X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Freassoc.c;h=c5262e1092d340b8848d0e60d217ee0dbadf9b87;hb=093a47b5cada6a1d0bf667ae8c857e78ca7c9233;hp=07e96ccfdc1e7c333ee378208817387cac7c9cd5;hpb=9d500f9aa4f5fe482e47417ce1190b9193856543;p=libfirm diff --git a/ir/opt/reassoc.c b/ir/opt/reassoc.c index 07e96ccfd..c5262e109 100644 --- a/ir/opt/reassoc.c +++ b/ir/opt/reassoc.c @@ -21,10 +21,10 @@ * @file * @brief Reassociation * @author Michael Beck - * @version $Id$ */ #include "config.h" +#include "iroptimize.h" #include "iropt_t.h" #include "irnode_t.h" #include "irgraph_t.h" @@ -47,11 +47,6 @@ DEBUG_ONLY(static firm_dbg_module_t *dbg;) -typedef struct _walker_t { - int changes; /**< set, if a reassociation take place */ - waitq *wq; /**< a wait queue */ -} walker_t; - typedef enum { NO_CONSTANT = 0, /**< node is not constant */ REAL_CONSTANT = 1, /**< node is a Const that is suitable for constant folding */ @@ -222,7 +217,7 @@ static ir_mode *get_mode_from_ops(ir_node *op1, ir_node *op2) static int reassoc_commutative(ir_node **node) { ir_node *n = *node; - ir_op *op = get_irn_op(n); + ir_op *op = get_irn_op(n); ir_node *block = get_nodes_block(n); ir_node *t1, *c1; @@ -259,6 +254,7 @@ static int reassoc_commutative(ir_node **node) */ ir_node *irn, *in[2]; ir_mode *mode, *mode_c1 = get_irn_mode(c1), *mode_c2 = get_irn_mode(c2); + ir_graph *irg = get_irn_irg(c1); /* It might happen, that c1 and c2 have different modes, for * instance Is and Iu. @@ -285,11 +281,11 @@ static int reassoc_commutative(ir_node **node) in[1] = c2; 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[1] = optimize_node(new_ir_node(NULL, irg, 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)); + irn = optimize_node(new_ir_node(NULL, irg, 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, @@ -306,6 +302,56 @@ static int reassoc_commutative(ir_node **node) } } } + if (get_irn_op(c1) == op) { + ir_node *t = c1; + c1 = t1; + t1 = t; + } + if (get_irn_op(t1) == op) { + ir_node *l = get_binop_left(t1); + ir_node *r = get_binop_right(t1); + const_class_t c_r; + + if (r == c1) { + ir_node *t = r; + r = l; + l = t; + } + c_r = get_const_class(r, block); + if (c_r != NO_CONSTANT) { + /* + * Beware: don't do the following op if a constant was + * placed below, else we will fall into a loop. + */ + return 0; + } + + if (l == c1) { + /* convert x .OP. (x .OP. y) => y .OP. (x .OP. x) */ + ir_mode *mode_res = get_irn_mode(n); + ir_mode *mode_c1 = get_irn_mode(c1); + ir_graph *irg = get_irn_irg(c1); + ir_node *irn, *in[2]; + + in[0] = c1; + in[1] = c1; + + in[1] = optimize_node(new_ir_node(NULL, irg, block, op, mode_c1, 2, in)); + in[0] = r; + + irn = optimize_node(new_ir_node(NULL, irg, block, op, mode_res, 2, in)); + + DBG((dbg, LEVEL_5, "Applied: %n .%s. (%n .%s. %n) => %n .%s. (%n .%s. %n)\n", + c1, get_irn_opname(n), l, get_irn_opname(n), r, + r, get_irn_opname(n), c1, get_irn_opname(n), c1)); + + if (n != irn) { + exchange(n, irn); + *node = irn; + return 1; + } + } + } return 0; } /* reassoc_commutative */ @@ -404,6 +450,7 @@ static int reassoc_commutative(ir_node **n) ir_mode *mode_right; ir_node *new_node; ir_node *in[2]; + ir_graph *irg = get_irn_irg(last); in[0] = last; in[1] = args[i]; @@ -417,11 +464,11 @@ static int reassoc_commutative(ir_node **n) } if (mode_right != mode) { assert(is_Add(node) && mode_is_reference(get_irn_mode(node))); - in[1] = new_r_Conv(current_ir_graph, commutative_block,in[1], mode); + in[1] = new_r_Conv(irg, commutative_block,in[1], mode); } /* TODO: produce useful debug info! */ - new_node = new_ir_node(NULL, current_ir_graph, commutative_block, + new_node = new_ir_node(NULL, irg, commutative_block, commutative_op, mode, 2, in); new_node = optimize_node(new_node); last = new_node; @@ -469,10 +516,11 @@ static int reassoc_Mul(ir_node **node) /* we can only multiplication rules on integer arithmetic */ if (mode_is_int(get_irn_mode(t1)) && mode_is_int(get_irn_mode(t2))) { + ir_graph *irg = get_irn_irg(t1); in[0] = new_rd_Mul(NULL, block, c, t1, mode); in[1] = new_rd_Mul(NULL, block, c, t2, mode); - irn = optimize_node(new_ir_node(NULL, current_ir_graph, block, op, mode, 2, in)); + irn = optimize_node(new_ir_node(NULL, irg, block, op, mode, 2, in)); /* In some cases it might happen that the new irn is equal the old one, for * instance in: @@ -499,12 +547,14 @@ static int reassoc_Mul(ir_node **node) /** * 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; +static int reassoc_Shl(ir_node **node) +{ + ir_node *n = *node; + ir_node *c = get_Shl_right(n); + ir_node *x, *blk, *irn; + ir_graph *irg; + ir_mode *mode; + ir_tarval *tv; if (! is_Const(c)) return 0; @@ -519,7 +569,8 @@ static int reassoc_Shl(ir_node **node) { return 0; blk = get_nodes_block(n); - c = new_Const(tv); + irg = get_irn_irg(blk); + c = new_r_Const(irg, tv); irn = new_rd_Mul(get_irn_dbg_info(n), blk, x, c, mode); if (irn != n) { @@ -535,56 +586,39 @@ static int reassoc_Shl(ir_node **node) { */ static void wq_walker(ir_node *n, void *env) { - walker_t *wenv = env; + waitq *const wq = (waitq*)env; set_irn_link(n, NULL); - if (is_no_Block(n)) { - ir_node *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. */ - return; - } - waitq_put(wenv->wq, n); - set_irn_link(n, wenv->wq); + if (!is_Block(n)) { + waitq_put(wq, n); + set_irn_link(n, wq); } } /* wq_walker */ /** * The walker for the reassociation. */ -static void do_reassociation(walker_t *wenv) +static void do_reassociation(waitq *const wq) { int i, res, changed; - ir_node *n, *blk; + ir_node *n; - while (! waitq_empty(wenv->wq)) { - n = waitq_get(wenv->wq); + while (! waitq_empty(wq)) { + n = (ir_node*)waitq_get(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); + ir_op *op = get_irn_op(n); + ir_mode *mode = get_irn_mode(n); res = 0; /* 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) + if (mode_is_float(mode) && get_irg_fp_model(get_irn_irg(n)) & fp_strict_algebraic) break; if (op->ops.reassociate) { @@ -595,15 +629,13 @@ static void do_reassociation(walker_t *wenv) } 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); + if (get_irn_link(pred) != wq) { + waitq_put(wq, pred); + set_irn_link(pred, wq); } } } @@ -617,7 +649,8 @@ static void do_reassociation(walker_t *wenv) * * 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) { +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; @@ -627,7 +660,7 @@ static ir_node *earliest_block(ir_node *a, ir_node *b, ir_node *curr_blk) { res = blk_b; else res = blk_a; - if (res == get_irg_start_block(current_ir_graph)) + if (res == get_irg_start_block(get_irn_irg(curr_blk))) return curr_blk; return res; } /* earliest_block */ @@ -641,7 +674,8 @@ static ir_node *earliest_block(ir_node *a, ir_node *b, ir_node *curr_blk) { * Handling SymConsts as const might be not a good idea for all * architectures ... */ -static int is_constant_expr(ir_node *irn) { +static int is_constant_expr(ir_node *irn) +{ ir_op *op; switch (get_irn_opcode(irn)) { @@ -664,7 +698,8 @@ static int is_constant_expr(ir_node *irn) { /** * Apply distributive Law for Mul and Add/Sub */ -static int reverse_rule_distributive(ir_node **node) { +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); @@ -746,12 +781,14 @@ transform: /** * Move Constants towards the root. */ -static int move_consts_up(ir_node **node) { +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; + ir_graph *irg; l = get_binop_left(n); r = get_binop_right(n); @@ -782,7 +819,8 @@ static int move_consts_up(ir_node **node) { dbg = dbg == get_irn_dbg_info(l) ? dbg : NULL; goto transform; } - } else if (get_irn_op(r) == op) { + } + if (get_irn_op(r) == op) { /* l .op. (a .op. b) */ a = get_binop_left(r); b = get_binop_right(r); @@ -827,7 +865,8 @@ transform: 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)); + irg = get_irn_irg(blk); + in[0] = irn = optimize_node(new_ir_node(dbg, irg, blk, op, mode, 2, in)); /* beware: optimize_node might have changed the opcode, check again */ if (is_Add(irn) || is_Sub(irn)) { @@ -836,7 +875,7 @@ transform: 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)); + irn = optimize_node(new_ir_node(dbg, irg, blk, op, mode, 2, in)); exchange(n, irn); *node = irn; @@ -846,13 +885,16 @@ transform: /** * 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); +static void reverse_rules(ir_node *node, void *env) +{ + (void)env; + + ir_graph *irg = get_irn_irg(node); + 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) + if (mode_is_float(mode) && get_irg_fp_model(irg) & fp_strict_algebraic) return; do { @@ -860,11 +902,11 @@ static void reverse_rules(ir_node *node, void *env) { res = 0; if (is_op_commutative(op)) { - wenv->changes |= res = move_consts_up(&node); + 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); + res = reverse_rule_distributive(&node); } } while (res); } @@ -872,93 +914,65 @@ static void reverse_rules(ir_node *node, void *env) { /* * do the reassociation */ -int optimize_reassociation(ir_graph *irg) +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"); - rem = current_ir_graph; - current_ir_graph = irg; - - /* we use dominance to detect dead blocks */ - assure_doms(irg); + assure_irg_properties(irg, + IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE + | IR_GRAPH_PROPERTY_CONSISTENT_LOOPINFO); #ifdef NEW_REASSOC - assure_irg_outs(irg); + assire_irg_properties(irg, IR_GRAPH_PROPERTY_CONSISTENT_OUTS); obstack_init(&commutative_args); #endif - /* - * Calculate loop info, so we could identify loop-invariant - * code and threat it like a constant. - * We only need control flow loops here but can handle generic - * INTRA info as well. - */ - state = get_irg_loopinfo_state(irg); - if ((state & loopinfo_inter) || - (state & (loopinfo_constructed | loopinfo_valid)) != (loopinfo_constructed | loopinfo_valid)) - construct_cf_backedges(irg); - - env.changes = 0; - env.wq = new_waitq(); + waitq *const 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); + irg_walk_graph(irg, NULL, wq_walker, wq); + do_reassociation(wq); /* reverse those rules that do not result in collapsed constants */ - irg_walk_graph(irg, NULL, reverse_rules, &env); + irg_walk_graph(irg, NULL, reverse_rules, NULL); } set_reassoc_running(0); - /* Handle graph state */ - if (env.changes) { - set_irg_outs_inconsistent(irg); - set_irg_loopinfo_inconsistent(irg); - } - #ifdef NEW_REASSOC obstack_free(&commutative_args, NULL); #endif - del_waitq(env.wq); - current_ir_graph = rem; - return env.changes; + del_waitq(wq); + + confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_CONTROL_FLOW); } /* optimize_reassociation */ /* create a pass for the reassociation */ -ir_graph_pass_t *optimize_reassociation_pass(const char *name) { - return def_graph_pass_ret(name ? name : "reassoc", optimize_reassociation); +ir_graph_pass_t *optimize_reassociation_pass(const char *name) +{ + return def_graph_pass(name ? name : "reassoc", optimize_reassociation); } /* optimize_reassociation_pass */ -/* Sets the default reassociation operation for an ir_op_ops. */ -ir_op_ops *firm_set_default_reassoc(ir_opcode code, ir_op_ops *ops) +static void register_node_reassoc_func(ir_op *op, reassociate_func func) { -#define CASE(a) case iro_##a: ops->reassociate = reassoc_##a; break - - switch (code) { - CASE(Mul); - CASE(Add); - CASE(Sub); - CASE(And); - CASE(Or); - CASE(Eor); - CASE(Shl); - default: - /* leave NULL */; - } + op->ops.reassociate = func; +} - return ops; -#undef CASE -} /* firm_set_default_reassoc */ +void ir_register_reassoc_node_ops(void) +{ + register_node_reassoc_func(op_Mul, reassoc_Mul); + register_node_reassoc_func(op_Add, reassoc_Add); + register_node_reassoc_func(op_Sub, reassoc_Sub); + register_node_reassoc_func(op_And, reassoc_And); + register_node_reassoc_func(op_Or, reassoc_Or); + register_node_reassoc_func(op_Eor, reassoc_Eor); + register_node_reassoc_func(op_Shl, reassoc_Shl); +} /* initialize the reassociation by adding operations to some opcodes */ void firm_init_reassociation(void)