X-Git-Url: http://git.sourceforge.jp/view?p=pf3gnuchains%2Fgcc-fork.git;a=blobdiff_plain;f=gcc%2Fmodulo-sched.c;h=9179de504aa7d220849fad0c9280b43932883644;hp=84f33364eb12481510daab350da931c3641fb606;hb=0f7f01cfb1510f422a8b27e58bfed9e345116b32;hpb=30f4db0d0ca544491708a54ec8bb771fd697c48c diff --git a/gcc/modulo-sched.c b/gcc/modulo-sched.c index 84f33364eb1..9179de504aa 100644 --- a/gcc/modulo-sched.c +++ b/gcc/modulo-sched.c @@ -1,5 +1,5 @@ /* Swing Modulo Scheduling implementation. - Copyright (C) 2004 + Copyright (C) 2004, 2005, 2006, 2007, 2008, 2009, 2010 Free Software Foundation, Inc. Contributed by Ayal Zaks and Mustafa Hagog @@ -7,7 +7,7 @@ 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 @@ -16,27 +16,24 @@ 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, 59 Temple Place - Suite 330, Boston, MA -02111-1307, USA. */ +along with GCC; see the file COPYING3. If not see +. */ #include "config.h" #include "system.h" #include "coretypes.h" #include "tm.h" -#include "toplev.h" +#include "diagnostic-core.h" #include "rtl.h" #include "tm_p.h" #include "hard-reg-set.h" -#include "basic-block.h" #include "regs.h" #include "function.h" #include "flags.h" #include "insn-config.h" #include "insn-attr.h" #include "except.h" -#include "toplev.h" #include "recog.h" #include "sched-int.h" #include "target.h" @@ -46,8 +43,11 @@ Software Foundation, 59 Temple Place - Suite 330, Boston, MA #include "expr.h" #include "params.h" #include "gcov-io.h" -#include "df.h" #include "ddg.h" +#include "timevar.h" +#include "tree-pass.h" +#include "dbgcnt.h" +#include "df.h" #ifdef INSN_SCHEDULING @@ -83,8 +83,21 @@ Software Foundation, 59 Temple Place - Suite 330, Boston, MA perform modulo variable expansion for live ranges that span more than II cycles (i.e. use register copies to prevent a def from overwriting itself before reaching the use). -*/ + SMS works with countable loops (1) whose control part can be easily + decoupled from the rest of the loop and (2) whose loop count can + be easily adjusted. This is because we peel a constant number of + iterations into a prologue and epilogue for which we want to avoid + emitting the control part, and a kernel which is to iterate that + constant number of iterations less than the original loop. So the + control part should be a set of insns clearly identified and having + its own iv, not otherwise used in the loop (at-least for now), which + initializes a register before the loop to the number of iterations. + Currently SMS relies on the do-loop pattern to recognize such loops, + where (1) the control part comprises of all insns defining and/or + using a certain 'count' register and (2) the loop count can be + adjusted by modifying this register prior to the loop. + TODO: Rely on cfgloop analysis instead. */ /* This page defines partial-schedule structures and functions for modulo scheduling. */ @@ -106,8 +119,6 @@ typedef struct ps_insn *ps_insn_ptr; #define PS_STAGE_COUNT(ps) ((PS_MAX_CYCLE (ps) - PS_MIN_CYCLE (ps) \ + 1 + (ps)->ii - 1) / (ps)->ii) -#define CFG_HOOKS cfg_layout_rtl_cfg_hooks - /* A single instruction in the partial schedule. */ struct ps_insn { @@ -146,42 +157,44 @@ struct partial_schedule ddg_ptr g; /* The DDG of the insns in the partial schedule. */ }; +/* We use this to record all the register replacements we do in + the kernel so we can undo SMS if it is not profitable. */ +struct undo_replace_buff_elem +{ + rtx insn; + rtx orig_reg; + rtx new_reg; + struct undo_replace_buff_elem *next; +}; + -partial_schedule_ptr create_partial_schedule (int ii, ddg_ptr, int history); -void free_partial_schedule (partial_schedule_ptr); -void reset_partial_schedule (partial_schedule_ptr, int new_ii); + +static partial_schedule_ptr create_partial_schedule (int ii, ddg_ptr, int history); +static void free_partial_schedule (partial_schedule_ptr); +static void reset_partial_schedule (partial_schedule_ptr, int new_ii); void print_partial_schedule (partial_schedule_ptr, FILE *); -ps_insn_ptr ps_add_node_check_conflicts (partial_schedule_ptr, - ddg_node_ptr node, int cycle, - sbitmap must_precede, - sbitmap must_follow); -void rotate_partial_schedule (partial_schedule_ptr, int); +static void verify_partial_schedule (partial_schedule_ptr, sbitmap); +static ps_insn_ptr ps_add_node_check_conflicts (partial_schedule_ptr, + ddg_node_ptr node, int cycle, + sbitmap must_precede, + sbitmap must_follow); +static void rotate_partial_schedule (partial_schedule_ptr, int); void set_row_column_for_ps (partial_schedule_ptr); +static void ps_insert_empty_row (partial_schedule_ptr, int, sbitmap); +static int compute_split_row (sbitmap, int, int, int, ddg_node_ptr); /* This page defines constants and structures for the modulo scheduling driver. */ -/* As in haifa-sched.c: */ -/* issue_rate is the number of insns that can be scheduled in the same - machine cycle. It can be defined in the config/mach/mach.h file, - otherwise we set it to 1. */ - -static int issue_rate; - -/* For printing statistics. */ -static FILE *stats_file; - -static int sms_order_nodes (ddg_ptr, int, int * result); +static int sms_order_nodes (ddg_ptr, int, int *, int *); static void set_node_sched_params (ddg_ptr); -static partial_schedule_ptr sms_schedule_by_order (ddg_ptr, int, int, - int *, FILE*); -static void permute_partial_schedule (partial_schedule_ptr ps, rtx last); -static void generate_prolog_epilog (partial_schedule_ptr, rtx, rtx, int); -static void duplicate_insns_of_cycles (partial_schedule_ptr ps, - int from_stage, int to_stage, - int is_prolog); - +static partial_schedule_ptr sms_schedule_by_order (ddg_ptr, int, int, int *); +static void permute_partial_schedule (partial_schedule_ptr, rtx); +static void generate_prolog_epilog (partial_schedule_ptr, struct loop *, + rtx, rtx); +static void duplicate_insns_of_cycles (partial_schedule_ptr, + int, int, int, rtx); #define SCHED_ASAP(x) (((node_sched_params_ptr)(x)->aux.info)->asap) #define SCHED_TIME(x) (((node_sched_params_ptr)(x)->aux.info)->time) @@ -222,7 +235,7 @@ typedef struct node_sched_params code in order to use sched_analyze() for computing the dependencies. They are used when initializing the sched_info structure. */ static const char * -sms_print_insn (rtx insn, int aligned ATTRIBUTE_UNUSED) +sms_print_insn (const_rtx insn, int aligned ATTRIBUTE_UNUSED) { static char tmp[80]; @@ -230,12 +243,6 @@ sms_print_insn (rtx insn, int aligned ATTRIBUTE_UNUSED) return tmp; } -static int -contributes_to_priority (rtx next, rtx insn) -{ - return BLOCK_NUM (next) == BLOCK_NUM (insn); -} - static void compute_jump_reg_dependencies (rtx insn ATTRIBUTE_UNUSED, regset cond_exec ATTRIBUTE_UNUSED, @@ -244,7 +251,17 @@ compute_jump_reg_dependencies (rtx insn ATTRIBUTE_UNUSED, { } -static struct sched_info sms_sched_info = +static struct common_sched_info_def sms_common_sched_info; + +static struct sched_deps_info_def sms_sched_deps_info = + { + compute_jump_reg_dependencies, + NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, + NULL, + 0, 0, 0 + }; + +static struct haifa_sched_info sms_sched_info = { NULL, NULL, @@ -252,82 +269,67 @@ static struct sched_info sms_sched_info = NULL, NULL, sms_print_insn, - contributes_to_priority, - compute_jump_reg_dependencies, + NULL, + NULL, /* insn_finishes_block_p */ NULL, NULL, NULL, NULL, - 0, 0, 0 -}; + 0, 0, + NULL, NULL, NULL, + 0 +}; -/* Return the register decremented and tested or zero if it is not a decrement - and branch jump insn (similar to doloop_condition_get). */ +/* Given HEAD and TAIL which are the first and last insns in a loop; + return the register which controls the loop. Return zero if it has + more than one occurrence in the loop besides the control part or the + do-loop pattern is not of the form we expect. */ static rtx -doloop_register_get (rtx insn, rtx *comp) +doloop_register_get (rtx head ATTRIBUTE_UNUSED, rtx tail ATTRIBUTE_UNUSED) { - rtx pattern, cmp, inc, reg, condition; - - if (!JUMP_P (insn)) - return NULL_RTX; - pattern = PATTERN (insn); - - /* The canonical doloop pattern we expect is: - - (parallel [(set (pc) (if_then_else (condition) - (label_ref (label)) - (pc))) - (set (reg) (plus (reg) (const_int -1))) - (additional clobbers and uses)]) - - where condition is further restricted to be - (ne (reg) (const_int 1)). */ - - if (GET_CODE (pattern) != PARALLEL) - return NULL_RTX; - - cmp = XVECEXP (pattern, 0, 0); - inc = XVECEXP (pattern, 0, 1); - /* Return the compare rtx. */ - *comp = cmp; +#ifdef HAVE_doloop_end + rtx reg, condition, insn, first_insn_not_to_check; - /* Check for (set (reg) (something)). */ - if (GET_CODE (inc) != SET || ! REG_P (SET_DEST (inc))) + if (!JUMP_P (tail)) return NULL_RTX; - /* Extract loop counter register. */ - reg = SET_DEST (inc); - - /* Check if something = (plus (reg) (const_int -1)). */ - if (GET_CODE (SET_SRC (inc)) != PLUS - || XEXP (SET_SRC (inc), 0) != reg - || XEXP (SET_SRC (inc), 1) != constm1_rtx) + /* TODO: Free SMS's dependence on doloop_condition_get. */ + condition = doloop_condition_get (tail); + if (! condition) return NULL_RTX; - /* Check for (set (pc) (if_then_else (condition) - (label_ref (label)) - (pc))). */ - if (GET_CODE (cmp) != SET - || SET_DEST (cmp) != pc_rtx - || GET_CODE (SET_SRC (cmp)) != IF_THEN_ELSE - || GET_CODE (XEXP (SET_SRC (cmp), 1)) != LABEL_REF - || XEXP (SET_SRC (cmp), 2) != pc_rtx) - return NULL_RTX; + if (REG_P (XEXP (condition, 0))) + reg = XEXP (condition, 0); + else if (GET_CODE (XEXP (condition, 0)) == PLUS + && REG_P (XEXP (XEXP (condition, 0), 0))) + reg = XEXP (XEXP (condition, 0), 0); + else + gcc_unreachable (); - /* Extract loop termination condition. */ - condition = XEXP (SET_SRC (cmp), 0); + /* Check that the COUNT_REG has no other occurrences in the loop + until the decrement. We assume the control part consists of + either a single (parallel) branch-on-count or a (non-parallel) + branch immediately preceded by a single (decrement) insn. */ + first_insn_not_to_check = (GET_CODE (PATTERN (tail)) == PARALLEL ? tail + : PREV_INSN (tail)); - /* Check if condition = (ne (reg) (const_int 1)), which is more - restrictive than the check in doloop_condition_get: - if ((GET_CODE (condition) != GE && GET_CODE (condition) != NE) - || GET_CODE (XEXP (condition, 1)) != CONST_INT). */ - if (GET_CODE (condition) != NE - || XEXP (condition, 1) != const1_rtx) - return NULL_RTX; + for (insn = head; insn != first_insn_not_to_check; insn = NEXT_INSN (insn)) + if (reg_mentioned_p (reg, insn)) + { + if (dump_file) + { + fprintf (dump_file, "SMS count_reg found "); + print_rtl_single (dump_file, reg); + fprintf (dump_file, " outside control in insn:\n"); + print_rtl_single (dump_file, insn); + } - if (XEXP (condition, 0) == reg) - return reg; + return NULL_RTX; + } + return reg; +#else return NULL_RTX; +#endif } /* Check if COUNT_REG is set to a constant in the PRE_HEADER block, so @@ -340,15 +342,19 @@ const_iteration_count (rtx count_reg, basic_block pre_header, { rtx insn; rtx head, tail; - get_block_head_tail (pre_header->index, &head, &tail); + + if (! pre_header) + return NULL_RTX; + + get_ebb_head_tail (pre_header, pre_header, &head, &tail); for (insn = tail; insn != PREV_INSN (head); insn = PREV_INSN (insn)) - if (INSN_P (insn) && single_set (insn) && + if (NONDEBUG_INSN_P (insn) && single_set (insn) && rtx_equal_p (count_reg, SET_DEST (single_set (insn)))) { rtx pat = single_set (insn); - if (GET_CODE (SET_SRC (pat)) == CONST_INT) + if (CONST_INT_P (SET_SRC (pat))) { *count = INTVAL (SET_SRC (pat)); return insn; @@ -366,7 +372,10 @@ const_iteration_count (rtx count_reg, basic_block pre_header, static int res_MII (ddg_ptr g) { - return (g->num_nodes / issue_rate); + if (targetm.sched.sms_res_mii) + return targetm.sched.sms_res_mii (g); + + return ((g->num_nodes - g->num_debug) / issue_rate); } @@ -398,66 +407,50 @@ set_node_sched_params (ddg_ptr g) } static void -print_node_sched_params (FILE * dump_file, int num_nodes) +print_node_sched_params (FILE *file, int num_nodes, ddg_ptr g) { int i; + if (! file) + return; for (i = 0; i < num_nodes; i++) { node_sched_params_ptr nsp = &node_sched_params[i]; rtx reg_move = nsp->first_reg_move; int j; - fprintf (dump_file, "Node %d:\n", i); - fprintf (dump_file, " asap = %d:\n", nsp->asap); - fprintf (dump_file, " time = %d:\n", nsp->time); - fprintf (dump_file, " nreg_moves = %d:\n", nsp->nreg_moves); + fprintf (file, "Node = %d; INSN = %d\n", i, + (INSN_UID (g->nodes[i].insn))); + fprintf (file, " asap = %d:\n", nsp->asap); + fprintf (file, " time = %d:\n", nsp->time); + fprintf (file, " nreg_moves = %d:\n", nsp->nreg_moves); for (j = 0; j < nsp->nreg_moves; j++) { - fprintf (dump_file, " reg_move = "); - print_rtl_single (dump_file, reg_move); + fprintf (file, " reg_move = "); + print_rtl_single (file, reg_move); reg_move = PREV_INSN (reg_move); } } } -/* Calculate an upper bound for II. SMS should not schedule the loop if it - requires more cycles than this bound. Currently set to the sum of the - longest latency edge for each node. Reset based on experiments. */ -static int -calculate_maxii (ddg_ptr g) -{ - int i; - int maxii = 0; - - for (i = 0; i < g->num_nodes; i++) - { - ddg_node_ptr u = &g->nodes[i]; - ddg_edge_ptr e; - int max_edge_latency = 0; - - for (e = u->out; e; e = e->next_out) - max_edge_latency = MAX (max_edge_latency, e->latency); - - maxii += max_edge_latency; - } - return maxii; -} - - -/* Given the partial schedule, generate register moves when the length - of the register live range is more than ii; the number of moves is - determined according to the following equation: - SCHED_TIME (use) - SCHED_TIME (def) { 1 broken loop-carried - nreg_moves = ----------------------------------- - { dependence. - ii { 0 if not. - This handles the modulo-variable-expansions (mve's) needed for the ps. */ -static void -generate_reg_moves (partial_schedule_ptr ps) +/* + Breaking intra-loop register anti-dependences: + Each intra-loop register anti-dependence implies a cross-iteration true + dependence of distance 1. Therefore, we can remove such false dependencies + and figure out if the partial schedule broke them by checking if (for a + true-dependence of distance 1): SCHED_TIME (def) < SCHED_TIME (use) and + if so generate a register move. The number of such moves is equal to: + SCHED_TIME (use) - SCHED_TIME (def) { 0 broken + nreg_moves = ----------------------------------- + 1 - { dependence. + ii { 1 if not. +*/ +static struct undo_replace_buff_elem * +generate_reg_moves (partial_schedule_ptr ps, bool rescan) { ddg_ptr g = ps->g; int ii = ps->ii; int i; + struct undo_replace_buff_elem *reg_move_replaces = NULL; for (i = 0; i < g->num_nodes; i++) { @@ -475,6 +468,9 @@ generate_reg_moves (partial_schedule_ptr ps) { int nreg_moves4e = (SCHED_TIME (e->dest) - SCHED_TIME (e->src)) / ii; + if (e->distance == 1) + nreg_moves4e = (SCHED_TIME (e->dest) - SCHED_TIME (e->src) + ii) / ii; + /* If dest precedes src in the schedule of the kernel, then dest will read before src writes and we can save one reg_copy. */ if (SCHED_ROW (e->dest) == SCHED_ROW (e->src) @@ -498,6 +494,9 @@ generate_reg_moves (partial_schedule_ptr ps) { int dest_copy = (SCHED_TIME (e->dest) - SCHED_TIME (e->src)) / ii; + if (e->distance == 1) + dest_copy = (SCHED_TIME (e->dest) - SCHED_TIME (e->src) + ii) / ii; + if (SCHED_ROW (e->dest) == SCHED_ROW (e->src) && SCHED_COLUMN (e->dest) < SCHED_COLUMN (e->src)) dest_copy--; @@ -509,25 +508,64 @@ generate_reg_moves (partial_schedule_ptr ps) /* Now generate the reg_moves, attaching relevant uses to them. */ SCHED_NREG_MOVES (u) = nreg_moves; old_reg = prev_reg = copy_rtx (SET_DEST (single_set (u->insn))); - last_reg_move = u->insn; + /* Insert the reg-moves right before the notes which precede + the insn they relates to. */ + last_reg_move = u->first_note; for (i_reg_move = 0; i_reg_move < nreg_moves; i_reg_move++) { - int i_use; + unsigned int i_use = 0; rtx new_reg = gen_reg_rtx (GET_MODE (prev_reg)); rtx reg_move = gen_move_insn (new_reg, prev_reg); + sbitmap_iterator sbi; - add_insn_before (reg_move, last_reg_move); + add_insn_before (reg_move, last_reg_move, NULL); last_reg_move = reg_move; if (!SCHED_FIRST_REG_MOVE (u)) SCHED_FIRST_REG_MOVE (u) = reg_move; - EXECUTE_IF_SET_IN_SBITMAP (uses_of_defs[i_reg_move], 0, i_use, - replace_rtx (g->nodes[i_use].insn, old_reg, new_reg)); + EXECUTE_IF_SET_IN_SBITMAP (uses_of_defs[i_reg_move], 0, i_use, sbi) + { + struct undo_replace_buff_elem *rep; + + rep = (struct undo_replace_buff_elem *) + xcalloc (1, sizeof (struct undo_replace_buff_elem)); + rep->insn = g->nodes[i_use].insn; + rep->orig_reg = old_reg; + rep->new_reg = new_reg; + + if (! reg_move_replaces) + reg_move_replaces = rep; + else + { + rep->next = reg_move_replaces; + reg_move_replaces = rep; + } + + replace_rtx (g->nodes[i_use].insn, old_reg, new_reg); + if (rescan) + df_insn_rescan (g->nodes[i_use].insn); + } prev_reg = new_reg; } + sbitmap_vector_free (uses_of_defs); + } + return reg_move_replaces; +} + +/* Free memory allocated for the undo buffer. */ +static void +free_undo_replace_buff (struct undo_replace_buff_elem *reg_move_replaces) +{ + + while (reg_move_replaces) + { + struct undo_replace_buff_elem *rep = reg_move_replaces; + + reg_move_replaces = reg_move_replaces->next; + free (rep); } } @@ -536,23 +574,27 @@ generate_reg_moves (partial_schedule_ptr ps) static void normalize_sched_times (partial_schedule_ptr ps) { - int i; - ddg_ptr g = ps->g; + int row; int amount = PS_MIN_CYCLE (ps); int ii = ps->ii; + ps_insn_ptr crr_insn; - for (i = 0; i < g->num_nodes; i++) - { - ddg_node_ptr u = &g->nodes[i]; - int normalized_time = SCHED_TIME (u) - amount; - - if (normalized_time < 0) - abort (); - - SCHED_TIME (u) = normalized_time; - SCHED_ROW (u) = normalized_time % ii; - SCHED_STAGE (u) = normalized_time / ii; - } + for (row = 0; row < ii; row++) + for (crr_insn = ps->rows[row]; crr_insn; crr_insn = crr_insn->next_in_row) + { + ddg_node_ptr u = crr_insn->node; + int normalized_time = SCHED_TIME (u) - amount; + + if (dump_file) + fprintf (dump_file, "crr_insn->node=%d, crr_insn->cycle=%d,\ + min_cycle=%d\n", crr_insn->node->cuid, SCHED_TIME + (u), ps->min_cycle); + gcc_assert (SCHED_TIME (u) >= ps->min_cycle); + gcc_assert (SCHED_TIME (u) <= ps->max_cycle); + SCHED_TIME (u) = normalized_time; + SCHED_ROW (u) = normalized_time % ii; + SCHED_STAGE (u) = normalized_time / ii; + } } /* Set SCHED_COLUMN of each node according to its position in PS. */ @@ -588,12 +630,9 @@ permute_partial_schedule (partial_schedule_ptr ps, rtx last) PREV_INSN (last)); } -/* Used to generate the prologue & epilogue. Duplicate the subset of - nodes whose stages are between FROM_STAGE and TO_STAGE (inclusive - of both), together with a prefix/suffix of their reg_moves. */ static void duplicate_insns_of_cycles (partial_schedule_ptr ps, int from_stage, - int to_stage, int for_prolog) + int to_stage, int for_prolog, rtx count_reg) { int row; ps_insn_ptr ps_ij; @@ -605,12 +644,19 @@ duplicate_insns_of_cycles (partial_schedule_ptr ps, int from_stage, int j, i_reg_moves; rtx reg_move = NULL_RTX; + /* Do not duplicate any insn which refers to count_reg as it + belongs to the control part. + TODO: This should be done by analyzing the control part of + the loop. */ + if (reg_mentioned_p (count_reg, u_node->insn)) + continue; + if (for_prolog) { /* SCHED_STAGE (u_node) >= from_stage == 0. Generate increasing number of reg_moves starting with the second occurrence of u_node, which is generated if its SCHED_STAGE <= to_stage. */ - i_reg_moves = to_stage - SCHED_STAGE (u_node); + i_reg_moves = to_stage - SCHED_STAGE (u_node) + 1; i_reg_moves = MAX (i_reg_moves, 0); i_reg_moves = MIN (i_reg_moves, SCHED_NREG_MOVES (u_node)); @@ -644,7 +690,6 @@ duplicate_insns_of_cycles (partial_schedule_ptr ps, int from_stage, for (j = 0; j < i_reg_moves; j++, reg_move = NEXT_INSN (reg_move)) emit_insn (copy_rtx (PATTERN (reg_move))); - if (SCHED_STAGE (u_node) >= from_stage && SCHED_STAGE (u_node) <= to_stage) duplicate_insn_chain (u_node->first_note, u_node->insn); @@ -654,167 +699,209 @@ duplicate_insns_of_cycles (partial_schedule_ptr ps, int from_stage, /* Generate the instructions (including reg_moves) for prolog & epilog. */ static void -generate_prolog_epilog (partial_schedule_ptr ps, rtx orig_loop_beg, - rtx orig_loop_end, int unknown_count) +generate_prolog_epilog (partial_schedule_ptr ps, struct loop *loop, + rtx count_reg, rtx count_init) { int i; int last_stage = PS_STAGE_COUNT (ps) - 1; edge e; - rtx c_reg = NULL_RTX; - rtx cmp = NULL_RTX; - rtx precond_jump = NULL_RTX; - rtx precond_exit_label = NULL_RTX; - rtx precond_exit_label_insn = NULL_RTX; - rtx last_epilog_insn = NULL_RTX; - rtx loop_exit_label = NULL_RTX; - rtx loop_exit_label_insn = NULL_RTX; - rtx orig_loop_bct = NULL_RTX; - - /* Loop header edge. */ - e = ps->g->bb->pred; - if (e->src == ps->g->bb) - e = e->pred_next; /* Generate the prolog, inserting its insns on the loop-entry edge. */ start_sequence (); - /* This is the place where we want to insert the precondition. */ - if (unknown_count) - precond_jump = emit_note (NOTE_INSN_DELETED); + if (!count_init) + { + /* Generate instructions at the beginning of the prolog to + adjust the loop count by STAGE_COUNT. If loop count is constant + (count_init), this constant is adjusted by STAGE_COUNT in + generate_prolog_epilog function. */ + rtx sub_reg = NULL_RTX; + + sub_reg = expand_simple_binop (GET_MODE (count_reg), MINUS, + count_reg, GEN_INT (last_stage), + count_reg, 1, OPTAB_DIRECT); + gcc_assert (REG_P (sub_reg)); + if (REGNO (sub_reg) != REGNO (count_reg)) + emit_move_insn (count_reg, sub_reg); + } for (i = 0; i < last_stage; i++) - duplicate_insns_of_cycles (ps, 0, i, 1); + duplicate_insns_of_cycles (ps, 0, i, 1, count_reg); + + /* Put the prolog on the entry edge. */ + e = loop_preheader_edge (loop); + split_edge_and_insert (e, get_insns ()); - /* No need to call insert_insn_on_edge; we prepared the sequence. */ - e->insns.r = get_insns (); end_sequence (); /* Generate the epilog, inserting its insns on the loop-exit edge. */ start_sequence (); for (i = 0; i < last_stage; i++) - duplicate_insns_of_cycles (ps, i + 1, last_stage, 0); + duplicate_insns_of_cycles (ps, i + 1, last_stage, 0, count_reg); - last_epilog_insn = emit_note (NOTE_INSN_DELETED); + /* Put the epilogue on the exit edge. */ + gcc_assert (single_exit (loop)); + e = single_exit (loop); + split_edge_and_insert (e, get_insns ()); + end_sequence (); +} - /* Emit the label where to put the original loop code. */ - if (unknown_count) +/* Return true if all the BBs of the loop are empty except the + loop header. */ +static bool +loop_single_full_bb_p (struct loop *loop) +{ + unsigned i; + basic_block *bbs = get_loop_body (loop); + + for (i = 0; i < loop->num_nodes ; i++) { - rtx label, cond; + rtx head, tail; + bool empty_bb = true; - precond_exit_label = gen_label_rtx (); - precond_exit_label_insn = emit_label (precond_exit_label); + if (bbs[i] == loop->header) + continue; - /* Put the original loop code. */ - reorder_insns_nobb (orig_loop_beg, orig_loop_end, precond_exit_label_insn); + /* Make sure that basic blocks other than the header + have only notes labels or jumps. */ + get_ebb_head_tail (bbs[i], bbs[i], &head, &tail); + for (; head != NEXT_INSN (tail); head = NEXT_INSN (head)) + { + if (NOTE_P (head) || LABEL_P (head) + || (INSN_P (head) && (DEBUG_INSN_P (head) || JUMP_P (head)))) + continue; + empty_bb = false; + break; + } - /* Change the label of the BCT to be the PRECOND_EXIT_LABEL. */ - orig_loop_bct = get_last_insn (); - c_reg = doloop_register_get (orig_loop_bct, &cmp); - label = XEXP (SET_SRC (cmp), 1); - cond = XEXP (SET_SRC (cmp), 0); + if (! empty_bb) + { + free (bbs); + return false; + } + } + free (bbs); + return true; +} + +/* A simple loop from SMS point of view; it is a loop that is composed of + either a single basic block or two BBs - a header and a latch. */ +#define SIMPLE_SMS_LOOP_P(loop) ((loop->num_nodes < 3 ) \ + && (EDGE_COUNT (loop->latch->preds) == 1) \ + && (EDGE_COUNT (loop->latch->succs) == 1)) + +/* Return true if the loop is in its canonical form and false if not. + i.e. SIMPLE_SMS_LOOP_P and have one preheader block, and single exit. */ +static bool +loop_canon_p (struct loop *loop) +{ - if (! c_reg || GET_CODE (cond) != NE) - abort (); + if (loop->inner || !loop_outer (loop)) + { + if (dump_file) + fprintf (dump_file, "SMS loop inner or !loop_outer\n"); + return false; + } - XEXP (label, 0) = precond_exit_label; - JUMP_LABEL (orig_loop_bct) = precond_exit_label_insn; - LABEL_NUSES (precond_exit_label_insn)++; + if (!single_exit (loop)) + { + if (dump_file) + { + rtx insn = BB_END (loop->header); - /* Generate the loop exit label. */ - loop_exit_label = gen_label_rtx (); - loop_exit_label_insn = emit_label (loop_exit_label); + fprintf (dump_file, "SMS loop many exits "); + fprintf (dump_file, " %s %d (file, line)\n", + insn_file (insn), insn_line (insn)); + } + return false; } - e = ps->g->bb->succ; - if (e->dest == ps->g->bb) - e = e->succ_next; + if (! SIMPLE_SMS_LOOP_P (loop) && ! loop_single_full_bb_p (loop)) + { + if (dump_file) + { + rtx insn = BB_END (loop->header); - e->insns.r = get_insns (); - end_sequence (); + fprintf (dump_file, "SMS loop many BBs. "); + fprintf (dump_file, " %s %d (file, line)\n", + insn_file (insn), insn_line (insn)); + } + return false; + } - commit_edge_insertions (); + return true; +} - if (unknown_count) +/* If there are more than one entry for the loop, + make it one by splitting the first entry edge and + redirecting the others to the new BB. */ +static void +canon_loop (struct loop *loop) +{ + edge e; + edge_iterator i; + + /* Avoid annoying special cases of edges going to exit + block. */ + FOR_EACH_EDGE (e, i, EXIT_BLOCK_PTR->preds) + if ((e->flags & EDGE_FALLTHRU) && (EDGE_COUNT (e->src->succs) > 1)) + split_edge (e); + + if (loop->latch == loop->header + || EDGE_COUNT (loop->latch->succs) > 1) { - rtx precond_insns, epilog_jump, insert_after_insn; - basic_block loop_exit_bb = BLOCK_FOR_INSN (loop_exit_label_insn); - basic_block epilog_bb = BLOCK_FOR_INSN (last_epilog_insn); - basic_block precond_bb = BLOCK_FOR_INSN (precond_jump); - basic_block orig_loop_bb = BLOCK_FOR_INSN (precond_exit_label_insn); - edge epilog_exit_edge = epilog_bb->succ; - - /* Do loop preconditioning to take care of cases were the loop count is - less than the stage count. Update the CFG properly. */ - insert_after_insn = precond_jump; - start_sequence (); - c_reg = doloop_register_get (ps->g->closing_branch->insn, &cmp); - emit_cmp_and_jump_insns (c_reg, GEN_INT (PS_STAGE_COUNT (ps)), LT, NULL, - GET_MODE (c_reg), 1, precond_exit_label); - precond_insns = get_insns (); - precond_jump = get_last_insn (); - end_sequence (); - reorder_insns (precond_insns, precond_jump, insert_after_insn); - - /* Generate a subtract instruction at the beginning of the prolog to - adjust the loop count by STAGE_COUNT. */ - emit_insn_after (gen_sub2_insn (c_reg, GEN_INT (PS_STAGE_COUNT (ps) - 1)), - precond_jump); - update_bb_for_insn (precond_bb); - delete_insn (insert_after_insn); - - /* Update label info for the precondition jump. */ - JUMP_LABEL (precond_jump) = precond_exit_label_insn; - LABEL_NUSES (precond_exit_label_insn)++; - - /* Update the CFG. */ - split_block (precond_bb, precond_jump); - make_edge (precond_bb, orig_loop_bb, 0); - - /* Add a jump at end of the epilog to the LOOP_EXIT_LABEL to jump over the - original loop copy and update the CFG. */ - epilog_jump = emit_jump_insn_after (gen_jump (loop_exit_label), - last_epilog_insn); - delete_insn (last_epilog_insn); - JUMP_LABEL (epilog_jump) = loop_exit_label_insn; - LABEL_NUSES (loop_exit_label_insn)++; - - redirect_edge_succ (epilog_exit_edge, loop_exit_bb); - epilog_exit_edge->flags &= ~EDGE_FALLTHRU; - emit_barrier_after (BB_END (epilog_bb)); + FOR_EACH_EDGE (e, i, loop->header->preds) + if (e->src == loop->latch) + break; + split_edge (e); } } -/* Return the line note insn preceding INSN, for debugging. Taken from - emit-rtl.c. */ -static rtx -find_line_note (rtx insn) +/* Setup infos. */ +static void +setup_sched_infos (void) { - for (; insn; insn = PREV_INSN (insn)) - if (NOTE_P (insn) - && NOTE_LINE_NUMBER (insn) >= 0) - break; + memcpy (&sms_common_sched_info, &haifa_common_sched_info, + sizeof (sms_common_sched_info)); + sms_common_sched_info.sched_pass_id = SCHED_SMS_PASS; + common_sched_info = &sms_common_sched_info; - return insn; + sched_deps_info = &sms_sched_deps_info; + current_sched_info = &sms_sched_info; } +/* Probability in % that the sms-ed loop rolls enough so that optimized + version may be entered. Just a guess. */ +#define PROB_SMS_ENOUGH_ITERATIONS 80 + +/* Used to calculate the upper bound of ii. */ +#define MAXII_FACTOR 2 + /* Main entry point, perform SMS scheduling on the loops of the function that consist of single basic blocks. */ -void -sms_schedule (FILE *dump_file) +static void +sms_schedule (void) { - static int passes = 0; rtx insn; ddg_ptr *g_arr, g; - basic_block bb, pre_header = NULL; int * node_order; - int maxii; - int i; + int maxii, max_asap; + loop_iterator li; partial_schedule_ptr ps; - int max_bb_index = last_basic_block; - struct df *df; - - stats_file = dump_file; + basic_block bb = NULL; + struct loop *loop; + basic_block condition_bb = NULL; + edge latch_edge; + gcov_type trip_count = 0; + + loop_optimizer_init (LOOPS_HAVE_PREHEADERS + | LOOPS_HAVE_RECORDED_EXITS); + if (number_of_loops () <= 1) + { + loop_optimizer_finalize (); + return; /* There are no loops to schedule. */ + } /* Initialize issue_rate. */ if (targetm.sched.issue_rate) @@ -822,257 +909,283 @@ sms_schedule (FILE *dump_file) int temp = reload_completed; reload_completed = 1; - issue_rate = (*targetm.sched.issue_rate) (); + issue_rate = targetm.sched.issue_rate (); reload_completed = temp; } else issue_rate = 1; /* Initialize the scheduler. */ - current_sched_info = &sms_sched_info; - sched_init (NULL); + setup_sched_infos (); + haifa_sched_init (); - /* Init Data Flow analysis, to be used in interloop dep calculation. */ - df = df_init (); - df_analyze (df, 0, DF_ALL); + /* Allocate memory to hold the DDG array one entry for each loop. + We use loop->num as index into this array. */ + g_arr = XCNEWVEC (ddg_ptr, number_of_loops ()); - /* Allocate memory to hold the DDG array. */ - g_arr = xcalloc (max_bb_index, sizeof (ddg_ptr)); + if (dump_file) + { + fprintf (dump_file, "\n\nSMS analysis phase\n"); + fprintf (dump_file, "===================\n\n"); + } /* Build DDGs for all the relevant loops and hold them in G_ARR - indexed by the loop BB index. */ - FOR_EACH_BB (bb) + indexed by the loop index. */ + FOR_EACH_LOOP (li, loop, 0) { rtx head, tail; - rtx count_reg, comp; - edge e, pre_header_edge; + rtx count_reg; - if (bb->index < 0) - continue; + /* For debugging. */ + if (dbg_cnt (sms_sched_loop) == false) + { + if (dump_file) + fprintf (dump_file, "SMS reached max limit... \n"); - /* Check if bb has two successors, one being itself. */ - e = bb->succ; - if (!e || !e->succ_next || e->succ_next->succ_next) - continue; + break; + } - if (e->dest != bb && e->succ_next->dest != bb) - continue; + if (dump_file) + { + rtx insn = BB_END (loop->header); - if ((e->flags & EDGE_COMPLEX) - || (e->succ_next->flags & EDGE_COMPLEX)) - continue; + fprintf (dump_file, "SMS loop num: %d, file: %s, line: %d\n", + loop->num, insn_file (insn), insn_line (insn)); - /* Check if bb has two predecessors, one being itself. */ - /* In view of above tests, suffices to check e->pred_next->pred_next? */ - e = bb->pred; - if (!e || !e->pred_next || e->pred_next->pred_next) - continue; + } - if (e->src != bb && e->pred_next->src != bb) - continue; + if (! loop_canon_p (loop)) + continue; - if ((e->flags & EDGE_COMPLEX) - || (e->pred_next->flags & EDGE_COMPLEX)) + if (! loop_single_full_bb_p (loop)) + { + if (dump_file) + fprintf (dump_file, "SMS not loop_single_full_bb_p\n"); continue; + } - /* For debugging. */ - if ((passes++ > MAX_SMS_LOOP_NUMBER) && (MAX_SMS_LOOP_NUMBER != -1)) - { - if (dump_file) - fprintf (dump_file, "SMS reached MAX_PASSES... \n"); - break; - } + bb = loop->header; - get_block_head_tail (bb->index, &head, &tail); - pre_header_edge = bb->pred; - if (bb->pred->src != bb) - pre_header_edge = bb->pred->pred_next; + get_ebb_head_tail (bb, bb, &head, &tail); + latch_edge = loop_latch_edge (loop); + gcc_assert (single_exit (loop)); + if (single_exit (loop)->count) + trip_count = latch_edge->count / single_exit (loop)->count; - /* Perfrom SMS only on loops that their average count is above threshold. */ - if (bb->count < pre_header_edge->count * SMS_LOOP_AVERAGE_COUNT_THRESHOLD) - { - if (stats_file) - { - rtx line_note = find_line_note (tail); + /* Perform SMS only on loops that their average count is above threshold. */ - if (line_note) - { - expanded_location xloc; - NOTE_EXPANDED_LOCATION (xloc, line_note); - fprintf (stats_file, "SMS bb %s %d (file, line)\n", - xloc.file, xloc.line); - } - fprintf (stats_file, "SMS single-bb-loop\n"); + if ( latch_edge->count + && (latch_edge->count < single_exit (loop)->count * SMS_LOOP_AVERAGE_COUNT_THRESHOLD)) + { + if (dump_file) + { + fprintf (dump_file, " %s %d (file, line)\n", + insn_file (tail), insn_line (tail)); + fprintf (dump_file, "SMS single-bb-loop\n"); if (profile_info && flag_branch_probabilities) { - fprintf (stats_file, "SMS loop-count "); - fprintf (stats_file, HOST_WIDEST_INT_PRINT_DEC, + fprintf (dump_file, "SMS loop-count "); + fprintf (dump_file, HOST_WIDEST_INT_PRINT_DEC, (HOST_WIDEST_INT) bb->count); - fprintf (stats_file, "\n"); - fprintf (stats_file, "SMS preheader-count "); - fprintf (stats_file, HOST_WIDEST_INT_PRINT_DEC, - (HOST_WIDEST_INT) pre_header_edge->count); - fprintf (stats_file, "\n"); - fprintf (stats_file, "SMS profile-sum-max "); - fprintf (stats_file, HOST_WIDEST_INT_PRINT_DEC, + fprintf (dump_file, "\n"); + fprintf (dump_file, "SMS trip-count "); + fprintf (dump_file, HOST_WIDEST_INT_PRINT_DEC, + (HOST_WIDEST_INT) trip_count); + fprintf (dump_file, "\n"); + fprintf (dump_file, "SMS profile-sum-max "); + fprintf (dump_file, HOST_WIDEST_INT_PRINT_DEC, (HOST_WIDEST_INT) profile_info->sum_max); - fprintf (stats_file, "\n"); + fprintf (dump_file, "\n"); } } continue; } /* Make sure this is a doloop. */ - if ( !(count_reg = doloop_register_get (tail, &comp))) + if ( !(count_reg = doloop_register_get (head, tail))) + { + if (dump_file) + fprintf (dump_file, "SMS doloop_register_get failed\n"); continue; + } - e = bb->pred; - if (e->src == bb) - pre_header = e->pred_next->src; - else - pre_header = e->src; - - /* Don't handle BBs with calls or barriers, or !single_set insns. */ - for (insn = head; insn != NEXT_INSN (tail); insn = NEXT_INSN (insn)) - if (CALL_P (insn) - || BARRIER_P (insn) - || (INSN_P (insn) && !JUMP_P (insn) - && !single_set (insn) && GET_CODE (PATTERN (insn)) != USE)) - break; + /* Don't handle BBs with calls or barriers, or !single_set insns, + or auto-increment insns (to avoid creating invalid reg-moves + for the auto-increment insns). + ??? Should handle auto-increment insns. + ??? Should handle insns defining subregs. */ + for (insn = head; insn != NEXT_INSN (tail); insn = NEXT_INSN (insn)) + { + rtx set; + + if (CALL_P (insn) + || BARRIER_P (insn) + || (NONDEBUG_INSN_P (insn) && !JUMP_P (insn) + && !single_set (insn) && GET_CODE (PATTERN (insn)) != USE) + || (FIND_REG_INC_NOTE (insn, NULL_RTX) != 0) + || (INSN_P (insn) && (set = single_set (insn)) + && GET_CODE (SET_DEST (set)) == SUBREG)) + break; + } if (insn != NEXT_INSN (tail)) { - if (stats_file) + if (dump_file) { if (CALL_P (insn)) - fprintf (stats_file, "SMS loop-with-call\n"); + fprintf (dump_file, "SMS loop-with-call\n"); else if (BARRIER_P (insn)) - fprintf (stats_file, "SMS loop-with-barrier\n"); - else - fprintf (stats_file, "SMS loop-with-not-single-set\n"); - print_rtl_single (stats_file, insn); + fprintf (dump_file, "SMS loop-with-barrier\n"); + else if (FIND_REG_INC_NOTE (insn, NULL_RTX) != 0) + fprintf (dump_file, "SMS reg inc\n"); + else if ((NONDEBUG_INSN_P (insn) && !JUMP_P (insn) + && !single_set (insn) && GET_CODE (PATTERN (insn)) != USE)) + fprintf (dump_file, "SMS loop-with-not-single-set\n"); + else + fprintf (dump_file, "SMS loop with subreg in lhs\n"); + print_rtl_single (dump_file, insn); } continue; } - if (! (g = create_ddg (bb, df, 0))) + if (! (g = create_ddg (bb, 0))) { - if (stats_file) - fprintf (stats_file, "SMS doloop\n"); + if (dump_file) + fprintf (dump_file, "SMS create_ddg failed\n"); continue; } - g_arr[bb->index] = g; - } - - /* Release Data Flow analysis data structures. */ - df_finish (df); + g_arr[loop->num] = g; + if (dump_file) + fprintf (dump_file, "...OK\n"); - /* Go over the built DDGs and perfrom SMS for each one of them. */ - for (i = 0; i < max_bb_index; i++) + } + if (dump_file) + { + fprintf (dump_file, "\nSMS transformation phase\n"); + fprintf (dump_file, "=========================\n\n"); + } + + /* We don't want to perform SMS on new loops - created by versioning. */ + FOR_EACH_LOOP (li, loop, 0) { rtx head, tail; - rtx count_reg, count_init, comp; - edge pre_header_edge; + rtx count_reg, count_init; int mii, rec_mii; - int stage_count = 0; + unsigned stage_count = 0; HOST_WIDEST_INT loop_count = 0; - if (! (g = g_arr[i])) + if (! (g = g_arr[loop->num])) continue; if (dump_file) - print_ddg (dump_file, g); + { + rtx insn = BB_END (loop->header); - get_block_head_tail (g->bb->index, &head, &tail); + fprintf (dump_file, "SMS loop num: %d, file: %s, line: %d\n", + loop->num, insn_file (insn), insn_line (insn)); - pre_header_edge = g->bb->pred; - if (g->bb->pred->src != g->bb) - pre_header_edge = g->bb->pred->pred_next; + print_ddg (dump_file, g); + } - if (stats_file) - { - rtx line_note = find_line_note (tail); + get_ebb_head_tail (loop->header, loop->header, &head, &tail); - if (line_note) - { - expanded_location xloc; - NOTE_EXPANDED_LOCATION (xloc, line_note); - fprintf (stats_file, "SMS bb %s %d (file, line)\n", - xloc.file, xloc.line); - } - fprintf (stats_file, "SMS single-bb-loop\n"); + latch_edge = loop_latch_edge (loop); + gcc_assert (single_exit (loop)); + if (single_exit (loop)->count) + trip_count = latch_edge->count / single_exit (loop)->count; + + if (dump_file) + { + fprintf (dump_file, " %s %d (file, line)\n", + insn_file (tail), insn_line (tail)); + fprintf (dump_file, "SMS single-bb-loop\n"); if (profile_info && flag_branch_probabilities) { - fprintf (stats_file, "SMS loop-count "); - fprintf (stats_file, HOST_WIDEST_INT_PRINT_DEC, + fprintf (dump_file, "SMS loop-count "); + fprintf (dump_file, HOST_WIDEST_INT_PRINT_DEC, (HOST_WIDEST_INT) bb->count); - fprintf (stats_file, "\n"); - fprintf (stats_file, "SMS preheader-count "); - fprintf (stats_file, HOST_WIDEST_INT_PRINT_DEC, - (HOST_WIDEST_INT) pre_header_edge->count); - fprintf (stats_file, "\n"); - fprintf (stats_file, "SMS profile-sum-max "); - fprintf (stats_file, HOST_WIDEST_INT_PRINT_DEC, + fprintf (dump_file, "\n"); + fprintf (dump_file, "SMS profile-sum-max "); + fprintf (dump_file, HOST_WIDEST_INT_PRINT_DEC, (HOST_WIDEST_INT) profile_info->sum_max); - fprintf (stats_file, "\n"); + fprintf (dump_file, "\n"); } - fprintf (stats_file, "SMS doloop\n"); - fprintf (stats_file, "SMS built-ddg %d\n", g->num_nodes); - fprintf (stats_file, "SMS num-loads %d\n", g->num_loads); - fprintf (stats_file, "SMS num-stores %d\n", g->num_stores); + fprintf (dump_file, "SMS doloop\n"); + fprintf (dump_file, "SMS built-ddg %d\n", g->num_nodes); + fprintf (dump_file, "SMS num-loads %d\n", g->num_loads); + fprintf (dump_file, "SMS num-stores %d\n", g->num_stores); } - /* Make sure this is a doloop. */ - if ( !(count_reg = doloop_register_get (tail, &comp))) - abort (); - /* This should be NULL_RTX if the count is unknown at compile time. */ - count_init = const_iteration_count (count_reg, pre_header, &loop_count); + /* In case of th loop have doloop register it gets special + handling. */ + count_init = NULL_RTX; + if ((count_reg = doloop_register_get (head, tail))) + { + basic_block pre_header; + + pre_header = loop_preheader_edge (loop)->src; + count_init = const_iteration_count (count_reg, pre_header, + &loop_count); + } + gcc_assert (count_reg); - if (stats_file && count_init) + if (dump_file && count_init) { - fprintf (stats_file, "SMS const-doloop "); - fprintf (stats_file, HOST_WIDEST_INT_PRINT_DEC, loop_count); - fprintf (stats_file, "\n"); + fprintf (dump_file, "SMS const-doloop "); + fprintf (dump_file, HOST_WIDEST_INT_PRINT_DEC, + loop_count); + fprintf (dump_file, "\n"); } - node_order = (int *) xmalloc (sizeof (int) * g->num_nodes); + node_order = XNEWVEC (int, g->num_nodes); mii = 1; /* Need to pass some estimate of mii. */ - rec_mii = sms_order_nodes (g, mii, node_order); + rec_mii = sms_order_nodes (g, mii, node_order, &max_asap); mii = MAX (res_MII (g), rec_mii); - maxii = (calculate_maxii (g) * SMS_MAX_II_FACTOR) / 100; + maxii = MAX (max_asap, MAXII_FACTOR * mii); - if (stats_file) - fprintf (stats_file, "SMS iis %d %d %d (rec_mii, mii, maxii)\n", + if (dump_file) + fprintf (dump_file, "SMS iis %d %d %d (rec_mii, mii, maxii)\n", rec_mii, mii, maxii); /* After sms_order_nodes and before sms_schedule_by_order, to copy over ASAP. */ set_node_sched_params (g); - ps = sms_schedule_by_order (g, mii, maxii, node_order, dump_file); + ps = sms_schedule_by_order (g, mii, maxii, node_order); - if (ps) + if (ps){ stage_count = PS_STAGE_COUNT (ps); + gcc_assert(stage_count >= 1); + } - if (stage_count == 0 || (count_init && (stage_count > loop_count))) + /* Stage count of 1 means that there is no interleaving between + iterations, let the scheduling passes do the job. */ + if (stage_count <= 1 + || (count_init && (loop_count <= stage_count)) + || (flag_branch_probabilities && (trip_count <= stage_count))) { if (dump_file) - fprintf (dump_file, "SMS failed... \n"); - if (stats_file) - fprintf (stats_file, "SMS sched-failed %d\n", stage_count); + { + fprintf (dump_file, "SMS failed... \n"); + fprintf (dump_file, "SMS sched-failed (stage-count=%d, loop-count=", stage_count); + fprintf (dump_file, HOST_WIDEST_INT_PRINT_DEC, loop_count); + fprintf (dump_file, ", trip-count="); + fprintf (dump_file, HOST_WIDEST_INT_PRINT_DEC, trip_count); + fprintf (dump_file, ")\n"); + } + continue; } else { - rtx orig_loop_beg = NULL_RTX; - rtx orig_loop_end = NULL_RTX; + struct undo_replace_buff_elem *reg_move_replaces; - if (stats_file) + if (dump_file) { - fprintf (stats_file, + fprintf (dump_file, "SMS succeeded %d %d (with ii, sc)\n", ps->ii, stage_count); print_partial_schedule (ps, dump_file); @@ -1081,61 +1194,68 @@ sms_schedule (FILE *dump_file) g->closing_branch->cuid, PS_MIN_CYCLE (ps) - 1); } - /* Save the original loop if we want to do loop preconditioning in - case the BCT count is not known. */ - if (! count_init) - { - int i; - - start_sequence (); - /* Copy the original loop code before modifying it - - so we can use it later. */ - for (i = 0; i < ps->g->num_nodes; i++) - duplicate_insn_chain (ps->g->nodes[i].first_note, - ps->g->nodes[i].insn); - - orig_loop_beg = get_insns (); - orig_loop_end = get_last_insn (); - end_sequence (); - } /* Set the stage boundaries. If the DDG is built with closing_branch_deps, the closing_branch was scheduled and should appear in the last (ii-1) row. Otherwise, we are free to schedule the branch, and we let nodes that were scheduled at the first PS_MIN_CYCLE cycle appear in the first - row; this should reduce stage_count to minimum. */ + row; this should reduce stage_count to minimum. + TODO: Revisit the issue of scheduling the insns of the + control part relative to the branch when the control part + has more than one insn. */ normalize_sched_times (ps); rotate_partial_schedule (ps, PS_MIN_CYCLE (ps)); set_columns_for_ps (ps); + canon_loop (loop); + + /* case the BCT count is not known , Do loop-versioning */ + if (count_reg && ! count_init) + { + rtx comp_rtx = gen_rtx_fmt_ee (GT, VOIDmode, count_reg, + GEN_INT(stage_count)); + unsigned prob = (PROB_SMS_ENOUGH_ITERATIONS + * REG_BR_PROB_BASE) / 100; + + loop_version (loop, comp_rtx, &condition_bb, + prob, prob, REG_BR_PROB_BASE - prob, + true); + } + + /* Set new iteration count of loop kernel. */ + if (count_reg && count_init) + SET_SRC (single_set (count_init)) = GEN_INT (loop_count + - stage_count + 1); + + /* Now apply the scheduled kernel to the RTL of the loop. */ permute_partial_schedule (ps, g->closing_branch->first_note); /* Mark this loop as software pipelined so the later scheduling passes doesn't touch it. */ if (! flag_resched_modulo_sched) - emit_note_before (NOTE_INSN_DISABLE_SCHED_OF_BLOCK, - g->closing_branch->insn); + g->bb->flags |= BB_DISABLE_SCHEDULE; + /* The life-info is not valid any more. */ + df_set_bb_dirty (g->bb); - generate_reg_moves (ps); + reg_move_replaces = generate_reg_moves (ps, true); if (dump_file) - print_node_sched_params (dump_file, g->num_nodes); - - /* Set new iteration count of loop kernel. */ - if (count_init) - SET_SRC (single_set (count_init)) = GEN_INT (loop_count - - stage_count + 1); - + print_node_sched_params (dump_file, g->num_nodes, g); /* Generate prolog and epilog. */ - generate_prolog_epilog (ps, orig_loop_beg, orig_loop_end, - count_init ? 0 : 1); + generate_prolog_epilog (ps, loop, count_reg, count_init); + + free_undo_replace_buff (reg_move_replaces); } + free_partial_schedule (ps); free (node_sched_params); free (node_order); free_ddg (g); } + free (g_arr); + /* Release scheduler data, needed until now because of DFA. */ - sched_finish (); + haifa_sched_finish (); + loop_optimizer_finalize (); } /* The SMS scheduling algorithm itself @@ -1214,205 +1334,698 @@ sms_schedule (FILE *dump_file) set to 0 to save compile time. */ #define DFA_HISTORY SMS_DFA_HISTORY +/* A threshold for the number of repeated unsuccessful attempts to insert + an empty row, before we flush the partial schedule and start over. */ +#define MAX_SPLIT_NUM 10 +/* Given the partial schedule PS, this function calculates and returns the + cycles in which we can schedule the node with the given index I. + NOTE: Here we do the backtracking in SMS, in some special cases. We have + noticed that there are several cases in which we fail to SMS the loop + because the sched window of a node is empty due to tight data-deps. In + such cases we want to unschedule some of the predecessors/successors + until we get non-empty scheduling window. It returns -1 if the + scheduling window is empty and zero otherwise. */ + +static int +get_sched_window (partial_schedule_ptr ps, int *nodes_order, int i, + sbitmap sched_nodes, int ii, int *start_p, int *step_p, int *end_p) +{ + int start, step, end; + ddg_edge_ptr e; + int u = nodes_order [i]; + ddg_node_ptr u_node = &ps->g->nodes[u]; + sbitmap psp = sbitmap_alloc (ps->g->num_nodes); + sbitmap pss = sbitmap_alloc (ps->g->num_nodes); + sbitmap u_node_preds = NODE_PREDECESSORS (u_node); + sbitmap u_node_succs = NODE_SUCCESSORS (u_node); + int psp_not_empty; + int pss_not_empty; + + /* 1. compute sched window for u (start, end, step). */ + sbitmap_zero (psp); + sbitmap_zero (pss); + psp_not_empty = sbitmap_a_and_b_cg (psp, u_node_preds, sched_nodes); + pss_not_empty = sbitmap_a_and_b_cg (pss, u_node_succs, sched_nodes); + + if (psp_not_empty && !pss_not_empty) + { + int early_start = INT_MIN; + + end = INT_MAX; + for (e = u_node->in; e != 0; e = e->next_in) + { + ddg_node_ptr v_node = e->src; + + if (dump_file) + { + fprintf (dump_file, "\nProcessing edge: "); + print_ddg_edge (dump_file, e); + fprintf (dump_file, + "\nScheduling %d (%d) in psp_not_empty," + " checking p %d (%d): ", u_node->cuid, + INSN_UID (u_node->insn), v_node->cuid, INSN_UID + (v_node->insn)); + } + + if (TEST_BIT (sched_nodes, v_node->cuid)) + { + int p_st = SCHED_TIME (v_node); + + early_start = + MAX (early_start, p_st + e->latency - (e->distance * ii)); + + if (dump_file) + fprintf (dump_file, + "pred st = %d; early_start = %d; latency: %d", + p_st, early_start, e->latency); + + if (e->data_type == MEM_DEP) + end = MIN (end, SCHED_TIME (v_node) + ii - 1); + } + else if (dump_file) + fprintf (dump_file, "the node is not scheduled\n"); + } + start = early_start; + end = MIN (end, early_start + ii); + /* Schedule the node close to it's predecessors. */ + step = 1; + + if (dump_file) + fprintf (dump_file, + "\nScheduling %d (%d) in a window (%d..%d) with step %d\n", + u_node->cuid, INSN_UID (u_node->insn), start, end, step); + } + + else if (!psp_not_empty && pss_not_empty) + { + int late_start = INT_MAX; + + end = INT_MIN; + for (e = u_node->out; e != 0; e = e->next_out) + { + ddg_node_ptr v_node = e->dest; + + if (dump_file) + { + fprintf (dump_file, "\nProcessing edge:"); + print_ddg_edge (dump_file, e); + fprintf (dump_file, + "\nScheduling %d (%d) in pss_not_empty," + " checking s %d (%d): ", u_node->cuid, + INSN_UID (u_node->insn), v_node->cuid, INSN_UID + (v_node->insn)); + } + + if (TEST_BIT (sched_nodes, v_node->cuid)) + { + int s_st = SCHED_TIME (v_node); + + late_start = MIN (late_start, + s_st - e->latency + (e->distance * ii)); + + if (dump_file) + fprintf (dump_file, + "succ st = %d; late_start = %d; latency = %d", + s_st, late_start, e->latency); + + if (e->data_type == MEM_DEP) + end = MAX (end, SCHED_TIME (v_node) - ii + 1); + if (dump_file) + fprintf (dump_file, "end = %d\n", end); + + } + else if (dump_file) + fprintf (dump_file, "the node is not scheduled\n"); + + } + start = late_start; + end = MAX (end, late_start - ii); + /* Schedule the node close to it's successors. */ + step = -1; + + if (dump_file) + fprintf (dump_file, + "\nScheduling %d (%d) in a window (%d..%d) with step %d\n", + u_node->cuid, INSN_UID (u_node->insn), start, end, step); + + } + + else if (psp_not_empty && pss_not_empty) + { + int early_start = INT_MIN; + int late_start = INT_MAX; + int count_preds = 0; + int count_succs = 0; + + start = INT_MIN; + end = INT_MAX; + for (e = u_node->in; e != 0; e = e->next_in) + { + ddg_node_ptr v_node = e->src; + + if (dump_file) + { + fprintf (dump_file, "\nProcessing edge:"); + print_ddg_edge (dump_file, e); + fprintf (dump_file, + "\nScheduling %d (%d) in psp_pss_not_empty," + " checking p %d (%d): ", u_node->cuid, INSN_UID + (u_node->insn), v_node->cuid, INSN_UID + (v_node->insn)); + } + + if (TEST_BIT (sched_nodes, v_node->cuid)) + { + int p_st = SCHED_TIME (v_node); + + early_start = MAX (early_start, + p_st + e->latency + - (e->distance * ii)); + + if (dump_file) + fprintf (dump_file, + "pred st = %d; early_start = %d; latency = %d", + p_st, early_start, e->latency); + + if (e->type == TRUE_DEP && e->data_type == REG_DEP) + count_preds++; + + if (e->data_type == MEM_DEP) + end = MIN (end, SCHED_TIME (v_node) + ii - 1); + } + else if (dump_file) + fprintf (dump_file, "the node is not scheduled\n"); + + } + for (e = u_node->out; e != 0; e = e->next_out) + { + ddg_node_ptr v_node = e->dest; + + if (dump_file) + { + fprintf (dump_file, "\nProcessing edge:"); + print_ddg_edge (dump_file, e); + fprintf (dump_file, + "\nScheduling %d (%d) in psp_pss_not_empty," + " checking s %d (%d): ", u_node->cuid, INSN_UID + (u_node->insn), v_node->cuid, INSN_UID + (v_node->insn)); + } + + if (TEST_BIT (sched_nodes, v_node->cuid)) + { + int s_st = SCHED_TIME (v_node); + + late_start = MIN (late_start, + s_st - e->latency + + (e->distance * ii)); + + if (dump_file) + fprintf (dump_file, + "succ st = %d; late_start = %d; latency = %d", + s_st, late_start, e->latency); + + if (e->type == TRUE_DEP && e->data_type == REG_DEP) + count_succs++; + + if (e->data_type == MEM_DEP) + start = MAX (start, SCHED_TIME (v_node) - ii + 1); + } + else if (dump_file) + fprintf (dump_file, "the node is not scheduled\n"); + + } + start = MAX (start, early_start); + end = MIN (end, MIN (early_start + ii, late_start + 1)); + step = 1; + /* If there are more successors than predecessors schedule the + node close to it's successors. */ + if (count_succs >= count_preds) + { + int old_start = start; + + start = end - 1; + end = old_start - 1; + step = -1; + } + } + else /* psp is empty && pss is empty. */ + { + start = SCHED_ASAP (u_node); + end = start + ii; + step = 1; + } + + *start_p = start; + *step_p = step; + *end_p = end; + sbitmap_free (psp); + sbitmap_free (pss); + + if ((start >= end && step == 1) || (start <= end && step == -1)) + { + if (dump_file) + fprintf (dump_file, "\nEmpty window: start=%d, end=%d, step=%d\n", + start, end, step); + return -1; + } + + return 0; +} + +/* Calculate MUST_PRECEDE/MUST_FOLLOW bitmaps of U_NODE; which is the + node currently been scheduled. At the end of the calculation + MUST_PRECEDE/MUST_FOLLOW contains all predecessors/successors of + U_NODE which are (1) already scheduled in the first/last row of + U_NODE's scheduling window, (2) whose dependence inequality with U + becomes an equality when U is scheduled in this same row, and (3) + whose dependence latency is zero. + + The first and last rows are calculated using the following parameters: + START/END rows - The cycles that begins/ends the traversal on the window; + searching for an empty cycle to schedule U_NODE. + STEP - The direction in which we traverse the window. + II - The initiation interval. */ + +static void +calculate_must_precede_follow (ddg_node_ptr u_node, int start, int end, + int step, int ii, sbitmap sched_nodes, + sbitmap must_precede, sbitmap must_follow) +{ + ddg_edge_ptr e; + int first_cycle_in_window, last_cycle_in_window; + + gcc_assert (must_precede && must_follow); + + /* Consider the following scheduling window: + {first_cycle_in_window, first_cycle_in_window+1, ..., + last_cycle_in_window}. If step is 1 then the following will be + the order we traverse the window: {start=first_cycle_in_window, + first_cycle_in_window+1, ..., end=last_cycle_in_window+1}, + or {start=last_cycle_in_window, last_cycle_in_window-1, ..., + end=first_cycle_in_window-1} if step is -1. */ + first_cycle_in_window = (step == 1) ? start : end - step; + last_cycle_in_window = (step == 1) ? end - step : start; + + sbitmap_zero (must_precede); + sbitmap_zero (must_follow); + + if (dump_file) + fprintf (dump_file, "\nmust_precede: "); + + /* Instead of checking if: + (SMODULO (SCHED_TIME (e->src), ii) == first_row_in_window) + && ((SCHED_TIME (e->src) + e->latency - (e->distance * ii)) == + first_cycle_in_window) + && e->latency == 0 + we use the fact that latency is non-negative: + SCHED_TIME (e->src) - (e->distance * ii) <= + SCHED_TIME (e->src) + e->latency - (e->distance * ii)) <= + first_cycle_in_window + and check only if + SCHED_TIME (e->src) - (e->distance * ii) == first_cycle_in_window */ + for (e = u_node->in; e != 0; e = e->next_in) + if (TEST_BIT (sched_nodes, e->src->cuid) + && ((SCHED_TIME (e->src) - (e->distance * ii)) == + first_cycle_in_window)) + { + if (dump_file) + fprintf (dump_file, "%d ", e->src->cuid); + + SET_BIT (must_precede, e->src->cuid); + } + + if (dump_file) + fprintf (dump_file, "\nmust_follow: "); + + /* Instead of checking if: + (SMODULO (SCHED_TIME (e->dest), ii) == last_row_in_window) + && ((SCHED_TIME (e->dest) - e->latency + (e->distance * ii)) == + last_cycle_in_window) + && e->latency == 0 + we use the fact that latency is non-negative: + SCHED_TIME (e->dest) + (e->distance * ii) >= + SCHED_TIME (e->dest) - e->latency + (e->distance * ii)) >= + last_cycle_in_window + and check only if + SCHED_TIME (e->dest) + (e->distance * ii) == last_cycle_in_window */ + for (e = u_node->out; e != 0; e = e->next_out) + if (TEST_BIT (sched_nodes, e->dest->cuid) + && ((SCHED_TIME (e->dest) + (e->distance * ii)) == + last_cycle_in_window)) + { + if (dump_file) + fprintf (dump_file, "%d ", e->dest->cuid); + + SET_BIT (must_follow, e->dest->cuid); + } + + if (dump_file) + fprintf (dump_file, "\n"); +} + +/* Return 1 if U_NODE can be scheduled in CYCLE. Use the following + parameters to decide if that's possible: + PS - The partial schedule. + U - The serial number of U_NODE. + NUM_SPLITS - The number of row splits made so far. + MUST_PRECEDE - The nodes that must precede U_NODE. (only valid at + the first row of the scheduling window) + MUST_FOLLOW - The nodes that must follow U_NODE. (only valid at the + last row of the scheduling window) */ + +static bool +try_scheduling_node_in_cycle (partial_schedule_ptr ps, ddg_node_ptr u_node, + int u, int cycle, sbitmap sched_nodes, + int *num_splits, sbitmap must_precede, + sbitmap must_follow) +{ + ps_insn_ptr psi; + bool success = 0; + + verify_partial_schedule (ps, sched_nodes); + psi = ps_add_node_check_conflicts (ps, u_node, cycle, + must_precede, must_follow); + if (psi) + { + SCHED_TIME (u_node) = cycle; + SET_BIT (sched_nodes, u); + success = 1; + *num_splits = 0; + if (dump_file) + fprintf (dump_file, "Scheduled w/o split in %d\n", cycle); + + } + + return success; +} + +/* This function implements the scheduling algorithm for SMS according to the + above algorithm. */ static partial_schedule_ptr -sms_schedule_by_order (ddg_ptr g, int mii, int maxii, int *nodes_order, FILE *dump_file) +sms_schedule_by_order (ddg_ptr g, int mii, int maxii, int *nodes_order) { int ii = mii; - int i, c, success; - int try_again_with_larger_ii = true; + int i, c, success, num_splits = 0; + int flush_and_start_over = true; int num_nodes = g->num_nodes; - ddg_edge_ptr e; int start, end, step; /* Place together into one struct? */ sbitmap sched_nodes = sbitmap_alloc (num_nodes); - sbitmap psp = sbitmap_alloc (num_nodes); - sbitmap pss = sbitmap_alloc (num_nodes); sbitmap must_precede = sbitmap_alloc (num_nodes); sbitmap must_follow = sbitmap_alloc (num_nodes); + sbitmap tobe_scheduled = sbitmap_alloc (num_nodes); partial_schedule_ptr ps = create_partial_schedule (ii, g, DFA_HISTORY); - while (try_again_with_larger_ii && ii < maxii) + sbitmap_ones (tobe_scheduled); + sbitmap_zero (sched_nodes); + + while (flush_and_start_over && (ii < maxii)) { + if (dump_file) - fprintf(dump_file, "Starting with ii=%d\n", ii); - try_again_with_larger_ii = false; + fprintf (dump_file, "Starting with ii=%d\n", ii); + flush_and_start_over = false; sbitmap_zero (sched_nodes); for (i = 0; i < num_nodes; i++) { int u = nodes_order[i]; - ddg_node_ptr u_node = &g->nodes[u]; - sbitmap u_node_preds = NODE_PREDECESSORS (u_node); - sbitmap u_node_succs = NODE_SUCCESSORS (u_node); - int psp_not_empty; - int pss_not_empty; + ddg_node_ptr u_node = &ps->g->nodes[u]; rtx insn = u_node->insn; - if (!INSN_P (insn)) - continue; + if (!NONDEBUG_INSN_P (insn)) + { + RESET_BIT (tobe_scheduled, u); + continue; + } if (JUMP_P (insn)) /* Closing branch handled later. */ + { + RESET_BIT (tobe_scheduled, u); + continue; + } + + if (TEST_BIT (sched_nodes, u)) continue; - /* 1. compute sched window for u (start, end, step). */ - sbitmap_zero (psp); - sbitmap_zero (pss); - psp_not_empty = sbitmap_a_and_b_cg (psp, u_node_preds, sched_nodes); - pss_not_empty = sbitmap_a_and_b_cg (pss, u_node_succs, sched_nodes); + /* Try to get non-empty scheduling window. */ + success = 0; + if (get_sched_window (ps, nodes_order, i, sched_nodes, ii, &start, + &step, &end) == 0) + { + if (dump_file) + fprintf (dump_file, "\nTrying to schedule node %d \ + INSN = %d in (%d .. %d) step %d\n", u, (INSN_UID + (g->nodes[u].insn)), start, end, step); + + gcc_assert ((step > 0 && start < end) + || (step < 0 && start > end)); + + calculate_must_precede_follow (u_node, start, end, step, ii, + sched_nodes, must_precede, + must_follow); + + for (c = start; c != end; c += step) + { + sbitmap tmp_precede = NULL; + sbitmap tmp_follow = NULL; + + if (c == start) + { + if (step == 1) + tmp_precede = must_precede; + else /* step == -1. */ + tmp_follow = must_follow; + } + if (c == end - step) + { + if (step == 1) + tmp_follow = must_follow; + else /* step == -1. */ + tmp_precede = must_precede; + } + + success = + try_scheduling_node_in_cycle (ps, u_node, u, c, + sched_nodes, + &num_splits, tmp_precede, + tmp_follow); + if (success) + break; + } + + verify_partial_schedule (ps, sched_nodes); + } + if (!success) + { + int split_row; + + if (ii++ == maxii) + break; + + if (num_splits >= MAX_SPLIT_NUM) + { + num_splits = 0; + flush_and_start_over = true; + verify_partial_schedule (ps, sched_nodes); + reset_partial_schedule (ps, ii); + verify_partial_schedule (ps, sched_nodes); + break; + } + + num_splits++; + /* The scheduling window is exclusive of 'end' + whereas compute_split_window() expects an inclusive, + ordered range. */ + if (step == 1) + split_row = compute_split_row (sched_nodes, start, end - 1, + ps->ii, u_node); + else + split_row = compute_split_row (sched_nodes, end + 1, start, + ps->ii, u_node); + + ps_insert_empty_row (ps, split_row, sched_nodes); + i--; /* Go back and retry node i. */ + + if (dump_file) + fprintf (dump_file, "num_splits=%d\n", num_splits); + } - if (psp_not_empty && !pss_not_empty) - { - int early_start = 0; + /* ??? If (success), check register pressure estimates. */ + } /* Continue with next node. */ + } /* While flush_and_start_over. */ + if (ii >= maxii) + { + free_partial_schedule (ps); + ps = NULL; + } + else + gcc_assert (sbitmap_equal (tobe_scheduled, sched_nodes)); - end = INT_MAX; - for (e = u_node->in; e != 0; e = e->next_in) - { - ddg_node_ptr v_node = e->src; - if (TEST_BIT (sched_nodes, v_node->cuid)) - { - int node_st = SCHED_TIME (v_node) - + e->latency - (e->distance * ii); + sbitmap_free (sched_nodes); + sbitmap_free (must_precede); + sbitmap_free (must_follow); + sbitmap_free (tobe_scheduled); - early_start = MAX (early_start, node_st); + return ps; +} - if (e->data_type == MEM_DEP) - end = MIN (end, SCHED_TIME (v_node) + ii - 1); - } - } - start = early_start; - end = MIN (end, early_start + ii); - step = 1; - } +/* This function inserts a new empty row into PS at the position + according to SPLITROW, keeping all already scheduled instructions + intact and updating their SCHED_TIME and cycle accordingly. */ +static void +ps_insert_empty_row (partial_schedule_ptr ps, int split_row, + sbitmap sched_nodes) +{ + ps_insn_ptr crr_insn; + ps_insn_ptr *rows_new; + int ii = ps->ii; + int new_ii = ii + 1; + int row; - else if (!psp_not_empty && pss_not_empty) - { - int late_start = INT_MAX; + verify_partial_schedule (ps, sched_nodes); - end = INT_MIN; - for (e = u_node->out; e != 0; e = e->next_out) - { - ddg_node_ptr v_node = e->dest; - if (TEST_BIT (sched_nodes, v_node->cuid)) - { - late_start = MIN (late_start, - SCHED_TIME (v_node) - e->latency - + (e->distance * ii)); - if (e->data_type == MEM_DEP) - end = MAX (end, SCHED_TIME (v_node) - ii + 1); - } - } - start = late_start; - end = MAX (end, late_start - ii); - step = -1; - } + /* We normalize sched_time and rotate ps to have only non-negative sched + times, for simplicity of updating cycles after inserting new row. */ + split_row -= ps->min_cycle; + split_row = SMODULO (split_row, ii); + if (dump_file) + fprintf (dump_file, "split_row=%d\n", split_row); - else if (psp_not_empty && pss_not_empty) - { - int early_start = 0; - int late_start = INT_MAX; + normalize_sched_times (ps); + rotate_partial_schedule (ps, ps->min_cycle); - start = INT_MIN; - end = INT_MAX; - for (e = u_node->in; e != 0; e = e->next_in) - { - ddg_node_ptr v_node = e->src; - - if (TEST_BIT (sched_nodes, v_node->cuid)) - { - early_start = MAX (early_start, - SCHED_TIME (v_node) + e->latency - - (e->distance * ii)); - if (e->data_type == MEM_DEP) - end = MIN (end, SCHED_TIME (v_node) + ii - 1); - } - } - for (e = u_node->out; e != 0; e = e->next_out) - { - ddg_node_ptr v_node = e->dest; - - if (TEST_BIT (sched_nodes, v_node->cuid)) - { - late_start = MIN (late_start, - SCHED_TIME (v_node) - e->latency - + (e->distance * ii)); - if (e->data_type == MEM_DEP) - start = MAX (start, SCHED_TIME (v_node) - ii + 1); - } - } - start = MAX (start, early_start); - end = MIN (end, MIN (early_start + ii, late_start + 1)); - step = 1; - } - else /* psp is empty && pss is empty. */ - { - start = SCHED_ASAP (u_node); - end = start + ii; - step = 1; - } + rows_new = (ps_insn_ptr *) xcalloc (new_ii, sizeof (ps_insn_ptr)); + for (row = 0; row < split_row; row++) + { + rows_new[row] = ps->rows[row]; + ps->rows[row] = NULL; + for (crr_insn = rows_new[row]; + crr_insn; crr_insn = crr_insn->next_in_row) + { + ddg_node_ptr u = crr_insn->node; + int new_time = SCHED_TIME (u) + (SCHED_TIME (u) / ii); - /* 2. Try scheduling u in window. */ - if (dump_file) - fprintf(dump_file, "Trying to schedule node %d in (%d .. %d) step %d\n", - u, start, end, step); - - /* use must_follow & must_precede bitmaps to determine order - of nodes within the cycle. */ - sbitmap_zero (must_precede); - sbitmap_zero (must_follow); - for (e = u_node->in; e != 0; e = e->next_in) - if (TEST_BIT (sched_nodes, e->src->cuid) - && e->latency == (ii * e->distance) - && start == SCHED_TIME (e->src)) - SET_BIT (must_precede, e->src->cuid); - - for (e = u_node->out; e != 0; e = e->next_out) - if (TEST_BIT (sched_nodes, e->dest->cuid) - && e->latency == (ii * e->distance) - && end == SCHED_TIME (e->dest)) - SET_BIT (must_follow, e->dest->cuid); - - success = 0; - if ((step > 0 && start < end) || (step < 0 && start > end)) - for (c = start; c != end; c += step) - { - ps_insn_ptr psi; - - psi = ps_add_node_check_conflicts (ps, u_node, c, - must_precede, - must_follow); - - if (psi) - { - SCHED_TIME (u_node) = c; - SET_BIT (sched_nodes, u); - success = 1; - if (dump_file) - fprintf(dump_file, "Schedule in %d\n", c); - break; - } - } - if (!success) - { - /* ??? Try backtracking instead of immediately ii++? */ - ii++; - try_again_with_larger_ii = true; - reset_partial_schedule (ps, ii); - break; - } - /* ??? If (success), check register pressure estimates. */ - } /* Continue with next node. */ - } /* While try_again_with_larger_ii. */ + SCHED_TIME (u) = new_time; + crr_insn->cycle = new_time; + SCHED_ROW (u) = new_time % new_ii; + SCHED_STAGE (u) = new_time / new_ii; + } - sbitmap_free (sched_nodes); - sbitmap_free (psp); - sbitmap_free (pss); + } - if (ii >= maxii) + rows_new[split_row] = NULL; + + for (row = split_row; row < ii; row++) { - free_partial_schedule (ps); - ps = NULL; + rows_new[row + 1] = ps->rows[row]; + ps->rows[row] = NULL; + for (crr_insn = rows_new[row + 1]; + crr_insn; crr_insn = crr_insn->next_in_row) + { + ddg_node_ptr u = crr_insn->node; + int new_time = SCHED_TIME (u) + (SCHED_TIME (u) / ii) + 1; + + SCHED_TIME (u) = new_time; + crr_insn->cycle = new_time; + SCHED_ROW (u) = new_time % new_ii; + SCHED_STAGE (u) = new_time / new_ii; + } } - return ps; + + /* Updating ps. */ + ps->min_cycle = ps->min_cycle + ps->min_cycle / ii + + (SMODULO (ps->min_cycle, ii) >= split_row ? 1 : 0); + ps->max_cycle = ps->max_cycle + ps->max_cycle / ii + + (SMODULO (ps->max_cycle, ii) >= split_row ? 1 : 0); + free (ps->rows); + ps->rows = rows_new; + ps->ii = new_ii; + gcc_assert (ps->min_cycle >= 0); + + verify_partial_schedule (ps, sched_nodes); + + if (dump_file) + fprintf (dump_file, "min_cycle=%d, max_cycle=%d\n", ps->min_cycle, + ps->max_cycle); +} + +/* Given U_NODE which is the node that failed to be scheduled; LOW and + UP which are the boundaries of it's scheduling window; compute using + SCHED_NODES and II a row in the partial schedule that can be split + which will separate a critical predecessor from a critical successor + thereby expanding the window, and return it. */ +static int +compute_split_row (sbitmap sched_nodes, int low, int up, int ii, + ddg_node_ptr u_node) +{ + ddg_edge_ptr e; + int lower = INT_MIN, upper = INT_MAX; + ddg_node_ptr crit_pred = NULL; + ddg_node_ptr crit_succ = NULL; + int crit_cycle; + + for (e = u_node->in; e != 0; e = e->next_in) + { + ddg_node_ptr v_node = e->src; + + if (TEST_BIT (sched_nodes, v_node->cuid) + && (low == SCHED_TIME (v_node) + e->latency - (e->distance * ii))) + if (SCHED_TIME (v_node) > lower) + { + crit_pred = v_node; + lower = SCHED_TIME (v_node); + } + } + + if (crit_pred != NULL) + { + crit_cycle = SCHED_TIME (crit_pred) + 1; + return SMODULO (crit_cycle, ii); + } + + for (e = u_node->out; e != 0; e = e->next_out) + { + ddg_node_ptr v_node = e->dest; + if (TEST_BIT (sched_nodes, v_node->cuid) + && (up == SCHED_TIME (v_node) - e->latency + (e->distance * ii))) + if (SCHED_TIME (v_node) < upper) + { + crit_succ = v_node; + upper = SCHED_TIME (v_node); + } + } + + if (crit_succ != NULL) + { + crit_cycle = SCHED_TIME (crit_succ); + return SMODULO (crit_cycle, ii); + } + + if (dump_file) + fprintf (dump_file, "Both crit_pred and crit_succ are NULL\n"); + + return SMODULO ((low + up + 1) / 2, ii); +} + +static void +verify_partial_schedule (partial_schedule_ptr ps, sbitmap sched_nodes) +{ + int row; + ps_insn_ptr crr_insn; + + for (row = 0; row < ps->ii; row++) + for (crr_insn = ps->rows[row]; crr_insn; crr_insn = crr_insn->next_in_row) + { + ddg_node_ptr u = crr_insn->node; + + gcc_assert (TEST_BIT (sched_nodes, u->cuid)); + /* ??? Test also that all nodes of sched_nodes are in ps, perhaps by + popcount (sched_nodes) == number of insns in ps. */ + gcc_assert (SCHED_TIME (u) >= ps->min_cycle); + gcc_assert (SCHED_TIME (u) <= ps->max_cycle); + } } @@ -1431,7 +2044,7 @@ typedef struct node_order_params * nopa; static void order_nodes_of_sccs (ddg_all_sccs_ptr, int * result); static int order_nodes_in_scc (ddg_ptr, sbitmap, sbitmap, int*, int); -static nopa calculate_order_params (ddg_ptr, int mii); +static nopa calculate_order_params (ddg_ptr, int, int *); static int find_max_asap (ddg_ptr, sbitmap); static int find_max_hv_min_mob (ddg_ptr, sbitmap); static int find_max_dv_min_mob (ddg_ptr, sbitmap); @@ -1454,30 +2067,40 @@ check_nodes_order (int *node_order, int num_nodes) sbitmap_zero (tmp); + if (dump_file) + fprintf (dump_file, "SMS final nodes order: \n"); + for (i = 0; i < num_nodes; i++) { int u = node_order[i]; - if (u >= num_nodes || u < 0 || TEST_BIT (tmp, u)) - abort (); + if (dump_file) + fprintf (dump_file, "%d ", u); + gcc_assert (u < num_nodes && u >= 0 && !TEST_BIT (tmp, u)); SET_BIT (tmp, u); } + if (dump_file) + fprintf (dump_file, "\n"); + sbitmap_free (tmp); } /* Order the nodes of G for scheduling and pass the result in NODE_ORDER. Also set aux.count of each node to ASAP. - Return the recMII for the given DDG. */ + Put maximal ASAP to PMAX_ASAP. Return the recMII for the given DDG. */ static int -sms_order_nodes (ddg_ptr g, int mii, int * node_order) +sms_order_nodes (ddg_ptr g, int mii, int * node_order, int *pmax_asap) { int i; int rec_mii = 0; ddg_all_sccs_ptr sccs = create_ddg_all_sccs (g); - nopa nops = calculate_order_params (g, mii); + nopa nops = calculate_order_params (g, mii, pmax_asap); + + if (dump_file) + print_sccs (dump_file, sccs, g); order_nodes_of_sccs (sccs, node_order); @@ -1513,7 +2136,7 @@ order_nodes_of_sccs (ddg_all_sccs_ptr all_sccs, int * node_order) sbitmap_zero (prev_sccs); sbitmap_ones (ones); - /* Perfrom the node ordering starting from the SCC with the highest recMII. + /* Perform the node ordering starting from the SCC with the highest recMII. For each SCC order the nodes according to their ASAP/ALAP/HEIGHT etc. */ for (i = 0; i < all_sccs->num_sccs; i++) { @@ -1549,7 +2172,7 @@ order_nodes_of_sccs (ddg_all_sccs_ptr all_sccs, int * node_order) /* MII is needed if we consider backarcs (that do not close recursive cycles). */ static struct node_order_params * -calculate_order_params (ddg_ptr g, int mii ATTRIBUTE_UNUSED) +calculate_order_params (ddg_ptr g, int mii ATTRIBUTE_UNUSED, int *pmax_asap) { int u; int max_asap; @@ -1600,18 +2223,31 @@ calculate_order_params (ddg_ptr g, int mii ATTRIBUTE_UNUSED) HEIGHT (e->dest) + e->latency); } } + if (dump_file) + { + fprintf (dump_file, "\nOrder params\n"); + for (u = 0; u < num_nodes; u++) + { + ddg_node_ptr u_node = &g->nodes[u]; + fprintf (dump_file, "node %d, ASAP: %d, ALAP: %d, HEIGHT: %d\n", u, + ASAP (u_node), ALAP (u_node), HEIGHT (u_node)); + } + } + + *pmax_asap = max_asap; return node_order_params_arr; } static int find_max_asap (ddg_ptr g, sbitmap nodes) { - int u; + unsigned int u = 0; int max_asap = -1; int result = -1; + sbitmap_iterator sbi; - EXECUTE_IF_SET_IN_SBITMAP (nodes, 0, u, + EXECUTE_IF_SET_IN_SBITMAP (nodes, 0, u, sbi) { ddg_node_ptr u_node = &g->nodes[u]; @@ -1620,19 +2256,20 @@ find_max_asap (ddg_ptr g, sbitmap nodes) max_asap = ASAP (u_node); result = u; } - }); + } return result; } static int find_max_hv_min_mob (ddg_ptr g, sbitmap nodes) { - int u; + unsigned int u = 0; int max_hv = -1; int min_mob = INT_MAX; int result = -1; + sbitmap_iterator sbi; - EXECUTE_IF_SET_IN_SBITMAP (nodes, 0, u, + EXECUTE_IF_SET_IN_SBITMAP (nodes, 0, u, sbi) { ddg_node_ptr u_node = &g->nodes[u]; @@ -1648,19 +2285,20 @@ find_max_hv_min_mob (ddg_ptr g, sbitmap nodes) min_mob = MOB (u_node); result = u; } - }); + } return result; } static int find_max_dv_min_mob (ddg_ptr g, sbitmap nodes) { - int u; + unsigned int u = 0; int max_dv = -1; int min_mob = INT_MAX; int result = -1; + sbitmap_iterator sbi; - EXECUTE_IF_SET_IN_SBITMAP (nodes, 0, u, + EXECUTE_IF_SET_IN_SBITMAP (nodes, 0, u, sbi) { ddg_node_ptr u_node = &g->nodes[u]; @@ -1676,7 +2314,7 @@ find_max_dv_min_mob (ddg_ptr g, sbitmap nodes) min_mob = MOB (u_node); result = u; } - }); + } return result; } @@ -1787,11 +2425,11 @@ order_nodes_in_scc (ddg_ptr g, sbitmap nodes_ordered, sbitmap scc, modulo scheduling. */ /* Create a partial schedule and allocate a memory to hold II rows. */ -partial_schedule_ptr + +static partial_schedule_ptr create_partial_schedule (int ii, ddg_ptr g, int history) { - partial_schedule_ptr ps = (partial_schedule_ptr) - xmalloc (sizeof (struct partial_schedule)); + partial_schedule_ptr ps = XNEW (struct partial_schedule); ps->rows = (ps_insn_ptr *) xcalloc (ii, sizeof (ps_insn_ptr)); ps->ii = ii; ps->history = history; @@ -1823,7 +2461,8 @@ free_ps_insns (partial_schedule_ptr ps) } /* Free all the memory allocated to the partial schedule. */ -void + +static void free_partial_schedule (partial_schedule_ptr ps) { if (!ps) @@ -1835,7 +2474,8 @@ free_partial_schedule (partial_schedule_ptr ps) /* Clear the rows array with its PS_INSNs, and create a new one with NEW_II rows. */ -void + +static void reset_partial_schedule (partial_schedule_ptr ps, int new_ii) { if (!ps) @@ -1862,7 +2502,7 @@ print_partial_schedule (partial_schedule_ptr ps, FILE *dump) { ps_insn_ptr ps_i = ps->rows[i]; - fprintf (dump, "\n[CYCLE %d ]: ", i); + fprintf (dump, "\n[ROW %d ]: ", i); while (ps_i) { fprintf (dump, "%d, ", @@ -1876,7 +2516,7 @@ print_partial_schedule (partial_schedule_ptr ps, FILE *dump) static ps_insn_ptr create_ps_insn (ddg_node_ptr node, int rest_count, int cycle) { - ps_insn_ptr ps_i = xmalloc (sizeof (struct ps_insn)); + ps_insn_ptr ps_i = XNEW (struct ps_insn); ps_i->node = node; ps_i->next_in_row = NULL; @@ -1890,7 +2530,7 @@ create_ps_insn (ddg_node_ptr node, int rest_count, int cycle) /* Removes the given PS_INSN from the partial schedule. Returns false if the node is not found in the partial schedule, else returns true. */ -static int +static bool remove_node_from_ps (partial_schedule_ptr ps, ps_insn_ptr ps_i) { int row; @@ -1939,16 +2579,16 @@ ps_insn_find_column (partial_schedule_ptr ps, ps_insn_ptr ps_i, row = SMODULO (ps_i->cycle, ps->ii); /* Find the first must follow and the last must precede - and insert the node immediatly after the must precede - but make sure that it there is no must follow after it. */ + and insert the node immediately after the must precede + but make sure that it there is no must follow after it. */ for (next_ps_i = ps->rows[row]; next_ps_i; next_ps_i = next_ps_i->next_in_row) { - if (TEST_BIT (must_follow, next_ps_i->node->cuid) + if (must_follow && TEST_BIT (must_follow, next_ps_i->node->cuid) && ! first_must_follow) first_must_follow = next_ps_i; - if (TEST_BIT (must_precede, next_ps_i->node->cuid)) + if (must_precede && TEST_BIT (must_precede, next_ps_i->node->cuid)) { /* If we have already met a node that must follow, then there is no possible column. */ @@ -1982,8 +2622,8 @@ ps_insn_find_column (partial_schedule_ptr ps, ps_insn_ptr ps_i, } /* Advances the PS_INSN one column in its current row; returns false - in failure and true in success. Bit N is set in MUST_FOLLOW if - the node with cuid N must be come after the node pointed to by + in failure and true in success. Bit N is set in MUST_FOLLOW if + the node with cuid N must be come after the node pointed to by PS_I when scheduled in the same cycle. */ static int ps_insn_advance_column (partial_schedule_ptr ps, ps_insn_ptr ps_i, @@ -2005,10 +2645,10 @@ ps_insn_advance_column (partial_schedule_ptr ps, ps_insn_ptr ps_i, /* Check if next_in_row is dependent on ps_i, both having same sched times (typically ANTI_DEP). If so, ps_i cannot skip over it. */ - if (TEST_BIT (must_follow, next_node->cuid)) + if (must_follow && TEST_BIT (must_follow, next_node->cuid)) return false; - /* Advace PS_I over its next_in_row in the doubly linked list. */ + /* Advance PS_I over its next_in_row in the doubly linked list. */ prev = ps_i->prev_in_row; next = ps_i->next_in_row; @@ -2031,9 +2671,9 @@ ps_insn_advance_column (partial_schedule_ptr ps, ps_insn_ptr ps_i, } /* Inserts a DDG_NODE to the given partial schedule at the given cycle. - Returns 0 if this is not possible and a PS_INSN otherwise. Bit N is - set in MUST_PRECEDE/MUST_FOLLOW if the node with cuid N must be come - before/after (respectively) the node pointed to by PS_I when scheduled + Returns 0 if this is not possible and a PS_INSN otherwise. Bit N is + set in MUST_PRECEDE/MUST_FOLLOW if the node with cuid N must be come + before/after (respectively) the node pointed to by PS_I when scheduled in the same cycle. */ static ps_insn_ptr add_node_to_ps (partial_schedule_ptr ps, ddg_node_ptr node, int cycle, @@ -2069,15 +2709,17 @@ advance_one_cycle (void) { if (targetm.sched.dfa_pre_cycle_insn) state_transition (curr_state, - (*targetm.sched.dfa_pre_cycle_insn) ()); + targetm.sched.dfa_pre_cycle_insn ()); state_transition (curr_state, NULL); if (targetm.sched.dfa_post_cycle_insn) state_transition (curr_state, - (*targetm.sched.dfa_post_cycle_insn) ()); + targetm.sched.dfa_post_cycle_insn ()); } + + /* Checks if PS has resource conflicts according to DFA, starting from FROM cycle to TO cycle; returns true if there are conflicts and false if there are no conflicts. Assumes DFA is being used. */ @@ -2101,7 +2743,7 @@ ps_has_conflicts (partial_schedule_ptr ps, int from, int to) { rtx insn = crr_insn->node->insn; - if (!INSN_P (insn)) + if (!NONDEBUG_INSN_P (insn)) continue; /* Check if there is room for the current insn. */ @@ -2109,14 +2751,14 @@ ps_has_conflicts (partial_schedule_ptr ps, int from, int to) return true; /* Update the DFA state and return with failure if the DFA found - recource conflicts. */ + resource conflicts. */ if (state_transition (curr_state, insn) >= 0) return true; if (targetm.sched.variable_issue) can_issue_more = - (*targetm.sched.variable_issue) (sched_dump, sched_verbose, - insn, can_issue_more); + targetm.sched.variable_issue (sched_dump, sched_verbose, + insn, can_issue_more); /* A naked CLOBBER or USE generates no instruction, so don't let them consume issue slots. */ else if (GET_CODE (PATTERN (insn)) != USE @@ -2132,8 +2774,8 @@ ps_has_conflicts (partial_schedule_ptr ps, int from, int to) /* Checks if the given node causes resource conflicts when added to PS at cycle C. If not the node is added to PS and returned; otherwise zero - is returned. Bit N is set in MUST_PRECEDE/MUST_FOLLOW if the node with - cuid N must be come before/after (respectively) the node pointed to by + is returned. Bit N is set in MUST_PRECEDE/MUST_FOLLOW if the node with + cuid N must be come before/after (respectively) the node pointed to by PS_I when scheduled in the same cycle. */ ps_insn_ptr ps_add_node_check_conflicts (partial_schedule_ptr ps, ddg_node_ptr n, @@ -2206,4 +2848,58 @@ rotate_partial_schedule (partial_schedule_ptr ps, int start_cycle) ps->min_cycle -= start_cycle; } -#endif /* INSN_SCHEDULING*/ +#endif /* INSN_SCHEDULING */ + +static bool +gate_handle_sms (void) +{ + return (optimize > 0 && flag_modulo_sched); +} + + +/* Run instruction scheduler. */ +/* Perform SMS module scheduling. */ +static unsigned int +rest_of_handle_sms (void) +{ +#ifdef INSN_SCHEDULING + basic_block bb; + + /* Collect loop information to be used in SMS. */ + cfg_layout_initialize (0); + sms_schedule (); + + /* Update the life information, because we add pseudos. */ + max_regno = max_reg_num (); + + /* Finalize layout changes. */ + FOR_EACH_BB (bb) + if (bb->next_bb != EXIT_BLOCK_PTR) + bb->aux = bb->next_bb; + free_dominance_info (CDI_DOMINATORS); + cfg_layout_finalize (); +#endif /* INSN_SCHEDULING */ + return 0; +} + +struct rtl_opt_pass pass_sms = +{ + { + RTL_PASS, + "sms", /* name */ + gate_handle_sms, /* gate */ + rest_of_handle_sms, /* execute */ + NULL, /* sub */ + NULL, /* next */ + 0, /* static_pass_number */ + TV_SMS, /* tv_id */ + 0, /* properties_required */ + 0, /* properties_provided */ + 0, /* properties_destroyed */ + TODO_dump_func, /* todo_flags_start */ + TODO_df_finish | TODO_verify_rtl_sharing | + TODO_dump_func | + TODO_ggc_collect /* todo_flags_finish */ + } +}; +