X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fia32%2Fia32_optimize.c;h=0d434a7159d31bbec25b629450fd02cf8adaff39;hb=a1a465eb2b3f54027b29f829423fffd0396937f4;hp=22fb8781e82a27321e98edbf5d4ef5c725c5480c;hpb=8c98b3f4fd1281d86ed76478a630237367a72a2d;p=libfirm diff --git a/ir/be/ia32/ia32_optimize.c b/ir/be/ia32/ia32_optimize.c index 22fb8781e..0d434a715 100644 --- a/ir/be/ia32/ia32_optimize.c +++ b/ir/be/ia32/ia32_optimize.c @@ -1,3 +1,13 @@ +/** + * Project: libFIRM + * File name: ir/be/ia32/ia32_optimize.c + * Purpose: Implements several optimizations for IA32 + * Author: Christian Wuerdig + * CVS-ID: $Id$ + * Copyright: (c) 2006 Universität Karlsruhe + * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE. + */ + #ifdef HAVE_CONFIG_H #include "config.h" #endif @@ -9,6 +19,9 @@ #include "iredges.h" #include "tv.h" #include "irgmod.h" +#include "irgwalk.h" +#include "height.h" +#include "irbitset.h" #include "../be_t.h" #include "../beabi.h" @@ -18,20 +31,49 @@ #include "ia32_new_nodes.h" #include "bearch_ia32_t.h" #include "gen_ia32_regalloc_if.h" /* the generated interface (register type and class defenitions) */ +#include "ia32_transform.h" +#include "ia32_dbg_stat.h" +#include "ia32_util.h" + +#define AGGRESSIVE_AM + +typedef enum { + IA32_AM_CAND_NONE = 0, + IA32_AM_CAND_LEFT = 1, + IA32_AM_CAND_RIGHT = 2, + IA32_AM_CAND_BOTH = 3 +} ia32_am_cand_t; #undef is_NoMem #define is_NoMem(irn) (get_irn_op(irn) == op_NoMem) -static int be_is_NoReg(be_abi_irg_t *babi, const ir_node *irn) { - if (be_abi_get_callee_save_irn(babi, &ia32_gp_regs[REG_XXX]) == irn || - be_abi_get_callee_save_irn(babi, &ia32_fp_regs[REG_XXXX]) == irn) - { - return 1; - } +typedef int is_op_func_t(const ir_node *n); +typedef ir_node *load_func_t(dbg_info *db, ir_graph *irg, ir_node *block, ir_node *base, ir_node *index, ir_node *mem); - return 0; +/** + * checks if a node represents the NOREG value + */ +static int be_is_NoReg(ia32_code_gen_t *cg, const ir_node *irn) { + be_abi_irg_t *babi = cg->birg->abi; + const arch_register_t *fp_noreg = USE_SSE2(cg) ? + &ia32_xmm_regs[REG_XMM_NOREG] : &ia32_vfp_regs[REG_VFP_NOREG]; + + return (be_abi_get_callee_save_irn(babi, &ia32_gp_regs[REG_GP_NOREG]) == irn) || + (be_abi_get_callee_save_irn(babi, fp_noreg) == irn); } + + +/************************************************* + * _____ _ _ + * / ____| | | | | + * | | ___ _ __ ___| |_ __ _ _ __ | |_ ___ + * | | / _ \| '_ \/ __| __/ _` | '_ \| __/ __| + * | |___| (_) | | | \__ \ || (_| | | | | |_\__ \ + * \_____\___/|_| |_|___/\__\__,_|_| |_|\__|___/ + * + *************************************************/ + /** * creates a unique ident by adding a number to a tag * @@ -47,8 +89,6 @@ static ident *unique_id(const char *tag) return new_id_from_str(str); } - - /** * Transforms a SymConst. * @@ -59,14 +99,24 @@ static ident *unique_id(const char *tag) * @return the created ia32 Const node */ static ir_node *gen_SymConst(ia32_transform_env_t *env) { - ir_node *cnst; dbg_info *dbg = env->dbg; ir_mode *mode = env->mode; ir_graph *irg = env->irg; ir_node *block = env->block; + ir_node *cnst; + + if (mode_is_float(mode)) { + FP_USED(env->cg); + if (USE_SSE2(env->cg)) + cnst = new_rd_ia32_xConst(dbg, irg, block, mode); + else + cnst = new_rd_ia32_vfConst(dbg, irg, block, mode); + } + else + cnst = new_rd_ia32_Const(dbg, irg, block, mode); - cnst = new_rd_ia32_Const(dbg, irg, block, mode); set_ia32_Const_attr(cnst, env->irn); + return cnst; } @@ -95,7 +145,7 @@ static ir_type *get_prim_type(pmap *types, ir_mode *mode) static entity *get_entity_for_tv(ia32_code_gen_t *cg, ir_node *cnst) { tarval *tv = get_Const_tarval(cnst); - pmap_entry *e = pmap_find(cg->tv_ent, tv); + pmap_entry *e = pmap_find(cg->isa->tv_ent, tv); entity *res; ir_graph *rem; @@ -103,9 +153,9 @@ static entity *get_entity_for_tv(ia32_code_gen_t *cg, ir_node *cnst) ir_mode *mode = get_irn_mode(cnst); ir_type *tp = get_Const_type(cnst); if (tp == firm_unknown_type) - tp = get_prim_type(cg->types, mode); + tp = get_prim_type(cg->isa->types, mode); - res = new_entity(get_glob_type(), unique_id("ia32FloatCnst_%u"), tp); + res = new_entity(get_glob_type(), unique_id(".LC%u"), tp); set_entity_ld_ident(res, get_entity_ident(res)); set_entity_visibility(res, visibility_local); @@ -118,6 +168,8 @@ static entity *get_entity_for_tv(ia32_code_gen_t *cg, ir_node *cnst) current_ir_graph = get_const_code_irg(); set_atomic_ent_value(res, new_Const_type(tv, tp)); current_ir_graph = rem; + + pmap_insert(cg->isa->tv_ent, tv, res); } else res = e->value; @@ -134,7 +186,7 @@ static entity *get_entity_for_tv(ia32_code_gen_t *cg, ir_node *cnst) * @return the created ia32 Const node */ static ir_node *gen_Const(ia32_transform_env_t *env) { - ir_node *cnst; + ir_node *cnst, *load; symconst_symbol sym; ir_graph *irg = env->irg; ir_node *block = env->block; @@ -143,154 +195,513 @@ static ir_node *gen_Const(ia32_transform_env_t *env) { ir_mode *mode = env->mode; if (mode_is_float(mode)) { + FP_USED(env->cg); + if (! USE_SSE2(env->cg)) { + cnst_classify_t clss = classify_Const(node); + + if (clss == CNST_NULL) + return new_rd_ia32_vfldz(dbg, irg, block, mode); + else if (clss == CNST_ONE) + return new_rd_ia32_vfld1(dbg, irg, block, mode); + } sym.entity_p = get_entity_for_tv(env->cg, node); - cnst = new_rd_SymConst(dbg, irg, block, sym, symconst_addr_ent); - env->irn = cnst; - cnst = gen_SymConst(env); + + cnst = new_rd_SymConst(dbg, irg, block, sym, symconst_addr_ent); + load = new_r_Load(irg, block, get_irg_no_mem(irg), cnst, mode); + load = new_r_Proj(irg, block, load, mode, pn_Load_res); + env->irn = cnst; + env->mode = mode_P; + cnst = gen_SymConst(env); + add_irn_dep(cnst, be_abi_get_start_barrier(env->cg->birg->abi)); + set_Load_ptr(get_Proj_pred(load), cnst); + cnst = load; } else { cnst = new_rd_ia32_Const(dbg, irg, block, get_irn_mode(node)); + add_irn_dep(cnst, be_abi_get_start_barrier(env->cg->birg->abi)); set_ia32_Const_attr(cnst, node); } + return cnst; } - - /** * Transforms (all) Const's into ia32_Const and places them in the * block where they are used (or in the cfg-pred Block in case of Phi's). - * Additionally all mode_P nodes are changed into mode_Is nodes. + * Additionally all reference nodes are changed into mode_Is nodes. + * NOTE: irn must be a firm constant! + */ +static void ia32_transform_const(ir_node *irn, void *env) { + ia32_code_gen_t *cg = env; + ir_node *cnst = NULL; + ia32_transform_env_t tenv; + + tenv.cg = cg; + tenv.irg = cg->irg; + tenv.mode = get_irn_mode(irn); + tenv.dbg = get_irn_dbg_info(irn); + tenv.irn = irn; + DEBUG_ONLY(tenv.mod = cg->mod;) + +#if 1 + /* place const either in the smallest dominator of all its users or the original block */ + if (cg->opt & IA32_OPT_PLACECNST) + tenv.block = node_users_smallest_common_dominator(irn, 1); + else + tenv.block = get_nodes_block(irn); +#else + /* Actually, there is no real sense in placing */ + /* the Consts in the successor of the start block. */ + { + ir_node *afterstart = NULL; + ir_node *startblock = get_irg_start_block(tenv.irg); + const ir_edge_t *edge; + + foreach_block_succ(startblock, edge) { + ir_node *block = get_edge_src_irn(edge); + if (block != startblock) { + afterstart = block; + break; + } + } + assert(afterstart != NULL); + tenv.block = afterstart; + } +#endif + + switch (get_irn_opcode(irn)) { + case iro_Const: + cnst = gen_Const(&tenv); + break; + case iro_SymConst: + cnst = gen_SymConst(&tenv); + break; + default: + assert(0 && "Wrong usage of ia32_transform_const!"); + } + + assert(cnst && "Could not create ia32 Const"); + + /* set the new ia32 const */ + exchange(irn, cnst); +} + +/** + * Transform all firm consts and assure, we visit each const only once. */ -void ia32_place_consts_set_modes(ir_node *irn, void *env) { - ia32_code_gen_t *cg = env; - ia32_transform_env_t tenv; - ir_mode *mode; - ir_node *pred, *cnst; - int i; - opcode opc; +static void ia32_place_consts_walker(ir_node *irn, void *env) { + ia32_code_gen_t *cg = env; + + if (! is_Const(irn) && ! is_SymConst(irn)) + return; + + ia32_transform_const(irn, cg); +} +/** + * Replace reference modes with mode_Iu and preserve store value modes. + */ +static void ia32_set_modes(ir_node *irn, void *env) { if (is_Block(irn)) return; - mode = get_irn_mode(irn); + /* transform all reference nodes into mode_Iu nodes */ + if (mode_is_reference(get_irn_mode(irn))) { + set_irn_mode(irn, mode_Iu); + } +} - /* transform all mode_P nodes into mode_Is nodes */ - if (mode == mode_P) { - set_irn_mode(irn, mode_Is); - mode = mode_Is; +/** + * Walks over the graph, transforms all firm consts into ia32 consts + * and places them into the "best" block. + * @param cg The ia32 codegenerator object + */ +static void ia32_transform_all_firm_consts(ia32_code_gen_t *cg) { + irg_walk_graph(cg->irg, NULL, ia32_place_consts_walker, cg); +} + +/* Place all consts and change pointer arithmetics into unsigned integer arithmetics. */ +void ia32_pre_transform_phase(ia32_code_gen_t *cg) { + /* + We need to transform the consts twice: + - the psi condition tree transformer needs existing constants to be ia32 constants + - the psi condition tree transformer inserts new firm constants which need to be transformed + */ + ia32_transform_all_firm_consts(cg); + irg_walk_graph(cg->irg, ia32_set_modes, ia32_transform_psi_cond_tree, cg); + ia32_transform_all_firm_consts(cg); +} + +/******************************************************************************************************** + * _____ _ _ ____ _ _ _ _ _ + * | __ \ | | | | / __ \ | | (_) (_) | | (_) + * | |__) |__ ___ _ __ | |__ ___ | | ___ | | | |_ __ | |_ _ _ __ ___ _ ______ _| |_ _ ___ _ __ + * | ___/ _ \/ _ \ '_ \| '_ \ / _ \| |/ _ \ | | | | '_ \| __| | '_ ` _ \| |_ / _` | __| |/ _ \| '_ \ + * | | | __/ __/ |_) | | | | (_) | | __/ | |__| | |_) | |_| | | | | | | |/ / (_| | |_| | (_) | | | | + * |_| \___|\___| .__/|_| |_|\___/|_|\___| \____/| .__/ \__|_|_| |_| |_|_/___\__,_|\__|_|\___/|_| |_| + * | | | | + * |_| |_| + ********************************************************************************************************/ + +/** + * NOTE: THESE PEEPHOLE OPTIMIZATIONS MUST BE CALLED AFTER SCHEDULING AND REGISTER ALLOCATION. + */ + +static int ia32_cnst_compare(ir_node *n1, ir_node *n2) { + return get_ia32_id_cnst(n1) == get_ia32_id_cnst(n2); +} + +/** + * Checks for potential CJmp/CJmpAM optimization candidates. + */ +static ir_node *ia32_determine_cjmp_cand(ir_node *irn, is_op_func_t *is_op_func) { + ir_node *cand = NULL; + ir_node *prev = sched_prev(irn); + + if (is_Block(prev)) { + if (get_Block_n_cfgpreds(prev) == 1) + prev = get_Block_cfgpred(prev, 0); + else + prev = NULL; } - tenv.block = get_nodes_block(irn); - tenv.cg = cg; - tenv.irg = cg->irg; - tenv.mod = cg->mod; + /* The predecessor must be a ProjX. */ + if (prev && is_Proj(prev) && get_irn_mode(prev) == mode_X) { + prev = get_Proj_pred(prev); - /* Loop over all predecessors and check for Sym/Const nodes */ - for (i = get_irn_arity(irn) - 1; i >= 0; --i) { - pred = get_irn_n(irn, i); - cnst = NULL; - opc = get_irn_opcode(pred); - tenv.irn = pred; - tenv.mode = get_irn_mode(pred); - tenv.dbg = get_irn_dbg_info(pred); + if (is_op_func(prev)) + cand = prev; + } - /* If it's a Phi, then we need to create the */ - /* new Const in it's predecessor block */ - if (is_Phi(irn)) { - tenv.block = get_Block_cfgpred_block(get_nodes_block(irn), i); - } + return cand; +} - /* put the const into the block where the original const was */ - if (! cg->opt.placecnst) { - tenv.block = get_nodes_block(pred); - } +static int is_TestJmp_cand(const ir_node *irn) { + return is_ia32_TestJmp(irn) || is_ia32_And(irn); +} - switch (opc) { - case iro_Const: - cnst = gen_Const(&tenv); - break; - case iro_SymConst: - cnst = gen_SymConst(&tenv); - break; - default: - break; +/** + * Checks if two consecutive arguments of cand matches + * the two arguments of irn (TestJmp). + */ +static int is_TestJmp_replacement(ir_node *cand, ir_node *irn) { + ir_node *in1 = get_irn_n(irn, 0); + ir_node *in2 = get_irn_n(irn, 1); + int i, n = get_irn_arity(cand); + int same_args = 0; + + for (i = 0; i < n - 1; i++) { + if (get_irn_n(cand, i) == in1 && + get_irn_n(cand, i + 1) == in2) + { + same_args = 1; + break; } + } + + if (same_args) + return ia32_cnst_compare(cand, irn); + + return 0; +} + +/** + * Tries to replace a TestJmp by a CJmp or CJmpAM (in case of And) + */ +static void ia32_optimize_TestJmp(ir_node *irn, ia32_code_gen_t *cg) { + ir_node *cand = ia32_determine_cjmp_cand(irn, is_TestJmp_cand); + int replace = 0; - /* if we found a const, then set it */ - if (cnst) { - set_irn_n(irn, i, cnst); + /* we found a possible candidate */ + replace = cand ? is_TestJmp_replacement(cand, irn) : 0; + + if (replace) { + DBG((cg->mod, LEVEL_1, "replacing %+F by ", irn)); + + if (is_ia32_And(cand)) + set_irn_op(irn, op_ia32_CJmpAM); + else + set_irn_op(irn, op_ia32_CJmp); + + DB((cg->mod, LEVEL_1, "%+F\n", irn)); + } +} + +static int is_CondJmp_cand(const ir_node *irn) { + return is_ia32_CondJmp(irn) || is_ia32_Sub(irn); +} + +/** + * Checks if the arguments of cand are the same of irn. + */ +static int is_CondJmp_replacement(ir_node *cand, ir_node *irn) { + int i, n = get_irn_arity(cand); + int same_args = 1; + + for (i = 0; i < n; i++) { + if (get_irn_n(cand, i) != get_irn_n(irn, i)) { + same_args = 0; + break; } } + + if (same_args) + return ia32_cnst_compare(cand, irn); + + return 0; } +/** + * Tries to replace a CondJmp by a CJmpAM + */ +static void ia32_optimize_CondJmp(ir_node *irn, ia32_code_gen_t *cg) { + ir_node *cand = ia32_determine_cjmp_cand(irn, is_CondJmp_cand); + int replace = 0; -/****************************************************************** - * _ _ __ __ _ - * /\ | | | | | \/ | | | - * / \ __| | __| |_ __ ___ ___ ___| \ / | ___ __| | ___ - * / /\ \ / _` |/ _` | '__/ _ \/ __/ __| |\/| |/ _ \ / _` |/ _ \ - * / ____ \ (_| | (_| | | | __/\__ \__ \ | | | (_) | (_| | __/ - * /_/ \_\__,_|\__,_|_| \___||___/___/_| |_|\___/ \__,_|\___| - * - ******************************************************************/ + /* we found a possible candidate */ + replace = cand ? is_CondJmp_replacement(cand, irn) : 0; -static int node_is_comm(const ir_node *irn) { - return is_ia32_irn(irn) ? is_ia32_commutative(irn) : 0; + if (replace) { + DBG((cg->mod, LEVEL_1, "replacing %+F by ", irn)); + DBG_OPT_CJMP(irn); + + set_irn_op(irn, op_ia32_CJmpAM); + + DB((cg->mod, LEVEL_1, "%+F\n", irn)); + } } -static int ia32_get_irn_n_edges(const ir_node *irn) { - const ir_edge_t *edge; - int cnt = 0; +// only optimize up to 48 stores behind IncSPs +#define MAXPUSH_OPTIMIZE 48 - foreach_out_edge(irn, edge) { - cnt++; +/** + * Tries to create pushs from IncSP,Store combinations + */ +static void ia32_create_Pushs(ir_node *irn, ia32_code_gen_t *cg) { + int i; + int offset; + ir_node *node; + ir_node *stores[MAXPUSH_OPTIMIZE]; + ir_node *block = get_nodes_block(irn); + ir_graph *irg = cg->irg; + ir_node *curr_sp; + ir_mode *spmode = get_irn_mode(irn); + + memset(stores, 0, sizeof(stores)); + + assert(be_is_IncSP(irn)); + + offset = be_get_IncSP_offset(irn); + if(offset < 4) + return; + + /* + * We first walk the schedule after the IncSP node as long as we find + * suitable stores that could be transformed to a push. + * We save them into the stores array which is sorted by the frame offset/4 + * attached to the node + */ + for(node = sched_next(irn); !sched_is_end(node); node = sched_next(node)) { + const char *am_offs; + ir_node *mem; + int offset = -1; + int n; + int storeslot; + + // it has to be a store + if(!is_ia32_Store(node)) + break; + + // it has to use our sp value + if(get_irn_n(node, 0) != irn) + continue; + // store has to be attached to NoMem + mem = get_irn_n(node, 3); + if(!is_NoMem(mem)) { + continue; + } + + if( (get_ia32_am_flavour(node) & ia32_am_IS) != 0) + break; + + am_offs = get_ia32_am_offs(node); + if(am_offs == NULL) { + offset = 0; + } else { + // the am_offs has to be of the form "+NUMBER" + if(sscanf(am_offs, "+%d%n", &offset, &n) != 1 || am_offs[n] != '\0') { + // we shouldn't have any cases in the compiler at the moment + // that produce something different from esp+XX + assert(0); + break; + } + } + + storeslot = offset / 4; + if(storeslot >= MAXPUSH_OPTIMIZE) + continue; + + // storing into the same slot twice is bad (and shouldn't happen...) + if(stores[storeslot] != NULL) + break; + + // storing at half-slots is bad + if(offset % 4 != 0) + break; + + stores[storeslot] = node; } - return cnt; + curr_sp = get_irn_n(irn, 0); + + // walk the stores in inverse order and create pushs for them + i = (offset / 4) - 1; + if(i >= MAXPUSH_OPTIMIZE) { + i = MAXPUSH_OPTIMIZE - 1; + } + + for( ; i >= 0; --i) { + const ir_edge_t *edge, *next; + const arch_register_t *spreg; + ir_node *push; + ir_node *val, *mem; + ir_node *store = stores[i]; + ir_node *noreg = ia32_new_NoReg_gp(cg); + + if(store == NULL || is_Bad(store)) + break; + + val = get_irn_n(store, 2); + mem = get_irn_n(store, 3); + spreg = arch_get_irn_register(cg->arch_env, curr_sp); + + // create a push + push = new_rd_ia32_Push(NULL, irg, block, noreg, noreg, val, curr_sp, mem); + if(get_ia32_immop_type(store) != ia32_ImmNone) { + copy_ia32_Immop_attr(push, store); + } + sched_add_before(irn, push); + + // create stackpointer proj + curr_sp = new_r_Proj(irg, block, push, spmode, pn_ia32_Push_stack); + arch_set_irn_register(cg->arch_env, curr_sp, spreg); + sched_add_before(irn, curr_sp); + + // rewire memprojs of the store + foreach_out_edge_safe(store, edge, next) { + ir_node *succ = get_edge_src_irn(edge); + + assert(is_Proj(succ) && get_Proj_proj(succ) == pn_ia32_Store_M); + set_irn_n(succ, 0, push); + } + + // we can remove the store now + set_irn_n(store, 0, new_Bad()); + set_irn_n(store, 1, new_Bad()); + set_irn_n(store, 2, new_Bad()); + set_irn_n(store, 3, new_Bad()); + sched_remove(store); + + offset -= 4; + } + + be_set_IncSP_offset(irn, offset); + + // can we remove the IncSP now? + if(offset == 0) { + const ir_edge_t *edge, *next; + + foreach_out_edge_safe(irn, edge, next) { + ir_node *arg = get_edge_src_irn(edge); + int pos = get_edge_src_pos(edge); + + set_irn_n(arg, pos, curr_sp); + } + + set_irn_n(irn, 0, new_Bad()); + sched_remove(irn); + } else { + set_irn_n(irn, 0, curr_sp); + } } /** - * Returns the first mode_M Proj connected to irn. + * Tries to optimize two following IncSP. */ -static ir_node *get_mem_proj(const ir_node *irn) { - const ir_edge_t *edge; - ir_node *src; +static void ia32_optimize_IncSP(ir_node *irn, ia32_code_gen_t *cg) { + ir_node *prev = be_get_IncSP_pred(irn); + int real_uses = get_irn_n_edges(prev); - assert(get_irn_mode(irn) == mode_T && "expected mode_T node"); + if (be_is_IncSP(prev) && real_uses == 1) { + /* first IncSP has only one IncSP user, kill the first one */ + int prev_offs = be_get_IncSP_offset(prev); + int curr_offs = be_get_IncSP_offset(irn); - foreach_out_edge(irn, edge) { - src = get_edge_src_irn(edge); + be_set_IncSP_offset(prev, prev_offs + curr_offs); - assert(is_Proj(src) && "Proj expected"); + /* Omit the optimized IncSP */ + be_set_IncSP_pred(irn, be_get_IncSP_pred(prev)); - if (get_irn_mode(src) == mode_M) - return src; + set_irn_n(prev, 0, new_Bad()); + sched_remove(prev); } - - return NULL; } /** - * Returns the first Proj with mode != mode_M connected to irn. + * Performs Peephole Optimizations. */ -static ir_node *get_res_proj(const ir_node *irn) { - const ir_edge_t *edge; - ir_node *src; +static void ia32_peephole_optimize_node(ir_node *irn, void *env) { + ia32_code_gen_t *cg = env; + + /* AMD CPUs want explicit compare before conditional jump */ + if (! ARCH_AMD(cg->opt_arch)) { + if (is_ia32_TestJmp(irn)) + ia32_optimize_TestJmp(irn, cg); + else if (is_ia32_CondJmp(irn)) + ia32_optimize_CondJmp(irn, cg); + } - assert(get_irn_mode(irn) == mode_T && "expected mode_T node"); + if (be_is_IncSP(irn)) { + // optimize_IncSP doesn't respect dependency edges yet... + //ia32_optimize_IncSP(irn, cg); + (void) ia32_optimize_IncSP; + ia32_create_Pushs(irn, cg); + } +} - foreach_out_edge(irn, edge) { - src = get_edge_src_irn(edge); +void ia32_peephole_optimization(ir_graph *irg, ia32_code_gen_t *cg) { + irg_walk_graph(irg, ia32_peephole_optimize_node, NULL, cg); +} - assert(is_Proj(src) && "Proj expected"); +/****************************************************************** + * _ _ __ __ _ + * /\ | | | | | \/ | | | + * / \ __| | __| |_ __ ___ ___ ___| \ / | ___ __| | ___ + * / /\ \ / _` |/ _` | '__/ _ \/ __/ __| |\/| |/ _ \ / _` |/ _ \ + * / ____ \ (_| | (_| | | | __/\__ \__ \ | | | (_) | (_| | __/ + * /_/ \_\__,_|\__,_|_| \___||___/___/_| |_|\___/ \__,_|\___| + * + ******************************************************************/ - if (get_irn_mode(src) != mode_M) - return src; +typedef struct { + ia32_code_gen_t *cg; + heights_t *h; +} ia32_am_opt_env_t; + +static int node_is_ia32_comm(const ir_node *irn) { + return is_ia32_irn(irn) ? is_ia32_commutative(irn) : 0; +} + +static int ia32_get_irn_n_edges(const ir_node *irn) { + const ir_edge_t *edge; + int cnt = 0; + + foreach_out_edge(irn, edge) { + cnt++; } - return NULL; + return cnt; } /** @@ -300,7 +711,7 @@ static ir_node *get_res_proj(const ir_node *irn) { * @param is_op_func The check-function * @return 1 if conditions are fulfilled, 0 otherwise */ -static int pred_is_specific_node(const ir_node *pred, int (*is_op_func)(const ir_node *n)) { +static int pred_is_specific_node(const ir_node *pred, is_op_func_t *is_op_func) { if (is_Proj(pred) && is_op_func(get_Proj_pred(pred))) { return 1; } @@ -330,47 +741,145 @@ static int pred_is_specific_nodeblock(const ir_node *bl, const ir_node *pred, return 0; } - - /** - * Checks if irn is a candidate for address calculation or address mode. + * Checks if irn is a candidate for address calculation. * - * address calculation (AC): * - none of the operand must be a Load within the same block OR * - all Loads must have more than one user OR * - the irn has a frame entity (it's a former FrameAddr) * + * @param block The block the Loads must/mustnot be in + * @param irn The irn to check + * return 1 if irn is a candidate, 0 otherwise + */ +static int is_addr_candidate(const ir_node *block, const ir_node *irn) { + ir_node *in, *left, *right; + int n, is_cand = 1; + + left = get_irn_n(irn, 2); + right = get_irn_n(irn, 3); + + in = left; + +#ifndef AGGRESSIVE_AM + if (pred_is_specific_nodeblock(block, in, is_ia32_Ld)) { + n = ia32_get_irn_n_edges(in); + is_cand = (n == 1) ? 0 : is_cand; /* load with only one user: don't create LEA */ + } + + in = right; + + if (pred_is_specific_nodeblock(block, in, is_ia32_Ld)) { + n = ia32_get_irn_n_edges(in); + is_cand = (n == 1) ? 0 : is_cand; /* load with only one user: don't create LEA */ + } +#else + (void) n; +#endif + + is_cand = get_ia32_frame_ent(irn) ? 1 : is_cand; + + return is_cand; +} + +/** + * Checks if irn is a candidate for address mode. + * * address mode (AM): * - at least one operand has to be a Load within the same block AND * - the load must not have other users than the irn AND * - the irn must not have a frame entity set * - * @param block The block the Loads must/not be in + * @param cg The ia32 code generator + * @param h The height information of the irg + * @param block The block the Loads must/mustnot be in * @param irn The irn to check - * @param check_addr 1 if to check for address calculation, 0 otherwise - * return 1 if irn is a candidate for AC or AM, 0 otherwise + * return 0 if irn is no candidate, 1 if left load can be used, 2 if right one, 3 for both */ -static int is_candidate(const ir_node *block, const ir_node *irn, int check_addr) { - ir_node *in; - int n, is_cand = check_addr; +static ia32_am_cand_t is_am_candidate(ia32_code_gen_t *cg, heights_t *h, const ir_node *block, ir_node *irn) { + ir_node *in, *load, *other, *left, *right; + int is_cand = 0, cand; + + if (is_ia32_Ld(irn) || is_ia32_St(irn) || is_ia32_Store8Bit(irn) || is_ia32_vfild(irn) || is_ia32_vfist(irn) || + is_ia32_GetST0(irn) || is_ia32_SetST0(irn) || is_ia32_xStoreSimple(irn)) + return 0; + + left = get_irn_n(irn, 2); + right = get_irn_n(irn, 3); - in = get_irn_n(irn, 2); + in = left; if (pred_is_specific_nodeblock(block, in, is_ia32_Ld)) { +#ifndef AGGRESSIVE_AM + int n; n = ia32_get_irn_n_edges(in); - is_cand = check_addr ? (n == 1 ? 0 : is_cand) : (n == 1 ? 1 : is_cand); + is_cand = (n == 1) ? 1 : is_cand; /* load with more than one user: no AM */ +#else + is_cand = 1; +#endif + + load = get_Proj_pred(in); + other = right; + + /* 8bit Loads are not supported (for binary ops), + * they cannot be used with every register */ + if (get_irn_arity(irn) != 4 && get_mode_size_bits(get_ia32_ls_mode(load)) < 16) { + assert(get_irn_arity(irn) == 5); + is_cand = 0; + } + + /* If there is a data dependency of other irn from load: cannot use AM */ + if (is_cand && get_nodes_block(other) == block) { + other = skip_Proj(other); + is_cand = heights_reachable_in_block(h, other, load) ? 0 : is_cand; + /* this could happen in loops */ + is_cand = heights_reachable_in_block(h, load, irn) ? 0 : is_cand; + } } - in = get_irn_n(irn, 3); + cand = is_cand ? IA32_AM_CAND_LEFT : IA32_AM_CAND_NONE; + in = right; + is_cand = 0; if (pred_is_specific_nodeblock(block, in, is_ia32_Ld)) { +#ifndef AGGRESSIVE_AM + int n; n = ia32_get_irn_n_edges(in); - is_cand = check_addr ? (n == 1 ? 0 : is_cand) : (n == 1 ? 1 : is_cand); + is_cand = (n == 1) ? 1 : is_cand; /* load with more than one user: no AM */ +#endif + + load = get_Proj_pred(in); + other = left; + + /* 8bit Loads are not supported, they cannot be used with every register */ + if (get_mode_size_bits(get_ia32_ls_mode(load)) < 16) + is_cand = 0; + + /* If there is a data dependency of other irn from load: cannot use load */ + if (is_cand && get_nodes_block(other) == block) { + other = skip_Proj(other); + is_cand = heights_reachable_in_block(h, other, load) ? 0 : is_cand; + /* this could happen in loops */ + is_cand = heights_reachable_in_block(h, load, irn) ? 0 : is_cand; + } } - is_cand = get_ia32_frame_ent(irn) ? (check_addr ? 1 : 0) : is_cand; + cand = is_cand ? (cand | IA32_AM_CAND_RIGHT) : cand; - return is_cand; + /* check some special cases */ + if (USE_SSE2(cg) && is_ia32_Conv_I2FP(irn)) { + /* SSE Conv I -> FP cvtsi2s(s|d) can only load 32 bit values */ + if (get_mode_size_bits(get_ia32_tgt_mode(irn)) != 32) + cand = IA32_AM_CAND_NONE; + } + else if (is_ia32_Conv_I2I(irn)) { + /* we cannot load an N bit value and implicitly convert it into an M bit value if N > M */ + if (get_mode_size_bits(get_ia32_src_mode(irn)) > get_mode_size_bits(get_ia32_tgt_mode(irn))) + cand = IA32_AM_CAND_NONE; + } + + /* if the irn has a frame entity: we do not use address mode */ + return get_ia32_frame_ent(irn) ? IA32_AM_CAND_NONE : cand; } /** @@ -383,35 +892,198 @@ static int load_store_addr_is_equal(const ir_node *load, const ir_node *store, int is_equal = (addr_b == get_irn_n(load, 0)) && (addr_i == get_irn_n(load, 1)); entity *lent = get_ia32_frame_ent(load); entity *sent = get_ia32_frame_ent(store); + ident *lid = get_ia32_am_sc(load); + ident *sid = get_ia32_am_sc(store); + char *loffs = get_ia32_am_offs(load); + char *soffs = get_ia32_am_offs(store); /* are both entities set and equal? */ - is_equal = lent && sent && (lent == sent); + if (is_equal && (lent || sent)) + is_equal = lent && sent && (lent == sent); + + /* are address mode idents set and equal? */ + if (is_equal && (lid || sid)) + is_equal = lid && sid && (lid == sid); + + /* are offsets set and equal */ + if (is_equal && (loffs || soffs)) + is_equal = loffs && soffs && strcmp(loffs, soffs) == 0; /* are the load and the store of the same mode? */ - is_equal = get_ia32_ls_mode(load) == get_ia32_ls_mode(store); + is_equal = is_equal ? get_ia32_ls_mode(load) == get_ia32_ls_mode(store) : 0; return is_equal; } +typedef enum _ia32_take_lea_attr { + IA32_LEA_ATTR_NONE = 0, + IA32_LEA_ATTR_BASE = (1 << 0), + IA32_LEA_ATTR_INDEX = (1 << 1), + IA32_LEA_ATTR_OFFS = (1 << 2), + IA32_LEA_ATTR_SCALE = (1 << 3), + IA32_LEA_ATTR_AMSC = (1 << 4), + IA32_LEA_ATTR_FENT = (1 << 5) +} ia32_take_lea_attr; + +/** + * Decides if we have to keep the LEA operand or if we can assimilate it. + */ +static int do_new_lea(ir_node *irn, ir_node *base, ir_node *index, ir_node *lea, + int have_am_sc, ia32_code_gen_t *cg) +{ + entity *irn_ent = get_ia32_frame_ent(irn); + entity *lea_ent = get_ia32_frame_ent(lea); + int ret_val = 0; + int is_noreg_base = be_is_NoReg(cg, base); + int is_noreg_index = be_is_NoReg(cg, index); + ia32_am_flavour_t am_flav = get_ia32_am_flavour(lea); + + /* If the Add and the LEA both have a different frame entity set: keep */ + if (irn_ent && lea_ent && (irn_ent != lea_ent)) + return IA32_LEA_ATTR_NONE; + else if (! irn_ent && lea_ent) + ret_val |= IA32_LEA_ATTR_FENT; + + /* If the Add and the LEA both have already an address mode symconst: keep */ + if (have_am_sc && get_ia32_am_sc(lea)) + return IA32_LEA_ATTR_NONE; + else if (get_ia32_am_sc(lea)) + ret_val |= IA32_LEA_ATTR_AMSC; + + /* Check the different base-index combinations */ + + if (! is_noreg_base && ! is_noreg_index) { + /* Assimilate if base is the lea and the LEA is just a Base + Offset calculation */ + if ((base == lea) && ! (am_flav & ia32_I ? 1 : 0)) { + if (am_flav & ia32_O) + ret_val |= IA32_LEA_ATTR_OFFS; + + ret_val |= IA32_LEA_ATTR_BASE; + } + else + return IA32_LEA_ATTR_NONE; + } + else if (! is_noreg_base && is_noreg_index) { + /* Base is set but index not */ + if (base == lea) { + /* Base points to LEA: assimilate everything */ + if (am_flav & ia32_O) + ret_val |= IA32_LEA_ATTR_OFFS; + if (am_flav & ia32_S) + ret_val |= IA32_LEA_ATTR_SCALE; + if (am_flav & ia32_I) + ret_val |= IA32_LEA_ATTR_INDEX; + + ret_val |= IA32_LEA_ATTR_BASE; + } + else if (am_flav & ia32_B ? 0 : 1) { + /* Base is not the LEA but the LEA is an index only calculation: assimilate */ + if (am_flav & ia32_O) + ret_val |= IA32_LEA_ATTR_OFFS; + if (am_flav & ia32_S) + ret_val |= IA32_LEA_ATTR_SCALE; + + ret_val |= IA32_LEA_ATTR_INDEX; + } + else + return IA32_LEA_ATTR_NONE; + } + else if (is_noreg_base && ! is_noreg_index) { + /* Index is set but not base */ + if (index == lea) { + /* Index points to LEA: assimilate everything */ + if (am_flav & ia32_O) + ret_val |= IA32_LEA_ATTR_OFFS; + if (am_flav & ia32_S) + ret_val |= IA32_LEA_ATTR_SCALE; + if (am_flav & ia32_B) + ret_val |= IA32_LEA_ATTR_BASE; + + ret_val |= IA32_LEA_ATTR_INDEX; + } + else if (am_flav & ia32_I ? 0 : 1) { + /* Index is not the LEA but the LEA is a base only calculation: assimilate */ + if (am_flav & ia32_O) + ret_val |= IA32_LEA_ATTR_OFFS; + if (am_flav & ia32_S) + ret_val |= IA32_LEA_ATTR_SCALE; + + ret_val |= IA32_LEA_ATTR_BASE; + } + else + return IA32_LEA_ATTR_NONE; + } + else { + assert(0 && "There must have been set base or index"); + } + + return ret_val; +} + +/** + * Adds res before irn into schedule if irn was scheduled. + * @param irn The schedule point + * @param res The node to be scheduled + */ +static INLINE void try_add_to_sched(ir_node *irn, ir_node *res) { + if (sched_is_scheduled(irn)) + sched_add_before(irn, res); +} + +/** + * Removes irn from schedule if it was scheduled. If irn is a mode_T node + * all it's Projs are removed as well. + * @param irn The irn to be removed from schedule + */ +static INLINE void try_remove_from_sched(ir_node *irn) { + int i, arity; + + if (sched_is_scheduled(irn)) { + if (get_irn_mode(irn) == mode_T) { + const ir_edge_t *edge; + foreach_out_edge(irn, edge) { + ir_node *proj = get_edge_src_irn(edge); + if (sched_is_scheduled(proj)) { + set_irn_n(proj, 0, new_Bad()); + sched_remove(proj); + } + } + } + arity = get_irn_arity(irn); + for(i = 0; i < arity; ++i) { + set_irn_n(irn, i, new_Bad()); + } + sched_remove(irn); + } +} /** * Folds Add or Sub to LEA if possible */ -static ir_node *fold_addr(be_abi_irg_t *babi, ir_node *irn, firm_dbg_module_t *mod, ir_node *noreg) { - ir_graph *irg = get_irn_irg(irn); - dbg_info *dbg = get_irn_dbg_info(irn); - ir_node *block = get_nodes_block(irn); - ir_node *res = irn; - char *offs = NULL; - char *offs_cnst = NULL; - char *offs_lea = NULL; - int scale = 0; - int isadd = 0; - int dolea = 0; - ir_node *left, *right, *temp; - ir_node *base, *index; +static ir_node *fold_addr(ia32_code_gen_t *cg, ir_node *irn, ir_node *noreg) { + ir_graph *irg = get_irn_irg(irn); + dbg_info *dbg = get_irn_dbg_info(irn); + ir_node *block = get_nodes_block(irn); + ir_node *res = irn; + ir_node *shift = NULL; + ir_node *lea_o = NULL; + ir_node *lea = NULL; + char *offs = NULL; + const char *offs_cnst = NULL; + char *offs_lea = NULL; + int scale = 0; + int isadd = 0; + int dolea = 0; + int have_am_sc = 0; + int am_sc_sign = 0; + ident *am_sc = NULL; + entity *lea_ent = NULL; + ir_node *left, *right, *temp; + ir_node *base, *index; + int consumed_left_shift; ia32_am_flavour_t am_flav; + DEBUG_ONLY(firm_dbg_module_t *mod = cg->mod;) if (is_ia32_Add(irn)) isadd = 1; @@ -420,7 +1092,7 @@ static ir_node *fold_addr(be_abi_irg_t *babi, ir_node *irn, firm_dbg_module_t *m right = get_irn_n(irn, 3); /* "normalize" arguments in case of add with two operands */ - if (isadd && ! be_is_NoReg(babi, right)) { + if (isadd && ! be_is_NoReg(cg, right)) { /* put LEA == ia32_am_O as right operand */ if (is_ia32_Lea(left) && get_ia32_am_flavour(left) == ia32_am_O) { set_irn_n(irn, 2, right); @@ -455,34 +1127,49 @@ static ir_node *fold_addr(be_abi_irg_t *babi, ir_node *irn, firm_dbg_module_t *m scale = 0; am_flav = 0; - /* check if operand is either const */ - if (get_ia32_cnst(irn)) { - DBG((mod, LEVEL_1, "\tfound op with imm")); + /* check for operation with immediate */ + if (is_ia32_ImmConst(irn)) { + DBG((mod, LEVEL_1, "\tfound op with imm const")); offs_cnst = get_ia32_cnst(irn); dolea = 1; } + else if (is_ia32_ImmSymConst(irn)) { + DBG((mod, LEVEL_1, "\tfound op with imm symconst")); - /* determine the operand which needs to be checked */ - if (be_is_NoReg(babi, right)) { - temp = left; - } - else { - temp = right; + have_am_sc = 1; + dolea = 1; + am_sc = get_ia32_id_cnst(irn); + am_sc_sign = is_ia32_am_sc_sign(irn); } - /* check if right operand is AMConst (LEA with ia32_am_O) */ - if (is_ia32_Lea(temp) && get_ia32_am_flavour(temp) == ia32_am_O) { + /* determine the operand which needs to be checked */ + temp = be_is_NoReg(cg, right) ? left : right; + + /* check if right operand is AMConst (LEA with ia32_am_O) */ + /* but we can only eat it up if there is no other symconst */ + /* because the linker won't accept two symconsts */ + if (! have_am_sc && is_ia32_Lea(temp) && get_ia32_am_flavour(temp) == ia32_am_O) { DBG((mod, LEVEL_1, "\tgot op with LEA am_O")); - offs_lea = get_ia32_am_offs(temp); - dolea = 1; + offs_lea = get_ia32_am_offs(temp); + am_sc = get_ia32_am_sc(temp); + am_sc_sign = is_ia32_am_sc_sign(temp); + have_am_sc = 1; + dolea = 1; + lea_o = temp; + + if (temp == base) + base = noreg; + else if (temp == right) + right = noreg; } if (isadd) { /* default for add -> make right operand to index */ - index = right; - dolea = 1; + index = right; + dolea = 1; + consumed_left_shift = -1; DBG((mod, LEVEL_1, "\tgot LEA candidate with index %+F\n", index)); @@ -490,56 +1177,75 @@ static ir_node *fold_addr(be_abi_irg_t *babi, ir_node *irn, firm_dbg_module_t *m temp = left; if (is_ia32_Lea(left)) { temp = right; + consumed_left_shift = 0; } /* check for SHL 1,2,3 */ if (pred_is_specific_node(temp, is_ia32_Shl)) { - temp = get_Proj_pred(temp); + temp = get_Proj_pred(temp); + shift = temp; if (get_ia32_Immop_tarval(temp)) { scale = get_tarval_long(get_ia32_Immop_tarval(temp)); if (scale <= 3) { - index = get_irn_n(temp, 2); + index = get_irn_n(temp, 2); + consumed_left_shift = consumed_left_shift < 0 ? 1 : 0; DBG((mod, LEVEL_1, "\tgot scaled index %+F\n", index)); } + else { + scale = 0; + shift = NULL; + } } } /* fix base */ - if (! be_is_NoReg(babi, index)) { + if (! be_is_NoReg(cg, index)) { /* if we have index, but left == right -> no base */ if (left == right) { base = noreg; } - else if (! is_ia32_Lea(left) && (index != right)) { - /* index != right -> we found a good Shl */ - /* left != LEA -> this Shl was the left operand */ - /* -> base is right operand */ - base = right; + else if (consumed_left_shift == 1) { + /* -> base is right operand */ + base = (right == lea_o) ? noreg : right; } } } /* Try to assimilate a LEA as left operand */ if (is_ia32_Lea(left) && (get_ia32_am_flavour(left) != ia32_am_O)) { - am_flav = get_ia32_am_flavour(left); - - /* If we have an Add with a real right operand (not NoReg) and */ - /* the LEA contains already an index calculation then we create */ - /* a new LEA. */ - /* If the LEA contains already a frame_entity then we also */ - /* create a new one otherwise we would loose it. */ - if ((isadd && ((!be_is_NoReg(babi, index) && (am_flav & ia32_am_I))) || get_ia32_frame_ent(left))) { + /* check if we can assimilate the LEA */ + int take_attr = do_new_lea(irn, base, index, left, have_am_sc, cg); + + if (take_attr == IA32_LEA_ATTR_NONE) { DBG((mod, LEVEL_1, "\tleave old LEA, creating new one\n")); } else { DBG((mod, LEVEL_1, "\tgot LEA as left operand ... assimilating\n")); - offs = get_ia32_am_offs(left); - base = get_irn_n(left, 0); - index = get_irn_n(left, 1); - scale = get_ia32_am_scale(left); + lea = left; /* for statistics */ + + if (take_attr & IA32_LEA_ATTR_OFFS) + offs = get_ia32_am_offs(left); + + if (take_attr & IA32_LEA_ATTR_AMSC) { + am_sc = get_ia32_am_sc(left); + have_am_sc = 1; + am_sc_sign = is_ia32_am_sc_sign(left); + } + + if (take_attr & IA32_LEA_ATTR_SCALE) + scale = get_ia32_am_scale(left); + + if (take_attr & IA32_LEA_ATTR_BASE) + base = get_irn_n(left, 0); + + if (take_attr & IA32_LEA_ATTR_INDEX) + index = get_irn_n(left, 1); + + if (take_attr & IA32_LEA_ATTR_FENT) + lea_ent = get_ia32_frame_ent(left); } } @@ -575,11 +1281,21 @@ static ir_node *fold_addr(be_abi_irg_t *babi, ir_node *irn, firm_dbg_module_t *m } } + /* set the address mode symconst */ + if (have_am_sc) { + set_ia32_am_sc(res, am_sc); + if (am_sc_sign) + set_ia32_am_sc_sign(res); + } + /* copy the frame entity (could be set in case of Add */ /* which was a FrameAddr) */ - set_ia32_frame_ent(res, get_ia32_frame_ent(irn)); + if (lea_ent) + set_ia32_frame_ent(res, lea_ent); + else + set_ia32_frame_ent(res, get_ia32_frame_ent(irn)); - if (is_ia32_use_frame(irn)) + if (get_ia32_frame_ent(res)) set_ia32_use_frame(res); /* set scale */ @@ -587,13 +1303,13 @@ static ir_node *fold_addr(be_abi_irg_t *babi, ir_node *irn, firm_dbg_module_t *m am_flav = ia32_am_N; /* determine new am flavour */ - if (offs || offs_cnst || offs_lea) { + if (offs || offs_cnst || offs_lea || have_am_sc) { am_flav |= ia32_O; } - if (! be_is_NoReg(babi, base)) { + if (! be_is_NoReg(cg, base)) { am_flav |= ia32_B; } - if (! be_is_NoReg(babi, index)) { + if (! be_is_NoReg(cg, index)) { am_flav |= ia32_I; } if (scale > 0) { @@ -603,10 +1319,51 @@ static ir_node *fold_addr(be_abi_irg_t *babi, ir_node *irn, firm_dbg_module_t *m set_ia32_op_type(res, ia32_AddrModeS); + SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(cg, irn)); + DBG((mod, LEVEL_1, "\tLEA [%+F + %+F * %d + %s]\n", base, index, scale, get_ia32_am_offs(res))); + /* we will exchange it, report here before the Proj is created */ + if (shift && lea && lea_o) { + try_remove_from_sched(shift); + try_remove_from_sched(lea); + try_remove_from_sched(lea_o); + DBG_OPT_LEA4(irn, lea_o, lea, shift, res); + } + else if (shift && lea) { + try_remove_from_sched(shift); + try_remove_from_sched(lea); + DBG_OPT_LEA3(irn, lea, shift, res); + } + else if (shift && lea_o) { + try_remove_from_sched(shift); + try_remove_from_sched(lea_o); + DBG_OPT_LEA3(irn, lea_o, shift, res); + } + else if (lea && lea_o) { + try_remove_from_sched(lea); + try_remove_from_sched(lea_o); + DBG_OPT_LEA3(irn, lea_o, lea, res); + } + else if (shift) { + try_remove_from_sched(shift); + DBG_OPT_LEA2(irn, shift, res); + } + else if (lea) { + try_remove_from_sched(lea); + DBG_OPT_LEA2(irn, lea, res); + } + else if (lea_o) { + try_remove_from_sched(lea_o); + DBG_OPT_LEA2(irn, lea_o, res); + } + else + DBG_OPT_LEA1(irn, res); + /* get the result Proj of the Add/Sub */ - irn = get_res_proj(irn); + try_add_to_sched(irn, res); + try_remove_from_sched(irn); + irn = ia32_get_res_proj(irn); assert(irn && "Couldn't find result proj"); @@ -617,34 +1374,79 @@ static ir_node *fold_addr(be_abi_irg_t *babi, ir_node *irn, firm_dbg_module_t *m return res; } + /** - * Optimizes a pattern around irn to address mode if possible. + * Merges a Load/Store node with a LEA. + * @param irn The Load/Store node + * @param lea The LEA */ -void ia32_optimize_am(ir_node *irn, void *env) { - ia32_code_gen_t *cg = env; - firm_dbg_module_t *mod = cg->mod; - ir_node *res = irn; - be_abi_irg_t *babi = cg->birg->abi; - dbg_info *dbg; - ir_mode *mode; - ir_node *block, *noreg_gp, *noreg_fp; - ir_node *left, *right, *temp; - ir_node *store, *load, *mem_proj; - ir_node *succ, *addr_b, *addr_i; - int check_am_src = 0; +static void merge_loadstore_lea(ir_node *irn, ir_node *lea) { + entity *irn_ent = get_ia32_frame_ent(irn); + entity *lea_ent = get_ia32_frame_ent(lea); - if (! is_ia32_irn(irn)) + /* If the irn and the LEA both have a different frame entity set: do not merge */ + if (irn_ent && lea_ent && (irn_ent != lea_ent)) return; + else if (! irn_ent && lea_ent) { + set_ia32_frame_ent(irn, lea_ent); + set_ia32_use_frame(irn); + } - dbg = get_irn_dbg_info(irn); - mode = get_irn_mode(irn); - block = get_nodes_block(irn); - noreg_gp = ia32_new_NoReg_gp(cg); - noreg_fp = ia32_new_NoReg_fp(cg); + /* get the AM attributes from the LEA */ + add_ia32_am_offs(irn, get_ia32_am_offs(lea)); + set_ia32_am_scale(irn, get_ia32_am_scale(lea)); + set_ia32_am_flavour(irn, get_ia32_am_flavour(lea)); - DBG((mod, LEVEL_1, "checking for AM\n")); + set_ia32_am_sc(irn, get_ia32_am_sc(lea)); + if (is_ia32_am_sc_sign(lea)) + set_ia32_am_sc_sign(irn); + + set_ia32_op_type(irn, is_ia32_Ld(irn) ? ia32_AddrModeS : ia32_AddrModeD); - /* 1st part: check for address calculations and transform the into Lea */ + /* set base and index */ + set_irn_n(irn, 0, get_irn_n(lea, 0)); + set_irn_n(irn, 1, get_irn_n(lea, 1)); + + try_remove_from_sched(lea); + + /* clear remat flag */ + set_ia32_flags(irn, get_ia32_flags(irn) & ~arch_irn_flags_rematerializable); + + if (is_ia32_Ld(irn)) + DBG_OPT_LOAD_LEA(lea, irn); + else + DBG_OPT_STORE_LEA(lea, irn); + +} + +/** + * Sets new_right index of irn to right and new_left index to left. + * Also exchange left and right + */ +static void exchange_left_right(ir_node *irn, ir_node **left, ir_node **right, int new_left, int new_right) { + ir_node *temp; + + set_irn_n(irn, new_right, *right); + set_irn_n(irn, new_left, *left); + + temp = *left; + *left = *right; + *right = temp; + + /* this is only needed for Compares, but currently ALL nodes + * have this attribute :-) */ + set_ia32_pncode(irn, get_inversed_pnc(get_ia32_pncode(irn))); +} + +/** + * Performs address calculation optimization (create LEAs if possible) + */ +static void optimize_lea(ir_node *irn, void *env) { + ia32_code_gen_t *cg = env; + ir_node *block, *noreg_gp, *left, *right; + + if (! is_ia32_irn(irn)) + return; /* Following cases can occur: */ /* - Sub (l, imm) -> LEA [base - offset] */ @@ -656,25 +1458,73 @@ void ia32_optimize_am(ir_node *irn, void *env) { /* with scale > 1 iff l/r == shl (1,2,3) */ if (is_ia32_Sub(irn) || is_ia32_Add(irn)) { - left = get_irn_n(irn, 2); - right = get_irn_n(irn, 3); + left = get_irn_n(irn, 2); + right = get_irn_n(irn, 3); + block = get_nodes_block(irn); + noreg_gp = ia32_new_NoReg_gp(cg); /* Do not try to create a LEA if one of the operands is a Load. */ /* check is irn is a candidate for address calculation */ - if (is_candidate(block, irn, 1)) { - DBG((mod, LEVEL_1, "\tfound address calculation candidate %+F ... ", irn)); - res = fold_addr(babi, irn, mod, noreg_gp); + if (is_addr_candidate(block, irn)) { + ir_node *res; + + DBG((cg->mod, LEVEL_1, "\tfound address calculation candidate %+F ... ", irn)); + res = fold_addr(cg, irn, noreg_gp); - if (res == irn) - DB((mod, LEVEL_1, "transformed into %+F\n", res)); + if (res != irn) + DB((cg->mod, LEVEL_1, "transformed into %+F\n", res)); else - DB((mod, LEVEL_1, "not transformed\n")); + DB((cg->mod, LEVEL_1, "not transformed\n")); + } + } + else if (is_ia32_Ld(irn) || is_ia32_St(irn) || is_ia32_Store8Bit(irn)) { + /* - Load -> LEA into Load } TODO: If the LEA is used by more than one Load/Store */ + /* - Store -> LEA into Store } it might be better to keep the LEA */ + left = get_irn_n(irn, 0); + + if (is_ia32_Lea(left)) { + const ir_edge_t *edge, *ne; + ir_node *src; + + /* merge all Loads/Stores connected to this LEA with the LEA */ + foreach_out_edge_safe(left, edge, ne) { + src = get_edge_src_irn(edge); + + if (src && (get_edge_src_pos(edge) == 0) && (is_ia32_Ld(src) || is_ia32_St(src) || is_ia32_Store8Bit(src))) { + DBG((cg->mod, LEVEL_1, "\nmerging %+F into %+F\n", left, irn)); + if (! is_ia32_got_lea(src)) + merge_loadstore_lea(src, left); + set_ia32_got_lea(src); + } + } } } +} - /* 2nd part: fold following patterns: */ - /* - Load -> LEA into Load } TODO: If the LEA is used by more than one Load/Store */ - /* - Store -> LEA into Store } it might be better to keep the LEA */ +/** + * Checks for address mode patterns and performs the + * necessary transformations. + * This function is called by a walker. + */ +static void optimize_am(ir_node *irn, void *env) { + ia32_am_opt_env_t *am_opt_env = env; + ia32_code_gen_t *cg = am_opt_env->cg; + heights_t *h = am_opt_env->h; + ir_node *block, *left, *right; + ir_node *store, *load, *mem_proj; + ir_node *succ, *addr_b, *addr_i; + int check_am_src = 0; + int need_exchange_on_fail = 0; + DEBUG_ONLY(firm_dbg_module_t *mod = cg->mod;) + + if (! is_ia32_irn(irn) || is_ia32_Ld(irn) || is_ia32_St(irn) || is_ia32_Store8Bit(irn)) + return; + + block = get_nodes_block(irn); + + DBG((mod, LEVEL_1, "checking for AM\n")); + + /* fold following patterns: */ /* - op -> Load into AMop with am_Source */ /* conditions: */ /* - op is am_Source capable AND */ @@ -688,231 +1538,277 @@ void ia32_optimize_am(ir_node *irn, void *env) { /* - the Load and Store are in the same block AND */ /* - nobody else uses the result of the op */ - if ((res == irn) && (get_ia32_am_support(irn) != ia32_am_None) && !is_ia32_Lea(irn)) { - /* 1st: check for Load/Store -> LEA */ - if (is_ia32_Ld(irn) || is_ia32_St(irn)) { - left = get_irn_n(irn, 0); + if ((get_ia32_am_support(irn) != ia32_am_None) && ! is_ia32_Lea(irn)) { + ia32_am_cand_t cand = is_am_candidate(cg, h, block, irn); + ia32_am_cand_t orig_cand = cand; - if (is_ia32_Lea(left)) { - DBG((mod, LEVEL_1, "\nmerging %+F into %+F\n", left, irn)); + /* cand == 1: load is left; cand == 2: load is right; */ - /* get the AM attributes from the LEA */ - add_ia32_am_offs(irn, get_ia32_am_offs(left)); - set_ia32_am_scale(irn, get_ia32_am_scale(left)); - set_ia32_am_flavour(irn, get_ia32_am_flavour(left)); + if (cand == IA32_AM_CAND_NONE) + return; - set_ia32_op_type(irn, is_ia32_St(irn) ? ia32_AddrModeD : ia32_AddrModeS); + DBG((mod, LEVEL_1, "\tfound address mode candidate %+F ... ", irn)); - /* set base and index */ - set_irn_n(irn, 0, get_irn_n(left, 0)); - set_irn_n(irn, 1, get_irn_n(left, 1)); - } + left = get_irn_n(irn, 2); + if (get_irn_arity(irn) == 4) { + /* it's an "unary" operation */ + right = left; + cand = IA32_AM_CAND_BOTH; + } + else { + right = get_irn_n(irn, 3); } - /* check if the node is an address mode candidate */ - else if (is_candidate(block, irn, 0)) { - DBG((mod, LEVEL_1, "\tfound address mode candidate %+F ... ", irn)); - left = get_irn_n(irn, 2); - if (get_irn_arity(irn) == 4) { - /* it's an "unary" operation */ - right = left; - } - else { - right = get_irn_n(irn, 3); - } + /* normalize commutative ops */ + if (node_is_ia32_comm(irn) && (cand == IA32_AM_CAND_RIGHT)) { + + /* Assure that left operand is always a Load if there is one */ + /* because non-commutative ops can only use Dest AM if the left */ + /* operand is a load, so we only need to check left operand. */ + + exchange_left_right(irn, &left, &right, 3, 2); + need_exchange_on_fail = 1; + + /* now: load is right */ + cand = IA32_AM_CAND_LEFT; + } - /* normalize commutative ops */ - if (node_is_comm(irn)) { - /* Assure that right operand is always a Load if there is one */ - /* because non-commutative ops can only use Dest AM if the right */ - /* operand is a load, so we only need to check right operand. */ - if (pred_is_specific_nodeblock(block, left, is_ia32_Ld)) - { - set_irn_n(irn, 2, right); - set_irn_n(irn, 3, left); - - temp = left; - left = right; - right = temp; + /* check for Store -> op -> Load */ + + /* Store -> op -> Load optimization is only possible if supported by op */ + /* and if right operand is a Load */ + if ((get_ia32_am_support(irn) & ia32_am_Dest) && (cand & IA32_AM_CAND_LEFT)) + { + /* An address mode capable op always has a result Proj. */ + /* If this Proj is used by more than one other node, we don't need to */ + /* check further, otherwise we check for Store and remember the address, */ + /* the Store points to. */ + + succ = ia32_get_res_proj(irn); + assert(succ && "Couldn't find result proj"); + + addr_b = NULL; + addr_i = NULL; + store = NULL; + + /* now check for users and Store */ + if (ia32_get_irn_n_edges(succ) == 1) { + succ = get_edge_src_irn(get_irn_out_edge_first(succ)); + + if (is_ia32_xStore(succ) || is_ia32_Store(succ)) { + store = succ; + addr_b = get_irn_n(store, 0); + addr_i = get_irn_n(store, 1); } } - /* check for Store -> op -> Load */ - - /* Store -> op -> Load optimization is only possible if supported by op */ - /* and if right operand is a Load */ - if ((get_ia32_am_support(irn) & ia32_am_Dest) && - pred_is_specific_nodeblock(block, right, is_ia32_Ld)) - { - - /* An address mode capable op always has a result Proj. */ - /* If this Proj is used by more than one other node, we don't need to */ - /* check further, otherwise we check for Store and remember the address, */ - /* the Store points to. */ - - succ = get_res_proj(irn); - assert(succ && "Couldn't find result proj"); - - addr_b = NULL; - addr_i = NULL; - store = NULL; - - /* now check for users and Store */ - if (ia32_get_irn_n_edges(succ) == 1) { - succ = get_edge_src_irn(get_irn_out_edge_first(succ)); - - if (is_ia32_fStore(succ) || is_ia32_Store(succ)) { - store = succ; - addr_b = get_irn_n(store, 0); - - /* Could be that the Store is connected to the address */ - /* calculating LEA while the Load is already transformed. */ - if (is_ia32_Lea(addr_b)) { - succ = addr_b; - addr_b = get_irn_n(succ, 0); - addr_i = get_irn_n(succ, 1); - } - else { - addr_i = noreg_gp; - } + if (store) { + /* we found a Store as single user: Now check for Load */ + + /* skip the Proj for easier access */ + load = is_Proj(right) ? (is_ia32_Load(get_Proj_pred(right)) ? get_Proj_pred(right) : NULL) : NULL; + + /* Extra check for commutative ops with two Loads */ + /* -> put the interesting Load left */ + if (load && node_is_ia32_comm(irn) && (cand == IA32_AM_CAND_BOTH)) { + if (load_store_addr_is_equal(load, store, addr_b, addr_i)) { + /* We exchange left and right, so it's easier to kill */ + /* the correct Load later and to handle unary operations. */ + exchange_left_right(irn, &left, &right, 3, 2); + need_exchange_on_fail ^= 1; } } - if (store) { - /* we found a Store as single user: Now check for Load */ - - /* Extra check for commutative ops with two Loads */ - /* -> put the interesting Load right */ - if (node_is_comm(irn) && - pred_is_specific_nodeblock(block, left, is_ia32_Ld)) - { - if ((addr_b == get_irn_n(get_Proj_pred(left), 0)) && - (addr_i == get_irn_n(get_Proj_pred(left), 1))) - { - /* We exchange left and right, so it's easier to kill */ - /* the correct Load later and to handle unary operations. */ - set_irn_n(irn, 2, right); - set_irn_n(irn, 3, left); - - temp = left; - left = right; - right = temp; - } + /* skip the Proj for easier access */ + load = get_Proj_pred(left); + + /* Compare Load and Store address */ + if (load_store_addr_is_equal(load, store, addr_b, addr_i)) { + /* Left Load is from same address, so we can */ + /* disconnect the Load and Store here */ + + /* set new base, index and attributes */ + set_irn_n(irn, 0, addr_b); + set_irn_n(irn, 1, addr_i); + add_ia32_am_offs(irn, get_ia32_am_offs(load)); + set_ia32_am_scale(irn, get_ia32_am_scale(load)); + set_ia32_am_flavour(irn, get_ia32_am_flavour(load)); + set_ia32_op_type(irn, ia32_AddrModeD); + set_ia32_frame_ent(irn, get_ia32_frame_ent(load)); + set_ia32_ls_mode(irn, get_ia32_ls_mode(load)); + + set_ia32_am_sc(irn, get_ia32_am_sc(load)); + if (is_ia32_am_sc_sign(load)) + set_ia32_am_sc_sign(irn); + + if (is_ia32_use_frame(load)) + set_ia32_use_frame(irn); + + /* connect to Load memory and disconnect Load */ + if (get_irn_arity(irn) == 5) { + /* binary AMop */ + set_irn_n(irn, 4, get_irn_n(load, 2)); + set_irn_n(irn, 2, ia32_get_admissible_noreg(cg, irn, 2)); + } + else { + /* unary AMop */ + set_irn_n(irn, 3, get_irn_n(load, 2)); + set_irn_n(irn, 2, ia32_get_admissible_noreg(cg, irn, 2)); } - /* skip the Proj for easier access */ - load = get_Proj_pred(right); + /* connect the memory Proj of the Store to the op */ + mem_proj = ia32_get_proj_for_mode(store, mode_M); + set_Proj_pred(mem_proj, irn); + set_Proj_proj(mem_proj, 1); - /* Compare Load and Store address */ - if (load_store_addr_is_equal(load, store, addr_b, addr_i)) { - /* Right Load is from same address, so we can */ - /* disconnect the Load and Store here */ - - /* set new base, index and attributes */ - set_irn_n(irn, 0, addr_b); - set_irn_n(irn, 1, addr_i); - add_ia32_am_offs(irn, get_ia32_am_offs(load)); - set_ia32_am_scale(irn, get_ia32_am_scale(load)); - set_ia32_am_flavour(irn, get_ia32_am_flavour(load)); - set_ia32_op_type(irn, ia32_AddrModeD); - set_ia32_frame_ent(irn, get_ia32_frame_ent(load)); - set_ia32_ls_mode(irn, get_ia32_ls_mode(load)); - - if (is_ia32_use_frame(load)) - set_ia32_use_frame(irn); - - /* connect to Load memory and disconnect Load */ - if (get_irn_arity(irn) == 5) { - /* binary AMop */ - set_irn_n(irn, 4, get_irn_n(load, 2)); - set_irn_n(irn, 3, noreg_gp); - } - else { - /* unary AMop */ - set_irn_n(irn, 3, get_irn_n(load, 2)); - set_irn_n(irn, 2, noreg_gp); - } - - /* connect the memory Proj of the Store to the op */ - mem_proj = get_mem_proj(store); - set_Proj_pred(mem_proj, irn); - set_Proj_proj(mem_proj, 1); - - DB((mod, LEVEL_1, "merged with %+F and %+F into dest AM\n", load, store)); - } - } /* if (store) */ - else if (get_ia32_am_support(irn) & ia32_am_Source) { - /* There was no store, check if we still can optimize for source address mode */ - check_am_src = 1; + /* clear remat flag */ + set_ia32_flags(irn, get_ia32_flags(irn) & ~arch_irn_flags_rematerializable); + + try_remove_from_sched(load); + try_remove_from_sched(store); + DBG_OPT_AM_D(load, store, irn); + + DB((mod, LEVEL_1, "merged with %+F and %+F into dest AM\n", load, store)); + + need_exchange_on_fail = 0; } - } /* if (support AM Dest) */ + } /* if (store) */ else if (get_ia32_am_support(irn) & ia32_am_Source) { - /* op doesn't support am AM Dest -> check for AM Source */ + /* There was no store, check if we still can optimize for source address mode */ check_am_src = 1; } + } /* if (support AM Dest) */ + else if (get_ia32_am_support(irn) & ia32_am_Source) { + /* op doesn't support am AM Dest -> check for AM Source */ + check_am_src = 1; + } - /* normalize commutative ops */ - if (node_is_comm(irn)) { - /* Assure that left operand is always a Load if there is one */ - /* because non-commutative ops can only use Source AM if the */ - /* left operand is a Load, so we only need to check the left */ - /* operand afterwards. */ - if (pred_is_specific_nodeblock(block, right, is_ia32_Ld)) { - set_irn_n(irn, 2, right); - set_irn_n(irn, 3, left); - - temp = left; - left = right; - right = temp; - } - } + /* was exchanged but optimize failed: exchange back */ + if (need_exchange_on_fail) { + exchange_left_right(irn, &left, &right, 3, 2); + cand = orig_cand; + } - /* optimize op -> Load iff Load is only used by this op */ - /* and left operand is a Load which only used by this irn */ - if (check_am_src && - pred_is_specific_nodeblock(block, left, is_ia32_Ld) && - (ia32_get_irn_n_edges(left) == 1)) - { - left = get_Proj_pred(left); - - addr_b = get_irn_n(left, 0); - addr_i = get_irn_n(left, 1); - - /* set new base, index and attributes */ - set_irn_n(irn, 0, addr_b); - set_irn_n(irn, 1, addr_i); - add_ia32_am_offs(irn, get_ia32_am_offs(left)); - set_ia32_am_scale(irn, get_ia32_am_scale(left)); - set_ia32_am_flavour(irn, get_ia32_am_flavour(left)); - set_ia32_op_type(irn, ia32_AddrModeS); - set_ia32_frame_ent(irn, get_ia32_frame_ent(left)); - set_ia32_ls_mode(irn, get_ia32_ls_mode(left)); - - if (is_ia32_use_frame(left)) - set_ia32_use_frame(irn); - - /* connect to Load memory */ - if (get_irn_arity(irn) == 5) { - /* binary AMop */ - set_irn_n(irn, 4, get_irn_n(left, 2)); - } - else { - /* unary AMop */ - set_irn_n(irn, 3, get_irn_n(left, 2)); - } + need_exchange_on_fail = 0; - /* disconnect from Load */ - set_irn_n(irn, 2, noreg_gp); + /* normalize commutative ops */ + if (check_am_src && node_is_ia32_comm(irn) && (cand == IA32_AM_CAND_LEFT)) { - /* If Load has a memory Proj, connect it to the op */ - mem_proj = get_mem_proj(left); - if (mem_proj) { - set_Proj_pred(mem_proj, irn); - set_Proj_proj(mem_proj, 1); - } + /* Assure that right operand is always a Load if there is one */ + /* because non-commutative ops can only use Source AM if the */ + /* right operand is a Load, so we only need to check the right */ + /* operand afterwards. */ + + exchange_left_right(irn, &left, &right, 3, 2); + need_exchange_on_fail = 1; + + /* now: load is left */ + cand = IA32_AM_CAND_RIGHT; + } + + /* optimize op -> Load iff Load is only used by this op */ + /* and right operand is a Load which only used by this irn */ + if (check_am_src && + (cand & IA32_AM_CAND_RIGHT) && + (ia32_get_irn_n_edges(right) == 1)) + { + ir_node *load = get_Proj_pred(right); + + addr_b = get_irn_n(load, 0); + addr_i = get_irn_n(load, 1); + + /* set new base, index and attributes */ + set_irn_n(irn, 0, addr_b); + set_irn_n(irn, 1, addr_i); + add_ia32_am_offs(irn, get_ia32_am_offs(load)); + set_ia32_am_scale(irn, get_ia32_am_scale(load)); + set_ia32_am_flavour(irn, get_ia32_am_flavour(load)); + set_ia32_op_type(irn, ia32_AddrModeS); + set_ia32_frame_ent(irn, get_ia32_frame_ent(load)); + set_ia32_ls_mode(irn, get_ia32_ls_mode(load)); + + set_ia32_am_sc(irn, get_ia32_am_sc(load)); + if (is_ia32_am_sc_sign(load)) + set_ia32_am_sc_sign(irn); + + /* clear remat flag */ + set_ia32_flags(irn, get_ia32_flags(irn) & ~arch_irn_flags_rematerializable); + + if (is_ia32_use_frame(load)) + set_ia32_use_frame(irn); + + /* connect to Load memory and disconnect Load */ + if (get_irn_arity(irn) == 5) { + /* binary AMop */ + set_irn_n(irn, 4, get_irn_n(load, 2)); + set_irn_n(irn, 3, ia32_get_admissible_noreg(cg, irn, 3)); + } else { + assert(get_irn_arity(irn) == 4); + /* unary AMop */ + set_irn_n(irn, 3, get_irn_n(load, 2)); + set_irn_n(irn, 2, ia32_get_admissible_noreg(cg, irn, 2)); + } + + /* this is only needed for Compares, but currently ALL nodes + * have this attribute :-) */ + set_ia32_pncode(irn, get_inversed_pnc(get_ia32_pncode(irn))); + + DBG_OPT_AM_S(load, irn); - DB((mod, LEVEL_1, "merged with %+F into source AM\n", left)); + /* If Load has a memory Proj, connect it to the op */ + mem_proj = ia32_get_proj_for_mode(load, mode_M); + if (mem_proj) { + set_Proj_pred(mem_proj, irn); + set_Proj_proj(mem_proj, 1); } + + try_remove_from_sched(load); + + DB((mod, LEVEL_1, "merged with %+F into source AM\n", load)); + } + else { + /* was exchanged but optimize failed: exchange back */ + if (need_exchange_on_fail) + exchange_left_right(irn, &left, &right, 3, 2); } } } + +/** + * Performs address mode optimization. + */ +void ia32_optimize_addressmode(ia32_code_gen_t *cg) { + /* if we are supposed to do AM or LEA optimization: recalculate edges */ + if (cg->opt & (IA32_OPT_DOAM | IA32_OPT_LEA)) { + edges_deactivate(cg->irg); + edges_activate(cg->irg); + } + else { + /* no optimizations at all */ + return; + } + + /* beware: we cannot optimize LEA and AM in one run because */ + /* LEA optimization adds new nodes to the irg which */ + /* invalidates the phase data */ + + if (cg->opt & IA32_OPT_LEA) { + irg_walk_blkwise_graph(cg->irg, NULL, optimize_lea, cg); + } + + if (cg->dump) + be_dump(cg->irg, "-lea", dump_ir_block_graph_sched); + + if (cg->opt & IA32_OPT_DOAM) { + /* we need height information for am optimization */ + heights_t *h = heights_new(cg->irg); + ia32_am_opt_env_t env; + + env.cg = cg; + env.h = h; + + irg_walk_blkwise_graph(cg->irg, NULL, optimize_am, &env); + + heights_free(h); + } +}