Commit 7a1aca9c by Richard Sandiford Committed by Richard Sandiford

sched-int.h (_haifa_insn_data): Move priority_status.

gcc/
	* sched-int.h (_haifa_insn_data): Move priority_status.
	Add model_index.
	(INSN_MODEL_INDEX): New macro.
	* haifa-sched.c (insn_delay): New function.
	(sched_regno_pressure_class): Update commentary.
	(mark_regno_birth_or_death): Pass the liveness bitmap and
	pressure array as arguments, instead of using curr_reg_live and
	curr_reg_pressure.  Only update the pressure if the bit in the
	liveness set has changed.
	(initiate_reg_pressure_info): Always trust the live-in set for
	SCHED_PRESSURE_MODEL.
	(initiate_bb_reg_pressure_info): Update call to
	mark_regno_birth_or_death.
	(dep_list_size): Take the list as argument.
	(calculate_reg_deaths): New function, extracted from...
	(setup_insn_reg_pressure_info): ...here.
	(MODEL_BAR): New macro.
	(model_pressure_data, model_insn_info, model_pressure_limit)
	(model_pressure_group): New structures.
	(model_schedule, model_worklist, model_insns, model_num_insns)
	(model_curr_point, model_before_pressure, model_next_priority):
	New variables.
	(MODEL_PRESSURE_DATA, MODEL_MAX_PRESSURE, MODEL_REF_PRESSURE)
	(MODEL_INSN_INFO, MODEL_INSN): New macros.
	(model_index, model_update_limit_points_in_group): New functions.
	(model_update_limit_points, model_last_use_except): Likewise.
	(model_start_update_pressure, model_update_pressure): Likewise.
	(model_recompute, model_spill_cost, model_excess_group_cost): Likewise.
	(model_excess_cost, model_dump_pressure_points): Likewise.
	(model_set_excess_costs): Likewise.
	(rank_for_schedule): Extend SCHED_PRIORITY_WEIGHTED ordering to
	SCHED_PRIORITY_MODEL.  Use insn_delay.  Use the order in the model
	schedule as an alternative tie-breaker.  Update the call to
	dep_list_size.
	(ready_sort): Call model_set_excess_costs.
	(update_register_pressure): Update call to mark_regno_birth_or_death.
	Rely on that function to check liveness rather than doing it here.
	(model_classify_pressure, model_order_p, model_add_to_worklist_at)
	(model_remove_from_worklist, model_add_to_worklist, model_promote_insn)
	(model_add_to_schedule, model_analyze_insns, model_init_pressure_group)
	(model_record_pressure, model_record_pressures): New functions.
	(model_record_final_pressures, model_add_successors_to_worklist)
	(model_promote_predecessors, model_choose_insn): Likewise.
	(model_reset_queue_indices, model_dump_pressure_summary): Likewise.
	(model_start_schedule, model_finalize_pressure_group): Likewise.
	(model_end_schedule): Likewise.
	(schedule_insn): Say when we're scheduling the next instruction
	in the model schedule.
	(schedule_insn): Handle SCHED_PRESSURE_MODEL.
	(queue_to_ready): Do not add instructions that are
	MAX_SCHED_READY_INSNS beyond the current point of the model schedule.
	Always allow the next instruction in the model schedule to be added.
	(debug_ready_list): Print the INSN_REG_PRESSURE_EXCESS_COST_CHANGE
	and delay for SCHED_PRESSURE_MODEL too.
	(prune_ready_list): Extend SCHED_PRIORITY_WEIGHTED handling to
	SCHED_PRIORITY_MODEL, but also take the DFA into account.
	(schedule_block): Call model_start_schedule and model_end_schedule.
	Extend SCHED_PRIORITY_WEIGHTED stall handling to SCHED_PRIORITY_MODEL.
	(sched_init): Extend INSN_REG_PRESSURE_EXCESS_COST_CHANGE handling
	to SCHED_PRESSURE_MODEL, but don't allocate saved_reg_live or
	region_ref_regs.
	(sched_finish): Update accordingly.
	(fix_tick_ready): Extend INSN_REG_PRESSURE_EXCESS_COST_CHANGE handling
	to SCHED_PRESSURE_MODEL.
	(add_jump_dependencies): Update call to dep_list_size.
	(haifa_finish_h_i_d): Fix leak of max_reg_pressure.
	(haifa_init_insn): Extend INSN_REG_PRESSURE_EXCESS_COST_CHANGE handling
	to SCHED_PRESSURE_MODEL.
	* sched-deps.c (init_insn_reg_pressure_info): Likewise, but don't
	allocate INSN_MAX_REG_PRESSURE for SCHED_PRESSURE_MODEL.
	(sched_analyze_insn): Extend INSN_REG_PRESSURE_EXCESS_COST_CHANGE
	handling to SCHED_PRESSURE_MODEL.

From-SVN: r186882
parent 60867e8c
2012-04-26 Richard Sandiford <richard.sandiford@linaro.org> 2012-04-26 Richard Sandiford <richard.sandiford@linaro.org>
* sched-int.h (_haifa_insn_data): Move priority_status.
Add model_index.
(INSN_MODEL_INDEX): New macro.
* haifa-sched.c (insn_delay): New function.
(sched_regno_pressure_class): Update commentary.
(mark_regno_birth_or_death): Pass the liveness bitmap and
pressure array as arguments, instead of using curr_reg_live and
curr_reg_pressure. Only update the pressure if the bit in the
liveness set has changed.
(initiate_reg_pressure_info): Always trust the live-in set for
SCHED_PRESSURE_MODEL.
(initiate_bb_reg_pressure_info): Update call to
mark_regno_birth_or_death.
(dep_list_size): Take the list as argument.
(calculate_reg_deaths): New function, extracted from...
(setup_insn_reg_pressure_info): ...here.
(MODEL_BAR): New macro.
(model_pressure_data, model_insn_info, model_pressure_limit)
(model_pressure_group): New structures.
(model_schedule, model_worklist, model_insns, model_num_insns)
(model_curr_point, model_before_pressure, model_next_priority):
New variables.
(MODEL_PRESSURE_DATA, MODEL_MAX_PRESSURE, MODEL_REF_PRESSURE)
(MODEL_INSN_INFO, MODEL_INSN): New macros.
(model_index, model_update_limit_points_in_group): New functions.
(model_update_limit_points, model_last_use_except): Likewise.
(model_start_update_pressure, model_update_pressure): Likewise.
(model_recompute, model_spill_cost, model_excess_group_cost): Likewise.
(model_excess_cost, model_dump_pressure_points): Likewise.
(model_set_excess_costs): Likewise.
(rank_for_schedule): Extend SCHED_PRIORITY_WEIGHTED ordering to
SCHED_PRIORITY_MODEL. Use insn_delay. Use the order in the model
schedule as an alternative tie-breaker. Update the call to
dep_list_size.
(ready_sort): Call model_set_excess_costs.
(update_register_pressure): Update call to mark_regno_birth_or_death.
Rely on that function to check liveness rather than doing it here.
(model_classify_pressure, model_order_p, model_add_to_worklist_at)
(model_remove_from_worklist, model_add_to_worklist, model_promote_insn)
(model_add_to_schedule, model_analyze_insns, model_init_pressure_group)
(model_record_pressure, model_record_pressures): New functions.
(model_record_final_pressures, model_add_successors_to_worklist)
(model_promote_predecessors, model_choose_insn): Likewise.
(model_reset_queue_indices, model_dump_pressure_summary): Likewise.
(model_start_schedule, model_finalize_pressure_group): Likewise.
(model_end_schedule): Likewise.
(schedule_insn): Say when we're scheduling the next instruction
in the model schedule.
(schedule_insn): Handle SCHED_PRESSURE_MODEL.
(queue_to_ready): Do not add instructions that are
MAX_SCHED_READY_INSNS beyond the current point of the model schedule.
Always allow the next instruction in the model schedule to be added.
(debug_ready_list): Print the INSN_REG_PRESSURE_EXCESS_COST_CHANGE
and delay for SCHED_PRESSURE_MODEL too.
(prune_ready_list): Extend SCHED_PRIORITY_WEIGHTED handling to
SCHED_PRIORITY_MODEL, but also take the DFA into account.
(schedule_block): Call model_start_schedule and model_end_schedule.
Extend SCHED_PRIORITY_WEIGHTED stall handling to SCHED_PRIORITY_MODEL.
(sched_init): Extend INSN_REG_PRESSURE_EXCESS_COST_CHANGE handling
to SCHED_PRESSURE_MODEL, but don't allocate saved_reg_live or
region_ref_regs.
(sched_finish): Update accordingly.
(fix_tick_ready): Extend INSN_REG_PRESSURE_EXCESS_COST_CHANGE handling
to SCHED_PRESSURE_MODEL.
(add_jump_dependencies): Update call to dep_list_size.
(haifa_finish_h_i_d): Fix leak of max_reg_pressure.
(haifa_init_insn): Extend INSN_REG_PRESSURE_EXCESS_COST_CHANGE handling
to SCHED_PRESSURE_MODEL.
* sched-deps.c (init_insn_reg_pressure_info): Likewise, but don't
allocate INSN_MAX_REG_PRESSURE for SCHED_PRESSURE_MODEL.
(sched_analyze_insn): Extend INSN_REG_PRESSURE_EXCESS_COST_CHANGE
handling to SCHED_PRESSURE_MODEL.
2012-04-26 Richard Sandiford <richard.sandiford@linaro.org>
* common.opt (fsched-pressure-algorithm=): New option. * common.opt (fsched-pressure-algorithm=): New option.
* flag-types.h (sched_pressure_algorithm): New enum. * flag-types.h (sched_pressure_algorithm): New enum.
* sched-int.h (sched_pressure_p): Replace with... * sched-int.h (sched_pressure_p): Replace with...
......
...@@ -398,6 +398,14 @@ basic_block (* sched_split_block) (basic_block, rtx); ...@@ -398,6 +398,14 @@ basic_block (* sched_split_block) (basic_block, rtx);
/* Create empty basic block after the specified block. */ /* Create empty basic block after the specified block. */
basic_block (* sched_create_empty_bb) (basic_block); basic_block (* sched_create_empty_bb) (basic_block);
/* Return the number of cycles until INSN is expected to be ready.
Return zero if it already is. */
static int
insn_delay (rtx insn)
{
return MAX (INSN_TICK (insn) - clock_var, 0);
}
static int static int
may_trap_exp (const_rtx x, int is_store) may_trap_exp (const_rtx x, int is_store)
{ {
...@@ -875,7 +883,7 @@ schedule_insns (void) ...@@ -875,7 +883,7 @@ schedule_insns (void)
enum sched_pressure_algorithm sched_pressure; enum sched_pressure_algorithm sched_pressure;
/* Map regno -> its pressure class. The map defined only when /* Map regno -> its pressure class. The map defined only when
SCHED_PRESSURE is SCHED_PRESSURE_WEIGHTED. */ SCHED_PRESSURE != SCHED_PRESSURE_NONE. */
enum reg_class *sched_regno_pressure_class; enum reg_class *sched_regno_pressure_class;
/* The current register pressure. Only elements corresponding pressure /* The current register pressure. Only elements corresponding pressure
...@@ -903,10 +911,12 @@ sched_init_region_reg_pressure_info (void) ...@@ -903,10 +911,12 @@ sched_init_region_reg_pressure_info (void)
bitmap_clear (region_ref_regs); bitmap_clear (region_ref_regs);
} }
/* Update current register pressure related info after birth (if /* PRESSURE[CL] describes the pressure on register class CL. Update it
BIRTH_P) or death of register REGNO. */ for the birth (if BIRTH_P) or death (if !BIRTH_P) of register REGNO.
static void LIVE tracks the set of live registers; if it is null, assume that
mark_regno_birth_or_death (int regno, bool birth_p) every birth or death is genuine. */
static inline void
mark_regno_birth_or_death (bitmap live, int *pressure, int regno, bool birth_p)
{ {
enum reg_class pressure_class; enum reg_class pressure_class;
...@@ -917,15 +927,15 @@ mark_regno_birth_or_death (int regno, bool birth_p) ...@@ -917,15 +927,15 @@ mark_regno_birth_or_death (int regno, bool birth_p)
{ {
if (birth_p) if (birth_p)
{ {
bitmap_set_bit (curr_reg_live, regno); if (!live || bitmap_set_bit (live, regno))
curr_reg_pressure[pressure_class] pressure[pressure_class]
+= (ira_reg_class_max_nregs += (ira_reg_class_max_nregs
[pressure_class][PSEUDO_REGNO_MODE (regno)]); [pressure_class][PSEUDO_REGNO_MODE (regno)]);
} }
else else
{ {
bitmap_clear_bit (curr_reg_live, regno); if (!live || bitmap_clear_bit (live, regno))
curr_reg_pressure[pressure_class] pressure[pressure_class]
-= (ira_reg_class_max_nregs -= (ira_reg_class_max_nregs
[pressure_class][PSEUDO_REGNO_MODE (regno)]); [pressure_class][PSEUDO_REGNO_MODE (regno)]);
} }
...@@ -936,13 +946,13 @@ mark_regno_birth_or_death (int regno, bool birth_p) ...@@ -936,13 +946,13 @@ mark_regno_birth_or_death (int regno, bool birth_p)
{ {
if (birth_p) if (birth_p)
{ {
bitmap_set_bit (curr_reg_live, regno); if (!live || bitmap_set_bit (live, regno))
curr_reg_pressure[pressure_class]++; pressure[pressure_class]++;
} }
else else
{ {
bitmap_clear_bit (curr_reg_live, regno); if (!live || bitmap_clear_bit (live, regno))
curr_reg_pressure[pressure_class]--; pressure[pressure_class]--;
} }
} }
} }
...@@ -960,8 +970,10 @@ initiate_reg_pressure_info (bitmap live) ...@@ -960,8 +970,10 @@ initiate_reg_pressure_info (bitmap live)
curr_reg_pressure[ira_pressure_classes[i]] = 0; curr_reg_pressure[ira_pressure_classes[i]] = 0;
bitmap_clear (curr_reg_live); bitmap_clear (curr_reg_live);
EXECUTE_IF_SET_IN_BITMAP (live, 0, j, bi) EXECUTE_IF_SET_IN_BITMAP (live, 0, j, bi)
if (current_nr_blocks == 1 || bitmap_bit_p (region_ref_regs, j)) if (sched_pressure == SCHED_PRESSURE_MODEL
mark_regno_birth_or_death (j, true); || current_nr_blocks == 1
|| bitmap_bit_p (region_ref_regs, j))
mark_regno_birth_or_death (curr_reg_live, curr_reg_pressure, j, true);
} }
/* Mark registers in X as mentioned in the current region. */ /* Mark registers in X as mentioned in the current region. */
...@@ -1015,7 +1027,8 @@ initiate_bb_reg_pressure_info (basic_block bb) ...@@ -1015,7 +1027,8 @@ initiate_bb_reg_pressure_info (basic_block bb)
if (regno == INVALID_REGNUM) if (regno == INVALID_REGNUM)
break; break;
if (! bitmap_bit_p (df_get_live_in (bb), regno)) if (! bitmap_bit_p (df_get_live_in (bb), regno))
mark_regno_birth_or_death (regno, true); mark_regno_birth_or_death (curr_reg_live, curr_reg_pressure,
regno, true);
} }
#endif #endif
} }
...@@ -1445,19 +1458,19 @@ contributes_to_priority_p (dep_t dep) ...@@ -1445,19 +1458,19 @@ contributes_to_priority_p (dep_t dep)
return true; return true;
} }
/* Compute the number of nondebug forward deps of an insn. */ /* Compute the number of nondebug deps in list LIST for INSN. */
static int static int
dep_list_size (rtx insn) dep_list_size (rtx insn, sd_list_types_def list)
{ {
sd_iterator_def sd_it; sd_iterator_def sd_it;
dep_t dep; dep_t dep;
int dbgcount = 0, nodbgcount = 0; int dbgcount = 0, nodbgcount = 0;
if (!MAY_HAVE_DEBUG_INSNS) if (!MAY_HAVE_DEBUG_INSNS)
return sd_lists_size (insn, SD_LIST_FORW); return sd_lists_size (insn, list);
FOR_EACH_DEP (insn, SD_LIST_FORW, sd_it, dep) FOR_EACH_DEP (insn, list, sd_it, dep)
{ {
if (DEBUG_INSN_P (DEP_CON (dep))) if (DEBUG_INSN_P (DEP_CON (dep)))
dbgcount++; dbgcount++;
...@@ -1465,7 +1478,7 @@ dep_list_size (rtx insn) ...@@ -1465,7 +1478,7 @@ dep_list_size (rtx insn)
nodbgcount++; nodbgcount++;
} }
gcc_assert (dbgcount + nodbgcount == sd_lists_size (insn, SD_LIST_FORW)); gcc_assert (dbgcount + nodbgcount == sd_lists_size (insn, list));
return nodbgcount; return nodbgcount;
} }
...@@ -1484,7 +1497,7 @@ priority (rtx insn) ...@@ -1484,7 +1497,7 @@ priority (rtx insn)
{ {
int this_priority = -1; int this_priority = -1;
if (dep_list_size (insn) == 0) if (dep_list_size (insn, SD_LIST_FORW) == 0)
/* ??? We should set INSN_PRIORITY to insn_cost when and insn has /* ??? We should set INSN_PRIORITY to insn_cost when and insn has
some forward deps but all of them are ignored by some forward deps but all of them are ignored by
contributes_to_priority hook. At the moment we set priority of contributes_to_priority hook. At the moment we set priority of
...@@ -1580,6 +1593,22 @@ do { if ((N_READY) == 2) \ ...@@ -1580,6 +1593,22 @@ do { if ((N_READY) == 2) \
qsort (READY, N_READY, sizeof (rtx), rank_for_schedule); } \ qsort (READY, N_READY, sizeof (rtx), rank_for_schedule); } \
while (0) while (0)
/* For each pressure class CL, set DEATH[CL] to the number of registers
in that class that die in INSN. */
static void
calculate_reg_deaths (rtx insn, int *death)
{
int i;
struct reg_use_data *use;
for (i = 0; i < ira_pressure_classes_num; i++)
death[ira_pressure_classes[i]] = 0;
for (use = INSN_REG_USE_LIST (insn); use != NULL; use = use->next_insn_use)
if (dying_use_p (use))
mark_regno_birth_or_death (0, death, use->regno, true);
}
/* Setup info about the current register pressure impact of scheduling /* Setup info about the current register pressure impact of scheduling
INSN at the current scheduling point. */ INSN at the current scheduling point. */
static void static void
...@@ -1591,24 +1620,12 @@ setup_insn_reg_pressure_info (rtx insn) ...@@ -1591,24 +1620,12 @@ setup_insn_reg_pressure_info (rtx insn)
enum reg_class cl; enum reg_class cl;
struct reg_pressure_data *pressure_info; struct reg_pressure_data *pressure_info;
int *max_reg_pressure; int *max_reg_pressure;
struct reg_use_data *use;
static int death[N_REG_CLASSES]; static int death[N_REG_CLASSES];
gcc_checking_assert (!DEBUG_INSN_P (insn)); gcc_checking_assert (!DEBUG_INSN_P (insn));
excess_cost_change = 0; excess_cost_change = 0;
for (i = 0; i < ira_pressure_classes_num; i++) calculate_reg_deaths (insn, death);
death[ira_pressure_classes[i]] = 0;
for (use = INSN_REG_USE_LIST (insn); use != NULL; use = use->next_insn_use)
if (dying_use_p (use))
{
cl = sched_regno_pressure_class[use->regno];
if (use->regno < FIRST_PSEUDO_REGISTER)
death[cl]++;
else
death[cl]
+= ira_reg_class_max_nregs[cl][PSEUDO_REGNO_MODE (use->regno)];
}
pressure_info = INSN_REG_PRESSURE (insn); pressure_info = INSN_REG_PRESSURE (insn);
max_reg_pressure = INSN_MAX_REG_PRESSURE (insn); max_reg_pressure = INSN_MAX_REG_PRESSURE (insn);
gcc_assert (pressure_info != NULL && max_reg_pressure != NULL); gcc_assert (pressure_info != NULL && max_reg_pressure != NULL);
...@@ -1630,272 +1647,1038 @@ setup_insn_reg_pressure_info (rtx insn) ...@@ -1630,272 +1647,1038 @@ setup_insn_reg_pressure_info (rtx insn)
INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insn) = excess_cost_change; INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insn) = excess_cost_change;
} }
/* Returns a positive value if x is preferred; returns a negative value if /* This is the first page of code related to SCHED_PRESSURE_MODEL.
y is preferred. Should never return 0, since that will make the sort It tries to make the scheduler take register pressure into account
unstable. */ without introducing too many unnecessary stalls. It hooks into the
main scheduling algorithm at several points:
static int - Before scheduling starts, model_start_schedule constructs a
rank_for_schedule (const void *x, const void *y) "model schedule" for the current block. This model schedule is
{ chosen solely to keep register pressure down. It does not take the
rtx tmp = *(const rtx *) y; target's pipeline or the original instruction order into account,
rtx tmp2 = *(const rtx *) x; except as a tie-breaker. It also doesn't work to a particular
int tmp_class, tmp2_class; pressure limit.
int val, priority_val, info_val;
if (MAY_HAVE_DEBUG_INSNS) This model schedule gives us an idea of what pressure can be
{ achieved for the block and gives us an example of a schedule that
/* Schedule debug insns as early as possible. */ keeps to that pressure. It also makes the final schedule less
if (DEBUG_INSN_P (tmp) && !DEBUG_INSN_P (tmp2)) dependent on the original instruction order. This is important
return -1; because the original order can either be "wide" (many values live
else if (!DEBUG_INSN_P (tmp) && DEBUG_INSN_P (tmp2)) at once, such as in user-scheduled code) or "narrow" (few values
return 1; live at once, such as after loop unrolling, where several
else if (DEBUG_INSN_P (tmp) && DEBUG_INSN_P (tmp2)) iterations are executed sequentially).
return INSN_LUID (tmp) - INSN_LUID (tmp2);
}
/* The insn in a schedule group should be issued the first. */ We do not apply this model schedule to the rtx stream. We simply
if (flag_sched_group_heuristic && record it in model_schedule. We also compute the maximum pressure,
SCHED_GROUP_P (tmp) != SCHED_GROUP_P (tmp2)) MP, that was seen during this schedule.
return SCHED_GROUP_P (tmp2) ? 1 : -1;
/* Make sure that priority of TMP and TMP2 are initialized. */ - Instructions are added to the ready queue even if they require
gcc_assert (INSN_PRIORITY_KNOWN (tmp) && INSN_PRIORITY_KNOWN (tmp2)); a stall. The length of the stall is instead computed as:
if (sched_pressure == SCHED_PRESSURE_WEIGHTED) MAX (INSN_TICK (INSN) - clock_var, 0)
{
int diff;
/* Prefer insn whose scheduling results in the smallest register (= insn_delay). This allows rank_for_schedule to choose between
pressure excess. */ introducing a deliberate stall or increasing pressure.
if ((diff = (INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp)
+ (INSN_TICK (tmp) > clock_var
? INSN_TICK (tmp) - clock_var : 0)
- INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp2)
- (INSN_TICK (tmp2) > clock_var
? INSN_TICK (tmp2) - clock_var : 0))) != 0)
return diff;
}
- Before sorting the ready queue, model_set_excess_costs assigns
a pressure-based cost to each ready instruction in the queue.
This is the instruction's INSN_REG_PRESSURE_EXCESS_COST_CHANGE
(ECC for short) and is effectively measured in cycles.
if (sched_pressure == SCHED_PRESSURE_WEIGHTED - rank_for_schedule ranks instructions based on:
&& (INSN_TICK (tmp2) > clock_var || INSN_TICK (tmp) > clock_var))
{
if (INSN_TICK (tmp) <= clock_var)
return -1;
else if (INSN_TICK (tmp2) <= clock_var)
return 1;
else
return INSN_TICK (tmp) - INSN_TICK (tmp2);
}
/* If we are doing backtracking in this schedule, prefer insns that ECC (insn) + insn_delay (insn)
have forward dependencies with negative cost against an insn that
was already scheduled. */
if (current_sched_info->flags & DO_BACKTRACKING)
{
priority_val = FEEDS_BACKTRACK_INSN (tmp2) - FEEDS_BACKTRACK_INSN (tmp);
if (priority_val)
return priority_val;
}
/* Prefer insn with higher priority. */ then as:
priority_val = INSN_PRIORITY (tmp2) - INSN_PRIORITY (tmp);
if (flag_sched_critical_path_heuristic && priority_val) insn_delay (insn)
return priority_val;
/* Prefer speculative insn with greater dependencies weakness. */ So, for example, an instruction X1 with an ECC of 1 that can issue
if (flag_sched_spec_insn_heuristic && spec_info) now will win over an instruction X0 with an ECC of zero that would
{ introduce a stall of one cycle. However, an instruction X2 with an
ds_t ds1, ds2; ECC of 2 that can issue now will lose to both X0 and X1.
dw_t dw1, dw2;
int dw;
ds1 = TODO_SPEC (tmp) & SPECULATIVE; - When an instruction is scheduled, model_recompute updates the model
if (ds1) schedule with the new pressures (some of which might now exceed the
dw1 = ds_weak (ds1); original maximum pressure MP). model_update_limit_points then searches
else for the new point of maximum pressure, if not already known. */
dw1 = NO_DEP_WEAK;
ds2 = TODO_SPEC (tmp2) & SPECULATIVE; /* Used to separate high-verbosity debug information for SCHED_PRESSURE_MODEL
if (ds2) from surrounding debug information. */
dw2 = ds_weak (ds2); #define MODEL_BAR \
else ";;\t\t+------------------------------------------------------\n"
dw2 = NO_DEP_WEAK;
dw = dw2 - dw1; /* Information about the pressure on a particular register class at a
if (dw > (NO_DEP_WEAK / 8) || dw < -(NO_DEP_WEAK / 8)) particular point of the model schedule. */
return dw; struct model_pressure_data {
} /* The pressure at this point of the model schedule, or -1 if the
point is associated with an instruction that has already been
scheduled. */
int ref_pressure;
info_val = (*current_sched_info->rank) (tmp, tmp2); /* The maximum pressure during or after this point of the model schedule. */
if(flag_sched_rank_heuristic && info_val) int max_pressure;
return info_val; };
/* Compare insns based on their relation to the last scheduled /* Per-instruction information that is used while building the model
non-debug insn. */ schedule. Here, "schedule" refers to the model schedule rather
if (flag_sched_last_insn_heuristic && last_nondebug_scheduled_insn) than the main schedule. */
{ struct model_insn_info {
dep_t dep1; /* The instruction itself. */
dep_t dep2; rtx insn;
rtx last = last_nondebug_scheduled_insn;
/* Classify the instructions into three classes: /* If this instruction is in model_worklist, these fields link to the
1) Data dependent on last schedule insn. previous (higher-priority) and next (lower-priority) instructions
2) Anti/Output dependent on last scheduled insn. in the list. */
3) Independent of last scheduled insn, or has latency of one. struct model_insn_info *prev;
Choose the insn from the highest numbered class if different. */ struct model_insn_info *next;
dep1 = sd_find_dep_between (last, tmp, true);
/* While constructing the schedule, QUEUE_INDEX describes whether an
instruction has already been added to the schedule (QUEUE_SCHEDULED),
is in model_worklist (QUEUE_READY), or neither (QUEUE_NOWHERE).
old_queue records the value that QUEUE_INDEX had before scheduling
started, so that we can restore it once the schedule is complete. */
int old_queue;
/* The relative importance of an unscheduled instruction. Higher
values indicate greater importance. */
unsigned int model_priority;
/* The length of the longest path of satisfied true dependencies
that leads to this instruction. */
unsigned int depth;
/* The length of the longest path of dependencies of any kind
that leads from this instruction. */
unsigned int alap;
/* The number of predecessor nodes that must still be scheduled. */
int unscheduled_preds;
};
if (dep1 == NULL || dep_cost (dep1) == 1) /* Information about the pressure limit for a particular register class.
tmp_class = 3; This structure is used when applying a model schedule to the main
else if (/* Data dependence. */ schedule. */
DEP_TYPE (dep1) == REG_DEP_TRUE) struct model_pressure_limit {
tmp_class = 1; /* The maximum register pressure seen in the original model schedule. */
else int orig_pressure;
tmp_class = 2;
dep2 = sd_find_dep_between (last, tmp2, true); /* The maximum register pressure seen in the current model schedule
(which excludes instructions that have already been scheduled). */
int pressure;
if (dep2 == NULL || dep_cost (dep2) == 1) /* The point of the current model schedule at which PRESSURE is first
tmp2_class = 3; reached. It is set to -1 if the value needs to be recomputed. */
else if (/* Data dependence. */ int point;
DEP_TYPE (dep2) == REG_DEP_TRUE) };
tmp2_class = 1;
else
tmp2_class = 2;
if ((val = tmp2_class - tmp_class)) /* Describes a particular way of measuring register pressure. */
return val; struct model_pressure_group {
} /* Index PCI describes the maximum pressure on ira_pressure_classes[PCI]. */
struct model_pressure_limit limits[N_REG_CLASSES];
/* Prefer the insn which has more later insns that depend on it. /* Index (POINT * ira_num_pressure_classes + PCI) describes the pressure
This gives the scheduler more freedom when scheduling later on register class ira_pressure_classes[PCI] at point POINT of the
instructions at the expense of added register pressure. */ current model schedule. A POINT of model_num_insns describes the
pressure at the end of the schedule. */
struct model_pressure_data *model;
};
val = (dep_list_size (tmp2) - dep_list_size (tmp)); /* Index POINT gives the instruction at point POINT of the model schedule.
This array doesn't change during main scheduling. */
static VEC (rtx, heap) *model_schedule;
if (flag_sched_dep_count_heuristic && val != 0) /* The list of instructions in the model worklist, sorted in order of
return val; decreasing priority. */
static struct model_insn_info *model_worklist;
/* If insns are equally good, sort by INSN_LUID (original insn order), /* Index I describes the instruction with INSN_LUID I. */
so that we make the sort stable. This minimizes instruction movement, static struct model_insn_info *model_insns;
thus minimizing sched's effect on debugging and cross-jumping. */
return INSN_LUID (tmp) - INSN_LUID (tmp2);
}
/* Resort the array A in which only element at index N may be out of order. */ /* The number of instructions in the model schedule. */
static int model_num_insns;
HAIFA_INLINE static void /* The index of the first instruction in model_schedule that hasn't yet been
swap_sort (rtx *a, int n) added to the main schedule, or model_num_insns if all of them have. */
{ static int model_curr_point;
rtx insn = a[n - 1];
int i = n - 2;
while (i >= 0 && rank_for_schedule (a + i, &insn) >= 0) /* Describes the pressure before each instruction in the model schedule. */
{ static struct model_pressure_group model_before_pressure;
a[i + 1] = a[i];
i -= 1;
}
a[i + 1] = insn;
}
/* Add INSN to the insn queue so that it can be executed at least /* The first unused model_priority value (as used in model_insn_info). */
N_CYCLES after the currently executing insn. Preserve insns static unsigned int model_next_priority;
chain for debugging purposes. REASON will be printed in debugging
output. */
HAIFA_INLINE static void
queue_insn (rtx insn, int n_cycles, const char *reason) /* The model_pressure_data for ira_pressure_classes[PCI] in GROUP
at point POINT of the model schedule. */
#define MODEL_PRESSURE_DATA(GROUP, POINT, PCI) \
(&(GROUP)->model[(POINT) * ira_pressure_classes_num + (PCI)])
/* The maximum pressure on ira_pressure_classes[PCI] in GROUP at or
after point POINT of the model schedule. */
#define MODEL_MAX_PRESSURE(GROUP, POINT, PCI) \
(MODEL_PRESSURE_DATA (GROUP, POINT, PCI)->max_pressure)
/* The pressure on ira_pressure_classes[PCI] in GROUP at point POINT
of the model schedule. */
#define MODEL_REF_PRESSURE(GROUP, POINT, PCI) \
(MODEL_PRESSURE_DATA (GROUP, POINT, PCI)->ref_pressure)
/* Information about INSN that is used when creating the model schedule. */
#define MODEL_INSN_INFO(INSN) \
(&model_insns[INSN_LUID (INSN)])
/* The instruction at point POINT of the model schedule. */
#define MODEL_INSN(POINT) \
(VEC_index (rtx, model_schedule, POINT))
/* Return INSN's index in the model schedule, or model_num_insns if it
doesn't belong to that schedule. */
static int
model_index (rtx insn)
{ {
int next_q = NEXT_Q_AFTER (q_ptr, n_cycles); if (INSN_MODEL_INDEX (insn) == 0)
rtx link = alloc_INSN_LIST (insn, insn_queue[next_q]); return model_num_insns;
int new_tick; return INSN_MODEL_INDEX (insn) - 1;
}
gcc_assert (n_cycles <= max_insn_queue_index); /* Make sure that GROUP->limits is up-to-date for the current point
gcc_assert (!DEBUG_INSN_P (insn)); of the model schedule. */
insn_queue[next_q] = link; static void
q_size += 1; model_update_limit_points_in_group (struct model_pressure_group *group)
{
int pci, max_pressure, point;
if (sched_verbose >= 2) for (pci = 0; pci < ira_pressure_classes_num; pci++)
{ {
fprintf (sched_dump, ";;\t\tReady-->Q: insn %s: ", /* We may have passed the final point at which the pressure in
(*current_sched_info->print_insn) (insn, 0)); group->limits[pci].pressure was reached. Update the limit if so. */
max_pressure = MODEL_MAX_PRESSURE (group, model_curr_point, pci);
group->limits[pci].pressure = max_pressure;
fprintf (sched_dump, "queued for %d cycles (%s).\n", n_cycles, reason); /* Find the point at which MAX_PRESSURE is first reached. We need
} to search in three cases:
QUEUE_INDEX (insn) = next_q; - We've already moved past the previous pressure point.
In this case we search forward from model_curr_point.
if (current_sched_info->flags & DO_BACKTRACKING) - We scheduled the previous point of maximum pressure ahead of
{ its position in the model schedule, but doing so didn't bring
new_tick = clock_var + n_cycles; the pressure point earlier. In this case we search forward
if (INSN_TICK (insn) == INVALID_TICK || INSN_TICK (insn) < new_tick) from that previous pressure point.
INSN_TICK (insn) = new_tick;
if (INSN_EXACT_TICK (insn) != INVALID_TICK - Scheduling an instruction early caused the maximum pressure
&& INSN_EXACT_TICK (insn) < clock_var + n_cycles) to decrease. In this case we will have set the pressure
{ point to -1, and we search forward from model_curr_point. */
must_backtrack = true; point = MAX (group->limits[pci].point, model_curr_point);
if (sched_verbose >= 2) while (point < model_num_insns
fprintf (sched_dump, ";;\t\tcausing a backtrack.\n"); && MODEL_REF_PRESSURE (group, point, pci) < max_pressure)
} point++;
group->limits[pci].point = point;
gcc_assert (MODEL_REF_PRESSURE (group, point, pci) == max_pressure);
gcc_assert (MODEL_MAX_PRESSURE (group, point, pci) == max_pressure);
} }
} }
/* Remove INSN from queue. */ /* Make sure that all register-pressure limits are up-to-date for the
current position in the model schedule. */
static void static void
queue_remove (rtx insn) model_update_limit_points (void)
{ {
gcc_assert (QUEUE_INDEX (insn) >= 0); model_update_limit_points_in_group (&model_before_pressure);
remove_free_INSN_LIST_elem (insn, &insn_queue[QUEUE_INDEX (insn)]);
q_size--;
QUEUE_INDEX (insn) = QUEUE_NOWHERE;
} }
/* Return a pointer to the bottom of the ready list, i.e. the insn /* Return the model_index of the last unscheduled use in chain USE
with the lowest priority. */ outside of USE's instruction. Return -1 if there are no other uses,
or model_num_insns if the register is live at the end of the block. */
rtx * static int
ready_lastpos (struct ready_list *ready) model_last_use_except (struct reg_use_data *use)
{ {
gcc_assert (ready->n_ready >= 1); struct reg_use_data *next;
return ready->vec + ready->first - ready->n_ready + 1; int last, index;
last = -1;
for (next = use->next_regno_use; next != use; next = next->next_regno_use)
if (NONDEBUG_INSN_P (next->insn)
&& QUEUE_INDEX (next->insn) != QUEUE_SCHEDULED)
{
index = model_index (next->insn);
if (index == model_num_insns)
return model_num_insns;
if (last < index)
last = index;
}
return last;
} }
/* Add an element INSN to the ready list so that it ends up with the /* An instruction with model_index POINT has just been scheduled, and it
lowest/highest priority depending on FIRST_P. */ adds DELTA to the pressure on ira_pressure_classes[PCI] after POINT - 1.
Update MODEL_REF_PRESSURE (GROUP, POINT, PCI) and
MODEL_MAX_PRESSURE (GROUP, POINT, PCI) accordingly. */
HAIFA_INLINE static void static void
ready_add (struct ready_list *ready, rtx insn, bool first_p) model_start_update_pressure (struct model_pressure_group *group,
int point, int pci, int delta)
{ {
if (!first_p) int next_max_pressure;
{
if (ready->first == ready->n_ready) if (point == model_num_insns)
{ {
memmove (ready->vec + ready->veclen - ready->n_ready, /* The instruction wasn't part of the model schedule; it was moved
ready_lastpos (ready), from a different block. Update the pressure for the end of
ready->n_ready * sizeof (rtx)); the model schedule. */
ready->first = ready->veclen - 1; MODEL_REF_PRESSURE (group, point, pci) += delta;
} MODEL_MAX_PRESSURE (group, point, pci) += delta;
ready->vec[ready->first - ready->n_ready] = insn;
} }
else else
{ {
if (ready->first == ready->veclen - 1) /* Record that this instruction has been scheduled. Nothing now
changes between POINT and POINT + 1, so get the maximum pressure
from the latter. If the maximum pressure decreases, the new
pressure point may be before POINT. */
MODEL_REF_PRESSURE (group, point, pci) = -1;
next_max_pressure = MODEL_MAX_PRESSURE (group, point + 1, pci);
if (MODEL_MAX_PRESSURE (group, point, pci) > next_max_pressure)
{ {
if (ready->n_ready) MODEL_MAX_PRESSURE (group, point, pci) = next_max_pressure;
/* ready_lastpos() fails when called with (ready->n_ready == 0). */ if (group->limits[pci].point == point)
memmove (ready->vec + ready->veclen - ready->n_ready - 1, group->limits[pci].point = -1;
ready_lastpos (ready),
ready->n_ready * sizeof (rtx));
ready->first = ready->veclen - 2;
} }
ready->vec[++(ready->first)] = insn;
} }
}
ready->n_ready++; /* Record that scheduling a later instruction has changed the pressure
at point POINT of the model schedule by DELTA (which might be 0).
Update GROUP accordingly. Return nonzero if these changes might
trigger changes to previous points as well. */
static int
model_update_pressure (struct model_pressure_group *group,
int point, int pci, int delta)
{
int ref_pressure, max_pressure, next_max_pressure;
/* If POINT hasn't yet been scheduled, update its pressure. */
ref_pressure = MODEL_REF_PRESSURE (group, point, pci);
if (ref_pressure >= 0 && delta != 0)
{
ref_pressure += delta;
MODEL_REF_PRESSURE (group, point, pci) = ref_pressure;
/* Check whether the maximum pressure in the overall schedule
has increased. (This means that the MODEL_MAX_PRESSURE of
every point <= POINT will need to increae too; see below.) */
if (group->limits[pci].pressure < ref_pressure)
group->limits[pci].pressure = ref_pressure;
/* If we are at maximum pressure, and the maximum pressure
point was previously unknown or later than POINT,
bring it forward. */
if (group->limits[pci].pressure == ref_pressure
&& !IN_RANGE (group->limits[pci].point, 0, point))
group->limits[pci].point = point;
/* If POINT used to be the point of maximum pressure, but isn't
any longer, we need to recalculate it using a forward walk. */
if (group->limits[pci].pressure > ref_pressure
&& group->limits[pci].point == point)
group->limits[pci].point = -1;
}
/* Update the maximum pressure at POINT. Changes here might also
affect the maximum pressure at POINT - 1. */
next_max_pressure = MODEL_MAX_PRESSURE (group, point + 1, pci);
max_pressure = MAX (ref_pressure, next_max_pressure);
if (MODEL_MAX_PRESSURE (group, point, pci) != max_pressure)
{
MODEL_MAX_PRESSURE (group, point, pci) = max_pressure;
return 1;
}
return 0;
}
/* INSN has just been scheduled. Update the model schedule accordingly. */
static void
model_recompute (rtx insn)
{
struct {
int last_use;
int regno;
} uses[FIRST_PSEUDO_REGISTER + MAX_RECOG_OPERANDS];
struct reg_use_data *use;
struct reg_pressure_data *reg_pressure;
int delta[N_REG_CLASSES];
int pci, point, mix, new_last, cl, ref_pressure, queue;
unsigned int i, num_uses, num_pending_births;
bool print_p;
/* The destinations of INSN were previously live from POINT onwards, but are
now live from model_curr_point onwards. Set up DELTA accordingly. */
point = model_index (insn);
reg_pressure = INSN_REG_PRESSURE (insn);
for (pci = 0; pci < ira_pressure_classes_num; pci++)
{
cl = ira_pressure_classes[pci];
delta[cl] = reg_pressure[pci].set_increase;
}
/* Record which registers previously died at POINT, but which now die
before POINT. Adjust DELTA so that it represents the effect of
this change after POINT - 1. Set NUM_PENDING_BIRTHS to the number of
registers that will be born in the range [model_curr_point, POINT). */
num_uses = 0;
num_pending_births = 0;
for (use = INSN_REG_USE_LIST (insn); use != NULL; use = use->next_insn_use)
{
new_last = model_last_use_except (use);
if (new_last < point)
{
gcc_assert (num_uses < ARRAY_SIZE (uses));
uses[num_uses].last_use = new_last;
uses[num_uses].regno = use->regno;
/* This register is no longer live after POINT - 1. */
mark_regno_birth_or_death (NULL, delta, use->regno, false);
num_uses++;
if (new_last >= 0)
num_pending_births++;
}
}
/* Update the MODEL_REF_PRESSURE and MODEL_MAX_PRESSURE for POINT.
Also set each group pressure limit for POINT. */
for (pci = 0; pci < ira_pressure_classes_num; pci++)
{
cl = ira_pressure_classes[pci];
model_start_update_pressure (&model_before_pressure,
point, pci, delta[cl]);
}
/* Walk the model schedule backwards, starting immediately before POINT. */
print_p = false;
if (point != model_curr_point)
do
{
point--;
insn = MODEL_INSN (point);
queue = QUEUE_INDEX (insn);
if (queue != QUEUE_SCHEDULED)
{
/* DELTA describes the effect of the move on the register pressure
after POINT. Make it describe the effect on the pressure
before POINT. */
i = 0;
while (i < num_uses)
{
if (uses[i].last_use == point)
{
/* This register is now live again. */
mark_regno_birth_or_death (NULL, delta,
uses[i].regno, true);
/* Remove this use from the array. */
uses[i] = uses[num_uses - 1];
num_uses--;
num_pending_births--;
}
else
i++;
}
if (sched_verbose >= 5)
{
char buf[2048];
if (!print_p)
{
fprintf (sched_dump, MODEL_BAR);
fprintf (sched_dump, ";;\t\t| New pressure for model"
" schedule\n");
fprintf (sched_dump, MODEL_BAR);
print_p = true;
}
print_pattern (buf, PATTERN (insn), 0);
fprintf (sched_dump, ";;\t\t| %3d %4d %-30s ",
point, INSN_UID (insn), buf);
for (pci = 0; pci < ira_pressure_classes_num; pci++)
{
cl = ira_pressure_classes[pci];
ref_pressure = MODEL_REF_PRESSURE (&model_before_pressure,
point, pci);
fprintf (sched_dump, " %s:[%d->%d]",
reg_class_names[ira_pressure_classes[pci]],
ref_pressure, ref_pressure + delta[cl]);
}
fprintf (sched_dump, "\n");
}
}
/* Adjust the pressure at POINT. Set MIX to nonzero if POINT - 1
might have changed as well. */
mix = num_pending_births;
for (pci = 0; pci < ira_pressure_classes_num; pci++)
{
cl = ira_pressure_classes[pci];
mix |= delta[cl];
mix |= model_update_pressure (&model_before_pressure,
point, pci, delta[cl]);
}
}
while (mix && point > model_curr_point);
if (print_p)
fprintf (sched_dump, MODEL_BAR);
}
/* model_spill_cost (CL, P, P') returns the cost of increasing the
pressure on CL from P to P'. We use this to calculate a "base ECC",
baseECC (CL, X), for each pressure class CL and each instruction X.
Supposing X changes the pressure on CL from P to P', and that the
maximum pressure on CL in the current model schedule is MP', then:
* if X occurs before or at the next point of maximum pressure in
the model schedule and P' > MP', then:
baseECC (CL, X) = model_spill_cost (CL, MP, P')
The idea is that the pressure after scheduling a fixed set of
instructions -- in this case, the set up to and including the
next maximum pressure point -- is going to be the same regardless
of the order; we simply want to keep the intermediate pressure
under control. Thus X has a cost of zero unless scheduling it
now would exceed MP'.
If all increases in the set are by the same amount, no zero-cost
instruction will ever cause the pressure to exceed MP'. However,
if X is instead moved past an instruction X' with pressure in the
range (MP' - (P' - P), MP'), the pressure at X' will increase
beyond MP'. Since baseECC is very much a heuristic anyway,
it doesn't seem worth the overhead of tracking cases like these.
The cost of exceeding MP' is always based on the original maximum
pressure MP. This is so that going 2 registers over the original
limit has the same cost regardless of whether it comes from two
separate +1 deltas or from a single +2 delta.
* if X occurs after the next point of maximum pressure in the model
schedule and P' > P, then:
baseECC (CL, X) = model_spill_cost (CL, MP, MP' + (P' - P))
That is, if we move X forward across a point of maximum pressure,
and if X increases the pressure by P' - P, then we conservatively
assume that scheduling X next would increase the maximum pressure
by P' - P. Again, the cost of doing this is based on the original
maximum pressure MP, for the same reason as above.
* if P' < P, P > MP, and X occurs at or after the next point of
maximum pressure, then:
baseECC (CL, X) = -model_spill_cost (CL, MAX (MP, P'), P)
That is, if we have already exceeded the original maximum pressure MP,
and if X might reduce the maximum pressure again -- or at least push
it further back, and thus allow more scheduling freedom -- it is given
a negative cost to reflect the improvement.
* otherwise,
baseECC (CL, X) = 0
In this case, X is not expected to affect the maximum pressure MP',
so it has zero cost.
We then create a combined value baseECC (X) that is the sum of
baseECC (CL, X) for each pressure class CL.
baseECC (X) could itself be used as the ECC value described above.
However, this is often too conservative, in the sense that it
tends to make high-priority instructions that increase pressure
wait too long in cases where introducing a spill would be better.
For this reason the final ECC is a priority-adjusted form of
baseECC (X). Specifically, we calculate:
P (X) = INSN_PRIORITY (X) - insn_delay (X) - baseECC (X)
baseP = MAX { P (X) | baseECC (X) <= 0 }
Then:
ECC (X) = MAX (MIN (baseP - P (X), baseECC (X)), 0)
Thus an instruction's effect on pressure is ignored if it has a high
enough priority relative to the ones that don't increase pressure.
Negative values of baseECC (X) do not increase the priority of X
itself, but they do make it harder for other instructions to
increase the pressure further.
This pressure cost is deliberately timid. The intention has been
to choose a heuristic that rarely interferes with the normal list
scheduler in cases where that scheduler would produce good code.
We simply want to curb some of its worst excesses. */
/* Return the cost of increasing the pressure in class CL from FROM to TO.
Here we use the very simplistic cost model that every register above
ira_available_class_regs[CL] has a spill cost of 1. We could use other
measures instead, such as one based on MEMORY_MOVE_COST. However:
(1) In order for an instruction to be scheduled, the higher cost
would need to be justified in a single saving of that many stalls.
This is overly pessimistic, because the benefit of spilling is
often to avoid a sequence of several short stalls rather than
a single long one.
(2) The cost is still arbitrary. Because we are not allocating
registers during scheduling, we have no way of knowing for
sure how many memory accesses will be required by each spill,
where the spills will be placed within the block, or even
which block(s) will contain the spills.
So a higher cost than 1 is often too conservative in practice,
forcing blocks to contain unnecessary stalls instead of spill code.
The simple cost below seems to be the best compromise. It reduces
the interference with the normal list scheduler, which helps make
it more suitable for a default-on option. */
static int
model_spill_cost (int cl, int from, int to)
{
from = MAX (from, ira_available_class_regs[cl]);
return MAX (to, from) - from;
}
/* Return baseECC (ira_pressure_classes[PCI], POINT), given that
P = curr_reg_pressure[ira_pressure_classes[PCI]] and that
P' = P + DELTA. */
static int
model_excess_group_cost (struct model_pressure_group *group,
int point, int pci, int delta)
{
int pressure, cl;
cl = ira_pressure_classes[pci];
if (delta < 0 && point >= group->limits[pci].point)
{
pressure = MAX (group->limits[pci].orig_pressure,
curr_reg_pressure[cl] + delta);
return -model_spill_cost (cl, pressure, curr_reg_pressure[cl]);
}
if (delta > 0)
{
if (point > group->limits[pci].point)
pressure = group->limits[pci].pressure + delta;
else
pressure = curr_reg_pressure[cl] + delta;
if (pressure > group->limits[pci].pressure)
return model_spill_cost (cl, group->limits[pci].orig_pressure,
pressure);
}
return 0;
}
/* Return baseECC (MODEL_INSN (INSN)). Dump the costs to sched_dump
if PRINT_P. */
static int
model_excess_cost (rtx insn, bool print_p)
{
int point, pci, cl, cost, this_cost, delta;
struct reg_pressure_data *insn_reg_pressure;
int insn_death[N_REG_CLASSES];
calculate_reg_deaths (insn, insn_death);
point = model_index (insn);
insn_reg_pressure = INSN_REG_PRESSURE (insn);
cost = 0;
if (print_p)
fprintf (sched_dump, ";;\t\t| %3d %4d | %4d %+3d |", point,
INSN_UID (insn), INSN_PRIORITY (insn), insn_delay (insn));
/* Sum up the individual costs for each register class. */
for (pci = 0; pci < ira_pressure_classes_num; pci++)
{
cl = ira_pressure_classes[pci];
delta = insn_reg_pressure[pci].set_increase - insn_death[cl];
this_cost = model_excess_group_cost (&model_before_pressure,
point, pci, delta);
cost += this_cost;
if (print_p)
fprintf (sched_dump, " %s:[%d base cost %d]",
reg_class_names[cl], delta, this_cost);
}
if (print_p)
fprintf (sched_dump, "\n");
return cost;
}
/* Dump the next points of maximum pressure for GROUP. */
static void
model_dump_pressure_points (struct model_pressure_group *group)
{
int pci, cl;
fprintf (sched_dump, ";;\t\t| pressure points");
for (pci = 0; pci < ira_pressure_classes_num; pci++)
{
cl = ira_pressure_classes[pci];
fprintf (sched_dump, " %s:[%d->%d at ", reg_class_names[cl],
curr_reg_pressure[cl], group->limits[pci].pressure);
if (group->limits[pci].point < model_num_insns)
fprintf (sched_dump, "%d:%d]", group->limits[pci].point,
INSN_UID (MODEL_INSN (group->limits[pci].point)));
else
fprintf (sched_dump, "end]");
}
fprintf (sched_dump, "\n");
}
/* Set INSN_REG_PRESSURE_EXCESS_COST_CHANGE for INSNS[0...COUNT-1]. */
static void
model_set_excess_costs (rtx *insns, int count)
{
int i, cost, priority_base, priority;
bool print_p;
/* Record the baseECC value for each instruction in the model schedule,
except that negative costs are converted to zero ones now rather thatn
later. Do not assign a cost to debug instructions, since they must
not change code-generation decisions. Experiments suggest we also
get better results by not assigning a cost to instructions from
a different block.
Set PRIORITY_BASE to baseP in the block comment above. This is the
maximum priority of the "cheap" instructions, which should always
include the next model instruction. */
priority_base = 0;
print_p = false;
for (i = 0; i < count; i++)
if (INSN_MODEL_INDEX (insns[i]))
{
if (sched_verbose >= 6 && !print_p)
{
fprintf (sched_dump, MODEL_BAR);
fprintf (sched_dump, ";;\t\t| Pressure costs for ready queue\n");
model_dump_pressure_points (&model_before_pressure);
fprintf (sched_dump, MODEL_BAR);
print_p = true;
}
cost = model_excess_cost (insns[i], print_p);
if (cost <= 0)
{
priority = INSN_PRIORITY (insns[i]) - insn_delay (insns[i]) - cost;
priority_base = MAX (priority_base, priority);
cost = 0;
}
INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insns[i]) = cost;
}
if (print_p)
fprintf (sched_dump, MODEL_BAR);
/* Use MAX (baseECC, 0) and baseP to calculcate ECC for each
instruction. */
for (i = 0; i < count; i++)
{
cost = INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insns[i]);
priority = INSN_PRIORITY (insns[i]) - insn_delay (insns[i]);
if (cost > 0 && priority > priority_base)
{
cost += priority_base - priority;
INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insns[i]) = MAX (cost, 0);
}
}
}
/* Returns a positive value if x is preferred; returns a negative value if
y is preferred. Should never return 0, since that will make the sort
unstable. */
static int
rank_for_schedule (const void *x, const void *y)
{
rtx tmp = *(const rtx *) y;
rtx tmp2 = *(const rtx *) x;
int tmp_class, tmp2_class;
int val, priority_val, info_val;
if (MAY_HAVE_DEBUG_INSNS)
{
/* Schedule debug insns as early as possible. */
if (DEBUG_INSN_P (tmp) && !DEBUG_INSN_P (tmp2))
return -1;
else if (!DEBUG_INSN_P (tmp) && DEBUG_INSN_P (tmp2))
return 1;
else if (DEBUG_INSN_P (tmp) && DEBUG_INSN_P (tmp2))
return INSN_LUID (tmp) - INSN_LUID (tmp2);
}
/* The insn in a schedule group should be issued the first. */
if (flag_sched_group_heuristic &&
SCHED_GROUP_P (tmp) != SCHED_GROUP_P (tmp2))
return SCHED_GROUP_P (tmp2) ? 1 : -1;
/* Make sure that priority of TMP and TMP2 are initialized. */
gcc_assert (INSN_PRIORITY_KNOWN (tmp) && INSN_PRIORITY_KNOWN (tmp2));
if (sched_pressure != SCHED_PRESSURE_NONE)
{
int diff;
/* Prefer insn whose scheduling results in the smallest register
pressure excess. */
if ((diff = (INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp)
+ insn_delay (tmp)
- INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp2)
- insn_delay (tmp2))))
return diff;
}
if (sched_pressure != SCHED_PRESSURE_NONE
&& (INSN_TICK (tmp2) > clock_var || INSN_TICK (tmp) > clock_var))
{
if (INSN_TICK (tmp) <= clock_var)
return -1;
else if (INSN_TICK (tmp2) <= clock_var)
return 1;
else
return INSN_TICK (tmp) - INSN_TICK (tmp2);
}
/* If we are doing backtracking in this schedule, prefer insns that
have forward dependencies with negative cost against an insn that
was already scheduled. */
if (current_sched_info->flags & DO_BACKTRACKING)
{
priority_val = FEEDS_BACKTRACK_INSN (tmp2) - FEEDS_BACKTRACK_INSN (tmp);
if (priority_val)
return priority_val;
}
/* Prefer insn with higher priority. */
priority_val = INSN_PRIORITY (tmp2) - INSN_PRIORITY (tmp);
if (flag_sched_critical_path_heuristic && priority_val)
return priority_val;
/* Prefer speculative insn with greater dependencies weakness. */
if (flag_sched_spec_insn_heuristic && spec_info)
{
ds_t ds1, ds2;
dw_t dw1, dw2;
int dw;
ds1 = TODO_SPEC (tmp) & SPECULATIVE;
if (ds1)
dw1 = ds_weak (ds1);
else
dw1 = NO_DEP_WEAK;
ds2 = TODO_SPEC (tmp2) & SPECULATIVE;
if (ds2)
dw2 = ds_weak (ds2);
else
dw2 = NO_DEP_WEAK;
dw = dw2 - dw1;
if (dw > (NO_DEP_WEAK / 8) || dw < -(NO_DEP_WEAK / 8))
return dw;
}
info_val = (*current_sched_info->rank) (tmp, tmp2);
if(flag_sched_rank_heuristic && info_val)
return info_val;
/* Compare insns based on their relation to the last scheduled
non-debug insn. */
if (flag_sched_last_insn_heuristic && last_nondebug_scheduled_insn)
{
dep_t dep1;
dep_t dep2;
rtx last = last_nondebug_scheduled_insn;
/* Classify the instructions into three classes:
1) Data dependent on last schedule insn.
2) Anti/Output dependent on last scheduled insn.
3) Independent of last scheduled insn, or has latency of one.
Choose the insn from the highest numbered class if different. */
dep1 = sd_find_dep_between (last, tmp, true);
if (dep1 == NULL || dep_cost (dep1) == 1)
tmp_class = 3;
else if (/* Data dependence. */
DEP_TYPE (dep1) == REG_DEP_TRUE)
tmp_class = 1;
else
tmp_class = 2;
dep2 = sd_find_dep_between (last, tmp2, true);
if (dep2 == NULL || dep_cost (dep2) == 1)
tmp2_class = 3;
else if (/* Data dependence. */
DEP_TYPE (dep2) == REG_DEP_TRUE)
tmp2_class = 1;
else
tmp2_class = 2;
if ((val = tmp2_class - tmp_class))
return val;
}
/* Prefer instructions that occur earlier in the model schedule. */
if (sched_pressure == SCHED_PRESSURE_MODEL)
{
int diff;
diff = model_index (tmp) - model_index (tmp2);
if (diff != 0)
return diff;
}
/* Prefer the insn which has more later insns that depend on it.
This gives the scheduler more freedom when scheduling later
instructions at the expense of added register pressure. */
val = (dep_list_size (tmp2, SD_LIST_FORW)
- dep_list_size (tmp, SD_LIST_FORW));
if (flag_sched_dep_count_heuristic && val != 0)
return val;
/* If insns are equally good, sort by INSN_LUID (original insn order),
so that we make the sort stable. This minimizes instruction movement,
thus minimizing sched's effect on debugging and cross-jumping. */
return INSN_LUID (tmp) - INSN_LUID (tmp2);
}
/* Resort the array A in which only element at index N may be out of order. */
HAIFA_INLINE static void
swap_sort (rtx *a, int n)
{
rtx insn = a[n - 1];
int i = n - 2;
while (i >= 0 && rank_for_schedule (a + i, &insn) >= 0)
{
a[i + 1] = a[i];
i -= 1;
}
a[i + 1] = insn;
}
/* Add INSN to the insn queue so that it can be executed at least
N_CYCLES after the currently executing insn. Preserve insns
chain for debugging purposes. REASON will be printed in debugging
output. */
HAIFA_INLINE static void
queue_insn (rtx insn, int n_cycles, const char *reason)
{
int next_q = NEXT_Q_AFTER (q_ptr, n_cycles);
rtx link = alloc_INSN_LIST (insn, insn_queue[next_q]);
int new_tick;
gcc_assert (n_cycles <= max_insn_queue_index);
gcc_assert (!DEBUG_INSN_P (insn));
insn_queue[next_q] = link;
q_size += 1;
if (sched_verbose >= 2)
{
fprintf (sched_dump, ";;\t\tReady-->Q: insn %s: ",
(*current_sched_info->print_insn) (insn, 0));
fprintf (sched_dump, "queued for %d cycles (%s).\n", n_cycles, reason);
}
QUEUE_INDEX (insn) = next_q;
if (current_sched_info->flags & DO_BACKTRACKING)
{
new_tick = clock_var + n_cycles;
if (INSN_TICK (insn) == INVALID_TICK || INSN_TICK (insn) < new_tick)
INSN_TICK (insn) = new_tick;
if (INSN_EXACT_TICK (insn) != INVALID_TICK
&& INSN_EXACT_TICK (insn) < clock_var + n_cycles)
{
must_backtrack = true;
if (sched_verbose >= 2)
fprintf (sched_dump, ";;\t\tcausing a backtrack.\n");
}
}
}
/* Remove INSN from queue. */
static void
queue_remove (rtx insn)
{
gcc_assert (QUEUE_INDEX (insn) >= 0);
remove_free_INSN_LIST_elem (insn, &insn_queue[QUEUE_INDEX (insn)]);
q_size--;
QUEUE_INDEX (insn) = QUEUE_NOWHERE;
}
/* Return a pointer to the bottom of the ready list, i.e. the insn
with the lowest priority. */
rtx *
ready_lastpos (struct ready_list *ready)
{
gcc_assert (ready->n_ready >= 1);
return ready->vec + ready->first - ready->n_ready + 1;
}
/* Add an element INSN to the ready list so that it ends up with the
lowest/highest priority depending on FIRST_P. */
HAIFA_INLINE static void
ready_add (struct ready_list *ready, rtx insn, bool first_p)
{
if (!first_p)
{
if (ready->first == ready->n_ready)
{
memmove (ready->vec + ready->veclen - ready->n_ready,
ready_lastpos (ready),
ready->n_ready * sizeof (rtx));
ready->first = ready->veclen - 1;
}
ready->vec[ready->first - ready->n_ready] = insn;
}
else
{
if (ready->first == ready->veclen - 1)
{
if (ready->n_ready)
/* ready_lastpos() fails when called with (ready->n_ready == 0). */
memmove (ready->vec + ready->veclen - ready->n_ready - 1,
ready_lastpos (ready),
ready->n_ready * sizeof (rtx));
ready->first = ready->veclen - 2;
}
ready->vec[++(ready->first)] = insn;
}
ready->n_ready++;
if (DEBUG_INSN_P (insn)) if (DEBUG_INSN_P (insn))
ready->n_debug++; ready->n_debug++;
...@@ -2001,6 +2784,9 @@ ready_sort (struct ready_list *ready) ...@@ -2001,6 +2784,9 @@ ready_sort (struct ready_list *ready)
if (!DEBUG_INSN_P (first[i])) if (!DEBUG_INSN_P (first[i]))
setup_insn_reg_pressure_info (first[i]); setup_insn_reg_pressure_info (first[i]);
} }
if (sched_pressure == SCHED_PRESSURE_MODEL
&& model_curr_point < model_num_insns)
model_set_excess_costs (first, ready->n_ready);
SCHED_SORT (first, ready->n_ready); SCHED_SORT (first, ready->n_ready);
} }
...@@ -2016,176 +2802,785 @@ adjust_priority (rtx prev) ...@@ -2016,176 +2802,785 @@ adjust_priority (rtx prev)
notes, which we removed in schedule_region. Nor did it try to notes, which we removed in schedule_region. Nor did it try to
take into account register pressure or anything useful like that. take into account register pressure or anything useful like that.
Revisit when we have a machine model to work with and not before. */ Revisit when we have a machine model to work with and not before. */
if (targetm.sched.adjust_priority)
INSN_PRIORITY (prev) =
targetm.sched.adjust_priority (prev, INSN_PRIORITY (prev));
}
/* Advance DFA state STATE on one cycle. */
void
advance_state (state_t state)
{
if (targetm.sched.dfa_pre_advance_cycle)
targetm.sched.dfa_pre_advance_cycle ();
if (targetm.sched.dfa_pre_cycle_insn)
state_transition (state,
targetm.sched.dfa_pre_cycle_insn ());
state_transition (state, NULL);
if (targetm.sched.dfa_post_cycle_insn)
state_transition (state,
targetm.sched.dfa_post_cycle_insn ());
if (targetm.sched.dfa_post_advance_cycle)
targetm.sched.dfa_post_advance_cycle ();
}
/* Advance time on one cycle. */
HAIFA_INLINE static void
advance_one_cycle (void)
{
advance_state (curr_state);
if (sched_verbose >= 6)
fprintf (sched_dump, ";;\tAdvanced a state.\n");
}
/* Update register pressure after scheduling INSN. */
static void
update_register_pressure (rtx insn)
{
struct reg_use_data *use;
struct reg_set_data *set;
gcc_checking_assert (!DEBUG_INSN_P (insn));
for (use = INSN_REG_USE_LIST (insn); use != NULL; use = use->next_insn_use)
if (dying_use_p (use))
mark_regno_birth_or_death (curr_reg_live, curr_reg_pressure,
use->regno, false);
for (set = INSN_REG_SET_LIST (insn); set != NULL; set = set->next_insn_set)
mark_regno_birth_or_death (curr_reg_live, curr_reg_pressure,
set->regno, true);
}
/* Set up or update (if UPDATE_P) max register pressure (see its
meaning in sched-int.h::_haifa_insn_data) for all current BB insns
after insn AFTER. */
static void
setup_insn_max_reg_pressure (rtx after, bool update_p)
{
int i, p;
bool eq_p;
rtx insn;
static int max_reg_pressure[N_REG_CLASSES];
save_reg_pressure ();
for (i = 0; i < ira_pressure_classes_num; i++)
max_reg_pressure[ira_pressure_classes[i]]
= curr_reg_pressure[ira_pressure_classes[i]];
for (insn = NEXT_INSN (after);
insn != NULL_RTX && ! BARRIER_P (insn)
&& BLOCK_FOR_INSN (insn) == BLOCK_FOR_INSN (after);
insn = NEXT_INSN (insn))
if (NONDEBUG_INSN_P (insn))
{
eq_p = true;
for (i = 0; i < ira_pressure_classes_num; i++)
{
p = max_reg_pressure[ira_pressure_classes[i]];
if (INSN_MAX_REG_PRESSURE (insn)[i] != p)
{
eq_p = false;
INSN_MAX_REG_PRESSURE (insn)[i]
= max_reg_pressure[ira_pressure_classes[i]];
}
}
if (update_p && eq_p)
break;
update_register_pressure (insn);
for (i = 0; i < ira_pressure_classes_num; i++)
if (max_reg_pressure[ira_pressure_classes[i]]
< curr_reg_pressure[ira_pressure_classes[i]])
max_reg_pressure[ira_pressure_classes[i]]
= curr_reg_pressure[ira_pressure_classes[i]];
}
restore_reg_pressure ();
}
/* Update the current register pressure after scheduling INSN. Update
also max register pressure for unscheduled insns of the current
BB. */
static void
update_reg_and_insn_max_reg_pressure (rtx insn)
{
int i;
int before[N_REG_CLASSES];
for (i = 0; i < ira_pressure_classes_num; i++)
before[i] = curr_reg_pressure[ira_pressure_classes[i]];
update_register_pressure (insn);
for (i = 0; i < ira_pressure_classes_num; i++)
if (curr_reg_pressure[ira_pressure_classes[i]] != before[i])
break;
if (i < ira_pressure_classes_num)
setup_insn_max_reg_pressure (insn, true);
}
/* Set up register pressure at the beginning of basic block BB whose
insns starting after insn AFTER. Set up also max register pressure
for all insns of the basic block. */
void
sched_setup_bb_reg_pressure_info (basic_block bb, rtx after)
{
gcc_assert (sched_pressure == SCHED_PRESSURE_WEIGHTED);
initiate_bb_reg_pressure_info (bb);
setup_insn_max_reg_pressure (after, false);
}
/* If doing predication while scheduling, verify whether INSN, which
has just been scheduled, clobbers the conditions of any
instructions that must be predicated in order to break their
dependencies. If so, remove them from the queues so that they will
only be scheduled once their control dependency is resolved. */
static void
check_clobbered_conditions (rtx insn)
{
HARD_REG_SET t;
int i;
if ((current_sched_info->flags & DO_PREDICATION) == 0)
return;
find_all_hard_reg_sets (insn, &t);
restart:
for (i = 0; i < ready.n_ready; i++)
{
rtx x = ready_element (&ready, i);
if (TODO_SPEC (x) == DEP_CONTROL && cond_clobbered_p (x, t))
{
ready_remove_insn (x);
goto restart;
}
}
for (i = 0; i <= max_insn_queue_index; i++)
{
rtx link;
int q = NEXT_Q_AFTER (q_ptr, i);
restart_queue:
for (link = insn_queue[q]; link; link = XEXP (link, 1))
{
rtx x = XEXP (link, 0);
if (TODO_SPEC (x) == DEP_CONTROL && cond_clobbered_p (x, t))
{
queue_remove (x);
goto restart_queue;
}
}
}
}
/* Return (in order):
- positive if INSN adversely affects the pressure on one
register class
- negative if INSN reduces the pressure on one register class
- 0 if INSN doesn't affect the pressure on any register class. */
static int
model_classify_pressure (struct model_insn_info *insn)
{
struct reg_pressure_data *reg_pressure;
int death[N_REG_CLASSES];
int pci, cl, sum;
calculate_reg_deaths (insn->insn, death);
reg_pressure = INSN_REG_PRESSURE (insn->insn);
sum = 0;
for (pci = 0; pci < ira_pressure_classes_num; pci++)
{
cl = ira_pressure_classes[pci];
if (death[cl] < reg_pressure[pci].set_increase)
return 1;
sum += reg_pressure[pci].set_increase - death[cl];
}
return sum;
}
/* Return true if INSN1 should come before INSN2 in the model schedule. */
static int
model_order_p (struct model_insn_info *insn1, struct model_insn_info *insn2)
{
unsigned int height1, height2;
unsigned int priority1, priority2;
/* Prefer instructions with a higher model priority. */
if (insn1->model_priority != insn2->model_priority)
return insn1->model_priority > insn2->model_priority;
/* Combine the length of the longest path of satisfied true dependencies
that leads to each instruction (depth) with the length of the longest
path of any dependencies that leads from the instruction (alap).
Prefer instructions with the greatest combined length. If the combined
lengths are equal, prefer instructions with the greatest depth.
The idea is that, if we have a set S of "equal" instructions that each
have ALAP value X, and we pick one such instruction I, any true-dependent
successors of I that have ALAP value X - 1 should be preferred over S.
This encourages the schedule to be "narrow" rather than "wide".
However, if I is a low-priority instruction that we decided to
schedule because of its model_classify_pressure, and if there
is a set of higher-priority instructions T, the aforementioned
successors of I should not have the edge over T. */
height1 = insn1->depth + insn1->alap;
height2 = insn2->depth + insn2->alap;
if (height1 != height2)
return height1 > height2;
if (insn1->depth != insn2->depth)
return insn1->depth > insn2->depth;
/* We have no real preference between INSN1 an INSN2 as far as attempts
to reduce pressure go. Prefer instructions with higher priorities. */
priority1 = INSN_PRIORITY (insn1->insn);
priority2 = INSN_PRIORITY (insn2->insn);
if (priority1 != priority2)
return priority1 > priority2;
/* Use the original rtl sequence as a tie-breaker. */
return insn1 < insn2;
}
/* Add INSN to the model worklist immediately after PREV. Add it to the
beginning of the list if PREV is null. */
static void
model_add_to_worklist_at (struct model_insn_info *insn,
struct model_insn_info *prev)
{
gcc_assert (QUEUE_INDEX (insn->insn) == QUEUE_NOWHERE);
QUEUE_INDEX (insn->insn) = QUEUE_READY;
insn->prev = prev;
if (prev)
{
insn->next = prev->next;
prev->next = insn;
}
else
{
insn->next = model_worklist;
model_worklist = insn;
}
if (insn->next)
insn->next->prev = insn;
}
/* Remove INSN from the model worklist. */
static void
model_remove_from_worklist (struct model_insn_info *insn)
{
gcc_assert (QUEUE_INDEX (insn->insn) == QUEUE_READY);
QUEUE_INDEX (insn->insn) = QUEUE_NOWHERE;
if (insn->prev)
insn->prev->next = insn->next;
else
model_worklist = insn->next;
if (insn->next)
insn->next->prev = insn->prev;
}
/* Add INSN to the model worklist. Start looking for a suitable position
between neighbors PREV and NEXT, testing at most MAX_SCHED_READY_INSNS
insns either side. A null PREV indicates the beginning of the list and
a null NEXT indicates the end. */
static void
model_add_to_worklist (struct model_insn_info *insn,
struct model_insn_info *prev,
struct model_insn_info *next)
{
int count;
count = MAX_SCHED_READY_INSNS;
if (count > 0 && prev && model_order_p (insn, prev))
do
{
count--;
prev = prev->prev;
}
while (count > 0 && prev && model_order_p (insn, prev));
else
while (count > 0 && next && model_order_p (next, insn))
{
count--;
prev = next;
next = next->next;
}
model_add_to_worklist_at (insn, prev);
}
/* INSN may now have a higher priority (in the model_order_p sense)
than before. Move it up the worklist if necessary. */
static void
model_promote_insn (struct model_insn_info *insn)
{
struct model_insn_info *prev;
int count;
prev = insn->prev;
count = MAX_SCHED_READY_INSNS;
while (count > 0 && prev && model_order_p (insn, prev))
{
count--;
prev = prev->prev;
}
if (prev != insn->prev)
{
model_remove_from_worklist (insn);
model_add_to_worklist_at (insn, prev);
}
}
/* Add INSN to the end of the model schedule. */
static void
model_add_to_schedule (rtx insn)
{
unsigned int point;
gcc_assert (QUEUE_INDEX (insn) == QUEUE_NOWHERE);
QUEUE_INDEX (insn) = QUEUE_SCHEDULED;
point = VEC_length (rtx, model_schedule);
VEC_quick_push (rtx, model_schedule, insn);
INSN_MODEL_INDEX (insn) = point + 1;
}
/* Analyze the instructions that are to be scheduled, setting up
MODEL_INSN_INFO (...) and model_num_insns accordingly. Add ready
instructions to model_worklist. */
static void
model_analyze_insns (void)
{
rtx start, end, iter;
sd_iterator_def sd_it;
dep_t dep;
struct model_insn_info *insn, *con;
model_num_insns = 0;
start = PREV_INSN (current_sched_info->next_tail);
end = current_sched_info->prev_head;
for (iter = start; iter != end; iter = PREV_INSN (iter))
if (NONDEBUG_INSN_P (iter))
{
insn = MODEL_INSN_INFO (iter);
insn->insn = iter;
FOR_EACH_DEP (iter, SD_LIST_FORW, sd_it, dep)
{
con = MODEL_INSN_INFO (DEP_CON (dep));
if (con->insn && insn->alap < con->alap + 1)
insn->alap = con->alap + 1;
}
insn->old_queue = QUEUE_INDEX (iter);
QUEUE_INDEX (iter) = QUEUE_NOWHERE;
insn->unscheduled_preds = dep_list_size (iter, SD_LIST_HARD_BACK);
if (insn->unscheduled_preds == 0)
model_add_to_worklist (insn, NULL, model_worklist);
if (targetm.sched.adjust_priority) model_num_insns++;
INSN_PRIORITY (prev) = }
targetm.sched.adjust_priority (prev, INSN_PRIORITY (prev));
} }
/* Advance DFA state STATE on one cycle. */ /* The global state describes the register pressure at the start of the
void model schedule. Initialize GROUP accordingly. */
advance_state (state_t state)
{
if (targetm.sched.dfa_pre_advance_cycle)
targetm.sched.dfa_pre_advance_cycle ();
if (targetm.sched.dfa_pre_cycle_insn) static void
state_transition (state, model_init_pressure_group (struct model_pressure_group *group)
targetm.sched.dfa_pre_cycle_insn ()); {
int pci, cl;
state_transition (state, NULL); for (pci = 0; pci < ira_pressure_classes_num; pci++)
{
cl = ira_pressure_classes[pci];
group->limits[pci].pressure = curr_reg_pressure[cl];
group->limits[pci].point = 0;
}
/* Use index model_num_insns to record the state after the last
instruction in the model schedule. */
group->model = XNEWVEC (struct model_pressure_data,
(model_num_insns + 1) * ira_pressure_classes_num);
}
if (targetm.sched.dfa_post_cycle_insn) /* Record that MODEL_REF_PRESSURE (GROUP, POINT, PCI) is PRESSURE.
state_transition (state, Update the maximum pressure for the whole schedule. */
targetm.sched.dfa_post_cycle_insn ());
if (targetm.sched.dfa_post_advance_cycle) static void
targetm.sched.dfa_post_advance_cycle (); model_record_pressure (struct model_pressure_group *group,
int point, int pci, int pressure)
{
MODEL_REF_PRESSURE (group, point, pci) = pressure;
if (group->limits[pci].pressure < pressure)
{
group->limits[pci].pressure = pressure;
group->limits[pci].point = point;
}
} }
/* Advance time on one cycle. */ /* INSN has just been added to the end of the model schedule. Record its
HAIFA_INLINE static void register-pressure information. */
advance_one_cycle (void)
static void
model_record_pressures (struct model_insn_info *insn)
{ {
advance_state (curr_state); struct reg_pressure_data *reg_pressure;
if (sched_verbose >= 6) int point, pci, cl, delta;
fprintf (sched_dump, ";;\tAdvanced a state.\n"); int death[N_REG_CLASSES];
point = model_index (insn->insn);
if (sched_verbose >= 2)
{
char buf[2048];
if (point == 0)
{
fprintf (sched_dump, "\n;;\tModel schedule:\n;;\n");
fprintf (sched_dump, ";;\t| idx insn | mpri hght dpth prio |\n");
}
print_pattern (buf, PATTERN (insn->insn), 0);
fprintf (sched_dump, ";;\t| %3d %4d | %4d %4d %4d %4d | %-30s ",
point, INSN_UID (insn->insn), insn->model_priority,
insn->depth + insn->alap, insn->depth,
INSN_PRIORITY (insn->insn), buf);
}
calculate_reg_deaths (insn->insn, death);
reg_pressure = INSN_REG_PRESSURE (insn->insn);
for (pci = 0; pci < ira_pressure_classes_num; pci++)
{
cl = ira_pressure_classes[pci];
delta = reg_pressure[pci].set_increase - death[cl];
if (sched_verbose >= 2)
fprintf (sched_dump, " %s:[%d,%+d]", reg_class_names[cl],
curr_reg_pressure[cl], delta);
model_record_pressure (&model_before_pressure, point, pci,
curr_reg_pressure[cl]);
}
if (sched_verbose >= 2)
fprintf (sched_dump, "\n");
} }
/* Update register pressure after scheduling INSN. */ /* All instructions have been added to the model schedule. Record the
final register pressure in GROUP and set up all MODEL_MAX_PRESSUREs. */
static void static void
update_register_pressure (rtx insn) model_record_final_pressures (struct model_pressure_group *group)
{ {
struct reg_use_data *use; int point, pci, max_pressure, ref_pressure, cl;
struct reg_set_data *set;
gcc_checking_assert (!DEBUG_INSN_P (insn)); for (pci = 0; pci < ira_pressure_classes_num; pci++)
{
/* Record the final pressure for this class. */
cl = ira_pressure_classes[pci];
point = model_num_insns;
ref_pressure = curr_reg_pressure[cl];
model_record_pressure (group, point, pci, ref_pressure);
for (use = INSN_REG_USE_LIST (insn); use != NULL; use = use->next_insn_use) /* Record the original maximum pressure. */
if (dying_use_p (use) && bitmap_bit_p (curr_reg_live, use->regno)) group->limits[pci].orig_pressure = group->limits[pci].pressure;
mark_regno_birth_or_death (use->regno, false);
for (set = INSN_REG_SET_LIST (insn); set != NULL; set = set->next_insn_set) /* Update the MODEL_MAX_PRESSURE for every point of the schedule. */
mark_regno_birth_or_death (set->regno, true); max_pressure = ref_pressure;
MODEL_MAX_PRESSURE (group, point, pci) = max_pressure;
while (point > 0)
{
point--;
ref_pressure = MODEL_REF_PRESSURE (group, point, pci);
max_pressure = MAX (max_pressure, ref_pressure);
MODEL_MAX_PRESSURE (group, point, pci) = max_pressure;
}
}
} }
/* Set up or update (if UPDATE_P) max register pressure (see its /* Update all successors of INSN, given that INSN has just been scheduled. */
meaning in sched-int.h::_haifa_insn_data) for all current BB insns
after insn AFTER. */
static void static void
setup_insn_max_reg_pressure (rtx after, bool update_p) model_add_successors_to_worklist (struct model_insn_info *insn)
{ {
int i, p; sd_iterator_def sd_it;
bool eq_p; struct model_insn_info *con;
rtx insn; dep_t dep;
static int max_reg_pressure[N_REG_CLASSES];
save_reg_pressure (); FOR_EACH_DEP (insn->insn, SD_LIST_FORW, sd_it, dep)
for (i = 0; i < ira_pressure_classes_num; i++)
max_reg_pressure[ira_pressure_classes[i]]
= curr_reg_pressure[ira_pressure_classes[i]];
for (insn = NEXT_INSN (after);
insn != NULL_RTX && ! BARRIER_P (insn)
&& BLOCK_FOR_INSN (insn) == BLOCK_FOR_INSN (after);
insn = NEXT_INSN (insn))
if (NONDEBUG_INSN_P (insn))
{ {
eq_p = true; con = MODEL_INSN_INFO (DEP_CON (dep));
for (i = 0; i < ira_pressure_classes_num; i++) /* Ignore debug instructions, and instructions from other blocks. */
if (con->insn)
{ {
p = max_reg_pressure[ira_pressure_classes[i]]; con->unscheduled_preds--;
if (INSN_MAX_REG_PRESSURE (insn)[i] != p)
/* Update the depth field of each true-dependent successor.
Increasing the depth gives them a higher priority than
before. */
if (DEP_TYPE (dep) == REG_DEP_TRUE && con->depth < insn->depth + 1)
{ {
eq_p = false; con->depth = insn->depth + 1;
INSN_MAX_REG_PRESSURE (insn)[i] if (QUEUE_INDEX (con->insn) == QUEUE_READY)
= max_reg_pressure[ira_pressure_classes[i]]; model_promote_insn (con);
} }
/* If this is a true dependency, or if there are no remaining
dependencies for CON (meaning that CON only had non-true
dependencies), make sure that CON is on the worklist.
We don't bother otherwise because it would tend to fill the
worklist with a lot of low-priority instructions that are not
yet ready to issue. */
if ((con->depth > 0 || con->unscheduled_preds == 0)
&& QUEUE_INDEX (con->insn) == QUEUE_NOWHERE)
model_add_to_worklist (con, insn, insn->next);
} }
if (update_p && eq_p)
break;
update_register_pressure (insn);
for (i = 0; i < ira_pressure_classes_num; i++)
if (max_reg_pressure[ira_pressure_classes[i]]
< curr_reg_pressure[ira_pressure_classes[i]])
max_reg_pressure[ira_pressure_classes[i]]
= curr_reg_pressure[ira_pressure_classes[i]];
} }
restore_reg_pressure ();
} }
/* Update the current register pressure after scheduling INSN. Update /* Give INSN a higher priority than any current instruction, then give
also max register pressure for unscheduled insns of the current unscheduled predecessors of INSN a higher priority still. If any of
BB. */ those predecessors are not on the model worklist, do the same for its
predecessors, and so on. */
static void static void
update_reg_and_insn_max_reg_pressure (rtx insn) model_promote_predecessors (struct model_insn_info *insn)
{ {
int i; struct model_insn_info *pro, *first;
int before[N_REG_CLASSES]; sd_iterator_def sd_it;
dep_t dep;
for (i = 0; i < ira_pressure_classes_num; i++) if (sched_verbose >= 7)
before[i] = curr_reg_pressure[ira_pressure_classes[i]]; fprintf (sched_dump, ";;\t+--- priority of %d = %d, priority of",
update_register_pressure (insn); INSN_UID (insn->insn), model_next_priority);
for (i = 0; i < ira_pressure_classes_num; i++) insn->model_priority = model_next_priority++;
if (curr_reg_pressure[ira_pressure_classes[i]] != before[i]) model_remove_from_worklist (insn);
break; model_add_to_worklist_at (insn, NULL);
if (i < ira_pressure_classes_num)
setup_insn_max_reg_pressure (insn, true);
}
/* Set up register pressure at the beginning of basic block BB whose first = NULL;
insns starting after insn AFTER. Set up also max register pressure for (;;)
for all insns of the basic block. */ {
void FOR_EACH_DEP (insn->insn, SD_LIST_HARD_BACK, sd_it, dep)
sched_setup_bb_reg_pressure_info (basic_block bb, rtx after) {
{ pro = MODEL_INSN_INFO (DEP_PRO (dep));
gcc_assert (sched_pressure == SCHED_PRESSURE_WEIGHTED); /* The first test is to ignore debug instructions, and instructions
initiate_bb_reg_pressure_info (bb); from other blocks. */
setup_insn_max_reg_pressure (after, false); if (pro->insn
&& pro->model_priority != model_next_priority
&& QUEUE_INDEX (pro->insn) != QUEUE_SCHEDULED)
{
pro->model_priority = model_next_priority;
if (sched_verbose >= 7)
fprintf (sched_dump, " %d", INSN_UID (pro->insn));
if (QUEUE_INDEX (pro->insn) == QUEUE_READY)
{
/* PRO is already in the worklist, but it now has
a higher priority than before. Move it at the
appropriate place. */
model_remove_from_worklist (pro);
model_add_to_worklist (pro, NULL, model_worklist);
}
else
{
/* PRO isn't in the worklist. Recursively process
its predecessors until we find one that is. */
pro->next = first;
first = pro;
}
}
}
if (!first)
break;
insn = first;
first = insn->next;
}
if (sched_verbose >= 7)
fprintf (sched_dump, " = %d\n", model_next_priority);
model_next_priority++;
} }
/* If doing predication while scheduling, verify whether INSN, which /* Pick one instruction from model_worklist and process it. */
has just been scheduled, clobbers the conditions of any
instructions that must be predicated in order to break their
dependencies. If so, remove them from the queues so that they will
only be scheduled once their control dependency is resolved. */
static void static void
check_clobbered_conditions (rtx insn) model_choose_insn (void)
{ {
HARD_REG_SET t; struct model_insn_info *insn, *fallback;
int i; int count;
if ((current_sched_info->flags & DO_PREDICATION) == 0) if (sched_verbose >= 7)
return; {
fprintf (sched_dump, ";;\t+--- worklist:\n");
insn = model_worklist;
count = MAX_SCHED_READY_INSNS;
while (count > 0 && insn)
{
fprintf (sched_dump, ";;\t+--- %d [%d, %d, %d, %d]\n",
INSN_UID (insn->insn), insn->model_priority,
insn->depth + insn->alap, insn->depth,
INSN_PRIORITY (insn->insn));
count--;
insn = insn->next;
}
}
find_all_hard_reg_sets (insn, &t); /* Look for a ready instruction whose model_classify_priority is zero
or negative, picking the highest-priority one. Adding such an
instruction to the schedule now should do no harm, and may actually
do some good.
restart: Failing that, see whether there is an instruction with the highest
for (i = 0; i < ready.n_ready; i++) extant model_priority that is not yet ready, but which would reduce
pressure if it became ready. This is designed to catch cases like:
(set (mem (reg R1)) (reg R2))
where the instruction is the last remaining use of R1 and where the
value of R2 is not yet available (or vice versa). The death of R1
means that this instruction already reduces pressure. It is of
course possible that the computation of R2 involves other registers
that are hard to kill, but such cases are rare enough for this
heuristic to be a win in general.
Failing that, just pick the highest-priority instruction in the
worklist. */
count = MAX_SCHED_READY_INSNS;
insn = model_worklist;
fallback = 0;
for (;;)
{ {
rtx x = ready_element (&ready, i); if (count == 0 || !insn)
if (TODO_SPEC (x) == DEP_CONTROL && cond_clobbered_p (x, t))
{ {
ready_remove_insn (x); insn = fallback ? fallback : model_worklist;
goto restart; break;
} }
if (insn->unscheduled_preds)
{
if (model_worklist->model_priority == insn->model_priority
&& !fallback
&& model_classify_pressure (insn) < 0)
fallback = insn;
} }
for (i = 0; i <= max_insn_queue_index; i++) else
{ {
rtx link; if (model_classify_pressure (insn) <= 0)
int q = NEXT_Q_AFTER (q_ptr, i); break;
}
count--;
insn = insn->next;
}
restart_queue: if (sched_verbose >= 7 && insn != model_worklist)
for (link = insn_queue[q]; link; link = XEXP (link, 1))
{
rtx x = XEXP (link, 0);
if (TODO_SPEC (x) == DEP_CONTROL && cond_clobbered_p (x, t))
{ {
queue_remove (x); if (insn->unscheduled_preds)
goto restart_queue; fprintf (sched_dump, ";;\t+--- promoting insn %d, with dependencies\n",
INSN_UID (insn->insn));
else
fprintf (sched_dump, ";;\t+--- promoting insn %d, which is ready\n",
INSN_UID (insn->insn));
} }
if (insn->unscheduled_preds)
/* INSN isn't yet ready to issue. Give all its predecessors the
highest priority. */
model_promote_predecessors (insn);
else
{
/* INSN is ready. Add it to the end of model_schedule and
process its successors. */
model_add_successors_to_worklist (insn);
model_remove_from_worklist (insn);
model_add_to_schedule (insn->insn);
model_record_pressures (insn);
update_register_pressure (insn->insn);
} }
}
/* Restore all QUEUE_INDEXs to the values that they had before
model_start_schedule was called. */
static void
model_reset_queue_indices (void)
{
unsigned int i;
rtx insn;
FOR_EACH_VEC_ELT (rtx, model_schedule, i, insn)
QUEUE_INDEX (insn) = MODEL_INSN_INFO (insn)->old_queue;
}
/* We have calculated the model schedule and spill costs. Print a summary
to sched_dump. */
static void
model_dump_pressure_summary (void)
{
int pci, cl;
fprintf (sched_dump, ";; Pressure summary:");
for (pci = 0; pci < ira_pressure_classes_num; pci++)
{
cl = ira_pressure_classes[pci];
fprintf (sched_dump, " %s:%d", reg_class_names[cl],
model_before_pressure.limits[pci].pressure);
} }
fprintf (sched_dump, "\n\n");
}
/* Initialize the SCHED_PRESSURE_MODEL information for the current
scheduling region. */
static void
model_start_schedule (void)
{
basic_block bb;
model_next_priority = 1;
model_schedule = VEC_alloc (rtx, heap, sched_max_luid);
model_insns = XCNEWVEC (struct model_insn_info, sched_max_luid);
bb = BLOCK_FOR_INSN (NEXT_INSN (current_sched_info->prev_head));
initiate_reg_pressure_info (df_get_live_in (bb));
model_analyze_insns ();
model_init_pressure_group (&model_before_pressure);
while (model_worklist)
model_choose_insn ();
gcc_assert (model_num_insns == (int) VEC_length (rtx, model_schedule));
if (sched_verbose >= 2)
fprintf (sched_dump, "\n");
model_record_final_pressures (&model_before_pressure);
model_reset_queue_indices ();
XDELETEVEC (model_insns);
model_curr_point = 0;
initiate_reg_pressure_info (df_get_live_in (bb));
if (sched_verbose >= 1)
model_dump_pressure_summary ();
}
/* Free the information associated with GROUP. */
static void
model_finalize_pressure_group (struct model_pressure_group *group)
{
XDELETEVEC (group->model);
}
/* Free the information created by model_start_schedule. */
static void
model_end_schedule (void)
{
model_finalize_pressure_group (&model_before_pressure);
VEC_free (rtx, heap, model_schedule);
} }
/* A structure that holds local state for the loop in schedule_block. */ /* A structure that holds local state for the loop in schedule_block. */
...@@ -2240,6 +3635,10 @@ schedule_insn (rtx insn) ...@@ -2240,6 +3635,10 @@ schedule_insn (rtx insn)
reg_class_names[ira_pressure_classes[i]], reg_class_names[ira_pressure_classes[i]],
pressure_info[i].set_increase, pressure_info[i].change); pressure_info[i].set_increase, pressure_info[i].change);
} }
if (sched_pressure == SCHED_PRESSURE_MODEL
&& model_curr_point < model_num_insns
&& model_index (insn) == model_curr_point)
fprintf (sched_dump, ":model %d", model_curr_point);
fputc ('\n', sched_dump); fputc ('\n', sched_dump);
} }
...@@ -2307,6 +3706,24 @@ schedule_insn (rtx insn) ...@@ -2307,6 +3706,24 @@ schedule_insn (rtx insn)
gcc_assert (QUEUE_INDEX (insn) == QUEUE_NOWHERE); gcc_assert (QUEUE_INDEX (insn) == QUEUE_NOWHERE);
QUEUE_INDEX (insn) = QUEUE_SCHEDULED; QUEUE_INDEX (insn) = QUEUE_SCHEDULED;
if (sched_pressure == SCHED_PRESSURE_MODEL
&& model_curr_point < model_num_insns
&& NONDEBUG_INSN_P (insn))
{
if (model_index (insn) == model_curr_point)
do
model_curr_point++;
while (model_curr_point < model_num_insns
&& (QUEUE_INDEX (MODEL_INSN (model_curr_point))
== QUEUE_SCHEDULED));
else
model_recompute (insn);
model_update_limit_points ();
update_register_pressure (insn);
if (sched_verbose >= 2)
print_curr_reg_pressure ();
}
gcc_assert (INSN_TICK (insn) >= MIN_TICK); gcc_assert (INSN_TICK (insn) >= MIN_TICK);
if (INSN_TICK (insn) > clock_var) if (INSN_TICK (insn) > clock_var)
/* INSN has been prematurely moved from the queue to the ready list. /* INSN has been prematurely moved from the queue to the ready list.
...@@ -3138,7 +4555,16 @@ queue_to_ready (struct ready_list *ready) ...@@ -3138,7 +4555,16 @@ queue_to_ready (struct ready_list *ready)
/* If the ready list is full, delay the insn for 1 cycle. /* If the ready list is full, delay the insn for 1 cycle.
See the comment in schedule_block for the rationale. */ See the comment in schedule_block for the rationale. */
if (!reload_completed if (!reload_completed
&& ready->n_ready - ready->n_debug > MAX_SCHED_READY_INSNS && (ready->n_ready - ready->n_debug > MAX_SCHED_READY_INSNS
|| (sched_pressure == SCHED_PRESSURE_MODEL
/* Limit pressure recalculations to MAX_SCHED_READY_INSNS
instructions too. */
&& model_index (insn) > (model_curr_point
+ MAX_SCHED_READY_INSNS)))
&& !(sched_pressure == SCHED_PRESSURE_MODEL
&& model_curr_point < model_num_insns
/* Always allow the next model instruction to issue. */
&& model_index (insn) == model_curr_point)
&& !SCHED_GROUP_P (insn) && !SCHED_GROUP_P (insn)
&& insn != skip_insn) && insn != skip_insn)
queue_insn (insn, 1, "ready full"); queue_insn (insn, 1, "ready full");
...@@ -3366,12 +4792,12 @@ debug_ready_list (struct ready_list *ready) ...@@ -3366,12 +4792,12 @@ debug_ready_list (struct ready_list *ready)
fprintf (sched_dump, " %s:%d", fprintf (sched_dump, " %s:%d",
(*current_sched_info->print_insn) (p[i], 0), (*current_sched_info->print_insn) (p[i], 0),
INSN_LUID (p[i])); INSN_LUID (p[i]));
if (sched_pressure == SCHED_PRESSURE_WEIGHTED) if (sched_pressure != SCHED_PRESSURE_NONE)
fprintf (sched_dump, "(cost=%d", fprintf (sched_dump, "(cost=%d",
INSN_REG_PRESSURE_EXCESS_COST_CHANGE (p[i])); INSN_REG_PRESSURE_EXCESS_COST_CHANGE (p[i]));
if (INSN_TICK (p[i]) > clock_var) if (INSN_TICK (p[i]) > clock_var)
fprintf (sched_dump, ":delay=%d", INSN_TICK (p[i]) - clock_var); fprintf (sched_dump, ":delay=%d", INSN_TICK (p[i]) - clock_var);
if (sched_pressure == SCHED_PRESSURE_WEIGHTED) if (sched_pressure != SCHED_PRESSURE_NONE)
fprintf (sched_dump, ")"); fprintf (sched_dump, ")");
} }
fprintf (sched_dump, "\n"); fprintf (sched_dump, "\n");
...@@ -4001,8 +5427,17 @@ prune_ready_list (state_t temp_state, bool first_cycle_insn_p, ...@@ -4001,8 +5427,17 @@ prune_ready_list (state_t temp_state, bool first_cycle_insn_p,
cost = 1; cost = 1;
reason = "asm"; reason = "asm";
} }
else if (sched_pressure == SCHED_PRESSURE_WEIGHTED) else if (sched_pressure != SCHED_PRESSURE_NONE)
{
if (sched_pressure == SCHED_PRESSURE_MODEL
&& INSN_TICK (insn) <= clock_var)
{
memcpy (temp_state, curr_state, dfa_state_size);
if (state_transition (temp_state, insn) >= 0)
INSN_TICK (insn) = clock_var + 1;
}
cost = 0; cost = 0;
}
else else
{ {
int delay_cost = 0; int delay_cost = 0;
...@@ -4175,6 +5610,9 @@ schedule_block (basic_block *target_bb) ...@@ -4175,6 +5610,9 @@ schedule_block (basic_block *target_bb)
in try_ready () (which is called through init_ready_list ()). */ in try_ready () (which is called through init_ready_list ()). */
(*current_sched_info->init_ready_list) (); (*current_sched_info->init_ready_list) ();
if (sched_pressure == SCHED_PRESSURE_MODEL)
model_start_schedule ();
/* The algorithm is O(n^2) in the number of ready insns at any given /* The algorithm is O(n^2) in the number of ready insns at any given
time in the worst case. Before reload we are more likely to have time in the worst case. Before reload we are more likely to have
big lists so truncate them to a reasonable size. */ big lists so truncate them to a reasonable size. */
...@@ -4420,7 +5858,7 @@ schedule_block (basic_block *target_bb) ...@@ -4420,7 +5858,7 @@ schedule_block (basic_block *target_bb)
else else
insn = ready_remove_first (&ready); insn = ready_remove_first (&ready);
if (sched_pressure == SCHED_PRESSURE_WEIGHTED if (sched_pressure != SCHED_PRESSURE_NONE
&& INSN_TICK (insn) > clock_var) && INSN_TICK (insn) > clock_var)
{ {
ready_add (&ready, insn, true); ready_add (&ready, insn, true);
...@@ -4671,6 +6109,9 @@ schedule_block (basic_block *target_bb) ...@@ -4671,6 +6109,9 @@ schedule_block (basic_block *target_bb)
} }
} }
if (sched_pressure == SCHED_PRESSURE_MODEL)
model_end_schedule ();
if (success) if (success)
{ {
commit_schedule (prev_head, tail, target_bb); commit_schedule (prev_head, tail, target_bb);
...@@ -4793,7 +6234,7 @@ sched_init (void) ...@@ -4793,7 +6234,7 @@ sched_init (void)
else else
sched_pressure = SCHED_PRESSURE_NONE; sched_pressure = SCHED_PRESSURE_NONE;
if (sched_pressure == SCHED_PRESSURE_WEIGHTED) if (sched_pressure != SCHED_PRESSURE_NONE)
ira_setup_eliminable_regset (); ira_setup_eliminable_regset ();
/* Initialize SPEC_INFO. */ /* Initialize SPEC_INFO. */
...@@ -4871,7 +6312,7 @@ sched_init (void) ...@@ -4871,7 +6312,7 @@ sched_init (void)
if (targetm.sched.init_global) if (targetm.sched.init_global)
targetm.sched.init_global (sched_dump, sched_verbose, get_max_uid () + 1); targetm.sched.init_global (sched_dump, sched_verbose, get_max_uid () + 1);
if (sched_pressure == SCHED_PRESSURE_WEIGHTED) if (sched_pressure != SCHED_PRESSURE_NONE)
{ {
int i, max_regno = max_reg_num (); int i, max_regno = max_reg_num ();
...@@ -4888,9 +6329,12 @@ sched_init (void) ...@@ -4888,9 +6329,12 @@ sched_init (void)
? ira_pressure_class_translate[REGNO_REG_CLASS (i)] ? ira_pressure_class_translate[REGNO_REG_CLASS (i)]
: ira_pressure_class_translate[reg_allocno_class (i)]); : ira_pressure_class_translate[reg_allocno_class (i)]);
curr_reg_live = BITMAP_ALLOC (NULL); curr_reg_live = BITMAP_ALLOC (NULL);
if (sched_pressure == SCHED_PRESSURE_WEIGHTED)
{
saved_reg_live = BITMAP_ALLOC (NULL); saved_reg_live = BITMAP_ALLOC (NULL);
region_ref_regs = BITMAP_ALLOC (NULL); region_ref_regs = BITMAP_ALLOC (NULL);
} }
}
curr_state = xmalloc (dfa_state_size); curr_state = xmalloc (dfa_state_size);
} }
...@@ -4988,14 +6432,17 @@ void ...@@ -4988,14 +6432,17 @@ void
sched_finish (void) sched_finish (void)
{ {
haifa_finish_h_i_d (); haifa_finish_h_i_d ();
if (sched_pressure == SCHED_PRESSURE_WEIGHTED) if (sched_pressure != SCHED_PRESSURE_NONE)
{ {
if (regstat_n_sets_and_refs != NULL) if (regstat_n_sets_and_refs != NULL)
regstat_free_n_sets_and_refs (); regstat_free_n_sets_and_refs ();
free (sched_regno_pressure_class); if (sched_pressure == SCHED_PRESSURE_WEIGHTED)
{
BITMAP_FREE (region_ref_regs); BITMAP_FREE (region_ref_regs);
BITMAP_FREE (saved_reg_live); BITMAP_FREE (saved_reg_live);
}
BITMAP_FREE (curr_reg_live); BITMAP_FREE (curr_reg_live);
free (sched_regno_pressure_class);
} }
free (curr_state); free (curr_state);
...@@ -5267,7 +6714,7 @@ fix_tick_ready (rtx next) ...@@ -5267,7 +6714,7 @@ fix_tick_ready (rtx next)
INSN_TICK (next) = tick; INSN_TICK (next) = tick;
delay = tick - clock_var; delay = tick - clock_var;
if (delay <= 0 || sched_pressure == SCHED_PRESSURE_WEIGHTED) if (delay <= 0 || sched_pressure != SCHED_PRESSURE_NONE)
delay = QUEUE_READY; delay = QUEUE_READY;
change_queue_index (next, delay); change_queue_index (next, delay);
...@@ -6522,7 +7969,7 @@ add_jump_dependencies (rtx insn, rtx jump) ...@@ -6522,7 +7969,7 @@ add_jump_dependencies (rtx insn, rtx jump)
if (insn == jump) if (insn == jump)
break; break;
if (dep_list_size (insn) == 0) if (dep_list_size (insn, SD_LIST_FORW) == 0)
{ {
dep_def _new_dep, *new_dep = &_new_dep; dep_def _new_dep, *new_dep = &_new_dep;
...@@ -6663,6 +8110,7 @@ haifa_finish_h_i_d (void) ...@@ -6663,6 +8110,7 @@ haifa_finish_h_i_d (void)
FOR_EACH_VEC_ELT (haifa_insn_data_def, h_i_d, i, data) FOR_EACH_VEC_ELT (haifa_insn_data_def, h_i_d, i, data)
{ {
free (data->max_reg_pressure);
free (data->reg_pressure); free (data->reg_pressure);
for (use = data->reg_use_list; use != NULL; use = next) for (use = data->reg_use_list; use != NULL; use = next)
{ {
...@@ -6693,7 +8141,7 @@ haifa_init_insn (rtx insn) ...@@ -6693,7 +8141,7 @@ haifa_init_insn (rtx insn)
/* Extend dependency caches by one element. */ /* Extend dependency caches by one element. */
extend_dependency_caches (1, false); extend_dependency_caches (1, false);
} }
if (sched_pressure == SCHED_PRESSURE_WEIGHTED) if (sched_pressure != SCHED_PRESSURE_NONE)
init_insn_reg_pressure_info (insn); init_insn_reg_pressure_info (insn);
} }
......
...@@ -2168,7 +2168,7 @@ init_insn_reg_pressure_info (rtx insn) ...@@ -2168,7 +2168,7 @@ init_insn_reg_pressure_info (rtx insn)
static struct reg_pressure_data *pressure_info; static struct reg_pressure_data *pressure_info;
rtx link; rtx link;
gcc_assert (sched_pressure == SCHED_PRESSURE_WEIGHTED); gcc_assert (sched_pressure != SCHED_PRESSURE_NONE);
if (! INSN_P (insn)) if (! INSN_P (insn))
return; return;
...@@ -2199,6 +2199,7 @@ init_insn_reg_pressure_info (rtx insn) ...@@ -2199,6 +2199,7 @@ init_insn_reg_pressure_info (rtx insn)
len = sizeof (struct reg_pressure_data) * ira_pressure_classes_num; len = sizeof (struct reg_pressure_data) * ira_pressure_classes_num;
pressure_info pressure_info
= INSN_REG_PRESSURE (insn) = (struct reg_pressure_data *) xmalloc (len); = INSN_REG_PRESSURE (insn) = (struct reg_pressure_data *) xmalloc (len);
if (sched_pressure == SCHED_PRESSURE_WEIGHTED)
INSN_MAX_REG_PRESSURE (insn) = (int *) xcalloc (ira_pressure_classes_num INSN_MAX_REG_PRESSURE (insn) = (int *) xcalloc (ira_pressure_classes_num
* sizeof (int), 1); * sizeof (int), 1);
for (i = 0; i < ira_pressure_classes_num; i++) for (i = 0; i < ira_pressure_classes_num; i++)
...@@ -2951,7 +2952,7 @@ sched_analyze_insn (struct deps_desc *deps, rtx x, rtx insn) ...@@ -2951,7 +2952,7 @@ sched_analyze_insn (struct deps_desc *deps, rtx x, rtx insn)
|| (NONJUMP_INSN_P (insn) && control_flow_insn_p (insn))) || (NONJUMP_INSN_P (insn) && control_flow_insn_p (insn)))
reg_pending_barrier = MOVE_BARRIER; reg_pending_barrier = MOVE_BARRIER;
if (sched_pressure == SCHED_PRESSURE_WEIGHTED) if (sched_pressure != SCHED_PRESSURE_NONE)
{ {
setup_insn_reg_uses (deps, insn); setup_insn_reg_uses (deps, insn);
init_insn_reg_pressure_info (insn); init_insn_reg_pressure_info (insn);
......
...@@ -794,6 +794,11 @@ struct _haifa_insn_data ...@@ -794,6 +794,11 @@ struct _haifa_insn_data
short cost; short cost;
/* '> 0' if priority is valid,
'== 0' if priority was not yet computed,
'< 0' if priority in invalid and should be recomputed. */
signed char priority_status;
/* Set if there's DEF-USE dependence between some speculatively /* Set if there's DEF-USE dependence between some speculatively
moved load insn and this one. */ moved load insn and this one. */
unsigned int fed_by_spec_load : 1; unsigned int fed_by_spec_load : 1;
...@@ -811,11 +816,6 @@ struct _haifa_insn_data ...@@ -811,11 +816,6 @@ struct _haifa_insn_data
their TODO_SPEC recomputed. */ their TODO_SPEC recomputed. */
unsigned int must_recompute_spec : 1; unsigned int must_recompute_spec : 1;
/* '> 0' if priority is valid,
'== 0' if priority was not yet computed,
'< 0' if priority in invalid and should be recomputed. */
signed char priority_status;
/* What speculations are necessary to apply to schedule the instruction. */ /* What speculations are necessary to apply to schedule the instruction. */
ds_t todo_spec; ds_t todo_spec;
...@@ -854,6 +854,7 @@ struct _haifa_insn_data ...@@ -854,6 +854,7 @@ struct _haifa_insn_data
/* Info about how scheduling the insn changes cost of register /* Info about how scheduling the insn changes cost of register
pressure excess (between source and target). */ pressure excess (between source and target). */
int reg_pressure_excess_cost_change; int reg_pressure_excess_cost_change;
int model_index;
}; };
typedef struct _haifa_insn_data haifa_insn_data_def; typedef struct _haifa_insn_data haifa_insn_data_def;
...@@ -876,6 +877,7 @@ extern VEC(haifa_insn_data_def, heap) *h_i_d; ...@@ -876,6 +877,7 @@ extern VEC(haifa_insn_data_def, heap) *h_i_d;
#define INSN_REG_PRESSURE_EXCESS_COST_CHANGE(INSN) \ #define INSN_REG_PRESSURE_EXCESS_COST_CHANGE(INSN) \
(HID (INSN)->reg_pressure_excess_cost_change) (HID (INSN)->reg_pressure_excess_cost_change)
#define INSN_PRIORITY_STATUS(INSN) (HID (INSN)->priority_status) #define INSN_PRIORITY_STATUS(INSN) (HID (INSN)->priority_status)
#define INSN_MODEL_INDEX(INSN) (HID (INSN)->model_index)
typedef struct _haifa_deps_insn_data haifa_deps_insn_data_def; typedef struct _haifa_deps_insn_data haifa_deps_insn_data_def;
typedef haifa_deps_insn_data_def *haifa_deps_insn_data_t; typedef haifa_deps_insn_data_def *haifa_deps_insn_data_t;
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment