/* is it a type preserving mov, with ok flags? */
static bool is_eligible_mov(struct ir3_instruction *instr, bool allow_flags)
{
if (is_same_type_mov(instr)) {
struct ir3_register *dst = instr->regs[0];
struct ir3_register *src = instr->regs[1];
struct ir3_instruction *src_instr = ssa(src);
/* only if mov src is SSA (not const/immed): */
if (!src_instr)
return false;
/* no indirect: */
if (dst->flags & IR3_REG_RELATIV)
return false;
if (src->flags & IR3_REG_RELATIV)
return false;
if (src->flags & IR3_REG_ARRAY)
return false;
if (!allow_flags)
if (src->flags & (IR3_REG_FABS | IR3_REG_FNEG |
IR3_REG_SABS | IR3_REG_SNEG | IR3_REG_BNOT))
return false;
/* TODO: remove this hack: */
if (src_instr->opc == OPC_META_FO)
return false;
return true;
}
return false;
}
static void console_int(uint32_t port, void *data)
{
XENCONS_RING_IDX cons, prod;
cons = console.intf->in_cons;
prod = console.intf->in_prod;
rmb();
if (prod == cons)
return;
uint32_t in_size = prod - cons;
uint8_t buf[in_size];
uint8_t *ptr = buf;
ssa(SYS_STATS_IO_INPUT, in_size);
while (prod > cons)
{
int idx = MASK_XENCONS_IDX(cons++, console.intf->in);
#ifdef DEBUG_CONSOLE
if (debug_key(console.intf->in[idx]))
{
in_size--;
continue;
}
#endif
*ptr++ = console.intf->in[idx];
}
console.intf->in_cons = prod;
wmb();
if (console.attached)
outlet_new_data(console.attached, buf, in_size);
}
/* Handle special case of eliminating output mov, and similar cases where
* there isn't a normal "consuming" instruction. In this case we cannot
* collapse flags (ie. output mov from const, or w/ abs/neg flags, cannot
* be eliminated)
*/
static struct ir3_instruction *
eliminate_output_mov(struct ir3_instruction *instr)
{
if (is_eligible_mov(instr, false)) {
struct ir3_register *reg = instr->regs[1];
if (!(reg->flags & IR3_REG_ARRAY)) {
struct ir3_instruction *src_instr = ssa(reg);
debug_assert(src_instr);
return src_instr;
}
}
return instr;
}
ArgGroup& ArgGroup::typedValue(int i) {
// If there's exactly one register argument slot left, the whole TypedValue
// goes on the stack instead of being split between a register and the
// stack.
if (m_gpArgs.size() == num_arg_regs() - 1) {
m_override = &m_stkArgs;
}
static_assert(offsetof(TypedValue, m_data) == 0, "");
static_assert(offsetof(TypedValue, m_type) == 8, "");
ssa(i).type(i);
m_override = nullptr;
return *this;
}
void goto_symext::symex_dead(statet &state)
{
const goto_programt::instructiont &instruction=*state.source.pc;
const codet &code=to_code(instruction.code);
if(code.operands().size()!=1)
throw "dead expects one operand";
if(code.op0().id()!=ID_symbol)
throw "dead expects symbol as first operand";
// We increase the L2 renaming to make these non-deterministic.
// We also prevent propagation of old values.
ssa_exprt ssa(to_symbol_expr(code.op0()));
state.rename(ssa, ns, goto_symex_statet::L1);
// in case of pointers, put something into the value set
if(ns.follow(code.op0().type()).id()==ID_pointer)
{
exprt failed=
get_failed_symbol(to_symbol_expr(code.op0()), ns);
exprt rhs;
if(failed.is_not_nil())
{
address_of_exprt address_of_expr;
address_of_expr.object()=failed;
address_of_expr.type()=code.op0().type();
rhs=address_of_expr;
}
else
rhs=exprt(ID_invalid);
state.rename(rhs, ns, goto_symex_statet::L1);
state.value_set.assign(ssa, rhs, ns, true, false);
}
ssa_exprt ssa_lhs=to_ssa_expr(ssa);
const irep_idt &l1_identifier=ssa_lhs.get_identifier();
// prevent propagation
state.propagation.remove(l1_identifier);
// L2 renaming
if(state.level2.current_names.find(l1_identifier)!=
state.level2.current_names.end())
state.level2.increase_counter(l1_identifier);
}
/* propagate register flags from src to dst.. negates need special
* handling to cancel each other out.
*/
static void combine_flags(unsigned *dstflags, struct ir3_instruction *src)
{
unsigned srcflags = src->regs[1]->flags;
/* if what we are combining into already has (abs) flags,
* we can drop (neg) from src:
*/
if (*dstflags & IR3_REG_FABS)
srcflags &= ~IR3_REG_FNEG;
if (*dstflags & IR3_REG_SABS)
srcflags &= ~IR3_REG_SNEG;
if (srcflags & IR3_REG_FABS)
*dstflags |= IR3_REG_FABS;
if (srcflags & IR3_REG_SABS)
*dstflags |= IR3_REG_SABS;
if (srcflags & IR3_REG_FNEG)
*dstflags ^= IR3_REG_FNEG;
if (srcflags & IR3_REG_SNEG)
*dstflags ^= IR3_REG_SNEG;
if (srcflags & IR3_REG_BNOT)
*dstflags ^= IR3_REG_BNOT;
*dstflags &= ~IR3_REG_SSA;
*dstflags |= srcflags & IR3_REG_SSA;
*dstflags |= srcflags & IR3_REG_CONST;
*dstflags |= srcflags & IR3_REG_IMMED;
*dstflags |= srcflags & IR3_REG_RELATIV;
*dstflags |= srcflags & IR3_REG_ARRAY;
*dstflags |= srcflags & IR3_REG_HIGH;
/* if src of the src is boolean we can drop the (abs) since we know
* the source value is already a postitive integer. This cleans
* up the absnegs that get inserted when converting between nir and
* native boolean (see ir3_b2n/n2b)
*/
struct ir3_instruction *srcsrc = ssa(src->regs[1]);
if (srcsrc && is_bool(srcsrc))
*dstflags &= ~IR3_REG_SABS;
}
void console_write(const char *msg, int len) {
static int was_cr = 0;
int sent = 0;
while (sent < len)
{
XENCONS_RING_IDX cons, prod;
cons = console.intf->out_cons;
rmb();
prod = console.intf->out_prod;
// while ((sent < len) && (prod - cons < sizeof(console.intf->out)))
// console.intf->out[MASK_XENCONS_IDX(prod++, console.intf->out)] = msg[sent++];
//
// It may be possible to use stty or ESC sequence instead of this nastiness
//
while ((sent < len) && (prod - cons < sizeof(console.intf->out))) {
if (msg[sent] == '\n' && !was_cr)
{
int idx = MASK_XENCONS_IDX(prod, console.intf->out);
console.intf->out[idx] = '\r';
prod++;
if (prod - cons >= sizeof(console.intf->out))
break;
}
was_cr = (msg[sent] == '\r');
int idx = MASK_XENCONS_IDX(prod, console.intf->out);
console.intf->out[idx] = msg[sent++];
prod++;
}
console.intf->out_prod = prod;
wmb();
event_kick(console.chan);
}
ssa(SYS_STATS_IO_OUTPUT, len);
}
/* is it a type preserving mov, with ok flags? */
static bool is_eligible_mov(struct ir3_instruction *instr, bool allow_flags)
{
if (is_same_type_mov(instr)) {
struct ir3_register *dst = instr->regs[0];
struct ir3_register *src = instr->regs[1];
struct ir3_instruction *src_instr = ssa(src);
/* only if mov src is SSA (not const/immed): */
if (!src_instr)
return false;
/* no indirect: */
if (dst->flags & IR3_REG_RELATIV)
return false;
if (src->flags & IR3_REG_RELATIV)
return false;
if (!allow_flags)
if (src->flags & (IR3_REG_FABS | IR3_REG_FNEG |
IR3_REG_SABS | IR3_REG_SNEG | IR3_REG_BNOT))
return false;
/* TODO: remove this hack: */
if (src_instr->opc == OPC_META_FO)
return false;
/* TODO: we currently don't handle left/right neighbors
* very well when inserting parallel-copies into phi..
* to avoid problems don't eliminate a mov coming out
* of phi..
*/
if (src_instr->opc == OPC_META_PHI)
return false;
return true;
}
return false;
}
proc_t *scheduler_next(proc_t *current, int reds_left)
{
set_phase(PHASE_NEXT);
uint32_t reds_used = SLICE_REDUCTIONS -reds_left;
ssi(SYS_STATS_CTX_SWITCHES);
ssa(SYS_STATS_REDUCTIONS, reds_used);
current->total_reds += reds_used;
proc_t *next_proc = 0;
uint64_t ticks = monotonic_clock(); // freeze time
assert(current->my_queue == MY_QUEUE_NONE);
#ifdef PROFILE_HARNESS
static uint64_t proc_started_ns = 0;
if (proc_started_ns != 0)
prof_slice_complete(current->pid,
current->result.what, current->cap.ip, proc_started_ns, ticks);
#endif
proc_t *expired;
while ((expired = wait_list_expired(&queues.on_timed_receive, ticks)) != 0)
{
expired->cap.ip = expired->result.jump_to;
if (scheduler_park_runnable_N(expired) < 0)
scheduler_exit_process(expired, A_NO_MEMORY);
}
int memory_exhausted = 0;
switch (current->result.what)
{
case SLICE_RESULT_YIELD:
if (scheduler_park_runnable_N(current) < 0)
memory_exhausted = 1;
break;
case SLICE_RESULT_WAIT:
if (current->result.until_when == LING_INFINITY)
{
if (proc_list_put_N(&queues.on_infinite_receive, current) < 0)
memory_exhausted = 1;
else
current->my_queue = MY_QUEUE_INF_WAIT;
}
else
{
if (wait_list_put_N(&queues.on_timed_receive,
current, current->result.until_when) < 0)
memory_exhausted = 1;
else
current->my_queue = MY_QUEUE_TIMED_WAIT;
}
break;
case SLICE_RESULT_DONE:
scheduler_exit_process(current, A_NORMAL);
break;
case SLICE_RESULT_PURGE_PROCS:
// purge_module() call may have detected processes lingering on the old
// code - terminate them
if (scheduler_park_runnable_N(current) < 0)
memory_exhausted = 1;
for (int i = 0; i < num_purged; i++)
if (scheduler_signal_exit_N(purgatory[i], current->pid, A_KILL) < 0)
memory_exhausted = 1;
num_purged = 0;
break;
case SLICE_RESULT_EXIT:
scheduler_exit_process(current, current->result.reason);
// what about the returned value when main function just returns?
break;
case SLICE_RESULT_EXIT2:
// only needed to implement erlang:exit/2
if (scheduler_park_runnable_N(current) < 0 ||
(scheduler_signal_exit_N(current->result.victim,
current->pid,
current->result.reason2) < 0))
memory_exhausted = 1;
break;
case SLICE_RESULT_ERROR:
scheduler_exit_process(current, current->result.reason);
// how is this different from SLICE_RESULT_EXIT?
break;
case SLICE_RESULT_THROW:
scheduler_exit_process(current, current->result.reason);
// how is this different from SLICE_RESULT_EXIT?
break;
default:
{
assert(current->result.what == SLICE_RESULT_OUTLET_CLOSE);
if (scheduler_park_runnable_N(current) < 0)
memory_exhausted = 1;
outlet_t *closing = current->result.closing;
//assert(is_atom(current->result.why));
outlet_close(closing, current->result.why);
break;
}
}
if (memory_exhausted)
scheduler_exit_process(current, A_NO_MEMORY);
do_pending:
ticks = monotonic_clock();
while ((expired = wait_list_expired(&queues.on_timed_receive, ticks)) != 0)
{
expired->cap.ip = expired->result.jump_to;
if (scheduler_park_runnable_N(expired) < 0)
scheduler_exit_process(expired, A_NO_MEMORY);
}
set_phase(PHASE_EVENTS);
// software events/timeouts
net_check_timeouts();
etimer_expired(ticks);
// 'hardware' events
int nr_fired = events_do_pending();
update_event_times(nr_fired, ticks);
set_phase(PHASE_NEXT);
// select_runnable
if (!proc_queue_is_empty(&queues.high_prio))
next_proc = proc_queue_get(&queues.high_prio);
else if (normal_count < NORMAL_ADVANTAGE)
{
if (!proc_queue_is_empty(&queues.normal_prio))
next_proc = proc_queue_get(&queues.normal_prio);
else if (!proc_queue_is_empty(&queues.low_prio))
next_proc = proc_queue_get(&queues.low_prio);
normal_count++;
}
else
{
if (!proc_queue_is_empty(&queues.low_prio))
next_proc = proc_queue_get(&queues.low_prio);
else if (!proc_queue_is_empty(&queues.normal_prio))
next_proc = proc_queue_get(&queues.normal_prio);
normal_count = 0;
}
if (next_proc == 0)
{
// no runnable processes; poll for events from all three sources
// Beware that events_poll() reports events 5us after they occur. If
// a new event is expected very soon we are better off polling event
// bits manually (using events_do_pending())
// Devote a portion of time until the next event to gc waiting processes
garbage_collect_waiting_processes(expect_event_in_ns /2);
if (expect_event_in_ns < MANUAL_POLLING_THRESHOLD)
goto do_pending;
uint64_t next_ticks = wait_list_timeout(&queues.on_timed_receive);
uint64_t closest_timeout = etimer_closest_timeout();
if (closest_timeout < next_ticks)
next_ticks = closest_timeout;
closest_timeout = lwip_closest_timeout();
if (closest_timeout < next_ticks)
next_ticks = closest_timeout;
scheduler_runtime_update();
events_poll(next_ticks); // LING_INFINITY is big enough
scheduler_runtime_start();
goto do_pending;
}
next_proc->my_queue = MY_QUEUE_NONE;
//TODO: update stats
#ifdef PROFILE_HARNESS
proc_started_ns = ticks;
#endif
set_phase(PHASE_ERLANG);
return next_proc;
}
/**
* Find instruction src's which are mov's that can be collapsed, replacing
* the mov dst with the mov src
*/
static void
instr_cp(struct ir3_cp_ctx *ctx, struct ir3_instruction *instr)
{
struct ir3_register *reg;
if (instr->regs_count == 0)
return;
if (ir3_instr_check_mark(instr))
return;
/* walk down the graph from each src: */
foreach_src_n(reg, n, instr) {
struct ir3_instruction *src = ssa(reg);
if (!src)
continue;
instr_cp(ctx, src);
/* TODO non-indirect access we could figure out which register
* we actually want and allow cp..
*/
if (reg->flags & IR3_REG_ARRAY)
continue;
reg_cp(ctx, instr, reg, n);
}
if (instr->regs[0]->flags & IR3_REG_ARRAY) {
struct ir3_instruction *src = ssa(instr->regs[0]);
if (src)
instr_cp(ctx, src);
}
if (instr->address) {
instr_cp(ctx, instr->address);
ir3_instr_set_address(instr, eliminate_output_mov(instr->address));
}
/* we can end up with extra cmps.s from frontend, which uses a
*
* cmps.s p0.x, cond, 0
*
* as a way to mov into the predicate register. But frequently 'cond'
* is itself a cmps.s/cmps.f/cmps.u. So detect this special case and
* just re-write the instruction writing predicate register to get rid
* of the double cmps.
*/
if ((instr->opc == OPC_CMPS_S) &&
(instr->regs[0]->num == regid(REG_P0, 0)) &&
ssa(instr->regs[1]) &&
(instr->regs[2]->flags & IR3_REG_IMMED) &&
(instr->regs[2]->iim_val == 0)) {
struct ir3_instruction *cond = ssa(instr->regs[1]);
switch (cond->opc) {
case OPC_CMPS_S:
case OPC_CMPS_F:
case OPC_CMPS_U:
instr->opc = cond->opc;
instr->flags = cond->flags;
instr->cat2 = cond->cat2;
instr->address = cond->address;
instr->regs[1] = cond->regs[1];
instr->regs[2] = cond->regs[2];
break;
default:
break;
}
}
}
/**
* Handle cp for a given src register. This additionally handles
* the cases of collapsing immedate/const (which replace the src
* register with a non-ssa src) or collapsing mov's from relative
* src (which needs to also fixup the address src reference by the
* instruction).
*/
static void
reg_cp(struct ir3_cp_ctx *ctx, struct ir3_instruction *instr,
struct ir3_register *reg, unsigned n)
{
struct ir3_instruction *src = ssa(reg);
/* don't propagate copies into a PHI, since we don't know if the
* src block executed:
*/
if (instr->opc == OPC_META_PHI)
return;
if (is_eligible_mov(src, true)) {
/* simple case, no immed/const/relativ, only mov's w/ ssa src: */
struct ir3_register *src_reg = src->regs[1];
unsigned new_flags = reg->flags;
combine_flags(&new_flags, src);
if (valid_flags(instr, n, new_flags)) {
if (new_flags & IR3_REG_ARRAY) {
debug_assert(!(reg->flags & IR3_REG_ARRAY));
reg->array = src_reg->array;
}
reg->flags = new_flags;
reg->instr = ssa(src_reg);
}
src = ssa(reg); /* could be null for IR3_REG_ARRAY case */
if (!src)
return;
} else if (is_same_type_mov(src) &&
/* cannot collapse const/immed/etc into meta instrs: */
!is_meta(instr)) {
/* immed/const/etc cases, which require some special handling: */
struct ir3_register *src_reg = src->regs[1];
unsigned new_flags = reg->flags;
combine_flags(&new_flags, src);
if (!valid_flags(instr, n, new_flags)) {
/* See if lowering an immediate to const would help. */
if (valid_flags(instr, n, (new_flags & ~IR3_REG_IMMED) | IR3_REG_CONST)) {
debug_assert(new_flags & IR3_REG_IMMED);
instr->regs[n + 1] = lower_immed(ctx, src_reg, new_flags);
return;
}
/* special case for "normal" mad instructions, we can
* try swapping the first two args if that fits better.
*
* the "plain" MAD's (ie. the ones that don't shift first
* src prior to multiply) can swap their first two srcs if
* src[0] is !CONST and src[1] is CONST:
*/
if ((n == 1) && is_mad(instr->opc) &&
!(instr->regs[0 + 1]->flags & (IR3_REG_CONST | IR3_REG_RELATIV)) &&
valid_flags(instr, 0, new_flags)) {
/* swap src[0] and src[1]: */
struct ir3_register *tmp;
tmp = instr->regs[0 + 1];
instr->regs[0 + 1] = instr->regs[1 + 1];
instr->regs[1 + 1] = tmp;
n = 0;
} else {
return;
}
}
/* Here we handle the special case of mov from
* CONST and/or RELATIV. These need to be handled
* specially, because in the case of move from CONST
* there is no src ir3_instruction so we need to
* replace the ir3_register. And in the case of
* RELATIV we need to handle the address register
* dependency.
*/
if (src_reg->flags & IR3_REG_CONST) {
/* an instruction cannot reference two different
* address registers:
*/
if ((src_reg->flags & IR3_REG_RELATIV) &&
conflicts(instr->address, reg->instr->address))
return;
/* This seems to be a hw bug, or something where the timings
* just somehow don't work out. This restriction may only
* apply if the first src is also CONST.
*/
if ((opc_cat(instr->opc) == 3) && (n == 2) &&
(src_reg->flags & IR3_REG_RELATIV) &&
(src_reg->array.offset == 0))
return;
src_reg = ir3_reg_clone(instr->block->shader, src_reg);
src_reg->flags = new_flags;
instr->regs[n+1] = src_reg;
if (src_reg->flags & IR3_REG_RELATIV)
ir3_instr_set_address(instr, reg->instr->address);
return;
}
if ((src_reg->flags & IR3_REG_RELATIV) &&
!conflicts(instr->address, reg->instr->address)) {
src_reg = ir3_reg_clone(instr->block->shader, src_reg);
src_reg->flags = new_flags;
instr->regs[n+1] = src_reg;
ir3_instr_set_address(instr, reg->instr->address);
return;
}
/* NOTE: seems we can only do immed integers, so don't
* need to care about float. But we do need to handle
* abs/neg *before* checking that the immediate requires
* few enough bits to encode:
*
* TODO: do we need to do something to avoid accidentally
* catching a float immed?
*/
if (src_reg->flags & IR3_REG_IMMED) {
int32_t iim_val = src_reg->iim_val;
debug_assert((opc_cat(instr->opc) == 1) ||
(opc_cat(instr->opc) == 6) ||
ir3_cat2_int(instr->opc));
if (new_flags & IR3_REG_SABS)
iim_val = abs(iim_val);
if (new_flags & IR3_REG_SNEG)
iim_val = -iim_val;
if (new_flags & IR3_REG_BNOT)
iim_val = ~iim_val;
/* other than category 1 (mov) we can only encode up to 10 bits: */
if ((instr->opc == OPC_MOV) ||
!((iim_val & ~0x3ff) && (-iim_val & ~0x3ff))) {
new_flags &= ~(IR3_REG_SABS | IR3_REG_SNEG | IR3_REG_BNOT);
src_reg = ir3_reg_clone(instr->block->shader, src_reg);
src_reg->flags = new_flags;
src_reg->iim_val = iim_val;
instr->regs[n+1] = src_reg;
} else if (valid_flags(instr, n, (new_flags & ~IR3_REG_IMMED) | IR3_REG_CONST)) {
/* See if lowering an immediate to const would help. */
instr->regs[n+1] = lower_immed(ctx, src_reg, new_flags);
}
return;
}
}
}
static struct ir3_instruction *instr_get(void *arr, int idx)
{
return ssa(((struct ir3_instruction *)arr)->regs[idx+1]);
}
bool fixSSA(Procedure& proc)
{
PhaseScope phaseScope(proc, "fixSSA");
// Collect the stack "variables". If there aren't any, then we don't have anything to do.
// That's a fairly common case.
HashMap<StackSlotValue*, Type> stackVariable;
for (Value* value : proc.values()) {
if (StackSlotValue* stack = value->as<StackSlotValue>()) {
if (stack->kind() == StackSlotKind::Anonymous)
stackVariable.add(stack, Void);
}
}
if (stackVariable.isEmpty())
return false;
// Make sure that we know how to optimize all of these. We only know how to handle Load and
// Store on anonymous variables.
for (Value* value : proc.values()) {
auto reject = [&] (Value* value) {
if (StackSlotValue* stack = value->as<StackSlotValue>())
stackVariable.remove(stack);
};
auto handleAccess = [&] (Value* access, Type type) {
StackSlotValue* stack = access->lastChild()->as<StackSlotValue>();
if (!stack)
return;
if (value->as<MemoryValue>()->offset()) {
stackVariable.remove(stack);
return;
}
auto result = stackVariable.find(stack);
if (result == stackVariable.end())
return;
if (result->value == Void) {
result->value = type;
return;
}
if (result->value == type)
return;
stackVariable.remove(result);
};
switch (value->opcode()) {
case Load:
// We're OK with loads from stack variables at an offset of zero.
handleAccess(value, value->type());
break;
case Store:
// We're OK with stores to stack variables, but not storing stack variables.
reject(value->child(0));
handleAccess(value, value->child(0)->type());
break;
default:
for (Value* child : value->children())
reject(child);
break;
}
}
Vector<StackSlotValue*> deadValues;
for (auto& entry : stackVariable) {
if (entry.value == Void)
deadValues.append(entry.key);
}
for (StackSlotValue* deadValue : deadValues) {
deadValue->replaceWithNop();
stackVariable.remove(deadValue);
}
if (stackVariable.isEmpty())
return false;
// We know that we have variables to optimize, so do that now.
breakCriticalEdges(proc);
SSACalculator ssa(proc);
// Create a SSACalculator::Variable for every stack variable.
Vector<StackSlotValue*> variableToStack;
HashMap<StackSlotValue*, SSACalculator::Variable*> stackToVariable;
for (auto& entry : stackVariable) {
StackSlotValue* stack = entry.key;
SSACalculator::Variable* variable = ssa.newVariable();
RELEASE_ASSERT(variable->index() == variableToStack.size());
variableToStack.append(stack);
stackToVariable.add(stack, variable);
}
// Create Defs for all of the stores to the stack variable.
for (BasicBlock* block : proc) {
for (Value* value : *block) {
if (value->opcode() != Store)
continue;
StackSlotValue* stack = value->child(1)->as<StackSlotValue>();
if (!stack)
continue;
if (SSACalculator::Variable* variable = stackToVariable.get(stack))
ssa.newDef(variable, block, value->child(0));
}
}
// Decide where Phis are to be inserted. This creates them but does not insert them.
ssa.computePhis(
[&] (SSACalculator::Variable* variable, BasicBlock* block) -> Value* {
StackSlotValue* stack = variableToStack[variable->index()];
Value* phi = proc.add<Value>(Phi, stackVariable.get(stack), stack->origin());
if (verbose) {
dataLog(
"Adding Phi for ", pointerDump(stack), " at ", *block, ": ",
deepDump(proc, phi), "\n");
}
return phi;
});
// Now perform the conversion.
InsertionSet insertionSet(proc);
HashMap<StackSlotValue*, Value*> mapping;
for (BasicBlock* block : proc.blocksInPreOrder()) {
mapping.clear();
for (auto& entry : stackToVariable) {
StackSlotValue* stack = entry.key;
SSACalculator::Variable* variable = entry.value;
SSACalculator::Def* def = ssa.reachingDefAtHead(block, variable);
if (def)
mapping.set(stack, def->value());
}
for (SSACalculator::Def* phiDef : ssa.phisForBlock(block)) {
StackSlotValue* stack = variableToStack[phiDef->variable()->index()];
insertionSet.insertValue(0, phiDef->value());
mapping.set(stack, phiDef->value());
}
for (unsigned valueIndex = 0; valueIndex < block->size(); ++valueIndex) {
Value* value = block->at(valueIndex);
value->performSubstitution();
switch (value->opcode()) {
case Load: {
if (StackSlotValue* stack = value->child(0)->as<StackSlotValue>()) {
if (Value* replacement = mapping.get(stack))
value->replaceWithIdentity(replacement);
}
break;
}
case Store: {
if (StackSlotValue* stack = value->child(1)->as<StackSlotValue>()) {
if (stackToVariable.contains(stack)) {
mapping.set(stack, value->child(0));
value->replaceWithNop();
}
}
break;
}
default:
break;
}
}
unsigned upsilonInsertionPoint = block->size() - 1;
Origin upsilonOrigin = block->last()->origin();
for (BasicBlock* successorBlock : block->successorBlocks()) {
for (SSACalculator::Def* phiDef : ssa.phisForBlock(successorBlock)) {
Value* phi = phiDef->value();
SSACalculator::Variable* variable = phiDef->variable();
StackSlotValue* stack = variableToStack[variable->index()];
Value* mappedValue = mapping.get(stack);
if (verbose) {
dataLog(
"Mapped value for ", *stack, " with successor Phi ", *phi, " at end of ",
*block, ": ", pointerDump(mappedValue), "\n");
}
if (!mappedValue)
mappedValue = insertionSet.insertBottom(upsilonInsertionPoint, phi);
insertionSet.insert<UpsilonValue>(
upsilonInsertionPoint, upsilonOrigin, mappedValue, phi);
}
}
insertionSet.execute(block);
}
// Finally, kill the stack slots.
for (StackSlotValue* stack : variableToStack)
stack->replaceWithNop();
if (verbose) {
dataLog("B3 after SSA conversion:\n");
dataLog(proc);
}
return true;
}
/**
* Find instruction src's which are mov's that can be collapsed, replacing
* the mov dst with the mov src
*/
static void
instr_cp(struct ir3_cp_ctx *ctx, struct ir3_instruction *instr)
{
struct ir3_register *reg;
if (instr->regs_count == 0)
return;
if (ir3_instr_check_mark(instr))
return;
/* walk down the graph from each src: */
foreach_src_n(reg, n, instr) {
struct ir3_instruction *src = ssa(reg);
if (!src)
continue;
instr_cp(ctx, src);
/* TODO non-indirect access we could figure out which register
* we actually want and allow cp..
*/
if (reg->flags & IR3_REG_ARRAY)
continue;
/* Don't CP absneg into meta instructions, that won't end well: */
if (is_meta(instr) && (src->opc != OPC_MOV))
continue;
reg_cp(ctx, instr, reg, n);
}
if (instr->regs[0]->flags & IR3_REG_ARRAY) {
struct ir3_instruction *src = ssa(instr->regs[0]);
if (src)
instr_cp(ctx, src);
}
if (instr->address) {
instr_cp(ctx, instr->address);
ir3_instr_set_address(instr, eliminate_output_mov(instr->address));
}
/* we can end up with extra cmps.s from frontend, which uses a
*
* cmps.s p0.x, cond, 0
*
* as a way to mov into the predicate register. But frequently 'cond'
* is itself a cmps.s/cmps.f/cmps.u. So detect this special case and
* just re-write the instruction writing predicate register to get rid
* of the double cmps.
*/
if ((instr->opc == OPC_CMPS_S) &&
(instr->regs[0]->num == regid(REG_P0, 0)) &&
ssa(instr->regs[1]) &&
(instr->regs[2]->flags & IR3_REG_IMMED) &&
(instr->regs[2]->iim_val == 0)) {
struct ir3_instruction *cond = ssa(instr->regs[1]);
switch (cond->opc) {
case OPC_CMPS_S:
case OPC_CMPS_F:
case OPC_CMPS_U:
instr->opc = cond->opc;
instr->flags = cond->flags;
instr->cat2 = cond->cat2;
instr->address = cond->address;
instr->regs[1] = cond->regs[1];
instr->regs[2] = cond->regs[2];
instr->barrier_class |= cond->barrier_class;
instr->barrier_conflict |= cond->barrier_conflict;
unuse(cond);
break;
default:
break;
}
}
/* Handle converting a sam.s2en (taking samp/tex idx params via
* register) into a normal sam (encoding immediate samp/tex idx)
* if they are immediate. This saves some instructions and regs
* in the common case where we know samp/tex at compile time:
*/
if (is_tex(instr) && (instr->flags & IR3_INSTR_S2EN) &&
!(ir3_shader_debug & IR3_DBG_FORCES2EN)) {
/* The first src will be a fan-in (collect), if both of it's
* two sources are mov from imm, then we can
*/
struct ir3_instruction *samp_tex = ssa(instr->regs[1]);
debug_assert(samp_tex->opc == OPC_META_FI);
struct ir3_instruction *samp = ssa(samp_tex->regs[1]);
struct ir3_instruction *tex = ssa(samp_tex->regs[2]);
if ((samp->opc == OPC_MOV) &&
(samp->regs[1]->flags & IR3_REG_IMMED) &&
(tex->opc == OPC_MOV) &&
(tex->regs[1]->flags & IR3_REG_IMMED)) {
instr->flags &= ~IR3_INSTR_S2EN;
instr->cat5.samp = samp->regs[1]->iim_val;
instr->cat5.tex = tex->regs[1]->iim_val;
instr->regs[1]->instr = NULL;
}
}
}
void dBasicBlocksGraph::ConvertToSSA ()
{
dConvertToSSASolver ssa (this);
ssa.Solve();
}
void cgCallBuiltin(IRLS& env, const IRInstruction* inst) {
auto const extra = inst->extra<CallBuiltin>();
auto const callee = extra->callee;
auto const returnType = inst->typeParam();
auto const funcReturnType = callee->returnType();
auto const returnByValue = callee->isReturnByValue();
auto const dstData = dstLoc(env, inst, 0).reg(0);
auto const dstType = dstLoc(env, inst, 0).reg(1);
auto& v = vmain(env);
// Whether `t' is passed in/out of C++ as String&/Array&/Object&.
auto const isReqPtrRef = [] (MaybeDataType t) {
return isStringType(t) || isArrayLikeType(t) ||
t == KindOfObject || t == KindOfResource;
};
if (FixupMap::eagerRecord(callee)) {
auto const sp = srcLoc(env, inst, 1).reg();
auto const spOffset = cellsToBytes(extra->spOffset.offset);
auto const& marker = inst->marker();
auto const pc = marker.fixupSk().unit()->entry() + marker.fixupBcOff();
auto const synced_sp = v.makeReg();
v << lea{sp[spOffset], synced_sp};
emitEagerSyncPoint(v, pc, rvmtl(), srcLoc(env, inst, 0).reg(), synced_sp);
}
int returnOffset = rds::kVmMInstrStateOff +
offsetof(MInstrState, tvBuiltinReturn);
auto args = argGroup(env, inst);
if (!returnByValue) {
if (isBuiltinByRef(funcReturnType)) {
if (isReqPtrRef(funcReturnType)) {
returnOffset += TVOFF(m_data);
}
// Pass the address of tvBuiltinReturn to the native function as the
// location where it can construct the return Array, String, Object, or
// Variant.
args.addr(rvmtl(), returnOffset);
args.indirect();
}
}
// The srcs past the first two (sp and fp) are the arguments to the callee.
auto srcNum = uint32_t{2};
// Add the this_ or self_ argument for HNI builtins.
if (callee->isMethod()) {
if (callee->isStatic()) {
args.ssa(srcNum);
++srcNum;
} else {
// Note that we don't support objects with vtables here (if they may need
// a $this pointer adjustment). This should be filtered out during irgen
// or before.
args.ssa(srcNum);
++srcNum;
}
}
// Add the func_num_args() value if needed.
if (callee->attrs() & AttrNumArgs) {
// If `numNonDefault' is negative, this is passed as an src.
if (extra->numNonDefault >= 0) {
args.imm((int64_t)extra->numNonDefault);
} else {
args.ssa(srcNum);
++srcNum;
}
}
// Add the positional arguments.
for (uint32_t i = 0; i < callee->numParams(); ++i, ++srcNum) {
auto const& pi = callee->params()[i];
// Non-pointer and NativeArg args are passed by value. String, Array,
// Object, and Variant are passed by const&, i.e. a pointer to stack memory
// holding the value, so we expect PtrToT types for these. Pointers to
// req::ptr types (String, Array, Object) need adjusting to point to
// &ptr->m_data.
if (TVOFF(m_data) && !pi.nativeArg && isReqPtrRef(pi.builtinType)) {
assertx(inst->src(srcNum)->type() <= TPtrToGen);
args.addr(srcLoc(env, inst, srcNum).reg(), TVOFF(m_data));
} else if (pi.nativeArg && !pi.builtinType && !callee->byRef(i)) {
// This condition indicates a MixedTV (i.e., TypedValue-by-value) arg.
args.typedValue(srcNum);
} else {
args.ssa(srcNum, pi.builtinType == KindOfDouble);
}
}
auto dest = [&] () -> CallDest {
if (isBuiltinByRef(funcReturnType)) {
if (!returnByValue) return kVoidDest; // indirect return
return funcReturnType
? callDest(dstData) // String, Array, or Object
: callDest(dstData, dstType); // Variant
}
return funcReturnType == KindOfDouble
? callDestDbl(env, inst)
: callDest(env, inst);
}();
cgCallHelper(v, env, CallSpec::direct(callee->nativeFuncPtr()),
dest, SyncOptions::Sync, args);
// For primitive return types (int, bool, double) and returnByValue, the
// return value is already in dstData/dstType.
if (returnType.isSimpleType() || returnByValue) return;
// For return by reference (String, Object, Array, Variant), the builtin
// writes the return value into MInstrState::tvBuiltinReturn, from where it
// has to be tested and copied.
if (returnType.isReferenceType()) {
// The return type is String, Array, or Object; fold nullptr to KindOfNull.
assertx(isBuiltinByRef(funcReturnType) && isReqPtrRef(funcReturnType));
v << load{rvmtl()[returnOffset], dstData};
if (dstType.isValid()) {
auto const sf = v.makeReg();
auto const rtype = v.cns(returnType.toDataType());
auto const nulltype = v.cns(KindOfNull);
v << testq{dstData, dstData, sf};
v << cmovb{CC_Z, sf, rtype, nulltype, dstType};
}
return;
}
if (returnType <= TCell || returnType <= TBoxedCell) {
// The return type is Variant; fold KindOfUninit to KindOfNull.
assertx(isBuiltinByRef(funcReturnType) && !isReqPtrRef(funcReturnType));
static_assert(KindOfUninit == 0, "KindOfUninit must be 0 for test");
v << load{rvmtl()[returnOffset + TVOFF(m_data)], dstData};
if (dstType.isValid()) {
auto const rtype = v.makeReg();
v << loadb{rvmtl()[returnOffset + TVOFF(m_type)], rtype};
auto const sf = v.makeReg();
auto const nulltype = v.cns(KindOfNull);
v << testb{rtype, rtype, sf};
v << cmovb{CC_Z, sf, rtype, nulltype, dstType};
}
return;
}
not_reached();
}
void goto_symext::symex_decl(statet &state, const symbol_exprt &expr)
{
// We increase the L2 renaming to make these non-deterministic.
// We also prevent propagation of old values.
ssa_exprt ssa(expr);
state.rename(ssa, ns, goto_symex_statet::L1);
const irep_idt &l1_identifier=ssa.get_identifier();
// rename type to L2
state.rename(ssa.type(), l1_identifier, ns);
ssa.update_type();
// in case of pointers, put something into the value set
if(ns.follow(expr.type()).id()==ID_pointer)
{
exprt failed=
get_failed_symbol(expr, ns);
exprt rhs;
if(failed.is_not_nil())
{
address_of_exprt address_of_expr;
address_of_expr.object()=failed;
address_of_expr.type()=expr.type();
rhs=address_of_expr;
}
else
rhs=exprt(ID_invalid);
state.rename(rhs, ns, goto_symex_statet::L1);
state.value_set.assign(ssa, rhs, ns, true, false);
}
// prevent propagation
state.propagation.remove(l1_identifier);
// L2 renaming
// inlining may yield multiple declarations of the same identifier
// within the same L1 context
if(state.level2.current_names.find(l1_identifier)==
state.level2.current_names.end())
state.level2.current_names[l1_identifier]=std::make_pair(ssa, 0);
state.level2.increase_counter(l1_identifier);
const bool record_events=state.record_events;
state.record_events=false;
state.rename(ssa, ns);
state.record_events=record_events;
// we hide the declaration of auxiliary variables
// and if the statement itself is hidden
bool hidden=
ns.lookup(expr.get_identifier()).is_auxiliary ||
state.top().hidden_function ||
state.source.pc->source_location.get_hide();
target.decl(
state.guard.as_expr(),
ssa,
state.source,
hidden?symex_targett::HIDDEN:symex_targett::STATE);
assert(state.dirty);
if((*state.dirty)(ssa.get_object_name()) &&
state.atomic_section_id==0)
target.shared_write(
state.guard.as_expr(),
ssa,
state.atomic_section_id,
state.source);
}
