X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firopt.c;h=a452d25192971c4e6da040a9c491581f34d6f1c7;hb=62f1d1d20663ad5ec9b8e7d2c2ee3d23bf014b62;hp=6e96bd751b8c9290c9e852c2679282e5a8e6368b;hpb=a142727aca4ef56cf754045fc0129b6d1b840a63;p=libfirm diff --git a/ir/ir/iropt.c b/ir/ir/iropt.c index 6e96bd751..a452d2519 100644 --- a/ir/ir/iropt.c +++ b/ir/ir/iropt.c @@ -27,9 +27,7 @@ # include "config.h" #endif -#ifdef HAVE_STRING_H #include -#endif #include "irnode_t.h" #include "irgraph_t.h" @@ -47,8 +45,8 @@ #include "irarch.h" #include "hashptr.h" #include "archop.h" -#include "opt_polymorphy.h" #include "opt_confirms.h" +#include "opt_polymorphy.h" #include "irtools.h" #include "xmalloc.h" @@ -147,8 +145,7 @@ static tarval *computed_value_Carry(ir_node *n) { 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)) + if (tarval_is_null(ta) || tarval_is_null(tb)) return get_mode_null(m); } return tarval_bad; @@ -168,7 +165,7 @@ static tarval *computed_value_Borrow(ir_node *n) { 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) { + } else if (tarval_is_null(ta)) { return get_mode_null(m); } return tarval_bad; @@ -193,19 +190,25 @@ static tarval *computed_value_Minus(ir_node *n) { static tarval *computed_value_Mul(ir_node *n) { ir_node *a = get_Mul_left(n); ir_node *b = get_Mul_right(n); + ir_mode *mode; tarval *ta = value_of(a); tarval *tb = value_of(b); - if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) { + mode = get_irn_mode(n); + if (mode != get_irn_mode(a)) { + /* n * n = 2n bit multiplication */ + ta = tarval_convert_to(ta, mode); + tb = tarval_convert_to(tb, mode); + } + + if (ta != tarval_bad && tb != tarval_bad) { return tarval_mul(ta, tb); } else { - /* a*0 = 0 or 0*b = 0: - calls computed_value recursive and returns the 0 with proper - mode. */ - if ((ta != tarval_bad) && (ta == get_mode_null(get_tarval_mode(ta)))) + /* a*0 = 0 or 0*b = 0 */ + if (ta == get_mode_null(mode)) return ta; - if ((tb != tarval_bad) && (tb == get_mode_null(get_tarval_mode(tb)))) + if (tb == get_mode_null(mode)) return tb; } return tarval_bad; @@ -306,12 +309,8 @@ static tarval *computed_value_And(ir_node *n) { if ((ta != tarval_bad) && (tb != tarval_bad)) { return tarval_and (ta, tb); } else { - tarval *v; - - if ( (classify_tarval ((v = ta)) == TV_CLASSIFY_NULL) - || (classify_tarval ((v = tb)) == TV_CLASSIFY_NULL)) { - return v; - } + if (tarval_is_null(ta)) return ta; + if (tarval_is_null(tb)) return tb; } return tarval_bad; } /* computed_value_And */ @@ -330,11 +329,8 @@ static tarval *computed_value_Or(ir_node *n) { if ((ta != tarval_bad) && (tb != tarval_bad)) { return tarval_or (ta, tb); } else { - tarval *v; - if ( (classify_tarval ((v = ta)) == TV_CLASSIFY_ALL_ONE) - || (classify_tarval ((v = tb)) == TV_CLASSIFY_ALL_ONE)) { - return v; - } + if (tarval_is_all_one(ta)) return ta; + if (tarval_is_all_one(tb)) return tb; } return tarval_bad; } /* computed_value_Or */ @@ -538,25 +534,25 @@ static tarval *computed_value_Proj_Cmp(ir_node *n) { ir_node *aba = skip_Id(skip_Proj(ab)); if ( ( (/* aa is ProjP and aaa is Alloc */ - (get_irn_op(aa) == op_Proj) - && (mode_is_reference(get_irn_mode(aa))) - && (get_irn_op(aaa) == op_Alloc)) + is_Proj(aa) + && mode_is_reference(get_irn_mode(aa)) + && is_Alloc(aaa)) && ( (/* ab is NULL */ - (get_irn_op(ab) == op_Const) - && (mode_is_reference(get_irn_mode(ab))) - && (get_Const_tarval(ab) == get_mode_null(get_irn_mode(ab)))) + is_Const(ab) + && mode_is_reference(get_irn_mode(ab)) + && is_Const_null(ab)) || (/* ab is other Alloc */ - (get_irn_op(ab) == op_Proj) - && (mode_is_reference(get_irn_mode(ab))) - && (get_irn_op(aba) == op_Alloc) + is_Proj(ab) + && mode_is_reference(get_irn_mode(ab)) + && is_Alloc(aba) && (aaa != aba)))) || (/* aa is NULL and aba is Alloc */ - (get_irn_op(aa) == op_Const) - && (mode_is_reference(get_irn_mode(aa))) - && (get_Const_tarval(aa) == get_mode_null(get_irn_mode(aa))) - && (get_irn_op(ab) == op_Proj) - && (mode_is_reference(get_irn_mode(ab))) - && (get_irn_op(aba) == op_Alloc))) + is_Const(aa) + && mode_is_reference(get_irn_mode(aa)) + && is_Const_null(aa) + && is_Proj(ab) + && mode_is_reference(get_irn_mode(ab)) + && is_Alloc(aba))) /* 3.: */ return new_tarval_from_long(proj_nr & pn_Cmp_Ne, mode_b); } @@ -738,7 +734,7 @@ static ir_node *equivalent_node_Block(ir_node *n) This should be true, as the block is matured before optimize is called. But what about Phi-cycles with the Phi0/Id that could not be resolved? Remaining Phi nodes are just Ids. */ - if ((n_preds == 1) && (get_irn_op(get_Block_cfgpred(n, 0)) == op_Jmp)) { + if (n_preds == 1 && is_Jmp(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. */ @@ -748,8 +744,7 @@ static ir_node *equivalent_node_Block(ir_node *n) n = predblock; DBG_OPT_STG(oldn, n); } - } else if ((n_preds == 1) && - (get_irn_op(skip_Proj(get_Block_cfgpred(n, 0))) == op_Cond)) { + } else if (n_preds == 1 && is_Cond(skip_Proj(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. */ @@ -765,10 +760,10 @@ static ir_node *equivalent_node_Block(ir_node *n) ir_node *a = get_Block_cfgpred(n, 0); ir_node *b = get_Block_cfgpred(n, 1); - if ((get_irn_op(a) == op_Proj) && - (get_irn_op(b) == op_Proj) && + if (is_Proj(a) && + is_Proj(b) && (get_Proj_pred(a) == get_Proj_pred(b)) && - (get_irn_op(get_Proj_pred(a)) == op_Cond) && + is_Cond(get_Proj_pred(a)) && (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. */ @@ -855,12 +850,10 @@ static ir_node *equivalent_node_neutral_zero(ir_node *n) * 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) { - if (get_irn_mode(on) == get_irn_mode(n)) { - n = on; + if (tarval_is_null(tv) && get_irn_mode(on) == get_irn_mode(n)) { + n = on; - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0); - } + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0); } return n; @@ -896,7 +889,7 @@ static ir_node *equivalent_node_Add(ir_node *n) { left = get_Add_left(n); right = get_Add_right(n); - if (get_irn_op(left) == op_Sub) { + if (is_Sub(left)) { if (get_Sub_right(left) == right) { /* (a - x) + x */ @@ -907,7 +900,7 @@ static ir_node *equivalent_node_Add(ir_node *n) { } } } - if (get_irn_op(right) == op_Sub) { + if (is_Sub(right)) { if (get_Sub_right(right) == left) { /* x + (a - x) */ @@ -931,7 +924,7 @@ 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 (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) { + if (is_Const(b) && is_Const_null(b)) { n = a; DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0); @@ -955,40 +948,23 @@ static ir_node *equivalent_node_left_zero(ir_node *n) { */ static ir_node *equivalent_node_Sub(ir_node *n) { ir_node *oldn = n; - ir_node *a, *b; + ir_node *b; ir_mode *mode = get_irn_mode(n); /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */ if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic)) return n; - a = get_Sub_left(n); b = get_Sub_right(n); /* Beware: modes might be different */ - if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) { + if (is_Const(b) && is_Const_null(b)) { + ir_node *a = get_Sub_left(n); if (mode == get_irn_mode(a)) { n = a; DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0); } - } else if (get_irn_op(a) == op_Add) { - if (mode_wrap_around(mode)) { - ir_node *left = get_Add_left(a); - ir_node *right = get_Add_right(a); - - if (left == b) { - if (mode == get_irn_mode(right)) { - n = right; - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); - } - } else if (right == b) { - if (mode == get_irn_mode(left)) { - n = left; - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); - } - } - } } return n; } /* equivalent_node_Sub */ @@ -1009,7 +985,7 @@ static ir_node *equivalent_node_idempotent_unop(ir_node *n) { /* optimize symmetric unop */ if (get_irn_op(pred) == get_irn_op(n)) { n = get_unop_op(pred); - DBG_OPT_ALGSIM2(oldn, pred, n); + DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY); } return n; } /* equivalent_node_idempotent_unop */ @@ -1017,7 +993,7 @@ static ir_node *equivalent_node_idempotent_unop(ir_node *n) { /** Optimize 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 @@ -1027,15 +1003,19 @@ static ir_node *equivalent_node_idempotent_unop(ir_node *n) { static ir_node *equivalent_node_Mul(ir_node *n) { ir_node *oldn = n; ir_node *a = get_Mul_left(n); - ir_node *b = get_Mul_right(n); - /* 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, FS_OPT_NEUTRAL_1); - } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { - n = a; - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1); + /* we can handle here only the n * n = n bit cases */ + if (get_irn_mode(n) == get_irn_mode(a)) { + ir_node *b = get_Mul_right(n); + + /* Mul is commutative and has again an other neutral element. */ + if (is_Const(a) && is_Const_one(a)) { + n = b; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1); + } else if (is_Const(b) && is_Const_one(b)) { + n = a; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1); + } } return n; } /* equivalent_node_Mul */ @@ -1048,13 +1028,15 @@ static ir_node *equivalent_node_Div(ir_node *n) { ir_node *b = get_Div_right(n); /* Div is not commutative. */ - if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */ + if (is_Const(b) && is_Const_one(b)) { /* div(x, 1) == x */ /* Turn Div into a tuple (mem, bad, a) */ ir_node *mem = get_Div_mem(n); + ir_node *blk = get_irn_n(n, -1); 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); + set_Tuple_pred(n, pn_Div_M, mem); + set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk)); + set_Tuple_pred(n, pn_Div_X_except, new_Bad()); /* no exception */ + set_Tuple_pred(n, pn_Div_res, a); } return n; } /* equivalent_node_Div */ @@ -1067,13 +1049,15 @@ static ir_node *equivalent_node_Quot(ir_node *n) { ir_node *b = get_Quot_right(n); /* Div is not commutative. */ - if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* Quot(x, 1) == x */ - /* Turn Quot into a tuple (mem, bad, a) */ + if (is_Const(b) && is_Const_one(b)) { /* Quot(x, 1) == x */ + /* Turn Quot into a tuple (mem, jmp, bad, a) */ ir_node *mem = get_Quot_mem(n); + ir_node *blk = get_irn_n(n, -1); turn_into_tuple(n, pn_Quot_max); - set_Tuple_pred(n, pn_Quot_M, mem); - set_Tuple_pred(n, pn_Quot_X_except, new_Bad()); /* no exception */ - set_Tuple_pred(n, pn_Quot_res, a); + set_Tuple_pred(n, pn_Quot_M, mem); + set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk)); + set_Tuple_pred(n, pn_Quot_X_except, new_Bad()); /* no exception */ + set_Tuple_pred(n, pn_Quot_res, a); } return n; } /* equivalent_node_Quot */ @@ -1082,20 +1066,22 @@ static ir_node *equivalent_node_Quot(ir_node *n) { * Optimize a / 1 = a. */ static ir_node *equivalent_node_DivMod(ir_node *n) { - ir_node *a = get_DivMod_left(n); ir_node *b = get_DivMod_right(n); /* Div is not commutative. */ - if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */ - /* Turn DivMod into a tuple (mem, bad, a, 0) */ + if (is_Const(b) && is_Const_one(b)) { /* div(x, 1) == x */ + /* Turn DivMod into a tuple (mem, jmp, bad, a, 0) */ + ir_node *a = get_DivMod_left(n); ir_node *mem = get_Div_mem(n); - ir_mode *mode = get_irn_mode(b); + ir_node *blk = get_irn_n(n, -1); + ir_mode *mode = get_DivMod_resmode(n); 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, new_Const(mode, get_mode_null(mode))); + set_Tuple_pred(n, pn_DivMod_M, mem); + set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk)); + 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, new_Const(mode, get_mode_null(mode))); } return n; } /* equivalent_node_DivMod */ @@ -1112,10 +1098,10 @@ 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) { + } else if (is_Const(a) && is_Const_null(a)) { n = b; DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR); - } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) { + } else if (is_Const(b) && is_Const_null(b)) { n = a; DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR); } @@ -1124,7 +1110,7 @@ static ir_node *equivalent_node_Or(ir_node *n) { } /* equivalent_node_Or */ /** - * Optimize a & 0b1...1 = 0b1...1 & a = a & a = a. + * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a. */ static ir_node *equivalent_node_And(ir_node *n) { ir_node *oldn = n; @@ -1135,13 +1121,35 @@ 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) { + return n; + } + if (is_Const(a) && is_Const_all_one(a)) { n = b; DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND); - } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ALL_ONE) { + return n; + } + if (is_Const(b) && is_Const_all_one(b)) { n = a; DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND); + return n; + } + if (is_Or(a)) { + if (b == get_Or_left(a) || b == get_Or_right(a)) { + /* (a|X) & a */ + n = b; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND); + return n; + } + } + if (is_Or(b)) { + if (a == get_Or_left(b) || a == get_Or_right(b)) { + /* a & (a|X) */ + n = a; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND); + return n; + } } + return n; } /* equivalent_node_And */ @@ -1157,12 +1165,23 @@ static ir_node *equivalent_node_Conv(ir_node *n) { ir_mode *a_mode = get_irn_mode(a); if (n_mode == a_mode) { /* No Conv necessary */ - /* leave strict floating point Conv's */ - if (get_Conv_strict(n)) - return n; + if (get_Conv_strict(n)) { + /* special case: the predecessor might be a also a Conv */ + if (is_Conv(a)) { + if (! get_Conv_strict(a)) { + /* first one is not strict, kick it */ + set_Conv_op(n, get_Conv_op(a)); + return n; + } + /* else both are strict conv, second is superflous */ + } else { + /* leave strict floating point Conv's */ + return n; + } + } n = a; DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV); - } else if (get_irn_op(a) == op_Conv) { /* Conv(Conv(b)) */ + } else if (is_Conv(a)) { /* Conv(Conv(b)) */ ir_mode *b_mode; b = get_Conv_op(a); @@ -1173,7 +1192,7 @@ static ir_node *equivalent_node_Conv(ir_node *n) { if (n_mode == mode_b) { n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */ DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV); - } else if (mode_is_int(n_mode) || mode_is_character(n_mode)) { + } else if (mode_is_int(n_mode)) { if (smaller_mode(b_mode, a_mode)){ n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */ DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV); @@ -1323,29 +1342,46 @@ static ir_node *equivalent_node_Sync(ir_node *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) { - ir_node *oldn = n; - ir_node *a = get_Proj_pred(n); +static ir_node *equivalent_node_Proj(ir_node *proj) { + ir_node *oldn = proj; + ir_node *a = get_Proj_pred(proj); - if ( get_irn_op(a) == op_Tuple) { + if (is_Tuple(a)) { /* Remove the Tuple/Proj combination. */ - if ( get_Proj_proj(n) <= get_Tuple_n_preds(a) ) { - n = get_Tuple_pred(a, get_Proj_proj(n)); - DBG_OPT_TUPLE(oldn, a, n); + if ( get_Proj_proj(proj) <= get_Tuple_n_preds(a) ) { + proj = get_Tuple_pred(a, get_Proj_proj(proj)); + DBG_OPT_TUPLE(oldn, a, proj); } else { - assert(0); /* This should not happen! */ - n = new_Bad(); + /* This should not happen! */ + assert(! "found a Proj with higher number than Tuple predecessors"); + proj = new_Bad(); } - } else if (get_irn_mode(n) == mode_X) { - if (is_Block_dead(get_nodes_block(skip_Proj(n)))) { + } else if (get_irn_mode(proj) == mode_X) { + if (is_Block_dead(get_nodes_block(skip_Proj(proj)))) { /* Remove dead control flow -- early gigo(). */ - n = new_Bad(); + proj = new_Bad(); } else if (get_opt_ldst_only_null_ptr_exceptions()) { - ir_op *op = get_irn_op(a); + if (is_Load(a)) { + /* get the Load address */ + ir_node *addr = get_Load_ptr(a); + ir_node *blk = get_irn_n(a, -1); + ir_node *confirm; - if (op == op_Load || op == op_Store) { + if (value_not_null(addr, &confirm)) { + if (confirm == NULL) { + /* this node may float if it did not depend on a Confirm */ + set_irn_pinned(a, op_pin_state_floats); + } + if (get_Proj_proj(proj) == pn_Load_X_except) { + DBG_OPT_EXC_REM(proj); + return new_Bad(); + } else + return new_r_Jmp(current_ir_graph, blk); + } + } else if (is_Store(a)) { /* get the load/store address */ - ir_node *addr = get_irn_n(a, 1); + ir_node *addr = get_Store_ptr(a); + ir_node *blk = get_irn_n(a, -1); ir_node *confirm; if (value_not_null(addr, &confirm)) { @@ -1353,14 +1389,17 @@ static ir_node *equivalent_node_Proj(ir_node *n) { /* 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(); + if (get_Proj_proj(proj) == pn_Store_X_except) { + DBG_OPT_EXC_REM(proj); + return new_Bad(); + } else + return new_r_Jmp(current_ir_graph, blk); } } } } - return n; + return proj; } /* equivalent_node_Proj */ /** @@ -1400,7 +1439,7 @@ static ir_node *equivalent_node_Mux(ir_node *n) n = get_Mux_true(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))) { + else if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) { ir_node *cmp = get_Proj_pred(sel); long proj_nr = get_Proj_proj(sel); ir_node *b = get_Mux_false(n); @@ -1414,10 +1453,11 @@ static ir_node *equivalent_node_Mux(ir_node *n) * with NaN's because -NaN == NaN. * However, if +0 and -0 is handled differently, we cannot use the first one. */ - if (get_irn_op(cmp) == op_Cmp && get_Cmp_left(cmp) == a) { - if (classify_Const(get_Cmp_right(cmp)) == CNST_NULL) { + if (is_Cmp(cmp) && get_Cmp_left(cmp) == a) { + ir_node *cmp_r = get_Cmp_right(cmp); + if (is_Const(cmp_r) && is_Const_null(cmp_r)) { /* Mux(a CMP 0, X, a) */ - if (get_irn_op(b) == op_Minus && get_Minus_op(b) == a) { + if (is_Minus(b) && get_Minus_op(b) == a) { /* Mux(a CMP 0, -a, a) */ if (proj_nr == pn_Cmp_Eq) { /* Mux(a == 0, -a, a) ==> -a */ @@ -1428,7 +1468,7 @@ static ir_node *equivalent_node_Mux(ir_node *n) n = a; DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); } - } else if (classify_Const(b) == CNST_NULL) { + } else if (is_Const(b) && is_Const_null(b)) { /* Mux(a CMP 0, 0, a) */ if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) { /* Mux(a != 0, 0, a) ==> a */ @@ -1470,7 +1510,7 @@ static ir_node *equivalent_node_Cmp(ir_node *n) { ir_node *left = get_Cmp_left(n); ir_node *right = get_Cmp_right(n); - if (get_irn_op(left) == op_Minus && get_irn_op(right) == op_Minus && + if (is_Minus(left) && is_Minus(right) && !mode_overflow_on_unary_Minus(get_irn_mode(left))) { left = get_Minus_op(left); right = get_Minus_op(right); @@ -1488,7 +1528,7 @@ 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)) { + if (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) { /* * rare case: two identical Confirms one after another, * replace the second one with the first. @@ -1518,12 +1558,14 @@ static ir_node *equivalent_node_CopyB(ir_node *n) { ir_node *b = get_CopyB_src(n); if (a == b) { - /* Turn CopyB into a tuple (mem, bad, bad) */ + /* Turn CopyB into a tuple (mem, jmp, bad, bad) */ ir_node *mem = get_CopyB_mem(n); + ir_node *blk = get_nodes_block(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()); + set_Tuple_pred(n, pn_CopyB_M, mem); + set_Tuple_pred(n, pn_CopyB_X_regular, new_r_Jmp(current_ir_graph, blk)); + set_Tuple_pred(n, pn_CopyB_X_except, new_Bad()); /* no exception */ + set_Tuple_pred(n, pn_CopyB_M_except, new_Bad()); } return n; } /* equivalent_node_CopyB */ @@ -1539,7 +1581,7 @@ static ir_node *equivalent_node_Bound(ir_node *n) { /* By definition lower < upper, so if idx == lower --> lower <= idx && idx < upper */ if (idx == lower) { - /* Turn Bound into a tuple (mem, bad, idx) */ + /* Turn Bound into a tuple (mem, jmp, bad, idx) */ ret_tuple = 1; } else { ir_node *pred = skip_Proj(idx); @@ -1555,20 +1597,22 @@ static ir_node *equivalent_node_Bound(ir_node *n) { /* * 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 + * add an exception Proj to a new location then. + * So, we must turn in into a tuple. */ ret_tuple = 1; } } } if (ret_tuple) { - /* Turn Bound into a tuple (mem, bad, idx) */ + /* Turn Bound into a tuple (mem, jmp, bad, idx) */ ir_node *mem = get_Bound_mem(n); + ir_node *blk = get_nodes_block(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); + set_Tuple_pred(n, pn_Bound_M, mem); + set_Tuple_pred(n, pn_Bound_X_regular, new_r_Jmp(current_ir_graph, blk)); /* no exception */ + set_Tuple_pred(n, pn_Bound_X_except, new_Bad()); /* no exception */ + set_Tuple_pred(n, pn_Bound_res, idx); } return n; } /* equivalent_node_Bound */ @@ -1641,30 +1685,6 @@ static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *op #undef CASE } /* firm_set_default_equivalent_node */ -/** - * Do node specific optimizations of nodes predecessors. - */ -static void optimize_preds(ir_node *n) { - switch (get_irn_opcode(n)) { - - case iro_Cmp: { /* We don't want Cast as input to Cmp. */ - ir_node *a = get_Cmp_left(n), *b = get_Cmp_right(n); - - if (get_irn_op(a) == op_Cast) { - a = get_Cast_op(a); - set_Cmp_left(n, a); - } - if (get_irn_op(b) == op_Cast) { - b = get_Cast_op(b); - set_Cmp_right(n, b); - } - break; - } - - default: break; - } /* end switch */ -} /* optimize_preds */ - /** * Returns non-zero if a node is a Phi node * with all predecessors constant. @@ -1785,7 +1805,7 @@ static ir_node *transform_node_AddSub(ir_node *n) { ir_node *right = get_binop_right(n); int ref_bits = get_mode_size_bits(mode); - if (get_irn_op(left) == op_Conv) { + if (is_Conv(left)) { ir_mode *mode = get_irn_mode(left); int bits = get_mode_size_bits(mode); @@ -1808,7 +1828,7 @@ static ir_node *transform_node_AddSub(ir_node *n) { } } - if (get_irn_op(right) == op_Conv) { + if (is_Conv(right)) { ir_mode *mode = get_irn_mode(right); int bits = get_mode_size_bits(mode); @@ -1849,18 +1869,17 @@ static ir_node *transform_node_AddSub(ir_node *n) { return c; \ } -#define HANDLE_UNOP_PHI(op,a,c) \ - c = NULL; \ - if (is_const_Phi(a)) { \ - /* check for Op(Phi) */ \ - c = apply_unop_on_phi(a, op); \ - } \ - if (c) { \ - DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \ - return c; \ +#define HANDLE_UNOP_PHI(op,a,c) \ + c = NULL; \ + if (is_const_Phi(a)) { \ + /* check for Op(Phi) */ \ + c = apply_unop_on_phi(a, op); \ + if (c) { \ + DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \ + return c; \ + } \ } - /** * Do the AddSub optimization, then Transform * Constant folding on Phi @@ -1888,7 +1907,8 @@ static ir_node *transform_node_Add(ir_node *n) { return n; if (mode_is_num(mode)) { - if (a == b) { + /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */ + if (!is_arch_dep_running() && a == b && mode_is_int(mode)) { ir_node *block = get_irn_n(n, -1); n = new_rd_Mul( @@ -1899,7 +1919,9 @@ static ir_node *transform_node_Add(ir_node *n) { new_r_Const_long(current_ir_graph, block, mode, 2), mode); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A); - } else if (get_irn_op(a) == op_Minus) { + return n; + } + if (is_Minus(a)) { n = new_rd_Sub( get_irn_dbg_info(n), current_ir_graph, @@ -1908,7 +1930,9 @@ static ir_node *transform_node_Add(ir_node *n) { get_Minus_op(a), mode); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B); - } else if (get_irn_op(b) == op_Minus) { + return n; + } + if (is_Minus(b)) { n = new_rd_Sub( get_irn_dbg_info(n), current_ir_graph, @@ -1917,79 +1941,132 @@ static ir_node *transform_node_Add(ir_node *n) { get_Minus_op(b), mode); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B); + return n; } - /* 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 (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( + if (! is_reassoc_running()) { + /* do NOT execute this code if reassociation is enabled, it does the inverse! */ + if (is_Mul(a)) { + 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_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( + 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); + return 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_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( + 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); + return n; + } + } + if (is_Mul(b)) { + 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_r_Const_long(current_ir_graph, blk, mode, 1), - mode), - mode); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A); + 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); + return n; + } + 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); + return n; + } + } + } + if (get_mode_arithmetic(mode) == irma_twos_complement) { + /* Here we rely on constants be on the RIGHT side */ + if (is_Not(a)) { + ir_node *op = get_Not_op(a); + + if (is_Const(b) && is_Const_one(b)) { + /* ~x + 1 = -x */ + ir_node *blk = get_irn_n(n, -1); + n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1); + return n; + } + if (op == b) { + /* ~x + x = -1 */ + ir_node *blk = get_irn_n(n, -1); + n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode)); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X); + return n; + } + } + if (is_Not(b)) { + ir_node *op = get_Not_op(b); + + if (op == a) { + /* x + ~x = -1 */ + ir_node *blk = get_irn_n(n, -1); + n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode)); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X); + return n; + } } } } return n; } /* transform_node_Add */ +/* returns -cnst */ +static ir_node *const_negate(ir_node *cnst) { + tarval *tv = tarval_neg(get_Const_tarval(cnst)); + dbg_info *dbgi = get_irn_dbg_info(cnst); + ir_graph *irg = get_irn_irg(cnst); + ir_node *block = get_nodes_block(cnst); + ir_mode *mode = get_irn_mode(cnst); + if (tv == tarval_bad) return NULL; + return new_rd_Const(dbgi, irg, block, mode, tv); +} + /** * Do the AddSub optimization, then Transform * Constant folding on Phi * Sub(0,a) -> Minus(a) * Sub(Mul(a, x), a) -> Mul(a, x-1) * Sub(Sub(x, y), b) -> Sub(x, Add(y,b)) + * Sub(Add(a, x), x) -> a + * Sub(x, Add(x, a)) -> -a + * Sub(x, Const) -> Add(x, -Const) */ static ir_node *transform_node_Sub(ir_node *n) { ir_mode *mode; @@ -2001,16 +2078,89 @@ static ir_node *transform_node_Sub(ir_node *n) { a = get_Sub_left(n); b = get_Sub_right(n); - HANDLE_BINOP_PHI(tarval_sub, a,b,c); - mode = get_irn_mode(n); +restart: + HANDLE_BINOP_PHI(tarval_sub, a,b,c); + /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */ if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic)) return n; + if (is_Const(b) && get_irn_mode(b) != mode_P) { + /* a - C -> a + (-C) */ + ir_node *cnst = const_negate(b); + if (cnst != NULL) { + ir_node *block = get_nodes_block(n); + dbg_info *dbgi = get_irn_dbg_info(n); + ir_graph *irg = get_irn_irg(n); + + n = new_rd_Add(dbgi, irg, block, a, cnst, mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD); + return n; + } + } + + if (is_Minus(a)) { /* (-a) - b -> -(a + b) */ + ir_graph *irg = current_ir_graph; + dbg_info *dbg = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_node *left = get_Minus_op(a); + ir_mode *mode = get_irn_mode(n); + ir_node *add = new_rd_Add(dbg, irg, block, left, b, mode); + + n = new_rd_Minus(dbg, irg, block, add, mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD); + return n; + } else if (is_Minus(b)) { /* a - (-b) -> a + b */ + ir_graph *irg = current_ir_graph; + dbg_info *dbg = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_node *right = get_Minus_op(b); + ir_mode *mode = get_irn_mode(n); + + n = new_rd_Add(dbg, irg, block, a, right, mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS); + return n; + } else if (is_Sub(b)) { /* a - (b - c) -> a + (c - b) */ + ir_graph *irg = current_ir_graph; + dbg_info *s_dbg = get_irn_dbg_info(b); + ir_node *s_block = get_nodes_block(b); + ir_node *s_left = get_Sub_right(b); + ir_node *s_right = get_Sub_left(b); + ir_mode *s_mode = get_irn_mode(b); + ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, s_left, s_right, s_mode); + dbg_info *a_dbg = get_irn_dbg_info(n); + ir_node *a_block = get_nodes_block(n); + ir_mode *a_mode = get_irn_mode(n); + + n = new_rd_Add(a_dbg, irg, a_block, a, sub, a_mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD); + return n; + } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */ + ir_node *m_right = get_Mul_right(b); + if (is_Const(m_right)) { + ir_node *cnst2 = const_negate(m_right); + if (cnst2 != NULL) { + ir_graph *irg = current_ir_graph; + dbg_info *m_dbg = get_irn_dbg_info(b); + ir_node *m_block = get_nodes_block(b); + ir_node *m_left = get_Mul_left(b); + ir_mode *m_mode = get_irn_mode(b); + ir_node *mul = new_rd_Mul(m_dbg, irg, m_block, m_left, cnst2, m_mode); + dbg_info *a_dbg = get_irn_dbg_info(n); + ir_node *a_block = get_nodes_block(n); + ir_mode *a_mode = get_irn_mode(n); + + n = new_rd_Add(a_dbg, irg, a_block, a, mul, a_mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD); + return n; + } + } + } + /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */ - if (mode_is_num(mode) && mode == get_irn_mode(a) && (classify_Const(a) == CNST_NULL)) { + if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) { n = new_rd_Minus( get_irn_dbg_info(n), current_ir_graph, @@ -2018,9 +2168,86 @@ static ir_node *transform_node_Sub(ir_node *n) { b, mode); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A); + return n; + } + if (is_Add(a)) { + if (mode_wrap_around(mode)) { + ir_node *left = get_Add_left(a); + ir_node *right = get_Add_right(a); + + /* FIXME: Does the Conv's work only for two complement or generally? */ + if (left == b) { + if (mode != get_irn_mode(right)) { + /* This Sub is an effective Cast */ + right = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), right, mode); + } + n = right; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); + return n; + } else if (right == b) { + if (mode != get_irn_mode(left)) { + /* This Sub is an effective Cast */ + left = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), left, mode); + } + n = left; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); + return n; + } + } + } + if (is_Add(b)) { + if (mode_wrap_around(mode)) { + ir_node *left = get_Add_left(b); + ir_node *right = get_Add_right(b); + + /* FIXME: Does the Conv's work only for two complement or generally? */ + if (left == a) { + ir_mode *r_mode = get_irn_mode(right); + + n = new_r_Minus(get_irn_irg(n), get_irn_n(n, -1), right, r_mode); + if (mode != r_mode) { + /* This Sub is an effective Cast */ + n = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), n, mode); + } + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); + return n; + } else if (right == a) { + ir_mode *l_mode = get_irn_mode(left); + + n = new_r_Minus(get_irn_irg(n), get_irn_n(n, -1), left, l_mode); + if (mode != l_mode) { + /* This Sub is an effective Cast */ + n = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), n, mode); + } + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); + return n; + } + } + } + if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) { + ir_mode *mode = get_irn_mode(a); + + if (mode == get_irn_mode(b)) { + ir_mode *ma, *mb; + + a = get_Conv_op(a); + b = get_Conv_op(b); + + /* check if it's allowed to skip the conv */ + ma = get_irn_mode(a); + mb = get_irn_mode(b); + + if (mode_is_reference(ma) && mode_is_reference(mb)) { + /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */ + set_Sub_left(n, a); + set_Sub_right(n, b); + + goto restart; + } + } } /* do NOT execute this code if reassociation is enabled, it does the inverse! */ - else if (get_opt_reassociation() && get_irn_op(a) == op_Mul) { + if (!is_reassoc_running() && is_Mul(a)) { ir_node *ma = get_Mul_left(a); ir_node *mb = get_Mul_right(a); @@ -2038,6 +2265,7 @@ static ir_node *transform_node_Sub(ir_node *n) { mode), mode); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A); + return n; } else if (mb == b) { ir_node *blk = get_irn_n(n, -1); n = new_rd_Mul( @@ -2052,22 +2280,25 @@ static ir_node *transform_node_Sub(ir_node *n) { mode), mode); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A); + return n; } - } else if (get_irn_op(a) == op_Sub) { - ir_node *x = get_Sub_left(a); - ir_node *y = get_Sub_right(a); - ir_node *blk = get_irn_n(n, -1); - ir_mode *m_b = get_irn_mode(b); - ir_mode *m_y = get_irn_mode(y); + } + if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */ + ir_node *x = get_Sub_left(a); + ir_node *y = get_Sub_right(a); + ir_node *blk = get_irn_n(n, -1); + ir_mode *m_b = get_irn_mode(b); + ir_mode *m_y = get_irn_mode(y); + ir_mode *add_mode; ir_node *add; /* Determine the right mode for the Add. */ if (m_b == m_y) - mode = m_b; + add_mode = m_b; else if (mode_is_reference(m_b)) - mode = m_b; + add_mode = m_b; else if (mode_is_reference(m_y)) - mode = m_y; + add_mode = m_y; else { /* * Both modes are different but none is reference, @@ -2078,16 +2309,72 @@ static ir_node *transform_node_Sub(ir_node *n) { return n; } - add = new_r_Add(current_ir_graph, blk, y, b, mode); + add = new_r_Add(current_ir_graph, blk, y, b, add_mode); - set_Sub_left(n, x); - set_Sub_right(n, add); - DBG_OPT_ALGSIM0(n, n, FS_OPT_SUB_SUB_X_Y_Z); + n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, x, add, mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z); + return n; } + if (get_mode_arithmetic(mode) == irma_twos_complement) { + if (is_Const(a) && is_Not(b)) { + /* c - ~X = X + (c+1) */ + tarval *tv = get_Const_tarval(a); + + tv = tarval_add(tv, get_mode_one(mode)); + if (tv != tarval_bad) { + ir_node *blk = get_irn_n(n, -1); + ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv); + n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, get_Not_op(b), c, mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X); + return n; + } + } + } return n; } /* transform_node_Sub */ +/** + * Several transformation done on n*n=2n bits mul. + * These transformations must be done here because new nodes may be produced. + */ +static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) { + ir_node *oldn = n; + ir_node *a = get_Mul_left(n); + ir_node *b = get_Mul_right(n); + tarval *ta = value_of(a); + tarval *tb = value_of(b); + ir_mode *smode = get_irn_mode(a); + + if (ta == get_mode_one(smode)) { + ir_node *blk = get_irn_n(n, -1); + n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, b, mode); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1); + return n; + } + else if (ta == get_mode_minus_one(smode)) { + ir_node *blk = get_irn_n(n, -1); + n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, b, smode); + n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1); + return n; + } + if (tb == get_mode_one(smode)) { + ir_node *blk = get_irn_n(a, -1); + n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, a, mode); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1); + return n; + } + else if (tb == get_mode_minus_one(smode)) { + ir_node *blk = get_irn_n(n, -1); + n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, a, smode); + n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1); + return n; + } + return n; +} + /** * Transform Mul(a,-1) into -a. * Do constant evaluation of Phi nodes. @@ -2095,13 +2382,18 @@ static ir_node *transform_node_Sub(ir_node *n) { */ static ir_node *transform_node_Mul(ir_node *n) { ir_node *c, *oldn = n; + ir_mode *mode = get_irn_mode(n); ir_node *a = get_Mul_left(n); ir_node *b = get_Mul_right(n); - ir_mode *mode; + + if (is_Bad(a) || is_Bad(b)) + return n; + + if (mode != get_irn_mode(a)) + return transform_node_Mul2n(n, mode); HANDLE_BINOP_PHI(tarval_mul, a,b,c); - mode = get_irn_mode(n); if (mode_is_signed(mode)) { ir_node *r = NULL; @@ -2115,6 +2407,64 @@ static ir_node *transform_node_Mul(ir_node *n) { return n; } } + if (is_Minus(a)) { + if (is_Const(b)) { /* (-a) * const -> a * -const */ + ir_node *cnst = const_negate(b); + if (cnst != NULL) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), cnst, mode); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1); + return n; + } + } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */ + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), get_Minus_op(b), mode); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS); + return n; + } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */ + ir_node *sub_l = get_Sub_left(b); + ir_node *sub_r = get_Sub_right(b); + dbg_info *dbgi = get_irn_dbg_info(n); + ir_graph *irg = current_ir_graph; + ir_node *block = get_nodes_block(n); + ir_node *new_b = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode); + n = new_rd_Mul(dbgi, irg, block, get_Minus_op(a), new_b, mode); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS); + return n; + } + } else if (is_Minus(b)) { + if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */ + ir_node *sub_l = get_Sub_left(a); + ir_node *sub_r = get_Sub_right(a); + dbg_info *dbgi = get_irn_dbg_info(n); + ir_graph *irg = current_ir_graph; + ir_node *block = get_nodes_block(n); + ir_node *new_a = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode); + n = new_rd_Mul(dbgi, irg, block, new_a, get_Minus_op(b), mode); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS); + return n; + } + } + if (get_mode_arithmetic(mode) == irma_ieee754) { + if (is_Const(a)) { + tarval *tv = get_Const_tarval(a); + if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) { + n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), b, b, mode); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A); + return n; + } + } + else if (is_Const(b)) { + tarval *tv = get_Const_tarval(b); + if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) { + n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, a, mode); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A); + return n; + } + } + } return arch_dep_replace_mul_with_shifts(n); } /* transform_node_Mul */ @@ -2123,25 +2473,56 @@ static ir_node *transform_node_Mul(ir_node *n) { */ static ir_node *transform_node_Div(ir_node *n) { tarval *tv = value_of(n); + ir_mode *mode = get_Div_resmode(n); ir_node *value = n; - /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */ - if (tv != tarval_bad) { value = new_Const(get_tarval_mode(tv), tv); DBG_OPT_CSTEVAL(n, value); - } else /* Try architecture dependent optimization */ - value = arch_dep_replace_div_by_const(n); + goto make_tuple; + } else { + ir_node *a = get_Div_left(n); + ir_node *b = get_Div_right(n); + ir_node *dummy; + + if (a == b && value_not_zero(a, &dummy)) { + /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */ + value = new_Const(mode, get_mode_one(mode)); + DBG_OPT_CSTEVAL(n, value); + goto make_tuple; + } else { + if (mode_is_signed(mode) && is_Const(b)) { + tarval *tv = get_Const_tarval(b); + + if (tv == get_mode_minus_one(mode)) { + /* a / -1 */ + value = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode); + DBG_OPT_CSTEVAL(n, value); + goto make_tuple; + } + } + /* 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); + ir_node *mem, *blk; + +make_tuple: + /* Turn Div into a tuple (mem, jmp, bad, value) */ + mem = get_Div_mem(n); + blk = get_irn_n(n, -1); + /* skip a potential Pin */ + if (is_Pin(mem)) + mem = get_Pin_op(mem); 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); + set_Tuple_pred(n, pn_Div_M, mem); + set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk)); + set_Tuple_pred(n, pn_Div_X_except, new_Bad()); + set_Tuple_pred(n, pn_Div_res, value); } return n; } /* transform_node_Div */ @@ -2151,25 +2532,56 @@ static ir_node *transform_node_Div(ir_node *n) { */ static ir_node *transform_node_Mod(ir_node *n) { tarval *tv = value_of(n); + ir_mode *mode = get_Mod_resmode(n); ir_node *value = n; - /* BEWARE: it is NOT possible to optimize a%a to 0, as this may cause a exception */ - if (tv != tarval_bad) { value = new_Const(get_tarval_mode(tv), tv); DBG_OPT_CSTEVAL(n, value); - } else /* Try architecture dependent optimization */ - value = arch_dep_replace_mod_by_const(n); + goto make_tuple; + } else { + ir_node *a = get_Mod_left(n); + ir_node *b = get_Mod_right(n); + ir_node *dummy; + + if (a == b && value_not_zero(a, &dummy)) { + /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */ + value = new_Const(mode, get_mode_null(mode)); + DBG_OPT_CSTEVAL(n, value); + goto make_tuple; + } else { + if (mode_is_signed(mode) && is_Const(b)) { + tarval *tv = get_Const_tarval(b); + + if (tv == get_mode_minus_one(mode)) { + /* a % -1 = 0 */ + value = new_Const(mode, get_mode_null(mode)); + DBG_OPT_CSTEVAL(n, value); + goto make_tuple; + } + } + /* 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); + ir_node *mem, *blk; + +make_tuple: + /* Turn Mod into a tuple (mem, jmp, bad, value) */ + mem = get_Mod_mem(n); + blk = get_irn_n(n, -1); + /* skip a potential Pin */ + if (is_Pin(mem)) + mem = get_Pin_op(mem); 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); + set_Tuple_pred(n, pn_Mod_M, mem); + set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk)); + set_Tuple_pred(n, pn_Mod_X_except, new_Bad()); + set_Tuple_pred(n, pn_Mod_res, value); } return n; } /* transform_node_Mod */ @@ -2178,60 +2590,120 @@ static ir_node *transform_node_Mod(ir_node *n) { * Transform a DivMod node. */ static ir_node *transform_node_DivMod(ir_node *n) { - int evaluated = 0; - + ir_node *dummy; ir_node *a = get_DivMod_left(n); ir_node *b = get_DivMod_right(n); - ir_mode *mode = get_irn_mode(a); + ir_mode *mode = get_DivMod_resmode(n); tarval *ta = value_of(a); tarval *tb = value_of(b); - - if (!(mode_is_int(mode) && mode_is_int(get_irn_mode(b)))) - return n; - - /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */ + int evaluated = 0; if (tb != tarval_bad) { if (tb == get_mode_one(get_tarval_mode(tb))) { - b = new_Const (mode, get_mode_null(mode)); - evaluated = 1; - + b = new_Const(mode, get_mode_null(mode)); DBG_OPT_CSTEVAL(n, b); + goto make_tuple; } else if (ta != tarval_bad) { tarval *resa, *resb; - resa = tarval_div (ta, tb); + resa = tarval_div(ta, tb); if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through Jmp for X result!? */ - resb = tarval_mod (ta, tb); + resb = tarval_mod(ta, tb); if (resb == tarval_bad) return n; /* Causes exception! */ - a = new_Const (mode, resa); - b = new_Const (mode, resb); - evaluated = 1; - + a = new_Const(mode, resa); + b = new_Const(mode, resb); DBG_OPT_CSTEVAL(n, a); DBG_OPT_CSTEVAL(n, b); + goto make_tuple; + } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) { + a = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode); + b = new_Const(mode, get_mode_null(mode)); + DBG_OPT_CSTEVAL(n, a); + DBG_OPT_CSTEVAL(n, b); + goto make_tuple; } else { /* Try architecture dependent optimization */ arch_dep_replace_divmod_by_const(&a, &b, n); evaluated = a != NULL; } - } else if (ta == get_mode_null(mode)) { + } else if (a == b) { + if (value_not_zero(a, &dummy)) { + /* a/a && a != 0 */ + a = new_Const(mode, get_mode_one(mode)); + b = new_Const(mode, get_mode_null(mode)); + DBG_OPT_CSTEVAL(n, a); + DBG_OPT_CSTEVAL(n, b); + goto make_tuple; + } else { + /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */ + return n; + } + } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) { /* 0 / non-Const = 0 */ b = a; - evaluated = 1; + goto make_tuple; } if (evaluated) { /* replace by tuple */ - ir_node *mem = get_DivMod_mem(n); + ir_node *mem, *blk; + +make_tuple: + mem = get_DivMod_mem(n); + /* skip a potential Pin */ + if (is_Pin(mem)) + mem = get_Pin_op(mem); + + blk = get_irn_n(n, -1); 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_M, mem); + set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk)); + 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); } return n; } /* transform_node_DivMod */ +/** + * Optimize x / c to x * (1/c) + */ +static ir_node *transform_node_Quot(ir_node *n) { + ir_mode *mode = get_Quot_resmode(n); + ir_node *oldn = n; + + if (get_mode_arithmetic(mode) == irma_ieee754) { + ir_node *b = get_Quot_right(n); + + if (is_Const(b)) { + tarval *tv = get_Const_tarval(b); + + tv = tarval_quo(get_mode_one(mode), tv); + + /* Do the transformation if the result is either exact or we are not + using strict rules. */ + if (tv != tarval_bad && + (tarval_ieee754_get_exact() || (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic) == 0)) { + ir_node *blk = get_irn_n(n, -1); + ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv); + ir_node *a = get_Quot_left(n); + ir_node *m = new_rd_Mul(get_irn_dbg_info(n), current_ir_graph, blk, a, c, mode); + ir_node *mem = get_Quot_mem(n); + + /* skip a potential Pin */ + if (is_Pin(mem)) + mem = get_Pin_op(mem); + turn_into_tuple(n, pn_Quot_max); + set_Tuple_pred(n, pn_Quot_M, mem); + set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk)); + set_Tuple_pred(n, pn_Quot_X_except, new_r_Bad(current_ir_graph)); + set_Tuple_pred(n, pn_Quot_res, m); + DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL); + } + } + } + return n; +} /* transform_node_Quot */ + /** * Optimize Abs(x) into x if x is Confirmed >= 0 * Optimize Abs(x) into -x if x is Confirmed <= 0 @@ -2301,6 +2773,111 @@ static ir_node *transform_node_Cond(ir_node *n) { return n; } /* transform_node_Cond */ +typedef ir_node* (*recursive_transform) (ir_node *n); + +/** + * makes use of distributive laws for and, or, eor + * and(a OP c, b OP c) -> and(a, b) OP c + * note, might return a different op than n + */ +static ir_node *transform_bitwise_distributive(ir_node *n, + recursive_transform trans_func) +{ + ir_node *oldn = n; + ir_node *a = get_binop_left(n); + ir_node *b = get_binop_right(n); + ir_op *op = get_irn_op(a); + ir_op *op_root = get_irn_op(n); + + if(op != get_irn_op(b)) + return n; + + if (op == op_Conv) { + ir_node *a_op = get_Conv_op(a); + ir_node *b_op = get_Conv_op(b); + ir_mode *a_mode = get_irn_mode(a_op); + ir_mode *b_mode = get_irn_mode(b_op); + if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) { + ir_node *blk = get_irn_n(n, -1); + + n = exact_copy(n); + set_binop_left(n, a_op); + set_binop_right(n, b_op); + set_irn_mode(n, a_mode); + n = trans_func(n); + n = new_r_Conv(current_ir_graph, blk, n, get_irn_mode(oldn)); + + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND); + return n; + } + } + + if (op == op_Eor) { + /* nothing to gain here */ + return n; + } + + if (op == op_Shrs || op == op_Shr || op == op_Shl + || op == op_And || op == op_Or || op == op_Eor) { + ir_node *a_left = get_binop_left(a); + ir_node *a_right = get_binop_right(a); + ir_node *b_left = get_binop_left(b); + ir_node *b_right = get_binop_right(b); + ir_node *c = NULL; + ir_node *op1, *op2; + + if (is_op_commutative(op)) { + if (a_left == b_left) { + c = a_left; + op1 = a_right; + op2 = b_right; + } else if(a_left == b_right) { + c = a_left; + op1 = a_right; + op2 = b_left; + } else if(a_right == b_left) { + c = a_right; + op1 = a_left; + op2 = b_right; + } + } + if(a_right == b_right) { + c = a_right; + op1 = a_left; + op2 = b_left; + } + + if (c != NULL) { + /* (a sop c) & (b sop c) => (a & b) sop c */ + ir_node *blk = get_irn_n(n, -1); + + ir_node *new_n = exact_copy(n); + set_binop_left(new_n, op1); + set_binop_right(new_n, op2); + new_n = trans_func(new_n); + + if(op_root == op_Eor && op == op_Or) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_graph *irg = current_ir_graph; + ir_mode *mode = get_irn_mode(c); + + c = new_rd_Not(dbgi, irg, blk, c, mode); + n = new_rd_And(dbgi, irg, blk, new_n, c, mode); + } else { + n = exact_copy(a); + set_irn_n(n, -1, blk); + set_binop_left(n, new_n); + set_binop_right(n, c); + } + + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND); + return n; + } + } + + return n; +} + /** * Transform an And. */ @@ -2308,8 +2885,130 @@ static ir_node *transform_node_And(ir_node *n) { ir_node *c, *oldn = n; ir_node *a = get_And_left(n); ir_node *b = get_And_right(n); + ir_mode *mode; HANDLE_BINOP_PHI(tarval_and, a,b,c); + + mode = get_irn_mode(n); + + /* we can evaluate 2 Projs of the same Cmp */ + if (mode == mode_b && is_Proj(a) && is_Proj(b)) { + ir_node *pred_a = get_Proj_pred(a); + ir_node *pred_b = get_Proj_pred(b); + if (pred_a == pred_b) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(pred_a); + pn_Cmp pn_a = get_Proj_proj(a); + pn_Cmp pn_b = get_Proj_proj(b); + /* yes, we can simply calculate with pncs */ + pn_Cmp new_pnc = pn_a & pn_b; + + return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b, new_pnc); + } + } + if (is_Or(a)) { + if (is_Not(b)) { + ir_node *op = get_Not_op(b); + if (is_And(op)) { + ir_node *ba = get_And_left(op); + ir_node *bb = get_And_right(op); + + /* it's enough to test the following cases due to normalization! */ + if (get_Or_left(a) == ba && get_Or_right(a) == bb) { + /* (a|b) & ~(a&b) = a^b */ + ir_node *block = get_nodes_block(n); + + n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, ba, bb, mode); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR); + return n; + } + } + } + } + if (is_Or(b)) { + if (is_Not(a)) { + ir_node *op = get_Not_op(a); + if (is_And(op)) { + ir_node *aa = get_And_left(op); + ir_node *ab = get_And_right(op); + + /* it's enough to test the following cases due to normalization! */ + if (get_Or_left(b) == aa && get_Or_right(b) == ab) { + /* (a|b) & ~(a&b) = a^b */ + ir_node *block = get_nodes_block(n); + + n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, aa, ab, mode); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR); + return n; + } + } + } + } + if (is_Eor(a)) { + ir_node *al = get_Eor_left(a); + ir_node *ar = get_Eor_right(a); + + if (al == b) { + /* (b ^ a) & b -> ~a & b */ + dbg_info *dbg = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + + ar = new_rd_Not(dbg, current_ir_graph, block, ar, mode); + n = new_rd_And(dbg, current_ir_graph, block, ar, b, mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT); + return n; + } + if (ar == b) { + /* (a ^ b) & b -> ~a & b */ + dbg_info *dbg = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + + al = new_rd_Not(dbg, current_ir_graph, block, al, mode); + n = new_rd_And(dbg, current_ir_graph, block, al, b, mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT); + return n; + } + } + if (is_Eor(b)) { + ir_node *bl = get_Eor_left(b); + ir_node *br = get_Eor_right(b); + + if (bl == a) { + /* a & (a ^ b) -> a & ~b */ + dbg_info *dbg = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + + br = new_rd_Not(dbg, current_ir_graph, block, br, mode); + n = new_rd_And(dbg, current_ir_graph, block, br, a, mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT); + return n; + } + if (br == a) { + /* a & (b ^ a) -> a & ~b */ + dbg_info *dbg = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + + bl = new_rd_Not(dbg, current_ir_graph, block, bl, mode); + n = new_rd_And(dbg, current_ir_graph, block, bl, a, mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT); + return n; + } + } + if (is_Not(a) && is_Not(b)) { + /* ~a & ~b = ~(a|b) */ + ir_node *block = get_nodes_block(n); + ir_mode *mode = get_irn_mode(n); + + a = get_Not_op(a); + b = get_Not_op(b); + n = new_rd_Or(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode); + n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN); + return n; + } + + n = transform_bitwise_distributive(n, transform_node_And); + return n; } /* transform_node_And */ @@ -2324,28 +3023,53 @@ static ir_node *transform_node_Eor(ir_node *n) { HANDLE_BINOP_PHI(tarval_eor, a,b,c); + /* we can evaluate 2 Projs of the same Cmp */ + if (mode == mode_b && is_Proj(a) && is_Proj(b)) { + ir_node *pred_a = get_Proj_pred(a); + ir_node *pred_b = get_Proj_pred(b); + if(pred_a == pred_b) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(pred_a); + pn_Cmp pn_a = get_Proj_proj(a); + pn_Cmp pn_b = get_Proj_proj(b); + /* yes, we can simply calculate with pncs */ + pn_Cmp new_pnc = pn_a ^ pn_b; + + return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b, + new_pnc); + } + } + if (a == b) { /* a ^ a = 0 */ 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, FS_OPT_EOR_A_A); - } else if ((mode == mode_b) - && (get_irn_op(a) == op_Proj) - && (get_irn_mode(a) == mode_b) - && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE) - && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) { + } else if (mode == mode_b && + is_Proj(a) && + is_Const(b) && is_Const_one(b) && + is_Cmp(get_Proj_pred(a))) { /* The Eor negates a Cmp. The Cmp has the negated result anyways! */ 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, 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_irn_n(n, -1), a, mode_b); - - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT); + } else if (is_Const(b)) { + if (is_Not(a)) { /* ~x ^ const -> x ^ ~const */ + ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(b))); + ir_node *not_op = get_Not_op(a); + dbg_info *dbg = get_irn_dbg_info(n); + ir_graph *irg = current_ir_graph; + ir_node *block = get_nodes_block(n); + ir_mode *mode = get_irn_mode(n); + n = new_rd_Eor(dbg, irg, block, not_op, cnst, mode); + return n; + } else if (is_Const_all_one(b)) { /* x ^ 1...1 -> ~1 */ + n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT); + } + } else { + n = transform_bitwise_distributive(n, transform_node_Eor); } return n; @@ -2356,31 +3080,140 @@ static ir_node *transform_node_Eor(ir_node *n) { */ static ir_node *transform_node_Not(ir_node *n) { ir_node *c, *oldn = n; - ir_node *a = get_Not_op(n); + ir_node *a = get_Not_op(n); + ir_mode *mode = get_irn_mode(n); HANDLE_UNOP_PHI(tarval_not,a,c); /* check for a boolean Not */ - if ( (get_irn_mode(n) == mode_b) - && (get_irn_op(a) == op_Proj) - && (get_irn_mode(a) == mode_b) - && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) { + if (mode == mode_b && + is_Proj(a) && + is_Cmp(get_Proj_pred(a))) { /* We negate a Cmp. The Cmp has the negated result anyways! */ 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; + } + if (is_Eor(a)) { + ir_node *eor_b = get_Eor_right(a); + if (is_Const(eor_b)) { /* ~(x ^ const) -> x ^ ~const */ + ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(eor_b))); + ir_node *eor_a = get_Eor_left(a); + dbg_info *dbg = get_irn_dbg_info(n); + ir_graph *irg = current_ir_graph; + ir_node *block = get_nodes_block(n); + ir_mode *mode = get_irn_mode(n); + n = new_rd_Eor(dbg, irg, block, eor_a, cnst, mode); + return n; + } + } + if (get_mode_arithmetic(mode) == irma_twos_complement) { + if (is_Minus(a)) { /* ~-x -> x + -1 */ + dbg_info *dbg = get_irn_dbg_info(n); + ir_graph *irg = current_ir_graph; + ir_node *block = get_nodes_block(n); + ir_node *add_l = get_Minus_op(a); + ir_node *add_r = new_rd_Const(dbg, irg, block, mode, get_mode_minus_one(mode)); + n = new_rd_Add(dbg, irg, block, add_l, add_r, mode); + } else if (is_Add(a)) { + ir_node *add_r = get_Add_right(a); + if (is_Const(add_r) && is_Const_all_one(add_r)) { + /* ~(x + -1) = -x */ + ir_node *op = get_Add_left(a); + ir_node *blk = get_irn_n(n, -1); + n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, get_irn_mode(n)); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1); + } + } } return n; } /* transform_node_Not */ /** * Transform a Minus. + * Optimize: + * -(~x) = x + 1 + * -(a-b) = b - a */ static ir_node *transform_node_Minus(ir_node *n) { ir_node *c, *oldn = n; ir_node *a = get_Minus_op(n); + ir_mode *mode; HANDLE_UNOP_PHI(tarval_neg,a,c); + + mode = get_irn_mode(a); + if (get_mode_arithmetic(mode) == irma_twos_complement) { + /* the following rules are only to twos-complement */ + if (is_Not(a)) { + /* -(~x) = x + 1 */ + ir_node *op = get_Not_op(a); + tarval *tv = get_mode_one(mode); + ir_node *blk = get_irn_n(n, -1); + ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv); + n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, op, c, mode); + DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT); + return n; + } + if (is_Shr(a)) { + ir_node *c = get_Shr_right(a); + + if (is_Const(c)) { + tarval *tv = get_Const_tarval(c); + + if (tarval_is_long(tv) && get_tarval_long(tv) == get_mode_size_bits(mode) - 1) { + /* -(a >>u (size-1)) = a >>s (size-1) */ + ir_node *v = get_Shr_left(a); + + n = new_rd_Shrs(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), v, c, mode); + DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE); + return n; + } + } + } + if (is_Shrs(a)) { + ir_node *c = get_Shrs_right(a); + + if (is_Const(c)) { + tarval *tv = get_Const_tarval(c); + + if (tarval_is_long(tv) && get_tarval_long(tv) == get_mode_size_bits(mode) - 1) { + /* -(a >>s (size-1)) = a >>u (size-1) */ + ir_node *v = get_Shrs_left(a); + + n = new_rd_Shr(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), v, c, mode); + DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE); + return n; + } + } + } + } + if (is_Sub(a)) { + /* - (a-b) = b - a */ + ir_node *la = get_Sub_left(a); + ir_node *ra = get_Sub_right(a); + ir_node *blk = get_irn_n(n, -1); + + n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, ra, la, mode); + DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB); + return n; + } + + if (is_Mul(a)) { /* -(a * const) -> a * -const */ + ir_node *mul_l = get_Mul_left(a); + ir_node *mul_r = get_Mul_right(a); + if (is_Const(mul_r)) { + tarval *tv = tarval_neg(get_Const_tarval(mul_r)); + ir_node *cnst = new_Const(mode, tv); + dbg_info *dbg = get_irn_dbg_info(a); + ir_graph *irg = current_ir_graph; + ir_node *block = get_nodes_block(a); + n = new_rd_Mul(dbg, irg, block, mul_l, cnst, mode); + return n; + } + } + return n; } /* transform_node_Minus */ @@ -2392,11 +3225,11 @@ static ir_node *transform_node_Cast(ir_node *n) { ir_node *pred = get_Cast_op(n); ir_type *tp = get_irn_type(n); - if (get_irn_op(pred) == op_Const && get_Const_type(pred) != tp) { + if (is_Const(pred) && get_Const_type(pred) != tp) { 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)) { + } else if (is_SymConst(pred) && get_SymConst_value_type(pred) != tp) { 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); @@ -2410,29 +3243,43 @@ static ir_node *transform_node_Cast(ir_node *n) { * 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 = get_Div_right(n); - ir_node *confirm; + ir_node *div = get_Proj_pred(proj); + ir_node *b = get_Div_right(div); + ir_node *confirm, *res, *new_mem; long proj_nr; if (value_not_zero(b, &confirm)) { /* div(x, y) && y != 0 */ + if (confirm == NULL) { + /* we are sure we have a Const != 0 */ + new_mem = get_Div_mem(div); + if (is_Pin(new_mem)) + new_mem = get_Pin_op(new_mem); + set_Div_mem(div, new_mem); + set_irn_pinned(div, op_pin_state_floats); + } + proj_nr = get_Proj_proj(proj); - if (proj_nr == pn_Div_X_except) { + switch (proj_nr) { + case pn_Div_X_regular: + return new_r_Jmp(current_ir_graph, get_irn_n(div, -1)); + + case pn_Div_X_except: /* we found an exception handler, remove it */ DBG_OPT_EXC_REM(proj); return new_Bad(); - } else if (proj_nr == pn_Div_M) { - ir_node *res = get_Div_mem(n); - ir_node *new_mem = get_irg_no_mem(current_ir_graph); + + case pn_Div_M: + res = get_Div_mem(div); + new_mem = get_irg_no_mem(current_ir_graph); if (confirm) { /* This node can only float up to the Confirm block */ new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem); } - set_irn_pinned(n, op_pin_state_floats); + set_irn_pinned(div, op_pin_state_floats); /* this is a Div without exception, we can remove the memory edge */ - set_Div_mem(n, new_mem); + set_Div_mem(div, new_mem); return res; } } @@ -2444,38 +3291,54 @@ static ir_node *transform_node_Proj_Div(ir_node *proj) { * 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 = get_Mod_right(n); - ir_node *confirm; + ir_node *mod = get_Proj_pred(proj); + ir_node *b = get_Mod_right(mod); + ir_node *confirm, *res, *new_mem; long proj_nr; if (value_not_zero(b, &confirm)) { /* mod(x, y) && y != 0 */ proj_nr = get_Proj_proj(proj); - if (proj_nr == pn_Mod_X_except) { + if (confirm == NULL) { + /* we are sure we have a Const != 0 */ + new_mem = get_Mod_mem(mod); + if (is_Pin(new_mem)) + new_mem = get_Pin_op(new_mem); + set_Mod_mem(mod, new_mem); + set_irn_pinned(mod, op_pin_state_floats); + } + + switch (proj_nr) { + + case pn_Mod_X_regular: + return new_r_Jmp(current_ir_graph, get_irn_n(mod, -1)); + + case pn_Mod_X_except: /* we found an exception handler, remove it */ DBG_OPT_EXC_REM(proj); return new_Bad(); - } else if (proj_nr == pn_Mod_M) { - ir_node *res = get_Mod_mem(n); - ir_node *new_mem = get_irg_no_mem(current_ir_graph); + + case pn_Mod_M: + res = get_Mod_mem(mod); + new_mem = get_irg_no_mem(current_ir_graph); if (confirm) { /* This node can only float up to the Confirm block */ new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem); } - set_irn_pinned(n, op_pin_state_floats); /* this is a Mod without exception, 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); + set_Mod_mem(mod, new_mem); return res; + case pn_Mod_res: + if (get_Mod_left(mod) == 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(mod, res); + return res; + } } } return proj; @@ -2486,38 +3349,55 @@ static ir_node *transform_node_Proj_Mod(ir_node *proj) { * 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 = get_DivMod_right(n); - ir_node *confirm; + ir_node *divmod = get_Proj_pred(proj); + ir_node *b = get_DivMod_right(divmod); + ir_node *confirm, *res, *new_mem; long proj_nr; if (value_not_zero(b, &confirm)) { /* DivMod(x, y) && y != 0 */ proj_nr = get_Proj_proj(proj); - if (proj_nr == pn_DivMod_X_except) { + if (confirm == NULL) { + /* we are sure we have a Const != 0 */ + new_mem = get_DivMod_mem(divmod); + if (is_Pin(new_mem)) + new_mem = get_Pin_op(new_mem); + set_DivMod_mem(divmod, new_mem); + set_irn_pinned(divmod, op_pin_state_floats); + } + + switch (proj_nr) { + + case pn_DivMod_X_regular: + return new_r_Jmp(current_ir_graph, get_irn_n(divmod, -1)); + + case pn_DivMod_X_except: /* we found an exception handler, remove it */ DBG_OPT_EXC_REM(proj); return new_Bad(); - } else if (proj_nr == pn_DivMod_M) { - ir_node *res = get_DivMod_mem(n); - ir_node *new_mem = get_irg_no_mem(current_ir_graph); + + case pn_DivMod_M: + res = get_DivMod_mem(divmod); + new_mem = get_irg_no_mem(current_ir_graph); if (confirm) { /* This node can only float up to the Confirm block */ new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem); } - set_irn_pinned(n, op_pin_state_floats); /* this is a DivMod without exception, we can remove the memory edge */ - set_DivMod_mem(n, get_irg_no_mem(current_ir_graph)); + set_DivMod_mem(divmod, new_mem); 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; + case pn_DivMod_res_mod: + if (get_DivMod_left(divmod) == 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(divmod, res); + return res; + } } } return proj; @@ -2558,217 +3438,404 @@ static ir_node *transform_node_Proj_Cond(ir_node *proj) { * Normalizes and optimizes Cmp nodes. */ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { - if (get_opt_reassociation()) { - ir_node *n = get_Proj_pred(proj); - ir_node *left = get_Cmp_left(n); - ir_node *right = get_Cmp_right(n); - ir_node *c = NULL; - tarval *tv = NULL; - int changed = 0; - ir_mode *mode = NULL; - long proj_nr = get_Proj_proj(proj); + ir_node *n = get_Proj_pred(proj); + ir_node *left = get_Cmp_left(n); + ir_node *right = get_Cmp_right(n); + ir_node *c = NULL; + tarval *tv = NULL; + int changed = 0; + ir_mode *mode = NULL; + long proj_nr = get_Proj_proj(proj); + + /* we can evaluate this direct */ + switch(proj_nr) { + case pn_Cmp_False: + return new_Const(mode_b, get_tarval_b_false()); + case pn_Cmp_True: + return new_Const(mode_b, get_tarval_b_true()); + case pn_Cmp_Leg: + if(!mode_is_float(get_irn_mode(left))) + return new_Const(mode_b, get_tarval_b_true()); + break; + default: + break; + } - /* - * First step: normalize the compare op - * by placing the constant on the right site - * or moving the lower address node to the left. - * We ignore the case that both are constants - * this case should be optimized away. - */ - if (get_irn_op(right) == op_Const) { - c = right; - } else if (get_irn_op(left) == op_Const) { - c = left; - left = right; - right = c; + /* Remove unnecessary conversions */ + /* TODO handle constants */ + if (is_Conv(left) && is_Conv(right)) { + ir_mode *mode = get_irn_mode(left); + ir_node *op_left = get_Conv_op(left); + ir_node *op_right = get_Conv_op(right); + ir_mode *mode_left = get_irn_mode(op_left); + ir_mode *mode_right = get_irn_mode(op_right); + + if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)) { + ir_graph *irg = current_ir_graph; + ir_node *block = get_nodes_block(n); + + if (mode_left == mode_right) { + left = op_left; + right = op_right; + changed |= 1; + } else if (smaller_mode(mode_left, mode_right)) { + left = new_r_Conv(irg, block, op_left, mode_right); + right = op_right; + changed |= 1; + } else if (smaller_mode(mode_right, mode_left)) { + left = op_left; + right = new_r_Conv(irg, block, op_right, mode_left); + changed |= 1; + } + } + } - proj_nr = get_inversed_pnc(proj_nr); - changed |= 1; - } else if (get_irn_idx(left) > get_irn_idx(right)) { - ir_node *t = left; + /* TODO extend to arbitrary constants */ + if (is_Conv(left) && is_Const(right) && is_Const_null(right)) { + ir_mode* mode = get_irn_mode(left); + ir_node* op = get_Conv_op(left); + ir_mode* op_mode = get_irn_mode(op); + + if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) && + (mode_is_signed(mode) || !mode_is_signed(op_mode))) { + ir_node *null = new_Const(op_mode, get_mode_null(op_mode)); + set_Cmp_left( n, op); + set_Cmp_right(n, null); + return proj; + } + } - left = right; - right = t; + /* remove Casts */ + if (is_Cast(left)) + left = get_Cast_op(left); + if (is_Cast(right)) + right = get_Cast_op(right); + + /* remove operation of both sides if possible */ + if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) { + ir_opcode lop = get_irn_opcode(left); + + if (lop == get_irn_opcode(right)) { + ir_node *ll, *lr, *rl, *rr; + + /* same operation on both sides, try to remove */ + switch (lop) { + case iro_Not: + case iro_Minus: + /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */ + left = get_unop_op(left); + right = get_unop_op(right); + changed |= 1; + break; + case iro_Add: + ll = get_Add_left(left); + lr = get_Add_right(left); + rl = get_Add_left(right); + rr = get_Add_right(right); + + if (ll == rl) { + /* X + a CMP X + b ==> a CMP b */ + left = lr; + right = rr; + changed |= 1; + } else if (ll == rr) { + /* X + a CMP b + X ==> a CMP b */ + left = lr; + right = rl; + changed |= 1; + } else if (lr == rl) { + /* a + X CMP X + b ==> a CMP b */ + left = ll; + right = rr; + changed |= 1; + } else if (lr == rr) { + /* a + X CMP b + X ==> a CMP b */ + left = ll; + right = rl; + changed |= 1; + } + break; + case iro_Sub: + ll = get_Sub_left(left); + lr = get_Sub_right(left); + rl = get_Sub_left(right); + rr = get_Sub_right(right); + + if (ll == rl) { + /* X - a CMP X - b ==> a CMP b */ + left = lr; + right = rr; + changed |= 1; + } else if (lr == rr) { + /* a - X CMP b - X ==> a CMP b */ + left = ll; + right = rl; + changed |= 1; + } + break; + case iro_Rot: + if (get_Rot_right(left) == get_Rot_right(right)) { + /* a ROT X CMP b ROT X */ + left = get_Rot_left(left); + right = get_Rot_left(right); + changed |= 1; + } + break; + default: + break; + } + } + } - proj_nr = get_inversed_pnc(proj_nr); - changed |= 1; + if (get_irn_mode(left) == mode_b) { + ir_graph *irg = current_ir_graph; + ir_node *block = get_nodes_block(n); + + switch (proj_nr) { + case pn_Cmp_Le: return new_r_Or( irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); + case pn_Cmp_Lt: return new_r_And(irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); + case pn_Cmp_Ge: return new_r_Or( irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); + case pn_Cmp_Gt: return new_r_And(irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); + case pn_Cmp_Lg: return new_r_Eor(irg, block, left, right, mode_b); + case pn_Cmp_Eq: return new_r_Not(irg, block, new_r_Eor(irg, block, left, right, mode_b), mode_b); } + } - /* - * Second step: Try to reduce the magnitude - * of a constant. This may help to generate better code - * later and may help to normalize more compares. - * Of course this is only possible for integer values. - */ - if (c) { - mode = get_irn_mode(c); - tv = get_Const_tarval(c); + if (!get_opt_reassociation()) + return proj; - if (tv != tarval_bad) { - /* the following optimization is possible on modes without Overflow - * on Unary Minus or on == and !=: - * -a CMP c ==> a swap(CMP) -c - * - * Beware: for two-complement Overflow may occur, so only == and != can - * be optimized, see this: - * -MININT < 0 =/=> MININT > 0 !!! - */ - if (get_opt_constant_folding() && get_irn_op(left) == op_Minus && - (!mode_overflow_on_unary_Minus(mode) || - (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) { - left = get_Minus_op(left); - tv = tarval_sub(get_mode_null(mode), tv); + /* + * First step: normalize the compare op + * by placing the constant on the right side + * or moving the lower address node to the left. + * We ignore the case that both are constants + * this case should be optimized away. + */ + if (is_Const(right)) { + c = right; + } else if (is_Const(left)) { + c = left; + left = right; + right = c; + + proj_nr = get_inversed_pnc(proj_nr); + changed |= 1; + } else if (get_irn_idx(left) > get_irn_idx(right)) { + ir_node *t = left; + + left = right; + right = t; + + proj_nr = get_inversed_pnc(proj_nr); + changed |= 1; + } + + /* + * Second step: Try to reduce the magnitude + * of a constant. This may help to generate better code + * later and may help to normalize more compares. + * Of course this is only possible for integer values. + */ + if (c) { + mode = get_irn_mode(c); + tv = get_Const_tarval(c); + + if (tv != tarval_bad) { + /* the following optimization is possible on modes without Overflow + * on Unary Minus or on == and !=: + * -a CMP c ==> a swap(CMP) -c + * + * Beware: for two-complement Overflow may occur, so only == and != can + * be optimized, see this: + * -MININT < 0 =/=> MININT > 0 !!! + */ + if (is_Minus(left) && + (!mode_overflow_on_unary_Minus(mode) || + (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) { + tv = tarval_neg(tv); + if (tv != tarval_bad) { + left = get_Minus_op(left); proj_nr = get_inversed_pnc(proj_nr); changed |= 2; } + } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) { + tv = tarval_not(tv); - /* for integer modes, we have more */ - 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_Uo) { - proj_nr &= ~pn_Cmp_Uo; - set_Proj_proj(proj, proj_nr); - } + if (tv != tarval_bad) { + left = get_Not_op(left); + changed |= 2; + } + } - /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */ - if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) && - tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) { - tv = tarval_sub(tv, get_mode_one(mode)); + /* for integer modes, we have more */ + 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_Uo) { + proj_nr &= ~pn_Cmp_Uo; + set_Proj_proj(proj, proj_nr); + } + + /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */ + if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) && + tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) { + tv = tarval_sub(tv, get_mode_one(mode)); + if (tv != tarval_bad) { proj_nr ^= pn_Cmp_Eq; changed |= 2; } - /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */ - else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) && - tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) { - tv = tarval_add(tv, get_mode_one(mode)); + } + /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */ + else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) && + tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) { + tv = tarval_add(tv, get_mode_one(mode)); + if (tv != tarval_bad) { proj_nr ^= pn_Cmp_Eq; changed |= 2; } + } - /* the following reassociations work only for == and != */ - if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) { + /* the following reassociations work only for == and != */ + if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) { - /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */ - if (classify_tarval(tv) == TV_CLASSIFY_NULL && get_irn_op(left) == op_Sub) { - right = get_Sub_right(left); - left = get_Sub_left(left); + /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */ + if (tarval_is_null(tv) && is_Sub(left)) { + right = get_Sub_right(left); + left = get_Sub_left(left); - tv = value_of(right); - changed = 1; - } + tv = value_of(right); + changed = 1; + } - if (tv != tarval_bad) { - ir_op *op = get_irn_op(left); + if (tv != tarval_bad) { + /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */ + if (is_Sub(left)) { + ir_node *c1 = get_Sub_right(left); + tarval *tv2 = value_of(c1); - /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */ - if (op == op_Sub) { - ir_node *c1 = get_Sub_right(left); - tarval *tv2 = value_of(c1); + if (tv2 != tarval_bad) { + tv2 = tarval_add(tv, value_of(c1)); if (tv2 != tarval_bad) { - tv2 = tarval_add(tv, value_of(c1)); - - if (tv2 != tarval_bad) { - left = get_Sub_left(left); - tv = tv2; - changed |= 2; - } + left = get_Sub_left(left); + tv = tv2; + changed |= 2; } } - /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */ - else if (op == op_Add) { - ir_node *a_l = get_Add_left(left); - ir_node *a_r = get_Add_right(left); - ir_node *a; - tarval *tv2; - - if (get_irn_op(a_l) == op_Const) { - a = a_r; - tv2 = value_of(a_l); - } else { - a = a_l; - tv2 = value_of(a_r); - } - - if (tv2 != tarval_bad) { - tv2 = tarval_sub(tv, tv2); - - if (tv2 != tarval_bad) { - left = a; - tv = tv2; - changed |= 2; - } - } + } + /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */ + else if (is_Add(left)) { + ir_node *a_l = get_Add_left(left); + ir_node *a_r = get_Add_right(left); + ir_node *a; + tarval *tv2; + + if (is_Const(a_l)) { + a = a_r; + tv2 = value_of(a_l); + } else { + a = a_l; + tv2 = value_of(a_r); } - /* -a == c ==> a == -c, -a != c ==> a != -c */ - else if (op == op_Minus) { - tarval *tv2 = tarval_sub(get_mode_null(mode), tv); + + if (tv2 != tarval_bad) { + tv2 = tarval_sub(tv, tv2); if (tv2 != tarval_bad) { - left = get_Minus_op(left); + left = a; tv = tv2; changed |= 2; } } } - } /* == or != */ - /* the following reassociations work only for <= */ - else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) { - if (tv != tarval_bad) { - ir_op *op = get_irn_op(left); - - /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */ - if (op == op_Abs) { + /* -a == c ==> a == -c, -a != c ==> a != -c */ + else if (is_Minus(left)) { + tarval *tv2 = tarval_sub(get_mode_null(mode), tv); + + if (tv2 != tarval_bad) { + left = get_Minus_op(left); + tv = tv2; + changed |= 2; } } } - } /* 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; + } /* == or != */ + /* the following reassociations work only for <= */ + else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) { + if (tv != tarval_bad) { + /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */ + if (get_irn_op(left) == op_Abs) { // TODO something is missing here } } } - } /* tarval != bad */ - } + } /* 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) && is_And(left)) { + if (tarval_is_single_bit(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 & 2) /* need a new Const */ + right = new_Const(mode, tv); + + if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) { + ir_node *op = get_Proj_pred(left); - if (changed) { - ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */ + if ((is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) || + (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) { + ir_node *c = get_binop_right(op); - if (changed & 2) /* need a new Const */ - right = new_Const(mode, tv); + if (is_Const(c)) { + tarval *tv = get_Const_tarval(c); - /* create a new compare */ - n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block, - left, right); + if (tarval_is_single_bit(tv)) { + /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */ + ir_node *v = get_binop_left(op); + ir_node *blk = get_irn_n(op, -1); + ir_mode *mode = get_irn_mode(v); - set_Proj_pred(proj, n); - set_Proj_proj(proj, proj_nr); + tv = tarval_sub(tv, get_mode_one(mode)); + left = new_rd_And(get_irn_dbg_info(op), current_ir_graph, blk, v, new_Const(mode, tv), mode); + changed |= 1; + } + } } } + + if (changed) { + ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */ + + /* create a new compare */ + n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block, left, right); + + set_Proj_pred(proj, n); + set_Proj_proj(proj, proj_nr); + } + return proj; } /* transform_node_Proj_Cmp */ @@ -2858,7 +3925,7 @@ static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) { assert(is_op_commutative(get_irn_op(binop))); - if (get_irn_op(op_a) == op_Const) { + if (is_Const(op_a)) { *a = op_b; *c = op_a; } else { @@ -2876,6 +3943,11 @@ static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) { * OR c2 ===> OR * AND c1 * OR + * + * + * value c2 value c1 + * AND c1 ===> OR if (c1 | c2) == 0x111..11 + * OR */ static ir_node *transform_node_Or_bf_store(ir_node *or) { ir_node *and, *c1; @@ -2887,63 +3959,78 @@ static ir_node *transform_node_Or_bf_store(ir_node *or) { tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2; - get_comm_Binop_Ops(or, &and, &c1); - if ((get_irn_op(c1) != op_Const) || (get_irn_op(and) != op_And)) - return or; + while (1) { + get_comm_Binop_Ops(or, &and, &c1); + if (!is_Const(c1) || !is_And(and)) + return or; - get_comm_Binop_Ops(and, &or_l, &c2); - if ((get_irn_op(c2) != op_Const) || (get_irn_op(or_l) != op_Or)) - return or; + get_comm_Binop_Ops(and, &or_l, &c2); + if (!is_Const(c2)) + return or; - get_comm_Binop_Ops(or_l, &and_l, &c3); - if ((get_irn_op(c3) != op_Const) || (get_irn_op(and_l) != op_And)) - return or; + tv1 = get_Const_tarval(c1); + tv2 = get_Const_tarval(c2); - get_comm_Binop_Ops(and_l, &value, &c4); - if (get_irn_op(c4) != op_Const) - return or; + tv = tarval_or(tv1, tv2); + if (tarval_is_all_one(tv)) { + /* the AND does NOT clear a bit with isn't set by the OR */ + set_Or_left(or, or_l); + set_Or_right(or, c1); - /* ok, found the pattern, check for conditions */ - assert(mode == get_irn_mode(and)); - assert(mode == get_irn_mode(or_l)); - assert(mode == get_irn_mode(and_l)); + /* check for more */ + continue; + } - tv1 = get_Const_tarval(c1); - tv2 = get_Const_tarval(c2); - tv3 = get_Const_tarval(c3); - tv4 = get_Const_tarval(c4); + if (!is_Or(or_l)) + return or; - tv = tarval_or(tv4, tv2); - if (classify_tarval(tv) != TV_CLASSIFY_ALL_ONE) { - /* have at least one 0 at the same bit position */ - return or; - } + get_comm_Binop_Ops(or_l, &and_l, &c3); + if (!is_Const(c3) || !is_And(and_l)) + return or; - n_tv4 = tarval_not(tv4); - if (tv3 != tarval_and(tv3, n_tv4)) { - /* bit in the or_mask is outside the and_mask */ - return or; - } + get_comm_Binop_Ops(and_l, &value, &c4); + if (!is_Const(c4)) + return or; - n_tv2 = tarval_not(tv2); - if (tv1 != tarval_and(tv1, n_tv2)) { - /* bit in the or_mask is outside the and_mask */ - return or; - } + /* ok, found the pattern, check for conditions */ + assert(mode == get_irn_mode(and)); + assert(mode == get_irn_mode(or_l)); + assert(mode == get_irn_mode(and_l)); + + tv3 = get_Const_tarval(c3); + tv4 = get_Const_tarval(c4); + + tv = tarval_or(tv4, tv2); + if (!tarval_is_all_one(tv)) { + /* have at least one 0 at the same bit position */ + return or; + } + + n_tv4 = tarval_not(tv4); + if (tv3 != tarval_and(tv3, n_tv4)) { + /* bit in the or_mask is outside the and_mask */ + return or; + } - /* ok, all conditions met */ - block = get_irn_n(or, -1); + n_tv2 = tarval_not(tv2); + if (tv1 != tarval_and(tv1, n_tv2)) { + /* bit in the or_mask is outside the and_mask */ + return or; + } - new_and = new_r_And(current_ir_graph, block, - value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode); + /* ok, all conditions met */ + 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); - new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1)); + new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1)); - set_Or_left(or, new_and); - set_Or_right(or, new_const); + set_Or_left(or, new_and); + set_Or_right(or, new_const); - /* check for more */ - return transform_node_Or_bf_store(or); + /* check for more */ + } } /* transform_node_Or_bf_store */ /** @@ -2961,16 +4048,16 @@ static ir_node *transform_node_Or_Rot(ir_node *or) { shl = get_binop_left(or); shr = get_binop_right(or); - if (get_irn_op(shl) == op_Shr) { - if (get_irn_op(shr) != op_Shl) + if (is_Shr(shl)) { + if (!is_Shl(shr)) return or; irn = shl; shl = shr; shr = irn; - } else if (get_irn_op(shl) != op_Shl) { + } else if (!is_Shl(shl)) { return or; - } else if (get_irn_op(shr) != op_Shr) { + } else if (!is_Shr(shr)) { return or; } x = get_Shl_left(shl); @@ -2979,7 +4066,7 @@ static ir_node *transform_node_Or_Rot(ir_node *or) { c1 = get_Shl_right(shl); c2 = get_Shr_right(shr); - if (get_irn_op(c1) == op_Const && get_irn_op(c2) == op_Const) { + if (is_Const(c1) && is_Const(c2)) { tv1 = get_Const_tarval(c1); if (! tarval_is_long(tv1)) return or; @@ -2999,7 +4086,7 @@ static ir_node *transform_node_Or_Rot(ir_node *or) { DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT); return n; - } else if (get_irn_op(c1) == op_Sub) { + } else if (is_Sub(c1)) { v = c2; sub = c1; @@ -3007,7 +4094,7 @@ static ir_node *transform_node_Or_Rot(ir_node *or) { return or; c1 = get_Sub_left(sub); - if (get_irn_op(c1) != op_Const) + if (!is_Const(c1)) return or; tv1 = get_Const_tarval(c1); @@ -3025,12 +4112,12 @@ static ir_node *transform_node_Or_Rot(ir_node *or) { DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT); return n; - } else if (get_irn_op(c2) == op_Sub) { + } else if (is_Sub(c2)) { v = c1; sub = c2; c1 = get_Sub_left(sub); - if (get_irn_op(c1) != op_Const) + if (!is_Const(c1)) return or; tv1 = get_Const_tarval(c1); @@ -3061,10 +4148,44 @@ static ir_node *transform_node_Or(ir_node *n) { ir_node *a = get_Or_left(n); ir_node *b = get_Or_right(n); + if (is_Not(a) && is_Not(b)) { + /* ~a | ~b = ~(a&b) */ + ir_node *block = get_nodes_block(n); + ir_mode *mode = get_irn_mode(n); + + a = get_Not_op(a); + b = get_Not_op(b); + n = new_rd_And(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode); + n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN); + return n; + } + + /* we can evaluate 2 Projs of the same Cmp */ + if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) { + ir_node *pred_a = get_Proj_pred(a); + ir_node *pred_b = get_Proj_pred(b); + if (pred_a == pred_b) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(pred_a); + pn_Cmp pn_a = get_Proj_proj(a); + pn_Cmp pn_b = get_Proj_proj(b); + /* yes, we can simply calculate with pncs */ + pn_Cmp new_pnc = pn_a | pn_b; + + return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b, + new_pnc); + } + } + HANDLE_BINOP_PHI(tarval_or, a,b,c); n = transform_node_Or_bf_store(n); n = transform_node_Or_Rot(n); + if (n != oldn) + return n; + + n = transform_bitwise_distributive(n, transform_node_Or); return n; } /* transform_node_Or */ @@ -3172,20 +4293,49 @@ static ir_node *transform_node_Shl(ir_node *n) { * 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); + int i, j, n_keepalives = get_End_n_keepalives(n); + ir_node **in; + + NEW_ARR_A(ir_node *, in, n_keepalives); - for (i = 0; i < n_keepalives; ++i) { + for (i = j = 0; i < n_keepalives; ++i) { ir_node *ka = get_End_keepalive(n, i); if (is_Block(ka)) { - if (is_Block_dead(ka)) { - set_End_keepalive(n, i, new_Bad()); + if (! is_Block_dead(ka)) { + in[j++] = ka; } - } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) - set_End_keepalive(n, i, new_Bad()); + continue; + } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) { + continue; + } + /* FIXME: beabi need to keep a Proj(M) */ + if (is_Phi(ka) || is_irn_keep(ka) || is_Proj(ka)) + in[j++] = ka; } + if (j != n_keepalives) + set_End_keepalives(n, j, in); return n; } /* transform_node_End */ +/** returns 1 if a == -b */ +static int is_negated_value(ir_node *a, ir_node *b) { + if(is_Minus(a) && get_Minus_op(a) == b) + return 1; + if(is_Minus(b) && get_Minus_op(b) == a) + return 1; + if(is_Sub(a) && is_Sub(b)) { + ir_node *a_left = get_Sub_left(a); + ir_node *a_right = get_Sub_right(a); + ir_node *b_left = get_Sub_left(b); + ir_node *b_right = get_Sub_right(b); + + if(a_left == b_right && a_right == b_left) + return 1; + } + + return 0; +} + /** * Optimize a Mux into some simpler cases. */ @@ -3193,122 +4343,88 @@ static ir_node *transform_node_Mux(ir_node *n) { ir_node *oldn = n, *sel = get_Mux_sel(n); ir_mode *mode = get_irn_mode(n); - if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(mode)) { - ir_node *cmp = get_Proj_pred(sel); - long proj_nr = get_Proj_proj(sel); - ir_node *f = get_Mux_false(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_irn_n(n, -1); - - /* - * Note: normalization puts the constant on the right site, - * so we check only one case. - * - * Note further that these optimization work even for floating point - * with NaN's because -NaN == NaN. - * However, if +0 and -0 is handled differently, we cannot use the first one. - */ - if (get_irn_op(f) == op_Minus && - get_Minus_op(f) == t && - get_Cmp_left(cmp) == t) { - - if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) { - /* Mux(a >=/> 0, -a, a) ==> Abs(a) */ - n = new_rd_Abs(get_irn_dbg_info(n), - current_ir_graph, - block, - t, mode); - 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) { - /* Mux(a <=/< 0, -a, a) ==> Minus(Abs(a)) */ - n = new_rd_Abs(get_irn_dbg_info(n), - current_ir_graph, - block, - t, mode); - n = new_rd_Minus(get_irn_dbg_info(n), - current_ir_graph, - block, - n, mode); - - DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS); - return n; + if (mode == mode_b) { + ir_node *t = get_Mux_true(n); + ir_node *f = get_Mux_false(n); + dbg_info *dbg = get_irn_dbg_info(n); + ir_node *block = get_irn_n(n, -1); + ir_graph *irg = current_ir_graph; + + if (is_Const(t)) { + tarval *tv_t = get_Const_tarval(t); + if (tv_t == tarval_b_true) { + if (is_Const(f)) { + assert(get_Const_tarval(f) == tarval_b_false); + return sel; + } else { + return new_rd_Or(dbg, irg, block, sel, f, mode_b); } - } else if (get_irn_op(t) == op_Minus && - get_Minus_op(t) == f && - get_Cmp_left(cmp) == f) { - - if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) { - /* Mux(a <=/< 0, a, -a) ==> Abs(a) */ - n = new_rd_Abs(get_irn_dbg_info(n), - current_ir_graph, - block, - f, mode); - 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) { - /* Mux(a >=/> 0, a, -a) ==> Minus(Abs(a)) */ - n = new_rd_Abs(get_irn_dbg_info(n), - current_ir_graph, - block, - f, mode); - n = new_rd_Minus(get_irn_dbg_info(n), - current_ir_graph, - block, - n, mode); - - DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS); - return n; + } else { + ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b); + assert(tv_t == tarval_b_false); + if (is_Const(f)) { + assert(get_Const_tarval(f) == tarval_b_true); + return not_sel; + } else { + return new_rd_And(dbg, irg, block, not_sel, f, mode_b); } } + } else if (is_Const(f)) { + tarval *tv_f = get_Const_tarval(f); + if (tv_f == tarval_b_true) { + ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b); + return new_rd_Or(dbg, irg, block, not_sel, t, mode_b); + } else { + assert(tv_f == tarval_b_false); + return new_rd_And(dbg, irg, block, sel, t, mode_b); + } + } + } - if (mode_is_int(mode) && mode_is_signed(mode) && - get_mode_arithmetic(mode) == irma_twos_complement) { - ir_node *x = get_Cmp_left(cmp); - - /* the following optimization works only with signed integer two-complement mode */ + if (is_Proj(sel) && !mode_honor_signed_zeros(mode)) { + ir_node *cmp = get_Proj_pred(sel); + long pn = get_Proj_proj(sel); + ir_node *f = get_Mux_false(n); + ir_node *t = get_Mux_true(n); - if (mode == get_irn_mode(x)) { - /* - * FIXME: this restriction is two rigid, as it would still - * work if mode(x) = Hs and mode == Is, but at least it removes - * all wrong cases. - */ - if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Le) && - classify_Const(t) == CNST_ALL_ONE && - classify_Const(f) == CNST_NULL) { - /* - * Mux(x:T Shrs(x, sizeof_bits(T) - 1) - * Conditions: - * T must be signed. - */ - n = new_rd_Shrs(get_irn_dbg_info(n), - current_ir_graph, block, x, - new_r_Const_long(current_ir_graph, block, mode_Iu, - get_mode_size_bits(mode) - 1), - mode); - DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR); + /* + * Note: normalization puts the constant on the right side, + * so we check only one case. + * + * Note further that these optimization work even for floating point + * with NaN's because -NaN == NaN. + * However, if +0 and -0 is handled differently, we cannot use the first + * one. + */ + if (is_Cmp(cmp)) { + ir_node *cmp_r = get_Cmp_right(cmp); + if (is_Const(cmp_r) && is_Const_null(cmp_r)) { + ir_node *block = get_irn_n(n, -1); + + if(is_negated_value(f, t)) { + ir_node *cmp_left = get_Cmp_left(cmp); + + /* Psi(a >= 0, a, -a) = Psi(a <= 0, -a, a) ==> Abs(a) */ + if ( (cmp_left == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)) + || (cmp_left == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))) + { + n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block, + cmp_left, mode); + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS); return n; - } else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Ge) && - classify_Const(t) == CNST_ONE && - classify_Const(f) == CNST_NULL) { - /* - * Mux(x:T >/>= 0, 0, 1) -> Shr(-x, sizeof_bits(T) - 1) - * Conditions: - * T must be signed. - */ - n = new_rd_Shr(get_irn_dbg_info(n), - current_ir_graph, block, - new_r_Minus(current_ir_graph, block, x, mode), - new_r_Const_long(current_ir_graph, block, mode_Iu, - get_mode_size_bits(mode) - 1), - mode); - DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR); + /* Psi(a <= 0, a, -a) = Psi(a >= 0, -a, a) ==> -Abs(a) */ + } else if ((cmp_left == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)) + || (cmp_left == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))) + { + n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block, + cmp_left, mode); + n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, + block, n, mode); + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS); return n; } } + } } } @@ -3325,6 +4441,42 @@ static ir_node *transform_node_Psi(ir_node *n) { return n; } /* transform_node_Psi */ +/** + * optimize sync nodes that have other syncs as input we simply add the inputs + * of the other sync to our own inputs + */ +static ir_node *transform_node_Sync(ir_node *n) { + int i, arity; + + arity = get_irn_arity(n); + for(i = 0; i < get_irn_arity(n); /*empty*/) { + int i2, arity2; + ir_node *in = get_irn_n(n, i); + if(!is_Sync(in)) { + ++i; + continue; + } + + /* set sync input 0 instead of the sync */ + set_irn_n(n, i, get_irn_n(in, 0)); + /* so we check this input again for syncs */ + + /* append all other inputs of the sync to our sync */ + arity2 = get_irn_arity(in); + for(i2 = 1; i2 < arity2; ++i2) { + ir_node *in_in = get_irn_n(in, i2); + add_irn_n(n, in_in); + /* increase arity so we also check the new inputs for syncs */ + arity++; + } + } + + /* rehash the sync node */ + add_identities(current_ir_graph->value_table, n); + + return n; +} + /** * Tries several [inplace] [optimizing] transformations and returns an * equivalent node. The difference to equivalent_node() is that these @@ -3332,8 +4484,19 @@ static ir_node *transform_node_Psi(ir_node *n) { * not be freed even if the equivalent node isn't the old one. */ static ir_node *transform_node(ir_node *n) { - if (n->op->ops.transform_node) - n = n->op->ops.transform_node(n); + ir_node *oldn; + + /* + * Transform_node is the only "optimizing transformation" that might + * return a node with a different opcode. We iterate HERE until fixpoint + * to get the final result. + */ + do { + oldn = n; + if (n->op->ops.transform_node) + n = n->op->ops.transform_node(n); + } while (oldn != n); + return n; } /* transform_node */ @@ -3360,6 +4523,7 @@ static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops CASE(Div); CASE(Mod); CASE(DivMod); + CASE(Quot); CASE(Abs); CASE(Cond); CASE(And); @@ -3377,6 +4541,7 @@ static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops CASE(End); CASE(Mux); CASE(Psi); + CASE(Sync); default: /* leave NULL */; } @@ -3400,7 +4565,7 @@ static int node_cmp_attr_Const(ir_node *a, ir_node *b) { /** Compares the attributes of two Proj nodes. */ static int node_cmp_attr_Proj(ir_node *a, ir_node *b) { - return get_irn_proj_attr (a) != get_irn_proj_attr (b); + return get_irn_proj_attr(a) != get_irn_proj_attr(b); } /* node_cmp_attr_Proj */ /** Compares the attributes of two Filter nodes. */ @@ -3410,21 +4575,25 @@ static int node_cmp_attr_Filter(ir_node *a, ir_node *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); + const alloc_attr *pa = get_irn_alloc_attr(a); + const alloc_attr *pb = get_irn_alloc_attr(b); + return (pa->where != pb->where) || (pa->type != pb->type); } /* node_cmp_attr_Alloc */ /** 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); + const free_attr *pa = get_irn_free_attr(a); + const free_attr *pb = get_irn_free_attr(b); + return (pa->where != pb->where) || (pa->type != pb->type); } /* node_cmp_attr_Free */ /** 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) - || (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); + const symconst_attr *pa = get_irn_symconst_attr(a); + const symconst_attr *pb = get_irn_symconst_attr(b); + return (pa->num != pb->num) + || (pa->sym.type_p != pb->sym.type_p) + || (pa->tp != pb->tp); } /* node_cmp_attr_SymConst */ /** Compares the attributes of two Call nodes. */ @@ -3434,16 +4603,25 @@ static int node_cmp_attr_Call(ir_node *a, ir_node *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) - || (get_irn_sel_attr(a).ent->name != get_irn_sel_attr(b).ent->name) - || (get_irn_sel_attr(a).ent->owner != get_irn_sel_attr(b).ent->owner) - || (get_irn_sel_attr(a).ent->ld_name != get_irn_sel_attr(b).ent->ld_name) - || (get_irn_sel_attr(a).ent->type != get_irn_sel_attr(b).ent->type); + const ir_entity *a_ent = get_Sel_entity(a); + const ir_entity *b_ent = get_Sel_entity(b); + return + (a_ent->kind != b_ent->kind) || + (a_ent->name != b_ent->name) || + (a_ent->owner != b_ent->owner) || + (a_ent->ld_name != b_ent->ld_name) || + (a_ent->type != b_ent->type); } /* node_cmp_attr_Sel */ /** 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); + /* we can only enter this function if both nodes have the same number of inputs, + hence it is enough to check if one of them is a Phi0 */ + if (is_Phi0(a)) { + /* check the Phi0 attribute */ + return get_irn_phi0_attr(a) != get_irn_phi0_attr(b); + } + return 0; } /* node_cmp_attr_Phi */ /** Compares the attributes of two Conv nodes. */ @@ -3462,12 +4640,19 @@ static int node_cmp_attr_Load(ir_node *a, ir_node *b) { get_Load_volatility(b) == volatility_is_volatile) /* NEVER do CSE on volatile Loads */ return 1; + /* do not CSE Loads with different alignment. Be conservative. */ + if (get_Load_align(a) != get_Load_align(b)) + return 1; return get_Load_mode(a) != get_Load_mode(b); } /* node_cmp_attr_Load */ /** Compares the attributes of two Store nodes. */ static int node_cmp_attr_Store(ir_node *a, ir_node *b) { + /* do not CSE Stores with different alignment. Be conservative. */ + if (get_Store_align(a) != get_Store_align(b)) + return 1; + /* NEVER do CSE on volatile Stores */ return (get_Store_volatility(a) == volatility_is_volatile || get_Store_volatility(b) == volatility_is_volatile); @@ -3478,6 +4663,52 @@ static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) { return (get_Confirm_cmp(a) != get_Confirm_cmp(b)); } /* node_cmp_attr_Confirm */ +/** Compares the attributes of two ASM nodes. */ +static int node_cmp_attr_ASM(ir_node *a, ir_node *b) { + int i, n; + const ir_asm_constraint *ca; + const ir_asm_constraint *cb; + ident **cla, **clb; + + if (get_ASM_text(a) != get_ASM_text(b)) + return 1; + + /* Should we really check the constraints here? Should be better, but is strange. */ + n = get_ASM_n_input_constraints(a); + if (n != get_ASM_n_input_constraints(b)) + return 0; + + ca = get_ASM_input_constraints(a); + cb = get_ASM_input_constraints(b); + for (i = 0; i < n; ++i) { + if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint) + return 1; + } + + n = get_ASM_n_output_constraints(a); + if (n != get_ASM_n_output_constraints(b)) + return 0; + + ca = get_ASM_output_constraints(a); + cb = get_ASM_output_constraints(b); + for (i = 0; i < n; ++i) { + if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint) + return 1; + } + + n = get_ASM_n_clobbers(a); + if (n != get_ASM_n_clobbers(b)) + return 0; + + cla = get_ASM_clobbers(a); + clb = get_ASM_clobbers(b); + for (i = 0; i < n; ++i) { + if (cla[i] != clb[i]) + return 1; + } + return 0; +} /* node_cmp_attr_ASM */ + /** * Set the default node attribute compare operation for an ir_op_ops. * @@ -3509,6 +4740,7 @@ static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops) CASE(Load); CASE(Store); CASE(Confirm); + CASE(ASM); default: /* leave NULL */; } @@ -3601,6 +4833,35 @@ void del_identities(pset *value_table) { del_pset(value_table); } /* del_identities */ +/** + * Normalize a node by putting constants (and operands with smaller + * node index) on the right + * + * @param n The node to normalize + */ +static void normalize_node(ir_node *n) { + if (get_opt_reassociation()) { + if (is_op_commutative(get_irn_op(n))) { + ir_node *l = get_binop_left(n); + ir_node *r = get_binop_right(n); + int l_idx = get_irn_idx(l); + int r_idx = get_irn_idx(r); + + /* For commutative operators perform a OP b == b OP a but keep + constants on the RIGHT side. This helps greatly in some optimizations. + Moreover we use the idx number to make the form deterministic. */ + if (is_irn_constlike(l)) + l_idx = -l_idx; + if (is_irn_constlike(r)) + r_idx = -r_idx; + if (l_idx < r_idx) { + set_binop_left(n, r); + set_binop_right(n, l); + } + } + } +} /* normalize_node */ + /** * Return the canonical node computing the same value as n. * @@ -3618,18 +4879,7 @@ static INLINE ir_node *identify(pset *value_table, ir_node *n) { if (!value_table) return n; - if (get_opt_reassociation()) { - if (is_op_commutative(get_irn_op(n))) { - ir_node *l = get_binop_left(n); - ir_node *r = get_binop_right(n); - - /* for commutative operators perform a OP b == b OP a */ - if (get_irn_idx(l) > get_irn_idx(r)) { - set_binop_left(n, r); - set_binop_right(n, l); - } - } - } + normalize_node(n); o = pset_find(value_table, n, ir_node_hash(n)); if (!o) return n; @@ -3663,21 +4913,9 @@ ir_node *identify_remember(pset *value_table, ir_node *n) { if (!value_table) return n; - if (get_opt_reassociation()) { - if (is_op_commutative(get_irn_op(n))) { - ir_node *l = get_binop_left(n); - ir_node *r = get_binop_right(n); - - /* for commutative operators perform a OP b == b OP a */ - if (l > r) { - set_binop_left(n, r); - set_binop_right(n, l); - } - } - } - + normalize_node(n); /* lookup or insert in hash table with given hash key. */ - o = pset_insert (value_table, n, ir_node_hash (n)); + o = pset_insert(value_table, n, ir_node_hash(n)); if (o != n) { DBG_OPT_CSE(n, o); @@ -3707,7 +4945,7 @@ void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) { * 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 ir_node *gigo(ir_node *node) { int i, irn_arity; ir_op *op = get_irn_op(node); @@ -3802,6 +5040,8 @@ ir_node *optimize_node(ir_node *n) { if (get_opt_constant_folding()) { /* neither constants nor Tuple values can be evaluated */ if (iro != iro_Const && (get_irn_mode(n) != mode_T)) { + unsigned fp_model = get_irg_fp_model(current_ir_graph); + int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0); /* try to evaluate */ tv = computed_value(n); if (tv != tarval_bad) { @@ -3834,13 +5074,15 @@ ir_node *optimize_node(ir_node *n) { /* evaluation was successful -- replace the node. */ irg_kill_node(current_ir_graph, n); - nw = new_Const(get_tarval_mode (tv), tv); + nw = new_Const(get_tarval_mode(tv), tv); - if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv)) + if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv)) set_Const_type(nw, old_tp); DBG_OPT_CSTEVAL(oldn, nw); + tarval_enable_fp_ops(old_fp_mode); return nw; } + tarval_enable_fp_ops(old_fp_mode); } } @@ -3852,8 +5094,6 @@ ir_node *optimize_node(ir_node *n) { (iro == iro_Block) ) /* Flags tested local. */ n = equivalent_node(n); - optimize_preds(n); /* do node specific optimizations of nodes predecessors. */ - /* Common Subexpression Elimination. * * Checks whether n is already available. @@ -3903,12 +5143,14 @@ ir_node *optimize_in_place_2(ir_node *n) { ir_node *oldn = n; ir_opcode iro = get_irn_opcode(n); - if (!get_opt_optimize() && (get_irn_op(n) != op_Phi)) return n; + if (!get_opt_optimize() && !is_Phi(n)) return n; /* 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) { + unsigned fp_model = get_irg_fp_model(current_ir_graph); + int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0); /* try to evaluate */ tv = computed_value(n); if (tv != tarval_bad) { @@ -3928,8 +5170,10 @@ ir_node *optimize_in_place_2(ir_node *n) { set_Const_type(n, old_tp); DBG_OPT_CSTEVAL(oldn, n); + tarval_enable_fp_ops(old_fp_mode); return n; } + tarval_enable_fp_ops(old_fp_mode); } } @@ -3941,8 +5185,6 @@ ir_node *optimize_in_place_2(ir_node *n) { (iro == iro_Block) ) /* Flags tested local. */ n = equivalent_node(n); - optimize_preds(n); /* do node specific optimizations of nodes predecessors. */ - /** common subexpression elimination **/ /* Checks whether n is already available. */ /* The block input is used to distinguish different subexpressions. Right