/* Update register liveness to reflect that X is now life (if rvalue is nonzero) or dead (if rvalue is zero) in INFO->x_block, and likewise Y in INFO->y_block. Return the number of registers the liveness of which changed in each block (as a negative number if registers became dead). */ static int note_local_live (struct equiv_info *info, rtx x, rtx y, int rvalue) { unsigned x_regno = REGNO (x); unsigned y_regno = REGNO (y); int x_nominal_nregs = (x_regno >= FIRST_PSEUDO_REGISTER ? 1 : hard_regno_nregs[x_regno][GET_MODE (x)]); int y_nominal_nregs = (y_regno >= FIRST_PSEUDO_REGISTER ? 1 : hard_regno_nregs[y_regno][GET_MODE (y)]); int x_change = assign_reg_reg_set (info->x_local_live, x, rvalue); int y_change = assign_reg_reg_set (info->y_local_live, y, rvalue); gcc_assert (x_nominal_nregs && y_nominal_nregs); gcc_assert (x_change * y_nominal_nregs == y_change * x_nominal_nregs); if (y_change) { if (reload_completed) { unsigned x_regno ATTRIBUTE_UNUSED = REGNO (x); unsigned y_regno = REGNO (y); enum machine_mode x_mode = GET_MODE (x); if (secondary_reload_class (0, REGNO_REG_CLASS (y_regno), x_mode, x) != NO_REGS #ifdef SECONDARY_MEMORY_NEEDED || SECONDARY_MEMORY_NEEDED (REGNO_REG_CLASS (y_regno), REGNO_REG_CLASS (x_regno), x_mode) #endif ) y_change *= IMPOSSIBLE_MOVE_FACTOR; } info->cur.input_count += y_change; info->cur.version++; } return x_change; }
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 void recompute_gain_for_pattern_seq (pattern_seq pseq) { matching_seq mseq; rtx x; int i; int hascall; HARD_REG_SET linkregs; /* Initialize data. */ SET_HARD_REG_SET (linkregs); pseq->link_reg = NULL_RTX; pseq->abstracted_length = 0; pseq->gain = -(seq_call_cost - seq_jump_cost + seq_return_cost); /* Determine ABSTRACTED_LENGTH and COST for matching sequences of PSEQ. ABSTRACTED_LENGTH may be less than MATCHING_LENGTH if sequences in the same block overlap. */ for (mseq = pseq->matching_seqs; mseq; mseq = mseq->next_matching_seq) { /* Determine ABSTRACTED_LENGTH. */ if (mseq->next_matching_seq) mseq->abstracted_length = (int)(mseq->next_matching_seq->idx - mseq->idx); else mseq->abstracted_length = mseq->matching_length; if (mseq->abstracted_length > mseq->matching_length) mseq->abstracted_length = mseq->matching_length; /* Compute the cost of sequence. */ RECOMPUTE_COST (mseq); /* If COST is big enough registers live in this matching sequence should not be used as a link register. Also set ABSTRACTED_LENGTH of PSEQ. */ if (mseq->cost > seq_call_cost) { clear_regs_live_in_seq (&linkregs, mseq->insn, mseq->abstracted_length); if (mseq->abstracted_length > pseq->abstracted_length) pseq->abstracted_length = mseq->abstracted_length; } } /* Modify ABSTRACTED_LENGTH of PSEQ if pattern sequence overlaps with one of the matching sequences. */ for (mseq = pseq->matching_seqs; mseq; mseq = mseq->next_matching_seq) { x = pseq->insn; for (i = 0; (i < pseq->abstracted_length) && (x != mseq->insn); i++) x = prev_insn_in_block (x); pseq->abstracted_length = i; } /* Compute the cost of pattern sequence. */ RECOMPUTE_COST (pseq); /* No gain if COST is too small. */ if (pseq->cost <= seq_call_cost) { pseq->gain = -1; return; } /* Ensure that no matching sequence is longer than the pattern sequence. */ for (mseq = pseq->matching_seqs; mseq; mseq = mseq->next_matching_seq) { if (mseq->abstracted_length > pseq->abstracted_length) { mseq->abstracted_length = pseq->abstracted_length; RECOMPUTE_COST (mseq); } /* Once the length is stabilizing the gain can be calculated. */ if (mseq->cost > seq_call_cost) pseq->gain += mseq->cost - seq_call_cost; } /* No need to do further work if there is no gain. */ if (pseq->gain <= 0) return; /* Should not use registers live in the pattern sequence as link register. */ clear_regs_live_in_seq (&linkregs, pseq->insn, pseq->abstracted_length); /* Determine whether pattern sequence contains a call_insn. */ hascall = 0; x = pseq->insn; for (i = 0; i < pseq->abstracted_length; i++) { if (CALL_P (x)) { hascall = 1; break; } x = prev_insn_in_block (x); } /* Should not use a register as a link register if - it is a fixed register, or - the sequence contains a call insn and the register is a call used register, or - the register needs to be saved if used in a function but was not used before (since saving it can invalidate already computed frame pointer offsets), or - the register cannot be used as a base register. */ for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) if (fixed_regs[i] #ifdef REGNO_OK_FOR_INDIRECT_JUMP_P || (!REGNO_OK_FOR_INDIRECT_JUMP_P (i, Pmode)) #else || (!ok_for_base_p_1 (i, Pmode, MEM, SCRATCH)) || (!reg_class_subset_p (REGNO_REG_CLASS (i), base_reg_class (VOIDmode, MEM, SCRATCH))) #endif || (hascall && call_used_regs[i]) || (!call_used_regs[i] && !df_regs_ever_live_p (i))) CLEAR_HARD_REG_BIT (linkregs, i); /* Find an appropriate register to be used as the link register. */ for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) if (TEST_HARD_REG_BIT (linkregs, i)) { pseq->link_reg = gen_rtx_REG (Pmode, i); break; } /* Abstraction is not possible if no link register is available, so set gain to 0. */ if (!pseq->link_reg) pseq->gain = 0; }
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; }
static int reload_cse_simplify_operands (rtx insn, rtx testreg) { int i, j; /* For each operand, all registers that are equivalent to it. */ HARD_REG_SET equiv_regs[MAX_RECOG_OPERANDS]; const char *constraints[MAX_RECOG_OPERANDS]; /* Vector recording how bad an alternative is. */ int *alternative_reject; /* Vector recording how many registers can be introduced by choosing this alternative. */ int *alternative_nregs; /* Array of vectors recording, for each operand and each alternative, which hard register to substitute, or -1 if the operand should be left as it is. */ int *op_alt_regno[MAX_RECOG_OPERANDS]; /* Array of alternatives, sorted in order of decreasing desirability. */ int *alternative_order; extract_insn (insn); if (recog_data.n_alternatives == 0 || recog_data.n_operands == 0) return 0; /* Figure out which alternative currently matches. */ if (! constrain_operands (1)) fatal_insn_not_found (insn); alternative_reject = XALLOCAVEC (int, recog_data.n_alternatives); alternative_nregs = XALLOCAVEC (int, recog_data.n_alternatives); alternative_order = XALLOCAVEC (int, recog_data.n_alternatives); memset (alternative_reject, 0, recog_data.n_alternatives * sizeof (int)); memset (alternative_nregs, 0, recog_data.n_alternatives * sizeof (int)); /* For each operand, find out which regs are equivalent. */ for (i = 0; i < recog_data.n_operands; i++) { cselib_val *v; struct elt_loc_list *l; rtx op; enum machine_mode mode; CLEAR_HARD_REG_SET (equiv_regs[i]); /* cselib blows up on CODE_LABELs. Trying to fix that doesn't seem right, so avoid the problem here. Likewise if we have a constant and the insn pattern doesn't tell us the mode we need. */ if (LABEL_P (recog_data.operand[i]) || (CONSTANT_P (recog_data.operand[i]) && recog_data.operand_mode[i] == VOIDmode)) continue; op = recog_data.operand[i]; mode = GET_MODE (op); #ifdef LOAD_EXTEND_OP if (MEM_P (op) && GET_MODE_BITSIZE (mode) < BITS_PER_WORD && LOAD_EXTEND_OP (mode) != UNKNOWN) { rtx set = single_set (insn); /* We might have multiple sets, some of which do implicit extension. Punt on this for now. */ if (! set) continue; /* If the destination is also a MEM or a STRICT_LOW_PART, no extension applies. Also, if there is an explicit extension, we don't have to worry about an implicit one. */ else if (MEM_P (SET_DEST (set)) || GET_CODE (SET_DEST (set)) == STRICT_LOW_PART || GET_CODE (SET_SRC (set)) == ZERO_EXTEND || GET_CODE (SET_SRC (set)) == SIGN_EXTEND) ; /* Continue ordinary processing. */ #ifdef CANNOT_CHANGE_MODE_CLASS /* If the register cannot change mode to word_mode, it follows that it cannot have been used in word_mode. */ else if (REG_P (SET_DEST (set)) && CANNOT_CHANGE_MODE_CLASS (GET_MODE (SET_DEST (set)), word_mode, REGNO_REG_CLASS (REGNO (SET_DEST (set))))) ; /* Continue ordinary processing. */ #endif /* If this is a straight load, make the extension explicit. */ else if (REG_P (SET_DEST (set)) && recog_data.n_operands == 2 && SET_SRC (set) == op && SET_DEST (set) == recog_data.operand[1-i]) { validate_change (insn, recog_data.operand_loc[i], gen_rtx_fmt_e (LOAD_EXTEND_OP (mode), word_mode, op), 1); validate_change (insn, recog_data.operand_loc[1-i], gen_rtx_REG (word_mode, REGNO (SET_DEST (set))), 1); if (! apply_change_group ()) return 0; return reload_cse_simplify_operands (insn, testreg); } else /* ??? There might be arithmetic operations with memory that are safe to optimize, but is it worth the trouble? */ continue; } #endif /* LOAD_EXTEND_OP */ v = cselib_lookup (op, recog_data.operand_mode[i], 0); if (! v) continue; for (l = v->locs; l; l = l->next) if (REG_P (l->loc)) SET_HARD_REG_BIT (equiv_regs[i], REGNO (l->loc)); } for (i = 0; i < recog_data.n_operands; i++) { enum machine_mode mode; int regno; const char *p; op_alt_regno[i] = XALLOCAVEC (int, recog_data.n_alternatives); for (j = 0; j < recog_data.n_alternatives; j++) op_alt_regno[i][j] = -1; p = constraints[i] = recog_data.constraints[i]; mode = recog_data.operand_mode[i]; /* Add the reject values for each alternative given by the constraints for this operand. */ j = 0; while (*p != '\0') { char c = *p++; if (c == ',') j++; else if (c == '?') alternative_reject[j] += 3; else if (c == '!') alternative_reject[j] += 300; } /* We won't change operands which are already registers. We also don't want to modify output operands. */ regno = true_regnum (recog_data.operand[i]); if (regno >= 0 || constraints[i][0] == '=' || constraints[i][0] == '+') continue; for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++) { int rclass = (int) NO_REGS; if (! TEST_HARD_REG_BIT (equiv_regs[i], regno)) continue; SET_REGNO (testreg, regno); PUT_MODE (testreg, mode); /* We found a register equal to this operand. Now look for all alternatives that can accept this register and have not been assigned a register they can use yet. */ j = 0; p = constraints[i]; for (;;) { char c = *p; switch (c) { case '=': case '+': case '?': case '#': case '&': case '!': case '*': case '%': case '0': case '1': case '2': case '3': case '4': case '5': case '6': case '7': case '8': case '9': case '<': case '>': case 'V': case 'o': case 'E': case 'F': case 'G': case 'H': case 's': case 'i': case 'n': case 'I': case 'J': case 'K': case 'L': case 'M': case 'N': case 'O': case 'P': case 'p': case 'X': case TARGET_MEM_CONSTRAINT: /* These don't say anything we care about. */ break; case 'g': case 'r': rclass = reg_class_subunion[(int) rclass][(int) GENERAL_REGS]; break; default: rclass = (reg_class_subunion [(int) rclass] [(int) REG_CLASS_FROM_CONSTRAINT ((unsigned char) c, p)]); break; case ',': case '\0': /* See if REGNO fits this alternative, and set it up as the replacement register if we don't have one for this alternative yet and the operand being replaced is not a cheap CONST_INT. */ if (op_alt_regno[i][j] == -1 && reg_fits_class_p (testreg, rclass, 0, mode) && (GET_CODE (recog_data.operand[i]) != CONST_INT || (rtx_cost (recog_data.operand[i], SET, optimize_bb_for_speed_p (BLOCK_FOR_INSN (insn))) > rtx_cost (testreg, SET, optimize_bb_for_speed_p (BLOCK_FOR_INSN (insn)))))) { alternative_nregs[j]++; op_alt_regno[i][j] = regno; } j++; rclass = (int) NO_REGS; break; } p += CONSTRAINT_LEN (c, p); if (c == '\0') break; } } } /* Record all alternatives which are better or equal to the currently matching one in the alternative_order array. */ for (i = j = 0; i < recog_data.n_alternatives; i++) if (alternative_reject[i] <= alternative_reject[which_alternative]) alternative_order[j++] = i; recog_data.n_alternatives = j; /* Sort it. Given a small number of alternatives, a dumb algorithm won't hurt too much. */ for (i = 0; i < recog_data.n_alternatives - 1; i++) { int best = i; int best_reject = alternative_reject[alternative_order[i]]; int best_nregs = alternative_nregs[alternative_order[i]]; int tmp; for (j = i + 1; j < recog_data.n_alternatives; j++) { int this_reject = alternative_reject[alternative_order[j]]; int this_nregs = alternative_nregs[alternative_order[j]]; if (this_reject < best_reject || (this_reject == best_reject && this_nregs > best_nregs)) { best = j; best_reject = this_reject; best_nregs = this_nregs; } } tmp = alternative_order[best]; alternative_order[best] = alternative_order[i]; alternative_order[i] = tmp; } /* Substitute the operands as determined by op_alt_regno for the best alternative. */ j = alternative_order[0]; for (i = 0; i < recog_data.n_operands; i++) { enum machine_mode mode = recog_data.operand_mode[i]; if (op_alt_regno[i][j] == -1) continue; validate_change (insn, recog_data.operand_loc[i], gen_rtx_REG (mode, op_alt_regno[i][j]), 1); } for (i = recog_data.n_dups - 1; i >= 0; i--) { int op = recog_data.dup_num[i]; enum machine_mode mode = recog_data.operand_mode[op]; if (op_alt_regno[op][j] == -1) continue; validate_change (insn, recog_data.dup_loc[i], gen_rtx_REG (mode, op_alt_regno[op][j]), 1); } return apply_change_group (); }
static int reload_cse_simplify_set (rtx set, rtx insn) { int did_change = 0; int dreg; rtx src; enum reg_class dclass; int old_cost; cselib_val *val; struct elt_loc_list *l; #ifdef LOAD_EXTEND_OP enum rtx_code extend_op = UNKNOWN; #endif bool speed = optimize_bb_for_speed_p (BLOCK_FOR_INSN (insn)); dreg = true_regnum (SET_DEST (set)); if (dreg < 0) return 0; src = SET_SRC (set); if (side_effects_p (src) || true_regnum (src) >= 0) return 0; dclass = REGNO_REG_CLASS (dreg); #ifdef LOAD_EXTEND_OP /* When replacing a memory with a register, we need to honor assumptions that combine made wrt the contents of sign bits. We'll do this by generating an extend instruction instead of a reg->reg copy. Thus the destination must be a register that we can widen. */ if (MEM_P (src) && GET_MODE_BITSIZE (GET_MODE (src)) < BITS_PER_WORD && (extend_op = LOAD_EXTEND_OP (GET_MODE (src))) != UNKNOWN && !REG_P (SET_DEST (set))) return 0; #endif val = cselib_lookup (src, GET_MODE (SET_DEST (set)), 0); if (! val) return 0; /* If memory loads are cheaper than register copies, don't change them. */ if (MEM_P (src)) old_cost = MEMORY_MOVE_COST (GET_MODE (src), dclass, 1); else if (REG_P (src)) old_cost = REGISTER_MOVE_COST (GET_MODE (src), REGNO_REG_CLASS (REGNO (src)), dclass); else old_cost = rtx_cost (src, SET, speed); for (l = val->locs; l; l = l->next) { rtx this_rtx = l->loc; int this_cost; if (CONSTANT_P (this_rtx) && ! references_value_p (this_rtx, 0)) { #ifdef LOAD_EXTEND_OP if (extend_op != UNKNOWN) { HOST_WIDE_INT this_val; /* ??? I'm lazy and don't wish to handle CONST_DOUBLE. Other constants, such as SYMBOL_REF, cannot be extended. */ if (GET_CODE (this_rtx) != CONST_INT) continue; this_val = INTVAL (this_rtx); switch (extend_op) { case ZERO_EXTEND: this_val &= GET_MODE_MASK (GET_MODE (src)); break; case SIGN_EXTEND: /* ??? In theory we're already extended. */ if (this_val == trunc_int_for_mode (this_val, GET_MODE (src))) break; default: gcc_unreachable (); } this_rtx = GEN_INT (this_val); } #endif this_cost = rtx_cost (this_rtx, SET, speed); } else if (REG_P (this_rtx)) { #ifdef LOAD_EXTEND_OP if (extend_op != UNKNOWN) { this_rtx = gen_rtx_fmt_e (extend_op, word_mode, this_rtx); this_cost = rtx_cost (this_rtx, SET, speed); } else #endif this_cost = REGISTER_MOVE_COST (GET_MODE (this_rtx), REGNO_REG_CLASS (REGNO (this_rtx)), dclass); } else continue; /* If equal costs, prefer registers over anything else. That tends to lead to smaller instructions on some machines. */ if (this_cost < old_cost || (this_cost == old_cost && REG_P (this_rtx) && !REG_P (SET_SRC (set)))) { #ifdef LOAD_EXTEND_OP if (GET_MODE_BITSIZE (GET_MODE (SET_DEST (set))) < BITS_PER_WORD && extend_op != UNKNOWN #ifdef CANNOT_CHANGE_MODE_CLASS && !CANNOT_CHANGE_MODE_CLASS (GET_MODE (SET_DEST (set)), word_mode, REGNO_REG_CLASS (REGNO (SET_DEST (set)))) #endif ) { rtx wide_dest = gen_rtx_REG (word_mode, REGNO (SET_DEST (set))); ORIGINAL_REGNO (wide_dest) = ORIGINAL_REGNO (SET_DEST (set)); validate_change (insn, &SET_DEST (set), wide_dest, 1); } #endif validate_unshare_change (insn, &SET_SRC (set), this_rtx, 1); old_cost = this_cost, did_change = 1; } } return did_change; }