X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firopt.c;h=2d222b5240ee3bf457368e5b070d130174efcf33;hb=cb91bddc9cacdab7c28e4336847bd3dc248aa549;hp=3dab4dd39cafe20893a7b666732aa16c44ecf806;hpb=59d73e15451ba87ac676c0d3b67e06a171db1e8e;p=libfirm diff --git a/ir/ir/iropt.c b/ir/ir/iropt.c index 3dab4dd39..2d222b524 100644 --- a/ir/ir/iropt.c +++ b/ir/ir/iropt.c @@ -53,16 +53,22 @@ value_of (ir_node *n) return tarval_bad; } +/** + * return the value of a Constant + */ static tarval *computed_value_Const(ir_node *n) { return get_Const_tarval(n); } +/** + * return the value of a 'sizeof' SymConst + */ static tarval *computed_value_SymConst(ir_node *n) { - if ((get_SymConst_kind(n) ==symconst_size) && + if ((get_SymConst_kind(n) == symconst_size) && (get_type_state(get_SymConst_type(n))) == layout_fixed) - return new_tarval_from_long(get_type_size_bytes(get_SymConst_type(n)), mode_Is); + return new_tarval_from_long(get_type_size_bytes(get_SymConst_type(n)), get_irn_mode(n)); return tarval_bad; } @@ -207,8 +213,8 @@ static tarval *computed_value_And(ir_node *n) } else { tarval *v; - if ( (tarval_classify ((v = computed_value (a))) == TV_CLASSIFY_NULL) - || (tarval_classify ((v = computed_value (b))) == TV_CLASSIFY_NULL)) { + if ( (classify_tarval ((v = computed_value (a))) == TV_CLASSIFY_NULL) + || (classify_tarval ((v = computed_value (b))) == TV_CLASSIFY_NULL)) { return v; } } @@ -227,8 +233,8 @@ static tarval *computed_value_Or(ir_node *n) return tarval_or (ta, tb); } else { tarval *v; - if ( (tarval_classify ((v = computed_value (a))) == TV_CLASSIFY_ALL_ONE) - || (tarval_classify ((v = computed_value (b))) == TV_CLASSIFY_ALL_ONE)) { + if ( (classify_tarval ((v = computed_value (a))) == TV_CLASSIFY_ALL_ONE) + || (classify_tarval ((v = computed_value (b))) == TV_CLASSIFY_ALL_ONE)) { return v; } } @@ -311,7 +317,7 @@ static tarval *computed_value_Rot(ir_node *n) tarval *tb = value_of(b); if ((ta != tarval_bad) && (tb != tarval_bad)) { - /* return tarval_rot (ta, tb); */ + return tarval_rot (ta, tb); } return tarval_bad; } @@ -346,12 +352,13 @@ static tarval *computed_value_Proj(ir_node *n) 3. The predecessors are Allocs or void* constants. Allocs never return NULL, they raise an exception. Therefore we can predict the Cmp result. */ - if (get_irn_op(a) == op_Cmp) { + switch (get_irn_opcode(a)) { + case iro_Cmp: aa = get_Cmp_left(a); ab = get_Cmp_right(a); if (aa == ab) { /* 1.: */ - /* This is a tric with the bits used for encoding the Cmp + /* This is a trick with the bits used for encoding the Cmp Proj numbers, the following statement is not the same: return new_tarval_from_long ((get_Proj_proj(n) == Eq), mode_b) */ return new_tarval_from_long ((get_Proj_proj(n) & Eq), mode_b); @@ -366,8 +373,8 @@ static tarval *computed_value_Proj(ir_node *n) return new_tarval_from_long (get_Proj_proj(n) & flags, mode_b); } } else { /* check for 3.: */ - ir_node *aaa = skip_nop(skip_Proj(aa)); - ir_node *aba = skip_nop(skip_Proj(ab)); + ir_node *aaa = skip_Id(skip_Proj(aa)); + ir_node *aba = skip_Id(skip_Proj(ab)); if ( ( (/* aa is ProjP and aaa is Alloc */ (get_irn_op(aa) == op_Proj) @@ -393,18 +400,54 @@ static tarval *computed_value_Proj(ir_node *n) return new_tarval_from_long (get_Proj_proj(n) & Ne, mode_b); } } - } else if (get_irn_op(a) == op_DivMod) { + break; + + case iro_DivMod: + { tarval *tb = value_of(b = get_DivMod_right(a)); tarval *ta = value_of(a = get_DivMod_left(a)); if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) { if (tb == get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */ return tarval_bad; - if (get_Proj_proj(n)== 0) /* Div */ + if (get_Proj_proj(n)== pn_DivMod_res_div) return tarval_div(ta, tb); - else /* Mod */ + else if (get_Proj_proj(n)== pn_DivMod_res_mod) return tarval_mod(ta, tb); } + break; + } + + case iro_Div: + { + tarval *tb = value_of(b = get_Div_right(a)); + tarval *ta = value_of(a = get_Div_left(a)); + + if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) { + if (tb == get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */ + return tarval_bad; + if (get_Proj_proj(n)== pn_Div_res) + return tarval_div(ta, tb); + } + break; + } + + case iro_Mod: + { + tarval *tb = value_of(b = get_Mod_right(a)); + tarval *ta = value_of(a = get_Mod_left(a)); + + if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) { + if (tb == get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */ + return tarval_bad; + if (get_Proj_proj(n)== pn_Mod_res) + return tarval_mod(ta, tb); + } + break; + } + + default: + return tarval_bad; } return tarval_bad; } @@ -412,6 +455,8 @@ static tarval *computed_value_Proj(ir_node *n) /** * If the parameter n can be computed, return its value, else tarval_bad. * Performs constant folding. + * + * @param n The node this should be evaluated */ tarval *computed_value(ir_node *n) { @@ -482,7 +527,7 @@ static ir_node *equivalent_node_Block(ir_node *n) { ir_node *oldn = n; - /* The Block constructor does not call optimize, but mature_block + /* The Block constructor does not call optimize, but mature_immBlock calls the optimization. */ assert(get_Block_matured(n)); @@ -494,7 +539,7 @@ static ir_node *equivalent_node_Block(ir_node *n) Remaining Phi nodes are just Ids. */ if ((get_Block_n_cfgpreds(n) == 1) && (get_irn_op(get_Block_cfgpred(n, 0)) == op_Jmp)) { - ir_node *predblock = get_nodes_Block(get_Block_cfgpred(n, 0)); + ir_node *predblock = get_nodes_block(get_Block_cfgpred(n, 0)); if (predblock == oldn) { /* Jmp jumps into the block it is in -- deal self cycle. */ n = new_Bad(); DBG_OPT_DEAD; @@ -504,7 +549,7 @@ static ir_node *equivalent_node_Block(ir_node *n) } else if ((get_Block_n_cfgpreds(n) == 1) && (get_irn_op(skip_Proj(get_Block_cfgpred(n, 0))) == op_Cond)) { - ir_node *predblock = get_nodes_Block(get_Block_cfgpred(n, 0)); + ir_node *predblock = get_nodes_block(get_Block_cfgpred(n, 0)); if (predblock == oldn) { /* Jmp jumps into the block it is in -- deal self cycle. */ n = new_Bad(); DBG_OPT_DEAD; @@ -524,7 +569,7 @@ static ir_node *equivalent_node_Block(ir_node *n) (get_irn_mode(get_Cond_selector(get_Proj_pred(a))) == mode_b)) { /* Also a single entry Block following a single exit Block. Phis have twice the same operand and will be optimized away. */ - n = get_nodes_Block(a); DBG_OPT_IFSIM; + n = get_nodes_block(a); DBG_OPT_IFSIM; } } else if (get_opt_unreachable_code() && (n != current_ir_graph->start_block) && @@ -551,7 +596,7 @@ static ir_node *equivalent_node_Jmp(ir_node *n) { /* GL: Why not same for op_Raise?? */ /* unreachable code elimination */ - if (is_Bad(get_nodes_Block(n))) + if (is_Bad(get_nodes_block(n))) n = new_Bad(); return n; @@ -607,7 +652,7 @@ static ir_node *equivalent_node_neutral_zero(ir_node *n) /* If this predecessors constant value is zero, the operation is unnecessary. Remove it: */ - if (tarval_classify (tv) == TV_CLASSIFY_NULL) { + if (classify_tarval (tv) == TV_CLASSIFY_NULL) { n = on; DBG_OPT_ALGSIM1; } @@ -628,7 +673,7 @@ static ir_node *equivalent_node_left_zero(ir_node *n) ir_node *a = get_binop_left(n); ir_node *b = get_binop_right(n); - if (tarval_classify(computed_value(b)) == TV_CLASSIFY_NULL) { + if (classify_tarval(computed_value(b)) == TV_CLASSIFY_NULL) { n = a; DBG_OPT_ALGSIM1; } @@ -647,10 +692,11 @@ static ir_node *equivalent_node_left_zero(ir_node *n) static ir_node *equivalent_node_symmetric_unop(ir_node *n) { ir_node *oldn = n; + ir_node *pred = get_unop_op(n); /* optimize symmetric unop */ - if (get_irn_op(get_unop_op(n)) == get_irn_op(n)) { - n = get_unop_op(get_unop_op(n)); DBG_OPT_ALGSIM2; + if (get_irn_op(pred) == get_irn_op(n)) { + n = get_unop_op(pred); DBG_OPT_ALGSIM2; } return n; } @@ -673,9 +719,9 @@ static ir_node *equivalent_node_Mul(ir_node *n) ir_node *b = get_Mul_right(n); /* Mul is commutative and has again an other neutral element. */ - if (tarval_classify (computed_value (a)) == TV_CLASSIFY_ONE) { + if (classify_tarval (computed_value (a)) == TV_CLASSIFY_ONE) { n = b; DBG_OPT_ALGSIM1; - } else if (tarval_classify (computed_value (b)) == TV_CLASSIFY_ONE) { + } else if (classify_tarval (computed_value (b)) == TV_CLASSIFY_ONE) { n = a; DBG_OPT_ALGSIM1; } return n; @@ -690,7 +736,7 @@ static ir_node *equivalent_node_Div(ir_node *n) ir_node *b = get_Div_right(n); /* Div is not commutative. */ - if (tarval_classify(computed_value(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */ + if (classify_tarval(computed_value(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */ /* Turn Div into a tuple (mem, bad, a) */ ir_node *mem = get_Div_mem(n); turn_into_tuple(n, 3); @@ -713,9 +759,9 @@ static ir_node *equivalent_node_And(ir_node *n) if (a == b) { n = a; /* And has it's own neutral element */ - } else if (tarval_classify(computed_value(a)) == TV_CLASSIFY_ALL_ONE) { + } else if (classify_tarval(computed_value(a)) == TV_CLASSIFY_ALL_ONE) { n = b; - } else if (tarval_classify(computed_value(b)) == TV_CLASSIFY_ALL_ONE) { + } else if (classify_tarval(computed_value(b)) == TV_CLASSIFY_ALL_ONE) { n = a; } if (n != oldn) DBG_OPT_ALGSIM1; @@ -851,65 +897,8 @@ static ir_node *equivalent_node_Phi(ir_node *n) } /** - * Optimize Loads after Store. - * - * @todo FAILS for volatile entities - */ -static ir_node *equivalent_node_Load(ir_node *n) -{ - ir_node *oldn = n; - - /* remove unnecessary Load. */ - ir_node *a = skip_Proj(get_Load_mem(n)); - ir_node *b = get_Load_ptr(n); - ir_node *c; - - /* TODO: check for volatile */ - if (get_irn_op(a) == op_Store && get_Store_ptr(a) == b) { - /* We load immediately after a store -- a read after write. */ - ir_node *mem = get_Load_mem(n); - - c = get_Store_value(a); - turn_into_tuple(n, 3); - set_Tuple_pred(n, pn_Load_M, mem); - set_Tuple_pred(n, pn_Load_res, c); - set_Tuple_pred(n, pn_Load_X_except, new_Bad()); DBG_OPT_RAW; - } - return n; -} - -/** - * Optimize store after store and load atfter store. - * - * @todo FAILS for volatile entities + * optimize Proj(Tuple) and gigo for ProjX in Bad block */ -static ir_node *equivalent_node_Store(ir_node *n) -{ - ir_node *oldn = n; - - /* remove unnecessary store. */ - ir_node *a = skip_Proj(get_Store_mem(n)); - ir_node *b = get_Store_ptr(n); - ir_node *c = skip_Proj(get_Store_value(n)); - - if (get_irn_op(a) == op_Store - && get_Store_ptr(a) == b - && skip_Proj(get_Store_value(a)) == c) { - /* We have twice exactly the same store -- a write after write. */ - n = a; DBG_OPT_WAW; - } else if (get_irn_op(c) == op_Load - && (a == c || skip_Proj(get_Load_mem(c)) == a) - && get_Load_ptr(c) == b ) { - /* We just loaded the value from the same memory, i.e., the store - doesn't change the memory -- a write after read. */ - a = get_Store_mem(n); - turn_into_tuple(n, 2); - set_Tuple_pred(n, pn_Store_M, a); - set_Tuple_pred(n, pn_Store_X_except, new_Bad()); DBG_OPT_WAR; - } - return n; -} - static ir_node *equivalent_node_Proj(ir_node *n) { ir_node *oldn = n; @@ -925,7 +914,7 @@ static ir_node *equivalent_node_Proj(ir_node *n) n = new_Bad(); } } else if (get_irn_mode(n) == mode_X && - is_Bad(get_nodes_Block(n))) { + is_Bad(get_nodes_block(n))) { /* Remove dead control flow -- early gigo. */ n = new_Bad(); } @@ -943,12 +932,6 @@ static ir_node *equivalent_node_Id(ir_node *n) return n; } -/* -case iro_Mod, Quot, DivMod - DivMod allocates new nodes --> it's treated in transform node. - What about Quot, DivMod? -*/ - /** * 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 @@ -993,8 +976,6 @@ static ir_op *firm_set_default_equivalent_node(ir_op *op) CASE(And); CASE(Conv); CASE(Phi); -// CASE(Load); /* dangerous */ -// CASE(Store); /* dangerous, see todo */ CASE(Proj); CASE(Id); default: @@ -1114,7 +1095,7 @@ static ir_node *transform_node_DivMod(ir_node *n) set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */ set_Tuple_pred(n, pn_DivMod_res_div, a); set_Tuple_pred(n, pn_DivMod_res_mod, b); - assert(get_nodes_Block(n)); + assert(get_nodes_block(n)); } return n; @@ -1133,7 +1114,7 @@ static ir_node *transform_node_Cond(ir_node *n) (get_opt_unreachable_code())) { /* It's a boolean Cond, branching on a boolean constant. Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */ - jmp = new_r_Jmp(current_ir_graph, get_nodes_Block(n)); + jmp = new_r_Jmp(current_ir_graph, get_nodes_block(n)); turn_into_tuple(n, 2); if (ta == tarval_b_true) { set_Tuple_pred(n, pn_Cond_false, new_Bad()); @@ -1143,7 +1124,7 @@ static ir_node *transform_node_Cond(ir_node *n) set_Tuple_pred(n, pn_Cond_true, new_Bad()); } /* We might generate an endless loop, so keep it alive. */ - add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_Block(n)); + add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n)); } else if ((ta != tarval_bad) && (get_irn_mode(a) == mode_Iu) && (get_Cond_kind(n) == dense) && @@ -1152,21 +1133,21 @@ static ir_node *transform_node_Cond(ir_node *n) Also this tuple might get really big... I generate the Jmp here, and remember it in link. Link is used when optimizing Proj. */ - set_irn_link(n, new_r_Jmp(current_ir_graph, get_nodes_Block(n))); + set_irn_link(n, new_r_Jmp(current_ir_graph, get_nodes_block(n))); /* We might generate an endless loop, so keep it alive. */ - add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_Block(n)); + add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n)); } else if ((get_irn_op(a) == op_Eor) && (get_irn_mode(a) == mode_b) - && (tarval_classify(computed_value(get_Eor_right(a))) == TV_CLASSIFY_ONE)) { + && (classify_tarval(computed_value(get_Eor_right(a))) == TV_CLASSIFY_ONE)) { /* The Eor is a negate. Generate a new Cond without the negate, simulate the negate by exchanging the results. */ - set_irn_link(n, new_r_Cond(current_ir_graph, get_nodes_Block(n), + set_irn_link(n, new_r_Cond(current_ir_graph, get_nodes_block(n), get_Eor_left(a))); } else if ((get_irn_op(a) == op_Not) && (get_irn_mode(a) == mode_b)) { /* A Not before the Cond. Generate a new Cond without the Not, simulate the Not by exchanging the results. */ - set_irn_link(n, new_r_Cond(current_ir_graph, get_nodes_Block(n), + set_irn_link(n, new_r_Cond(current_ir_graph, get_nodes_block(n), get_Not_op(a))); } return n; @@ -1180,16 +1161,16 @@ static ir_node *transform_node_Eor(ir_node *n) if ((get_irn_mode(n) == mode_b) && (get_irn_op(a) == op_Proj) && (get_irn_mode(a) == mode_b) - && (tarval_classify (computed_value (b)) == TV_CLASSIFY_ONE) + && (classify_tarval (computed_value (b)) == TV_CLASSIFY_ONE) && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) /* The Eor negates a Cmp. The Cmp has the negated result anyways! */ - n = new_r_Proj(current_ir_graph, get_nodes_Block(n), get_Proj_pred(a), + n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a), mode_b, get_negated_pnc(get_Proj_proj(a))); else if ((get_irn_mode(n) == mode_b) - && (tarval_classify (computed_value (b)) == TV_CLASSIFY_ONE)) + && (classify_tarval (computed_value (b)) == TV_CLASSIFY_ONE)) /* The Eor is a Not. Replace it by a Not. */ /* ????!!!Extend to bitfield 1111111. */ - n = new_r_Not(current_ir_graph, get_nodes_Block(n), a, mode_b); + n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode_b); return n; } @@ -1206,7 +1187,7 @@ static ir_node *transform_node_Not(ir_node *n) && (get_irn_mode(a) == mode_b) && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) /* We negate a Cmp. The Cmp has the negated result anyways! */ - n = new_r_Proj(current_ir_graph, get_nodes_Block(n), get_Proj_pred(a), + n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a), mode_b, get_negated_pnc(get_Proj_proj(a))); return n; @@ -1214,7 +1195,8 @@ static ir_node *transform_node_Not(ir_node *n) /** * Transform a Div/Mod/DivMod with a non-zero constant. Must be - * done here to avoid that this optimization runs more than once... + * done here instead of equivalent node because in creates new + * nodes. */ static ir_node *transform_node_Proj(ir_node *proj) { @@ -1227,7 +1209,7 @@ static ir_node *transform_node_Proj(ir_node *proj) b = get_Div_right(n); tb = computed_value(b); - if (tb != tarval_bad && tarval_classify(tb) != TV_CLASSIFY_NULL) { /* div(x, c) && c != 0 */ + if (tb != tarval_bad && classify_tarval(tb) != TV_CLASSIFY_NULL) { /* div(x, c) && c != 0 */ ir_node *div, *proj; ir_node *a = get_Div_left(n); ir_node *mem = get_Div_mem(n); @@ -1236,9 +1218,9 @@ static ir_node *transform_node_Proj(ir_node *proj) set_optimize(0); { div = new_rd_Div(get_irn_dbg_info(n), current_ir_graph, - get_nodes_Block(n), get_irg_initial_mem(current_ir_graph), a, b); + get_nodes_block(n), get_irg_initial_mem(current_ir_graph), a, b); - proj = new_r_Proj(current_ir_graph, get_nodes_Block(n), div, get_irn_mode(a), pn_Div_res); + proj = new_r_Proj(current_ir_graph, get_nodes_block(n), div, get_irn_mode(a), pn_Div_res); } set_optimize(rem); @@ -1252,7 +1234,7 @@ static ir_node *transform_node_Proj(ir_node *proj) b = get_Mod_right(n); tb = computed_value(b); - if (tb != tarval_bad && tarval_classify(tb) != TV_CLASSIFY_NULL) { /* mod(x, c) && c != 0 */ + if (tb != tarval_bad && classify_tarval(tb) != TV_CLASSIFY_NULL) { /* mod(x, c) && c != 0 */ ir_node *mod, *proj; ir_node *a = get_Mod_left(n); ir_node *mem = get_Mod_mem(n); @@ -1261,9 +1243,9 @@ static ir_node *transform_node_Proj(ir_node *proj) set_optimize(0); { mod = new_rd_Mod(get_irn_dbg_info(n), current_ir_graph, - get_nodes_Block(n), get_irg_initial_mem(current_ir_graph), a, b); + get_nodes_block(n), get_irg_initial_mem(current_ir_graph), a, b); - proj = new_r_Proj(current_ir_graph, get_nodes_Block(n), mod, get_irn_mode(a), pn_Mod_res); + proj = new_r_Proj(current_ir_graph, get_nodes_block(n), mod, get_irn_mode(a), pn_Mod_res); } set_optimize(rem); @@ -1277,7 +1259,7 @@ static ir_node *transform_node_Proj(ir_node *proj) b = get_DivMod_right(n); tb = computed_value(b); - if (tb != tarval_bad && tarval_classify(tb) != TV_CLASSIFY_NULL) { /* DivMod(x, c) && c != 0 */ + if (tb != tarval_bad && classify_tarval(tb) != TV_CLASSIFY_NULL) { /* DivMod(x, c) && c != 0 */ ir_node *div_mod, *proj_div, *proj_mod; ir_node *a = get_Mod_left(n); ir_node *mem = get_Mod_mem(n); @@ -1286,10 +1268,10 @@ static ir_node *transform_node_Proj(ir_node *proj) set_optimize(0); { div_mod = new_rd_DivMod(get_irn_dbg_info(n), current_ir_graph, - get_nodes_Block(n), get_irg_initial_mem(current_ir_graph), a, b); + get_nodes_block(n), get_irg_initial_mem(current_ir_graph), a, b); - proj_div = new_r_Proj(current_ir_graph, get_nodes_Block(n), div_mod, get_irn_mode(a), pn_DivMod_res_div); - proj_mod = new_r_Proj(current_ir_graph, get_nodes_Block(n), div_mod, get_irn_mode(a), pn_DivMod_res_mod); + proj_div = new_r_Proj(current_ir_graph, get_nodes_block(n), div_mod, get_irn_mode(a), pn_DivMod_res_div); + proj_mod = new_r_Proj(current_ir_graph, get_nodes_block(n), div_mod, get_irn_mode(a), pn_DivMod_res_mod); } set_optimize(rem); @@ -1323,6 +1305,10 @@ static ir_node *transform_node_Proj(ir_node *proj) } return proj; + case iro_Tuple: + /* should not happen, but if it doest will optimize */ + break; + default: /* do nothing */ return proj; @@ -1332,31 +1318,6 @@ static ir_node *transform_node_Proj(ir_node *proj) return equivalent_node_Proj(proj); } -/** - * Transform a Store before a Store to the same address... - * Both nodes must be in the same block. - * - * @todo Check for volatile! Moreover, what if the first store - * has a exception handler while the other has not? - */ -static ir_node *transform_node_Store(ir_node *store) -{ - ir_node *pred = skip_Proj(get_Store_mem(store)); - ir_node *ptr = get_Store_ptr(store); - - if (get_irn_op(pred) == op_Store && - get_Store_ptr(pred) == ptr && - get_nodes_block(pred) == get_nodes_block(store)) { - /* the Store n is useless, as it is overwritten by the store store */ - ir_node *mem = get_Store_mem(pred); - - turn_into_tuple(pred, 2); - set_Tuple_pred(pred, pn_Store_M, mem); - set_Tuple_pred(pred, pn_Store_X_except, new_Bad()); - } - return store; -} - /** * returns the operands of a commutative bin-op, if one operand is * a const, it is returned as the second one. @@ -1426,7 +1387,7 @@ static ir_node *transform_node_Or(ir_node *or) tv4 = get_Const_tarval(c4); tv = tarval_or(tv4, tv2); - if (tarval_classify(tv) != TV_CLASSIFY_ALL_ONE) { + if (classify_tarval(tv) != TV_CLASSIFY_ALL_ONE) { /* have at least one 0 at the same bit position */ return or; } @@ -1489,7 +1450,6 @@ static ir_op *firm_set_default_transform_node(ir_op *op) CASE(Eor); CASE(Not); CASE(Proj); -// CASE(Store); /* dangerous, see todo */ CASE(Or); default: op->transform_node = NULL; @@ -1568,6 +1528,23 @@ static int node_cmp_attr_Cast(ir_node *a, ir_node *b) return get_Cast_type(a) != get_Cast_type(b); } +static int node_cmp_attr_Load(ir_node *a, ir_node *b) +{ + if (get_Load_volatility(a) == volatility_is_volatile || + get_Load_volatility(b) == volatility_is_volatile) + /* NEVER do CSE on volatile Loads */ + return 1; + + return get_Load_mode(a) != get_Load_mode(b); +} + +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_Load_volatility(b) == volatility_is_volatile); +} + /** * set the default node attribute compare operation */ @@ -1590,6 +1567,8 @@ static ir_op *firm_set_default_node_cmp_attr(ir_op *op) CASE(Sel); CASE(Phi); CASE(Cast); + CASE(Load); + CASE(Store); default: op->node_cmp_attr = NULL; } @@ -1621,8 +1600,8 @@ vt_cmp (const void *elt, const void *key) if (irn_arity_a != get_irn_arity(b)) return 1; - /* for block-local cse and pinned nodes: */ - if (!get_opt_global_cse() || (get_op_pinned(get_irn_op(a)) == pinned)) { + /* for block-local cse and op_pin_state_pinned nodes: */ + if (!get_opt_global_cse() || (get_op_pinned(get_irn_op(a)) == op_pin_state_pinned)) { if (get_irn_n(a, -1) != get_irn_n(b, -1)) return 1; } @@ -1724,7 +1703,7 @@ identify (pset *value_table, ir_node *n) } /** - * During construction we set the pinned flag in the graph right when the + * During construction we set the op_pin_state_pinned flag in the graph right when the * optimization is performed. The flag turning on procedure global cse could * be changed between two allocations. This way we are safe. */ @@ -1733,7 +1712,7 @@ identify_cons (pset *value_table, ir_node *n) { ir_node *old = n; n = identify(value_table, n); if (get_irn_n(old, -1) != get_irn_n(n, -1)) - set_irg_pinned(current_ir_graph, floats); + set_irg_pinned(current_ir_graph, op_pin_state_floats); return n; } @@ -1836,7 +1815,7 @@ optimize_node (ir_node *n) tv = computed_value (n); if ((get_irn_mode(n) != mode_T) && (tv != tarval_bad)) { /* - * we MUST copy the node here temparary, because it's still needed + * we MUST copy the node here temporary, because it's still needed * for DBG_OPT_ALGSIM0 */ int node_size = offsetof(ir_node, attr) + n->op->attr_size; @@ -1867,7 +1846,7 @@ optimize_node (ir_node *n) /** common subexpression elimination **/ /* Checks whether n is already available. */ /* The block input is used to distinguish different subexpressions. Right - now all nodes are pinned to blocks, i.e., the cse only finds common + now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common subexpressions within a block. */ if (get_opt_cse()) n = identify_cons (current_ir_graph->value_table, n); @@ -1938,7 +1917,6 @@ optimize_in_place_2 (ir_node *n) } /* remove unnecessary nodes */ - /*if (get_opt_constant_folding()) */ if (get_opt_constant_folding() || (iro == iro_Phi) || /* always optimize these nodes. */ (iro == iro_Id) || /* ... */ @@ -1951,7 +1929,7 @@ optimize_in_place_2 (ir_node *n) /** common subexpression elimination **/ /* Checks whether n is already available. */ /* The block input is used to distinguish different subexpressions. Right - now all nodes are pinned to blocks, i.e., the cse only finds common + now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common subexpressions within a block. */ if (get_opt_cse()) { n = identify (current_ir_graph->value_table, n); @@ -1990,9 +1968,10 @@ optimize_in_place (ir_node *n) assert(get_irg_phase_state(current_ir_graph) != phase_building); if (get_opt_global_cse()) - set_irg_pinned(current_ir_graph, floats); + set_irg_pinned(current_ir_graph, op_pin_state_floats); if (get_irg_outs_state(current_ir_graph) == outs_consistent) set_irg_outs_inconsistent(current_ir_graph); + /* Maybe we could also test whether optimizing the node can change the control graph. */ if (get_irg_dom_state(current_ir_graph) == dom_consistent)