bool cfg_layout_can_duplicate_bb_p (basic_block bb) { /* Do not attempt to duplicate tablejumps, as we need to unshare the dispatch table. This is difficult to do, as the instructions computing jump destination may be hoisted outside the basic block. */ if (tablejump_p (BB_END (bb), NULL, NULL)) return false; /* Do not duplicate blocks containing insns that can't be copied. */ if (targetm.cannot_copy_insn_p) { rtx insn = BB_HEAD (bb); while (1) { if (INSN_P (insn) && targetm.cannot_copy_insn_p (insn)) return false; if (insn == BB_END (bb)) break; insn = NEXT_INSN (insn); } } return true; }
/* INSN is being scheduled after LAST. Update counters. */ static void begin_schedule_ready (rtx insn, rtx last) { sched_rgn_n_insns++; if (BLOCK_FOR_INSN (insn) == last_bb /* INSN is a jump in the last block, ... */ && control_flow_insn_p (insn) /* that is going to be moved over some instructions. */ && last != PREV_INSN (insn)) { edge e; basic_block bb; /* An obscure special case, where we do have partially dead instruction scheduled after last control flow instruction. In this case we can create new basic block. It is always exactly one basic block last in the sequence. */ e = find_fallthru_edge (last_bb->succs); gcc_checking_assert (!e || !(e->flags & EDGE_COMPLEX)); gcc_checking_assert (BLOCK_FOR_INSN (insn) == last_bb && !IS_SPECULATION_CHECK_P (insn) && BB_HEAD (last_bb) != insn && BB_END (last_bb) == insn); { rtx x; x = NEXT_INSN (insn); if (e) gcc_checking_assert (NOTE_P (x) || LABEL_P (x)); else gcc_checking_assert (BARRIER_P (x)); } if (e) { bb = split_edge (e); gcc_assert (NOTE_INSN_BASIC_BLOCK_P (BB_END (bb))); } else /* Create an empty unreachable block after the INSN. */ bb = create_basic_block (NEXT_INSN (insn), NULL_RTX, last_bb); /* split_edge () creates BB before E->DEST. Keep in mind, that this operation extends scheduling region till the end of BB. Hence, we need to shift NEXT_TAIL, so haifa-sched.c won't go out of the scheduling region. */ current_sched_info->next_tail = NEXT_INSN (BB_END (bb)); gcc_assert (current_sched_info->next_tail); /* Append new basic block to the end of the ebb. */ sched_init_only_bb (bb, last_bb); gcc_assert (last_bb == bb); } }
static void erase_matching_seqs (void) { seq_block sb; matching_seq mseq; rtx insn; basic_block bb; rtx retlabel, saveinsn, callinsn; int i; for (sb = seq_blocks; sb; sb = sb->next_seq_block) { for (mseq = sb->matching_seqs; mseq; mseq = mseq->next_matching_seq) { insn = mseq->insn; bb = BLOCK_FOR_INSN (insn); /* Get the label after the sequence. This will be the return address. The label will be referenced using a symbol_ref so protect it from deleting. */ retlabel = block_label_after (insn); LABEL_PRESERVE_P (retlabel) = 1; /* Delete the insns of the sequence. */ for (i = 0; i < sb->length; i++) insn = prev_insn_in_block (insn); delete_basic_block (split_block_and_df_analyze (bb, insn)); /* Emit an insn saving the return address to the link register before the deleted sequence. */ saveinsn = emit_insn_after (gen_move_insn (pattern_seqs->link_reg, gen_symbol_ref_rtx_for_label (retlabel)), BB_END (bb)); BLOCK_FOR_INSN (saveinsn) = bb; /* Emit a jump to the appropriate part of the pattern sequence after the save insn. Also update the basic block. */ callinsn = emit_jump_insn_after (gen_jump (sb->label), saveinsn); JUMP_LABEL (callinsn) = sb->label; LABEL_NUSES (sb->label)++; BLOCK_FOR_INSN (callinsn) = bb; BB_END (bb) = callinsn; /* Maintain control flow and liveness information. */ SET_REGNO_REG_SET (df_get_live_out (bb), REGNO (pattern_seqs->link_reg)); emit_barrier_after (BB_END (bb)); make_single_succ_edge (bb, BLOCK_FOR_INSN (sb->label), 0); IOR_REG_SET (df_get_live_out (bb), df_get_live_in (BLOCK_FOR_INSN (sb->label))); make_edge (BLOCK_FOR_INSN (seq_blocks->label), BLOCK_FOR_INSN (retlabel), EDGE_ABNORMAL); } } }
static void split_pattern_seq (void) { rtx insn; basic_block bb; rtx retlabel, retjmp, saveinsn; int i; seq_block sb; insn = pattern_seqs->insn; bb = BLOCK_FOR_INSN (insn); /* Get the label after the sequence. This will be the return address. The label will be referenced using a symbol_ref so protect it from deleting. */ retlabel = block_label_after (insn); LABEL_PRESERVE_P (retlabel) = 1; /* Emit an indirect jump via the link register after the sequence acting as the return insn. Also emit a barrier and update the basic block. */ if (!find_reg_note (BB_END (bb), REG_NORETURN, NULL)) retjmp = emit_jump_insn_after (gen_indirect_jump (pattern_seqs->link_reg), BB_END (bb)); emit_barrier_after (BB_END (bb)); /* Replace all outgoing edges with a new one to the block of RETLABEL. */ while (EDGE_COUNT (bb->succs) != 0) remove_edge (EDGE_SUCC (bb, 0)); make_edge (bb, BLOCK_FOR_INSN (retlabel), EDGE_ABNORMAL); /* Split the sequence according to SEQ_BLOCKS and cache the label of the resulting basic blocks. */ i = 0; for (sb = seq_blocks; sb; sb = sb->next_seq_block) { for (; i < sb->length; i++) insn = prev_insn_in_block (insn); sb->label = block_label (split_block_and_df_analyze (bb, insn)); } /* Emit an insn saving the return address to the link register before the sequence. */ saveinsn = emit_insn_after (gen_move_insn (pattern_seqs->link_reg, gen_symbol_ref_rtx_for_label (retlabel)), BB_END (bb)); /* Update liveness info. */ SET_REGNO_REG_SET (df_get_live_out (bb), REGNO (pattern_seqs->link_reg)); }
/* Provide a slim dump the instruction chain starting at FIRST to F, honoring the dump flags given in FLAGS. Currently, TDF_BLOCKS and TDF_DETAILS include more information on the basic blocks. */ void print_rtl_slim_with_bb (FILE *f, rtx first, int flags) { basic_block current_bb = NULL; rtx insn; for (insn = first; NULL != insn; insn = NEXT_INSN (insn)) { if ((flags & TDF_BLOCKS) && (INSN_P (insn) || GET_CODE (insn) == NOTE) && BLOCK_FOR_INSN (insn) && !current_bb) { current_bb = BLOCK_FOR_INSN (insn); dump_bb_info (current_bb, true, false, flags, ";; ", f); } dump_insn_slim (f, insn); if ((flags & TDF_BLOCKS) && current_bb && insn == BB_END (current_bb)) { dump_bb_info (current_bb, false, true, flags, ";; ", f); current_bb = NULL; } } }
/* Same as above, but stop at LAST or when COUNT == 0. If COUNT < 0 it will stop only at LAST or NULL rtx. */ void print_rtl_slim (FILE *f, rtx first, rtx last, int count, int flags) { basic_block current_bb = NULL; rtx insn, tail; tail = last ? NEXT_INSN (last) : NULL_RTX; for (insn = first; (insn != NULL) && (insn != tail) && (count != 0); insn = NEXT_INSN (insn)) { if ((flags & TDF_BLOCKS) && (INSN_P (insn) || GET_CODE (insn) == NOTE) && BLOCK_FOR_INSN (insn) && !current_bb) { current_bb = BLOCK_FOR_INSN (insn); dump_bb_info (current_bb, true, false, flags, ";; ", f); } dump_insn_slim (f, insn); if ((flags & TDF_BLOCKS) && current_bb && insn == BB_END (current_bb)) { dump_bb_info (current_bb, false, true, flags, ";; ", f); current_bb = NULL; } if (count > 0) count--; } }
static void add_test (rtx cond, basic_block bb, basic_block dest) { rtx seq, jump, label; enum machine_mode mode; rtx op0 = XEXP (cond, 0), op1 = XEXP (cond, 1); enum rtx_code code = GET_CODE (cond); mode = GET_MODE (XEXP (cond, 0)); if (mode == VOIDmode) mode = GET_MODE (XEXP (cond, 1)); start_sequence (); op0 = force_operand (op0, NULL_RTX); op1 = force_operand (op1, NULL_RTX); label = block_label (dest); do_compare_rtx_and_jump (op0, op1, code, 0, mode, NULL_RTX, NULL_RTX, label); jump = get_last_insn (); JUMP_LABEL (jump) = label; /* The jump is supposed to handle an unlikely special case. */ REG_NOTES (jump) = gen_rtx_EXPR_LIST (REG_BR_PROB, const0_rtx, REG_NOTES (jump)); LABEL_NUSES (label)++; seq = get_insns (); end_sequence (); emit_insn_after (seq, BB_END (bb)); }
static rtx block_label_after (rtx insn) { basic_block bb = BLOCK_FOR_INSN (insn); if ((insn == BB_END (bb)) && (bb->next_bb != EXIT_BLOCK_PTR)) return block_label (bb->next_bb); else return block_label (split_block_and_df_analyze (bb, insn)); }
static rtx next_active_insn_bb (basic_block bb, rtx insn) { for (insn = NEXT_INSN (insn); insn != NEXT_INSN (BB_END (bb)); insn = NEXT_INSN (insn)) if (active_insn_p (insn)) return insn; return NULL_RTX; }
static void record_effective_endpoints (void) { rtx next_insn; basic_block bb; rtx insn; for (insn = get_insns (); insn && NOTE_P (insn) && NOTE_LINE_NUMBER (insn) != NOTE_INSN_BASIC_BLOCK; insn = NEXT_INSN (insn)) continue; /* No basic blocks at all? */ gcc_assert (insn); if (PREV_INSN (insn)) cfg_layout_function_header = unlink_insn_chain (get_insns (), PREV_INSN (insn)); else cfg_layout_function_header = NULL_RTX; next_insn = get_insns (); FOR_EACH_BB (bb) { rtx end; if (PREV_INSN (BB_HEAD (bb)) && next_insn != BB_HEAD (bb)) bb->il.rtl->header = unlink_insn_chain (next_insn, PREV_INSN (BB_HEAD (bb))); end = skip_insns_after_block (bb); if (NEXT_INSN (BB_END (bb)) && BB_END (bb) != end) bb->il.rtl->footer = unlink_insn_chain (NEXT_INSN (BB_END (bb)), end); next_insn = NEXT_INSN (BB_END (bb)); } cfg_layout_function_footer = next_insn; if (cfg_layout_function_footer) cfg_layout_function_footer = unlink_insn_chain (cfg_layout_function_footer, get_last_insn ()); }
static int count_insns (basic_block bb) { rtx insn; int n = 0; for (insn = BB_HEAD (bb); insn != NEXT_INSN (BB_END (bb)); insn = NEXT_INSN (insn)) if (active_insn_p (insn)) n++; return n; }
static struct seginfo * new_seginfo (int mode, rtx_insn *insn, int bb, HARD_REG_SET regs_live) { struct seginfo *ptr; gcc_assert (!NOTE_INSN_BASIC_BLOCK_P (insn) || insn == BB_END (NOTE_BASIC_BLOCK (insn))); ptr = XNEW (struct seginfo); ptr->mode = mode; ptr->insn_ptr = insn; ptr->bbnum = bb; ptr->next = NULL; COPY_HARD_REG_SET (ptr->regs_live, regs_live); return ptr; }
bool cfg_layout_can_duplicate_bb_p (basic_block bb) { edge s; if (bb == EXIT_BLOCK_PTR || bb == ENTRY_BLOCK_PTR) return false; /* Duplicating fallthru block to exit would require adding a jump and splitting the real last BB. */ for (s = bb->succ; s; s = s->succ_next) if (s->dest == EXIT_BLOCK_PTR && s->flags & EDGE_FALLTHRU) return false; /* Do not attempt to duplicate tablejumps, as we need to unshare the dispatch table. This is difficult to do, as the instructions computing jump destination may be hoisted outside the basic block. */ if (tablejump_p (BB_END (bb), NULL, NULL)) return false; /* Do not duplicate blocks containing insns that can't be copied. */ if (targetm.cannot_copy_insn_p) { rtx insn = BB_HEAD (bb); while (1) { if (INSN_P (insn) && (*targetm.cannot_copy_insn_p) (insn)) return false; if (insn == BB_END (bb)) break; insn = NEXT_INSN (insn); } } return true; }
static instr * display_stored_regs(instr * pro_pc, unsigned char * sp) { instr * ret_pc = 0; int reg; unsigned long value; printk("Prologue [<%p>], Frame %p:\n", pro_pc, sp); while (!BB_END(*pro_pc)) if (STK_PUSH_MATCH(*pro_pc)) { reg = (*pro_pc & MEM_REG) >> 21; value = *(unsigned long *)(sp + (*pro_pc & MEM_OFF)); if (reg == 26) ret_pc = (instr *)value; printk("\t\t%s / 0x%016lx\n", reg_name[reg], value); }
static bool copy_bb_p (basic_block bb, int code_may_grow) { int size = 0; int max_size = uncond_jump_length; rtx insn; int n_succ; edge e; if (!bb->frequency) return false; if (!bb->pred || !bb->pred->pred_next) return false; if (!cfg_layout_can_duplicate_bb_p (bb)) return false; /* Avoid duplicating blocks which have many successors (PR/13430). */ n_succ = 0; for (e = bb->succ; e; e = e->succ_next) { n_succ++; if (n_succ > 8) return false; } if (code_may_grow && maybe_hot_bb_p (bb)) max_size *= 8; for (insn = BB_HEAD (bb); insn != NEXT_INSN (BB_END (bb)); insn = NEXT_INSN (insn)) { if (INSN_P (insn)) size += get_attr_length (insn); } if (size <= max_size) return true; if (rtl_dump_file) { fprintf (rtl_dump_file, "Block %d can't be copied because its size = %d.\n", bb->index, size); } return false; }
basic_block cfg_layout_duplicate_bb (basic_block bb) { rtx insn; basic_block new_bb; insn = duplicate_insn_chain (BB_HEAD (bb), BB_END (bb)); new_bb = create_basic_block (insn, insn ? get_last_insn () : NULL, EXIT_BLOCK_PTR->prev_bb); BB_COPY_PARTITION (new_bb, bb); if (bb->il.rtl->header) { insn = bb->il.rtl->header; while (NEXT_INSN (insn)) insn = NEXT_INSN (insn); insn = duplicate_insn_chain (bb->il.rtl->header, insn); if (insn) new_bb->il.rtl->header = unlink_insn_chain (insn, get_last_insn ()); } if (bb->il.rtl->footer) { insn = bb->il.rtl->footer; while (NEXT_INSN (insn)) insn = NEXT_INSN (insn); insn = duplicate_insn_chain (bb->il.rtl->footer, insn); if (insn) new_bb->il.rtl->footer = unlink_insn_chain (insn, get_last_insn ()); } if (bb->il.rtl->global_live_at_start) { new_bb->il.rtl->global_live_at_start = ALLOC_REG_SET (®_obstack); new_bb->il.rtl->global_live_at_end = ALLOC_REG_SET (®_obstack); COPY_REG_SET (new_bb->il.rtl->global_live_at_start, bb->il.rtl->global_live_at_start); COPY_REG_SET (new_bb->il.rtl->global_live_at_end, bb->il.rtl->global_live_at_end); } return new_bb; }
static basic_block make_forwarder_block (basic_block bb, int redirect_latch, int redirect_nonlatch, edge except, int conn_latch) { edge e, next_e, fallthru; basic_block dummy; rtx insn; insn = PREV_INSN (first_insn_after_basic_block_note (bb)); /* For empty block split_block will return NULL. */ if (BB_END (bb) == insn) emit_note_after (NOTE_INSN_DELETED, insn); fallthru = split_block (bb, insn); dummy = fallthru->src; bb = fallthru->dest; bb->aux = xmalloc (sizeof (int)); HEADER_BLOCK (dummy) = 0; HEADER_BLOCK (bb) = 1; /* Redirect back edges we want to keep. */ for (e = dummy->pred; e; e = next_e) { next_e = e->pred_next; if (e == except || !((redirect_latch && LATCH_EDGE (e)) || (redirect_nonlatch && !LATCH_EDGE (e)))) { dummy->frequency -= EDGE_FREQUENCY (e); dummy->count -= e->count; if (dummy->frequency < 0) dummy->frequency = 0; if (dummy->count < 0) dummy->count = 0; redirect_edge_with_latch_update (e, bb); } } alloc_aux_for_edge (fallthru, sizeof (int)); LATCH_EDGE (fallthru) = conn_latch; return dummy; }
static void clear_regs_live_in_seq (HARD_REG_SET * regs, rtx insn, int length) { basic_block bb; regset_head live; HARD_REG_SET hlive; rtx x; int i; /* Initialize liveness propagation. */ bb = BLOCK_FOR_INSN (insn); INIT_REG_SET (&live); bitmap_copy (&live, DF_LR_OUT (bb)); df_simulate_initialize_backwards (bb, &live); /* Propagate until INSN if found. */ for (x = BB_END (bb); x != insn; x = PREV_INSN (x)) df_simulate_one_insn_backwards (bb, x, &live); /* Clear registers live after INSN. */ renumbered_reg_set_to_hard_reg_set (&hlive, &live); AND_COMPL_HARD_REG_SET (*regs, hlive); /* Clear registers live in and before the sequence. */ for (i = 0; i < length;) { rtx prev = PREV_INSN (x); df_simulate_one_insn_backwards (bb, x, &live); if (INSN_P (x)) { renumbered_reg_set_to_hard_reg_set (&hlive, &live); AND_COMPL_HARD_REG_SET (*regs, hlive); i++; } x = prev; } /* Free unused data. */ CLEAR_REG_SET (&live); }
static int optimize_mode_switching (void) { int e; basic_block bb; bool need_commit = false; static const int num_modes[] = NUM_MODES_FOR_MODE_SWITCHING; #define N_ENTITIES ARRAY_SIZE (num_modes) int entity_map[N_ENTITIES]; struct bb_info *bb_info[N_ENTITIES]; int i, j; int n_entities = 0; int max_num_modes = 0; bool emitted ATTRIBUTE_UNUSED = false; basic_block post_entry = 0; basic_block pre_exit = 0; struct edge_list *edge_list = 0; /* These bitmaps are used for the LCM algorithm. */ sbitmap *kill, *del, *insert, *antic, *transp, *comp; sbitmap *avin, *avout; for (e = N_ENTITIES - 1; e >= 0; e--) if (OPTIMIZE_MODE_SWITCHING (e)) { int entry_exit_extra = 0; /* Create the list of segments within each basic block. If NORMAL_MODE is defined, allow for two extra blocks split from the entry and exit block. */ if (targetm.mode_switching.entry && targetm.mode_switching.exit) entry_exit_extra = 3; bb_info[n_entities] = XCNEWVEC (struct bb_info, last_basic_block_for_fn (cfun) + entry_exit_extra); entity_map[n_entities++] = e; if (num_modes[e] > max_num_modes) max_num_modes = num_modes[e]; } if (! n_entities) return 0; /* Make sure if MODE_ENTRY is defined MODE_EXIT is defined. */ gcc_assert ((targetm.mode_switching.entry && targetm.mode_switching.exit) || (!targetm.mode_switching.entry && !targetm.mode_switching.exit)); if (targetm.mode_switching.entry && targetm.mode_switching.exit) { /* Split the edge from the entry block, so that we can note that there NORMAL_MODE is supplied. */ post_entry = split_edge (single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun))); pre_exit = create_pre_exit (n_entities, entity_map, num_modes); } df_analyze (); /* Create the bitmap vectors. */ antic = sbitmap_vector_alloc (last_basic_block_for_fn (cfun), n_entities * max_num_modes); transp = sbitmap_vector_alloc (last_basic_block_for_fn (cfun), n_entities * max_num_modes); comp = sbitmap_vector_alloc (last_basic_block_for_fn (cfun), n_entities * max_num_modes); avin = sbitmap_vector_alloc (last_basic_block_for_fn (cfun), n_entities * max_num_modes); avout = sbitmap_vector_alloc (last_basic_block_for_fn (cfun), n_entities * max_num_modes); kill = sbitmap_vector_alloc (last_basic_block_for_fn (cfun), n_entities * max_num_modes); bitmap_vector_ones (transp, last_basic_block_for_fn (cfun)); bitmap_vector_clear (antic, last_basic_block_for_fn (cfun)); bitmap_vector_clear (comp, last_basic_block_for_fn (cfun)); for (j = n_entities - 1; j >= 0; j--) { int e = entity_map[j]; int no_mode = num_modes[e]; struct bb_info *info = bb_info[j]; rtx_insn *insn; /* Determine what the first use (if any) need for a mode of entity E is. This will be the mode that is anticipatable for this block. Also compute the initial transparency settings. */ FOR_EACH_BB_FN (bb, cfun) { struct seginfo *ptr; int last_mode = no_mode; bool any_set_required = false; HARD_REG_SET live_now; info[bb->index].mode_out = info[bb->index].mode_in = no_mode; REG_SET_TO_HARD_REG_SET (live_now, df_get_live_in (bb)); /* Pretend the mode is clobbered across abnormal edges. */ { edge_iterator ei; edge eg; FOR_EACH_EDGE (eg, ei, bb->preds) if (eg->flags & EDGE_COMPLEX) break; if (eg) { rtx_insn *ins_pos = BB_HEAD (bb); if (LABEL_P (ins_pos)) ins_pos = NEXT_INSN (ins_pos); gcc_assert (NOTE_INSN_BASIC_BLOCK_P (ins_pos)); if (ins_pos != BB_END (bb)) ins_pos = NEXT_INSN (ins_pos); ptr = new_seginfo (no_mode, ins_pos, bb->index, live_now); add_seginfo (info + bb->index, ptr); for (i = 0; i < no_mode; i++) clear_mode_bit (transp[bb->index], j, i); } } FOR_BB_INSNS (bb, insn) { if (INSN_P (insn)) { int mode = targetm.mode_switching.needed (e, insn); rtx link; if (mode != no_mode && mode != last_mode) { any_set_required = true; last_mode = mode; ptr = new_seginfo (mode, insn, bb->index, live_now); add_seginfo (info + bb->index, ptr); for (i = 0; i < no_mode; i++) clear_mode_bit (transp[bb->index], j, i); } if (targetm.mode_switching.after) last_mode = targetm.mode_switching.after (e, last_mode, insn); /* Update LIVE_NOW. */ for (link = REG_NOTES (insn); link; link = XEXP (link, 1)) if (REG_NOTE_KIND (link) == REG_DEAD) reg_dies (XEXP (link, 0), &live_now); note_stores (PATTERN (insn), reg_becomes_live, &live_now); for (link = REG_NOTES (insn); link; link = XEXP (link, 1)) if (REG_NOTE_KIND (link) == REG_UNUSED) reg_dies (XEXP (link, 0), &live_now); } } info[bb->index].computing = last_mode; /* Check for blocks without ANY mode requirements. N.B. because of MODE_AFTER, last_mode might still be different from no_mode, in which case we need to mark the block as nontransparent. */ if (!any_set_required) { ptr = new_seginfo (no_mode, BB_END (bb), bb->index, live_now); add_seginfo (info + bb->index, ptr); if (last_mode != no_mode) for (i = 0; i < no_mode; i++) clear_mode_bit (transp[bb->index], j, i); } } if (targetm.mode_switching.entry && targetm.mode_switching.exit) { int mode = targetm.mode_switching.entry (e); info[post_entry->index].mode_out = info[post_entry->index].mode_in = no_mode; if (pre_exit) { info[pre_exit->index].mode_out = info[pre_exit->index].mode_in = no_mode; } if (mode != no_mode) { bb = post_entry; /* By always making this nontransparent, we save an extra check in make_preds_opaque. We also need this to avoid confusing pre_edge_lcm when antic is cleared but transp and comp are set. */ for (i = 0; i < no_mode; i++) clear_mode_bit (transp[bb->index], j, i); /* Insert a fake computing definition of MODE into entry blocks which compute no mode. This represents the mode on entry. */ info[bb->index].computing = mode; if (pre_exit) info[pre_exit->index].seginfo->mode = targetm.mode_switching.exit (e); } } /* Set the anticipatable and computing arrays. */ for (i = 0; i < no_mode; i++) { int m = targetm.mode_switching.priority (entity_map[j], i); FOR_EACH_BB_FN (bb, cfun) { if (info[bb->index].seginfo->mode == m) set_mode_bit (antic[bb->index], j, m); if (info[bb->index].computing == m) set_mode_bit (comp[bb->index], j, m); } } } /* Calculate the optimal locations for the placement mode switches to modes with priority I. */ FOR_EACH_BB_FN (bb, cfun) bitmap_not (kill[bb->index], transp[bb->index]); edge_list = pre_edge_lcm_avs (n_entities * max_num_modes, transp, comp, antic, kill, avin, avout, &insert, &del); for (j = n_entities - 1; j >= 0; j--) { int no_mode = num_modes[entity_map[j]]; /* Insert all mode sets that have been inserted by lcm. */ for (int ed = NUM_EDGES (edge_list) - 1; ed >= 0; ed--) { edge eg = INDEX_EDGE (edge_list, ed); eg->aux = (void *)(intptr_t)-1; for (i = 0; i < no_mode; i++) { int m = targetm.mode_switching.priority (entity_map[j], i); if (mode_bit_p (insert[ed], j, m)) { eg->aux = (void *)(intptr_t)m; break; } } } FOR_EACH_BB_FN (bb, cfun) { struct bb_info *info = bb_info[j]; int last_mode = no_mode; /* intialize mode in availability for bb. */ for (i = 0; i < no_mode; i++) if (mode_bit_p (avout[bb->index], j, i)) { if (last_mode == no_mode) last_mode = i; if (last_mode != i) { last_mode = no_mode; break; } } info[bb->index].mode_out = last_mode; /* intialize mode out availability for bb. */ last_mode = no_mode; for (i = 0; i < no_mode; i++) if (mode_bit_p (avin[bb->index], j, i)) { if (last_mode == no_mode) last_mode = i; if (last_mode != i) { last_mode = no_mode; break; } } info[bb->index].mode_in = last_mode; for (i = 0; i < no_mode; i++) if (mode_bit_p (del[bb->index], j, i)) info[bb->index].seginfo->mode = no_mode; } /* Now output the remaining mode sets in all the segments. */ /* In case there was no mode inserted. the mode information on the edge might not be complete. Update mode info on edges and commit pending mode sets. */ need_commit |= commit_mode_sets (edge_list, entity_map[j], bb_info[j]); /* Reset modes for next entity. */ clear_aux_for_edges (); FOR_EACH_BB_FN (bb, cfun) { struct seginfo *ptr, *next; int cur_mode = bb_info[j][bb->index].mode_in; for (ptr = bb_info[j][bb->index].seginfo; ptr; ptr = next) { next = ptr->next; if (ptr->mode != no_mode) { rtx_insn *mode_set; rtl_profile_for_bb (bb); start_sequence (); targetm.mode_switching.emit (entity_map[j], ptr->mode, cur_mode, ptr->regs_live); mode_set = get_insns (); end_sequence (); /* modes kill each other inside a basic block. */ cur_mode = ptr->mode; /* Insert MODE_SET only if it is nonempty. */ if (mode_set != NULL_RTX) { emitted = true; if (NOTE_INSN_BASIC_BLOCK_P (ptr->insn_ptr)) /* We need to emit the insns in a FIFO-like manner, i.e. the first to be emitted at our insertion point ends up first in the instruction steam. Because we made sure that NOTE_INSN_BASIC_BLOCK is only used for initially empty basic blocks, we can achieve this by appending at the end of the block. */ emit_insn_after (mode_set, BB_END (NOTE_BASIC_BLOCK (ptr->insn_ptr))); else emit_insn_before (mode_set, ptr->insn_ptr); } default_rtl_profile (); } free (ptr); } } free (bb_info[j]); } free_edge_list (edge_list); /* Finished. Free up all the things we've allocated. */ sbitmap_vector_free (del); sbitmap_vector_free (insert); sbitmap_vector_free (kill); sbitmap_vector_free (antic); sbitmap_vector_free (transp); sbitmap_vector_free (comp); sbitmap_vector_free (avin); sbitmap_vector_free (avout); if (need_commit) commit_edge_insertions (); if (targetm.mode_switching.entry && targetm.mode_switching.exit) cleanup_cfg (CLEANUP_NO_INSN_DEL); else if (!need_commit && !emitted) return 0; return 1; }
static basic_block create_pre_exit (int n_entities, int *entity_map, const int *num_modes) { edge eg; edge_iterator ei; basic_block pre_exit; /* The only non-call predecessor at this stage is a block with a fallthrough edge; there can be at most one, but there could be none at all, e.g. when exit is called. */ pre_exit = 0; FOR_EACH_EDGE (eg, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds) if (eg->flags & EDGE_FALLTHRU) { basic_block src_bb = eg->src; rtx_insn *last_insn; rtx ret_reg; gcc_assert (!pre_exit); /* If this function returns a value at the end, we have to insert the final mode switch before the return value copy to its hard register. */ if (EDGE_COUNT (EXIT_BLOCK_PTR_FOR_FN (cfun)->preds) == 1 && NONJUMP_INSN_P ((last_insn = BB_END (src_bb))) && GET_CODE (PATTERN (last_insn)) == USE && GET_CODE ((ret_reg = XEXP (PATTERN (last_insn), 0))) == REG) { int ret_start = REGNO (ret_reg); int nregs = hard_regno_nregs[ret_start][GET_MODE (ret_reg)]; int ret_end = ret_start + nregs; bool short_block = false; bool multi_reg_return = false; bool forced_late_switch = false; rtx_insn *before_return_copy; do { rtx_insn *return_copy = PREV_INSN (last_insn); rtx return_copy_pat, copy_reg; int copy_start, copy_num; int j; if (NONDEBUG_INSN_P (return_copy)) { /* When using SJLJ exceptions, the call to the unregister function is inserted between the clobber of the return value and the copy. We do not want to split the block before this or any other call; if we have not found the copy yet, the copy must have been deleted. */ if (CALL_P (return_copy)) { short_block = true; break; } return_copy_pat = PATTERN (return_copy); switch (GET_CODE (return_copy_pat)) { case USE: /* Skip USEs of multiple return registers. __builtin_apply pattern is also handled here. */ if (GET_CODE (XEXP (return_copy_pat, 0)) == REG && (targetm.calls.function_value_regno_p (REGNO (XEXP (return_copy_pat, 0))))) { multi_reg_return = true; last_insn = return_copy; continue; } break; case ASM_OPERANDS: /* Skip barrier insns. */ if (!MEM_VOLATILE_P (return_copy_pat)) break; /* Fall through. */ case ASM_INPUT: case UNSPEC_VOLATILE: last_insn = return_copy; continue; default: break; } /* If the return register is not (in its entirety) likely spilled, the return copy might be partially or completely optimized away. */ return_copy_pat = single_set (return_copy); if (!return_copy_pat) { return_copy_pat = PATTERN (return_copy); if (GET_CODE (return_copy_pat) != CLOBBER) break; else if (!optimize) { /* This might be (clobber (reg [<result>])) when not optimizing. Then check if the previous insn is the clobber for the return register. */ copy_reg = SET_DEST (return_copy_pat); if (GET_CODE (copy_reg) == REG && !HARD_REGISTER_NUM_P (REGNO (copy_reg))) { if (INSN_P (PREV_INSN (return_copy))) { return_copy = PREV_INSN (return_copy); return_copy_pat = PATTERN (return_copy); if (GET_CODE (return_copy_pat) != CLOBBER) break; } } } } copy_reg = SET_DEST (return_copy_pat); if (GET_CODE (copy_reg) == REG) copy_start = REGNO (copy_reg); else if (GET_CODE (copy_reg) == SUBREG && GET_CODE (SUBREG_REG (copy_reg)) == REG) copy_start = REGNO (SUBREG_REG (copy_reg)); else { /* When control reaches end of non-void function, there are no return copy insns at all. This avoids an ice on that invalid function. */ if (ret_start + nregs == ret_end) short_block = true; break; } if (!targetm.calls.function_value_regno_p (copy_start)) copy_num = 0; else copy_num = hard_regno_nregs[copy_start][GET_MODE (copy_reg)]; /* If the return register is not likely spilled, - as is the case for floating point on SH4 - then it might be set by an arithmetic operation that needs a different mode than the exit block. */ for (j = n_entities - 1; j >= 0; j--) { int e = entity_map[j]; int mode = targetm.mode_switching.needed (e, return_copy); if (mode != num_modes[e] && mode != targetm.mode_switching.exit (e)) break; } if (j >= 0) { /* __builtin_return emits a sequence of loads to all return registers. One of them might require another mode than MODE_EXIT, even if it is unrelated to the return value, so we want to put the final mode switch after it. */ if (multi_reg_return && targetm.calls.function_value_regno_p (copy_start)) forced_late_switch = true; /* For the SH4, floating point loads depend on fpscr, thus we might need to put the final mode switch after the return value copy. That is still OK, because a floating point return value does not conflict with address reloads. */ if (copy_start >= ret_start && copy_start + copy_num <= ret_end && OBJECT_P (SET_SRC (return_copy_pat))) forced_late_switch = true; break; } if (copy_num == 0) { last_insn = return_copy; continue; } if (copy_start >= ret_start && copy_start + copy_num <= ret_end) nregs -= copy_num; else if (!multi_reg_return || !targetm.calls.function_value_regno_p (copy_start)) break; last_insn = return_copy; } /* ??? Exception handling can lead to the return value copy being already separated from the return value use, as in unwind-dw2.c . Similarly, conditionally returning without a value, and conditionally using builtin_return can lead to an isolated use. */ if (return_copy == BB_HEAD (src_bb)) { short_block = true; break; } last_insn = return_copy; } while (nregs); /* If we didn't see a full return value copy, verify that there is a plausible reason for this. If some, but not all of the return register is likely spilled, we can expect that there is a copy for the likely spilled part. */ gcc_assert (!nregs || forced_late_switch || short_block || !(targetm.class_likely_spilled_p (REGNO_REG_CLASS (ret_start))) || (nregs != hard_regno_nregs[ret_start][GET_MODE (ret_reg)]) /* For multi-hard-register floating point values, sometimes the likely-spilled part is ordinarily copied first, then the other part is set with an arithmetic operation. This doesn't actually cause reload failures, so let it pass. */ || (GET_MODE_CLASS (GET_MODE (ret_reg)) != MODE_INT && nregs != 1)); if (!NOTE_INSN_BASIC_BLOCK_P (last_insn)) { before_return_copy = emit_note_before (NOTE_INSN_DELETED, last_insn); /* Instructions preceding LAST_INSN in the same block might require a different mode than MODE_EXIT, so if we might have such instructions, keep them in a separate block from pre_exit. */ src_bb = split_block (src_bb, PREV_INSN (before_return_copy))->dest; } else before_return_copy = last_insn; pre_exit = split_block (src_bb, before_return_copy)->src; } else { pre_exit = split_edge (eg); } } return pre_exit; }
static basic_block expand_gimple_tailcall (basic_block bb, tree stmt, bool *can_fallthru) { rtx last2, last; edge e; edge_iterator ei; int probability; gcov_type count; last2 = last = get_last_insn (); expand_expr_stmt (stmt); for (last = NEXT_INSN (last); last; last = NEXT_INSN (last)) if (CALL_P (last) && SIBLING_CALL_P (last)) goto found; maybe_dump_rtl_for_tree_stmt (stmt, last2); *can_fallthru = true; return NULL; found: /* ??? Wouldn't it be better to just reset any pending stack adjust? Any instructions emitted here are about to be deleted. */ do_pending_stack_adjust (); /* Remove any non-eh, non-abnormal edges that don't go to exit. */ /* ??? I.e. the fallthrough edge. HOWEVER! If there were to be EH or abnormal edges, we shouldn't have created a tail call in the first place. So it seems to me we should just be removing all edges here, or redirecting the existing fallthru edge to the exit block. */ probability = 0; count = 0; for (ei = ei_start (bb->succs); (e = ei_safe_edge (ei)); ) { if (!(e->flags & (EDGE_ABNORMAL | EDGE_EH))) { if (e->dest != EXIT_BLOCK_PTR) { e->dest->count -= e->count; e->dest->frequency -= EDGE_FREQUENCY (e); if (e->dest->count < 0) e->dest->count = 0; if (e->dest->frequency < 0) e->dest->frequency = 0; } count += e->count; probability += e->probability; remove_edge (e); } else ei_next (&ei); } /* This is somewhat ugly: the call_expr expander often emits instructions after the sibcall (to perform the function return). These confuse the find_sub_basic_blocks code, so we need to get rid of these. */ last = NEXT_INSN (last); gcc_assert (BARRIER_P (last)); *can_fallthru = false; while (NEXT_INSN (last)) { /* For instance an sqrt builtin expander expands if with sibcall in the then and label for `else`. */ if (LABEL_P (NEXT_INSN (last))) { *can_fallthru = true; break; } delete_insn (NEXT_INSN (last)); } e = make_edge (bb, EXIT_BLOCK_PTR, EDGE_ABNORMAL | EDGE_SIBCALL); e->probability += probability; e->count += count; BB_END (bb) = last; update_bb_for_insn (bb); if (NEXT_INSN (last)) { bb = create_basic_block (NEXT_INSN (last), get_last_insn (), bb); last = BB_END (bb); if (BARRIER_P (last)) BB_END (bb) = PREV_INSN (last); } maybe_dump_rtl_for_tree_stmt (stmt, last2); return bb; }
static rtx may_unswitch_on (basic_block bb, struct loop *loop, rtx *cinsn) { rtx test, at, op[2], stest; struct rtx_iv iv; unsigned i; enum machine_mode mode; /* BB must end in a simple conditional jump. */ if (EDGE_COUNT (bb->succs) != 2) return NULL_RTX; if (!any_condjump_p (BB_END (bb))) return NULL_RTX; /* With branches inside loop. */ if (!flow_bb_inside_loop_p (loop, EDGE_SUCC (bb, 0)->dest) || !flow_bb_inside_loop_p (loop, EDGE_SUCC (bb, 1)->dest)) return NULL_RTX; /* It must be executed just once each iteration (because otherwise we are unable to update dominator/irreducible loop information correctly). */ if (!just_once_each_iteration_p (loop, bb)) return NULL_RTX; /* Condition must be invariant. */ test = get_condition (BB_END (bb), &at, true, false); if (!test) return NULL_RTX; for (i = 0; i < 2; i++) { op[i] = XEXP (test, i); if (CONSTANT_P (op[i])) continue; if (!iv_analyze (at, op[i], &iv)) return NULL_RTX; if (iv.step != const0_rtx || iv.first_special) return NULL_RTX; op[i] = get_iv_value (&iv, const0_rtx); } mode = GET_MODE (op[0]); if (mode == VOIDmode) mode = GET_MODE (op[1]); if (GET_MODE_CLASS (mode) == MODE_CC) { if (at != BB_END (bb)) return NULL_RTX; if (!rtx_equal_p (op[0], XEXP (test, 0)) || !rtx_equal_p (op[1], XEXP (test, 1))) return NULL_RTX; *cinsn = BB_END (bb); return test; } stest = simplify_gen_relational (GET_CODE (test), SImode, mode, op[0], op[1]); if (stest == const0_rtx || stest == const_true_rtx) return stest; return canon_condition (gen_rtx_fmt_ee (GET_CODE (test), SImode, op[0], op[1])); }
static struct loop * unswitch_loop (struct loop *loop, basic_block unswitch_on, rtx cond, rtx cinsn) { edge entry, latch_edge, true_edge, false_edge, e; basic_block switch_bb, unswitch_on_alt; struct loop *nloop; int irred_flag, prob; rtx seq; /* Some sanity checking. */ gcc_assert (flow_bb_inside_loop_p (loop, unswitch_on)); gcc_assert (EDGE_COUNT (unswitch_on->succs) == 2); gcc_assert (just_once_each_iteration_p (loop, unswitch_on)); gcc_assert (!loop->inner); gcc_assert (flow_bb_inside_loop_p (loop, EDGE_SUCC (unswitch_on, 0)->dest)); gcc_assert (flow_bb_inside_loop_p (loop, EDGE_SUCC (unswitch_on, 1)->dest)); entry = loop_preheader_edge (loop); /* Make a copy. */ irred_flag = entry->flags & EDGE_IRREDUCIBLE_LOOP; entry->flags &= ~EDGE_IRREDUCIBLE_LOOP; if (!duplicate_loop_to_header_edge (loop, entry, 1, NULL, NULL, NULL, 0)) return NULL; entry->flags |= irred_flag; /* Record the block with condition we unswitch on. */ unswitch_on_alt = get_bb_copy (unswitch_on); true_edge = BRANCH_EDGE (unswitch_on_alt); false_edge = FALLTHRU_EDGE (unswitch_on); latch_edge = single_succ_edge (get_bb_copy (loop->latch)); /* Create a block with the condition. */ prob = true_edge->probability; switch_bb = create_empty_bb (EXIT_BLOCK_PTR->prev_bb); seq = compare_and_jump_seq (XEXP (cond, 0), XEXP (cond, 1), GET_CODE (cond), block_label (true_edge->dest), prob, cinsn); emit_insn_after (seq, BB_END (switch_bb)); e = make_edge (switch_bb, true_edge->dest, 0); e->probability = prob; e->count = latch_edge->count * prob / REG_BR_PROB_BASE; e = make_edge (switch_bb, FALLTHRU_EDGE (unswitch_on)->dest, EDGE_FALLTHRU); e->probability = false_edge->probability; e->count = latch_edge->count * (false_edge->probability) / REG_BR_PROB_BASE; if (irred_flag) { switch_bb->flags |= BB_IRREDUCIBLE_LOOP; EDGE_SUCC (switch_bb, 0)->flags |= EDGE_IRREDUCIBLE_LOOP; EDGE_SUCC (switch_bb, 1)->flags |= EDGE_IRREDUCIBLE_LOOP; } else { switch_bb->flags &= ~BB_IRREDUCIBLE_LOOP; EDGE_SUCC (switch_bb, 0)->flags &= ~EDGE_IRREDUCIBLE_LOOP; EDGE_SUCC (switch_bb, 1)->flags &= ~EDGE_IRREDUCIBLE_LOOP; } /* Loopify from the copy of LOOP body, constructing the new loop. */ nloop = loopify (latch_edge, single_pred_edge (get_bb_copy (loop->header)), switch_bb, BRANCH_EDGE (switch_bb), FALLTHRU_EDGE (switch_bb), true, prob, REG_BR_PROB_BASE - prob); copy_loop_info (loop, nloop); /* Remove branches that are now unreachable in new loops. */ remove_path (true_edge); remove_path (false_edge); /* Preserve the simple loop preheaders. */ split_edge (loop_preheader_edge (loop)); split_edge (loop_preheader_edge (nloop)); return nloop; }
void debug_bb_slim (struct basic_block_def *bb) { print_rtl_slim (stderr, BB_HEAD (bb), BB_END (bb), -1, 32); }
static void collect_pattern_seqs (void) { htab_iterator hti0, hti1, hti2; p_hash_bucket hash_bucket; p_hash_elem e0, e1; #if defined STACK_REGS || defined HAVE_cc0 basic_block bb; bitmap_head dont_collect; /* Extra initialization step to ensure that no stack registers (if present) or cc0 code (if present) are live across abnormal edges. Set a flag in DONT_COLLECT for an insn if a stack register is live after the insn or the insn is cc0 setter or user. */ bitmap_initialize (&dont_collect, NULL); #ifdef STACK_REGS FOR_EACH_BB (bb) { regset_head live; rtx insn; rtx prev; /* Initialize liveness propagation. */ INIT_REG_SET (&live); bitmap_copy (&live, DF_LR_OUT (bb)); df_simulate_initialize_backwards (bb, &live); /* Propagate liveness info and mark insns where a stack reg is live. */ insn = BB_END (bb); for (insn = BB_END (bb); ; insn = prev) { prev = PREV_INSN (insn); if (INSN_P (insn)) { int reg; for (reg = FIRST_STACK_REG; reg <= LAST_STACK_REG; reg++) { if (REGNO_REG_SET_P (&live, reg)) { bitmap_set_bit (&dont_collect, INSN_UID (insn)); break; } } } if (insn == BB_HEAD (bb)) break; df_simulate_one_insn_backwards (bb, insn, &live); insn = prev; } /* Free unused data. */ CLEAR_REG_SET (&live); } #endif #ifdef HAVE_cc0 /* Mark CC0 setters and users as ineligible for collection into sequences. This is an over-conservative fix, since it is OK to include a cc0_setter, but only if we also include the corresponding cc0_user, and vice versa. */ FOR_EACH_BB (bb) { rtx insn; rtx next_tail; next_tail = NEXT_INSN (BB_END (bb)); for (insn = BB_HEAD (bb); insn != next_tail; insn = NEXT_INSN (insn)) { if (INSN_P (insn) && reg_mentioned_p (cc0_rtx, PATTERN (insn))) bitmap_set_bit (&dont_collect, INSN_UID (insn)); } } #endif #endif /* defined STACK_REGS || defined HAVE_cc0 */ /* Initialize PATTERN_SEQS to empty. */ pattern_seqs = 0; /* Try to match every abstractable insn with every other insn in the same HASH_BUCKET. */ FOR_EACH_HTAB_ELEMENT (hash_buckets, hash_bucket, p_hash_bucket, hti0) if (htab_elements (hash_bucket->seq_candidates) > 1) FOR_EACH_HTAB_ELEMENT (hash_bucket->seq_candidates, e0, p_hash_elem, hti1) FOR_EACH_HTAB_ELEMENT (hash_bucket->seq_candidates, e1, p_hash_elem, hti2) if (e0 != e1 #if defined STACK_REGS || defined HAVE_cc0 && !bitmap_bit_p (&dont_collect, INSN_UID (e0->insn)) && !bitmap_bit_p (&dont_collect, INSN_UID (e1->insn)) #endif ) match_seqs (e0, e1); #if defined STACK_REGS || defined HAVE_cc0 /* Free unused data. */ bitmap_clear (&dont_collect); #endif }
static basic_block rotate_loop (edge back_edge, struct trace *trace, int trace_n) { basic_block bb; /* Information about the best end (end after rotation) of the loop. */ basic_block best_bb = NULL; edge best_edge = NULL; int best_freq = -1; gcov_type best_count = -1; /* The best edge is preferred when its destination is not visited yet or is a start block of some trace. */ bool is_preferred = false; /* Find the most frequent edge that goes out from current trace. */ bb = back_edge->dest; do { edge e; for (e = bb->succ; e; e = e->succ_next) if (e->dest != EXIT_BLOCK_PTR && e->dest->rbi->visited != trace_n && (e->flags & EDGE_CAN_FALLTHRU) && !(e->flags & EDGE_COMPLEX)) { if (is_preferred) { /* The best edge is preferred. */ if (!e->dest->rbi->visited || bbd[e->dest->index].start_of_trace >= 0) { /* The current edge E is also preferred. */ int freq = EDGE_FREQUENCY (e); if (freq > best_freq || e->count > best_count) { best_freq = freq; best_count = e->count; best_edge = e; best_bb = bb; } } } else { if (!e->dest->rbi->visited || bbd[e->dest->index].start_of_trace >= 0) { /* The current edge E is preferred. */ is_preferred = true; best_freq = EDGE_FREQUENCY (e); best_count = e->count; best_edge = e; best_bb = bb; } else { int freq = EDGE_FREQUENCY (e); if (!best_edge || freq > best_freq || e->count > best_count) { best_freq = freq; best_count = e->count; best_edge = e; best_bb = bb; } } } } bb = bb->rbi->next; } while (bb != back_edge->dest); if (best_bb) { /* Rotate the loop so that the BEST_EDGE goes out from the last block of the trace. */ if (back_edge->dest == trace->first) { trace->first = best_bb->rbi->next; } else { basic_block prev_bb; for (prev_bb = trace->first; prev_bb->rbi->next != back_edge->dest; prev_bb = prev_bb->rbi->next) ; prev_bb->rbi->next = best_bb->rbi->next; /* Try to get rid of uncond jump to cond jump. */ if (prev_bb->succ && !prev_bb->succ->succ_next) { basic_block header = prev_bb->succ->dest; /* Duplicate HEADER if it is a small block containing cond jump in the end. */ if (any_condjump_p (BB_END (header)) && copy_bb_p (header, 0)) { copy_bb (header, prev_bb->succ, prev_bb, trace_n); } } } } else { /* We have not found suitable loop tail so do no rotation. */ best_bb = back_edge->src; } best_bb->rbi->next = NULL; return best_bb; }
static bool copyprop_hardreg_forward_1 (basic_block bb, struct value_data *vd) { bool anything_changed = false; rtx insn; for (insn = BB_HEAD (bb); ; insn = NEXT_INSN (insn)) { int n_ops, i, alt, predicated; bool is_asm, any_replacements; rtx set; bool replaced[MAX_RECOG_OPERANDS]; bool changed = false; struct kill_set_value_data ksvd; if (!NONDEBUG_INSN_P (insn)) { if (DEBUG_INSN_P (insn)) { rtx loc = INSN_VAR_LOCATION_LOC (insn); if (!VAR_LOC_UNKNOWN_P (loc)) replace_oldest_value_addr (&INSN_VAR_LOCATION_LOC (insn), ALL_REGS, GET_MODE (loc), ADDR_SPACE_GENERIC, insn, vd); } if (insn == BB_END (bb)) break; else continue; } set = single_set (insn); extract_insn (insn); if (! constrain_operands (1)) fatal_insn_not_found (insn); preprocess_constraints (); alt = which_alternative; n_ops = recog_data.n_operands; is_asm = asm_noperands (PATTERN (insn)) >= 0; /* Simplify the code below by rewriting things to reflect matching constraints. Also promote OP_OUT to OP_INOUT in predicated instructions. */ predicated = GET_CODE (PATTERN (insn)) == COND_EXEC; for (i = 0; i < n_ops; ++i) { int matches = recog_op_alt[i][alt].matches; if (matches >= 0) recog_op_alt[i][alt].cl = recog_op_alt[matches][alt].cl; if (matches >= 0 || recog_op_alt[i][alt].matched >= 0 || (predicated && recog_data.operand_type[i] == OP_OUT)) recog_data.operand_type[i] = OP_INOUT; } /* Apply changes to earlier DEBUG_INSNs if possible. */ if (vd->n_debug_insn_changes) note_uses (&PATTERN (insn), cprop_find_used_regs, vd); /* For each earlyclobber operand, zap the value data. */ for (i = 0; i < n_ops; i++) if (recog_op_alt[i][alt].earlyclobber) kill_value (recog_data.operand[i], vd); /* Within asms, a clobber cannot overlap inputs or outputs. I wouldn't think this were true for regular insns, but scan_rtx treats them like that... */ note_stores (PATTERN (insn), kill_clobbered_value, vd); /* Kill all auto-incremented values. */ /* ??? REG_INC is useless, since stack pushes aren't done that way. */ for_each_rtx (&PATTERN (insn), kill_autoinc_value, vd); /* Kill all early-clobbered operands. */ for (i = 0; i < n_ops; i++) if (recog_op_alt[i][alt].earlyclobber) kill_value (recog_data.operand[i], vd); /* Special-case plain move instructions, since we may well be able to do the move from a different register class. */ if (set && REG_P (SET_SRC (set))) { rtx src = SET_SRC (set); unsigned int regno = REGNO (src); enum machine_mode mode = GET_MODE (src); unsigned int i; rtx new_rtx; /* If we are accessing SRC in some mode other that what we set it in, make sure that the replacement is valid. */ if (mode != vd->e[regno].mode) { if (hard_regno_nregs[regno][mode] > hard_regno_nregs[regno][vd->e[regno].mode]) goto no_move_special_case; /* And likewise, if we are narrowing on big endian the transformation is also invalid. */ if (hard_regno_nregs[regno][mode] < hard_regno_nregs[regno][vd->e[regno].mode] && (GET_MODE_SIZE (vd->e[regno].mode) > UNITS_PER_WORD ? WORDS_BIG_ENDIAN : BYTES_BIG_ENDIAN)) goto no_move_special_case; } /* If the destination is also a register, try to find a source register in the same class. */ if (REG_P (SET_DEST (set))) { new_rtx = find_oldest_value_reg (REGNO_REG_CLASS (regno), src, vd); if (new_rtx && validate_change (insn, &SET_SRC (set), new_rtx, 0)) { if (dump_file) fprintf (dump_file, "insn %u: replaced reg %u with %u\n", INSN_UID (insn), regno, REGNO (new_rtx)); changed = true; goto did_replacement; } /* We need to re-extract as validate_change clobbers recog_data. */ extract_insn (insn); if (! constrain_operands (1)) fatal_insn_not_found (insn); preprocess_constraints (); } /* Otherwise, try all valid registers and see if its valid. */ for (i = vd->e[regno].oldest_regno; i != regno; i = vd->e[i].next_regno) { new_rtx = maybe_mode_change (vd->e[i].mode, vd->e[regno].mode, mode, i, regno); if (new_rtx != NULL_RTX) { if (validate_change (insn, &SET_SRC (set), new_rtx, 0)) { ORIGINAL_REGNO (new_rtx) = ORIGINAL_REGNO (src); REG_ATTRS (new_rtx) = REG_ATTRS (src); REG_POINTER (new_rtx) = REG_POINTER (src); if (dump_file) fprintf (dump_file, "insn %u: replaced reg %u with %u\n", INSN_UID (insn), regno, REGNO (new_rtx)); changed = true; goto did_replacement; } /* We need to re-extract as validate_change clobbers recog_data. */ extract_insn (insn); if (! constrain_operands (1)) fatal_insn_not_found (insn); preprocess_constraints (); } } } no_move_special_case: any_replacements = false; /* For each input operand, replace a hard register with the eldest live copy that's in an appropriate register class. */ for (i = 0; i < n_ops; i++) { replaced[i] = false; /* Don't scan match_operand here, since we've no reg class information to pass down. Any operands that we could substitute in will be represented elsewhere. */ if (recog_data.constraints[i][0] == '\0') continue; /* Don't replace in asms intentionally referencing hard regs. */ if (is_asm && REG_P (recog_data.operand[i]) && (REGNO (recog_data.operand[i]) == ORIGINAL_REGNO (recog_data.operand[i]))) continue; if (recog_data.operand_type[i] == OP_IN) { if (recog_op_alt[i][alt].is_address) replaced[i] = replace_oldest_value_addr (recog_data.operand_loc[i], recog_op_alt[i][alt].cl, VOIDmode, ADDR_SPACE_GENERIC, insn, vd); else if (REG_P (recog_data.operand[i])) replaced[i] = replace_oldest_value_reg (recog_data.operand_loc[i], recog_op_alt[i][alt].cl, insn, vd); else if (MEM_P (recog_data.operand[i])) replaced[i] = replace_oldest_value_mem (recog_data.operand[i], insn, vd); } else if (MEM_P (recog_data.operand[i])) replaced[i] = replace_oldest_value_mem (recog_data.operand[i], insn, vd); /* If we performed any replacement, update match_dups. */ if (replaced[i]) { int j; rtx new_rtx; new_rtx = *recog_data.operand_loc[i]; recog_data.operand[i] = new_rtx; for (j = 0; j < recog_data.n_dups; j++) if (recog_data.dup_num[j] == i) validate_unshare_change (insn, recog_data.dup_loc[j], new_rtx, 1); any_replacements = true; } } if (any_replacements) { if (! apply_change_group ()) { for (i = 0; i < n_ops; i++) if (replaced[i]) { rtx old = *recog_data.operand_loc[i]; recog_data.operand[i] = old; } if (dump_file) fprintf (dump_file, "insn %u: reg replacements not verified\n", INSN_UID (insn)); } else changed = true; } did_replacement: if (changed) { anything_changed = true; /* If something changed, perhaps further changes to earlier DEBUG_INSNs can be applied. */ if (vd->n_debug_insn_changes) note_uses (&PATTERN (insn), cprop_find_used_regs, vd); } ksvd.vd = vd; ksvd.ignore_set_reg = NULL_RTX; /* Clobber call-clobbered registers. */ if (CALL_P (insn)) { unsigned int set_regno = INVALID_REGNUM; unsigned int set_nregs = 0; unsigned int regno; rtx exp; hard_reg_set_iterator hrsi; for (exp = CALL_INSN_FUNCTION_USAGE (insn); exp; exp = XEXP (exp, 1)) { rtx x = XEXP (exp, 0); if (GET_CODE (x) == SET) { rtx dest = SET_DEST (x); kill_value (dest, vd); set_value_regno (REGNO (dest), GET_MODE (dest), vd); copy_value (dest, SET_SRC (x), vd); ksvd.ignore_set_reg = dest; set_regno = REGNO (dest); set_nregs = hard_regno_nregs[set_regno][GET_MODE (dest)]; break; } } EXECUTE_IF_SET_IN_HARD_REG_SET (regs_invalidated_by_call, 0, regno, hrsi) if (regno < set_regno || regno >= set_regno + set_nregs) kill_value_regno (regno, 1, vd); /* If SET was seen in CALL_INSN_FUNCTION_USAGE, and SET_SRC of the SET isn't in regs_invalidated_by_call hard reg set, but instead among CLOBBERs on the CALL_INSN, we could wrongly assume the value in it is still live. */ if (ksvd.ignore_set_reg) note_stores (PATTERN (insn), kill_clobbered_value, vd); } /* Notice stores. */ note_stores (PATTERN (insn), kill_set_value, &ksvd); /* Notice copies. */ if (set && REG_P (SET_DEST (set)) && REG_P (SET_SRC (set))) copy_value (SET_DEST (set), SET_SRC (set), vd); if (insn == BB_END (bb)) break; } return anything_changed; }
void find_comparison_dom_walker::before_dom_children (basic_block bb) { struct comparison *last_cmp; rtx_insn *insn, *next, *last_clobber; bool last_cmp_valid; bool need_purge = false; bitmap killed; killed = BITMAP_ALLOC (NULL); /* The last comparison that was made. Will be reset to NULL once the flags are clobbered. */ last_cmp = NULL; /* True iff the last comparison has not been clobbered, nor have its inputs. Used to eliminate duplicate compares. */ last_cmp_valid = false; /* The last insn that clobbered the flags, if that insn is of a form that may be valid for eliminating a following compare. To be reset to NULL once the flags are set otherwise. */ last_clobber = NULL; /* Propagate the last live comparison throughout the extended basic block. */ if (single_pred_p (bb)) { last_cmp = (struct comparison *) single_pred (bb)->aux; if (last_cmp) last_cmp_valid = last_cmp->inputs_valid; } for (insn = BB_HEAD (bb); insn; insn = next) { rtx src; next = (insn == BB_END (bb) ? NULL : NEXT_INSN (insn)); if (!NONDEBUG_INSN_P (insn)) continue; /* Compute the set of registers modified by this instruction. */ bitmap_clear (killed); df_simulate_find_defs (insn, killed); src = conforming_compare (insn); if (src) { rtx eh_note = NULL; if (cfun->can_throw_non_call_exceptions) eh_note = find_reg_note (insn, REG_EH_REGION, NULL); if (last_cmp_valid && can_eliminate_compare (src, eh_note, last_cmp)) { if (eh_note) need_purge = true; delete_insn (insn); continue; } last_cmp = XCNEW (struct comparison); last_cmp->insn = insn; last_cmp->prev_clobber = last_clobber; last_cmp->in_a = XEXP (src, 0); last_cmp->in_b = XEXP (src, 1); last_cmp->eh_note = eh_note; last_cmp->orig_mode = GET_MODE (src); all_compares.safe_push (last_cmp); /* It's unusual, but be prepared for comparison patterns that also clobber an input, or perhaps a scratch. */ last_clobber = NULL; last_cmp_valid = true; } /* Notice if this instruction kills the flags register. */ else if (bitmap_bit_p (killed, targetm.flags_regnum)) { /* See if this insn could be the "clobber" that eliminates a future comparison. */ last_clobber = (arithmetic_flags_clobber_p (insn) ? insn : NULL); /* In either case, the previous compare is no longer valid. */ last_cmp = NULL; last_cmp_valid = false; } /* Notice if this instruction uses the flags register. */ else if (last_cmp) find_flags_uses_in_insn (last_cmp, insn); /* Notice if any of the inputs to the comparison have changed. */ if (last_cmp_valid && (bitmap_bit_p (killed, REGNO (last_cmp->in_a)) || (REG_P (last_cmp->in_b) && bitmap_bit_p (killed, REGNO (last_cmp->in_b))))) last_cmp_valid = false; }
static void combine_stack_adjustments_for_block (basic_block bb) { HOST_WIDE_INT last_sp_adjust = 0; rtx last_sp_set = NULL_RTX; rtx last2_sp_set = NULL_RTX; struct csa_reflist *reflist = NULL; rtx insn, next, set; struct record_stack_refs_data data; bool end_of_block = false; for (insn = BB_HEAD (bb); !end_of_block ; insn = next) { end_of_block = insn == BB_END (bb); next = NEXT_INSN (insn); if (! INSN_P (insn)) continue; set = single_set_for_csa (insn); if (set) { rtx dest = SET_DEST (set); rtx src = SET_SRC (set); /* Find constant additions to the stack pointer. */ if (dest == stack_pointer_rtx && GET_CODE (src) == PLUS && XEXP (src, 0) == stack_pointer_rtx && CONST_INT_P (XEXP (src, 1))) { HOST_WIDE_INT this_adjust = INTVAL (XEXP (src, 1)); /* If we've not seen an adjustment previously, record it now and continue. */ if (! last_sp_set) { last_sp_set = insn; last_sp_adjust = this_adjust; continue; } /* If not all recorded refs can be adjusted, or the adjustment is now too large for a constant addition, we cannot merge the two stack adjustments. Also we need to be careful to not move stack pointer such that we create stack accesses outside the allocated area. We can combine an allocation into the first insn, or a deallocation into the second insn. We can not combine an allocation followed by a deallocation. The only somewhat frequent occurrence of the later is when a function allocates a stack frame but does not use it. For this case, we would need to analyze rtl stream to be sure that allocated area is really unused. This means not only checking the memory references, but also all registers or global memory references possibly containing a stack frame address. Perhaps the best way to address this problem is to teach gcc not to allocate stack for objects never used. */ /* Combine an allocation into the first instruction. */ if (STACK_GROWS_DOWNWARD ? this_adjust <= 0 : this_adjust >= 0) { if (try_apply_stack_adjustment (last_sp_set, reflist, last_sp_adjust + this_adjust, this_adjust)) { /* It worked! */ maybe_move_args_size_note (last_sp_set, insn, false); delete_insn (insn); last_sp_adjust += this_adjust; continue; } } /* Otherwise we have a deallocation. Do not combine with a previous allocation. Combine into the second insn. */ else if (STACK_GROWS_DOWNWARD ? last_sp_adjust >= 0 : last_sp_adjust <= 0) { if (try_apply_stack_adjustment (insn, reflist, last_sp_adjust + this_adjust, -last_sp_adjust)) { /* It worked! */ maybe_move_args_size_note (insn, last_sp_set, true); delete_insn (last_sp_set); last_sp_set = insn; last_sp_adjust += this_adjust; free_csa_reflist (reflist); reflist = NULL; continue; } } /* Combination failed. Restart processing from here. If deallocation+allocation conspired to cancel, we can delete the old deallocation insn. */ if (last_sp_set) { if (last_sp_adjust == 0) { maybe_move_args_size_note (insn, last_sp_set, true); delete_insn (last_sp_set); } else last2_sp_set = last_sp_set; } free_csa_reflist (reflist); reflist = NULL; last_sp_set = insn; last_sp_adjust = this_adjust; continue; } /* Find a store with pre-(dec|inc)rement or pre-modify of exactly the previous adjustment and turn it into a simple store. This is equivalent to anticipating the stack adjustment so this must be an allocation. */ if (MEM_P (dest) && ((STACK_GROWS_DOWNWARD ? (GET_CODE (XEXP (dest, 0)) == PRE_DEC && last_sp_adjust == (HOST_WIDE_INT) GET_MODE_SIZE (GET_MODE (dest))) : (GET_CODE (XEXP (dest, 0)) == PRE_INC && last_sp_adjust == -(HOST_WIDE_INT) GET_MODE_SIZE (GET_MODE (dest)))) || ((STACK_GROWS_DOWNWARD ? last_sp_adjust >= 0 : last_sp_adjust <= 0) && GET_CODE (XEXP (dest, 0)) == PRE_MODIFY && GET_CODE (XEXP (XEXP (dest, 0), 1)) == PLUS && XEXP (XEXP (XEXP (dest, 0), 1), 0) == stack_pointer_rtx && GET_CODE (XEXP (XEXP (XEXP (dest, 0), 1), 1)) == CONST_INT && INTVAL (XEXP (XEXP (XEXP (dest, 0), 1), 1)) == -last_sp_adjust)) && XEXP (XEXP (dest, 0), 0) == stack_pointer_rtx && !reg_mentioned_p (stack_pointer_rtx, src) && memory_address_p (GET_MODE (dest), stack_pointer_rtx) && try_apply_stack_adjustment (insn, reflist, 0, -last_sp_adjust)) { if (last2_sp_set) maybe_move_args_size_note (last2_sp_set, last_sp_set, false); else maybe_move_args_size_note (insn, last_sp_set, true); delete_insn (last_sp_set); free_csa_reflist (reflist); reflist = NULL; last_sp_set = NULL_RTX; last_sp_adjust = 0; continue; } } data.insn = insn; data.reflist = reflist; if (!CALL_P (insn) && last_sp_set && !for_each_rtx (&PATTERN (insn), record_stack_refs, &data)) { reflist = data.reflist; continue; } reflist = data.reflist; /* Otherwise, we were not able to process the instruction. Do not continue collecting data across such a one. */ if (last_sp_set && (CALL_P (insn) || reg_mentioned_p (stack_pointer_rtx, PATTERN (insn)))) { if (last_sp_set && last_sp_adjust == 0) { force_move_args_size_note (bb, last2_sp_set, last_sp_set); delete_insn (last_sp_set); } free_csa_reflist (reflist); reflist = NULL; last2_sp_set = NULL_RTX; last_sp_set = NULL_RTX; last_sp_adjust = 0; } } if (last_sp_set && last_sp_adjust == 0) { force_move_args_size_note (bb, last2_sp_set, last_sp_set); delete_insn (last_sp_set); } if (reflist) free_csa_reflist (reflist); }
void find_comparison_dom_walker::before_dom_children (basic_block bb) { struct comparison *last_cmp; rtx insn, next, last_clobber; bool last_cmp_valid; bool need_purge = false; bitmap killed; killed = BITMAP_ALLOC (NULL); /* The last comparison that was made. Will be reset to NULL once the flags are clobbered. */ last_cmp = NULL; /* True iff the last comparison has not been clobbered, nor have its inputs. Used to eliminate duplicate compares. */ last_cmp_valid = false; /* The last insn that clobbered the flags, if that insn is of a form that may be valid for eliminating a following compare. To be reset to NULL once the flags are set otherwise. */ last_clobber = NULL; /* Propagate the last live comparison throughout the extended basic block. */ if (single_pred_p (bb)) { last_cmp = (struct comparison *) single_pred (bb)->aux; if (last_cmp) last_cmp_valid = last_cmp->inputs_valid; } for (insn = BB_HEAD (bb); insn; insn = next) { rtx src; next = (insn == BB_END (bb) ? NULL_RTX : NEXT_INSN (insn)); if (!NONDEBUG_INSN_P (insn)) continue; /* Compute the set of registers modified by this instruction. */ bitmap_clear (killed); df_simulate_find_defs (insn, killed); src = conforming_compare (insn); if (src) { enum machine_mode src_mode = GET_MODE (src); rtx eh_note = NULL; if (flag_non_call_exceptions) eh_note = find_reg_note (insn, REG_EH_REGION, NULL); if (!last_cmp_valid) goto dont_delete; /* Take care that it's in the same EH region. */ if (flag_non_call_exceptions && !rtx_equal_p (eh_note, last_cmp->eh_note)) goto dont_delete; /* Make sure the compare is redundant with the previous. */ if (!rtx_equal_p (last_cmp->in_a, XEXP (src, 0)) || !rtx_equal_p (last_cmp->in_b, XEXP (src, 1))) goto dont_delete; /* New mode must be compatible with the previous compare mode. */ { enum machine_mode new_mode = targetm.cc_modes_compatible (last_cmp->orig_mode, src_mode); if (new_mode == VOIDmode) goto dont_delete; if (new_mode != last_cmp->orig_mode) { rtx x, flags = gen_rtx_REG (src_mode, targetm.flags_regnum); /* Generate new comparison for substitution. */ x = gen_rtx_COMPARE (new_mode, XEXP (src, 0), XEXP (src, 1)); x = gen_rtx_SET (VOIDmode, flags, x); if (!validate_change (last_cmp->insn, &PATTERN (last_cmp->insn), x, false)) goto dont_delete; last_cmp->orig_mode = new_mode; } } /* All tests and substitutions succeeded! */ if (eh_note) need_purge = true; delete_insn (insn); continue; dont_delete: last_cmp = XCNEW (struct comparison); last_cmp->insn = insn; last_cmp->prev_clobber = last_clobber; last_cmp->in_a = XEXP (src, 0); last_cmp->in_b = XEXP (src, 1); last_cmp->eh_note = eh_note; last_cmp->orig_mode = src_mode; all_compares.safe_push (last_cmp); /* It's unusual, but be prepared for comparison patterns that also clobber an input, or perhaps a scratch. */ last_clobber = NULL; last_cmp_valid = true; } /* Notice if this instruction kills the flags register. */ else if (bitmap_bit_p (killed, targetm.flags_regnum)) { /* See if this insn could be the "clobber" that eliminates a future comparison. */ last_clobber = (arithmetic_flags_clobber_p (insn) ? insn : NULL); /* In either case, the previous compare is no longer valid. */ last_cmp = NULL; last_cmp_valid = false; continue; } /* Notice if this instruction uses the flags register. */ else if (last_cmp) find_flags_uses_in_insn (last_cmp, insn); /* Notice if any of the inputs to the comparison have changed. */ if (last_cmp_valid && (bitmap_bit_p (killed, REGNO (last_cmp->in_a)) || (REG_P (last_cmp->in_b) && bitmap_bit_p (killed, REGNO (last_cmp->in_b))))) last_cmp_valid = false; }