#include "irnode_t.h"
#include "ircons_t.h"
#include "irloop_t.h"
-#include "phiclass.h"
+#include "phiclass_t.h"
#include "iredges.h"
#include "execfreq.h"
+#include "irvrfy.h"
#include <lpp/lpp.h>
#include <lpp/lpp_net.h>
#include "bechordal_t.h"
-#define BIGM 100000.0
-
-#define DUMP_SOLUTION
-#define DUMP_ILP
+//#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 ONLY_BRIGGS_REMATS /* only remats without parameters (or only with ignored params) */
#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_PRE_REMAT /* allow values to die after a pre remat */
-#define CHECK_POST_REMAT /* check pressure after post remats (conservative but otherwise we can temporarily exceed the register pressure) */
+#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 NO_MEMCOPIES
+//#define VERIFY_DOMINANCE
+#define WITH_MEMOPERANDS
#define SOLVE
//#define SOLVE_LOCAL
#define LPP_SERVER "i44pc52"
#define LPP_SOLVER "cplex"
-#define COST_LOAD 10
-#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;
ir_node *keep;
#endif
set *values; /**< for collecting all definitions of values before running ssa-construction */
- set *execfreqs;
pset *spills;
set *interferences;
ir_node *m_unknown;
+#ifdef WITH_MEMOPERANDS
+ set *memoperands;
+#endif
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;
+#ifdef WITH_MEMOPERANDS
+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;
+#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)
+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));
}
+#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, 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);
+}
+#endif
+
+
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));
}
}
static double
-execution_frequency(const spill_ilp_t * si, const ir_node * irn)
+execution_frequency(const spill_ilp_t *si, const ir_node * irn)
{
#define FUDGE 0.001
- if(si->execfreqs) {
- if(is_Block(irn)) {
- return get_block_execfreq(si->execfreqs, irn) + FUDGE;
- } else {
- return get_block_execfreq(si->execfreqs, get_nodes_block(irn)) + FUDGE;
- }
- } else {
- if(is_Block(irn))
- return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
- else
- return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
- }
+#ifndef EXECFREQ_LOOPDEPH
+ return get_block_execfreq(si->chordal_env->exec_freq, get_block(irn)) + FUDGE;
+#else
+ if(is_Block(irn))
+ return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
+ else
+ return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
+#endif
}
static double
} else {
return arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, irn);
}
-
}
/**
static INLINE int
is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
{
- int i,
- n;
+ int n;
const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
int remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
ir_fprintf(stderr, " Node %+F is not rematerializable\n", irn);
#endif
- for (i = 0, n = get_irn_arity(irn); i < n && remat; ++i) {
- ir_node *op = get_irn_n(irn, i);
+ for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
+ ir_node *op = get_irn_n(irn, n);
remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
// if(!remat)
remat->inverse = 0;
} else {
arch_inverse_t inverse;
- int i,
- n;
+ int n;
/* get the index of the operand we want to retrieve by the inverse op */
- for (i = 0, n = get_irn_arity(op); i < n; ++i) {
- ir_node *arg = get_irn_n(op, i);
+ for (n = get_irn_arity(op)-1; n>=0; --n) {
+ ir_node *arg = get_irn_n(op, n);
if(arg == dest_value) break;
}
- if(i == n) return NULL;
+ if(n<0) return NULL;
- DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", i, op));
+ DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", n, op));
/* else ask the backend to give an inverse op */
- if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, i, &inverse, si->obst)) {
+ if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
int i;
DBG((si->dbg, LEVEL_4, "\t backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
assert(inverse.n > 0 && "inverse op should have at least one node");
- for(i=0; i<inverse.n; ++i) {
+ for(i=inverse.n-1; i>=0; --i) {
pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
}
{
remat_info_t *remat_info,
query;
- int i,
- n;
+ int n;
assert(remat->op);
assert(remat->value);
pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
/* insert the remat into the remats_be_operand set of each argument of the remat op */
- for (i = 0, n = get_irn_arity(remat->op); i < n; ++i) {
- ir_node *arg = get_irn_n(remat->op, i);
+ for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
+ ir_node *arg = get_irn_n(remat->op, n);
query.irn = arg;
query.remats = NULL;
}
}
+#ifdef NO_SINGLE_USE_REMATS
static int
get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
{
return i;
}
+#endif
+
+#ifdef ONLY_BRIGGS_REMATS
+static int
+get_irn_n_nonignore_args(const spill_ilp_t * si, const ir_node * irn)
+{
+ int n;
+ unsigned int ret = 0;
+
+ for(n=get_irn_arity(irn)-1; n>=0; --n) {
+ if(has_reg_class(si, irn)) ++ret;
+ }
+
+ return ret;
+}
+#endif
static INLINE void
get_remats_from_op(spill_ilp_t * si, const ir_node * op)
{
- int i,
- n;
+ int n;
remat_t *remat;
+ if( has_reg_class(si, op)
#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)) {
+ && (get_irn_n_nonremat_edges(si, op) > 1)
#endif
+#ifdef ONLY_BRIGGS_REMATS
+ && (get_irn_n_nonignore_args(si, op) == 0)
+#endif
+ ) {
remat = get_remat_from_op(si, op, op);
if(remat) {
add_remat(si, remat);
}
}
-#ifdef COLLECT_INVERSE_REMATS
+#if defined(COLLECT_INVERSE_REMATS) && !defined(ONLY_BRIGGS_REMATS)
/* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
for each arg */
- for (i = 0, n = get_irn_arity(op); i < n; ++i) {
- ir_node *arg = get_irn_n(op, i);
+ 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 */
{
const ir_node *op = remat->op;
const ir_node *prev;
- int i,
- n,
+ int n,
res = 1;
if(is_Block(pos)) {
}
#endif
- for(i=0, n=get_irn_arity(op); i<n && res; ++i) {
- const ir_node *arg = get_irn_n(op, i);
+ 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) {
bb = is_Block(pos)?pos:get_nodes_block(pos);
copy = exact_copy(irn);
+
+ _set_phi_class(copy, NULL);
set_nodes_block(copy, bb);
sched_put_before(si, pos, copy);
bb = is_Block(pos)?pos:get_nodes_block(pos);
copy = exact_copy(irn);
+
+ _set_phi_class(copy, NULL);
set_nodes_block(copy, bb);
sched_put_after(pos, copy);
return copy;
}
-static void
-insert_remat_after(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
+static ir_node *
+insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
{
char buf[256];
copy = insert_copy_after(si, remat->op, pos);
-// ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", remat->value, pos);
ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
op = obstack_alloc(si->obst, sizeof(*op));
op->is_remat = 1;
} else {
proj_copy = NULL;
}
+
+ return copy;
}
+
+ return NULL;
}
-static void
-insert_remat_before(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
+static ir_node *
+insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
{
char buf[256];
copy = insert_copy_before(si, remat->op, pos);
-// ir_snprintf(buf, sizeof(buf), "remat_%N_%N", remat->value, pos);
ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
op = obstack_alloc(si->obst, sizeof(*op));
op->is_remat = 1;
} else {
proj_copy = NULL;
}
+
+ return copy;
}
+
+ return NULL;
}
static int
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);
}
}
op_t *op;
pset *args;
ir_node *arg;
+ pset *remat_args;
next = sched_prev(irn);
pset_insert_ptr(live, arg);
}
}
+ /* delete defined value from live set */
+ if(has_reg_class(si, irn)) {
+ pset_remove_ptr(live, irn);
+ }
- /* insert all possible remats after irn */
+
+ remat_args = pset_new_ptr_default();
+
+ /* insert all possible remats before irn */
pset_foreach(args, arg) {
remat_info_t *remat_info,
query;
remat_t *remat;
- /* continue if the operand has the wrong reg class */
+ /* continue if the operand has the wrong reg class
+ */
if(!has_reg_class(si, arg))
continue;
continue;
}
- /* do not place post remats after jumps */
- if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
+ if(remat_info->remats) {
+ pset_foreach(remat_info->remats, remat) {
+ ir_node *remat_irn = NULL;
- if(remat_info->remats_by_operand) {
- pset_foreach(remat_info->remats_by_operand, remat) {
- /* 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));
+ 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)) {
- insert_remat_after(si, remat, irn, live);
- }
+ if(pset_find_ptr(live, remat->value)) {
+ remat_irn = insert_remat_before(si, remat, irn, live);
+ }
#else
- insert_remat_after(si, remat, irn, live);
+ 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);
+
+ if(!has_reg_class(si, remat_arg)) continue;
+
+ pset_insert_ptr(remat_args, remat_arg);
+ }
}
}
}
}
- /* delete defined value from live set */
- if(has_reg_class(si, irn)) {
- pset_remove_ptr(live, irn);
+ /* now we add remat args to op's args because they could also die at this op */
+ pset_foreach(args,arg) {
+ if(pset_find_ptr(remat_args, arg)) {
+ pset_remove_ptr(remat_args, arg);
+ }
+ }
+ pset_foreach(remat_args,arg) {
+ pset_insert_ptr(args, arg);
}
- /* insert all possible remats before irn */
+ /* insert all possible remats after irn */
pset_foreach(args, arg) {
remat_info_t *remat_info,
query;
remat_t *remat;
- /* continue if the operand has the wrong reg class
- */
+ /* continue if the operand has the wrong reg class */
if(!has_reg_class(si, arg))
continue;
continue;
}
- if(remat_info->remats) {
- pset_foreach(remat_info->remats, remat) {
- DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
+ /* do not place post remats after jumps */
+ if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
+
+ if(remat_info->remats_by_operand) {
+ pset_foreach(remat_info->remats_by_operand, remat) {
+ /* 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)) {
- insert_remat_before(si, remat, irn, live);
- }
+ if(pset_find_ptr(live, remat->value)) {
+ insert_remat_after(si, remat, irn, live);
+ }
#else
- insert_remat_before(si, remat, irn, live);
+ insert_remat_after(si, remat, irn, live);
#endif
+ }
}
}
}
+ del_pset(remat_args);
del_pset(args);
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);
}
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;
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);
}
}
}
}
+/**
+ * Inserts ILP-constraints and variables for memory copying before the given position
+ */
+static void
+insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
+{
+ const ir_node *succ;
+ const ir_edge_t *edge;
+ spill_bb_t *spill_bb = get_irn_link(block);
+ ir_node *phi;
+ int pos;
+ ilp_cst_t cst;
+ ilp_var_t copyreg;
+ char buf[256];
+ ir_node *tmp;
+
+
+ assert(edges_activated(current_ir_graph));
+
+ edge = get_block_succ_first(block);
+ if(!edge) return;
+
+ succ = edge->src;
+ pos = edge->pos;
+
+ edge = get_block_succ_next(block, edge);
+ /* next block can only contain phis, if this is a merge edge */
+ if(edge) return;
+
+ ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
+ copyreg = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
+
+ ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
+ cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
+
+ pset_foreach(live, tmp) {
+ spill_t *spill;
+#if 0
+ op_t *op = get_irn_link(irn);
+ lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
+#endif
+ spill = set_find_spill(spill_bb->ilp, tmp);
+ assert(spill);
+
+ lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
+ }
+ lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
+
+ sched_foreach(succ, phi) {
+ const ir_node *to_copy;
+ op_t *to_copy_op;
+ spill_t *to_copy_spill;
+ op_t *phi_op = get_irn_link(phi);
+ ilp_var_t reload = ILP_UNDEF;
+
+
+ if(!is_Phi(phi)) break;
+ if(!has_reg_class(si, phi)) continue;
+
+ to_copy = get_irn_n(phi, pos);
+
+ to_copy_op = get_irn_link(to_copy);
+
+ to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
+ assert(to_copy_spill);
+
+ if(spill_bb->reloads) {
+ keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
+
+ if(keyval) {
+ reload = PTR_TO_INT(keyval->val);
+ }
+ }
+
+ 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);
+ lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
+ if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
+ lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
+ foreach_pre_remat(si, block, tmp) {
+ op_t *remat_op = get_irn_link(tmp);
+ if(remat_op->attr.remat.remat->value == to_copy) {
+ lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
+ }
+ }
+
+ 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);
+ lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
+ lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
+ }
+}
+
+
/**
* Walk all irg blocks and emit this ILP
*/
char buf[256];
ilp_cst_t cst;
spill_bb_t *spill_bb = get_irn_link(bb);
- int i;
ir_node *tmp;
spill_t *spill;
- irn_live_t *li;
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);
/* maybe we should also assure that reg_out >= live_range etc. */
}
+#ifndef NO_MEMCOPIES
+ insert_mem_copy_position(si, live, bb);
+#endif
+
/*
* start new live ranges for values used by remats at end of block
* and assure the remat args are available
d = 0;
ilp_cst_t check_pre,
check_post;
-#ifdef CHECK_POST_REMAT
- ilp_cst_t check_post_remat;
-#endif
set *args;
+ 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))
/* skip projs */
if(is_Proj(irn)) continue;
-#if 0
- /* remove defined values from live set */
- if(has_reg_class(si, irn)) {
- assert(pset_find_ptr(live, irn));
- pset_remove_ptr(live, irn);
- }
-#endif
+ /*
+ * init set of irn's arguments
+ * and all possibly used values around this op
+ * and values defined by post remats
+ */
+ args = new_set(cmp_keyval, get_irn_arity(irn));
+ used = pset_new_ptr(pset_count(live) + get_irn_arity(irn));
+ remat_defs = pset_new_ptr(pset_count(live));
- /* init set of irn's arguments */
- args = new_set(cmp_keyval, get_irn_arity(irn));
for (n=get_irn_arity(irn)-1; n>=0; --n) {
ir_node *irn_arg = get_irn_n(irn, n);
if(has_reg_class(si, irn_arg)) {
set_insert_keyval(args, irn_arg, (void*)n);
+ pset_insert_ptr(used, irn_arg);
+ }
+ }
+ foreach_post_remat(irn, tmp) {
+ op_t *remat_op = get_irn_link(tmp);
+
+ pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
+
+ 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)) {
+ pset_insert_ptr(used, remat_arg);
+ }
+ }
+ }
+ foreach_pre_remat(si, irn, tmp) {
+ 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)) {
+ pset_insert_ptr(used, remat_arg);
+ }
}
}
* I N E P I L O G O F irn
**********************************/
+ /* ensure each dying value is used by only one post remat */
+ pset_foreach(used, tmp) {
+ ir_node *value = tmp;
+ op_t *value_op = get_irn_link(value);
+ ir_node *remat;
+ int n_remats = 0;
+
+ cst = ILP_UNDEF;
+ foreach_post_remat(irn, remat) {
+ op_t *remat_op = get_irn_link(remat);
+
+ for(n=get_irn_arity(remat)-1; n>=0; --n) {
+ ir_node *remat_arg = get_irn_n(remat, n);
+
+ /* if value is used by this remat add it to constraint */
+ if(remat_arg == value) {
+ 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_uniq(si->lpp, buf, lpp_less, 1.0);
+ }
+
+ n_remats++;
+ lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
+ break;
+ }
+ }
+ }
+
+ // 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) {
- ir_node *value = tmp;//remat_op->attr.remat.remat->value;
+ ir_node *value = tmp;
op_t *value_op = get_irn_link(value);
if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
ilp_var_t prev_lr = ILP_UNDEF;
ir_node *remat;
- cst = ILP_UNDEF;
- foreach_post_remat(irn, remat) {
- op_t *remat_op = get_irn_link(remat);
+ if(pset_find_ptr(remat_defs, value)) {
- /* if value is being rematerialized by this remat */
- if(value == remat_op->attr.remat.remat->value) {
- if(cst == ILP_UNDEF) {
- /* 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);
+ /* next_live_range <= prev_live_range + sum remat2s */
+ ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
+ 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);
+ ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
+ prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 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);
- }
+ 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);
- lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
+ foreach_post_remat(irn, remat) {
+ op_t *remat_op = get_irn_link(remat);
+
+ /* if value is being rematerialized by this remat */
+ if(value == remat_op->attr.remat.remat->value) {
+ lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
+ }
}
- }
- if(prev_lr != ILP_UNDEF) {
value_op->attr.live_range.ilp = prev_lr;
value_op->attr.live_range.op = irn;
}
}
}
- /* requirements for post remats and start live ranges from L/U for values dying here */
+ /* requirements for post remats and start live ranges from L/U' for values dying here */
foreach_post_remat(irn, tmp) {
op_t *remat_op = get_irn_link(tmp);
int n;
if(!has_reg_class(si, remat_arg)) continue;
/* only for values in L\U (TODO and D?), the others are handled with post_use */
- if(!set_find_keyval(args, remat_arg)) {
- /* remat <= live_rang(remat_arg) */
+ if(!pset_find_ptr(used, remat_arg)) {
+ /* remat <= live_range(remat_arg) */
ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
}
}
-#ifdef CHECK_POST_REMAT
- /* check the register pressure after the epilog */
- ir_snprintf(buf, sizeof(buf), "check_post_remat_%N", irn);
- check_post_remat = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
-
- /* iterate over L\U */
- pset_foreach(live, tmp) {
- if(!set_find_keyval(args, tmp)) {
- /* if a live value is not used by irn */
- tmp_op = get_irn_link(tmp);
- lpp_set_factor_fast(si->lpp, check_post_remat, tmp_op->attr.live_range.ilp, 1.0);
- }
- }
- /* iterate over remats in epilog and remove possibly defined values again from check_post_remat */
- foreach_post_remat(irn, tmp) {
- op_t *remat_op = get_irn_link(tmp);
- const ir_node *value = remat_op->attr.remat.remat->value;
- op_t *val_op = get_irn_link(value);
-
- assert(remat_op->is_remat && !remat_op->attr.remat.pre);
-
- /* values that are defined by remats are not counted */
-#ifdef REMAT_WHILE_LIVE
- assert(val_op->attr.live_range.ilp);
- lpp_set_factor_fast(si->lpp, check_post_remat, val_op->attr.live_range.ilp, 0.0);
-#else
- if(val_op->attr.live_range.ilp != ILP_UNDEF) {
- lpp_set_factor_fast(si->lpp, check_post_remat, val_op->attr.live_range.ilp, 0.0);
- }
-#endif /* REMAT_WHILE_LIVE */
-
- /* instead add these post remats to the check */
- lpp_set_factor_fast(si->lpp, check_post_remat, remat_op->attr.remat.ilp, 1.0);
- }
-
-#if 0 /* defs are still in live */
- /* get count of values in my register class defined by irn */
- /* also add defined values to check_post_remat; do this before iterating over args */
- pset_foreach(defs,tmp) {
- op_t *def_op = get_irn_link(tmp);
-
- lpp_set_factor_fast(si->lpp, check_post_remat, def_op->attr.live_range.ilp, 1.0);
- }
-#endif
-#endif /* CHECK_POST_REMAT */
-
-
d = pset_count(defs);
DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
/* check the register pressure in the epilog */
- /* sum_{L\U} lr + sum_{U} post_use <= k - |D| */
+ /* 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 */
+ /* add L\U' to check_post */
pset_foreach(live, tmp) {
- if(!set_find_keyval(args, tmp) && !pset_find_ptr(defs, tmp)) {
+ if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
/* if a live value is not used by irn */
tmp_op = get_irn_link(tmp);
lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
}
/***********************************************************
- * I T E R A T I O N O V E R A R G S F O R E P I L O G
+ * I T E R A T I O N O V E R U S E S F O R E P I L O G
**********************************************************/
- set_foreach(args, keyval) {
- ilp_var_t next_lr;
+ pset_foreach(used, tmp) {
+ ilp_var_t prev_lr;
ilp_var_t post_use;
int p = 0;
spill_t *spill;
- ir_node *arg = keyval->key;
+ ir_node *arg = tmp;
op_t *arg_op = get_irn_link(arg);
+ ir_node *remat;
spill = add_to_spill_bb(si, bb, arg);
- /* new live range for each argument */
+ /* new live range for each used value */
ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
- next_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
-
- i = (int)keyval->val;
+ prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
/* the epilog stuff - including post_use, check_post, check_post_remat */
ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
if(pset_find_ptr(live, arg)) {
DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
+ /* 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);
-#ifdef CHECK_POST_REMAT
- //lpp_set_factor_fast(si->lpp, check_post_remat, arg_op->attr.live_range.ilp, 1.0);
- lpp_set_factor_fast(si->lpp, check_post_remat, post_use, 1.0);
-#endif
}
- /*forall remat2 which use arg add a similar cst*/
- foreach_post_remat(irn, tmp) {
+ /* if value is not an arg of op and not possibly defined by post remat
+ * then it may only die and not become live
+ */
+ 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_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_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;
- 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);
+ for (n=get_irn_arity(remat)-1; n>=0; --n) {
+ ir_node *remat_arg = get_irn_n(remat, n);
+ op_t *remat_op = get_irn_link(remat);
if(remat_arg == arg) {
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);
}
}
}
- /* new live range begins for each argument */
- arg_op->attr.live_range.ilp = next_lr;
+#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;
+ /*if(!pset_find_ptr(live, arg)) {
+ pset_insert_ptr(live, arg);
+ add_to_spill_bb(si, bb, arg);
+ }*/
pset_insert_ptr(live, arg);
+
}
/* just to be sure */
check_post = ILP_UNDEF;
-#ifdef CHECK_POST_REMAT
- check_post_remat = ILP_UNDEF;
-#endif
/* 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;
+ 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;
+#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 */
}
- /* TODO allow new live ranges even if value does not die after remat??? MAY_DIE_AT_PRE_REMAT */
-
/* requirements for remats */
/* start new live ranges for values used by remats */
foreach_pre_remat(si, irn, tmp) {
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;
- /* if value is becoming live through use by remat */
- if(!pset_find_ptr(live, remat_arg)) {
- ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
- prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
-
- arg_op->attr.live_range.ilp = prev_lr;
- arg_op->attr.live_range.op = irn;
-
- DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat %+F\n", remat_arg, tmp));
-
- pset_insert_ptr(live, remat_arg);
- add_to_spill_bb(si, bb, remat_arg);
- }
-
/* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
}
+ del_pset(remat_defs);
+ del_pset(used);
del_set(args);
del_pset(defs);
defs = pset_new_ptr_default();
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);
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 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;
}
}
+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, const ir_node * bb, const ir_node * a, const 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)) {
- //DBG((si->dbg, LEVEL_1, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
- ir_fprintf(stderr, "\tvalues interfere in %+F: %+F, %+F\n", 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
-find_copy_path(spill_ilp_t * si, ir_node * irn, ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
+static int
+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);
+ 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);
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
-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();
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;
cst;
const ir_node *a = interference->a;
const ir_node *b = interference->b;
- struct list_head *pos;
/* 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);
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);
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);
}
-
-static void
-memcopyinsertor(spill_ilp_t * si)
-{
- /* weise Spillkontexte zu. Sorge bei Phis dafuer, dass gleiche
- * Kontexte zusammenfliessen (Operanden und Ergebnis hat gleichen
- * Kontext)
- */
-
-
-
-
-
-}
-
-
-
-
static INLINE int
is_zero(double x)
{
static void
dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
{
- DUMP_NODE_EDGE_FUNC old = get_dump_node_edge_hook();
+ DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
dump_consts_local(0);
set_dump_node_edge_hook(sched_pressure_edge_hook);
dump_ir_block_graph(irg, suffix);
- set_dump_node_edge_hook(old);
+ set_dump_node_edge_hook(old_edge_hook);
}
static void
{
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);
}
}
}
/** 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);
}
}
}
#endif
}
+static pset *
+get_spills_for_value(spill_ilp_t * si, const ir_node * value)
+{
+ pset *spills = pset_new_ptr_default();
+
+ const ir_node *next;
+ defs_t *defs;
+
+ defs = set_find_def(si->values, value);
+
+ if(defs && defs->spills) {
+ for(next = defs->spills; next; next = get_irn_link(next)) {
+ pset_insert_ptr(spills, next);
+ }
+ }
+
+ return spills;
+}
+
/**
* @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);
* @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;
pset_insert_ptr(si->spills, mem_phi);
#endif
+
return mem_phi;
}
defs->remats = remat;
}
-#if 0
-static void
-collect_spills(spill_ilp_t * si, ir_node * value, pset * spills, pset * visited)
-{
- ir_node *next;
- defs_t *defs;
-
- defs = set_find_def(si->values, value);
-
- if(defs && defs->spills) {
- for(next = defs->spills; next; next = get_irn_link(next)) {
- pset_insert_ptr(spills, next);
- }
- } else if (is_Phi(value)) {
- /* recursion */
- if(!pset_find_ptr(visited, value)) {
- int i,
- n;
-
- pset_insert_ptr(visited, value);
- for(i=0, n=get_irn_arity(value); i<n; ++i) {
- ir_node *arg = get_irn_n(value, i);
-
- collect_spills(si, arg, spills, visited);
- }
- }
- } else {
-// assert(0 && "Phi operand not spilled");
- }
-}
-#endif
-
-static pset *
-get_spills_for_value(spill_ilp_t * si, ir_node * value)
-{
- pset *spills = pset_new_ptr_default();
-// pset *visited = pset_new_ptr_default();
-
-// collect_spills(si, value, spills, visited);
-// del_pset(visited);
- ir_node *next;
- defs_t *defs;
-
- defs = set_find_def(si->values, value);
-
- if(defs && defs->spills) {
- for(next = defs->spills; next; next = get_irn_link(next)) {
- pset_insert_ptr(spills, next);
- }
- }
-
- return spills;
-}
/**
* 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,
DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
defs = set_find_def(si->values, value);
- /* get a spill of this value */
-#if 0
- if((!defs || !defs->spills) && is_Phi(value)) {
- pset *spills;
-
- spills = get_spills_for_value(si, value);
- spill = pset_first(spills);
- del_pset(spills);
-
- if(!defs) {
- defs = set_insert_def(si->values, value);
- }
- defs->spills = spill;
- set_irn_link(spill, NULL);
- } else {
- spill = defs->spills;
- }
-#endif
spill = defs->spills;
assert(spill && "no spill placed before reload");
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;
del_pset(spills_to_do);
}
+static ir_node *
+insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
+{
+ ir_node *insert_pos = bb;
+ ir_node *spill;
+ const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
+
+ /* find last definition of arg value in block */
+ ir_node *next;
+ defs_t *defs;
+ int last = 0;
+
+ defs = set_find_def(si->values, value);
+
+ if(defs && defs->remats) {
+ for(next = defs->remats; next; next = get_irn_link(next)) {
+ if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
+ last = sched_get_time_step(next);
+ insert_pos = next;
+ }
+ }
+ }
+
+ if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
+ last = sched_get_time_step(value);
+ insert_pos = value;
+ }
+
+ 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);
+
+ return spill;
+}
+
static void
phim_fixer(spill_ilp_t *si) {
defs_t *defs;
set_foreach(si->values, defs) {
const ir_node *phi = defs->value;
- ir_node *phi_m = NULL;
- ir_node *next = defs->spills;
- int n;
+ op_t *op = get_irn_link(phi);
+ ir_node *phi_m = NULL;
+ ir_node *next = defs->spills;
+ int n;
if(!is_Phi(phi)) continue;
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);
- /* get a spill of this value */
- ir_node *spill = val_defs->spills;
+ /* a spill of this value */
+ ir_node *spill;
- assert(spill && "no spill placed before PhiM");
+#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(!is_zero(name->value)) {
+ spill = insert_mem_copy(si, pred, value);
+ } 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;
- irn_live_t *li;
/* reloads at end of block */
if(spill_bb->reloads) {
ir_node *prev = sched_block_last_noncf(si, bb);
op_t *prev_op = get_irn_link(prev);
+ while(be_is_Spill(prev)) {
+ prev = sched_prev(prev);
+ }
+
+ prev_op = get_irn_link(prev);
+
/* insert reload before pre-remats */
- while(!sched_is_end(prev) && !be_is_Reload(prev) && !be_is_Spill(prev)
+ while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
&& prev_op->is_remat && prev_op->attr.remat.pre) {
insert_pos = prev;
- prev = sched_prev(insert_pos);
+ do {
+ prev = sched_prev(prev);
+ } while(be_is_Spill(prev));
+
prev_op = get_irn_link(prev);
+
}
reload = insert_reload(si, irn, insert_pos);
if(!is_zero(name->value)) {
ir_node *reload;
ir_node *insert_pos = irn;
- ir_node *prev = insert_pos;
+ ir_node *prev = sched_prev(insert_pos);
op_t *prev_op;
- do {
+ while(be_is_Spill(prev)) {
prev = sched_prev(prev);
- } while(be_is_Spill(prev));
+ }
prev_op = get_irn_link(prev);
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;
}
}
+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)
{
/* 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();
// 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);
/* 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();
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);
be_free_dominance_frontiers(dfi);
}
+
static void
writeback_results(spill_ilp_t * si)
{
delete_unnecessary_remats(si);
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);
- phim_fixer(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! */
clean_remat_info(si);
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)) ||
+ /* 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, walker_reload_mover, NULL, si);
}
+
+/**
+ * Walk all irg blocks and check for interfering spills inside of phi classes
+ */
+static void
+luke_meminterferencechecker(ir_node * bb, void * data)
+{
+ spill_ilp_t *si = (spill_ilp_t*)data;
+ int l1, l2;
+
+ 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 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(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);
+ }
+ }
+ }
+ }
+ }
+}
+
+static void
+verify_phiclasses(spill_ilp_t * si)
+{
+ /* analyze phi classes */
+ phi_class_compute(si->chordal_env->irg);
+
+ DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
+ irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
+}
+
+static void
+walker_spillslotassigner(ir_node * irn, void * data)
+{
+ void *cls;
+
+ if(!be_is_Spill(irn)) return;
+
+ /* set spill context to phi class if it has one ;) */
+
+#if 0
+ // Matze: not needed anymore
+ cls = get_phi_class(irn);
+ if(cls)
+ be_set_Spill_context(irn, cls);
+ else
+ be_set_Spill_context(irn, irn);
+#endif
+}
+
+
+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 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.cls = chordal_env->cls;
si.lpp = new_lpp(problem_name, lpp_minimize);
si.remat_info = new_set(cmp_remat_info, 4096);
- si.interferences = new_set(cmp_interference, 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();
-#ifndef EXECFREQ_LOOPDEPH
- si.execfreqs = compute_execfreq(chordal_env->irg);
-#else
- si.execfreqs = NULL;
-#endif
+ 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
#ifdef SOLVE
- DBG((si.dbg, LEVEL_1, "\tSolving %F\n", chordal_env->irg));
+ DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
#ifdef ILP_TIMEOUT
lpp_set_time_limit(si.lpp, ILP_TIMEOUT);
#endif
#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);
+#ifndef NO_MEMCOPIES
+ verify_phiclasses(&si);
+ assign_spillslots(&si);
+#endif
+
irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
dump_pressure_graph(&si, dump_suffix2);
- // TODO fix temporarily exceeded regpressure due to remat2s
-
- // TODO insert copys to fix interferences in memory
-
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);
- del_pset(si.spills);
-#ifndef EXECFREQ_LOOPDEPH
- free_execfreq(si.execfreqs);
+#ifdef WITH_MEMOPERANDS
+ del_set(si.memoperands);
#endif
+ del_pset(si.spills);
free_lpp(si.lpp);
obstack_free(&obst, NULL);
DBG((si.dbg, LEVEL_1, "\tdone.\n"));