X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firopt.c;h=698ba038408c4d5ca7608fe6c96943052094b0b4;hb=df144e685b61ac41bf4d82f172515aa81bb1a2f3;hp=ad79fce7250255dde564a3a2deecee85fa4a0c9e;hpb=055909393f5f1ae98f9f2a33e0fa39326d9d7393;p=libfirm diff --git a/ir/ir/iropt.c b/ir/ir/iropt.c index ad79fce72..698ba0384 100644 --- a/ir/ir/iropt.c +++ b/ir/ir/iropt.c @@ -1,28 +1,33 @@ /* - * Project: libFIRM - * File name: ir/ir/iropt.c - * Purpose: iropt --- optimizations intertwined with IR construction. - * Author: Christian Schaefer - * Modified by: Goetz Lindenmaier, Michael Beck - * Created: - * CVS-ID: $Id$ - * Copyright: (c) 1998-2006 Universität Karlsruhe - * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE. + * Copyright (C) 1995-2007 University of Karlsruhe. All right reserved. + * + * This file is part of libFirm. + * + * This file may be distributed and/or modified under the terms of the + * GNU General Public License version 2 as published by the Free Software + * Foundation and appearing in the file LICENSE.GPL included in the + * packaging of this file. + * + * Licensees holding valid libFirm Professional Edition licenses may use + * this file in accordance with the libFirm Commercial License. + * Agreement provided with the Software. + * + * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE + * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR + * PURPOSE. */ +/** + * @file + * @brief iropt --- optimizations intertwined with IR construction. + * @author Christian Schaefer, Goetz Lindenmaier, Michael Beck + * @version $Id$ + */ #ifdef HAVE_CONFIG_H # include "config.h" #endif -#ifdef HAVE_ALLOCA_H -#include -#endif -#ifdef HAVE_MALLOC_H -#include -#endif -#ifdef HAVE_STRING_H #include -#endif #include "irnode_t.h" #include "irgraph_t.h" @@ -40,9 +45,10 @@ #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" /* Make types visible to allow most efficient access */ #include "entity_t.h" @@ -947,40 +953,23 @@ 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) { + 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 */ @@ -1001,7 +990,7 @@ 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 */ @@ -1009,7 +998,7 @@ 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 @@ -1043,10 +1032,12 @@ static ir_node *equivalent_node_Div(ir_node *n) { if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */ /* Turn Div into a tuple (mem, bad, a) */ ir_node *mem = get_Div_mem(n); + 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 */ @@ -1060,12 +1051,14 @@ static ir_node *equivalent_node_Quot(ir_node *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) */ + /* 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 */ @@ -1074,20 +1067,22 @@ 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) */ + /* 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 */ @@ -1149,9 +1144,20 @@ 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)) */ @@ -1165,7 +1171,7 @@ 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); @@ -1315,29 +1321,48 @@ 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 (get_irn_op(a) == op_Tuple) { /* 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 (op == op_Load || op == op_Store) { + if (op == op_Load) { + /* get the Load address */ + ir_node *addr = get_Load_ptr(a); + ir_node *blk = get_irn_n(a, -1); + ir_node *confirm; + + 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 (op == op_Store) { /* 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)) { @@ -1345,14 +1370,17 @@ 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 */ /** @@ -1510,12 +1538,14 @@ 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 */ @@ -1531,7 +1561,7 @@ 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); @@ -1547,20 +1577,22 @@ 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 */ @@ -1880,7 +1912,7 @@ static ir_node *transform_node_Add(ir_node *n) { return n; if (mode_is_num(mode)) { - if (a == b) { + if (a == b && mode_is_int(mode)) { ir_node *block = get_irn_n(n, -1); n = new_rd_Mul( @@ -1972,6 +2004,15 @@ static ir_node *transform_node_Add(ir_node *n) { DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A); } } + /* Here we rely on constants be on the RIGHT side */ + else if (get_mode_arithmetic(mode) == irma_twos_complement && + is_Not(a) && classify_Const(b) == CNST_ONE) { + /* ~x + 1 = -x */ + ir_node *op = get_Not_op(a); + 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; } /* transform_node_Add */ @@ -1993,16 +2034,61 @@ 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_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 word 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); + } 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); + } + } + } else 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; + } + } + } /* 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)) { + else if (mode_is_num(mode) && mode == get_irn_mode(a) && (classify_Const(a) == CNST_NULL)) { n = new_rd_Minus( get_irn_dbg_info(n), current_ir_graph, @@ -2076,7 +2162,6 @@ static ir_node *transform_node_Sub(ir_node *n) { set_Sub_right(n, add); DBG_OPT_ALGSIM0(n, n, FS_OPT_SUB_SUB_X_Y_Z); } - return n; } /* transform_node_Sub */ @@ -2107,6 +2192,24 @@ static ir_node *transform_node_Mul(ir_node *n) { 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 */ @@ -2115,25 +2218,56 @@ 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 */ @@ -2143,25 +2277,56 @@ 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 */ @@ -2170,60 +2335,120 @@ 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 @@ -2293,15 +2518,123 @@ 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 + */ +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. */ static ir_node *transform_node_And(ir_node *n) { - ir_node *c, *oldn = n; + ir_node *c, *oldn; ir_node *a = get_And_left(n); ir_node *b = get_And_right(n); HANDLE_BINOP_PHI(tarval_and, a,b,c); + + n = transform_bitwise_distributive(n, transform_node_And); + return n; } /* transform_node_And */ @@ -2338,6 +2671,8 @@ static ir_node *transform_node_Eor(ir_node *n) { 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 { + n = transform_bitwise_distributive(n, transform_node_Eor); } return n; @@ -2349,30 +2684,54 @@ 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_op *op_a = get_irn_op(a); 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) + && (op_a == op_Proj) && (get_irn_mode(a) == mode_b) && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) { /* We negate a Cmp. The Cmp has the negated result anyways! */ n = new_r_Proj(current_ir_graph, get_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 (op_a == op_Sub && classify_Const(get_Sub_right(a)) == CNST_ONE) { + /* ~(x-1) = -x */ + ir_node *op = get_Sub_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 */ 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 && 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; } /* transform_node_Minus */ @@ -2402,29 +2761,34 @@ 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 */ 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; } } @@ -2436,38 +2800,46 @@ 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) { + 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); + set_irn_pinned(mod, 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, get_irg_no_mem(current_ir_graph)); 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; @@ -2478,38 +2850,47 @@ 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) { + 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); + set_irn_pinned(divmod, 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, get_irg_no_mem(current_ir_graph)); return res; - } else if (proj_nr == pn_DivMod_res_mod && get_DivMod_left(n) == b) { - /* a % a = 0 if a != 0 */ - ir_mode *mode = get_irn_mode(proj); - ir_node *res = new_Const(mode, get_mode_null(mode)); - DBG_OPT_CSTEVAL(n, res); - return res; + 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; @@ -2609,10 +2990,12 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { (!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); + tv = tarval_neg(tv); - proj_nr = get_inversed_pnc(proj_nr); - changed |= 2; + if (tv != tarval_bad) { + proj_nr = get_inversed_pnc(proj_nr); + changed |= 2; + } } /* for integer modes, we have more */ @@ -2629,16 +3012,20 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) { tv = tarval_sub(tv, get_mode_one(mode)); - proj_nr ^= pn_Cmp_Eq; - changed |= 2; + 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)); - proj_nr ^= pn_Cmp_Eq; - changed |= 2; + if (tv != tarval_bad) { + proj_nr ^= pn_Cmp_Eq; + changed |= 2; + } } /* the following reassociations work only for == and != */ @@ -2650,7 +3037,9 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { left = get_Sub_left(left); tv = value_of(right); - changed = 1; + if (tv != tarval_bad) { + changed = 1; + } } if (tv != tarval_bad) { @@ -2730,7 +3119,7 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { */ if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && (get_irn_op(left) == op_And)) { - if (is_single_bit_tarval(tv)) { + 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); @@ -2868,6 +3257,11 @@ 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; @@ -2879,63 +3273,78 @@ 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 (classify_tarval(tv) == TV_CLASSIFY_ALL_ONE) { + /* the AND does NOT clear a bit with isn't set be 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)); - /* ok, all conditions met */ - block = get_irn_n(or, -1); + tv3 = get_Const_tarval(c3); + tv4 = get_Const_tarval(c4); - new_and = new_r_And(current_ir_graph, block, - value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode); + tv = tarval_or(tv4, tv2); + if (classify_tarval(tv) != TV_CLASSIFY_ALL_ONE) { + /* have at least one 0 at the same bit position */ + return or; + } + + n_tv4 = tarval_not(tv4); + if (tv3 != tarval_and(tv3, n_tv4)) { + /* bit in the or_mask is outside the and_mask */ + return or; + } - new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1)); + n_tv2 = tarval_not(tv2); + if (tv1 != tarval_and(tv1, n_tv2)) { + /* bit in the or_mask is outside the and_mask */ + return or; + } + + /* ok, all conditions met */ + block = get_irn_n(or, -1); - set_Or_left(or, new_and); - set_Or_right(or, new_const); + new_and = new_r_And(current_ir_graph, block, + value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode); - /* check for more */ - return transform_node_Or_bf_store(or); + new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1)); + + set_Or_left(or, new_and); + set_Or_right(or, new_const); + + /* check for more */ + } } /* transform_node_Or_bf_store */ /** @@ -3057,6 +3466,10 @@ static ir_node *transform_node_Or(ir_node *n) { 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 */ @@ -3164,17 +3577,27 @@ 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 */ @@ -3185,6 +3608,44 @@ 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 (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 { + 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 (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); @@ -3324,8 +3785,19 @@ 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 */ @@ -3352,6 +3824,7 @@ 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); @@ -3392,7 +3865,7 @@ 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. */ @@ -3402,21 +3875,25 @@ 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. */ @@ -3426,16 +3903,25 @@ 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. */ @@ -3454,12 +3940,19 @@ 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); @@ -3470,6 +3963,52 @@ 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. * @@ -3501,6 +4040,7 @@ 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 */; } @@ -3659,9 +4199,17 @@ ir_node *identify_remember(pset *value_table, ir_node *n) { 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) { + 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); } @@ -3669,7 +4217,7 @@ ir_node *identify_remember(pset *value_table, ir_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);