#include "irvrfy.h"
#include <lpp/lpp.h>
+#include <lpp/mps.h>
#include <lpp/lpp_net.h>
#include <lpp/lpp_cplex.h>
//#include <lc_pset.h>
#include "bechordal_t.h"
-#define BIGM 100000.0
-
-#define DUMP_SOLUTION
-#define DUMP_ILP
-//#define KEEPALIVE /* keep alive all inserted remats and dump graph with remats */
-#define COLLECT_REMATS /* enable rematerialization */
-#define COLLECT_INVERSE_REMATS /* enable placement of inverse remats */
-#define REMAT_WHILE_LIVE /* only remat values that are live */
-//#define NO_ENLARGE_L1V3N355 /* do not remat after the death of some operand */
-//#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
-#define MAY_DIE_AT_REMAT /* allow values to die after a pre remat */
-#define NO_SINGLE_USE_REMATS /* do not repair schedule */
-//#define KEEPALIVE_SPILLS
-//#define KEEPALIVE_RELOADS
-#define GOODWIN_REDUCTION
-//#define NO_MEMCOPIES
-//#define VERIFY_DOMINANCE
+#ifdef WITH_LIBCORE
+#include <libcore/lc_opts.h>
+#include <libcore/lc_opts_enum.h>
+#endif /* WITH_LIBCORE */
+
+#define DUMP_PROBLEM 1
+#define DUMP_MPS 2
+#define DUMP_SOLUTION 4
+
+#define KEEPALIVE_REMATS 1
+#define KEEPALIVE_SPILLS 2
+#define KEEPALIVE_RELOADS 4
+
+#define VERIFY_MEMINTERF 1
+#define VERIFY_DOMINANCE 2
+
+#define REMATS_NONE 0
+#define REMATS_BRIGGS 1
+#define REMATS_NOINVERSE 2
+#define REMATS_ALL 3
+
+static int opt_dump_flags = 0;
+static int opt_log = 0;
+static int opt_keep_alive = 0;
+static int opt_goodwin = 1;
+static int opt_memcopies = 1;
+static int opt_memoperands = 1;
+static int opt_verify = VERIFY_MEMINTERF;
+static int opt_remats = REMATS_ALL;
+static int opt_repair_schedule = 0;
+static int opt_no_enlarge_liveness = 0;
+static int opt_remat_while_live = 1;
+static int opt_timeout = 300;
+static double opt_cost_reload = 8.0;
+static double opt_cost_memoperand = 7.0;
+static double opt_cost_spill = 50.0;
+static double opt_cost_remat = 1.0;
+
+
+#ifdef WITH_LIBCORE
+static const lc_opt_enum_mask_items_t dump_items[] = {
+ { "problem", DUMP_PROBLEM },
+ { "mps", DUMP_MPS },
+ { "solution", DUMP_SOLUTION },
+ { NULL, 0 }
+};
+
+static lc_opt_enum_mask_var_t dump_var = {
+ &opt_dump_flags, dump_items
+};
+
+static const lc_opt_enum_mask_items_t keepalive_items[] = {
+ { "remats", KEEPALIVE_REMATS },
+ { "spills", KEEPALIVE_SPILLS },
+ { "reloads", KEEPALIVE_RELOADS },
+ { NULL, 0 }
+};
+
+static lc_opt_enum_mask_var_t keep_alive_var = {
+ &opt_keep_alive, keepalive_items
+};
+
+static const lc_opt_enum_mask_items_t remats_items[] = {
+ { "none", REMATS_NONE },
+ { "briggs", REMATS_BRIGGS },
+ { "noinverse", REMATS_NOINVERSE },
+ { "all", REMATS_ALL },
+ { NULL, 0 }
+};
+
+static lc_opt_enum_mask_var_t remats_var = {
+ &opt_remats, remats_items
+};
+
+static const lc_opt_table_entry_t options[] = {
+ LC_OPT_ENT_ENUM_MASK("keepalive", "keep alive remats, spills or reloads", &keep_alive_var),
+
+ LC_OPT_ENT_BOOL ("goodwin", "activate goodwin reduction", &opt_goodwin),
+ LC_OPT_ENT_BOOL ("memcopies", "activate memcopy handling", &opt_memcopies),
+ LC_OPT_ENT_BOOL ("memoperands", "activate memoperands", &opt_memoperands),
+ LC_OPT_ENT_ENUM_INT ("remats", "type of remats to insert (none, briggs, noinverse or all)",&remats_var),
+ LC_OPT_ENT_BOOL ("repair_schedule", "repair the schedule by rematting once used nodes",&opt_repair_schedule),
+ LC_OPT_ENT_BOOL ("no_enlage_liveness", "do not enlarge liveness of operands of remats",&opt_no_enlarge_liveness),
+ LC_OPT_ENT_BOOL ("remat_while_live", "remat only values that can be used by real ops", &opt_remat_while_live),
+
+ LC_OPT_ENT_ENUM_MASK("dump", "dump problem, mps or solution", &dump_var),
+ LC_OPT_ENT_BOOL ("log", "activate the lpp log", &opt_log),
+ LC_OPT_ENT_INT ("timeout", "ILP solver timeout", &opt_timeout),
+
+ LC_OPT_ENT_DBL ("cost_reload", "cost of a reload", &opt_cost_reload),
+ LC_OPT_ENT_DBL ("cost_memoperand", "cost of a memory operand", &opt_cost_memoperand),
+ LC_OPT_ENT_DBL ("cost_spill", "cost of a spill instruction", &opt_cost_spill),
+ LC_OPT_ENT_DBL ("cost_remat", "cost of a rematerialization", &opt_cost_remat),
+ { NULL }
+};
+
+void be_spill_remat_register_options(lc_opt_entry_t *grp)
+{
+ lc_opt_entry_t *my_grp = lc_opt_get_grp(grp, "remat");
+ lc_opt_add_table(my_grp, options);
+}
+#endif
+
+
+//#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
+//#define SCHEDULE_PHIM /* insert phim nodes into schedule */
#define SOLVE
//#define SOLVE_LOCAL
#define LPP_SERVER "i44pc52"
#define LPP_SOLVER "cplex"
-#define COST_LOAD 8
-#define COST_STORE 50
-#define COST_REMAT 1
-#define ILP_TIMEOUT 300
-#define MAX_PATHS 16
+#define MAX_PATHS INT_MAX
#define ILP_UNDEF -1
typedef struct _spill_ilp_t {
set *remat_info;
pset *all_possible_remats;
pset *inverse_ops;
-#ifdef KEEPALIVE
ir_node *keep;
-#endif
set *values; /**< for collecting all definitions of values before running ssa-construction */
pset *spills;
set *interferences;
ir_node *m_unknown;
+ set *memoperands;
DEBUG_ONLY(firm_dbg_module_t * dbg);
} spill_ilp_t;
typedef struct _remat_t {
const ir_node *op; /**< for copy_irn */
const ir_node *value; /**< the value which is being recomputed by this remat */
- ir_node *proj; /**< not NULL if the above op produces a tuple */
+ const ir_node *proj; /**< not NULL if the above op produces a tuple */
int cost; /**< cost of this remat */
int inverse; /**< nonzero if this is an inverse remat */
} remat_t;
union {
struct {
ilp_var_t ilp;
- remat_t *remat; /** the remat this op belongs to */
+ const remat_t *remat; /** the remat this op belongs to */
int pre; /** 1, if this is a pressure-increasing remat */
} remat;
struct {
} op_t;
typedef struct _defs_t {
- ir_node *value;
- ir_node *spills; /**< points to the first spill for this value (linked by link field) */
- ir_node *remats; /**< points to the first definition for this value (linked by link field) */
+ const ir_node *value;
+ ir_node *spills; /**< points to the first spill for this value (linked by link field) */
+ ir_node *remats; /**< points to the first definition for this value (linked by link field) */
} defs_t;
typedef struct _remat_info_t {
} keyval_t;
typedef struct _spill_t {
- ir_node *irn;
- ilp_var_t reg_in;
- ilp_var_t mem_in;
- ilp_var_t reg_out;
- ilp_var_t mem_out;
- ilp_var_t spill;
+ ir_node *irn;
+ ilp_var_t reg_in;
+ ilp_var_t mem_in;
+ ilp_var_t reg_out;
+ ilp_var_t mem_out;
+ ilp_var_t spill;
} spill_t;
+typedef struct _memoperand_t {
+ ir_node *irn; /**< the irn */
+ unsigned int pos; /**< the position of the argument */
+ ilp_var_t ilp; /**< the ilp var for this memory operand */
+} memoperand_t;
+
static INLINE int
has_reg_class(const spill_ilp_t * si, const ir_node * irn)
{
return !(p->irn == q->irn);
}
+static int
+cmp_memoperands(const void *a, const void *b, size_t size)
+{
+ const memoperand_t *p = a;
+ const memoperand_t *q = b;
+
+ return !(p->irn == q->irn && p->pos == q->pos);
+}
+
static keyval_t *
-set_find_keyval(set * set, void * key)
+set_find_keyval(set * set, const void * key)
{
keyval_t query;
}
static defs_t *
-set_find_def(set * set, ir_node * value)
+set_find_def(set * set, const ir_node * value)
{
defs_t query;
}
static defs_t *
-set_insert_def(set * set, ir_node * value)
+set_insert_def(set * set, const ir_node * value)
{
defs_t query;
return set_insert(set, &query, sizeof(query), HASH_PTR(value));
}
+static memoperand_t *
+set_insert_memoperand(set * set, ir_node * irn, unsigned int pos, ilp_var_t ilp)
+{
+ memoperand_t query;
+
+ query.irn = irn;
+ query.pos = pos;
+ query.ilp = ilp;
+ return set_insert(set, &query, sizeof(query), HASH_PTR(irn)+pos);
+}
+
+static memoperand_t *
+set_find_memoperand(set * set, const ir_node * irn, unsigned int pos)
+{
+ memoperand_t query;
+
+ query.irn = (ir_node*)irn;
+ query.pos = pos;
+ return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos);
+}
+
+
static spill_t *
-set_find_spill(set * set, ir_node * value)
+set_find_spill(set * set, const ir_node * value)
{
spill_t query;
- query.irn = value;
+ query.irn = (ir_node*)value;
return set_find(set, &query, sizeof(query), HASH_PTR(value));
}
get_cost(const spill_ilp_t * si, const ir_node * irn)
{
if(be_is_Spill(irn)) {
- return COST_STORE;
+ return opt_cost_spill;
} else if(be_is_Reload(irn)){
- return COST_LOAD;
+ return opt_cost_reload;
} else {
return arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, irn);
}
const ir_node *proj = NULL;
if(is_Proj(dest_value)) {
- op = get_irn_n(op, 0);
+ op = get_Proj_pred(op);
proj = dest_value;
}
return i;
}
+static int
+get_irn_n_nonignore_args(const spill_ilp_t * si, const ir_node * irn)
+{
+ int n;
+ int ret = 0;
+
+// if(is_Proj(irn))
+// irn = get_Proj_pred(irn);
+
+ for(n=get_irn_arity(irn)-1; n>=0; --n) {
+ const ir_node *arg = get_irn_n(irn, n);
+
+ if(has_reg_class(si, arg)) ++ret;
+ }
+
+ return ret;
+}
+
static INLINE void
get_remats_from_op(spill_ilp_t * si, const ir_node * op)
{
int n;
remat_t *remat;
-#ifdef NO_SINGLE_USE_REMATS
- if(has_reg_class(si, op) && (get_irn_n_nonremat_edges(si, op) > 1)) {
-#else
- if(has_reg_class(si, op)) {
-#endif
+ if( has_reg_class(si, op)
+ && (opt_repair_schedule || get_irn_n_nonremat_edges(si, op) > 1)
+ && (opt_remats != REMATS_BRIGGS || get_irn_n_nonignore_args(si, op) == 0)
+ ) {
remat = get_remat_from_op(si, op, op);
if(remat) {
add_remat(si, remat);
}
}
-#ifdef COLLECT_INVERSE_REMATS
- /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
- for each arg */
- for (n = get_irn_arity(op)-1; n>=0; --n) {
- ir_node *arg = get_irn_n(op, n);
+ if(opt_remats == REMATS_ALL) {
+ /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
+ for each arg */
+ for (n = get_irn_arity(op)-1; n>=0; --n) {
+ ir_node *arg = get_irn_n(op, n);
- if(has_reg_class(si, arg)) {
- /* try to get an inverse remat */
- remat = get_remat_from_op(si, arg, op);
- if(remat) {
- add_remat(si, remat);
+ if(has_reg_class(si, arg)) {
+ /* try to get an inverse remat */
+ remat = get_remat_from_op(si, arg, op);
+ if(remat) {
+ add_remat(si, remat);
+ }
}
}
}
-#endif
-
}
static INLINE int
for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
const ir_node *arg = get_irn_n(op, n);
-#ifdef NO_ENLARGE_L1V3N355
- if(has_reg_class(si, arg) && live) {
- res &= pset_find_ptr(live, arg)?1:0;
+ if(opt_no_enlarge_liveness) {
+ if(has_reg_class(si, arg) && live) {
+ res &= pset_find_ptr((pset*)live, arg)?1:0;
+ } else {
+ res &= value_is_defined_before(si, pos, arg);
+ }
} else {
res &= value_is_defined_before(si, pos, arg);
}
-#else
- res &= value_is_defined_before(si, pos, arg);
-#endif
}
return res;
}
static ir_node *
-insert_remat_after(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
+insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
{
char buf[256];
*proj_copy;
op_t *op;
- DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
+ DBG((si->dbg, LEVEL_3, "\t >inserting remat2 %+F\n", remat->op));
copy = insert_copy_after(si, remat->op, pos);
op->is_remat = 1;
op->attr.remat.remat = remat;
op->attr.remat.pre = 0;
- op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
+ op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
set_irn_link(copy, op);
pset_insert_ptr(si->all_possible_remats, copy);
}
static ir_node *
-insert_remat_before(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
+insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
{
char buf[256];
op->is_remat = 1;
op->attr.remat.remat = remat;
op->attr.remat.pre = 1;
- op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
+ op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
set_irn_link(copy, op);
pset_insert_ptr(si->all_possible_remats, copy);
return edge ? 2 : 1;
}
+static int
+is_start_block(const ir_node * bb)
+{
+ return get_irg_start_block(get_irn_irg(bb)) == bb;
+}
+
+static int
+is_before_frame(const ir_node * bb, const ir_node * irn)
+{
+ const ir_node *frame = get_irg_frame(get_irn_irg(bb));
+
+ if(is_start_block(bb) && sched_get_time_step(frame) >= sched_get_time_step(irn))
+ return 1;
+ else
+ return 0;
+}
+
static int
is_merge_edge(const ir_node * bb)
{
-#ifdef GOODWIN_REDUCTION
- return get_block_n_succs(bb) == 1;
-#else
- return 1;
-#endif
+ if(is_start_block(bb))
+ return 0;
+
+ if(opt_goodwin)
+ return get_block_n_succs(bb) == 1;
+ else
+ return 1;
}
static int
is_diverge_edge(const ir_node * bb)
{
-#ifdef GOODWIN_REDUCTION
- return get_Block_n_cfgpreds(bb) == 1;
-#else
- return 1;
-#endif
+ if(is_start_block(bb))
+ return 0;
+
+ if(opt_goodwin)
+ return get_Block_n_cfgpreds(bb) == 1;
+ else
+ return 1;
}
static void
ir_node *remat_irn = NULL;
DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
-#ifdef REMAT_WHILE_LIVE
- if(pset_find_ptr(live, remat->value)) {
+ if(opt_remat_while_live) {
+ if(pset_find_ptr(live, remat->value)) {
+ remat_irn = insert_remat_before(si, remat, irn, live);
+ }
+ } else {
remat_irn = insert_remat_before(si, remat, irn, live);
}
-#else
- remat_irn = insert_remat_before(si, remat, irn, live);
-#endif
if(remat_irn) {
for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
ir_node *remat_arg = get_irn_n(remat_irn, n);
/* do not insert remats producing the same value as one of the operands */
if(!pset_find_ptr(args, remat->value)) {
DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
-#ifdef REMAT_WHILE_LIVE
- if(pset_find_ptr(live, remat->value)) {
+ if(opt_remat_while_live) {
+ if(pset_find_ptr(live, remat->value)) {
+ insert_remat_after(si, remat, irn, live);
+ }
+ } else {
insert_remat_after(si, remat, irn, live);
}
-#else
- insert_remat_after(si, remat, irn, live);
-#endif
}
}
}
irn = next;
}
- be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_in, i) {
- ir_node *value = be_lv_get_irn(si->lv, bb, i);
+ /* add remats at end if successor has multiple predecessors */
+ if(is_merge_edge(bb)) {
+ pset *live_out = pset_new_ptr_default();
+ ir_node *value;
+
+ be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
+ value = be_lv_get_irn(si->lv, bb, i);
- /* add remats at end if successor has multiple predecessors */
- if(is_merge_edge(bb)) {
- /* add remats at end of block */
if (be_is_live_end(si->lv, bb, value) && has_reg_class(si, value)) {
- remat_info_t *remat_info,
- query;
- remat_t *remat;
+ pset_insert_ptr(live_out, value);
+ }
+ }
- query.irn = value;
- query.remats = NULL;
- query.remats_by_operand = NULL;
- remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
+ /* add remats at end of block */
+ pset_foreach(live_out, value) {
+ remat_info_t *remat_info,
+ query;
+ remat_t *remat;
+
+ query.irn = value;
+ query.remats = NULL;
+ query.remats_by_operand = NULL;
+ remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
- if(remat_info && remat_info->remats) {
- pset_foreach(remat_info->remats, remat) {
- DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
+ if(remat_info && remat_info->remats) {
+ pset_foreach(remat_info->remats, remat) {
+ DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
- insert_remat_before(si, remat, bb, NULL);
- }
+ insert_remat_before(si, remat, bb, live_out);
}
}
}
- if(is_diverge_edge(bb)) {
- /* add remat2s at beginning of block */
+ del_pset(live_out);
+ }
+
+ if(is_diverge_edge(bb)) {
+ pset *live_in = pset_new_ptr_default();
+ ir_node *value;
+
+ be_lv_foreach(si->lv, bb, be_lv_state_in, i) {
+ value = be_lv_get_irn(si->lv, bb, i);
+
+ if (has_reg_class(si, value)) {
+ pset_insert_ptr(live_in, value);
+ }
+ }
+
+ /* add remat2s at beginning of block */
+ pset_foreach(live_in, value) {
if ((be_is_live_in(si->lv, bb, value) || (is_Phi(value) && get_nodes_block(value)==bb)) && has_reg_class(si, value)) {
remat_info_t *remat_info,
query;
query.remats_by_operand = NULL;
remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
- if(remat_info && remat_info->remats) {
- pset_foreach(remat_info->remats, remat) {
- DBG((si->dbg, LEVEL_4, "\t considering remat %+F at beginning of block %+F\n", remat->op, bb));
+ if(remat_info && remat_info->remats_by_operand) {
+ pset_foreach(remat_info->remats_by_operand, remat) {
+ DBG((si->dbg, LEVEL_4, "\t considering remat2 %+F at beginning of block %+F\n", remat->op, bb));
/* put the remat here if all its args are available */
- insert_remat_after(si, remat, bb, NULL);
+ insert_remat_after(si, remat, bb, live_in);
}
}
}
}
+ del_pset(live_in);
}
}
}
ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
- //cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
- cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
+ //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
spill_t query,
*spill;
double spill_cost;
+ int default_spilled;
/* handle values used by control flow nodes later separately */
query.irn = irn;
spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
- spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
+ spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
- spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
+ spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
- spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
+ spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
- spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
+ /* by default spill value right after definition */
+ default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
+ spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
if(is_merge_edge(bb)) {
ilp_var_t reload;
ilp_cst_t rel_cst;
ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
- reload = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
+ reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
/* reload <= mem_out */
- rel_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
+ rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
- }
+ }
spill->reg_in = ILP_UNDEF;
spill->mem_in = ILP_UNDEF;
ilp_cst_t end_use_req,
rel_cst;
ilp_var_t reload;
+ int default_spilled;
query.irn = irn;
spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
- spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
+ spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
- spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
+ spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
/* if irn is used at the end of the block, then it is live anyway */
//lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
- spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
+ spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
- spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
+ default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
+ spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
+ /* reload for use be control flow op */
ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
- reload = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
+ reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
/* reload <= mem_out */
- rel_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
+ rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
spill->mem_in = ILP_UNDEF;
ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
- end_use_req = lpp_add_cst(si->lpp, buf, lpp_equal, 1);
+ end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
}
next_post_remat(const ir_node * irn)
{
op_t *op;
+ ir_node *next;
if(is_Block(irn)) {
- irn = sched_block_first_nonphi(irn);
+ next = sched_block_first_nonphi(irn);
} else {
- irn = sched_next_op(irn);
+ next = sched_next_op(irn);
}
- if(sched_is_end(irn))
+ if(sched_is_end(next))
return NULL;
- op = (op_t*)get_irn_link(irn);
+ op = get_irn_link(next);
if(op->is_remat && !op->attr.remat.pre) {
- return irn;
+ return next;
}
return NULL;
spill_t *spill,
query;
char buf[256];
+ int default_spilled;
query.irn = irn;
spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
if(!spill) {
- double spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
+ double spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
spill->mem_in = ILP_UNDEF;
ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
- spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
+ spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
- spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
+ default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
+ spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
}
return spill;
if(edge) return;
ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
- copyreg = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
+ copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
- cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
pset_foreach(live, tmp) {
spill_t *spill;
}
ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
- cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
/* copy - reg_out - reload - remat - live_range <= 0 */
lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
}
ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
- cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
/* copy - reg_out - copyreg <= 0 */
lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
ir_node *tmp;
spill_t *spill;
pset *defs = pset_new_ptr_default();
+ const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
live = pset_new_ptr_default();
assert(!op->is_remat);
ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
- op->attr.live_range.ilp = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
+ op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
op->attr.live_range.op = bb;
ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
- cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
/* reg_out - reload - remat - live_range <= 0 */
lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
/* maybe we should also assure that reg_out >= live_range etc. */
}
-#ifndef NO_MEMCOPIES
- insert_mem_copy_position(si, live, bb);
-#endif
+ if(opt_memcopies)
+ insert_mem_copy_position(si, live, bb);
/*
* start new live ranges for values used by remats at end of block
/* if value is becoming live through use by remat */
if(!pset_find_ptr(live, remat_arg)) {
ir_snprintf(buf, sizeof(buf), "lr_%N_end%N", remat_arg, bb);
- prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
+ prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
arg_op->attr.live_range.ilp = prev_lr;
arg_op->attr.live_range.op = bb;
pset *used;
pset *remat_defs;
keyval_t *keyval;
+ ilp_cst_t one_memoperand;
/* iterate only until first phi */
if(is_Phi(irn))
if(n_remats == 0) {
/* sum remat2s <= 1 + n_remats*live_range */
ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
- cst = lpp_add_cst(si->lpp, buf, lpp_less, 1.0);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
}
n_remats++;
/* remat + \sum live_range(remat_arg) <= |args| */
ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
- cst = lpp_add_cst(si->lpp, buf, lpp_less, pset_count(remat_args));
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
pset_foreach(remat_args, remat_arg) {
/* next_live_range <= prev_live_range + sum remat2s */
ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
- cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
- prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
+ prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
ilp_var_t lr;
ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
- lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
+ lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
arg_op->attr.live_range.ilp = lr;
arg_op->attr.live_range.op = irn;
/* check the register pressure in the epilog */
/* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
- check_post = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - d);
+ check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
/* add L\U' to check_post */
pset_foreach(live, tmp) {
/* new live range for each used value */
ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
- prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
+ prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, is_before_frame(bb, irn)?1.0:0.0);
/* the epilog stuff - including post_use, check_post, check_post_remat */
ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
- post_use = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
+ post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
/* post_use >= next_lr + remat */
ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
- cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
if(!set_find_keyval(args, arg)) {
/* post_use <= prev_lr */
ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
- cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
/* next_lr <= prev_lr */
ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
- cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
}
}
-
/* forall post remat which use arg add a similar cst */
foreach_post_remat(irn, remat) {
int n;
DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
- cst = lpp_add_cst(si->lpp, buf, lpp_greater, 0.0);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
+ lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
+ lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
+ }
+ }
+ }
+
+ if(opt_memoperands) {
+ for(n = get_irn_arity(irn)-1; n>=0; --n) {
+ if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
+ ilp_var_t memoperand;
+
+ ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
+ memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_memoperand*execution_frequency(si, bb), 0.0);
+ set_insert_memoperand(si->memoperands, irn, n, memoperand);
+
+ ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
+
+ lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
- lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
+// if(arg_op->attr.live_range.ilp != ILP_UNDEF)
+// lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
}
}
}
/* check the register pressure in the prolog */
/* sum_{L\U} lr <= k - |U| */
ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
- check_pre = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - u);
+ check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
/* for the prolog remove defined values from the live set */
pset_foreach(defs, tmp) {
pset_remove_ptr(live, tmp);
}
+ if(opt_memoperands) {
+ ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
+ one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
+ }
+
/***********************************************************
* I T E R A T I O N O V E R A R G S F O R P R O L O G
**********************************************************/
set_foreach(args, keyval) {
- spill_t *spill;
- ir_node *arg = keyval->key;
- int i = PTR_TO_INT(keyval->val);
- op_t *arg_op = get_irn_link(arg);
+ spill_t *spill;
+ const ir_node *arg = keyval->key;
+ int i = PTR_TO_INT(keyval->val);
+ op_t *arg_op = get_irn_link(arg);
+ ilp_cst_t requirements;
+ int n_memoperands;
spill = set_find_spill(spill_bb->ilp, arg);
assert(spill);
ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
- op->attr.live_range.args.reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
+ op->attr.live_range.args.reloads[i] = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), is_before_frame(bb, irn)?0.0:1.0);
/* reload <= mem_out */
ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
- cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
/* requirement: arg must be in register for use */
/* reload + remat + live_range == 1 */
ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
- cst = lpp_add_cst(si->lpp, buf, lpp_equal, 1.0);
+ requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
- lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
- lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
+ lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
+ lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
foreach_pre_remat(si, irn, tmp) {
op_t *remat_op = get_irn_link(tmp);
if(remat_op->attr.remat.remat->value == arg) {
- lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
+ lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
+ }
+ }
+
+ if(opt_memoperands) {
+ n_memoperands = 0;
+ for(n = get_irn_arity(irn)-1; n>=0; --n) {
+ if(get_irn_n(irn, n) == arg) {
+ n_memoperands++;
+ }
+ }
+ for(n = get_irn_arity(irn)-1; n>=0; --n) {
+ if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
+ memoperand_t *memoperand;
+ memoperand = set_find_memoperand(si->memoperands, irn, n);
+
+ /* memoperand <= mem_out */
+ ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
+ lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
+ lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
+
+ /* the memoperand is only sufficient if it is used once by the op */
+ if(n_memoperands == 1)
+ lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
+
+ lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
+
+ /* we have one more free register if we use a memory operand */
+ lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
+ }
}
}
}
for (n=get_irn_arity(tmp)-1; n>=0; --n) {
ir_node *remat_arg = get_irn_n(tmp, n);
op_t *arg_op = get_irn_link(remat_arg);
- ilp_var_t prev_lr;
if(!has_reg_class(si, remat_arg)) continue;
set_foreach(spill_bb->ilp, spill) {
ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
- cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
if(pset_find_ptr(live, spill->irn)) {
+ int default_spilled;
DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
- spill->mem_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
+ default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn);
+ spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled);
lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
- if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
+ if(opt_memcopies && is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
int n;
op_t *op = get_irn_link(spill->irn);
/* copies are not for free */
ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
- var = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE * freq);
+ var = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_spill * freq, 1.0);
for(m=n; m>=0; --m) {
const ir_node *arg2 = get_irn_n(spill->irn, m);
}
}
+#if 0
/* copy <= mem_in */
ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
- cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
lpp_set_factor_fast(si->lpp, cst, var, 1.0);
lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
+#endif
}
}
}
}
+ foreach_post_remat(bb, tmp) {
+ int n;
+
+ for (n=get_irn_arity(tmp)-1; n>=0; --n) {
+ ir_node *remat_arg = get_irn_n(tmp, n);
+
+ if(!has_reg_class(si, remat_arg)) continue;
+
+ /* if value is becoming live through use by remat2 */
+ if(!pset_find_ptr(live, remat_arg)) {
+ op_t *remat_arg_op = get_irn_link(remat_arg);
+ ilp_cst_t nomem;
+
+ DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 at bb start %+F\n", remat_arg, tmp));
+
+ pset_insert_ptr(live, remat_arg);
+ spill = add_to_spill_bb(si, bb, remat_arg);
+ remat_arg_op->attr.live_range.ilp = ILP_UNDEF;
+
+ /* we need reg_in and mem_in for this value; they will be referenced later */
+ ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", remat_arg, bb);
+ spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
+ ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", remat_arg, bb);
+ spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
+
+
+ /* optimization: all memory stuff should be 0, for we do not want to insert reloads for remats */
+ ir_snprintf(buf, sizeof(buf), "nomem_%N_%N", remat_arg, bb);
+ nomem = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0.0);
+
+ lpp_set_factor_fast(si->lpp, nomem, spill->spill, 1.0);
+ if(spill_bb->reloads) {
+ keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
+
+ if(keyval) {
+ ilp_var_t reload = PTR_TO_INT(keyval->val);
+ lpp_set_factor_fast(si->lpp, nomem, reload, 1.0);
+ }
+ }
+ }
+ }
+ }
/* L\U is empty at bb start */
/* arg is live throughout epilog if it is reg_in into this block */
/* check the register pressure at the beginning of the block
* including remats
*/
+ /* reg_in entspricht post_use */
+
ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
- cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
pset_foreach(live, irn) {
ilp_cst_t nospill;
assert(spill);
ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
- spill->reg_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
+ spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
/* spill + mem_in <= 1 */
ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
- nospill = lpp_add_cst(si->lpp, buf, lpp_less, 1);
+ nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
- }
- foreach_post_remat(bb, irn) {
- op_t *remat_op = get_irn_link(irn);
-
- DBG((si->dbg, LEVEL_4, "\t next post remat: %+F\n", irn));
- assert(remat_op->is_remat && !remat_op->attr.remat.pre);
-
- lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
- }
-
- /* forall post remats add requirements */
- foreach_post_remat(bb, tmp) {
- int n;
-
- for (n=get_irn_arity(tmp)-1; n>=0; --n) {
- ir_node *remat_arg = get_irn_n(tmp, n);
- op_t *remat_op = get_irn_link(tmp);
-
- if(!has_reg_class(si, remat_arg)) continue;
-
- spill = set_find_spill(spill_bb->ilp, remat_arg);
- assert(spill);
-
- /* remat <= reg_in_argument */
- ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
- cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
- lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
- lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
- }
- }
+ } /* post_remats are NOT included in register pressure check because
+ they do not increase regpressure */
/* mem_in/reg_in for live_in values, especially phis and their arguments */
pset_foreach(live, irn) {
if(has_reg_class(si, phi_arg)) {
/* mem_in < mem_out_arg + copy */
ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
- mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
+ mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
/* reg_in < reg_out_arg */
ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
- reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
+ reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
assert(spill_p);
lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
- lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
+ if(opt_memcopies)
+ lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
+
lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
}
}
spill_t *spill_p;
ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
- mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
+ mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
- reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
+ reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
}
}
+ foreach_post_remat(bb, tmp) {
+ int n;
+
+ for (n=get_irn_arity(tmp)-1; n>=0; --n) {
+ ir_node *remat_arg = get_irn_n(tmp, n);
+ op_t *remat_op = get_irn_link(tmp);
+
+ if(!has_reg_class(si, remat_arg)) continue;
+
+ spill = set_find_spill(spill_bb->ilp, remat_arg);
+ assert(spill);
+
+ /* remat <= reg_in_argument */
+ ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
+ cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
+ lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
+ lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
+ }
+ }
+
+ pset_foreach(live, irn) {
+ const op_t *op = get_irn_link(irn);
+ const ir_node *remat;
+ int n_remats = 0;
+
+ if(op->attr.live_range.ilp == ILP_UNDEF) continue;
+
+ cst = ILP_UNDEF;
+
+ foreach_post_remat(bb, remat) {
+ int n;
+
+ for (n=get_irn_arity(remat)-1; n>=0; --n) {
+ const ir_node *arg = get_irn_n(remat, n);
+
+ if(arg == irn) {
+ const op_t *remat_op = get_irn_link(remat);
+
+ if(cst == ILP_UNDEF) {
+ /* \sum post_remat <= 1 + #post_remats * next(lr) */
+ ir_snprintf(buf, sizeof(buf), "remat2_%N_%N_arg_%N", remat, bb, irn);
+ cst = lpp_add_cst(si->lpp, buf, lpp_less, 1.0);
+ }
+ lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
+ ++n_remats;
+ break;
+ }
+ }
+ }
+ if(n_remats) {
+ lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, n_remats);
+ }
+ }
+
/* first live ranges from reg_ins */
pset_foreach(live, irn) {
op_t *op = get_irn_link(irn);
- spill = set_find_spill(spill_bb->ilp, irn);
- assert(spill && spill->irn == irn);
+ if(op->attr.live_range.ilp != ILP_UNDEF) {
- ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
- cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
- lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
- lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
+ spill = set_find_spill(spill_bb->ilp, irn);
+ assert(spill && spill->irn == irn);
- foreach_post_remat(bb, tmp) {
- op_t *remat_op = get_irn_link(tmp);
+ ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
+ lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
+ lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
- if(remat_op->attr.remat.remat->value == irn) {
- lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
+ foreach_post_remat(bb, tmp) {
+ op_t *remat_op = get_irn_link(tmp);
+
+ if(remat_op->attr.remat.remat->value == irn) {
+ lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
+ }
}
}
}
assert(spill);
ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
- cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
assert(!is_Proj(tmp));
if(op->is_remat) {
- ir_node *value = op->attr.remat.remat->value;
+ const ir_node *value = op->attr.remat.remat->value;
if(value == irn) {
/* only collect remats up to the first real use of a value */
}
}
}
-next_live:
+next_live: ;
}
del_pset(live);
}
static int
-values_interfere_in_block(const spill_ilp_t * si, ir_node * bb, ir_node * a, ir_node * b)
+values_interfere_in_block(const spill_ilp_t * si, const ir_node * bb, const ir_node * a, const ir_node * b)
{
const ir_edge_t *edge;
if(get_nodes_block(user) == bb
&& !is_Phi(user)
&& b != user
+ && !pset_find_ptr(si->inverse_ops, user)
&& value_dominates(b, user))
return 1;
}
ir_node *a = be_lv_get_irn(si->lv, bb, l1);
op_t *a_op = get_irn_link(a);
- if(a_op->is_remat) continue;
/* a is only interesting if it is in my register class and if it is inside a phi class */
if (has_reg_class(si, a) && get_phi_class(a)) {
+ if(a_op->is_remat || pset_find_ptr(si->inverse_ops, a))
+ continue;
+
for(l2=_be_lv_next_irn(si->lv, bb, 0xff, l1+1); l2>=0; l2=_be_lv_next_irn(si->lv, bb, 0xff, l2+1)) {
ir_node *b = be_lv_get_irn(si->lv, bb, l2);
op_t *b_op = get_irn_link(b);
- if(b_op->is_remat) continue;
/* a and b are only interesting if they are in the same phi class */
if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
+ if(b_op->is_remat || pset_find_ptr(si->inverse_ops, b))
+ continue;
+
if(values_interfere_in_block(si, bb, a, b)) {
DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
set_insert_interference(si, si->interferences, a, b, bb);
void *ptr;
ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
- cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
* @parameter visited contains a set of nodes already visited on this path
*/
static int
-find_copy_path(spill_ilp_t * si, ir_node * irn, ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
+find_copy_path(spill_ilp_t * si, const ir_node * irn, const ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
{
- ir_edge_t *edge;
- op_t *op = get_irn_link(irn);
- pset *visited_users = pset_new_ptr_default();
- int paths = 0;
+ const ir_edge_t *edge;
+ op_t *op = get_irn_link(irn);
+ pset *visited_users = pset_new_ptr_default();
+ int paths = 0;
- if(op->is_remat) return;
+ if(op->is_remat) return 0;
pset_insert_ptr(visited, irn);
paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
pset_remove(copies, INT_TO_PTR(copy), copy);
- /*if(paths > MAX_PATHS) {
+ if(paths > MAX_PATHS) {
if(pset_count(copies) == 0) {
ilp_cst_t cst;
char buf[256];
ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
- cst = lpp_add_cst(si->lpp, buf, lpp_equal, 0);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
}
} else if(pset_count(copies) == 0) {
paths = 0;
- }*/
+ }
}
}
paths += find_copy_path(si, user, target, any_interfere, copies, visited);
pset_remove(copies, INT_TO_PTR(copy), copy);
- /*if(paths > MAX_PATHS) {
+ if(paths > MAX_PATHS) {
if(pset_count(copies) == 0) {
ilp_cst_t cst;
char buf[256];
ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
- cst = lpp_add_cst(si->lpp, buf, lpp_equal, 0);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
}
} else if(pset_count(copies) == 0) {
paths = 0;
- }*/
+ }
}
}
}
static void
-gen_copy_constraints(spill_ilp_t * si, ir_node * a, ir_node * b, ilp_var_t any_interfere)
+gen_copy_constraints(spill_ilp_t * si, const ir_node * a, const ir_node * b, ilp_var_t any_interfere)
{
pset * copies = pset_new_ptr_default();
pset * visited = pset_new_ptr_default();
/* any_interf <= \sum interf */
ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
- any_interfere_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
- any_interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
+ any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
+ any_interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
const ir_node *bb = irnlist->irn;
spill_bb_t *spill_bb = get_irn_link(bb);
spill_t *spilla,
- *spillb,
- query;
+ *spillb;
char buf[256];
- query.irn = a;
spilla = set_find_spill(spill_bb->ilp, a);
assert(spilla);
- query.irn = b;
spillb = set_find_spill(spill_bb->ilp, b);
assert(spillb);
/* 2: - mem_in_a - spill_a + interfere <= 0 */
/* 3: - mem_in_b - spill_b + interfere <= 0 */
ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
- interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
+ interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
- cst = lpp_add_cst(si->lpp, buf, lpp_less, 1);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
- cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
- cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
/* any_interfere >= interf */
ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
- cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
return fabs(x) < 0.00001;
}
-#ifdef KEEPALIVE
static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
{
spill_ilp_t *si = get_irg_link(current_ir_graph);
be_dump(irg, suffix, dump_ir_block_graph_sched);
set_dump_node_vcgattr_hook(NULL);
}
-#endif
/**
* Edge hook to dump the schedule edges with annotated register pressure.
be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
}
-#ifdef KEEPALIVE
static void
connect_all_remats_with_keep(spill_ilp_t * si)
{
obstack_free(si->obst, ins);
}
}
-#endif
static void
connect_all_spills_with_keep(spill_ilp_t * si)
}
/** insert a spill at an arbitrary position */
-ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert, ir_node *ctx)
+ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert)
{
ir_node *bl = is_Block(insert)?insert:get_nodes_block(insert);
ir_graph *irg = get_irn_irg(bl);
const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
- spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn, ctx);
+ spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn);
/*
* search the right insertion point. a spill of a phi cannot be put
pset_foreach(remat_info->remats, remat)
{
if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
- set_irn_n(remat->proj, -1, bad);
- set_irn_n(remat->proj, 0, bad);
+ set_irn_n((ir_node*)remat->proj, -1, bad);
+ set_irn_n((ir_node*)remat->proj, 0, bad);
}
if(get_irn_n_edges(remat->op) == 0) {
for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
- set_irn_n(remat->op, n, bad);
+ set_irn_n((ir_node*)remat->op, n, bad);
}
}
}
static void
delete_unnecessary_remats(spill_ilp_t * si)
{
-#ifdef KEEPALIVE
- int n;
- ir_node *bad = get_irg_bad(si->chordal_env->irg);
+ if(opt_keep_alive & KEEPALIVE_REMATS) {
+ int n;
+ ir_node *bad = get_irg_bad(si->chordal_env->irg);
- if(si->keep) {
- ir_node *end = get_irg_end(si->chordal_env->irg);
- ir_node **keeps;
+ if(si->keep) {
+// ir_node *end = get_irg_end(si->chordal_env->irg);
+// ir_node **keeps;
- for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
- ir_node *keep_arg = get_irn_n(si->keep, n);
- op_t *arg_op = get_irn_link(keep_arg);
- lpp_name_t *name;
+ for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
+ ir_node *keep_arg = get_irn_n(si->keep, n);
+ op_t *arg_op = get_irn_link(keep_arg);
+ lpp_name_t *name;
- assert(arg_op->is_remat);
+ assert(arg_op->is_remat);
- name = si->lpp->vars[arg_op->attr.remat.ilp];
+ name = si->lpp->vars[arg_op->attr.remat.ilp];
- if(is_zero(name->value)) {
- DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
- /* TODO check whether reload is preferred over remat (could be bug) */
- delete_remat(si, keep_arg);
- } else {
- if(!arg_op->attr.remat.remat->inverse) {
- if(arg_op->attr.remat.pre) {
- DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
- } else {
- DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
- }
+ if(is_zero(name->value)) {
+ DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
+ /* TODO check whether reload is preferred over remat (could be bug) */
+ delete_remat(si, keep_arg);
} else {
- if(arg_op->attr.remat.pre) {
- DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
+ if(!arg_op->attr.remat.remat->inverse) {
+ if(arg_op->attr.remat.pre) {
+ DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
+ } else {
+ DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
+ }
} else {
- DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
+ if(arg_op->attr.remat.pre) {
+ DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
+ } else {
+ DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
+ }
}
}
- }
- set_irn_n(si->keep, n, bad);
- }
+ set_irn_n(si->keep, n, bad);
+ }
#if 0
- for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
- ir_node *end_arg = get_End_keepalive(end, i);
+ for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
+ ir_node *end_arg = get_End_keepalive(end, i);
- if(end_arg != si->keep) {
- obstack_grow(si->obst, &end_arg, sizeof(end_arg));
+ if(end_arg != si->keep) {
+ obstack_grow(si->obst, &end_arg, sizeof(end_arg));
+ }
}
- }
- keeps = obstack_finish(si->obst);
- set_End_keepalives(end, n-1, keeps);
- obstack_free(si->obst, keeps);
+ keeps = obstack_finish(si->obst);
+ set_End_keepalives(end, n-1, keeps);
+ obstack_free(si->obst, keeps);
#endif
+ } else {
+ DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
+ }
} else {
- DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
- }
-#else
- ir_node *remat;
+ ir_node *remat;
- pset_foreach(si->all_possible_remats, remat) {
- op_t *remat_op = get_irn_link(remat);
- lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
+ pset_foreach(si->all_possible_remats, remat) {
+ op_t *remat_op = get_irn_link(remat);
+ lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
- if(is_zero(name->value)) {
- DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
- /* TODO check whether reload is preferred over remat (could be bug) */
- delete_remat(si, remat);
- } else {
- if(!remat_op->attr.remat.remat->inverse) {
- if(remat_op->attr.remat.pre) {
- DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
- } else {
- DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
- }
+ if(is_zero(name->value)) {
+ DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
+ /* TODO check whether reload is preferred over remat (could be bug) */
+ delete_remat(si, remat);
} else {
- if(remat_op->attr.remat.pre) {
- DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
+ if(!remat_op->attr.remat.remat->inverse) {
+ if(remat_op->attr.remat.pre) {
+ DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
+ } else {
+ DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
+ }
} else {
- DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
+ if(remat_op->attr.remat.pre) {
+ DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
+ } else {
+ DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
+ }
}
}
}
}
-#endif
}
static pset *
-get_spills_for_value(spill_ilp_t * si, ir_node * value)
+get_spills_for_value(spill_ilp_t * si, const ir_node * value)
{
pset *spills = pset_new_ptr_default();
- ir_node *next;
- defs_t *defs;
+ const ir_node *next;
+ defs_t *defs;
defs = set_find_def(si->values, value);
return spills;
}
-static pset *
-get_remats_for_value(spill_ilp_t * si, ir_node * value)
-{
- pset *remats = pset_new_ptr_default();
-
- ir_node *next;
- defs_t *defs;
-
- pset_insert_ptr(remats, value);
- defs = set_find_def(si->values, value);
-
- if(defs && defs->remats) {
- for(next = defs->remats; next; next = get_irn_link(next)) {
- pset_insert_ptr(remats, next);
- }
- }
-
- return remats;
-}
-
-
/**
* @param before The node after which the spill will be placed in the schedule
*/
-/* TODO set context properly */
static ir_node *
-insert_spill(spill_ilp_t * si, ir_node * irn, ir_node * value, ir_node * before)
+insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
{
defs_t *defs;
ir_node *spill;
DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
- spill = be_spill2(arch_env, irn, before, irn);
+ spill = be_spill2(arch_env, irn, before);
defs = set_insert_def(si->values, value);
assert(defs);
set_irn_link(spill, defs->spills);
defs->spills = spill;
-#ifdef KEEPALIVE_SPILLS
- pset_insert_ptr(si->spills, spill);
-#endif
+ if(opt_keep_alive & KEEPALIVE_SPILLS)
+ pset_insert_ptr(si->spills, spill);
return spill;
}
* @param before The Phi node which has to be spilled
*/
static ir_node *
-insert_mem_phi(spill_ilp_t * si, const ir_node * phi)
+insert_mem_phi(spill_ilp_t * si, ir_node * phi)
{
ir_node *mem_phi;
ir_node **ins;
defs_t *defs;
int n;
- op_t *op = get_irn_link(phi);
NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
set_irn_link(mem_phi, defs->spills);
defs->spills = mem_phi;
+#ifdef SCHEDULE_PHIM
sched_add_after(phi, mem_phi);
-
-#ifdef KEEPALIVE_SPILLS
- pset_insert_ptr(si->spills, mem_phi);
#endif
+ if(opt_keep_alive & KEEPALIVE_SPILLS)
+ pset_insert_ptr(si->spills, mem_phi);
+
return mem_phi;
}
* Add reload before operation and add to list of defs
*/
static ir_node *
-insert_reload(spill_ilp_t * si, const ir_node * value, const ir_node * after)
+insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
{
defs_t *defs;
ir_node *reload,
return reload;
}
+void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
+{
+ defs_t *defs;
+ ir_node *reload;
+ ir_node *value = get_irn_n(memoperand->irn, memoperand->pos);
+ ir_node *spill;
+ const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
+
+ DBG((si->dbg, LEVEL_2, "\t inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
+
+ defs = set_find_def(si->values, value);
+
+ spill = defs->spills;
+ assert(spill && "no spill placed before reload");
+
+ reload = be_reload(arch_env, si->cls, memoperand->irn, get_irn_mode(value), spill);
+
+ arch_perform_memory_operand(arch_env, memoperand->irn, reload, memoperand->pos);
+ sched_remove(reload);
+}
+
+void insert_memoperands(spill_ilp_t * si)
+{
+ memoperand_t *memoperand;
+ lpp_name_t *name;
+
+ set_foreach(si->memoperands, memoperand) {
+ name = si->lpp->vars[memoperand->ilp];
+ if(!is_zero(name->value)) {
+ perform_memory_operand(si, memoperand);
+ }
+ }
+}
+
static void
walker_spill_placer(ir_node * bb, void * data) {
spill_ilp_t *si = (spill_ilp_t*)data;
}
static ir_node *
-insert_mem_copy(spill_ilp_t * si, const ir_node * bb, const ir_node * value)
+insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
{
ir_node *insert_pos = bb;
ir_node *spill;
DBG((si->dbg, LEVEL_2, "\t inserting mem copy for value %+F after %+F\n", value, insert_pos));
- spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos, value);
+ spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos);
return spill;
}
if(!phi_m) continue;
for(n=get_irn_arity(phi)-1; n>=0; --n) {
- const ir_node *value = get_irn_n(phi, n);
+ ir_node *value = get_irn_n(phi, n);
defs_t *val_defs = set_find_def(si->values, value);
- ir_node *arg = get_irn_n(phi_m, n);
/* a spill of this value */
ir_node *spill;
-#ifndef NO_MEMCOPIES
- ir_node *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
- lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
+ if(opt_memcopies) {
+ ir_node *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
+ lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
- if(!is_zero(name->value)) {
- spill = insert_mem_copy(si, pred, value);
+ if(!is_zero(name->value)) {
+ spill = insert_mem_copy(si, pred, value);
+ } else {
+ spill = val_defs->spills;
+ }
} else {
spill = val_defs->spills;
}
-#else
- spill = val_defs->spills;
-#endif
+
assert(spill && "no spill placed before PhiM");
set_irn_n(phi_m, n, spill);
}
spill_ilp_t *si = (spill_ilp_t*)data;
ir_node *irn;
spill_bb_t *spill_bb = get_irn_link(bb);
- int i;
/* reloads at end of block */
if(spill_bb->reloads) {
reload = insert_reload(si, irn, insert_pos);
-#ifdef KEEPALIVE_RELOADS
- pset_insert_ptr(si->spills, reload);
-#endif
+ if(opt_keep_alive & KEEPALIVE_RELOADS)
+ pset_insert_ptr(si->spills, reload);
}
}
}
set_irn_n(irn, n, reload);
-#ifdef KEEPALIVE_RELOADS
- pset_insert_ptr(si->spills, reload);
-#endif
+ if(opt_keep_alive & KEEPALIVE_RELOADS)
+ pset_insert_ptr(si->spills, reload);
}
}
}
walker_kill_unused(ir_node * bb, void * data)
{
struct kill_helper *kh = data;
- const ir_node *bad = get_irg_bad(get_irn_irg(bb));
+ ir_node *bad = get_irg_bad(get_irn_irg(bb));
ir_node *irn;
lc_bitset_free(kh.used);
}
-static void
+void
print_irn_pset(pset * p)
{
ir_node *irn;
}
}
-static void
+void
dump_phi_class(spill_ilp_t * si, pset * phiclass, const char * file)
{
FILE *f = fopen(file, "w");
/* then fix uses of spills */
set_foreach(si->values, defs) {
- pset *reloads;
- pset *spills;
- ir_node *next = defs->remats;
+ pset *reloads;
+ pset *spills;
+ const ir_node *next = defs->remats;
int remats = 0;
reloads = pset_new_ptr_default();
/* first fix uses of remats and reloads */
set_foreach(si->values, defs) {
- pset *nodes;
- ir_node *next = defs->remats;
+ pset *nodes;
+ const ir_node *next = defs->remats;
if(next) {
nodes = pset_new_ptr_default();
si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
+ if(opt_memoperands)
+ insert_memoperands(si);
phim_fixer(si);
/* clean the remat info! there are still back-edges leading there! */
if( sched_is_end(irn) ||
(be_is_Reload(irn) && has_reg_class(si, irn)) ||
/* do not move reload before its spill */
- (irn == be_get_Reload_mem(reload)) ) break;
+ (irn == be_get_Reload_mem(reload)) ||
+ /* do not move before phi */
+ is_Phi(irn)) break;
set_irn_link(irn, INT_TO_PTR(pressure+1));
DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
}
-static void
-walker_spillslotassigner(ir_node * irn, void * data)
-{
- spill_ilp_t *si = (spill_ilp_t*)data;
- void *cls;
-
- if(!be_is_Spill(irn)) return;
-
- /* set spill context to phi class if it has one ;) */
-
- cls = get_phi_class(irn);
- if(cls)
- be_set_Spill_context(irn, cls);
- else
- be_set_Spill_context(irn, irn);
-}
-
-
-static void
-assign_spillslots(spill_ilp_t * si)
-{
- DBG((si->dbg, LEVEL_2, "\t calling spill slot assigner\n"));
- irg_walk_graph(si->chordal_env->irg, walker_spillslotassigner, NULL, si);
-}
-
void
be_spill_remat(const be_chordal_env_t * chordal_env)
{
+ char buf[256];
char problem_name[256];
char dump_suffix[256];
char dump_suffix2[256];
FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
-#ifdef VERIFY_DOMINANCE
- be_check_dominance(chordal_env->irg);
-#endif
+ if(opt_verify & VERIFY_DOMINANCE)
+ be_check_dominance(chordal_env->irg);
obstack_init(&obst);
si.chordal_env = chordal_env;
si.lpp = new_lpp(problem_name, lpp_minimize);
si.remat_info = new_set(cmp_remat_info, 4096);
si.interferences = new_set(cmp_interference, 32);
+ si.memoperands = new_set(cmp_memoperands, 128);
si.all_possible_remats = pset_new_ptr_default();
si.spills = pset_new_ptr_default();
si.inverse_ops = pset_new_ptr_default();
si.lv = chordal_env->lv;
-#ifdef KEEPALIVE
si.keep = NULL;
-#endif
si.n_regs = get_n_regs(&si);
set_irg_link(chordal_env->irg, &si);
be_analyze_regpressure(chordal_env, "-pre");
-#ifdef COLLECT_REMATS
- /* collect remats */
- DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
- irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
-#endif
+ if(opt_remats) {
+ /* collect remats */
+ DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
+ irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
+ }
/* insert possible remats */
DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
-#ifdef KEEPALIVE
- DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
- connect_all_remats_with_keep(&si);
- /* dump graph with inserted remats */
- dump_graph_with_remats(chordal_env->irg, dump_suffix);
-#endif
+ if(opt_keep_alive & KEEPALIVE_REMATS) {
+ DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
+ connect_all_remats_with_keep(&si);
+ /* dump graph with inserted remats */
+ dump_graph_with_remats(chordal_env->irg, dump_suffix);
+ }
/* insert copies for phi arguments not in my regclass */
irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
-#ifndef NO_MEMCOPIES
- DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
- memcopyhandler(&si);
-#endif
+ if(opt_memcopies) {
+ DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
+ memcopyhandler(&si);
+ }
-#ifdef DUMP_ILP
- {
+ if(opt_dump_flags & DUMP_PROBLEM) {
FILE *f;
- char buf[256];
-
ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
if ((f = fopen(buf, "wt")) != NULL) {
lpp_dump_plain(si.lpp, f);
fclose(f);
}
}
-#endif
+
+ if(opt_dump_flags & DUMP_MPS) {
+ FILE *f;
+
+ ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name);
+ if((f = fopen(buf, "wt")) != NULL) {
+ mps_write_mps(si.lpp, s_mps_fixed, f);
+ fclose(f);
+ }
+
+ ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name);
+ if((f = fopen(buf, "wt")) != NULL) {
+ mps_write_mst(si.lpp, s_mps_fixed, f);
+ fclose(f);
+ }
+ }
+
+ lpp_check_startvals(si.lpp);
#ifdef SOLVE
- DBG((si.dbg, LEVEL_1, "\tSolving %F\n", chordal_env->irg));
-#ifdef ILP_TIMEOUT
- lpp_set_time_limit(si.lpp, ILP_TIMEOUT);
-#endif
+ DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
+ lpp_set_time_limit(si.lpp, opt_timeout);
+
+ if(opt_log)
+ lpp_set_log(si.lpp, stdout);
#ifdef SOLVE_LOCAL
lpp_solve_cplex(si.lpp);
DBG((si.dbg, LEVEL_1, "\t%s: iterations: %d, solution time: %g, objective function: %g\n", problem_name, si.lpp->iterations, si.lpp->sol_time, is_zero(si.lpp->objval)?0.0:si.lpp->objval));
-#ifdef DUMP_SOLUTION
- {
+ if(opt_dump_flags & DUMP_SOLUTION) {
FILE *f;
char buf[256];
fclose(f);
}
}
-#endif
writeback_results(&si);
kill_all_unused_values_in_schedule(&si);
-#if defined(KEEPALIVE_SPILLS) || defined(KEEPALIVE_RELOADS)
- be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
-#endif
+ if(opt_keep_alive & (KEEPALIVE_SPILLS | KEEPALIVE_RELOADS))
+ be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
// move reloads upwards
be_liveness_recompute(si.lv);
irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
move_reloads_upward(&si);
-#ifndef NO_MEMCOPIES
- verify_phiclasses(&si);
- assign_spillslots(&si);
-#endif
+ if(opt_memcopies) {
+ verify_phiclasses(&si);
+ }
irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
be_analyze_regpressure(chordal_env, "-post");
-#ifdef VERIFY_DOMINANCE
- be_check_dominance(chordal_env->irg);
-#endif
+ if(opt_verify & VERIFY_DOMINANCE)
+ be_check_dominance(chordal_env->irg);
free_dom(chordal_env->irg);
del_set(si.interferences);
del_pset(si.inverse_ops);
del_pset(si.all_possible_remats);
+ del_set(si.memoperands);
del_pset(si.spills);
free_lpp(si.lpp);
obstack_free(&obst, NULL);
projects / libfirm / blobdiff