X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbespillremat.c;h=69c216dce0e8b1b34282ce743d7b57d9a2301638;hb=92dc34b4a3a1880d8e39cc87aaa2bdfcdaf517f2;hp=374de5acb5bfd5e9b7b10f99d2b1fe6b37753b80;hpb=1e38a03dcea3b49eef3c6f7efcd711dbab692d1a;p=libfirm diff --git a/ir/be/bespillremat.c b/ir/be/bespillremat.c index 374de5acb..69c216dce 100644 --- a/ir/be/bespillremat.c +++ b/ir/be/bespillremat.c @@ -29,11 +29,13 @@ #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 +#include #include #include //#include @@ -52,53 +54,138 @@ #include "bechordal_t.h" -#define BIGM 100000.0 - -#define DUMP_SOLUTION -#define DUMP_ILP -//#define KEEPALIVE /* keep alive all inserted remats and dump graph with remats */ -#define COLLECT_REMATS /* enable rematerialization */ -#define COLLECT_INVERSE_REMATS /* enable placement of inverse remats */ -#define REMAT_WHILE_LIVE /* only remat values that are live */ -//#define NO_ENLARGE_L1V3N355 /* do not remat after the death of some operand */ -//#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */ -#define MAY_DIE_AT_REMAT /* allow values to die after a pre remat */ -#define NO_SINGLE_USE_REMATS /* do not repair schedule */ -//#define KEEPALIVE_SPILLS -//#define KEEPALIVE_RELOADS -#define GOODWIN_REDUCTION -#define NO_MEMCOPIES +#ifdef WITH_LIBCORE +#include +#include +#endif /* WITH_LIBCORE */ + +#define DUMP_PROBLEM 1 +#define DUMP_MPS 2 +#define DUMP_SOLUTION 4 + +#define KEEPALIVE_REMATS 1 +#define KEEPALIVE_SPILLS 2 +#define KEEPALIVE_RELOADS 4 + +#define VERIFY_MEMINTERF 1 +#define VERIFY_DOMINANCE 2 + +#define REMATS_NONE 0 +#define REMATS_BRIGGS 1 +#define REMATS_NOINVERSE 2 +#define REMATS_ALL 3 + +static int opt_dump_flags = 0; +static int opt_log = 0; +static int opt_keep_alive = 0; +static int opt_goodwin = 1; +static int opt_memcopies = 1; +static int opt_memoperands = 1; +static int opt_verify = VERIFY_MEMINTERF; +static int opt_remats = REMATS_ALL; +static int opt_repair_schedule = 0; +static int opt_no_enlarge_liveness = 0; +static int opt_remat_while_live = 1; +static int opt_timeout = 300; +static double opt_cost_reload = 8.0; +static double opt_cost_memoperand = 7.0; +static double opt_cost_spill = 50.0; +static double opt_cost_remat = 1.0; + + +#ifdef WITH_LIBCORE +static const lc_opt_enum_mask_items_t dump_items[] = { + { "problem", DUMP_PROBLEM }, + { "mps", DUMP_MPS }, + { "solution", DUMP_SOLUTION }, + { NULL, 0 } +}; + +static lc_opt_enum_mask_var_t dump_var = { + &opt_dump_flags, dump_items +}; + +static const lc_opt_enum_mask_items_t keepalive_items[] = { + { "remats", KEEPALIVE_REMATS }, + { "spills", KEEPALIVE_SPILLS }, + { "reloads", KEEPALIVE_RELOADS }, + { NULL, 0 } +}; + +static lc_opt_enum_mask_var_t keep_alive_var = { + &opt_keep_alive, keepalive_items +}; + +static const lc_opt_enum_mask_items_t remats_items[] = { + { "none", REMATS_NONE }, + { "briggs", REMATS_BRIGGS }, + { "noinverse", REMATS_NOINVERSE }, + { "all", REMATS_ALL }, + { NULL, 0 } +}; + +static lc_opt_enum_mask_var_t remats_var = { + &opt_remats, remats_items +}; + +static const lc_opt_table_entry_t options[] = { + LC_OPT_ENT_ENUM_MASK("keepalive", "keep alive remats, spills or reloads", &keep_alive_var), + + LC_OPT_ENT_BOOL ("goodwin", "activate goodwin reduction", &opt_goodwin), + LC_OPT_ENT_BOOL ("memcopies", "activate memcopy handling", &opt_memcopies), + LC_OPT_ENT_BOOL ("memoperands", "activate memoperands", &opt_memoperands), + LC_OPT_ENT_ENUM_INT ("remats", "type of remats to insert (none, briggs, noinverse or all)",&remats_var), + LC_OPT_ENT_BOOL ("repair_schedule", "repair the schedule by rematting once used nodes",&opt_repair_schedule), + LC_OPT_ENT_BOOL ("no_enlage_liveness", "do not enlarge liveness of operands of remats",&opt_no_enlarge_liveness), + LC_OPT_ENT_BOOL ("remat_while_live", "remat only values that can be used by real ops", &opt_remat_while_live), + + LC_OPT_ENT_ENUM_MASK("dump", "dump problem, mps or solution", &dump_var), + LC_OPT_ENT_BOOL ("log", "activate the lpp log", &opt_log), + LC_OPT_ENT_INT ("timeout", "ILP solver timeout", &opt_timeout), + + LC_OPT_ENT_DBL ("cost_reload", "cost of a reload", &opt_cost_reload), + LC_OPT_ENT_DBL ("cost_memoperand", "cost of a memory operand", &opt_cost_memoperand), + LC_OPT_ENT_DBL ("cost_spill", "cost of a spill instruction", &opt_cost_spill), + LC_OPT_ENT_DBL ("cost_remat", "cost of a rematerialization", &opt_cost_remat), + { NULL } +}; + +void be_spill_remat_register_options(lc_opt_entry_t *grp) +{ + lc_opt_entry_t *my_grp = lc_opt_get_grp(grp, "remat"); + lc_opt_add_table(my_grp, options); +} +#endif + + +//#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */ +//#define SCHEDULE_PHIM /* insert phim nodes into schedule */ #define SOLVE //#define SOLVE_LOCAL #define LPP_SERVER "i44pc52" #define LPP_SOLVER "cplex" -#define COST_LOAD 10 -#define COST_STORE 50 -#define COST_REMAT 1 - -#define ILP_TIMEOUT 120 +#define MAX_PATHS INT_MAX #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 *all_possible_remats; pset *inverse_ops; -#ifdef KEEPALIVE 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; + set *memoperands; DEBUG_ONLY(firm_dbg_module_t * dbg); } spill_ilp_t; @@ -113,7 +200,7 @@ typedef struct _spill_bb_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; @@ -128,7 +215,7 @@ typedef struct _op_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 { @@ -143,9 +230,9 @@ typedef struct _op_t { } 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 { @@ -160,14 +247,20 @@ typedef struct _keyval_t { } keyval_t; typedef struct _spill_t { - ir_node *irn; - ilp_var_t reg_in; - ilp_var_t mem_in; - ilp_var_t reg_out; - ilp_var_t mem_out; - ilp_var_t spill; + ir_node *irn; + ilp_var_t reg_in; + ilp_var_t mem_in; + ilp_var_t reg_out; + ilp_var_t mem_out; + ilp_var_t spill; } spill_t; +typedef struct _memoperand_t { + ir_node *irn; /**< the irn */ + unsigned int pos; /**< the position of the argument */ + ilp_var_t ilp; /**< the ilp var for this memory operand */ +} memoperand_t; + static INLINE int has_reg_class(const spill_ilp_t * si, const ir_node * irn) { @@ -207,8 +300,17 @@ cmp_spill(const void *a, const void *b, size_t size) return !(p->irn == q->irn); } +static int +cmp_memoperands(const void *a, const void *b, size_t size) +{ + const memoperand_t *p = a; + const memoperand_t *q = b; + + return !(p->irn == q->irn && p->pos == q->pos); +} + static keyval_t * -set_find_keyval(set * set, void * key) +set_find_keyval(set * set, const void * key) { keyval_t query; @@ -227,7 +329,7 @@ set_insert_keyval(set * set, void * key, void * val) } static defs_t * -set_find_def(set * set, ir_node * value) +set_find_def(set * set, const ir_node * value) { defs_t query; @@ -236,7 +338,7 @@ set_find_def(set * set, ir_node * value) } static defs_t * -set_insert_def(set * set, ir_node * value) +set_insert_def(set * set, const ir_node * value) { defs_t query; @@ -246,12 +348,34 @@ set_insert_def(set * set, ir_node * value) return set_insert(set, &query, sizeof(query), HASH_PTR(value)); } +static memoperand_t * +set_insert_memoperand(set * set, ir_node * irn, unsigned int pos, ilp_var_t ilp) +{ + memoperand_t query; + + query.irn = irn; + query.pos = pos; + query.ilp = ilp; + return set_insert(set, &query, sizeof(query), HASH_PTR(irn)+pos); +} + +static memoperand_t * +set_find_memoperand(set * set, const ir_node * irn, unsigned int pos) +{ + memoperand_t query; + + query.irn = (ir_node*)irn; + query.pos = pos; + return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos); +} + + static spill_t * -set_find_spill(set * set, ir_node * value) +set_find_spill(set * set, const ir_node * value) { spill_t query; - query.irn = value; + query.irn = (ir_node*)value; return set_find(set, &query, sizeof(query), HASH_PTR(value)); } @@ -289,34 +413,29 @@ cmp_keyval(const void *a, const void *b, size_t size) } 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 get_cost(const spill_ilp_t * si, const ir_node * irn) { if(be_is_Spill(irn)) { - return COST_STORE; + return opt_cost_spill; } else if(be_is_Reload(irn)){ - return COST_LOAD; + return opt_cost_reload; } else { return arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, irn); } - } /** @@ -360,7 +479,7 @@ get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * const ir_node *proj = NULL; if(is_Proj(dest_value)) { - op = get_irn_n(op, 0); + op = get_Proj_pred(op); proj = dest_value; } @@ -477,39 +596,55 @@ get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn) return i; } +static int +get_irn_n_nonignore_args(const spill_ilp_t * si, const ir_node * irn) +{ + int n; + int ret = 0; + +// if(is_Proj(irn)) +// irn = get_Proj_pred(irn); + + for(n=get_irn_arity(irn)-1; n>=0; --n) { + const ir_node *arg = get_irn_n(irn, n); + + if(has_reg_class(si, arg)) ++ret; + } + + return ret; +} + static INLINE void get_remats_from_op(spill_ilp_t * si, const ir_node * op) { int n; remat_t *remat; -#ifdef NO_SINGLE_USE_REMATS - if(has_reg_class(si, op) && (get_irn_n_nonremat_edges(si, op) > 1)) { -#else - if(has_reg_class(si, op)) { -#endif + if( has_reg_class(si, op) + && (opt_repair_schedule || get_irn_n_nonremat_edges(si, op) > 1) + && (opt_remats != REMATS_BRIGGS || get_irn_n_nonignore_args(si, op) == 0) + ) { remat = get_remat_from_op(si, op, op); if(remat) { add_remat(si, remat); } } -#ifdef COLLECT_INVERSE_REMATS - /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg) - for each arg */ - for (n = get_irn_arity(op)-1; n>=0; --n) { - ir_node *arg = get_irn_n(op, n); + if(opt_remats == REMATS_ALL) { + /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg) + for each arg */ + for (n = get_irn_arity(op)-1; n>=0; --n) { + ir_node *arg = get_irn_n(op, n); - if(has_reg_class(si, arg)) { - /* try to get an inverse remat */ - remat = get_remat_from_op(si, arg, op); - if(remat) { - add_remat(si, remat); + if(has_reg_class(si, arg)) { + /* try to get an inverse remat */ + remat = get_remat_from_op(si, arg, op); + if(remat) { + add_remat(si, remat); + } } } } -#endif - } static INLINE int @@ -659,15 +794,15 @@ can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * for(n=get_irn_arity(op)-1; n>=0 && res; --n) { const ir_node *arg = get_irn_n(op, n); -#ifdef NO_ENLARGE_L1V3N355 - if(has_reg_class(si, arg) && live) { - res &= pset_find_ptr(live, arg)?1:0; + if(opt_no_enlarge_liveness) { + if(has_reg_class(si, arg) && live) { + res &= pset_find_ptr((pset*)live, arg)?1:0; + } else { + res &= value_is_defined_before(si, pos, arg); + } } else { res &= value_is_defined_before(si, pos, arg); } -#else - res &= value_is_defined_before(si, pos, arg); -#endif } return res; @@ -718,6 +853,8 @@ insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos) 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); @@ -735,6 +872,8 @@ insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos) 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); @@ -742,7 +881,7 @@ insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos) } static ir_node * -insert_remat_after(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live) +insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live) { char buf[256]; @@ -751,7 +890,7 @@ insert_remat_after(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, *proj_copy; op_t *op; - DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op)); + DBG((si->dbg, LEVEL_3, "\t >inserting remat2 %+F\n", remat->op)); copy = insert_copy_after(si, remat->op, pos); @@ -760,7 +899,7 @@ insert_remat_after(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, op->is_remat = 1; op->attr.remat.remat = remat; op->attr.remat.pre = 0; - op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos)); + op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0); set_irn_link(copy, op); pset_insert_ptr(si->all_possible_remats, copy); @@ -780,7 +919,7 @@ insert_remat_after(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, } static ir_node * -insert_remat_before(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live) +insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live) { char buf[256]; @@ -798,7 +937,7 @@ insert_remat_before(spill_ilp_t * si, const remat_t * remat, const ir_node * pos op->is_remat = 1; op->attr.remat.remat = remat; op->attr.remat.pre = 1; - op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos)); + op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0); set_irn_link(copy, op); pset_insert_ptr(si->all_possible_remats, copy); @@ -831,24 +970,45 @@ get_block_n_succs(const ir_node *block) { return edge ? 2 : 1; } +static int +is_start_block(const ir_node * bb) +{ + return get_irg_start_block(get_irn_irg(bb)) == bb; +} + +static int +is_before_frame(const ir_node * bb, const ir_node * irn) +{ + const ir_node *frame = get_irg_frame(get_irn_irg(bb)); + + if(is_start_block(bb) && sched_get_time_step(frame) >= sched_get_time_step(irn)) + return 1; + else + return 0; +} + static int is_merge_edge(const ir_node * bb) { -#ifdef GOODWIN_REDUCTION - return get_block_n_succs(bb) == 1; -#else - return 1; -#endif + if(is_start_block(bb)) + return 0; + + if(opt_goodwin) + return get_block_n_succs(bb) == 1; + else + return 1; } static int is_diverge_edge(const ir_node * bb) { -#ifdef GOODWIN_REDUCTION - return get_Block_n_cfgpreds(bb) == 1; -#else - return 1; -#endif + if(is_start_block(bb)) + return 0; + + if(opt_goodwin) + return get_Block_n_cfgpreds(bb) == 1; + else + return 1; } static void @@ -892,17 +1052,16 @@ walker_remat_insertor(ir_node * bb, void * data) 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); } } @@ -994,13 +1153,13 @@ walker_remat_insertor(ir_node * bb, void * data) ir_node *remat_irn = NULL; DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg)); -#ifdef REMAT_WHILE_LIVE - if(pset_find_ptr(live, remat->value)) { + if(opt_remat_while_live) { + if(pset_find_ptr(live, remat->value)) { + remat_irn = insert_remat_before(si, remat, irn, live); + } + } else { remat_irn = insert_remat_before(si, remat, irn, live); } -#else - remat_irn = insert_remat_before(si, remat, irn, live); -#endif if(remat_irn) { for(n=get_irn_arity(remat_irn)-1; n>=0; --n) { ir_node *remat_arg = get_irn_n(remat_irn, n); @@ -1051,13 +1210,13 @@ walker_remat_insertor(ir_node * bb, void * data) /* do not insert remats producing the same value as one of the operands */ if(!pset_find_ptr(args, remat->value)) { DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg)); -#ifdef REMAT_WHILE_LIVE - if(pset_find_ptr(live, remat->value)) { + if(opt_remat_while_live) { + if(pset_find_ptr(live, remat->value)) { + insert_remat_after(si, remat, irn, live); + } + } else { insert_remat_after(si, remat, irn, live); } -#else - insert_remat_after(si, remat, irn, live); -#endif } } } @@ -1068,34 +1227,56 @@ walker_remat_insertor(ir_node * bb, void * data) irn = next; } - live_foreach(bb, li) { - ir_node *value = (ir_node *) li->irn; + /* add remats at end if successor has multiple predecessors */ + if(is_merge_edge(bb)) { + pset *live_out = pset_new_ptr_default(); + ir_node *value; - /* 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)) { - remat_info_t *remat_info, - query; - remat_t *remat; + be_lv_foreach(si->lv, bb, be_lv_state_end, i) { + value = be_lv_get_irn(si->lv, bb, i); - query.irn = value; - query.remats = NULL; - query.remats_by_operand = NULL; - remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value)); + if (be_is_live_end(si->lv, bb, value) && has_reg_class(si, value)) { + pset_insert_ptr(live_out, value); + } + } + + /* add remats at end of block */ + pset_foreach(live_out, value) { + remat_info_t *remat_info, + query; + remat_t *remat; + + query.irn = value; + query.remats = NULL; + query.remats_by_operand = NULL; + remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value)); - if(remat_info && remat_info->remats) { - pset_foreach(remat_info->remats, remat) { - DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb)); + if(remat_info && remat_info->remats) { + pset_foreach(remat_info->remats, remat) { + DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb)); - insert_remat_before(si, remat, bb, NULL); - } + insert_remat_before(si, remat, bb, live_out); } } } - if(is_diverge_edge(bb)) { - /* add remat2s at beginning of block */ - if ((live_is_in(li) || (is_Phi(value) && get_nodes_block(value)==bb)) && has_reg_class(si, value)) { + del_pset(live_out); + } + + if(is_diverge_edge(bb)) { + pset *live_in = pset_new_ptr_default(); + ir_node *value; + + be_lv_foreach(si->lv, bb, be_lv_state_in, i) { + value = be_lv_get_irn(si->lv, bb, i); + + if (has_reg_class(si, value)) { + pset_insert_ptr(live_in, value); + } + } + + /* add remat2s at beginning of block */ + pset_foreach(live_in, value) { + if ((be_is_live_in(si->lv, bb, value) || (is_Phi(value) && get_nodes_block(value)==bb)) && has_reg_class(si, value)) { remat_info_t *remat_info, query; remat_t *remat; @@ -1105,17 +1286,18 @@ walker_remat_insertor(ir_node * bb, void * data) query.remats_by_operand = NULL; remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value)); - if(remat_info && remat_info->remats) { - pset_foreach(remat_info->remats, remat) { - DBG((si->dbg, LEVEL_4, "\t considering remat %+F at beginning of block %+F\n", remat->op, bb)); + if(remat_info && remat_info->remats_by_operand) { + pset_foreach(remat_info->remats_by_operand, remat) { + DBG((si->dbg, LEVEL_4, "\t considering remat2 %+F at beginning of block %+F\n", remat->op, bb)); /* put the remat here if all its args are available */ - insert_remat_after(si, remat, bb, NULL); + insert_remat_after(si, remat, bb, live_in); } } } } + del_pset(live_in); } } @@ -1126,21 +1308,21 @@ static void 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); @@ -1166,8 +1348,8 @@ luke_endwalker(ir_node * bb, void * data) } 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)); @@ -1184,6 +1366,7 @@ luke_endwalker(ir_node * bb, void * data) spill_t query, *spill; double spill_cost; + int default_spilled; /* handle values used by control flow nodes later separately */ @@ -1192,31 +1375,33 @@ luke_endwalker(ir_node * bb, void * data) query.irn = irn; spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn)); - spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb); + spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb); ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb); - spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0); + spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0); lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0); ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb); - spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0); + spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0); ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb); - spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost); + /* by default spill value right after definition */ + default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn); + spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled); if(is_merge_edge(bb)) { ilp_var_t reload; ilp_cst_t rel_cst; ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn); - reload = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb)); + reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0); set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload)); /* reload <= mem_out */ - rel_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0); + rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0); lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0); lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0); - } + } spill->reg_in = ILP_UNDEF; spill->mem_in = ILP_UNDEF; @@ -1229,29 +1414,32 @@ luke_endwalker(ir_node * bb, void * data) ilp_cst_t end_use_req, rel_cst; ilp_var_t reload; + int default_spilled; query.irn = irn; spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn)); - spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb); + spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb); ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb); - spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0); + spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0); /* if irn is used at the end of the block, then it is live anyway */ //lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0); ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb); - spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0); + spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0); ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb); - spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost); + default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn); + spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled); + /* reload for use be control flow op */ ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn); - reload = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb)); + reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0); set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload)); /* reload <= mem_out */ - rel_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0); + rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0); lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0); lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0); @@ -1259,7 +1447,7 @@ luke_endwalker(ir_node * bb, void * data) 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); } @@ -1271,19 +1459,20 @@ static ir_node * 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; @@ -1350,11 +1539,12 @@ add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn) spill_t *spill, query; char buf[256]; + int default_spilled; query.irn = irn; spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn)); if(!spill) { - double spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb); + double spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb); spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn)); @@ -1363,10 +1553,11 @@ add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn) spill->mem_in = ILP_UNDEF; ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb); - spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0); + spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0); ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb); - spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost); + default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn); + spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled); } return spill; @@ -1375,13 +1566,13 @@ add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn) 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); } } @@ -1434,10 +1625,10 @@ insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block) if(edge) return; ir_snprintf(buf, sizeof(buf), "copyreg_%N", block); - copyreg = lpp_add_var(si->lpp, buf, lpp_binary, 0.0); + copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0); ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block); - cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs); pset_foreach(live, tmp) { spill_t *spill; @@ -1478,8 +1669,8 @@ insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block) } } - 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); @@ -1493,8 +1684,8 @@ insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block) } } - ir_snprintf(buf, sizeof(buf), "copyreq_%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); @@ -1519,6 +1710,7 @@ luke_blockwalker(ir_node * bb, void * data) ir_node *tmp; spill_t *spill; pset *defs = pset_new_ptr_default(); + const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env; live = pset_new_ptr_default(); @@ -1550,11 +1742,11 @@ luke_blockwalker(ir_node * bb, void * data) assert(!op->is_remat); ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb); - op->attr.live_range.ilp = lpp_add_var(si->lpp, buf, lpp_binary, 0.0); + op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0); op->attr.live_range.op = bb; ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn); - cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0); /* reg_out - reload - remat - live_range <= 0 */ lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0); @@ -1569,9 +1761,8 @@ luke_blockwalker(ir_node * bb, void * data) /* maybe we should also assure that reg_out >= live_range etc. */ } -#ifndef NO_MEMCOPIES - insert_mem_copy_position(si, live, bb); -#endif + if(opt_memcopies) + insert_mem_copy_position(si, live, bb); /* * start new live ranges for values used by remats at end of block @@ -1591,7 +1782,7 @@ luke_blockwalker(ir_node * bb, void * data) /* if value is becoming live through use by remat */ if(!pset_find_ptr(live, remat_arg)) { ir_snprintf(buf, sizeof(buf), "lr_%N_end%N", remat_arg, bb); - prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0); + prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0); arg_op->attr.live_range.ilp = prev_lr; arg_op->attr.live_range.op = bb; @@ -1642,6 +1833,7 @@ luke_blockwalker(ir_node * bb, void * data) pset *used; pset *remat_defs; keyval_t *keyval; + ilp_cst_t one_memoperand; /* iterate only until first phi */ if(is_Phi(irn)) @@ -1702,7 +1894,7 @@ luke_blockwalker(ir_node * bb, void * data) **********************************/ /* 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; @@ -1720,7 +1912,7 @@ luke_blockwalker(ir_node * bb, void * data) 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++; @@ -1730,12 +1922,45 @@ luke_blockwalker(ir_node * bb, void * data) } } - 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) { @@ -1750,10 +1975,10 @@ luke_blockwalker(ir_node * bb, void * data) /* next_live_range <= prev_live_range + sum remat2s */ ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn); - cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0); ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn); - prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0); + prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0); lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0); lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0); @@ -1786,7 +2011,7 @@ luke_blockwalker(ir_node * bb, void * data) /* only for values in L\U (TODO and D?), the others are handled with post_use */ if(!pset_find_ptr(used, remat_arg)) { - /* remat <= live_rang(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); @@ -1795,7 +2020,7 @@ luke_blockwalker(ir_node * bb, void * data) ilp_var_t lr; ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn); - lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0); + lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0); arg_op->attr.live_range.ilp = lr; arg_op->attr.live_range.op = irn; @@ -1822,7 +2047,7 @@ luke_blockwalker(ir_node * bb, void * data) /* 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) { @@ -1851,11 +2076,11 @@ luke_blockwalker(ir_node * bb, void * data) /* new live range for each used value */ ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn); - prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0); + prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, is_before_frame(bb, irn)?1.0:0.0); /* the epilog stuff - including post_use, check_post, check_post_remat */ ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn); - post_use = lpp_add_var(si->lpp, buf, lpp_binary, 0.0); + post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0); lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0); @@ -1865,7 +2090,7 @@ luke_blockwalker(ir_node * bb, void * data) /* 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); @@ -1877,21 +2102,20 @@ luke_blockwalker(ir_node * bb, void * data) 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; @@ -1904,9 +2128,29 @@ luke_blockwalker(ir_node * bb, void * data) DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn)); ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++); - cst = lpp_add_cst(si->lpp, buf, lpp_greater, 0.0); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0); + lpp_set_factor_fast(si->lpp, cst, post_use, -1.0); + lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0); + } + } + } + + if(opt_memoperands) { + for(n = get_irn_arity(irn)-1; n>=0; --n) { + if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) { + ilp_var_t memoperand; + + ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n); + memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_memoperand*execution_frequency(si, bb), 0.0); + set_insert_memoperand(si->memoperands, irn, n, memoperand); + + ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0); + + lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0); lpp_set_factor_fast(si->lpp, cst, post_use, 1.0); - lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0); +// if(arg_op->attr.live_range.ilp != ILP_UNDEF) +// lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0); } } } @@ -1936,47 +2180,84 @@ luke_blockwalker(ir_node * bb, void * data) /* check the register pressure in the prolog */ /* sum_{L\U} lr <= k - |U| */ ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn); - check_pre = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - u); + check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u); /* for the prolog remove defined values from the live set */ pset_foreach(defs, tmp) { pset_remove_ptr(live, tmp); } + if(opt_memoperands) { + ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn); + one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0); + } + /*********************************************************** * I T E R A T I O N O V E R A R G S F O R P R O L O G **********************************************************/ set_foreach(args, keyval) { - spill_t *spill; - ir_node *arg = keyval->key; - int i = PTR_TO_INT(keyval->val); - op_t *arg_op = get_irn_link(arg); + spill_t *spill; + const ir_node *arg = keyval->key; + int i = PTR_TO_INT(keyval->val); + op_t *arg_op = get_irn_link(arg); + ilp_cst_t requirements; + int n_memoperands; spill = set_find_spill(spill_bb->ilp, arg); assert(spill); ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn); - op->attr.live_range.args.reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb)); + op->attr.live_range.args.reloads[i] = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), is_before_frame(bb, irn)?0.0:1.0); /* reload <= mem_out */ ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn); - cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0); lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0); lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0); /* requirement: arg must be in register for use */ /* reload + remat + live_range == 1 */ ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg); - cst = lpp_add_cst(si->lpp, buf, lpp_equal, 1.0); + requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0); - lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0); - lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0); + lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0); + lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0); foreach_pre_remat(si, irn, tmp) { op_t *remat_op = get_irn_link(tmp); if(remat_op->attr.remat.remat->value == arg) { - lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0); + lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0); + } + } + + if(opt_memoperands) { + n_memoperands = 0; + for(n = get_irn_arity(irn)-1; n>=0; --n) { + if(get_irn_n(irn, n) == arg) { + n_memoperands++; + } + } + for(n = get_irn_arity(irn)-1; n>=0; --n) { + if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) { + memoperand_t *memoperand; + memoperand = set_find_memoperand(si->memoperands, irn, n); + + /* memoperand <= mem_out */ + ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0); + lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0); + lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0); + + /* the memoperand is only sufficient if it is used once by the op */ + if(n_memoperands == 1) + lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0); + + lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0); + + /* we have one more free register if we use a memory operand */ + lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0); + } } } } @@ -2000,7 +2281,6 @@ luke_blockwalker(ir_node * bb, void * data) 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; @@ -2064,19 +2344,21 @@ luke_blockwalker(ir_node * bb, void * data) set_foreach(spill_bb->ilp, spill) { ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb); - cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0); lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0); lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0); if(pset_find_ptr(live, spill->irn)) { + int default_spilled; DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb)); ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb); - spill->mem_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0); + default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn); + spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled); lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0); - if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) { + if(opt_memcopies && is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) { int n; op_t *op = get_irn_link(spill->irn); @@ -2105,7 +2387,7 @@ luke_blockwalker(ir_node * bb, void * data) /* copies are not for free */ ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn); - var = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE * freq); + var = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_spill * freq, 1.0); for(m=n; m>=0; --m) { const ir_node *arg2 = get_irn_n(spill->irn, m); @@ -2115,16 +2397,60 @@ luke_blockwalker(ir_node * bb, void * data) } } +#if 0 /* copy <= mem_in */ ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn); - cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0); lpp_set_factor_fast(si->lpp, cst, var, 1.0); lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0); +#endif } } } } + foreach_post_remat(bb, tmp) { + int n; + + for (n=get_irn_arity(tmp)-1; n>=0; --n) { + ir_node *remat_arg = get_irn_n(tmp, n); + + if(!has_reg_class(si, remat_arg)) continue; + + /* if value is becoming live through use by remat2 */ + if(!pset_find_ptr(live, remat_arg)) { + op_t *remat_arg_op = get_irn_link(remat_arg); + ilp_cst_t nomem; + + DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 at bb start %+F\n", remat_arg, tmp)); + + pset_insert_ptr(live, remat_arg); + spill = add_to_spill_bb(si, bb, remat_arg); + remat_arg_op->attr.live_range.ilp = ILP_UNDEF; + + /* we need reg_in and mem_in for this value; they will be referenced later */ + ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", remat_arg, bb); + spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0); + ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", remat_arg, bb); + spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0); + + + /* optimization: all memory stuff should be 0, for we do not want to insert reloads for remats */ + ir_snprintf(buf, sizeof(buf), "nomem_%N_%N", remat_arg, bb); + nomem = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0.0); + + lpp_set_factor_fast(si->lpp, nomem, spill->spill, 1.0); + if(spill_bb->reloads) { + keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg); + + if(keyval) { + ilp_var_t reload = PTR_TO_INT(keyval->val); + lpp_set_factor_fast(si->lpp, nomem, reload, 1.0); + } + } + } + } + } /* L\U is empty at bb start */ /* arg is live throughout epilog if it is reg_in into this block */ @@ -2132,8 +2458,10 @@ luke_blockwalker(ir_node * bb, void * data) /* check the register pressure at the beginning of the block * including remats */ + /* reg_in entspricht post_use */ + ir_snprintf(buf, sizeof(buf), "check_start_%N", bb); - cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs); pset_foreach(live, irn) { ilp_cst_t nospill; @@ -2142,47 +2470,19 @@ luke_blockwalker(ir_node * bb, void * data) assert(spill); ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb); - spill->reg_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0); + spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0); lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0); /* spill + mem_in <= 1 */ ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb); - nospill = lpp_add_cst(si->lpp, buf, lpp_less, 1); + nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1); lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0); lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0); - } - foreach_post_remat(bb, irn) { - op_t *remat_op = get_irn_link(irn); - - DBG((si->dbg, LEVEL_4, "\t next post remat: %+F\n", irn)); - assert(remat_op->is_remat && !remat_op->attr.remat.pre); - - lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0); - } - - /* forall post remats add requirements */ - foreach_post_remat(bb, tmp) { - int n; - - for (n=get_irn_arity(tmp)-1; n>=0; --n) { - ir_node *remat_arg = get_irn_n(tmp, n); - op_t *remat_op = get_irn_link(tmp); - - if(!has_reg_class(si, remat_arg)) continue; - - spill = set_find_spill(spill_bb->ilp, remat_arg); - assert(spill); - - /* remat <= reg_in_argument */ - ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg); - cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0); - lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0); - lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0); - } - } + } /* post_remats are NOT included in register pressure check because + they do not increase regpressure */ /* mem_in/reg_in for live_in values, especially phis and their arguments */ pset_foreach(live, irn) { @@ -2200,16 +2500,20 @@ luke_blockwalker(ir_node * bb, void * data) 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); @@ -2218,6 +2522,9 @@ luke_blockwalker(ir_node * bb, void * data) assert(spill_p); lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0); + if(opt_memcopies) + lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0); + lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0); } } @@ -2232,9 +2539,9 @@ luke_blockwalker(ir_node * bb, void * data) 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); @@ -2248,29 +2555,86 @@ luke_blockwalker(ir_node * bb, void * data) } } - /* first live ranges from reg_ins */ - pset_foreach(live, irn) { - op_t *op = get_irn_link(irn); + foreach_post_remat(bb, tmp) { + int n; - spill = set_find_spill(spill_bb->ilp, irn); - assert(spill && spill->irn == irn); + 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); - ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb); - cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0); - lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0); - lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0); + if(!has_reg_class(si, remat_arg)) continue; - foreach_post_remat(bb, tmp) { - op_t *remat_op = get_irn_link(tmp); + spill = set_find_spill(spill_bb->ilp, remat_arg); + assert(spill); - if(remat_op->attr.remat.remat->value == irn) { - lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0); - } + /* remat <= reg_in_argument */ + ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg); + cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0); + lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0); + lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0); } } - /* walk forward now and compute constraints for placing spills */ - /* this must only be done for values that are not defined in this block */ + pset_foreach(live, irn) { + const op_t *op = get_irn_link(irn); + const ir_node *remat; + int n_remats = 0; + + if(op->attr.live_range.ilp == ILP_UNDEF) continue; + + cst = ILP_UNDEF; + + foreach_post_remat(bb, remat) { + int n; + + for (n=get_irn_arity(remat)-1; n>=0; --n) { + const ir_node *arg = get_irn_n(remat, n); + + if(arg == irn) { + const op_t *remat_op = get_irn_link(remat); + + if(cst == ILP_UNDEF) { + /* \sum post_remat <= 1 + #post_remats * next(lr) */ + ir_snprintf(buf, sizeof(buf), "remat2_%N_%N_arg_%N", remat, bb, irn); + cst = lpp_add_cst(si->lpp, buf, lpp_less, 1.0); + } + lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0); + ++n_remats; + break; + } + } + } + if(n_remats) { + lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, n_remats); + } + } + + /* first live ranges from reg_ins */ + pset_foreach(live, irn) { + op_t *op = get_irn_link(irn); + + if(op->attr.live_range.ilp != ILP_UNDEF) { + + spill = set_find_spill(spill_bb->ilp, irn); + assert(spill && spill->irn == irn); + + ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0); + lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0); + lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0); + + foreach_post_remat(bb, tmp) { + op_t *remat_op = get_irn_link(tmp); + + if(remat_op->attr.remat.remat->value == irn) { + lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0); + } + } + } + } + + /* walk forward now and compute constraints for placing spills */ + /* this must only be done for values that are not defined in this block */ /* TODO are these values at start of block? if yes, just check whether this is a diverge edge and skip the loop */ pset_foreach(live, irn) { /* @@ -2284,7 +2648,7 @@ luke_blockwalker(ir_node * bb, void * data) 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); @@ -2297,10 +2661,10 @@ luke_blockwalker(ir_node * bb, void * data) 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 { @@ -2311,14 +2675,13 @@ luke_blockwalker(ir_node * bb, void * data) 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); @@ -2383,7 +2746,7 @@ set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, i } 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; @@ -2403,7 +2766,7 @@ values_interfere_in_block(ir_node * bb, ir_node * a, ir_node * b) /* 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) { @@ -2411,6 +2774,7 @@ values_interfere_in_block(ir_node * bb, ir_node * a, ir_node * b) if(get_nodes_block(user) == bb && !is_Phi(user) && b != user + && !pset_find_ptr(si->inverse_ops, user) && value_dominates(b, user)) return 1; } @@ -2425,26 +2789,29 @@ static void 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; + if(a_op->is_remat || pset_find_ptr(si->inverse_ops, a)) + continue; + + for(l2=_be_lv_next_irn(si->lv, bb, 0xff, l1+1); l2>=0; l2=_be_lv_next_irn(si->lv, bb, 0xff, l2+1)) { + ir_node *b = be_lv_get_irn(si->lv, bb, l2); op_t *b_op = get_irn_link(b); - if(b_op->is_remat) continue; /* a and b are only interesting if they are in the same phi class */ if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) { - if(values_interfere_in_block(bb, a, b)) { + if(b_op->is_remat || pset_find_ptr(si->inverse_ops, b)) + continue; + + if(values_interfere_in_block(si, bb, a, b)) { DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b)); set_insert_interference(si, si->interferences, a, b, bb); } @@ -2465,7 +2832,7 @@ write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere) 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); @@ -2479,18 +2846,21 @@ write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere) * @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) { @@ -2498,19 +2868,49 @@ find_copy_path(spill_ilp_t * si, ir_node * irn, ir_node * target, ilp_var_t any_ 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 */ @@ -2522,27 +2922,55 @@ find_copy_path(spill_ilp_t * si, ir_node * irn, ir_node * target, ilp_var_t any_ 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(); @@ -2567,8 +2995,6 @@ memcopyhandler(spill_ilp_t * si) 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; @@ -2581,8 +3007,8 @@ memcopyhandler(spill_ilp_t * si) /* any_interf <= \sum interf */ ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b); - any_interfere_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0); - any_interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0); + any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0); + any_interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0); lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0); @@ -2590,15 +3016,12 @@ memcopyhandler(spill_ilp_t * si) 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); @@ -2607,10 +3030,10 @@ memcopyhandler(spill_ilp_t * si) /* 2: - mem_in_a - spill_a + interfere <= 0 */ /* 3: - mem_in_b - spill_b + interfere <= 0 */ ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b); - interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0); + interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0); ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b); - cst = lpp_add_cst(si->lpp, buf, lpp_less, 1); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1); lpp_set_factor_fast(si->lpp, cst, interfere, -1.0); if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0); @@ -2619,14 +3042,14 @@ memcopyhandler(spill_ilp_t * si) 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); @@ -2637,7 +3060,7 @@ memcopyhandler(spill_ilp_t * si) /* 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); @@ -2649,31 +3072,12 @@ memcopyhandler(spill_ilp_t * si) } - -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) { return fabs(x) < 0.00001; } -#ifdef KEEPALIVE static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l) { spill_ilp_t *si = get_irg_link(current_ir_graph); @@ -2714,7 +3118,6 @@ dump_graph_with_remats(ir_graph * irg, const char * suffix) be_dump(irg, suffix, dump_ir_block_graph_sched); set_dump_node_vcgattr_hook(NULL); } -#endif /** * Edge hook to dump the schedule edges with annotated register pressure. @@ -2737,12 +3140,12 @@ sched_pressure_edge_hook(FILE *F, ir_node *irn) 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 @@ -2750,15 +3153,14 @@ walker_pressure_annotator(ir_node * bb, void * data) { 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); } } @@ -2795,7 +3197,6 @@ dump_pressure_graph(spill_ilp_t * si, const char *suffix) be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure); } -#ifdef KEEPALIVE static void connect_all_remats_with_keep(spill_ilp_t * si) { @@ -2820,7 +3221,6 @@ connect_all_remats_with_keep(spill_ilp_t * si) obstack_free(si->obst, ins); } } -#endif static void connect_all_spills_with_keep(spill_ilp_t * si) @@ -2849,7 +3249,7 @@ connect_all_spills_with_keep(spill_ilp_t * si) } /** insert a spill at an arbitrary position */ -ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert, ir_node *ctx) +ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert) { ir_node *bl = is_Block(insert)?insert:get_nodes_block(insert); ir_graph *irg = get_irn_irg(bl); @@ -2860,7 +3260,7 @@ ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert, ir 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 @@ -2911,13 +3311,13 @@ clean_remat_info(spill_ilp_t * si) 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); } } } @@ -2930,100 +3330,97 @@ clean_remat_info(spill_ilp_t * si) static void delete_unnecessary_remats(spill_ilp_t * si) { -#ifdef KEEPALIVE - int n; - ir_node *bad = get_irg_bad(si->chordal_env->irg); + if(opt_keep_alive & KEEPALIVE_REMATS) { + int n; + ir_node *bad = get_irg_bad(si->chordal_env->irg); - if(si->keep) { - ir_node *end = get_irg_end(si->chordal_env->irg); - ir_node **keeps; + if(si->keep) { +// ir_node *end = get_irg_end(si->chordal_env->irg); +// ir_node **keeps; - for (n=get_irn_arity(si->keep)-1; n>=0; --n) { - ir_node *keep_arg = get_irn_n(si->keep, n); - op_t *arg_op = get_irn_link(keep_arg); - lpp_name_t *name; + for (n=get_irn_arity(si->keep)-1; n>=0; --n) { + ir_node *keep_arg = get_irn_n(si->keep, n); + op_t *arg_op = get_irn_link(keep_arg); + lpp_name_t *name; - assert(arg_op->is_remat); + assert(arg_op->is_remat); - name = si->lpp->vars[arg_op->attr.remat.ilp]; + name = si->lpp->vars[arg_op->attr.remat.ilp]; - if(is_zero(name->value)) { - DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg)); - /* TODO check whether reload is preferred over remat (could be bug) */ - delete_remat(si, keep_arg); - } else { - if(!arg_op->attr.remat.remat->inverse) { - if(arg_op->attr.remat.pre) { - DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg)); - } else { - DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg)); - } + if(is_zero(name->value)) { + DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg)); + /* TODO check whether reload is preferred over remat (could be bug) */ + delete_remat(si, keep_arg); } else { - if(arg_op->attr.remat.pre) { - DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg)); + if(!arg_op->attr.remat.remat->inverse) { + if(arg_op->attr.remat.pre) { + DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg)); + } else { + DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg)); + } } else { - DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg)); + if(arg_op->attr.remat.pre) { + DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg)); + } else { + DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg)); + } } } - } - set_irn_n(si->keep, n, bad); - } + set_irn_n(si->keep, n, bad); + } #if 0 - for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) { - ir_node *end_arg = get_End_keepalive(end, i); + for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) { + ir_node *end_arg = get_End_keepalive(end, i); - if(end_arg != si->keep) { - obstack_grow(si->obst, &end_arg, sizeof(end_arg)); + if(end_arg != si->keep) { + obstack_grow(si->obst, &end_arg, sizeof(end_arg)); + } } - } - keeps = obstack_finish(si->obst); - set_End_keepalives(end, n-1, keeps); - obstack_free(si->obst, keeps); + keeps = obstack_finish(si->obst); + set_End_keepalives(end, n-1, keeps); + obstack_free(si->obst, keeps); #endif + } else { + DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n")); + } } else { - DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n")); - } -#else - ir_node *remat; + ir_node *remat; - pset_foreach(si->all_possible_remats, remat) { - op_t *remat_op = get_irn_link(remat); - lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp]; + pset_foreach(si->all_possible_remats, remat) { + op_t *remat_op = get_irn_link(remat); + lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp]; - if(is_zero(name->value)) { - DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat)); - /* TODO check whether reload is preferred over remat (could be bug) */ - delete_remat(si, remat); - } else { - if(!remat_op->attr.remat.remat->inverse) { - if(remat_op->attr.remat.pre) { - DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat)); - } else { - DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat)); - } + if(is_zero(name->value)) { + DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat)); + /* TODO check whether reload is preferred over remat (could be bug) */ + delete_remat(si, remat); } else { - if(remat_op->attr.remat.pre) { - DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat)); + if(!remat_op->attr.remat.remat->inverse) { + if(remat_op->attr.remat.pre) { + DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat)); + } else { + DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat)); + } } else { - DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat)); + if(remat_op->attr.remat.pre) { + DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat)); + } else { + DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat)); + } } } } } -#endif } static pset * -get_spills_for_value(spill_ilp_t * si, ir_node * value) +get_spills_for_value(spill_ilp_t * si, const ir_node * value) { pset *spills = pset_new_ptr_default(); -// pset *visited = pset_new_ptr_default(); -// collect_spills(si, value, spills, visited); -// del_pset(visited); - ir_node *next; - defs_t *defs; + const ir_node *next; + defs_t *defs; defs = set_find_def(si->values, value); @@ -3036,32 +3433,11 @@ get_spills_for_value(spill_ilp_t * si, ir_node * value) return spills; } -static pset * -get_remats_for_value(spill_ilp_t * si, ir_node * value) -{ - pset *remats = pset_new_ptr_default(); - - ir_node *next; - defs_t *defs; - - defs = set_find_def(si->values, value); - - if(defs && defs->remats) { - for(next = defs->remats; next; next = get_irn_link(next)) { - pset_insert_ptr(remats, next); - } - } - - return remats; -} - - /** * @param before The node after which the spill will be placed in the schedule */ -/* TODO set context properly */ static ir_node * -insert_spill(spill_ilp_t * si, ir_node * irn, ir_node * value, ir_node * before) +insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before) { defs_t *defs; ir_node *spill; @@ -3069,7 +3445,7 @@ insert_spill(spill_ilp_t * si, ir_node * irn, ir_node * value, ir_node * before) 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); @@ -3078,53 +3454,8 @@ insert_spill(spill_ilp_t * si, ir_node * irn, ir_node * value, ir_node * before) set_irn_link(spill, defs->spills); defs->spills = spill; -#ifdef KEEPALIVE_SPILLS - pset_insert_ptr(si->spills, spill); -#endif - - return spill; -} - -static ir_node * -insert_mem_copy(spill_ilp_t * si, const ir_node * bb, const ir_node * arg) -{ - ir_node *prev = sched_block_last_noncf(si, bb); - ir_node *insert_pos = sched_next(prev); - op_t *prev_op = get_irn_link(prev); - pset *remats = get_remats_for_value(si, arg); - ir_node *spill; - const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env; - - /* start from end of block and search a position for memcopy (spill) until block's last op begins */ - while(be_is_Spill(prev)) { - prev = sched_prev(prev); - } - - prev_op = get_irn_link(prev); - - while(!sched_is_end(prev) && !is_Phi(prev) - && prev_op->is_remat && prev_op->attr.remat.pre) { - - insert_pos = prev; - - if(pset_find_ptr(remats, prev)) { - insert_pos = sched_next(insert_pos); - break; - } - - do { - prev = sched_prev(prev); - } while(be_is_Spill(prev)); - - prev_op = get_irn_link(prev); - } - insert_pos = sched_prev(insert_pos); - - DBG((si->dbg, LEVEL_2, "\t inserting mem copy for value %+F after %+F\n", arg, insert_pos)); - - spill = be_spill2(arch_env, arg, insert_pos, arg); - - del_pset(remats); + if(opt_keep_alive & KEEPALIVE_SPILLS) + pset_insert_ptr(si->spills, spill); return spill; } @@ -3133,33 +3464,18 @@ insert_mem_copy(spill_ilp_t * si, const ir_node * bb, const ir_node * arg) * @param before The Phi node which has to be spilled */ static ir_node * -insert_mem_phi(spill_ilp_t * si, const ir_node * phi) +insert_mem_phi(spill_ilp_t * si, ir_node * phi) { ir_node *mem_phi; ir_node **ins; defs_t *defs; int n; - op_t *op = get_irn_link(phi); NEW_ARR_A(ir_node*, ins, get_irn_arity(phi)); -#ifndef NO_MEMCOPIES - for(n=get_irn_arity(phi)-1; n>=0; --n) { - ir_node *arg = get_irn_n(phi, n); - ir_node *bb = 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)) { - ins[n] = insert_mem_copy(si, bb, arg); - } else { - ins[n] = si->m_unknown; - } - } -#else for(n=get_irn_arity(phi)-1; n>=0; --n) { ins[n] = si->m_unknown; } -#endif mem_phi = new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M); @@ -3170,12 +3486,13 @@ insert_mem_phi(spill_ilp_t * si, const ir_node * phi) set_irn_link(mem_phi, defs->spills); defs->spills = mem_phi; +#ifdef SCHEDULE_PHIM sched_add_after(phi, mem_phi); - -#ifdef KEEPALIVE_SPILLS - pset_insert_ptr(si->spills, mem_phi); #endif + if(opt_keep_alive & KEEPALIVE_SPILLS) + pset_insert_ptr(si->spills, mem_phi); + return mem_phi; } @@ -3199,43 +3516,12 @@ insert_remat(spill_ilp_t * si, ir_node * remat) 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); idbg, 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"); @@ -3276,6 +3544,40 @@ insert_reload(spill_ilp_t * si, const ir_node * value, const ir_node * after) return reload; } +void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand) +{ + defs_t *defs; + ir_node *reload; + ir_node *value = get_irn_n(memoperand->irn, memoperand->pos); + ir_node *spill; + const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env; + + DBG((si->dbg, LEVEL_2, "\t inserting memory operand for value %+F at %+F\n", value, memoperand->irn)); + + defs = set_find_def(si->values, value); + + spill = defs->spills; + assert(spill && "no spill placed before reload"); + + reload = be_reload(arch_env, si->cls, memoperand->irn, get_irn_mode(value), spill); + + arch_perform_memory_operand(arch_env, memoperand->irn, reload, memoperand->pos); + sched_remove(reload); +} + +void insert_memoperands(spill_ilp_t * si) +{ + memoperand_t *memoperand; + lpp_name_t *name; + + set_foreach(si->memoperands, memoperand) { + name = si->lpp->vars[memoperand->ilp]; + if(!is_zero(name->value)) { + perform_memory_operand(si, memoperand); + } + } +} + static void walker_spill_placer(ir_node * bb, void * data) { spill_ilp_t *si = (spill_ilp_t*)data; @@ -3349,6 +3651,40 @@ walker_spill_placer(ir_node * bb, void * 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) { @@ -3356,9 +3692,10 @@ phim_fixer(spill_ilp_t *si) { 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; @@ -3373,17 +3710,28 @@ phim_fixer(spill_ilp_t *si) { if(!phi_m) continue; for(n=get_irn_arity(phi)-1; n>=0; --n) { - const ir_node *value = get_irn_n(phi, n); + ir_node *value = get_irn_n(phi, n); defs_t *val_defs = set_find_def(si->values, value); - ir_node *arg = get_irn_n(phi_m, n); - /* 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"); - if(is_Unknown(arg)) - set_irn_n(phi_m, n, spill); + if(opt_memcopies) { + ir_node *pred = get_Block_cfgpred_block(get_nodes_block(phi), n); + lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]]; + + if(!is_zero(name->value)) { + spill = insert_mem_copy(si, pred, value); + } else { + spill = val_defs->spills; + } + } else { + spill = val_defs->spills; + } + + assert(spill && "no spill placed before PhiM"); + set_irn_n(phi_m, n, spill); } } } @@ -3393,8 +3741,6 @@ walker_reload_placer(ir_node * bb, void * data) { 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) { @@ -3430,20 +3776,11 @@ walker_reload_placer(ir_node * bb, void * data) { prev_op = get_irn_link(prev); } -// /* insert reload before pre-remats */ -// while(!sched_is_end(prev) && !be_is_Reload(prev) //FIXME && !be_is_Spill(prev) -// && !is_Phi(prev) && prev_op->is_remat && prev_op->attr.remat.pre) { -// insert_pos = prev; -// -// prev = sched_prev(insert_pos); -// prev_op = get_irn_link(prev); -// } reload = insert_reload(si, irn, insert_pos); -#ifdef KEEPALIVE_RELOADS - pset_insert_ptr(si->spills, reload); -#endif + if(opt_keep_alive & KEEPALIVE_RELOADS) + pset_insert_ptr(si->spills, reload); } } } @@ -3498,9 +3835,8 @@ walker_reload_placer(ir_node * bb, void * data) { set_irn_n(irn, n, reload); -#ifdef KEEPALIVE_RELOADS - pset_insert_ptr(si->spills, reload); -#endif + if(opt_keep_alive & KEEPALIVE_RELOADS) + pset_insert_ptr(si->spills, reload); } } } @@ -3528,7 +3864,7 @@ static void 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; @@ -3569,7 +3905,7 @@ kill_all_unused_values_in_schedule(spill_ilp_t * si) lc_bitset_free(kh.used); } -static void +void print_irn_pset(pset * p) { ir_node *irn; @@ -3579,6 +3915,47 @@ print_irn_pset(pset * p) } } +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) { @@ -3590,9 +3967,9 @@ 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(); @@ -3613,7 +3990,7 @@ 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); @@ -3622,8 +3999,8 @@ rewire_uses(spill_ilp_t * si) /* 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(); @@ -3636,7 +4013,7 @@ rewire_uses(spill_ilp_t * si) 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); @@ -3660,8 +4037,10 @@ 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); + if(opt_memoperands) + insert_memoperands(si); + phim_fixer(si); /* clean the remat info! there are still back-edges leading there! */ clean_remat_info(si); @@ -3711,7 +4090,12 @@ walker_reload_mover(ir_node * bb, void * data) 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)); @@ -3741,25 +4125,24 @@ static void 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)) { - ir_fprintf(stderr, "Spills interfere in %+F: %+F, %+F\n", 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); } } } @@ -3777,38 +4160,13 @@ verify_phiclasses(spill_ilp_t * si) irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si); } -static void -walker_spillslotassigner(ir_node * irn, void * data) -{ - spill_ilp_t *si = (spill_ilp_t*)data; - void *cls; - - if(!be_is_Spill(irn)) return; - - /* set spill context to phi class if it has one ;) */ - - cls = get_phi_class(irn); - if(cls) - be_set_Spill_context(irn, cls); - else - be_set_Spill_context(irn, irn); -} - - -static void -assign_spillslots(spill_ilp_t * si) -{ - DBG((si->dbg, LEVEL_2, "\t calling spill slot assigner\n")); - irg_walk_graph(si->chordal_env->irg, walker_spillslotassigner, NULL, si); -} - void be_spill_remat(const be_chordal_env_t * chordal_env) { + char buf[256]; char problem_name[256]; char dump_suffix[256]; char dump_suffix2[256]; - char dump_suffix3[256]; struct obstack obst; spill_ilp_t si; @@ -3819,6 +4177,9 @@ be_spill_remat(const be_chordal_env_t * chordal_env) FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat"); DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name)); + if(opt_verify & VERIFY_DOMINANCE) + be_check_dominance(chordal_env->irg); + obstack_init(&obst); si.chordal_env = chordal_env; si.obst = &obst; @@ -3826,17 +4187,12 @@ be_spill_remat(const be_chordal_env_t * chordal_env) si.lpp = new_lpp(problem_name, lpp_minimize); si.remat_info = new_set(cmp_remat_info, 4096); si.interferences = new_set(cmp_interference, 32); + si.memoperands = new_set(cmp_memoperands, 128); si.all_possible_remats = pset_new_ptr_default(); si.spills = pset_new_ptr_default(); si.inverse_ops = pset_new_ptr_default(); -#ifndef EXECFREQ_LOOPDEPH - si.execfreqs = compute_execfreq(chordal_env->irg); -#else - si.execfreqs = NULL; -#endif -#ifdef KEEPALIVE + si.lv = chordal_env->lv; si.keep = NULL; -#endif si.n_regs = get_n_regs(&si); set_irg_link(chordal_env->irg, &si); @@ -3847,30 +4203,30 @@ be_spill_remat(const be_chordal_env_t * chordal_env) be_analyze_regpressure(chordal_env, "-pre"); -#ifdef COLLECT_REMATS - /* collect remats */ - DBG((si.dbg, LEVEL_1, "Collecting remats\n")); - irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si); -#endif + if(opt_remats) { + /* collect remats */ + DBG((si.dbg, LEVEL_1, "Collecting remats\n")); + irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si); + } /* insert possible remats */ DBG((si.dbg, LEVEL_1, "Inserting possible remats\n")); irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si); DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats))); -#ifdef KEEPALIVE - DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n")); - connect_all_remats_with_keep(&si); - /* dump graph with inserted remats */ - dump_graph_with_remats(chordal_env->irg, dump_suffix); -#endif + if(opt_keep_alive & KEEPALIVE_REMATS) { + DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n")); + connect_all_remats_with_keep(&si); + /* dump graph with inserted remats */ + dump_graph_with_remats(chordal_env->irg, dump_suffix); + } /* insert copies for phi arguments not in my regclass */ irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si); /* recompute liveness */ DBG((si.dbg, LEVEL_1, "Recomputing liveness\n")); - be_liveness(chordal_env->irg); + be_liveness_recompute(si.lv); /* build the ILP */ @@ -3881,29 +4237,44 @@ be_spill_remat(const be_chordal_env_t * chordal_env) DBG((si.dbg, LEVEL_2, "\t blockwalker\n")); irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si); -#ifndef NO_MEMCOPIES - DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n")); - memcopyhandler(&si); -#endif + if(opt_memcopies) { + DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n")); + memcopyhandler(&si); + } -#ifdef DUMP_ILP - { + if(opt_dump_flags & DUMP_PROBLEM) { FILE *f; - char buf[256]; - ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name); if ((f = fopen(buf, "wt")) != NULL) { lpp_dump_plain(si.lpp, f); fclose(f); } } -#endif + + if(opt_dump_flags & DUMP_MPS) { + FILE *f; + + ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name); + if((f = fopen(buf, "wt")) != NULL) { + mps_write_mps(si.lpp, s_mps_fixed, f); + fclose(f); + } + + ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name); + if((f = fopen(buf, "wt")) != NULL) { + mps_write_mst(si.lpp, s_mps_fixed, f); + fclose(f); + } + } + + lpp_check_startvals(si.lpp); #ifdef SOLVE - DBG((si.dbg, LEVEL_1, "\tSolving %F\n", chordal_env->irg)); -#ifdef ILP_TIMEOUT - lpp_set_time_limit(si.lpp, ILP_TIMEOUT); -#endif + DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next)); + lpp_set_time_limit(si.lpp, opt_timeout); + + if(opt_log) + lpp_set_log(si.lpp, stdout); #ifdef SOLVE_LOCAL lpp_solve_cplex(si.lpp); @@ -3915,8 +4286,7 @@ be_spill_remat(const be_chordal_env_t * chordal_env) DBG((si.dbg, LEVEL_1, "\t%s: iterations: %d, solution time: %g, objective function: %g\n", problem_name, si.lpp->iterations, si.lpp->sol_time, is_zero(si.lpp->objval)?0.0:si.lpp->objval)); -#ifdef DUMP_SOLUTION - { + if(opt_dump_flags & DUMP_SOLUTION) { FILE *f; char buf[256]; @@ -3930,7 +4300,6 @@ be_spill_remat(const be_chordal_env_t * chordal_env) fclose(f); } } -#endif writeback_results(&si); @@ -3938,19 +4307,17 @@ be_spill_remat(const be_chordal_env_t * chordal_env) kill_all_unused_values_in_schedule(&si); -#if defined(KEEPALIVE_SPILLS) || defined(KEEPALIVE_RELOADS) - be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph); -#endif + if(opt_keep_alive & (KEEPALIVE_SPILLS | KEEPALIVE_RELOADS)) + be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph); // move reloads upwards - be_liveness(chordal_env->irg); - //irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si); - //move_reloads_upward(&si); + be_liveness_recompute(si.lv); + irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si); + move_reloads_upward(&si); -#ifndef NO_MEMCOPIES - verify_phiclasses(&si); - assign_spillslots(&si); -#endif + if(opt_memcopies) { + verify_phiclasses(&si); + } irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si); @@ -3958,14 +4325,15 @@ be_spill_remat(const be_chordal_env_t * chordal_env) be_analyze_regpressure(chordal_env, "-post"); + if(opt_verify & VERIFY_DOMINANCE) + be_check_dominance(chordal_env->irg); + free_dom(chordal_env->irg); del_set(si.interferences); del_pset(si.inverse_ops); del_pset(si.all_possible_remats); + del_set(si.memoperands); del_pset(si.spills); -#ifndef EXECFREQ_LOOPDEPH - free_execfreq(si.execfreqs); -#endif free_lpp(si.lpp); obstack_free(&obst, NULL); DBG((si.dbg, LEVEL_1, "\tdone.\n"));