add SUBVL CSR set

[riscv-isa-sim.git] / riscv / processor.cc

// See LICENSE for license details.

#include "processor.h"
#include "extension.h"
#include "common.h"
#include "config.h"
#include "simif.h"
#include "mmu.h"
#include "disasm.h"
#include <cinttypes>
#include <cmath>
#include <cstdlib>
#include <iostream>
#include <assert.h>
#include <limits.h>
#include <stdexcept>
#include <algorithm>
#ifdef SPIKE_SIMPLEV
#include "sv_insn_redirect.h"
#endif

#undef STATE
#define STATE state

processor_t::processor_t(const char* isa, simif_t* sim, uint32_t id,
        bool halt_on_reset)
  : debug(false), halt_request(false), sim(sim), ext(NULL), id(id),
  halt_on_reset(halt_on_reset), last_pc(1), executions(1)
#ifdef SPIKE_SIMPLEV
    , s(this)
#endif
{
  parse_isa_string(isa);
  register_base_instructions();

#ifdef SPIKE_SIMPLEV
  mmu = new sv_mmu_t(sim, this);
#else
  mmu = new mmu_t(sim, this);
#endif

  disassembler = new disassembler_t(max_xlen);
  if (ext)
    for (auto disasm_insn : ext->get_disasms())
      disassembler->add_insn(disasm_insn);

  reset();
}

processor_t::~processor_t()
{
#ifdef RISCV_ENABLE_HISTOGRAM
  if (histogram_enabled)
  {
    fprintf(stderr, "PC Histogram size:%zu\n", pc_histogram.size());
    for (auto it : pc_histogram)
      fprintf(stderr, "%0" PRIx64 " %" PRIu64 "\n", it.first, it.second);
  }
#endif

  delete mmu;
  delete disassembler;
}

static void bad_isa_string(const char* isa)
{
  fprintf(stderr, "error: bad --isa option %s\n", isa);
  abort();
}

void processor_t::parse_isa_string(const char* str)
{
  std::string lowercase, tmp;
  for (const char *r = str; *r; r++)
    lowercase += std::tolower(*r);

  const char* p = lowercase.c_str();
  const char* all_subsets = "imafdqc";

  max_xlen = 64;
  state.misa = reg_t(2) << 62;

  if (strncmp(p, "rv32", 4) == 0)
    max_xlen = 32, state.misa = reg_t(1) << 30, p += 4;
  else if (strncmp(p, "rv64", 4) == 0)
    p += 4;
  else if (strncmp(p, "rv", 2) == 0)
    p += 2;

  if (!*p) {
    p = "imafdc";
  } else if (*p == 'g') { // treat "G" as "IMAFD"
    tmp = std::string("imafd") + (p+1);
    p = &tmp[0];
  } else if (*p != 'i') {
    bad_isa_string(str);
  }

  isa_string = "rv" + std::to_string(max_xlen) + p;
  state.misa |= 1L << ('s' - 'a'); // advertise support for supervisor mode
  state.misa |= 1L << ('u' - 'a'); // advertise support for user mode

  while (*p) {
    state.misa |= 1L << (*p - 'a');

    if (auto next = strchr(all_subsets, *p)) {
      all_subsets = next + 1;
      p++;
    } else if (*p == 'x') {
      const char* ext = p+1, *end = ext;
      while (islower(*end))
        end++;
      register_extension(find_extension(std::string(ext, end - ext).c_str())());
      p = end;
    } else {
      bad_isa_string(str);
    }
  }

  if (supports_extension('D') && !supports_extension('F'))
    bad_isa_string(str);

  if (supports_extension('Q') && !supports_extension('D'))
    bad_isa_string(str);

  if (supports_extension('Q') && max_xlen < 64)
    bad_isa_string(str);

  max_isa = state.misa;
}

void state_t::reset(reg_t max_isa)
{
  memset(this, 0, sizeof(*this));
  misa = max_isa;
  prv = PRV_M;
  pc = DEFAULT_RSTVEC;
  tselect = 0;
  for (unsigned int i = 0; i < num_triggers; i++)
    mcontrol[i].type = 2;
#ifdef SPIKE_SIMPLEV
  // set SV CSR banks to default (full) sizes
  msv.state_size = 1;
  ssv.state_size = 1;
  usv.state_size = 3;
#endif
}

void sv_shape_t::setup_map()
{
    int order[3] = {};
    int lims[3] = {xsz, ysz, zsz};
    int idxs[3] = {0,0,0};

    switch (permute) {
        case SV_SHAPE_PERM_XYZ: order[0] = 0; order[1] = 1; order[2] = 2; break;
        case SV_SHAPE_PERM_XZY: order[0] = 0; order[1] = 2; order[2] = 1; break;
        case SV_SHAPE_PERM_YXZ: order[0] = 1; order[1] = 0; order[2] = 2; break;
        case SV_SHAPE_PERM_YZX: order[0] = 1; order[1] = 2; order[2] = 0; break;
        case SV_SHAPE_PERM_ZXY: order[0] = 2; order[1] = 0; order[2] = 1; break;
        case SV_SHAPE_PERM_ZYX: order[0] = 2; order[1] = 1; order[2] = 0; break;
        default: throw trap_illegal_instruction(0);
    }
    for (int i = 0; i < 128; i++)
    {
        uint8_t new_idx = idxs[0] + idxs[1] * xsz + idxs[2] * xsz * ysz;
        map[i] = new_idx;
        for (int j = 0; j < 3; j++)
        {
            idxs[order[j]] = idxs[order[j]] + 1;
            if (idxs[order[j]] != lims[order[j]]) {
                break;
            }
            idxs[order[j]] = 0;
        }
    }
}

int state_t::sv_csr_sz()
{
    if (prv == PRV_M)
        return SV_MCSR_SZ;
    if (prv == PRV_S)
        return SV_SCSR_SZ;
    return SV_UCSR_SZ;
}
sv_csr_t &state_t::sv()
{
    if (prv == PRV_M)
        return get_msv();
    if (prv == PRV_S)
        return get_ssv();
    return get_usv();
}

sv_shape_t* state_t::get_shape(reg_t reg, bool pred)
{
    if (prv == PRV_M || prv == PRV_S || reg == 0) {
        return NULL;
    }
    for (int i = 0; i < 3; i++) {
        if (remap[i].regidx == reg && remap[i].pred == pred) {
            return &shape[i];
        }
    }
    return NULL;
}

void processor_t::set_debug(bool value)
{
  debug = value;
  if (ext)
    ext->set_debug(value);
}

void processor_t::set_histogram(bool value)
{
  histogram_enabled = value;
#ifndef RISCV_ENABLE_HISTOGRAM
  if (value) {
    fprintf(stderr, "PC Histogram support has not been properly enabled;");
    fprintf(stderr, " please re-build the riscv-isa-run project using \"configure --enable-histogram\".\n");
  }
#endif
}

void processor_t::reset()
{
  state.reset(max_isa);
  state.dcsr.halt = halt_on_reset;
  halt_on_reset = false;
  set_csr(CSR_MSTATUS, state.mstatus);

  if (ext)
    ext->reset(); // reset the extension

  if (sim)
    sim->proc_reset(id);
}

// Count number of contiguous 0 bits starting from the LSB.
static int ctz(reg_t val)
{
  int res = 0;
  if (val)
    while ((val & 1) == 0)
      val >>= 1, res++;
  return res;
}

void processor_t::take_interrupt(reg_t pending_interrupts)
{
  reg_t mie = get_field(state.mstatus, MSTATUS_MIE);
  reg_t m_enabled = state.prv < PRV_M || (state.prv == PRV_M && mie);
  reg_t enabled_interrupts = pending_interrupts & ~state.mideleg & -m_enabled;

  reg_t sie = get_field(state.mstatus, MSTATUS_SIE);
  reg_t s_enabled = state.prv < PRV_S || (state.prv == PRV_S && sie);
  // M-ints have highest priority; consider S-ints only if no M-ints pending
  if (enabled_interrupts == 0)
    enabled_interrupts = pending_interrupts & state.mideleg & -s_enabled;

  if (state.dcsr.cause == 0 && enabled_interrupts) {
    // nonstandard interrupts have highest priority
    if (enabled_interrupts >> IRQ_M_EXT)
      enabled_interrupts = enabled_interrupts >> IRQ_M_EXT << IRQ_M_EXT;
    // external interrupts have next-highest priority
    else if (enabled_interrupts & (MIP_MEIP | MIP_SEIP))
      enabled_interrupts = enabled_interrupts & (MIP_MEIP | MIP_SEIP);
    // software interrupts have next-highest priority
    else if (enabled_interrupts & (MIP_MSIP | MIP_SSIP))
      enabled_interrupts = enabled_interrupts & (MIP_MSIP | MIP_SSIP);
    // timer interrupts have next-highest priority
    else if (enabled_interrupts & (MIP_MTIP | MIP_STIP))
      enabled_interrupts = enabled_interrupts & (MIP_MTIP | MIP_STIP);
    else
      abort();

    throw trap_t(((reg_t)1 << (max_xlen-1)) | ctz(enabled_interrupts));
  }
}

static int xlen_to_uxl(int xlen)
{
  if (xlen == 32)
    return 1;
  if (xlen == 64)
    return 2;
  abort();
}

reg_t processor_t::legalize_privilege(reg_t prv)
{
  assert(prv <= PRV_M);

  if (!supports_extension('U'))
    return PRV_M;

  if (prv == PRV_H || !supports_extension('S'))
    return PRV_U;

  return prv;
}

void processor_t::set_privilege(reg_t prv)
{
  mmu->flush_tlb();
  state.prv = legalize_privilege(prv);
}

void processor_t::enter_debug_mode(uint8_t cause)
{
  state.dcsr.cause = cause;
  state.dcsr.prv = state.prv;
  set_privilege(PRV_M);
  state.dpc = state.pc;
  state.pc = DEBUG_ROM_ENTRY;
}

void processor_t::take_trap(trap_t& t, reg_t epc)
{
  if (debug) {
    fprintf(stderr, "core %3d: exception %s, epc 0x%016" PRIx64 "\n",
            id, t.name(), epc);
    if (t.has_tval())
      fprintf(stderr, "core %3d:           tval 0x%016" PRIx64 "\n", id,
          t.get_tval());
  }

  if (state.dcsr.cause) {
    if (t.cause() == CAUSE_BREAKPOINT) {
      state.pc = DEBUG_ROM_ENTRY;
    } else {
      state.pc = DEBUG_ROM_TVEC;
    }
    return;
  }

  if (t.cause() == CAUSE_BREAKPOINT && (
              (state.prv == PRV_M && state.dcsr.ebreakm) ||
              (state.prv == PRV_S && state.dcsr.ebreaks) ||
              (state.prv == PRV_U && state.dcsr.ebreaku))) {
    enter_debug_mode(DCSR_CAUSE_SWBP);
    return;
  }

  // by default, trap to M-mode, unless delegated to S-mode
  reg_t bit = t.cause();
  reg_t deleg = state.medeleg;
  bool interrupt = (bit & ((reg_t)1 << (max_xlen-1))) != 0;
  if (interrupt)
    deleg = state.mideleg, bit &= ~((reg_t)1 << (max_xlen-1));
  if (state.prv <= PRV_S && bit < max_xlen && ((deleg >> bit) & 1)) {
    // handle the trap in S-mode
    state.pc = state.stvec;
    state.scause = t.cause();
    state.sepc = epc;
    state.stval = t.get_tval();

    reg_t s = state.mstatus;
    s = set_field(s, MSTATUS_SPIE, get_field(s, MSTATUS_SIE));
    s = set_field(s, MSTATUS_SPP, state.prv);
    s = set_field(s, MSTATUS_SIE, 0);
    set_csr(CSR_MSTATUS, s);
    set_privilege(PRV_S);
  } else {
    reg_t vector = (state.mtvec & 1) && interrupt ? 4*bit : 0;
    state.pc = (state.mtvec & ~(reg_t)1) + vector;
    state.mepc = epc;
    state.mcause = t.cause();
    state.mtval = t.get_tval();

    reg_t s = state.mstatus;
    s = set_field(s, MSTATUS_MPIE, get_field(s, MSTATUS_MIE));
    s = set_field(s, MSTATUS_MPP, state.prv);
    s = set_field(s, MSTATUS_MIE, 0);
    set_csr(CSR_MSTATUS, s);
    set_privilege(PRV_M);
  }
}

void processor_t::disasm(insn_t insn)
{
  uint64_t bits = insn.bits() & ((1ULL << (8 * insn_length(insn.bits()))) - 1);
  if (last_pc != state.pc || last_bits != bits) {
    if (executions != 1) {
      fprintf(stderr, "core %3d: Executed %" PRIx64 " times\n", id, executions);
    }

    fprintf(stderr, "core %3d: 0x%016" PRIx64 " (0x%08" PRIx64 ") %s\n",
            id, state.pc, bits, disassembler->disassemble(insn).c_str());
    last_pc = state.pc;
    last_bits = bits;
    executions = 1;
  } else {
    executions++;
  }
}

int processor_t::paddr_bits()
{
  assert(xlen == max_xlen);
  return max_xlen == 64 ? 50 : 34;
}

void state_t::get_csr_start_end(int &start, int &end)
{
    start = sv().state_bank * 4;
    end = start + (1 << (sv().state_size+1));
    start = std::min(sv_csr_sz(), start);
    end = std::min(sv_csr_sz(), end);
    fprintf(stderr, "sv state csr start/end: %d %d\n", start, end);
}

void state_t::sv_csr_reg_unpack()
{
    // okaaay and now "unpack" the CAM to make it easier to use.  this
    // approach is not designed to be efficient right now.  optimise later
    // first clear the old tables
    memset(sv().sv_int_tb, 0, sizeof(sv().sv_int_tb));
    memset(sv().sv_fp_tb, 0, sizeof(sv().sv_fp_tb));
    // now walk the CAM and unpack it
    int start = 0;
    int end = 0;
    get_csr_start_end(start, end);
    for (int i = start; i < end; i++)
    {
        union sv_reg_csr_entry *c = &sv().sv_csrs[i];
        uint64_t idx = c->b.regkey;
        sv_reg_entry *r;
        if (c->u == 0)
        {
            break;
        }
        // XXX damn.  this basically duplicates sv_insn_t::get_regentry.
        if (c->b.type == 1)
        {
            r = &sv().sv_int_tb[idx];
        }
        else
        {
            r = &sv().sv_fp_tb[idx];
        }
        r->elwidth = c->b.elwidth;
        r->regidx = c->b.regidx;
        r->isvec = c->b.isvec;
        r->active = true;
        fprintf(stderr, "setting REGCFG type:%d isvec:%d %d %d\n",
                        c->b.type, r->isvec, (int)idx, (int)r->regidx);
    }
}

void state_t::sv_csr_pred_unpack()
{
    memset(sv().sv_pred_int_tb, 0, sizeof(sv().sv_pred_int_tb));
    memset(sv().sv_pred_fp_tb, 0, sizeof(sv().sv_pred_fp_tb));
    int start = 0;
    int end = 0;
    get_csr_start_end(start, end);
    for (int i = start; i < end; i++)
    {
        union sv_pred_csr_entry *c = &sv().sv_pred_csrs[i];
        uint64_t idx = c->b.regkey;
        if (c->u == 0)
        {
            break;
        }
        sv_pred_entry *r;
        // XXX damn.  this basically duplicates sv_insn_t::get_predentry.
        if (c->b.type == 1)
        {
            r = &sv().sv_pred_int_tb[idx];
        }
        else
        {
            r = &sv().sv_pred_fp_tb[idx];
        }
        r->regidx = c->b.regidx;
        r->zero   = c->b.zero;
        r->inv    = c->b.inv;
        r->ffirst = c->b.ffirst;
        r->active = true;
        fprintf(stderr, "setting PREDCFG %d type:%d zero:%d %d %d\n",
                        i, c->b.type, r->zero, (int)idx, (int)r->regidx);
    }
}

reg_t processor_t::set_csr(int which, reg_t val, bool imm_mode)
{
  reg_t old_val = get_csr(which);
  val = _zext_xlen(val);
  reg_t delegable_ints = MIP_SSIP | MIP_STIP | MIP_SEIP
                       | ((ext != NULL) << IRQ_COP);
  reg_t all_ints = delegable_ints | MIP_MSIP | MIP_MTIP;
  fprintf(stderr, "set CSR %x %lx\n", which, val);
  switch (which)
  {
#ifdef SPIKE_SIMPLEV
    case CSR_USVMVL:
      state.sv().mvl = std::min(val+1, (uint64_t)64); // limited to XLEN width
      old_val = state.sv().mvl - 1;
      // TODO XXX throw exception if val == 0
      fprintf(stderr, "set MVL %lx\n", state.sv().mvl);
      break;
    case CSR_USVSTATE:
    {
      // bits 0-5: mvl - 6-11: vl - 12-17: srcoffs - 18-23: destoffs
      set_csr(CSR_USVMVL, get_field(val, SV_STATE_VL )+1);
      set_csr(CSR_USVVL , get_field(val, SV_STATE_MVL)+1);
      set_csr(CSR_USVSUBVL , get_field(val, SV_STATE_SUBVL)+1);
      // decode (and limit) src/dest VL offsets
      reg_t srcoffs = get_field(val, SV_STATE_SRCOFFS);
      reg_t destoffs = get_field(val, SV_STATE_DESTOFFS);
      state.sv().srcoffs  = std::min(srcoffs , state.sv().vl-1);
      state.sv().destoffs = std::min(destoffs, state.sv().vl-1);
      // decode (and limit) src/dest SUBVL offsets
      reg_t subsrcoffs = get_field(val, SV_STATE_SSVOFFS);
      reg_t subdestoffs = get_field(val, SV_STATE_DSVOFFS);
      state.sv().ssvoffs  = std::min(subsrcoffs , state.sv().subvl-1);
      state.sv().dsvoffs = std::min(subdestoffs, state.sv().subvl-1);
      //int state_bank = get_field(val, SV_STATE_BANK);
      //int state_size = get_field(val, SV_STATE_SIZE);
      //set_csr(CSR_USVCFG, state_bank | (state_size << 3));
      break;
    }
    case CSR_USVCFG:
    {
      int old_bank = state.sv().state_bank;
      int old_size = state.sv().state_size;
      state.sv().state_bank = get_field(val, SV_CFG_BANK);
      state.sv().state_size = get_field(val, SV_CFG_SIZE);
      if (old_bank != state.sv().state_bank ||
          old_size != state.sv().state_size)
      {
        // if the bank or size is changed, the csrs that are enabled
        // also changes.  easiest thing in software: recalculate them all
        state.sv_csr_pred_unpack();
        state.sv_csr_reg_unpack();
      }
      break;
    }
    case CSR_USVSUBVL:
      state.sv().subvl = std::max(1, std::min(4, (int)val));
      old_val = state.sv().subvl;
      // TODO XXX throw exception if val attempted to be set == 0
      fprintf(stderr, "set VL %lx\n", state.sv().vl);
      break;
    case CSR_USVVL:
      state.sv().vl = std::min(state.sv().mvl, val + 1);
      old_val = state.sv().mvl - 1;
      // TODO XXX throw exception if val == 0
      fprintf(stderr, "set VL %lx\n", state.sv().vl);
      break;
    case CSR_SVREGTOP:
    case CSR_SVREGBOT:
    {
      bool top = (which == CSR_SVREGTOP);
      uint64_t v = (uint64_t)val;
      fprintf(stderr, "set SVREG %d %lx\n", top, v);
      int start = 0;
      int end = 0;
      state.get_csr_start_end(start, end);
      uint64_t res_old = 0;
      int num_entries = val & 0xf;
      int max_xlen_entries = (xlen == 64) ? 4 : 2;
      if (!imm_mode) {
          num_entries = max_xlen_entries;
      }
      // read 2 16-bit entries for RV32, 4 16-bit entries for RV64
      int popidx = 0;
      for (int i = 0; i < num_entries; i++) {
          uint64_t svcfg = 0;
          if (!imm_mode) {
              uint64_t mask = 0xffffUL << (i*16UL);
              svcfg = get_field(v, mask);
              fprintf(stderr, "SVREG mask %lx cfg %lx\n", mask, svcfg);
              if (!svcfg && i > 0) {
                break;
              }
          }
          // see regpush on how this works.
          uint64_t res = state.sv().regpush(svcfg, end, top);
          if (res != 0) {
              res_old |= res << (popidx * 16UL);
              popidx += 1;
              if (popidx == max_xlen_entries) {
                break;
              }
          }
      }
      old_val = res_old;
      state.sv_csr_reg_unpack();
      break;
    }
    case CSR_SVPREDCFG0:
    case CSR_SVPREDCFG1:
    case CSR_SVPREDCFG2:
    case CSR_SVPREDCFG3:
    case CSR_SVPREDCFG4:
    case CSR_SVPREDCFG5:
    case CSR_SVPREDCFG6:
    case CSR_SVPREDCFG7:
    {
      // comments removed as it's near-identical to the regs version
      // TODO: macro-ify
      uint64_t v = (uint64_t)val;
      int tbidx = (which - CSR_SVPREDCFG0) * 2;
      fprintf(stderr, "set PREDCFG %d %lx\n", tbidx, v);
      state.sv().sv_pred_csrs[tbidx].u = get_field(v, 0xffff);
      state.sv().sv_pred_csrs[tbidx+1].u = get_field(v, 0xffff0000);
      int clroffset = 2;
      if (xlen == 64)
      {
          state.sv().sv_pred_csrs[tbidx+2].u = get_field(v, 0xffffUL<<32);
          state.sv().sv_pred_csrs[tbidx+3].u = get_field(v, 0xffffUL<<48);
          clroffset = 4;
      }
      for (int i = tbidx+clroffset; i < 16; i++)
      {
          state.sv().sv_pred_csrs[i].u = 0;
      }
      state.sv_csr_pred_unpack();
      break;
    }
    case CSR_UREMAP:
    {
      state.remap[0].regidx = get_field(val, SV_REMAP_REGIDX0);
      state.remap[1].regidx = get_field(val, SV_REMAP_REGIDX1);
      state.remap[2].regidx = get_field(val, SV_REMAP_REGIDX2);
      state.remap[0].pred = get_field(val, SV_REMAP_PRED0);
      state.remap[1].pred = get_field(val, SV_REMAP_PRED1);
      state.remap[2].pred = get_field(val, SV_REMAP_PRED2);
      state.remap[0].shape = get_field(val, SV_REMAP_SHAPE0);
      state.remap[1].shape = get_field(val, SV_REMAP_SHAPE1);
      state.remap[2].shape = get_field(val, SV_REMAP_SHAPE2);
      break;
    }
    case CSR_USHAPE0:
    case CSR_USHAPE1:
    case CSR_USHAPE2:
    {
      int shapeidx = which - CSR_USHAPE0;
      state.shape[shapeidx].xsz = get_field(val, SV_SHAPE_XDIM) + 1;
      state.shape[shapeidx].ysz = get_field(val, SV_SHAPE_YDIM) + 1;
      state.shape[shapeidx].zsz = get_field(val, SV_SHAPE_ZDIM) + 1;
      state.shape[shapeidx].offs = (get_field(val, (1<<7 )) ? 0x1 : 0) |
                                   (get_field(val, (1<<15)) ? 0x2 : 0) |
                                   (get_field(val, (1<<23)) ? 0x4 : 0);
      state.shape[shapeidx].permute = get_field(val, SV_SHAPE_PERM);
      state.shape[shapeidx].setup_map();
      fprintf(stderr, "sv shape %d x %d y %d z %d offs %d perm %d\n",
              shapeidx,
              state.shape[shapeidx].xsz,
              state.shape[shapeidx].ysz,
              state.shape[shapeidx].zsz,
              state.shape[shapeidx].offs,
              state.shape[shapeidx].permute);
      break;
    }
#endif
    case CSR_FFLAGS:
      dirty_fp_state;
      state.fflags = val & (FSR_AEXC >> FSR_AEXC_SHIFT);
      break;
    case CSR_FRM:
      dirty_fp_state;
      state.frm = val & (FSR_RD >> FSR_RD_SHIFT);
      break;
    case CSR_FCSR:
      dirty_fp_state;
      state.fflags = (val & FSR_AEXC) >> FSR_AEXC_SHIFT;
      state.frm = (val & FSR_RD) >> FSR_RD_SHIFT;
      break;
    case CSR_MSTATUS: {
      if ((val ^ state.mstatus) &
          (MSTATUS_MPP | MSTATUS_MPRV | MSTATUS_SUM | MSTATUS_MXR))
        mmu->flush_tlb();

      reg_t mask = MSTATUS_SIE | MSTATUS_SPIE | MSTATUS_MIE | MSTATUS_MPIE
                 | MSTATUS_FS | MSTATUS_MPRV | MSTATUS_SUM
                 | MSTATUS_MXR | MSTATUS_TW | MSTATUS_TVM
                 | MSTATUS_TSR | MSTATUS_UXL | MSTATUS_SXL |
                 (ext ? MSTATUS_XS : 0);

      reg_t requested_mpp = legalize_privilege(get_field(val, MSTATUS_MPP));
      state.mstatus = set_field(state.mstatus, MSTATUS_MPP, requested_mpp);
      if (supports_extension('S'))
        mask |= MSTATUS_SPP;

      state.mstatus = (state.mstatus & ~mask) | (val & mask);

      bool dirty = (state.mstatus & MSTATUS_FS) == MSTATUS_FS;
      dirty |= (state.mstatus & MSTATUS_XS) == MSTATUS_XS;
      if (max_xlen == 32)
        state.mstatus = set_field(state.mstatus, MSTATUS32_SD, dirty);
      else
        state.mstatus = set_field(state.mstatus, MSTATUS64_SD, dirty);

      state.mstatus = set_field(state.mstatus, MSTATUS_UXL, xlen_to_uxl(max_xlen));
      state.mstatus = set_field(state.mstatus, MSTATUS_UXL, xlen_to_uxl(max_xlen));
      state.mstatus = set_field(state.mstatus, MSTATUS_SXL, xlen_to_uxl(max_xlen));
      // U-XLEN == S-XLEN == M-XLEN
      xlen = max_xlen;
      break;
    }
    case CSR_MIP: {
      reg_t mask = MIP_SSIP | MIP_STIP;
      state.mip = (state.mip & ~mask) | (val & mask);
      break;
    }
    case CSR_MIE:
      state.mie = (state.mie & ~all_ints) | (val & all_ints);
      break;
    case CSR_MIDELEG:
      state.mideleg = (state.mideleg & ~delegable_ints) | (val & delegable_ints);
      break;
    case CSR_MEDELEG: {
      reg_t mask =
        (1 << CAUSE_MISALIGNED_FETCH) |
        (1 << CAUSE_BREAKPOINT) |
        (1 << CAUSE_USER_ECALL) |
        (1 << CAUSE_FETCH_PAGE_FAULT) |
        (1 << CAUSE_LOAD_PAGE_FAULT) |
        (1 << CAUSE_STORE_PAGE_FAULT);
      state.medeleg = (state.medeleg & ~mask) | (val & mask);
      break;
    }
    case CSR_MINSTRET:
    case CSR_MCYCLE:
      if (xlen == 32)
        state.minstret = (state.minstret >> 32 << 32) | (val & 0xffffffffU);
      else
        state.minstret = val;
      // The ISA mandates that if an instruction writes instret, the write
      // takes precedence over the increment to instret.  However, Spike
      // unconditionally increments instret after executing an instruction.
      // Correct for this artifact by decrementing instret here.
      state.minstret--;
      break;
    case CSR_MINSTRETH:
    case CSR_MCYCLEH:
      state.minstret = (val << 32) | (state.minstret << 32 >> 32);
      state.minstret--; // See comment above.
      break;
    case CSR_SCOUNTEREN:
      state.scounteren = val;
      break;
    case CSR_MCOUNTEREN:
      state.mcounteren = val;
      break;
    case CSR_SSTATUS: {
      reg_t mask = SSTATUS_SIE | SSTATUS_SPIE | SSTATUS_SPP | SSTATUS_FS
                 | SSTATUS_XS | SSTATUS_SUM | SSTATUS_MXR;
      return set_csr(CSR_MSTATUS, (state.mstatus & ~mask) | (val & mask));
    }
    case CSR_SIP: {
      reg_t mask = MIP_SSIP & state.mideleg;
      return set_csr(CSR_MIP, (state.mip & ~mask) | (val & mask));
    }
    case CSR_SIE:
      return set_csr(CSR_MIE,
                     (state.mie & ~state.mideleg) | (val & state.mideleg));
    case CSR_SATP: {
      mmu->flush_tlb();
      if (max_xlen == 32)
        state.satp = val & (SATP32_PPN | SATP32_MODE);
      if (max_xlen == 64 && (get_field(val, SATP64_MODE) == SATP_MODE_OFF ||
                             get_field(val, SATP64_MODE) == SATP_MODE_SV39 ||
                             get_field(val, SATP64_MODE) == SATP_MODE_SV48))
        state.satp = val & (SATP64_PPN | SATP64_MODE);
      break;
    }
    case CSR_SEPC: state.sepc = val & ~(reg_t)1; break;
    case CSR_STVEC: state.stvec = val >> 2 << 2; break;
    case CSR_SSCRATCH: state.sscratch = val; break;
    case CSR_SCAUSE: state.scause = val; break;
    case CSR_STVAL: state.stval = val; break;
    case CSR_MEPC: state.mepc = val & ~(reg_t)1; break;
    case CSR_MTVEC: state.mtvec = val & ~(reg_t)2; break;
    case CSR_MSCRATCH: state.mscratch = val; break;
    case CSR_MCAUSE: state.mcause = val; break;
    case CSR_MTVAL: state.mtval = val; break;
    case CSR_MISA: {
      // the write is ignored if increasing IALIGN would misalign the PC
      if (!(val & (1L << ('C' - 'A'))) && (state.pc & 2))
        break;

      if (!(val & (1L << ('F' - 'A'))))
        val &= ~(1L << ('D' - 'A'));

      // allow MAFDC bits in MISA to be modified
      reg_t mask = 0;
      mask |= 1L << ('M' - 'A');
      mask |= 1L << ('A' - 'A');
      mask |= 1L << ('F' - 'A');
      mask |= 1L << ('D' - 'A');
      mask |= 1L << ('C' - 'A');
      mask &= max_isa;

      state.misa = (val & mask) | (state.misa & ~mask);
      break;
    }
    case CSR_TSELECT:
      if (val < state.num_triggers) {
        state.tselect = val;
      }
      break;
    case CSR_TDATA1:
      {
        mcontrol_t *mc = &state.mcontrol[state.tselect];
        if (mc->dmode && !state.dcsr.cause) {
          break;
        }
        mc->dmode = get_field(val, MCONTROL_DMODE(xlen));
        mc->select = get_field(val, MCONTROL_SELECT);
        mc->timing = get_field(val, MCONTROL_TIMING);
        mc->action = (mcontrol_action_t) get_field(val, MCONTROL_ACTION);
        mc->chain = get_field(val, MCONTROL_CHAIN);
        mc->match = (mcontrol_match_t) get_field(val, MCONTROL_MATCH);
        mc->m = get_field(val, MCONTROL_M);
        mc->h = get_field(val, MCONTROL_H);
        mc->s = get_field(val, MCONTROL_S);
        mc->u = get_field(val, MCONTROL_U);
        mc->execute = get_field(val, MCONTROL_EXECUTE);
        mc->store = get_field(val, MCONTROL_STORE);
        mc->load = get_field(val, MCONTROL_LOAD);
        // Assume we're here because of csrw.
        if (mc->execute)
          mc->timing = 0;
        trigger_updated();
      }
      break;
    case CSR_TDATA2:
      if (state.mcontrol[state.tselect].dmode && !state.dcsr.cause) {
        break;
      }
      if (state.tselect < state.num_triggers) {
        state.tdata2[state.tselect] = val;
      }
      break;
    case CSR_DCSR:
      state.dcsr.prv = get_field(val, DCSR_PRV);
      state.dcsr.step = get_field(val, DCSR_STEP);
      // TODO: ndreset and fullreset
      state.dcsr.ebreakm = get_field(val, DCSR_EBREAKM);
      state.dcsr.ebreakh = get_field(val, DCSR_EBREAKH);
      state.dcsr.ebreaks = get_field(val, DCSR_EBREAKS);
      state.dcsr.ebreaku = get_field(val, DCSR_EBREAKU);
      state.dcsr.halt = get_field(val, DCSR_HALT);
      break;
    case CSR_DPC:
      state.dpc = val & ~(reg_t)1;
      break;
    case CSR_DSCRATCH:
      state.dscratch = val;
      break;
  }
  return old_val;
}

reg_t processor_t::get_csr(int which)
{
  uint32_t ctr_en = -1;
  if (state.prv < PRV_M)
    ctr_en &= state.mcounteren;
  if (state.prv < PRV_S)
    ctr_en &= state.scounteren;
  bool ctr_ok = (ctr_en >> (which & 31)) & 1;

  if (ctr_ok) {
    if (which >= CSR_HPMCOUNTER3 && which <= CSR_HPMCOUNTER31)
      return 0;
    if (xlen == 32 && which >= CSR_HPMCOUNTER3H && which <= CSR_HPMCOUNTER31H)
      return 0;
  }
  if (which >= CSR_MHPMCOUNTER3 && which <= CSR_MHPMCOUNTER31)
    return 0;
  if (xlen == 32 && which >= CSR_MHPMCOUNTER3H && which <= CSR_MHPMCOUNTER31H)
    return 0;
  if (which >= CSR_MHPMEVENT3 && which <= CSR_MHPMEVENT31)
    return 0;

  switch (which)
  {
#ifdef SPIKE_SIMPLEV
    case CSR_USVVL:
      return state.sv().vl;
    case CSR_USVCFG:
      return (state.sv().state_bank)      | (state.sv().state_size<<3);
    case CSR_USVSTATE:
      return (state.sv().vl-1)            | ((state.sv().mvl-1)<<6) |
             (state.sv().srcoffs<<12)     | (state.sv().destoffs<<18) |
             (state.sv().state_bank<<24)  | (state.sv().state_size<<26);
    case CSR_USVMVL:
      return state.sv().mvl;
    case CSR_SVREGTOP:
    case CSR_SVREGBOT:
      return 0;// XXX TODO: return correct entry
    case CSR_SVPREDCFG0:
    case CSR_SVPREDCFG1:
    case CSR_SVPREDCFG2:
    case CSR_SVPREDCFG3:
    case CSR_SVPREDCFG4:
    case CSR_SVPREDCFG5:
    case CSR_SVPREDCFG6:
    case CSR_SVPREDCFG7:
      return 0;// XXX TODO: return correct entry
    case CSR_UREMAP:
      return 0;// XXX TODO: return correct entry
    case CSR_USHAPE0:
    case CSR_USHAPE1:
    case CSR_USHAPE2:
      return 0;// XXX TODO: return correct entry
#endif
    case CSR_FFLAGS:
      require_fp;
      if (!supports_extension('F'))
        break;
      return state.fflags;
    case CSR_FRM:
      require_fp;
      if (!supports_extension('F'))
        break;
      return state.frm;
    case CSR_FCSR:
      require_fp;
      if (!supports_extension('F'))
        break;
      return (state.fflags << FSR_AEXC_SHIFT) | (state.frm << FSR_RD_SHIFT);
    case CSR_INSTRET:
    case CSR_CYCLE:
      if (ctr_ok)
        return state.minstret;
      break;
    case CSR_MINSTRET:
    case CSR_MCYCLE:
      return state.minstret;
    case CSR_INSTRETH:
    case CSR_CYCLEH:
      if (ctr_ok && xlen == 32)
        return state.minstret >> 32;
      break;
    case CSR_MINSTRETH:
    case CSR_MCYCLEH:
      if (xlen == 32)
        return state.minstret >> 32;
      break;
    case CSR_SCOUNTEREN: return state.scounteren;
    case CSR_MCOUNTEREN: return state.mcounteren;
    case CSR_SSTATUS: {
      reg_t mask = SSTATUS_SIE | SSTATUS_SPIE | SSTATUS_SPP | SSTATUS_FS
                 | SSTATUS_XS | SSTATUS_SUM | SSTATUS_MXR | SSTATUS_UXL;
      reg_t sstatus = state.mstatus & mask;
      if ((sstatus & SSTATUS_FS) == SSTATUS_FS ||
          (sstatus & SSTATUS_XS) == SSTATUS_XS)
        sstatus |= (xlen == 32 ? SSTATUS32_SD : SSTATUS64_SD);
      return sstatus;
    }
    case CSR_SIP: return state.mip & state.mideleg;
    case CSR_SIE: return state.mie & state.mideleg;
    case CSR_SEPC: return state.sepc & pc_alignment_mask();
    case CSR_STVAL: return state.stval;
    case CSR_STVEC: return state.stvec;
    case CSR_SCAUSE:
      if (max_xlen > xlen)
        return state.scause | ((state.scause >> (max_xlen-1)) << (xlen-1));
      return state.scause;
    case CSR_SATP:
      if (get_field(state.mstatus, MSTATUS_TVM))
        require_privilege(PRV_M);
      return state.satp;
    case CSR_SSCRATCH: return state.sscratch;
    case CSR_MSTATUS: return state.mstatus;
    case CSR_MIP: return state.mip;
    case CSR_MIE: return state.mie;
    case CSR_MEPC: return state.mepc & pc_alignment_mask();
    case CSR_MSCRATCH: return state.mscratch;
    case CSR_MCAUSE: return state.mcause;
    case CSR_MTVAL: return state.mtval;
    case CSR_MISA: return state.misa;
    case CSR_MARCHID: return 0;
    case CSR_MIMPID: return 0;
    case CSR_MVENDORID: return 0;
    case CSR_MHARTID: return id;
    case CSR_MTVEC: return state.mtvec;
    case CSR_MEDELEG: return state.medeleg;
    case CSR_MIDELEG: return state.mideleg;
    case CSR_TSELECT: return state.tselect;
    case CSR_TDATA1:
      if (state.tselect < state.num_triggers) {
        reg_t v = 0;
        mcontrol_t *mc = &state.mcontrol[state.tselect];
        v = set_field(v, MCONTROL_TYPE(xlen), mc->type);
        v = set_field(v, MCONTROL_DMODE(xlen), mc->dmode);
        v = set_field(v, MCONTROL_MASKMAX(xlen), mc->maskmax);
        v = set_field(v, MCONTROL_SELECT, mc->select);
        v = set_field(v, MCONTROL_TIMING, mc->timing);
        v = set_field(v, MCONTROL_ACTION, mc->action);
        v = set_field(v, MCONTROL_CHAIN, mc->chain);
        v = set_field(v, MCONTROL_MATCH, mc->match);
        v = set_field(v, MCONTROL_M, mc->m);
        v = set_field(v, MCONTROL_H, mc->h);
        v = set_field(v, MCONTROL_S, mc->s);
        v = set_field(v, MCONTROL_U, mc->u);
        v = set_field(v, MCONTROL_EXECUTE, mc->execute);
        v = set_field(v, MCONTROL_STORE, mc->store);
        v = set_field(v, MCONTROL_LOAD, mc->load);
        return v;
      } else {
        return 0;
      }
      break;
    case CSR_TDATA2:
      if (state.tselect < state.num_triggers) {
        return state.tdata2[state.tselect];
      } else {
        return 0;
      }
      break;
    case CSR_TDATA3: return 0;
    case CSR_DCSR:
      {
        uint32_t v = 0;
        v = set_field(v, DCSR_XDEBUGVER, 1);
        v = set_field(v, DCSR_EBREAKM, state.dcsr.ebreakm);
        v = set_field(v, DCSR_EBREAKH, state.dcsr.ebreakh);
        v = set_field(v, DCSR_EBREAKS, state.dcsr.ebreaks);
        v = set_field(v, DCSR_EBREAKU, state.dcsr.ebreaku);
        v = set_field(v, DCSR_STOPCYCLE, 0);
        v = set_field(v, DCSR_STOPTIME, 0);
        v = set_field(v, DCSR_CAUSE, state.dcsr.cause);
        v = set_field(v, DCSR_STEP, state.dcsr.step);
        v = set_field(v, DCSR_PRV, state.dcsr.prv);
        return v;
      }
    case CSR_DPC:
      return state.dpc & pc_alignment_mask();
    case CSR_DSCRATCH:
      return state.dscratch;
  }
  throw trap_illegal_instruction(0);
}

reg_t illegal_instruction(processor_t* p, insn_t insn, reg_t pc)
{
  throw trap_illegal_instruction(0);
}

insn_func_t processor_t::decode_insn(insn_t insn)
{
  // look up opcode in hash table
  size_t idx = insn.bits() % OPCODE_CACHE_SIZE;
  insn_desc_t desc = opcode_cache[idx];

  if (unlikely(insn.bits() != desc.match)) {
    // fall back to linear search
    insn_desc_t* p = &instructions[0];
    while ((insn.bits() & p->mask) != p->match)
      p++;
    desc = *p;

    if (p->mask != 0 && p > &instructions[0]) {
      if (p->match != (p-1)->match && p->match != (p+1)->match) {
        // move to front of opcode list to reduce miss penalty
        while (--p >= &instructions[0])
          *(p+1) = *p;
        instructions[0] = desc;
      }
    }

    opcode_cache[idx] = desc;
    opcode_cache[idx].match = insn.bits();
  }

  return xlen == 64 ? desc.rv64 : desc.rv32;
}

void processor_t::register_insn(insn_desc_t desc)
{
  instructions.push_back(desc);
}

void processor_t::build_opcode_map()
{
  struct cmp {
    bool operator()(const insn_desc_t& lhs, const insn_desc_t& rhs) {
      if (lhs.match == rhs.match)
        return lhs.mask > rhs.mask;
      return lhs.match > rhs.match;
    }
  };
  std::sort(instructions.begin(), instructions.end(), cmp());

  for (size_t i = 0; i < OPCODE_CACHE_SIZE; i++)
    opcode_cache[i] = {0, 0, &illegal_instruction, &illegal_instruction};
}

void processor_t::register_extension(extension_t* x)
{
  for (auto insn : x->get_instructions())
    register_insn(insn);
  build_opcode_map();
  for (auto disasm_insn : x->get_disasms())
    disassembler->add_insn(disasm_insn);
  if (ext != NULL)
    throw std::logic_error("only one extension may be registered");
  ext = x;
  x->set_processor(this);
}

void processor_t::register_base_instructions()
{
  #define DECLARE_INSN(name, match, mask) \
    insn_bits_t name##_match = (match), name##_mask = (mask);
  #include "encoding.h"
  #undef DECLARE_INSN

  #define DEFINE_INSN(name) \
    REGISTER_INSN(this, name, name##_match, name##_mask)
  #include "insn_list.h"
  #undef DEFINE_INSN

  register_insn({0, 0, &illegal_instruction, &illegal_instruction});
  build_opcode_map();
}

bool processor_t::load(reg_t addr, size_t len, uint8_t* bytes)
{
  switch (addr)
  {
    case 0:
      if (len <= 4) {
        memset(bytes, 0, len);
        bytes[0] = get_field(state.mip, MIP_MSIP);
        return true;
      }
      break;
  }

  return false;
}

bool processor_t::store(reg_t addr, size_t len, const uint8_t* bytes)
{
  switch (addr)
  {
    case 0:
      if (len <= 4) {
        state.mip = set_field(state.mip, MIP_MSIP, bytes[0]);
        return true;
      }
      break;
  }

  return false;
}

void processor_t::trigger_updated()
{
  mmu->flush_tlb();
  mmu->check_triggers_fetch = false;
  mmu->check_triggers_load = false;
  mmu->check_triggers_store = false;

  for (unsigned i = 0; i < state.num_triggers; i++) {
    if (state.mcontrol[i].execute) {
      mmu->check_triggers_fetch = true;
    }
    if (state.mcontrol[i].load) {
      mmu->check_triggers_load = true;
    }
    if (state.mcontrol[i].store) {
      mmu->check_triggers_store = true;
    }
  }
}