X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbespillremat.c;h=99caaf4987327bd7f85de324650c1d6738f14f80;hb=c1fdf770d4d000dd5cf22daead32369342c5f5d1;hp=6b01d9789569eb9fadbedccdfe0ada74830c3d13;hpb=bc33949b64462b26348c9120e81b16293550a8ae;p=libfirm diff --git a/ir/be/bespillremat.c b/ir/be/bespillremat.c index 6b01d9789..99caaf498 100644 --- a/ir/be/bespillremat.c +++ b/ir/be/bespillremat.c @@ -1,21 +1,36 @@ -/** vim: set sw=4 ts=4: - * @file bespillremat.c - * @date 2006-04-06 - * @author Adam M. Szalkowski & Sebastian Hack +/* + * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved. * - * ILP based spilling & rematerialization + * This file is part of libFirm. * - * Copyright (C) 2006 Universitaet Karlsruhe - * Released under the GPL + * This file may be distributed and/or modified under the terms of the + * GNU General Public License version 2 as published by the Free Software + * Foundation and appearing in the file LICENSE.GPL included in the + * packaging of this file. + * + * Licensees holding valid libFirm Professional Edition licenses may use + * this file in accordance with the libFirm Commercial License. + * Agreement provided with the Software. + * + * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE + * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR + * PURPOSE. + */ + +/** + * @file + * @brief ILP based spilling & rematerialization + * @author Adam M. Szalkowski + * @date 06.04.2006 + * @version $Id$ */ -#ifdef HAVE_CONFIG_H #include "config.h" -#endif #ifdef WITH_ILP #include +#include "array_t.h" #include "hashptr.h" #include "debug.h" #include "obst.h" @@ -29,76 +44,165 @@ #include "irnode_t.h" #include "ircons_t.h" #include "irloop_t.h" +#include "irnodeset.h" #include "phiclass.h" -#include "iredges.h" +#include "iredges_t.h" #include "execfreq.h" +#include "irvrfy.h" +#include "irbackedge_t.h" +#include "irprofile.h" #include +#include #include #include -//#include -#include #include "be_t.h" +#include "beirg_t.h" #include "belive_t.h" #include "besched_t.h" -#include "beirgmod.h" -#include "bearch.h" +#include "bessaconstr.h" +#include "bearch_t.h" +#include "beintlive_t.h" +#include "beabi.h" #include "benode_t.h" #include "beutil.h" #include "bespillremat.h" #include "bespill.h" #include "bepressurestat.h" - +#include "bespilloptions.h" #include "bechordal_t.h" +#include "bemodule.h" + +#include "lc_opts.h" +#include "lc_opts_enum.h" + +#define DUMP_PROBLEM 1 +#define DUMP_MPS 2 +#define DUMP_SOLUTION 4 +#define DUMP_STATS 8 +#define DUMP_PRESSURE 16 + +#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 unsigned opt_dump_flags = 0; +static int opt_log = 0; +static unsigned 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 unsigned 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 = 15.0; +static double opt_cost_remat = 1.0; + + +static const lc_opt_enum_mask_items_t dump_items[] = { + { "problem", DUMP_PROBLEM }, + { "mps", DUMP_MPS }, + { "solution", DUMP_SOLUTION }, + { "stats", DUMP_STATS }, + { "pressure", DUMP_PRESSURE }, + { 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 } +}; -#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_PRE_REMAT /* allow values to die after a pre remat */ -#define CHECK_POST_REMAT /* check pressure after post remats (conservative but otherwise we can temporarily exceed the register pressure) */ -#define NO_SINGLE_USE_REMATS /* do not repair schedule */ -//#define KEEPALIVE_SPILLS -//#define KEEPALIVE_RELOADS -#define GOODWIN_REDUCTION +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 inserted nodes", &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", &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", "only remat where rematted value was live", &opt_remat_while_live), + + LC_OPT_ENT_ENUM_MASK("dump", "dump problem, solution or statistical data", &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), + LC_OPT_LAST +}; + +//#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; - spill_env_t *senv; + be_irg_t *birg; + 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; + phi_classes_t *pc; +#ifndef SCHEDULE_PHIM + pset *phims; +#endif DEBUG_ONLY(firm_dbg_module_t * dbg); } spill_ilp_t; @@ -106,15 +210,14 @@ typedef int ilp_var_t; typedef int ilp_cst_t; typedef struct _spill_bb_t { - set *ilp; - pset *copys_needed; - ilp_var_t *reloads; + set *ilp; + set *reloads; } 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; @@ -129,21 +232,24 @@ 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 { ilp_var_t ilp; ir_node *op; /** the operation this live range belongs to */ - ilp_var_t *reloads; + union { + ilp_var_t *reloads; + ilp_var_t *copies; + } args; } live_range; } attr; } 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 { @@ -158,18 +264,24 @@ 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) { - return chordal_has_class(si->chordal_env, irn); + return arch_irn_consider_in_reg_alloc(si->cls, irn); } #if 0 @@ -190,7 +302,7 @@ static int cmp_remat(const void *a, const void *b) { const remat_t *r = a; - const remat_t *s = a; + const remat_t *s = b; return !(r == s || r->op == s->op); } @@ -200,13 +312,24 @@ cmp_spill(const void *a, const void *b, size_t size) { const spill_t *p = a; const spill_t *q = b; + (void) size; // return !(p->irn == q->irn && p->bb == q->bb); 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; + (void) size; + + 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; @@ -225,7 +348,7 @@ 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; @@ -234,7 +357,7 @@ 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; @@ -244,19 +367,41 @@ 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)); } #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s))) #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s))) #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i))) -#define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i))) +#define foreach_pre_remat(s,i) for((i)=next_pre_remat((s)); (i); (i)=next_pre_remat((i))) #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i))) static int @@ -264,6 +409,7 @@ cmp_remat_info(const void *a, const void *b, size_t size) { const remat_info_t *p = a; const remat_info_t *q = b; + (void) size; return !(p->irn == q->irn); } @@ -273,6 +419,7 @@ cmp_defs(const void *a, const void *b, size_t size) { const defs_t *p = a; const defs_t *q = b; + (void) size; return !(p->value == q->value); } @@ -282,39 +429,37 @@ cmp_keyval(const void *a, const void *b, size_t size) { const keyval_t *p = a; const keyval_t *q = b; + (void) size; return !(p->key == q->key); } 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; - } + if(ir_profile_has_data()) + return ((double)ir_profile_get_block_execcount(get_block_const(irn))) + FUDGE; + +#ifndef EXECFREQ_LOOPDEPH + return get_block_execfreq(si->birg->exec_freq, get_block_const(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) +static double get_cost(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); + return arch_get_op_estimated_cost(irn); } - } /** @@ -323,19 +468,17 @@ get_cost(const spill_ilp_t * si, const ir_node * irn) static INLINE int is_rematerializable(const spill_ilp_t * si, const ir_node * irn) { - int i, - n; - const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env; - int remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0; + int n; + int remat = (arch_irn_get_flags(irn) & arch_irn_flags_rematerializable) != 0; #if 0 if(!remat) ir_fprintf(stderr, " Node %+F is not rematerializable\n", irn); #endif - for (i = 0, n = get_irn_arity(irn); i < n && remat; ++i) { - ir_node *op = get_irn_n(irn, i); - remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem); + for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) { + ir_node *op = get_irn_n(irn, n); + remat &= has_reg_class(si, op) || arch_irn_get_flags(op) & arch_irn_flags_ignore || is_NoMem(op); // if(!remat) // ir_fprintf(stderr, " Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn); @@ -359,43 +502,42 @@ 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; } if(!is_rematerializable(si, op)) return NULL; - remat = obstack_alloc(si->obst, sizeof(*remat)); - remat->op = op; - remat->cost = get_cost(si, op); - remat->value = dest_value; - remat->proj = proj; + remat = obstack_alloc(si->obst, sizeof(*remat)); + remat->op = op; + remat->cost = (int)get_cost(op); + remat->value = dest_value; + remat->proj = proj; remat->inverse = 0; } else { arch_inverse_t inverse; - int i, - n; + int n; /* get the index of the operand we want to retrieve by the inverse op */ - for (i = 0, n = get_irn_arity(op); i < n; ++i) { - ir_node *arg = get_irn_n(op, i); + for (n = get_irn_arity(op)-1; n>=0; --n) { + ir_node *arg = get_irn_n(op, n); if(arg == dest_value) break; } - if(i == n) return NULL; + if(n<0) return NULL; - DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", i, op)); + DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", n, op)); /* else ask the backend to give an inverse op */ - if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, i, &inverse, si->obst)) { + if(arch_get_inverse(op, n, &inverse, si->obst)) { int i; DBG((si->dbg, LEVEL_4, "\t backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs)); assert(inverse.n > 0 && "inverse op should have at least one node"); - for(i=0; i=0; --i) { pset_insert_ptr(si->inverse_ops, inverse.nodes[i]); } @@ -407,7 +549,9 @@ get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL; remat->inverse = 1; - assert(is_Proj(remat->proj)); + // Matze: commented this out, this doesn't seem to be true if + // the inverse is a simple operation with only 1 result... + //assert(is_Proj(remat->proj)); } else { assert(0 && "I can not handle remats with more than 2 nodes"); } @@ -430,8 +574,7 @@ add_remat(const spill_ilp_t * si, const remat_t * remat) { remat_info_t *remat_info, query; - int i, - n; + int n; assert(remat->op); assert(remat->value); @@ -447,8 +590,8 @@ add_remat(const spill_ilp_t * si, const remat_t * remat) pset_insert(remat_info->remats, remat, HASH_PTR(remat->op)); /* insert the remat into the remats_be_operand set of each argument of the remat op */ - for (i = 0, n = get_irn_arity(remat->op); i < n; ++i) { - ir_node *arg = get_irn_n(remat->op, i); + for (n = get_irn_arity(remat->op)-1; n>=0; --n) { + ir_node *arg = get_irn_n(remat->op, n); query.irn = arg; query.remats = NULL; @@ -478,40 +621,55 @@ 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 i, - n; + 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 (i = 0, n = get_irn_arity(op); i < n; ++i) { - ir_node *arg = get_irn_n(op, i); + 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 @@ -520,6 +678,7 @@ value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_no ir_node *block; ir_node *def_block = get_nodes_block(val); int ret; + (void) si; if(val == pos) return 0; @@ -546,10 +705,9 @@ value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_no return ret; } -static INLINE ir_node * -sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb) +static INLINE ir_node *sched_block_last_noncf(const ir_node * bb) { - return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env); + return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, NULL); } /** @@ -564,6 +722,7 @@ sched_block_first_nonphi(const ir_node * bb) static int sched_skip_proj_predicator(const ir_node * irn, void * data) { + (void) data; return (is_Proj(irn)); } @@ -611,21 +770,70 @@ sched_put_after(ir_node * insert, ir_node * irn) } else { insert = sched_next_op(insert); } + sched_reset(irn); sched_add_before(insert, irn); } -static void -sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn) +static void sched_put_before(ir_node * insert, ir_node * irn) { if(is_Block(insert)) { - insert = sched_block_last_noncf(si, insert); + insert = sched_block_last_noncf(insert); } else { insert = sched_next_nonproj(insert, 0); insert = sched_prev(insert); } + sched_reset(irn); sched_add_after(insert, irn); } +static ir_node * +next_post_remat(const ir_node * irn) +{ + op_t *op; + ir_node *next; + + if(is_Block(irn)) { + next = sched_block_first_nonphi(irn); + } else { + next = sched_next_op(irn); + } + + if(sched_is_end(next)) + return NULL; + + op = get_irn_link(next); + if(op->is_remat && !op->attr.remat.pre) { + return next; + } + + return NULL; +} + + +static ir_node *next_pre_remat(const ir_node * irn) +{ + op_t *op; + ir_node *ret; + + if(is_Block(irn)) { + ret = sched_block_last_noncf(irn); + ret = sched_next(ret); + ret = sched_prev_op(ret); + } else { + ret = sched_prev_op(irn); + } + + if(sched_is_end(ret) || is_Phi(ret)) + return NULL; + + op = (op_t*)get_irn_link(ret); + if(op->is_remat && op->attr.remat.pre) { + return ret; + } + + return NULL; +} + /** * Tells you whether a @p remat can be placed before the irn @p pos */ @@ -634,12 +842,11 @@ can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * { const ir_node *op = remat->op; const ir_node *prev; - int i, - n, + int n, res = 1; if(is_Block(pos)) { - prev = sched_block_last_noncf(si, pos); + prev = sched_block_last_noncf(pos); prev = sched_next_nonproj(prev, 0); } else { prev = sched_prev_op(pos); @@ -659,18 +866,18 @@ can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * } #endif - for(i=0, n=get_irn_arity(op); i=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; @@ -721,8 +928,10 @@ 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(si->pc, copy, NULL); set_nodes_block(copy, bb); - sched_put_before(si, pos, copy); + sched_put_before(pos, copy); return copy; } @@ -738,14 +947,16 @@ 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(si->pc, copy, NULL); set_nodes_block(copy, bb); sched_put_after(pos, copy); return copy; } -static void -insert_remat_after(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live) +static ir_node * +insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live) { char buf[256]; @@ -754,17 +965,16 @@ 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); -// ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", remat->value, pos); ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos); op = obstack_alloc(si->obst, sizeof(*op)); op->is_remat = 1; 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); @@ -776,11 +986,15 @@ insert_remat_after(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, } else { proj_copy = NULL; } + + return copy; } + + return NULL; } -static void -insert_remat_before(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live) +static ir_node * +insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live) { char buf[256]; @@ -793,13 +1007,12 @@ insert_remat_before(spill_ilp_t * si, const remat_t * remat, const ir_node * pos copy = insert_copy_before(si, remat->op, pos); -// ir_snprintf(buf, sizeof(buf), "remat_%N_%N", remat->value, pos); ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos); op = obstack_alloc(si->obst, sizeof(*op)); op->is_remat = 1; 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); @@ -811,7 +1024,11 @@ insert_remat_before(spill_ilp_t * si, const remat_t * remat, const ir_node * pos } else { proj_copy = NULL; } + + return copy; } + + return NULL; } static int @@ -828,24 +1045,116 @@ 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_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 +get_live_end(spill_ilp_t * si, ir_node * bb, pset * live) +{ + ir_node *irn; + int i; + + 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)) { + pset_insert_ptr(live, irn); + } + } + + irn = sched_last(bb); + + /* all values eaten by control flow operations are also live until the end of the block */ + sched_foreach_reverse(bb, irn) { + int i; + + if (!sched_skip_cf_predicator(irn, NULL)) break; + + for(i=get_irn_arity(irn)-1; i>=0; --i) { + ir_node *arg = get_irn_n(irn,i); + + if(has_reg_class(si, arg)) { + pset_insert_ptr(live, arg); + } + } + } + /* + * find values that are used by remats at end of block + * and insert them into live set + */ + foreach_pre_remat(bb, irn) { + int n; + + for (n=get_irn_arity(irn)-1; n>=0; --n) { + ir_node *remat_arg = get_irn_n(irn, n); + + if(!has_reg_class(si, remat_arg)) continue; + + /* if value is becoming live through use by remat */ + if(!pset_find_ptr(live, remat_arg)) { + DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat at end of block %+F\n", remat_arg, irn)); + + pset_insert_ptr(live, remat_arg); + } + } + } +} + +static void +walker_regclass_copy_insertor(ir_node * irn, void * data) +{ + spill_ilp_t *si = data; + + if(is_Phi(irn) && has_reg_class(si, irn)) { + int n; + + for(n=get_irn_arity(irn)-1; n>=0; --n) { + ir_node *phi_arg = get_irn_n(irn, n); + ir_node *bb = get_Block_cfgpred_block(get_nodes_block(irn), n); + + if(!has_reg_class(si, phi_arg)) { + ir_node *copy = be_new_Copy(si->cls, si->birg->irg, bb, phi_arg); + ir_node *pos = sched_block_last_noncf(bb); + op_t *op = obstack_alloc(si->obst, sizeof(*op)); + + DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn)); + sched_add_after(pos, copy); + set_irn_n(irn, n, copy); + + op->is_remat = 0; + op->attr.live_range.args.reloads = NULL; + op->attr.live_range.ilp = ILP_UNDEF; + set_irn_link(copy, op); + } + } + } } /** @@ -856,86 +1165,66 @@ static void walker_remat_insertor(ir_node * bb, void * data) { spill_ilp_t *si = data; - spill_bb_t *spill_bb; ir_node *irn; - int i, - n; - irn_live_t *li; - pset *live = pset_new_ptr_default(); + int n, i; + pset *live; + pset *post_remats; + remat_t *remat; + + /* skip start block, no remats to do there */ + if(is_start_block(bb)) return; DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb)); - live_foreach(bb, li) { - ir_node *value = (ir_node *) li->irn; + live = pset_new_ptr_default(); + 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); } } - spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb)); - set_irn_link(bb, spill_bb); - irn = sched_last(bb); while(!sched_is_end(irn)) { ir_node *next; - op_t *op; pset *args; ir_node *arg; + pset *used; next = sched_prev(irn); - DBG((si->dbg, LEVEL_5, "\t at %+F (next: %+F)\n", irn, next)); + /* delete defined value from live set */ + if(has_reg_class(si, irn)) { + pset_remove_ptr(live, irn); + } if(is_Phi(irn) || is_Proj(irn)) { - op_t *op; - - if(has_reg_class(si, irn)) { - pset_remove_ptr(live, irn); - } - - op = obstack_alloc(si->obst, sizeof(*op)); - op->is_remat = 0; - op->attr.live_range.reloads = NULL; - op->attr.live_range.ilp = ILP_UNDEF; - set_irn_link(irn, op); - irn = next; continue; } - op = obstack_alloc(si->obst, sizeof(*op)); - op->is_remat = 0; - op->attr.live_range.ilp = ILP_UNDEF; - op->attr.live_range.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.reloads) * get_irn_arity(irn)); - memset(op->attr.live_range.reloads, 0xFF, sizeof(*op->attr.live_range.reloads) * get_irn_arity(irn)); - set_irn_link(irn, op); - args = pset_new_ptr_default(); + used = pset_new_ptr_default(); - /* collect arguments of op */ - for (i = 0, n = get_irn_arity(irn); i < n; ++i) { - ir_node *arg = get_irn_n(irn, i); + /* collect arguments of op and set args of op already live in epilog */ + for (n = get_irn_arity(irn)-1; n>=0; --n) { + ir_node *arg = get_irn_n(irn, n); pset_insert_ptr(args, arg); - } - - /* set args of op live in epilog */ - pset_foreach(args, arg) { if(has_reg_class(si, arg)) { pset_insert_ptr(live, arg); + pset_insert_ptr(used, arg); } } - /* insert all possible remats after irn */ + /* insert all possible remats before irn */ pset_foreach(args, arg) { remat_info_t *remat_info, query; - remat_t *remat; - /* continue if the operand has the wrong reg class - */ + /* continue if the operand has the wrong reg class */ if(!has_reg_class(si, arg)) continue; @@ -948,39 +1237,41 @@ walker_remat_insertor(ir_node * bb, void * data) continue; } - /* do not place post remats after jumps */ - if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue; + if(remat_info->remats) { + pset_foreach(remat_info->remats, remat) { + ir_node *remat_irn = NULL; + + DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg)); + remat_irn = insert_remat_before(si, remat, irn, live); - if(remat_info->remats_by_operand) { - pset_foreach(remat_info->remats_by_operand, remat) { - /* do not insert remats producing the same value as one of the operands */ - if(!pset_find_ptr(args, remat->value)) { - DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg)); -#ifdef REMAT_WHILE_LIVE - if(pset_find_ptr(live, remat->value)) { - insert_remat_after(si, remat, irn, live); + if(remat_irn) { + for(n=get_irn_arity(remat_irn)-1; n>=0; --n) { + ir_node *remat_arg = get_irn_n(remat_irn, n); + + /* collect args of remats which are not args of op */ + if(has_reg_class(si, remat_arg) && !pset_find_ptr(args, remat_arg)) { + pset_insert_ptr(used, remat_arg); + } } -#else - insert_remat_after(si, remat, irn, live); -#endif } } } } - /* delete defined value from live set */ - if(has_reg_class(si, irn)) { - pset_remove_ptr(live, irn); + /* do not place post remats after jumps */ + if (sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) { + del_pset(used); + del_pset(args); + break; } - /* insert all possible remats before irn */ - pset_foreach(args, arg) { + /* insert all possible remats after irn */ + post_remats = pset_new_ptr_default(); + pset_foreach(used, arg) { remat_info_t *remat_info, query; - remat_t *remat; - /* continue if the operand has the wrong reg class - */ + /* continue if the operand has the wrong reg class */ if(!has_reg_class(si, arg)) continue; @@ -993,75 +1284,167 @@ walker_remat_insertor(ir_node * bb, void * data) continue; } - if(remat_info->remats) { - pset_foreach(remat_info->remats, remat) { - DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg)); -#ifdef REMAT_WHILE_LIVE - if(pset_find_ptr(live, remat->value)) { - insert_remat_before(si, remat, irn, live); + if(remat_info->remats_by_operand) { + pset_foreach(remat_info->remats_by_operand, remat) { + /* do not insert remats producing the same value as one of the operands */ + if(!pset_find_ptr(args, remat->value)) { + DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg)); + + /* only remat values that can be used by real ops */ + if(!opt_remat_while_live || pset_find_ptr(live, remat->value)) { + pset_insert_ptr(post_remats, remat); + } } -#else - insert_remat_before(si, remat, irn, live); -#endif } } } + pset_foreach(post_remats, remat) { + insert_remat_after(si, remat, irn, live); + } + del_pset(post_remats); + del_pset(used); del_pset(args); 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; - - 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)); - - insert_remat_before(si, remat, bb, NULL); - } - } - } - } - 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)) { - remat_info_t *remat_info, - query; - remat_t *remat; + get_live_end(si, bb, live_out); - query.irn = value; - query.remats = NULL; - query.remats_by_operand = NULL; - remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value)); + /* add remats at end of block */ + pset_foreach(live_out, value) { + remat_info_t *remat_info, + query; - 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)); + 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)); + + insert_remat_before(si, remat, bb, live_out); + } + } + } + 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 phis to live_in */ + sched_foreach(bb, value) { + if(!is_Phi(value)) break; + + if(has_reg_class(si, value)) { + pset_insert_ptr(live_in, value); + } + } + + /* add remat2s at beginning of block */ + post_remats = pset_new_ptr_default(); + pset_foreach(live_in, value) { + remat_info_t *remat_info, + query; + + query.irn = value; + query.remats = NULL; + query.remats_by_operand = NULL; + remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value)); - /* put the remat here if all its args are available */ - insert_remat_after(si, remat, bb, NULL); + 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 and result is still live */ + if(!opt_remat_while_live || pset_find_ptr(live_in, remat->value)) { + pset_insert_ptr(post_remats, remat); } } } } + pset_foreach(post_remats, remat) { + insert_remat_after(si, remat, bb, live_in); + } + del_pset(post_remats); + del_pset(live_in); + } +} + +static int +can_be_copied(const ir_node * bb, const ir_node * irn) +{ + const ir_edge_t *edge = get_block_succ_first(bb); + const ir_node *next_bb = edge->src; + int pos = edge->pos; + const ir_node *phi; + + assert(is_merge_edge(bb)); + + sched_foreach(next_bb, phi) { + const ir_node *phi_arg; + + if(!is_Phi(phi)) break; + + phi_arg = get_irn_n(phi, pos); + + if(phi_arg == irn) { + return 1; + } + } + return 0; +} + +/** + * Initialize additional node info + */ +static void +luke_initializer(ir_node * bb, void * data) +{ + spill_ilp_t *si = (spill_ilp_t*)data; + spill_bb_t *spill_bb; + ir_node *irn; + + spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb)); + set_irn_link(bb, spill_bb); + + sched_foreach(bb, irn) { + op_t *op; + + op = obstack_alloc(si->obst, sizeof(*op)); + op->is_remat = 0; + op->attr.live_range.ilp = ILP_UNDEF; + if(is_Phi(irn)) { + if(opt_memcopies) { + op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn)); + memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn)); + } + } else if(!is_Proj(irn)) { + op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn)); + memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn)); + } else { + op->attr.live_range.args.reloads = NULL; + } + set_irn_link(irn, op); } } + /** * Preparation of blocks' ends for Luke Blockwalker(tm)(R) */ @@ -1069,128 +1452,162 @@ 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)) { - op_t *op; - + 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)) { pset_insert_ptr(live, irn); - op = get_irn_link(irn); - assert(!op->is_remat); + } + } + /* + * find values that are used by remats at end of block + * and insert them into live set + */ + foreach_pre_remat(bb, irn) { + int n; + + for (n=get_irn_arity(irn)-1; n>=0; --n) { + ir_node *remat_arg = get_irn_n(irn, n); + + if(has_reg_class(si, remat_arg)) { + pset_insert_ptr(live, remat_arg); + } } } /* collect values used by cond jumps etc. at bb end (use_end) -> always live */ - /* their reg_out is unimportant because it can always be set */ + /* their reg_out must always be set */ sched_foreach_reverse(bb, irn) { - int i, - n; + int n; - if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break; + if (!sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) break; + + for (n=get_irn_arity(irn)-1; n>=0; --n) { + ir_node *irn_arg = get_irn_n(irn, n); - for (i = 0, n = get_irn_arity(irn); i < n; ++i) { - ir_node *irn_arg = get_irn_n(irn, i); if(has_reg_class(si, irn_arg)) { - pset_insert_ptr(use_end, irn); + pset_insert_ptr(use_end, irn_arg); } } } ir_snprintf(buf, sizeof(buf), "check_end_%N", bb); - 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, 16); + spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end)); - 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)) { - spill_t query, - *spill; + /* if this is a merge edge we can reload at the end of this block */ + if(is_merge_edge(bb)) { + spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end)); + } else if(pset_count(use_end)){ + spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end)); + } else { + spill_bb->reloads = NULL; + } - query.irn = irn; - spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn)); + pset_foreach(live,irn) { + spill_t query, + *spill; + double spill_cost; + int default_spilled; - ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb); - spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0); - /* if irn is used at the end of the block, then it is live anyway */ - if(!pset_find_ptr(use_end, irn)) - 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); + /* handle values used by control flow nodes later separately */ + if(pset_find_ptr(use_end, irn)) continue; - ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb); - spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE*execution_frequency(si, bb)); + query.irn = irn; + spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn)); - spill->reg_in = ILP_UNDEF; - spill->mem_in = ILP_UNDEF; - } - } + spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb); - del_pset(live); - del_pset(use_end); -} + ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb); + 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); -static ir_node * -next_post_remat(const ir_node * irn) -{ - op_t *op; + ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb); + spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0); - if(is_Block(irn)) { - irn = sched_block_first_nonphi(irn); - } else { - irn = sched_next_op(irn); - } + ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb); + /* 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(sched_is_end(irn)) - return NULL; + if(is_merge_edge(bb)) { + ilp_var_t reload; + ilp_cst_t rel_cst; - op = (op_t*)get_irn_link(irn); - if(op->is_remat && !op->attr.remat.pre) { - return irn; + ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn); + reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), can_be_copied(bb, irn)); + set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload)); + + /* reload <= mem_out */ + 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; } - return NULL; -} + pset_foreach(use_end,irn) { + spill_t query, + *spill; + double spill_cost; + 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)); -static ir_node * -next_pre_remat(const spill_ilp_t * si, const ir_node * irn) -{ - op_t *op; - ir_node *ret; + spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb); - if(is_Block(irn)) { - ret = sched_block_last_noncf(si, irn); - ret = sched_next(ret); - ret = sched_prev_op(ret); - } else { - ret = sched_prev_op(irn); - } + ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb); + spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0); - if(sched_is_end(ret) || is_Phi(ret)) - return NULL; + ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb); + spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0); - op = (op_t*)get_irn_link(ret); - if(op->is_remat && op->attr.remat.pre) { - return ret; + ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb); + 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_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_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; + + ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb); + 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); } - return NULL; + del_pset(live); + del_pset(use_end); } +#ifndef NDEBUG /** * Find a remat of value @p value in the epilog of @p pos */ @@ -1218,25 +1635,7 @@ find_post_remat(const ir_node * value, const ir_node * pos) return NULL; } - -/** - * Find a remat of value @p value in the prolog of @p pos - */ -static ir_node * -find_pre_remat(const spill_ilp_t * si, const ir_node * value, const ir_node * pos) -{ - while((pos = next_pre_remat(si,pos)) != NULL) { - op_t *op; - - op = get_irn_link(pos); - assert(op->is_remat && op->attr.remat.pre); - - if(op->attr.remat.remat->value == value) - return (ir_node*)pos; - } - - return NULL; -} +#endif static spill_t * add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn) @@ -1245,10 +1644,13 @@ 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:opt_cost_spill*execution_frequency(si, bb); + spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn)); spill->reg_out = ILP_UNDEF; @@ -1256,313 +1658,527 @@ 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, COST_STORE*execution_frequency(si, bb)); + 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; } /** - * Walk all irg blocks and emit this ILP + * Inserts ILP-constraints and variables for memory copying before the given position */ static void -luke_blockwalker(ir_node * bb, void * data) +insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block) { - spill_ilp_t *si = (spill_ilp_t*)data; - ir_node *irn; - irn_live_t *li; - pset *live; - char buf[256]; - ilp_cst_t cst; - spill_bb_t *spill_bb = get_irn_link(bb); - int i; - ir_node *tmp; - spill_t *spill; + const ir_node *succ; + const ir_edge_t *edge; + spill_bb_t *spill_bb = get_irn_link(block); + ir_node *phi; + int pos; + ilp_cst_t cst; + ilp_var_t copyreg; + char buf[256]; + ir_node *tmp; - live = pset_new_ptr_default(); + assert(edges_activated(current_ir_graph)); - /* do something at the end of the block */ + edge = get_block_succ_first(block); + if(!edge) return; - /* init live values at end of block */ - live_foreach(bb, li) { - ir_node *irn = (ir_node *) li->irn; + succ = edge->src; + pos = edge->pos; - if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) { - pset_insert_ptr(live, irn); - } - } + edge = get_block_succ_next(block, edge); + /* next block can only contain phis, if this is a merge edge */ + if(edge) return; - if(is_merge_edge(bb)) { - spill_bb->reloads = obstack_alloc(si->obst, pset_count(live) * sizeof(*spill_bb->reloads)); - memset(spill_bb->reloads, 0xFF, pset_count(live) * sizeof(*spill_bb->reloads)); - } else { - spill_bb->reloads = NULL; - } + ir_snprintf(buf, sizeof(buf), "copyreg_%N", block); + copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0); - i=0; - live_foreach(bb, li) { - ir_node *irn = (ir_node *) li->irn; - op_t *op; + ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs); - if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) { - spill = set_find_spill(spill_bb->ilp, irn); - assert(spill); + pset_foreach(live, tmp) { + spill_t *spill; +#if 0 + op_t *op = get_irn_link(irn); + lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0); +#endif + spill = set_find_spill(spill_bb->ilp, tmp); + assert(spill); - if(spill_bb->reloads) { - ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn); - spill_bb->reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb)); + lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0); + } + lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0); - /* reload <= mem_out */ - cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0); - lpp_set_factor_fast(si->lpp, cst, spill_bb->reloads[i], 1.0); - lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0); - } + sched_foreach(succ, phi) { + const ir_node *to_copy; + op_t *to_copy_op; + spill_t *to_copy_spill; + op_t *phi_op = get_irn_link(phi); + ilp_var_t reload = ILP_UNDEF; - op = get_irn_link(irn); - 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.op = bb; + if(!is_Phi(phi)) break; + if(!has_reg_class(si, phi)) continue; - ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn); - cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0); + to_copy = get_irn_n(phi, pos); + to_copy_op = get_irn_link(to_copy); - /* reg_out - reload - remat - live_range <= 0 */ - lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0); - if(spill_bb->reloads) lpp_set_factor_fast(si->lpp, cst, spill_bb->reloads[i], -1.0); - lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0); - foreach_pre_remat(si, 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); - } + to_copy_spill = set_find_spill(spill_bb->ilp, to_copy); + assert(to_copy_spill); + + if(spill_bb->reloads) { + keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy); + + if(keyval) { + reload = PTR_TO_INT(keyval->val); } + } - ++i; + ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0); + + /* copy - reg_out - reload - remat - live_range <= 0 */ + lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0); + lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0); + if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0); + lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0); + foreach_pre_remat(block, tmp) { + op_t *remat_op = get_irn_link(tmp); + if(remat_op->attr.remat.remat->value == to_copy) { + lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0); + } } + + ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0); + + /* copy - reg_out - copyreg <= 0 */ + lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0); + lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0); + lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0); } - DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb)); +} - sched_foreach_reverse(bb, irn) { - op_t *op; - op_t *tmp_op; - int n, - k, - d = 0; - ilp_cst_t check_pre, - check_post; -#ifdef CHECK_POST_REMAT - ilp_cst_t check_post_remat; -#endif - set *args = new_set(cmp_keyval, get_irn_arity(irn)); - keyval_t *keyval; - if(is_Phi(irn)) - break; +/** + * Walk all irg blocks and emit this ILP + */ +static void +luke_blockwalker(ir_node * bb, void * data) +{ + spill_ilp_t *si = (spill_ilp_t*)data; + ir_node *irn; + pset *live; + char buf[256]; + ilp_cst_t cst; + spill_bb_t *spill_bb = get_irn_link(bb); + ir_node *tmp; + spill_t *spill; + pset *defs = pset_new_ptr_default(); - op = get_irn_link(irn); - /* skip remats */ - if(op->is_remat) continue; - DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn)); + live = pset_new_ptr_default(); - if(has_reg_class(si, irn)) { - assert(pset_find_ptr(live, irn)); - pset_remove_ptr(live, irn); - } + /**************************************** + * B A S I C B L O C K E N D + ***************************************/ - /* init set of irn's arguments */ - for (i = 0, n = get_irn_arity(irn); i < n; ++i) { - ir_node *irn_arg = get_irn_n(irn, i); - if(has_reg_class(si, irn_arg)) { - set_insert_keyval(args, irn_arg, (void*)i); - } - } -#ifdef CHECK_POST_REMAT - /* check the register pressure after the epilog */ - ir_snprintf(buf, sizeof(buf), "check_post_remat_%N", irn); - check_post_remat = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs); + /* init live values at end of block */ + get_live_end(si, bb, live); - /* iterate over L\U */ - pset_foreach(live, tmp) { - if(!set_find_keyval(args, tmp)) { - /* if a live value is not used by irn */ - tmp_op = get_irn_link(tmp); -// assert(tmp_op->attr.live_range.op != irn); - lpp_set_factor_fast(si->lpp, check_post_remat, tmp_op->attr.live_range.ilp, 1.0); + pset_foreach(live, irn) { + op_t *op; + ilp_var_t reload = ILP_UNDEF; + + spill = set_find_spill(spill_bb->ilp, irn); + assert(spill); + + if(spill_bb->reloads) { + keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn); + + if(keyval) { + reload = PTR_TO_INT(keyval->val); } } - /* iterate over following remats and remove possibly defined values again from check_post_remat */ - foreach_post_remat(irn, tmp) { - op_t *remat_op = get_irn_link(tmp); - const ir_node *value = remat_op->attr.remat.remat->value; - op_t *val_op = get_irn_link(value); - assert(remat_op->is_remat && !remat_op->attr.remat.pre); + op = get_irn_link(irn); + assert(!op->is_remat); - /* values that are defined by remat2s are not counted */ -#ifdef REMAT_WHILE_LIVE - assert(val_op->attr.live_range.ilp); - lpp_set_factor_fast(si->lpp, check_post_remat, val_op->attr.live_range.ilp, 0.0); -#else - if(val_op->attr.live_range.ilp != ILP_UNDEF) { - lpp_set_factor_fast(si->lpp, check_post_remat, val_op->attr.live_range.ilp, 0.0); + ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb); + 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_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); + if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0); + lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0); + foreach_pre_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); + } + } + ir_snprintf(buf, sizeof(buf), "reg_out2_%N_%N", bb, irn); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0); + + /* value may only die at bb end if it is used for a mem copy */ + /* reg_out + \sum copy - reload - remat - live_range >= 0 */ + lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0); + if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0); + lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0); + foreach_pre_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); } -#endif /* REMAT_WHILE_LIVE */ } -#endif /* CHECK_POST_REMAT */ + if(is_merge_edge(bb)) { + const ir_edge_t *edge = get_block_succ_first(bb); + const ir_node *next_bb = edge->src; + int pos = edge->pos; + const ir_node *phi; + sched_foreach(next_bb, phi) { + const ir_node *phi_arg; - /* new live ranges for values from L\U defined by remat2s or used by remats */ - pset_foreach(live, tmp) { - ir_node *value = tmp;//remat_op->attr.remat.remat->value; - op_t *value_op = get_irn_link(value); + if(!is_Phi(phi)) break; - if(!set_find_keyval(args, value)) { - ilp_var_t prev_lr = ILP_UNDEF; - ir_node *remat; - cst = ILP_UNDEF; - - foreach_post_remat(irn, remat) { - op_t *remat_op = get_irn_link(remat); - - /* if value is being rematerialized by this remat */ - if(value == remat_op->attr.remat.remat->value) { - if(cst == ILP_UNDEF) { - /* next_live_range <= prev_live_range + sum remat2s */ - ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn); - cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0); - ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn); - prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0); - lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0); - lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0); - } + phi_arg = get_irn_n(phi, pos); - lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0); - } - } + if(phi_arg == irn) { + op_t *phi_op = get_irn_link(phi); + ilp_var_t copy = phi_op->attr.live_range.args.copies[pos]; -#ifdef MAY_DIE_AT_PRE_REMAT - if(cst == ILP_UNDEF) { - foreach_pre_remat(si, irn, remat) { - int i, - n; - - for (i = 0, n = get_irn_arity(remat); i < n; ++i) { - ir_node *remat_arg = get_irn_n(remat, i); - - /* if value is being used by this remat */ - if(value == remat_arg) { - /* next_live_range <= prev_live_range */ - ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn); - prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0); - - ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn); - cst = lpp_add_cst(si->lpp, buf, lpp_less, 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); - goto fertig; - } - /* TODO check afterwards whether lr dies after a pre-remat (should not happen) */ + lpp_set_factor_fast(si->lpp, cst, copy, 1.0); + } + } + } + } + + if(opt_memcopies) + insert_mem_copy_position(si, live, bb); + + /* + * assure the remat args are available + */ + foreach_pre_remat(bb, tmp) { + op_t *remat_op = get_irn_link(tmp); + int n; + + 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); + + if(!has_reg_class(si, remat_arg)) continue; + + spill = set_find_spill(spill_bb->ilp, remat_arg); + assert(spill); + + /* arguments of remats have to be live until the very end of the block + * remat = reg_out(remat_arg) and (reload(remat_arg) or live_range(remat_arg)), + * no remats, they could be in wrong order + */ + + ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg); + cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0); + + lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 3.0); + lpp_set_factor_fast(si->lpp, cst, spill->reg_out, -2.0); + lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0); + + /* use reload placed for this argument */ + 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, cst, reload, -1.0); + } + } + } + } + DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb)); + + + + + /************************************** + * B A S I C B L O C K B O D Y + **************************************/ + + sched_foreach_reverse_from(sched_block_last_noncf(bb), irn) { + op_t *op; + op_t *tmp_op; + int n, + u = 0, + d = 0; + ilp_cst_t check_pre, + check_post; + set *args; + pset *used; + pset *remat_defs; + keyval_t *keyval; + ilp_cst_t one_memoperand = -1; + + /* iterate only until first phi */ + if(is_Phi(irn)) + break; + + op = get_irn_link(irn); + /* skip remats */ + if(op->is_remat) continue; + + DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn)); + + /* collect defined values */ + if(has_reg_class(si, irn)) { + pset_insert_ptr(defs, irn); + } + + /* skip projs */ + if(is_Proj(irn)) continue; + + /* + * init set of irn's arguments + * and all possibly used values around this op + * and values defined by post remats + */ + args = new_set(cmp_keyval, get_irn_arity(irn)); + used = pset_new_ptr(pset_count(live) + get_irn_arity(irn)); + remat_defs = pset_new_ptr(pset_count(live)); + + if(!is_start_block(bb) || !be_is_Barrier(irn)) { + for (n=get_irn_arity(irn)-1; n>=0; --n) { + ir_node *irn_arg = get_irn_n(irn, n); + if(has_reg_class(si, irn_arg)) { + set_insert_keyval(args, irn_arg, (void*)n); + pset_insert_ptr(used, irn_arg); + } + } + foreach_post_remat(irn, tmp) { + op_t *remat_op = get_irn_link(tmp); + + pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value); + + for (n=get_irn_arity(tmp)-1; n>=0; --n) { + ir_node *remat_arg = get_irn_n(tmp, n); + if(has_reg_class(si, remat_arg)) { + pset_insert_ptr(used, remat_arg); + } + } + } + foreach_pre_remat(irn, tmp) { + for (n=get_irn_arity(tmp)-1; n>=0; --n) { + ir_node *remat_arg = get_irn_n(tmp, n); + if(has_reg_class(si, remat_arg)) { + pset_insert_ptr(used, remat_arg); + } + } + } + } + + /********************************** + * I N E P I L O G O F irn + **********************************/ + + /* ensure each dying value is used by only one post remat */ + pset_foreach(used, tmp) { + ir_node *value = tmp; + op_t *value_op = get_irn_link(value); + ir_node *remat; + int n_remats = 0; + + cst = ILP_UNDEF; + foreach_post_remat(irn, remat) { + op_t *remat_op = get_irn_link(remat); + + for(n=get_irn_arity(remat)-1; n>=0; --n) { + ir_node *remat_arg = get_irn_n(remat, n); + + /* if value is used by this remat add it to constraint */ + if(remat_arg == value) { + if(n_remats == 0) { + /* sum remat2s <= 1 + n_remats*live_range */ + ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0); } + + n_remats++; + lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0); + break; } } -fertig: -#endif + } + + if(pset_find_ptr(live, value) && cst != ILP_UNDEF) { + lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats); + } + } + + /* ensure at least one value dies at post remat */ + foreach_post_remat(irn, tmp) { + op_t *remat_op = get_irn_link(tmp); + pset *remat_args = pset_new_ptr(get_irn_arity(tmp)); + ir_node *remat_arg; + + for(n=get_irn_arity(tmp)-1; n>=0; --n) { + remat_arg = get_irn_n(tmp, n); + + if(has_reg_class(si, remat_arg)) { + + /* does arg always die at this op? */ + if(!pset_find_ptr(live, remat_arg)) + goto skip_one_must_die; + + pset_insert_ptr(remat_args, remat_arg); + } + } + + /* remat + \sum live_range(remat_arg) <= |args| */ + ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args)); + lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0); + + pset_foreach(remat_args, remat_arg) { + op_t *arg_op = get_irn_link(remat_arg); + + lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0); + } + +skip_one_must_die: + del_pset(remat_args); + } + + /* new live ranges for values from L\U defined by post remats */ + pset_foreach(live, tmp) { + ir_node *value = tmp; + op_t *value_op = get_irn_link(value); + + if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) { + ilp_var_t prev_lr = ILP_UNDEF; + ir_node *remat; + + if(pset_find_ptr(remat_defs, value)) { + + /* next_live_range <= prev_live_range + sum remat2s */ + ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0); + + ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn); + prev_lr = lpp_add_var_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); + + foreach_post_remat(irn, remat) { + op_t *remat_op = get_irn_link(remat); + + /* if value is being rematerialized by this remat */ + if(value == remat_op->attr.remat.remat->value) { + lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0); + } + } - if(prev_lr != ILP_UNDEF) { value_op->attr.live_range.ilp = prev_lr; value_op->attr.live_range.op = irn; } } } - /* get count of values in my register class defined by irn */ - /* also add defined values to check_post_remat; do this before iterating over args */ - if(get_irn_mode(irn) == mode_T) { - ir_node *proj = sched_next(irn); - op_t *proj_op = get_irn_link(proj); + /* requirements for post remats and start live ranges from L/U' for values dying here */ + foreach_post_remat(irn, tmp) { + op_t *remat_op = get_irn_link(tmp); + int n; - while(is_Proj(proj)) { - if(has_reg_class(si, proj)) { - ++d; -#ifdef CHECK_POST_REMAT - lpp_set_factor_fast(si->lpp, check_post_remat, proj_op->attr.live_range.ilp, 1.0); -#endif + 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); + + if(!has_reg_class(si, remat_arg)) continue; + + /* only for values in L\U (TODO and D?), the others are handled with post_use */ + if(!pset_find_ptr(used, remat_arg)) { + /* remat <= live_range(remat_arg) */ + ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg); + cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0); + + /* if value is becoming live through use by remat2 */ + if(!pset_find_ptr(live, remat_arg)) { + ilp_var_t lr; + + ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn); + 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; + + DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp)); + + pset_insert_ptr(live, remat_arg); + add_to_spill_bb(si, bb, remat_arg); + } + + lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0); + lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0); } - proj = sched_next(proj); - proj_op = get_irn_link(proj); - } - } else { - if(has_reg_class(si, irn)) { - d = 1; -#ifdef CHECK_POST_REMAT - lpp_set_factor_fast(si->lpp, check_post_remat, op->attr.live_range.ilp, 1.0); -#endif } } + + d = pset_count(defs); DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d)); /* count how many regs irn needs for arguments */ - k = set_count(args); + u = set_count(args); - /* check the register pressure in the prolog */ - /* sum_{L\U} lr <= n - |U| */ - ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn); - check_pre = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - k); /* 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); - set_foreach(args, keyval) { - ilp_var_t next_lr; - op_t *arg_op; + /* add L\U' to check_post */ + pset_foreach(live, tmp) { + if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) { + /* if a live value is not used by irn */ + tmp_op = get_irn_link(tmp); + lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0); + } + } + + /*********************************************************** + * I T E R A T I O N O V E R U S E S F O R E P I L O G + **********************************************************/ + + + pset_foreach(used, tmp) { + ilp_var_t prev_lr; ilp_var_t post_use; int p = 0; spill_t *spill; - ir_node *arg = keyval->key; + ir_node *arg = tmp; + op_t *arg_op = get_irn_link(arg); + ir_node *remat; spill = add_to_spill_bb(si, bb, arg); + /* new live range for each used value */ ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn); - next_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0); - - i = (int)keyval->val; - assert(iattr.live_range.reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb)); - - /* reload <= mem_out */ - cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0); - lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.reloads[i], 1.0); - lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0); + prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0); - arg_op = get_irn_link(arg); - - /* 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); - - lpp_set_factor_fast(si->lpp, cst, next_lr, 1.0); - lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.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); - } - } - - /* the epilog stuff - including post_use, post, post_remat */ + /* 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); @@ -1570,70 +2186,229 @@ fertig: if(pset_find_ptr(live, arg)) { DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn)); + /* post_use >= next_lr + remat */ ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++); - cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0); lpp_set_factor_fast(si->lpp, cst, post_use, -1.0); lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0); - -#ifdef CHECK_POST_REMAT - lpp_set_factor_fast(si->lpp, check_post_remat, arg_op->attr.live_range.ilp, 1.0); -#endif } - /*forall remat2 which use arg add a similar cst*/ - foreach_post_remat(irn, tmp) { - int i, - n; + /* forall post remat which use arg add a similar cst */ + foreach_post_remat(irn, remat) { + int n; - for (i = 0, n = get_irn_arity(tmp); i < n; ++i) { - ir_node *remat_arg = get_irn_n(tmp, i); - op_t *remat_op = get_irn_link(tmp); + for (n=get_irn_arity(remat)-1; n>=0; --n) { + ir_node *remat_arg = get_irn_n(remat, n); + op_t *remat_op = get_irn_link(remat); if(remat_arg == arg) { DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn)); + /* post_use >= remat */ 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 value is not an arg of op and not possibly defined by post remat + * then it may only die and not become live + */ + if(!set_find_keyval(args, arg)) { + /* post_use <= prev_lr */ + ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0); + lpp_set_factor_fast(si->lpp, cst, post_use, 1.0); + lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0); + + if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) { + /* next_lr <= prev_lr */ + ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0); + lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0); + lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0); + } + } + + if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) { + for(n = get_irn_arity(irn)-1; n>=0; --n) { + if (get_irn_n(irn, n) == arg && + arch_possible_memory_operand(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); } } } - /* new live range begins for each argument */ - arg_op->attr.live_range.ilp = next_lr; + /* new live range begins for each used value */ + arg_op->attr.live_range.ilp = prev_lr; arg_op->attr.live_range.op = irn; pset_insert_ptr(live, arg); } - /* start new live ranges for values used by remats */ - foreach_pre_remat(si, irn, tmp) { - int i, - n; + /* just to be sure */ + check_post = ILP_UNDEF; + + /* allow original defintions to be removed */ + if(opt_repair_schedule) { + pset_foreach(defs, tmp) { + op_t *tmp_op = get_irn_link(tmp); + spill_t *spill = set_find_spill(spill_bb->ilp, tmp); +#if 1 + ilp_var_t delete; + assert(spill); + + ir_snprintf(buf, sizeof(buf), "delete_%N", tmp); + delete = lpp_add_var_default(si->lpp, buf, lpp_binary, -1.0 * get_cost(irn) * execution_frequency(si, bb), 0.0); + + /* op may not be killed if its first live_range is 1 */ + ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0); + lpp_set_factor_fast(si->lpp, cst, delete, 1.0); + lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0); + + /* op may not be killed if it is spilled after the definition */ + ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0); + lpp_set_factor_fast(si->lpp, cst, delete, 1.0); + lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0); +#else + ilp_var_t keep; + assert(spill); + + ir_snprintf(buf, sizeof(buf), "keep_%N", tmp); + keep = lpp_add_var_default(si->lpp, buf, lpp_binary, get_cost(irn) * execution_frequency(si, bb), 1.0); + + /* op may not be killed if its first live_range is 1 */ + ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0); + lpp_set_factor_fast(si->lpp, cst, keep, 1.0); + lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, -1.0); + + /* op may not be killed if it is spilled after the definition */ + ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0); + lpp_set_factor_fast(si->lpp, cst, keep, 1.0); + lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0); +#endif + } + } else { +#if 0 + pset_foreach(defs, tmp) { + op_t *tmp_op = get_irn_link(tmp); + spill_t *spill = set_find_spill(spill_bb->ilp, tmp); + assert(spill); + + /* live_range or spill should be 1 + TODO: lr should be live until first use */ + ir_snprintf(buf, sizeof(buf), "nokillorig_%N", tmp); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 1.0); + lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0); + lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0); + } +#endif + } + - for (i = 0, n = get_irn_arity(tmp); i < n; ++i) { - ir_node *remat_arg = get_irn_n(tmp, i); - op_t *arg_op = get_irn_link(remat_arg); - ilp_var_t prev_lr; + /****************** + * P R O L O G + ******************/ + + /* 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_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 && (!is_start_block(bb) || be_is_Barrier(irn))) { + 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 + **********************************************************/ - if(!has_reg_class(si, remat_arg)) continue; - /* if value is becoming live through use by remat */ - if(!pset_find_ptr(live, remat_arg)) { - ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn); - prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0); + set_foreach(args, keyval) { + 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); - arg_op->attr.live_range.ilp = prev_lr; - arg_op->attr.live_range.op = irn; + ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn); + op->attr.live_range.args.reloads[i] = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0); - DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat %+F\n", remat_arg, tmp)); + /* reload <= mem_out */ + ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn); + 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); - /* TODO ist das hier die richtige Stelle???? */ - pset_insert_ptr(live, remat_arg); - add_to_spill_bb(si, bb, remat_arg); + /* requirement: arg must be in register for use */ + /* reload + remat + live_range == 1 */ + ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg); + requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 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(irn, tmp) { + op_t *remat_op = get_irn_link(tmp); + if(remat_op->attr.remat.remat->value == arg) { + lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0); + } + } + + if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) { + 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(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); + } } - /* TODO check afterwards whether lr dies after a pre-remat (should not happen) */ } } @@ -1642,20 +2417,17 @@ fertig: if(!set_find_keyval(args, tmp)) { /* if a live value is not used by irn */ tmp_op = get_irn_link(tmp); -// assert(tmp_op->attr.live_range.op != irn); lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0); - lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0); } } /* requirements for remats */ - foreach_pre_remat(si, irn, tmp) { + foreach_pre_remat(irn, tmp) { op_t *remat_op = get_irn_link(tmp); - int i, - n; + int n; - for (i = 0, n = get_irn_arity(tmp); i < n; ++i) { - ir_node *remat_arg = get_irn_n(tmp, i); + 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); if(!has_reg_class(si, remat_arg)) continue; @@ -1671,343 +2443,811 @@ fertig: if((keyval = set_find_keyval(args, remat_arg)) != NULL) { int index = (int)keyval->val; - lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.reloads[index], -1.0); + lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0); } } } - /* requirements for remats2 - * - * TODO unsure if this does the right thing. - * should insert values into set if they do not become live through remat and - * op - */ - foreach_post_remat(irn, tmp) { - op_t *remat_op = get_irn_link(tmp); - int i, - n; - for (i = 0, n = get_irn_arity(tmp); i < n; ++i) { - ir_node *remat_arg = get_irn_n(tmp, i); - op_t *arg_op = get_irn_link(remat_arg); - if(!has_reg_class(si, remat_arg)) continue; - /* only for values in L\U, the others are handled with post_use */ - if(!set_find_keyval(args, remat_arg)) { - /* remat <= live_rang(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); + /************************* + * D O N E W I T H O P + *************************/ - /* if value is becoming live through use by remat2 */ - if(!pset_find_ptr(live, remat_arg)) { - ilp_var_t lr; + DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn)); - ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn); - lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0); + pset_foreach(live, tmp) { + assert(has_reg_class(si, tmp)); + } - arg_op->attr.live_range.ilp = lr; - arg_op->attr.live_range.op = irn; +#ifndef NDEBUG + for (n=get_irn_arity(irn)-1; n>=0; --n) { + ir_node *arg = get_irn_n(irn, n); - DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp)); + assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op"); + } +#endif - pset_insert_ptr(live, remat_arg); - add_to_spill_bb(si, bb, remat_arg); + del_pset(remat_defs); + del_pset(used); + del_set(args); + del_pset(defs); + defs = pset_new_ptr_default(); + + /* skip everything above barrier in start block */ + if(is_start_block(bb) && be_is_Barrier(irn)) { + assert(pset_count(live) == 0); + break; + } + + } + del_pset(defs); + + + + /*************************************** + * B E G I N N I N G O F B L O C K + ***************************************/ + + + /* we are now at the beginning of the basic block, there are only \Phis in front of us */ + DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb)); + + pset_foreach(live, irn) { + assert(is_Phi(irn) || get_nodes_block(irn) != bb); + } + + /* construct mem_outs for all values */ + set_foreach(spill_bb->ilp, spill) { + ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb); + 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); + 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(opt_memcopies && is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) { + int n; + op_t *op = get_irn_link(spill->irn); + + for(n=get_irn_arity(spill->irn)-1; n>=0; --n) { + const ir_node *arg = get_irn_n(spill->irn, n); + double freq=0.0; + int m; + ilp_var_t var; + + + /* argument already done? */ + if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue; + + /* get sum of execution frequencies of blocks with the same phi argument */ + for(m=n; m>=0; --m) { + const ir_node *arg2 = get_irn_n(spill->irn, m); + + if(arg==arg2) { + freq += execution_frequency(si, get_Block_cfgpred_block(bb, m)); + } } - lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0); - lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0); + /* copies are not for free */ + ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn); + var = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_spill * freq, 1.0); + + for(m=n; m>=0; --m) { + const ir_node *arg2 = get_irn_n(spill->irn, m); + + if(arg==arg2) { + op->attr.live_range.args.copies[m] = var; + } + } + +#if 0 + /* copy <= mem_in */ + ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn); + 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 } } } + } -#ifdef CHECK_POST_REMAT - /* iterate over following remats and add them to check_post_remat */ - foreach_post_remat(irn, tmp) { - op_t *remat_op = get_irn_link(tmp); + foreach_post_remat(bb, tmp) { + int n; + op_t *remat_op = get_irn_link(tmp); + pset *remat_args = pset_new_ptr(get_irn_arity(tmp)); + ir_node *remat_arg; - assert(remat_op->is_remat && !remat_op->attr.remat.pre); + for (n=get_irn_arity(tmp)-1; n>=0; --n) { + remat_arg = get_irn_n(tmp, n); - lpp_set_factor_fast(si->lpp, check_post_remat, remat_op->attr.remat.ilp, 1.0); + if(has_reg_class(si, remat_arg)) { + pset_insert_ptr(remat_args, remat_arg); + } } -#endif + /* remat + \sum live_range(remat_arg) <= |args| */ + ir_snprintf(buf, sizeof(buf), "one_must_die_%N", 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) { + if(pset_find_ptr(live, remat_arg)) { + op_t *remat_arg_op = get_irn_link(remat_arg); + lpp_set_factor_fast(si->lpp, cst, remat_arg_op->attr.live_range.ilp, 1.0); + } + } + del_pset(remat_args); + } - DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn)); + foreach_post_remat(bb, tmp) { + int n; - pset_foreach(live, tmp) { - assert(has_reg_class(si, tmp)); + for(n=get_irn_arity(tmp)-1; n>=0; --n) { + ir_node *remat_arg = get_irn_n(tmp, n); + + /* if value is becoming live through use by remat2 */ + if(has_reg_class(si, remat_arg) && !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); + } } + } - for (i = 0, n = get_irn_arity(irn); i < n; ++i) { - ir_node *arg = get_irn_n(irn, i); + /* L\U is empty at bb start */ + /* arg is live throughout epilog if it is reg_in into this block */ - assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op"); + /* 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_uniq(si->lpp, buf, lpp_less, si->n_regs); + + pset_foreach(live, irn) { + ilp_cst_t nospill; + + spill = set_find_spill(spill_bb->ilp, irn); + assert(spill); + + ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb); + 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_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); + + } /* 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) { + int p = 0, + n; + + spill = set_find_spill(spill_bb->ilp, irn); + assert(spill && spill->irn == irn); + + if(is_Phi(irn) && get_nodes_block(irn) == bb) { + for (n=get_Phi_n_preds(irn)-1; n>=0; --n) { + ilp_cst_t mem_in, + reg_in; + ir_node *phi_arg = get_Phi_pred(irn, n); + 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_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_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); + + spill_p = set_find_spill(spill_bb_p->ilp, phi_arg); + 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); + } + } + } else { + /* else assure the value arrives on all paths in the same resource */ + + for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) { + ilp_cst_t mem_in, + reg_in; + 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; + + ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p); + 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_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); + + spill_p = set_find_spill(spill_bb_p->ilp, irn); + assert(spill_p); + + lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0); + lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0); + } } + } - del_set(args); + 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); + + ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg); + cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0); + lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0); + lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0); + } } + pset_foreach(live, irn) { + const op_t *op = get_irn_link(irn); + const ir_node *remat; + int n_remats = 0; + cst = ILP_UNDEF; - /* do something at the beginning of the block */ + foreach_post_remat(bb, remat) { + int n; - /* we are now at the beginning of the basic block, there are only \Phis in front of us */ - DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb)); + 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 remat2s <= 1 + n_remats*live_range */ + ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", irn, bb); + cst = lpp_add_cst_uniq(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(cst != ILP_UNDEF && op->attr.live_range.ilp != ILP_UNDEF) { + lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -n_remats); + } + } + /* first live ranges from reg_ins */ pset_foreach(live, irn) { - assert(is_Phi(irn) || get_nodes_block(irn) != bb); + 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); + } + } + } } - /* construct mem_outs for all values */ + /* 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) { + /* + * if value is defined in this block we can anways place the spill directly after the def + * -> no constraint necessary + */ + if(!is_Phi(irn) && get_nodes_block(irn) == bb) { + assert(0); + } + + + spill = set_find_spill(spill_bb->ilp, irn); + assert(spill); + + ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb); + 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); + + if(!is_Phi(irn)) { + sched_foreach_op(bb, tmp) { + op_t *op = get_irn_link(tmp); + + if(is_Phi(tmp)) continue; + assert(!is_Proj(tmp)); + + if(op->is_remat) { + const ir_node *value = op->attr.remat.remat->value; + + if(value == irn) { + /* 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 { + int n; + + for (n=get_irn_arity(tmp)-1; n>=0; --n) { + ir_node *arg = get_irn_n(tmp, n); + + if(arg == irn) { + /* if a value is used stop collecting remats */ + goto next_live; + } + } + } + } + } +next_live: ; + } + + del_pset(live); +} + +typedef struct _irnlist_t { + struct list_head list; + ir_node *irn; +} irnlist_t; + +typedef struct _interference_t { + struct list_head blocklist; + ir_node *a; + ir_node *b; +} interference_t; + +static int +cmp_interference(const void *a, const void *b, size_t size) +{ + const interference_t *p = a; + const interference_t *q = b; + (void) size; + + return !(p->a == q->a && p->b == q->b); +} + +static interference_t * +set_find_interference(set * set, ir_node * a, ir_node * b) +{ + interference_t query; + + query.a = (a>b)?a:b; + query.b = (a>b)?b:a; + + return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b))); +} + +static interference_t * +set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb) +{ + interference_t query, + *result; + irnlist_t *list = obstack_alloc(si->obst, sizeof(*list)); + + list->irn = bb; + + result = set_find_interference(set, a, b); + if(result) { + + list_add(&list->list, &result->blocklist); + return result; + } + + query.a = (a>b)?a:b; + query.b = (a>b)?b:a; + + result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b))); + + INIT_LIST_HEAD(&result->blocklist); + list_add(&list->list, &result->blocklist); + + return result; +} + +static +int 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; - 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); + if (get_nodes_block(a) != bb && get_nodes_block(b) != bb) { + /* both values are live in, so they interfere */ + return 1; + } - lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0); - lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0); + /* ensure a dominates b */ + if (value_dominates(b, a)) { + const ir_node *t; + t = b; + b = a; + a = t; + } + assert(get_nodes_block(b) == bb && "at least b should be defined here in this block"); - if(pset_find_ptr(live, spill->irn)) { - 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); + /* the following code is stolen from bera.c */ + if (be_is_live_end(si->lv, bb, a)) + return 1; - lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0); - } + foreach_out_edge(a, edge) { + const ir_node *user = edge->src; + if (get_nodes_block(user) == bb + && ! is_Phi(user) + && b != user + && ! pset_find_ptr(si->inverse_ops, user) + && value_dominates(b, user)) + return 1; } + return 0; +} - /* L\U is empty at bb start */ - /* arg is live throughout epilog if it is reg_in into this block */ +/** + * Walk all irg blocks and collect interfering values inside of phi classes + */ +static void +luke_interferencewalker(ir_node * bb, void * data) +{ + spill_ilp_t *si = (spill_ilp_t*)data; + int l1, l2; - /* check the register pressure at the beginning of the block - * including remats - */ - ir_snprintf(buf, sizeof(buf), "check_start_%N", bb); - cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs); + 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); - pset_foreach(live, irn) { - spill = set_find_spill(spill_bb->ilp, irn); - 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); + /* 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(si->pc, a)) { + if (a_op->is_remat || pset_find_ptr(si->inverse_ops, a)) + continue; - lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0); - } - foreach_post_remat(bb, irn) { - op_t *remat_op = get_irn_link(irn); + for (l2 = _be_lv_next_irn(si->lv, bb, 0xff, l1 + 1); l2 >= 0; l2 = _be_lv_next_irn(si->lv, bb, 0xff, l2 + 1)) { + ir_node *b = be_lv_get_irn(si->lv, bb, l2); + op_t *b_op = get_irn_link(b); - DBG((si->dbg, LEVEL_4, "\t next post remat: %+F\n", irn)); - assert(remat_op->is_remat && !remat_op->attr.remat.pre); + /* a and b are only interesting if they are in the same phi class */ + if (has_reg_class(si, b) && get_phi_class(si->pc, a) == get_phi_class(si->pc, b)) { + if (b_op->is_remat || pset_find_ptr(si->inverse_ops, b)) + continue; - lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0); + 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); + } + } + } + } } +} - /* forall remat2 add requirements */ - foreach_post_remat(bb, tmp) { - int i, - n; +static unsigned int copy_path_id = 0; - for (i = 0, n = get_irn_arity(tmp); i < n; ++i) { - ir_node *remat_arg = get_irn_n(tmp, i); - op_t *remat_op = get_irn_link(tmp); +static void +write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere) +{ + ilp_cst_t cst; + ilp_var_t copy; + char buf[256]; + void *ptr; - if(!has_reg_class(si, remat_arg)) continue; + ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++); + cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0); - spill = set_find_spill(spill_bb->ilp, remat_arg); - assert(spill); + lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0); - /* TODO verify this is placed correctly */ - ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg); - cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0); - lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0); - lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0); - } + pset_foreach(copies, ptr) { + copy = PTR_TO_INT(ptr); + lpp_set_factor_fast(si->lpp, cst, copy, -1.0); } +} - /* mem_in/reg_in for live_in values, especially phis and their arguments */ - pset_foreach(live, irn) { - int p = 0, - i, - n; +/** + * @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 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) +{ + const ir_edge_t *edge; + op_t *op = get_irn_link(irn); + pset *visited_users = pset_new_ptr_default(); + int paths = 0; - spill = set_find_spill(spill_bb->ilp, irn); - assert(spill && spill->irn == irn); + if(op->is_remat) return 0; - if(is_Phi(irn) && get_nodes_block(irn) == bb) { - for (i = 0, n = get_Phi_n_preds(irn); i < n; ++i) { - ilp_cst_t mem_in, - reg_in; - ir_node *phi_arg = get_Phi_pred(irn, i); - ir_node *bb_p = get_Block_cfgpred_block(bb, i); - spill_bb_t *spill_bb_p = get_irn_link(bb_p); - spill_t *spill_p; + pset_insert_ptr(visited, 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)) { - 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); - 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); + if(is_Phi(irn)) { + int n; + pset *visited_operands = pset_new_ptr(get_irn_arity(irn)); - 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); + /* visit all operands */ + for(n=get_irn_arity(irn)-1; n>=0; --n) { + ir_node *arg = get_irn_n(irn, n); + ilp_var_t copy = op->attr.live_range.args.copies[n]; - spill_p = set_find_spill(spill_bb_p->ilp, phi_arg); - assert(spill_p); + if(!has_reg_class(si, arg)) continue; + if(pset_find_ptr(visited_operands, arg)) continue; + pset_insert_ptr(visited_operands, arg); - lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0); - lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0); + 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); + 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; } } - } else { - /* else assure the value arrives on all paths in the same resource */ - - for (i = 0, n = get_Block_n_cfgpreds(bb); i < n; ++i) { - ilp_cst_t mem_in, - reg_in; - ir_node *bb_p = get_Block_cfgpred_block(bb, i); - spill_bb_t *spill_bb_p = get_irn_link(bb_p); - 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); - 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); - - 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); + } - spill_p = set_find_spill(spill_bb_p->ilp, irn); - assert(spill_p); + del_pset(visited_operands); + } - lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0); - lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0); + /* visit all uses which are phis */ + foreach_out_edge(irn, edge) { + ir_node *user = edge->src; + int pos = edge->pos; + op_t *op = get_irn_link(user); + ilp_var_t copy; + + 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); + 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; } } } - /* first live ranges from reg_ins */ - pset_foreach(live, irn) { - op_t *op = get_irn_link(irn); + del_pset(visited_users); + pset_remove_ptr(visited, irn); + return paths; +} - spill = set_find_spill(spill_bb->ilp, irn); - assert(spill && spill->irn == irn); +static void +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(); - 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); + find_copy_path(si, a, b, any_interfere, copies, visited); - foreach_post_remat(bb, tmp) { - op_t *remat_op = get_irn_link(tmp); + del_pset(visited); + del_pset(copies); +} - 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) { - spill = set_find_spill(spill_bb->ilp, irn); - assert(spill); +static void +memcopyhandler(spill_ilp_t * si) +{ + interference_t *interference; + char buf[256]; + /* teste Speicherwerte auf Interferenz */ - ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb); - cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0); + DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n")); + irg_block_walk_graph(si->birg->irg, luke_interferencewalker, NULL, si); - 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); + /* now lets emit the ILP unequations for the crap */ + set_foreach(si->interferences, interference) { + irnlist_t *irnlist; + ilp_var_t interfere, any_interfere; + ilp_cst_t any_interfere_cst, cst; + const ir_node *a = interference->a; + const ir_node *b = interference->b; - sched_foreach_op(bb, tmp) { - op_t *op = get_irn_link(tmp); + /* any_interf <= \sum interf */ + ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b); + 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); - if(is_Phi(tmp)) continue; - assert(!is_Proj(tmp)); + lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0); - if(op->is_remat) { - ir_node *value = op->attr.remat.remat->value; + list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) { + const ir_node *bb = irnlist->irn; + spill_bb_t *spill_bb = get_irn_link(bb); + spill_t *spilla, + *spillb; + char buf[256]; - if(value == irn) { - /* only collect remats up to the first use of a value */ - lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0); - } - } else { - int i, - n; + spilla = set_find_spill(spill_bb->ilp, a); + assert(spilla); - for (i = 0, n = get_irn_arity(tmp); i < n; ++i) { - ir_node *arg = get_irn_n(tmp, i); + spillb = set_find_spill(spill_bb->ilp, b); + assert(spillb); - if(arg == irn) { - /* if a value is used stop collecting remats */ - cst = ILP_UNDEF; - } - break; - } - } - if(cst == ILP_UNDEF) break; - } - } + /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */ + /* 1: mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */ + /* 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_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_uniq(si->lpp, buf, lpp_less, 1); - /* if a value is used by a mem-phi, then mem_in of this value is 0 (has to be spilled again into a different slot) - mem_in(phi) -> not mem_in(orig_value) TODO: how does this depend on a certain predecessor? - */ + 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); + if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0); + lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0); - /* mem_in of mem-phi has associated costs (but first one is free) */ - /* define n_mem_copies as positive integer in each predecessor block, - #mem_in into this block from predecessor block - 1 weighted with SPILL_COST*execfreq(predecessor) - TODO - */ + ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b); + 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); - del_pset(live); -} + ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b); + 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); + lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0); -#if 0 - * Speicherkopienminimierung: teste Speicherwerte auf Interferenz - * und weise Spillkontexte zu. Sorge bei Phis dafuer, dass gleiche - * Kontexte zusammenfliessen (Operanden und Ergebnis hat gleichen - * Kontext) -#endif -static INLINE int -is_zero(double x) -{ - return fabs(x) < 0.00001; -} + lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0); -#if 0 -static int -is_spilled(const spill_ilp_t * si, const live_range_t * lr) -{ - return !is_zero(lpp_get_var_sol(si->lpp, lr->in_mem_var)); + /* any_interfere >= interf */ + ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb); + 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); + } + + /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */ + gen_copy_constraints(si,a,b,any_interfere); + } } -#endif -static int -is_mem_phi(const ir_node * phi, void *data) + +static INLINE int +is_zero(double x) { - spill_ilp_t *si = data; -// return is_spilled(si, get_use_head(si, phi)->closest_use); - return 0; + return fabs(x) < 0.00001; } -#ifdef KEEPALIVE +/** + * node attribute hook for changing colors + */ static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l) { spill_ilp_t *si = get_irg_link(current_ir_graph); + (void) l; if(pset_find_ptr(si->all_possible_remats, n)) { op_t *op = (op_t*)get_irn_link(n); @@ -2045,7 +3285,6 @@ 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. @@ -2068,12 +3307,12 @@ 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 @@ -2081,16 +3320,14 @@ walker_pressure_annotator(ir_node * bb, void * data) { spill_ilp_t *si = data; ir_node *irn; - irn_live_t *li; - int i, - 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); } } @@ -2110,8 +3347,8 @@ walker_pressure_annotator(ir_node * bb, void * data) if(!is_Proj(irn)) projs = 0; - for (i = 0, n = get_irn_arity(irn); i < n; ++i) { - ir_node *arg = get_irn_n(irn, i); + for (n=get_irn_arity(irn)-1; n>=0; --n) { + ir_node *arg = get_irn_n(irn, n); if(has_reg_class(si, arg)) pset_insert_ptr(live, arg); } @@ -2124,10 +3361,9 @@ walker_pressure_annotator(ir_node * bb, void * data) static void dump_pressure_graph(spill_ilp_t * si, const char *suffix) { - be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure); + be_dump(si->birg->irg, suffix, dump_ir_block_graph_sched_pressure); } -#ifdef KEEPALIVE static void connect_all_remats_with_keep(spill_ilp_t * si) { @@ -2147,12 +3383,11 @@ connect_all_remats_with_keep(spill_ilp_t * si) ++pos; } - si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins); + si->keep = be_new_Keep(si->cls, si->birg->irg, get_irg_end_block(si->birg->irg), n_remats, ins); obstack_free(si->obst, ins); } } -#endif static void connect_all_spills_with_keep(spill_ilp_t * si) @@ -2174,25 +3409,24 @@ connect_all_spills_with_keep(spill_ilp_t * si) ++pos; } - keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins); + keep = be_new_Keep(si->cls, si->birg->irg, get_irg_end_block(si->birg->irg), n_spills, ins); obstack_free(si->obst, ins); } } /** 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) +static ir_node *be_spill2(ir_node *irn, ir_node *insert) { - ir_node *bl = is_Block(insert)?insert:get_nodes_block(insert); + ir_node *bl = is_Block(insert) ? insert : get_nodes_block(insert); ir_graph *irg = get_irn_irg(bl); - ir_node *frame = get_irg_frame(irg); - ir_node *spill; - ir_node *next; - - 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); + ir_node *frame = get_irg_frame(irg); + ir_node *spill; + ir_node *next; + const arch_register_class_t *cls = arch_get_irn_reg_class(irn, -1); + const arch_register_class_t *cls_frame = arch_get_irn_reg_class(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 @@ -2206,7 +3440,7 @@ ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert, ir * which is its default initialization (see above). */ - if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert)) + if (bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert)) insert = frame; for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert)) @@ -2218,26 +3452,24 @@ ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert, ir static void delete_remat(spill_ilp_t * si, ir_node * remat) { - int i, - n; - ir_node *bad = get_irg_bad(si->chordal_env->irg); + int n; + ir_node *bad = get_irg_bad(si->birg->irg); sched_remove(remat); /* kill links to operands */ - for (i = -1, n = get_irn_arity(remat); i < n; ++i) { - set_irn_n(remat, i, bad); + for (n=get_irn_arity(remat)-1; n>=-1; --n) { + set_irn_n(remat, n, bad); } } static void clean_remat_info(spill_ilp_t * si) { - int i, - n; + int n; remat_t *remat; remat_info_t *remat_info; - ir_node *bad = get_irg_bad(si->chordal_env->irg); + ir_node *bad = get_irg_bad(si->birg->irg); set_foreach(si->remat_info, remat_info) { if(!remat_info->remats) continue; @@ -2245,13 +3477,19 @@ 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); + if(sched_is_scheduled(remat->proj)) { + sched_remove((ir_node*)remat->proj); + } + 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 (i = -1, n = get_irn_arity(remat->op); i < n; ++i) { - set_irn_n(remat->op, i, bad); + if(sched_is_scheduled(remat->op)) { + sched_remove((ir_node*)remat->op); + } + for (n=get_irn_arity(remat->op)-1; n>=-1; --n) { + set_irn_n((ir_node*)remat->op, n, bad); } } } @@ -2264,105 +3502,119 @@ clean_remat_info(spill_ilp_t * si) static void delete_unnecessary_remats(spill_ilp_t * si) { -#ifdef KEEPALIVE - int i, - 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->birg->irg); - if(si->keep) { - ir_node *end = get_irg_end(si->chordal_env->irg); - ir_node **keeps; + if(si->keep) { + 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 (i = 0, n = get_irn_arity(si->keep); i < n; ++i) { - ir_node *keep_arg = get_irn_n(si->keep, i); - op_t *arg_op = get_irn_link(keep_arg); - lpp_name_t *name; + assert(arg_op->is_remat); + + 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)); + } + } else { + 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)); + } + } + } - assert(arg_op->is_remat); + set_irn_n(si->keep, n, bad); + } + } else { + DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n")); + } + } else { + ir_node *remat; - name = si->lpp->vars[arg_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", keep_arg)); + 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, keep_arg); + delete_remat(si, remat); } 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)); + 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", keep_arg)); + DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat)); } } else { - if(arg_op->attr.remat.pre) { - DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg)); + 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", keep_arg)); + DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat)); } } } - - set_irn_n(si->keep, i, bad); - } -#if 0 - 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)); - } } - 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 - ir_node *remat; +} + +static pset * +get_spills_for_value(spill_ilp_t * si, const ir_node * value) +{ + pset *spills = pset_new_ptr_default(); - 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]; + const ir_node *next; + defs_t *defs; - 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)); - } - } else { - 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)); - } - } + 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); } } -#endif + + return spills; +} + +static ir_node * +new_r_PhiM_nokeep(ir_graph * irg, ir_node *block, int arity, ir_node **in) +{ + ir_node *res; + + assert( get_irn_arity(block) == arity ); + + res = new_ir_node(NULL, irg, block, op_Phi, mode_M, arity, in); + res->attr.phi.u.backedge = new_backedge_arr(irg->obst, arity); + + return res; } /** * @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; - const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env; 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(irn, before); defs = set_insert_def(si->values, value); assert(defs); @@ -2371,9 +3623,8 @@ 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 + if(opt_keep_alive & KEEPALIVE_SPILLS) + pset_insert_ptr(si->spills, spill); return spill; } @@ -2382,21 +3633,20 @@ insert_spill(spill_ilp_t * si, ir_node * irn, ir_node * value, ir_node * before) * @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 i, - n; + int n; NEW_ARR_A(ir_node*, ins, get_irn_arity(phi)); - for(i=0,n=get_irn_arity(phi); im_unknown; + for(n=get_irn_arity(phi)-1; n>=0; --n) { + ins[n] = si->m_unknown; } - mem_phi = new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M); + mem_phi = new_r_PhiM_nokeep(si->birg->irg, get_nodes_block(phi), get_irn_arity(phi), ins); defs = set_insert_def(si->values, phi); assert(defs); @@ -2405,125 +3655,57 @@ 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); +#else + pset_insert_ptr(si->phims, mem_phi); +#endif -#ifdef KEEPALIVE_SPILLS - pset_insert_ptr(si->spills, mem_phi); -#endif - - return mem_phi; -} - -/** - * Add remat to list of defs, destroys link field! - */ -static void -insert_remat(spill_ilp_t * si, ir_node * remat) -{ - defs_t *defs; - op_t *remat_op = get_irn_link(remat); - - assert(remat_op->is_remat); - - defs = set_insert_def(si->values, remat_op->attr.remat.remat->value); - assert(defs); - - /* enter into the linked list */ - set_irn_link(remat, defs->remats); - 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); ispills, mem_phi); - collect_spills(si, arg, spills, visited); - } - } - } else { -// assert(0 && "Phi operand not spilled"); - } + + return mem_phi; } -#endif -static pset * -get_spills_for_value(spill_ilp_t * si, ir_node * value) +/** + * Add remat to list of defs, destroys link field! + */ +static void +insert_remat(spill_ilp_t * si, ir_node * remat) { - 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; + op_t *remat_op = get_irn_link(remat); - defs = set_find_def(si->values, value); + assert(remat_op->is_remat); - if(defs && defs->spills) { - for(next = defs->spills; next; next = get_irn_link(next)) { - pset_insert_ptr(spills, next); - } - } + defs = set_insert_def(si->values, remat_op->attr.remat.remat->value); + assert(defs); - return spills; + /* enter into the linked list */ + set_irn_link(remat, defs->remats); + defs->remats = remat; } + /** * Add reload before operation and add to list of defs */ static ir_node * -insert_reload(spill_ilp_t * si, const ir_node * value, const ir_node * after) +insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after) { defs_t *defs; ir_node *reload, *spill; - const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env; DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after)); defs = set_find_def(si->values, value); - /* get a spill of this value */ -#if 0 - if((!defs || !defs->spills) && is_Phi(value)) { - pset *spills; - - spills = get_spills_for_value(si, value); - - spill = pset_first(spills); - del_pset(spills); - if(!defs) { - defs = set_insert_def(si->values, value); - } - defs->spills = spill; - set_irn_link(spill, NULL); - } else { - spill = defs->spills; - } -#endif spill = defs->spills; assert(spill && "no spill placed before reload"); - reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill); + reload = be_reload(si->cls, after, get_irn_mode(value), spill); /* enter into the linked list */ set_irn_link(reload, defs->remats); @@ -2532,6 +3714,35 @@ 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 *value = get_irn_n(memoperand->irn, memoperand->pos); + ir_node *spill; + + 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"); + + arch_perform_memory_operand(memoperand->irn, spill, memoperand->pos); +} + +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; @@ -2556,6 +3767,13 @@ walker_spill_placer(ir_node * bb, void * data) { name = si->lpp->vars[spill->spill]; if(!is_zero(name->value)) { + /* place spill directly after definition */ + if(get_nodes_block(spill->irn) == bb) { + insert_spill(si, spill->irn, spill->irn, spill->irn); + continue; + } + + /* place spill at bb start */ if(spill->reg_in > 0) { name = si->lpp->vars[spill->reg_in]; if(!is_zero(name->value)) { @@ -2563,6 +3781,7 @@ walker_spill_placer(ir_node * bb, void * data) { continue; } } + /* place spill after a remat */ pset_insert_ptr(spills_to_do, spill->irn); } } @@ -2597,16 +3816,50 @@ 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; + + /* 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(is_Block(insert_pos)?value:insert_pos, insert_pos); + + return spill; +} + static void phim_fixer(spill_ilp_t *si) { defs_t *defs; set_foreach(si->values, defs) { const ir_node *phi = defs->value; - ir_node *phi_m = NULL; - ir_node *next = defs->spills; - int i, - 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; @@ -2620,16 +3873,29 @@ phim_fixer(spill_ilp_t *si) { } if(!phi_m) continue; - for(i=0,n=get_irn_arity(phi); i=0; --n) { + ir_node *value = get_irn_n(phi, n); defs_t *val_defs = set_find_def(si->values, value); - /* get a spill of this value */ - ir_node *spill = val_defs->spills; + /* a spill of this value */ + ir_node *spill; - assert(spill && "no spill placed before PhiM"); - set_irn_n(phi_m, i, 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); } } } @@ -2639,9 +3905,51 @@ 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) { + keyval_t *keyval; + + set_foreach(spill_bb->reloads, keyval) { + ir_node *irn = (ir_node*)keyval->key; + ilp_var_t reload = PTR_TO_INT(keyval->val); + lpp_name_t *name; + + name = si->lpp->vars[reload]; + if(!is_zero(name->value)) { + ir_node *reload; + ir_node *insert_pos = bb; + ir_node *prev = sched_block_last_noncf(bb); + op_t *prev_op = get_irn_link(prev); + + while(be_is_Spill(prev)) { + prev = sched_prev(prev); + } + + prev_op = get_irn_link(prev); + + /* insert reload before pre-remats */ + while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev) + && prev_op->is_remat && prev_op->attr.remat.pre) { + insert_pos = prev; + + do { + prev = sched_prev(prev); + } while(be_is_Spill(prev)); + + prev_op = get_irn_link(prev); + + } + + reload = insert_reload(si, irn, insert_pos); + + if(opt_keep_alive & KEEPALIVE_RELOADS) + pset_insert_ptr(si->spills, reload); + } + } + } + + /* walk and insert more reloads and collect remats */ sched_foreach_reverse(bb, irn) { op_t *op = get_irn_link(irn); @@ -2655,22 +3963,22 @@ walker_reload_placer(ir_node * bb, void * data) { } else { int n; - for (i = 0, n = get_irn_arity(irn); i < n; ++i) { - ir_node *arg = get_irn_n(irn, i); + for (n=get_irn_arity(irn)-1; n>=0; --n) { + ir_node *arg = get_irn_n(irn, n); - if(op->attr.live_range.reloads && op->attr.live_range.reloads[i] != ILP_UNDEF) { + if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) { lpp_name_t *name; - name = si->lpp->vars[op->attr.live_range.reloads[i]]; + name = si->lpp->vars[op->attr.live_range.args.reloads[n]]; if(!is_zero(name->value)) { ir_node *reload; ir_node *insert_pos = irn; - ir_node *prev = insert_pos; + ir_node *prev = sched_prev(insert_pos); op_t *prev_op; - do { + while(be_is_Spill(prev)) { prev = sched_prev(prev); - } while(be_is_Spill(prev)); + } prev_op = get_irn_link(prev); @@ -2689,66 +3997,31 @@ walker_reload_placer(ir_node * bb, void * data) { reload = insert_reload(si, arg, insert_pos); - set_irn_n(irn, i, reload); - -#ifdef KEEPALIVE_RELOADS - pset_insert_ptr(si->spills, reload); -#endif - } - } - } - } - } - - /* reloads at end of block */ - if(spill_bb->reloads) { - i=0; - live_foreach(bb, li) { - ir_node *irn = (ir_node *) li->irn; - - if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) { - lpp_name_t *name; + assert(reload && "no reload returned"); + set_irn_n(irn, n, reload); - name = si->lpp->vars[spill_bb->reloads[i]]; - if(!is_zero(name->value)) { - ir_node *reload; - ir_node *insert_pos = bb; - ir_node *prev = sched_prev(insert_pos); - op_t *prev_op = get_irn_link(prev); - - /* insert reload before pre-remats */ - while(!sched_is_end(prev) && !be_is_Reload(prev) && !be_is_Spill(prev) - && prev_op->is_remat && prev_op->attr.remat.pre) { - insert_pos = prev; - - prev = sched_prev(insert_pos); - prev_op = get_irn_link(prev); + if(opt_keep_alive & KEEPALIVE_RELOADS) + pset_insert_ptr(si->spills, reload); } - - reload = insert_reload(si, irn, insert_pos); - -#ifdef KEEPALIVE_RELOADS - pset_insert_ptr(si->spills, reload); -#endif } - ++i; } } } del_set(spill_bb->ilp); + if(spill_bb->reloads) del_set(spill_bb->reloads); } static void walker_collect_used(ir_node * irn, void * data) { - lc_bitset_t *used = data; + bitset_t *used = data; - lc_bitset_set(used, get_irn_idx(irn)); + bitset_set(used, get_irn_idx(irn)); } struct kill_helper { - lc_bitset_t *used; + bitset_t *used; spill_ilp_t *si; }; @@ -2756,47 +4029,71 @@ 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; for(irn=sched_first(bb); !sched_is_end(irn);) { ir_node *next = sched_next(irn); - int i, - n; + int n; - if(!lc_bitset_is_set(kh->used, get_irn_idx(irn))) { + if(!bitset_is_set(kh->used, get_irn_idx(irn))) { if(be_is_Spill(irn) || be_is_Reload(irn)) { - DBG((kh->si->dbg, LEVEL_1, "\t SUBOPTIMAL! %+F IS UNUSED (cost: %g)\n", irn, get_cost(kh->si, irn)*execution_frequency(kh->si, bb))); + DBG((kh->si->dbg, LEVEL_1, "\t SUBOPTIMAL! %+F IS UNUSED (cost: %g)\n", irn, get_cost(irn) * execution_frequency(kh->si, bb))); +#if 0 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?"); +#endif } sched_remove(irn); set_nodes_block(irn, bad); - for (i = 0, n = get_irn_arity(irn); i < n; ++i) { - set_irn_n(irn, i, bad); + for (n=get_irn_arity(irn)-1; n>=0; --n) { + set_irn_n(irn, n, bad); } } irn = next; } } +#ifndef SCHEDULE_PHIM +static void +kill_unused_phims(spill_ilp_t * si, struct kill_helper * kh) +{ + ir_node *phi; + ir_node *bad = get_irg_bad(si->birg->irg); + int n; + + pset_foreach(si->phims, phi) { + if(!bitset_is_set(kh->used, get_irn_idx(phi))) { + + set_nodes_block(phi, bad); + for (n=get_irn_arity(phi)-1; n>=0; --n) { + set_irn_n(phi, n, bad); + } + } + } +} +#endif + static void kill_all_unused_values_in_schedule(spill_ilp_t * si) { - struct kill_helper kh; + struct kill_helper kh; - kh.used = lc_bitset_malloc(get_irg_last_idx(si->chordal_env->irg)); + kh.used = bitset_malloc(get_irg_last_idx(si->birg->irg)); kh.si = si; - irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used); - irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh); + irg_walk_graph(si->birg->irg, walker_collect_used, NULL, kh.used); +#ifndef SCHEDULE_PHIM + kill_unused_phims(si, &kh); +#endif + irg_block_walk_graph(si->birg->irg, walker_kill_unused, NULL, &kh); - lc_bitset_free(kh.used); + bitset_free(kh.used); } -static void +void print_irn_pset(pset * p) { ir_node *irn; @@ -2806,20 +4103,64 @@ print_irn_pset(pset * p) } } +void +dump_phi_class(spill_ilp_t *si, ir_node **phiclass, const char * file) +{ + FILE *f = fopen(file, "w"); + ir_node *irn; + interference_t *interference; + int i; + + set_break(si->interferences); + + ir_fprintf(f, "digraph phiclass {\n"); + + for (i = ARR_LEN(phiclass) - 1; i >= 0; --i) { + irn = phiclass[i]; + if (is_Phi(irn)) + ir_fprintf(f, " %F%N [shape=box]\n", irn, irn); + } + + for (i = ARR_LEN(phiclass) - 1; i >= 0; --i) { + int n; + + irn = phiclass[i]; + 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(si->pc, (ir_node *)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) { - dom_front_info_t *dfi = be_compute_dominance_frontiers(si->chordal_env->irg); defs_t *defs; - pset *ignore = pset_new_ptr(1); + ir_nodeset_t ignore; - pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg)); + ir_nodeset_init(&ignore); + ir_nodeset_insert(&ignore, get_irg_end(si->birg->irg)); /* 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(); @@ -2836,11 +4177,25 @@ rewire_uses(spill_ilp_t * si) spills = get_spills_for_value(si, defs->value); DBG((si->dbg, LEVEL_2, "\t %d remats, %d reloads, and %d spills for value %+F\n", remats, pset_count(reloads), pset_count(spills), defs->value)); if(pset_count(spills) > 1) { + be_ssa_construction_env_t senv; + ir_node *node; //assert(pset_count(reloads) > 0); // print_irn_pset(spills); // print_irn_pset(reloads); - be_ssa_constr_set_ignore(dfi, spills, ignore); + be_ssa_construction_init(&senv, si->birg); + be_ssa_construction_set_ignore_uses(&senv, &ignore); + pset_foreach(spills, node) { + be_ssa_construction_add_copy(&senv, node); + } + pset_foreach(spills, node) { + be_ssa_construction_fix_users(&senv, node); + } + be_ssa_construction_update_liveness_phis(&senv, si->lv); + pset_foreach(spills, node) { + be_liveness_update(si->lv, node); + } + be_ssa_construction_destroy(&senv); } del_pset(reloads); @@ -2849,32 +4204,55 @@ 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; + const ir_node *next = defs->remats; + int orig_kept = 0; if(next) { - nodes = pset_new_ptr_default(); - pset_insert_ptr(nodes, defs->value); + be_ssa_construction_env_t senv; + + be_ssa_construction_init(&senv, si->birg); + + if(sched_is_scheduled(defs->value)) { + be_ssa_construction_add_copy(&senv, (ir_node*) defs->value); + orig_kept = 1; + } + + next = defs->remats; + while(next) { + be_ssa_construction_add_copy(&senv, (ir_node*) next); + next = get_irn_link(next); + } + + if(sched_is_scheduled(defs->value)) { + be_ssa_construction_fix_users(&senv, (ir_node*) defs->value); + } + next = defs->remats; while(next) { - pset_insert_ptr(nodes, next); + be_ssa_construction_fix_users(&senv, (ir_node*) next); next = get_irn_link(next); } - 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_construction_update_liveness_phis(&senv, si->lv); + if(sched_is_scheduled(defs->value)) { + be_liveness_update(si->lv, (ir_node*) defs->value); + } + + next = defs->remats; + while(next) { + be_liveness_update(si->lv, (ir_node*) next); + next = get_irn_link(next); } - del_pset(nodes); + be_ssa_construction_destroy(&senv); } } + ir_nodeset_destroy(&ignore); // remove_unused_defs(si); - - be_free_dominance_frontiers(dfi); } + static void writeback_results(spill_ilp_t * si) { @@ -2884,10 +4262,12 @@ writeback_results(spill_ilp_t * si) DBG((si->dbg, LEVEL_1, "Applying results\n")); 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); + si->m_unknown = new_r_Unknown(si->birg->irg, mode_M); + irg_block_walk_graph(si->birg->irg, walker_spill_placer, NULL, si); + irg_block_walk_graph(si->birg->irg, walker_reload_placer, NULL, si); + if(opt_memoperands) + insert_memoperands(si); phim_fixer(si); - irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si); /* clean the remat info! there are still back-edges leading there! */ clean_remat_info(si); @@ -2902,18 +4282,22 @@ writeback_results(spill_ilp_t * si) static int get_n_regs(spill_ilp_t * si) { - int arch_n_regs = arch_register_class_n_regs(si->cls); - int free = 0; - int i; + int arch_n_regs = arch_register_class_n_regs(si->cls); - for(i=0; icls->regs[i], ignore)) { - free++; - } - } + bitset_t *arch_regs = bitset_malloc(arch_n_regs); + bitset_t *abi_regs = bitset_malloc(arch_n_regs); + + arch_put_non_ignore_regs(si->cls, arch_regs); + be_abi_put_ignore_regs(si->birg->abi, si->cls, abi_regs); + + bitset_andnot(arch_regs, abi_regs); + arch_n_regs = bitset_popcnt(arch_regs); + + bitset_free(arch_regs); + bitset_free(abi_regs); - DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", free, si->cls->name)); - return free; + DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", arch_n_regs, si->cls->name)); + return arch_n_regs; } static void @@ -2937,7 +4321,12 @@ 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)); @@ -2956,102 +4345,177 @@ walker_reload_mover(ir_node * bb, void * data) static void move_reloads_upward(spill_ilp_t * si) { - irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si); + irg_block_walk_graph(si->birg->irg, walker_reload_mover, NULL, si); +} + + +/** + * Walk all irg blocks and check for interfering spills inside of phi classes + */ +static void +luke_meminterferencechecker(ir_node * bb, void * data) +{ + spill_ilp_t *si = (spill_ilp_t*)data; + int l1, l2; + + be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) { + ir_node *a = be_lv_get_irn(si->lv, bb, l1); + + if (! be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) + continue; + + /* a is only interesting if it is in my register class and if it is inside a phi class */ + if (has_reg_class(si, a) && get_phi_class(si->pc, a)) { + for (l2 = _be_lv_next_irn(si->lv, bb, 0xff, l1 + 1); l2 >= 0; l2 = _be_lv_next_irn(si->lv, bb, 0xff, l2 + 1)) { + ir_node *b = be_lv_get_irn(si->lv, bb, l2); + + if (! be_is_Spill(b) && (! is_Phi(b) || get_irn_mode(b) != mode_T)) + continue; + + /* a and b are only interesting if they are in the same phi class */ + if (has_reg_class(si, b) && get_phi_class(si->pc, a) == get_phi_class(si->pc, b)) { + if (values_interfere_in_block(si, bb, a, b)) { + ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b); + } + } + } + } + } +} + +static void +verify_phiclasses(spill_ilp_t * si) +{ + /* analyze phi classes */ + phi_class_free(si->pc); + si->pc = phi_class_new_from_irg(si->birg->irg, 0); + + DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n")); + irg_block_walk_graph(si->birg->irg, luke_meminterferencechecker, NULL, si); } void -be_spill_remat(const be_chordal_env_t * chordal_env) +be_spill_remat(be_irg_t *birg, const arch_register_class_t *cls) { + 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; + ir_graph *irg = be_get_birg_irg(birg); - ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name); - ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name); - ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name); - ir_snprintf(dump_suffix3, sizeof(dump_suffix3), "-%s-reloads_moved", chordal_env->cls->name); + ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", irg, cls->name); + ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", cls->name); + ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", cls->name); 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(irg); + + be_assure_liveness(birg); + obstack_init(&obst); - si.chordal_env = chordal_env; - si.obst = &obst; - si.senv = be_new_spill_env(chordal_env); - si.cls = chordal_env->cls; - si.lpp = new_lpp(problem_name, lpp_minimize); - si.remat_info = new_set(cmp_remat_info, 4096); + si.obst = &obst; + si.birg = birg; + si.cls = cls; + 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.keep = NULL; -#endif - si.n_regs = get_n_regs(&si); + si.spills = pset_new_ptr_default(); + si.inverse_ops = pset_new_ptr_default(); + si.lv = birg->lv; + si.keep = NULL; + si.n_regs = get_n_regs(&si); - set_irg_link(chordal_env->irg, &si); - compute_doms(chordal_env->irg); + set_irg_link(irg, &si); + compute_doms(irg); - be_analyze_regpressure(chordal_env, "-pre"); + /* compute phi classes */ + // phi_class_compute(irg); -#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_dump_flags & DUMP_STATS) + be_analyze_regpressure(birg, cls, "-pre"); + + DBG((si.dbg, LEVEL_2, "\t initializing\n")); + irg_block_walk_graph(irg, luke_initializer, NULL, &si); + + if(opt_remats) { + /* collect remats */ + DBG((si.dbg, LEVEL_1, "Collecting remats\n")); + irg_walk_graph(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); + irg_block_walk_graph(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(irg, dump_suffix); + } + /* insert copies for phi arguments not in my regclass */ + irg_walk_graph(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 */ - DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n")); DBG((si.dbg, LEVEL_2, "\t endwalker\n")); - irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si); + irg_block_walk_graph(irg, luke_endwalker, NULL, &si); DBG((si.dbg, LEVEL_2, "\t blockwalker\n")); - irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si); + irg_block_walk_graph(irg, luke_blockwalker, NULL, &si); -#ifdef DUMP_ILP - { - FILE *f; - char buf[256]; + si.pc = phi_class_new_from_irg(birg->irg, 0); + if (opt_memcopies) { + DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n")); + memcopyhandler(&si); + } + if (opt_dump_flags & DUMP_PROBLEM) { + FILE *f; 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); @@ -3061,10 +4525,9 @@ be_spill_remat(const be_chordal_env_t * chordal_env) assert(lpp_is_sol_valid(si.lpp) && "solution of ILP must be valid"); - 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)); + DBG((si.dbg, LEVEL_1, "\t%s: iterations: %d, solution time: %g, objective function: %g, best bound: %g\n", problem_name, si.lpp->iterations, si.lpp->sol_time, is_zero(si.lpp->objval)?0.0:si.lpp->objval, is_zero(si.lpp->best_bound)?0.0:si.lpp->best_bound)); -#ifdef DUMP_SOLUTION - { + if(opt_dump_flags & DUMP_SOLUTION) { FILE *f; char buf[256]; @@ -3078,50 +4541,76 @@ be_spill_remat(const be_chordal_env_t * chordal_env) fclose(f); } } -#endif +#ifndef SCHEDULE_PHIM + si.phims = pset_new_ptr_default(); +#endif writeback_results(&si); + #endif /* SOLVE */ 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); +#if !defined(SCHEDULE_PHIM) && defined(SOLVE) + del_pset(si.phims); #endif - be_liveness(chordal_env->irg); - irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si); + if(opt_keep_alive & (KEEPALIVE_SPILLS | KEEPALIVE_RELOADS)) + be_dump(irg, "-spills-placed", dump_ir_block_graph); - dump_pressure_graph(&si, dump_suffix2); + // move reloads upwards + be_liveness_recompute(si.lv); + irg_block_walk_graph(irg, walker_pressure_annotator, NULL, &si); + move_reloads_upward(&si); - // TODO fix temporarily exceeded regpressure due to remat2s + if(opt_memcopies) { + verify_phiclasses(&si); + } - // TODO insert copys to fix interferences in memory + irg_block_walk_graph(irg, walker_pressure_annotator, NULL, &si); - // move reloads upwards - move_reloads_upward(&si); - irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si); - dump_pressure_graph(&si, dump_suffix3); + if(opt_dump_flags & DUMP_PRESSURE) + dump_pressure_graph(&si, dump_suffix2); - be_analyze_regpressure(chordal_env, "-post"); + if(opt_dump_flags & DUMP_STATS) + be_analyze_regpressure(birg, cls, "-post"); - free_dom(chordal_env->irg); + if(opt_verify & VERIFY_DOMINANCE) + be_check_dominance(irg); + + free_dom(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); + phi_class_free(si.pc); obstack_free(&obst, NULL); -// exit(0); + DBG((si.dbg, LEVEL_1, "\tdone.\n")); +} + +void be_init_spillremat(void) +{ + static be_spiller_t remat_spiller = { + be_spill_remat + }; + lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be"); + lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra"); + lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal"); + lc_opt_entry_t *remat_grp = lc_opt_get_grp(chordal_grp, "remat"); + + be_register_spiller("remat", &remat_spiller); + lc_opt_add_table(remat_grp, options); } +BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillremat); + #else /* WITH_ILP */ -static void -only_that_you_can_compile_without_WITH_ILP_defined(void) +static __attribute__((unused)) +void only_that_you_can_compile_without_WITH_ILP_defined(void) { }