X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fia32%2Fia32_optimize.c;h=f67ae7e19b2cf8e889a3c9c627e73097ed886fe6;hb=c2ead6606a9c2f318a628b152ec0ce16cbbe4483;hp=16b478c3bbc92bc0dc8b8d538f0cadc33b85b57e;hpb=277830bf1592378d809bc22b0a0c3c9c5bdbdd8c;p=libfirm diff --git a/ir/be/ia32/ia32_optimize.c b/ir/be/ia32/ia32_optimize.c index 16b478c3b..f67ae7e19 100644 --- a/ir/be/ia32/ia32_optimize.c +++ b/ir/be/ia32/ia32_optimize.c @@ -1,5 +1,5 @@ /* - * Copyright (C) 1995-2007 University of Karlsruhe. All right reserved. + * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved. * * This file is part of libFirm. * @@ -38,6 +38,8 @@ #include "irgwalk.h" #include "height.h" #include "irbitset.h" +#include "irprintf.h" +#include "error.h" #include "../be_t.h" #include "../beabi.h" @@ -46,52 +48,268 @@ #include "../bepeephole.h" #include "ia32_new_nodes.h" +#include "ia32_optimize.h" #include "bearch_ia32_t.h" #include "gen_ia32_regalloc_if.h" #include "ia32_transform.h" #include "ia32_dbg_stat.h" #include "ia32_util.h" +#include "ia32_architecture.h" DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;) static const arch_env_t *arch_env; static ia32_code_gen_t *cg; -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); +/** + * Returns non-zero if the given node produces + * a zero flag. + * + * @param node the node to check + * @param pn if >= 0, the projection number of the used result + */ +static int produces_zero_flag(ir_node *node, int pn) +{ + ir_node *count; + const ia32_immediate_attr_t *imm_attr; + + if (!is_ia32_irn(node)) + return 0; + + if (pn >= 0) { + if (pn != pn_ia32_res) + return 0; + } + + switch (get_ia32_irn_opcode(node)) { + case iro_ia32_Add: + case iro_ia32_Adc: + case iro_ia32_And: + case iro_ia32_Or: + case iro_ia32_Xor: + case iro_ia32_Sub: + case iro_ia32_Sbb: + case iro_ia32_Neg: + case iro_ia32_Inc: + case iro_ia32_Dec: + return 1; + + case iro_ia32_ShlD: + case iro_ia32_ShrD: + case iro_ia32_Shl: + case iro_ia32_Shr: + case iro_ia32_Sar: + assert(n_ia32_ShlD_count == n_ia32_ShrD_count); + assert(n_ia32_Shl_count == n_ia32_Shr_count + && n_ia32_Shl_count == n_ia32_Sar_count); + if (is_ia32_ShlD(node) || is_ia32_ShrD(node)) { + count = get_irn_n(node, n_ia32_ShlD_count); + } else { + count = get_irn_n(node, n_ia32_Shl_count); + } + /* when shift count is zero the flags are not affected, so we can only + * do this for constants != 0 */ + if (!is_ia32_Immediate(count)) + return 0; + + imm_attr = get_ia32_immediate_attr_const(count); + if (imm_attr->symconst != NULL) + return 0; + if ((imm_attr->offset & 0x1f) == 0) + return 0; + return 1; + + default: + break; + } + return 0; +} /** - * checks if a node represents the NOREG value + * If the given node has not mode_T, creates a mode_T version (with a result Proj). + * + * @param node the node to change + * + * @return the new mode_T node (if the mode was changed) or node itself */ -static INLINE int be_is_NoReg(ia32_code_gen_t *cg, const ir_node *irn) { - return irn == cg->noreg_gp || irn == cg->noreg_xmm || irn == cg->noreg_vfp; +static ir_node *turn_into_mode_t(ir_node *node) +{ + ir_node *block; + ir_node *res_proj; + ir_node *new_node; + const arch_register_t *reg; + + if(get_irn_mode(node) == mode_T) + return node; + + assert(get_irn_mode(node) == mode_Iu); + + new_node = exact_copy(node); + set_irn_mode(new_node, mode_T); + + block = get_nodes_block(new_node); + res_proj = new_r_Proj(current_ir_graph, block, new_node, mode_Iu, + pn_ia32_res); + + reg = arch_get_irn_register(arch_env, node); + arch_set_irn_register(arch_env, res_proj, reg); + + be_peephole_before_exchange(node, res_proj); + sched_add_before(node, new_node); + sched_remove(node); + exchange(node, res_proj); + be_peephole_after_exchange(res_proj); + + return new_node; } -/******************************************************************************************************** - * _____ _ _ ____ _ _ _ _ _ - * | __ \ | | | | / __ \ | | (_) (_) | | (_) - * | |__) |__ ___ _ __ | |__ ___ | | ___ | | | |_ __ | |_ _ _ __ ___ _ ______ _| |_ _ ___ _ __ - * | ___/ _ \/ _ \ '_ \| '_ \ / _ \| |/ _ \ | | | | '_ \| __| | '_ ` _ \| |_ / _` | __| |/ _ \| '_ \ - * | | | __/ __/ |_) | | | | (_) | | __/ | |__| | |_) | |_| | | | | | | |/ / (_| | |_| | (_) | | | | - * |_| \___|\___| .__/|_| |_|\___/|_|\___| \____/| .__/ \__|_|_| |_| |_|_/___\__,_|\__|_|\___/|_| |_| - * | | | | - * |_| |_| - ********************************************************************************************************/ +/** + * Peephole optimization for Test instructions. + * We can remove the Test, if a zero flags was produced which is still + * live. + */ +static void peephole_ia32_Test(ir_node *node) +{ + ir_node *left = get_irn_n(node, n_ia32_Test_left); + ir_node *right = get_irn_n(node, n_ia32_Test_right); + ir_node *flags_proj; + ir_node *block; + ir_mode *flags_mode; + int pn = -1; + ir_node *schedpoint; + const ir_edge_t *edge; + + assert(n_ia32_Test_left == n_ia32_Test8Bit_left + && n_ia32_Test_right == n_ia32_Test8Bit_right); + + /* we need a test for 0 */ + if(left != right) + return; + + block = get_nodes_block(node); + if(get_nodes_block(left) != block) + return; + + if(is_Proj(left)) { + pn = get_Proj_proj(left); + left = get_Proj_pred(left); + } + + /* happens rarely, but if it does code will panic' */ + if (is_ia32_Unknown_GP(left)) + return; + + /* walk schedule up and abort when we find left or some other node destroys + the flags */ + schedpoint = sched_prev(node); + while(schedpoint != left) { + schedpoint = sched_prev(schedpoint); + if(arch_irn_is(arch_env, schedpoint, modify_flags)) + return; + if(schedpoint == block) + panic("couldn't find left"); + } + + /* make sure only Lg/Eq tests are used */ + foreach_out_edge(node, edge) { + ir_node *user = get_edge_src_irn(edge); + int pnc = get_ia32_condcode(user); + + if(pnc != pn_Cmp_Eq && pnc != pn_Cmp_Lg) { + return; + } + } + + if(!produces_zero_flag(left, pn)) + return; + + left = turn_into_mode_t(left); + + flags_mode = ia32_reg_classes[CLASS_ia32_flags].mode; + flags_proj = new_r_Proj(current_ir_graph, block, left, flags_mode, + pn_ia32_flags); + arch_set_irn_register(arch_env, flags_proj, &ia32_flags_regs[REG_EFLAGS]); + + assert(get_irn_mode(node) != mode_T); + + be_peephole_before_exchange(node, flags_proj); + exchange(node, flags_proj); + sched_remove(node); + be_peephole_after_exchange(flags_proj); +} /** - * NOTE: THESE PEEPHOLE OPTIMIZATIONS MUST BE CALLED AFTER SCHEDULING AND REGISTER ALLOCATION. + * AMD Athlon works faster when RET is not destination of + * conditional jump or directly preceded by other jump instruction. + * Can be avoided by placing a Rep prefix before the return. */ +static void peephole_ia32_Return(ir_node *node) { + ir_node *block, *irn; + + if (!ia32_cg_config.use_pad_return) + return; + + block = get_nodes_block(node); + + if (get_Block_n_cfgpreds(block) == 1) { + ir_node *pred = get_Block_cfgpred(block, 0); -// only optimize up to 48 stores behind IncSPs + if (is_Jmp(pred)) { + /* The block of the return has only one predecessor, + which jumps directly to this block. + This jump will be encoded as a fall through, so we + ignore it here. + However, the predecessor might be empty, so it must be + ensured that empty blocks are gone away ... */ + return; + } + } + + /* check if this return is the first on the block */ + sched_foreach_reverse_from(node, irn) { + switch (get_irn_opcode(irn)) { + case beo_Return: + /* the return node itself, ignore */ + continue; + case beo_Barrier: + /* ignore the barrier, no code generated */ + continue; + case beo_IncSP: + /* arg, IncSP 0 nodes might occur, ignore these */ + if (be_get_IncSP_offset(irn) == 0) + continue; + return; + case iro_Phi: + continue; + default: + return; + } + } + /* yep, return is the first real instruction in this block */ +#if 0 + { + /* add an rep prefix to the return */ + ir_node *rep = new_rd_ia32_RepPrefix(get_irn_dbg_info(node), current_ir_graph, block); + keep_alive(rep); + sched_add_before(node, rep); + } +#else + /* ensure, that the 3 byte return is generated */ + be_Return_set_emit_pop(node, 1); +#endif +} + +/* only optimize up to 48 stores behind IncSPs */ #define MAXPUSH_OPTIMIZE 48 /** - * Tries to create pushs from IncSP,Store combinations + * Tries to create Push's from IncSP, Store combinations. + * The Stores are replaced by Push's, the IncSP is modified + * (possibly into IncSP 0, but not removed). */ -static void ia32_create_Pushs(ir_node *irn) +static void peephole_IncSP_Store_to_push(ir_node *irn) { - int i; - int offset; + int i, maxslot, inc_ofs; ir_node *node; ir_node *stores[MAXPUSH_OPTIMIZE]; ir_node *block = get_nodes_block(irn); @@ -103,8 +321,8 @@ static void ia32_create_Pushs(ir_node *irn) assert(be_is_IncSP(irn)); - offset = be_get_IncSP_offset(irn); - if(offset < 4) + inc_ofs = be_get_IncSP_offset(irn); + if (inc_ofs < 4) return; /* @@ -113,62 +331,63 @@ static void ia32_create_Pushs(ir_node *irn) * 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)) { + maxslot = -1; + for (node = sched_next(irn); !sched_is_end(node); node = sched_next(node)) { ir_node *mem; int offset; int storeslot; - // it has to be a store - if(!is_ia32_Store(node)) + /* it has to be a Store */ + if (!is_ia32_Store(node)) break; - // it has to use our sp value - if(get_irn_n(node, n_ia32_base) != irn) + /* it has to use our sp value */ + if (get_irn_n(node, n_ia32_base) != irn) continue; - // store has to be attached to NoMem + /* Store has to be attached to NoMem */ mem = get_irn_n(node, n_ia32_mem); - if(!is_NoMem(mem)) { + if (!is_NoMem(mem)) continue; - } /* unfortunately we can't support the full AMs possible for push at the * moment. TODO: fix this */ - if(get_ia32_am_scale(node) > 0 || !is_ia32_NoReg_GP(get_irn_n(node, n_ia32_index))) + if (get_ia32_am_scale(node) > 0 || !is_ia32_NoReg_GP(get_irn_n(node, n_ia32_index))) break; offset = get_ia32_am_offs_int(node); + /* we should NEVER access uninitialized stack BELOW the current SP */ + assert(offset >= 0); - storeslot = offset / 4; - if(storeslot >= MAXPUSH_OPTIMIZE) - continue; + offset = inc_ofs - 4 - offset; - // storing into the same slot twice is bad (and shouldn't happen...) - if(stores[storeslot] != NULL) + /* storing at half-slots is bad */ + if ((offset & 3) != 0) break; - // storing at half-slots is bad - if(offset % 4 != 0) + if (offset < 0 || offset >= MAXPUSH_OPTIMIZE * 4) + continue; + storeslot = offset >> 2; + + /* storing into the same slot twice is bad (and shouldn't happen...) */ + if (stores[storeslot] != NULL) break; stores[storeslot] = node; + if (storeslot > maxslot) + maxslot = storeslot; } curr_sp = be_get_IncSP_pred(irn); - // 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) { + /* walk through the stores and create Pushs for them */ + for (i = 0; i <= maxslot; ++i) { const arch_register_t *spreg; ir_node *push; ir_node *val, *mem, *mem_proj; ir_node *store = stores[i]; ir_node *noreg = ia32_new_NoReg_gp(cg); - if(store == NULL || is_Bad(store)) + if (store == NULL) break; val = get_irn_n(store, n_ia32_unary_op); @@ -177,110 +396,187 @@ static void ia32_create_Pushs(ir_node *irn) push = new_rd_ia32_Push(get_irn_dbg_info(store), irg, block, noreg, noreg, mem, val, curr_sp); - set_ia32_am_support(push, ia32_am_Source, ia32_am_unary); - sched_add_before(irn, push); - // create stackpointer proj + /* 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); - // create memory proj + /* create memory Proj */ mem_proj = new_r_Proj(irg, block, push, mode_M, pn_ia32_Push_M); - // use the memproj now + /* use the memproj now */ exchange(store, mem_proj); - // we can remove the store now + /* we can remove the store now */ sched_remove(store); - offset -= 4; + inc_ofs -= 4; } - be_set_IncSP_offset(irn, offset); - - // can we remove the IncSP now? - if(offset == 0) { - const ir_edge_t *edge, *next; + be_set_IncSP_offset(irn, inc_ofs); + be_set_IncSP_pred(irn, curr_sp); +} - foreach_out_edge_safe(irn, edge, next) { - ir_node *arg = get_edge_src_irn(edge); - int pos = get_edge_src_pos(edge); +/** + * Find a free GP register if possible, else return NULL. + */ +static const arch_register_t *get_free_gp_reg(void) +{ + int i; - set_irn_n(arg, pos, curr_sp); - } + for(i = 0; i < N_ia32_gp_REGS; ++i) { + const arch_register_t *reg = &ia32_gp_regs[i]; + if(arch_register_type_is(reg, ignore)) + continue; - set_irn_n(irn, 0, new_Bad()); - sched_remove(irn); - } else { - set_irn_n(irn, 0, curr_sp); + if(be_peephole_get_value(CLASS_ia32_gp, i) == NULL) + return &ia32_gp_regs[i]; } + + return NULL; } /** - * Tries to optimize two following IncSP. + * Creates a Pop instruction before the given schedule point. + * + * @param dbgi debug info + * @param irg the graph + * @param block the block + * @param stack the previous stack value + * @param schedpoint the new node is added before this node + * @param reg the register to pop + * + * @return the new stack value */ -static void ia32_optimize_IncSP(ir_node *node) +static ir_node *create_pop(dbg_info *dbgi, ir_graph *irg, ir_node *block, + ir_node *stack, ir_node *schedpoint, + const arch_register_t *reg) { - int pred_offs; - int curr_offs; - int offs; - ir_node *pred = be_get_IncSP_pred(node); - ir_node *predpred; + const arch_register_t *esp = &ia32_gp_regs[REG_ESP]; + ir_node *pop; + ir_node *keep; + ir_node *val; + ir_node *in[1]; - if(!be_is_IncSP(pred)) - return; + pop = new_rd_ia32_Pop(dbgi, irg, block, new_NoMem(), stack); - if(get_irn_n_edges(pred) > 1) - return; + stack = new_r_Proj(irg, block, pop, mode_Iu, pn_ia32_Pop_stack); + arch_set_irn_register(arch_env, stack, esp); + val = new_r_Proj(irg, block, pop, mode_Iu, pn_ia32_Pop_res); + arch_set_irn_register(arch_env, val, reg); - pred_offs = be_get_IncSP_offset(pred); - curr_offs = be_get_IncSP_offset(node); + sched_add_before(schedpoint, pop); - if(pred_offs == BE_STACK_FRAME_SIZE_EXPAND) { - if(curr_offs != BE_STACK_FRAME_SIZE_SHRINK) { - return; - } - offs = 0; - } else if(pred_offs == BE_STACK_FRAME_SIZE_SHRINK) { - if(curr_offs != BE_STACK_FRAME_SIZE_EXPAND) { - return; - } - offs = 0; - } else if(curr_offs == BE_STACK_FRAME_SIZE_EXPAND - || curr_offs == BE_STACK_FRAME_SIZE_SHRINK) { - return; - } else { - offs = curr_offs + pred_offs; - } + in[0] = val; + keep = be_new_Keep(&ia32_reg_classes[CLASS_ia32_gp], irg, block, 1, in); + sched_add_before(schedpoint, keep); - be_set_IncSP_offset(node, offs); + return stack; +} - /* rewire dependency edges */ - predpred = be_get_IncSP_pred(pred); - edges_reroute_kind(pred, predpred, EDGE_KIND_DEP, current_ir_graph); +/** + * Creates a Push instruction before the given schedule point. + * + * @param dbgi debug info + * @param irg the graph + * @param block the block + * @param stack the previous stack value + * @param schedpoint the new node is added before this node + * @param reg the register to pop + * + * @return the new stack value + */ +static ir_node *create_push(dbg_info *dbgi, ir_graph *irg, ir_node *block, + ir_node *stack, ir_node *schedpoint, + const arch_register_t *reg) +{ + const arch_register_t *esp = &ia32_gp_regs[REG_ESP]; + ir_node *noreg, *nomem, *push, *val; + + val = new_rd_ia32_ProduceVal(NULL, irg, block); + arch_set_irn_register(arch_env, val, reg); + sched_add_before(schedpoint, val); + + noreg = ia32_new_NoReg_gp(cg); + nomem = get_irg_no_mem(irg); + push = new_rd_ia32_Push(dbgi, irg, block, noreg, noreg, nomem, val, stack); + sched_add_before(schedpoint, push); - /* Omit the IncSP */ - be_set_IncSP_pred(node, predpred); - sched_remove(pred); - be_kill_node(pred); + stack = new_r_Proj(irg, block, push, mode_Iu, pn_ia32_Push_stack); + arch_set_irn_register(arch_env, stack, esp); + + return stack; } /** - * Performs Peephole Optimizations. + * Optimize an IncSp by replacing it with Push/Pop. */ -static void ia32_peephole_optimize_node(ir_node *node, void *env) +static void peephole_be_IncSP(ir_node *node) { - (void) env; - if (be_is_IncSP(node)) { - ia32_optimize_IncSP(node); + const arch_register_t *esp = &ia32_gp_regs[REG_ESP]; + const arch_register_t *reg; + ir_graph *irg = current_ir_graph; + dbg_info *dbgi; + ir_node *block; + ir_node *stack; + int offset; + + /* first optimize incsp->incsp combinations */ + be_peephole_IncSP_IncSP(node); + + /* transform IncSP->Store combinations to Push where possible */ + peephole_IncSP_Store_to_push(node); + + if (arch_get_irn_register(arch_env, node) != esp) + return; + + /* replace IncSP -4 by Pop freereg when possible */ + offset = be_get_IncSP_offset(node); + if ((offset != -8 || ia32_cg_config.use_add_esp_8) && + (offset != -4 || ia32_cg_config.use_add_esp_4) && + (offset != +4 || ia32_cg_config.use_sub_esp_4) && + (offset != +8 || ia32_cg_config.use_sub_esp_8)) + return; + + if (offset < 0) { + /* we need a free register for pop */ + reg = get_free_gp_reg(); + if (reg == NULL) + return; + + dbgi = get_irn_dbg_info(node); + block = get_nodes_block(node); + stack = be_get_IncSP_pred(node); - if (cg->opt & IA32_OPT_PUSHARGS) - ia32_create_Pushs(node); + stack = create_pop(dbgi, irg, block, stack, node, reg); + + if (offset == -8) { + stack = create_pop(dbgi, irg, block, stack, node, reg); + } + } else { + dbgi = get_irn_dbg_info(node); + block = get_nodes_block(node); + stack = be_get_IncSP_pred(node); + reg = &ia32_gp_regs[REG_EAX]; + + stack = create_push(dbgi, irg, block, stack, node, reg); + + if (offset == +8) { + stack = create_push(dbgi, irg, block, stack, node, reg); + } } + + be_peephole_before_exchange(node, stack); + sched_remove(node); + exchange(node, stack); + be_peephole_after_exchange(stack); } -static ir_node *optimize_ia32_Const(ir_node *node) +/** + * Peephole optimisation for ia32_Const's + */ +static void peephole_ia32_Const(ir_node *node) { const ia32_immediate_attr_t *attr = get_ia32_immediate_attr_const(node); const arch_register_t *reg; @@ -292,11 +588,13 @@ static ir_node *optimize_ia32_Const(ir_node *node) ir_node *noreg; /* try to transform a mov 0, reg to xor reg reg */ - if(attr->offset != 0 || attr->symconst != NULL) - return NULL; - /* xor destroys the flags, so noone must be using them */ - if(be_peephole_get_value(CLASS_ia32_flags, REG_EFLAGS) != NULL) - return NULL; + if (attr->offset != 0 || attr->symconst != NULL) + return; + if (ia32_cg_config.use_mov_0) + return; + /* xor destroys the flags, so no-one must be using them */ + if (be_peephole_get_value(CLASS_ia32_flags, REG_EFLAGS) != NULL) + return; reg = arch_get_irn_register(arch_env, node); assert(be_peephole_get_reg_value(reg) == NULL); @@ -314,29 +612,313 @@ static ir_node *optimize_ia32_Const(ir_node *node) sched_add_before(node, produceval); sched_add_before(node, xor); + + be_peephole_before_exchange(node, xor); exchange(node, xor); sched_remove(node); + be_peephole_after_exchange(xor); +} - return xor; +static INLINE int is_noreg(ia32_code_gen_t *cg, const ir_node *node) +{ + return node == cg->noreg_gp; +} + +static ir_node *create_immediate_from_int(ia32_code_gen_t *cg, int val) +{ + ir_graph *irg = current_ir_graph; + ir_node *start_block = get_irg_start_block(irg); + ir_node *immediate = new_rd_ia32_Immediate(NULL, irg, start_block, NULL, + 0, val); + arch_set_irn_register(cg->arch_env, immediate, &ia32_gp_regs[REG_GP_NOREG]); + + return immediate; +} + +static ir_node *create_immediate_from_am(ia32_code_gen_t *cg, + const ir_node *node) +{ + ir_graph *irg = get_irn_irg(node); + ir_node *block = get_nodes_block(node); + int offset = get_ia32_am_offs_int(node); + int sc_sign = is_ia32_am_sc_sign(node); + ir_entity *entity = get_ia32_am_sc(node); + ir_node *res; + + res = new_rd_ia32_Immediate(NULL, irg, block, entity, sc_sign, offset); + arch_set_irn_register(cg->arch_env, res, &ia32_gp_regs[REG_GP_NOREG]); + return res; } -static void register_peephole_optimisation(ir_op *op, peephole_opt_func func) +static int is_am_one(const ir_node *node) { + int offset = get_ia32_am_offs_int(node); + ir_entity *entity = get_ia32_am_sc(node); + + return offset == 1 && entity == NULL; +} + +static int is_am_minus_one(const ir_node *node) +{ + int offset = get_ia32_am_offs_int(node); + ir_entity *entity = get_ia32_am_sc(node); + + return offset == -1 && entity == NULL; +} + +/** + * Transforms a LEA into an Add or SHL if possible. + */ +static void peephole_ia32_Lea(ir_node *node) +{ + const arch_env_t *arch_env = cg->arch_env; + ir_graph *irg = current_ir_graph; + ir_node *base; + ir_node *index; + const arch_register_t *base_reg; + const arch_register_t *index_reg; + const arch_register_t *out_reg; + int scale; + int has_immediates; + ir_node *op1; + ir_node *op2; + dbg_info *dbgi; + ir_node *block; + ir_node *res; + ir_node *noreg; + ir_node *nomem; + + assert(is_ia32_Lea(node)); + + /* we can only do this if are allowed to globber the flags */ + if(be_peephole_get_value(CLASS_ia32_flags, REG_EFLAGS) != NULL) + return; + + base = get_irn_n(node, n_ia32_Lea_base); + index = get_irn_n(node, n_ia32_Lea_index); + + if(is_noreg(cg, base)) { + base = NULL; + base_reg = NULL; + } else { + base_reg = arch_get_irn_register(arch_env, base); + } + if(is_noreg(cg, index)) { + index = NULL; + index_reg = NULL; + } else { + index_reg = arch_get_irn_register(arch_env, index); + } + + if(base == NULL && index == NULL) { + /* we shouldn't construct these in the first place... */ +#ifdef DEBUG_libfirm + ir_fprintf(stderr, "Optimisation warning: found immediate only lea\n"); +#endif + return; + } + + out_reg = arch_get_irn_register(arch_env, node); + scale = get_ia32_am_scale(node); + assert(!is_ia32_need_stackent(node) || get_ia32_frame_ent(node) != NULL); + /* check if we have immediates values (frame entities should already be + * expressed in the offsets) */ + if(get_ia32_am_offs_int(node) != 0 || get_ia32_am_sc(node) != NULL) { + has_immediates = 1; + } else { + has_immediates = 0; + } + + /* we can transform leas where the out register is the same as either the + * base or index register back to an Add or Shl */ + if(out_reg == base_reg) { + if(index == NULL) { +#ifdef DEBUG_libfirm + if(!has_immediates) { + ir_fprintf(stderr, "Optimisation warning: found lea which is " + "just a copy\n"); + } +#endif + op1 = base; + goto make_add_immediate; + } + if(scale == 0 && !has_immediates) { + op1 = base; + op2 = index; + goto make_add; + } + /* can't create an add */ + return; + } else if(out_reg == index_reg) { + if(base == NULL) { + if(has_immediates && scale == 0) { + op1 = index; + goto make_add_immediate; + } else if(!has_immediates && scale > 0) { + op1 = index; + op2 = create_immediate_from_int(cg, scale); + goto make_shl; + } else if(!has_immediates) { +#ifdef DEBUG_libfirm + ir_fprintf(stderr, "Optimisation warning: found lea which is " + "just a copy\n"); +#endif + } + } else if(scale == 0 && !has_immediates) { + op1 = index; + op2 = base; + goto make_add; + } + /* can't create an add */ + return; + } else { + /* can't create an add */ + return; + } + +make_add_immediate: + if(ia32_cg_config.use_incdec) { + if(is_am_one(node)) { + dbgi = get_irn_dbg_info(node); + block = get_nodes_block(node); + res = new_rd_ia32_Inc(dbgi, irg, block, op1); + arch_set_irn_register(arch_env, res, out_reg); + goto exchange; + } + if(is_am_minus_one(node)) { + dbgi = get_irn_dbg_info(node); + block = get_nodes_block(node); + res = new_rd_ia32_Dec(dbgi, irg, block, op1); + arch_set_irn_register(arch_env, res, out_reg); + goto exchange; + } + } + op2 = create_immediate_from_am(cg, node); + +make_add: + dbgi = get_irn_dbg_info(node); + block = get_nodes_block(node); + noreg = ia32_new_NoReg_gp(cg); + nomem = new_NoMem(); + res = new_rd_ia32_Add(dbgi, irg, block, noreg, noreg, nomem, op1, op2); + arch_set_irn_register(arch_env, res, out_reg); + set_ia32_commutative(res); + goto exchange; + +make_shl: + dbgi = get_irn_dbg_info(node); + block = get_nodes_block(node); + noreg = ia32_new_NoReg_gp(cg); + nomem = new_NoMem(); + res = new_rd_ia32_Shl(dbgi, irg, block, op1, op2); + arch_set_irn_register(arch_env, res, out_reg); + goto exchange; + +exchange: + SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(cg, node)); + + /* add new ADD/SHL to schedule */ + DBG_OPT_LEA2ADD(node, res); + + /* exchange the Add and the LEA */ + be_peephole_before_exchange(node, res); + sched_add_before(node, res); + sched_remove(node); + exchange(node, res); + be_peephole_after_exchange(res); +} + +/** + * Split a Imul mem, imm into a Load mem and Imul reg, imm if possible. + */ +static void peephole_ia32_Imul_split(ir_node *imul) { + const ir_node *right = get_irn_n(imul, n_ia32_IMul_right); + const arch_register_t *reg; + ir_node *load, *block, *base, *index, *mem, *res, *noreg; + dbg_info *dbgi; + ir_graph *irg; + + if (! is_ia32_Immediate(right) || get_ia32_op_type(imul) != ia32_AddrModeS) { + /* no memory, imm form ignore */ + return; + } + /* we need a free register */ + reg = get_free_gp_reg(); + if (reg == NULL) + return; + + /* fine, we can rebuild it */ + dbgi = get_irn_dbg_info(imul); + block = get_nodes_block(imul); + irg = current_ir_graph; + base = get_irn_n(imul, n_ia32_IMul_base); + index = get_irn_n(imul, n_ia32_IMul_index); + mem = get_irn_n(imul, n_ia32_IMul_mem); + load = new_rd_ia32_Load(dbgi, irg, block, base, index, mem); + + /* copy all attributes */ + set_irn_pinned(load, get_irn_pinned(imul)); + set_ia32_op_type(load, ia32_AddrModeS); + set_ia32_ls_mode(load, get_ia32_ls_mode(imul)); + + set_ia32_am_scale(load, get_ia32_am_scale(imul)); + set_ia32_am_sc(load, get_ia32_am_sc(imul)); + set_ia32_am_offs_int(load, get_ia32_am_offs_int(imul)); + if (is_ia32_am_sc_sign(imul)) + set_ia32_am_sc_sign(load); + if (is_ia32_use_frame(imul)) + set_ia32_use_frame(load); + set_ia32_frame_ent(load, get_ia32_frame_ent(imul)); + + sched_add_before(imul, load); + + mem = new_rd_Proj(dbgi, irg, block, load, mode_M, pn_ia32_Load_M); + res = new_rd_Proj(dbgi, irg, block, load, mode_Iu, pn_ia32_Load_res); + + arch_set_irn_register(arch_env, res, reg); + be_peephole_after_exchange(res); + + set_irn_n(imul, n_ia32_IMul_mem, mem); + noreg = get_irn_n(imul, n_ia32_IMul_left); + set_irn_n(imul, n_ia32_IMul_left, res); + set_ia32_op_type(imul, ia32_Normal); +} + +/** + * Replace xorps r,r and xorpd r,r by pxor r,r + */ +static void peephole_ia32_xZero(ir_node *xor) { + set_irn_op(xor, op_ia32_xPzero); +} + +/** + * Register a peephole optimisation function. + */ +static void register_peephole_optimisation(ir_op *op, peephole_opt_func func) { assert(op->ops.generic == NULL); - op->ops.generic = (void*) func; + op->ops.generic = (op_func)func; } -void ia32_peephole_optimization(ir_graph *irg, ia32_code_gen_t *new_cg) +/* Perform peephole-optimizations. */ +void ia32_peephole_optimization(ia32_code_gen_t *new_cg) { cg = new_cg; arch_env = cg->arch_env; /* register peephole optimisations */ clear_irp_opcodes_generic_func(); - register_peephole_optimisation(op_ia32_Const, optimize_ia32_Const); + register_peephole_optimisation(op_ia32_Const, peephole_ia32_Const); + register_peephole_optimisation(op_be_IncSP, peephole_be_IncSP); + register_peephole_optimisation(op_ia32_Lea, peephole_ia32_Lea); + register_peephole_optimisation(op_ia32_Test, peephole_ia32_Test); + register_peephole_optimisation(op_ia32_Test8Bit, peephole_ia32_Test); + register_peephole_optimisation(op_be_Return, peephole_ia32_Return); + if (! ia32_cg_config.use_imul_mem_imm32) + register_peephole_optimisation(op_ia32_IMul, peephole_ia32_Imul_split); + if (ia32_cg_config.use_pxor) + register_peephole_optimisation(op_ia32_xZero, peephole_ia32_xZero); be_peephole_opt(cg->birg); - irg_walk_graph(irg, ia32_peephole_optimize_node, NULL, NULL); } /** @@ -367,15 +949,24 @@ static INLINE void try_kill(ir_node *node) static void optimize_conv_store(ir_node *node) { ir_node *pred; + ir_node *pred_proj; ir_mode *conv_mode; ir_mode *store_mode; if(!is_ia32_Store(node) && !is_ia32_Store8Bit(node)) return; - pred = get_irn_n(node, 2); + assert(n_ia32_Store_val == n_ia32_Store8Bit_val); + pred_proj = get_irn_n(node, n_ia32_Store_val); + if(is_Proj(pred_proj)) { + pred = get_Proj_pred(pred_proj); + } else { + pred = pred_proj; + } if(!is_ia32_Conv_I2I(pred) && !is_ia32_Conv_I2I8Bit(pred)) return; + if(get_ia32_op_type(pred) != ia32_Normal) + return; /* the store only stores the lower bits, so we only need the conv * it it shrinks the mode */ @@ -384,9 +975,11 @@ static void optimize_conv_store(ir_node *node) if(get_mode_size_bits(conv_mode) < get_mode_size_bits(store_mode)) return; - set_irn_n(node, 2, get_irn_n(pred, 2)); - if(get_irn_n_edges(pred) == 0) { - be_kill_node(pred); + set_irn_n(node, n_ia32_Store_val, get_irn_n(pred, n_ia32_Conv_I2I_val)); + if(get_irn_n_edges(pred_proj) == 0) { + be_kill_node(pred_proj); + if(pred != pred_proj) + be_kill_node(pred); } } @@ -399,7 +992,8 @@ static void optimize_load_conv(ir_node *node) if (!is_ia32_Conv_I2I(node) && !is_ia32_Conv_I2I8Bit(node)) return; - pred = get_irn_n(node, 2); + assert(n_ia32_Conv_I2I_val == n_ia32_Conv_I2I8Bit_val); + pred = get_irn_n(node, n_ia32_Conv_I2I_val); if(!is_Proj(pred)) return; @@ -439,6 +1033,8 @@ static void optimize_conv_conv(ir_node *node) { ir_node *pred_proj, *pred, *result_conv; ir_mode *pred_mode, *conv_mode; + int conv_mode_bits; + int pred_mode_bits; if (!is_ia32_Conv_I2I(node) && !is_ia32_Conv_I2I8Bit(node)) return; @@ -455,11 +1051,17 @@ static void optimize_conv_conv(ir_node *node) /* we know that after a conv, the upper bits are sign extended * so we only need the 2nd conv if it shrinks the mode */ - conv_mode = get_ia32_ls_mode(node); - pred_mode = get_ia32_ls_mode(pred); - /* if 2nd conv is smaller then first conv, then we can always take the 2nd - * conv */ - if(get_mode_size_bits(conv_mode) <= get_mode_size_bits(pred_mode)) { + conv_mode = get_ia32_ls_mode(node); + conv_mode_bits = get_mode_size_bits(conv_mode); + pred_mode = get_ia32_ls_mode(pred); + pred_mode_bits = get_mode_size_bits(pred_mode); + + if(conv_mode_bits == pred_mode_bits + && get_mode_sign(conv_mode) == get_mode_sign(pred_mode)) { + result_conv = pred_proj; + } else if(conv_mode_bits <= pred_mode_bits) { + /* if 2nd conv is smaller then first conv, then we can always take the + * 2nd conv */ if(get_irn_n_edges(pred_proj) == 1) { result_conv = pred_proj; set_ia32_ls_mode(pred, conv_mode); @@ -470,8 +1072,7 @@ static void optimize_conv_conv(ir_node *node) set_ia32_in_req_all(pred, get_ia32_in_req_all(node)); } } else { - /* TODO: construct syncs/stuff here but we'll probably end up with - * 2 statements anyway */ + /* we don't want to end up with 2 loads, so we better do nothing */ if(get_irn_mode(pred) == mode_T) { return; } @@ -502,8 +1103,10 @@ static void optimize_conv_conv(ir_node *node) /* kill the conv */ exchange(node, result_conv); - if(get_irn_n_edges(pred) == 0) { - be_kill_node(pred); + if(get_irn_n_edges(pred_proj) == 0) { + be_kill_node(pred_proj); + if(pred != pred_proj) + be_kill_node(pred); } optimize_conv_conv(result_conv); }