X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbespillbelady.c;h=9fcba3c3b46a14d8a61fd85bb6c4db94d51763bf;hb=320595b2d2f4b0bc4f7bea706c9c575146607dbf;hp=6f0b45a60e8e1bd7f75704eb5e3c25df46bc025e;hpb=0b1140d1887f497ef413709e5cda2608f2cd1017;p=libfirm diff --git a/ir/be/bespillbelady.c b/ir/be/bespillbelady.c index 6f0b45a60..9fcba3c3b 100644 --- a/ir/be/bespillbelady.c +++ b/ir/be/bespillbelady.c @@ -24,9 +24,7 @@ * @date 20.09.2005 * @version $Id$ */ -#ifdef HAVE_CONFIG_H #include "config.h" -#endif #include @@ -41,20 +39,19 @@ #include "ircons_t.h" #include "irprintf.h" #include "irnodeset.h" -#include "xmalloc.h" #include "beutil.h" -#include "bearch_t.h" +#include "bearch.h" #include "beuses.h" -#include "besched_t.h" +#include "besched.h" #include "beirgmod.h" #include "belive_t.h" -#include "benode_t.h" +#include "benode.h" #include "bechordal_t.h" -#include "bespilloptions.h" -#include "beloopana.h" -#include "beirg_t.h" #include "bespill.h" +#include "beloopana.h" +#include "beirg.h" +#include "bespillutil.h" #include "bemodule.h" #define DBG_SPILL 1 @@ -86,7 +83,6 @@ typedef struct _workset_t { } workset_t; static struct obstack obst; -static const arch_env_t *arch_env; static const arch_register_class_t *cls; static const be_lv_t *lv; static be_loopana_t *loop_ana; @@ -100,9 +96,9 @@ static unsigned instr_nr; /**< current instruction number static spill_env_t *senv; /**< see bespill.h */ static ir_node **blocklist; -static bool move_spills = true; -static bool respectloopdepth = true; -static bool improve_known_preds = true; +static int move_spills = true; +static int respectloopdepth = true; +static int improve_known_preds = true; /* factor to weight the different costs of reloading/rematerializing a node (see bespill.h be_get_reload_costs_no_weight) */ static int remat_bonus = 10; @@ -122,26 +118,12 @@ static int loc_compare(const void *a, const void *b) return p->time - q->time; } -void workset_print(const workset_t *w) -{ - int i; - - for(i = 0; i < w->len; ++i) { - ir_fprintf(stderr, "%+F %d\n", w->vals[i].node, w->vals[i].time); - } -} - /** * Alloc a new workset on obstack @p ob with maximum size @p max */ static workset_t *new_workset(void) { - workset_t *res; - size_t size = sizeof(*res) + n_regs * sizeof(res->vals[0]); - - res = obstack_alloc(&obst, size); - memset(res, 0, size); - return res; + return OALLOCFZ(&obst, workset_t, vals, n_regs); } /** @@ -149,10 +131,8 @@ static workset_t *new_workset(void) */ static workset_t *workset_clone(workset_t *workset) { - workset_t *res; - size_t size = sizeof(*res) + n_regs * sizeof(res->vals[0]); - res = obstack_alloc(&obst, size); - memcpy(res, workset, size); + workset_t *res = OALLOCF(&obst, workset_t, vals, n_regs); + memcpy(res, workset, sizeof(*res) + n_regs * sizeof(res->vals[0])); return res; } @@ -185,7 +165,7 @@ static void workset_insert(workset_t *workset, ir_node *val, bool spilled) loc_t *loc; int i; /* check for current regclass */ - assert(arch_irn_consider_in_reg_alloc(arch_env, cls, val)); + assert(arch_irn_consider_in_reg_alloc(cls, val)); /* check if val is already contained */ for (i = 0; i < workset->len; ++i) { @@ -218,10 +198,10 @@ static void workset_clear(workset_t *workset) /** * Removes the value @p val from the workset if present. */ -static INLINE void workset_remove(workset_t *workset, ir_node *val) +static inline void workset_remove(workset_t *workset, ir_node *val) { int i; - for(i = 0; i < workset->len; ++i) { + for (i = 0; i < workset->len; ++i) { if (workset->vals[i].node == val) { workset->vals[i] = workset->vals[--workset->len]; return; @@ -229,7 +209,7 @@ static INLINE void workset_remove(workset_t *workset, ir_node *val) } } -static INLINE const loc_t *workset_contains(const workset_t *ws, +static inline const loc_t *workset_contains(const workset_t *ws, const ir_node *val) { int i; @@ -248,7 +228,7 @@ static INLINE const loc_t *workset_contains(const workset_t *ws, * @p v A variable to put the current value in * @p i An integer for internal use */ -#define workset_foreach(ws, v, i) for(i=0; \ +#define workset_foreach(ws, v, i) for (i=0; \ v=(i < ws->len) ? ws->vals[i].node : NULL, i < ws->len; \ ++i) @@ -257,7 +237,7 @@ static INLINE const loc_t *workset_contains(const workset_t *ws, #define workset_set_length(ws, length) (ws)->len = length #define workset_get_length(ws) ((ws)->len) #define workset_get_val(ws, i) ((ws)->vals[i].node) -#define workset_sort(ws) qsort((ws)->vals, (ws)->len, sizeof((ws)->vals[0]), loc_compare); +#define workset_sort(ws) do { qsort((ws)->vals, (ws)->len, sizeof((ws)->vals[0]), loc_compare); } while(0) typedef struct _block_info_t { @@ -266,12 +246,9 @@ typedef struct _block_info_t } block_info_t; -static void *new_block_info(void) +static block_info_t *new_block_info(void) { - block_info_t *res = obstack_alloc(&obst, sizeof(res[0])); - memset(res, 0, sizeof(res[0])); - - return res; + return OALLOCZ(&obst, block_info_t); } #define get_block_info(block) ((block_info_t *)get_irn_link(block)) @@ -280,15 +257,14 @@ static void *new_block_info(void) /** * @return The distance to the next use or 0 if irn has dont_spill flag set */ -static INLINE unsigned get_distance(ir_node *from, unsigned from_step, +static inline unsigned get_distance(ir_node *from, unsigned from_step, const ir_node *def, int skip_from_uses) { be_next_use_t use; - int flags = arch_irn_get_flags(arch_env, def); unsigned costs; unsigned time; - assert(! (flags & arch_irn_flags_ignore)); + assert(!arch_irn_is_ignore(def)); use = be_get_next_use(uses, from, from_step, def, skip_from_uses); time = use.time; @@ -296,7 +272,7 @@ static INLINE unsigned get_distance(ir_node *from, unsigned from_step, return USES_INFINITY; /* We have to keep nonspillable nodes in the workingset */ - if (flags & arch_irn_flags_dont_spill) + if (arch_irn_get_flags(skip_Proj_const(def)) & arch_irn_flags_dont_spill) return 0; /* give some bonus to rematerialisable nodes */ @@ -320,8 +296,8 @@ static INLINE unsigned get_distance(ir_node *from, unsigned from_step, */ static void displace(workset_t *new_vals, int is_usage) { - ir_node **to_insert = alloca(n_regs * sizeof(to_insert[0])); - bool *spilled = alloca(n_regs * sizeof(spilled[0])); + ir_node **to_insert = ALLOCAN(ir_node*, n_regs); + bool *spilled = ALLOCAN(bool, n_regs); ir_node *val; int i; int len; @@ -360,14 +336,6 @@ static void displace(workset_t *new_vals, int is_usage) /* Only make more free room if we do not have enough */ if (spills_needed > 0) { - ir_node *curr_bb = NULL; - workset_t *ws_start = NULL; - - if (move_spills) { - curr_bb = get_nodes_block(instr); - ws_start = get_block_info(curr_bb)->start_workset; - } - DB((dbg, DBG_DECIDE, " disposing %d values\n", spills_needed)); /* calculate current next-use distance for live values */ @@ -484,13 +452,13 @@ static loc_t to_take_or_not_to_take(ir_node* first, ir_node *node, loc.node = node; loc.spilled = false; - if (!arch_irn_consider_in_reg_alloc(arch_env, cls, node)) { + if (!arch_irn_consider_in_reg_alloc(cls, node)) { loc.time = USES_INFINITY; return loc; } /* We have to keep nonspillable nodes in the workingset */ - if (arch_irn_get_flags(arch_env, node) & arch_irn_flags_dont_spill) { + if (arch_irn_get_flags(skip_Proj_const(node)) & arch_irn_flags_dont_spill) { loc.time = 0; DB((dbg, DBG_START, " %+F taken (dontspill node)\n", node, loc.time)); return loc; @@ -512,7 +480,7 @@ static loc_t to_take_or_not_to_take(ir_node* first, ir_node *node, DB((dbg, DBG_START, " %+F taken (%u, live in all preds)\n", node, loc.time)); return loc; - } else if(available == AVAILABLE_NOWHERE) { + } else if (available == AVAILABLE_NOWHERE) { DB((dbg, DBG_START, " %+F not taken (%u, live in no pred)\n", node, loc.time)); loc.time = USES_INFINITY; @@ -555,9 +523,9 @@ static void decide_start_workset(const ir_node *block) /* check predecessors */ arity = get_irn_arity(block); - pred_worksets = alloca(sizeof(pred_worksets[0]) * arity); + pred_worksets = ALLOCAN(workset_t*, arity); all_preds_known = true; - for(i = 0; i < arity; ++i) { + for (i = 0; i < arity; ++i) { ir_node *pred_block = get_Block_cfgpred_block(block, i); block_info_t *pred_info = get_block_info(pred_block); @@ -582,7 +550,7 @@ static void decide_start_workset(const ir_node *block) if (! is_Phi(node)) break; - if (!arch_irn_consider_in_reg_alloc(arch_env, cls, node)) + if (!arch_irn_consider_in_reg_alloc(cls, node)) continue; if (all_preds_known) { @@ -642,21 +610,23 @@ static void decide_start_workset(const ir_node *block) int p, arity; loc_t *loc = & delayed[i]; - /* don't use values which are dead in a known predecessors - * to not induce unnecessary reloads */ - arity = get_irn_arity(block); - for (p = 0; p < arity; ++p) { - ir_node *pred_block = get_Block_cfgpred_block(block, p); - block_info_t *pred_info = get_block_info(pred_block); - - if (pred_info == NULL) - continue; - - if (!workset_contains(pred_info->end_workset, loc->node)) { - DB((dbg, DBG_START, - " delayed %+F not live at pred %+F\n", loc->node, - pred_block)); - goto skip_delayed; + if (!is_Phi(loc->node)) { + /* don't use values which are dead in a known predecessors + * to not induce unnecessary reloads */ + arity = get_irn_arity(block); + for (p = 0; p < arity; ++p) { + ir_node *pred_block = get_Block_cfgpred_block(block, p); + block_info_t *pred_info = get_block_info(pred_block); + + if (pred_info == NULL) + continue; + + if (!workset_contains(pred_info->end_workset, loc->node)) { + DB((dbg, DBG_START, + " delayed %+F not live at pred %+F\n", loc->node, + pred_block)); + goto skip_delayed; + } } } @@ -706,7 +676,7 @@ static void decide_start_workset(const ir_node *block) /* determine spill status of the values: If there's 1 pred block (which * is no backedge) where the value is spilled then we must set it to * spilled here. */ - for(i = 0; i < ws_count; ++i) { + for (i = 0; i < ws_count; ++i) { loc_t *loc = &ws->vals[i]; ir_node *value = loc->node; bool spilled; @@ -721,7 +691,7 @@ static void decide_start_workset(const ir_node *block) /* determine if value was spilled on any predecessor */ spilled = false; - for(n = 0; n < arity; ++n) { + for (n = 0; n < arity; ++n) { workset_t *pred_workset = pred_worksets[n]; int p_len; int p; @@ -730,7 +700,7 @@ static void decide_start_workset(const ir_node *block) continue; p_len = workset_get_length(pred_workset); - for(p = 0; p < p_len; ++p) { + for (p = 0; p < p_len; ++p) { loc_t *l = &pred_workset->vals[p]; if (l->node != value) @@ -815,9 +785,9 @@ static void process_block(ir_node *block) /* allocate all values _used_ by this instruction */ workset_clear(new_vals); - for(i = 0, arity = get_irn_arity(irn); i < arity; ++i) { + for (i = 0, arity = get_irn_arity(irn); i < arity; ++i) { ir_node *in = get_irn_n(irn, i); - if (!arch_irn_consider_in_reg_alloc(arch_env, cls, in)) + if (!arch_irn_consider_in_reg_alloc(cls, in)) continue; /* (note that "spilled" is irrelevant here) */ @@ -832,12 +802,12 @@ static void process_block(ir_node *block) foreach_out_edge(irn, edge) { ir_node *proj = get_edge_src_irn(edge); - if (!arch_irn_consider_in_reg_alloc(arch_env, cls, proj)) + if (!arch_irn_consider_in_reg_alloc(cls, proj)) continue; workset_insert(new_vals, proj, false); } } else { - if (!arch_irn_consider_in_reg_alloc(arch_env, cls, irn)) + if (!arch_irn_consider_in_reg_alloc(cls, irn)) continue; workset_insert(new_vals, irn, false); } @@ -870,10 +840,14 @@ static void fix_block_borders(ir_node *block, void *data) DB((dbg, DBG_FIX, "\n")); DB((dbg, DBG_FIX, "Fixing %+F\n", block)); + arity = get_irn_arity(block); + /* can happen for endless loops */ + if (arity == 0) + return; + start_workset = get_block_info(block)->start_workset; /* process all pred blocks */ - arity = get_irn_arity(block); for (i = 0; i < arity; ++i) { ir_node *pred = get_Block_cfgpred_block(block, i); workset_t *pred_end_workset = get_block_info(pred)->end_workset; @@ -927,7 +901,7 @@ static void fix_block_borders(ir_node *block, void *data) assert(!l->spilled); /* we might have unknowns as argument for the phi */ - if (!arch_irn_consider_in_reg_alloc(arch_env, cls, node)) + if (!arch_irn_consider_in_reg_alloc(cls, node)) continue; } @@ -953,12 +927,6 @@ static void fix_block_borders(ir_node *block, void *data) } } -static void add_block(ir_node *block, void *data) -{ - (void) data; - ARR_APP1(ir_node*, blocklist, block); -} - static void be_spill_belady(be_irg_t *birg, const arch_register_class_t *rcls) { int i; @@ -982,7 +950,6 @@ static void be_spill_belady(be_irg_t *birg, const arch_register_class_t *rcls) /* init belady env */ stat_ev_tim_push(); obstack_init(&obst); - arch_env = birg->main_env->arch_env; cls = rcls; lv = be_get_birg_liveness(birg); n_regs = cls->n_regs - be_put_ignore_regs(birg, cls, NULL); @@ -990,8 +957,7 @@ static void be_spill_belady(be_irg_t *birg, const arch_register_class_t *rcls) uses = be_begin_uses(irg, lv); loop_ana = be_new_loop_pressure(birg, cls); senv = be_new_spill_env(birg); - blocklist = NEW_ARR_F(ir_node*, 0); - irg_block_edges_walk(get_irg_start_block(irg), NULL, add_block, NULL); + blocklist = be_get_cfgpostorder(irg); stat_ev_tim_pop("belady_time_init"); stat_ev_tim_push(); @@ -1020,6 +986,7 @@ static void be_spill_belady(be_irg_t *birg, const arch_register_class_t *rcls) obstack_free(&obst, NULL); } +BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady); void be_init_spillbelady(void) { static be_spiller_t belady_spiller = { @@ -1032,5 +999,3 @@ void be_init_spillbelady(void) be_register_spiller("belady", &belady_spiller); FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady"); } - -BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady);