X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firopt.c;h=8150b4496ab5cf381eafe56051faa79809eea81e;hb=9be70b7ef6120836707d76050372a89247cec2a6;hp=a8e2d1f5a7b67154886c83da0e48a74decc9a8d5;hpb=0e9428a1fd32b6084e0c561fa8056912069613d9;p=libfirm diff --git a/ir/ir/iropt.c b/ir/ir/iropt.c index a8e2d1f5a..8150b4496 100644 --- a/ir/ir/iropt.c +++ b/ir/ir/iropt.c @@ -6,6 +6,8 @@ ** iropt --- optimizations intertwined with IR construction. */ +/* $Id$ */ + #ifdef HAVE_CONFIG_H # include #endif @@ -18,21 +20,22 @@ # include "irvrfy.h" # include "tv.h" # include "tune.h" -# include "debinfo.h" +# include "dbginfo_t.h" +# include "iropt_dbg.c" /* Make types visible to allow most efficient access */ # include "entity_t.h" -/* Trivial inlineable routine for copy propagation. - Does follow Ids, needed to optimize inlined code. */ -static inline ir_node * +/* Trivial INLINEable routine for copy propagation. + Does follow Ids, needed to optimize INLINEd code. */ +static INLINE ir_node * follow_Id (ir_node *n) { while (get_irn_op (n) == op_Id) n = get_Id_pred (n); return n; } -static inline tarval * +static INLINE tarval * value_of (ir_node *n) { if ((n != NULL) && (get_irn_op(n) == op_Const)) @@ -70,6 +73,12 @@ computed_value (ir_node *n) switch (get_irn_opcode(n)) { case iro_Const: res = get_Const_tarval(n); + break; + case iro_SymConst: + if ((get_SymConst_kind(n) == size) && + (get_type_state(get_SymConst_type(n))) == layout_fixed) + res = tarval_from_long (mode_i, get_type_size(get_SymConst_type(n))); + break; case iro_Add: if (ta && tb && (get_irn_mode(a) == get_irn_mode(b)) && (get_irn_mode(a) != mode_p)) { @@ -171,8 +180,8 @@ computed_value (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, 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. 3. The predecessors are Allocs or void* constants. Allocs never @@ -232,8 +241,6 @@ computed_value (ir_node *n) else /* Mod */ res = tarval_mod(ta, tb); } - } else { - /* printf(" # comp_val: Proj node, not optimized\n"); */ } } break; @@ -256,7 +263,7 @@ 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; @@ -268,13 +275,14 @@ different_identity (ir_node *a, ir_node *b) new nodes. It is therefore safe to free N if the node returned is not N. If a node returns a Tuple we can not just skip it. If the size of the in array fits, we transform n into a tuple (e.g., Div). */ -static ir_node * +ir_node * equivalent_node (ir_node *n) { int ins; ir_node *a = NULL; /* to shutup gcc */ ir_node *b = NULL; /* to shutup gcc */ ir_node *c = NULL; /* to shutup gcc */ + ir_node *oldn = n; ins = get_irn_arity (n); @@ -294,17 +302,34 @@ equivalent_node (ir_node *n) calls the optimization. */ assert(get_Block_matured(n)); - /* A single entry Block following a single exit Block can be merged, - if it is not the Start block. */ + /* Straightening: a single entry Block following a single exit Block + can be merged, if it is not the Start block. */ /* !!! Beware, all Phi-nodes of n must have been optimized away. This should be true, as the block is matured before optimize is called. But what about Phi-cycles with the Phi0/Id that could not be resolved? Remaining Phi nodes are just Ids. */ - if (get_Block_n_cfgpreds(n) == 1 - && get_irn_op(get_Block_cfgpred(n, 0)) == op_Jmp) { - n = get_nodes_Block(get_Block_cfgpred(n, 0)); - - } else if ((n != current_ir_graph->start_block) && + if ((get_Block_n_cfgpreds(n) == 1) && + (get_irn_op(get_Block_cfgpred(n, 0)) == op_Jmp) && + (get_opt_control_flow())) { + n = get_nodes_Block(get_Block_cfgpred(n, 0)); DBG_OPT_STG; + + } else if ((get_Block_n_cfgpreds(n) == 2) && + (get_opt_control_flow())) { + /* Test whether Cond jumps twice to this block + @@@ we could do this also with two loops finding two preds from several ones. */ + a = get_Block_cfgpred(n, 0); + 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)) { + /* 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; + } + } else if (get_opt_unreachable_code() && + (n != current_ir_graph->start_block) && (n != current_ir_graph->end_block) ) { int i; /* If all inputs are dead, this block is dead too, except if it is @@ -323,29 +348,28 @@ equivalent_node (ir_node *n) /* unreachable code elimination */ if (is_Bad(get_nodes_Block(n))) n = new_Bad(); break; - /* We do not evaluate Cond here as we replace it by a new node, a Jmp. - See cases for iro_Cond and iro_Proj in transform_node. */ - /** remove stuff as x+0, x*1 x&true ... constant expression evaluation **/ + /* We do not evaluate Cond here as we replace it by a new node, a Jmp. + See cases for iro_Cond and iro_Proj in transform_node. */ + /** remove stuff as x+0, x*1 x&true ... constant expression evaluation **/ case iro_Or: if (a == b) {n = a; break;} case iro_Add: - case iro_Eor: - { tarval *tv; - ir_node *on; - /* After running compute_node there is only one constant predecessor. - Find this predecessors value and remember the other node: */ - if ((tv = computed_value (a))) { - on = b; - } else if ((tv = computed_value (b))) { - on = a; - } else break; - - /* If this predecessors constant value is zero, the operation is - unnecessary. Remove it: */ - if (tarval_classify (tv) == 0) { - n = on; - } + case iro_Eor: { + tarval *tv; + ir_node *on; + /* After running compute_node there is only one constant predecessor. + Find this predecessors value and remember the other node: */ + if ((tv = computed_value (a))) { + on = b; + } else if ((tv = computed_value (b))) { + on = a; + } else break; + + /* If this predecessors constant value is zero, the operation is + unnecessary. Remove it: */ + if (tarval_classify (tv) == 0) { + n = on; DBG_OPT_ALGSIM1; } - break; + } break; case iro_Sub: case iro_Shl: case iro_Shr: @@ -353,23 +377,24 @@ equivalent_node (ir_node *n) case iro_Rot: /* these operations are not commutative. Test only one predecessor. */ if (tarval_classify (computed_value (b)) == 0) { - n = a; + n = a; DBG_OPT_ALGSIM1; /* Test if b > #bits of a ==> return 0 / divide b by #bits --> transform node? */ } break; case iro_Not: /* NotNot x == x */ case iro_Minus: /* --x == x */ /* ??? Is this possible or can --x raise an - out of bounds exception if min =! max? */ - if (get_irn_op(get_unop_op(n)) == get_irn_op(n)) - n = get_unop_op(get_unop_op(n)); + out of bounds exception if min =! max? */ + if (get_irn_op(get_unop_op(n)) == get_irn_op(n)) { + n = get_unop_op(get_unop_op(n)); DBG_OPT_ALGSIM2 + } break; case iro_Mul: /* Mul is commutative and has again an other neutral element. */ if (tarval_classify (computed_value (a)) == 1) { - n = b; + n = b; DBG_OPT_ALGSIM1 } else if (tarval_classify (computed_value (b)) == 1) { - n = a; + n = a; DBG_OPT_ALGSIM1 } break; case iro_Div: @@ -383,26 +408,28 @@ equivalent_node (ir_node *n) set_Tuple_pred(n, 2, a); } break; - /* GL: Why are they skipped? DivMod allocates new nodes --> it's - teated in transform node. + /* case iro_Mod, Quot, DivMod - */ + DivMod allocates new nodes --> it's treated in transform node. + What about Quot, DivMod? + */ case iro_And: - if (a == b) n = a; - /* And has it's own neutral element */ - else if (tarval_classify (computed_value (a)) == -1) { + if (a == b) { + n = a; /* And has it's own neutral element */ + } else if (tarval_classify (computed_value (a)) == -1) { n = b; } else if (tarval_classify (computed_value (b)) == -1) { n = a; } + if (n != oldn) DBG_OPT_ALGSIM1; break; case iro_Conv: if (get_irn_mode(n) == get_irn_mode(a)) { /* No Conv necessary */ - n = a; + n = a; DBG_OPT_ALGSIM3; } else if (get_irn_mode(n) == mode_b) { if (get_irn_op(a) == op_Conv && get_irn_mode (get_Conv_op(a)) == mode_b) { - n = get_Conv_op(a); /* Convb(Conv*(xxxb(...))) == xxxb(...) */ + n = get_Conv_op(a); /* Convb(Conv*(xxxb(...))) == xxxb(...) */ DBG_OPT_ALGSIM2; } } break; @@ -413,11 +440,10 @@ equivalent_node (ir_node *n) - no Phi in start block. - remove Id operators that are inputs to Phi - fold Phi-nodes, iff they have only one predecessor except - themselves. + themselves. */ int i, n_preds; - ir_node *block = NULL; /* to shutup gcc */ ir_node *first_val = NULL; /* to shutup gcc */ ir_node *scnd_val = NULL; /* to shutup gcc */ @@ -425,18 +451,13 @@ equivalent_node (ir_node *n) n_preds = get_Phi_n_preds(n); block = get_nodes_Block(n); - assert(get_irn_op (block) == op_Block); + /* @@@ 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 */ + (block == current_ir_graph->start_block)) /* There should be no Phi nodes */ + return new_Bad(); /* in the Start Block. */ - /* there should be no Phi nodes in the Start region. */ - if (block == current_ir_graph->start_block) { - n = new_Bad(); - break; - } - - if (n_preds == 0) { /* Phi of dead Region without predecessors. */ - /* GL: why not return new_Bad? */ - break; - } + if (n_preds == 0) break; /* Phi of dead Region without predecessors. */ #if 0 /* first we test for a special case: */ @@ -447,9 +468,9 @@ equivalent_node (ir_node *n) ir_node *a = follow_Id (get_Phi_pred(n, 0)); ir_node *b = follow_Id (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)) == follow_Id(get_irn_n(b, 0))) - && (get_irn_n(a, 1) == get_irn_n (b, 1)) + && (get_irn_op(b) == op_Confirm) + && (follow_Id (get_irn_n(a, 0)) == follow_Id(get_irn_n(b, 0))) + && (get_irn_n(a, 1) == get_irn_n (b, 1)) && (a->data.num == (~b->data.num & irpn_True) )) { n = follow_Id (get_irn_n(a, 0)); break; @@ -463,8 +484,8 @@ equivalent_node (ir_node *n) /* skip Id's */ set_Phi_pred(n, i, first_val); if ( (first_val != n) /* not self pointer */ - && (get_irn_op(first_val) != op_Bad) /* value not dead */ - && !(is_Bad (get_Block_cfgpred(block, i))) ) { /* not dead control flow */ + && (get_irn_op(first_val) != op_Bad) /* value not dead */ + && !(is_Bad (get_Block_cfgpred(block, i))) ) { /* not dead control flow */ break; /* then found first value. */ } } @@ -482,17 +503,17 @@ equivalent_node (ir_node *n) set_Phi_pred(n, i, scnd_val); if ( (scnd_val != n) && (scnd_val != first_val) - && (get_irn_op(scnd_val) != op_Bad) - && !(is_Bad (get_Block_cfgpred(block, i))) ) { + && (get_irn_op(scnd_val) != op_Bad) + && !(is_Bad (get_Block_cfgpred(block, i))) ) { break; } } /* Fold, if no multiple distinct non-self-referencing inputs */ if (i >= n_preds) { - n = first_val; + n = first_val; DBG_OPT_PHI; } else { - /* skip the remaining Ids. */ + /* skip the remaining Ids. */ while (++i < n_preds) { set_Phi_pred(n, i, follow_Id(get_Phi_pred(n, i))); } @@ -504,20 +525,20 @@ equivalent_node (ir_node *n) { #if 0 /* Is an illegal transformation: different nodes can represent the same pointer value!! */ - a = skip_Proj(get_Load_mem(n)); - b = get_Load_ptr(n); - - if (get_irn_op(a) == op_Store) { - if ( different_identity (b, get_Store_ptr(a))) { - /* load and store use different pointers, therefore load - needs not take store's memory but the state before. */ - set_Load_mem (n, get_Store_mem(a)); - } else if (( 0 /* ???didn't get cryptic test that returns 0 */ )) { - } - } + a = skip_Proj(get_Load_mem(n)); + b = get_Load_ptr(n); + + if (get_irn_op(a) == op_Store) { + if ( different_identity (b, get_Store_ptr(a))) { + /* load and store use different pointers, therefore load + needs not take store's memory but the state before. */ + set_Load_mem (n, get_Store_mem(a)); + } else if (( 0 /* ???didn't get cryptic test that returns 0 */ )) { + } + } #endif } - break; + break; case iro_Store: /* remove unnecessary store. */ { @@ -529,17 +550,17 @@ equivalent_node (ir_node *n) && get_Store_ptr(a) == b && skip_Proj(get_Store_value(a)) == c) { /* We have twice exactly the same store -- a write after write. */ - n = a; + n = a; DBG_OPT_WAW; } else if (get_irn_op(c) == op_Load && (a == c || skip_Proj(get_Load_mem(c)) == a) - && get_Load_ptr(c) == b ) - /* !!!??? and a cryptic test */ { + && get_Load_ptr(c) == b ) { /* We just loaded the value from the same memory, i.e., the store doesn't change the memory -- a write after read. */ + a = get_Store_mem(n); turn_into_tuple(n, 2); set_Tuple_pred(n, 0, a); - set_Tuple_pred(n, 1, new_Bad()); - } + set_Tuple_pred(n, 1, new_Bad()); DBG_OPT_WAR; + } } break; @@ -550,21 +571,21 @@ equivalent_node (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)); + n = get_Tuple_pred(a, get_Proj_proj(n)); DBG_OPT_TUPLE; } else { assert(0); /* This should not happen! */ n = new_Bad(); } } else if (get_irn_mode(n) == mode_X && is_Bad(get_nodes_Block(n))) { - /* Remove dead control flow. */ + /* Remove dead control flow -- early gigo. */ n = new_Bad(); } } break; case iro_Id: - n = follow_Id (n); + n = follow_Id (n); DBG_OPT_ID; break; default: break; @@ -574,7 +595,7 @@ equivalent_node (ir_node *n) } /* end equivalent_node() */ -/* tries several [inplace] [optimizing] transformations and returns a +/* tries several [inplace] [optimizing] transformations and returns an equivalent node. The difference to equivalent_node is that these transformations _do_ generate new nodes, and thus the old node must not be freed even if the equivalent node isn't the old one. */ @@ -586,6 +607,28 @@ transform_node (ir_node *n) tarval *ta, *tb; switch (get_irn_opcode(n)) { + case iro_Div: { + ta = computed_value(n); + if (ta) { + /* Turn Div into a tuple (mem, bad, value) */ + ir_node *mem = get_Div_mem(n); + turn_into_tuple(n, 3); + set_Tuple_pred(n, 0, mem); + set_Tuple_pred(n, 1, new_Bad()); + set_Tuple_pred(n, 2, new_Const(get_tv_mode(ta), ta)); + } + } break; + case iro_Mod: { + ta = computed_value(n); + if (ta) { + /* Turn Div into a tuple (mem, bad, value) */ + ir_node *mem = get_Mod_mem(n); + turn_into_tuple(n, 3); + set_Tuple_pred(n, 0, mem); + set_Tuple_pred(n, 1, new_Bad()); + set_Tuple_pred(n, 2, new_Const(get_tv_mode(ta), ta)); + } + } break; case iro_DivMod: { int evaluated = 0; @@ -595,8 +638,8 @@ transform_node (ir_node *n) b = get_DivMod_right(n); mode = get_irn_mode(a); - if (!( mode_is_int(get_irn_mode(a)) - && mode_is_int(get_irn_mode(b)))) + if (!(mode_is_int(get_irn_mode(a)) && + mode_is_int(get_irn_mode(b)))) break; if (a == b) { @@ -646,33 +689,42 @@ transform_node (ir_node *n) a = get_Cond_selector(n); ta = value_of(a); - if (ta && (get_irn_mode(a) == mode_b)) { + if (ta && + (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 (tv_val_b(ta) == 1) /* GL: I hope this returns 1 if true */ { - set_Tuple_pred(n, 0, new_Bad()); - set_Tuple_pred(n, 1, jmp); + set_Tuple_pred(n, 0, new_Bad()); + set_Tuple_pred(n, 1, jmp); } else { - set_Tuple_pred(n, 0, jmp); - set_Tuple_pred(n, 1, new_Bad()); + set_Tuple_pred(n, 0, jmp); + set_Tuple_pred(n, 1, new_Bad()); } - } else if (ta && (get_irn_mode(a) == mode_I) && (get_Cond_kind(n) == dense)) { + /* 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 && + (get_irn_mode(a) == mode_I) && + (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 when optimizing Proj. */ set_irn_link(n, new_r_Jmp(current_ir_graph, get_nodes_Block(n))); - } else if ( (get_irn_op(get_Cond_selector(n)) == op_Eor) - && (get_irn_mode(get_Cond_selector(n)) == mode_b) - && (tarval_classify(computed_value(get_Eor_right(a))) == 1)) { + /* 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(get_Cond_selector(n)) == op_Eor) + && (get_irn_mode(get_Cond_selector(n)) == mode_b) + && (tarval_classify(computed_value(get_Eor_right(a))) == 1)) { /* 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))); - } else if ( (get_irn_op(get_Cond_selector(n)) == op_Not) - && (get_irn_mode(get_Cond_selector(n)) == mode_b)) { + } else if ((get_irn_op(get_Cond_selector(n)) == op_Not) + && (get_irn_mode(get_Cond_selector(n)) == 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), @@ -684,26 +736,26 @@ transform_node (ir_node *n) case iro_Proj: { a = get_Proj_pred(n); - if ( (get_irn_op(a) == op_Cond) - && get_irn_link(a) - && get_irn_op(get_irn_link(a)) == op_Cond) { - /* Use the better Cond if the Proj projs from a Cond which get's - its result from an Eor/Not. */ - assert ( ( (get_irn_op(get_Cond_selector(a)) == op_Eor) - || (get_irn_op(get_Cond_selector(a)) == op_Not)) - && (get_irn_mode(get_Cond_selector(a)) == mode_b) + if ((get_irn_op(a) == op_Cond) + && get_irn_link(a) + && get_irn_op(get_irn_link(a)) == op_Cond) { + /* Use the better Cond if the Proj projs from a Cond which get's + its result from an Eor/Not. */ + assert (((get_irn_op(get_Cond_selector(a)) == op_Eor) + || (get_irn_op(get_Cond_selector(a)) == op_Not)) + && (get_irn_mode(get_Cond_selector(a)) == mode_b) && (get_irn_op(get_irn_link(a)) == op_Cond) - && (get_Cond_selector(get_irn_link(a)) == - get_Eor_left(get_Cond_selector(a)))); + && (get_Cond_selector(get_irn_link(a)) == get_Eor_left(get_Cond_selector(a)))); set_Proj_pred(n, get_irn_link(a)); if (get_Proj_proj(n) == 0) set_Proj_proj(n, 1); else set_Proj_proj(n, 0); - } else if ( (get_irn_op(a) == op_Cond) - && (get_irn_mode(get_Cond_selector(a)) == mode_I) + } else if ((get_irn_op(a) == op_Cond) + && (get_irn_mode(get_Cond_selector(a)) == mode_I) && value_of(a) - && (get_Cond_kind(a) == dense)) { + && (get_Cond_kind(a) == dense) + && (get_opt_unreachable_code())) { /* The Cond is a Switch on a Constant */ if (get_Proj_proj(n) == tv_val_CHIL(value_of(a))) { /* The always taken branch, reuse the existing Jmp. */ @@ -713,7 +765,7 @@ transform_node (ir_node *n) n = get_irn_link(a); } else {/* Not taken control flow, but be careful with the default! */ if (get_Proj_proj(n) < a->attr.c.default_proj){ - /* a never taken branch */ + /* a never taken branch */ n = new_Bad(); } else { a->attr.c.default_proj = get_Proj_proj(n); @@ -725,15 +777,15 @@ transform_node (ir_node *n) a = get_Eor_left(n); b = get_Eor_right(n); - if ( (get_irn_mode(n) == mode_b) + if ((get_irn_mode(n) == mode_b) && (get_irn_op(a) == op_Proj) - && (get_irn_mode(a) == mode_b) + && (get_irn_mode(a) == mode_b) && (tarval_classify (computed_value (b)) == 1) && (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) + else if ((get_irn_mode(n) == mode_b) && (tarval_classify (computed_value (b)) == 1)) /* The Eor is a Not. Replace it by a Not. */ /* ????!!!Extend to bitfield 1111111. */ @@ -745,7 +797,7 @@ transform_node (ir_node *n) if ( (get_irn_mode(n) == mode_b) && (get_irn_op(a) == op_Proj) - && (get_irn_mode(a) == mode_b) + && (get_irn_mode(a) == mode_b) && (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), @@ -759,9 +811,8 @@ transform_node (ir_node *n) /* **************** Common Subexpression Elimination **************** */ -/* Compare function for two nodes in the hash table. Gets two */ -/* nodes as parameters. */ -/* @@@ a+b != b+a ? */ +/* Compare function for two nodes in the hash table. Gets two */ +/* nodes as parameters. Returns 0 if the nodes are a cse. */ static int vt_cmp (const void *elt, const void *key) { @@ -777,29 +828,27 @@ vt_cmp (const void *elt, const void *key) (get_irn_mode(a) != get_irn_mode(b))) return 1; /* compare if a's in and b's in are equal */ - /* GL: we optimize only nodes with in arrays of fixed sizes. - if (get_irn_arity (a) != -2) { - ins = get_irn_arity (a); - if (ins != get_irn_arity (b)) return 1; - ain = get_irn_in (a); - bin = get_irn_in (b); - } - */ if (get_irn_arity (a) != get_irn_arity(b)) return 1; - /* compare a->in[0..ins] with b->in[0..ins], i.e., include the block. */ - /* do if (*ain++ != *bin++) return 1; while (ins--); */ - for (i = -1; i < get_irn_arity(a); i++) - if (get_irn_n(a, i) != get_irn_n(b, i)) + /* for block-local cse and pinned nodes: */ + if (!get_opt_global_cse() || (get_op_pinned(get_irn_op(a)) == pinned)) { + if (get_irn_n(a, -1) != get_irn_n(b, -1)) return 1; + } + /* compare a->in[0..ins] with b->in[0..ins] */ + for (i = 0; i < get_irn_arity(a); i++) + if (get_irn_n(a, i) != get_irn_n(b, i)) + return 1; switch (get_irn_opcode(a)) { case iro_Const: return get_irn_const_attr (a) != get_irn_const_attr (b); case iro_Proj: return get_irn_proj_attr (a) != get_irn_proj_attr (b); + case iro_Filter: + return get_Filter_proj(a) != get_Filter_proj(b); case iro_Alloc: 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); @@ -815,8 +864,7 @@ vt_cmp (const void *elt, const void *key) || (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).ltyp != get_irn_sel_attr(b).ltyp); + || (get_irn_sel_attr(a).ent->type != get_irn_sel_attr(b).ent->type); case iro_Phi: return get_irn_phi_attr (a) != get_irn_phi_attr (b); default: ; @@ -861,36 +909,51 @@ del_identities (pset *value_table) /* Return the canonical node computing the same value as n. Looks up the node in a hash table. */ -static inline ir_node * +static INLINE ir_node * identify (pset *value_table, ir_node *n) { ir_node *o = NULL; if (!value_table) return n; - switch (get_irn_opcode (n)) { - case iro_Add: - case iro_Mul: - case iro_Or: - case iro_And: - case iro_Eor: - { - /* 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 (get_opt_reassociation()) { + switch (get_irn_opcode (n)) { + case iro_Add: + case iro_Mul: + case iro_Or: + case iro_And: + case iro_Eor: + { + /* 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); + } } + break; + default: break; } - break; - default: break; } + o = pset_find (value_table, n, ir_node_hash (n)); if (!o) return n; return o; } +/* During construction we set the pinned flag in the graph right when the + optimizatin is performed. The flag turning on procedure global cse could + be changed between two allocations. This way we are safe. */ +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, floats); + return n; +} + /* Return the canonical node computing the same value as n. Looks up the node in a hash table, enters it in the table if it isn't there yet. */ @@ -916,7 +979,7 @@ add_identities (pset *value_table, ir_node *node) { /* garbage in, garbage out. If a node has a dead input, i.e., the Bad node is input to the node, return the Bad node. */ -static inline ir_node * +static INLINE ir_node * gigo (ir_node *node) { int i; @@ -957,19 +1020,13 @@ optimize (ir_node *n) /* Allways optimize Phi nodes: part of the construction. */ if ((!get_optimize()) && (get_irn_op(n) != op_Phi)) return n; - /* if not optimize return n */ - if (n == NULL) { - printf(" attention: empty node!!! \n"); - return n; - } - /* constant expression evaluation / constant folding */ if (get_opt_constant_folding()) { /* constants can not be evaluated */ if (get_irn_op(n) != op_Const) { /* try to evaluate */ tv = computed_value (n); - if (tv != NULL) { + if ((get_irn_mode(n) != mode_T) && (tv != NULL)) { /* evaluation was succesful -- replace the node. */ obstack_free (current_ir_graph->obst, n); return new_Const (get_tv_mode (tv), tv); @@ -978,7 +1035,11 @@ optimize (ir_node *n) } /* remove unnecessary nodes */ - if (get_opt_constant_folding() || get_irn_op(n) == op_Phi) + if (get_opt_constant_folding() || + (get_irn_op(n) == op_Phi) || /* always optimize these nodes. */ + (get_irn_op(n) == op_Id) || + (get_irn_op(n) == op_Proj) || + (get_irn_op(n) == op_Block) ) /* Flags tested local. */ n = equivalent_node (n); /** common subexpression elimination **/ @@ -987,35 +1048,32 @@ optimize (ir_node *n) now all nodes are pinned to blocks, i.e., the cse only finds common subexpressions within a block. */ if (get_opt_cse()) - n = identify (current_ir_graph->value_table, n); - /* identify found a cse, so deallocate the old node. */ + n = identify_cons (current_ir_graph->value_table, n); + if (n != old_n) { + /* We found an existing, better node, so we can deallocate the old node. */ obstack_free (current_ir_graph->obst, old_n); - /* The AmRoq fiasco returns n here. Martin's version doesn't. */ } /* Some more constant expression evaluation that does not allow to free the node. */ - if (get_opt_constant_folding()) + if (get_opt_constant_folding() || + (get_irn_op(n) == op_Cond) || + (get_irn_op(n) == op_Proj)) /* Flags tested local. */ n = transform_node (n); - /* Remove nodes with dead (Bad) input. */ - if (get_opt_unreachable_code()) - n = gigo (n); + /* Remove nodes with dead (Bad) input. + Run always for transformation induced Bads. */ + n = gigo (n); + /* Now we can verify the node, as it has no dead inputs any more. */ irn_vrfy(n); /* Now we have a legal, useful node. Enter it in hash table for cse */ - if (get_opt_cse()) { + if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) { n = identify_remember (current_ir_graph->value_table, n); } -#if 0 /* GL: what's the use of this?? */ - if ((current_ir_graph->state & irgs_building) && IR_KEEP_ALIVE (n)) { - assert (~current_ir_graph->state & irgs_keep_alives_in_arr); - pdeq_putr (current_ir_graph->keep.living, n); - } -#endif return n; } @@ -1024,38 +1082,43 @@ optimize (ir_node *n) nodes lying on the obstack. Remove these by a dead node elimination, i.e., a copying garbage collection. */ ir_node * -optimize_in_place (ir_node *n) +optimize_in_place_2 (ir_node *n) { tarval *tv; ir_node *old_n = n; - if (!get_optimize()) return n; + if (!get_optimize() && (get_irn_op(n) != op_Phi)) return n; /* if not optimize return n */ if (n == NULL) { + assert(0); /* Here this is possible. Why? */ return n; } + /* constant expression evaluation / constant folding */ if (get_opt_constant_folding()) { /* constants can not be evaluated */ if (get_irn_op(n) != op_Const) { /* try to evaluate */ tv = computed_value (n); - if (tv != NULL) { + if ((get_irn_mode(n) != mode_T) && (tv != NULL)) { /* evaluation was succesful -- replace the node. */ n = new_Const (get_tv_mode (tv), tv); - deb_info_copy(n, old_n, id_from_str("const_eval", 10)); + __dbg_info_merge_pair(n, old_n, dbg_const_eval); return n; - /* xprintf("* optimize: computed node %I\n", n->op->name);*/ } } } /* remove unnecessary nodes */ /*if (get_opt_constant_folding()) */ - if (get_opt_constant_folding() || get_irn_op(n) == op_Phi) + if (get_opt_constant_folding() || + (get_irn_op(n) == op_Phi) || /* always optimize these nodes. */ + (get_irn_op(n) == op_Id) || /* ... */ + (get_irn_op(n) == op_Proj) || /* ... */ + (get_irn_op(n) == op_Block) ) /* Flags tested local. */ n = equivalent_node (n); /** common subexpression elimination **/ @@ -1063,21 +1126,20 @@ optimize_in_place (ir_node *n) /* The block input is used to distinguish different subexpressions. Right now all nodes are pinned to blocks, i.e., the cse only finds common subexpressions within a block. */ - if (get_opt_cse()) + if (get_opt_cse()) { n = identify (current_ir_graph->value_table, n); - - /* identify found a cse, so deallocate the old node. */ - if (n != old_n) { - /* The AmRoq fiasco returns n here. Martin's version doesn't. */ } /* Some more constant expression evaluation. */ - if (get_opt_constant_folding()) + if (get_opt_constant_folding() || + (get_irn_op(n) == op_Cond) || + (get_irn_op(n) == op_Proj)) /* Flags tested local. */ n = transform_node (n); - /* Remove nodes with dead (Bad) input. */ - if (get_opt_unreachable_code()) - n = gigo (n); + /* Remove nodes with dead (Bad) input. + Run always for transformation induced Bads. */ + n = gigo (n); + /* Now we can verify the node, as it has no dead inputs any more. */ irn_vrfy(n); @@ -1089,3 +1151,19 @@ optimize_in_place (ir_node *n) return n; } + +/* Wrapper for external use, set proper status bits after optimization */ +ir_node * +optimize_in_place (ir_node *n) { + /* Handle graph state */ + assert(get_irg_phase_state(current_ir_graph) != phase_building); + if (get_opt_global_cse()) + set_irg_pinned(current_ir_graph, floats); + if (get_irg_outs_state(current_ir_graph) == outs_consistent) + set_irg_outs_inconsistent(current_ir_graph); + /* Maybe we could also test whether optimizing the node can + change the control graph. */ + if (get_irg_dom_state(current_ir_graph) == dom_consistent) + set_irg_dom_inconsistent(current_ir_graph); + return optimize_in_place_2 (n); +}