X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firopt.c;h=b9a1246f9e858cfddfdf88bc5f5075ed980e5038;hb=428ca393b64f346987ae1abf54017cdbff4a2605;hp=6d862ab45ec4362812a84a6a8ae0565c46a34d3b;hpb=54e122b5201b5efc22978c9c5039571e7270dcde;p=libfirm diff --git a/ir/ir/iropt.c b/ir/ir/iropt.c index 6d862ab45..b9a1246f9 100644 --- a/ir/ir/iropt.c +++ b/ir/ir/iropt.c @@ -18,10 +18,10 @@ # include "irgraph_t.h" # include "irmode_t.h" # include "iropt_t.h" -# include "ircons.h" +# include "ircons_t.h" # include "irgmod.h" # include "irvrfy.h" -# include "tv.h" +# include "tv_t.h" # include "dbginfo_t.h" # include "iropt_dbg.h" # include "irflag_t.h" @@ -53,16 +53,22 @@ value_of (ir_node *n) return tarval_bad; } +/** + * return the value of a Constant + */ static tarval *computed_value_Const(ir_node *n) { return get_Const_tarval(n); } +/** + * return the value of a 'sizeof' SymConst + */ static tarval *computed_value_SymConst(ir_node *n) { - if ((get_SymConst_kind(n) == size) && + if ((get_SymConst_kind(n) == symconst_size) && (get_type_state(get_SymConst_type(n))) == layout_fixed) - return new_tarval_from_long (get_type_size(get_SymConst_type(n)), mode_Is); + return new_tarval_from_long(get_type_size_bytes(get_SymConst_type(n)), get_irn_mode(n)); return tarval_bad; } @@ -207,8 +213,8 @@ static tarval *computed_value_And(ir_node *n) } else { tarval *v; - if ( (tarval_classify ((v = computed_value (a))) == TV_CLASSIFY_NULL) - || (tarval_classify ((v = computed_value (b))) == TV_CLASSIFY_NULL)) { + if ( (classify_tarval ((v = computed_value (a))) == TV_CLASSIFY_NULL) + || (classify_tarval ((v = computed_value (b))) == TV_CLASSIFY_NULL)) { return v; } } @@ -227,8 +233,8 @@ static tarval *computed_value_Or(ir_node *n) return tarval_or (ta, tb); } else { tarval *v; - if ( (tarval_classify ((v = computed_value (a))) == TV_CLASSIFY_ALL_ONE) - || (tarval_classify ((v = computed_value (b))) == TV_CLASSIFY_ALL_ONE)) { + if ( (classify_tarval ((v = computed_value (a))) == TV_CLASSIFY_ALL_ONE) + || (classify_tarval ((v = computed_value (b))) == TV_CLASSIFY_ALL_ONE)) { return v; } } @@ -311,7 +317,7 @@ static tarval *computed_value_Rot(ir_node *n) tarval *tb = value_of(b); if ((ta != tarval_bad) && (tb != tarval_bad)) { - /* return tarval_rot (ta, tb); */ + return tarval_rot (ta, tb); } return tarval_bad; } @@ -346,12 +352,13 @@ static tarval *computed_value_Proj(ir_node *n) 3. The predecessors are Allocs or void* constants. Allocs never return NULL, they raise an exception. Therefore we can predict the Cmp result. */ - if (get_irn_op(a) == op_Cmp) { + switch (get_irn_opcode(a)) { + case iro_Cmp: aa = get_Cmp_left(a); ab = get_Cmp_right(a); if (aa == ab) { /* 1.: */ - /* This is a tric with the bits used for encoding the Cmp + /* 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); @@ -366,8 +373,8 @@ static tarval *computed_value_Proj(ir_node *n) return new_tarval_from_long (get_Proj_proj(n) & flags, mode_b); } } else { /* check for 3.: */ - ir_node *aaa = skip_nop(skip_Proj(aa)); - ir_node *aba = skip_nop(skip_Proj(ab)); + 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) @@ -393,18 +400,54 @@ static tarval *computed_value_Proj(ir_node *n) return new_tarval_from_long (get_Proj_proj(n) & Ne, mode_b); } } - } else if (get_irn_op(a) == op_DivMod) { + 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)== 0) /* Div */ + 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); + } + 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); - else /* Mod */ + } + 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); } + break; + } + + default: + return tarval_bad; } return tarval_bad; } @@ -413,7 +456,7 @@ static tarval *computed_value_Proj(ir_node *n) * If the parameter n can be computed, return its value, else tarval_bad. * Performs constant folding. * - * GL: Only if n is arithmetic operator? + * @param n The node this should be evaluated */ tarval *computed_value(ir_node *n) { @@ -484,7 +527,7 @@ static ir_node *equivalent_node_Block(ir_node *n) { ir_node *oldn = n; - /* The Block constructor does not call optimize, but mature_block + /* The Block constructor does not call optimize, but mature_immBlock calls the optimization. */ assert(get_Block_matured(n)); @@ -494,16 +537,26 @@ static ir_node *equivalent_node_Block(ir_node *n) 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) && - (get_opt_control_flow_straightening())) { - n = get_nodes_Block(get_Block_cfgpred(n, 0)); - if (n == oldn) - /* Jmp jumps into the block it is in -- deal self cycle. */ - n = new_Bad(); DBG_OPT_STG; - - } else if ((get_Block_n_cfgpreds(n) == 2) && - (get_opt_control_flow_weak_simplification())) { + if ((get_Block_n_cfgpreds(n) == 1) && + (get_irn_op(get_Block_cfgpred(n, 0)) == op_Jmp)) { + 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; + } else if (get_opt_control_flow_straightening()) { + n = predblock; DBG_OPT_STG; + } + } + else if ((get_Block_n_cfgpreds(n) == 1) && + (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; + } + } + else if ((get_Block_n_cfgpreds(n) == 2) && + (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); @@ -516,7 +569,7 @@ static ir_node *equivalent_node_Block(ir_node *n) (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; + n = get_nodes_block(a); DBG_OPT_IFSIM; } } else if (get_opt_unreachable_code() && (n != current_ir_graph->start_block) && @@ -535,11 +588,15 @@ static ir_node *equivalent_node_Block(ir_node *n) return n; } +/** + * Returns a equivalent node for a Jmp, a Bad :-) + * Of course this only happens if the Block of the Jmp is Bad. + */ 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_Bad(get_nodes_block(n))) n = new_Bad(); return n; @@ -552,6 +609,9 @@ 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; @@ -568,7 +628,8 @@ static ir_node *equivalent_node_Or(ir_node *n) } /** - * optimize operations that are commutative and have neutral 0. + * optimize operations that are commutative and have neutral 0, + * so a op 0 = 0 op a = a. */ static ir_node *equivalent_node_neutral_zero(ir_node *n) { @@ -582,35 +643,28 @@ 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 = computed_value(a)) != tarval_bad) { on = b; - } else if ((tv = computed_value (b)) != tarval_bad) { + } else if ((tv = computed_value(b)) != tarval_bad) { on = a; } else return n; /* If this predecessors constant value is zero, the operation is unnecessary. Remove it: */ - if (tarval_classify (tv) == TV_CLASSIFY_NULL) { + if (classify_tarval (tv) == TV_CLASSIFY_NULL) { n = on; DBG_OPT_ALGSIM1; } return n; } -static ir_node *equivalent_node_Add(ir_node *n) -{ - return equivalent_node_neutral_zero(n); -} - -static ir_node *equivalent_node_Eor(ir_node *n) -{ - return equivalent_node_neutral_zero(n); -} +#define equivalent_node_Add equivalent_node_neutral_zero +#define equivalent_node_Eor equivalent_node_neutral_zero /** - * optimize operations that are not commutative but have neutral 0 on left. - * Test only one predecessor. + * optimize operations that are not commutative but have neutral 0 on left, + * so a op 0 = a. */ static ir_node *equivalent_node_left_zero(ir_node *n) { @@ -619,38 +673,22 @@ 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 (tarval_classify (computed_value (b)) == TV_CLASSIFY_NULL) { + if (classify_tarval(computed_value(b)) == TV_CLASSIFY_NULL) { n = a; DBG_OPT_ALGSIM1; } return n; } -static ir_node *equivalent_node_Sub(ir_node *n) -{ - return equivalent_node_left_zero(n); -} - -static ir_node *equivalent_node_Shl(ir_node *n) -{ - return equivalent_node_left_zero(n); -} - -static ir_node *equivalent_node_Shr(ir_node *n) -{ - return equivalent_node_left_zero(n); -} - -static ir_node *equivalent_node_Shrs(ir_node *n) -{ - return equivalent_node_left_zero(n); -} - -static ir_node *equivalent_node_Rot(ir_node *n) -{ - return equivalent_node_left_zero(n); -} +#define equivalent_node_Sub equivalent_node_left_zero +#define equivalent_node_Shl equivalent_node_left_zero +#define equivalent_node_Shr equivalent_node_left_zero +#define equivalent_node_Shrs equivalent_node_left_zero +#define equivalent_node_Rot equivalent_node_left_zero +/** + * Er, a "symmetic unop", ie op(op(n)) = n. + */ static ir_node *equivalent_node_symmetric_unop(ir_node *n) { ir_node *oldn = n; @@ -662,19 +700,16 @@ static ir_node *equivalent_node_symmetric_unop(ir_node *n) return n; } -static ir_node *equivalent_node_Not(ir_node *n) -{ - /* NotNot x == x */ - return equivalent_node_symmetric_unop(n); -} +/* NotNot x == x */ +#define equivalent_node_Not equivalent_node_symmetric_unop -static ir_node *equivalent_node_Minus(ir_node *n) -{ - /* --x == x */ /* ??? Is this possible or can --x raise an - out of bounds exception if min =! max? */ - return equivalent_node_symmetric_unop(n); -} +/* --x == x */ /* ??? Is this possible or can --x raise an + out of bounds exception if min =! max? */ +#define equivalent_node_Minus equivalent_node_symmetric_unop +/** + * Optimize a * 1 = 1 * a = a. + */ static ir_node *equivalent_node_Mul(ir_node *n) { ir_node *oldn = n; @@ -683,21 +718,24 @@ 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 (tarval_classify (computed_value (a)) == TV_CLASSIFY_ONE) { + if (classify_tarval (computed_value (a)) == TV_CLASSIFY_ONE) { n = b; DBG_OPT_ALGSIM1; - } else if (tarval_classify (computed_value (b)) == TV_CLASSIFY_ONE) { + } else if (classify_tarval (computed_value (b)) == TV_CLASSIFY_ONE) { n = a; DBG_OPT_ALGSIM1; } return n; } +/** + * Optimize a / 1 = a. + */ static ir_node *equivalent_node_Div(ir_node *n) { ir_node *a = get_Div_left(n); ir_node *b = get_Div_right(n); /* Div is not commutative. */ - if (tarval_classify (computed_value (b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */ + if (classify_tarval(computed_value(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); @@ -708,6 +746,9 @@ static ir_node *equivalent_node_Div(ir_node *n) return n; } +/** + * Optimize a & 0b1...1 = 0b1...1 & a = a & a = a. + */ static ir_node *equivalent_node_And(ir_node *n) { ir_node *oldn = n; @@ -717,15 +758,18 @@ static ir_node *equivalent_node_And(ir_node *n) if (a == b) { n = a; /* And has it's own neutral element */ - } else if (tarval_classify (computed_value (a)) == TV_CLASSIFY_ALL_ONE) { + } else if (classify_tarval(computed_value(a)) == TV_CLASSIFY_ALL_ONE) { n = b; - } else if (tarval_classify (computed_value (b)) == TV_CLASSIFY_ALL_ONE) { + } else if (classify_tarval(computed_value(b)) == TV_CLASSIFY_ALL_ONE) { n = a; } if (n != oldn) DBG_OPT_ALGSIM1; return n; } +/** + * Try to remove useless conv's: + */ static ir_node *equivalent_node_Conv(ir_node *n) { ir_node *oldn = n; @@ -777,7 +821,7 @@ static ir_node *equivalent_node_Phi(ir_node *n) n_preds = get_Phi_n_preds(n); - block = get_nodes_Block(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 */ @@ -792,8 +836,8 @@ static ir_node *equivalent_node_Phi(ir_node *n) value that is known at a certain point. This is useful for dataflow analysis. */ if (n_preds == 2) { - ir_node *a = follow_Id (get_Phi_pred(n, 0)); - ir_node *b = follow_Id (get_Phi_pred(n, 1)); + 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)) @@ -804,15 +848,20 @@ static ir_node *equivalent_node_Phi(ir_node *n) } #endif + /* If the Block has a Bad pred, we also have one. */ + for (i = 0; i < n_preds; ++i) + if (is_Bad (get_Block_cfgpred(block, i))) + set_Phi_pred(n, i, new_Bad()); + /* Find first non-self-referencing input */ for (i = 0; i < n_preds; ++i) { - first_val = follow_Id(get_Phi_pred(n, i)); - /* skip Id's */ - set_Phi_pred(n, i, first_val); + first_val = get_Phi_pred(n, i); 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 */ - break; /* then found first value. */ +#if 1 + && (get_irn_op(first_val) != op_Bad) +#endif + ) { /* value not dead */ + break; /* then found first value. */ } } @@ -824,13 +873,13 @@ static ir_node *equivalent_node_Phi(ir_node *n) /* follow_Id () for rest of inputs, determine if any of these are non-self-referencing */ while (++i < n_preds) { - scnd_val = follow_Id(get_Phi_pred(n, i)); - /* skip Id's */ - set_Phi_pred(n, i, scnd_val); + scnd_val = get_Phi_pred(n, i); if ( (scnd_val != n) && (scnd_val != first_val) +#if 1 && (get_irn_op(scnd_val) != op_Bad) - && !(is_Bad (get_Block_cfgpred(block, i))) ) { +#endif + ) { break; } } @@ -839,33 +888,46 @@ static ir_node *equivalent_node_Phi(ir_node *n) if (i >= n_preds) { n = first_val; DBG_OPT_PHI; } else { - /* skip the remaining Ids. */ - while (++i < n_preds) { - set_Phi_pred(n, i, follow_Id(get_Phi_pred(n, i))); - } + /* skip the remaining Ids (done in get_Phi_pred). */ + /* superfluous, since we walk all to propagate Block's Bads. + while (++i < n_preds) get_Phi_pred(n, i); */ } return n; } +/** + * Optimize Loads after Store. + * + * @todo FAILS for volatile entities + */ static ir_node *equivalent_node_Load(ir_node *n) { -#if 0 /* Is an illegal transformation: different nodes can - represent the same pointer value!! */ - ir_node *a = skip_Proj(get_Load_mem(n)); - ir_node *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 - return n; + ir_node *oldn = n; + + /* remove unnecessary Load. */ + ir_node *a = skip_Proj(get_Load_mem(n)); + ir_node *b = get_Load_ptr(n); + ir_node *c; + + /* TODO: check for volatile */ + if (get_irn_op(a) == op_Store && get_Store_ptr(a) == b) { + /* We load immediately after a store -- a read after write. */ + ir_node *mem = get_Load_mem(n); + + c = get_Store_value(a); + turn_into_tuple(n, 3); + set_Tuple_pred(n, pn_Load_M, mem); + set_Tuple_pred(n, pn_Load_res, c); + set_Tuple_pred(n, pn_Load_X_except, new_Bad()); DBG_OPT_RAW; + } + return n; } +/** + * Optimize store after store and load atfter store. + * + * @todo FAILS for volatile entities + */ static ir_node *equivalent_node_Store(ir_node *n) { ir_node *oldn = n; @@ -904,22 +966,28 @@ static ir_node *equivalent_node_Proj(ir_node *n) if ( get_Proj_proj(n) <= get_Tuple_n_preds(a) ) { n = get_Tuple_pred(a, get_Proj_proj(n)); DBG_OPT_TUPLE; } else { - assert(0); /* This should not happen! */ - n = new_Bad(); +// assert(0); /* This should not happen! */ +// n = new_Bad(); + dump_ir_block_graph(current_ir_graph, "-CRASH"); + printf(">>>%d\n", get_irn_node_nr(n)); + exit(1); } } else if (get_irn_mode(n) == mode_X && - is_Bad(get_nodes_Block(n))) { + is_Bad(get_nodes_block(n))) { /* Remove dead control flow -- early gigo. */ n = new_Bad(); } return n; } +/** + * Remove Id's. + */ 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; return n; } @@ -937,7 +1005,7 @@ case iro_Mod, Quot, DivMod * in array fits, we transform n into a tuple (e.g., Div). */ ir_node * -equivalent_node (ir_node *n) +equivalent_node(ir_node *n) { if (n->op->equivalent_node) return n->op->equivalent_node(n); @@ -973,8 +1041,8 @@ static ir_op *firm_set_default_equivalent_node(ir_op *op) CASE(And); CASE(Conv); CASE(Phi); - CASE(Load); - CASE(Store); + CASE(Load); /* dangerous */ + CASE(Store); /* dangerous, see todo */ CASE(Proj); CASE(Id); default: @@ -1020,31 +1088,35 @@ optimize_preds(ir_node *n) { static ir_node *transform_node_Div(ir_node *n) { - tarval *ta = computed_value(n); + tarval *tv = computed_value(n); - if (ta != tarval_bad) { + /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */ + + if (tv != tarval_bad) { /* 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(ta), ta)); + set_Tuple_pred(n, pn_Div_res, new_Const(get_tarval_mode(tv), tv)); } return n; } static ir_node *transform_node_Mod(ir_node *n) { - tarval *ta = computed_value(n); + tarval *tv = computed_value(n); + + /* BEWARE: it is NOT possible to optimize a%a to 0, as this may cause a exception */ - if (ta != tarval_bad) { + if (tv != tarval_bad) { /* 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(ta), ta)); + set_Tuple_pred(n, pn_Mod_res, new_Const(get_tarval_mode(tv), tv)); } return n; } @@ -1056,37 +1128,32 @@ static ir_node *transform_node_DivMod(ir_node *n) ir_node *a = get_DivMod_left(n); ir_node *b = get_DivMod_right(n); ir_mode *mode = get_irn_mode(a); + tarval *ta = value_of(a); + tarval *tb = value_of(b); if (!(mode_is_int(mode) && mode_is_int(get_irn_mode(b)))) return n; - if (a == b) { - a = new_Const(mode, get_mode_one(mode)); - b = new_Const(mode, get_mode_null(mode)); - evaluated = 1; - } else { - tarval *ta = value_of(a); - tarval *tb = value_of(b); - - if (tb != tarval_bad) { - if (tb == get_mode_one(get_tarval_mode(tb))) { - b = new_Const (mode, get_mode_null(mode)); - evaluated = 1; - } else if (ta != tarval_bad) { - tarval *resa, *resb; - resa = tarval_div (ta, tb); - if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through - 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 if (ta == get_mode_null(mode)) { - b = a; + /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */ + + if (tb != tarval_bad) { + if (tb == get_mode_one(get_tarval_mode(tb))) { + b = new_Const (mode, get_mode_null(mode)); + evaluated = 1; + } else if (ta != tarval_bad) { + tarval *resa, *resb; + resa = tarval_div (ta, tb); + if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through + 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 if (ta == get_mode_null(mode)) { + b = a; + evaluated = 1; } if (evaluated) { /* replace by tuple */ ir_node *mem = get_DivMod_mem(n); @@ -1095,7 +1162,7 @@ static ir_node *transform_node_DivMod(ir_node *n) 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, b); - assert(get_nodes_Block(n)); + assert(get_nodes_block(n)); } return n; @@ -1114,7 +1181,7 @@ static ir_node *transform_node_Cond(ir_node *n) (get_opt_unreachable_code())) { /* It's a boolean Cond, branching on a boolean constant. Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */ - jmp = new_r_Jmp(current_ir_graph, get_nodes_Block(n)); + jmp = new_r_Jmp(current_ir_graph, get_nodes_block(n)); turn_into_tuple(n, 2); if (ta == tarval_b_true) { set_Tuple_pred(n, pn_Cond_false, new_Bad()); @@ -1124,7 +1191,7 @@ static ir_node *transform_node_Cond(ir_node *n) set_Tuple_pred(n, pn_Cond_true, new_Bad()); } /* We might generate an endless loop, so keep it alive. */ - add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_Block(n)); + 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) && @@ -1133,21 +1200,21 @@ static ir_node *transform_node_Cond(ir_node *n) 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))); + set_irn_link(n, new_r_Jmp(current_ir_graph, get_nodes_block(n))); /* We might generate an endless loop, so keep it alive. */ - add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_Block(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) - && (tarval_classify(computed_value(get_Eor_right(a))) == TV_CLASSIFY_ONE)) { + && (classify_tarval(computed_value(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), + set_irn_link(n, new_r_Cond(current_ir_graph, get_nodes_block(n), get_Eor_left(a))); } else if ((get_irn_op(a) == op_Not) && (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), + set_irn_link(n, new_r_Cond(current_ir_graph, get_nodes_block(n), get_Not_op(a))); } return n; @@ -1161,20 +1228,23 @@ 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) - && (tarval_classify (computed_value (b)) == TV_CLASSIFY_ONE) + && (classify_tarval (computed_value (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), + 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) - && (tarval_classify (computed_value (b)) == TV_CLASSIFY_ONE)) + && (classify_tarval (computed_value (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); + n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode_b); return n; } +/** + * Transfor a boolean Not. + */ static ir_node *transform_node_Not(ir_node *n) { ir_node *a = get_Not_op(n); @@ -1184,12 +1254,257 @@ static ir_node *transform_node_Not(ir_node *n) && (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), + n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a), mode_b, get_negated_pnc(get_Proj_proj(a))); return n; } +/** + * Transform a Div/Mod/DivMod with a non-zero constant. Must be + * done here to avoid that this optimization runs more than once... + */ +static ir_node *transform_node_Proj(ir_node *proj) +{ + ir_node *n = get_Proj_pred(proj); + ir_node *b; + tarval *tb; + + switch (get_irn_opcode(n)) { + case iro_Div: + b = get_Div_right(n); + tb = computed_value(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(); + + 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); + + proj = new_r_Proj(current_ir_graph, get_nodes_block(n), div, get_irn_mode(a), pn_Div_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); + + 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(); + + 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); + + proj = new_r_Proj(current_ir_graph, get_nodes_block(n), mod, get_irn_mode(a), pn_Mod_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); + + 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); + + 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); + } + break; + + case iro_Cond: + if (get_opt_unreachable_code()) { + b = get_Cond_selector(n); + tb = computed_value(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(); + } + } + } + return proj; + + default: + /* do nothing */ + return proj; + } + + /* we have added a Tuple, optimize it for the current Proj away */ + return equivalent_node_Proj(proj); +} + +/** + * Transform a Store before a Store to the same address... + * Both nodes must be in the same block. + * + * @todo Check for volatile! Moreover, what if the first store + * has a exception handler while the other has not? + */ +static ir_node *transform_node_Store(ir_node *store) +{ + ir_node *pred = skip_Proj(get_Store_mem(store)); + ir_node *ptr = get_Store_ptr(store); + + if (get_irn_op(pred) == op_Store && + get_Store_ptr(pred) == ptr && + get_nodes_block(pred) == get_nodes_block(store)) { + /* the Store n is useless, as it is overwritten by the store store */ + ir_node *mem = get_Store_mem(pred); + + turn_into_tuple(pred, 2); + set_Tuple_pred(pred, pn_Store_M, mem); + set_Tuple_pred(pred, pn_Store_X_except, new_Bad()); + } + return store; +} + +/** + * returns the operands of a commutative bin-op, if one operand is + * a const, it is returned as the second one. + */ +static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) +{ + ir_node *op_a = get_binop_left(binop); + ir_node *op_b = get_binop_right(binop); + + assert(is_op_commutative(get_irn_op(binop))); + + if (get_irn_op(op_a) == op_Const) { + *a = op_b; + *c = op_a; + } + else { + *a = op_a; + *c = op_b; + } +} + +/** + * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible. + * Such pattern may arise in bitfield stores. + * + * value c4 value c4 & c2 + * AND c3 AND c1 | c3 + * OR c2 ===> OR + * AND c1 + * OR + */ +static ir_node *transform_node_Or(ir_node *or) +{ + ir_node *and, *c1; + ir_node *or_l, *c2; + ir_node *and_l, *c3; + ir_node *value, *c4; + ir_node *new_and, *new_const, *block; + ir_mode *mode = get_irn_mode(or); + + tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2; + + get_comm_Binop_Ops(or, &and, &c1); + if ((get_irn_op(c1) != op_Const) || (get_irn_op(and) != op_And)) + return or; + + get_comm_Binop_Ops(and, &or_l, &c2); + if ((get_irn_op(c2) != op_Const) || (get_irn_op(or_l) != op_Or)) + return or; + + get_comm_Binop_Ops(or_l, &and_l, &c3); + if ((get_irn_op(c3) != op_Const) || (get_irn_op(and_l) != op_And)) + return or; + + get_comm_Binop_Ops(and_l, &value, &c4); + if (get_irn_op(c4) != op_Const) + return or; + + /* ok, found the pattern, check for conditions */ + assert(mode == get_irn_mode(and)); + assert(mode == get_irn_mode(or_l)); + assert(mode == get_irn_mode(and_l)); + + tv1 = get_Const_tarval(c1); + tv2 = get_Const_tarval(c2); + tv3 = get_Const_tarval(c3); + tv4 = get_Const_tarval(c4); + + tv = tarval_or(tv4, tv2); + if (classify_tarval(tv) != TV_CLASSIFY_ALL_ONE) { + /* have at least one 0 at the same bit position */ + return or; + } + + n_tv4 = tarval_not(tv4); + if (tv3 != tarval_and(tv3, n_tv4)) { + /* bit in the or_mask is outside the and_mask */ + return or; + } + + n_tv2 = tarval_not(tv2); + if (tv1 != tarval_and(tv1, n_tv2)) { + /* bit in the or_mask is outside the and_mask */ + return or; + } + + /* ok, all conditions met */ + block = get_nodes_block(or); + + new_and = new_r_And(current_ir_graph, block, + value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode); + + new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1)); + + set_Or_left(or, new_and); + set_Or_right(or, new_const); + + /* check for more */ + return transform_node_Or(or); +} /** * Tries several [inplace] [optimizing] transformations and returns an @@ -1221,6 +1536,9 @@ static ir_op *firm_set_default_transform_node(ir_op *op) CASE(Cond); CASE(Eor); CASE(Not); + CASE(Proj); + CASE(Store); /* dangerous, see todo */ + CASE(Or); default: op->transform_node = NULL; } @@ -1266,7 +1584,7 @@ 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).tori.typ != get_irn_symconst_attr(b).tori.typ); + || (get_irn_symconst_attr(a).sym.type_p != get_irn_symconst_attr(b).sym.type_p); } static int node_cmp_attr_Call(ir_node *a, ir_node *b) @@ -1351,8 +1669,8 @@ vt_cmp (const void *elt, const void *key) if (irn_arity_a != get_irn_arity(b)) return 1; - /* for block-local cse and pinned nodes: */ - if (!get_opt_global_cse() || (get_op_pinned(get_irn_op(a)) == pinned)) { + /* 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)) return 1; } @@ -1372,27 +1690,38 @@ vt_cmp (const void *elt, const void *key) return 0; } -/** +/* * Calculate a hash value of a node. */ -static unsigned +unsigned ir_node_hash (ir_node *node) { unsigned h; int i, irn_arity; - /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */ - h = irn_arity = get_irn_arity(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); + } 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); + } else { - /* consider all in nodes... except the block. */ - for (i = 0; i < irn_arity; i++) { - h = 9*h + (unsigned long)get_irn_n(node, i); - } + /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */ + h = irn_arity = get_irn_arity(node); - /* ...mode,... */ - h = 9*h + (unsigned long) get_irn_mode (node); - /* ...and code */ - h = 9*h + (unsigned long) get_irn_op (node); + /* consider all in nodes... except the block. */ + for (i = 0; i < irn_arity; i++) { + h = 9*h + (unsigned)get_irn_n(node, i); + } + + /* ...mode,... */ + h = 9*h + (unsigned) get_irn_mode (node); + /* ...and code */ + h = 9*h + (unsigned) get_irn_op (node); + } return h; } @@ -1412,6 +1741,10 @@ del_identities (pset *value_table) /** * Return the canonical node computing the same value as n. * Looks up the node in a hash table. + * + * For Const nodes this is performed in the constructor, too. Const + * nodes are extremely time critical because of their frequent use in + * constant string arrays. */ static INLINE ir_node * identify (pset *value_table, ir_node *n) @@ -1439,8 +1772,8 @@ identify (pset *value_table, ir_node *n) } /** - * During construction we set the pinned flag in the graph right when the - * optimizatin is performed. The flag turning on procedure global cse could + * During construction we set the op_pin_state_pinned flag in the graph right when the + * optimization is performed. The flag turning on procedure global cse could * be changed between two allocations. This way we are safe. */ static INLINE ir_node * @@ -1448,7 +1781,7 @@ 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); + set_irg_pinned(current_ir_graph, op_pin_state_floats); return n; } @@ -1554,8 +1887,12 @@ optimize_node (ir_node *n) * we MUST copy the node here temparary, because it's still needed * for DBG_OPT_ALGSIM0 */ - ir_node x = *n; - oldn = &x; + int node_size = offsetof(ir_node, attr) + n->op->attr_size; + ir_node *x = alloca(node_size); + + memcpy(x, n, node_size); + oldn = x; + /* evaluation was successful -- replace the node. */ obstack_free (current_ir_graph->obst, n); n = new_Const (get_tarval_mode (tv), tv); @@ -1578,7 +1915,7 @@ optimize_node (ir_node *n) /** common subexpression elimination **/ /* Checks whether n is already available. */ /* The block input is used to distinguish different subexpressions. Right - now all nodes are pinned to blocks, i.e., the cse only finds common + now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common subexpressions within a block. */ if (get_opt_cse()) n = identify_cons (current_ir_graph->value_table, n); @@ -1662,7 +1999,7 @@ optimize_in_place_2 (ir_node *n) /** common subexpression elimination **/ /* Checks whether n is already available. */ /* The block input is used to distinguish different subexpressions. Right - now all nodes are pinned to blocks, i.e., the cse only finds common + now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common subexpressions within a block. */ if (get_opt_cse()) { n = identify (current_ir_graph->value_table, n); @@ -1701,7 +2038,7 @@ optimize_in_place (ir_node *n) assert(get_irg_phase_state(current_ir_graph) != phase_building); if (get_opt_global_cse()) - set_irg_pinned(current_ir_graph, floats); + set_irg_pinned(current_ir_graph, op_pin_state_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