X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firopt.c;h=c5291cbf9c0beeb8404cebedf190c614300e8607;hb=3c3425a50a1d721b74a015c6812257e32feeac85;hp=b67e8ed4c2ba6a196e3b459f725c110198ee93b5;hpb=b7bf9774e5e704382bdf3eb7c09bc5755d4f067e;p=libfirm diff --git a/ir/ir/iropt.c b/ir/ir/iropt.c index b67e8ed4c..c5291cbf9 100644 --- a/ir/ir/iropt.c +++ b/ir/ir/iropt.c @@ -21,7 +21,6 @@ * @file * @brief iropt --- optimizations intertwined with IR construction. * @author Christian Schaefer, Goetz Lindenmaier, Michael Beck - * @version $Id$ */ #include "config.h" @@ -44,7 +43,6 @@ #include "irhooks.h" #include "irarch.h" #include "hashptr.h" -#include "opt_polymorphy.h" #include "irtools.h" #include "irhooks.h" #include "array_t.h" @@ -53,7 +51,6 @@ #include "bitfiddle.h" #include "be.h" -/* Make types visible to allow most efficient access */ #include "entity_t.h" static bool is_Or_Eor_Add(const ir_node *node) @@ -85,7 +82,6 @@ static ir_tarval *default_value_of(const ir_node *n) value_of_func value_of_ptr = default_value_of; -/* * Set a new value_of function. */ void set_value_of_func(value_of_func func) { if (func != NULL) @@ -588,6 +584,15 @@ ir_relation ir_get_possible_cmp_relations(const ir_node *left, /* Alloc nodes never return null (but throw an exception) */ if (is_Alloc(left) && tarval_is_null(tv_r)) possible &= ~ir_relation_equal; + /* stuff known through confirm nodes */ + if (is_Confirm(left) && get_Confirm_bound(left) == right) { + possible &= get_Confirm_relation(left); + } + if (is_Confirm(right) && get_Confirm_bound(right) == left) { + ir_relation relation = get_Confirm_relation(right); + relation = get_inversed_relation(relation); + possible &= relation; + } return possible; } @@ -609,6 +614,17 @@ static ir_tarval *compute_cmp(const ir_node *cmp) return computed_value_Cmp_Confirm(cmp, left, right, relation); } +/** + * some people want to call compute_cmp directly, in this case we have to + * test the constant folding flag again + */ +static ir_tarval *compute_cmp_ext(const ir_node *cmp) +{ + if (!get_opt_constant_folding()) + return tarval_bad; + return compute_cmp(cmp); +} + /** * Return the value of a Cmp. * @@ -618,7 +634,7 @@ static ir_tarval *compute_cmp(const ir_node *cmp) static ir_tarval *computed_value_Cmp(const ir_node *cmp) { /* we can't construct Constb after lowering mode_b nodes */ - if (is_irg_state(get_irn_irg(cmp), IR_GRAPH_STATE_MODEB_LOWERED)) + if (irg_is_constrained(get_irn_irg(cmp), IR_GRAPH_CONSTRAINT_MODEB_LOWERED)) return tarval_bad; return compute_cmp(cmp); @@ -720,60 +736,6 @@ ir_tarval *computed_value(const ir_node *n) return tarval_bad; } -/** - * Set the default computed_value evaluator in an ir_op_ops. - * - * @param code the opcode for the default operation - * @param ops the operations initialized - * - * @return - * The operations. - */ -static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops) -{ -#define CASE(a) \ - case iro_##a: \ - ops->computed_value = computed_value_##a; \ - break -#define CASE_PROJ(a) \ - case iro_##a: \ - ops->computed_value_Proj = computed_value_Proj_##a; \ - break - - switch (code) { - CASE(Add); - CASE(And); - CASE(Borrow); - CASE(Carry); - CASE(Cmp); - CASE(Confirm); - CASE(Const); - CASE(Conv); - CASE(Eor); - CASE(Minus); - CASE(Mul); - CASE(Mux); - CASE(Not); - CASE(Or); - CASE(Proj); - CASE(Rotl); - CASE(Shl); - CASE(Shr); - CASE(Shrs); - CASE(Sub); - CASE(SymConst); - CASE_PROJ(Div); - CASE_PROJ(Mod); - default: - /* leave NULL */ - break; - } - - return ops; -#undef CASE_PROJ -#undef CASE -} - /** * Optimize operations that are commutative and have neutral 0, * so a op 0 = 0 op a = a. @@ -937,11 +899,6 @@ static ir_node *equivalent_node_left_zero(ir_node *n) return n; } -#define equivalent_node_Shl equivalent_node_left_zero -#define equivalent_node_Shr equivalent_node_left_zero -#define equivalent_node_Shrs equivalent_node_left_zero -#define equivalent_node_Rotl equivalent_node_left_zero - /** * Optimize a - 0 and (a + x) - x (for modes with wrap-around). * @@ -1002,13 +959,6 @@ static ir_node *equivalent_node_idempotent_unop(ir_node *n) return n; } -/** Optimize Not(Not(x)) == x. */ -#define equivalent_node_Not equivalent_node_idempotent_unop - -/** -(-x) == x ??? Is this possible or can --x raise an - out of bounds exception if min =! max? */ -#define equivalent_node_Minus equivalent_node_idempotent_unop - /** * Optimize a * 1 = 1 * a = a. */ @@ -1106,9 +1056,9 @@ static ir_node *equivalent_node_And(ir_node *n) /* Check Conv(all_one) & Const = all_one */ ir_tarval *one = get_mode_all_one(convopmode); ir_tarval *conv = tarval_convert_to(one, mode); - ir_tarval *and = tarval_and(conv, tv); + ir_tarval *tand = tarval_and(conv, tv); - if (tarval_is_all_one(and)) { + if (tarval_is_all_one(tand)) { /* Conv(X) & Const = X */ n = a; DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND); @@ -1249,7 +1199,7 @@ restart: if (mode_is_int(n_mode) && get_mode_arithmetic(a_mode) == irma_ieee754) { /* ConvI(ConvF(I)) -> I, iff float mantissa >= int mode */ unsigned int_mantissa = get_mode_size_bits(n_mode) - (mode_is_signed(n_mode) ? 1 : 0); - unsigned float_mantissa = tarval_ieee754_get_mantissa_size(a_mode); + unsigned float_mantissa = get_mode_mantissa_size(a_mode); if (float_mantissa >= int_mantissa) { n = b; @@ -1477,7 +1427,7 @@ static ir_node *equivalent_node_Mux(ir_node *n) if (ts == tarval_bad && is_Cmp(sel)) { /* try again with a direct call to compute_cmp, as we don't care * about the MODEB_LOWERED flag here */ - ts = compute_cmp(sel); + ts = compute_cmp_ext(sel); } /* Mux(true, f, t) == t */ @@ -1608,59 +1558,6 @@ ir_node *equivalent_node(ir_node *n) return n; } -/** - * Sets the default equivalent node operation for an ir_op_ops. - * - * @param code the opcode for the default operation - * @param ops the operations initialized - * - * @return - * The operations. - */ -static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops) -{ -#define CASE(a) \ - case iro_##a: \ - ops->equivalent_node = equivalent_node_##a; \ - break -#define CASE_PROJ(a) \ - case iro_##a: \ - ops->equivalent_node_Proj = equivalent_node_Proj_##a; \ - break - - switch (code) { - CASE(Eor); - CASE(Add); - CASE(Shl); - CASE(Shr); - CASE(Shrs); - CASE(Rotl); - CASE(Sub); - CASE(Not); - CASE(Minus); - CASE(Mul); - CASE(Or); - CASE(And); - CASE(Conv); - CASE(Phi); - CASE_PROJ(Tuple); - CASE_PROJ(Div); - CASE_PROJ(CopyB); - CASE_PROJ(Bound); - CASE(Proj); - CASE(Id); - CASE(Mux); - CASE(Confirm); - default: - /* leave NULL */ - break; - } - - return ops; -#undef CASE -#undef CASE_PROJ -} - /** * Returns non-zero if a node is a Phi node * with all predecessors constant. @@ -2149,7 +2046,7 @@ static ir_node *transform_node_Or_bf_store(ir_node *irn_or) } /* ok, all conditions met */ - block = get_irn_n(irn_or, -1); + block = get_nodes_block(irn_or); irg = get_irn_irg(block); new_and = new_r_And(block, value, new_r_Const(irg, tarval_and(tv4, tv2)), mode); @@ -2375,7 +2272,7 @@ static ir_node *transform_node_bitop_shift(ir_node *n) ir_tarval *tv2; ir_tarval *tv_bitop; - if (!is_irg_state(irg, IR_GRAPH_STATE_NORMALISATION2)) + if (!irg_is_constrained(irg, IR_GRAPH_CONSTRAINT_NORMALISATION2)) return n; assert(is_And(n) || is_Or(n) || is_Eor(n) || is_Or_Eor_Add(n)); @@ -2461,6 +2358,55 @@ static bool complement_values(const ir_node *a, const ir_node *b) return false; } +typedef ir_tarval *(tv_fold_binop_func)(ir_tarval *a, ir_tarval *b); + +/** + * for associative operations fold: + * op(op(x, c0), c1) to op(x, op(c0, c1)) with constants folded. + * This is a "light" version of the reassociation phase + */ +static ir_node *fold_constant_associativity(ir_node *node, + tv_fold_binop_func fold) +{ + ir_graph *irg; + ir_op *op; + ir_node *left; + ir_node *right = get_binop_right(node); + ir_node *left_right; + ir_node *left_left; + ir_tarval *c0; + ir_tarval *c1; + ir_tarval *new_c; + ir_node *new_const; + ir_node *new_node; + if (!is_Const(right)) + return node; + + op = get_irn_op(node); + left = get_binop_left(node); + if (get_irn_op(left) != op) + return node; + + left_right = get_binop_right(left); + if (!is_Const(left_right)) + return node; + + left_left = get_binop_left(left); + c0 = get_Const_tarval(left_right); + c1 = get_Const_tarval(right); + irg = get_irn_irg(node); + if (get_tarval_mode(c0) != get_tarval_mode(c1)) + return node; + new_c = fold(c0, c1); + if (new_c == tarval_bad) + return node; + new_const = new_r_Const(irg, new_c); + new_node = exact_copy(node); + set_binop_left(new_node, left_left); + set_binop_right(new_node, new_const); + return new_node; +} + /** * Transform an Or. */ @@ -2472,6 +2418,10 @@ static ir_node *transform_node_Or_(ir_node *n) ir_node *c; ir_mode *mode; + n = fold_constant_associativity(n, tarval_or); + if (n != oldn) + return n; + if (is_Not(a) && is_Not(b)) { /* ~a | ~b = ~(a&b) */ ir_node *block = get_nodes_block(n); @@ -2512,9 +2462,9 @@ static ir_node *transform_node_Or_(ir_node *n) ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode); ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode); ir_node *conv = new_rd_Conv(dbgi, block, xora, b_mode); - ir_node *or = new_rd_Or(dbgi, block, conv, xorb, b_mode); + ir_node *orn = new_rd_Or(dbgi, block, conv, xorb, b_mode); ir_node *zero = create_zero_const(irg, b_mode); - return new_rd_Cmp(dbgi, block, or, zero, ir_relation_less_greater); + return new_rd_Cmp(dbgi, block, orn, zero, ir_relation_less_greater); } if (values_in_mode(get_irn_mode(b_left), get_irn_mode(a_left))) { ir_graph *irg = get_irn_irg(n); @@ -2525,9 +2475,9 @@ static ir_node *transform_node_Or_(ir_node *n) ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode); ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode); ir_node *conv = new_rd_Conv(dbgi, block, xorb, a_mode); - ir_node *or = new_rd_Or(dbgi, block, xora, conv, a_mode); + ir_node *orn = new_rd_Or(dbgi, block, xora, conv, a_mode); ir_node *zero = create_zero_const(irg, a_mode); - return new_rd_Cmp(dbgi, block, or, zero, ir_relation_less_greater); + return new_rd_Cmp(dbgi, block, orn, zero, ir_relation_less_greater); } } } @@ -2576,6 +2526,10 @@ static ir_node *transform_node_Eor_(ir_node *n) ir_mode *mode = get_irn_mode(n); ir_node *c; + n = fold_constant_associativity(n, tarval_eor); + if (n != oldn) + return n; + /* we can combine the relations of two compares with the same operands */ if (is_Cmp(a) && is_Cmp(b)) { ir_node *a_left = get_Cmp_left(a); @@ -2643,8 +2597,6 @@ static ir_node *transform_node_Eor(ir_node *n) return transform_node_Eor_(n); } - - /** * Do the AddSub optimization, then Transform * Constant folding on Phi @@ -2662,6 +2614,10 @@ static ir_node *transform_node_Add(ir_node *n) ir_node *c; ir_node *oldn = n; + n = fold_constant_associativity(n, tarval_add); + if (n != oldn) + return n; + n = transform_node_AddSub(n); if (n != oldn) return n; @@ -2680,6 +2636,21 @@ static ir_node *transform_node_Add(ir_node *n) } } + if (is_Const(b) && get_mode_arithmetic(mode) == irma_twos_complement) { + ir_tarval *tv = get_Const_tarval(b); + ir_tarval *min = get_mode_min(mode); + /* if all bits are set, then this has the same effect as a Not. + * Note that the following == gives false for different modes which + * is exactly what we want */ + if (tv == min) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_graph *irg = get_irn_irg(n); + ir_node *block = get_nodes_block(n); + ir_node *cnst = new_r_Const(irg, min); + return new_rd_Eor(dbgi, block, a, cnst, mode); + } + } + HANDLE_BINOP_PHI((eval_func) tarval_add, a, b, c, mode); /* for FP the following optimizations are only allowed if @@ -2694,7 +2665,7 @@ static ir_node *transform_node_Add(ir_node *n) ir_graph *irg = get_irn_irg(n); /* the following code leads to endless recursion when Mul are replaced * by a simple instruction chain */ - if (!is_irg_state(irg, IR_GRAPH_STATE_ARCH_DEP) + if (!irg_is_constrained(irg, IR_GRAPH_CONSTRAINT_ARCH_DEP) && a == b && mode_is_int(mode)) { ir_node *block = get_nodes_block(n); @@ -3083,8 +3054,8 @@ restart: ir_node *block = get_nodes_block(n); ir_mode *mode = get_irn_mode(n); ir_node *notn = new_rd_Not(dbgi, block, and_right, mode); - ir_node *and = new_rd_And(dbgi, block, a, notn, mode); - return and; + ir_node *andn = new_rd_And(dbgi, block, a, notn, mode); + return andn; } } } @@ -3121,7 +3092,7 @@ static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) } if (tb == get_mode_one(smode)) { /* (L)a * (L)1 = (L)a */ - ir_node *blk = get_irn_n(a, -1); + ir_node *blk = get_nodes_block(a); n = new_rd_Conv(get_irn_dbg_info(n), blk, a, mode); DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1); return n; @@ -3149,7 +3120,8 @@ static ir_node *transform_node_Mul(ir_node *n) ir_node *a = get_Mul_left(n); ir_node *b = get_Mul_right(n); - if (is_Bad(a) || is_Bad(b)) + n = fold_constant_associativity(n, tarval_mul); + if (n != oldn) return n; if (mode != get_irn_mode(a)) @@ -3230,10 +3202,11 @@ static ir_node *transform_node_Mul(ir_node *n) return n; } } - if (get_mode_arithmetic(mode) == irma_ieee754) { + if (get_mode_arithmetic(mode) == irma_ieee754 + || get_mode_arithmetic(mode) == irma_x86_extended_float) { if (is_Const(a)) { ir_tarval *tv = get_Const_tarval(a); - if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv) + if (tarval_get_exponent(tv) == 1 && tarval_zero_mantissa(tv) && !tarval_is_negative(tv)) { /* 2.0 * b = b + b */ n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), b, b, mode); @@ -3243,7 +3216,7 @@ static ir_node *transform_node_Mul(ir_node *n) } else if (is_Const(b)) { ir_tarval *tv = get_Const_tarval(b); - if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv) + if (tarval_get_exponent(tv) == 1 && tarval_zero_mantissa(tv) && !tarval_is_negative(tv)) { /* a * 2.0 = a + a */ n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), a, a, mode); @@ -3467,7 +3440,6 @@ make_tuple: */ static ir_node *transform_node_Cond(ir_node *n) { - ir_node *a = get_Cond_selector(n); ir_graph *irg = get_irn_irg(n); ir_tarval *ta; @@ -3477,15 +3449,11 @@ static ir_node *transform_node_Cond(ir_node *n) if (get_irg_pinned(irg) == op_pin_state_floats) return n; - /* we do not handle switches here */ - if (get_irn_mode(a) != mode_b) - return n; - ta = value_of(a); if (ta == tarval_bad && is_Cmp(a)) { /* try again with a direct call to compute_cmp, as we don't care * about the MODEB_LOWERED flag here */ - ta = compute_cmp(a); + ta = compute_cmp_ext(a); } if (ta != tarval_bad && get_irn_mode(a) == mode_b) { @@ -3503,7 +3471,49 @@ static ir_node *transform_node_Cond(ir_node *n) } /* We might generate an endless loop, so keep it alive. */ add_End_keepalive(get_irg_end(irg), blk); - clear_irg_state(irg, IR_GRAPH_STATE_NO_UNREACHABLE_CODE); + clear_irg_properties(irg, IR_GRAPH_PROPERTY_NO_UNREACHABLE_CODE); + } + return n; +} + +static ir_node *transform_node_Switch(ir_node *n) +{ + ir_node *op = get_Switch_selector(n); + ir_tarval *val = value_of(op); + if (val != tarval_bad) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_graph *irg = get_irn_irg(n); + unsigned n_outs = get_Switch_n_outs(n); + ir_node *block = get_nodes_block(n); + ir_node *bad = new_r_Bad(irg, mode_X); + ir_node **in = XMALLOCN(ir_node*, n_outs); + const ir_switch_table *table = get_Switch_table(n); + size_t n_entries = ir_switch_table_get_n_entries(table); + long jmp_pn = 0; + size_t i; + unsigned o; + for (i = 0; i < n_entries; ++i) { + const ir_switch_table_entry *entry + = ir_switch_table_get_entry_const(table, i); + ir_tarval *min = entry->min; + ir_tarval *max = entry->max; + if (entry->pn == 0) + continue; + if ((min == max && min == val) + || (tarval_cmp(val, min) != ir_relation_less + && tarval_cmp(val, max) != ir_relation_greater)) { + jmp_pn = entry->pn; + break; + } + } + for (o = 0; o < n_outs; ++o) { + if (o == (unsigned)jmp_pn) { + in[o] = new_rd_Jmp(dbgi, block); + } else { + in[o] = bad; + } + } + return new_r_Tuple(block, (int)n_outs, in); } return n; } @@ -3535,7 +3545,7 @@ static ir_node *transform_node_shift_bitop(ir_node *n) ir_tarval *tv2; ir_tarval *tv_shift; - if (is_irg_state(irg, IR_GRAPH_STATE_NORMALISATION2)) + if (irg_is_constrained(irg, IR_GRAPH_CONSTRAINT_NORMALISATION2)) return n; assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n)); @@ -3608,6 +3618,10 @@ static ir_node *transform_node_And(ir_node *n) ir_node *b = get_And_right(n); ir_mode *mode; + n = fold_constant_associativity(n, tarval_and); + if (n != oldn) + return n; + if (is_Cmp(a) && is_Cmp(b)) { ir_node *a_left = get_Cmp_left(a); ir_node *a_right = get_Cmp_right(a); @@ -3635,10 +3649,10 @@ static ir_node *transform_node_And(ir_node *n) ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode); ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode); ir_node *conv = new_rd_Conv(dbgi, block, xora, b_mode); - ir_node *or = new_rd_Or(dbgi, block, conv, xorb, b_mode); + ir_node *orn = new_rd_Or(dbgi, block, conv, xorb, b_mode); ir_graph *irg = get_irn_irg(n); ir_node *zero = create_zero_const(irg, b_mode); - return new_rd_Cmp(dbgi, block, or, zero, ir_relation_equal); + return new_rd_Cmp(dbgi, block, orn, zero, ir_relation_equal); } if (values_in_mode(get_irn_mode(b_left), get_irn_mode(a_left))) { dbg_info *dbgi = get_irn_dbg_info(n); @@ -3648,10 +3662,10 @@ static ir_node *transform_node_And(ir_node *n) ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode); ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode); ir_node *conv = new_rd_Conv(dbgi, block, xorb, a_mode); - ir_node *or = new_rd_Or(dbgi, block, xora, conv, a_mode); + ir_node *orn = new_rd_Or(dbgi, block, xora, conv, a_mode); ir_graph *irg = get_irn_irg(n); ir_node *zero = create_zero_const(irg, a_mode); - return new_rd_Cmp(dbgi, block, or, zero, ir_relation_equal); + return new_rd_Cmp(dbgi, block, orn, zero, ir_relation_equal); } } } @@ -3963,27 +3977,25 @@ static ir_node *transform_node_Minus(ir_node *n) */ static ir_node *transform_node_Proj_Load(ir_node *proj) { - if (get_opt_ldst_only_null_ptr_exceptions()) { - if (get_irn_mode(proj) == mode_X) { - ir_node *load = get_Proj_pred(proj); + if (get_irn_mode(proj) == mode_X) { + ir_node *load = get_Proj_pred(proj); - /* get the Load address */ - const ir_node *addr = get_Load_ptr(load); - const ir_node *confirm; + /* get the Load address */ + const ir_node *addr = get_Load_ptr(load); + const ir_node *confirm; - if (value_not_null(addr, &confirm)) { - if (confirm == NULL) { - /* this node may float if it did not depend on a Confirm */ - set_irn_pinned(load, op_pin_state_floats); - } - if (get_Proj_proj(proj) == pn_Load_X_except) { - ir_graph *irg = get_irn_irg(proj); - DBG_OPT_EXC_REM(proj); - return new_r_Bad(irg, mode_X); - } else { - ir_node *blk = get_nodes_block(load); - return new_r_Jmp(blk); - } + if (value_not_null(addr, &confirm)) { + if (confirm == NULL) { + /* this node may float if it did not depend on a Confirm */ + set_irn_pinned(load, op_pin_state_floats); + } + if (get_Proj_proj(proj) == pn_Load_X_except) { + ir_graph *irg = get_irn_irg(proj); + DBG_OPT_EXC_REM(proj); + return new_r_Bad(irg, mode_X); + } else { + ir_node *blk = get_nodes_block(load); + return new_r_Jmp(blk); } } } @@ -3995,27 +4007,25 @@ static ir_node *transform_node_Proj_Load(ir_node *proj) */ static ir_node *transform_node_Proj_Store(ir_node *proj) { - if (get_opt_ldst_only_null_ptr_exceptions()) { - if (get_irn_mode(proj) == mode_X) { - ir_node *store = get_Proj_pred(proj); + if (get_irn_mode(proj) == mode_X) { + ir_node *store = get_Proj_pred(proj); - /* get the load/store address */ - const ir_node *addr = get_Store_ptr(store); - const ir_node *confirm; + /* get the load/store address */ + const ir_node *addr = get_Store_ptr(store); + const ir_node *confirm; - if (value_not_null(addr, &confirm)) { - if (confirm == NULL) { - /* this node may float if it did not depend on a Confirm */ - set_irn_pinned(store, op_pin_state_floats); - } - if (get_Proj_proj(proj) == pn_Store_X_except) { - ir_graph *irg = get_irn_irg(proj); - DBG_OPT_EXC_REM(proj); - return new_r_Bad(irg, mode_X); - } else { - ir_node *blk = get_nodes_block(store); - return new_r_Jmp(blk); - } + if (value_not_null(addr, &confirm)) { + if (confirm == NULL) { + /* this node may float if it did not depend on a Confirm */ + set_irn_pinned(store, op_pin_state_floats); + } + if (get_Proj_proj(proj) == pn_Store_X_except) { + ir_graph *irg = get_irn_irg(proj); + DBG_OPT_EXC_REM(proj); + return new_r_Bad(irg, mode_X); + } else { + ir_node *blk = get_nodes_block(store); + return new_r_Jmp(blk); } } } @@ -4102,7 +4112,7 @@ static ir_node *transform_node_Proj_Mod(ir_node *proj) switch (proj_nr) { case pn_Mod_X_regular: - return new_r_Jmp(get_irn_n(mod, -1)); + return new_r_Jmp(get_nodes_block(mod)); case pn_Mod_X_except: { ir_graph *irg = get_irn_irg(proj); @@ -4139,87 +4149,6 @@ static ir_node *transform_node_Proj_Mod(ir_node *proj) return proj; } -/** - * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases. - */ -static ir_node *transform_node_Proj_Cond(ir_node *proj) -{ - ir_node *n = get_Proj_pred(proj); - ir_node *b = get_Cond_selector(n); - - if (mode_is_int(get_irn_mode(b))) { - ir_tarval *tb = value_of(b); - - if (tb != tarval_bad) { - /* we have a constant switch */ - long num = get_Proj_proj(proj); - - if (num != get_Cond_default_proj(n)) { /* we cannot optimize default Proj's yet */ - if (get_tarval_long(tb) == num) { - /* Do NOT create a jump here, or we will have 2 control flow ops - * in a block. This case is optimized away in optimize_cf(). */ - return proj; - } else { - ir_graph *irg = get_irn_irg(proj); - /* this case will NEVER be taken, kill it */ - clear_irg_state(irg, IR_GRAPH_STATE_NO_UNREACHABLE_CODE); - return new_r_Bad(irg, mode_X); - } - } - } else { - long num = get_Proj_proj(proj); - vrp_attr *b_vrp = vrp_get_info(b); - if (num != get_Cond_default_proj(n) && b_vrp) { - /* Try handling with vrp data. We only remove dead parts. */ - ir_tarval *tp = new_tarval_from_long(num, get_irn_mode(b)); - - if (b_vrp->range_type == VRP_RANGE) { - ir_relation cmp_result = tarval_cmp(b_vrp->range_bottom, tp); - ir_relation cmp_result2 = tarval_cmp(b_vrp->range_top, tp); - - if ((cmp_result & ir_relation_greater) == cmp_result - && (cmp_result2 & ir_relation_less) == cmp_result2) { - ir_graph *irg = get_irn_irg(proj); - clear_irg_state(irg, IR_GRAPH_STATE_NO_UNREACHABLE_CODE); - return new_r_Bad(irg, mode_X); - } - } else if (b_vrp->range_type == VRP_ANTIRANGE) { - ir_relation cmp_result = tarval_cmp(b_vrp->range_bottom, tp); - ir_relation cmp_result2 = tarval_cmp(b_vrp->range_top, tp); - - if ((cmp_result & ir_relation_less_equal) == cmp_result - && (cmp_result2 & ir_relation_greater_equal) == cmp_result2) { - ir_graph *irg = get_irn_irg(proj); - clear_irg_state(irg, IR_GRAPH_STATE_NO_UNREACHABLE_CODE); - return new_r_Bad(irg, mode_X); - } - } - - if (!(tarval_cmp( - tarval_and( b_vrp->bits_set, tp), - b_vrp->bits_set - ) == ir_relation_equal)) { - ir_graph *irg = get_irn_irg(proj); - clear_irg_state(irg, IR_GRAPH_STATE_NO_UNREACHABLE_CODE); - return new_r_Bad(irg, mode_X); - } - - if (!(tarval_cmp( - tarval_and( - tarval_not(tp), - tarval_not(b_vrp->bits_not_set)), - tarval_not(b_vrp->bits_not_set)) - == ir_relation_equal)) { - ir_graph *irg = get_irn_irg(proj); - clear_irg_state(irg, IR_GRAPH_STATE_NO_UNREACHABLE_CODE); - return new_r_Bad(irg, mode_X); - } - } - } - } - return proj; -} - /** * return true if the operation returns a value with exactly 1 bit set */ @@ -4242,6 +4171,26 @@ static bool is_single_bit(const ir_node *node) return false; } +/** + * checks if node just flips a bit in another node and returns that other node + * if so. @p tv should be a value having just 1 bit set + */ +static ir_node *flips_bit(const ir_node *node, ir_tarval *tv) +{ + if (is_Not(node)) + return get_Not_op(node); + if (is_Eor(node)) { + ir_node *right = get_Eor_right(node); + if (is_Const(right)) { + ir_tarval *right_tv = get_Const_tarval(right); + ir_mode *mode = get_irn_mode(node); + if (tarval_and(right_tv, tv) != get_mode_null(mode)) + return get_Eor_left(node); + } + } + return NULL; +} + /** * Normalizes and optimizes Cmp nodes. */ @@ -4489,25 +4438,53 @@ static ir_node *transform_node_Cmp(ir_node *n) } } - /* Cmp(And(1bit, val), 1bit) "bit-testing" can be replaced - * by the simpler Cmp(And(1bit), val), 0) negated pnc */ - if (mode_is_int(mode) && is_And(left) - && (relation == ir_relation_equal + if (mode_is_int(mode) && is_And(left)) { + /* a complicated Cmp(And(1bit, val), 1bit) "bit-testing" can be replaced + * by the simpler Cmp(And(1bit, val), 0) negated pnc */ + if (relation == ir_relation_equal || (mode_is_signed(mode) && relation == ir_relation_less_greater) - || (!mode_is_signed(mode) && (relation & ir_relation_less_equal) == ir_relation_less))) { - ir_node *and0 = get_And_left(left); - ir_node *and1 = get_And_right(left); - if (and1 == right) { - ir_node *tmp = and0; - and0 = and1; - and1 = tmp; - } - if (and0 == right && is_single_bit(and0)) { - ir_graph *irg = get_irn_irg(n); - relation = - relation == ir_relation_equal ? ir_relation_less_greater : ir_relation_equal; - right = create_zero_const(irg, mode); - changed |= 1; + || (!mode_is_signed(mode) && (relation & ir_relation_less_equal) == ir_relation_less)) { + ir_node *and0 = get_And_left(left); + ir_node *and1 = get_And_right(left); + if (and1 == right) { + ir_node *tmp = and0; + and0 = and1; + and1 = tmp; + } + if (and0 == right && is_single_bit(and0)) { + ir_graph *irg = get_irn_irg(n); + relation = + relation == ir_relation_equal ? ir_relation_less_greater + : ir_relation_equal; + right = create_zero_const(irg, mode); + changed |= 1; + goto is_bittest; + } + } + + if (is_Const(right) && is_Const_null(right) && + (relation == ir_relation_equal + || (relation == ir_relation_less_greater) + || (!mode_is_signed(mode) && relation == ir_relation_greater))) { +is_bittest: { + /* instead of flipping the bit before the bit-test operation negate + * pnc */ + ir_node *and0 = get_And_left(left); + ir_node *and1 = get_And_right(left); + if (is_Const(and1)) { + ir_tarval *tv = get_Const_tarval(and1); + if (tarval_is_single_bit(tv)) { + ir_node *flipped = flips_bit(and0, tv); + if (flipped != NULL) { + dbg_info *dbgi = get_irn_dbg_info(left); + ir_node *block = get_nodes_block(left); + relation = get_negated_relation(relation); + left = new_rd_And(dbgi, block, flipped, and1, mode); + changed |= 1; + } + } + } + } } } @@ -4571,6 +4548,43 @@ static ir_node *transform_node_Cmp(ir_node *n) if (tv != tarval_bad) { ir_mode *mode = get_irn_mode(right); + /* cmp(mux(x, cf, ct), c2) can be eliminated: + * cmp(ct,c2) | cmp(cf,c2) | result + * -----------|------------|-------- + * true | true | True + * false | false | False + * true | false | x + * false | true | not(x) + */ + if (is_Mux(left)) { + ir_node *mux_true = get_Mux_true(left); + ir_node *mux_false = get_Mux_false(left); + if (is_Const(mux_true) && is_Const(mux_false)) { + /* we can fold true/false constant separately */ + ir_tarval *tv_true = get_Const_tarval(mux_true); + ir_tarval *tv_false = get_Const_tarval(mux_false); + ir_relation r_true = tarval_cmp(tv_true, tv); + ir_relation r_false = tarval_cmp(tv_false, tv); + if (r_true != ir_relation_false + || r_false != ir_relation_false) { + bool rel_true = (r_true & relation) != 0; + bool rel_false = (r_false & relation) != 0; + ir_node *cond = get_Mux_sel(left); + if (rel_true == rel_false) { + relation = rel_true ? ir_relation_true + : ir_relation_false; + } else if (rel_true) { + return cond; + } else { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_node *notn = new_rd_Not(dbgi, block, cond, mode_b); + return notn; + } + } + } + } + /* TODO extend to arbitrary constants */ if (is_Conv(left) && tarval_is_null(tv)) { ir_node *op = get_Conv_op(left); @@ -4891,7 +4905,7 @@ static ir_node *transform_node_Cmp(ir_node *n) if (tarval_is_single_bit(tv)) { /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */ ir_node *v = get_binop_left(op); - ir_node *blk = get_irn_n(op, -1); + ir_node *blk = get_nodes_block(op); ir_graph *irg = get_irn_irg(op); ir_mode *mode = get_irn_mode(v); @@ -5017,7 +5031,7 @@ static ir_node *transform_node_Proj(ir_node *proj) /** * Test whether a block is unreachable * Note: That this only returns true when - * IR_GRAPH_STATE_OPTIMIZE_UNREACHABLE_CODE is set. + * IR_GRAPH_CONSTRAINT_OPTIMIZE_UNREACHABLE_CODE is set. * This is important, as you easily end up producing invalid constructs in the * unreachable code when optimizing away edges into the unreachable code. * So only set this flag when you iterate localopts to the fixpoint. @@ -5028,7 +5042,7 @@ static ir_node *transform_node_Proj(ir_node *proj) static bool is_block_unreachable(const ir_node *block) { const ir_graph *irg = get_irn_irg(block); - if (!is_irg_state(irg, IR_GRAPH_STATE_OPTIMIZE_UNREACHABLE_CODE)) + if (!irg_is_constrained(irg, IR_GRAPH_CONSTRAINT_OPTIMIZE_UNREACHABLE_CODE)) return false; return get_Block_dom_depth(block) < 0; } @@ -5040,7 +5054,7 @@ static ir_node *transform_node_Block(ir_node *block) ir_node *bad = NULL; int i; - if (!is_irg_state(irg, IR_GRAPH_STATE_OPTIMIZE_UNREACHABLE_CODE)) + if (!irg_is_constrained(irg, IR_GRAPH_CONSTRAINT_OPTIMIZE_UNREACHABLE_CODE)) return block; for (i = 0; i < arity; ++i) { @@ -5105,7 +5119,6 @@ static ir_node *transform_node_Phi(ir_node *phi) return phi; /* Move the Pin nodes "behind" the Phi. */ - block = get_irn_n(phi, -1); new_phi = new_r_Phi(block, n, in, mode_M); return new_r_Pin(block, new_phi); } @@ -5149,7 +5162,6 @@ static ir_node *transform_node_Phi(ir_node *phi) return phi; /* move the Confirm nodes "behind" the Phi */ - block = get_irn_n(phi, -1); new_phi = new_r_Phi(block, n, in, get_irn_mode(phi)); return new_r_Confirm(block, new_phi, bound, relation); } @@ -5157,9 +5169,6 @@ static ir_node *transform_node_Phi(ir_node *phi) return phi; } -/* forward */ -static ir_node *transform_node(ir_node *n); - /** * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl. * @@ -5248,7 +5257,7 @@ static ir_node *transform_node_shift(ir_node *n) DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT); - return transform_node(irn); + return irn; } /** @@ -5769,6 +5778,92 @@ bool ir_is_optimizable_mux(const ir_node *sel, const ir_node *mux_false, return false; } +/** + * Optimize a Mux(c, 0, 1) node (sometimes called a "set" instruction) + */ +static ir_node *transform_Mux_set(ir_node *n) +{ + ir_node *cond = get_Mux_sel(n); + ir_mode *dest_mode; + ir_mode *mode; + ir_node *left; + ir_node *right; + ir_relation relation; + bool need_not; + dbg_info *dbgi; + ir_node *block; + ir_graph *irg; + ir_node *a; + ir_node *b; + unsigned bits; + ir_tarval *tv; + ir_node *shift_cnt; + ir_node *res; + + if (!is_Cmp(cond)) + return n; + left = get_Cmp_left(cond); + mode = get_irn_mode(left); + if (!mode_is_int(mode) && !mode_is_reference(mode)) + return n; + dest_mode = get_irn_mode(n); + if (!mode_is_int(dest_mode) && !mode_is_reference(dest_mode)) + return n; + right = get_Cmp_right(cond); + relation = get_Cmp_relation(cond) & ~ir_relation_unordered; + if (get_mode_size_bits(mode) >= get_mode_size_bits(dest_mode) + && !(mode_is_signed(mode) && is_Const(right) && is_Const_null(right) + && relation != ir_relation_greater)) + return n; + + need_not = false; + switch (relation) { + case ir_relation_less: + /* a < b -> (a - b) >> 31 */ + a = left; + b = right; + break; + case ir_relation_less_equal: + /* a <= b -> ~(a - b) >> 31 */ + a = right; + b = left; + need_not = true; + break; + case ir_relation_greater: + /* a > b -> (b - a) >> 31 */ + a = right; + b = left; + break; + case ir_relation_greater_equal: + /* a >= b -> ~(a - b) >> 31 */ + a = left; + b = right; + need_not = true; + break; + default: + return n; + } + + dbgi = get_irn_dbg_info(n); + block = get_nodes_block(n); + irg = get_irn_irg(block); + bits = get_mode_size_bits(dest_mode); + tv = new_tarval_from_long(bits-1, mode_Iu); + shift_cnt = new_rd_Const(dbgi, irg, tv); + + if (mode != dest_mode) { + a = new_rd_Conv(dbgi, block, a, dest_mode); + b = new_rd_Conv(dbgi, block, b, dest_mode); + } + + res = new_rd_Sub(dbgi, block, a, b, dest_mode); + if (need_not) { + res = new_rd_Not(dbgi, block, res, dest_mode); + } + res = new_rd_Shr(dbgi, block, res, shift_cnt, dest_mode); + return res; +} + /** * Optimize a Mux into some simpler cases. */ @@ -5817,12 +5912,18 @@ static ir_node *transform_node_Mux(ir_node *n) relation = get_negated_relation(relation); sel = new_rd_Cmp(seldbgi, block, get_Cmp_left(sel), get_Cmp_right(sel), relation); - n = new_rd_Mux(get_irn_dbg_info(n), get_nodes_block(n), sel, f, t, mode); + return new_rd_Mux(get_irn_dbg_info(n), get_nodes_block(n), sel, f, t, mode); + } + + if (is_Const(f) && is_Const_null(f) && is_Const(t) && is_Const_one(t)) { + n = transform_Mux_set(n); + if (n != oldn) + return n; } /* the following optimisations create new mode_b nodes, so only do them * before mode_b lowering */ - if (!is_irg_state(irg, IR_GRAPH_STATE_MODEB_LOWERED)) { + if (!irg_is_constrained(irg, IR_GRAPH_CONSTRAINT_MODEB_LOWERED)) { if (is_Mux(t)) { ir_node* block = get_nodes_block(n); ir_node* c0 = sel; @@ -5831,23 +5932,15 @@ static ir_node *transform_node_Mux(ir_node *n) ir_node* f1 = get_Mux_false(t); if (f == f1) { /* Mux(cond0, Mux(cond1, x, y), y) => Mux(cond0 && cond1, x, y) */ - ir_node* and_ = new_r_And(block, c0, c1, mode_b); - ir_node* new_mux = new_r_Mux(block, and_, f1, t1, mode); - n = new_mux; - sel = and_; - f = f1; - t = t1; - DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE); + ir_node* and_ = new_r_And(block, c0, c1, mode_b); + DBG_OPT_ALGSIM0(oldn, t1, FS_OPT_MUX_COMBINE); + return new_r_Mux(block, and_, f1, t1, mode); } else if (f == t1) { /* Mux(cond0, Mux(cond1, x, y), x) */ ir_node* not_c1 = new_r_Not(block, c1, mode_b); ir_node* and_ = new_r_And(block, c0, not_c1, mode_b); - ir_node* new_mux = new_r_Mux(block, and_, t1, f1, mode); - n = new_mux; - sel = and_; - f = t1; - t = f1; - DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE); + DBG_OPT_ALGSIM0(oldn, f1, FS_OPT_MUX_COMBINE); + return new_r_Mux(block, and_, t1, f1, mode); } } else if (is_Mux(f)) { ir_node* block = get_nodes_block(n); @@ -5857,23 +5950,15 @@ static ir_node *transform_node_Mux(ir_node *n) ir_node* f1 = get_Mux_false(f); if (t == t1) { /* Mux(cond0, x, Mux(cond1, x, y)) -> typical if (cond0 || cond1) x else y */ - ir_node* or_ = new_r_Or(block, c0, c1, mode_b); - ir_node* new_mux = new_r_Mux(block, or_, f1, t1, mode); - n = new_mux; - sel = or_; - f = f1; - t = t1; - DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE); + ir_node* or_ = new_r_Or(block, c0, c1, mode_b); + DBG_OPT_ALGSIM0(oldn, f1, FS_OPT_MUX_COMBINE); + return new_r_Mux(block, or_, f1, t1, mode); } else if (t == f1) { /* Mux(cond0, x, Mux(cond1, y, x)) */ ir_node* not_c1 = new_r_Not(block, c1, mode_b); ir_node* or_ = new_r_Or(block, c0, not_c1, mode_b); - ir_node* new_mux = new_r_Mux(block, or_, t1, f1, mode); - n = new_mux; - sel = or_; - f = t1; - t = f1; - DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE); + DBG_OPT_ALGSIM0(oldn, t1, FS_OPT_MUX_COMBINE); + return new_r_Mux(block, or_, t1, f1, mode); } } @@ -5914,28 +5999,6 @@ static ir_node *transform_node_Mux(ir_node *n) } } } - - /* more normalization: Mux(sel, 0, 1) is simply a conv from the mode_b - * value to integer. */ - if (is_Const(t) && is_Const(f) && mode_is_int(mode)) { - ir_tarval *a = get_Const_tarval(t); - ir_tarval *b = get_Const_tarval(f); - - if (tarval_is_one(a) && tarval_is_null(b)) { - ir_node *block = get_nodes_block(n); - ir_node *conv = new_r_Conv(block, sel, mode); - n = conv; - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV); - return n; - } else if (tarval_is_null(a) && tarval_is_one(b)) { - ir_node *block = get_nodes_block(n); - ir_node *not_ = new_r_Not(block, sel, mode_b); - ir_node *conv = new_r_Conv(block, not_, mode); - n = conv; - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV); - return n; - } - } } if (is_Cmp(sel) && mode_is_int(mode) && is_cmp_equality_zero(sel)) { @@ -6032,7 +6095,8 @@ static ir_node *transform_node_Sync(ir_node *n) ++arity; break; } - if (get_Sync_pred(n, k) == pred_pred) break; + if (get_Sync_pred(n, k) == pred_pred) + break; } } } @@ -6188,387 +6252,186 @@ static ir_node *transform_node_Call(ir_node *call) */ static ir_node *transform_node(ir_node *n) { - ir_node *oldn; - - /* - * Transform_node is the only "optimizing transformation" that might - * return a node with a different opcode. We iterate HERE until fixpoint - * to get the final result. - */ - do { - oldn = n; - if (n->op->ops.transform_node != NULL) - n = n->op->ops.transform_node(n); - } while (oldn != n); - - return n; -} - -/** - * Sets the default transform node operation for an ir_op_ops. - * - * @param code the opcode for the default operation - * @param ops the operations initialized - * - * @return - * The operations. - */ -static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops) -{ -#define CASE(a) \ - case iro_##a: \ - ops->transform_node = transform_node_##a; \ - break -#define CASE_PROJ(a) \ - case iro_##a: \ - ops->transform_node_Proj = transform_node_Proj_##a; \ - break -#define CASE_PROJ_EX(a) \ - case iro_##a: \ - ops->transform_node = transform_node_##a; \ - ops->transform_node_Proj = transform_node_Proj_##a; \ - break - - switch (code) { - CASE(Add); - CASE(And); - CASE(Block); - CASE(Call); - CASE(Cmp); - CASE(Conv); - CASE(End); - CASE(Eor); - CASE(Minus); - CASE(Mul); - CASE(Mux); - CASE(Not); - CASE(Or); - CASE(Phi); - CASE(Proj); - CASE(Rotl); - CASE(Sel); - CASE(Shl); - CASE(Shr); - CASE(Shrs); - CASE(Sub); - CASE(Sync); - CASE_PROJ(Bound); - CASE_PROJ(CopyB); - CASE_PROJ(Store); - CASE_PROJ_EX(Cond); - CASE_PROJ_EX(Div); - CASE_PROJ_EX(Load); - CASE_PROJ_EX(Mod); - default: - break; - } - - return ops; -#undef CASE_PROJ_EX -#undef CASE_PROJ -#undef CASE -} - - -/* **************** Common Subexpression Elimination **************** */ - -/** The size of the hash table used, should estimate the number of nodes - in a graph. */ -#define N_IR_NODES 512 - -/** Compares two exception attributes */ -static int node_cmp_exception(const ir_node *a, const ir_node *b) -{ - const except_attr *ea = &a->attr.except; - const except_attr *eb = &b->attr.except; - return ea->pin_state != eb->pin_state; -} - -/** Compares the attributes of two Const nodes. */ -static int node_cmp_attr_Const(const ir_node *a, const ir_node *b) -{ - return get_Const_tarval(a) != get_Const_tarval(b); -} - -/** Compares the attributes of two Proj nodes. */ -static int node_cmp_attr_Proj(const ir_node *a, const ir_node *b) -{ - return a->attr.proj.proj != b->attr.proj.proj; -} - -/** Compares the attributes of two Alloc nodes. */ -static int node_cmp_attr_Alloc(const ir_node *a, const ir_node *b) -{ - const alloc_attr *pa = &a->attr.alloc; - const alloc_attr *pb = &b->attr.alloc; - if (pa->where != pb->where || pa->type != pb->type) - return 1; - return node_cmp_exception(a, b); -} - -/** Compares the attributes of two Free nodes. */ -static int node_cmp_attr_Free(const ir_node *a, const ir_node *b) -{ - const free_attr *pa = &a->attr.free; - const free_attr *pb = &b->attr.free; - return (pa->where != pb->where) || (pa->type != pb->type); -} - -/** Compares the attributes of two SymConst nodes. */ -static int node_cmp_attr_SymConst(const ir_node *a, const ir_node *b) -{ - const symconst_attr *pa = &a->attr.symc; - const symconst_attr *pb = &b->attr.symc; - return (pa->kind != pb->kind) - || (pa->sym.type_p != pb->sym.type_p); -} - -/** Compares the attributes of two Call nodes. */ -static int node_cmp_attr_Call(const ir_node *a, const ir_node *b) -{ - const call_attr *pa = &a->attr.call; - const call_attr *pb = &b->attr.call; - if (pa->type != pb->type || pa->tail_call != pb->tail_call) - return 1; - return node_cmp_exception(a, b); -} - -/** Compares the attributes of two Sel nodes. */ -static int node_cmp_attr_Sel(const ir_node *a, const ir_node *b) -{ - const ir_entity *a_ent = get_Sel_entity(a); - const ir_entity *b_ent = get_Sel_entity(b); - return a_ent != b_ent; -} - -/** Compares the attributes of two Phi nodes. */ -static int node_cmp_attr_Phi(const ir_node *a, const ir_node *b) -{ - /* we can only enter this function if both nodes have the same number of inputs, - hence it is enough to check if one of them is a Phi0 */ - if (is_Phi0(a)) { - /* check the Phi0 pos attribute */ - return a->attr.phi.u.pos != b->attr.phi.u.pos; - } - return 0; -} - -/** Compares the attributes of two Conv nodes. */ -static int node_cmp_attr_Conv(const ir_node *a, const ir_node *b) -{ - return get_Conv_strict(a) != get_Conv_strict(b); -} - -/** Compares the attributes of two Cast nodes. */ -static int node_cmp_attr_Cast(const ir_node *a, const ir_node *b) -{ - return get_Cast_type(a) != get_Cast_type(b); -} - -/** Compares the attributes of two Load nodes. */ -static int node_cmp_attr_Load(const ir_node *a, const ir_node *b) -{ - if (get_Load_volatility(a) == volatility_is_volatile || - get_Load_volatility(b) == volatility_is_volatile) - /* NEVER do CSE on volatile Loads */ - return 1; - /* do not CSE Loads with different alignment. Be conservative. */ - if (get_Load_unaligned(a) != get_Load_unaligned(b)) - return 1; - if (get_Load_mode(a) != get_Load_mode(b)) - return 1; - return node_cmp_exception(a, b); -} - -/** Compares the attributes of two Store nodes. */ -static int node_cmp_attr_Store(const ir_node *a, const ir_node *b) -{ - /* do not CSE Stores with different alignment. Be conservative. */ - if (get_Store_unaligned(a) != get_Store_unaligned(b)) - return 1; - /* NEVER do CSE on volatile Stores */ - if (get_Store_volatility(a) == volatility_is_volatile || - get_Store_volatility(b) == volatility_is_volatile) - return 1; - return node_cmp_exception(a, b); -} - -static int node_cmp_attr_CopyB(const ir_node *a, const ir_node *b) -{ - if (get_CopyB_type(a) != get_CopyB_type(b)) - return 1; - - return node_cmp_exception(a, b); -} - -static int node_cmp_attr_Bound(const ir_node *a, const ir_node *b) -{ - return node_cmp_exception(a, b); -} - -/** Compares the attributes of two Div nodes. */ -static int node_cmp_attr_Div(const ir_node *a, const ir_node *b) -{ - const div_attr *ma = &a->attr.div; - const div_attr *mb = &b->attr.div; - if (ma->resmode != mb->resmode || ma->no_remainder != mb->no_remainder) - return 1; - return node_cmp_exception(a, b); -} - -/** Compares the attributes of two Mod nodes. */ -static int node_cmp_attr_Mod(const ir_node *a, const ir_node *b) -{ - const mod_attr *ma = &a->attr.mod; - const mod_attr *mb = &b->attr.mod; - if (ma->resmode != mb->resmode) - return 1; - return node_cmp_exception(a, b); -} - -static int node_cmp_attr_Cmp(const ir_node *a, const ir_node *b) -{ - const cmp_attr *ma = &a->attr.cmp; - const cmp_attr *mb = &b->attr.cmp; - return ma->relation != mb->relation; -} - -/** Compares the attributes of two Confirm nodes. */ -static int node_cmp_attr_Confirm(const ir_node *a, const ir_node *b) -{ - const confirm_attr *ma = &a->attr.confirm; - const confirm_attr *mb = &b->attr.confirm; - return ma->relation != mb->relation; -} - -/** Compares the attributes of two Builtin nodes. */ -static int node_cmp_attr_Builtin(const ir_node *a, const ir_node *b) -{ - if (get_Builtin_kind(a) != get_Builtin_kind(b)) - return 1; - if (get_Builtin_type(a) != get_Builtin_type(b)) - return 1; - return node_cmp_exception(a, b); -} - -/** Compares the attributes of two ASM nodes. */ -static int node_cmp_attr_ASM(const ir_node *a, const ir_node *b) -{ - int i, n; - const ir_asm_constraint *ca; - const ir_asm_constraint *cb; - ident **cla, **clb; - - if (get_ASM_text(a) != get_ASM_text(b)) - return 1; + ir_node *old_n; + unsigned iro; +restart: + old_n = n; + iro = get_irn_opcode_(n); + /* constant expression evaluation / constant folding */ + if (get_opt_constant_folding()) { + /* neither constants nor Tuple values can be evaluated */ + if (iro != iro_Const && get_irn_mode(n) != mode_T) { + /* try to evaluate */ + ir_tarval *tv = computed_value(n); + if (tv != tarval_bad) { + /* evaluation was successful -- replace the node. */ + ir_graph *irg = get_irn_irg(n); - /* Should we really check the constraints here? Should be better, but is strange. */ - n = get_ASM_n_input_constraints(a); - if (n != get_ASM_n_input_constraints(b)) - return 1; + n = new_r_Const(irg, tv); - ca = get_ASM_input_constraints(a); - cb = get_ASM_input_constraints(b); - for (i = 0; i < n; ++i) { - if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint - || ca[i].mode != cb[i].mode) - return 1; + DBG_OPT_CSTEVAL(old_n, n); + return n; + } + } } - n = get_ASM_n_output_constraints(a); - if (n != get_ASM_n_output_constraints(b)) - return 1; - - ca = get_ASM_output_constraints(a); - cb = get_ASM_output_constraints(b); - for (i = 0; i < n; ++i) { - if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint - || ca[i].mode != cb[i].mode) - return 1; + /* remove unnecessary nodes */ + if (get_opt_constant_folding() || + (iro == iro_Phi) || /* always optimize these nodes. */ + (iro == iro_Id) || /* ... */ + (iro == iro_Proj) || /* ... */ + (iro == iro_Block)) { /* Flags tested local. */ + n = equivalent_node(n); + if (n != old_n) + goto restart; } - n = get_ASM_n_clobbers(a); - if (n != get_ASM_n_clobbers(b)) - return 1; - - cla = get_ASM_clobbers(a); - clb = get_ASM_clobbers(b); - for (i = 0; i < n; ++i) { - if (cla[i] != clb[i]) - return 1; + /* Some more constant expression evaluation. */ + if (get_opt_algebraic_simplification() || + (iro == iro_Cond) || + (iro == iro_Proj)) { /* Flags tested local. */ + if (n->op->ops.transform_node != NULL) { + n = n->op->ops.transform_node(n); + if (n != old_n) { + goto restart; + } + } } - return node_cmp_exception(a, b); + return n; } -/** Compares the inexistent attributes of two Dummy nodes. */ -static int node_cmp_attr_Dummy(const ir_node *a, const ir_node *b) -{ - (void) a; - (void) b; - /* Dummy nodes never equal by definition */ - return 1; +static void register_computed_value_func(ir_op *op, computed_value_func func) +{ + assert(op->ops.computed_value == NULL || op->ops.computed_value == func); + op->ops.computed_value = func; +} + +static void register_computed_value_func_proj(ir_op *op, + computed_value_func func) +{ + assert(op->ops.computed_value_Proj == NULL + || op->ops.computed_value_Proj == func); + op->ops.computed_value_Proj = func; +} + +static void register_equivalent_node_func(ir_op *op, equivalent_node_func func) +{ + assert(op->ops.equivalent_node == NULL || op->ops.equivalent_node == func); + op->ops.equivalent_node = func; +} + +static void register_equivalent_node_func_proj(ir_op *op, + equivalent_node_func func) +{ + assert(op->ops.equivalent_node_Proj == NULL + || op->ops.equivalent_node_Proj == func); + op->ops.equivalent_node_Proj = func; +} + +static void register_transform_node_func(ir_op *op, transform_node_func func) +{ + assert(op->ops.transform_node == NULL || op->ops.transform_node == func); + op->ops.transform_node = func; +} + +static void register_transform_node_func_proj(ir_op *op, + transform_node_func func) +{ + assert(op->ops.transform_node_Proj == NULL + || op->ops.transform_node_Proj == func); + op->ops.transform_node_Proj = func; +} + +void ir_register_opt_node_ops(void) +{ + register_computed_value_func(op_Add, computed_value_Add); + register_computed_value_func(op_And, computed_value_And); + register_computed_value_func(op_Borrow, computed_value_Borrow); + register_computed_value_func(op_Carry, computed_value_Carry); + register_computed_value_func(op_Cmp, computed_value_Cmp); + register_computed_value_func(op_Confirm, computed_value_Confirm); + register_computed_value_func(op_Const, computed_value_Const); + register_computed_value_func(op_Conv, computed_value_Conv); + register_computed_value_func(op_Eor, computed_value_Eor); + register_computed_value_func(op_Minus, computed_value_Minus); + register_computed_value_func(op_Mul, computed_value_Mul); + register_computed_value_func(op_Mux, computed_value_Mux); + register_computed_value_func(op_Not, computed_value_Not); + register_computed_value_func(op_Or, computed_value_Or); + register_computed_value_func(op_Proj, computed_value_Proj); + register_computed_value_func(op_Rotl, computed_value_Rotl); + register_computed_value_func(op_Shl, computed_value_Shl); + register_computed_value_func(op_Shr, computed_value_Shr); + register_computed_value_func(op_Shrs, computed_value_Shrs); + register_computed_value_func(op_Sub, computed_value_Sub); + register_computed_value_func(op_SymConst, computed_value_SymConst); + register_computed_value_func_proj(op_Div, computed_value_Proj_Div); + register_computed_value_func_proj(op_Mod, computed_value_Proj_Mod); + + register_equivalent_node_func(op_Add, equivalent_node_Add); + register_equivalent_node_func(op_And, equivalent_node_And); + register_equivalent_node_func(op_Confirm, equivalent_node_Confirm); + register_equivalent_node_func(op_Conv, equivalent_node_Conv); + register_equivalent_node_func(op_Eor, equivalent_node_Eor); + register_equivalent_node_func(op_Id, equivalent_node_Id); + register_equivalent_node_func(op_Minus, equivalent_node_idempotent_unop); + register_equivalent_node_func(op_Mul, equivalent_node_Mul); + register_equivalent_node_func(op_Mux, equivalent_node_Mux); + register_equivalent_node_func(op_Not, equivalent_node_idempotent_unop); + register_equivalent_node_func(op_Or, equivalent_node_Or); + register_equivalent_node_func(op_Phi, equivalent_node_Phi); + register_equivalent_node_func(op_Proj, equivalent_node_Proj); + register_equivalent_node_func(op_Rotl, equivalent_node_left_zero); + register_equivalent_node_func(op_Shl, equivalent_node_left_zero); + register_equivalent_node_func(op_Shr, equivalent_node_left_zero); + register_equivalent_node_func(op_Shrs, equivalent_node_left_zero); + register_equivalent_node_func(op_Sub, equivalent_node_Sub); + register_equivalent_node_func_proj(op_Bound, equivalent_node_Proj_Bound); + register_equivalent_node_func_proj(op_CopyB, equivalent_node_Proj_CopyB); + register_equivalent_node_func_proj(op_Div, equivalent_node_Proj_Div); + register_equivalent_node_func_proj(op_Tuple, equivalent_node_Proj_Tuple); + + register_transform_node_func(op_Add, transform_node_Add); + register_transform_node_func(op_And, transform_node_And); + register_transform_node_func(op_Block, transform_node_Block); + register_transform_node_func(op_Call, transform_node_Call); + register_transform_node_func(op_Cmp, transform_node_Cmp); + register_transform_node_func(op_Cond, transform_node_Cond); + register_transform_node_func(op_Conv, transform_node_Conv); + register_transform_node_func(op_Div, transform_node_Div); + register_transform_node_func(op_End, transform_node_End); + register_transform_node_func(op_Eor, transform_node_Eor); + register_transform_node_func(op_Load, transform_node_Load); + register_transform_node_func(op_Minus, transform_node_Minus); + register_transform_node_func(op_Mod, transform_node_Mod); + register_transform_node_func(op_Mul, transform_node_Mul); + register_transform_node_func(op_Mux, transform_node_Mux); + register_transform_node_func(op_Not, transform_node_Not); + register_transform_node_func(op_Or, transform_node_Or); + register_transform_node_func(op_Phi, transform_node_Phi); + register_transform_node_func(op_Proj, transform_node_Proj); + register_transform_node_func(op_Rotl, transform_node_Rotl); + register_transform_node_func(op_Shl, transform_node_Shl); + register_transform_node_func(op_Shrs, transform_node_Shrs); + register_transform_node_func(op_Shr, transform_node_Shr); + register_transform_node_func(op_Sub, transform_node_Sub); + register_transform_node_func(op_Switch, transform_node_Switch); + register_transform_node_func(op_Sync, transform_node_Sync); + register_transform_node_func_proj(op_Bound, transform_node_Proj_Bound); + register_transform_node_func_proj(op_CopyB, transform_node_Proj_CopyB); + register_transform_node_func_proj(op_Div, transform_node_Proj_Div); + register_transform_node_func_proj(op_Load, transform_node_Proj_Load); + register_transform_node_func_proj(op_Mod, transform_node_Proj_Mod); + register_transform_node_func_proj(op_Store, transform_node_Proj_Store); } -static int node_cmp_attr_InstOf(const ir_node *a, const ir_node *b) -{ - if (get_InstOf_type(a) != get_InstOf_type(b)) - return 1; - return node_cmp_exception(a, b); -} - -/** - * Set the default node attribute compare operation for an ir_op_ops. - * - * @param code the opcode for the default operation - * @param ops the operations initialized - * - * @return - * The operations. - */ -static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops) -{ -#define CASE(a) \ - case iro_##a: \ - ops->node_cmp_attr = node_cmp_attr_##a; \ - break - - switch (code) { - CASE(ASM); - CASE(Alloc); - CASE(Bound); - CASE(Builtin); - CASE(Call); - CASE(Cast); - CASE(Cmp); - CASE(Confirm); - CASE(Const); - CASE(Conv); - CASE(CopyB); - CASE(Div); - CASE(Dummy); - CASE(Free); - CASE(InstOf); - CASE(Load); - CASE(Mod); - CASE(Phi); - CASE(Proj); - CASE(Sel); - CASE(Store); - CASE(SymConst); - default: - /* leave NULL */ - break; - } +/* **************** Common Subexpression Elimination **************** */ - return ops; -#undef CASE -} +/** The size of the hash table used, should estimate the number of nodes + in a graph. */ +#define N_IR_NODES 512 -/* - * Compare function for two nodes in the value table. Gets two - * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression. - */ int identities_cmp(const void *elt, const void *key) { ir_node *a = (ir_node *)elt; @@ -6591,7 +6454,7 @@ int identities_cmp(const void *elt, const void *key) if (get_irn_pinned(a) == op_pin_state_pinned) { /* for pinned nodes, the block inputs must be equal */ - if (get_irn_n(a, -1) != get_irn_n(b, -1)) + if (get_nodes_block(a) != get_nodes_block(b)) return 1; } else { ir_node *block_a = get_nodes_block(a); @@ -6633,17 +6496,11 @@ int identities_cmp(const void *elt, const void *key) return 0; } -/* - * Calculate a hash value of a node. - * - * @param node The IR-node - */ unsigned ir_node_hash(const ir_node *node) { return node->op->ops.hash(node); } - void new_identities(ir_graph *irg) { if (irg->value_table != NULL) @@ -6657,8 +6514,15 @@ void del_identities(ir_graph *irg) del_pset(irg->value_table); } -/* Normalize a node by putting constants (and operands with larger - * node index) on the right (operator side). */ +static int cmp_node_nr(const void *a, const void *b) +{ + ir_node **p1 = (ir_node**)a; + ir_node **p2 = (ir_node**)b; + long n1 = get_irn_node_nr(*p1); + long n2 = get_irn_node_nr(*p2); + return (n1>n2) - (n1value_table, ir_node*, node) { + foreach_pset(irg->value_table, ir_node, node) { visit(node, env); } current_ir_graph = rem; } -/** - * These optimizations deallocate nodes from the obstack. - * It can only be called if it is guaranteed that no other nodes - * reference this one, i.e., right after construction of a node. - * - * @param n The node to optimize - */ ir_node *optimize_node(ir_node *n) { ir_node *oldn = n; @@ -6834,12 +6702,13 @@ ir_node *optimize_node(ir_node *n) free the node. */ iro = get_irn_opcode(n); if (get_opt_algebraic_simplification() || - (iro == iro_Cond) || - (iro == iro_Proj)) /* Flags tested local. */ + (iro == iro_Cond) || + (iro == iro_Proj)) { /* Flags tested local. */ n = transform_node(n); + } /* Now we have a legal, useful node. Enter it in hash table for CSE */ - if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) { + if (get_opt_cse()) { ir_node *o = n; n = identify_remember(o); if (o != n) @@ -6849,75 +6718,40 @@ ir_node *optimize_node(ir_node *n) return n; } - -/** - * These optimizations never deallocate nodes (in place). This can cause dead - * nodes lying on the obstack. Remove these by a dead node elimination, - * i.e., a copying garbage collection. - */ ir_node *optimize_in_place_2(ir_node *n) { - ir_tarval *tv; - ir_node *oldn = n; - unsigned iro = get_irn_opcode(n); - if (!get_opt_optimize() && !is_Phi(n)) return n; - if (iro == iro_Deleted) + if (is_Deleted(n)) return n; - /* constant expression evaluation / constant folding */ - if (get_opt_constant_folding()) { - /* neither constants nor Tuple values can be evaluated */ - if (iro != iro_Const && get_irn_mode(n) != mode_T) { - /* try to evaluate */ - tv = computed_value(n); - if (tv != tarval_bad) { - /* evaluation was successful -- replace the node. */ - ir_graph *irg = get_irn_irg(n); - - n = new_r_Const(irg, tv); - - DBG_OPT_CSTEVAL(oldn, n); - return n; - } - } - } - - /* remove unnecessary nodes */ - if (get_opt_constant_folding() || - (iro == iro_Phi) || /* always optimize these nodes. */ - (iro == iro_Id) || /* ... */ - (iro == iro_Proj) || /* ... */ - (iro == iro_Block) ) /* Flags tested local. */ - n = equivalent_node(n); - /** common subexpression elimination **/ /* Checks whether n is already available. */ - /* The block input is used to distinguish different subexpressions. Right - now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common - subexpressions within a block. */ + /* The block input is used to distinguish different subexpressions. + * Right now all nodes are op_pin_state_pinned to blocks, i.e., the cse + * only finds common subexpressions within a block. */ if (get_opt_cse()) { ir_node *o = n; - n = identify_remember(o); - if (o != n) + n = identify_remember(n); + if (n != o) { DBG_OPT_CSE(o, n); + /* we have another existing node now, we do not optimize it here */ + return n; + } } - /* Some more constant expression evaluation. */ - iro = get_irn_opcode(n); - if (get_opt_constant_folding() || - (iro == iro_Cond) || - (iro == iro_Proj)) /* Flags tested local. */ - n = transform_node(n); + n = transform_node(n); /* Now we can verify the node, as it has no dead inputs any more. */ irn_verify(n); /* Now we have a legal, useful node. Enter it in hash table for cse. - Blocks should be unique anyways. (Except the successor of start: - is cse with the start block!) */ - if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) { + * + * Note: This is only necessary because some of the optimisations + * operate in-place (set_XXX_bla, turn_into_tuple, ...) which is considered + * bad practice and should be fixed sometime. + */ + if (get_opt_cse()) { ir_node *o = n; n = identify_remember(o); if (o != n) @@ -6927,9 +6761,6 @@ ir_node *optimize_in_place_2(ir_node *n) return n; } -/** - * Wrapper for external use, set proper status bits after optimization. - */ ir_node *optimize_in_place(ir_node *n) { ir_graph *irg = get_irn_irg(n); @@ -6941,80 +6772,6 @@ ir_node *optimize_in_place(ir_node *n) /* FIXME: Maybe we could also test whether optimizing the node can change the control graph. */ - clear_irg_state(irg, IR_GRAPH_STATE_CONSISTENT_DOMINANCE); + clear_irg_properties(irg, IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE); return optimize_in_place_2(n); } - -/** - * Calculate a hash value of a Const node. - */ -static unsigned hash_Const(const ir_node *node) -{ - unsigned h; - - /* special value for const, as they only differ in their tarval. */ - h = HASH_PTR(node->attr.con.tarval); - - return h; -} - -/** - * Calculate a hash value of a SymConst node. - */ -static unsigned hash_SymConst(const ir_node *node) -{ - unsigned h; - - /* all others are pointers */ - h = HASH_PTR(node->attr.symc.sym.type_p); - - return h; -} - -/** - * Set the default hash operation in an ir_op_ops. - * - * @param code the opcode for the default operation - * @param ops the operations initialized - * - * @return - * The operations. - */ -static ir_op_ops *firm_set_default_hash(unsigned code, ir_op_ops *ops) -{ -#define CASE(a) \ - case iro_##a: \ - ops->hash = hash_##a; \ - break - - /* hash function already set */ - if (ops->hash != NULL) - return ops; - - switch (code) { - CASE(Const); - CASE(SymConst); - default: - /* use input/mode default hash if no function was given */ - ops->hash = firm_default_hash; - } - - return ops; -#undef CASE -} - -/* - * Sets the default operation for an ir_ops. - */ -ir_op_ops *firm_set_default_operations(unsigned code, ir_op_ops *ops) -{ - ops = firm_set_default_hash(code, ops); - ops = firm_set_default_computed_value(code, ops); - ops = firm_set_default_equivalent_node(code, ops); - ops = firm_set_default_transform_node(code, ops); - ops = firm_set_default_node_cmp_attr(code, ops); - ops = firm_set_default_get_type_attr(code, ops); - ops = firm_set_default_get_entity_attr(code, ops); - - return ops; -}