X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fia32%2Fia32_finish.c;h=fab1312af5f7aad4b1ceebe1d5774b5df6ea421a;hb=3f807bf48426a29da4129ff29c44a4b4690c45f6;hp=81a63d142fc2b7ab02dd7ad5ca6051326760e84c;hpb=74dd7c3a1b43d9a789bcc9ac76e0a0215a881249;p=libfirm diff --git a/ir/be/ia32/ia32_finish.c b/ir/be/ia32/ia32_finish.c index 81a63d142..fab1312af 100644 --- a/ir/be/ia32/ia32_finish.c +++ b/ir/be/ia32/ia32_finish.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. * @@ -44,6 +44,7 @@ #include "ia32_finish.h" #include "ia32_new_nodes.h" #include "ia32_map_regs.h" +#include "ia32_common_transform.h" #include "ia32_transform.h" #include "ia32_dbg_stat.h" #include "ia32_optimize.h" @@ -59,28 +60,21 @@ static void ia32_transform_sub_to_neg_add(ir_node *irn, ia32_code_gen_t *cg) { ir_graph *irg; ir_node *in1, *in2, *noreg, *nomem, *res; ir_node *noreg_fp, *block; - ir_mode *mode = get_irn_mode(irn); - dbg_info *dbg = get_irn_dbg_info(irn); - const arch_register_t *in1_reg, *in2_reg, *out_reg, **slots; - int i, arity; + dbg_info *dbg; + const arch_register_t *in1_reg, *in2_reg, *out_reg; - /* Return if not a Sub or xSub */ - if (!is_ia32_Sub(irn) && !is_ia32_xSub(irn)) - return; /* fix_am will solve this for AddressMode variants */ - if(get_ia32_op_type(irn) != ia32_Normal) + if (get_ia32_op_type(irn) != ia32_Normal) return; - noreg = ia32_new_NoReg_gp(cg); - noreg_fp = ia32_new_NoReg_fp(cg); - nomem = new_rd_NoMem(cg->irg); - in1 = get_irn_n(irn, n_ia32_binary_left); - in2 = get_irn_n(irn, n_ia32_binary_right); - in1_reg = arch_get_irn_register(cg->arch_env, in1); - in2_reg = arch_get_irn_register(cg->arch_env, in2); - out_reg = get_ia32_out_reg(irn, 0); - - assert(get_irn_mode(irn) != mode_T); + noreg = ia32_new_NoReg_gp(cg); + noreg_fp = ia32_new_NoReg_xmm(cg); + nomem = new_rd_NoMem(cg->irg); + in1 = get_irn_n(irn, n_ia32_binary_left); + in2 = get_irn_n(irn, n_ia32_binary_right); + in1_reg = arch_get_irn_register(cg->arch_env, in1); + in2_reg = arch_get_irn_register(cg->arch_env, in2); + out_reg = get_ia32_out_reg(irn, 0); irg = cg->irg; block = get_nodes_block(irn); @@ -89,273 +83,161 @@ static void ia32_transform_sub_to_neg_add(ir_node *irn, ia32_code_gen_t *cg) { if (out_reg != in2_reg) return; + dbg = get_irn_dbg_info(irn); + /* generate the neg src2 */ - if(mode_is_float(mode)) { + if (is_ia32_xSub(irn)) { int size; ir_entity *entity; ir_mode *op_mode = get_ia32_ls_mode(irn); + assert(get_irn_mode(irn) != mode_T); + res = new_rd_ia32_xXor(dbg, irg, block, noreg, noreg, nomem, in2, noreg_fp); size = get_mode_size_bits(op_mode); entity = ia32_gen_fp_known_const(size == 32 ? ia32_SSIGN : ia32_DSIGN); set_ia32_am_sc(res, entity); set_ia32_op_type(res, ia32_AddrModeS); set_ia32_ls_mode(res, op_mode); - } else { - res = new_rd_ia32_Neg(dbg, irg, block, in2); - } - arch_set_irn_register(cg->arch_env, res, in2_reg); - /* add to schedule */ - sched_add_before(irn, res); + arch_set_irn_register(cg->arch_env, res, in2_reg); + + /* add to schedule */ + sched_add_before(irn, res); - /* generate the add */ - if (mode_is_float(mode)) { + /* generate the add */ res = new_rd_ia32_xAdd(dbg, irg, block, noreg, noreg, nomem, res, in1); - set_ia32_am_support(res, ia32_am_Source, ia32_am_binary); set_ia32_ls_mode(res, get_ia32_ls_mode(irn)); - } else { - res = new_rd_ia32_Add(dbg, irg, block, noreg, noreg, nomem, res, in1); - set_ia32_am_support(res, ia32_am_Full, ia32_am_binary); - set_ia32_commutative(res); - } - - SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(cg, irn)); - /* copy register */ - slots = get_ia32_slots(res); - slots[0] = in2_reg; - - /* exchange the add and the sub */ - edges_reroute(irn, res, irg); - /* add to schedule */ - sched_add_before(irn, res); + /* exchange the add and the sub */ + edges_reroute(irn, res, irg); - /* remove the old sub */ - sched_remove(irn); - arity = get_irn_arity(irn); - for(i = 0; i < arity; ++i) { - set_irn_n(irn, i, new_Bad()); - } - - DBG_OPT_SUB2NEGADD(irn, res); -} - -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 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. - * THIS FUNCTIONS MUST BE CALLED AFTER REGISTER ALLOCATION. - */ -static void ia32_transform_lea_to_add_or_shl(ir_node *node, ia32_code_gen_t *cg) -{ - 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; - - if(!is_ia32_Lea(node)) - 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; + /* add to schedule */ + sched_add_before(irn, res); } 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"); + ir_node *res_proj = NULL; + ir_node *flags_proj = NULL; + const ir_edge_t *edge; + + if (get_irn_mode(irn) == mode_T) { + /* collect the Proj uses */ + foreach_out_edge(irn, edge) { + ir_node *proj = get_edge_src_irn(edge); + long pn = get_Proj_proj(proj); + if(pn == pn_ia32_Sub_res) { + assert(res_proj == NULL); + res_proj = proj; + } else { + assert(pn == pn_ia32_Sub_flags); + assert(flags_proj == NULL); + flags_proj = proj; + } } -#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 + + if (flags_proj == NULL) { + res = new_rd_ia32_Neg(dbg, irg, block, in2); + arch_set_irn_register(cg->arch_env, res, in2_reg); + + /* add to schedule */ + sched_add_before(irn, res); + + /* generate the add */ + res = new_rd_ia32_Add(dbg, irg, block, noreg, noreg, nomem, res, in1); + arch_set_irn_register(cg->arch_env, res, out_reg); + set_ia32_commutative(res); + + /* exchange the add and the sub */ + edges_reroute(irn, res, irg); + + /* add to schedule */ + sched_add_before(irn, res); + } else { + ir_node *stc, *cmc, *not, *adc; + ir_node *adc_flags; + + /* + * ARG, the above technique does NOT set the flags right. + * So, we must produce the following code: + * t1 = ~b + * t2 = a + ~b + Carry + * Complement Carry + * + * a + -b = a + (~b + 1) would set the carry flag IF a == b ... + */ + not = new_rd_ia32_Not(dbg, irg, block, in2); + arch_set_irn_register(cg->arch_env, not, in2_reg); + sched_add_before(irn, not); + + stc = new_rd_ia32_Stc(dbg, irg, block); + arch_set_irn_register(cg->arch_env, stc, + &ia32_flags_regs[REG_EFLAGS]); + sched_add_before(irn, stc); + + adc = new_rd_ia32_Adc(dbg, irg, block, noreg, noreg, nomem, not, + in1, stc); + arch_set_irn_register(cg->arch_env, adc, out_reg); + sched_add_before(irn, adc); + + set_irn_mode(adc, mode_T); + adc_flags = new_r_Proj(irg, block, adc, mode_Iu, pn_ia32_Adc_flags); + arch_set_irn_register(cg->arch_env, adc_flags, + &ia32_flags_regs[REG_EFLAGS]); + + cmc = new_rd_ia32_Cmc(dbg, irg, block, adc_flags); + arch_set_irn_register(cg->arch_env, cmc, + &ia32_flags_regs[REG_EFLAGS]); + sched_add_before(irn, cmc); + + exchange(flags_proj, cmc); + if (res_proj != NULL) { + set_Proj_pred(res_proj, adc); + set_Proj_proj(res_proj, pn_ia32_Adc_res); } - } 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(cg->isa->opt & IA32_OPT_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; + res = adc; } } - 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 */ - sched_add_before(node, res); - - DBG_OPT_LEA2ADD(node, res); + SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(cg, irn)); - /* remove the old LEA */ - sched_remove(node); + /* remove the old sub */ + sched_remove(irn); + kill_node(irn); - /* exchange the Add and the LEA */ - exchange(node, res); + DBG_OPT_SUB2NEGADD(irn, res); } static INLINE int need_constraint_copy(ir_node *irn) { + /* the 3 operand form of IMul needs no constraint copy */ + if(is_ia32_IMul(irn)) { + ir_node *right = get_irn_n(irn, n_ia32_IMul_right); + if(is_ia32_Immediate(right)) + return 0; + } + return ! is_ia32_Lea(irn) && ! is_ia32_Conv_I2I(irn) && ! is_ia32_Conv_I2I8Bit(irn) && ! is_ia32_CMov(irn); } +/** + * Returns the index of the "same" register. + * On the x86, we should have only one. + */ +static int get_first_same(const arch_register_req_t* req) +{ + const unsigned other = req->other_same; + int i; + + for (i = 0; i < 32; ++i) { + if (other & (1U << i)) return i; + } + assert(! "same position not found"); + return 32; +} + /** * Insert copies for all ia32 nodes where the should_be_same requirement * is not fulfilled. @@ -370,17 +252,7 @@ static void assure_should_be_same_requirements(ia32_code_gen_t *cg, const arch_register_t *out_reg, *in_reg; int n_res, i; ir_node *in_node, *block; - ia32_op_type_t op_tp; - - if(!is_ia32_irn(node)) - return; - - /* some nodes are just a bit less efficient, but need no fixing if the - * should be same requirement is not fulfilled */ - if(!need_constraint_copy(node)) - return; - op_tp = get_ia32_op_type(node); reqs = get_ia32_out_req_all(node); n_res = get_ia32_n_res(node); block = get_nodes_block(node); @@ -395,13 +267,13 @@ static void assure_should_be_same_requirements(ia32_code_gen_t *cg, ir_node *perm_proj1; ir_node *uses_out_reg; const arch_register_req_t *req = reqs[i]; - const arch_register_class_t *class; + const arch_register_class_t *cls; int uses_out_reg_pos; if (!arch_register_req_is(req, should_be_same)) continue; - same_pos = req->other_same[0]; + same_pos = get_first_same(req); /* get in and out register */ out_reg = get_ia32_out_reg(node, i); @@ -414,8 +286,8 @@ static void assure_should_be_same_requirements(ia32_code_gen_t *cg, /* unknowns can be changed to any register we want on emitting */ if (is_unknown_reg(in_reg)) continue; - class = arch_register_get_class(in_reg); - assert(class == arch_register_get_class(out_reg)); + cls = arch_register_get_class(in_reg); + assert(cls == arch_register_get_class(out_reg)); /* check if any other input operands uses the out register */ arity = get_irn_arity(node); @@ -447,7 +319,7 @@ static void assure_should_be_same_requirements(ia32_code_gen_t *cg, * (the register can't be live since the operation will override it * anyway) */ if(uses_out_reg == NULL) { - ir_node *copy = be_new_Copy(class, irg, block, in_node); + ir_node *copy = be_new_Copy(cls, irg, block, in_node); DBG_OPT_2ADDRCPY(copy); /* destination is the out register */ @@ -465,7 +337,7 @@ static void assure_should_be_same_requirements(ia32_code_gen_t *cg, } /* for commutative nodes we can simply swap the left/right */ - if(is_ia32_commutative(node) && uses_out_reg_pos == n_ia32_binary_right) { + if (uses_out_reg_pos == n_ia32_binary_right && is_ia32_commutative(node)) { ia32_swap_left_right(node); DBG((dbg, LEVEL_1, "swapped left/right input of %+F to resolve " "should be same constraint\n", node)); @@ -480,7 +352,7 @@ static void assure_should_be_same_requirements(ia32_code_gen_t *cg, * after! the operation as we will override the register. */ in[0] = in_node; in[1] = uses_out_reg; - perm = be_new_Perm(class, irg, block, 2, in); + perm = be_new_Perm(cls, irg, block, 2, in); perm_proj0 = new_r_Proj(irg, block, perm, get_irn_mode(in[0]), 0); perm_proj1 = new_r_Proj(irg, block, perm, get_irn_mode(in[1]), 1); @@ -533,8 +405,8 @@ static void fix_am_source(ir_node *irn, void *env) { if (get_ia32_am_arity(irn) != ia32_am_binary) return; - base = get_irn_n(irn, 0); - index = get_irn_n(irn, 1); + base = get_irn_n(irn, n_ia32_base); + index = get_irn_n(irn, n_ia32_index); reg_base = arch_get_irn_register(arch_env, base); reg_index = arch_get_irn_register(arch_env, index); @@ -548,7 +420,7 @@ static void fix_am_source(ir_node *irn, void *env) { if (arch_register_req_is(reqs[i], should_be_same)) { /* get in and out register */ const arch_register_t *out_reg = get_ia32_out_reg(irn, i); - int same_pos = reqs[i]->other_same[0]; + int same_pos = get_first_same(reqs[i]); ir_node *same_node = get_irn_n(irn, same_pos); const arch_register_t *same_reg = arch_get_irn_register(arch_env, same_node); @@ -561,6 +433,7 @@ static void fix_am_source(ir_node *irn, void *env) { ir_node *load_res; ir_node *mem; int pnres; + int pnmem; /* should_be same constraint is fullfilled, nothing to do */ if(out_reg == same_reg) @@ -578,11 +451,13 @@ static void fix_am_source(ir_node *irn, void *env) { if (same_cls == &ia32_reg_classes[CLASS_ia32_gp]) { load = new_rd_ia32_Load(dbgi, irg, block, base, index, mem); pnres = pn_ia32_Load_res; + pnmem = pn_ia32_Load_M; proj_mode = mode_Iu; } else if (same_cls == &ia32_reg_classes[CLASS_ia32_xmm]) { load = new_rd_ia32_xLoad(dbgi, irg, block, base, index, mem, get_ia32_ls_mode(irn)); pnres = pn_ia32_xLoad_res; + pnmem = pn_ia32_xLoad_M; proj_mode = mode_E; } else { panic("cannot turn back address mode for this register class"); @@ -607,11 +482,11 @@ static void fix_am_source(ir_node *irn, void *env) { foreach_out_edge_safe(irn, edge, next) { ir_node *node = get_edge_src_irn(edge); int pn = get_Proj_proj(node); - if(pn == 0) { + if (pn == pn_ia32_res) { exchange(node, irn); - } else { - assert(pn == 1); + } else if (pn == pn_ia32_mem) { set_Proj_pred(node, load); + set_Proj_proj(node, pnmem); } } set_irn_mode(irn, mode_Iu); @@ -645,16 +520,20 @@ static void ia32_finish_irg_walker(ir_node *block, void *env) { next = sched_next(irn); /* check if there is a sub which need to be transformed */ - ia32_transform_sub_to_neg_add(irn, cg); - - /* transform a LEA into an Add if possible */ - ia32_transform_lea_to_add_or_shl(irn, cg); + if (is_ia32_Sub(irn) || is_ia32_xSub(irn)) { + ia32_transform_sub_to_neg_add(irn, cg); + } } /* second: insert copies and finish irg */ for (irn = sched_first(block); ! sched_is_end(irn); irn = next) { next = sched_next(irn); - assure_should_be_same_requirements(cg, irn); + if (is_ia32_irn(irn)) { + /* some nodes are just a bit less efficient, but need no fixing if the + * should be same requirement is not fulfilled */ + if (need_constraint_copy(irn)) + assure_should_be_same_requirements(cg, irn); + } } }