| Line 43... |
Line 43... |
#include "op_support.h"
|
#include "op_support.h"
|
|
|
#include "rec_i386.h"
|
#include "rec_i386.h"
|
|
|
/* Stuff that is really a `micro' operation but is rather big (or for some other
|
/* Stuff that is really a `micro' operation but is rather big (or for some other
|
* reason (like calling exit()) */
|
* reason like calling exit()) */
|
|
|
void upd_reg_from_t(oraddr_t pc)
|
void upd_reg_from_t(oraddr_t pc, int bound)
|
{
|
{
|
int reg;
|
int reg;
|
|
|
reg = cpu_state.curr_page->ts[(pc & (PAGE_SIZE - 1)) / 2];
|
pc = ((pc & (PAGE_SIZE - 1)) / 4);
|
|
|
|
if(bound) {
|
|
reg = cpu_state.curr_page->ts_bound[pc + 1];
|
|
} else
|
|
reg = cpu_state.curr_page->ts_during[pc];
|
|
|
if(reg & 0x1f)
|
if(reg & 0x1f)
|
cpu_state.reg[reg & 0x1f] = cpu_state.t0;
|
cpu_state.reg[reg & 0x1f] = cpu_state.t0;
|
|
|
if((reg >> 5) & 0x1f)
|
if((reg >> 5) & 0x1f)
|
| Line 63... |
Line 68... |
cpu_state.reg[(reg >> 10) & 0x1f] = cpu_state.t2;
|
cpu_state.reg[(reg >> 10) & 0x1f] = cpu_state.t2;
|
}
|
}
|
|
|
void op_support_nop_exit(void)
|
void op_support_nop_exit(void)
|
{
|
{
|
upd_reg_from_t(get_pc());
|
upd_reg_from_t(get_pc(), 0);
|
PRINTF("exit(%"PRIdREG")\n", cpu_state.reg[3]);
|
PRINTF("exit(%"PRIxREG")\n", cpu_state.reg[3]);
|
fprintf(stderr, "@reset : cycles %lld, insn #%lld\n",
|
fprintf(stderr, "@reset : cycles %lld, insn #%lld\n",
|
runtime.sim.reset_cycles, runtime.cpu.reset_instructions);
|
runtime.sim.reset_cycles, runtime.cpu.reset_instructions);
|
fprintf(stderr, "@exit : cycles %lld, insn #%lld\n", runtime.sim.cycles,
|
fprintf(stderr, "@exit : cycles %lld, insn #%lld\n", runtime.sim.cycles,
|
runtime.cpu.instructions);
|
runtime.cpu.instructions);
|
fprintf(stderr, " diff : cycles %lld, insn #%lld\n",
|
fprintf(stderr, " diff : cycles %lld, insn #%lld\n",
|
runtime.sim.cycles - runtime.sim.reset_cycles,
|
runtime.sim.cycles - runtime.sim.reset_cycles,
|
runtime.cpu.instructions - runtime.cpu.reset_instructions);
|
runtime.cpu.instructions - runtime.cpu.reset_instructions);
|
/* FIXME: Implement emulation of a stalled cpu
|
/* FIXME: Implement emulation of a stalled cpu
|
if (config.debug.gdb_enabled)
|
if (config.debug.gdb_enabled)
|
set_stall_state (1);
|
set_stall_state (1);
|
else
|
else {
|
runtime.sim.cont_run = 0;
|
handle_sim_command();
|
|
sim_done();
|
|
}
|
*/
|
*/
|
exit(0);
|
exit(0);
|
}
|
}
|
|
|
void op_support_nop_reset(void)
|
void op_support_nop_reset(void)
|
| Line 92... |
Line 99... |
runtime.cpu.reset_instructions = runtime.cpu.instructions;
|
runtime.cpu.reset_instructions = runtime.cpu.instructions;
|
}
|
}
|
|
|
void op_support_nop_printf(void)
|
void op_support_nop_printf(void)
|
{
|
{
|
upd_reg_from_t(get_pc());
|
upd_reg_from_t(get_pc(), 0);
|
simprintf(cpu_state.reg[4], cpu_state.reg[3]);
|
simprintf(cpu_state.reg[4], cpu_state.reg[3]);
|
}
|
}
|
|
|
void op_support_nop_report(void)
|
void op_support_nop_report(void)
|
{
|
{
|
upd_reg_from_t(get_pc());
|
upd_reg_from_t(get_pc(), 0);
|
PRINTF("report(0x%"PRIxREG");\n", cpu_state.reg[3]);
|
PRINTF("report(0x%"PRIxREG");\n", cpu_state.reg[3]);
|
}
|
}
|
|
|
void op_support_nop_report_imm(int imm)
|
void op_support_nop_report_imm(int imm)
|
{
|
{
|
upd_reg_from_t(get_pc());
|
upd_reg_from_t(get_pc(), 0);
|
PRINTF("report %i (0x%"PRIxREG");\n", imm, cpu_state.reg[3]);
|
PRINTF("report %i (0x%"PRIxREG");\n", imm, cpu_state.reg[3]);
|
}
|
}
|
|
|
/* Handles a jump */
|
/* Handles a jump */
|
/* addr is a VIRTUAL address */
|
/* addr is a VIRTUAL address */
|
| Line 116... |
Line 123... |
* simulator (most likely without -fomit-frame-pointer) and thus env will point
|
* simulator (most likely without -fomit-frame-pointer) and thus env will point
|
* to some bogus value. */
|
* to some bogus value. */
|
void do_jump(oraddr_t addr)
|
void do_jump(oraddr_t addr)
|
{
|
{
|
struct dyn_page *target_dp;
|
struct dyn_page *target_dp;
|
struct x_ref *xref;
|
|
oraddr_t phys_page;
|
oraddr_t phys_page;
|
|
|
|
/* Temporaries are always shipped out */
|
|
cpu_state.ts_current = 1;
|
|
|
/* The pc is set to the location of the jump in op_set_pc_preemt(_check) and
|
/* The pc is set to the location of the jump in op_set_pc_preemt(_check) and
|
* then it is incermented by 4 when the scheduler is run. If a scheduled job
|
* then it is incermented by 4 when the scheduler is run. If a scheduled job
|
* so happens to raise an exception cpu_state.delay_insn will still be set and
|
* so happens to raise an exception cpu_state.delay_insn will still be set and
|
* so except_handle will do its pc adjusting magic (ie. -4 from it) and every-
|
* so except_handle will do its pc adjusting magic (ie. -4 from it) and every-
|
* thing ends up just working right, except when a scheduled job does not
|
* thing ends up just working right, except when a scheduled job does not
|
| Line 154... |
Line 163... |
/* Since writes to the 0x0-0xff range do not dirtyfy a page recompile the 0x0
|
/* Since writes to the 0x0-0xff range do not dirtyfy a page recompile the 0x0
|
* page if the jump is to that location */
|
* page if the jump is to that location */
|
if(phys_page < 0x100)
|
if(phys_page < 0x100)
|
target_dp->dirty = 1;
|
target_dp->dirty = 1;
|
|
|
/* Check if this location is cross-referenced */
|
|
if(!(xref = find_host_x_ref(target_dp->xrefs, phys_page))) {
|
|
target_dp->dirty = 1;
|
|
xref = add_to_xrefs(target_dp, phys_page);
|
|
if(cpu_state.curr_page)
|
|
add_to_held_xrefs(cpu_state.curr_page, xref);
|
|
} else {
|
|
/* Only increment reference count if this page didn't already */
|
|
if(cpu_state.curr_page && !find_held_x_ref(cpu_state.curr_page->held_xrefs,
|
|
phys_page)) {
|
|
xref->ref++;
|
|
add_to_held_xrefs(cpu_state.curr_page, xref);
|
|
}
|
|
}
|
|
|
|
if(target_dp->dirty)
|
if(target_dp->dirty)
|
recompile_page(target_dp);
|
recompile_page(target_dp);
|
|
|
cpu_state.curr_page = target_dp;
|
cpu_state.curr_page = target_dp;
|
|
|
| Line 184... |
Line 178... |
else
|
else
|
upd_cycles_dec(target_dp->delayr);
|
upd_cycles_dec(target_dp->delayr);
|
|
|
cpu_state.ts_current = 0;
|
cpu_state.ts_current = 0;
|
|
|
/* Initially this (and do_rfe/handle_except) returned the address that we
|
/* Initially this returned the address that we should jump to and then the
|
* should jump to and then the recompiled code performed the jump. This was
|
* recompiled code performed the jump. This was no problem if the jump was
|
* no problem if the jump was trully an interpage jump or if the location
|
* trully an interpage jump or if the location didn't need recompileation. If
|
* didn't need recompileation. If the jump is page local and the page needs
|
* the jump is page local and the page needs recompileation there is a very
|
* recompileation there is a very high probability that the page will move in
|
* high probability that the page will move in memory and then the return
|
* memory and then the return address that is on the stack will point to
|
* address that is on the stack will point to memory that has already been
|
* memory that has already been freed, sometimes leading to crashes */
|
* freed, sometimes leading to crashes */
|
/* This looks like it could really be simpler, but no it can't. The only
|
/* This looks like it could really be simpler, but no it can't. The only
|
* issue here is the stack: it has to be unwound. This function is called
|
* issue here is the stack: it has to be unwound. This function is called
|
* from except_handle, which generally ends up quite high on the stack... */
|
* from except_handle, which generally ends up quite high on the stack... */
|
or_longjmp(xref->dyn_addr);
|
enter_dyn_code(phys_page, target_dp);
|
}
|
}
|
|
|
/* l.rfe is a hard instruction to emulate. One could just call
|
/* Wrapper around analysis() that contains all the recompiler specific stuff */
|
* do_jump(cpu_state.sprs[SPR_EPCR_BASE]), but then the location that we jump to
|
void op_support_analysis(void)
|
* will get cross referenced and because the page that contains the exception
|
|
* handlers is very rearly marked as dirty it will accumulate alot of held
|
|
* cross references over time. */
|
|
void do_rfe(void)
|
|
{
|
{
|
struct dyn_page *target_dp;
|
upd_sim_cycles();
|
struct x_ref *xref;
|
if(ADDR_PAGE(cpu_state.pc) != cpu_state.pc)
|
oraddr_t phys_page;
|
upd_reg_from_t(cpu_state.pc - (cpu_state.delay_insn ? 4 : 0), 0);
|
int already_held = 0;
|
|
|
|
set_pc(cpu_state.sprs[SPR_EPCR_BASE]);
|
|
|
|
phys_page = immu_translate(cpu_state.sprs[SPR_EPCR_BASE]);
|
|
|
|
/* Same reason as in do_jump() */
|
|
runtime.sim.mem_cycles = 0;
|
|
|
|
/* op_do_sched has run by the time this is run, which makes the pc point to
|
|
* the instruction after l.rfe. */
|
|
printf("Returning from exception to %"PRIxADDR" from %"PRIxADDR"\n",
|
|
phys_page, cpu_state.sprs[SPR_PPC]);
|
|
|
|
target_dp = find_dynd_page(phys_page);
|
|
|
|
if(!target_dp)
|
|
target_dp = new_dp(phys_page);
|
|
|
|
/* Since writes to the 0x0-0xff range do not dirtyfy a page recompile the 0x0
|
|
* page if the jump is to that location */
|
|
if(phys_page < 0x100)
|
|
target_dp->dirty = 1;
|
|
|
|
/* Check if this location is cross-referenced */
|
|
if(!(xref = find_host_x_ref(target_dp->xrefs, phys_page))) {
|
|
xref = add_to_xrefs(target_dp, phys_page);
|
|
/* Calling dirtyfy_page is real tempting but if we get to the situation were
|
|
* the l.rfe instruction and the location to which it returns to are on the
|
|
* same page then all the exception cross references will get removed and
|
|
* this will result in excessive recompileations of this page */
|
|
target_dp->dirty = 1;
|
|
|
|
/* There is alot of code (especially in linux) that do loops like this:
|
|
* int a;
|
|
* // Stuff such that b gets on another page than a
|
|
* int b;
|
|
* for(i = 0; i < (some big value); i++) {
|
|
* a = b;
|
|
* // Some more stuff
|
|
* }
|
|
* Here a DTLB miss will happen on every acess to a and b and l.rfe will
|
|
* always return to the same locations but since the previous l.rfe to this
|
|
* page was to a different location the page will get recompiled each time a
|
|
* or b is acessed. This is why the last NUM_RFE_HELD returns are `cached'.
|
|
*/
|
|
if(++cpu_state.rfe_held_xref_pos == NUM_RFE_HELD)
|
|
cpu_state.rfe_held_xref_pos = 0;
|
|
|
|
if(cpu_state.rfe_held_xrefs[cpu_state.rfe_held_xref_pos])
|
|
cpu_state.rfe_held_xrefs[cpu_state.rfe_held_xref_pos]->ref--;
|
|
|
|
cpu_state.rfe_held_xrefs[cpu_state.rfe_held_xref_pos] = xref;
|
|
} else {
|
|
/* Make sure we increase this cross reference's reference count, since it is
|
|
* decremented below. */
|
|
xref->ref++;
|
|
already_held = 1;
|
|
}
|
|
|
|
if(target_dp->dirty)
|
|
recompile_page(target_dp);
|
|
|
|
if(already_held)
|
|
xref->ref--;
|
|
|
|
cpu_state.curr_page = target_dp;
|
|
|
|
/* FIXME: If the page is backed by more than one type of memory, this will
|
|
* produce wrong results */
|
|
if(cpu_state.sprs[SPR_SR] & SPR_SR_IME)
|
|
/* Add the mmu hit delay to the cycle counter */
|
|
upd_cycles_dec(target_dp->delayr - config.immu.hitdelay);
|
|
else
|
else
|
upd_cycles_dec(target_dp->delayr);
|
upd_reg_from_t(cpu_state.pc, 0);
|
|
runtime.cpu.instructions++;
|
cpu_state.ts_current = 0;
|
analysis(&cpu_state.iqueue);
|
|
|
/* See the comment at the end of do_jump */
|
|
or_longjmp(xref->dyn_addr);
|
|
}
|
|
|
|
/* Handles an exception. */
|
|
void handle_except(oraddr_t except)
|
|
{
|
|
struct dyn_page *target_dp;
|
|
struct x_ref *xref;
|
|
|
|
/* NOTE: It is known when this code will be run. It is therefore not
|
|
* necessary to have to plough through cpu_state.curr_page->ts to store the
|
|
* temporaries. On the other hand, except_handle is also called from the
|
|
* scheduler, therefore we don't know when it is called and we can't move the
|
|
* temporaries to their permanent storeage in the recompiled code. */
|
|
|
|
/* op_do_sched has run by the time we run this, which makes the pc point to
|
|
* the next instruction. */
|
|
printf("Exception %"PRIxADDR" (%s) from %"PRIxADDR"\n", except,
|
|
except_name(except), get_pc() - 4);
|
|
|
|
set_pc(except);
|
|
|
|
target_dp = find_dynd_page(except);
|
|
|
|
if(!target_dp)
|
|
target_dp = new_dp(except);
|
|
|
|
/* Check if this location is cross-referenced */
|
|
if(!(xref = find_host_x_ref(target_dp->xrefs, except))) {
|
|
/* See the comment in do_rfe for why dirtyfy page is not called */
|
|
target_dp->dirty = 1;
|
|
xref = add_to_xrefs(target_dp, except);
|
|
} else {
|
|
/* If this cross reference is scheduled for removal increment its reference
|
|
* count */
|
|
if(!xref->ref)
|
|
xref->ref++;
|
|
}
|
|
|
|
if(target_dp->dirty)
|
|
recompile_page(target_dp);
|
|
|
|
cpu_state.curr_page = target_dp;
|
|
|
|
/* FIXME: If the page is backed by more than one type of memory, this will
|
|
* produce wrong results */
|
|
/* Address translation is disabled above (no need to add hitdelay) */
|
|
upd_cycles_dec(target_dp->delayr);
|
|
|
|
cpu_state.ts_current = 0;
|
|
|
|
/* See the comment at the end of do_jump */
|
|
or_longjmp(xref->dyn_addr);
|
|
}
|
}
|
|
|
|
|
No newline at end of file
|
No newline at end of file
|
