diff options
Diffstat (limited to 'gcc/modulo-sched.c')
| -rw-r--r-- | gcc/modulo-sched.c | 38 |
1 files changed, 28 insertions, 10 deletions
diff --git a/gcc/modulo-sched.c b/gcc/modulo-sched.c index 57879baf799..e623ea1cb53 100644 --- a/gcc/modulo-sched.c +++ b/gcc/modulo-sched.c @@ -339,6 +339,10 @@ const_iteration_count (rtx count_reg, basic_block pre_header, { rtx insn; rtx head, tail; + + if (! pre_header) + return NULL_RTX; + get_block_head_tail (pre_header->index, &head, &tail); for (insn = tail; insn != PREV_INSN (head); insn = PREV_INSN (insn)) @@ -401,6 +405,8 @@ print_node_sched_params (FILE * dump_file, int num_nodes) { int i; + if (! dump_file) + return; for (i = 0; i < num_nodes; i++) { node_sched_params_ptr nsp = &node_sched_params[i]; @@ -443,14 +449,17 @@ calculate_maxii (ddg_ptr g) 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. */ +/* + 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 - { dependecnce. + ii { 1 if not. +*/ static void generate_reg_moves (partial_schedule_ptr ps) { @@ -474,6 +483,9 @@ 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) @@ -497,6 +509,9 @@ 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--; @@ -540,7 +555,8 @@ normalize_sched_times (partial_schedule_ptr ps) int amount = PS_MIN_CYCLE (ps); int ii = ps->ii; - for (i = 0; i < g->num_nodes; i++) + /* Don't include the closing branch assuming that it is the last node. */ + for (i = 0; i < g->num_nodes - 1; i++) { ddg_node_ptr u = &g->nodes[i]; int normalized_time = SCHED_TIME (u) - amount; @@ -609,7 +625,7 @@ duplicate_insns_of_cycles (partial_schedule_ptr ps, int from_stage, /* 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)); @@ -1109,6 +1125,8 @@ sms_schedule (FILE *dump_file) scheduling passes doesn't touch it. */ if (! flag_resched_modulo_sched) g->bb->flags |= BB_DISABLE_SCHEDULE; + /* The life-info is not valid any more. */ + g->bb->flags |= BB_DIRTY; generate_reg_moves (ps); if (dump_file) |