X-Git-Url: http://git.sourceforge.jp/view?a=blobdiff_plain;f=gcc%2Fsched-rgn.c;h=004064ecf420b8bbf137d1f9919d91735c7f3983;hb=7d2bb3573bb5b43bdc4e6c47d38e1fab1552eda8;hp=47d0a9a17a30432d5e412bb5e84608201f5bd6cc;hpb=f2aa2401262b4d4cd3efa813546d9143facaf865;p=pf3gnuchains%2Fgcc-fork.git diff --git a/gcc/sched-rgn.c b/gcc/sched-rgn.c index 47d0a9a17a3..004064ecf42 100644 --- a/gcc/sched-rgn.c +++ b/gcc/sched-rgn.c @@ -1,6 +1,7 @@ /* Instruction scheduling pass. - Copyright (C) 1992, 1993, 1994, 1995, 1996, 1997, 1998, - 1999, 2000, 2001, 2002, 2003, 2004, 2005 Free Software Foundation, Inc. + Copyright (C) 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, + 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008 + Free Software Foundation, Inc. Contributed by Michael Tiemann (tiemann@cygnus.com) Enhanced by, and currently maintained by, Jim Wilson (wilson@cygnus.com) @@ -8,7 +9,7 @@ This file is part of GCC. GCC is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free -Software Foundation; either version 2, or (at your option) any later +Software Foundation; either version 3, or (at your option) any later version. GCC is distributed in the hope that it will be useful, but WITHOUT ANY @@ -17,9 +18,8 @@ FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. You should have received a copy of the GNU General Public License -along with GCC; see the file COPYING. If not, write to the Free -Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA -02110-1301, USA. */ +along with GCC; see the file COPYING3. If not see +. */ /* This pass implements list scheduling within basic blocks. It is run twice: (1) after flow analysis, but before register allocation, @@ -64,100 +64,68 @@ Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA #include "cfglayout.h" #include "params.h" #include "sched-int.h" +#include "sel-sched.h" #include "target.h" #include "timevar.h" #include "tree-pass.h" - -/* Define when we want to do count REG_DEAD notes before and after scheduling - for sanity checking. We can't do that when conditional execution is used, - as REG_DEAD exist only for unconditional deaths. */ - -#if !defined (HAVE_conditional_execution) && defined (ENABLE_CHECKING) -#define CHECK_DEAD_NOTES 1 -#else -#define CHECK_DEAD_NOTES 0 -#endif - +#include "dbgcnt.h" #ifdef INSN_SCHEDULING + /* Some accessor macros for h_i_d members only used within this file. */ -#define INSN_REF_COUNT(INSN) (h_i_d[INSN_UID (INSN)].ref_count) -#define FED_BY_SPEC_LOAD(insn) (h_i_d[INSN_UID (insn)].fed_by_spec_load) -#define IS_LOAD_INSN(insn) (h_i_d[INSN_UID (insn)].is_load_insn) +#define FED_BY_SPEC_LOAD(INSN) (HID (INSN)->fed_by_spec_load) +#define IS_LOAD_INSN(INSN) (HID (insn)->is_load_insn) /* nr_inter/spec counts interblock/speculative motion for the function. */ static int nr_inter, nr_spec; static int is_cfg_nonregular (void); -static bool sched_is_disabled_for_current_region_p (void); - -/* A region is the main entity for interblock scheduling: insns - are allowed to move between blocks in the same region, along - control flow graph edges, in the 'up' direction. */ -typedef struct -{ - /* Number of extended basic blocks in region. */ - int rgn_nr_blocks; - /* cblocks in the region (actually index in rgn_bb_table). */ - int rgn_blocks; - /* Dependencies for this region are already computed. Basically, indicates, - that this is a recovery block. */ - unsigned int dont_calc_deps : 1; - /* This region has at least one non-trivial ebb. */ - unsigned int has_real_ebb : 1; -} -region; /* Number of regions in the procedure. */ -static int nr_regions; +int nr_regions = 0; /* Table of region descriptions. */ -static region *rgn_table; +region *rgn_table = NULL; /* Array of lists of regions' blocks. */ -static int *rgn_bb_table; +int *rgn_bb_table = NULL; /* Topological order of blocks in the region (if b2 is reachable from b1, block_to_bb[b2] > block_to_bb[b1]). Note: A basic block is always referred to by either block or b, while its topological order name (in the region) is referred to by bb. */ -static int *block_to_bb; +int *block_to_bb = NULL; /* The number of the region containing a block. */ -static int *containing_rgn; +int *containing_rgn = NULL; + +/* ebb_head [i] - is index in rgn_bb_table of the head basic block of i'th ebb. + Currently we can get a ebb only through splitting of currently + scheduling block, therefore, we don't need ebb_head array for every region, + hence, its sufficient to hold it for current one only. */ +int *ebb_head = NULL; /* The minimum probability of reaching a source block so that it will be considered for speculative scheduling. */ static int min_spec_prob; -#define RGN_NR_BLOCKS(rgn) (rgn_table[rgn].rgn_nr_blocks) -#define RGN_BLOCKS(rgn) (rgn_table[rgn].rgn_blocks) -#define RGN_DONT_CALC_DEPS(rgn) (rgn_table[rgn].dont_calc_deps) -#define RGN_HAS_REAL_EBB(rgn) (rgn_table[rgn].has_real_ebb) -#define BLOCK_TO_BB(block) (block_to_bb[block]) -#define CONTAINING_RGN(block) (containing_rgn[block]) - -void debug_regions (void); -static void find_single_block_region (void); +static void find_single_block_region (bool); static void find_rgns (void); -static void extend_rgns (int *, int *, sbitmap, int *); static bool too_large (int, int *, int *); -extern void debug_live (int, int); - /* Blocks of the current region being scheduled. */ -static int current_nr_blocks; -static int current_blocks; +int current_nr_blocks; +int current_blocks; -static int rgn_n_insns; +/* A speculative motion requires checking live information on the path + from 'source' to 'target'. The split blocks are those to be checked. + After a speculative motion, live information should be modified in + the 'update' blocks. -/* The mapping from ebb to block. */ -/* ebb_head [i] - is index in rgn_bb_table, while - EBB_HEAD (i) - is basic block index. - BASIC_BLOCK (EBB_HEAD (i)) - head of ebb. */ -#define BB_TO_BLOCK(ebb) (rgn_bb_table[ebb_head[ebb]]) -#define EBB_FIRST_BB(ebb) BASIC_BLOCK (BB_TO_BLOCK (ebb)) -#define EBB_LAST_BB(ebb) BASIC_BLOCK (rgn_bb_table[ebb_head[ebb + 1] - 1]) + Lists of split and update blocks for each candidate of the current + target are in array bblst_table. */ +static basic_block *bblst_table; +static int bblst_size, bblst_last; /* Target info declarations. @@ -183,23 +151,14 @@ typedef struct candidate; static candidate *candidate_table; - -/* A speculative motion requires checking live information on the path - from 'source' to 'target'. The split blocks are those to be checked. - After a speculative motion, live information should be modified in - the 'update' blocks. - - Lists of split and update blocks for each candidate of the current - target are in array bblst_table. */ -static basic_block *bblst_table; -static int bblst_size, bblst_last; - -#define IS_VALID(src) ( candidate_table[src].is_valid ) -#define IS_SPECULATIVE(src) ( candidate_table[src].is_speculative ) +#define IS_VALID(src) (candidate_table[src].is_valid) +#define IS_SPECULATIVE(src) (candidate_table[src].is_speculative) +#define IS_SPECULATIVE_INSN(INSN) \ + (IS_SPECULATIVE (BLOCK_TO_BB (BLOCK_NUM (INSN)))) #define SRC_PROB(src) ( candidate_table[src].src_prob ) /* The bb being currently scheduled. */ -static int target_bb; +int target_bb; /* List of edges. */ typedef struct @@ -214,7 +173,6 @@ static int edgelst_last; static void extract_edgelst (sbitmap, edgelst *); - /* Target info functions. */ static void split_edges (int, int, edgelst *); static void compute_trg_info (int); @@ -260,24 +218,11 @@ static edgeset *pot_split; /* For every bb, a set of its ancestor edges. */ static edgeset *ancestor_edges; -/* Array of EBBs sizes. Currently we can get a ebb only through - splitting of currently scheduling block, therefore, we don't need - ebb_head array for every region, its sufficient to hold it only - for current one. */ -static int *ebb_head; - -static void compute_dom_prob_ps (int); - #define INSN_PROBABILITY(INSN) (SRC_PROB (BLOCK_TO_BB (BLOCK_NUM (INSN)))) -#define IS_SPECULATIVE_INSN(INSN) (IS_SPECULATIVE (BLOCK_TO_BB (BLOCK_NUM (INSN)))) -#define INSN_BB(INSN) (BLOCK_TO_BB (BLOCK_NUM (INSN))) /* Speculative scheduling functions. */ static int check_live_1 (int, rtx); static void update_live_1 (int, rtx); -static int check_live (rtx, int); -static void update_live (rtx, int); -static void set_spec_fed (rtx); static int is_pfree (rtx, int, int); static int find_conditional_protection (rtx, int); static int is_conditionally_protected (rtx, int, int); @@ -285,12 +230,10 @@ static int is_prisky (rtx, int, int); static int is_exception_free (rtx, int, int); static bool sets_likely_spilled (rtx); -static void sets_likely_spilled_1 (rtx, rtx, void *); +static void sets_likely_spilled_1 (rtx, const_rtx, void *); static void add_branch_dependences (rtx, rtx); -static void compute_block_backward_dependences (int); -void debug_dependencies (void); +static void compute_block_dependences (int); -static void init_regions (void); static void schedule_region (int); static rtx concat_INSN_LIST (rtx, rtx); static void concat_insn_mem_list (rtx, rtx, rtx *, rtx *); @@ -321,29 +264,49 @@ is_cfg_nonregular (void) return 1; /* If we have exception handlers, then we consider the cfg not well - structured. ?!? We should be able to handle this now that flow.c - computes an accurate cfg for EH. */ + structured. ?!? We should be able to handle this now that we + compute an accurate cfg for EH. */ if (current_function_has_exception_handlers ()) return 1; - /* If we have non-jumping insns which refer to labels, then we consider - the cfg not well structured. */ + /* If we have insns which refer to labels as non-jumped-to operands, + then we consider the cfg not well structured. */ FOR_EACH_BB (b) FOR_BB_INSNS (b, insn) { - /* Check for labels referred by non-jump insns. */ - if (NONJUMP_INSN_P (insn) || CALL_P (insn)) - { - rtx note = find_reg_note (insn, REG_LABEL, NULL_RTX); - if (note - && ! (JUMP_P (NEXT_INSN (insn)) - && find_reg_note (NEXT_INSN (insn), REG_LABEL, - XEXP (note, 0)))) - return 1; - } + rtx note, next, set, dest; + /* If this function has a computed jump, then we consider the cfg not well structured. */ - else if (JUMP_P (insn) && computed_jump_p (insn)) + if (JUMP_P (insn) && computed_jump_p (insn)) + return 1; + + if (!INSN_P (insn)) + continue; + + note = find_reg_note (insn, REG_LABEL_OPERAND, NULL_RTX); + if (note == NULL_RTX) + continue; + + /* For that label not to be seen as a referred-to label, this + must be a single-set which is feeding a jump *only*. This + could be a conditional jump with the label split off for + machine-specific reasons or a casesi/tablejump. */ + next = next_nonnote_insn (insn); + if (next == NULL_RTX + || !JUMP_P (next) + || (JUMP_LABEL (next) != XEXP (note, 0) + && find_reg_note (next, REG_LABEL_TARGET, + XEXP (note, 0)) == NULL_RTX) + || BLOCK_FOR_INSN (insn) != BLOCK_FOR_INSN (next)) + return 1; + + set = single_set (insn); + if (set == NULL_RTX) + return 1; + + dest = SET_DEST (set); + if (!REG_P (dest) || !dead_or_set_p (next, dest)) return 1; } @@ -414,28 +377,160 @@ debug_regions (void) } } +/* Print the region's basic blocks. */ + +void +debug_region (int rgn) +{ + int bb; + + fprintf (stderr, "\n;; ------------ REGION %d ----------\n\n", rgn); + fprintf (stderr, ";;\trgn %d nr_blocks %d:\n", rgn, + rgn_table[rgn].rgn_nr_blocks); + fprintf (stderr, ";;\tbb/block: "); + + /* We don't have ebb_head initialized yet, so we can't use + BB_TO_BLOCK (). */ + current_blocks = RGN_BLOCKS (rgn); + + for (bb = 0; bb < rgn_table[rgn].rgn_nr_blocks; bb++) + fprintf (stderr, " %d/%d ", bb, rgn_bb_table[current_blocks + bb]); + + fprintf (stderr, "\n\n"); + + for (bb = 0; bb < rgn_table[rgn].rgn_nr_blocks; bb++) + { + debug_bb_n_slim (rgn_bb_table[current_blocks + bb]); + fprintf (stderr, "\n"); + } + + fprintf (stderr, "\n"); + +} + +/* True when a bb with index BB_INDEX contained in region RGN. */ +static bool +bb_in_region_p (int bb_index, int rgn) +{ + int i; + + for (i = 0; i < rgn_table[rgn].rgn_nr_blocks; i++) + if (rgn_bb_table[current_blocks + i] == bb_index) + return true; + + return false; +} + +/* Dump region RGN to file F using dot syntax. */ +void +dump_region_dot (FILE *f, int rgn) +{ + int i; + + fprintf (f, "digraph Region_%d {\n", rgn); + + /* We don't have ebb_head initialized yet, so we can't use + BB_TO_BLOCK (). */ + current_blocks = RGN_BLOCKS (rgn); + + for (i = 0; i < rgn_table[rgn].rgn_nr_blocks; i++) + { + edge e; + edge_iterator ei; + int src_bb_num = rgn_bb_table[current_blocks + i]; + struct basic_block_def *bb = BASIC_BLOCK (src_bb_num); + + FOR_EACH_EDGE (e, ei, bb->succs) + if (bb_in_region_p (e->dest->index, rgn)) + fprintf (f, "\t%d -> %d\n", src_bb_num, e->dest->index); + } + fprintf (f, "}\n"); +} + +/* The same, but first open a file specified by FNAME. */ +void +dump_region_dot_file (const char *fname, int rgn) +{ + FILE *f = fopen (fname, "wt"); + dump_region_dot (f, rgn); + fclose (f); +} + /* Build a single block region for each basic block in the function. This allows for using the same code for interblock and basic block scheduling. */ static void -find_single_block_region (void) +find_single_block_region (bool ebbs_p) { - basic_block bb; + basic_block bb, ebb_start; + int i = 0; nr_regions = 0; - FOR_EACH_BB (bb) - { - rgn_bb_table[nr_regions] = bb->index; - RGN_NR_BLOCKS (nr_regions) = 1; - RGN_BLOCKS (nr_regions) = nr_regions; - RGN_DONT_CALC_DEPS (nr_regions) = 0; - RGN_HAS_REAL_EBB (nr_regions) = 0; - CONTAINING_RGN (bb->index) = nr_regions; - BLOCK_TO_BB (bb->index) = 0; - nr_regions++; - } + if (ebbs_p) { + int probability_cutoff; + if (profile_info && flag_branch_probabilities) + probability_cutoff = PARAM_VALUE (TRACER_MIN_BRANCH_PROBABILITY_FEEDBACK); + else + probability_cutoff = PARAM_VALUE (TRACER_MIN_BRANCH_PROBABILITY); + probability_cutoff = REG_BR_PROB_BASE / 100 * probability_cutoff; + + FOR_EACH_BB (ebb_start) + { + RGN_NR_BLOCKS (nr_regions) = 0; + RGN_BLOCKS (nr_regions) = i; + RGN_DONT_CALC_DEPS (nr_regions) = 0; + RGN_HAS_REAL_EBB (nr_regions) = 0; + + for (bb = ebb_start; ; bb = bb->next_bb) + { + edge e; + edge_iterator ei; + + rgn_bb_table[i] = bb->index; + RGN_NR_BLOCKS (nr_regions)++; + CONTAINING_RGN (bb->index) = nr_regions; + BLOCK_TO_BB (bb->index) = i - RGN_BLOCKS (nr_regions); + i++; + + if (bb->next_bb == EXIT_BLOCK_PTR + || LABEL_P (BB_HEAD (bb->next_bb))) + break; + + FOR_EACH_EDGE (e, ei, bb->succs) + if ((e->flags & EDGE_FALLTHRU) != 0) + break; + if (! e) + break; + if (e->probability <= probability_cutoff) + break; + } + + ebb_start = bb; + nr_regions++; + } + } + else + FOR_EACH_BB (bb) + { + rgn_bb_table[nr_regions] = bb->index; + RGN_NR_BLOCKS (nr_regions) = 1; + RGN_BLOCKS (nr_regions) = nr_regions; + RGN_DONT_CALC_DEPS (nr_regions) = 0; + RGN_HAS_REAL_EBB (nr_regions) = 0; + + CONTAINING_RGN (bb->index) = nr_regions; + BLOCK_TO_BB (bb->index) = 0; + nr_regions++; + } +} + +/* Estimate number of the insns in the BB. */ +static int +rgn_estimate_number_of_insns (basic_block bb) +{ + return INSN_LUID (BB_END (bb)) - INSN_LUID (BB_HEAD (bb)); } /* Update number of blocks and the estimate for number of insns @@ -446,8 +541,8 @@ static bool too_large (int block, int *num_bbs, int *num_insns) { (*num_bbs)++; - (*num_insns) += (INSN_LUID (BB_END (BASIC_BLOCK (block))) - - INSN_LUID (BB_HEAD (BASIC_BLOCK (block)))); + (*num_insns) += (common_sched_info->estimate_number_of_insns + (BASIC_BLOCK (block))); return ((*num_bbs > PARAM_VALUE (PARAM_MAX_SCHED_REGION_BLOCKS)) || (*num_insns > PARAM_VALUE (PARAM_MAX_SCHED_REGION_INSNS))); @@ -500,7 +595,7 @@ too_large (int block, int *num_bbs, int *num_insns) of edge tables. That would simplify it somewhat. */ static void -find_rgns (void) +haifa_find_rgns (void) { int *max_hdr, *dfs_nr, *degree; char no_loops = 1; @@ -691,7 +786,7 @@ find_rgns (void) extend_regions_p = PARAM_VALUE (PARAM_MAX_SCHED_EXTEND_REGIONS_ITERS) > 0; if (extend_regions_p) { - degree1 = xmalloc (last_basic_block * sizeof (int)); + degree1 = XNEWVEC (int, last_basic_block); extended_rgn_header = sbitmap_alloc (last_basic_block); sbitmap_zero (extended_rgn_header); } @@ -756,8 +851,7 @@ find_rgns (void) /* Estimate # insns, and count # blocks in the region. */ num_bbs = 1; - num_insns = (INSN_LUID (BB_END (bb)) - - INSN_LUID (BB_HEAD (bb))); + num_insns = common_sched_info->estimate_number_of_insns (bb); /* Find all loop latches (blocks with back edges to the loop header) or all the leaf blocks in the cfg has no loops. @@ -961,6 +1055,19 @@ find_rgns (void) sbitmap_free (in_stack); } + +/* Wrapper function. + If FLAG_SEL_SCHED_PIPELINING is set, then use custom function to form + regions. Otherwise just call find_rgns_haifa. */ +static void +find_rgns (void) +{ + if (sel_sched_p () && flag_sel_sched_pipelining) + sel_find_rgns (); + else + haifa_find_rgns (); +} + static int gather_region_statistics (int **); static void print_region_statistics (int *, int, int *, int); @@ -981,7 +1088,7 @@ gather_region_statistics (int **rsp) if (nr_blocks > a_sz) { - a = xrealloc (a, nr_blocks * sizeof (*a)); + a = XRESIZEVEC (int, a, nr_blocks); do a[a_sz++] = 0; while (a_sz != nr_blocks); @@ -1030,7 +1137,7 @@ print_region_statistics (int *s1, int s1_sz, int *s2, int s2_sz) LOOP_HDR - mapping from block to the containing loop (two blocks can reside within one region if they have the same loop header). */ -static void +void extend_rgns (int *degree, int *idxp, sbitmap header, int *loop_hdr) { int *order, i, rescan = 0, idx = *idxp, iter = 0, max_iter, *max_hdr; @@ -1038,10 +1145,10 @@ extend_rgns (int *degree, int *idxp, sbitmap header, int *loop_hdr) max_iter = PARAM_VALUE (PARAM_MAX_SCHED_EXTEND_REGIONS_ITERS); - max_hdr = xmalloc (last_basic_block * sizeof (*max_hdr)); + max_hdr = XNEWVEC (int, last_basic_block); - order = xmalloc (last_basic_block * sizeof (*order)); - post_order_compute (order, false); + order = XNEWVEC (int, last_basic_block); + post_order_compute (order, false, false); for (i = nblocks - 1; i >= 0; i--) { @@ -1064,7 +1171,8 @@ extend_rgns (int *degree, int *idxp, sbitmap header, int *loop_hdr) CFG should be traversed until no further changes are made. On each iteration the set of the region heads is extended (the set of those blocks that have max_hdr[bbi] == bbi). This set is upper bounded by the - set of all basic blocks, thus the algorithm is guaranteed to terminate. */ + set of all basic blocks, thus the algorithm is guaranteed to + terminate. */ while (rescan && iter < max_iter) { @@ -1356,13 +1464,26 @@ static void compute_trg_info (int trg) { candidate *sp; - edgelst el; + edgelst el = { NULL, 0 }; int i, j, k, update_idx; basic_block block; sbitmap visited; edge_iterator ei; edge e; + candidate_table = XNEWVEC (candidate, current_nr_blocks); + + bblst_last = 0; + /* bblst_table holds split blocks and update blocks for each block after + the current one in the region. split blocks and update blocks are + the TO blocks of region edges, so there can be at most rgn_nr_edges + of them. */ + bblst_size = (current_nr_blocks - target_bb) * rgn_nr_edges; + bblst_table = XNEWVEC (basic_block, bblst_size); + + edgelst_last = 0; + edgelst_table = XNEWVEC (edge, rgn_nr_edges); + /* Define some of the fields for the target bb as well. */ sp = candidate_table + trg; sp->is_valid = 1; @@ -1449,6 +1570,15 @@ compute_trg_info (int trg) sbitmap_free (visited); } +/* Free the computed target info. */ +static void +free_trg_info (void) +{ + free (candidate_table); + free (bblst_table); + free (edgelst_table); +} + /* Print candidates info, for debugging purposes. Callable from debugger. */ void @@ -1501,6 +1631,8 @@ debug_candidates (int trg) /* Functions for speculative scheduling. */ +static bitmap_head not_in_df; + /* Return 0 if x is a set of a register alive in the beginning of one of the split-blocks of src, otherwise return 1. */ @@ -1552,18 +1684,15 @@ check_live_1 (int src, rtx x) for (i = 0; i < candidate_table[src].split_bbs.nr_members; i++) { basic_block b = candidate_table[src].split_bbs.first_member[i]; + int t = bitmap_bit_p (¬_in_df, b->index); /* We can have split blocks, that were recently generated. - such blocks are always outside current region. */ - gcc_assert (glat_start[b->index] - || CONTAINING_RGN (b->index) - != CONTAINING_RGN (BB_TO_BLOCK (src))); - if (!glat_start[b->index] - || REGNO_REG_SET_P (glat_start[b->index], - regno + j)) - { - return 0; - } + Such blocks are always outside current region. */ + gcc_assert (!t || (CONTAINING_RGN (b->index) + != CONTAINING_RGN (BB_TO_BLOCK (src)))); + + if (t || REGNO_REG_SET_P (df_get_live_in (b), regno + j)) + return 0; } } } @@ -1573,15 +1702,13 @@ check_live_1 (int src, rtx x) for (i = 0; i < candidate_table[src].split_bbs.nr_members; i++) { basic_block b = candidate_table[src].split_bbs.first_member[i]; + int t = bitmap_bit_p (¬_in_df, b->index); - gcc_assert (glat_start[b->index] - || CONTAINING_RGN (b->index) - != CONTAINING_RGN (BB_TO_BLOCK (src))); - if (!glat_start[b->index] - || REGNO_REG_SET_P (glat_start[b->index], regno)) - { - return 0; - } + gcc_assert (!t || (CONTAINING_RGN (b->index) + != CONTAINING_RGN (BB_TO_BLOCK (src)))); + + if (t || REGNO_REG_SET_P (df_get_live_in (b), regno)) + return 0; } } } @@ -1637,7 +1764,7 @@ update_live_1 (int src, rtx x) { basic_block b = candidate_table[src].update_bbs.first_member[i]; - SET_REGNO_REG_SET (glat_start[b->index], regno + j); + SET_REGNO_REG_SET (df_get_live_in (b), regno + j); } } } @@ -1647,7 +1774,7 @@ update_live_1 (int src, rtx x) { basic_block b = candidate_table[src].update_bbs.first_member[i]; - SET_REGNO_REG_SET (glat_start[b->index], regno); + SET_REGNO_REG_SET (df_get_live_in (b), regno); } } } @@ -1711,12 +1838,13 @@ update_live (rtx insn, int src) static void set_spec_fed (rtx load_insn) { - rtx link; + sd_iterator_def sd_it; + dep_t dep; - for (link = INSN_DEPEND (load_insn); link; link = XEXP (link, 1)) - if (GET_MODE (link) == VOIDmode) - FED_BY_SPEC_LOAD (XEXP (link, 0)) = 1; -} /* set_spec_fed */ + FOR_EACH_DEP (load_insn, SD_LIST_FORW, sd_it, dep) + if (DEP_TYPE (dep) == REG_DEP_TRUE) + FED_BY_SPEC_LOAD (DEP_CON (dep)) = 1; +} /* On the path from the insn to load_insn_bb, find a conditional branch depending on insn, that guards the speculative load. */ @@ -1724,17 +1852,19 @@ branch depending on insn, that guards the speculative load. */ static int find_conditional_protection (rtx insn, int load_insn_bb) { - rtx link; + sd_iterator_def sd_it; + dep_t dep; /* Iterate through DEF-USE forward dependences. */ - for (link = INSN_DEPEND (insn); link; link = XEXP (link, 1)) + FOR_EACH_DEP (insn, SD_LIST_FORW, sd_it, dep) { - rtx next = XEXP (link, 0); + rtx next = DEP_CON (dep); + if ((CONTAINING_RGN (BLOCK_NUM (next)) == CONTAINING_RGN (BB_TO_BLOCK (load_insn_bb))) && IS_REACHABLE (INSN_BB (next), load_insn_bb) && load_insn_bb != INSN_BB (next) - && GET_MODE (link) == VOIDmode + && DEP_TYPE (dep) == REG_DEP_TRUE && (JUMP_P (next) || find_conditional_protection (next, load_insn_bb))) return 1; @@ -1744,29 +1874,30 @@ find_conditional_protection (rtx insn, int load_insn_bb) /* Returns 1 if the same insn1 that participates in the computation of load_insn's address is feeding a conditional branch that is - guarding on load_insn. This is true if we find a the two DEF-USE + guarding on load_insn. This is true if we find two DEF-USE chains: insn1 -> ... -> conditional-branch insn1 -> ... -> load_insn, - and if a flow path exist: + and if a flow path exists: insn1 -> ... -> conditional-branch -> ... -> load_insn, and if insn1 is on the path region-entry -> ... -> bb_trg -> ... load_insn. - Locate insn1 by climbing on LOG_LINKS from load_insn. - Locate the branch by following INSN_DEPEND from insn1. */ + Locate insn1 by climbing on INSN_BACK_DEPS from load_insn. + Locate the branch by following INSN_FORW_DEPS from insn1. */ static int is_conditionally_protected (rtx load_insn, int bb_src, int bb_trg) { - rtx link; + sd_iterator_def sd_it; + dep_t dep; - for (link = LOG_LINKS (load_insn); link; link = XEXP (link, 1)) + FOR_EACH_DEP (load_insn, SD_LIST_BACK, sd_it, dep) { - rtx insn1 = XEXP (link, 0); + rtx insn1 = DEP_PRO (dep); /* Must be a DEF-USE dependence upon non-branch. */ - if (GET_MODE (link) != VOIDmode + if (DEP_TYPE (dep) != REG_DEP_TRUE || JUMP_P (insn1)) continue; @@ -1809,28 +1940,29 @@ is_conditionally_protected (rtx load_insn, int bb_src, int bb_trg) static int is_pfree (rtx load_insn, int bb_src, int bb_trg) { - rtx back_link; + sd_iterator_def back_sd_it; + dep_t back_dep; candidate *candp = candidate_table + bb_src; if (candp->split_bbs.nr_members != 1) /* Must have exactly one escape block. */ return 0; - for (back_link = LOG_LINKS (load_insn); - back_link; back_link = XEXP (back_link, 1)) + FOR_EACH_DEP (load_insn, SD_LIST_BACK, back_sd_it, back_dep) { - rtx insn1 = XEXP (back_link, 0); + rtx insn1 = DEP_PRO (back_dep); - if (GET_MODE (back_link) == VOIDmode) + if (DEP_TYPE (back_dep) == REG_DEP_TRUE) + /* Found a DEF-USE dependence (insn1, load_insn). */ { - /* Found a DEF-USE dependence (insn1, load_insn). */ - rtx fore_link; + sd_iterator_def fore_sd_it; + dep_t fore_dep; - for (fore_link = INSN_DEPEND (insn1); - fore_link; fore_link = XEXP (fore_link, 1)) + FOR_EACH_DEP (insn1, SD_LIST_FORW, fore_sd_it, fore_dep) { - rtx insn2 = XEXP (fore_link, 0); - if (GET_MODE (fore_link) == VOIDmode) + rtx insn2 = DEP_CON (fore_dep); + + if (DEP_TYPE (fore_dep) == REG_DEP_TRUE) { /* Found a DEF-USE dependence (insn1, insn2). */ if (haifa_classify_insn (insn2) != PFREE_CANDIDATE) @@ -1863,7 +1995,7 @@ is_prisky (rtx load_insn, int bb_src, int bb_trg) if (FED_BY_SPEC_LOAD (load_insn)) return 1; - if (LOG_LINKS (load_insn) == NULL) + if (sd_lists_empty_p (load_insn, SD_LIST_BACK)) /* Dependence may 'hide' out of the region. */ return 1; @@ -1930,21 +2062,15 @@ static int can_schedule_ready_p (rtx); static void begin_schedule_ready (rtx, rtx); static ds_t new_ready (rtx, ds_t); static int schedule_more_p (void); -static const char *rgn_print_insn (rtx, int); +static const char *rgn_print_insn (const_rtx, int); static int rgn_rank (rtx, rtx); -static int contributes_to_priority (rtx, rtx); static void compute_jump_reg_dependencies (rtx, regset, regset, regset); /* Functions for speculative scheduling. */ -static void add_remove_insn (rtx, int); -static void extend_regions (void); -static void add_block1 (basic_block, basic_block); -static void fix_recovery_cfg (int, int, int); +static void rgn_add_remove_insn (rtx, int); +static void rgn_add_block (basic_block, basic_block); +static void rgn_fix_recovery_cfg (int, int, int); static basic_block advance_target_bb (basic_block, rtx); -static void check_dead_notes1 (int, sbitmap); -#ifdef ENABLE_CHECKING -static int region_head_or_leaf_p (basic_block, int); -#endif /* Return nonzero if there are more insns that should be scheduled. */ @@ -1971,26 +2097,11 @@ init_ready_list (void) /* Print debugging information. */ if (sched_verbose >= 5) - debug_dependencies (); + debug_rgn_dependencies (target_bb); /* Prepare current target block info. */ if (current_nr_blocks > 1) - { - candidate_table = XNEWVEC (candidate, current_nr_blocks); - - bblst_last = 0; - /* bblst_table holds split blocks and update blocks for each block after - the current one in the region. split blocks and update blocks are - the TO blocks of region edges, so there can be at most rgn_nr_edges - of them. */ - bblst_size = (current_nr_blocks - target_bb) * rgn_nr_edges; - bblst_table = XNEWVEC (basic_block, bblst_size); - - edgelst_last = 0; - edgelst_table = XNEWVEC (edge, rgn_nr_edges); - - compute_trg_info (target_bb); - } + compute_trg_info (target_bb); /* Initialize ready list with all 'ready' insns in target block. Count number of insns in the target block being scheduled. */ @@ -2097,10 +2208,24 @@ new_ready (rtx next, ds_t ts) if (not_ex_free /* We are here because is_exception_free () == false. But we possibly can handle that with control speculation. */ - && current_sched_info->flags & DO_SPECULATION) - /* Here we got new control-speculative instruction. */ - ts = set_dep_weak (ts, BEGIN_CONTROL, MAX_DEP_WEAK); + && sched_deps_info->generate_spec_deps + && spec_info->mask & BEGIN_CONTROL) + { + ds_t new_ds; + + /* Add control speculation to NEXT's dependency type. */ + new_ds = set_dep_weak (ts, BEGIN_CONTROL, MAX_DEP_WEAK); + + /* Check if NEXT can be speculated with new dependency type. */ + if (sched_insn_is_legitimate_for_speculation_p (next, new_ds)) + /* Here we got new control-speculative instruction. */ + ts = new_ds; + else + /* NEXT isn't ready yet. */ + ts = (ts & ~SPECULATIVE) | HARD_DEP; + } else + /* NEXT isn't ready yet. */ ts = (ts & ~SPECULATIVE) | HARD_DEP; } } @@ -2114,7 +2239,7 @@ new_ready (rtx next, ds_t ts) to be formatted so that multiple output lines will line up nicely. */ static const char * -rgn_print_insn (rtx insn, int aligned) +rgn_print_insn (const_rtx insn, int aligned) { static char tmp[80]; @@ -2165,7 +2290,7 @@ rgn_rank (rtx insn1, rtx insn2) return nonzero if we should include this dependence in priority calculations. */ -static int +int contributes_to_priority (rtx next, rtx insn) { /* NEXT and INSN reside in one ebb. */ @@ -2187,10 +2312,36 @@ compute_jump_reg_dependencies (rtx insn ATTRIBUTE_UNUSED, add_branch_dependences. */ } +/* This variable holds common_sched_info hooks and data relevant to + the interblock scheduler. */ +static struct common_sched_info_def rgn_common_sched_info; + + +/* This holds data for the dependence analysis relevant to + the interblock scheduler. */ +static struct sched_deps_info_def rgn_sched_deps_info; + +/* This holds constant data used for initializing the above structure + for the Haifa scheduler. */ +static const struct sched_deps_info_def rgn_const_sched_deps_info = + { + compute_jump_reg_dependencies, + NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, + 0, 0, 0 + }; + +/* Same as above, but for the selective scheduler. */ +static const struct sched_deps_info_def rgn_const_sel_sched_deps_info = + { + compute_jump_reg_dependencies, + NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, + 0, 0, 0 + }; + /* Used in schedule_insns to initialize current_sched_info for scheduling regions (or single basic blocks). */ -static struct sched_info region_sched_info = +static const struct haifa_sched_info rgn_const_sched_info = { init_ready_list, can_schedule_ready_p, @@ -2199,26 +2350,29 @@ static struct sched_info region_sched_info = rgn_rank, rgn_print_insn, contributes_to_priority, - compute_jump_reg_dependencies, NULL, NULL, NULL, NULL, - 0, 0, 0, + 0, 0, - add_remove_insn, + rgn_add_remove_insn, begin_schedule_ready, - add_block1, advance_target_bb, - fix_recovery_cfg, -#ifdef ENABLE_CHECKING - region_head_or_leaf_p, -#endif - SCHED_RGN | USE_GLAT -#ifdef ENABLE_CHECKING - | DETACH_LIFE_INFO -#endif + SCHED_RGN }; +/* This variable holds the data and hooks needed to the Haifa scheduler backend + for the interblock scheduler frontend. */ +static struct haifa_sched_info rgn_sched_info; + +/* Returns maximum priority that an insn was assigned to. */ + +int +get_rgn_sched_max_insns_priority (void) +{ + return rgn_sched_info.sched_max_insns_priority; +} + /* Determine if PAT sets a CLASS_LIKELY_SPILLED_P register. */ static bool @@ -2230,7 +2384,7 @@ sets_likely_spilled (rtx pat) } static void -sets_likely_spilled_1 (rtx x, rtx pat, void *data) +sets_likely_spilled_1 (rtx x, const_rtx pat, void *data) { bool *ret = (bool *) data; @@ -2241,9 +2395,12 @@ sets_likely_spilled_1 (rtx x, rtx pat, void *data) *ret = true; } +/* A bitmap to note insns that participate in any dependency. Used in + add_branch_dependences. */ +static sbitmap insn_referenced; + /* Add dependences so that branches are scheduled to run last in their block. */ - static void add_branch_dependences (rtx head, rtx tail) { @@ -2284,11 +2441,12 @@ add_branch_dependences (rtx head, rtx tail) { if (!NOTE_P (insn)) { - if (last != 0 && !find_insn_list (insn, LOG_LINKS (last))) + if (last != 0 + && sd_find_dep_between (insn, last, false) == NULL) { if (! sched_insns_conditions_mutex_p (last, insn)) add_dependence (last, insn, REG_DEP_ANTI); - INSN_REF_COUNT (insn)++; + SET_BIT (insn_referenced, INSN_LUID (insn)); } CANT_MOVE (insn) = 1; @@ -2310,12 +2468,11 @@ add_branch_dependences (rtx head, rtx tail) { insn = prev_nonnote_insn (insn); - if (INSN_REF_COUNT (insn) != 0) + if (TEST_BIT (insn_referenced, INSN_LUID (insn))) continue; if (! sched_insns_conditions_mutex_p (last, insn)) add_dependence (last, insn, REG_DEP_ANTI); - INSN_REF_COUNT (insn) = 1; } #ifdef HAVE_conditional_execution @@ -2383,10 +2540,10 @@ static struct deps *bb_deps; static rtx concat_INSN_LIST (rtx copy, rtx old) { - rtx new = old; + rtx new_rtx = old; for (; copy ; copy = XEXP (copy, 1)) - new = alloc_INSN_LIST (XEXP (copy, 0), new); - return new; + new_rtx = alloc_INSN_LIST (XEXP (copy, 0), new_rtx); + return new_rtx; } static void @@ -2408,6 +2565,57 @@ concat_insn_mem_list (rtx copy_insns, rtx copy_mems, rtx *old_insns_p, *old_mems_p = new_mems; } +/* Join PRED_DEPS to the SUCC_DEPS. */ +void +deps_join (struct deps *succ_deps, struct deps *pred_deps) +{ + unsigned reg; + reg_set_iterator rsi; + + /* The reg_last lists are inherited by successor. */ + EXECUTE_IF_SET_IN_REG_SET (&pred_deps->reg_last_in_use, 0, reg, rsi) + { + struct deps_reg *pred_rl = &pred_deps->reg_last[reg]; + struct deps_reg *succ_rl = &succ_deps->reg_last[reg]; + + succ_rl->uses = concat_INSN_LIST (pred_rl->uses, succ_rl->uses); + succ_rl->sets = concat_INSN_LIST (pred_rl->sets, succ_rl->sets); + succ_rl->clobbers = concat_INSN_LIST (pred_rl->clobbers, + succ_rl->clobbers); + succ_rl->uses_length += pred_rl->uses_length; + succ_rl->clobbers_length += pred_rl->clobbers_length; + } + IOR_REG_SET (&succ_deps->reg_last_in_use, &pred_deps->reg_last_in_use); + + /* Mem read/write lists are inherited by successor. */ + concat_insn_mem_list (pred_deps->pending_read_insns, + pred_deps->pending_read_mems, + &succ_deps->pending_read_insns, + &succ_deps->pending_read_mems); + concat_insn_mem_list (pred_deps->pending_write_insns, + pred_deps->pending_write_mems, + &succ_deps->pending_write_insns, + &succ_deps->pending_write_mems); + + succ_deps->last_pending_memory_flush + = concat_INSN_LIST (pred_deps->last_pending_memory_flush, + succ_deps->last_pending_memory_flush); + + succ_deps->pending_read_list_length += pred_deps->pending_read_list_length; + succ_deps->pending_write_list_length += pred_deps->pending_write_list_length; + succ_deps->pending_flush_length += pred_deps->pending_flush_length; + + /* last_function_call is inherited by successor. */ + succ_deps->last_function_call + = concat_INSN_LIST (pred_deps->last_function_call, + succ_deps->last_function_call); + + /* sched_before_next_call is inherited by successor. */ + succ_deps->sched_before_next_call + = concat_INSN_LIST (pred_deps->sched_before_next_call, + succ_deps->sched_before_next_call); +} + /* After computing the dependencies for block BB, propagate the dependencies found in TMP_DEPS to the successors of the block. */ static void @@ -2420,59 +2628,13 @@ propagate_deps (int bb, struct deps *pred_deps) /* bb's structures are inherited by its successors. */ FOR_EACH_EDGE (e, ei, block->succs) { - struct deps *succ_deps; - unsigned reg; - reg_set_iterator rsi; - /* Only bbs "below" bb, in the same region, are interesting. */ if (e->dest == EXIT_BLOCK_PTR || CONTAINING_RGN (block->index) != CONTAINING_RGN (e->dest->index) || BLOCK_TO_BB (e->dest->index) <= bb) continue; - succ_deps = bb_deps + BLOCK_TO_BB (e->dest->index); - - /* The reg_last lists are inherited by successor. */ - EXECUTE_IF_SET_IN_REG_SET (&pred_deps->reg_last_in_use, 0, reg, rsi) - { - struct deps_reg *pred_rl = &pred_deps->reg_last[reg]; - struct deps_reg *succ_rl = &succ_deps->reg_last[reg]; - - succ_rl->uses = concat_INSN_LIST (pred_rl->uses, succ_rl->uses); - succ_rl->sets = concat_INSN_LIST (pred_rl->sets, succ_rl->sets); - succ_rl->clobbers = concat_INSN_LIST (pred_rl->clobbers, - succ_rl->clobbers); - succ_rl->uses_length += pred_rl->uses_length; - succ_rl->clobbers_length += pred_rl->clobbers_length; - } - IOR_REG_SET (&succ_deps->reg_last_in_use, &pred_deps->reg_last_in_use); - - /* Mem read/write lists are inherited by successor. */ - concat_insn_mem_list (pred_deps->pending_read_insns, - pred_deps->pending_read_mems, - &succ_deps->pending_read_insns, - &succ_deps->pending_read_mems); - concat_insn_mem_list (pred_deps->pending_write_insns, - pred_deps->pending_write_mems, - &succ_deps->pending_write_insns, - &succ_deps->pending_write_mems); - - succ_deps->last_pending_memory_flush - = concat_INSN_LIST (pred_deps->last_pending_memory_flush, - succ_deps->last_pending_memory_flush); - - succ_deps->pending_lists_length += pred_deps->pending_lists_length; - succ_deps->pending_flush_length += pred_deps->pending_flush_length; - - /* last_function_call is inherited by successor. */ - succ_deps->last_function_call - = concat_INSN_LIST (pred_deps->last_function_call, - succ_deps->last_function_call); - - /* sched_before_next_call is inherited by successor. */ - succ_deps->sched_before_next_call - = concat_INSN_LIST (pred_deps->sched_before_next_call, - succ_deps->sched_before_next_call); + deps_join (bb_deps + BLOCK_TO_BB (e->dest->index), pred_deps); } /* These lists should point to the right place, for correct @@ -2489,7 +2651,7 @@ propagate_deps (int bb, struct deps *pred_deps) pred_deps->pending_write_mems = 0; } -/* Compute backward dependences inside bb. In a multiple blocks region: +/* Compute dependences inside bb. In a multiple blocks region: (1) a bb is analyzed after its predecessors, and (2) the lists in effect at the end of bb (after analyzing for bb) are inherited by bb's successors. @@ -2507,7 +2669,7 @@ propagate_deps (int bb, struct deps *pred_deps) similar, and the result is interblock dependences in the region. */ static void -compute_block_backward_dependences (int bb) +compute_block_dependences (int bb) { rtx head, tail; struct deps tmp_deps; @@ -2517,14 +2679,33 @@ compute_block_backward_dependences (int bb) /* Do the analysis for this block. */ gcc_assert (EBB_FIRST_BB (bb) == EBB_LAST_BB (bb)); get_ebb_head_tail (EBB_FIRST_BB (bb), EBB_LAST_BB (bb), &head, &tail); + sched_analyze (&tmp_deps, head, tail); - add_branch_dependences (head, tail); + + /* Selective scheduling handles control dependencies by itself. */ + if (!sel_sched_p ()) + add_branch_dependences (head, tail); if (current_nr_blocks > 1) propagate_deps (bb, &tmp_deps); /* Free up the INSN_LISTs. */ free_deps (&tmp_deps); + + if (targetm.sched.dependencies_evaluation_hook) + targetm.sched.dependencies_evaluation_hook (head, tail); +} + +/* Free dependencies of instructions inside BB. */ +static void +free_block_dependencies (int bb) +{ + rtx head; + rtx tail; + + get_ebb_head_tail (EBB_FIRST_BB (bb), EBB_LAST_BB (bb), &head, &tail); + + sched_free_deps (head, tail, true); } /* Remove all INSN_LISTs and EXPR_LISTs from the pending lists and add @@ -2544,86 +2725,96 @@ free_pending_lists (void) } } -/* Print dependences for debugging, callable from debugger. */ - +/* Print dependences for debugging starting from FROM_BB. + Callable from debugger. */ +/* Print dependences for debugging starting from FROM_BB. + Callable from debugger. */ void -debug_dependencies (void) +debug_rgn_dependencies (int from_bb) { int bb; - fprintf (sched_dump, ";; --------------- forward dependences: ------------ \n"); - for (bb = 0; bb < current_nr_blocks; bb++) + fprintf (sched_dump, + ";; --------------- forward dependences: ------------ \n"); + + for (bb = from_bb; bb < current_nr_blocks; bb++) { rtx head, tail; - rtx next_tail; - rtx insn; - gcc_assert (EBB_FIRST_BB (bb) == EBB_LAST_BB (bb)); get_ebb_head_tail (EBB_FIRST_BB (bb), EBB_LAST_BB (bb), &head, &tail); - next_tail = NEXT_INSN (tail); fprintf (sched_dump, "\n;; --- Region Dependences --- b %d bb %d \n", BB_TO_BLOCK (bb), bb); - fprintf (sched_dump, ";; %7s%6s%6s%6s%6s%6s%14s\n", - "insn", "code", "bb", "dep", "prio", "cost", - "reservation"); - fprintf (sched_dump, ";; %7s%6s%6s%6s%6s%6s%14s\n", - "----", "----", "--", "---", "----", "----", - "-----------"); + debug_dependencies (head, tail); + } +} + +/* Print dependencies information for instructions between HEAD and TAIL. + ??? This function would probably fit best in haifa-sched.c. */ +void debug_dependencies (rtx head, rtx tail) +{ + rtx insn; + rtx next_tail = NEXT_INSN (tail); - for (insn = head; insn != next_tail; insn = NEXT_INSN (insn)) - { - rtx link; + fprintf (sched_dump, ";; %7s%6s%6s%6s%6s%6s%14s\n", + "insn", "code", "bb", "dep", "prio", "cost", + "reservation"); + fprintf (sched_dump, ";; %7s%6s%6s%6s%6s%6s%14s\n", + "----", "----", "--", "---", "----", "----", + "-----------"); - if (! INSN_P (insn)) + for (insn = head; insn != next_tail; insn = NEXT_INSN (insn)) + { + if (! INSN_P (insn)) + { + int n; + fprintf (sched_dump, ";; %6d ", INSN_UID (insn)); + if (NOTE_P (insn)) { - int n; - fprintf (sched_dump, ";; %6d ", INSN_UID (insn)); - if (NOTE_P (insn)) - { - n = NOTE_LINE_NUMBER (insn); - if (n < 0) - fprintf (sched_dump, "%s\n", GET_NOTE_INSN_NAME (n)); - else - { - expanded_location xloc; - NOTE_EXPANDED_LOCATION (xloc, insn); - fprintf (sched_dump, "line %d, file %s\n", - xloc.line, xloc.file); - } - } - else - fprintf (sched_dump, " {%s}\n", GET_RTX_NAME (GET_CODE (insn))); - continue; + n = NOTE_KIND (insn); + fprintf (sched_dump, "%s\n", GET_NOTE_INSN_NAME (n)); } - - fprintf (sched_dump, - ";; %s%5d%6d%6d%6d%6d%6d ", - (SCHED_GROUP_P (insn) ? "+" : " "), - INSN_UID (insn), - INSN_CODE (insn), - INSN_BB (insn), - INSN_DEP_COUNT (insn), - INSN_PRIORITY (insn), - insn_cost (insn, 0, 0)); - - if (recog_memoized (insn) < 0) - fprintf (sched_dump, "nothing"); else - print_reservation (sched_dump, insn); - - fprintf (sched_dump, "\t: "); - for (link = INSN_DEPEND (insn); link; link = XEXP (link, 1)) - fprintf (sched_dump, "%d ", INSN_UID (XEXP (link, 0))); - fprintf (sched_dump, "\n"); + fprintf (sched_dump, " {%s}\n", GET_RTX_NAME (GET_CODE (insn))); + continue; } + + fprintf (sched_dump, + ";; %s%5d%6d%6d%6d%6d%6d ", + (SCHED_GROUP_P (insn) ? "+" : " "), + INSN_UID (insn), + INSN_CODE (insn), + BLOCK_NUM (insn), + sched_emulate_haifa_p ? -1 : sd_lists_size (insn, SD_LIST_BACK), + (sel_sched_p () ? (sched_emulate_haifa_p ? -1 + : INSN_PRIORITY (insn)) + : INSN_PRIORITY (insn)), + (sel_sched_p () ? (sched_emulate_haifa_p ? -1 + : insn_cost (insn)) + : insn_cost (insn))); + + if (recog_memoized (insn) < 0) + fprintf (sched_dump, "nothing"); + else + print_reservation (sched_dump, insn); + + fprintf (sched_dump, "\t: "); + { + sd_iterator_def sd_it; + dep_t dep; + + FOR_EACH_DEP (insn, SD_LIST_FORW, sd_it, dep) + fprintf (sched_dump, "%d ", INSN_UID (DEP_CON (dep))); + } + fprintf (sched_dump, "\n"); } + fprintf (sched_dump, "\n"); } /* Returns true if all the basic blocks of the current region have NOTE_DISABLE_SCHED_OF_BLOCK which means not to schedule that region. */ -static bool +bool sched_is_disabled_for_current_region_p (void) { int bb; @@ -2635,6 +2826,47 @@ sched_is_disabled_for_current_region_p (void) return true; } +/* Free all region dependencies saved in INSN_BACK_DEPS and + INSN_RESOLVED_BACK_DEPS. The Haifa scheduler does this on the fly + when scheduling, so this function is supposed to be called from + the selective scheduling only. */ +void +free_rgn_deps (void) +{ + int bb; + + for (bb = 0; bb < current_nr_blocks; bb++) + { + rtx head, tail; + + gcc_assert (EBB_FIRST_BB (bb) == EBB_LAST_BB (bb)); + get_ebb_head_tail (EBB_FIRST_BB (bb), EBB_LAST_BB (bb), &head, &tail); + + sched_free_deps (head, tail, false); + } +} + +static int rgn_n_insns; + +/* Compute insn priority for a current region. */ +void +compute_priorities (void) +{ + int bb; + + current_sched_info->sched_max_insns_priority = 0; + for (bb = 0; bb < current_nr_blocks; bb++) + { + rtx head, tail; + + gcc_assert (EBB_FIRST_BB (bb) == EBB_LAST_BB (bb)); + get_ebb_head_tail (EBB_FIRST_BB (bb), EBB_LAST_BB (bb), &head, &tail); + + rgn_n_insns += set_priorities (head, tail); + } + current_sched_info->sched_max_insns_priority++; +} + /* Schedule a region. A region is either an inner loop, a loop-free subroutine, or a single basic block. Each bb in the region is scheduled after its flow predecessors. */ @@ -2642,80 +2874,254 @@ sched_is_disabled_for_current_region_p (void) static void schedule_region (int rgn) { - basic_block block; - edge_iterator ei; - edge e; int bb; int sched_rgn_n_insns = 0; rgn_n_insns = 0; - /* Set variables for the current region. */ - current_nr_blocks = RGN_NR_BLOCKS (rgn); - current_blocks = RGN_BLOCKS (rgn); - - /* See comments in add_block1, for what reasons we allocate +1 element. */ - ebb_head = xrealloc (ebb_head, (current_nr_blocks + 1) * sizeof (*ebb_head)); - for (bb = 0; bb <= current_nr_blocks; bb++) - ebb_head[bb] = current_blocks + bb; + + rgn_setup_region (rgn); /* Don't schedule region that is marked by NOTE_DISABLE_SCHED_OF_BLOCK. */ if (sched_is_disabled_for_current_region_p ()) return; - if (!RGN_DONT_CALC_DEPS (rgn)) + sched_rgn_compute_dependencies (rgn); + + sched_rgn_local_init (rgn); + + /* Set priorities. */ + compute_priorities (); + + sched_extend_ready_list (rgn_n_insns); + + /* Now we can schedule all blocks. */ + for (bb = 0; bb < current_nr_blocks; bb++) { - init_deps_global (); + basic_block first_bb, last_bb, curr_bb; + rtx head, tail; - /* Initializations for region data dependence analysis. */ - bb_deps = XNEWVEC (struct deps, current_nr_blocks); - for (bb = 0; bb < current_nr_blocks; bb++) - init_deps (bb_deps + bb); + first_bb = EBB_FIRST_BB (bb); + last_bb = EBB_LAST_BB (bb); - /* Compute LOG_LINKS. */ - for (bb = 0; bb < current_nr_blocks; bb++) - compute_block_backward_dependences (bb); + get_ebb_head_tail (first_bb, last_bb, &head, &tail); - /* Compute INSN_DEPEND. */ - for (bb = current_nr_blocks - 1; bb >= 0; bb--) - { - rtx head, tail; + if (no_real_insns_p (head, tail)) + { + gcc_assert (first_bb == last_bb); + continue; + } + + current_sched_info->prev_head = PREV_INSN (head); + current_sched_info->next_tail = NEXT_INSN (tail); - gcc_assert (EBB_FIRST_BB (bb) == EBB_LAST_BB (bb)); - get_ebb_head_tail (EBB_FIRST_BB (bb), EBB_LAST_BB (bb), &head, &tail); + remove_notes (head, tail); - compute_forward_dependences (head, tail); + unlink_bb_notes (first_bb, last_bb); + + target_bb = bb; + + gcc_assert (flag_schedule_interblock || current_nr_blocks == 1); + current_sched_info->queue_must_finish_empty = current_nr_blocks == 1; - if (targetm.sched.dependencies_evaluation_hook) - targetm.sched.dependencies_evaluation_hook (head, tail); + curr_bb = first_bb; + if (dbg_cnt (sched_block)) + { + schedule_block (&curr_bb); + gcc_assert (EBB_FIRST_BB (bb) == first_bb); + sched_rgn_n_insns += sched_n_insns; + } + else + { + sched_rgn_n_insns += rgn_n_insns; } - free_pending_lists (); + /* Clean up. */ + if (current_nr_blocks > 1) + free_trg_info (); + } - finish_deps_global (); + /* Sanity check: verify that all region insns were scheduled. */ + gcc_assert (sched_rgn_n_insns == rgn_n_insns); - free (bb_deps); + sched_finish_ready_list (); + + /* Done with this region. */ + sched_rgn_local_finish (); + + /* Free dependencies. */ + for (bb = 0; bb < current_nr_blocks; ++bb) + free_block_dependencies (bb); + + gcc_assert (haifa_recovery_bb_ever_added_p + || deps_pools_are_empty_p ()); +} + +/* Initialize data structures for region scheduling. */ + +void +sched_rgn_init (bool single_blocks_p) +{ + min_spec_prob = ((PARAM_VALUE (PARAM_MIN_SPEC_PROB) * REG_BR_PROB_BASE) + / 100); + + nr_inter = 0; + nr_spec = 0; + + extend_regions (); + + CONTAINING_RGN (ENTRY_BLOCK) = -1; + CONTAINING_RGN (EXIT_BLOCK) = -1; + + /* Compute regions for scheduling. */ + if (single_blocks_p + || n_basic_blocks == NUM_FIXED_BLOCKS + 1 + || !flag_schedule_interblock + || is_cfg_nonregular ()) + { + find_single_block_region (sel_sched_p ()); } else - /* This is a recovery block. It is always a single block region. */ - gcc_assert (current_nr_blocks == 1); - - /* Set priorities. */ - current_sched_info->sched_max_insns_priority = 0; - for (bb = 0; bb < current_nr_blocks; bb++) { - rtx head, tail; + /* Compute the dominators and post dominators. */ + if (!sel_sched_p ()) + calculate_dominance_info (CDI_DOMINATORS); + + /* Find regions. */ + find_rgns (); + + if (sched_verbose >= 3) + debug_regions (); + + /* For now. This will move as more and more of haifa is converted + to using the cfg code. */ + if (!sel_sched_p ()) + free_dominance_info (CDI_DOMINATORS); + } + + gcc_assert (0 < nr_regions && nr_regions <= n_basic_blocks); + + RGN_BLOCKS (nr_regions) = (RGN_BLOCKS (nr_regions - 1) + + RGN_NR_BLOCKS (nr_regions - 1)); +} + +/* Free data structures for region scheduling. */ +void +sched_rgn_finish (void) +{ + /* Reposition the prologue and epilogue notes in case we moved the + prologue/epilogue insns. */ + if (reload_completed) + reposition_prologue_and_epilogue_notes (); + + if (sched_verbose) + { + if (reload_completed == 0 + && flag_schedule_interblock) + { + fprintf (sched_dump, + "\n;; Procedure interblock/speculative motions == %d/%d \n", + nr_inter, nr_spec); + } + else + gcc_assert (nr_inter <= 0); + fprintf (sched_dump, "\n\n"); + } + + nr_regions = 0; + + free (rgn_table); + rgn_table = NULL; + + free (rgn_bb_table); + rgn_bb_table = NULL; + + free (block_to_bb); + block_to_bb = NULL; + + free (containing_rgn); + containing_rgn = NULL; + + free (ebb_head); + ebb_head = NULL; +} + +/* Setup global variables like CURRENT_BLOCKS and CURRENT_NR_BLOCK to + point to the region RGN. */ +void +rgn_setup_region (int rgn) +{ + int bb; + + /* Set variables for the current region. */ + current_nr_blocks = RGN_NR_BLOCKS (rgn); + current_blocks = RGN_BLOCKS (rgn); + + /* EBB_HEAD is a region-scope structure. But we realloc it for + each region to save time/memory/something else. + See comments in add_block1, for what reasons we allocate +1 element. */ + ebb_head = XRESIZEVEC (int, ebb_head, current_nr_blocks + 1); + for (bb = 0; bb <= current_nr_blocks; bb++) + ebb_head[bb] = current_blocks + bb; +} + +/* Compute instruction dependencies in region RGN. */ +void +sched_rgn_compute_dependencies (int rgn) +{ + if (!RGN_DONT_CALC_DEPS (rgn)) + { + int bb; + + if (sel_sched_p ()) + sched_emulate_haifa_p = 1; + + init_deps_global (); + + /* Initializations for region data dependence analysis. */ + bb_deps = XNEWVEC (struct deps, current_nr_blocks); + for (bb = 0; bb < current_nr_blocks; bb++) + init_deps (bb_deps + bb); + + /* Initialize bitmap used in add_branch_dependences. */ + insn_referenced = sbitmap_alloc (sched_max_luid); + sbitmap_zero (insn_referenced); - gcc_assert (EBB_FIRST_BB (bb) == EBB_LAST_BB (bb)); - get_ebb_head_tail (EBB_FIRST_BB (bb), EBB_LAST_BB (bb), &head, &tail); + /* Compute backward dependencies. */ + for (bb = 0; bb < current_nr_blocks; bb++) + compute_block_dependences (bb); + + sbitmap_free (insn_referenced); + free_pending_lists (); + finish_deps_global (); + free (bb_deps); - rgn_n_insns += set_priorities (head, tail); + /* We don't want to recalculate this twice. */ + RGN_DONT_CALC_DEPS (rgn) = 1; + + if (sel_sched_p ()) + sched_emulate_haifa_p = 0; } - current_sched_info->sched_max_insns_priority++; + else + /* (This is a recovery block. It is always a single block region.) + OR (We use selective scheduling.) */ + gcc_assert (current_nr_blocks == 1 || sel_sched_p ()); +} +/* Init region data structures. Returns true if this region should + not be scheduled. */ +void +sched_rgn_local_init (int rgn) +{ + int bb; + /* Compute interblock info: probabilities, split-edges, dominators, etc. */ if (current_nr_blocks > 1) { + basic_block block; + edge e; + edge_iterator ei; + prob = XNEWVEC (int, current_nr_blocks); dom = sbitmap_vector_alloc (current_nr_blocks, current_nr_blocks); @@ -2762,316 +3168,98 @@ schedule_region (int rgn) e->aux = NULL; } } - - /* Now we can schedule all blocks. */ - for (bb = 0; bb < current_nr_blocks; bb++) - { - basic_block first_bb, last_bb, curr_bb; - rtx head, tail; - int b = BB_TO_BLOCK (bb); - - first_bb = EBB_FIRST_BB (bb); - last_bb = EBB_LAST_BB (bb); - - get_ebb_head_tail (first_bb, last_bb, &head, &tail); - - if (no_real_insns_p (head, tail)) - { - gcc_assert (first_bb == last_bb); - continue; - } - - current_sched_info->prev_head = PREV_INSN (head); - current_sched_info->next_tail = NEXT_INSN (tail); - - if (write_symbols != NO_DEBUG) - { - save_line_notes (b, head, tail); - } - - /* rm_other_notes only removes notes which are _inside_ the - block---that is, it won't remove notes before the first real insn - or after the last real insn of the block. So if the first insn - has a REG_SAVE_NOTE which would otherwise be emitted before the - insn, it is redundant with the note before the start of the - block, and so we have to take it out. */ - if (INSN_P (head)) - { - rtx note; - - for (note = REG_NOTES (head); note; note = XEXP (note, 1)) - if (REG_NOTE_KIND (note) == REG_SAVE_NOTE) - remove_note (head, note); - } - else - /* This means that first block in ebb is empty. - It looks to me as an impossible thing. There at least should be - a recovery check, that caused the splitting. */ - gcc_unreachable (); - - /* Remove remaining note insns from the block, save them in - note_list. These notes are restored at the end of - schedule_block (). */ - rm_other_notes (head, tail); - - unlink_bb_notes (first_bb, last_bb); - - target_bb = bb; - - gcc_assert (flag_schedule_interblock || current_nr_blocks == 1); - current_sched_info->queue_must_finish_empty = current_nr_blocks == 1; - - curr_bb = first_bb; - schedule_block (&curr_bb, rgn_n_insns); - gcc_assert (EBB_FIRST_BB (bb) == first_bb); - sched_rgn_n_insns += sched_n_insns; - - /* Clean up. */ - if (current_nr_blocks > 1) - { - free (candidate_table); - free (bblst_table); - free (edgelst_table); - } - } - - /* Sanity check: verify that all region insns were scheduled. */ - gcc_assert (sched_rgn_n_insns == rgn_n_insns); - - /* Restore line notes. */ - if (write_symbols != NO_DEBUG) - { - for (bb = 0; bb < current_nr_blocks; bb++) - { - rtx head, tail; - - get_ebb_head_tail (EBB_FIRST_BB (bb), EBB_LAST_BB (bb), &head, &tail); - restore_line_notes (head, tail); - } - } - - /* Done with this region. */ - - if (current_nr_blocks > 1) - { - free (prob); - sbitmap_vector_free (dom); - sbitmap_vector_free (pot_split); - sbitmap_vector_free (ancestor_edges); - free (rgn_edges); - } } -/* Indexed by region, holds the number of death notes found in that region. - Used for consistency checks. */ -static int *deaths_in_region; - -/* Initialize data structures for region scheduling. */ - -static void -init_regions (void) +/* Free data computed for the finished region. */ +void +sched_rgn_local_free (void) { - sbitmap blocks; - int rgn; - - nr_regions = 0; - rgn_table = 0; - rgn_bb_table = 0; - block_to_bb = 0; - containing_rgn = 0; - extend_regions (); + free (prob); + sbitmap_vector_free (dom); + sbitmap_vector_free (pot_split); + sbitmap_vector_free (ancestor_edges); + free (rgn_edges); +} - /* Compute regions for scheduling. */ - if (reload_completed - || n_basic_blocks == NUM_FIXED_BLOCKS + 1 - || !flag_schedule_interblock - || is_cfg_nonregular ()) +/* Free data computed for the finished region. */ +void +sched_rgn_local_finish (void) +{ + if (current_nr_blocks > 1 && !sel_sched_p ()) { - find_single_block_region (); + sched_rgn_local_free (); } - else - { - /* Compute the dominators and post dominators. */ - calculate_dominance_info (CDI_DOMINATORS); +} - /* Find regions. */ - find_rgns (); +/* Setup scheduler infos. */ +void +rgn_setup_common_sched_info (void) +{ + memcpy (&rgn_common_sched_info, &haifa_common_sched_info, + sizeof (rgn_common_sched_info)); - if (sched_verbose >= 3) - debug_regions (); + rgn_common_sched_info.fix_recovery_cfg = rgn_fix_recovery_cfg; + rgn_common_sched_info.add_block = rgn_add_block; + rgn_common_sched_info.estimate_number_of_insns + = rgn_estimate_number_of_insns; + rgn_common_sched_info.sched_pass_id = SCHED_RGN_PASS; - /* For now. This will move as more and more of haifa is converted - to using the cfg code in flow.c. */ - free_dominance_info (CDI_DOMINATORS); - } - RGN_BLOCKS (nr_regions) = RGN_BLOCKS (nr_regions - 1) + - RGN_NR_BLOCKS (nr_regions - 1); + common_sched_info = &rgn_common_sched_info; +} +/* Setup all *_sched_info structures (for the Haifa frontend + and for the dependence analysis) in the interblock scheduler. */ +void +rgn_setup_sched_infos (void) +{ + if (!sel_sched_p ()) + memcpy (&rgn_sched_deps_info, &rgn_const_sched_deps_info, + sizeof (rgn_sched_deps_info)); + else + memcpy (&rgn_sched_deps_info, &rgn_const_sel_sched_deps_info, + sizeof (rgn_sched_deps_info)); - if (CHECK_DEAD_NOTES) - { - blocks = sbitmap_alloc (last_basic_block); - deaths_in_region = XNEWVEC (int, nr_regions); - /* Remove all death notes from the subroutine. */ - for (rgn = 0; rgn < nr_regions; rgn++) - check_dead_notes1 (rgn, blocks); + sched_deps_info = &rgn_sched_deps_info; - sbitmap_free (blocks); - } - else - count_or_remove_death_notes (NULL, 1); + memcpy (&rgn_sched_info, &rgn_const_sched_info, sizeof (rgn_sched_info)); + current_sched_info = &rgn_sched_info; } /* The one entry point in this file. */ - void schedule_insns (void) { - sbitmap large_region_blocks, blocks; int rgn; - int any_large_regions; - basic_block bb; /* Taking care of this degenerate case makes the rest of this code simpler. */ if (n_basic_blocks == NUM_FIXED_BLOCKS) return; - nr_inter = 0; - nr_spec = 0; + rgn_setup_common_sched_info (); + rgn_setup_sched_infos (); - /* We need current_sched_info in init_dependency_caches, which is - invoked via sched_init. */ - current_sched_info = ®ion_sched_info; + haifa_sched_init (); + sched_rgn_init (reload_completed); - sched_init (); + bitmap_initialize (¬_in_df, 0); + bitmap_clear (¬_in_df); - min_spec_prob = ((PARAM_VALUE (PARAM_MIN_SPEC_PROB) * REG_BR_PROB_BASE) - / 100); - - init_regions (); - - /* EBB_HEAD is a region-scope structure. But we realloc it for - each region to save time/memory/something else. */ - ebb_head = 0; - /* Schedule every region in the subroutine. */ for (rgn = 0; rgn < nr_regions; rgn++) - schedule_region (rgn); - - free(ebb_head); - - /* Update life analysis for the subroutine. Do single block regions - first so that we can verify that live_at_start didn't change. Then - do all other blocks. */ - /* ??? There is an outside possibility that update_life_info, or more - to the point propagate_block, could get called with nonzero flags - more than once for one basic block. This would be kinda bad if it - were to happen, since REG_INFO would be accumulated twice for the - block, and we'd have twice the REG_DEAD notes. - - I'm fairly certain that this _shouldn't_ happen, since I don't think - that live_at_start should change at region heads. Not sure what the - best way to test for this kind of thing... */ - - if (current_sched_info->flags & DETACH_LIFE_INFO) - /* this flag can be set either by the target or by ENABLE_CHECKING. */ - attach_life_info (); - - allocate_reg_life_data (); - - any_large_regions = 0; - large_region_blocks = sbitmap_alloc (last_basic_block); - sbitmap_zero (large_region_blocks); - FOR_EACH_BB (bb) - SET_BIT (large_region_blocks, bb->index); - - blocks = sbitmap_alloc (last_basic_block); - sbitmap_zero (blocks); - - /* Update life information. For regions consisting of multiple blocks - we've possibly done interblock scheduling that affects global liveness. - For regions consisting of single blocks we need to do only local - liveness. */ - for (rgn = 0; rgn < nr_regions; rgn++) - if (RGN_NR_BLOCKS (rgn) > 1 - /* Or the only block of this region has been split. */ - || RGN_HAS_REAL_EBB (rgn) - /* New blocks (e.g. recovery blocks) should be processed - as parts of large regions. */ - || !glat_start[rgn_bb_table[RGN_BLOCKS (rgn)]]) - any_large_regions = 1; - else - { - SET_BIT (blocks, rgn_bb_table[RGN_BLOCKS (rgn)]); - RESET_BIT (large_region_blocks, rgn_bb_table[RGN_BLOCKS (rgn)]); - } - - /* Don't update reg info after reload, since that affects - regs_ever_live, which should not change after reload. */ - update_life_info (blocks, UPDATE_LIFE_LOCAL, - (reload_completed ? PROP_DEATH_NOTES - : (PROP_DEATH_NOTES | PROP_REG_INFO))); - if (any_large_regions) - { - update_life_info (large_region_blocks, UPDATE_LIFE_GLOBAL, - (reload_completed ? PROP_DEATH_NOTES - : (PROP_DEATH_NOTES | PROP_REG_INFO))); - -#ifdef ENABLE_CHECKING - check_reg_live (true); -#endif - } - - if (CHECK_DEAD_NOTES) - { - /* Verify the counts of basic block notes in single basic block - regions. */ - for (rgn = 0; rgn < nr_regions; rgn++) - if (RGN_NR_BLOCKS (rgn) == 1) - { - sbitmap_zero (blocks); - SET_BIT (blocks, rgn_bb_table[RGN_BLOCKS (rgn)]); - - gcc_assert (deaths_in_region[rgn] - == count_or_remove_death_notes (blocks, 0)); - } - free (deaths_in_region); - } - - /* Reposition the prologue and epilogue notes in case we moved the - prologue/epilogue insns. */ - if (reload_completed) - reposition_prologue_and_epilogue_notes (get_insns ()); - - if (sched_verbose) - { - if (reload_completed == 0 && flag_schedule_interblock) - { - fprintf (sched_dump, - "\n;; Procedure interblock/speculative motions == %d/%d \n", - nr_inter, nr_spec); - } - else - gcc_assert (nr_inter <= 0); - fprintf (sched_dump, "\n\n"); - } + if (dbg_cnt (sched_region)) + schedule_region (rgn); /* Clean up. */ - free (rgn_table); - free (rgn_bb_table); - free (block_to_bb); - free (containing_rgn); + sched_rgn_finish (); + bitmap_clear (¬_in_df); - sched_finish (); - - sbitmap_free (blocks); - sbitmap_free (large_region_blocks); + haifa_sched_finish (); } /* INSN has been added to/removed from current region. */ static void -add_remove_insn (rtx insn, int remove_p) +rgn_add_remove_insn (rtx insn, int remove_p) { if (!remove_p) rgn_n_insns++; @@ -3088,7 +3276,7 @@ add_remove_insn (rtx insn, int remove_p) } /* Extend internal data structures. */ -static void +void extend_regions (void) { rgn_table = XRESIZEVEC (region, rgn_table, n_basic_blocks); @@ -3097,40 +3285,37 @@ extend_regions (void) containing_rgn = XRESIZEVEC (int, containing_rgn, last_basic_block); } -/* BB was added to ebb after AFTER. */ -static void -add_block1 (basic_block bb, basic_block after) +void +rgn_make_new_region_out_of_new_block (basic_block bb) { - extend_regions (); + int i; - if (after == 0 || after == EXIT_BLOCK_PTR) - { - int i; - - i = RGN_BLOCKS (nr_regions); - /* I - first free position in rgn_bb_table. */ + i = RGN_BLOCKS (nr_regions); + /* I - first free position in rgn_bb_table. */ - rgn_bb_table[i] = bb->index; - RGN_NR_BLOCKS (nr_regions) = 1; - RGN_DONT_CALC_DEPS (nr_regions) = after == EXIT_BLOCK_PTR; - RGN_HAS_REAL_EBB (nr_regions) = 0; - CONTAINING_RGN (bb->index) = nr_regions; - BLOCK_TO_BB (bb->index) = 0; + rgn_bb_table[i] = bb->index; + RGN_NR_BLOCKS (nr_regions) = 1; + RGN_HAS_REAL_EBB (nr_regions) = 0; + RGN_DONT_CALC_DEPS (nr_regions) = 0; + CONTAINING_RGN (bb->index) = nr_regions; + BLOCK_TO_BB (bb->index) = 0; - nr_regions++; + nr_regions++; - RGN_BLOCKS (nr_regions) = i + 1; + RGN_BLOCKS (nr_regions) = i + 1; +} - if (CHECK_DEAD_NOTES) - { - sbitmap blocks = sbitmap_alloc (last_basic_block); - deaths_in_region = xrealloc (deaths_in_region, nr_regions * - sizeof (*deaths_in_region)); +/* BB was added to ebb after AFTER. */ +static void +rgn_add_block (basic_block bb, basic_block after) +{ + extend_regions (); + bitmap_set_bit (¬_in_df, bb->index); - check_dead_notes1 (nr_regions - 1, blocks); - - sbitmap_free (blocks); - } + if (after == 0 || after == EXIT_BLOCK_PTR) + { + rgn_make_new_region_out_of_new_block (bb); + RGN_DONT_CALC_DEPS (nr_regions - 1) = (after == EXIT_BLOCK_PTR); } else { @@ -3147,9 +3332,15 @@ add_block1 (basic_block bb, basic_block after) is _always_ valid for access. */ i = BLOCK_TO_BB (after->index) + 1; - for (pos = ebb_head[i]; rgn_bb_table[pos] != after->index; pos--); + pos = ebb_head[i] - 1; + /* Now POS is the index of the last block in the region. */ + + /* Find index of basic block AFTER. */ + for (; rgn_bb_table[pos] != after->index; pos--); + pos++; gcc_assert (pos > ebb_head[i - 1]); + /* i - ebb right after "AFTER". */ /* ebb_head[i] - VALID. */ @@ -3177,9 +3368,6 @@ add_block1 (basic_block bb, basic_block after) for (++i; i <= nr_regions; i++) RGN_BLOCKS (i)++; - - /* We don't need to call check_dead_notes1 () because this new block - is just a split of the old. We don't want to count anything twice. */ } } @@ -3187,7 +3375,7 @@ add_block1 (basic_block bb, basic_block after) For parameter meaning please refer to sched-int.h: struct sched_info: fix_recovery_cfg. */ static void -fix_recovery_cfg (int bbi, int check_bbi, int check_bb_nexti) +rgn_fix_recovery_cfg (int bbi, int check_bbi, int check_bb_nexti) { int old_pos, new_pos, i; @@ -3229,56 +3417,13 @@ advance_target_bb (basic_block bb, rtx insn) return bb->next_bb; } -/* Count and remove death notes in region RGN, which consists of blocks - with indecies in BLOCKS. */ -static void -check_dead_notes1 (int rgn, sbitmap blocks) -{ - int b; - - sbitmap_zero (blocks); - for (b = RGN_NR_BLOCKS (rgn) - 1; b >= 0; --b) - SET_BIT (blocks, rgn_bb_table[RGN_BLOCKS (rgn) + b]); - - deaths_in_region[rgn] = count_or_remove_death_notes (blocks, 1); -} - -#ifdef ENABLE_CHECKING -/* Return non zero, if BB is head or leaf (depending of LEAF_P) block in - current region. For more information please refer to - sched-int.h: struct sched_info: region_head_or_leaf_p. */ -static int -region_head_or_leaf_p (basic_block bb, int leaf_p) -{ - if (!leaf_p) - return bb->index == rgn_bb_table[RGN_BLOCKS (CONTAINING_RGN (bb->index))]; - else - { - int i; - edge e; - edge_iterator ei; - - i = CONTAINING_RGN (bb->index); - - FOR_EACH_EDGE (e, ei, bb->succs) - if (e->dest != EXIT_BLOCK_PTR - && CONTAINING_RGN (e->dest->index) == i - /* except self-loop. */ - && e->dest != bb) - return 0; - - return 1; - } -} -#endif /* ENABLE_CHECKING */ - #endif static bool gate_handle_sched (void) { #ifdef INSN_SCHEDULING - return flag_schedule_insns; + return flag_schedule_insns && dbg_cnt (sched_func); #else return 0; #endif @@ -3289,10 +3434,11 @@ static unsigned int rest_of_handle_sched (void) { #ifdef INSN_SCHEDULING - /* Do control and data sched analysis, - and write some of the results to dump file. */ - - schedule_insns (); + if (flag_selective_scheduling + && ! maybe_skip_selective_scheduling ()) + run_selective_scheduling (); + else + schedule_insns (); #endif return 0; } @@ -3301,7 +3447,8 @@ static bool gate_handle_sched2 (void) { #ifdef INSN_SCHEDULING - return optimize > 0 && flag_schedule_insns_after_reload; + return optimize > 0 && flag_schedule_insns_after_reload + && dbg_cnt (sched2_func); #else return 0; #endif @@ -3312,27 +3459,26 @@ static unsigned int rest_of_handle_sched2 (void) { #ifdef INSN_SCHEDULING - /* Do control and data sched analysis again, - and write some more of the results to dump file. */ - - split_all_insns (1); - - if (flag_sched2_use_superblocks || flag_sched2_use_traces) + if (flag_selective_scheduling2 + && ! maybe_skip_selective_scheduling ()) + run_selective_scheduling (); + else { - schedule_ebbs (); - /* No liveness updating code yet, but it should be easy to do. - reg-stack recomputes the liveness when needed for now. */ - count_or_remove_death_notes (NULL, 1); - cleanup_cfg (CLEANUP_EXPENSIVE); + /* Do control and data sched analysis again, + and write some more of the results to dump file. */ + if (flag_sched2_use_superblocks || flag_sched2_use_traces) + schedule_ebbs (); + else + schedule_insns (); } - else - schedule_insns (); #endif return 0; } -struct tree_opt_pass pass_sched = +struct rtl_opt_pass pass_sched = { + { + RTL_PASS, "sched1", /* name */ gate_handle_sched, /* gate */ rest_of_handle_sched, /* execute */ @@ -3344,13 +3490,17 @@ struct tree_opt_pass pass_sched = 0, /* properties_provided */ 0, /* properties_destroyed */ 0, /* todo_flags_start */ + TODO_df_finish | TODO_verify_rtl_sharing | TODO_dump_func | - TODO_ggc_collect, /* todo_flags_finish */ - 'S' /* letter */ + TODO_verify_flow | + TODO_ggc_collect /* todo_flags_finish */ + } }; -struct tree_opt_pass pass_sched2 = +struct rtl_opt_pass pass_sched2 = { + { + RTL_PASS, "sched2", /* name */ gate_handle_sched2, /* gate */ rest_of_handle_sched2, /* execute */ @@ -3362,8 +3512,10 @@ struct tree_opt_pass pass_sched2 = 0, /* properties_provided */ 0, /* properties_destroyed */ 0, /* todo_flags_start */ + TODO_df_finish | TODO_verify_rtl_sharing | TODO_dump_func | - TODO_ggc_collect, /* todo_flags_finish */ - 'R' /* letter */ + TODO_verify_flow | + TODO_ggc_collect /* todo_flags_finish */ + } };