | SIMM);
-RD = temp;
+RD = p->set_pcr(insn.rtype.rs1, p->get_pcr(insn.rtype.rs1) | SIMM);
diff --git a/riscv/insns/sh.h b/riscv/insns/sh.h
index 235e50e..69234dc 100644
--- a/riscv/insns/sh.h
+++ b/riscv/insns/sh.h
@@ -1 +1 @@
-mmu.store_uint16(BTYPE_EADDR, RS2);
+MMU.store_uint16(BTYPE_EADDR, RS2);
diff --git a/riscv/insns/slli.h b/riscv/insns/slli.h
index bfaf430..151d970 100644
--- a/riscv/insns/slli.h
+++ b/riscv/insns/slli.h
@@ -3,6 +3,6 @@ if(xpr64)
 else
 {
   if(SHAMT & 0x20)
-    throw trap_illegal_instruction;
+    throw trap_illegal_instruction();
   RD = sext32(RS1 << SHAMT);
 }
diff --git a/riscv/insns/srai.h b/riscv/insns/srai.h
index bb17d27..7360d5f 100644
--- a/riscv/insns/srai.h
+++ b/riscv/insns/srai.h
@@ -3,6 +3,6 @@ if(xpr64)
 else
 {
   if(SHAMT & 0x20)
-    throw trap_illegal_instruction;
+    throw trap_illegal_instruction();
   RD = sext32(int32_t(RS1) >> SHAMT);
 }
diff --git a/riscv/insns/srli.h b/riscv/insns/srli.h
index 5378fd1..f5b8c02 100644
--- a/riscv/insns/srli.h
+++ b/riscv/insns/srli.h
@@ -3,6 +3,6 @@ if(xpr64)
 else
 {
   if(SHAMT & 0x20)
-    throw trap_illegal_instruction;
+    throw trap_illegal_instruction();
   RD = sext32((uint32_t)RS1 >> SHAMT);
 }
diff --git a/riscv/insns/sw.h b/riscv/insns/sw.h
index 008d8c0..81ca71d 100644
--- a/riscv/insns/sw.h
+++ b/riscv/insns/sw.h
@@ -1 +1 @@
-mmu.store_uint32(BTYPE_EADDR, RS2);
+MMU.store_uint32(BTYPE_EADDR, RS2);
diff --git a/riscv/insns/syscall.h b/riscv/insns/syscall.h
index 2c7199d..b31b7e6 100644
--- a/riscv/insns/syscall.h
+++ b/riscv/insns/syscall.h
@@ -1 +1 @@
-throw trap_syscall;
+throw trap_syscall();
diff --git a/riscv/interactive.cc b/riscv/interactive.cc
index fe70760..33efcfb 100644
--- a/riscv/interactive.cc
+++ b/riscv/interactive.cc
@@ -113,39 +113,39 @@ void sim_t::interactive_quit(const std::string& cmd, const std::vector<std::stri
 reg_t sim_t::get_pc(const std::vector<std::string>& args)
 {
   if(args.size() != 1)
-    throw trap_illegal_instruction;
+    throw trap_illegal_instruction();
 
   int p = atoi(args[0].c_str());
   if(p >= (int)num_cores())
-    throw trap_illegal_instruction;
+    throw trap_illegal_instruction();
 
-  return procs[p]->pc;
+  return procs[p]->state.pc;
 }
 
 reg_t sim_t::get_reg(const std::vector<std::string>& args)
 {
   if(args.size() != 2)
-    throw trap_illegal_instruction;
+    throw trap_illegal_instruction();
 
   int p = atoi(args[0].c_str());
   int r = atoi(args[1].c_str());
   if(p >= (int)num_cores() || r >= NXPR)
-    throw trap_illegal_instruction;
+    throw trap_illegal_instruction();
 
-  return procs[p]->XPR[r];
+  return procs[p]->state.XPR[r];
 }
 
 reg_t sim_t::get_freg(const std::vector<std::string>& args)
 {
   if(args.size() != 2)
-    throw trap_illegal_instruction;
+    throw trap_illegal_instruction();
 
   int p = atoi(args[0].c_str());
   int r = atoi(args[1].c_str());
   if(p >= (int)num_cores() || r >= NFPR)
-    throw trap_illegal_instruction;
+    throw trap_illegal_instruction();
 
-  return procs[p]->FPR[r];
+  return procs[p]->state.FPR[r];
 }
 
 void sim_t::interactive_reg(const std::string& cmd, const std::vector<std::string>& args)
@@ -177,15 +177,16 @@ void sim_t::interactive_fregd(const std::string& cmd, const std::vector<std::str
 reg_t sim_t::get_mem(const std::vector<std::string>& args)
 {
   if(args.size() != 1 && args.size() != 2)
-    throw trap_illegal_instruction;
+    throw trap_illegal_instruction();
 
   std::string addr_str = args[0];
+  mmu_t* mmu = debug_mmu;
   if(args.size() == 2)
   {
     int p = atoi(args[0].c_str());
     if(p >= (int)num_cores())
-      throw trap_illegal_instruction;
-    mmu->set_ptbr(procs[p]->mmu.get_ptbr());
+      throw trap_illegal_instruction();
+    mmu = procs[p]->get_mmu();
     addr_str = args[1];
   }
 
@@ -220,12 +221,12 @@ void sim_t::interactive_mem(const std::string& cmd, const std::vector<std::strin
 void sim_t::interactive_str(const std::string& cmd, const std::vector<std::string>& args)
 {
   if(args.size() != 1)
-    throw trap_illegal_instruction;
+    throw trap_illegal_instruction();
 
   reg_t addr = strtol(args[0].c_str(),NULL,16);
 
   char ch;
-  while((ch = mmu->load_uint8(addr++)))
+  while((ch = debug_mmu->load_uint8(addr++)))
     putchar(ch);
 
   putchar('\n');
diff --git a/riscv/mmu.cc b/riscv/mmu.cc
index 01cbb97..96884d6 100644
--- a/riscv/mmu.cc
+++ b/riscv/mmu.cc
@@ -5,8 +5,7 @@
 #include "processor.h"
 
 mmu_t::mmu_t(char* _mem, size_t _memsz)
- : mem(_mem), memsz(_memsz), badvaddr(0),
-   ptbr(0), proc(NULL)
+ : mem(_mem), memsz(_memsz), proc(NULL)
 {
   flush_tlb();
 }
@@ -27,7 +26,6 @@ void mmu_t::flush_tlb()
   memset(tlb_store_tag, -1, sizeof(tlb_store_tag));
 
   flush_icache();
-  yield_load_reservation();
 }
 
 reg_t mmu_t::refill_tlb(reg_t addr, reg_t bytes, bool store, bool fetch)
@@ -38,7 +36,7 @@ reg_t mmu_t::refill_tlb(reg_t addr, reg_t bytes, bool store, bool fetch)
   reg_t pte = walk(addr);
 
   reg_t pte_perm = pte & PTE_PERM;
-  if (proc == NULL || (proc->sr & SR_S))
+  if (proc == NULL || (proc->state.sr & SR_S))
     pte_perm = (pte_perm/(PTE_SX/PTE_UX)) & PTE_PERM;
   pte_perm |= pte & PTE_V;
 
@@ -46,10 +44,11 @@ reg_t mmu_t::refill_tlb(reg_t addr, reg_t bytes, bool store, bool fetch)
   if(unlikely((pte_perm & perm) != perm))
   {
     if (fetch)
-      throw trap_instruction_access_fault;
+      throw trap_instruction_access_fault();
 
-    badvaddr = addr;
-    throw store ? trap_store_access_fault : trap_load_access_fault;
+    if (store)
+      throw trap_store_access_fault(addr);
+    throw trap_load_access_fault(addr);
   }
 
   reg_t pgoff = addr & (PGSIZE-1);
@@ -77,14 +76,14 @@ pte_t mmu_t::walk(reg_t addr)
   int shift = 8*sizeof(reg_t) - VA_BITS;
   if (((sreg_t)addr << shift >> shift) != (sreg_t)addr)
     ;
-  else if (proc == NULL || !(proc->sr & SR_VM))
+  else if (proc == NULL || !(proc->state.sr & SR_VM))
   {
     if(addr < memsz)
       pte = PTE_V | PTE_PERM | ((addr >> PGSHIFT) << PGSHIFT);
   }
   else
   {
-    reg_t base = ptbr;
+    reg_t base = proc->get_state()->ptbr;
     reg_t ptd;
 
     int ptshift = (LEVELS-1)*PTIDXBITS;
diff --git a/riscv/mmu.h b/riscv/mmu.h
index bb35f70..b69de9c 100644
--- a/riscv/mmu.h
+++ b/riscv/mmu.h
@@ -33,16 +33,9 @@ public:
   #define load_func(type) \
     type##_t load_##type(reg_t addr) { \
       if(unlikely(addr % sizeof(type##_t))) \
-      { \
-        badvaddr = addr; \
-        throw trap_load_address_misaligned; \
-      } \
+        throw trap_load_address_misaligned(addr); \
       reg_t paddr = translate(addr, sizeof(type##_t), false, false); \
       return *(type##_t*)(mem + paddr); \
-    } \
-    type##_t load_reserved_##type(reg_t addr) { \
-      load_reservation = addr; \
-      return load_##type(addr); \
     }
 
   // load value from memory at aligned address; zero extend to register width
@@ -61,18 +54,9 @@ public:
   #define store_func(type) \
     void store_##type(reg_t addr, type##_t val) { \
       if(unlikely(addr % sizeof(type##_t))) \
-      { \
-        badvaddr = addr; \
-        throw trap_store_address_misaligned; \
-      } \
+        throw trap_store_address_misaligned(addr); \
       reg_t paddr = translate(addr, sizeof(type##_t), true, false); \
       *(type##_t*)(mem + paddr) = val; \
-    } \
-    reg_t store_conditional_##type(reg_t addr, type##_t val) { \
-      if (addr == load_reservation) { \
-        store_##type(addr, val); \
-        return 0; \
-      } else return 1; \
     }
 
   // store value to memory at aligned address
@@ -111,23 +95,16 @@ public:
     return icache_data[idx];
   }
 
-  reg_t get_badvaddr() { return badvaddr; }
-  reg_t get_ptbr() { return ptbr; }
-  void set_ptbr(reg_t addr) { ptbr = addr & ~(PGSIZE-1); }
   void set_processor(processor_t* p) { proc = p; flush_tlb(); }
 
   void flush_tlb();
   void flush_icache();
-  void yield_load_reservation() { load_reservation = -1; }
 
   void register_memtracer(memtracer_t*);
 
 private:
   char* mem;
   size_t memsz;
-  reg_t load_reservation;
-  reg_t badvaddr;
-  reg_t ptbr;
   processor_t* proc;
   memtracer_list_t tracer;
 
diff --git a/riscv/processor.cc b/riscv/processor.cc
index 4dff486..0510a3d 100644
--- a/riscv/processor.cc
+++ b/riscv/processor.cc
@@ -19,7 +19,9 @@ processor_t::processor_t(sim_t* _sim, mmu_t* _mmu, uint32_t _id)
   mmu.set_processor(this);
 
   #define DECLARE_INSN(name, match, mask) \
-    register_insn(match, mask, (insn_func_t)&processor_t::rv32_##name, (insn_func_t)&processor_t::rv64_##name);
+    extern reg_t rv32_##name(processor_t*, insn_t, reg_t); \
+    extern reg_t rv64_##name(processor_t*, insn_t, reg_t); \
+    register_insn(match, mask, rv32_##name, rv64_##name);
   #include "opcodes.h"
   #undef DECLARE_INSN
 }
@@ -28,17 +30,15 @@ processor_t::~processor_t()
 {
 }
 
-void processor_t::reset(bool value)
+void state_t::reset()
 {
-  if (run == !value)
-    return;
-  run = !value;
-
   // the ISA guarantees on boot that the PC is 0x2000 and the the processor
   // is in supervisor mode, and in 64-bit mode, if supported, with traps
   // and virtual memory disabled.
-  sr = 0;
-  set_pcr(PCR_SR, SR_S | SR_S64 | SR_IM);
+  sr = SR_S;
+#ifdef RISCV_ENABLE_64BIT
+  sr |= SR_S64;
+#endif
   pc = 0x2000;
 
   // the following state is undefined upon boot-up,
@@ -55,23 +55,36 @@ void processor_t::reset(bool value)
   count = 0;
   compare = 0;
   cycle = 0;
-  set_fsr(0);
+  fsr = 0;
+
+  load_reservation = -1;
+}
+
+void processor_t::reset(bool value)
+{
+  if (run == !value)
+    return;
+  run = !value;
+
+  state.reset();
 }
 
-void processor_t::set_fsr(uint32_t val)
+uint32_t processor_t::set_fsr(uint32_t val)
 {
-  fsr = val & ~FSR_ZERO; // clear FSR bits that read as zero
+  uint32_t old_fsr = state.fsr;
+  state.fsr = val & ~FSR_ZERO; // clear FSR bits that read as zero
+  return old_fsr;
 }
 
 void processor_t::take_interrupt()
 {
-  uint32_t interrupts = (sr & SR_IP) >> SR_IP_SHIFT;
-  interrupts &= (sr & SR_IM) >> SR_IM_SHIFT;
+  uint32_t interrupts = (state.sr & SR_IP) >> SR_IP_SHIFT;
+  interrupts &= (state.sr & SR_IM) >> SR_IM_SHIFT;
 
-  if(interrupts && (sr & SR_EI))
-    for(int i = 0; ; i++, interrupts >>= 1)
-      if(interrupts & 1)
-        throw interrupt_t(i);
+  if (interrupts && (state.sr & SR_EI))
+    for (int i = 0; ; i++, interrupts >>= 1)
+      if (interrupts & 1)
+        throw trap_t((1ULL << ((state.sr & SR_S64) ? 63 : 31)) + i);
 }
 
 void processor_t::step(size_t n, bool noisy)
@@ -80,20 +93,19 @@ void processor_t::step(size_t n, bool noisy)
     return;
 
   size_t i = 0;
+  reg_t npc = state.pc;
+  mmu_t& _mmu = mmu;
+
   try
   {
     take_interrupt();
 
-    mmu_t& _mmu = mmu;
-    reg_t npc = pc;
-
     // execute_insn fetches and executes one instruction
     #define execute_insn(noisy) \
       do { \
         mmu_t::insn_fetch_t fetch = _mmu.load_insn(npc); \
         if(noisy) disasm(fetch.insn, npc); \
         npc = fetch.func(this, fetch.insn, npc); \
-        pc = npc; \
       } while(0)
 
     if(noisy) for( ; i < n; i++) // print out instructions as we go
@@ -111,46 +123,45 @@ void processor_t::step(size_t n, bool noisy)
       for( ; i < n; i++)
         execute_insn(false);
     }
+
+    state.pc = npc;
   }
-  catch(trap_t t)
-  {
-    // an exception occurred in the target processor
-    take_trap(t,noisy);
-  }
-  catch(interrupt_t t)
+  catch(trap_t& t)
   {
-    take_trap((1ULL << ((sr & SR_S64) ? 63 : 31)) + t.i, noisy);
+    take_trap(npc, t, noisy);
   }
 
-  cycle += i;
+  state.cycle += i;
 
   // update timer and possibly register a timer interrupt
-  uint32_t old_count = count;
-  count += i;
-  if(old_count < compare && uint64_t(old_count) + i >= compare)
+  uint32_t old_count = state.count;
+  state.count += i;
+  if(old_count < state.compare && uint64_t(old_count) + i >= state.compare)
     set_interrupt(IRQ_TIMER, true);
 }
 
-void processor_t::take_trap(reg_t t, bool noisy)
+void processor_t::take_trap(reg_t pc, trap_t& t, bool noisy)
 {
   if(noisy)
   {
-    if ((sreg_t)t < 0)
+    if ((sreg_t)t.cause() < 0)
       fprintf(stderr, "core %3d: interrupt %d, epc 0x%016" PRIx64 "\n",
-              id, uint8_t(t), pc);
+              id, uint8_t(t.cause()), pc);
     else
       fprintf(stderr, "core %3d: trap %s, epc 0x%016" PRIx64 "\n",
-              id, trap_name(trap_t(t)), pc);
+              id, t.name(), pc);
   }
 
-  // switch to supervisor, set previous supervisor bit, disable traps
-  set_pcr(PCR_SR, (((sr & ~SR_EI) | SR_S) & ~SR_PS & ~SR_PEI) |
-                  ((sr & SR_S) ? SR_PS : 0) |
-                  ((sr & SR_EI) ? SR_PEI : 0));
-  cause = t;
-  epc = pc;
-  pc = evec;
-  badvaddr = mmu.get_badvaddr();
+  // switch to supervisor, set previous supervisor bit, disable interrupts
+  set_pcr(PCR_SR, (((state.sr & ~SR_EI) | SR_S) & ~SR_PS & ~SR_PEI) |
+                  ((state.sr & SR_S) ? SR_PS : 0) |
+                  ((state.sr & SR_EI) ? SR_PEI : 0));
+  yield_load_reservation();
+  state.cause = t.cause();
+  state.epc = pc;
+  state.pc = state.evec;
+
+  t.side_effects(&state); // might set badvaddr etc.
 }
 
 void processor_t::deliver_ipi()
@@ -164,42 +175,44 @@ void processor_t::disasm(insn_t insn, reg_t pc)
   // the disassembler is stateless, so we share it
   static disassembler disasm;
   fprintf(stderr, "core %3d: 0x%016" PRIx64 " (0x%08" PRIxFAST32 ") %s\n",
-          id, pc, insn.bits, disasm.disassemble(insn).c_str());
+          id, state.pc, insn.bits, disasm.disassemble(insn).c_str());
 }
 
-void processor_t::set_pcr(int which, reg_t val)
+reg_t processor_t::set_pcr(int which, reg_t val)
 {
+  reg_t old_pcr = get_pcr(which);
+
   switch (which)
   {
     case PCR_SR:
-      sr = (val & ~SR_IP) | (sr & SR_IP);
+      state.sr = (val & ~SR_IP) | (state.sr & SR_IP);
 #ifndef RISCV_ENABLE_64BIT
-      sr &= ~(SR_S64 | SR_U64);
+      state.sr &= ~(SR_S64 | SR_U64);
 #endif
 #ifndef RISCV_ENABLE_FPU
-      sr &= ~SR_EF;
+      state.sr &= ~SR_EF;
 #endif
 #ifndef RISCV_ENABLE_VEC
-      sr &= ~SR_EV;
+      state.sr &= ~SR_EV;
 #endif
-      sr &= ~SR_ZERO;
+      state.sr &= ~SR_ZERO;
       mmu.flush_tlb();
       break;
     case PCR_EPC:
-      epc = val;
+      state.epc = val;
       break;
     case PCR_EVEC: 
-      evec = val;
+      state.evec = val;
       break;
     case PCR_COUNT:
-      count = val;
+      state.count = val;
       break;
     case PCR_COMPARE:
       set_interrupt(IRQ_TIMER, false);
-      compare = val;
+      state.compare = val;
       break;
     case PCR_PTBR:
-      mmu.set_ptbr(val);
+      state.ptbr = val & ~(PGSIZE-1);
       break;
     case PCR_SEND_IPI:
       sim.send_ipi(val);
@@ -208,20 +221,22 @@ void processor_t::set_pcr(int which, reg_t val)
       set_interrupt(IRQ_IPI, val & 1);
       break;
     case PCR_K0:
-      pcr_k0 = val;
+      state.pcr_k0 = val;
       break;
     case PCR_K1:
-      pcr_k1 = val;
+      state.pcr_k1 = val;
       break;
     case PCR_TOHOST:
-      if (tohost == 0)
-        tohost = val;
+      if (state.tohost == 0)
+        state.tohost = val;
       break;
     case PCR_FROMHOST:
       set_interrupt(IRQ_HOST, val != 0);
-      fromhost = val;
+      state.fromhost = val;
       break;
   }
+
+  return old_pcr;
 }
 
 reg_t processor_t::get_pcr(int which)
@@ -229,37 +244,38 @@ reg_t processor_t::get_pcr(int which)
   switch (which)
   {
     case PCR_SR:
-      return sr;
+      return state.sr;
     case PCR_EPC:
-      return epc;
+      return state.epc;
     case PCR_BADVADDR:
-      return badvaddr;
+      return state.badvaddr;
     case PCR_EVEC:
-      return evec;
+      return state.evec;
     case PCR_COUNT:
-      return count;
+      return state.count;
     case PCR_COMPARE:
-      return compare;
+      return state.compare;
     case PCR_CAUSE:
-      return cause;
+      return state.cause;
     case PCR_PTBR:
-      return mmu.get_ptbr();
+      return state.ptbr;
     case PCR_ASID:
       return 0;
     case PCR_FATC:
       mmu.flush_tlb();
+      return 0;
     case PCR_HARTID:
       return id;
     case PCR_IMPL:
       return 1;
     case PCR_K0:
-      return pcr_k0;
+      return state.pcr_k0;
     case PCR_K1:
-      return pcr_k1;
+      return state.pcr_k1;
     case PCR_TOHOST:
-      return tohost;
+      return state.tohost;
     case PCR_FROMHOST:
-      return fromhost;
+      return state.fromhost;
   }
   return -1;
 }
@@ -268,27 +284,26 @@ void processor_t::set_interrupt(int which, bool on)
 {
   uint32_t mask = (1 << (which + SR_IP_SHIFT)) & SR_IP;
   if (on)
-    sr |= mask;
+    state.sr |= mask;
   else
-    sr &= ~mask;
+    state.sr &= ~mask;
+}
+
+static reg_t illegal_instruction(processor_t* p, insn_t insn, reg_t pc)
+{
+  throw trap_illegal_instruction();
 }
 
 insn_func_t processor_t::decode_insn(insn_t insn)
 {
-  bool rv64 = (sr & SR_S) ? (sr & SR_S64) : (sr & SR_U64);
+  bool rv64 = (state.sr & SR_S) ? (state.sr & SR_S64) : (state.sr & SR_U64);
 
   auto key = insn.bits & ((1L << opcode_bits)-1);
-  auto it = opcode_map.find(key);
   for (auto it = opcode_map.find(key); it != opcode_map.end() && it->first == key; ++it)
     if ((insn.bits & it->second.mask) == it->second.match)
       return rv64 ? it->second.rv64 : it->second.rv32;
 
-  return &processor_t::illegal_instruction;
-}
-
-reg_t processor_t::illegal_instruction(insn_t insn, reg_t pc)
-{
-  throw trap_illegal_instruction;
+  return &illegal_instruction;
 }
 
 void processor_t::register_insn(uint32_t match, uint32_t mask, insn_func_t rv32, insn_func_t rv64)
diff --git a/riscv/processor.h b/riscv/processor.h
index e21901c..7b7f314 100644
--- a/riscv/processor.h
+++ b/riscv/processor.h
@@ -5,7 +5,6 @@
 
 #include "decode.h"
 #include <cstring>
-#include "trap.h"
 #include "config.h"
 #include <map>
 
@@ -13,30 +12,14 @@ class processor_t;
 class mmu_t;
 typedef reg_t (*insn_func_t)(processor_t*, insn_t, reg_t);
 class sim_t;
+class trap_t;
 
-// this class represents one processor in a RISC-V machine.
-class processor_t
+// architectural state of a RISC-V hart
+struct state_t
 {
-public:
-  processor_t(sim_t* _sim, mmu_t* _mmu, uint32_t _id);
-  ~processor_t();
-
-  void reset(bool value);
-  void step(size_t n, bool noisy); // run for n cycles
-  void deliver_ipi(); // register an interprocessor interrupt
-  bool running() { return run; }
-  void set_pcr(int which, reg_t val);
-  void set_interrupt(int which, bool on);
-  reg_t get_pcr(int which);
-  mmu_t* get_mmu() { return &mmu; }
-
-  void register_insn(uint32_t match, uint32_t mask, insn_func_t rv32, insn_func_t rv64);
+  void reset();
 
-private:
-  sim_t& sim;
-  mmu_t& mmu; // main memory is always accessed via the mmu
-
-  // user-visible architected state
+  // user-visible state
   reg_t pc;
   regfile_t<reg_t, NXPR, true> XPR;
   regfile_t<freg_t, NFPR, false> FPR;
@@ -46,17 +29,47 @@ private:
   reg_t epc;
   reg_t badvaddr;
   reg_t evec;
+  reg_t ptbr;
   reg_t pcr_k0;
   reg_t pcr_k1;
   reg_t cause;
   reg_t tohost;
   reg_t fromhost;
-  uint32_t id;
   uint32_t sr; // only modify the status register using set_pcr()
   uint32_t fsr;
   uint32_t count;
   uint32_t compare;
 
+  reg_t load_reservation;
+};
+
+// this class represents one processor in a RISC-V machine.
+class processor_t
+{
+public:
+  processor_t(sim_t* _sim, mmu_t* _mmu, uint32_t _id);
+  ~processor_t();
+
+  void reset(bool value);
+  void step(size_t n, bool noisy); // run for n cycles
+  void deliver_ipi(); // register an interprocessor interrupt
+  bool running() { return run; }
+  reg_t set_pcr(int which, reg_t val);
+  uint32_t set_fsr(uint32_t val); // set the floating-point status register
+  void set_interrupt(int which, bool on);
+  reg_t get_pcr(int which);
+  uint32_t get_fsr() { return state.fsr; }
+  mmu_t* get_mmu() { return &mmu; }
+  state_t* get_state() { return &state; }
+  void yield_load_reservation() { state.load_reservation = (reg_t)-1; }
+
+  void register_insn(uint32_t match, uint32_t mask, insn_func_t rv32, insn_func_t rv64);
+
+private:
+  sim_t& sim;
+  mmu_t& mmu; // main memory is always accessed via the mmu
+  state_t state;
+  uint32_t id;
   bool run; // !reset
 
   struct opcode_map_entry_t
@@ -70,22 +83,14 @@ private:
   std::multimap<uint32_t, opcode_map_entry_t> opcode_map;
 
   void take_interrupt(); // take a trap if any interrupts are pending
-  void set_fsr(uint32_t val); // set the floating-point status register
-  void take_trap(reg_t t, bool noisy); // take an exception
+  void take_trap(reg_t pc, trap_t& t, bool noisy); // take an exception
   void disasm(insn_t insn, reg_t pc); // disassemble and print an instruction
 
   friend class sim_t;
   friend class mmu_t;
   friend class htif_isasim_t;
 
-  #define DECLARE_INSN(name, match, mask) \
-    reg_t rv32_ ## name(insn_t insn, reg_t pc); \
-    reg_t rv64_ ## name(insn_t insn, reg_t pc);
-  #include "opcodes.h"
-  #undef DECLARE_INSN
-
   insn_func_t decode_insn(insn_t insn);
-  reg_t illegal_instruction(insn_t insn, reg_t pc);
 };
 
 #endif
diff --git a/riscv/sim.cc b/riscv/sim.cc
index 6534900..2ac3266 100644
--- a/riscv/sim.cc
+++ b/riscv/sim.cc
@@ -18,9 +18,9 @@ static void handle_signal(int sig)
   signal(sig, &handle_signal);
 }
 
-sim_t::sim_t(size_t _nprocs, size_t mem_mb, const std::vector<std::string>& args)
-  : htif(new htif_isasim_t(this, args)),
-    procs(_nprocs), current_step(0), current_proc(0), debug(false)
+sim_t::sim_t(size_t nprocs, size_t mem_mb, const std::vector<std::string>& args)
+  : htif(new htif_isasim_t(this, args)), procs(std::max(nprocs, size_t(1))),
+    current_step(0), current_proc(0), debug(false)
 {
   signal(SIGINT, &handle_signal);
   // allocate target machine's memory, shrinking it as necessary
@@ -38,11 +38,9 @@ sim_t::sim_t(size_t _nprocs, size_t mem_mb, const std::vector<std::string>& args
     fprintf(stderr, "warning: only got %lu bytes of target mem (wanted %lu)\n",
             (unsigned long)memsz, (unsigned long)memsz0);
 
-  mmu = new mmu_t(mem, memsz);
+  debug_mmu = new mmu_t(mem, memsz);
 
-  if (_nprocs == 0)
-    _nprocs = 1;
-  for (size_t i = 0; i < _nprocs; i++)
+  for (size_t i = 0; i < procs.size(); i++)
     procs[i] = new processor_t(this, new mmu_t(mem, memsz), i);
 }
 
@@ -54,7 +52,7 @@ sim_t::~sim_t()
     delete procs[i];
     delete pmmu;
   }
-  delete mmu;
+  delete debug_mmu;
   free(mem);
 }
 
@@ -100,7 +98,7 @@ void sim_t::step(size_t n, bool noisy)
     if (current_step == INTERLEAVE)
     {
       current_step = 0;
-      procs[current_proc]->mmu.yield_load_reservation();
+      procs[current_proc]->yield_load_reservation();
       if (++current_proc == procs.size())
         current_proc = 0;
     }
@@ -117,7 +115,7 @@ bool sim_t::running()
 
 void sim_t::stop()
 {
-  procs[0]->tohost = 1;
+  procs[0]->state.tohost = 1;
   while (!htif->done())
     htif->tick();
 }
diff --git a/riscv/sim.h b/riscv/sim.h
index a53f3f0..bcb7c1c 100644
--- a/riscv/sim.h
+++ b/riscv/sim.h
@@ -38,7 +38,7 @@ private:
   std::auto_ptr<htif_isasim_t> htif;
   char* mem; // main memory
   size_t memsz; // memory size in bytes
-  mmu_t* mmu; // debug port into main memory
+  mmu_t* debug_mmu;  // debug port into main memory
   std::vector<processor_t*> procs;
 
   void step(size_t n, bool noisy); // step through simulation
diff --git a/riscv/trap.cc b/riscv/trap.cc
index ba0d867..ead3a00 100644
--- a/riscv/trap.cc
+++ b/riscv/trap.cc
@@ -1,12 +1,15 @@
-// See LICENSE for license details.
-
 #include "trap.h"
+#include "processor.h"
+#include <cstdio>
 
-const char* trap_name(trap_t t)
+const char* trap_t::name()
 {
-  #define DECLARE_TRAP(x) "trap_"#x
-  static const char* names[] = { TRAP_LIST };
-  #undef DECLARE_TRAP
+  const char* fmt = uint8_t(which) == which ? "trap #%u" : "interrupt #%u";
+  sprintf(_name, fmt, uint8_t(which));
+  return _name;
+}
 
-  return (unsigned)t >= sizeof(names)/sizeof(names[0]) ? "unknown" : names[t];
+void mem_trap_t::side_effects(state_t* state)
+{
+  state->badvaddr = badvaddr;
 }
diff --git a/riscv/trap.h b/riscv/trap.h
index 6448e51..a7a823b 100644
--- a/riscv/trap.h
+++ b/riscv/trap.h
@@ -3,35 +3,58 @@
 #ifndef _RISCV_TRAP_H
 #define _RISCV_TRAP_H
 
-#define TRAP_LIST \
-  DECLARE_TRAP(instruction_address_misaligned), \
-  DECLARE_TRAP(instruction_access_fault), \
-  DECLARE_TRAP(illegal_instruction), \
-  DECLARE_TRAP(privileged_instruction), \
-  DECLARE_TRAP(fp_disabled), \
-  DECLARE_TRAP(reserved0), \
-  DECLARE_TRAP(syscall), \
-  DECLARE_TRAP(breakpoint), \
-  DECLARE_TRAP(load_address_misaligned), \
-  DECLARE_TRAP(store_address_misaligned), \
-  DECLARE_TRAP(load_access_fault), \
-  DECLARE_TRAP(store_access_fault), \
-  DECLARE_TRAP(vector_disabled), \
-  DECLARE_TRAP(vector_bank), \
-  DECLARE_TRAP(vector_illegal_instruction), \
-  DECLARE_TRAP(reserved1), \
-
-#define DECLARE_TRAP(x) trap_##x
-enum trap_t
+#include "decode.h"
+
+class state_t;
+
+class trap_t
+{
+ public:
+  trap_t(reg_t which) : which(which) {}
+  virtual const char* name();
+  virtual void side_effects(state_t* state) {}
+  reg_t cause() { return which; }
+ private:
+  char _name[16];
+  reg_t which;
+};
+
+class mem_trap_t : public trap_t
 {
-  TRAP_LIST
-  NUM_TRAPS
+ public:
+  mem_trap_t(reg_t which, reg_t badvaddr)
+    : trap_t(which), badvaddr(badvaddr) {}
+  void side_effects(state_t* state);
+ private:
+  reg_t badvaddr;
+};
+
+#define DECLARE_TRAP(n, x) class trap_##x : public trap_t { \
+ public: \
+  trap_##x() : trap_t(n) {} \
+  const char* name() { return "trap_"#x; } \
 };
-#undef DECLARE_TRAP
 
-struct interrupt_t { interrupt_t(int which) : i(which) {} int i; };
-struct halt_t {}; // thrown to stop the processor from running
+#define DECLARE_MEM_TRAP(n, x) class trap_##x : public mem_trap_t { \
+ public: \
+  trap_##x(reg_t badvaddr) : mem_trap_t(n, badvaddr) {} \
+  const char* name() { return "trap_"#x; } \
+};
 
-extern "C" const char* trap_name(trap_t t);
+DECLARE_TRAP(0, instruction_address_misaligned)
+DECLARE_TRAP(1, instruction_access_fault)
+DECLARE_TRAP(2, illegal_instruction)
+DECLARE_TRAP(3, privileged_instruction)
+DECLARE_TRAP(4, fp_disabled)
+DECLARE_TRAP(5, reserved0)
+DECLARE_TRAP(6, syscall)
+DECLARE_TRAP(7, breakpoint)
+DECLARE_MEM_TRAP(8, load_address_misaligned)
+DECLARE_MEM_TRAP(9, store_address_misaligned)
+DECLARE_MEM_TRAP(10, load_access_fault)
+DECLARE_MEM_TRAP(11, store_access_fault)
+DECLARE_TRAP(12, vector_disabled)
+DECLARE_TRAP(13, vector_bank)
+DECLARE_TRAP(14, vector_illegal_instruction)
 
 #endif
-- 
2.30.2