X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Freassoc.c;h=8cfcd8c2ecfed349f15eb90771454c6c2b84fe34;hb=bbc6291b6a4ed4fe5bd8d1c2ddd57f4e6a0c8a74;hp=9f16ce889f126fe9822fdef77d7cad8e9355b007;hpb=0952881a2714f8ec2e87a6f4829bd55f41a0c68d;p=libfirm diff --git a/ir/opt/reassoc.c b/ir/opt/reassoc.c index 9f16ce889..8cfcd8c2e 100644 --- a/ir/opt/reassoc.c +++ b/ir/opt/reassoc.c @@ -1,36 +1,52 @@ /* - * Project: libFIRM - * File name: ir/opt/reassoc.c - * Purpose: Reassociation - * Author: Michael Beck - * Created: - * CVS-ID: $Id$ - * Copyright: (c) 1998-2007 Universität Karlsruhe - * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE. + * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved. + * + * This file is part of libFirm. + * + * This file may be distributed and/or modified under the terms of the + * GNU General Public License version 2 as published by the Free Software + * Foundation and appearing in the file LICENSE.GPL included in the + * packaging of this file. + * + * Licensees holding valid libFirm Professional Edition licenses may use + * this file in accordance with the libFirm Commercial License. + * Agreement provided with the Software. + * + * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE + * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR + * PURPOSE. + */ + +/** + * @file + * @brief Reassociation + * @author Michael Beck + * @version $Id$ */ #ifdef HAVE_CONFIG_H #include "config.h" #endif +#include "iropt_t.h" #include "irnode_t.h" #include "irgraph_t.h" #include "irmode_t.h" -#include "iropt_t.h" #include "ircons_t.h" #include "irgmod.h" -#include "dbginfo.h" #include "iropt_dbg.h" #include "irflag_t.h" #include "irgwalk.h" #include "reassoc_t.h" #include "irhooks.h" #include "irloop.h" +#include "pdeq.h" #include "debug.h" DEBUG_ONLY(static firm_dbg_module_t *dbg;) typedef struct _walker_t { - int changes; /* set, if a reassociation take place */ + int changes; /**< set, if a reassociation take place */ + waitq *wq; /**< a wait queue */ } walker_t; typedef enum { @@ -113,7 +129,7 @@ static void get_comm_Binop_ops(ir_node *binop, ir_node **a, ir_node **c) } /* get_comm_Binop_ops */ /** - * reassociate a Sub: x - c = (-c) + x + * reassociate a Sub: x - c = x + (-c) */ static int reassoc_Sub(ir_node **in) { @@ -129,16 +145,13 @@ static int reassoc_Sub(ir_node **in) block = get_nodes_block(n); /* handles rule R6: - * convert x - c => (-c) + x - * - * As there is NO real Minus in Firm it makes no sense to do this - * for non-real constants yet. + * convert x - c => x + (-c) */ if (get_const_class(right, block) == REAL_CONSTANT) { ir_node *left = get_Sub_left(n); ir_mode *mode; dbg_info *dbi; - ir_node *irn, *c; + ir_node *irn; switch (get_const_class(left, block)) { case REAL_CONSTANT: @@ -155,17 +168,19 @@ static int reassoc_Sub(ir_node **in) /* already constant, nothing to do */ return 0; } + mode = get_irn_mode(n); dbi = get_irn_dbg_info(n); /* Beware of SubP(P, Is) */ - c = new_r_Const(current_ir_graph, block, rmode, get_mode_null(rmode)); - irn = new_rd_Sub(dbi, current_ir_graph, block, c, right, rmode); - - irn = new_rd_Add(dbi, current_ir_graph, block, left, irn, get_irn_mode(n)); + irn = new_rd_Minus(dbi, current_ir_graph, block, right, rmode); + irn = new_rd_Add(dbi, current_ir_graph, block, left, irn, mode); DBG((dbg, LEVEL_5, "Applied: %n - %n => %n + (-%n)\n", - get_Sub_left(n), c, get_Sub_left(n), c)); + get_Sub_left(n), right, get_Sub_left(n), right)); + + if (n == irn) + return 0; exchange(n, irn); *in = irn; @@ -235,9 +250,9 @@ static int reassoc_commutative(ir_node **node) return 0; } - if ((c_c1 != NO_CONSTANT) & (c_c2 != NO_CONSTANT)) { + if ((c_c1 != NO_CONSTANT) /* & (c_c2 != NO_CONSTANT) */) { /* handles rules R7, R8, R9, R10: - * convert c1 .OP. (c2 .OP. x) => (c1 .OP. c2) .OP. x + * convert c1 .OP. (c2 .OP. x) => x .OP. (c1 .OP. c2) */ ir_node *irn, *in[2]; ir_mode *mode, *mode_c1 = get_irn_mode(c1), *mode_c2 = get_irn_mode(c2); @@ -265,16 +280,16 @@ static int reassoc_commutative(ir_node **node) in[0] = c1; in[1] = c2; - mode = get_mode_from_ops(in[0], in[1]); - in[0] = optimize_node(new_ir_node(NULL, current_ir_graph, block, op, mode, 2, in)); - in[1] = t2; + mode = get_mode_from_ops(in[0], in[1]); + in[1] = optimize_node(new_ir_node(NULL, current_ir_graph, block, op, mode, 2, in)); + in[0] = t2; mode = get_mode_from_ops(in[0], in[1]); irn = optimize_node(new_ir_node(NULL, current_ir_graph, block, op, mode, 2, in)); - DBG((dbg, LEVEL_5, "Applied: %n .%s. (%n .%s. %n) => (%n .%s. %n) .%s. %n\n", - c1, get_irn_opname(n), c2, get_irn_opname(n), - t2, c1, get_irn_opname(n), c2, get_irn_opname(n), t2)); + DBG((dbg, LEVEL_5, "Applied: %n .%s. (%n .%s. %n) => %n .%s. (%n .%s. %n)\n", + c1, get_irn_opname(n), c2, get_irn_opname(n), t2, + t2, get_irn_opname(n), c1, get_irn_opname(n), c2)); /* * In some rare cases it can really happen that we get the same node back. * This might be happen in dead loops, were the Phi nodes are already gone away. @@ -296,7 +311,7 @@ static int reassoc_commutative(ir_node **node) #define reassoc_Eor reassoc_commutative /** - * reassociate using distributive law for Mul and Add/Sub + * Reassociate using commutative law for Mul and distributive law for Mul and Add/Sub: */ static int reassoc_Mul(ir_node **node) { @@ -324,7 +339,6 @@ static int reassoc_Mul(ir_node **node) in[0] = new_rd_Mul(NULL, current_ir_graph, block, c, t1, mode); in[1] = new_rd_Mul(NULL, current_ir_graph, block, c, t2, mode); - mode = get_mode_from_ops(in[0], in[1]); irn = optimize_node(new_ir_node(NULL, current_ir_graph, block, op, mode, 2, in)); /* In some cases it might happen that the new irn is equal the old one, for @@ -349,14 +363,48 @@ static int reassoc_Mul(ir_node **node) return 0; } /* reassoc_Mul */ +/** + * Reassociate Shl. We transform Shl(x, const) into Mul's if possible. + */ +static int reassoc_Shl(ir_node **node) { + ir_node *n = *node; + ir_node *c = get_Shl_right(n); + ir_node *x, *blk, *irn; + ir_mode *mode; + tarval *tv; + + if (! is_Const(c)) + return 0; + + x = get_Shl_left(n); + mode = get_irn_mode(x); + + tv = get_mode_one(mode); + tv = tarval_shl(tv, get_Const_tarval(c)); + + if (tv == tarval_bad) + return 0; + + blk = get_nodes_block(n); + c = new_r_Const(current_ir_graph, blk, mode, tv); + irn = new_rd_Mul(get_irn_dbg_info(n), current_ir_graph, blk, x, c, mode); + + if (irn != n) { + exchange(n, irn); + *node = irn; + return 1; + } + return 0; +} /* reassoc_Shl */ + /** * The walker for the reassociation. */ -static void do_reassociation(ir_node *n, void *env) +static void wq_walker(ir_node *n, void *env) { walker_t *wenv = env; - int res; + set_irn_link(n, NULL); if (is_no_Block(n)) { ir_node *blk = get_nodes_block(n); @@ -366,10 +414,37 @@ static void do_reassociation(ir_node *n, void *env) which or cf_opt do not guarantee yet. */ return; } + waitq_put(wenv->wq, n); + set_irn_link(n, wenv->wq); + } +} /* wq_walker */ + +/** + * The walker for the reassociation. + */ +static void do_reassociation(walker_t *wenv) +{ + int i, res, changed; + ir_node *n, *blk; + + + while (! waitq_empty(wenv->wq)) { + n = waitq_get(wenv->wq); + set_irn_link(n, NULL); + + blk = get_nodes_block(n); + if (is_Block_dead(blk) || get_Block_dom_depth(blk) < 0) { + /* We are in a dead block, do not optimize or we may fall into an endless + loop. We check this here instead of requiring that all dead blocks are removed + which or cf_opt do not guarantee yet. */ + continue; + } + hook_reassociate(1); /* reassociation must run until a fixpoint is reached. */ + changed = 0; do { ir_op *op = get_irn_op(n); ir_mode *mode = get_irn_mode(n); @@ -383,14 +458,285 @@ static void do_reassociation(ir_node *n, void *env) if (op->ops.reassociate) { res = op->ops.reassociate(&n); - wenv->changes |= res; + changed |= res; } } while (res == 1); - hook_reassociate(0); + + wenv->changes |= changed; + + if (changed) { + for (i = get_irn_arity(n) - 1; i >= 0; --i) { + ir_node *pred = get_irn_n(n, i); + + if (get_irn_link(pred) != wenv->wq) { + waitq_put(wenv->wq, pred); + set_irn_link(pred, wenv->wq); + } + } + } } } /* do_reassociation */ +/** + * Returns the earliest were a,b are available. + * Note that we know that a, b both dominate + * the block of the previous operation, so one must dominate the other. + * + * If the earliest block is the start block, return curr_blk instead + */ +static ir_node *earliest_block(ir_node *a, ir_node *b, ir_node *curr_blk) { + ir_node *blk_a = get_nodes_block(a); + ir_node *blk_b = get_nodes_block(b); + ir_node *res; + + /* if blk_a != blk_b, one must dominate the other */ + if (block_dominates(blk_a, blk_b)) + res = blk_b; + else + res = blk_a; + if (res == get_irg_start_block(current_ir_graph)) + return curr_blk; + return res; +} /* earliest_block */ + +/** + * Checks whether a node is a Constant expression. + * The following trees are constant expressions: + * + * Const, SymConst, Const + SymConst + * + * Handling SymConsts as const might be not a good idea for all + * architectures ... + */ +static int is_constant_expr(ir_node *irn) { + ir_op *op; + + switch (get_irn_opcode(irn)) { + case iro_Const: + case iro_SymConst: + return 1; + case iro_Add: + op = get_irn_op(get_Add_left(irn)); + if (op != op_Const && op != op_SymConst) + return 0; + op = get_irn_op(get_Add_right(irn)); + if (op != op_Const && op != op_SymConst) + return 0; + return 1; + default: + return 0; + } +} /* is_constant_expr */ + +/** + * Apply distributive Law for Mul and Add/Sub + */ +static int reverse_rule_distributive(ir_node **node) { + ir_node *n = *node; + ir_node *left = get_binop_left(n); + ir_node *right = get_binop_right(n); + ir_node *x, *blk, *curr_blk; + ir_node *a, *b, *irn; + ir_op *op; + ir_mode *mode; + dbg_info *dbg; + + op = get_irn_op(left); + if (op != get_irn_op(right)) + return 0; + + if (op == op_Shl) { + x = get_Shl_right(left); + + if (x == get_Shl_right(right)) { + /* (a << x) +/- (b << x) ==> (a +/- b) << x */ + a = get_Shl_left(left); + b = get_Shl_left(right); + goto transform; + } + } else if (op == op_Mul) { + x = get_Mul_left(left); + + if (x == get_Mul_left(right)) { + /* (x * a) +/- (x * b) ==> (a +/- b) * x */ + a = get_Mul_right(left); + b = get_Mul_right(right); + goto transform; + } else if (x == get_Mul_right(right)) { + /* (x * a) +/- (b * x) ==> (a +/- b) * x */ + a = get_Mul_right(left); + b = get_Mul_left(right); + goto transform; + } + + x = get_Mul_right(left); + + if (x == get_Mul_right(right)) { + /* (a * x) +/- (b * x) ==> (a +/- b) * x */ + a = get_Mul_left(left); + b = get_Mul_left(right); + goto transform; + } else if (x == get_Mul_left(right)) { + /* (a * x) +/- (x * b) ==> (a +/- b) * x */ + a = get_Mul_left(left); + b = get_Mul_right(right); + goto transform; + } + } + return 0; + +transform: + curr_blk = get_nodes_block(n); + + blk = earliest_block(a, b, curr_blk); + + dbg = get_irn_dbg_info(n); + mode = get_irn_mode(n); + + if (is_Add(n)) + irn = new_rd_Add(dbg, current_ir_graph, blk, a, b, mode); + else + irn = new_rd_Sub(dbg, current_ir_graph, blk, a, b, mode); + + blk = earliest_block(irn, x, curr_blk); + + if (op == op_Mul) + irn = new_rd_Mul(dbg, current_ir_graph, blk, irn, x, mode); + else + irn = new_rd_Shl(dbg, current_ir_graph, blk, irn, x, mode); + + exchange(n, irn); + *node = irn; + return 1; +} /* reverse_rule_distributive */ + +/** + * Move Constants towards the root. + */ +static int move_consts_up(ir_node **node) { + ir_node *n = *node; + ir_op *op; + ir_node *l, *r, *a, *b, *c, *blk, *irn, *in[2]; + ir_mode *mode, *ma, *mb; + dbg_info *dbg; + + l = get_binop_left(n); + r = get_binop_right(n); + + /* check if one is already a constant expression */ + if (is_constant_expr(l) || is_constant_expr(r)) + return 0; + + dbg = get_irn_dbg_info(n); + op = get_irn_op(n); + if (get_irn_op(l) == op) { + /* (a .op. b) .op. r */ + a = get_binop_left(l); + b = get_binop_right(l); + + if (is_constant_expr(a)) { + /* (C .op. b) .op. r ==> (r .op. b) .op. C */ + c = a; + a = r; + blk = get_nodes_block(l); + dbg = dbg == get_irn_dbg_info(l) ? dbg : NULL; + goto transform; + } else if (is_constant_expr(b)) { + /* (a .op. C) .op. r ==> (a .op. r) .op. C */ + c = b; + b = r; + blk = get_nodes_block(l); + dbg = dbg == get_irn_dbg_info(l) ? dbg : NULL; + goto transform; + } + } else if (get_irn_op(r) == op) { + /* l .op. (a .op. b) */ + a = get_binop_left(r); + b = get_binop_right(r); + + if (is_constant_expr(a)) { + /* l .op. (C .op. b) ==> (l .op. b) .op. C */ + c = a; + a = l; + blk = get_nodes_block(r); + dbg = dbg == get_irn_dbg_info(r) ? dbg : NULL; + goto transform; + } else if (is_constant_expr(b)) { + /* l .op. (a .op. C) ==> (a .op. l) .op. C */ + c = b; + b = l; + blk = get_nodes_block(r); + dbg = dbg == get_irn_dbg_info(r) ? dbg : NULL; + goto transform; + } + } + return 0; + +transform: + /* In some cases a and b might be both of different integer mode, and c a SymConst. + * in that case we could either + * 1.) cast into unsigned mode + * 2.) ignore + * we implement the second here + */ + ma = get_irn_mode(a); + mb = get_irn_mode(b); + if (ma != mb && mode_is_int(ma) && mode_is_int(mb)) + return 0; + + /* check if (a .op. b) can be calculated in the same block is the old instruction */ + if (! block_dominates(get_nodes_block(a), blk)) + return 0; + if (! block_dominates(get_nodes_block(b), blk)) + return 0; + /* ok */ + in[0] = a; + in[1] = b; + + mode = get_mode_from_ops(a, b); + in[0] = irn = optimize_node(new_ir_node(dbg, current_ir_graph, blk, op, mode, 2, in)); + + /* beware: optimize_node might have changed the opcode, check again */ + if (is_Add(irn) || is_Sub(irn)) { + reverse_rule_distributive(&in[0]); + } + in[1] = c; + + mode = get_mode_from_ops(in[0], in[1]); + irn = optimize_node(new_ir_node(dbg, current_ir_graph, blk, op, mode, 2, in)); + + exchange(n, irn); + *node = irn; + return 1; +} /* move_consts_up */ + +/** + * Apply the rules in reverse order, removing code that was not collapsed + */ +static void reverse_rules(ir_node *node, void *env) { + walker_t *wenv = env; + ir_mode *mode = get_irn_mode(node); + int res; + + /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */ + if (mode_is_float(mode) && get_irg_fp_model(current_ir_graph) & fp_strict_algebraic) + return; + + do { + ir_op *op = get_irn_op(node); + + res = 0; + if (is_op_commutative(op)) { + wenv->changes |= res = move_consts_up(&node); + } + /* beware: move_consts_up might have changed the opcode, check again */ + if (is_Add(node) || is_Sub(node)) { + wenv->changes |= res = reverse_rule_distributive(&node); + } + } while (res); +} + /* * do the reassociation */ @@ -398,14 +744,14 @@ void optimize_reassociation(ir_graph *irg) { walker_t env; irg_loopinfo_state state; + ir_graph *rem; assert(get_irg_phase_state(irg) != phase_building); assert(get_irg_pinned(irg) != op_pin_state_floats && "Reassociation needs pinned graph to work properly"); - /* reassociation needs constant folding */ - if (!get_opt_reassociation() || !get_opt_constant_folding()) - return; + rem = current_ir_graph; + current_ir_graph = irg; /* we use dominance to detect dead blocks */ assure_doms(irg); @@ -422,15 +768,28 @@ void optimize_reassociation(ir_graph *irg) construct_cf_backedges(irg); env.changes = 0; + env.wq = new_waitq(); + + /* disable some optimizations while reassoc is running to prevent endless loops */ + set_reassoc_running(1); + { + /* now we have collected enough information, optimize */ + irg_walk_graph(irg, NULL, wq_walker, &env); + do_reassociation(&env); - /* now we have collected enough information, optimize */ - irg_walk_graph(irg, NULL, do_reassociation, &env); + /* reverse those rules that do not result in collapsed constants */ + irg_walk_graph(irg, NULL, reverse_rules, &env); + } + set_reassoc_running(0); /* Handle graph state */ if (env.changes) { set_irg_outs_inconsistent(irg); set_irg_loopinfo_inconsistent(irg); } + + del_waitq(env.wq); + current_ir_graph = rem; } /* optimize_reassociation */ /* Sets the default reassociation operation for an ir_op_ops. */ @@ -445,6 +804,7 @@ ir_op_ops *firm_set_default_reassoc(ir_opcode code, ir_op_ops *ops) CASE(And); CASE(Or); CASE(Eor); + CASE(Shl); default: /* leave NULL */; }