X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firopt.c;h=0c6d0490ad907d5d3e6266247b69f04f72e16e56;hb=23ea16875f38dc0c067cce24f64b5c70f3acc496;hp=cb0fca8c0213d7bb5d08e122c5bb00ffd05e19ad;hpb=cb8a909742e7ceb33f7d8ce0c36b424077d3fe59;p=libfirm diff --git a/ir/ir/iropt.c b/ir/ir/iropt.c index cb0fca8c0..0c6d0490a 100644 --- a/ir/ir/iropt.c +++ b/ir/ir/iropt.c @@ -24,24 +24,25 @@ #include #endif -# include "irnode_t.h" -# include "irgraph_t.h" -# include "iredges_t.h" -# include "irmode_t.h" -# include "iropt_t.h" -# include "ircons_t.h" -# include "irgmod.h" -# include "irvrfy.h" -# include "tv_t.h" -# include "dbginfo_t.h" -# include "iropt_dbg.h" -# include "irflag_t.h" -# include "irhooks.h" -# include "irarch.h" -# include "hashptr.h" -# include "archop.h" -# include "opt_polymorphy.h" -# include "opt_confirms.h" +#include "irnode_t.h" +#include "irgraph_t.h" +#include "iredges_t.h" +#include "irmode_t.h" +#include "iropt_t.h" +#include "ircons_t.h" +#include "irgmod.h" +#include "irvrfy.h" +#include "tv_t.h" +#include "dbginfo_t.h" +#include "iropt_dbg.h" +#include "irflag_t.h" +#include "irhooks.h" +#include "irarch.h" +#include "hashptr.h" +#include "archop.h" +#include "opt_polymorphy.h" +#include "opt_confirms.h" +#include "irtools.h" /* Make types visible to allow most efficient access */ # include "entity_t.h" @@ -59,9 +60,22 @@ static tarval *computed_value_Const(ir_node *n) */ static tarval *computed_value_SymConst(ir_node *n) { - if ((get_SymConst_kind(n) == symconst_size) && - (get_type_state(get_SymConst_type(n))) == layout_fixed) - return new_tarval_from_long(get_type_size_bytes(get_SymConst_type(n)), get_irn_mode(n)); + ir_type *type; + + switch (get_SymConst_kind(n)) { + case symconst_type_size: + type = get_SymConst_type(n); + if (get_type_state(type) == layout_fixed) + return new_tarval_from_long(get_type_size_bytes(type), get_irn_mode(n)); + break; + case symconst_type_align: + type = get_SymConst_type(n); + if (get_type_state(type) == layout_fixed) + return new_tarval_from_long(get_type_alignment_bytes(type), get_irn_mode(n)); + break; + default: + break; + } return tarval_bad; } @@ -106,6 +120,51 @@ static tarval *computed_value_Sub(ir_node *n) return tarval_bad; } +/** + * return the value of a Carry + * Special : a op 0, 0 op b + */ +static tarval *computed_value_Carry(ir_node *n) +{ + ir_node *a = get_binop_left(n); + ir_node *b = get_binop_right(n); + ir_mode *m = get_irn_mode(n); + + tarval *ta = value_of(a); + tarval *tb = value_of(b); + + if ((ta != tarval_bad) && (tb != tarval_bad)) { + tarval_add(ta, tb); + return tarval_carry() ? get_mode_one(m) : get_mode_null(m); + } else { + if ( (classify_tarval(ta) == TV_CLASSIFY_NULL) + || (classify_tarval(tb) == TV_CLASSIFY_NULL)) + return get_mode_null(m); + } + return tarval_bad; +} + +/** + * return the value of a Borrow + * Special : a op 0 + */ +static tarval *computed_value_Borrow(ir_node *n) +{ + ir_node *a = get_binop_left(n); + ir_node *b = get_binop_right(n); + ir_mode *m = get_irn_mode(n); + + tarval *ta = value_of(a); + tarval *tb = value_of(b); + + if ((ta != tarval_bad) && (tb != tarval_bad)) { + return tarval_cmp(ta, tb) == pn_Cmp_Lt ? get_mode_one(m) : get_mode_null(m); + } else if (classify_tarval(ta) == TV_CLASSIFY_NULL) { + return get_mode_null(m); + } + return tarval_bad; +} + /** * return the value of an unary Minus */ @@ -512,7 +571,6 @@ static tarval *computed_value_Proj_Cmp(ir_node *n) return new_tarval_from_long(proj_nr & pn_Cmp_Ne, mode_b); } } - return computed_value_Cmp_Confirm(a, aa, ab, proj_nr); } @@ -573,6 +631,18 @@ static tarval *computed_value_Mux(ir_node *n) return tarval_bad; } +/** + * Calculate the value of a Psi: can be evaluated, if a condition is true + * and all previous conditions are false. If all conditions are false + * we evaluate to the default one. + */ +static tarval *computed_value_Psi(ir_node *n) +{ + if (is_Mux(n)) + return computed_value_Mux(n); + return tarval_bad; +} + /** * calculate the value of a Confirm: can be evaluated, * if it has the form Confirm(x, '=', Const). @@ -591,22 +661,28 @@ static tarval *computed_value_Confirm(ir_node *n) */ tarval *computed_value(ir_node *n) { - if (n->op->computed_value) - return n->op->computed_value(n); + if (n->op->ops.computed_value) + return n->op->ops.computed_value(n); return tarval_bad; } /** - * set the default computed_value evaluator + * set the default computed_value evaluator in an ir_op_ops. + * + * @param code the opcode for the default operation + * @param ops the operations initialized + * + * @return + * The operations. */ -static ir_op *firm_set_default_computed_value(ir_op *op) +static ir_op_ops *firm_set_default_computed_value(opcode code, ir_op_ops *ops) { #define CASE(a) \ case iro_##a: \ - op->computed_value = computed_value_##a; \ + ops->computed_value = computed_value_##a; \ break - switch (op->code) { + switch (code) { CASE(Const); CASE(SymConst); CASE(Add); @@ -625,37 +701,21 @@ static ir_op *firm_set_default_computed_value(ir_op *op) CASE(Shr); CASE(Shrs); CASE(Rot); + CASE(Carry); + CASE(Borrow); CASE(Conv); CASE(Proj); CASE(Mux); + CASE(Psi); CASE(Confirm); default: - op->computed_value = NULL; + /* leave NULL */; } - return op; + return ops; #undef CASE } -#if 0 -/* returns 1 if the a and b are pointers to different locations. */ -static bool -different_identity (ir_node *a, ir_node *b) -{ - assert (mode_is_reference(get_irn_mode (a)) - && mode_is_reference(get_irn_mode (b))); - - if (get_irn_op (a) == op_Proj && get_irn_op(b) == op_Proj) { - 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) - return 1; - } - return 0; -} -#endif - /** * Returns a equivalent block for another block. * If the block has only one predecessor, this is @@ -723,19 +783,19 @@ 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); + n = get_nodes_block(get_Proj_pred(a)); DBG_OPT_IFSIM1(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_cfg = get_Block_n_cfgpreds(n); + (n != get_irg_start_block(current_ir_graph)) && + (n != get_irg_end_block(current_ir_graph)) ) { + int i; /* 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 + the start or end block. This is one step of unreachable code elimination */ - for (i = 0; i < n_cfg; i++) { + for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) { ir_node *pred = get_Block_cfgpred(n, i); ir_node *pred_blk; @@ -749,8 +809,10 @@ static ir_node *equivalent_node_Block(ir_node *n) break; } } - if (i == n_cfg) + if (i < 0) { n = set_Block_dead(n); + DBG_OPT_DEAD_BLOCK(oldn, n); + } } return n; @@ -758,11 +820,10 @@ static ir_node *equivalent_node_Block(ir_node *n) /** * Returns a equivalent node for a Jmp, a Bad :-) - * Of course this only happens if the Block of the Jmp is Bad. + * Of course this only happens if the Block of the Jmp is dead. */ static ir_node *equivalent_node_Jmp(ir_node *n) { - /* GL: Why not same for op_Raise?? */ /* unreachable code elimination */ if (is_Block_dead(get_nodes_block(n))) n = new_Bad(); @@ -770,6 +831,10 @@ static ir_node *equivalent_node_Jmp(ir_node *n) return n; } +/** Raise is handled in the same way as Jmp. */ +#define equivalent_node_Raise equivalent_node_Jmp + + /* We do not evaluate Cond here as we replace it by a new node, a Jmp. See transform_node_Proj_Cond(). */ @@ -814,6 +879,9 @@ static ir_node *equivalent_node_neutral_zero(ir_node *n) return n; } +/** + * Eor is commutative and has neutral 0. + */ #define equivalent_node_Eor equivalent_node_neutral_zero /* @@ -829,6 +897,11 @@ static ir_node *equivalent_node_Add(ir_node *n) { ir_node *oldn = n; ir_node *left, *right; + ir_mode *mode = get_irn_mode(n); + + /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */ + if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic)) + return n; n = equivalent_node_neutral_zero(n); if (n != oldn) @@ -842,7 +915,7 @@ static ir_node *equivalent_node_Add(ir_node *n) /* (a - x) + x */ n = get_Sub_left(left); - if (get_irn_mode(oldn) == get_irn_mode(n)) { + if (mode == get_irn_mode(n)) { DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB); return n; } @@ -853,7 +926,7 @@ static ir_node *equivalent_node_Add(ir_node *n) /* x + (a - x) */ n = get_Sub_left(right); - if (get_irn_mode(oldn) == get_irn_mode(n)) { + if (mode == get_irn_mode(n)) { DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB); return n; } @@ -899,40 +972,43 @@ static ir_node *equivalent_node_left_zero(ir_node *n) static ir_node *equivalent_node_Sub(ir_node *n) { ir_node *oldn = n; + ir_node *a, *b; + ir_mode *mode = get_irn_mode(n); - ir_node *a = get_Sub_left(n); - ir_node *b = get_Sub_right(n); + /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */ + if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic)) + return n; + + a = get_Sub_left(n); + b = get_Sub_right(n); /* Beware: modes might be different */ if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) { - if (get_irn_mode(n) == get_irn_mode(a)) { + if (mode == get_irn_mode(a)) { n = a; DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0); } } else if (get_irn_op(a) == op_Add) { - ir_mode *mode = get_irn_mode(n); - if (mode_wrap_around(mode)) { ir_node *left = get_Add_left(a); ir_node *right = get_Add_right(a); if (left == b) { - if (get_irn_mode(n) == get_irn_mode(right)) { + if (mode == get_irn_mode(right)) { n = right; DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); } } else if (right == b) { - if (get_irn_mode(n) == get_irn_mode(left)) { + if (mode == get_irn_mode(left)) { n = left; DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); } } } } - return n; } @@ -940,9 +1016,10 @@ static ir_node *equivalent_node_Sub(ir_node *n) /** * Optimize an "idempotent unary op", ie op(op(n)) = n. * - * @fixme -(-a) == a, but might overflow two times. - * We handle it anyway here but the better way would be a - * flag. This would be needed for Pascal for instance. + * @todo + * -(-a) == a, but might overflow two times. + * We handle it anyway here but the better way would be a + * flag. This would be needed for Pascal for instance. */ static ir_node *equivalent_node_idempotent_unop(ir_node *n) { @@ -957,10 +1034,10 @@ static ir_node *equivalent_node_idempotent_unop(ir_node *n) return n; } -/* Not(Not(x)) == x */ +/** Not(Not(x)) == x */ #define equivalent_node_Not equivalent_node_idempotent_unop -/* --x == x */ /* ??? Is this possible or can --x raise an +/** --x == x ??? Is this possible or can --x raise an out of bounds exception if min =! max? */ #define equivalent_node_Minus equivalent_node_idempotent_unop @@ -970,7 +1047,6 @@ static ir_node *equivalent_node_idempotent_unop(ir_node *n) static ir_node *equivalent_node_Mul(ir_node *n) { ir_node *oldn = n; - ir_node *a = get_Mul_left(n); ir_node *b = get_Mul_right(n); @@ -997,7 +1073,7 @@ static ir_node *equivalent_node_Div(ir_node *n) 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); + turn_into_tuple(n, pn_Div_max); 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_res, a); @@ -1005,6 +1081,25 @@ static ir_node *equivalent_node_Div(ir_node *n) return n; } +/** + * Optimize a / 1.0 = a. + */ +static ir_node *equivalent_node_Quot(ir_node *n) { + ir_node *a = get_Quot_left(n); + ir_node *b = get_Quot_right(n); + + /* Div is not commutative. */ + if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* Quot(x, 1) == x */ + /* Turn Quot into a tuple (mem, bad, a) */ + ir_node *mem = get_Quot_mem(n); + turn_into_tuple(n, pn_Quot_max); + set_Tuple_pred(n, pn_Quot_M, mem); + set_Tuple_pred(n, pn_Quot_X_except, new_Bad()); /* no exception */ + set_Tuple_pred(n, pn_Quot_res, a); + } + return n; +} + /** * Optimize a / 1 = a. */ @@ -1019,7 +1114,7 @@ static ir_node *equivalent_node_DivMod(ir_node *n) ir_node *mem = get_Div_mem(n); ir_mode *mode = get_irn_mode(b); - turn_into_tuple(n, 4); + turn_into_tuple(n, pn_DivMod_max); 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); @@ -1088,6 +1183,9 @@ 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 */ + /* leave strict floating point Conv's */ + if (get_Conv_strict(n)) + return n; n = a; DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV); } else if (get_irn_op(a) == op_Conv) { /* Conv(Conv(b)) */ @@ -1128,12 +1226,13 @@ static ir_node *equivalent_node_Cast(ir_node *n) { return n; } -/* Several optimizations: +/** + 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) { int i, n_preds; @@ -1141,7 +1240,6 @@ static ir_node *equivalent_node_Phi(ir_node *n) ir_node *oldn = n; ir_node *block = NULL; /* to shutup gcc */ ir_node *first_val = NULL; /* to shutup gcc */ - ir_node *scnd_val = NULL; /* to shutup gcc */ if (!get_opt_normalize()) return n; @@ -1151,8 +1249,8 @@ static ir_node *equivalent_node_Phi(ir_node *n) /* @@@ fliegt 'raus, sollte aber doch immer wahr sein!!! assert(get_irn_arity(block) == n_preds && "phi in wrong block!"); */ 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. */ + (block == get_irg_start_block(current_ir_graph))) /* There should be no Phi nodes */ + return new_Bad(); /* in the Start Block. */ if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */ @@ -1178,12 +1276,10 @@ static ir_node *equivalent_node_Phi(ir_node *n) return new_Bad(); } - scnd_val = NULL; - - /* follow_Id () for rest of inputs, determine if any of these + /* search for rest of inputs, determine if any of these are non-self-referencing */ while (++i < n_preds) { - scnd_val = get_Phi_pred(n, i); + ir_node *scnd_val = get_Phi_pred(n, i); if ( (scnd_val != n) && (scnd_val != first_val) #if 1 @@ -1198,16 +1294,63 @@ static ir_node *equivalent_node_Phi(ir_node *n) /* Fold, if no multiple distinct non-self-referencing inputs */ n = first_val; DBG_OPT_PHI(oldn, n); - } else { - /* 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 Proj(Tuple) and gigo() for ProjX in Bad block + Several optimizations: + - no Sync in start block. + - fold Sync-nodes, iff they have only one predecessor except + themselves. + */ +static ir_node *equivalent_node_Sync(ir_node *n) +{ + int i, n_preds; + + ir_node *oldn = n; + ir_node *first_val = NULL; /* to shutup gcc */ + + if (!get_opt_normalize()) return n; + + n_preds = get_Sync_n_preds(n); + + /* Find first non-self-referencing input */ + for (i = 0; i < n_preds; ++i) { + first_val = get_Sync_pred(n, i); + if ((first_val != n) /* not self pointer */ && + (! is_Bad(first_val)) + ) { /* value not dead */ + break; /* then found first value. */ + } + } + + if (i >= n_preds) + /* A totally Bad or self-referencing Sync (we didn't break the above loop) */ + return new_Bad(); + + /* search the rest of inputs, determine if any of these + are non-self-referencing */ + while (++i < n_preds) { + ir_node *scnd_val = get_Sync_pred(n, i); + if ((scnd_val != n) && + (scnd_val != first_val) && + (! is_Bad(scnd_val)) + ) + break; + } + + if (i >= n_preds) { + /* Fold, if no multiple distinct non-self-referencing inputs */ + n = first_val; + DBG_OPT_SYNC(oldn, n); + } + return n; +} + +/** + * optimize Proj(Tuple) and gigo() for ProjX in Bad block, + * ProjX(Load) and ProjX(Store) */ static ir_node *equivalent_node_Proj(ir_node *n) { @@ -1224,12 +1367,28 @@ static ir_node *equivalent_node_Proj(ir_node *n) assert(0); /* This should not happen! */ n = new_Bad(); } - } else if (get_irn_mode(n) == mode_X) { + } + else if (get_irn_mode(n) == mode_X) { if (is_Block_dead(get_nodes_block(skip_Proj(n)))) { /* Remove dead control flow -- early gigo(). */ n = new_Bad(); } + else if (get_opt_ldst_only_null_ptr_exceptions()) { + ir_op *op = get_irn_op(a); + + if (op == op_Load || op == op_Store) { + /* get the load/store address */ + ir_node *addr = get_irn_n(a, 1); + if (value_not_null(addr)) { + /* this node may float if it did not depend on a Confirm */ + set_irn_pinned(a, op_pin_state_floats); + DBG_OPT_EXC_REM(n); + return new_Bad(); + } + } + } } + return n; } @@ -1317,7 +1476,17 @@ static ir_node *equivalent_node_Mux(ir_node *n) } } } + return n; +} +/** + * Returns a equivalent node of a Psi: if a condition is true + * and all previous conditions are false we know its value. + * If all conditions are false its value is the default one. + */ +static ir_node *equivalent_node_Psi(ir_node *n) { + if (is_Mux(n)) + return equivalent_node_Mux(n); return n; } @@ -1345,21 +1514,99 @@ static ir_node *equivalent_node_Cmp(ir_node *n) /** * Remove Confirm nodes if setting is on. + * Replace Confirms(x, '=', Constlike) by Constlike. */ static ir_node *equivalent_node_Confirm(ir_node *n) { - if (get_Confirm_cmp(n) == pn_Cmp_Eq) { + ir_node *pred = get_Confirm_value(n); + pn_Cmp pnc = get_Confirm_cmp(n); + + if (get_irn_op(pred) == op_Confirm && pnc == get_Confirm_cmp(pred)) { + /* + * rare case: two identical Confirms one after another, + * replace the second one with the first. + */ + n = pred; + } + if (pnc == pn_Cmp_Eq) { ir_node *bound = get_Confirm_bound(n); - ir_op *op = get_irn_op(bound); /* * Optimize a rare case: - * Confirm(x, '=', Const) ==> Const + * Confirm(x, '=', Constlike) ==> Constlike */ - if (op == op_Const || op == op_SymConst) + if (is_irn_constlike(bound)) { + DBG_OPT_CONFIRM(n, bound); return bound; + } + } + return get_opt_remove_confirm() ? get_Confirm_value(n) : n; +} + +/** + * Optimize CopyB(mem, x, x) into a Nop + */ +static ir_node *equivalent_node_CopyB(ir_node *n) +{ + ir_node *a = get_CopyB_dst(n); + ir_node *b = get_CopyB_src(n); + + if (a == b) { + /* Turn CopyB into a tuple (mem, bad, bad) */ + ir_node *mem = get_CopyB_mem(n); + turn_into_tuple(n, pn_CopyB_max); + set_Tuple_pred(n, pn_CopyB_M, mem); + set_Tuple_pred(n, pn_CopyB_X_except, new_Bad()); /* no exception */ + set_Tuple_pred(n, pn_CopyB_M_except, new_Bad()); } - return get_opt_remove_Confirm() ? get_Confirm_value(n) : n; + return n; +} + +/** + * Optimize Bounds(idx, idx, upper) into idx. + */ +static ir_node *equivalent_node_Bound(ir_node *n) +{ + ir_node *idx = get_Bound_index(n); + ir_node *lower = get_Bound_lower(n); + int ret_tuple = 0; + + /* By definition lower < upper, so if idx == lower --> + lower <= idx && idx < upper */ + if (idx == lower) { + /* Turn Bound into a tuple (mem, bad, idx) */ + ret_tuple = 1; + } + else { + ir_node *pred = skip_Proj(idx); + + if (get_irn_op(pred) == op_Bound) { + /* + * idx was Bounds_check previously, it is still valid if + * lower <= pred_lower && pred_upper <= upper. + */ + ir_node *upper = get_Bound_upper(n); + if (get_Bound_lower(pred) == lower && + get_Bound_upper(pred) == upper) { + /* + * One could expect that we simply return the previous + * Bound here. However, this would be wrong, as we could + * add an exception Proj to a new location than. + * So, we must turn in into a tuple + */ + ret_tuple = 1; + } + } + } + if (ret_tuple) { + /* Turn Bound into a tuple (mem, bad, idx) */ + ir_node *mem = get_Bound_mem(n); + turn_into_tuple(n, pn_Bound_max); + set_Tuple_pred(n, pn_Bound_M, mem); + set_Tuple_pred(n, pn_Bound_X_except, new_Bad()); /* no exception */ + set_Tuple_pred(n, pn_Bound_res, idx); + } + return n; } /** @@ -1372,24 +1619,31 @@ static ir_node *equivalent_node_Confirm(ir_node *n) ir_node * equivalent_node(ir_node *n) { - if (n->op->equivalent_node) - return n->op->equivalent_node(n); + if (n->op->ops.equivalent_node) + return n->op->ops.equivalent_node(n); return n; } /** - * set the default equivalent node operation + * sets the default equivalent node operation for an ir_op_ops. + * + * @param code the opcode for the default operation + * @param ops the operations initialized + * + * @return + * The operations. */ -static ir_op *firm_set_default_equivalent_node(ir_op *op) +static ir_op_ops *firm_set_default_equivalent_node(opcode code, ir_op_ops *ops) { #define CASE(a) \ case iro_##a: \ - op->equivalent_node = equivalent_node_##a; \ + ops->equivalent_node = equivalent_node_##a; \ break - switch (op->code) { + switch (code) { CASE(Block); CASE(Jmp); + CASE(Raise); CASE(Or); CASE(Add); CASE(Eor); @@ -1402,21 +1656,26 @@ static ir_op *firm_set_default_equivalent_node(ir_op *op) CASE(Minus); CASE(Mul); CASE(Div); + CASE(Quot); CASE(DivMod); CASE(And); CASE(Conv); CASE(Cast); CASE(Phi); + CASE(Sync); CASE(Proj); CASE(Id); CASE(Mux); + CASE(Psi); CASE(Cmp); CASE(Confirm); + CASE(CopyB); + CASE(Bound); default: - op->equivalent_node = NULL; + /* leave NULL */; } - return op; + return ops; #undef CASE } @@ -1453,6 +1712,114 @@ optimize_preds(ir_node *n) { } /* end switch */ } +/** + * Returns non-zero if a node is a Phi node + * with all predecessors constant. + */ +static int is_const_Phi(ir_node *n) { + int i; + + if (! is_Phi(n)) + return 0; + for (i = get_irn_arity(n) - 1; i >= 0; --i) + if (! is_Const(get_irn_n(n, i))) + return 0; + return 1; +} + +/** + * Apply an evaluator on a binop with a constant operators (and one Phi). + * + * @param phi the Phi node + * @param other the other operand + * @param eval an evaluator function + * @param left if non-zero, other is the left operand, else the right + * + * @return a new Phi node if the conversion was successful, NULL else + */ +static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(tarval *, tarval *), int left) { + tarval *tv; + void **res; + ir_node *pred; + ir_mode *mode; + ir_graph *irg; + int i, n = get_irn_arity(phi); + + NEW_ARR_A(void *, res, n); + if (left) { + for (i = 0; i < n; ++i) { + pred = get_irn_n(phi, i); + tv = get_Const_tarval(pred); + tv = eval(other, tv); + + if (tv == tarval_bad) { + /* folding failed, bad */ + return NULL; + } + res[i] = tv; + } + } + else { + for (i = 0; i < n; ++i) { + pred = get_irn_n(phi, i); + tv = get_Const_tarval(pred); + tv = eval(tv, other); + + if (tv == tarval_bad) { + /* folding failed, bad */ + return 0; + } + res[i] = tv; + } + } + mode = get_irn_mode(phi); + irg = current_ir_graph; + for (i = 0; i < n; ++i) { + pred = get_irn_n(phi, i); + res[i] = new_r_Const_type(irg, get_nodes_block(pred), + mode, res[i], get_Const_type(pred)); + } + return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode); +} + +/** + * Apply an evaluator on a unop with a constant operator (a Phi). + * + * @param phi the Phi node + * @param eval an evaluator function + * + * @return a new Phi node if the conversion was successful, NULL else + */ +static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) { + tarval *tv; + void **res; + ir_node *pred; + ir_mode *mode; + ir_graph *irg; + int i, n = get_irn_arity(phi); + + NEW_ARR_A(void *, res, n); + for (i = 0; i < n; ++i) { + pred = get_irn_n(phi, i); + tv = get_Const_tarval(pred); + tv = eval(tv); + + if (tv == tarval_bad) { + /* folding failed, bad */ + return 0; + } + res[i] = tv; + } + mode = get_irn_mode(phi); + irg = current_ir_graph; + for (i = 0; i < n; ++i) { + pred = get_irn_n(phi, i); + res[i] = new_r_Const_type(irg, get_nodes_block(pred), + mode, res[i], get_Const_type(pred)); + } + return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode); +} + /** * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)). @@ -1516,8 +1883,38 @@ static ir_node *transform_node_AddSub(ir_node *n) return n; } +#define HANDLE_BINOP_PHI(op,a,b,c) \ + c = NULL; \ + if (is_Const(b) && is_const_Phi(a)) { \ + /* check for Op(Phi, Const) */ \ + c = apply_binop_on_phi(a, get_Const_tarval(b), op, 0); \ + } \ + else if (is_Const(a) && is_const_Phi(b)) { \ + /* check for Op(Const, Phi) */ \ + c = apply_binop_on_phi(b, get_Const_tarval(a), op, 1); \ + } \ + if (c) { \ + DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \ + return c; \ + } + +#define HANDLE_UNOP_PHI(op,a,c) \ + c = NULL; \ + if (is_const_Phi(a)) { \ + /* check for Op(Phi) */ \ + c = apply_unop_on_phi(a, op); \ + } \ + if (c) { \ + DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \ + return c; \ + } + + /** - * Do the AddSub optimization, then Transform Add(a,a) into Mul(a, 2) + * Do the AddSub optimization, then Transform + * Constant folding on Phi + * Add(a,a) -> Mul(a, 2) + * Add(Mul(a, x), a) -> Mul(a, x+1) * if the mode is integer or float. * Transform Add(a,-b) into Sub(a,b). * Reassociation might fold this further. @@ -1525,15 +1922,23 @@ static ir_node *transform_node_AddSub(ir_node *n) static ir_node *transform_node_Add(ir_node *n) { ir_mode *mode; - ir_node *oldn = n; + ir_node *a, *b, *c, *oldn = n; n = transform_node_AddSub(n); + a = get_Add_left(n); + b = get_Add_right(n); + + HANDLE_BINOP_PHI(tarval_add, a,b,c); + mode = get_irn_mode(n); - if (mode_is_num(mode)) { - ir_node *a = get_Add_left(n); - if (a == get_Add_right(n)) { + /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */ + if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic)) + return n; + + if (mode_is_num(mode)) { + if (a == b) { ir_node *block = get_irn_n(n, -1); n = new_rd_Mul( @@ -1545,28 +1950,88 @@ static ir_node *transform_node_Add(ir_node *n) mode); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A); } - else { - ir_node *b = get_Add_right(n); - - if (get_irn_op(a) == op_Minus) { - n = new_rd_Sub( - get_irn_dbg_info(n), - current_ir_graph, - get_irn_n(n, -1), - b, - get_Minus_op(a), - mode); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B); + else if (get_irn_op(a) == op_Minus) { + n = new_rd_Sub( + get_irn_dbg_info(n), + current_ir_graph, + get_irn_n(n, -1), + b, + get_Minus_op(a), + mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B); + } + else if (get_irn_op(b) == op_Minus) { + n = new_rd_Sub( + get_irn_dbg_info(n), + current_ir_graph, + get_irn_n(n, -1), + a, + get_Minus_op(b), + mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B); + } + /* do NOT execute this code if reassociation is enabled, it does the inverse! */ + else if (!get_opt_reassociation() && get_irn_op(a) == op_Mul) { + ir_node *ma = get_Mul_left(a); + ir_node *mb = get_Mul_right(a); + + if (b == ma) { + ir_node *blk = get_irn_n(n, -1); + n = new_rd_Mul( + get_irn_dbg_info(n), current_ir_graph, blk, + ma, + new_rd_Add( + get_irn_dbg_info(n), current_ir_graph, blk, + mb, + new_r_Const_long(current_ir_graph, blk, mode, 1), + mode), + mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A); } - else if (get_irn_op(b) == op_Minus) { - n = new_rd_Sub( - get_irn_dbg_info(n), - current_ir_graph, - get_irn_n(n, -1), - a, - get_Minus_op(b), - mode); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B); + else if (b == mb) { + ir_node *blk = get_irn_n(n, -1); + n = new_rd_Mul( + get_irn_dbg_info(n), current_ir_graph, blk, + mb, + new_rd_Add( + get_irn_dbg_info(n), current_ir_graph, blk, + ma, + new_r_Const_long(current_ir_graph, blk, mode, 1), + mode), + mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A); + } + } + /* do NOT execute this code if reassociation is enabled, it does the inverse! */ + else if (!get_opt_reassociation() && get_irn_op(b) == op_Mul) { + ir_node *ma = get_Mul_left(b); + ir_node *mb = get_Mul_right(b); + + if (a == ma) { + ir_node *blk = get_irn_n(n, -1); + n = new_rd_Mul( + get_irn_dbg_info(n), current_ir_graph, blk, + ma, + new_rd_Add( + get_irn_dbg_info(n), current_ir_graph, blk, + mb, + new_r_Const_long(current_ir_graph, blk, mode, 1), + mode), + mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A); + } + else if (a == mb) { + ir_node *blk = get_irn_n(n, -1); + n = new_rd_Mul( + get_irn_dbg_info(n), current_ir_graph, blk, + mb, + new_rd_Add( + get_irn_dbg_info(n), current_ir_graph, blk, + ma, + new_r_Const_long(current_ir_graph, blk, mode, 1), + mode), + mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A); } } } @@ -1574,41 +2039,127 @@ static ir_node *transform_node_Add(ir_node *n) } /** - * Do the AddSub optimization, then Transform Sub(0,a) into Minus(a). + * Do the AddSub optimization, then Transform + * Constant folding on Phi + * Sub(0,a) -> Minus(a) + * Sub(Mul(a, x), a) -> Mul(a, x-1) + * Sub(Sub(x, y), b) -> Sub(x, Add(y,b)) */ static ir_node *transform_node_Sub(ir_node *n) { ir_mode *mode; ir_node *oldn = n; + ir_node *a, *b, *c; n = transform_node_AddSub(n); + a = get_Sub_left(n); + b = get_Sub_right(n); + + HANDLE_BINOP_PHI(tarval_sub, a,b,c); + mode = get_irn_mode(n); - if (mode_is_num(mode) && (classify_Const(get_Sub_left(n)) == CNST_NULL)) { + + /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */ + if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic)) + return n; + + if (mode_is_num(mode) && (classify_Const(a) == CNST_NULL)) { n = new_rd_Minus( get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), - get_Sub_right(n), + b, mode); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A); } + /* do NOT execute this code if reassociation is enabled, it does the inverse! */ + else if (get_opt_reassociation() && get_irn_op(a) == op_Mul) { + ir_node *ma = get_Mul_left(a); + ir_node *mb = get_Mul_right(a); + + if (ma == b) { + ir_node *blk = get_irn_n(n, -1); + n = new_rd_Mul( + get_irn_dbg_info(n), + current_ir_graph, blk, + ma, + new_rd_Sub( + get_irn_dbg_info(n), + current_ir_graph, blk, + mb, + new_r_Const_long(current_ir_graph, blk, mode, 1), + mode), + mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A); + } + else if (mb == b) { + ir_node *blk = get_irn_n(n, -1); + n = new_rd_Mul( + get_irn_dbg_info(n), + current_ir_graph, blk, + mb, + new_rd_Sub( + get_irn_dbg_info(n), + current_ir_graph, blk, + ma, + new_r_Const_long(current_ir_graph, blk, mode, 1), + mode), + mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A); + } + } + else if (get_irn_op(a) == op_Sub) { + ir_node *x = get_Sub_left(a); + ir_node *y = get_Sub_right(a); + ir_node *blk = get_irn_n(n, -1); + ir_mode *m_b = get_irn_mode(b); + ir_mode *m_y = get_irn_mode(y); + ir_node *add; + + /* Determine the right mode for the Add. */ + if (m_b == m_y) + mode = m_b; + else if (mode_is_reference(m_b)) + mode = m_b; + else if (mode_is_reference(m_y)) + mode = m_y; + else { + /* + * Both modes are different but none is reference, + * happens for instance in SubP(SubP(P, Iu), Is). + * We have two possibilities here: Cast or ignore. + * Currently we ignore this case. + */ + return n; + } + + add = new_r_Add(current_ir_graph, blk, y, b, mode); + + set_Sub_left(n, x); + set_Sub_right(n, add); + DBG_OPT_ALGSIM0(n, n, FS_OPT_SUB_SUB_X_Y_Z); + } return n; } /** * Transform Mul(a,-1) into -a. + * Do constant evaluation of Phi nodes. * Do architecture dependent optimizations on Mul nodes */ static ir_node *transform_node_Mul(ir_node *n) { - ir_node *oldn = n; - ir_mode *mode = get_irn_mode(n); + ir_node *c, *oldn = n; + ir_node *a = get_Mul_left(n); + ir_node *b = get_Mul_right(n); + ir_mode *mode; + HANDLE_BINOP_PHI(tarval_mul, a,b,c); + + mode = get_irn_mode(n); if (mode_is_signed(mode)) { ir_node *r = NULL; - ir_node *a = get_Mul_left(n); - ir_node *b = get_Mul_right(n); if (value_of(a) == get_mode_minus_one(mode)) r = b; @@ -1624,7 +2175,7 @@ static ir_node *transform_node_Mul(ir_node *n) { } /** - * transform a Div Node + * Transform a Div Node. */ static ir_node *transform_node_Div(ir_node *n) { @@ -1645,7 +2196,7 @@ static ir_node *transform_node_Div(ir_node *n) /* Turn Div into a tuple (mem, bad, value) */ ir_node *mem = get_Div_mem(n); - turn_into_tuple(n, 3); + turn_into_tuple(n, pn_Div_max); 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, value); @@ -1654,7 +2205,7 @@ static ir_node *transform_node_Div(ir_node *n) } /** - * transform a Mod node + * Transform a Mod node. */ static ir_node *transform_node_Mod(ir_node *n) { @@ -1675,7 +2226,7 @@ static ir_node *transform_node_Mod(ir_node *n) /* Turn Mod into a tuple (mem, bad, value) */ ir_node *mem = get_Mod_mem(n); - turn_into_tuple(n, 3); + turn_into_tuple(n, pn_Mod_max); 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, value); @@ -1684,7 +2235,7 @@ static ir_node *transform_node_Mod(ir_node *n) } /** - * transform a DivMod node + * Transform a DivMod node. */ static ir_node *transform_node_DivMod(ir_node *n) { @@ -1734,7 +2285,7 @@ static ir_node *transform_node_DivMod(ir_node *n) if (evaluated) { /* replace by tuple */ ir_node *mem = get_DivMod_mem(n); - turn_into_tuple(n, 4); + turn_into_tuple(n, pn_DivMod_max); 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); @@ -1780,7 +2331,7 @@ static ir_node *transform_node_Abs(ir_node *n) } /** - * transform a Cond node + * Transform a Cond node. */ static ir_node *transform_node_Cond(ir_node *n) { @@ -1800,7 +2351,7 @@ static ir_node *transform_node_Cond(ir_node *n) /* 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)); - turn_into_tuple(n, 2); + turn_into_tuple(n, pn_Cond_max); if (ta == tarval_b_true) { set_Tuple_pred(n, pn_Cond_false, new_Bad()); set_Tuple_pred(n, pn_Cond_true, jmp); @@ -1810,44 +2361,35 @@ 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())) { - /* 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))); - /* 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(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))); - } 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), - get_Not_op(a))); } return n; } +/** + * Transform an And. + */ +static ir_node *transform_node_And(ir_node *n) +{ + ir_node *c, *oldn = n; + ir_node *a = get_And_left(n); + ir_node *b = get_And_right(n); + + HANDLE_BINOP_PHI(tarval_and, a,b,c); + return n; +} + /** * Transform an Eor. */ static ir_node *transform_node_Eor(ir_node *n) { - ir_node *oldn = n; + ir_node *c, *oldn = n; ir_node *a = get_Eor_left(n); ir_node *b = get_Eor_right(n); ir_mode *mode = get_irn_mode(n); + HANDLE_BINOP_PHI(tarval_eor, a,b,c); + if (a == b) { /* a ^ a = 0 */ n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), @@ -1878,13 +2420,16 @@ static ir_node *transform_node_Eor(ir_node *n) } /** - * Transform a boolean Not. + * Transform a Not. */ static ir_node *transform_node_Not(ir_node *n) { - ir_node *oldn = n; + ir_node *c, *oldn = n; ir_node *a = get_Not_op(n); + HANDLE_UNOP_PHI(tarval_not,a,c); + + /* check for a boolean Not */ if ( (get_irn_mode(n) == mode_b) && (get_irn_op(a) == op_Proj) && (get_irn_mode(a) == mode_b) @@ -1894,7 +2439,18 @@ static ir_node *transform_node_Not(ir_node *n) mode_b, get_negated_pnc(get_Proj_proj(a), mode_b)); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP); } + return n; +} + +/** + * Transform a Minus. + */ +static ir_node *transform_node_Minus(ir_node *n) +{ + ir_node *c, *oldn = n; + ir_node *a = get_Minus_op(n); + HANDLE_UNOP_PHI(tarval_neg,a,c); return n; } @@ -1904,7 +2460,7 @@ static ir_node *transform_node_Not(ir_node *n) static ir_node *transform_node_Cast(ir_node *n) { ir_node *oldn = n; ir_node *pred = get_Cast_op(n); - type *tp = get_irn_type(n); + ir_type *tp = get_irn_type(n); if (get_irn_op(pred) == op_Const && get_Const_type(pred) != tp) { n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred), @@ -1933,18 +2489,23 @@ static ir_node *transform_node_Proj_Div(ir_node *proj) /* div(x, y) && y != 0 */ proj_nr = get_Proj_proj(proj); - /* this node may float */ + /* this node may float if it did not depend on a Confirm */ set_irn_pinned(n, op_pin_state_floats); if (proj_nr == pn_Div_X_except) { /* we found an exception handler, remove it */ + DBG_OPT_EXC_REM(proj); return new_Bad(); - } else { - /* the memory Proj can be removed */ + } + else if (proj_nr == pn_Div_M) { ir_node *res = get_Div_mem(n); - set_Div_mem(n, get_irg_no_mem(current_ir_graph)); - if (proj_nr == pn_Div_M) - return res; + /* the memory Proj can only be removed if we divide by a + real constant, but the node never produce a new memory */ + if (value_of(b) != tarval_bad) { + /* this is a Div by a const, we can remove the memory edge */ + set_Div_mem(n, get_irg_no_mem(current_ir_graph)); + } + return res; } } return proj; @@ -1964,18 +2525,30 @@ static ir_node *transform_node_Proj_Mod(ir_node *proj) /* mod(x, y) && y != 0 */ proj_nr = get_Proj_proj(proj); - /* this node may float */ + /* this node may float if it did not depend on a Confirm */ set_irn_pinned(n, op_pin_state_floats); if (proj_nr == pn_Mod_X_except) { /* we found an exception handler, remove it */ + DBG_OPT_EXC_REM(proj); return new_Bad(); - } else { - /* the memory Proj can be removed */ + } else if (proj_nr == pn_Mod_M) { ir_node *res = get_Mod_mem(n); - set_Mod_mem(n, get_irg_no_mem(current_ir_graph)); - if (proj_nr == pn_Mod_M) - return res; + /* the memory Proj can only be removed if we divide by a + real constant, but the node never produce a new memory */ + if (value_of(b) != tarval_bad) { + /* this is a Mod by a const, we can remove the memory edge */ + set_Mod_mem(n, get_irg_no_mem(current_ir_graph)); + } + return res; + } + else if (proj_nr == pn_Mod_res && get_Mod_left(n) == b) { + /* a % a = 0 if a != 0 */ + ir_mode *mode = get_irn_mode(proj); + ir_node *res = new_Const(mode, get_mode_null(mode)); + + DBG_OPT_CSTEVAL(n, res); + return res; } } return proj; @@ -1995,19 +2568,31 @@ static ir_node *transform_node_Proj_DivMod(ir_node *proj) /* DivMod(x, y) && y != 0 */ proj_nr = get_Proj_proj(proj); - /* this node may float */ + /* this node may float if it did not depend on a Confirm */ set_irn_pinned(n, op_pin_state_floats); if (proj_nr == pn_DivMod_X_except) { /* we found an exception handler, remove it */ + DBG_OPT_EXC_REM(proj); return new_Bad(); } - else { - /* the memory Proj can be removed */ + else if (proj_nr == pn_DivMod_M) { ir_node *res = get_DivMod_mem(n); - set_DivMod_mem(n, get_irg_no_mem(current_ir_graph)); - if (proj_nr == pn_DivMod_M) - return res; + /* the memory Proj can only be removed if we divide by a + real constant, but the node never produce a new memory */ + if (value_of(b) != tarval_bad) { + /* this is a DivMod by a const, we can remove the memory edge */ + set_DivMod_mem(n, get_irg_no_mem(current_ir_graph)); + } + return res; + } + else if (proj_nr == pn_DivMod_res_mod && get_DivMod_left(n) == b) { + /* a % a = 0 if a != 0 */ + ir_mode *mode = get_irn_mode(proj); + ir_node *res = new_Const(mode, get_mode_null(mode)); + + DBG_OPT_CSTEVAL(n, res); + return res; } } return proj; @@ -2021,21 +2606,24 @@ static ir_node *transform_node_Proj_Cond(ir_node *proj) if (get_opt_unreachable_code()) { ir_node *n = get_Proj_pred(proj); ir_node *b = get_Cond_selector(n); - tarval *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 (mode_is_int(get_irn_mode(b))) { + tarval *tb = value_of(b); - 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 { - /* this case will NEVER be taken, kill it */ - return new_Bad(); + if (tb != tarval_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 { + /* this case will NEVER be taken, kill it */ + return new_Bad(); + } } } } @@ -2075,7 +2663,7 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) proj_nr = get_inversed_pnc(proj_nr); changed |= 1; } - else if (left > right) { + else if (get_irn_idx(left) > get_irn_idx(right)) { ir_node *t = left; left = right; @@ -2118,8 +2706,8 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) if (mode_is_int(mode)) { /* Ne includes Unordered which is not possible on integers. * However, frontends often use this wrong, so fix it here */ - if (proj_nr == pn_Cmp_Ne) { - proj_nr = pn_Cmp_Lg; + if (proj_nr & pn_Cmp_Uo) { + proj_nr &= ~pn_Cmp_Uo; set_Proj_proj(proj, proj_nr); } @@ -2219,7 +2807,32 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) } } } /* mode_is_int */ - } + + /* + * optimization for AND: + * Optimize: + * And(x, C) == C ==> And(x, C) != 0 + * And(x, C) != C ==> And(X, C) == 0 + * + * if C is a single Bit constant. + */ + if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && + (get_irn_op(left) == op_And)) { + if (is_single_bit_tarval(tv)) { + /* check for Constant's match. We have check hare the tarvals, + because our const might be changed */ + ir_node *la = get_And_left(left); + ir_node *ra = get_And_right(left); + if ((is_Const(la) && get_Const_tarval(la) == tv) || + (is_Const(ra) && get_Const_tarval(ra) == tv)) { + /* fine: do the transformation */ + tv = get_mode_null(get_tarval_mode(tv)); + proj_nr ^= pn_Cmp_Leg; + changed |= 2; + } + } + } + } /* tarval != bad */ } if (changed) { @@ -2273,6 +2886,49 @@ static ir_node *transform_node_Proj(ir_node *proj) } } +/** + * Move Confirms down through Phi nodes. + */ +static ir_node *transform_node_Phi(ir_node *phi) { + int i, n; + ir_mode *mode = get_irn_mode(phi); + + if (mode_is_reference(mode)) { + n = get_irn_arity(phi); + + /* Beware of Phi0 */ + if (n > 0) { + ir_node *pred = get_irn_n(phi, 0); + ir_node *bound, *new_Phi, *block, **in; + pn_Cmp pnc; + + if (! is_Confirm(pred)) + return phi; + + bound = get_Confirm_bound(pred); + pnc = get_Confirm_cmp(pred); + + NEW_ARR_A(ir_node *, in, n); + in[0] = get_Confirm_value(pred); + + for (i = 1; i < n; ++i) { + pred = get_irn_n(phi, i); + + if (! is_Confirm(pred) || + get_Confirm_bound(pred) != bound || + get_Confirm_cmp(pred) != pnc) + return phi; + in[i] = get_Confirm_value(pred); + } + /* move the Confirm nodes "behind" the Phi */ + block = get_irn_n(phi, -1); + new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi)); + return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc); + } + } + return phi; +} + /** * returns the operands of a commutative bin-op, if one operand is * a const, it is returned as the second one. @@ -2486,16 +3142,23 @@ static ir_node *transform_node_Or_Rot(ir_node *or) } /** - * Optimize an Or + * Transform an Or. */ -static ir_node *transform_node_Or(ir_node *or) +static ir_node *transform_node_Or(ir_node *n) { - or = transform_node_Or_bf_store(or); - or = transform_node_Or_Rot(or); + ir_node *c, *oldn = n; + ir_node *a = get_Or_left(n); + ir_node *b = get_Or_right(n); - return or; + HANDLE_BINOP_PHI(tarval_or, a,b,c); + + n = transform_node_Or_bf_store(n); + n = transform_node_Or_Rot(n); + + return n; } + /* forward */ static ir_node *transform_node(ir_node *n); @@ -2559,9 +3222,44 @@ static ir_node *transform_node_shift(ir_node *n) return n; } -#define transform_node_Shr transform_node_shift -#define transform_node_Shrs transform_node_shift -#define transform_node_Shl transform_node_shift +/** + * Transform a Shr. + */ +static ir_node *transform_node_Shr(ir_node *n) +{ + ir_node *c, *oldn = n; + ir_node *a = get_Shr_left(n); + ir_node *b = get_Shr_right(n); + + HANDLE_BINOP_PHI(tarval_shr, a, b, c); + return transform_node_shift(n); +} + +/** + * Transform a Shrs. + */ +static ir_node *transform_node_Shrs(ir_node *n) +{ + ir_node *c, *oldn = n; + ir_node *a = get_Shrs_left(n); + ir_node *b = get_Shrs_right(n); + + HANDLE_BINOP_PHI(tarval_shrs, a, b, c); + return transform_node_shift(n); +} + +/** + * Transform a Shl. + */ +static ir_node *transform_node_Shl(ir_node *n) +{ + ir_node *c, *oldn = n; + ir_node *a = get_Shl_left(n); + ir_node *b = get_Shl_right(n); + + HANDLE_BINOP_PHI(tarval_shl, a, b, c); + return transform_node_shift(n); +} /** * Remove dead blocks and nodes in dead blocks @@ -2717,6 +3415,16 @@ static ir_node *transform_node_Mux(ir_node *n) return arch_transform_node_Mux(n); } +/** + * Optimize a Psi into some simpler cases. + */ +static ir_node *transform_node_Psi(ir_node *n) { + if (is_Mux(n)) + return transform_node_Mux(n); + + return n; +} + /** * Tries several [inplace] [optimizing] transformations and returns an * equivalent node. The difference to equivalent_node() is that these @@ -2725,22 +3433,28 @@ static ir_node *transform_node_Mux(ir_node *n) */ static ir_node *transform_node(ir_node *n) { - if (n->op->transform_node) - n = n->op->transform_node(n); + if (n->op->ops.transform_node) + n = n->op->ops.transform_node(n); return n; } /** - * set the default transform node operation + * sSets the default transform node operation for an ir_op_ops. + * + * @param code the opcode for the default operation + * @param ops the operations initialized + * + * @return + * The operations. */ -static ir_op *firm_set_default_transform_node(ir_op *op) +static ir_op_ops *firm_set_default_transform_node(opcode code, ir_op_ops *ops) { #define CASE(a) \ case iro_##a: \ - op->transform_node = transform_node_##a; \ + ops->transform_node = transform_node_##a; \ break - switch (op->code) { + switch (code) { CASE(Add); CASE(Sub); CASE(Mul); @@ -2749,22 +3463,26 @@ static ir_op *firm_set_default_transform_node(ir_op *op) CASE(DivMod); CASE(Abs); CASE(Cond); + CASE(And); + CASE(Or); CASE(Eor); + CASE(Minus); CASE(Not); CASE(Cast); CASE(Proj); + CASE(Phi); CASE(Sel); - CASE(Or); CASE(Shr); CASE(Shrs); CASE(Shl); CASE(End); CASE(Mux); + CASE(Psi); default: - op->transform_node = NULL; + /* leave NULL */; } - return op; + return ops; #undef CASE } @@ -2783,8 +3501,7 @@ static int node_cmp_attr_Const(ir_node *a, ir_node *b) } /** Compares the attributes of two Proj nodes. */ -static int node_cmp_attr_Proj(ir_node *a, ir_node *b) -{ +static int node_cmp_attr_Proj(ir_node *a, ir_node *b) { return get_irn_proj_attr (a) != get_irn_proj_attr (b); } @@ -2817,8 +3534,7 @@ static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) } /** Compares the attributes of two Call nodes. */ -static int node_cmp_attr_Call(ir_node *a, ir_node *b) -{ +static int node_cmp_attr_Call(ir_node *a, ir_node *b) { return (get_irn_call_attr(a) != get_irn_call_attr(b)); } @@ -2833,14 +3549,17 @@ static int node_cmp_attr_Sel(ir_node *a, ir_node *b) } /** Compares the attributes of two Phi nodes. */ -static int node_cmp_attr_Phi(ir_node *a, ir_node *b) -{ +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 Conv nodes. */ +static int node_cmp_attr_Conv(ir_node *a, ir_node *b) { + return get_Conv_strict(a) != get_Conv_strict(b); +} + /** Compares the attributes of two Cast nodes. */ -static int node_cmp_attr_Cast(ir_node *a, ir_node *b) -{ +static int node_cmp_attr_Cast(ir_node *a, ir_node *b) { return get_Cast_type(a) != get_Cast_type(b); } @@ -2870,16 +3589,22 @@ static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) } /** - * set the default node attribute compare operation + * Set the default node attribute compare operation for an ir_op_ops. + * + * @param code the opcode for the default operation + * @param ops the operations initialized + * + * @return + * The operations. */ -static ir_op *firm_set_default_node_cmp_attr(ir_op *op) +static ir_op_ops *firm_set_default_node_cmp_attr(opcode code, ir_op_ops *ops) { -#define CASE(a) \ - case iro_##a: \ - op->node_cmp_attr = node_cmp_attr_##a; \ +#define CASE(a) \ + case iro_##a: \ + ops->node_cmp_attr = node_cmp_attr_##a; \ break - switch (op->code) { + switch (code) { CASE(Const); CASE(Proj); CASE(Filter); @@ -2889,24 +3614,24 @@ static ir_op *firm_set_default_node_cmp_attr(ir_op *op) CASE(Call); CASE(Sel); CASE(Phi); + CASE(Conv); CASE(Cast); CASE(Load); CASE(Store); CASE(Confirm); default: - op->node_cmp_attr = NULL; + /* leave NULL */; } - return op; + return ops; #undef CASE } -/** +/* * 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) +int identities_cmp(const void *elt, const void *key) { ir_node *a, *b; int i, irn_arity_a; @@ -2939,8 +3664,8 @@ vt_cmp (const void *elt, const void *key) * here, we already now that the nodes are identical except their * attributes */ - if (a->op->node_cmp_attr) - return a->op->node_cmp_attr(a, b); + if (a->op->ops.node_cmp_attr) + return a->op->ops.node_cmp_attr(a, b); return 0; } @@ -2948,8 +3673,7 @@ vt_cmp (const void *elt, const void *key) /* * Calculate a hash value of a node. */ -unsigned -ir_node_hash (ir_node *node) +unsigned ir_node_hash(ir_node *node) { unsigned h; int i, irn_arity; @@ -2960,7 +3684,7 @@ ir_node_hash (ir_node *node) 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 = HASH_PTR(node->attr.i.sym.type_p); + h = HASH_PTR(node->attr.symc.sym.type_p); h = 9*h + HASH_PTR(get_irn_mode(node)); } else { @@ -2968,7 +3692,7 @@ ir_node_hash (ir_node *node) h = irn_arity = get_irn_intra_arity(node); /* consider all in nodes... except the block if not a control flow. */ - for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) { + for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) { h = 9*h + HASH_PTR(get_irn_intra_n(node, i)); } @@ -2981,13 +3705,11 @@ ir_node_hash (ir_node *node) return h; } -pset * -new_identities(void) { - return new_pset(vt_cmp, N_IR_NODES); +pset *new_identities(void) { + return new_pset(identities_cmp, N_IR_NODES); } -void -del_identities(pset *value_table) { +void del_identities(pset *value_table) { del_pset(value_table); } @@ -2999,8 +3721,7 @@ del_identities(pset *value_table) { * 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) +static INLINE ir_node *identify(pset *value_table, ir_node *n) { ir_node *o = NULL; @@ -3012,14 +3733,14 @@ identify (pset *value_table, ir_node *n) ir_node *r = get_binop_right(n); /* for commutative operators perform a OP b == b OP a */ - if (l > r) { + if (get_irn_idx(l) > get_irn_idx(r)) { set_binop_left(n, r); set_binop_right(n, l); } } } - o = pset_find (value_table, n, ir_node_hash (n)); + o = pset_find(value_table, n, ir_node_hash(n)); if (!o) return n; DBG_OPT_CSE(n, o); @@ -3032,8 +3753,7 @@ identify (pset *value_table, ir_node *n) * 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 * -identify_cons (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); @@ -3042,13 +3762,12 @@ identify_cons (pset *value_table, ir_node *n) { 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. */ -static ir_node * -identify_remember (pset *value_table, ir_node *n) +ir_node *identify_remember(pset *value_table, ir_node *n) { ir_node *o = NULL; @@ -3077,18 +3796,28 @@ identify_remember (pset *value_table, ir_node *n) return o; } -void -add_identities (pset *value_table, ir_node *node) { - if (get_opt_cse() && (get_irn_opcode(node) != iro_Block)) - identify_remember (value_table, node); +/* Add a node to the identities value table. */ +void add_identities(pset *value_table, ir_node *node) { + if (get_opt_cse() && is_no_Block(node)) + identify_remember(value_table, node); +} + +/* Visit each node in the value table of a graph. */ +void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) { + ir_node *node; + ir_graph *rem = current_ir_graph; + + current_ir_graph = irg; + foreach_pset(irg->value_table, node) + visit(node, env); + current_ir_graph = rem; } /** - * garbage in, garbage out. If a node has a dead input, i.e., the + * 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 * -gigo (ir_node *node) +static INLINE ir_node *gigo(ir_node *node) { int i, irn_arity; ir_op *op = get_irn_op(node); @@ -3106,7 +3835,8 @@ gigo (ir_node *node) if (is_Block(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(); } @@ -3114,7 +3844,7 @@ gigo (ir_node *node) /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the blocks predecessors is dead. */ - if ( op != op_Block && op != op_Phi && op != op_Tuple) { + if (op != op_Block && op != op_Phi && op != op_Tuple) { irn_arity = get_irn_arity(node); /* @@ -3125,9 +3855,19 @@ gigo (ir_node *node) return new_Bad(); for (i = 0; i < irn_arity; i++) { - if (is_Bad(get_irn_n(node, i))) { + ir_node *pred = get_irn_n(node, i); + + if (is_Bad(pred)) return new_Bad(); - } +#if 0 + /* Propagating Unknowns here seems to be a bad idea, because + sometimes we need a node as a input and did not want that + it kills it's user. + However, it might be useful to move this into a later phase + (if you think that optimizing such code is useful). */ + if (is_Unknown(pred) && mode_is_data(get_irn_mode(node))) + return new_Unknown(get_irn_mode(node)); +#endif } } #if 0 @@ -3146,14 +3886,14 @@ gigo (ir_node *node) return node; } - /** * These optimizations deallocate nodes from the obstack. * It can only be called if it is guaranteed that no other nodes * reference this one, i.e., right after construction of a node. + * + * current_ir_graph must be set to the graph of the node! */ -ir_node * -optimize_node(ir_node *n) +ir_node *optimize_node(ir_node *n) { tarval *tv; ir_node *oldn = n; @@ -3170,7 +3910,7 @@ optimize_node(ir_node *n) tv = computed_value(n); if (tv != tarval_bad) { ir_node *nw; - type *old_tp = get_irn_type(n); + ir_type *old_tp = get_irn_type(n); int i, arity = get_irn_arity(n); int node_size; @@ -3197,7 +3937,7 @@ optimize_node(ir_node *n) edges_node_deleted(n, current_ir_graph); /* evaluation was successful -- replace the node. */ - obstack_free(current_ir_graph->obst, n); + irg_kill_node(current_ir_graph, n); nw = new_Const(get_tarval_mode (tv), tv); if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv)) @@ -3232,8 +3972,7 @@ optimize_node(ir_node *n) edges_node_deleted(oldn, current_ir_graph); /* We found an existing, better node, so we can deallocate the old node. */ - obstack_free (current_ir_graph->obst, oldn); - + irg_kill_node(current_ir_graph, oldn); return n; } @@ -3264,8 +4003,7 @@ optimize_node(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_2 (ir_node *n) +ir_node *optimize_in_place_2(ir_node *n) { tarval *tv; ir_node *oldn = n; @@ -3281,7 +4019,7 @@ optimize_in_place_2 (ir_node *n) tv = computed_value(n); if (tv != tarval_bad) { /* evaluation was successful -- replace the node. */ - type *old_tp = get_irn_type(n); + ir_type *old_tp = get_irn_type(n); int i, arity = get_irn_arity(n); /* @@ -3347,8 +4085,7 @@ optimize_in_place_2 (ir_node *n) /** * Wrapper for external use, set proper status bits after optimization. */ -ir_node * -optimize_in_place (ir_node *n) +ir_node *optimize_in_place(ir_node *n) { /* Handle graph state */ assert(get_irg_phase_state(current_ir_graph) != phase_building); @@ -3358,23 +4095,24 @@ optimize_in_place (ir_node *n) 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 + /* FIXME: 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); + set_irg_doms_inconsistent(current_ir_graph); return optimize_in_place_2 (n); } -/** - * set the default ir op operations +/* + * Sets the default operation for an ir_ops. */ -ir_op *firm_set_default_operations(ir_op *op) +ir_op_ops *firm_set_default_operations(opcode code, ir_op_ops *ops) { - op = firm_set_default_computed_value(op); - op = firm_set_default_equivalent_node(op); - op = firm_set_default_transform_node(op); - op = firm_set_default_node_cmp_attr(op); - op = firm_set_default_get_type(op); - - return op; + ops = firm_set_default_computed_value(code, ops); + ops = firm_set_default_equivalent_node(code, ops); + ops = firm_set_default_transform_node(code, ops); + ops = firm_set_default_node_cmp_attr(code, ops); + ops = firm_set_default_get_type(code, ops); + ops = firm_set_default_get_type_attr(code, ops); + ops = firm_set_default_get_entity_attr(code, ops); + + return ops; }