X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firopt.c;h=e4647288f4cefdafc37ede8baccd7bdc4e9de2bd;hb=e1c33a238578342a072e1c95ff12eefe6d0acd37;hp=0f9c6411e93d3ddaccccbf058ce950a610849919;hpb=498dcd223be5ec2f021e75a4cf9ce948cc87c994;p=libfirm diff --git a/ir/ir/iropt.c b/ir/ir/iropt.c index 0f9c6411e..e4647288f 100644 --- a/ir/ir/iropt.c +++ b/ir/ir/iropt.c @@ -3,7 +3,7 @@ * File name: ir/ir/iropt.c * Purpose: iropt --- optimizations intertwined with IR construction. * Author: Christian Schaefer - * Modified by: Goetz Lindenmaier + * Modified by: Goetz Lindenmaier, Michael Beck * Created: * CVS-ID: $Id$ * Copyright: (c) 1998-2005 Universität Karlsruhe @@ -24,38 +24,28 @@ #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 "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" /* Make types visible to allow most efficient access */ # include "entity_t.h" -/** - * Trivial INLINEable routine for copy propagation. - * Does follow Ids, needed to optimize INLINEd code. - */ -static INLINE ir_node * -follow_Id (ir_node *n) -{ - while (get_irn_op (n) == op_Id) n = get_Id_pred (n); - return n; -} - /** * return the value of a Constant */ @@ -69,9 +59,22 @@ 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; } @@ -116,6 +119,51 @@ 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 */ @@ -522,7 +570,7 @@ static tarval *computed_value_Proj_Cmp(ir_node *n) return new_tarval_from_long(proj_nr & pn_Cmp_Ne, mode_b); } } - return tarval_bad; + return computed_value_Cmp_Confirm(a, aa, ab, proj_nr); } /** @@ -582,6 +630,28 @@ 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). + */ +static tarval *computed_value_Confirm(ir_node *n) +{ + return get_Confirm_cmp(n) == pn_Cmp_Eq ? + value_of(get_Confirm_bound(n)) : tarval_bad; +} + /** * If the parameter n can be computed, return its value, else tarval_bad. * Performs constant folding. @@ -590,22 +660,28 @@ static tarval *computed_value_Mux(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); @@ -624,36 +700,21 @@ 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 @@ -690,7 +751,7 @@ static ir_node *equivalent_node_Block(ir_node *n) if (predblock == oldn) { /* Jmp jumps into the block it is in -- deal self cycle. */ n = set_Block_dead(n); - DBG_OPT_DEAD(oldn, n); + DBG_OPT_DEAD_BLOCK(oldn, n); } else if (get_opt_control_flow_straightening()) { n = predblock; DBG_OPT_STG(oldn, n); @@ -698,17 +759,19 @@ static ir_node *equivalent_node_Block(ir_node *n) } else if ((n_preds == 1) && (get_irn_op(skip_Proj(get_Block_cfgpred(n, 0))) == op_Cond)) { - ir_node *predblock = get_nodes_block(get_Block_cfgpred(n, 0)); + ir_node *predblock = get_Block_cfgpred_block(n, 0); if (predblock == oldn) { /* Jmp jumps into the block it is in -- deal self cycle. */ n = set_Block_dead(n); - DBG_OPT_DEAD(oldn, n); + DBG_OPT_DEAD_BLOCK(oldn, n); } } else if ((n_preds == 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. */ + * The more general case which more than 2 predecessors is handles + * in optimize_cf(), we handle only this special case for speed here. + */ ir_node *a = get_Block_cfgpred(n, 0); ir_node *b = get_Block_cfgpred(n, 1); @@ -719,23 +782,24 @@ 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(oldn, a, b, n); + 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); + } + else if (get_opt_unreachable_code() && + (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; if (is_Bad(pred)) continue; - pred_blk = get_nodes_block(pred); + pred_blk = get_nodes_block(skip_Proj(pred)); if (is_Block_dead(pred_blk)) continue; @@ -744,8 +808,10 @@ 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; @@ -753,11 +819,10 @@ 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(); @@ -765,12 +830,12 @@ static ir_node *equivalent_node_Jmp(ir_node *n) return n; } -static ir_node *equivalent_node_Cond(ir_node *n) -{ - /* We do not evaluate Cond here as we replace it by a new node, a Jmp. - See cases for iro_Cond and iro_Proj in transform_node(). */ - 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(). */ /** * optimize operations that are commutative and have neutral 0, @@ -796,19 +861,74 @@ static ir_node *equivalent_node_neutral_zero(ir_node *n) return n; /* If this predecessors constant value is zero, the operation is - unnecessary. Remove it: */ + * unnecessary. Remove it. + * + * Beware: If n is a Add, the mode of on and n might be different + * which happens in this rare construction: NULL + 3. + * Then, a Conv would be needed which we cannot include here. + */ if (classify_tarval (tv) == TV_CLASSIFY_NULL) { - n = on; + if (get_irn_mode(on) == get_irn_mode(n)) { + n = on; - DBG_OPT_ALGSIM1(oldn, a, b, n); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0); + } } return n; } -#define equivalent_node_Add equivalent_node_neutral_zero +/** + * Eor is commutative and has neutral 0. + */ #define equivalent_node_Eor equivalent_node_neutral_zero +/* + * Optimize a - 0 and (a - x) + x (for modes with wrap-around). + * + * The second one looks strange, but this construct + * is used heavily in the LCC sources :-). + * + * Beware: The Mode of an Add may be different than the mode of its + * predecessors, so we could not return a predecessors in all cases. + */ +static ir_node *equivalent_node_Add(ir_node *n) +{ + ir_node *oldn = n; + ir_node *left, *right; + + n = equivalent_node_neutral_zero(n); + if (n != oldn) + return n; + + left = get_Add_left(n); + right = get_Add_right(n); + + if (get_irn_op(left) == op_Sub) { + if (get_Sub_right(left) == right) { + /* (a - x) + x */ + + n = get_Sub_left(left); + if (get_irn_mode(oldn) == get_irn_mode(n)) { + DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB); + return n; + } + } + } + if (get_irn_op(right) == op_Sub) { + if (get_Sub_right(right) == left) { + /* x + (a - x) */ + + n = get_Sub_left(right); + if (get_irn_mode(oldn) == get_irn_mode(n)) { + DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB); + return n; + } + } + } + return n; +} + /** * optimize operations that are not commutative but have neutral 0 on left, * so a op 0 = a. @@ -823,24 +943,74 @@ static ir_node *equivalent_node_left_zero(ir_node *n) if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) { n = a; - DBG_OPT_ALGSIM1(oldn, a, b, n); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0); } return 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 +/** + * Optimize a - 0 and (a + x) - x (for modes with wrap-around). + * + * The second one looks strange, but this construct + * is used heavily in the LCC sources :-). + * + * Beware: The Mode of a Sub may be different than the mode of its + * predecessors, so we could not return a predecessors in all cases. + */ +static ir_node *equivalent_node_Sub(ir_node *n) +{ + ir_node *oldn = n; + + ir_node *a = get_Sub_left(n); + ir_node *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)) { + 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)) { + 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)) { + n = left; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); + } + } + } + } + + return 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) { @@ -855,10 +1025,10 @@ 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 @@ -875,10 +1045,10 @@ static ir_node *equivalent_node_Mul(ir_node *n) /* Mul is commutative and has again an other neutral element. */ if (classify_tarval(value_of(a)) == TV_CLASSIFY_ONE) { n = b; - DBG_OPT_ALGSIM1(oldn, a, b, n); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1); } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { n = a; - DBG_OPT_ALGSIM1(oldn, a, b, n); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1); } return n; } @@ -895,7 +1065,7 @@ 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); @@ -917,7 +1087,7 @@ 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); @@ -938,12 +1108,13 @@ static ir_node *equivalent_node_Or(ir_node *n) if (a == b) { n = a; /* Or has it's own neutral element */ + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR); } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_NULL) { n = b; - DBG_OPT_ALGSIM1(oldn, a, b, n); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR); } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) { n = a; - DBG_OPT_ALGSIM1(oldn, a, b, n); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR); } return n; @@ -961,12 +1132,13 @@ static ir_node *equivalent_node_And(ir_node *n) if (a == b) { n = a; /* And has it's own neutral element */ + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND); } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_ALL_ONE) { n = b; - DBG_OPT_ALGSIM1(oldn, a, b, n); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND); } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ALL_ONE) { n = a; - DBG_OPT_ALGSIM1(oldn, a, b, n); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND); } return n; } @@ -985,7 +1157,7 @@ static ir_node *equivalent_node_Conv(ir_node *n) if (n_mode == a_mode) { /* No Conv necessary */ n = a; - DBG_OPT_ALGSIM3(oldn, a, n); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV); } else if (get_irn_op(a) == op_Conv) { /* Conv(Conv(b)) */ ir_mode *b_mode; @@ -996,12 +1168,12 @@ static ir_node *equivalent_node_Conv(ir_node *n) if (n_mode == b_mode) { if (n_mode == mode_b) { n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */ - DBG_OPT_ALGSIM1(oldn, a, b, n); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV); } else if (mode_is_int(n_mode) || mode_is_character(n_mode)) { if (smaller_mode(b_mode, a_mode)){ n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */ - DBG_OPT_ALGSIM1(oldn, a, b, n); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV); } } } @@ -1014,18 +1186,23 @@ static ir_node *equivalent_node_Conv(ir_node *n) * is already the type of the Cast. */ static ir_node *equivalent_node_Cast(ir_node *n) { + ir_node *oldn = n; ir_node *pred = get_Cast_op(n); - if (get_irn_type(pred) == get_Cast_type(n)) + + if (get_irn_type(pred) == get_Cast_type(n)) { n = pred; + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST); + } 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; @@ -1033,7 +1210,6 @@ 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; @@ -1043,79 +1219,109 @@ 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. */ -#if 0 - /* first we test for a special case: */ - /* Confirm is a special node fixing additional information for a - value that is known at a certain point. This is useful for - dataflow analysis. */ - if (n_preds == 2) { - ir_node *a = get_Phi_pred(n, 0); - ir_node *b = get_Phi_pred(n, 1); - if ( (get_irn_op(a) == op_Confirm) - && (get_irn_op(b) == op_Confirm) - && follow_Id (get_irn_n(a, 0) == get_irn_n(b, 0)) - && (get_irn_n(a, 1) == get_irn_n (b, 1)) - && (a->data.num == (~b->data.num & irpn_True) )) { - return get_irn_n(a, 0); - } - } -#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))) + 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) { + for (i = 0; i < n_preds; ++i) { first_val = get_Phi_pred(n, i); if ( (first_val != n) /* not self pointer */ #if 1 - && (get_irn_op(first_val) != op_Bad) + && (! is_Bad(first_val)) #endif ) { /* value not dead */ break; /* then found first value. */ } } - /* A totally Bad or self-referencing Phi (we didn't break the above loop) */ - if (i >= n_preds) { return new_Bad(); } - - scnd_val = NULL; + if (i >= n_preds) { + /* A totally Bad or self-referencing Phi (we didn't break the above loop) */ + return new_Bad(); + } - /* 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 - && (get_irn_op(scnd_val) != op_Bad) + && (! is_Bad(scnd_val)) #endif ) { break; } } - /* Fold, if no multiple distinct non-self-referencing inputs */ if (i >= n_preds) { + /* Fold, if no multiple distinct non-self-referencing inputs */ n = first_val; - DBG_OPT_PHI(oldn, first_val, n); - } else { - /* skip the remaining Ids (done in get_Phi_pred). */ - /* superfluous, since we walk all to propagate Block's Bads. - while (++i < n_preds) get_Phi_pred(n, i); */ + DBG_OPT_PHI(oldn, n); } 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. + @fixme: are there loop's in Sync's + */ +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) { @@ -1132,11 +1338,28 @@ 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 && - is_Block_dead(get_nodes_block(n))) { - /* Remove dead control flow -- early gigo. */ - n = new_Bad(); } + 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; } @@ -1147,7 +1370,10 @@ static ir_node *equivalent_node_Id(ir_node *n) { ir_node *oldn = n; - n = follow_Id(n); + do { + n = get_Id_pred(n); + } while (get_irn_op(n) == op_Id); + DBG_OPT_ID(oldn, n); return n; } @@ -1161,19 +1387,19 @@ static ir_node *equivalent_node_Mux(ir_node *n) tarval *ts = value_of(sel); /* Mux(true, f, t) == t */ - if (ts == get_tarval_b_true()) { + if (ts == tarval_b_true) { n = get_Mux_true(n); - DBG_OPT_ALGSIM0(oldn, n); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C); } /* Mux(false, f, t) == f */ - else if (ts == get_tarval_b_false()) { + else if (ts == tarval_b_false) { n = get_Mux_false(n); - DBG_OPT_ALGSIM0(oldn, n); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C); } /* Mux(v, x, x) == x */ else if (get_Mux_false(n) == get_Mux_true(n)) { n = get_Mux_true(n); - DBG_OPT_ALGSIM0(oldn, n); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ); } else if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(get_irn_mode(n))) { ir_node *cmp = get_Proj_pred(sel); @@ -1197,12 +1423,12 @@ static ir_node *equivalent_node_Mux(ir_node *n) if (proj_nr == pn_Cmp_Eq) { /* Mux(a == 0, -a, a) ==> -a */ n = b; - DBG_OPT_ALGSIM0(oldn, n); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) { /* Mux(a != 0, -a, a) ==> a */ n = a; - DBG_OPT_ALGSIM0(oldn, n); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); } } else if (classify_Const(b) == CNST_NULL) { @@ -1210,18 +1436,28 @@ static ir_node *equivalent_node_Mux(ir_node *n) if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) { /* Mux(a != 0, 0, a) ==> a */ n = a; - DBG_OPT_ALGSIM0(oldn, n); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); } else if (proj_nr == pn_Cmp_Eq) { /* Mux(a == 0, 0, a) ==> 0 */ n = b; - DBG_OPT_ALGSIM0(oldn, n); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); } } } } } + 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; } @@ -1247,6 +1483,103 @@ static ir_node *equivalent_node_Cmp(ir_node *n) return n; } +/** + * Remove Confirm nodes if setting is on. + * Replace Confirms(x, '=', Constlike) by Constlike. + */ +static ir_node *equivalent_node_Confirm(ir_node *n) +{ + 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); + + /* + * Optimize a rare case: + * Confirm(x, '=', Constlike) ==> Constlike + */ + 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 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; +} + /** * equivalent_node() returns a node equivalent to input n. It skips all nodes that * perform no actual computation, as, e.g., the Id nodes. It does not create @@ -1257,25 +1590,31 @@ static ir_node *equivalent_node_Cmp(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(Cond); + CASE(Raise); CASE(Or); CASE(Add); CASE(Eor); @@ -1293,15 +1632,20 @@ static ir_op *firm_set_default_equivalent_node(ir_op *op) 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 } @@ -1338,6 +1682,18 @@ optimize_preds(ir_node *n) { } /* end switch */ } +/** + * Returns non-zero if all Phi predecessors are constants + */ +static int is_const_Phi(ir_node *phi) { + int i; + + for (i = get_irn_arity(phi) - 1; i >= 0; --i) + if (! is_Const(get_irn_n(phi, i))) + return 0; + return 1; +} + /** * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)). @@ -1402,7 +1758,9 @@ static ir_node *transform_node_AddSub(ir_node *n) } /** - * Do the AddSub optimization, then Transform Add(a,a) into Mul(a, 2) + * Do the AddSub optimization, then Transform + * 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. @@ -1417,9 +1775,10 @@ static ir_node *transform_node_Add(ir_node *n) mode = get_irn_mode(n); if (mode_is_num(mode)) { ir_node *a = get_Add_left(n); + ir_node *b = get_Add_right(n); - if (a == get_Add_right(n)) { - ir_node *block = get_nodes_block(n); + if (a == b) { + ir_node *block = get_irn_n(n, -1); n = new_rd_Mul( get_irn_dbg_info(n), @@ -1428,30 +1787,90 @@ static ir_node *transform_node_Add(ir_node *n) a, new_r_Const_long(current_ir_graph, block, mode, 2), mode); - DBG_OPT_ALGSIM0(oldn, n); + 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_nodes_block(n), - b, - get_Minus_op(a), - mode); - DBG_OPT_ALGSIM0(oldn, n); + 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_nodes_block(n), - a, - get_Minus_op(b), - mode); - DBG_OPT_ALGSIM0(oldn, n); + 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); } } } @@ -1459,32 +1878,73 @@ 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 + * Sub(0,a) -> Minus(a) + * Sub(Mul(a, x), a) -> Mul(a, x-1) */ static ir_node *transform_node_Sub(ir_node *n) { ir_mode *mode; ir_node *oldn = n; + ir_node *a, *b; n = transform_node_AddSub(n); mode = get_irn_mode(n); - if (mode_is_num(mode) && (classify_Const(get_Sub_left(n)) == CNST_NULL)) { + a = get_Sub_left(n); + b = get_Sub_right(n); + if (mode_is_num(mode) && (classify_Const(a) == CNST_NULL)) { n = new_rd_Minus( get_irn_dbg_info(n), current_ir_graph, - get_nodes_block(n), - get_Sub_right(n), + get_irn_n(n, -1), + b, mode); - DBG_OPT_ALGSIM0(oldn, n); + 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); + } } return n; } /** - Transform Mul(a,-1) into -a. - * Do architecture dependend optimizations on Mul nodes + * Transform Mul(a,-1) into -a. + * Do architecture dependent optimizations on Mul nodes */ static ir_node *transform_node_Mul(ir_node *n) { ir_node *oldn = n; @@ -1500,8 +1960,8 @@ static ir_node *transform_node_Mul(ir_node *n) { else if (value_of(b) == get_mode_minus_one(mode)) r = a; if (r) { - n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), r, mode); - DBG_OPT_ALGSIM1(oldn, a, b, n); + n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1); return n; } } @@ -1523,14 +1983,14 @@ static ir_node *transform_node_Div(ir_node *n) DBG_OPT_CSTEVAL(n, value); } - else /* Try architecture dependend optimization */ + else /* Try architecture dependent optimization */ value = arch_dep_replace_div_by_const(n); if (value != n) { /* Turn Div into a tuple (mem, bad, value) */ ir_node *mem = get_Div_mem(n); - turn_into_tuple(n, 3); + 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); @@ -1553,14 +2013,14 @@ static ir_node *transform_node_Mod(ir_node *n) DBG_OPT_CSTEVAL(n, value); } - else /* Try architecture dependend optimization */ + else /* Try architecture dependent optimization */ value = arch_dep_replace_mod_by_const(n); if (value != n) { /* Turn Mod into a tuple (mem, bad, value) */ ir_node *mem = get_Mod_mem(n); - turn_into_tuple(n, 3); + 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); @@ -1607,7 +2067,7 @@ static ir_node *transform_node_DivMod(ir_node *n) DBG_OPT_CSTEVAL(n, a); DBG_OPT_CSTEVAL(n, b); } - else { /* Try architecture dependend optimization */ + else { /* Try architecture dependent optimization */ arch_dep_replace_divmod_by_const(&a, &b, n); evaluated = a != NULL; } @@ -1619,125 +2079,46 @@ 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); set_Tuple_pred(n, pn_DivMod_res_mod, b); - assert(get_nodes_block(n)); } return n; } - /** - * Optimize Abs(x) => x if x is Confirmed >= 0 - * Optimize Abs(x) => -x if x is Confirmed <= 0 + * Optimize Abs(x) into x if x is Confirmed >= 0 + * Optimize Abs(x) into -x if x is Confirmed <= 0 */ static ir_node *transform_node_Abs(ir_node *n) { - ir_node *oldn = n; - ir_node *a = get_Abs_op(n); - tarval *tv, *c; - ir_mode *mode; - pn_Cmp cmp, ncmp; - int do_minus = 0; - - if (get_irn_op(a) != op_Confirm) - return n; - - tv = value_of(get_Confirm_bound(n)); - if (tv == tarval_bad) - return n; - - mode = get_irn_mode(n); - - /* - * We can handle only >=, >, <, <= cases. - * We could handle == too, but this should not happen. - * - * Note that for integer modes we have a slightly better - * optimization possibilities, so we handle this - * different. - */ - cmp = get_Confirm_cmp(n); - - switch (cmp) { - case pn_Cmp_Lt: - /* - * must be x < c <= 1 to be useful if integer mode and -0 = 0 - * x < c <= 0 to be useful else - */ - case pn_Cmp_Le: - /* - * must be x <= c < 1 to be useful if integer mode and -0 = 0 - * x <= c < 0 to be useful else - */ - c = mode_is_int(mode) && mode_honor_signed_zeros(mode) ? - get_mode_one(mode) : get_mode_null(mode); - - ncmp = tarval_cmp(tv, c) ^ pn_Cmp_Eq; - if (cmp != ncmp) - return n; + ir_node *oldn = n; + ir_node *a = get_Abs_op(n); + value_classify sign = classify_value_sign(a); - /* yep, we can replace by -x */ - do_minus = 1; - break; + if (sign == VALUE_NEGATIVE) { + ir_mode *mode = get_irn_mode(n); - case pn_Cmp_Ge: - /* - * must be x >= c > -1 to be useful if integer mode - * x >= c >= 0 to be useful else - */ - case pn_Cmp_Gt: /* - * must be x > c >= -1 to be useful if integer mode - * x > c >= 0 to be useful else + * We can replace the Abs by -x here. + * We even could add a new Confirm here. + * + * Note that -x would create a new node, so we could + * not run it in the equivalent_node() context. */ - if (mode_is_int(mode)) { - c = get_mode_minus_one(mode); - - ncmp = tarval_cmp(tv, c) ^ pn_Cmp_Eq; - if (cmp != ncmp) - return n; - } - else { - c = get_mode_minus_one(mode); - - ncmp = tarval_cmp(tv, c); - - if (ncmp != pn_Cmp_Eq && ncmp != pn_Cmp_Gt) - return n; - } - - /* yep, we can replace by x */ - do_minus = 0; - break; - - default: - return n; - } - - /* - * Ok, when we get here, we can replace the Abs - * by either x or -x. In the second case we eve could - * add a new Confirm. - * - * Note that -x would create a new node, so we could - * not run it in the equivalent_node() context. - */ - if (do_minus) { - ir_node *c; - n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, - get_nodes_block(n), a, mode); + get_irn_n(n, -1), a, mode); - c = new_Const(mode, tarval_neg(tv)); - n = new_r_Confirm(current_ir_graph, get_nodes_block(n), n, c, get_inversed_pnc(cmp)); + DBG_OPT_CONFIRM(oldn, n); } - else { + else if (sign == VALUE_POSITIVE) { + /* n is positive, Abs is not needed */ n = a; + + DBG_OPT_CONFIRM(oldn, n); } return n; @@ -1754,13 +2135,17 @@ static ir_node *transform_node_Cond(ir_node *n) ir_node *a = get_Cond_selector(n); tarval *ta = value_of(a); + /* we need block info which is not available in floating irgs */ + if (get_irg_pinned(current_ir_graph) == op_pin_state_floats) + return n; + if ((ta != tarval_bad) && (get_irn_mode(a) == mode_b) && (get_opt_unreachable_code())) { /* It's a boolean Cond, branching on a boolean constant. Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */ 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); @@ -1770,30 +2155,6 @@ 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; } @@ -1810,9 +2171,9 @@ static ir_node *transform_node_Eor(ir_node *n) if (a == b) { /* a ^ a = 0 */ - n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), + n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), mode, get_mode_null(mode)); - DBG_OPT_ALGSIM0(oldn, n); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A); } else if ((mode == mode_b) && (get_irn_op(a) == op_Proj) @@ -1820,18 +2181,18 @@ static ir_node *transform_node_Eor(ir_node *n) && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE) && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) { /* The Eor negates a Cmp. The Cmp has the negated result anyways! */ - n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a), - mode_b, get_negated_pnc(get_Proj_proj(a))); + n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a), + mode_b, get_negated_pnc(get_Proj_proj(a), mode)); - DBG_OPT_ALGSIM0(oldn, n); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL); } else if ((mode == mode_b) && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)) { /* The Eor is a Not. Replace it by a Not. */ /* ????!!!Extend to bitfield 1111111. */ - n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode_b); + n = new_r_Not(current_ir_graph, get_irn_n(n, -1), a, mode_b); - DBG_OPT_ALGSIM0(oldn, n); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT); } return n; @@ -1850,28 +2211,28 @@ 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), - mode_b, get_negated_pnc(get_Proj_proj(a))); - DBG_OPT_ALGSIM0(oldn, n); + n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a), + mode_b, get_negated_pnc(get_Proj_proj(a), mode_b)); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP); } return n; } /** - * Transform a Cast of a Const into a new Const + * Transform a Cast_type(Const) into a new Const_type */ 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_nodes_block(pred), get_irn_mode(pred), + n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred), get_Const_tarval(pred), tp); DBG_OPT_CSTEVAL(oldn, n); } else if ((get_irn_op(pred) == op_SymConst) && (get_SymConst_value_type(pred) != tp)) { - n = new_rd_SymConst_type(NULL, current_ir_graph, get_nodes_block(pred), get_SymConst_symbol(pred), + n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_SymConst_symbol(pred), get_SymConst_kind(pred), tp); DBG_OPT_CSTEVAL(oldn, n); } @@ -1880,106 +2241,123 @@ static ir_node *transform_node_Cast(ir_node *n) { } /** - * Transform a Proj(Div) with a non-zero constant. + * Transform a Proj(Div) with a non-zero value. * Removes the exceptions and routes the memory to the NoMem node. */ static ir_node *transform_node_Proj_Div(ir_node *proj) { ir_node *n = get_Proj_pred(proj); - ir_node *b; - tarval *tb; + ir_node *b = get_Div_right(n); long proj_nr; - b = get_Div_right(n); - tb = value_of(b); - - if (tb != tarval_bad && classify_tarval(tb) != TV_CLASSIFY_NULL) { - /* div(x, c) && c != 0 */ + if (value_not_zero(b)) { + /* 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; } /** - * Transform a Proj(Mod) with a non-zero constant. + * Transform a Proj(Mod) with a non-zero value. * Removes the exceptions and routes the memory to the NoMem node. */ static ir_node *transform_node_Proj_Mod(ir_node *proj) { ir_node *n = get_Proj_pred(proj); - ir_node *b; - tarval *tb; + ir_node *b = get_Mod_right(n); long proj_nr; - b = get_Mod_right(n); - tb = value_of(b); - - if (tb != tarval_bad && classify_tarval(tb) != TV_CLASSIFY_NULL) { - /* mod(x, c) && c != 0 */ + if (value_not_zero(b)) { + /* 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; } /** - * Transform a Proj(DivMod) with a non-zero constant. + * Transform a Proj(DivMod) with a non-zero value. * Removes the exceptions and routes the memory to the NoMem node. */ static ir_node *transform_node_Proj_DivMod(ir_node *proj) { ir_node *n = get_Proj_pred(proj); - ir_node *b; - tarval *tb; + ir_node *b = get_DivMod_right(n); long proj_nr; - b = get_DivMod_right(n); - tb = value_of(b); - - if (tb != tarval_bad && classify_tarval(tb) != TV_CLASSIFY_NULL) { - /* DivMod(x, c) && c != 0 */ + if (value_not_zero(b)) { + /* 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; @@ -1993,21 +2371,24 @@ 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(); + } } } } @@ -2047,7 +2428,7 @@ 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; @@ -2090,8 +2471,8 @@ 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); } @@ -2191,11 +2572,36 @@ 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) { - ir_node *block = get_nodes_block(n); + ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */ if (changed & 2) /* need a new Const */ right = new_Const(mode, tv); @@ -2245,6 +2651,42 @@ 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; + pn_Cmp pnc; + + if (! is_Confirm(pred)) + return phi; + + bound = get_Confirm_bound(pred); + pnc = get_Confirm_cmp(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; + } + return new_r_Confirm(current_ir_graph, get_irn_n(phi, -1), 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. @@ -2332,7 +2774,7 @@ static ir_node *transform_node_Or_bf_store(ir_node *or) } /* ok, all conditions met */ - block = get_nodes_block(or); + block = get_irn_n(or, -1); new_and = new_r_And(current_ir_graph, block, value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode); @@ -2395,11 +2837,11 @@ static ir_node *transform_node_Or_Rot(ir_node *or) return or; /* yet, condition met */ - block = get_nodes_block(or); + block = get_irn_n(or, -1); n = new_r_Rot(current_ir_graph, block, x, c1, mode); - DBG_OPT_ALGSIM1(or, shl, shr, n); + DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT); return n; } else if (get_irn_op(c1) == op_Sub) { @@ -2426,7 +2868,7 @@ static ir_node *transform_node_Or_Rot(ir_node *or) /* a Rot right is not supported, so use a rot left */ n = new_r_Rot(current_ir_graph, block, x, sub, mode); - DBG_OPT_ALGSIM0(or, n); + DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT); return n; } else if (get_irn_op(c2) == op_Sub) { @@ -2445,12 +2887,12 @@ static ir_node *transform_node_Or_Rot(ir_node *or) return or; /* yet, condition met */ - block = get_nodes_block(or); + block = get_irn_n(or, -1); /* a Rot Left */ n = new_r_Rot(current_ir_graph, block, x, v, mode); - DBG_OPT_ALGSIM0(or, n); + DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT); return n; } @@ -2472,7 +2914,9 @@ static ir_node *transform_node_Or(ir_node *or) static ir_node *transform_node(ir_node *n); /** - * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl + * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl. + * + * Should be moved to reassociation? */ static ir_node *transform_node_shift(ir_node *n) { @@ -2515,14 +2959,14 @@ static ir_node *transform_node_shift(ir_node *n) if (flag) { /* ok, we can replace it */ - ir_node *in[2], *irn, *block = get_nodes_block(n); + ir_node *in[2], *irn, *block = get_irn_n(n, -1); in[0] = get_binop_left(left); in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res); irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in); - DBG_OPT_ALGSIM0(n, irn); + DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT); return transform_node(irn); } @@ -2534,21 +2978,27 @@ static ir_node *transform_node_shift(ir_node *n) #define transform_node_Shl transform_node_shift /** - * Remove dead blocks in keepalive list. We do not generate a new End node. + * Remove dead blocks and nodes in dead blocks + * in keep alive list. We do not generate a new End node. */ static ir_node *transform_node_End(ir_node *n) { int i, n_keepalives = get_End_n_keepalives(n); for (i = 0; i < n_keepalives; ++i) { ir_node *ka = get_End_keepalive(n, i); - if (is_Block(ka) && is_Block_dead(ka)) + if (is_Block(ka)) { + if (is_Block_dead(ka)) { + set_End_keepalive(n, i, new_Bad()); + } + } + else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) set_End_keepalive(n, i, new_Bad()); } return n; } /** - * Optimize a Mux into some simplier cases. + * Optimize a Mux into some simpler cases. */ static ir_node *transform_node_Mux(ir_node *n) { @@ -2562,7 +3012,7 @@ static ir_node *transform_node_Mux(ir_node *n) ir_node *t = get_Mux_true(n); if (get_irn_op(cmp) == op_Cmp && classify_Const(get_Cmp_right(cmp)) == CNST_NULL) { - ir_node *block = get_nodes_block(n); + ir_node *block = get_irn_n(n, -1); /* * Note: normalization puts the constant on the right site, @@ -2582,7 +3032,7 @@ static ir_node *transform_node_Mux(ir_node *n) current_ir_graph, block, t, mode); - DBG_OPT_ALGSIM1(oldn, cmp, sel, n); + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS); return n; } else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) { @@ -2596,7 +3046,7 @@ static ir_node *transform_node_Mux(ir_node *n) block, n, mode); - DBG_OPT_ALGSIM1(oldn, cmp, sel, n); + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS); return n; } } @@ -2610,7 +3060,7 @@ static ir_node *transform_node_Mux(ir_node *n) current_ir_graph, block, f, mode); - DBG_OPT_ALGSIM1(oldn, cmp, sel, n); + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS); return n; } else if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) { @@ -2624,7 +3074,7 @@ static ir_node *transform_node_Mux(ir_node *n) block, n, mode); - DBG_OPT_ALGSIM1(oldn, cmp, sel, n); + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS); return n; } } @@ -2654,7 +3104,7 @@ static ir_node *transform_node_Mux(ir_node *n) new_r_Const_long(current_ir_graph, block, mode_Iu, get_mode_size_bits(mode) - 1), mode); - DBG_OPT_ALGSIM1(oldn, cmp, sel, n); + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR); return n; } else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Ge) && @@ -2671,7 +3121,7 @@ static ir_node *transform_node_Mux(ir_node *n) new_r_Const_long(current_ir_graph, block, mode_Iu, get_mode_size_bits(mode) - 1), mode); - DBG_OPT_ALGSIM1(oldn, cmp, sel, n); + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR); return n; } } @@ -2681,6 +3131,16 @@ 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 @@ -2689,22 +3149,28 @@ 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); @@ -2717,6 +3183,7 @@ static ir_op *firm_set_default_transform_node(ir_op *op) CASE(Not); CASE(Cast); CASE(Proj); + CASE(Phi); CASE(Sel); CASE(Or); CASE(Shr); @@ -2724,11 +3191,12 @@ static ir_op *firm_set_default_transform_node(ir_op *op) CASE(Shl); CASE(End); CASE(Mux); + CASE(Psi); default: - op->transform_node = NULL; + /* leave NULL */; } - return op; + return ops; #undef CASE } @@ -2749,33 +3217,33 @@ 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) { - return get_irn_proj_attr (a) != get_irn_proj_attr (b); + return get_irn_proj_attr (a) != get_irn_proj_attr (b); } /** Compares the attributes of two Filter nodes. */ static int node_cmp_attr_Filter(ir_node *a, ir_node *b) { - return get_Filter_proj(a) != get_Filter_proj(b); + return get_Filter_proj(a) != get_Filter_proj(b); } /** Compares the attributes of two Alloc nodes. */ static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) { - return (get_irn_alloc_attr(a).where != get_irn_alloc_attr(b).where) - || (get_irn_alloc_attr(a).type != get_irn_alloc_attr(b).type); + return (get_irn_alloc_attr(a).where != get_irn_alloc_attr(b).where) + || (get_irn_alloc_attr(a).type != get_irn_alloc_attr(b).type); } /** Compares the attributes of two Free nodes. */ static int node_cmp_attr_Free(ir_node *a, ir_node *b) { - return (get_irn_free_attr(a).where != get_irn_free_attr(b).where) - || (get_irn_free_attr(a).type != get_irn_free_attr(b).type); + return (get_irn_free_attr(a).where != get_irn_free_attr(b).where) + || (get_irn_free_attr(a).type != get_irn_free_attr(b).type); } /** Compares the attributes of two SymConst nodes. */ static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) { - return (get_irn_symconst_attr(a).num != get_irn_symconst_attr(b).num) + return (get_irn_symconst_attr(a).num != get_irn_symconst_attr(b).num) || (get_irn_symconst_attr(a).sym.type_p != get_irn_symconst_attr(b).sym.type_p) || (get_irn_symconst_attr(a).tp != get_irn_symconst_attr(b).tp); } @@ -2783,13 +3251,13 @@ 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) { - return (get_irn_call_attr(a) != get_irn_call_attr(b)); + return (get_irn_call_attr(a) != get_irn_call_attr(b)); } /** Compares the attributes of two Sel nodes. */ static int node_cmp_attr_Sel(ir_node *a, ir_node *b) { - return (get_irn_sel_attr(a).ent->kind != get_irn_sel_attr(b).ent->kind) + return (get_irn_sel_attr(a).ent->kind != get_irn_sel_attr(b).ent->kind) || (get_irn_sel_attr(a).ent->name != get_irn_sel_attr(b).ent->name) || (get_irn_sel_attr(a).ent->owner != get_irn_sel_attr(b).ent->owner) || (get_irn_sel_attr(a).ent->ld_name != get_irn_sel_attr(b).ent->ld_name) @@ -2799,13 +3267,13 @@ 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) { - return get_irn_phi_attr (a) != get_irn_phi_attr (b); + return get_irn_phi_attr (a) != get_irn_phi_attr (b); } /** Compares the attributes of two Cast nodes. */ static int node_cmp_attr_Cast(ir_node *a, ir_node *b) { - return get_Cast_type(a) != get_Cast_type(b); + return get_Cast_type(a) != get_Cast_type(b); } /** Compares the attributes of two Load nodes. */ @@ -2824,20 +3292,32 @@ static int node_cmp_attr_Store(ir_node *a, ir_node *b) { /* NEVER do CSE on volatile Stores */ return (get_Store_volatility(a) == volatility_is_volatile || - get_Store_volatility(b) == volatility_is_volatile); + get_Store_volatility(b) == volatility_is_volatile); +} + +/** Compares the attributes of two Confirm nodes. */ +static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) +{ + return (get_Confirm_cmp(a) != get_Confirm_cmp(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); @@ -2850,20 +3330,20 @@ static ir_op *firm_set_default_node_cmp_attr(ir_op *op) 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; @@ -2896,8 +3376,8 @@ 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; } @@ -2940,7 +3420,7 @@ ir_node_hash (ir_node *node) pset * new_identities(void) { - return new_pset(vt_cmp, N_IR_NODES); + return new_pset(identities_cmp, N_IR_NODES); } void @@ -2956,8 +3436,7 @@ 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; @@ -2969,14 +3448,14 @@ 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); @@ -2989,8 +3468,7 @@ 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); @@ -2999,13 +3477,12 @@ 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; @@ -3034,33 +3511,47 @@ 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 * 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); + ir_op *op = get_irn_op(node); /* remove garbage blocks by looking at control flow that leaves the block and replacing the control flow by Bad. */ if (get_irn_mode(node) == mode_X) { - ir_node *block = get_nodes_block(node); - if (!get_Block_matured(block)) return node; /* Don't optimize nodes in immature blocks. */ - if (op == op_End) return node; /* Don't optimize End, may have Bads. */ + ir_node *block = get_nodes_block(skip_Proj(node)); - if (get_irn_op(block) == op_Block && get_Block_matured(block)) { + /* Don't optimize nodes in immature blocks. */ + if (!get_Block_matured(block)) return node; + /* Don't optimize End, may have Bads. */ + if (op == op_End) return 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(); } @@ -3071,13 +3562,27 @@ gigo (ir_node *node) if ( op != op_Block && op != op_Phi && op != op_Tuple) { irn_arity = get_irn_arity(node); - if (is_Block_dead(get_nodes_block(node))) + /* + * Beware: we can only read the block of a non-floating node. + */ + if (is_irn_pinned_in_irg(node) && + is_Block_dead(get_nodes_block(node))) return new_Bad(); for (i = 0; i < irn_arity; i++) { - if (is_Bad(get_irn_n(node, i))) { + 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, i might be useful to move this into a later phase + (it you thing 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 @@ -3096,112 +3601,115 @@ 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) -{ - tarval *tv; - ir_node *oldn = n; - opcode iro = get_irn_opcode(n); - - type *old_tp = get_irn_type(n); - { - int i, arity = get_irn_arity(n); - for (i = 0; i < arity && !old_tp; ++i) - old_tp = get_irn_type(get_irn_n(n, i)); - } - - /* Always optimize Phi nodes: part of the construction. */ - if ((!get_opt_optimize()) && (iro != iro_Phi)) return n; +ir_node *optimize_node(ir_node *n) +{ + tarval *tv; + ir_node *oldn = n; + opcode iro = get_irn_opcode(n); - /* constant expression evaluation / constant folding */ - if (get_opt_constant_folding()) { - /* constants can not be evaluated */ - if (iro != iro_Const) { - /* try to evaluate */ - tv = computed_value(n); - if ((get_irn_mode(n) != mode_T) && (tv != tarval_bad)) { - ir_node *nw; + /* Always optimize Phi nodes: part of the construction. */ + if ((!get_opt_optimize()) && (iro != iro_Phi)) return n; - /* - * we MUST copy the node here temporary, because it's still needed - * for DBG_OPT_CSTEVAL - */ - int node_size = offsetof(ir_node, attr) + n->op->attr_size; - oldn = alloca(node_size); + /* constant expression evaluation / constant folding */ + if (get_opt_constant_folding()) { + /* neither constants nor Tuple values can be evaluated */ + if (iro != iro_Const && (get_irn_mode(n) != mode_T)) { + /* try to evaluate */ + tv = computed_value(n); + if (tv != tarval_bad) { + ir_node *nw; + ir_type *old_tp = get_irn_type(n); + int i, arity = get_irn_arity(n); + int node_size; - memcpy(oldn, n, node_size); - CLONE_ARR_A(ir_node *, oldn->in, n->in); + /* + * Try to recover the type of the new expression. + */ + for (i = 0; i < arity && !old_tp; ++i) + old_tp = get_irn_type(get_irn_n(n, i)); - /* ARG, copy the in array, we need it for statistics */ - memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0])); + /* + * we MUST copy the node here temporary, because it's still needed + * for DBG_OPT_CSTEVAL + */ + node_size = offsetof(ir_node, attr) + n->op->attr_size; + oldn = alloca(node_size); + memcpy(oldn, n, node_size); + CLONE_ARR_A(ir_node *, oldn->in, n->in); - edges_node_deleted(n, current_ir_graph); + /* ARG, copy the in array, we need it for statistics */ + memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0])); - /* evaluation was successful -- replace the node. */ - obstack_free (current_ir_graph->obst, n); - nw = new_Const (get_tarval_mode (tv), tv); + /* note the inplace edges module */ + edges_node_deleted(n, current_ir_graph); - if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv)) - set_Const_type(nw, old_tp); - DBG_OPT_CSTEVAL(oldn, nw); - return nw; - } - } - } + /* evaluation was successful -- replace the node. */ + irg_kill_node(current_ir_graph, n); + nw = new_Const(get_tarval_mode (tv), tv); - /* remove unnecessary nodes */ - if (get_opt_constant_folding() || - (iro == iro_Phi) || /* always optimize these nodes. */ - (iro == iro_Id) || - (iro == iro_Proj) || - (iro == iro_Block) ) /* Flags tested local. */ - n = equivalent_node (n); + if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv)) + set_Const_type(nw, old_tp); + DBG_OPT_CSTEVAL(oldn, nw); + return nw; + } + } + } - optimize_preds(n); /* do node specific optimizations of nodes predecessors. */ + /* remove unnecessary nodes */ + if (get_opt_constant_folding() || + (iro == iro_Phi) || /* always optimize these nodes. */ + (iro == iro_Id) || + (iro == iro_Proj) || + (iro == iro_Block) ) /* Flags tested local. */ + n = equivalent_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 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); + optimize_preds(n); /* do node specific optimizations of nodes predecessors. */ - if (n != oldn) { - edges_node_deleted(oldn, current_ir_graph); + /* Common Subexpression Elimination. + * + * Checks whether n is already available. + * The block input is used to distinguish different subexpressions. Right + * 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); - /* We found an existing, better node, so we can deallocate the old node. */ - obstack_free (current_ir_graph->obst, oldn); + if (n != oldn) { + edges_node_deleted(oldn, current_ir_graph); - return n; - } + /* We found an existing, better node, so we can deallocate the old node. */ + irg_kill_node(current_ir_graph, oldn); + return n; + } - /* Some more constant expression evaluation that does not allow to - free the node. */ - iro = get_irn_opcode(n); - if (get_opt_constant_folding() || - (iro == iro_Cond) || - (iro == iro_Proj) || - (iro == iro_Sel)) /* Flags tested local. */ - n = transform_node (n); + /* Some more constant expression evaluation that does not allow to + free the node. */ + iro = get_irn_opcode(n); + if (get_opt_constant_folding() || + (iro == iro_Cond) || + (iro == iro_Proj) || + (iro == iro_Sel)) /* Flags tested local. */ + n = transform_node (n); - /* Remove nodes with dead (Bad) input. - Run always for transformation induced Bads. */ - n = gigo (n); + /* Remove nodes with dead (Bad) input. + Run always for transformation induced Bads. */ + n = gigo (n); - /* Now we have a legal, useful node. Enter it in hash table for cse */ - if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) { - n = identify_remember (current_ir_graph->value_table, n); - } + /* Now we have a legal, useful node. Enter it in hash table for CSE */ + if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) { + n = identify_remember (current_ir_graph->value_table, n); + } - return n; + return n; } @@ -3210,41 +3718,35 @@ 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; opcode iro = get_irn_opcode(n); - type *old_tp = get_irn_type(n); - { - int i, arity = get_irn_arity(n); - for (i = 0; i < arity && !old_tp; ++i) - old_tp = get_irn_type(get_irn_n(n, i)); - } - if (!get_opt_optimize() && (get_irn_op(n) != op_Phi)) return n; - /* if not optimize return n */ - if (n == NULL) { - assert(0); - /* Here this is possible. Why? */ - return n; - } - /* constant expression evaluation / constant folding */ if (get_opt_constant_folding()) { - /* constants can not be evaluated */ - if (iro != iro_Const) { + /* neither constants nor Tuple values can be evaluated */ + if (iro != iro_Const && get_irn_mode(n) != mode_T) { /* try to evaluate */ tv = computed_value(n); - if ((get_irn_mode(n) != mode_T) && (tv != tarval_bad)) { + if (tv != tarval_bad) { /* evaluation was successful -- replace the node. */ - n = new_Const (get_tarval_mode (tv), tv); + ir_type *old_tp = get_irn_type(n); + int i, arity = get_irn_arity(n); + + /* + * Try to recover the type of the new expression. + */ + for (i = 0; i < arity && !old_tp; ++i) + old_tp = get_irn_type(get_irn_n(n, i)); + + n = new_Const(get_tarval_mode(tv), tv); - if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv)) - set_Const_type(n, old_tp); + if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv)) + set_Const_type(n, old_tp); DBG_OPT_CSTEVAL(oldn, n); return n; @@ -3258,7 +3760,7 @@ optimize_in_place_2 (ir_node *n) (iro == iro_Id) || /* ... */ (iro == iro_Proj) || /* ... */ (iro == iro_Block) ) /* Flags tested local. */ - n = equivalent_node (n); + n = equivalent_node(n); optimize_preds(n); /* do node specific optimizations of nodes predecessors. */ @@ -3268,7 +3770,7 @@ optimize_in_place_2 (ir_node *n) 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); + n = identify(current_ir_graph->value_table, n); } /* Some more constant expression evaluation. */ @@ -3277,11 +3779,11 @@ optimize_in_place_2 (ir_node *n) (iro == iro_Cond) || (iro == iro_Proj) || (iro == iro_Sel)) /* Flags tested local. */ - n = transform_node (n); + n = transform_node(n); /* Remove nodes with dead (Bad) input. Run always for transformation induced Bads. */ - n = gigo (n); + n = gigo(n); /* Now we can verify the node, as it has no dead inputs any more. */ irn_vrfy(n); @@ -3290,7 +3792,7 @@ optimize_in_place_2 (ir_node *n) Blocks should be unique anyways. (Except the successor of start: is cse with the start block!) */ if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) - n = identify_remember (current_ir_graph->value_table, n); + n = identify_remember(current_ir_graph->value_table, n); return n; } @@ -3298,8 +3800,7 @@ 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); @@ -3309,23 +3810,24 @@ 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); + 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 op; + return ops; }