/* Perform doloop optimizations Copyright (C) 2004, 2005, 2006, 2007, 2008, 2010 Free Software Foundation, Inc. Based on code by Michael P. Hayes (m.hayes@elec.canterbury.ac.nz) This file is part of GCC. GCC is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 3, or (at your option) any later version. GCC is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. You should have received a copy of the GNU General Public License along with GCC; see the file COPYING3. If not see . */ #include "config.h" #include "system.h" #include "coretypes.h" #include "tm.h" #include "rtl.h" #include "flags.h" #include "expr.h" #include "hard-reg-set.h" #include "basic-block.h" #include "diagnostic-core.h" #include "tm_p.h" #include "cfgloop.h" #include "params.h" #include "target.h" #include "dumpfile.h" /* This module is used to modify loops with a determinable number of iterations to use special low-overhead looping instructions. It first validates whether the loop is well behaved and has a determinable number of iterations (either at compile or run-time). It then modifies the loop to use a low-overhead looping pattern as follows: 1. A pseudo register is allocated as the loop iteration counter. 2. The number of loop iterations is calculated and is stored in the loop counter. 3. At the end of the loop, the jump insn is replaced by the doloop_end pattern. The compare must remain because it might be used elsewhere. If the loop-variable or condition register are used elsewhere, they will be eliminated by flow. 4. An optional doloop_begin pattern is inserted at the top of the loop. TODO The optimization should only performed when either the biv used for exit condition is unused at all except for the exit test, or if we do not have to change its value, since otherwise we have to add a new induction variable, which usually will not pay up (unless the cost of the doloop pattern is somehow extremely lower than the cost of compare & jump, or unless the bct register cannot be used for anything else but doloop -- ??? detect these cases). */ #ifdef HAVE_doloop_end /* Return the loop termination condition for PATTERN or zero if it is not a decrement and branch jump insn. */ rtx doloop_condition_get (rtx doloop_pat) { rtx cmp; rtx inc; rtx reg; rtx inc_src; rtx condition; rtx pattern; rtx cc_reg = NULL_RTX; rtx reg_orig = NULL_RTX; /* The canonical doloop pattern we expect has one of the following forms: 1) (parallel [(set (pc) (if_then_else (condition) (label_ref (label)) (pc))) (set (reg) (plus (reg) (const_int -1))) (additional clobbers and uses)]) The branch must be the first entry of the parallel (also required by jump.c), and the second entry of the parallel must be a set of the loop counter register. Some targets (IA-64) wrap the set of the loop counter in an if_then_else too. 2) (set (reg) (plus (reg) (const_int -1)) (set (pc) (if_then_else (reg != 0) (label_ref (label)) (pc))). Some targets (ARM) do the comparison before the branch, as in the following form: 3) (parallel [(set (cc) (compare ((plus (reg) (const_int -1), 0))) (set (reg) (plus (reg) (const_int -1)))]) (set (pc) (if_then_else (cc == NE) (label_ref (label)) (pc))) */ pattern = PATTERN (doloop_pat); if (GET_CODE (pattern) != PARALLEL) { rtx cond; rtx prev_insn = prev_nondebug_insn (doloop_pat); rtx cmp_arg1, cmp_arg2; rtx cmp_orig; /* In case the pattern is not PARALLEL we expect two forms of doloop which are cases 2) and 3) above: in case 2) the decrement immediately precedes the branch, while in case 3) the compare and decrement instructions immediately precede the branch. */ if (prev_insn == NULL_RTX || !INSN_P (prev_insn)) return 0; cmp = pattern; if (GET_CODE (PATTERN (prev_insn)) == PARALLEL) { /* The third case: the compare and decrement instructions immediately precede the branch. */ cmp_orig = XVECEXP (PATTERN (prev_insn), 0, 0); if (GET_CODE (cmp_orig) != SET) return 0; if (GET_CODE (SET_SRC (cmp_orig)) != COMPARE) return 0; cmp_arg1 = XEXP (SET_SRC (cmp_orig), 0); cmp_arg2 = XEXP (SET_SRC (cmp_orig), 1); if (cmp_arg2 != const0_rtx || GET_CODE (cmp_arg1) != PLUS) return 0; reg_orig = XEXP (cmp_arg1, 0); if (XEXP (cmp_arg1, 1) != GEN_INT (-1) || !REG_P (reg_orig)) return 0; cc_reg = SET_DEST (cmp_orig); inc = XVECEXP (PATTERN (prev_insn), 0, 1); } else inc = PATTERN (prev_insn); /* We expect the condition to be of the form (reg != 0) */ cond = XEXP (SET_SRC (cmp), 0); if (GET_CODE (cond) != NE || XEXP (cond, 1) != const0_rtx) return 0; } else { cmp = XVECEXP (pattern, 0, 0); inc = XVECEXP (pattern, 0, 1); } /* Check for (set (reg) (something)). */ if (GET_CODE (inc) != SET) return 0; reg = SET_DEST (inc); if (! REG_P (reg)) return 0; /* Check if something = (plus (reg) (const_int -1)). On IA-64, this decrement is wrapped in an if_then_else. */ inc_src = SET_SRC (inc); if (GET_CODE (inc_src) == IF_THEN_ELSE) inc_src = XEXP (inc_src, 1); if (GET_CODE (inc_src) != PLUS || XEXP (inc_src, 0) != reg || XEXP (inc_src, 1) != constm1_rtx) return 0; /* Check for (set (pc) (if_then_else (condition) (label_ref (label)) (pc))). */ if (GET_CODE (cmp) != SET || SET_DEST (cmp) != pc_rtx || GET_CODE (SET_SRC (cmp)) != IF_THEN_ELSE || GET_CODE (XEXP (SET_SRC (cmp), 1)) != LABEL_REF || XEXP (SET_SRC (cmp), 2) != pc_rtx) return 0; /* Extract loop termination condition. */ condition = XEXP (SET_SRC (cmp), 0); /* We expect a GE or NE comparison with 0 or 1. */ if ((GET_CODE (condition) != GE && GET_CODE (condition) != NE) || (XEXP (condition, 1) != const0_rtx && XEXP (condition, 1) != const1_rtx)) return 0; if ((XEXP (condition, 0) == reg) /* For the third case: */ || ((cc_reg != NULL_RTX) && (XEXP (condition, 0) == cc_reg) && (reg_orig == reg)) || (GET_CODE (XEXP (condition, 0)) == PLUS && XEXP (XEXP (condition, 0), 0) == reg)) { if (GET_CODE (pattern) != PARALLEL) /* For the second form we expect: (set (reg) (plus (reg) (const_int -1)) (set (pc) (if_then_else (reg != 0) (label_ref (label)) (pc))). is equivalent to the following: (parallel [(set (pc) (if_then_else (reg != 1) (label_ref (label)) (pc))) (set (reg) (plus (reg) (const_int -1))) (additional clobbers and uses)]) For the third form we expect: (parallel [(set (cc) (compare ((plus (reg) (const_int -1)), 0)) (set (reg) (plus (reg) (const_int -1)))]) (set (pc) (if_then_else (cc == NE) (label_ref (label)) (pc))) which is equivalent to the following: (parallel [(set (cc) (compare (reg, 1)) (set (reg) (plus (reg) (const_int -1))) (set (pc) (if_then_else (NE == cc) (label_ref (label)) (pc))))]) So we return the second form instead for the two cases. */ condition = gen_rtx_fmt_ee (NE, VOIDmode, inc_src, const1_rtx); return condition; } /* ??? If a machine uses a funny comparison, we could return a canonicalized form here. */ return 0; } /* Return nonzero if the loop specified by LOOP is suitable for the use of special low-overhead looping instructions. DESC describes the number of iterations of the loop. */ static bool doloop_valid_p (struct loop *loop, struct niter_desc *desc) { basic_block *body = get_loop_body (loop), bb; rtx insn; unsigned i; bool result = true; /* Check for loops that may not terminate under special conditions. */ if (!desc->simple_p || desc->assumptions || desc->infinite) { /* There are some cases that would require a special attention. For example if the comparison is LEU and the comparison value is UINT_MAX then the loop will not terminate. Similarly, if the comparison code is GEU and the comparison value is 0, the loop will not terminate. If the absolute increment is not 1, the loop can be infinite even with LTU/GTU, e.g. for (i = 3; i > 0; i -= 2) ??? We could compute these conditions at run-time and have a additional jump around the loop to ensure an infinite loop. However, it is very unlikely that this is the intended behavior of the loop and checking for these rare boundary conditions would pessimize all other code. If the loop is executed only a few times an extra check to restart the loop could use up most of the benefits of using a count register loop. Note however, that normally, this restart branch would never execute, so it could be predicted well by the CPU. We should generate the pessimistic code by default, and have an option, e.g. -funsafe-loops that would enable count-register loops in this case. */ if (dump_file) fprintf (dump_file, "Doloop: Possible infinite iteration case.\n"); result = false; goto cleanup; } for (i = 0; i < loop->num_nodes; i++) { bb = body[i]; for (insn = BB_HEAD (bb); insn != NEXT_INSN (BB_END (bb)); insn = NEXT_INSN (insn)) { /* Different targets have different necessities for low-overhead looping. Call the back end for each instruction within the loop to let it decide whether the insn prohibits a low-overhead loop. It will then return the cause for it to emit to the dump file. */ const char * invalid = targetm.invalid_within_doloop (insn); if (invalid) { if (dump_file) fprintf (dump_file, "Doloop: %s\n", invalid); result = false; goto cleanup; } } } result = true; cleanup: free (body); return result; } /* Adds test of COND jumping to DEST on edge *E and set *E to the new fallthru edge. If the condition is always false, do not do anything. If it is always true, redirect E to DEST and return false. In all other cases, true is returned. */ static bool add_test (rtx cond, edge *e, 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); basic_block bb; 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, -1); jump = get_last_insn (); if (!jump || !JUMP_P (jump)) { /* The condition is always false and the jump was optimized out. */ end_sequence (); return true; } seq = get_insns (); end_sequence (); /* There always is at least the jump insn in the sequence. */ gcc_assert (seq != NULL_RTX); bb = split_edge_and_insert (*e, seq); *e = single_succ_edge (bb); if (any_uncondjump_p (jump)) { /* The condition is always true. */ delete_insn (jump); redirect_edge_and_branch_force (*e, dest); return false; } JUMP_LABEL (jump) = label; /* The jump is supposed to handle an unlikely special case. */ add_reg_note (jump, REG_BR_PROB, const0_rtx); LABEL_NUSES (label)++; make_edge (bb, dest, (*e)->flags & ~EDGE_FALLTHRU); return true; } /* Modify the loop to use the low-overhead looping insn where LOOP describes the loop, DESC describes the number of iterations of the loop, and DOLOOP_INSN is the low-overhead looping insn to emit at the end of the loop. CONDITION is the condition separated from the DOLOOP_SEQ. COUNT is the number of iterations of the LOOP. */ static void doloop_modify (struct loop *loop, struct niter_desc *desc, rtx doloop_seq, rtx condition, rtx count) { rtx counter_reg; rtx tmp, noloop = NULL_RTX; rtx sequence; rtx jump_insn; rtx jump_label; int nonneg = 0; bool increment_count; basic_block loop_end = desc->out_edge->src; enum machine_mode mode; rtx true_prob_val; double_int iterations; jump_insn = BB_END (loop_end); if (dump_file) { fprintf (dump_file, "Doloop: Inserting doloop pattern ("); if (desc->const_iter) fprintf (dump_file, HOST_WIDEST_INT_PRINT_DEC, desc->niter); else fputs ("runtime", dump_file); fputs (" iterations).\n", dump_file); } /* Get the probability of the original branch. If it exists we would need to update REG_BR_PROB of the new jump_insn. */ true_prob_val = find_reg_note (jump_insn, REG_BR_PROB, NULL_RTX); /* Discard original jump to continue loop. The original compare result may still be live, so it cannot be discarded explicitly. */ delete_insn (jump_insn); counter_reg = XEXP (condition, 0); if (GET_CODE (counter_reg) == PLUS) counter_reg = XEXP (counter_reg, 0); mode = GET_MODE (counter_reg); increment_count = false; switch (GET_CODE (condition)) { case NE: /* Currently only NE tests against zero and one are supported. */ noloop = XEXP (condition, 1); if (noloop != const0_rtx) { gcc_assert (noloop == const1_rtx); increment_count = true; } break; case GE: /* Currently only GE tests against zero are supported. */ gcc_assert (XEXP (condition, 1) == const0_rtx); noloop = constm1_rtx; /* The iteration count does not need incrementing for a GE test. */ increment_count = false; /* Determine if the iteration counter will be non-negative. Note that the maximum value loaded is iterations_max - 1. */ if (max_loop_iterations (loop, &iterations) && (iterations.ule (double_int_one.llshift (GET_MODE_PRECISION (mode) - 1, GET_MODE_PRECISION (mode))))) nonneg = 1; break; /* Abort if an invalid doloop pattern has been generated. */ default: gcc_unreachable (); } if (increment_count) count = simplify_gen_binary (PLUS, mode, count, const1_rtx); /* Insert initialization of the count register into the loop header. */ start_sequence (); tmp = force_operand (count, counter_reg); convert_move (counter_reg, tmp, 1); sequence = get_insns (); end_sequence (); emit_insn_after (sequence, BB_END (loop_preheader_edge (loop)->src)); if (desc->noloop_assumptions) { rtx ass = copy_rtx (desc->noloop_assumptions); basic_block preheader = loop_preheader_edge (loop)->src; basic_block set_zero = split_edge (loop_preheader_edge (loop)); basic_block new_preheader = split_edge (loop_preheader_edge (loop)); edge te; /* Expand the condition testing the assumptions and if it does not pass, reset the count register to 0. */ redirect_edge_and_branch_force (single_succ_edge (preheader), new_preheader); set_immediate_dominator (CDI_DOMINATORS, new_preheader, preheader); set_zero->count = 0; set_zero->frequency = 0; te = single_succ_edge (preheader); for (; ass; ass = XEXP (ass, 1)) if (!add_test (XEXP (ass, 0), &te, set_zero)) break; if (ass) { /* We reached a condition that is always true. This is very hard to reproduce (such a loop does not roll, and thus it would most likely get optimized out by some of the preceding optimizations). In fact, I do not have any testcase for it. However, it would also be very hard to show that it is impossible, so we must handle this case. */ set_zero->count = preheader->count; set_zero->frequency = preheader->frequency; } if (EDGE_COUNT (set_zero->preds) == 0) { /* All the conditions were simplified to false, remove the unreachable set_zero block. */ delete_basic_block (set_zero); } else { /* Reset the counter to zero in the set_zero block. */ start_sequence (); convert_move (counter_reg, noloop, 0); sequence = get_insns (); end_sequence (); emit_insn_after (sequence, BB_END (set_zero)); set_immediate_dominator (CDI_DOMINATORS, set_zero, recompute_dominator (CDI_DOMINATORS, set_zero)); } set_immediate_dominator (CDI_DOMINATORS, new_preheader, recompute_dominator (CDI_DOMINATORS, new_preheader)); } /* Some targets (eg, C4x) need to initialize special looping registers. */ #ifdef HAVE_doloop_begin { rtx init; unsigned level = get_loop_level (loop) + 1; double_int iter; rtx iter_rtx; if (!max_loop_iterations (loop, &iter) || !iter.fits_shwi ()) iter_rtx = const0_rtx; else iter_rtx = GEN_INT (iter.to_shwi()); init = gen_doloop_begin (counter_reg, desc->const_iter ? desc->niter_expr : const0_rtx, iter_rtx, GEN_INT (level)); if (init) { start_sequence (); emit_insn (init); sequence = get_insns (); end_sequence (); emit_insn_after (sequence, BB_END (loop_preheader_edge (loop)->src)); } } #endif /* Insert the new low-overhead looping insn. */ emit_jump_insn_after (doloop_seq, BB_END (loop_end)); jump_insn = BB_END (loop_end); jump_label = block_label (desc->in_edge->dest); JUMP_LABEL (jump_insn) = jump_label; LABEL_NUSES (jump_label)++; /* Ensure the right fallthru edge is marked, for case we have reversed the condition. */ desc->in_edge->flags &= ~EDGE_FALLTHRU; desc->out_edge->flags |= EDGE_FALLTHRU; /* Add a REG_NONNEG note if the actual or estimated maximum number of iterations is non-negative. */ if (nonneg) add_reg_note (jump_insn, REG_NONNEG, NULL_RTX); /* Update the REG_BR_PROB note. */ if (true_prob_val) { /* Seems safer to use the branch probability. */ add_reg_note (jump_insn, REG_BR_PROB, GEN_INT (desc->in_edge->probability)); } } /* Process loop described by LOOP validating that the loop is suitable for conversion to use a low overhead looping instruction, replacing the jump insn where suitable. Returns true if the loop was successfully modified. */ static bool doloop_optimize (struct loop *loop) { enum machine_mode mode; rtx doloop_seq, doloop_pat, doloop_reg; rtx iterations, count; rtx iterations_max; rtx start_label; rtx condition; unsigned level, est_niter; int max_cost; struct niter_desc *desc; unsigned word_mode_size; unsigned HOST_WIDE_INT word_mode_max; double_int iter; if (dump_file) fprintf (dump_file, "Doloop: Processing loop %d.\n", loop->num); iv_analysis_loop_init (loop); /* Find the simple exit of a LOOP. */ desc = get_simple_loop_desc (loop); /* Check that loop is a candidate for a low-overhead looping insn. */ if (!doloop_valid_p (loop, desc)) { if (dump_file) fprintf (dump_file, "Doloop: The loop is not suitable.\n"); return false; } mode = desc->mode; est_niter = 3; if (desc->const_iter) est_niter = desc->niter; /* If the estimate on number of iterations is reliable (comes from profile feedback), use it. Do not use it normally, since the expected number of iterations of an unrolled loop is 2. */ if (loop->header->count) est_niter = expected_loop_iterations (loop); if (est_niter < 3) { if (dump_file) fprintf (dump_file, "Doloop: Too few iterations (%u) to be profitable.\n", est_niter); return false; } max_cost = COSTS_N_INSNS (PARAM_VALUE (PARAM_MAX_ITERATIONS_COMPUTATION_COST)); if (set_src_cost (desc->niter_expr, optimize_loop_for_speed_p (loop)) > max_cost) { if (dump_file) fprintf (dump_file, "Doloop: number of iterations too costly to compute.\n"); return false; } count = copy_rtx (desc->niter_expr); iterations = desc->const_iter ? desc->niter_expr : const0_rtx; if (!max_loop_iterations (loop, &iter) || !iter.fits_shwi ()) iterations_max = const0_rtx; else iterations_max = GEN_INT (iter.to_shwi()); level = get_loop_level (loop) + 1; /* Generate looping insn. If the pattern FAILs then give up trying to modify the loop since there is some aspect the back-end does not like. */ start_label = block_label (desc->in_edge->dest); doloop_reg = gen_reg_rtx (mode); doloop_seq = gen_doloop_end (doloop_reg, iterations, iterations_max, GEN_INT (level), start_label); word_mode_size = GET_MODE_PRECISION (word_mode); word_mode_max = ((unsigned HOST_WIDE_INT) 1 << (word_mode_size - 1) << 1) - 1; if (! doloop_seq && mode != word_mode /* Before trying mode different from the one in that # of iterations is computed, we must be sure that the number of iterations fits into the new mode. */ && (word_mode_size >= GET_MODE_PRECISION (mode) || iter.ule (double_int::from_shwi (word_mode_max)))) { if (word_mode_size > GET_MODE_PRECISION (mode)) { count = simplify_gen_unary (ZERO_EXTEND, word_mode, count, mode); iterations = simplify_gen_unary (ZERO_EXTEND, word_mode, iterations, mode); iterations_max = simplify_gen_unary (ZERO_EXTEND, word_mode, iterations_max, mode); } else { count = lowpart_subreg (word_mode, count, mode); iterations = lowpart_subreg (word_mode, iterations, mode); iterations_max = lowpart_subreg (word_mode, iterations_max, mode); } PUT_MODE (doloop_reg, word_mode); doloop_seq = gen_doloop_end (doloop_reg, iterations, iterations_max, GEN_INT (level), start_label); } if (! doloop_seq) { if (dump_file) fprintf (dump_file, "Doloop: Target unwilling to use doloop pattern!\n"); return false; } /* If multiple instructions were created, the last must be the jump instruction. Also, a raw define_insn may yield a plain pattern. */ doloop_pat = doloop_seq; if (INSN_P (doloop_pat)) { while (NEXT_INSN (doloop_pat) != NULL_RTX) doloop_pat = NEXT_INSN (doloop_pat); if (!JUMP_P (doloop_pat)) doloop_pat = NULL_RTX; } if (! doloop_pat || ! (condition = doloop_condition_get (doloop_pat))) { if (dump_file) fprintf (dump_file, "Doloop: Unrecognizable doloop pattern!\n"); return false; } doloop_modify (loop, desc, doloop_seq, condition, count); return true; } /* This is the main entry point. Process all loops using doloop_optimize. */ void doloop_optimize_loops (void) { loop_iterator li; struct loop *loop; FOR_EACH_LOOP (li, loop, 0) { doloop_optimize (loop); } iv_analysis_done (); #ifdef ENABLE_CHECKING verify_loop_structure (); #endif } #endif /* HAVE_doloop_end */