X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fia32%2Fia32_finish.c;h=1e62ce5caa6d811a42cf7e49a553cc5fa6aac9bb;hb=b35afc79238e7731b8b3beecc69468cb719735f5;hp=d8ba800f051f03cb1ed21d78d701c763ad662269;hpb=86698cdbe5ac4bb3aea1722631fe767a3d766644;p=libfirm diff --git a/ir/be/ia32/ia32_finish.c b/ir/be/ia32/ia32_finish.c index d8ba800f0..1e62ce5ca 100644 --- a/ir/be/ia32/ia32_finish.c +++ b/ir/be/ia32/ia32_finish.c @@ -21,7 +21,6 @@ * @file * @brief This file implements functions to finalize the irg for emit. * @author Christian Wuerdig - * @version $Id$ */ #include "config.h" @@ -34,14 +33,13 @@ #include "pdeq.h" #include "error.h" -#include "../bearch_t.h" -#include "../besched_t.h" -#include "../benode_t.h" +#include "bearch.h" +#include "besched.h" +#include "benode.h" #include "bearch_ia32_t.h" #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" @@ -51,38 +49,40 @@ DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;) /** - * Transforms a Sub or xSub into Neg--Add iff OUT_REG == SRC2_REG. + * Transforms a Sub or xSub into Neg--Add iff OUT_REG != SRC1_REG && OUT_REG == SRC2_REG. * THIS FUNCTIONS MUST BE CALLED AFTER REGISTER ALLOCATION. */ -static void ia32_transform_sub_to_neg_add(ir_node *irn, ia32_code_gen_t *cg) +static void ia32_transform_sub_to_neg_add(ir_node *irn) { ir_graph *irg; ir_node *in1, *in2, *noreg, *nomem, *res; ir_node *noreg_fp, *block; - dbg_info *dbg; + dbg_info *dbgi; const arch_register_t *in1_reg, *in2_reg, *out_reg; /* fix_am will solve this for AddressMode variants */ if (get_ia32_op_type(irn) != ia32_Normal) return; - noreg = ia32_new_NoReg_gp(cg); - noreg_fp = ia32_new_NoReg_xmm(cg); - nomem = new_rd_NoMem(cg->irg); + irg = get_irn_irg(irn); + noreg = ia32_new_NoReg_gp(irg); + noreg_fp = ia32_new_NoReg_xmm(irg); + nomem = get_irg_no_mem(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(in1); in2_reg = arch_get_irn_register(in2); - out_reg = arch_irn_get_register(irn, 0); + out_reg = arch_get_irn_register_out(irn, 0); - irg = cg->irg; - block = get_nodes_block(irn); + if (out_reg == in1_reg) + return; /* in case of sub and OUT == SRC2 we can transform the sequence into neg src2 -- add */ if (out_reg != in2_reg) return; - dbg = get_irn_dbg_info(irn); + block = get_nodes_block(irn); + dbgi = get_irn_dbg_info(irn); /* generate the neg src2 */ if (is_ia32_xSub(irn)) { @@ -92,7 +92,7 @@ static void ia32_transform_sub_to_neg_add(ir_node *irn, ia32_code_gen_t *cg) assert(get_irn_mode(irn) != mode_T); - res = new_rd_ia32_xXor(dbg, irg, block, noreg, noreg, nomem, in2, noreg_fp); + res = new_bd_ia32_xXor(dbgi, 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); @@ -105,56 +105,31 @@ static void ia32_transform_sub_to_neg_add(ir_node *irn, ia32_code_gen_t *cg) sched_add_before(irn, res); /* generate the add */ - res = new_rd_ia32_xAdd(dbg, irg, block, noreg, noreg, nomem, res, in1); + res = new_bd_ia32_xAdd(dbgi, block, noreg, noreg, nomem, res, in1); set_ia32_ls_mode(res, get_ia32_ls_mode(irn)); - - /* exchange the add and the sub */ - edges_reroute(irn, res, irg); - - /* add to schedule */ - sched_add_before(irn, res); } else { - ir_node *res_proj = NULL; - ir_node *flags_proj = NULL; - const ir_edge_t *edge; + ir_node *flags_proj = NULL; + ir_node *carry; 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); + if (pn == pn_ia32_flags) { assert(flags_proj == NULL); flags_proj = proj; + break; } } } - if (flags_proj == NULL) { - res = new_rd_ia32_Neg(dbg, irg, block, in2); - arch_set_irn_register(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(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; - + if (is_ia32_Sbb(irn)) { + /* Feed borrow (in CF) as carry (via CMC) into NOT+ADC. */ + carry = get_irn_n(irn, n_ia32_Sbb_eflags); + carry = new_bd_ia32_Cmc(dbgi, block, carry); + goto carry; + } else if (flags_proj != 0) { /* * ARG, the above technique does NOT set the flags right. * So, we must produce the following code: @@ -162,39 +137,57 @@ static void ia32_transform_sub_to_neg_add(ir_node *irn, ia32_code_gen_t *cg) * t2 = a + ~b + Carry * Complement Carry * - * a + -b = a + (~b + 1) would set the carry flag IF a == b ... + * a + -b = a + (~b + 1) would set the carry flag wrong IFF both a and b are zero. */ - not = new_rd_ia32_Not(dbg, irg, block, in2); - arch_set_irn_register(not, in2_reg); - sched_add_before(irn, not); + ir_node *cmc; + ir_node *nnot; + ir_node *adc; + ir_node *adc_flags; - stc = new_rd_ia32_Stc(dbg, irg, block); - arch_set_irn_register(stc, &ia32_flags_regs[REG_EFLAGS]); - sched_add_before(irn, stc); + carry = new_bd_ia32_Stc(dbgi, block); - adc = new_rd_ia32_Adc(dbg, irg, block, noreg, noreg, nomem, not, - in1, stc); - arch_set_irn_register(adc, out_reg); - sched_add_before(irn, adc); +carry: + nnot = new_bd_ia32_Not(dbgi, block, in2); + arch_set_irn_register(nnot, in2_reg); + sched_add_before(irn, nnot); + + arch_set_irn_register(carry, &ia32_registers[REG_EFLAGS]); + sched_add_before(irn, carry); - set_irn_mode(adc, mode_T); - adc_flags = new_r_Proj(irg, block, adc, mode_Iu, pn_ia32_Adc_flags); - arch_set_irn_register(adc_flags, &ia32_flags_regs[REG_EFLAGS]); + adc = new_bd_ia32_Adc(dbgi, block, noreg, noreg, nomem, nnot, in1, carry); + arch_set_irn_register(adc, out_reg); + set_ia32_commutative(adc); - cmc = new_rd_ia32_Cmc(dbg, irg, block, adc_flags); - arch_set_irn_register(cmc, &ia32_flags_regs[REG_EFLAGS]); - sched_add_before(irn, cmc); + if (flags_proj != NULL) { + set_irn_mode(adc, mode_T); + adc_flags = new_r_Proj(adc, mode_Iu, pn_ia32_Adc_flags); + arch_set_irn_register(adc_flags, &ia32_registers[REG_EFLAGS]); - exchange(flags_proj, cmc); - if (res_proj != NULL) { - set_Proj_pred(res_proj, adc); - set_Proj_proj(res_proj, pn_ia32_Adc_res); + cmc = new_bd_ia32_Cmc(dbgi, block, adc_flags); + arch_set_irn_register(cmc, &ia32_registers[REG_EFLAGS]); + sched_add_after(irn, cmc); + exchange(flags_proj, cmc); } res = adc; + } else { + res = new_bd_ia32_Neg(dbgi, block, in2); + arch_set_irn_register(res, in2_reg); + + /* add to schedule */ + sched_add_before(irn, res); + + /* generate the add */ + res = new_bd_ia32_Add(dbgi, block, noreg, noreg, nomem, res, in1); + arch_set_irn_register(res, out_reg); + set_ia32_commutative(res); } } + /* exchange the add and the sub */ + edges_reroute(irn, res); + sched_add_before(irn, res); + set_irn_mode(res, get_irn_mode(irn)); SET_IA32_ORIG_NODE(res, irn); @@ -219,7 +212,8 @@ static inline int need_constraint_copy(ir_node *irn) case iro_ia32_Lea: case iro_ia32_Conv_I2I: case iro_ia32_Conv_I2I8Bit: - case iro_ia32_CMov: + case iro_ia32_CMovcc: + case iro_ia32_Minus64Bit: return 0; default: @@ -239,8 +233,7 @@ static int get_first_same(const arch_register_req_t* req) for (i = 0; i < 32; ++i) { if (other & (1U << i)) return i; } - assert(! "same position not found"); - return 32; + panic("same position not found"); } /** @@ -248,30 +241,21 @@ static int get_first_same(const arch_register_req_t* req) * is not fulfilled. * Transform Sub into Neg -- Add if IN2 == OUT */ -static void assure_should_be_same_requirements(ia32_code_gen_t *cg, - ir_node *node) +static void assure_should_be_same_requirements(ir_node *node) { - ir_graph *irg = cg->irg; - const arch_register_req_t **reqs; const arch_register_t *out_reg, *in_reg; int n_res, i; ir_node *in_node, *block; - reqs = get_ia32_out_req_all(node); - n_res = arch_irn_get_n_outs(node); + n_res = arch_get_irn_n_outs(node); block = get_nodes_block(node); /* check all OUT requirements, if there is a should_be_same */ for (i = 0; i < n_res; i++) { int i2, arity; int same_pos; - ir_node *perm; - ir_node *in[2]; - ir_node *perm_proj0; - ir_node *perm_proj1; ir_node *uses_out_reg; - const arch_register_req_t *req = reqs[i]; - const arch_register_class_t *cls; + const arch_register_req_t *req = arch_get_irn_register_req_out(node, i); int uses_out_reg_pos; if (!arch_register_req_is(req, should_be_same)) @@ -280,18 +264,14 @@ static void assure_should_be_same_requirements(ia32_code_gen_t *cg, same_pos = get_first_same(req); /* get in and out register */ - out_reg = arch_irn_get_register(node, i); - in_node = get_irn_n(node, same_pos); - in_reg = arch_get_irn_register(in_node); + out_reg = arch_get_irn_register_out(node, i); + in_node = get_irn_n(node, same_pos); + in_reg = arch_get_irn_register(in_node); /* requirement already fulfilled? */ if (in_reg == out_reg) continue; - /* unknowns can be changed to any register we want on emitting */ - if (is_unknown_reg(in_reg)) - continue; - cls = arch_register_get_class(in_reg); - assert(cls == arch_register_get_class(out_reg)); + assert(in_reg->reg_class == out_reg->reg_class); /* check if any other input operands uses the out register */ arity = get_irn_arity(node); @@ -299,14 +279,14 @@ static void assure_should_be_same_requirements(ia32_code_gen_t *cg, uses_out_reg_pos = -1; for (i2 = 0; i2 < arity; ++i2) { ir_node *in = get_irn_n(node, i2); - const arch_register_t *in_reg; + const arch_register_t *other_in_reg; if (!mode_is_data(get_irn_mode(in))) continue; - in_reg = arch_get_irn_register(in); + other_in_reg = arch_get_irn_register(in); - if (in_reg != out_reg) + if (other_in_reg != out_reg) continue; if (uses_out_reg != NULL && in != uses_out_reg) { @@ -323,7 +303,7 @@ 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(cls, irg, block, in_node); + ir_node *copy = be_new_Copy(block, in_node); DBG_OPT_2ADDRCPY(copy); /* destination is the out register */ @@ -350,38 +330,7 @@ static void assure_should_be_same_requirements(ia32_code_gen_t *cg, continue; } -#ifdef DEBUG_libfirm - ir_fprintf(stderr, "Note: need perm to resolve should_be_same constraint at %+F (this is unsafe and should not happen in theory...)\n", node); -#endif - /* the out reg is used as node input: we need to permutate our input - * and the other (this is allowed, since the other node can't be live - * after! the operation as we will override the register. */ - in[0] = in_node; - in[1] = uses_out_reg; - 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); - - arch_set_irn_register(perm_proj0, out_reg); - arch_set_irn_register(perm_proj1, in_reg); - - sched_add_before(node, perm); - - DBG((dbg, LEVEL_1, - "created perm %+F for should be same argument at input %d of %+F (need permutate with %+F)\n", - perm, same_pos, node, uses_out_reg)); - - /* use the perm results */ - for (i2 = 0; i2 < arity; ++i2) { - ir_node *in = get_irn_n(node, i2); - - if (in == in_node) { - set_irn_n(node, i2, perm_proj0); - } else if (in == uses_out_reg) { - set_irn_n(node, i2, perm_proj1); - } - } + panic("Unresolved should_be_same constraint"); } } @@ -395,7 +344,6 @@ static void assure_should_be_same_requirements(ia32_code_gen_t *cg, */ static void fix_am_source(ir_node *irn) { - const arch_register_req_t **reqs; int n_res, i; /* check only ia32 nodes with source address mode */ @@ -405,22 +353,22 @@ static void fix_am_source(ir_node *irn) if (get_ia32_am_support(irn) != ia32_am_binary) return; - reqs = get_ia32_out_req_all(irn); - n_res = arch_irn_get_n_outs(irn); + n_res = arch_get_irn_n_outs(irn); for (i = 0; i < n_res; i++) { - const arch_register_t *out_reg; - int same_pos; - ir_node *same_node; - const arch_register_t *same_reg; - ir_node *load_res; + const arch_register_req_t *req = arch_get_irn_register_req_out(irn, i); + const arch_register_t *out_reg; + int same_pos; + ir_node *same_node; + const arch_register_t *same_reg; + ir_node *load_res; - if (!arch_register_req_is(reqs[i], should_be_same)) + if (!arch_register_req_is(req, should_be_same)) continue; /* get in and out register */ - out_reg = arch_irn_get_register(irn, i); - same_pos = get_first_same(reqs[i]); + out_reg = arch_get_irn_register_out(irn, i); + same_pos = get_first_same(req); same_node = get_irn_n(irn, same_pos); same_reg = arch_get_irn_register(same_node); @@ -434,7 +382,7 @@ static void fix_am_source(ir_node *irn) out_reg != arch_get_irn_register(get_irn_n(irn, n_ia32_index))) continue; - load_res = turn_back_am(irn); + load_res = ia32_turn_back_am(irn); arch_set_irn_register(load_res, out_reg); DBG((dbg, LEVEL_3, @@ -449,8 +397,8 @@ static void fix_am_source(ir_node *irn) */ static void ia32_finish_irg_walker(ir_node *block, void *env) { - ia32_code_gen_t *cg = env; ir_node *irn, *next; + (void) env; /* first: turn back AM source if necessary */ for (irn = sched_first(block); ! sched_is_end(irn); irn = next) { @@ -459,13 +407,11 @@ static void ia32_finish_irg_walker(ir_node *block, void *env) } for (irn = sched_first(block); ! sched_is_end(irn); irn = next) { - ia32_code_gen_t *cg = env; - next = sched_next(irn); /* check if there is a sub which need to be transformed */ - if (is_ia32_Sub(irn) || is_ia32_xSub(irn)) { - ia32_transform_sub_to_neg_add(irn, cg); + if (is_ia32_Sub(irn) || is_ia32_Sbb(irn) || is_ia32_xSub(irn)) { + ia32_transform_sub_to_neg_add(irn); } } @@ -476,7 +422,7 @@ static void ia32_finish_irg_walker(ir_node *block, void *env) /* 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); + assure_should_be_same_requirements(irn); } } } @@ -486,7 +432,7 @@ static void ia32_finish_irg_walker(ir_node *block, void *env) */ static void ia32_push_on_queue_walker(ir_node *block, void *env) { - waitq *wq = env; + waitq *wq = (waitq*)env; waitq_put(wq, block); } @@ -494,7 +440,7 @@ static void ia32_push_on_queue_walker(ir_node *block, void *env) /** * Add Copy nodes for not fulfilled should_be_equal constraints */ -void ia32_finish_irg(ir_graph *irg, ia32_code_gen_t *cg) +void ia32_finish_irg(ir_graph *irg) { waitq *wq = new_waitq(); @@ -502,8 +448,8 @@ void ia32_finish_irg(ir_graph *irg, ia32_code_gen_t *cg) irg_block_walk_graph(irg, NULL, ia32_push_on_queue_walker, wq); while (! waitq_empty(wq)) { - ir_node *block = waitq_get(wq); - ia32_finish_irg_walker(block, cg); + ir_node *block = (ir_node*)waitq_get(wq); + ia32_finish_irg_walker(block, NULL); } del_waitq(wq); }