X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firopt.c;h=3672f0412b5784784e75823a3dddaf0142841583;hb=60c9f323685e93f335d0f26aad2da647b669886d;hp=a2f6792251ad42d7c05c78c62bebe40592e271d6;hpb=362a79a8ed6471288da3449c5b112c49321b0c42;p=libfirm diff --git a/ir/ir/iropt.c b/ir/ir/iropt.c index a2f679225..3672f0412 100644 --- a/ir/ir/iropt.c +++ b/ir/ir/iropt.c @@ -11,7 +11,17 @@ */ #ifdef HAVE_CONFIG_H -# include +# include "config.h" +#endif + +#ifdef HAVE_ALLOCA_H +#include +#endif +#ifdef HAVE_MALLOC_H +#include +#endif +#ifdef HAVE_STRING_H +#include #endif # include "irnode_t.h" @@ -27,10 +37,17 @@ # include "irflag_t.h" # include "firmstat.h" # include "irarch.h" +# include "hashptr.h" /* Make types visible to allow most efficient access */ # include "entity_t.h" +# ifdef DO_HEAPANALYSIS +/* heapanal can't cope with NoMems */ +# else /* if defined DO_HEAPANALYSIS */ +# define USE_NOMEM +# endif /* defined DO_HEAPANALYSIS */ + /** * Trivial INLINEable routine for copy propagation. * Does follow Ids, needed to optimize INLINEd code. @@ -42,18 +59,6 @@ follow_Id (ir_node *n) return n; } -/** - * Returns the tarval of a Const node or tarval_bad for all other nodes. - */ -static INLINE tarval * -value_of (ir_node *n) -{ - if ((n != NULL) && (get_irn_op(n) == op_Const)) - return get_Const_tarval(n); /* might return tarval_bad */ - else - return tarval_bad; -} - /** * return the value of a Constant */ @@ -84,11 +89,9 @@ static tarval *computed_value_Add(ir_node *n) tarval *ta = value_of(a); tarval *tb = value_of(b); - if ((ta != tarval_bad) && (tb != tarval_bad) - && (get_irn_mode(a) == get_irn_mode(b)) - && !(get_mode_sort(get_irn_mode(a)) == irms_reference)) { + if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) return tarval_add(ta, tb); - } + return tarval_bad; } @@ -104,17 +107,15 @@ static tarval *computed_value_Sub(ir_node *n) tarval *tb; /* a - a */ - if (a == b) + if (a == b && !is_Bad(a)) return get_tarval_null(get_irn_mode(n)); ta = value_of(a); tb = value_of(b); - if ((ta != tarval_bad) && (tb != tarval_bad) - && (get_irn_mode(a) == get_irn_mode(b)) - && !(get_mode_sort(get_irn_mode(a)) == irms_reference)) { + if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) return tarval_sub(ta, tb); - } + return tarval_bad; } @@ -149,14 +150,10 @@ static tarval *computed_value_Mul(ir_node *n) /* a*0 = 0 or 0*b = 0: calls computed_value recursive and returns the 0 with proper mode. */ - tarval *v; - - if ( ( ((v = ta) != tarval_bad) - && (v == get_mode_null(get_tarval_mode(v))) ) - || ( ((v = tb) != tarval_bad) - && (v == get_mode_null(get_tarval_mode(v))) )) { - return v; - } + if ((ta != tarval_bad) && (ta == get_mode_null(get_tarval_mode(ta)))) + return ta; + if ((tb != tarval_bad) && (tb == get_mode_null(get_tarval_mode(tb)))) + return tb; } return tarval_bad; } @@ -190,8 +187,8 @@ static tarval *do_computed_value_Div(ir_node *a, ir_node *b) tarval *tb = value_of(b); /* Compute c1 / c2 or 0 / a, a != 0 */ - if ((ta != tarval_bad) && (tb != get_mode_null(get_irn_mode(b)))) { - if (tb != tarval_bad) /* div by zero: return tarval_bad */ + if (ta != tarval_bad) { + if ((tb != tarval_bad) && (tb != get_mode_null(get_irn_mode(b)))) /* div by zero: return tarval_bad */ return tarval_div(ta, tb); else if (ta == get_mode_null(get_tarval_mode(ta))) /* 0 / b == 0 */ return ta; @@ -266,8 +263,8 @@ static tarval *computed_value_And(ir_node *n) } else { tarval *v; - if ( (classify_tarval ((v = computed_value (a))) == TV_CLASSIFY_NULL) - || (classify_tarval ((v = computed_value (b))) == TV_CLASSIFY_NULL)) { + if ( (classify_tarval ((v = ta)) == TV_CLASSIFY_NULL) + || (classify_tarval ((v = tb)) == TV_CLASSIFY_NULL)) { return v; } } @@ -290,8 +287,8 @@ static tarval *computed_value_Or(ir_node *n) return tarval_or (ta, tb); } else { tarval *v; - if ( (classify_tarval ((v = computed_value (a))) == TV_CLASSIFY_ALL_ONE) - || (classify_tarval ((v = computed_value (b))) == TV_CLASSIFY_ALL_ONE)) { + if ( (classify_tarval ((v = ta)) == TV_CLASSIFY_ALL_ONE) + || (classify_tarval ((v = tb)) == TV_CLASSIFY_ALL_ONE)) { return v; } } @@ -306,8 +303,13 @@ static tarval *computed_value_Eor(ir_node *n) ir_node *a = get_Eor_left(n); ir_node *b = get_Eor_right(n); - tarval *ta = value_of(a); - tarval *tb = value_of(b); + tarval *ta, *tb; + + if (a == b) + return get_tarval_null(get_irn_mode(n)); + + ta = value_of(a); + tb = value_of(b); if ((ta != tarval_bad) && (tb != tarval_bad)) { return tarval_eor (ta, tb); @@ -440,14 +442,15 @@ static tarval *computed_value_Proj(ir_node *n) ab = get_Cmp_right(a); proj_nr = get_Proj_proj(n); - if (aa == ab) { /* 1.: */ + if (aa == ab && !mode_is_float(get_irn_mode(aa))) { /* 1.: */ + /* BEWARE: a == a is NOT always True for floating Point!!! */ /* This is a trick with the bits used for encoding the Cmp Proj numbers, the following statement is not the same: return new_tarval_from_long (proj_nr == Eq, mode_b) */ return new_tarval_from_long (proj_nr & Eq, mode_b); } else { - tarval *taa = computed_value (aa); - tarval *tab = computed_value (ab); + tarval *taa = value_of(aa); + tarval *tab = value_of(ab); if ((taa != tarval_bad) && (tab != tarval_bad)) { /* 2.: */ /* strange checks... */ @@ -510,6 +513,26 @@ static tarval *computed_value_Proj(ir_node *n) return tarval_bad; } +/** + * calculate the value of a Mux: can be evaluated, if the + * sel and the right input are known + */ +static tarval *computed_value_Mux(ir_node *n) +{ + ir_node *sel = get_Mux_sel(n); + tarval *ts = value_of(sel); + + if (ts == get_tarval_b_true()) { + ir_node *v = get_Mux_true(n); + return value_of(v); + } + else if (ts == get_tarval_b_false()) { + ir_node *v = get_Mux_false(n); + return value_of(v); + } + return tarval_bad; +} + /** * If the parameter n can be computed, return its value, else tarval_bad. * Performs constant folding. @@ -554,6 +577,7 @@ static ir_op *firm_set_default_computed_value(ir_op *op) CASE(Rot); CASE(Conv); CASE(Proj); + CASE(Mux); default: op->computed_value = NULL; } @@ -600,7 +624,7 @@ static ir_node *equivalent_node_Block(ir_node *n) ir_node *predblock = get_nodes_block(get_Block_cfgpred(n, 0)); if (predblock == oldn) { /* Jmp jumps into the block it is in -- deal self cycle. */ - n = new_Bad(); + n = set_Block_dead(n); DBG_OPT_DEAD(oldn, n); } else if (get_opt_control_flow_straightening()) { n = predblock; @@ -612,7 +636,7 @@ static ir_node *equivalent_node_Block(ir_node *n) ir_node *predblock = get_nodes_block(get_Block_cfgpred(n, 0)); if (predblock == oldn) { /* Jmp jumps into the block it is in -- deal self cycle. */ - n = new_Bad(); + n = set_Block_dead(n); DBG_OPT_DEAD(oldn, n); } } @@ -636,15 +660,27 @@ static ir_node *equivalent_node_Block(ir_node *n) } else if (get_opt_unreachable_code() && (n != current_ir_graph->start_block) && (n != current_ir_graph->end_block) ) { - int i; + int i, n_cfg = get_Block_n_cfgpreds(n); + /* If all inputs are dead, this block is dead too, except if it is the start or end block. This is a step of unreachable code elimination */ - for (i = 0; i < get_Block_n_cfgpreds(n); i++) { - if (!is_Bad(get_Block_cfgpred(n, i))) break; + for (i = 0; i < n_cfg; i++) { + ir_node *pred = get_Block_cfgpred(n, i); + ir_node *pred_blk; + + if (is_Bad(pred)) continue; + pred_blk = get_nodes_block(pred); + + if (is_Block_dead(pred_blk)) continue; + + if (pred_blk != n) { + /* really found a living input */ + break; + } } - if (i == get_Block_n_cfgpreds(n)) - n = new_Bad(); + if (i == n_cfg) + n = set_Block_dead(n); } return n; @@ -658,7 +694,7 @@ static ir_node *equivalent_node_Jmp(ir_node *n) { /* GL: Why not same for op_Raise?? */ /* unreachable code elimination */ - if (is_Bad(get_nodes_block(n))) + if (is_Block_dead(get_nodes_block(n))) n = new_Bad(); return n; @@ -671,26 +707,6 @@ static ir_node *equivalent_node_Cond(ir_node *n) return n; } -/** - * Use algebraic simplification a v a = a. - */ -static ir_node *equivalent_node_Or(ir_node *n) -{ - ir_node *oldn = n; - - ir_node *a = get_Or_left(n); - ir_node *b = get_Or_right(n); - - /* remove a v a */ - if (a == b) { - n = a; - - DBG_OPT_ALGSIM1(oldn, a, b, n); - } - - return n; -} - /** * optimize operations that are commutative and have neutral 0, * so a op 0 = 0 op a = a. @@ -707,9 +723,9 @@ static ir_node *equivalent_node_neutral_zero(ir_node *n) /* After running compute_node there is only one constant predecessor. Find this predecessors value and remember the other node: */ - if ((tv = computed_value(a)) != tarval_bad) { + if ((tv = value_of(a)) != tarval_bad) { on = b; - } else if ((tv = computed_value(b)) != tarval_bad) { + } else if ((tv = value_of(b)) != tarval_bad) { on = a; } else return n; @@ -739,7 +755,7 @@ static ir_node *equivalent_node_left_zero(ir_node *n) ir_node *a = get_binop_left(n); ir_node *b = get_binop_right(n); - if (classify_tarval(computed_value(b)) == TV_CLASSIFY_NULL) { + if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) { n = a; DBG_OPT_ALGSIM1(oldn, a, b, n); @@ -788,10 +804,10 @@ static ir_node *equivalent_node_Mul(ir_node *n) ir_node *b = get_Mul_right(n); /* Mul is commutative and has again an other neutral element. */ - if (classify_tarval (computed_value (a)) == TV_CLASSIFY_ONE) { + if (classify_tarval(value_of(a)) == TV_CLASSIFY_ONE) { n = b; DBG_OPT_ALGSIM1(oldn, a, b, n); - } else if (classify_tarval (computed_value (b)) == TV_CLASSIFY_ONE) { + } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { n = a; DBG_OPT_ALGSIM1(oldn, a, b, n); } @@ -807,7 +823,7 @@ static ir_node *equivalent_node_Div(ir_node *n) ir_node *b = get_Div_right(n); /* Div is not commutative. */ - if (classify_tarval(computed_value(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */ + if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */ /* Turn Div into a tuple (mem, bad, a) */ ir_node *mem = get_Div_mem(n); turn_into_tuple(n, 3); @@ -827,7 +843,7 @@ static ir_node *equivalent_node_DivMod(ir_node *n) ir_node *b = get_DivMod_right(n); /* Div is not commutative. */ - if (classify_tarval(computed_value(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */ + if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */ /* Turn DivMod into a tuple (mem, bad, a, 0) */ ir_node *mem = get_Div_mem(n); ir_mode *mode = get_irn_mode(b); @@ -841,6 +857,29 @@ static ir_node *equivalent_node_DivMod(ir_node *n) return n; } +/** + * Use algebraic simplification a | a = a | 0 = 0 | a = a. + */ +static ir_node *equivalent_node_Or(ir_node *n) +{ + ir_node *oldn = n; + + ir_node *a = get_Or_left(n); + ir_node *b = get_Or_right(n); + + if (a == b) { + n = a; /* Or has it's own neutral element */ + } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_NULL) { + n = b; + DBG_OPT_ALGSIM1(oldn, a, b, n); + } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) { + n = a; + DBG_OPT_ALGSIM1(oldn, a, b, n); + } + + return n; +} + /** * Optimize a & 0b1...1 = 0b1...1 & a = a & a = a. */ @@ -853,10 +892,10 @@ static ir_node *equivalent_node_And(ir_node *n) if (a == b) { n = a; /* And has it's own neutral element */ - } else if (classify_tarval(computed_value(a)) == TV_CLASSIFY_ALL_ONE) { + } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_ALL_ONE) { n = b; DBG_OPT_ALGSIM1(oldn, a, b, n); - } else if (classify_tarval(computed_value(b)) == TV_CLASSIFY_ALL_ONE) { + } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ALL_ONE) { n = a; DBG_OPT_ALGSIM1(oldn, a, b, n); } @@ -888,12 +927,12 @@ static ir_node *equivalent_node_Conv(ir_node *n) if (n_mode == b_mode) { if (n_mode == mode_b) { n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */ - DBG_OPT_ALGSIM1(oldn, a, b, n); + DBG_OPT_ALGSIM1(oldn, a, b, n); } else if (mode_is_int(n_mode) || mode_is_character(n_mode)) { if (smaller_mode(b_mode, a_mode)){ n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */ - DBG_OPT_ALGSIM1(oldn, a, b, n); + DBG_OPT_ALGSIM1(oldn, a, b, n); } } } @@ -901,14 +940,25 @@ static ir_node *equivalent_node_Conv(ir_node *n) return n; } +/** + * A Cast may be removed if the type of the previous node + * is already to type of the Cast. + */ +static ir_node *equivalent_node_Cast(ir_node *n) { + ir_node *pred = get_Cast_op(n); + if (get_irn_type(pred) == get_Cast_type(n)) + n = pred; + return n; +} + +/* Several optimizations: + - no Phi in start block. + - remove Id operators that are inputs to Phi + - fold Phi-nodes, iff they have only one predecessor except + themselves. +*/ static ir_node *equivalent_node_Phi(ir_node *n) { - /* Several optimizations: - - no Phi in start block. - - remove Id operators that are inputs to Phi - - fold Phi-nodes, iff they have only one predecessor except - themselves. - */ int i, n_preds; ir_node *oldn = n; @@ -923,7 +973,7 @@ static ir_node *equivalent_node_Phi(ir_node *n) block = get_nodes_block(n); /* @@@ fliegt 'raus, sollte aber doch immer wahr sein!!! assert(get_irn_arity(block) == n_preds && "phi in wrong block!"); */ - if ((is_Bad(block)) || /* Control dead */ + if ((is_Block_dead(block)) || /* Control dead */ (block == current_ir_graph->start_block)) /* There should be no Phi nodes */ return new_Bad(); /* in the Start Block. */ @@ -1014,7 +1064,7 @@ static ir_node *equivalent_node_Proj(ir_node *n) n = new_Bad(); } } else if (get_irn_mode(n) == mode_X && - is_Bad(get_nodes_block(n))) { + is_Block_dead(get_nodes_block(n))) { /* Remove dead control flow -- early gigo. */ n = new_Bad(); } @@ -1033,6 +1083,22 @@ static ir_node *equivalent_node_Id(ir_node *n) return n; } +/** + * optimize a Mux + */ +static ir_node *equivalent_node_Mux(ir_node *n) +{ + ir_node *sel = get_Mux_sel(n); + tarval *ts = value_of(sel); + + if (ts == get_tarval_b_true()) + return get_Mux_true(n); + else if (ts == get_tarval_b_false()) + return get_Mux_false(n); + + return n; +} + /** * equivalent_node() returns a node equivalent to input n. It skips all nodes that * perform no actual computation, as, e.g., the Id nodes. It does not create @@ -1077,9 +1143,11 @@ static ir_op *firm_set_default_equivalent_node(ir_op *op) CASE(DivMod); CASE(And); CASE(Conv); + CASE(Cast); CASE(Phi); CASE(Proj); CASE(Id); + CASE(Mux); default: op->equivalent_node = NULL; } @@ -1121,14 +1189,79 @@ optimize_preds(ir_node *n) { } /* end switch */ } -static ir_node *transform_node_Mul(ir_node *n) -{ +/** + * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and + * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)) if possible. + */ +static ir_node *transform_node_AddSub(ir_node *n) +{ + ir_mode *mode = get_irn_mode(n); + + if (mode_is_reference(mode)) { + ir_node *left = get_binop_left(n); + ir_node *right = get_binop_right(n); + int ref_bits = get_mode_size_bits(mode); + + if (get_irn_op(left) == op_Conv) { + ir_mode *mode = get_irn_mode(left); + int bits = get_mode_size_bits(mode); + + if (ref_bits == bits && + mode_is_int(mode) && + get_mode_arithmetic(mode) == irma_twos_complement) { + ir_node *pre = get_Conv_op(left); + ir_mode *pre_mode = get_irn_mode(pre); + + if (mode_is_int(pre_mode) && + get_mode_size_bits(pre_mode) == bits && + get_mode_arithmetic(pre_mode) == irma_twos_complement) { + /* ok, this conv just changes to sign, moreover the calculation + * is done with same number of bits as our address mode, so + * we can ignore the conv as address calculation can be viewed + * as either signed or unsigned + */ + set_binop_left(n, pre); + } + } + } + + if (get_irn_op(right) == op_Conv) { + ir_mode *mode = get_irn_mode(right); + int bits = get_mode_size_bits(mode); + + if (ref_bits == bits && + mode_is_int(mode) && + get_mode_arithmetic(mode) == irma_twos_complement) { + ir_node *pre = get_Conv_op(right); + ir_mode *pre_mode = get_irn_mode(pre); + + if (mode_is_int(pre_mode) && + get_mode_size_bits(pre_mode) == bits && + get_mode_arithmetic(pre_mode) == irma_twos_complement) { + /* ok, this conv just changes to sign, moreover the calculation + * is done with same number of bits as our address mode, so + * we can ignore the conv as address calculation can be viewed + * as either signed or unsigned + */ + set_binop_right(n, pre); + } + } + } + } + return n; +} + +#define transform_node_Add transform_node_AddSub +#define transform_node_Sub transform_node_AddSub + +/** Do architecture dependend optimizations on Mul nodes */ +static ir_node *transform_node_Mul(ir_node *n) { return arch_dep_replace_mul_with_shifts(n); } static ir_node *transform_node_Div(ir_node *n) { - tarval *tv = computed_value(n); + tarval *tv = value_of(n); ir_node *value = n; /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */ @@ -1136,7 +1269,7 @@ static ir_node *transform_node_Div(ir_node *n) if (tv != tarval_bad) value = new_Const(get_tarval_mode(tv), tv); else /* Try architecture dependand optimization */ - value = arch_dep_replace_div_with_shifts(n); + value = arch_dep_replace_div_by_const(n); if (value != n) { /* Turn Div into a tuple (mem, bad, value) */ @@ -1152,7 +1285,7 @@ static ir_node *transform_node_Div(ir_node *n) static ir_node *transform_node_Mod(ir_node *n) { - tarval *tv = computed_value(n); + tarval *tv = value_of(n); ir_node *value = n; /* BEWARE: it is NOT possible to optimize a%a to 0, as this may cause a exception */ @@ -1160,7 +1293,7 @@ static ir_node *transform_node_Mod(ir_node *n) if (tv != tarval_bad) value = new_Const(get_tarval_mode(tv), tv); else /* Try architecture dependand optimization */ - value = arch_dep_replace_mod_with_shifts(n); + value = arch_dep_replace_mod_by_const(n); if (value != n) { /* Turn Mod into a tuple (mem, bad, value) */ @@ -1205,7 +1338,7 @@ static ir_node *transform_node_DivMod(ir_node *n) evaluated = 1; } else { /* Try architecture dependand optimization */ - arch_dep_replace_divmod_with_shifts(&a, &b, n); + arch_dep_replace_divmod_by_const(&a, &b, n); evaluated = a != NULL; } } else if (ta == get_mode_null(mode)) { @@ -1264,7 +1397,7 @@ static ir_node *transform_node_Cond(ir_node *n) add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n)); } else if ((get_irn_op(a) == op_Eor) && (get_irn_mode(a) == mode_b) - && (classify_tarval(computed_value(get_Eor_right(a))) == TV_CLASSIFY_ONE)) { + && (classify_tarval(value_of(get_Eor_right(a))) == TV_CLASSIFY_ONE)) { /* The Eor is a negate. Generate a new Cond without the negate, simulate the negate by exchanging the results. */ set_irn_link(n, new_r_Cond(current_ir_graph, get_nodes_block(n), @@ -1279,6 +1412,9 @@ static ir_node *transform_node_Cond(ir_node *n) return n; } +/** + * Transform an Eor. + */ static ir_node *transform_node_Eor(ir_node *n) { ir_node *a = get_Eor_left(n); @@ -1287,13 +1423,13 @@ static ir_node *transform_node_Eor(ir_node *n) if ((get_irn_mode(n) == mode_b) && (get_irn_op(a) == op_Proj) && (get_irn_mode(a) == mode_b) - && (classify_tarval (computed_value (b)) == TV_CLASSIFY_ONE) + && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE) && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) /* The Eor negates a Cmp. The Cmp has the negated result anyways! */ n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a), mode_b, get_negated_pnc(get_Proj_proj(a))); else if ((get_irn_mode(n) == mode_b) - && (classify_tarval (computed_value (b)) == TV_CLASSIFY_ONE)) + && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)) /* The Eor is a Not. Replace it by a Not. */ /* ????!!!Extend to bitfield 1111111. */ n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode_b); @@ -1319,11 +1455,28 @@ static ir_node *transform_node_Not(ir_node *n) return n; } +/** + * Transform a Cast of a Const into a new Const + */ +static ir_node *transform_node_Cast(ir_node *n) { + ir_node *pred = get_Cast_op(n); + type *tp = get_irn_type(pred); + + if (get_irn_op(pred) == op_Const && get_Const_type(pred) != tp) { + n = new_rd_Const_type(NULL, current_ir_graph, get_nodes_block(pred), get_irn_mode(pred), + get_Const_tarval(pred), tp); + } else if ((get_irn_op(pred) == op_SymConst) && (get_SymConst_value_type(pred) != tp)) { + n = new_rd_SymConst_type(NULL, current_ir_graph, get_nodes_block(pred), get_SymConst_symbol(pred), + get_SymConst_kind(pred), tp); + } + return n; +} + /** * Transform a Div/Mod/DivMod with a non-zero constant. Must be - * done here instead of equivalent node because in creates new + * done here instead of equivalent node because it creates new * nodes. - * Removes the exceptions and routes the memory to the initial mem. + * Removes the exceptions and routes the memory to the NoMem node. * * Further, it optimizes jump tables by removing all impossible cases. */ @@ -1337,61 +1490,74 @@ static ir_node *transform_node_Proj(ir_node *proj) switch (get_irn_opcode(n)) { case iro_Div: b = get_Div_right(n); - tb = computed_value(b); + tb = value_of(b); if (tb != tarval_bad && classify_tarval(tb) != TV_CLASSIFY_NULL) { /* div(x, c) && c != 0 */ proj_nr = get_Proj_proj(proj); + /* this node may float */ + set_irn_pinned(n, op_pin_state_floats); + if (proj_nr == pn_Div_X_except) { /* we found an exception handler, remove it */ return new_Bad(); - } - else { - /* the memory Proj can be removed */ + } else { + /* the memory Proj can be removed */ ir_node *res = get_Div_mem(n); - set_Div_mem(n, get_irg_initial_mem(current_ir_graph)); - if (proj_nr == pn_Div_M) +# ifdef USE_NOMEM + set_Div_mem(n, get_irg_no_mem(current_ir_graph)); +# endif /* defined USE_NOMEM */ + if (proj_nr == pn_Div_M) return res; } } break; case iro_Mod: b = get_Mod_right(n); - tb = computed_value(b); + tb = value_of(b); if (tb != tarval_bad && classify_tarval(tb) != TV_CLASSIFY_NULL) { /* mod(x, c) && c != 0 */ proj_nr = get_Proj_proj(proj); + /* this node may float */ + set_irn_pinned(n, op_pin_state_floats); + if (proj_nr == pn_Mod_X_except) { /* we found an exception handler, remove it */ return new_Bad(); - } - else { - /* the memory Proj can be removed */ + } else { + /* the memory Proj can be removed */ ir_node *res = get_Mod_mem(n); - set_Mod_mem(n, get_irg_initial_mem(current_ir_graph)); +# ifdef USE_NOMEM + set_Mod_mem(n, get_irg_no_mem(current_ir_graph)); +# endif /* defined USE_NOMEM */ if (proj_nr == pn_Mod_M) - return res; + return res; } } break; case iro_DivMod: b = get_DivMod_right(n); - tb = computed_value(b); + tb = value_of(b); if (tb != tarval_bad && classify_tarval(tb) != TV_CLASSIFY_NULL) { /* DivMod(x, c) && c != 0 */ proj_nr = get_Proj_proj(proj); + /* this node may float */ + set_irn_pinned(n, op_pin_state_floats); + if (proj_nr == pn_DivMod_X_except) { /* we found an exception handler, remove it */ return new_Bad(); } else { - /* the memory Proj can be removed */ + /* the memory Proj can be removed */ ir_node *res = get_DivMod_mem(n); - set_DivMod_mem(n, get_irg_initial_mem(current_ir_graph)); +# ifdef USE_NOMEM + set_DivMod_mem(n, get_irg_no_mem(current_ir_graph)); +# endif /* defined USE_NOMEM */ if (proj_nr == pn_DivMod_M) - return res; + return res; } } break; @@ -1399,7 +1565,7 @@ static ir_node *transform_node_Proj(ir_node *proj) case iro_Cond: if (get_opt_unreachable_code()) { b = get_Cond_selector(n); - tb = computed_value(b); + tb = value_of(b); if (tb != tarval_bad && mode_is_int(get_tarval_mode(tb))) { /* we have a constant switch */ @@ -1532,6 +1698,65 @@ static ir_node *transform_node_Or(ir_node *or) return transform_node_Or(or); } +/* forward */ +static ir_node *transform_node(ir_node *n); + +/** + * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl + */ +static ir_node * transform_node_shift(ir_node *n) +{ + ir_node *left; + tarval *tv1, *tv2, *res; + ir_mode *mode; + int modulo_shf, flag; + + left = get_binop_left(n); + + /* different operations */ + if (get_irn_op(left) != get_irn_op(n)) + return n; + + tv1 = value_of(get_binop_right(n)); + if (tv1 == tarval_bad) + return n; + + tv2 = value_of(get_binop_right(left)); + if (tv2 == tarval_bad) + return n; + + res = tarval_add(tv1, tv2); + + /* beware: a simple replacement works only, if res < modulo shift */ + mode = get_irn_mode(n); + + flag = 0; + + modulo_shf = get_mode_modulo_shift(mode); + if (modulo_shf > 0) { + tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res)); + + if (tarval_cmp(res, modulo) & Lt) + flag = 1; + } + else + flag = 1; + + if (flag) { + /* ok, we can replace it */ + ir_node *in[2], *irn, *block = get_nodes_block(n); + + in[0] = get_binop_left(left); + in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res); + + irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in); + + return transform_node(irn); + } + return n; +} + + /** * Tries several [inplace] [optimizing] transformations and returns an * equivalent node. The difference to equivalent_node() is that these @@ -1556,6 +1781,8 @@ static ir_op *firm_set_default_transform_node(ir_op *op) break switch (op->code) { + CASE(Add); + CASE(Sub); CASE(Mul); CASE(Div); CASE(Mod); @@ -1563,8 +1790,14 @@ static ir_op *firm_set_default_transform_node(ir_op *op) CASE(Cond); CASE(Eor); CASE(Not); + CASE(Cast); CASE(Proj); CASE(Or); + case iro_Shr: + case iro_Shrs: + case iro_Shl: + op->transform_node = transform_node_shift; + break; default: op->transform_node = NULL; } @@ -1616,7 +1849,8 @@ static int node_cmp_attr_Free(ir_node *a, ir_node *b) static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) { return (get_irn_symconst_attr(a).num != get_irn_symconst_attr(b).num) - || (get_irn_symconst_attr(a).sym.type_p != get_irn_symconst_attr(b).sym.type_p); + || (get_irn_symconst_attr(a).sym.type_p != get_irn_symconst_attr(b).sym.type_p) + || (get_irn_symconst_attr(a).tp != get_irn_symconst_attr(b).tp); } /** Compares the attributes of two Call nodes. */ @@ -1625,12 +1859,6 @@ static int node_cmp_attr_Call(ir_node *a, ir_node *b) return (get_irn_call_attr(a) != get_irn_call_attr(b)); } -/** Compares the attributes of two FuncCall nodes. */ -static int node_cmp_attr_FuncCall(ir_node *a, ir_node *b) -{ - return (get_irn_funccall_attr(a) != get_irn_funccall_attr(b)); -} - /** Compares the attributes of two Sel nodes. */ static int node_cmp_attr_Sel(ir_node *a, ir_node *b) { @@ -1669,7 +1897,7 @@ static int node_cmp_attr_Store(ir_node *a, ir_node *b) { /* NEVER do CSE on volatile Stores */ return (get_Store_volatility(a) == volatility_is_volatile || - get_Load_volatility(b) == volatility_is_volatile); + get_Store_volatility(b) == volatility_is_volatile); } /** @@ -1690,7 +1918,6 @@ static ir_op *firm_set_default_node_cmp_attr(ir_op *op) CASE(Free); CASE(SymConst); CASE(Call); - CASE(FuncCall); CASE(Sel); CASE(Phi); CASE(Cast); @@ -1722,20 +1949,20 @@ vt_cmp (const void *elt, const void *key) if ((get_irn_op(a) != get_irn_op(b)) || (get_irn_mode(a) != get_irn_mode(b))) return 1; - /* compare if a's in and b's in are equal */ - irn_arity_a = get_irn_arity (a); - if (irn_arity_a != get_irn_arity(b)) + /* compare if a's in and b's in are of equal length */ + irn_arity_a = get_irn_intra_arity (a); + if (irn_arity_a != get_irn_intra_arity(b)) return 1; /* for block-local cse and op_pin_state_pinned nodes: */ - if (!get_opt_global_cse() || (get_op_pinned(get_irn_op(a)) == op_pin_state_pinned)) { - if (get_irn_n(a, -1) != get_irn_n(b, -1)) + if (!get_opt_global_cse() || (get_irn_pinned(a) == op_pin_state_pinned)) { + if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1)) return 1; } /* compare a->in[0..ins] with b->in[0..ins] */ for (i = 0; i < irn_arity_a; i++) - if (get_irn_n(a, i) != get_irn_n(b, i)) + if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i)) return 1; /* @@ -1759,41 +1986,39 @@ ir_node_hash (ir_node *node) if (node->op == op_Const) { /* special value for const, as they only differ in their tarval. */ - h = ((unsigned) node->attr.con.tv)>>3 ; - h = 9*h + (unsigned)get_irn_mode(node); + h = HASH_PTR(node->attr.con.tv); + h = 9*h + HASH_PTR(get_irn_mode(node)); } else if (node->op == op_SymConst) { /* special value for const, as they only differ in their symbol. */ - h = ((unsigned) node->attr.i.sym.type_p)>>3 ; - h = 9*h + (unsigned)get_irn_mode(node); + h = HASH_PTR(node->attr.i.sym.type_p); + h = 9*h + HASH_PTR(get_irn_mode(node)); } else { /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */ - h = irn_arity = get_irn_arity(node); + h = irn_arity = get_irn_intra_arity(node); - /* consider all in nodes... except the block. */ - for (i = 0; i < irn_arity; i++) { - h = 9*h + (unsigned)get_irn_n(node, i); + /* consider all in nodes... except the block if not a control flow. */ + for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) { + h = 9*h + HASH_PTR(get_irn_intra_n(node, i)); } /* ...mode,... */ - h = 9*h + (unsigned) get_irn_mode (node); + h = 9*h + HASH_PTR(get_irn_mode(node)); /* ...and code */ - h = 9*h + (unsigned) get_irn_op (node); + h = 9*h + HASH_PTR(get_irn_op(node)); } return h; } pset * -new_identities (void) -{ - return new_pset (vt_cmp, N_IR_NODES); +new_identities(void) { + return new_pset(vt_cmp, N_IR_NODES); } void -del_identities (pset *value_table) -{ - del_pset (value_table); +del_identities(pset *value_table) { + del_pset(value_table); } /** @@ -1903,6 +2128,7 @@ gigo (ir_node *node) if (get_irn_mode(node) == mode_X) { ir_node *block = get_nodes_block(node); if (op == op_End) return node; /* Don't optimize End, may have Bads. */ + if (get_irn_op(block) == op_Block && get_Block_matured(block)) { irn_arity = get_irn_arity(block); for (i = 0; i < irn_arity; i++) { @@ -1916,7 +2142,11 @@ gigo (ir_node *node) blocks predecessors is dead. */ if ( op != op_Block && op != op_Phi && op != op_Tuple) { irn_arity = get_irn_arity(node); - for (i = -1; i < irn_arity; i++) { + + if (is_Block_dead(get_nodes_block(node))) + return new_Bad(); + + for (i = 0; i < irn_arity; i++) { if (is_Bad(get_irn_n(node, i))) { return new_Bad(); } @@ -1951,7 +2181,14 @@ optimize_node (ir_node *n) ir_node *oldn = n; opcode iro = get_irn_opcode(n); - /* Always optimize Phi nodes: part of the construction. */ + type *old_tp = get_irn_type(n); + { + int i, arity = get_irn_arity(n); + for (i = 0; i < arity && !old_tp; ++i) + old_tp = get_irn_type(get_irn_n(n, i)); + } + + /* Allways optimize Phi nodes: part of the construction. */ if ((!get_opt_optimize()) && (iro != iro_Phi)) return n; /* constant expression evaluation / constant folding */ @@ -1959,26 +2196,27 @@ optimize_node (ir_node *n) /* constants can not be evaluated */ if (iro != iro_Const) { /* try to evaluate */ - tv = computed_value (n); + tv = computed_value(n); if ((get_irn_mode(n) != mode_T) && (tv != tarval_bad)) { /* * we MUST copy the node here temporary, because it's still needed - * for DBG_OPT_ALGSIM0 + * for DBG_OPT_CSTEVAL */ int node_size = offsetof(ir_node, attr) + n->op->attr_size; oldn = alloca(node_size); memcpy(oldn, n, node_size); - CLONE_ARR_A(ir_node *, oldn->in, n->in); + CLONE_ARR_A(ir_node *, oldn->in, n->in); - /* ARG, copy the in array, we need it for statistics */ - memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0])); + /* ARG, copy the in array, we need it for statistics */ + memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0])); /* evaluation was successful -- replace the node. */ obstack_free (current_ir_graph->obst, n); n = new_Const (get_tarval_mode (tv), tv); - - DBG_OPT_ALGSIM0(oldn, n); + if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv)) + set_Const_type(n, old_tp); + DBG_OPT_CSTEVAL(oldn, n); return n; } } @@ -2042,6 +2280,13 @@ optimize_in_place_2 (ir_node *n) ir_node *oldn = n; opcode iro = get_irn_opcode(n); + type *old_tp = get_irn_type(n); + { + int i, arity = get_irn_arity(n); + for (i = 0; i < arity && !old_tp; ++i) + old_tp = get_irn_type(get_irn_n(n, i)); + } + if (!get_opt_optimize() && (get_irn_op(n) != op_Phi)) return n; /* if not optimize return n */ @@ -2051,18 +2296,20 @@ optimize_in_place_2 (ir_node *n) return n; } - /* constant expression evaluation / constant folding */ if (get_opt_constant_folding()) { /* constants can not be evaluated */ if (iro != iro_Const) { /* try to evaluate */ - tv = computed_value (n); + tv = computed_value(n); if ((get_irn_mode(n) != mode_T) && (tv != tarval_bad)) { /* evaluation was successful -- replace the node. */ n = new_Const (get_tarval_mode (tv), tv); - DBG_OPT_ALGSIM0(oldn, n); + if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv)) + set_Const_type(n, old_tp); + + DBG_OPT_CSTEVAL(oldn, n); return n; } }