X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firopt.c;h=3672f0412b5784784e75823a3dddaf0142841583;hb=60c9f323685e93f335d0f26aad2da647b669886d;hp=28c59504f1d117b435509ad141abefdf3ca0b956;hpb=6c7d78ce0d866934e9d541dfaf5caaa646dc4d8e;p=libfirm diff --git a/ir/ir/iropt.c b/ir/ir/iropt.c index 28c59504f..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" @@ -26,10 +36,18 @@ # include "iropt_dbg.h" # 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. @@ -41,18 +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 */ @@ -72,6 +78,9 @@ static tarval *computed_value_SymConst(ir_node *n) return tarval_bad; } +/** + * return the value of an Add + */ static tarval *computed_value_Add(ir_node *n) { ir_node *a = get_Add_left(n); @@ -80,30 +89,39 @@ 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; } +/** + * return the value of a Sub + * Special case: a - a + */ static tarval *computed_value_Sub(ir_node *n) { ir_node *a = get_Sub_left(n); ir_node *b = get_Sub_right(n); + tarval *ta; + tarval *tb; - tarval *ta = value_of(a); - tarval *tb = value_of(b); + /* a - a */ + 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; } +/** + * return the value of an unary Minus + */ static tarval *computed_value_Minus(ir_node *n) { ir_node *a = get_Minus_op(n); @@ -115,6 +133,9 @@ static tarval *computed_value_Minus(ir_node *n) return tarval_bad; } +/** + * return the value of a Mul + */ static tarval *computed_value_Mul(ir_node *n) { ir_node *a = get_Mul_left(n); @@ -129,18 +150,17 @@ 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; } +/** + * return the value of a floating point Quot + */ static tarval *computed_value_Quot(ir_node *n) { ir_node *a = get_Quot_left(n); @@ -157,38 +177,64 @@ static tarval *computed_value_Quot(ir_node *n) return tarval_bad; } -static tarval *computed_value_Div(ir_node *n) +/** + * calculate the value of an integer Div of two nodes + * Special case: 0 / b + */ +static tarval *do_computed_value_Div(ir_node *a, ir_node *b) { - ir_node *a = get_Div_left(n); - ir_node *b = get_Div_right(n); - tarval *ta = value_of(a); tarval *tb = value_of(b); - /* This was missing in original implementation. Why? */ - if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) { - if (tb != get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */ + /* Compute c1 / c2 or 0 / a, a != 0 */ + 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; } return tarval_bad; } -static tarval *computed_value_Mod(ir_node *n) +/** + * return the value of an integer Div + */ +static tarval *computed_value_Div(ir_node *n) { - ir_node *a = get_Mod_left(n); - ir_node *b = get_Mod_right(n); + return do_computed_value_Div(get_Div_left(n), get_Div_right(n)); +} +/** + * calculate the value of an integer Mod of two nodes + * Special case: a % 1 + */ +static tarval *do_computed_value_Mod(ir_node *a, ir_node *b) +{ tarval *ta = value_of(a); tarval *tb = value_of(b); - /* This was missing in original implementation. Why? */ - if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) { - if (tb != get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */ + /* Compute c1 % c2 or a % 1 */ + if (tb != tarval_bad) { + if ((ta != tarval_bad) && (tb != get_mode_null(get_tarval_mode(tb)))) /* div by zero: return tarval_bad */ return tarval_mod(ta, tb); + else if (tb == get_mode_one(get_tarval_mode(tb))) /* x mod 1 == 0 */ + return get_mode_null(get_irn_mode(a)); } + return tarval_bad; } +/** + * return the value of an integer Mod + */ +static tarval *computed_value_Mod(ir_node *n) +{ + return do_computed_value_Mod(get_Mod_left(n), get_Mod_right(n)); +} + +/** + * return the value of an Abs + */ static tarval *computed_value_Abs(ir_node *n) { ir_node *a = get_Abs_op(n); @@ -200,6 +246,10 @@ static tarval *computed_value_Abs(ir_node *n) return tarval_bad; } +/** + * return the value of an And + * Special case: a & 0, 0 & b + */ static tarval *computed_value_And(ir_node *n) { ir_node *a = get_And_left(n); @@ -213,14 +263,18 @@ 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; } } return tarval_bad; } +/** + * return the value of an Or + * Special case: a | 1...1, 1...1 | b + */ static tarval *computed_value_Or(ir_node *n) { ir_node *a = get_Or_left(n); @@ -233,21 +287,29 @@ 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; } } return tarval_bad; } +/** + * return the value of an Eor + */ 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); @@ -255,6 +317,9 @@ static tarval *computed_value_Eor(ir_node *n) return tarval_bad; } +/** + * return the value of a Not + */ static tarval *computed_value_Not(ir_node *n) { ir_node *a = get_Not_op(n); @@ -266,6 +331,9 @@ static tarval *computed_value_Not(ir_node *n) return tarval_bad; } +/** + * return the value of a Shl + */ static tarval *computed_value_Shl(ir_node *n) { ir_node *a = get_Shl_left(n); @@ -280,6 +348,9 @@ static tarval *computed_value_Shl(ir_node *n) return tarval_bad; } +/** + * return the value of a Shr + */ static tarval *computed_value_Shr(ir_node *n) { ir_node *a = get_Shr_left(n); @@ -294,6 +365,9 @@ static tarval *computed_value_Shr(ir_node *n) return tarval_bad; } +/** + * return the value of a Shrs + */ static tarval *computed_value_Shrs(ir_node *n) { ir_node *a = get_Shrs_left(n); @@ -308,6 +382,9 @@ static tarval *computed_value_Shrs(ir_node *n) return tarval_bad; } +/** + * return the value of a Rot + */ static tarval *computed_value_Rot(ir_node *n) { ir_node *a = get_Rot_left(n); @@ -322,6 +399,9 @@ static tarval *computed_value_Rot(ir_node *n) return tarval_bad; } +/** + * return the value of a Conv + */ static tarval *computed_value_Conv(ir_node *n) { ir_node *a = get_Conv_op(n); @@ -333,10 +413,14 @@ static tarval *computed_value_Conv(ir_node *n) return tarval_bad; } +/** + * return the value of a Proj, handle Proj(Cmp), Proj(Div), Proj(Mod), Proj(DivMod) + */ static tarval *computed_value_Proj(ir_node *n) { - ir_node *a = get_Proj_pred(n), *b; + ir_node *a = get_Proj_pred(n); ir_node *aa, *ab; + long proj_nr; /* Optimize Cmp nodes. This performs a first step of unreachable code elimination. @@ -345,106 +429,83 @@ static tarval *computed_value_Proj(ir_node *n) only 1 is used. There are several case where we can evaluate a Cmp node: 1. The nodes compared are both the same. If we compare for - equal, greater equal, ... this will return true, else it - will return false. This step relies on cse. + equal, greater equal, ... this will return true, else it + will return false. This step relies on cse. 2. The predecessors of Cmp are target values. We can evaluate - the Cmp. + the Cmp. 3. The predecessors are Allocs or void* constants. Allocs never - return NULL, they raise an exception. Therefore we can predict - the Cmp result. */ + return NULL, they raise an exception. Therefore we can predict + the Cmp result. */ switch (get_irn_opcode(a)) { case iro_Cmp: aa = get_Cmp_left(a); 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 ((get_Proj_proj(n) == Eq), mode_b) */ - return new_tarval_from_long ((get_Proj_proj(n) & Eq), mode_b); + 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... */ - pnc_number flags = tarval_cmp (taa, tab); - if (flags != False) { - return new_tarval_from_long (get_Proj_proj(n) & flags, mode_b); - } + /* strange checks... */ + pnc_number flags = tarval_cmp (taa, tab); + if (flags != False) { + return new_tarval_from_long (proj_nr & flags, mode_b); + } } else { /* check for 3.: */ - ir_node *aaa = skip_Id(skip_Proj(aa)); - ir_node *aba = skip_Id(skip_Proj(ab)); - - if ( ( (/* aa is ProjP and aaa is Alloc */ - (get_irn_op(aa) == op_Proj) - && (mode_is_reference(get_irn_mode(aa))) - && (get_irn_op(aaa) == op_Alloc)) - && ( (/* ab is constant void */ - (get_irn_op(ab) == op_Const) - && (mode_is_reference(get_irn_mode(ab))) - && (get_Const_tarval(ab) == get_mode_null(get_irn_mode(ab)))) - || (/* ab is other Alloc */ - (get_irn_op(ab) == op_Proj) - && (mode_is_reference(get_irn_mode(ab))) - && (get_irn_op(aba) == op_Alloc) - && (aaa != aba)))) - || (/* aa is void and aba is Alloc */ - (get_irn_op(aa) == op_Const) - && (mode_is_reference(get_irn_mode(aa))) - && (get_Const_tarval(aa) == get_mode_null(get_irn_mode(aa))) - && (get_irn_op(ab) == op_Proj) - && (mode_is_reference(get_irn_mode(ab))) - && (get_irn_op(aba) == op_Alloc))) - /* 3.: */ - return new_tarval_from_long (get_Proj_proj(n) & Ne, mode_b); + ir_node *aaa = skip_Id(skip_Proj(aa)); + ir_node *aba = skip_Id(skip_Proj(ab)); + + if ( ( (/* aa is ProjP and aaa is Alloc */ + (get_irn_op(aa) == op_Proj) + && (mode_is_reference(get_irn_mode(aa))) + && (get_irn_op(aaa) == op_Alloc)) + && ( (/* ab is constant void */ + (get_irn_op(ab) == op_Const) + && (mode_is_reference(get_irn_mode(ab))) + && (get_Const_tarval(ab) == get_mode_null(get_irn_mode(ab)))) + || (/* ab is other Alloc */ + (get_irn_op(ab) == op_Proj) + && (mode_is_reference(get_irn_mode(ab))) + && (get_irn_op(aba) == op_Alloc) + && (aaa != aba)))) + || (/* aa is void and aba is Alloc */ + (get_irn_op(aa) == op_Const) + && (mode_is_reference(get_irn_mode(aa))) + && (get_Const_tarval(aa) == get_mode_null(get_irn_mode(aa))) + && (get_irn_op(ab) == op_Proj) + && (mode_is_reference(get_irn_mode(ab))) + && (get_irn_op(aba) == op_Alloc))) + /* 3.: */ + return new_tarval_from_long (proj_nr & Ne, mode_b); } } break; case iro_DivMod: - { - tarval *tb = value_of(b = get_DivMod_right(a)); - tarval *ta = value_of(a = get_DivMod_left(a)); - - if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) { - if (tb == get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */ - return tarval_bad; - if (get_Proj_proj(n)== pn_DivMod_res_div) - return tarval_div(ta, tb); - else if (get_Proj_proj(n)== pn_DivMod_res_mod) - return tarval_mod(ta, tb); - } + /* compute either the Div or the Mod part */ + proj_nr = get_Proj_proj(n); + if (proj_nr == pn_DivMod_res_div) + return do_computed_value_Div(get_DivMod_left(a), get_DivMod_right(a)); + else if (proj_nr == pn_DivMod_res_mod) + return do_computed_value_Mod(get_DivMod_left(a), get_DivMod_right(a)); break; - } case iro_Div: - { - tarval *tb = value_of(b = get_Div_right(a)); - tarval *ta = value_of(a = get_Div_left(a)); - - if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) { - if (tb == get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */ - return tarval_bad; - if (get_Proj_proj(n)== pn_Div_res) - return tarval_div(ta, tb); - } + if (get_Proj_proj(n) == pn_Div_res) + return computed_value(a); break; - } case iro_Mod: - { - tarval *tb = value_of(b = get_Mod_right(a)); - tarval *ta = value_of(a = get_Mod_left(a)); - - if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) { - if (tb == get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */ - return tarval_bad; - if (get_Proj_proj(n)== pn_Mod_res) - return tarval_mod(ta, tb); - } + if (get_Proj_proj(n) == pn_Mod_res) + return computed_value(a); break; - } default: return tarval_bad; @@ -452,6 +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. @@ -470,9 +551,9 @@ tarval *computed_value(ir_node *n) */ static ir_op *firm_set_default_computed_value(ir_op *op) { -#define CASE(a) \ - case iro_##a: \ - op->computed_value = computed_value_##a; \ +#define CASE(a) \ + case iro_##a: \ + op->computed_value = computed_value_##a; \ break switch (op->code) { @@ -496,6 +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; } @@ -516,7 +598,7 @@ different_identity (ir_node *a, ir_node *b) ir_node *a1 = get_Proj_pred (a); ir_node *b1 = get_Proj_pred (b); if (a1 != b1 && get_irn_op (a1) == op_Alloc - && get_irn_op (b1) == op_Alloc) + && get_irn_op (b1) == op_Alloc) return 1; } return 0; @@ -542,47 +624,63 @@ 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(); DBG_OPT_DEAD; + n = set_Block_dead(n); + DBG_OPT_DEAD(oldn, n); } else if (get_opt_control_flow_straightening()) { - n = predblock; DBG_OPT_STG; + n = predblock; + DBG_OPT_STG(oldn, n); } } else if ((get_Block_n_cfgpreds(n) == 1) && - (get_irn_op(skip_Proj(get_Block_cfgpred(n, 0))) == op_Cond)) { + (get_irn_op(skip_Proj(get_Block_cfgpred(n, 0))) == op_Cond)) { 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(); DBG_OPT_DEAD; + n = set_Block_dead(n); + DBG_OPT_DEAD(oldn, n); } } else if ((get_Block_n_cfgpreds(n) == 2) && - (get_opt_control_flow_weak_simplification())) { + (get_opt_control_flow_weak_simplification())) { /* Test whether Cond jumps twice to this block @@@ we could do this also with two loops finding two preds from several ones. */ ir_node *a = get_Block_cfgpred(n, 0); ir_node *b = get_Block_cfgpred(n, 1); if ((get_irn_op(a) == op_Proj) && - (get_irn_op(b) == op_Proj) && - (get_Proj_pred(a) == get_Proj_pred(b)) && - (get_irn_op(get_Proj_pred(a)) == op_Cond) && - (get_irn_mode(get_Cond_selector(get_Proj_pred(a))) == mode_b)) { + (get_irn_op(b) == op_Proj) && + (get_Proj_pred(a) == get_Proj_pred(b)) && + (get_irn_op(get_Proj_pred(a)) == op_Cond) && + (get_irn_mode(get_Cond_selector(get_Proj_pred(a))) == mode_b)) { /* Also a single entry Block following a single exit Block. Phis have - twice the same operand and will be optimized away. */ - n = get_nodes_block(a); DBG_OPT_IFSIM; + twice the same operand and will be optimized away. */ + n = get_nodes_block(a); + DBG_OPT_IFSIM(oldn, a, b, n); } } else if (get_opt_unreachable_code() && - (n != current_ir_graph->start_block) && - (n != current_ir_graph->end_block) ) { - int i; + (n != current_ir_graph->start_block) && + (n != current_ir_graph->end_block) ) { + 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; @@ -596,7 +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; @@ -609,24 +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; - } - - return n; -} - /** * optimize operations that are commutative and have neutral 0, * so a op 0 = 0 op a = a. @@ -643,9 +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; @@ -653,7 +733,9 @@ static ir_node *equivalent_node_neutral_zero(ir_node *n) /* If this predecessors constant value is zero, the operation is unnecessary. Remove it: */ if (classify_tarval (tv) == TV_CLASSIFY_NULL) { - n = on; DBG_OPT_ALGSIM1; + n = on; + + DBG_OPT_ALGSIM1(oldn, a, b, n); } return n; @@ -673,8 +755,10 @@ 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) { - n = a; DBG_OPT_ALGSIM1; + if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) { + n = a; + + DBG_OPT_ALGSIM1(oldn, a, b, n); } return n; @@ -692,10 +776,12 @@ static ir_node *equivalent_node_left_zero(ir_node *n) static ir_node *equivalent_node_symmetric_unop(ir_node *n) { ir_node *oldn = n; + ir_node *pred = get_unop_op(n); /* optimize symmetric unop */ - if (get_irn_op(get_unop_op(n)) == get_irn_op(n)) { - n = get_unop_op(get_unop_op(n)); DBG_OPT_ALGSIM2; + if (get_irn_op(pred) == get_irn_op(n)) { + n = get_unop_op(pred); + DBG_OPT_ALGSIM2(oldn, pred, n); } return n; } @@ -704,7 +790,7 @@ static ir_node *equivalent_node_symmetric_unop(ir_node *n) #define equivalent_node_Not equivalent_node_symmetric_unop /* --x == x */ /* ??? Is this possible or can --x raise an - out of bounds exception if min =! max? */ + out of bounds exception if min =! max? */ #define equivalent_node_Minus equivalent_node_symmetric_unop /** @@ -718,10 +804,12 @@ 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) { - n = b; DBG_OPT_ALGSIM1; - } else if (classify_tarval (computed_value (b)) == TV_CLASSIFY_ONE) { - n = a; DBG_OPT_ALGSIM1; + if (classify_tarval(value_of(a)) == TV_CLASSIFY_ONE) { + n = b; + DBG_OPT_ALGSIM1(oldn, a, b, n); + } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { + n = a; + DBG_OPT_ALGSIM1(oldn, a, b, n); } return n; } @@ -735,17 +823,63 @@ 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); set_Tuple_pred(n, pn_Div_M, mem); - set_Tuple_pred(n, pn_Div_X_except, new_Bad()); /* no exception */ + set_Tuple_pred(n, pn_Div_X_except, new_Bad()); /* no exception */ set_Tuple_pred(n, pn_Div_res, a); } return n; } +/** + * Optimize a / 1 = a. + */ +static ir_node *equivalent_node_DivMod(ir_node *n) +{ + ir_node *a = get_DivMod_left(n); + ir_node *b = get_DivMod_right(n); + + /* Div is not commutative. */ + 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); + + turn_into_tuple(n, 4); + set_Tuple_pred(n, pn_DivMod_M, mem); + set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */ + set_Tuple_pred(n, pn_DivMod_res_div, a); + set_Tuple_pred(n, pn_DivMod_res_mod, new_Const(mode, get_mode_null(mode))); + } + 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. */ @@ -758,12 +892,13 @@ 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; - } else if (classify_tarval(computed_value(b)) == TV_CLASSIFY_ALL_ONE) { + DBG_OPT_ALGSIM1(oldn, a, b, n); + } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ALL_ONE) { n = a; + DBG_OPT_ALGSIM1(oldn, a, b, n); } - if (n != oldn) DBG_OPT_ALGSIM1; return n; } @@ -780,7 +915,8 @@ static ir_node *equivalent_node_Conv(ir_node *n) ir_mode *a_mode = get_irn_mode(a); if (n_mode == a_mode) { /* No Conv necessary */ - n = a; DBG_OPT_ALGSIM3; + n = a; + DBG_OPT_ALGSIM3(oldn, a, n); } else if (get_irn_op(a) == op_Conv) { /* Conv(Conv(b)) */ ir_mode *b_mode; @@ -790,26 +926,39 @@ 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; + n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */ + 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; - } + if (smaller_mode(b_mode, a_mode)){ + n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */ + DBG_OPT_ALGSIM1(oldn, a, b, 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; @@ -824,9 +973,9 @@ 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. */ + return new_Bad(); /* in the Start Block. */ if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */ @@ -839,10 +988,10 @@ static ir_node *equivalent_node_Phi(ir_node *n) ir_node *a = get_Phi_pred(n, 0); ir_node *b = get_Phi_pred(n, 1); if ( (get_irn_op(a) == op_Confirm) - && (get_irn_op(b) == op_Confirm) - && follow_Id (get_irn_n(a, 0) == get_irn_n(b, 0)) - && (get_irn_n(a, 1) == get_irn_n (b, 1)) - && (a->data.num == (~b->data.num & irpn_True) )) { + && (get_irn_op(b) == op_Confirm) + && follow_Id (get_irn_n(a, 0) == get_irn_n(b, 0)) + && (get_irn_n(a, 1) == get_irn_n (b, 1)) + && (a->data.num == (~b->data.num & irpn_True) )) { return get_irn_n(a, 0); } } @@ -858,9 +1007,9 @@ static ir_node *equivalent_node_Phi(ir_node *n) first_val = get_Phi_pred(n, i); if ( (first_val != n) /* not self pointer */ #if 1 - && (get_irn_op(first_val) != op_Bad) + && (get_irn_op(first_val) != op_Bad) #endif - ) { /* value not dead */ + ) { /* value not dead */ break; /* then found first value. */ } } @@ -875,18 +1024,19 @@ static ir_node *equivalent_node_Phi(ir_node *n) while (++i < n_preds) { scnd_val = get_Phi_pred(n, i); if ( (scnd_val != n) - && (scnd_val != first_val) + && (scnd_val != first_val) #if 1 - && (get_irn_op(scnd_val) != op_Bad) + && (get_irn_op(scnd_val) != op_Bad) #endif - ) { + ) { break; } } /* Fold, if no multiple distinct non-self-referencing inputs */ if (i >= n_preds) { - n = first_val; DBG_OPT_PHI; + n = first_val; + DBG_OPT_PHI(oldn, first_val, n); } else { /* skip the remaining Ids (done in get_Phi_pred). */ /* superfluous, since we walk all to propagate Block's Bads. @@ -907,13 +1057,14 @@ static ir_node *equivalent_node_Proj(ir_node *n) if ( get_irn_op(a) == op_Tuple) { /* Remove the Tuple/Proj combination. */ if ( get_Proj_proj(n) <= get_Tuple_n_preds(a) ) { - n = get_Tuple_pred(a, get_Proj_proj(n)); DBG_OPT_TUPLE; + n = get_Tuple_pred(a, get_Proj_proj(n)); + DBG_OPT_TUPLE(oldn, a, n); } else { assert(0); /* This should not happen! */ 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(); } @@ -927,7 +1078,24 @@ static ir_node *equivalent_node_Id(ir_node *n) { ir_node *oldn = n; - n = follow_Id(n); DBG_OPT_ID; + n = follow_Id(n); + DBG_OPT_ID(oldn, 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; } @@ -951,9 +1119,9 @@ equivalent_node(ir_node *n) */ static ir_op *firm_set_default_equivalent_node(ir_op *op) { -#define CASE(a) \ - case iro_##a: \ - op->equivalent_node = equivalent_node_##a; \ +#define CASE(a) \ + case iro_##a: \ + op->equivalent_node = equivalent_node_##a; \ break switch (op->code) { @@ -972,11 +1140,14 @@ static ir_op *firm_set_default_equivalent_node(ir_op *op) CASE(Minus); CASE(Mul); CASE(Div); + CASE(DivMod); CASE(And); CASE(Conv); + CASE(Cast); CASE(Phi); CASE(Proj); CASE(Id); + CASE(Mux); default: op->equivalent_node = NULL; } @@ -1018,37 +1189,120 @@ optimize_preds(ir_node *n) { } /* end switch */ } +/** + * 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 */ - if (tv != tarval_bad) { + if (tv != tarval_bad) + value = new_Const(get_tarval_mode(tv), tv); + else /* Try architecture dependand optimization */ + value = arch_dep_replace_div_by_const(n); + + if (value != n) { /* Turn Div into a tuple (mem, bad, value) */ ir_node *mem = get_Div_mem(n); turn_into_tuple(n, 3); set_Tuple_pred(n, pn_Div_M, mem); set_Tuple_pred(n, pn_Div_X_except, new_Bad()); - set_Tuple_pred(n, pn_Div_res, new_Const(get_tarval_mode(tv), tv)); + set_Tuple_pred(n, pn_Div_res, value); } return 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 */ - if (tv != tarval_bad) { + if (tv != tarval_bad) + value = new_Const(get_tarval_mode(tv), tv); + else /* Try architecture dependand optimization */ + value = arch_dep_replace_mod_by_const(n); + + if (value != n) { /* Turn Mod into a tuple (mem, bad, value) */ ir_node *mem = get_Mod_mem(n); + turn_into_tuple(n, 3); set_Tuple_pred(n, pn_Mod_M, mem); set_Tuple_pred(n, pn_Mod_X_except, new_Bad()); - set_Tuple_pred(n, pn_Mod_res, new_Const(get_tarval_mode(tv), tv)); + set_Tuple_pred(n, pn_Mod_res, value); } return n; } @@ -1076,17 +1330,23 @@ static ir_node *transform_node_DivMod(ir_node *n) tarval *resa, *resb; resa = tarval_div (ta, tb); if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through - Jmp for X result!? */ + Jmp for X result!? */ resb = tarval_mod (ta, tb); if (resb == tarval_bad) return n; /* Causes exception! */ a = new_Const (mode, resa); b = new_Const (mode, resb); evaluated = 1; } + else { /* Try architecture dependand optimization */ + arch_dep_replace_divmod_by_const(&a, &b, n); + evaluated = a != NULL; + } } else if (ta == get_mode_null(mode)) { + /* 0 / non-Const = 0 */ b = a; evaluated = 1; } + if (evaluated) { /* replace by tuple */ ir_node *mem = get_DivMod_mem(n); turn_into_tuple(n, 4); @@ -1112,7 +1372,7 @@ static ir_node *transform_node_Cond(ir_node *n) (get_irn_mode(a) == mode_b) && (get_opt_unreachable_code())) { /* It's a boolean Cond, branching on a boolean constant. - Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */ + Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */ jmp = new_r_Jmp(current_ir_graph, get_nodes_block(n)); turn_into_tuple(n, 2); if (ta == tarval_b_true) { @@ -1125,9 +1385,9 @@ 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(current_ir_graph), get_nodes_block(n)); } else if ((ta != tarval_bad) && - (get_irn_mode(a) == mode_Iu) && - (get_Cond_kind(n) == dense) && - (get_opt_unreachable_code())) { + (get_irn_mode(a) == mode_Iu) && + (get_Cond_kind(n) == dense) && + (get_opt_unreachable_code())) { /* I don't want to allow Tuples smaller than the biggest Proj. Also this tuple might get really big... I generate the Jmp here, and remember it in link. Link is used @@ -1136,22 +1396,25 @@ 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(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)) { + && (get_irn_mode(a) == mode_b) + && (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), - get_Eor_left(a))); + get_Eor_left(a))); } else if ((get_irn_op(a) == op_Not) - && (get_irn_mode(a) == mode_b)) { + && (get_irn_mode(a) == mode_b)) { /* A Not before the Cond. Generate a new Cond without the Not, simulate the Not by exchanging the results. */ set_irn_link(n, new_r_Cond(current_ir_graph, get_nodes_block(n), - get_Not_op(a))); + get_Not_op(a))); } return n; } +/** + * Transform an Eor. + */ static ir_node *transform_node_Eor(ir_node *n) { ir_node *a = get_Eor_left(n); @@ -1160,13 +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))); + 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); @@ -1175,7 +1438,7 @@ static ir_node *transform_node_Eor(ir_node *n) } /** - * Transfor a boolean Not. + * Transform a boolean Not. */ static ir_node *transform_node_Not(ir_node *n) { @@ -1187,125 +1450,142 @@ static ir_node *transform_node_Not(ir_node *n) && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) /* We negate 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))); + mode_b, get_negated_pnc(get_Proj_proj(a))); 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 NoMem node. + * + * Further, it optimizes jump tables by removing all impossible cases. */ static ir_node *transform_node_Proj(ir_node *proj) { ir_node *n = get_Proj_pred(proj); ir_node *b; tarval *tb; + long proj_nr; switch (get_irn_opcode(n)) { case iro_Div: - b = get_Div_right(n); - tb = computed_value(b); + b = get_Div_right(n); + tb = value_of(b); if (tb != tarval_bad && classify_tarval(tb) != TV_CLASSIFY_NULL) { /* div(x, c) && c != 0 */ - ir_node *div, *proj; - ir_node *a = get_Div_left(n); - ir_node *mem = get_Div_mem(n); - int rem = get_optimize(); + proj_nr = get_Proj_proj(proj); - set_optimize(0); - { - div = new_rd_Div(get_irn_dbg_info(n), current_ir_graph, - get_nodes_block(n), get_irg_initial_mem(current_ir_graph), a, b); + /* this node may float */ + set_irn_pinned(n, op_pin_state_floats); - proj = new_r_Proj(current_ir_graph, get_nodes_block(n), div, get_irn_mode(a), pn_Div_res); + if (proj_nr == pn_Div_X_except) { + /* we found an exception handler, remove it */ + return new_Bad(); + } else { + /* the memory Proj can be removed */ + ir_node *res = get_Div_mem(n); +# 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; } - set_optimize(rem); - - turn_into_tuple(n, 3); - set_Tuple_pred(n, pn_Mod_M, mem); - set_Tuple_pred(n, pn_Mod_X_except, new_Bad()); - set_Tuple_pred(n, pn_Mod_res, proj); } break; case iro_Mod: - b = get_Mod_right(n); - tb = computed_value(b); + b = get_Mod_right(n); + tb = value_of(b); if (tb != tarval_bad && classify_tarval(tb) != TV_CLASSIFY_NULL) { /* mod(x, c) && c != 0 */ - ir_node *mod, *proj; - ir_node *a = get_Mod_left(n); - ir_node *mem = get_Mod_mem(n); - int rem = get_optimize(); + proj_nr = get_Proj_proj(proj); - set_optimize(0); - { - mod = new_rd_Mod(get_irn_dbg_info(n), current_ir_graph, - get_nodes_block(n), get_irg_initial_mem(current_ir_graph), a, b); + /* this node may float */ + set_irn_pinned(n, op_pin_state_floats); - proj = new_r_Proj(current_ir_graph, get_nodes_block(n), mod, get_irn_mode(a), pn_Mod_res); + if (proj_nr == pn_Mod_X_except) { + /* we found an exception handler, remove it */ + return new_Bad(); + } else { + /* the memory Proj can be removed */ + ir_node *res = get_Mod_mem(n); +# 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; } - set_optimize(rem); - - turn_into_tuple(n, 3); - set_Tuple_pred(n, pn_Mod_M, mem); - set_Tuple_pred(n, pn_Mod_X_except, new_Bad()); - set_Tuple_pred(n, pn_Mod_res, proj); } break; case iro_DivMod: - b = get_DivMod_right(n); - tb = computed_value(b); + b = get_DivMod_right(n); + tb = value_of(b); if (tb != tarval_bad && classify_tarval(tb) != TV_CLASSIFY_NULL) { /* DivMod(x, c) && c != 0 */ - ir_node *div_mod, *proj_div, *proj_mod; - ir_node *a = get_Mod_left(n); - ir_node *mem = get_Mod_mem(n); - int rem = get_optimize(); - - set_optimize(0); - { - div_mod = new_rd_DivMod(get_irn_dbg_info(n), current_ir_graph, - get_nodes_block(n), get_irg_initial_mem(current_ir_graph), a, b); - - proj_div = new_r_Proj(current_ir_graph, get_nodes_block(n), div_mod, get_irn_mode(a), pn_DivMod_res_div); - proj_mod = new_r_Proj(current_ir_graph, get_nodes_block(n), div_mod, get_irn_mode(a), pn_DivMod_res_mod); - } - set_optimize(rem); + proj_nr = get_Proj_proj(proj); - turn_into_tuple(n, 4); - set_Tuple_pred(n, pn_DivMod_M, mem); - set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); - set_Tuple_pred(n, pn_DivMod_res_div, proj_div); - set_Tuple_pred(n, pn_DivMod_res_mod, proj_mod); + /* 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 */ + ir_node *res = get_DivMod_mem(n); +# 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; + } } break; 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 */ - long num = get_Proj_proj(proj); - - if (num != get_Cond_defaultProj(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 - return new_Bad(); - } + /* we have a constant switch */ + long num = get_Proj_proj(proj); + + if (num != get_Cond_defaultProj(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 + return new_Bad(); + } } } return proj; case iro_Tuple: - /* should not happen, but if it doest will optimize */ + /* should not happen, but if it does will be optimized away */ break; default: @@ -1418,6 +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 @@ -1436,20 +1775,29 @@ static ir_node *transform_node(ir_node *n) */ static ir_op *firm_set_default_transform_node(ir_op *op) { -#define CASE(a) \ - case iro_##a: \ - op->transform_node = transform_node_##a; \ +#define CASE(a) \ + case iro_##a: \ + op->transform_node = transform_node_##a; \ break switch (op->code) { + CASE(Add); + CASE(Sub); + CASE(Mul); CASE(Div); CASE(Mod); CASE(DivMod); 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; } @@ -1465,76 +1813,101 @@ static ir_op *firm_set_default_transform_node(ir_op *op) in a graph. */ #define N_IR_NODES 512 +/** Compares the attributes of two Const nodes. */ static int node_cmp_attr_Const(ir_node *a, ir_node *b) { return (get_Const_tarval(a) != get_Const_tarval(b)) || (get_Const_type(a) != get_Const_type(b)); } +/** Compares the attributes of two Proj nodes. */ static int node_cmp_attr_Proj(ir_node *a, ir_node *b) { return get_irn_proj_attr (a) != get_irn_proj_attr (b); } +/** Compares the attributes of two Filter nodes. */ static int node_cmp_attr_Filter(ir_node *a, ir_node *b) { return get_Filter_proj(a) != get_Filter_proj(b); } +/** Compares the attributes of two Alloc nodes. */ static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) { return (get_irn_alloc_attr(a).where != get_irn_alloc_attr(b).where) - || (get_irn_alloc_attr(a).type != get_irn_alloc_attr(b).type); + || (get_irn_alloc_attr(a).type != get_irn_alloc_attr(b).type); } +/** Compares the attributes of two Free nodes. */ static int node_cmp_attr_Free(ir_node *a, ir_node *b) { return (get_irn_free_attr(a) != get_irn_free_attr(b)); } +/** Compares the attributes of two SymConst nodes. */ 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. */ static int node_cmp_attr_Call(ir_node *a, ir_node *b) { return (get_irn_call_attr(a) != get_irn_call_attr(b)); } -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) { - return (get_irn_sel_attr(a).ent->kind != get_irn_sel_attr(b).ent->kind) - || (get_irn_sel_attr(a).ent->name != get_irn_sel_attr(b).ent->name) - || (get_irn_sel_attr(a).ent->owner != get_irn_sel_attr(b).ent->owner) + return (get_irn_sel_attr(a).ent->kind != get_irn_sel_attr(b).ent->kind) + || (get_irn_sel_attr(a).ent->name != get_irn_sel_attr(b).ent->name) + || (get_irn_sel_attr(a).ent->owner != get_irn_sel_attr(b).ent->owner) || (get_irn_sel_attr(a).ent->ld_name != get_irn_sel_attr(b).ent->ld_name) - || (get_irn_sel_attr(a).ent->type != get_irn_sel_attr(b).ent->type); + || (get_irn_sel_attr(a).ent->type != get_irn_sel_attr(b).ent->type); } +/** Compares the attributes of two Phi nodes. */ static int node_cmp_attr_Phi(ir_node *a, ir_node *b) { return get_irn_phi_attr (a) != get_irn_phi_attr (b); } +/** Compares the attributes of two Cast nodes. */ static int node_cmp_attr_Cast(ir_node *a, 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(ir_node *a, 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; + + return get_Load_mode(a) != get_Load_mode(b); +} + +/** Compares the attributes of two Store nodes. */ +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_Store_volatility(b) == volatility_is_volatile); +} + /** * set the default node attribute compare operation */ static ir_op *firm_set_default_node_cmp_attr(ir_op *op) { -#define CASE(a) \ - case iro_##a: \ - op->node_cmp_attr = node_cmp_attr_##a; \ +#define CASE(a) \ + case iro_##a: \ + op->node_cmp_attr = node_cmp_attr_##a; \ break switch (op->code) { @@ -1545,10 +1918,11 @@ 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); + CASE(Load); + CASE(Store); default: op->node_cmp_attr = NULL; } @@ -1575,20 +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; /* @@ -1612,41 +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); } /** @@ -1664,14 +2036,15 @@ identify (pset *value_table, ir_node *n) if (!value_table) return n; - /* TODO: use a generic commutative attribute */ if (get_opt_reassociation()) { if (is_op_commutative(get_irn_op(n))) { + ir_node *l = get_binop_left(n); + ir_node *r = get_binop_right(n); + /* for commutative operators perform a OP b == b OP a */ - if (get_binop_left(n) > get_binop_right(n)) { - ir_node *h = get_binop_left(n); - set_binop_left(n, get_binop_right(n)); - set_binop_right(n, h); + if (l > r) { + set_binop_left(n, r); + set_binop_right(n, l); } } } @@ -1679,6 +2052,8 @@ identify (pset *value_table, ir_node *n) o = pset_find (value_table, n, ir_node_hash (n)); if (!o) return n; + DBG_OPT_CSE(n, o); + return o; } @@ -1690,6 +2065,7 @@ identify (pset *value_table, ir_node *n) static INLINE ir_node * identify_cons (pset *value_table, ir_node *n) { ir_node *old = n; + n = identify(value_table, n); if (get_irn_n(old, -1) != get_irn_n(n, -1)) set_irg_pinned(current_ir_graph, op_pin_state_floats); @@ -1702,23 +2078,39 @@ identify_cons (pset *value_table, ir_node *n) { * if it isn't there yet. */ static ir_node * -identify_remember (pset *value_table, ir_node *node) +identify_remember (pset *value_table, ir_node *n) { ir_node *o = NULL; - if (!value_table) return node; + if (!value_table) return n; + + if (get_opt_reassociation()) { + if (is_op_commutative(get_irn_op(n))) { + ir_node *l = get_binop_left(n); + ir_node *r = get_binop_right(n); + + /* for commutative operators perform a OP b == b OP a */ + if (l > r) { + set_binop_left(n, r); + set_binop_right(n, l); + } + } + } /* lookup or insert in hash table with given hash key. */ - o = pset_insert (value_table, node, ir_node_hash (node)); + o = pset_insert (value_table, n, ir_node_hash (n)); - if (o == node) return node; + if (o != n) { + DBG_OPT_CSE(n, o); + } return o; } void add_identities (pset *value_table, ir_node *node) { - identify_remember (value_table, node); + if (get_opt_cse() && (get_irn_opcode(node) != iro_Block)) + identify_remember (value_table, node); } /** @@ -1736,10 +2128,11 @@ 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++) { - if (!is_Bad(get_irn_n(block, i))) break; + if (!is_Bad(get_irn_n(block, i))) break; } if (i == irn_arity) return new_Bad(); } @@ -1749,7 +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(); } @@ -1784,6 +2181,13 @@ optimize_node (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)); + } + /* Allways optimize Phi nodes: part of the construction. */ if ((!get_opt_optimize()) && (iro != iro_Phi)) return n; @@ -1792,23 +2196,28 @@ 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 - */ - int node_size = offsetof(ir_node, attr) + n->op->attr_size; - ir_node *x = alloca(node_size); + /* + * we MUST copy the node here temporary, because it's still needed + * 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); - memcpy(x, n, node_size); - oldn = x; + /* 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; - return n; + obstack_free (current_ir_graph->obst, n); + n = new_Const (get_tarval_mode (tv), tv); + 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; } } } @@ -1871,6 +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 */ @@ -1880,18 +2296,21 @@ 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; - return n; + n = new_Const (get_tarval_mode (tv), tv); + + 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; } } }