#endif
}
+// This is expected to be inlined by the compiler so each use of execute_insn
+// includes a duplicated body of the function to get separate fetch.func
+// function calls.
static reg_t execute_insn(processor_t* p, reg_t pc, insn_fetch_t fetch)
{
commit_log_stash_privilege(p->get_state());
return npc;
}
+bool processor_t::slow_path()
+{
+ return debug || state.single_step != state.STEP_NONE || state.dcsr.cause;
+}
+
// fetch/decode/execute loop
void processor_t::step(size_t n)
{
if (state.dcsr.cause == DCSR_CAUSE_NONE) {
- // TODO: get_interrupt() isn't super fast. Does that matter?
- if (sim->debug_module.get_interrupt(id)) {
+ if (halt_request) {
enter_debug_mode(DCSR_CAUSE_DEBUGINT);
- } else if (state.dcsr.halt) {
+ } // !!!The halt bit in DCSR is deprecated.
+ else if (state.dcsr.halt) {
enter_debug_mode(DCSR_CAUSE_HALT);
}
- } else {
- // In Debug Mode, just do 11 steps at a time. Otherwise we're going to be
- // spinning the rest of the time anyway.
- n = std::min(n, (size_t) 11);
}
while (n > 0) {
if (unlikely(invalid_pc(pc))) { \
switch (pc) { \
case PC_SERIALIZE_BEFORE: state.serialized = true; break; \
- case PC_SERIALIZE_AFTER: instret++; break; \
+ case PC_SERIALIZE_AFTER: n = ++instret; break; \
default: abort(); \
} \
pc = state.pc; \
try
{
- take_interrupt();
+ take_pending_interrupt();
- // When we might single step, use the slow loop instead of the fast one.
- if (unlikely(debug || state.single_step != state.STEP_NONE || state.dcsr.cause))
+ if (unlikely(slow_path()))
{
while (instret < n)
{
- // TODO: implement this for the main loop also. To keep
- // performance good, probably go into this version when entering
- // debug mode or something.
if (unlikely(state.single_step == state.STEP_STEPPING)) {
state.single_step = state.STEP_STEPPED;
- } else if (unlikely(state.single_step == state.STEP_STEPPED)) {
- state.single_step = state.STEP_NONE;
- enter_debug_mode(DCSR_CAUSE_STEP);
- // enter_debug_mode changed state.pc, so we can't just continue.
- break;
}
insn_fetch_t fetch = mmu->load_insn(pc);
if (debug && !state.serialized)
disasm(fetch.insn);
pc = execute_insn(this, pc, fetch);
+ bool serialize_before = (pc == PC_SERIALIZE_BEFORE);
+
advance_pc();
+
+ if (unlikely(state.single_step == state.STEP_STEPPED) && !serialize_before) {
+ state.single_step = state.STEP_NONE;
+ enter_debug_mode(DCSR_CAUSE_STEP);
+ // enter_debug_mode changed state.pc, so we can't just continue.
+ break;
+ }
+
+ if (unlikely(state.pc >= DEBUG_START &&
+ state.pc < DEBUG_END)) {
+ // We're waiting for the debugger to tell us something.
+ return;
+ }
+
+
+
}
}
else while (instret < n)
{
+ // This code uses a modified Duff's Device to improve the performance
+ // of executing instructions. While typical Duff's Devices are used
+ // for software pipelining, the switch statement below primarily
+ // benefits from separate call points for the fetch.func function call
+ // found in each execute_insn. This function call is an indirect jump
+ // that depends on the current instruction. By having an indirect jump
+ // dedicated for each icache entry, you improve the performance of the
+ // host's next address predictor. Each case in the switch statement
+ // allows for the program flow to contine to the next case if it
+ // corresponds to the next instruction in the program and instret is
+ // still less than n.
+ //
+ // According to Andrew Waterman's recollection, this optimization
+ // resulted in approximately a 2x performance increase.
+ //
+ // If there is support for compressed instructions, the mmu and the
+ // switch statement get more complicated. Each branch target is stored
+ // in the index corresponding to mmu->icache_index(), but consecutive
+ // non-branching instructions are stored in consecutive indices even if
+ // mmu->icache_index() specifies a different index (which is the case
+ // for 32-bit instructions in the presence of compressed instructions).
+
+ // This figures out where to jump to in the switch statement
size_t idx = _mmu->icache_index(pc);
+
+ // This gets the cached decoded instruction form the MMU. If the MMU
+ // does not have the current pc cached, it will refill the MMU and
+ // return the correct entry. ic_entry->data.func is the C++ function
+ // corresponding to the instruction.
auto ic_entry = _mmu->access_icache(pc);
+ // This macro is included in "icache.h" included within the switch
+ // statement below. The indirect jump corresponding to the instruction
+ // is located within the execute_insn() function call.
#define ICACHE_ACCESS(i) { \
insn_fetch_t fetch = ic_entry->data; \
ic_entry++; \
state.pc = pc; \
}
+ // This switch statement implements the modified Duff's device as
+ // explained above.
switch (idx) {
+ // "icache.h" is generated by the gen_icache script
#include "icache.h"
}
{
take_trap(t, pc);
n = instret;
+
+ if (unlikely(state.single_step == state.STEP_STEPPED)) {
+ state.single_step = state.STEP_NONE;
+ enter_debug_mode(DCSR_CAUSE_STEP);
+ }
+ }
+ catch (trigger_matched_t& t)
+ {
+ if (mmu->matched_trigger) {
+ // This exception came from the MMU. That means the instruction hasn't
+ // fully executed yet. We start it again, but this time it won't throw
+ // an exception because matched_trigger is already set. (All memory
+ // instructions are idempotent so restarting is safe.)
+
+ insn_fetch_t fetch = mmu->load_insn(pc);
+ pc = execute_insn(this, pc, fetch);
+ advance_pc();
+
+ delete mmu->matched_trigger;
+ mmu->matched_trigger = NULL;
+ }
+ switch (state.mcontrol[t.index].action) {
+ case ACTION_DEBUG_MODE:
+ enter_debug_mode(DCSR_CAUSE_HWBP);
+ break;
+ case ACTION_DEBUG_EXCEPTION: {
+ mem_trap_t trap(CAUSE_BREAKPOINT, t.address);
+ take_trap(trap, pc);
+ break;
+ }
+ default:
+ abort();
+ }
}
state.minstret += instret;