#include "phiclass_t.h"
#include "iredges.h"
#include "execfreq.h"
+#include "irvrfy.h"
#include <lpp/lpp.h>
#include <lpp/lpp_net.h>
//#define KEEPALIVE_RELOADS
#define GOODWIN_REDUCTION
//#define NO_MEMCOPIES
+//#define VERIFY_DOMINANCE
+#define WITH_MEMOPERANDS
#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 120
+#define COST_LOAD 8
+#define COST_MEMOPERAND 7
+#define COST_STORE 50
+#define COST_REMAT 1
+#define ILP_TIMEOUT 300
+#define MAX_PATHS 16
#define ILP_UNDEF -1
typedef struct _spill_ilp_t {
const arch_register_class_t *cls;
int n_regs;
const be_chordal_env_t *chordal_env;
+ be_lv_t *lv;
lpp_t *lpp;
struct obstack *obst;
set *remat_info;
pset *spills;
set *interferences;
ir_node *m_unknown;
+#ifdef WITH_MEMOPERANDS
+ set *memoperands;
+#endif
DEBUG_ONLY(firm_dbg_module_t * dbg);
} spill_ilp_t;
ilp_var_t spill;
} spill_t;
+#ifdef WITH_MEMOPERANDS
+typedef struct _memoperand_t {
+ const 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;
+#endif
+
static INLINE int
has_reg_class(const spill_ilp_t * si, const ir_node * irn)
{
return !(p->irn == q->irn);
}
+#ifdef WITH_MEMOPERANDS
+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);
+}
+#endif
+
static keyval_t *
set_find_keyval(set * set, void * key)
{
return set_insert(set, &query, sizeof(query), HASH_PTR(value));
}
+#ifdef WITH_MEMOPERANDS
+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, ir_node * irn, unsigned int pos)
+{
+ memoperand_t query;
+
+ query.irn = irn;
+ query.pos = pos;
+ return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos);
+}
+#endif
+
+
static spill_t *
set_find_spill(set * set, ir_node * value)
{
} else {
return arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, irn);
}
-
}
/**
spill_ilp_t *si = data;
spill_bb_t *spill_bb;
ir_node *irn;
- int n;
- irn_live_t *li;
+ int n, i;
pset *live = pset_new_ptr_default();
DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
- live_foreach(bb, li) {
- ir_node *value = (ir_node *) li->irn;
+ be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
+ ir_node *value = be_lv_get_irn(si->lv, bb, i);
/* add remats at end of block */
- if (live_is_end(li) && has_reg_class(si, value)) {
+ if (has_reg_class(si, value)) {
pset_insert_ptr(live, value);
}
}
irn = next;
}
- live_foreach(bb, li) {
- ir_node *value = (ir_node *) li->irn;
+ 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)) {
/* add remats at end of block */
- if (live_is_end(li) && has_reg_class(si, value)) {
+ if (be_is_live_end(si->lv, bb, value) && has_reg_class(si, value)) {
remat_info_t *remat_info,
query;
remat_t *remat;
}
if(is_diverge_edge(bb)) {
/* add remat2s at beginning of block */
- if ((live_is_in(li) || (is_Phi(value) && get_nodes_block(value)==bb)) && has_reg_class(si, 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;
remat_t *remat;
luke_endwalker(ir_node * bb, void * data)
{
spill_ilp_t *si = (spill_ilp_t*)data;
- irn_live_t *li;
pset *live;
pset *use_end;
char buf[256];
ilp_cst_t cst;
ir_node *irn;
spill_bb_t *spill_bb = get_irn_link(bb);
+ int i;
live = pset_new_ptr_default();
use_end = pset_new_ptr_default();
- live_foreach(bb, li) {
- irn = (ir_node *) li->irn;
- if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
+ be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
+ irn = be_lv_get_irn(si->lv, bb, i);
+ if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
op_t *op;
pset_insert_ptr(live, irn);
}
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));
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;
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);
}
static void
get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
{
- irn_live_t *li;
ir_node *irn;
+ int i;
- live_foreach(bb, li) {
- irn = (ir_node *) li->irn;
+ be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
+ irn = be_lv_get_irn(si->lv, bb, i);
- if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
+ if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
pset_insert_ptr(live, irn);
}
}
copyreg = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
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", block, to_copy);
- cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
+ ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
+ 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", block, to_copy);
- cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
+ ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
+ 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();
+#ifdef WITH_MEMOPERANDS
+ const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
+#endif
live = pset_new_ptr_default();
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);
pset *used;
pset *remat_defs;
keyval_t *keyval;
+#ifdef WITH_MEMOPERANDS
+ ilp_cst_t one_memoperand;
+#endif
/* iterate only until first phi */
if(is_Phi(irn))
**********************************/
/* ensure each dying value is used by only one post remat */
- pset_foreach(live, tmp) {
+ pset_foreach(used, tmp) {
ir_node *value = tmp;
op_t *value_op = get_irn_link(value);
ir_node *remat;
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++;
}
}
- if(value_op->attr.live_range.ilp != ILP_UNDEF && cst != ILP_UNDEF) {
+ // value_op->attr.live_range.ilp != ILP_UNDEF
+ if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
}
}
+ /* ensure at least one value dies at post remat */
+ foreach_post_remat(irn, tmp) {
+ op_t *remat_op = get_irn_link(tmp);
+ pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
+ ir_node *remat_arg;
+
+ for(n=get_irn_arity(tmp)-1; n>=0; --n) {
+ remat_arg = get_irn_n(tmp, n);
+
+ if(has_reg_class(si, remat_arg)) {
+
+ /* does arg always die at this op? */
+ if(!pset_find_ptr(live, remat_arg))
+ goto skip_one_must_die;
+
+ pset_insert_ptr(remat_args, remat_arg);
+ }
+ }
+ /* remat + \sum live_range(remat_arg) <= |args| */
+ ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
+ 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) {
+ op_t *arg_op = get_irn_link(remat_arg);
+
+ lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
+ }
+
+skip_one_must_die:
+ del_pset(remat_args);
+ }
/* new live ranges for values from L\U defined by post remats */
pset_foreach(live, tmp) {
/* 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);
/* 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) {
/* 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);
- 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);
+ 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);
}
}
}
+#ifdef WITH_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(si->lpp, buf, lpp_binary, COST_MEMOPERAND*execution_frequency(si, bb));
+ 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);
+// if(arg_op->attr.live_range.ilp != ILP_UNDEF)
+// lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
+ }
+ }
+#endif
+
/* new live range begins for each used value */
arg_op->attr.live_range.ilp = prev_lr;
arg_op->attr.live_range.op = irn;
/* 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);
}
+#ifdef WITH_MEMOPERANDS
+ ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
+ one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
+#endif
+
/***********************************************************
* 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;
+ ir_node *arg = keyval->key;
+ int i = PTR_TO_INT(keyval->val);
+ op_t *arg_op = get_irn_link(arg);
+ ilp_cst_t requirements;
+#ifdef WITH_MEMOPERANDS
+ int n_memoperands;
+#endif
spill = set_find_spill(spill_bb->ilp, arg);
assert(spill);
/* 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);
}
}
+
+#ifdef WITH_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);
+ }
+ }
+#endif
}
/* iterate over L\U */
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);
/* 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);
}
* including remats
*/
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;
/* 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);
ir_node *bb_p = get_Block_cfgpred_block(bb, n);
spill_bb_t *spill_bb_p = get_irn_link(bb_p);
spill_t *spill_p;
+ op_t *op = get_irn_link(irn);
/* although the phi is in the right regclass one or more of
* its arguments can be in a different one or at least to
* ignore
*/
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);
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);
assert(spill && spill->irn == irn);
ir_snprintf(buf, sizeof(buf), "first_lr_%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, op->attr.live_range.ilp, 1.0);
lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -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);
ir_node *value = op->attr.remat.remat->value;
if(value == irn) {
- /* only collect remats up to the first use of a value */
+ /* only collect remats up to the first real use of a value */
lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
}
} else {
if(arg == irn) {
/* if a value is used stop collecting remats */
- cst = ILP_UNDEF;
+ goto next_live;
}
- break;
}
}
- if(cst == ILP_UNDEF) break;
}
}
+ /* if the semicolon is missing, the microsoft compiler complains */
+next_live: ;
}
del_pset(live);
}
static int
-values_interfere_in_block(ir_node * bb, ir_node * a, ir_node * b)
+values_interfere_in_block(const spill_ilp_t * si, ir_node * bb, ir_node * a, ir_node * b)
{
const ir_edge_t *edge;
/* the following code is stolen from bera.c */
- if(is_live_end(bb, a))
+ if(be_is_live_end(si->lv, bb, a))
return 1;
foreach_out_edge(a, edge) {
luke_interferencewalker(ir_node * bb, void * data)
{
spill_ilp_t *si = (spill_ilp_t*)data;
- irn_live_t *li1,
- *li2;
+ int l1, l2;
- live_foreach(bb, li1) {
- ir_node *a = (ir_node *) li1->irn;
+ be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
+ 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)) {
- for(li2=li1->next; li2; li2 = li2->next) {
- ir_node *b = (ir_node *) li2->irn;
+ 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(values_interfere_in_block(bb, a, b)) {
+ 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 copies contains a path of copies which lead us to irn
* @parameter visited contains a set of nodes already visited on this path
*/
-static void
+static int
find_copy_path(spill_ilp_t * si, ir_node * irn, 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;
if(op->is_remat) return;
if(is_Phi(irn)) {
int n;
+ pset *visited_operands = pset_new_ptr(get_irn_arity(irn));
/* visit all operands */
for(n=get_irn_arity(irn)-1; n>=0; --n) {
ilp_var_t copy = op->attr.live_range.args.copies[n];
if(!has_reg_class(si, arg)) continue;
+ if(pset_find_ptr(visited_operands, arg)) continue;
+ pset_insert_ptr(visited_operands, arg);
if(arg == target) {
+ if(++paths > MAX_PATHS && pset_count(copies) != 0) {
+ del_pset(visited_operands);
+ del_pset(visited_users);
+ pset_remove_ptr(visited, irn);
+ return paths;
+ }
pset_insert(copies, INT_TO_PTR(copy), copy);
write_copy_path_cst(si, copies, any_interfere);
pset_remove(copies, INT_TO_PTR(copy), copy);
- } else {
- if(!pset_find_ptr(visited, arg)) {
- pset_insert(copies, INT_TO_PTR(copy), copy);
- find_copy_path(si, arg, target, any_interfere, copies, visited);
- pset_remove(copies, INT_TO_PTR(copy), copy);
- }
+ } else if(!pset_find_ptr(visited, arg)) {
+ pset_insert(copies, INT_TO_PTR(copy), copy);
+ paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
+ pset_remove(copies, INT_TO_PTR(copy), copy);
+
+ /*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_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));
+
+ paths = 0;
+ } else {
+ del_pset(visited_operands);
+ del_pset(visited_users);
+ pset_remove_ptr(visited, irn);
+ return paths;
+ }
+ } else if(pset_count(copies) == 0) {
+ paths = 0;
+ }*/
}
}
+
+ del_pset(visited_operands);
}
/* visit all uses which are phis */
if(!is_Phi(user)) continue;
if(!has_reg_class(si, user)) continue;
+ if(pset_find_ptr(visited_users, user)) continue;
+ pset_insert_ptr(visited_users, user);
copy = op->attr.live_range.args.copies[pos];
if(user == target) {
+ if(++paths > MAX_PATHS && pset_count(copies) != 0) {
+ del_pset(visited_users);
+ pset_remove_ptr(visited, irn);
+ return paths;
+ }
pset_insert(copies, INT_TO_PTR(copy), copy);
write_copy_path_cst(si, copies, any_interfere);
pset_remove(copies, INT_TO_PTR(copy), copy);
- } else {
- if(!pset_find_ptr(visited, user)) {
- pset_insert(copies, INT_TO_PTR(copy), copy);
- find_copy_path(si, user, target, any_interfere, copies, visited);
- pset_remove(copies, INT_TO_PTR(copy), copy);
- }
+ } else if(!pset_find_ptr(visited, user)) {
+ pset_insert(copies, INT_TO_PTR(copy), copy);
+ paths += find_copy_path(si, user, target, any_interfere, copies, visited);
+ pset_remove(copies, INT_TO_PTR(copy), copy);
+
+ /*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_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));
+
+ paths = 0;
+ } else {
+ del_pset(visited_users);
+ pset_remove_ptr(visited, irn);
+ return paths;
+ }
+ } else if(pset_count(copies) == 0) {
+ paths = 0;
+ }*/
}
}
+ del_pset(visited_users);
pset_remove_ptr(visited, irn);
+ return paths;
}
static void
DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
-// phi_class_free(si->chordal_env->irg);
-
/* now lets emit the ILP unequations for the crap */
set_foreach(si->interferences, interference) {
irnlist_t *irnlist;
/* 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_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
any_interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
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);
{
spill_ilp_t *si = data;
ir_node *irn;
- irn_live_t *li;
- int n;
+ int n, i;
pset *live = pset_new_ptr_default();
int projs = 0;
- live_foreach(bb, li) {
- irn = (ir_node *) li->irn;
+ be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
+ irn = be_lv_get_irn(si->lv, bb, i);
- if (live_is_end(li) && has_reg_class(si, irn)) {
+ if (has_reg_class(si, irn)) {
pset_insert_ptr(live, irn);
}
}
return reload;
}
+#ifdef WITH_MEMOPERANDS
+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);
+ }
+ }
+}
+#endif
+
static void
walker_spill_placer(ir_node * bb, void * data) {
spill_ilp_t *si = (spill_ilp_t*)data;
defs_t *val_defs = set_find_def(si->values, value);
ir_node *arg = get_irn_n(phi_m, n);
- /* get a spill of this value */
- ir_node *spill = val_defs->spills;
+ /* a spill of this value */
+ ir_node *spill;
#ifndef NO_MEMCOPIES
if(!is_zero(name->value)) {
spill = insert_mem_copy(si, pred, value);
} else {
- assert(spill && "no spill placed before PhiM");
+ spill = val_defs->spills;
}
#else
- assert(spill && "no spill placed before PhiM");
+ spill = val_defs->spills;
#endif
+ assert(spill && "no spill placed before PhiM");
set_irn_n(phi_m, n, spill);
}
}
ir_node *irn;
spill_bb_t *spill_bb = get_irn_link(bb);
int i;
- irn_live_t *li;
/* reloads at end of block */
if(spill_bb->reloads) {
}
}
+static void
+dump_phi_class(spill_ilp_t * si, pset * phiclass, const char * file)
+{
+ FILE *f = fopen(file, "w");
+ ir_node *irn;
+ interference_t *interference;
+
+ pset_break(phiclass);
+ set_break(si->interferences);
+
+ ir_fprintf(f, "digraph phiclass {\n");
+
+ pset_foreach(phiclass, irn) {
+ if(is_Phi(irn))
+ ir_fprintf(f, " %F%N [shape=box]\n",irn,irn);
+ }
+
+ pset_foreach(phiclass, irn) {
+ int n;
+
+ if(!is_Phi(irn)) continue;
+
+ for(n=get_irn_arity(irn)-1; n>=0; --n) {
+ ir_node *arg = get_irn_n(irn, n);
+
+ ir_fprintf(f, " %F%N -> %F%N\n",irn,irn,arg,arg);
+ }
+ }
+
+ set_foreach(si->interferences, interference) {
+ const ir_node *a = interference->a;
+ const ir_node *b = interference->b;
+ if(get_phi_class(a) == phiclass) {
+ ir_fprintf(f, " %F%N -> %F%N [color=red,dir=none,style=bold]\n",a,a,b,b);
+ }
+ }
+
+ ir_fprintf(f, "}");
+ fclose(f);
+}
+
static void
rewire_uses(spill_ilp_t * si)
{
// print_irn_pset(spills);
// print_irn_pset(reloads);
- be_ssa_constr_set_ignore(dfi, spills, ignore);
+ be_ssa_constr_set_ignore(dfi, si->lv, spills, ignore);
}
del_pset(reloads);
if(pset_count(nodes) > 1) {
DBG((si->dbg, LEVEL_4, "\t %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
- be_ssa_constr_set(dfi, nodes);
+ be_ssa_constr_set(dfi, si->lv, nodes);
}
del_pset(nodes);
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);
+#ifdef WITH_MEMOPERANDS
+ insert_memoperands(si);
+#endif
phim_fixer(si);
/* clean the remat info! there are still back-edges leading there! */
pressure = (int)get_irn_link(irn);
while(pressure < si->n_regs) {
- if(sched_is_end(irn) || (be_is_Reload(irn) && has_reg_class(si, irn))) break;
+ 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;
set_irn_link(irn, INT_TO_PTR(pressure+1));
DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
luke_meminterferencechecker(ir_node * bb, void * data)
{
spill_ilp_t *si = (spill_ilp_t*)data;
- irn_live_t *li1,
- *li2;
+ int l1, l2;
- live_foreach(bb, li1) {
- ir_node *a = (ir_node *) li1->irn;
+ be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
+ ir_node *a = be_lv_get_irn(si->lv, bb, l1);
if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
- /* a is only interesting if it is inside a phi class */
- if (get_phi_class(a)) {
- for(li2=li1->next; li2; li2 = li2->next) {
- ir_node *b = (ir_node *) li2->irn;
+ /* 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)) {
+ 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);
if(!be_is_Spill(b) && (!is_Phi(b) || get_irn_mode(b) != mode_T)) continue;
/* a and b are only interesting if they are in the same phi class */
- if(get_phi_class(a) == get_phi_class(b)) {
- if(values_interfere_in_block(bb, a, b)) {
+ if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
+ if(values_interfere_in_block(si, bb, a, b)) {
ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
}
}
char problem_name[256];
char dump_suffix[256];
char dump_suffix2[256];
- char dump_suffix3[256];
struct obstack obst;
spill_ilp_t si;
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
+
obstack_init(&obst);
si.chordal_env = chordal_env;
si.obst = &obst;
si.lpp = new_lpp(problem_name, lpp_minimize);
si.remat_info = new_set(cmp_remat_info, 4096);
si.interferences = new_set(cmp_interference, 32);
+#ifdef WITH_MEMOPERANDS
+ si.memoperands = new_set(cmp_memoperands, 128);
+#endif
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
/* recompute liveness */
DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
- be_liveness(chordal_env->irg);
+ be_liveness_recompute(si.lv);
/* build the ILP */
#endif
// move reloads upwards
- be_liveness(chordal_env->irg);
+ be_liveness_recompute(si.lv);
irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
move_reloads_upward(&si);
be_analyze_regpressure(chordal_env, "-post");
+#ifdef VERIFY_DOMINANCE
+ be_check_dominance(chordal_env->irg);
+#endif
+
free_dom(chordal_env->irg);
del_set(si.interferences);
del_pset(si.inverse_ops);
del_pset(si.all_possible_remats);
+#ifdef WITH_MEMOPERANDS
+ del_set(si.memoperands);
+#endif
del_pset(si.spills);
free_lpp(si.lpp);
obstack_free(&obst, NULL);
projects / libfirm / blobdiff