X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fia32%2Fia32_transform.c;h=7817ef1749f1fc79b410c741cf426c95afa56f56;hb=b69fead7cd3ec7bf5e70c68b2b28afbad2192885;hp=368a262fe0c028c48e979f33c610c48aa7cf0b43;hpb=80a705d46ee27a26fa20d179f07e6c3c59694fe8;p=libfirm diff --git a/ir/be/ia32/ia32_transform.c b/ir/be/ia32/ia32_transform.c index 368a262fe..7817ef174 100644 --- a/ir/be/ia32/ia32_transform.c +++ b/ir/be/ia32/ia32_transform.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. * @@ -55,6 +55,7 @@ #include "../beutil.h" #include "../beirg_t.h" #include "../betranshlp.h" +#include "../be_t.h" #include "bearch_ia32_t.h" #include "ia32_nodes_attr.h" @@ -65,6 +66,7 @@ #include "ia32_optimize.h" #include "ia32_util.h" #include "ia32_address_mode.h" +#include "ia32_architecture.h" #include "gen_ia32_regalloc_if.h" @@ -95,7 +97,6 @@ DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;) static ia32_code_gen_t *env_cg = NULL; static ir_node *initial_fpcw = NULL; static heights_t *heights = NULL; -static transform_config_t transform_config; extern ir_op *get_op_Mulh(void); @@ -124,16 +125,6 @@ typedef ir_node *construct_binop_float_func(dbg_info *db, ir_graph *irg, typedef ir_node *construct_unop_func(dbg_info *db, ir_graph *irg, ir_node *block, ir_node *op); -/**************************************************************************************************** - * _ _ __ _ _ - * | | | | / _| | | (_) - * _ __ ___ __| | ___ | |_ _ __ __ _ _ __ ___| |_ ___ _ __ _ __ ___ __ _| |_ _ ___ _ __ - * | '_ \ / _ \ / _` |/ _ \ | __| '__/ _` | '_ \/ __| _/ _ \| '__| '_ ` _ \ / _` | __| |/ _ \| '_ \ - * | | | | (_) | (_| | __/ | |_| | | (_| | | | \__ \ || (_) | | | | | | | | (_| | |_| | (_) | | | | - * |_| |_|\___/ \__,_|\___| \__|_| \__,_|_| |_|___/_| \___/|_| |_| |_| |_|\__,_|\__|_|\___/|_| |_| - * - ****************************************************************************************************/ - static ir_node *try_create_Immediate(ir_node *node, char immediate_constraint_type); @@ -191,20 +182,63 @@ static ir_type *get_prim_type(pmap *types, ir_mode *mode) } /** - * Get an atomic entity that is initialized with a tarval + * Creates an immediate. + * + * @param symconst if set, create a SymConst immediate + * @param symconst_sign sign for the symconst + * @param val integer value for the immediate + */ +static ir_node *create_Immediate(ir_entity *symconst, int symconst_sign, long 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, + symconst, symconst_sign, val); + arch_set_irn_register(env_cg->arch_env, immediate, &ia32_gp_regs[REG_GP_NOREG]); + + return immediate; +} + +/** + * Get an atomic entity that is initialized with a tarval forming + * a given constant. + * + * @param cnst the node representing the constant */ static ir_entity *create_float_const_entity(ir_node *cnst) { ia32_isa_t *isa = env_cg->isa; - tarval *tv = get_Const_tarval(cnst); - pmap_entry *e = pmap_find(isa->tv_ent, tv); + tarval *key = get_Const_tarval(cnst); + pmap_entry *e = pmap_find(isa->tv_ent, key); ir_entity *res; ir_graph *rem; - if (! e) { - ir_mode *mode = get_irn_mode(cnst); - ir_type *tp = get_Const_type(cnst); - if (tp == firm_unknown_type) + if (e == NULL) { + tarval *tv = key; + ir_mode *mode = get_tarval_mode(tv); + ir_type *tp; + + if (! ia32_cg_config.use_sse2) { + /* try to reduce the mode to produce smaller sized entities */ + if (mode != mode_F) { + if (tarval_ieee754_can_conv_lossless(tv, mode_F)) { + mode = mode_F; + tv = tarval_convert_to(tv, mode); + } else if (mode != mode_D) { + if (tarval_ieee754_can_conv_lossless(tv, mode_D)) { + mode = mode_D; + tv = tarval_convert_to(tv, mode); + } + } + } + } + + if (mode == get_irn_mode(cnst)) { + /* mode was not changed */ + tp = get_Const_type(cnst); + if (tp == firm_unknown_type) + tp = get_prim_type(isa->types, mode); + } else tp = get_prim_type(isa->types, mode); res = new_entity(get_glob_type(), unique_id(".LC%u"), tp); @@ -221,7 +255,7 @@ static ir_entity *create_float_const_entity(ir_node *cnst) set_atomic_ent_value(res, new_Const_type(tv, tp)); current_ir_graph = rem; - pmap_insert(isa->tv_ent, tv, res); + pmap_insert(isa->tv_ent, key, res); } else { res = e->value; } @@ -247,10 +281,37 @@ static int is_Const_Minus_1(ir_node *node) { static int is_simple_x87_Const(ir_node *node) { tarval *tv = get_Const_tarval(node); + if (tarval_is_null(tv) || tarval_is_one(tv)) + return 1; + + /* TODO: match all the other float constants */ + return 0; +} + +/** + * returns true if constant can be created with a simple float command + */ +static int is_simple_sse_Const(ir_node *node) +{ + tarval *tv = get_Const_tarval(node); + ir_mode *mode = get_tarval_mode(tv); + + if (mode == mode_F) + return 1; - if(tarval_is_null(tv) || tarval_is_one(tv)) + if (tarval_is_null(tv) || tarval_is_one(tv)) return 1; + if (mode == mode_D) { + unsigned val = get_tarval_sub_bits(tv, 0) | + (get_tarval_sub_bits(tv, 1) << 8) | + (get_tarval_sub_bits(tv, 2) << 16) | + (get_tarval_sub_bits(tv, 3) << 24); + if (val == 0) + /* lower 32bit are zero, really a 32bit constant */ + return 1; + } + /* TODO: match all the other float constants */ return 0; } @@ -265,6 +326,8 @@ static ir_node *gen_Const(ir_node *node) { dbg_info *dbgi = get_irn_dbg_info(node); ir_mode *mode = get_irn_mode(node); + assert(is_Const(node)); + if (mode_is_float(mode)) { ir_node *res = NULL; ir_node *noreg = ia32_new_NoReg_gp(env_cg); @@ -272,16 +335,63 @@ static ir_node *gen_Const(ir_node *node) { ir_node *load; ir_entity *floatent; - if (USE_SSE2(env_cg)) { - if (is_Const_null(node)) { + if (ia32_cg_config.use_sse2) { + tarval *tv = get_Const_tarval(node); + if (tarval_is_null(tv)) { load = new_rd_ia32_xZero(dbgi, irg, block); set_ia32_ls_mode(load, mode); res = load; + } else if (tarval_is_one(tv)) { + int cnst = mode == mode_F ? 26 : 55; + ir_node *imm1 = create_Immediate(NULL, 0, cnst); + ir_node *imm2 = create_Immediate(NULL, 0, 2); + ir_node *pslld, *psrld; + + load = new_rd_ia32_xAllOnes(dbgi, irg, block); + set_ia32_ls_mode(load, mode); + pslld = new_rd_ia32_xPslld(dbgi, irg, block, load, imm1); + set_ia32_ls_mode(pslld, mode); + psrld = new_rd_ia32_xPsrld(dbgi, irg, block, pslld, imm2); + set_ia32_ls_mode(psrld, mode); + res = psrld; + } else if (mode == mode_F) { + /* we can place any 32bit constant by using a movd gp, sse */ + unsigned val = get_tarval_sub_bits(tv, 0) | + (get_tarval_sub_bits(tv, 1) << 8) | + (get_tarval_sub_bits(tv, 2) << 16) | + (get_tarval_sub_bits(tv, 3) << 24); + ir_node *cnst = new_rd_ia32_Const(dbgi, irg, block, NULL, 0, val); + load = new_rd_ia32_xMovd(dbgi, irg, block, cnst); + set_ia32_ls_mode(load, mode); + res = load; } else { + if (mode == mode_D) { + unsigned val = get_tarval_sub_bits(tv, 0) | + (get_tarval_sub_bits(tv, 1) << 8) | + (get_tarval_sub_bits(tv, 2) << 16) | + (get_tarval_sub_bits(tv, 3) << 24); + if (val == 0) { + ir_node *imm32 = create_Immediate(NULL, 0, 32); + ir_node *cnst, *psllq; + + /* fine, lower 32bit are zero, produce 32bit value */ + val = get_tarval_sub_bits(tv, 4) | + (get_tarval_sub_bits(tv, 5) << 8) | + (get_tarval_sub_bits(tv, 6) << 16) | + (get_tarval_sub_bits(tv, 7) << 24); + cnst = new_rd_ia32_Const(dbgi, irg, block, NULL, 0, val); + load = new_rd_ia32_xMovd(dbgi, irg, block, cnst); + set_ia32_ls_mode(load, mode); + psllq = new_rd_ia32_xPsllq(dbgi, irg, block, load, imm32); + set_ia32_ls_mode(psllq, mode); + res = psllq; + goto end; + } + } floatent = create_float_const_entity(node); load = new_rd_ia32_xLoad(dbgi, irg, block, noreg, noreg, nomem, - mode); + mode); set_ia32_op_type(load, ia32_AddrModeS); set_ia32_am_sc(load, floatent); set_ia32_flags(load, get_ia32_flags(load) | arch_irn_flags_rematerializable); @@ -291,9 +401,11 @@ static ir_node *gen_Const(ir_node *node) { if (is_Const_null(node)) { load = new_rd_ia32_vfldz(dbgi, irg, block); res = load; + set_ia32_ls_mode(load, mode); } else if (is_Const_one(node)) { load = new_rd_ia32_vfld1(dbgi, irg, block); res = load; + set_ia32_ls_mode(load, mode); } else { floatent = create_float_const_entity(node); @@ -302,12 +414,11 @@ static ir_node *gen_Const(ir_node *node) { set_ia32_am_sc(load, floatent); set_ia32_flags(load, get_ia32_flags(load) | arch_irn_flags_rematerializable); res = new_r_Proj(irg, block, load, mode_vfp, pn_ia32_vfld_res); + /* take the mode from the entity */ + set_ia32_ls_mode(load, get_type_mode(get_entity_type(floatent))); } - set_ia32_ls_mode(load, mode); } - - SET_IA32_ORIG_NODE(load, ia32_get_old_node_name(env_cg, node)); - +end: /* Const Nodes before the initial IncSP are a bad idea, because * they could be spilled and we have no SP ready at that point yet. * So add a dependency to the initial frame pointer calculation to @@ -319,15 +430,15 @@ static ir_node *gen_Const(ir_node *node) { SET_IA32_ORIG_NODE(load, ia32_get_old_node_name(env_cg, node)); return res; - } else { + } else { /* non-float mode */ ir_node *cnst; tarval *tv = get_Const_tarval(node); long val; tv = tarval_convert_to(tv, mode_Iu); - if(tv == get_tarval_bad() || tv == get_tarval_undefined() - || tv == NULL) { + if (tv == get_tarval_bad() || tv == get_tarval_undefined() || + tv == NULL) { panic("couldn't convert constant tarval (%+F)", node); } val = get_tarval_long(tv); @@ -359,7 +470,7 @@ static ir_node *gen_SymConst(ir_node *node) { ir_node *noreg = ia32_new_NoReg_gp(env_cg); ir_node *nomem = new_NoMem(); - if (USE_SSE2(env_cg)) + if (ia32_cg_config.use_sse2) cnst = new_rd_ia32_xLoad(dbgi, irg, block, noreg, noreg, nomem, mode_E); else cnst = new_rd_ia32_vfld(dbgi, irg, block, noreg, noreg, nomem, mode_E); @@ -463,36 +574,58 @@ const char *ia32_get_old_node_name(ia32_code_gen_t *cg, ir_node *irn) { } #endif /* NDEBUG */ -int ia32_use_source_address_mode(ir_node *block, ir_node *node, ir_node *other) +/** + * return true if the node is a Proj(Load) and could be used in source address + * mode for another node. Will return only true if the @p other node is not + * dependent on the memory of the Load (for binary operations use the other + * input here, for unary operations use NULL). + */ +static int ia32_use_source_address_mode(ir_node *block, ir_node *node, + ir_node *other, ir_node *other2, match_flags_t flags) { - ir_mode *mode = get_irn_mode(node); ir_node *load; long pn; /* float constants are always available */ - if(is_Const(node) && mode_is_float(mode) - && !is_simple_x87_Const(node) && get_irn_n_edges(node) == 1) { - return 1; + if (is_Const(node)) { + ir_mode *mode = get_irn_mode(node); + if (mode_is_float(mode)) { + if (ia32_cg_config.use_sse2) { + if (is_simple_sse_Const(node)) + return 0; + } else { + if (is_simple_x87_Const(node)) + return 0; + } + if (get_irn_n_edges(node) > 1) + return 0; + return 1; + } } - if(!is_Proj(node)) + if (!is_Proj(node)) return 0; load = get_Proj_pred(node); pn = get_Proj_proj(node); - if(!is_Load(load) || pn != pn_Load_res) + if (!is_Load(load) || pn != pn_Load_res) return 0; - if(get_nodes_block(load) != block) + if (get_nodes_block(load) != block) return 0; /* we only use address mode if we're the only user of the load */ - if(get_irn_n_edges(node) > 1) + if (get_irn_n_edges(node) != (flags & match_two_users ? 2 : 1)) return 0; - - if(other != NULL && get_Load_mode(load) != get_irn_mode(other)) + /* in some edge cases with address mode we might reach the load normally + * and through some AM sequence, if it is already materialized then we + * can't create an AM node from it */ + if (be_is_transformed(node)) return 0; /* don't do AM if other node inputs depend on the load (via mem-proj) */ - if(other != NULL && get_nodes_block(other) == block - && heights_reachable_in_block(heights, other, load)) + if (other != NULL && get_nodes_block(other) == block && + heights_reachable_in_block(heights, other, load)) + return 0; + if (other2 != NULL && get_nodes_block(other2) == block && + heights_reachable_in_block(heights, other2, load)) return 0; return 1; @@ -507,10 +640,24 @@ struct ia32_address_mode_t { ir_node *new_op1; ir_node *new_op2; op_pin_state pinned; - unsigned commutative:1; - unsigned ins_permuted:1; + unsigned commutative : 1; + unsigned ins_permuted : 1; }; +static void build_address_ptr(ia32_address_t *addr, ir_node *ptr, ir_node *mem) +{ + ir_node *noreg_gp; + + /* construct load address */ + memset(addr, 0, sizeof(addr[0])); + ia32_create_address_mode(addr, ptr, /*force=*/0); + + noreg_gp = ia32_new_NoReg_gp(env_cg); + addr->base = addr->base ? be_transform_node(addr->base) : noreg_gp; + addr->index = addr->index ? be_transform_node(addr->index) : noreg_gp; + addr->mem = be_transform_node(mem); +} + static void build_address(ia32_address_mode_t *am, ir_node *node) { ir_node *noreg_gp = ia32_new_NoReg_gp(env_cg); @@ -519,17 +666,15 @@ static void build_address(ia32_address_mode_t *am, ir_node *node) ir_node *ptr; ir_node *mem; ir_node *new_mem; - ir_node *base; - ir_node *index; - if(is_Const(node)) { + if (is_Const(node)) { ir_entity *entity = create_float_const_entity(node); addr->base = noreg_gp; addr->index = noreg_gp; addr->mem = new_NoMem(); addr->symconst_ent = entity; addr->use_frame = 1; - am->ls_mode = get_irn_mode(node); + am->ls_mode = get_type_mode(get_entity_type(entity)); am->pinned = op_pin_state_floats; return; } @@ -544,27 +689,13 @@ static void build_address(ia32_address_mode_t *am, ir_node *node) /* construct load address */ ia32_create_address_mode(addr, ptr, /*force=*/0); - base = addr->base; - index = addr->index; - - if(base == NULL) { - base = noreg_gp; - } else { - base = be_transform_node(base); - } - if(index == NULL) { - index = noreg_gp; - } else { - index = be_transform_node(index); - } - - addr->base = base; - addr->index = index; + addr->base = addr->base ? be_transform_node(addr->base) : noreg_gp; + addr->index = addr->index ? be_transform_node(addr->index) : noreg_gp; addr->mem = new_mem; } -static void set_address(ir_node *node, ia32_address_t *addr) +static void set_address(ir_node *node, const ia32_address_t *addr) { set_ia32_am_scale(node, addr->scale); set_ia32_am_sc(node, addr->symconst_ent); @@ -576,16 +707,19 @@ static void set_address(ir_node *node, ia32_address_t *addr) set_ia32_frame_ent(node, addr->frame_entity); } -static void set_am_attributes(ir_node *node, ia32_address_mode_t *am) +/** + * Apply attributes of a given address mode to a node. + */ +static void set_am_attributes(ir_node *node, const ia32_address_mode_t *am) { set_address(node, &am->addr); set_ia32_op_type(node, am->op_type); set_ia32_ls_mode(node, am->ls_mode); - if(am->pinned == op_pin_state_pinned && get_irn_pinned(node) != op_pin_state_pinned) { + if (am->pinned == op_pin_state_pinned) { set_irn_pinned(node, am->pinned); } - if(am->commutative) + if (am->commutative) set_ia32_commutative(node); } @@ -626,76 +760,115 @@ ir_node *ia32_skip_downconv(ir_node *node) { return node; } +#if 0 +static ir_node *create_upconv(ir_node *node, ir_node *orig_node) +{ + ir_mode *mode = get_irn_mode(node); + ir_node *block; + ir_mode *tgt_mode; + dbg_info *dbgi; + + if(mode_is_signed(mode)) { + tgt_mode = mode_Is; + } else { + tgt_mode = mode_Iu; + } + block = get_nodes_block(node); + dbgi = get_irn_dbg_info(node); -typedef enum { - match_commutative = 1 << 0, - match_am_and_immediates = 1 << 1, - match_no_am = 1 << 2, - match_8_bit_am = 1 << 3, - match_16_bit_am = 1 << 4, - match_no_immediate = 1 << 5, - match_force_32bit_op = 1 << 6, - match_skip_input_conv = 1 << 7 -} match_flags_t; + return create_I2I_Conv(mode, tgt_mode, dbgi, block, node, orig_node); +} +#endif +/** + * matches operands of a node into ia32 addressing/operand modes. This covers + * usage of source address mode, immediates, operations with non 32-bit modes, + * ... + * The resulting data is filled into the @p am struct. block is the block + * of the node whose arguments are matched. op1, op2 are the first and second + * input that are matched (op1 may be NULL). other_op is another unrelated + * input that is not matched! but which is needed sometimes to check if AM + * for op1/op2 is legal. + * @p flags describes the supported modes of the operation in detail. + */ static void match_arguments(ia32_address_mode_t *am, ir_node *block, - ir_node *op1, ir_node *op2, match_flags_t flags) + ir_node *op1, ir_node *op2, ir_node *other_op, + match_flags_t flags) { - ia32_address_t *addr = &am->addr; - ir_node *noreg_gp = ia32_new_NoReg_gp(env_cg); - ir_node *new_op1; - ir_node *new_op2; - ir_mode *mode = get_irn_mode(op2); + ia32_address_t *addr = &am->addr; + ir_mode *mode = get_irn_mode(op2); + int mode_bits = get_mode_size_bits(mode); + ir_node *noreg_gp, *new_op1, *new_op2; int use_am; unsigned commutative; int use_am_and_immediates; int use_immediate; - int skip_input_conv; - int mode_bits = get_mode_size_bits(mode); memset(am, 0, sizeof(am[0])); commutative = (flags & match_commutative) != 0; use_am_and_immediates = (flags & match_am_and_immediates) != 0; - use_am = ! (flags & match_no_am); - use_immediate = !(flags & match_no_immediate); - skip_input_conv = (flags & match_skip_input_conv) != 0; + use_am = (flags & match_am) != 0; + use_immediate = (flags & match_immediate) != 0; + assert(!use_am_and_immediates || use_immediate); assert(op2 != NULL); assert(!commutative || op1 != NULL); + assert(use_am || !(flags & match_8bit_am)); + assert(use_am || !(flags & match_16bit_am)); + + if (mode_bits == 8) { + if (!(flags & match_8bit_am)) + use_am = 0; + /* we don't automatically add upconvs yet */ + assert((flags & match_mode_neutral) || (flags & match_8bit)); + } else if (mode_bits == 16) { + if (!(flags & match_16bit_am)) + use_am = 0; + /* we don't automatically add upconvs yet */ + assert((flags & match_mode_neutral) || (flags & match_16bit)); + } - if(mode_bits == 8 && !(flags & match_8_bit_am)) { - use_am = 0; - } else if(mode_bits == 16 && !(flags & match_16_bit_am)) { - use_am = 0; + /* we can simply skip downconvs for mode neutral nodes: the upper bits + * can be random for these operations */ + if (flags & match_mode_neutral) { + op2 = ia32_skip_downconv(op2); + if (op1 != NULL) { + op1 = ia32_skip_downconv(op1); + } } - op2 = ia32_skip_downconv(op2); - if(op1 != NULL) - op1 = ia32_skip_downconv(op1); + /* match immediates. firm nodes are normalized: constants are always on the + * op2 input */ + new_op2 = NULL; + if (!(flags & match_try_am) && use_immediate) { + new_op2 = try_create_Immediate(op2, 0); + } - new_op2 = (use_immediate ? try_create_Immediate(op2, 0) : NULL); - if(new_op2 == NULL && use_am && ia32_use_source_address_mode(block, op2, op1)) { + noreg_gp = ia32_new_NoReg_gp(env_cg); + if (new_op2 == NULL && + use_am && ia32_use_source_address_mode(block, op2, op1, other_op, flags)) { build_address(am, op2); new_op1 = (op1 == NULL ? NULL : be_transform_node(op1)); - if(mode_is_float(mode)) { + if (mode_is_float(mode)) { new_op2 = ia32_new_NoReg_vfp(env_cg); } else { new_op2 = noreg_gp; } am->op_type = ia32_AddrModeS; - } else if(commutative && (new_op2 == NULL || use_am_and_immediates) && - use_am && ia32_use_source_address_mode(block, op1, op2)) { + } else if (commutative && (new_op2 == NULL || use_am_and_immediates) && + use_am && + ia32_use_source_address_mode(block, op1, op2, other_op, flags)) { ir_node *noreg; build_address(am, op1); - if(mode_is_float(mode)) { + if (mode_is_float(mode)) { noreg = ia32_new_NoReg_vfp(env_cg); } else { noreg = noreg_gp; } - if(new_op2 != NULL) { + if (new_op2 != NULL) { new_op1 = noreg; } else { new_op1 = be_transform_node(op2); @@ -704,21 +877,26 @@ static void match_arguments(ia32_address_mode_t *am, ir_node *block, } am->op_type = ia32_AddrModeS; } else { + if (flags & match_try_am) { + am->new_op1 = NULL; + am->new_op2 = NULL; + am->op_type = ia32_Normal; + return; + } + new_op1 = (op1 == NULL ? NULL : be_transform_node(op1)); - if(new_op2 == NULL) + if (new_op2 == NULL) new_op2 = be_transform_node(op2); am->op_type = ia32_Normal; - if(flags & match_force_32bit_op) { + am->ls_mode = get_irn_mode(op2); + if (flags & match_mode_neutral) am->ls_mode = mode_Iu; - } else { - am->ls_mode = get_irn_mode(op2); - } } - if(addr->base == NULL) + if (addr->base == NULL) addr->base = noreg_gp; - if(addr->index == NULL) + if (addr->index == NULL) addr->index = noreg_gp; - if(addr->mem == NULL) + if (addr->mem == NULL) addr->mem = new_NoMem(); am->new_op1 = new_op1; @@ -728,11 +906,10 @@ static void match_arguments(ia32_address_mode_t *am, ir_node *block, static ir_node *fix_mem_proj(ir_node *node, ia32_address_mode_t *am) { - ir_graph *irg = current_ir_graph; ir_mode *mode; ir_node *load; - if(am->mem_proj == NULL) + if (am->mem_proj == NULL) return node; /* we have to create a mode_T so the old MemProj can attach to us */ @@ -742,9 +919,9 @@ static ir_node *fix_mem_proj(ir_node *node, ia32_address_mode_t *am) mark_irn_visited(load); be_set_transformed_node(load, node); - if(mode != mode_T) { + if (mode != mode_T) { set_irn_mode(node, mode_T); - return new_rd_Proj(NULL, irg, get_nodes_block(node), node, mode, pn_ia32_res); + return new_rd_Proj(NULL, current_ir_graph, get_nodes_block(node), node, mode, pn_ia32_res); } else { return node; } @@ -753,6 +930,7 @@ static ir_node *fix_mem_proj(ir_node *node, ia32_address_mode_t *am) /** * Construct a standard binary operation, set AM and immediate if required. * + * @param node The original node for which the binop is created * @param op1 The first operand * @param op2 The second operand * @param func The node constructor function @@ -761,23 +939,22 @@ static ir_node *fix_mem_proj(ir_node *node, ia32_address_mode_t *am) static ir_node *gen_binop(ir_node *node, ir_node *op1, ir_node *op2, construct_binop_func *func, match_flags_t flags) { - ir_node *block = get_nodes_block(node); - ir_node *new_block = be_transform_node(block); - ir_graph *irg = current_ir_graph; - dbg_info *dbgi = get_irn_dbg_info(node); - ir_node *new_node; + dbg_info *dbgi; + ir_node *block, *new_block, *new_node; ia32_address_mode_t am; ia32_address_t *addr = &am.addr; - flags |= match_force_32bit_op; - - match_arguments(&am, block, op1, op2, flags); + block = get_nodes_block(node); + match_arguments(&am, block, op1, op2, NULL, flags); - new_node = func(dbgi, irg, new_block, addr->base, addr->index, addr->mem, - am.new_op1, am.new_op2); + dbgi = get_irn_dbg_info(node); + new_block = be_transform_node(block); + new_node = func(dbgi, current_ir_graph, new_block, + addr->base, addr->index, addr->mem, + am.new_op1, am.new_op2); set_am_attributes(new_node, &am); /* we can't use source address mode anymore when using immediates */ - if(is_ia32_Immediate(am.new_op1) || is_ia32_Immediate(am.new_op2)) + if (is_ia32_Immediate(am.new_op1) || is_ia32_Immediate(am.new_op2)) set_ia32_am_support(new_node, ia32_am_None, ia32_am_arity_none); SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node)); @@ -791,12 +968,12 @@ enum { n_ia32_l_binop_right, n_ia32_l_binop_eflags }; -COMPILETIME_ASSERT(n_ia32_l_binop_left == n_ia32_l_Adc_left, n_Adc_left) -COMPILETIME_ASSERT(n_ia32_l_binop_right == n_ia32_l_Adc_right, n_Adc_right) -COMPILETIME_ASSERT(n_ia32_l_binop_eflags == n_ia32_l_Adc_eflags, n_Adc_eflags) -COMPILETIME_ASSERT(n_ia32_l_binop_left == n_ia32_l_Sbb_left, n_Sbb_left) -COMPILETIME_ASSERT(n_ia32_l_binop_right == n_ia32_l_Sbb_right, n_Sbb_right) -COMPILETIME_ASSERT(n_ia32_l_binop_eflags == n_ia32_l_Sbb_eflags, n_Sbb_eflags) +COMPILETIME_ASSERT(n_ia32_l_binop_left == n_ia32_l_Adc_left, n_Adc_left) +COMPILETIME_ASSERT(n_ia32_l_binop_right == n_ia32_l_Adc_right, n_Adc_right) +COMPILETIME_ASSERT(n_ia32_l_binop_eflags == n_ia32_l_Adc_eflags, n_Adc_eflags) +COMPILETIME_ASSERT(n_ia32_l_binop_left == n_ia32_l_Sbb_minuend, n_Sbb_minuend) +COMPILETIME_ASSERT(n_ia32_l_binop_right == n_ia32_l_Sbb_subtrahend, n_Sbb_subtrahend) +COMPILETIME_ASSERT(n_ia32_l_binop_eflags == n_ia32_l_Sbb_eflags, n_Sbb_eflags) /** * Construct a binary operation which also consumes the eflags. @@ -810,21 +987,21 @@ static ir_node *gen_binop_flags(ir_node *node, construct_binop_flags_func *func, match_flags_t flags) { ir_node *src_block = get_nodes_block(node); - ir_node *block = be_transform_node(src_block); ir_node *op1 = get_irn_n(node, n_ia32_l_binop_left); ir_node *op2 = get_irn_n(node, n_ia32_l_binop_right); - ir_node *eflags = get_irn_n(node, n_ia32_l_binop_eflags); - ir_node *new_eflags = be_transform_node(eflags); - ir_graph *irg = current_ir_graph; - dbg_info *dbgi = get_irn_dbg_info(node); - ir_node *new_node; + dbg_info *dbgi; + ir_node *block, *new_node, *eflags, *new_eflags; ia32_address_mode_t am; ia32_address_t *addr = &am.addr; - match_arguments(&am, src_block, op1, op2, flags); + match_arguments(&am, src_block, op1, op2, NULL, flags); - new_node = func(dbgi, irg, block, addr->base, addr->index, - addr->mem, am.new_op1, am.new_op2, new_eflags); + dbgi = get_irn_dbg_info(node); + block = be_transform_node(src_block); + eflags = get_irn_n(node, n_ia32_l_binop_eflags); + new_eflags = be_transform_node(eflags); + new_node = func(dbgi, current_ir_graph, block, addr->base, addr->index, + addr->mem, am.new_op1, am.new_op2, new_eflags); set_am_attributes(new_node, &am); /* we can't use source address mode anymore when using immediates */ if(is_ia32_Immediate(am.new_op1) || is_ia32_Immediate(am.new_op2)) @@ -836,43 +1013,10 @@ static ir_node *gen_binop_flags(ir_node *node, construct_binop_flags_func *func, return new_node; } -/** - * Construct a standard binary operation, set AM and immediate if required. - * - * @param op1 The first operand - * @param op2 The second operand - * @param func The node constructor function - * @return The constructed ia32 node. - */ -static ir_node *gen_binop_sse_float(ir_node *node, ir_node *op1, ir_node *op2, - construct_binop_func *func, - match_flags_t flags) -{ - ir_node *block = get_nodes_block(node); - ir_node *new_block = be_transform_node(block); - dbg_info *dbgi = get_irn_dbg_info(node); - ir_graph *irg = current_ir_graph; - ir_node *new_node; - ia32_address_mode_t am; - ia32_address_t *addr = &am.addr; - - match_arguments(&am, block, op1, op2, flags); - - new_node = func(dbgi, irg, new_block, addr->base, addr->index, addr->mem, - am.new_op1, am.new_op2); - set_am_attributes(new_node, &am); - - SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node)); - - new_node = fix_mem_proj(new_node, &am); - - return new_node; -} - static ir_node *get_fpcw(void) { ir_node *fpcw; - if(initial_fpcw != NULL) + if (initial_fpcw != NULL) return initial_fpcw; fpcw = be_abi_get_ignore_irn(env_cg->birg->abi, @@ -894,18 +1038,24 @@ static ir_node *gen_binop_x87_float(ir_node *node, ir_node *op1, ir_node *op2, construct_binop_float_func *func, match_flags_t flags) { - ir_graph *irg = current_ir_graph; - dbg_info *dbgi = get_irn_dbg_info(node); - ir_node *block = get_nodes_block(node); - ir_node *new_block = be_transform_node(block); - ir_node *new_node; + ir_mode *mode = get_irn_mode(node); + dbg_info *dbgi; + ir_node *block, *new_block, *new_node; ia32_address_mode_t am; ia32_address_t *addr = &am.addr; - match_arguments(&am, block, op1, op2, flags); + /* cannot use address mode with long double on x87 */ + if (get_mode_size_bits(mode) > 64) + flags &= ~match_am; - new_node = func(dbgi, irg, new_block, addr->base, addr->index, addr->mem, - am.new_op1, am.new_op2, get_fpcw()); + block = get_nodes_block(node); + match_arguments(&am, block, op1, op2, NULL, flags); + + dbgi = get_irn_dbg_info(node); + new_block = be_transform_node(block); + new_node = func(dbgi, current_ir_graph, new_block, + addr->base, addr->index, addr->mem, + am.new_op1, am.new_op2, get_fpcw()); set_am_attributes(new_node, &am); SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node)); @@ -924,26 +1074,35 @@ static ir_node *gen_binop_x87_float(ir_node *node, ir_node *op1, ir_node *op2, * @return The constructed ia32 node. */ static ir_node *gen_shift_binop(ir_node *node, ir_node *op1, ir_node *op2, - construct_shift_func *func) + construct_shift_func *func, + match_flags_t flags) { - dbg_info *dbgi = get_irn_dbg_info(node); - ir_graph *irg = current_ir_graph; - ir_node *block = get_nodes_block(node); - ir_node *new_block = be_transform_node(block); - ir_node *new_op1 = be_transform_node(op1); - ir_node *new_op2; - ir_node *new_node; + dbg_info *dbgi; + ir_node *block, *new_block, *new_op1, *new_op2, *new_node; + + assert(! mode_is_float(get_irn_mode(node))); + assert(flags & match_immediate); + assert((flags & ~(match_mode_neutral | match_immediate)) == 0); - assert(! mode_is_float(get_irn_mode(node)) - && "Shift/Rotate with float not supported"); + if (flags & match_mode_neutral) { + op1 = ia32_skip_downconv(op1); + } else if (get_mode_size_bits(get_irn_mode(node)) != 32) { + panic("right shifting of non-32bit values not supported, yet"); + } + new_op1 = be_transform_node(op1); + /* the shift amount can be any mode that is bigger than 5 bits, since all + * other bits are ignored anyway */ while (is_Conv(op2) && get_irn_n_edges(op2) == 1) { op2 = get_Conv_op(op2); assert(get_mode_size_bits(get_irn_mode(op2)) >= 5); } new_op2 = create_immediate_or_transform(op2, 0); - new_node = func(dbgi, irg, new_block, new_op1, new_op2); + dbgi = get_irn_dbg_info(node); + block = get_nodes_block(node); + new_block = be_transform_node(block); + new_node = func(dbgi, current_ir_graph, new_block, new_op1, new_op2); SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node)); /* lowered shift instruction may have a dependency operand, handle it here */ @@ -964,15 +1123,22 @@ static ir_node *gen_shift_binop(ir_node *node, ir_node *op1, ir_node *op2, * @param func The node constructor function * @return The constructed ia32 node. */ -static ir_node *gen_unop(ir_node *node, ir_node *op, construct_unop_func *func) +static ir_node *gen_unop(ir_node *node, ir_node *op, construct_unop_func *func, + match_flags_t flags) { - ir_node *block = be_transform_node(get_nodes_block(node)); - ir_node *new_op = be_transform_node(op); - ir_node *new_node = NULL; - ir_graph *irg = current_ir_graph; - dbg_info *dbgi = get_irn_dbg_info(node); + dbg_info *dbgi; + ir_node *block, *new_block, *new_op, *new_node; - new_node = func(dbgi, irg, block, new_op); + assert(flags == 0 || flags == match_mode_neutral); + if (flags & match_mode_neutral) { + op = ia32_skip_downconv(op); + } + + new_op = be_transform_node(op); + dbgi = get_irn_dbg_info(node); + block = get_nodes_block(node); + new_block = be_transform_node(block); + new_node = func(dbgi, current_ir_graph, new_block, new_op); SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node)); @@ -982,29 +1148,32 @@ static ir_node *gen_unop(ir_node *node, ir_node *op, construct_unop_func *func) static ir_node *create_lea_from_address(dbg_info *dbgi, ir_node *block, ia32_address_t *addr) { - ir_graph *irg = current_ir_graph; - ir_node *base = addr->base; - ir_node *index = addr->index; - ir_node *res; + ir_node *base, *index, *res; - if(base == NULL) { + base = addr->base; + if (base == NULL) { base = ia32_new_NoReg_gp(env_cg); } else { base = be_transform_node(base); } - if(index == NULL) { + index = addr->index; + if (index == NULL) { index = ia32_new_NoReg_gp(env_cg); } else { index = be_transform_node(index); } - res = new_rd_ia32_Lea(dbgi, irg, block, base, index); + res = new_rd_ia32_Lea(dbgi, current_ir_graph, block, base, index); set_address(res, addr); return res; } +/** + * Returns non-zero if a given address mode has a symbolic or + * numerical offset != 0. + */ static int am_has_immediates(const ia32_address_t *addr) { return addr->offset != 0 || addr->symconst_ent != NULL @@ -1017,25 +1186,25 @@ static int am_has_immediates(const ia32_address_t *addr) * @return the created ia32 Add node */ static ir_node *gen_Add(ir_node *node) { - ir_graph *irg = current_ir_graph; - dbg_info *dbgi = get_irn_dbg_info(node); - ir_node *block = get_nodes_block(node); - ir_node *new_block = be_transform_node(block); - ir_node *op1 = get_Add_left(node); - ir_node *op2 = get_Add_right(node); - ir_mode *mode = get_irn_mode(node); - ir_node *new_node; - ir_node *add_immediate_op; + ir_mode *mode = get_irn_mode(node); + ir_node *op1 = get_Add_left(node); + ir_node *op2 = get_Add_right(node); + dbg_info *dbgi; + ir_node *block, *new_block, *new_node, *add_immediate_op; ia32_address_t addr; ia32_address_mode_t am; if (mode_is_float(mode)) { - if (USE_SSE2(env_cg)) - return gen_binop_sse_float(node, op1, op2, new_rd_ia32_xAdd, match_commutative); + if (ia32_cg_config.use_sse2) + return gen_binop(node, op1, op2, new_rd_ia32_xAdd, + match_commutative | match_am); else - return gen_binop_x87_float(node, op1, op2, new_rd_ia32_vfadd, match_commutative); + return gen_binop_x87_float(node, op1, op2, new_rd_ia32_vfadd, + match_commutative | match_am); } + ia32_mark_non_am(node); + op2 = ia32_skip_downconv(op2); op1 = ia32_skip_downconv(op1); @@ -1049,8 +1218,14 @@ static ir_node *gen_Add(ir_node *node) { memset(&addr, 0, sizeof(addr)); ia32_create_address_mode(&addr, node, /*force=*/1); add_immediate_op = NULL; + + dbgi = get_irn_dbg_info(node); + block = get_nodes_block(node); + new_block = be_transform_node(block); + /* a constant? */ if(addr.base == NULL && addr.index == NULL) { + ir_graph *irg = current_ir_graph; new_node = new_rd_ia32_Const(dbgi, irg, new_block, addr.symconst_ent, addr.symconst_sign, addr.offset); add_irn_dep(new_node, get_irg_frame(irg)); @@ -1079,11 +1254,12 @@ static ir_node *gen_Add(ir_node *node) { } /* test if we can use source address mode */ - match_arguments(&am, block, op1, op2, - match_commutative | match_force_32bit_op | match_skip_input_conv); + match_arguments(&am, block, op1, op2, NULL, match_commutative + | match_mode_neutral | match_am | match_immediate | match_try_am); /* construct an Add with source address mode */ if (am.op_type == ia32_AddrModeS) { + ir_graph *irg = current_ir_graph; ia32_address_t *am_addr = &am.addr; new_node = new_rd_ia32_Add(dbgi, irg, new_block, am_addr->base, am_addr->index, am_addr->mem, am.new_op1, @@ -1113,19 +1289,16 @@ static ir_node *gen_Mul(ir_node *node) { ir_mode *mode = get_irn_mode(node); if (mode_is_float(mode)) { - if (USE_SSE2(env_cg)) - return gen_binop_sse_float(node, op1, op2, new_rd_ia32_xMul, match_commutative); + if (ia32_cg_config.use_sse2) + return gen_binop(node, op1, op2, new_rd_ia32_xMul, + match_commutative | match_am); else - return gen_binop_x87_float(node, op1, op2, new_rd_ia32_vfmul, match_commutative); + return gen_binop_x87_float(node, op1, op2, new_rd_ia32_vfmul, + match_commutative | match_am); } - - /* - for the lower 32bit of the result it doesn't matter whether we use - signed or unsigned multiplication so we use IMul as it has fewer - constraints - */ return gen_binop(node, op1, op2, new_rd_ia32_IMul, - match_commutative | match_skip_input_conv | match_force_32bit_op); + match_commutative | match_am | match_mode_neutral | + match_immediate | match_am_and_immediates); } /** @@ -1144,17 +1317,15 @@ static ir_node *gen_Mulh(ir_node *node) ir_mode *mode = get_irn_mode(node); ir_node *op1 = get_Mulh_left(node); ir_node *op2 = get_Mulh_right(node); - ir_node *proj_EDX; + ir_node *proj_res_high; ir_node *new_node; - match_flags_t flags; ia32_address_mode_t am; ia32_address_t *addr = &am.addr; - flags = match_force_32bit_op | match_commutative | match_no_immediate; - assert(!mode_is_float(mode) && "Mulh with float not supported"); + assert(get_mode_size_bits(mode) == 32); - match_arguments(&am, block, op1, op2, flags); + match_arguments(&am, block, op1, op2, NULL, match_commutative | match_am); if (mode_is_signed(mode)) { new_node = new_rd_ia32_IMul1OP(dbgi, irg, new_block, addr->base, @@ -1176,11 +1347,11 @@ static ir_node *gen_Mulh(ir_node *node) fix_mem_proj(new_node, &am); - assert(pn_ia32_IMul1OP_EDX == pn_ia32_Mul_EDX); - proj_EDX = new_rd_Proj(dbgi, irg, block, new_node, - mode_Iu, pn_ia32_IMul1OP_EDX); + assert(pn_ia32_IMul1OP_res_high == pn_ia32_Mul_res_high); + proj_res_high = new_rd_Proj(dbgi, irg, block, new_node, + mode_Iu, pn_ia32_IMul1OP_res_high); - return proj_EDX; + return proj_res_high; } @@ -1217,9 +1388,9 @@ static ir_node *gen_And(ir_node *node) { return res; } } - return gen_binop(node, op1, op2, new_rd_ia32_And, - match_commutative | match_force_32bit_op | match_skip_input_conv); + match_commutative | match_mode_neutral | match_am + | match_immediate); } @@ -1234,8 +1405,8 @@ static ir_node *gen_Or(ir_node *node) { ir_node *op2 = get_Or_right(node); assert (! mode_is_float(get_irn_mode(node))); - return gen_binop(node, op1, op2, new_rd_ia32_Or, - match_commutative | match_skip_input_conv | match_force_32bit_op); + return gen_binop(node, op1, op2, new_rd_ia32_Or, match_commutative + | match_mode_neutral | match_am | match_immediate); } @@ -1250,8 +1421,8 @@ static ir_node *gen_Eor(ir_node *node) { ir_node *op2 = get_Eor_right(node); assert(! mode_is_float(get_irn_mode(node))); - return gen_binop(node, op1, op2, new_rd_ia32_Xor, - match_commutative | match_skip_input_conv | match_force_32bit_op); + return gen_binop(node, op1, op2, new_rd_ia32_Xor, match_commutative + | match_mode_neutral | match_am | match_immediate); } @@ -1266,131 +1437,122 @@ static ir_node *gen_Sub(ir_node *node) { ir_mode *mode = get_irn_mode(node); if (mode_is_float(mode)) { - if (USE_SSE2(env_cg)) - return gen_binop_sse_float(node, op1, op2, new_rd_ia32_xSub, 0); + if (ia32_cg_config.use_sse2) + return gen_binop(node, op1, op2, new_rd_ia32_xSub, match_am); else - return gen_binop_x87_float(node, op1, op2, new_rd_ia32_vfsub, 0); + return gen_binop_x87_float(node, op1, op2, new_rd_ia32_vfsub, + match_am); } - if(is_Const(op2)) { + if (is_Const(op2)) { ir_fprintf(stderr, "Optimisation warning: found sub with const (%+F)\n", node); } - return gen_binop(node, op1, op2, new_rd_ia32_Sub, - match_force_32bit_op | match_skip_input_conv); + return gen_binop(node, op1, op2, new_rd_ia32_Sub, match_mode_neutral + | match_am | match_immediate); } -typedef enum { flavour_Div = 1, flavour_Mod, flavour_DivMod } ia32_op_flavour_t; - /** * Generates an ia32 DivMod with additional infrastructure for the * register allocator if needed. - * - * @param dividend -no comment- :) - * @param divisor -no comment- :) - * @param dm_flav flavour_Div/Mod/DivMod - * @return The created ia32 DivMod node */ -static ir_node *generate_DivMod(ir_node *node, ir_node *dividend, - ir_node *divisor, ia32_op_flavour_t dm_flav) +static ir_node *create_Div(ir_node *node) { - ir_node *block = be_transform_node(get_nodes_block(node)); - ir_node *new_dividend = be_transform_node(dividend); - ir_node *new_divisor = be_transform_node(divisor); - ir_graph *irg = current_ir_graph; - dbg_info *dbgi = get_irn_dbg_info(node); - ir_node *noreg = ia32_new_NoReg_gp(env_cg); - ir_node *res, *proj_div, *proj_mod; + ir_graph *irg = current_ir_graph; + dbg_info *dbgi = get_irn_dbg_info(node); + ir_node *block = get_nodes_block(node); + ir_node *new_block = be_transform_node(block); + ir_node *mem; + ir_node *new_mem; + ir_node *op1; + ir_node *op2; + ir_node *new_node; ir_mode *mode; ir_node *sign_extension; - ir_node *mem, *new_mem; - int has_exc; + ia32_address_mode_t am; + ia32_address_t *addr = &am.addr; /* the upper bits have random contents for smaller modes */ - - proj_div = proj_mod = NULL; - has_exc = 0; - switch (dm_flav) { - case flavour_Div: - mem = get_Div_mem(node); - mode = get_Div_resmode(node); - proj_div = be_get_Proj_for_pn(node, pn_Div_res); - has_exc = be_get_Proj_for_pn(node, pn_Div_X_except) != NULL; - break; - case flavour_Mod: - mem = get_Mod_mem(node); - mode = get_Mod_resmode(node); - proj_mod = be_get_Proj_for_pn(node, pn_Mod_res); - has_exc = be_get_Proj_for_pn(node, pn_Mod_X_except) != NULL; - break; - case flavour_DivMod: - mem = get_DivMod_mem(node); - mode = get_DivMod_resmode(node); - proj_div = be_get_Proj_for_pn(node, pn_DivMod_res_div); - proj_mod = be_get_Proj_for_pn(node, pn_DivMod_res_mod); - has_exc = be_get_Proj_for_pn(node, pn_DivMod_X_except) != NULL; - break; - default: - panic("invalid divmod flavour!"); + switch (get_irn_opcode(node)) { + case iro_Div: + op1 = get_Div_left(node); + op2 = get_Div_right(node); + mem = get_Div_mem(node); + mode = get_Div_resmode(node); + break; + case iro_Mod: + op1 = get_Mod_left(node); + op2 = get_Mod_right(node); + mem = get_Mod_mem(node); + mode = get_Mod_resmode(node); + break; + case iro_DivMod: + op1 = get_DivMod_left(node); + op2 = get_DivMod_right(node); + mem = get_DivMod_mem(node); + mode = get_DivMod_resmode(node); + break; + default: + panic("invalid divmod node %+F", node); } - new_mem = be_transform_node(mem); - assert(get_mode_size_bits(mode) == 32); + match_arguments(&am, block, op1, op2, NULL, match_am); + + /* Beware: We don't need a Sync, if the memory predecessor of the Div node + is the memory of the consumed address. We can have only the second op as address + in Div nodes, so check only op2. */ + if(!is_NoMem(mem) && skip_Proj(mem) != skip_Proj(op2)) { + new_mem = be_transform_node(mem); + if(!is_NoMem(addr->mem)) { + ir_node *in[2]; + in[0] = new_mem; + in[1] = addr->mem; + new_mem = new_rd_Sync(dbgi, irg, new_block, 2, in); + } + } else { + new_mem = addr->mem; + } if (mode_is_signed(mode)) { - /* in signed mode, we need to sign extend the dividend */ - ir_node *produceval = new_rd_ia32_ProduceVal(dbgi, irg, block); + ir_node *produceval = new_rd_ia32_ProduceVal(dbgi, irg, new_block); add_irn_dep(produceval, get_irg_frame(irg)); - sign_extension = new_rd_ia32_Cltd(dbgi, irg, block, new_dividend, - produceval); + sign_extension = new_rd_ia32_Cltd(dbgi, irg, new_block, am.new_op1, + produceval); + + new_node = new_rd_ia32_IDiv(dbgi, irg, new_block, addr->base, + addr->index, new_mem, am.new_op2, + am.new_op1, sign_extension); } else { - sign_extension = new_rd_ia32_Const(dbgi, irg, block, NULL, 0, 0); - set_ia32_flags(sign_extension, get_ia32_flags(sign_extension) | arch_irn_flags_modify_flags); + sign_extension = new_rd_ia32_Const(dbgi, irg, new_block, NULL, 0, 0); add_irn_dep(sign_extension, get_irg_frame(irg)); - } - if (mode_is_signed(mode)) { - res = new_rd_ia32_IDiv(dbgi, irg, block, noreg, noreg, new_mem, - new_dividend, sign_extension, new_divisor); - } else { - res = new_rd_ia32_Div(dbgi, irg, block, noreg, noreg, new_mem, - new_dividend, sign_extension, new_divisor); + new_node = new_rd_ia32_Div(dbgi, irg, new_block, addr->base, + addr->index, new_mem, am.new_op2, + am.new_op1, sign_extension); } - set_ia32_exc_label(res, has_exc); - set_irn_pinned(res, get_irn_pinned(node)); + set_irn_pinned(new_node, get_irn_pinned(node)); - SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node)); + set_am_attributes(new_node, &am); + SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node)); - return res; + new_node = fix_mem_proj(new_node, &am); + + return new_node; } -/** - * Wrapper for generate_DivMod. Sets flavour_Mod. - * - */ static ir_node *gen_Mod(ir_node *node) { - return generate_DivMod(node, get_Mod_left(node), - get_Mod_right(node), flavour_Mod); + return create_Div(node); } -/** - * Wrapper for generate_DivMod. Sets flavour_Div. - * - */ static ir_node *gen_Div(ir_node *node) { - return generate_DivMod(node, get_Div_left(node), - get_Div_right(node), flavour_Div); + return create_Div(node); } -/** - * Wrapper for generate_DivMod. Sets flavour_DivMod. - */ static ir_node *gen_DivMod(ir_node *node) { - return generate_DivMod(node, get_DivMod_left(node), - get_DivMod_right(node), flavour_DivMod); + return create_Div(node); } @@ -1402,13 +1564,13 @@ static ir_node *gen_DivMod(ir_node *node) { */ static ir_node *gen_Quot(ir_node *node) { - ir_node *op1 = get_Quot_left(node); - ir_node *op2 = get_Quot_right(node); + ir_node *op1 = get_Quot_left(node); + ir_node *op2 = get_Quot_right(node); - if (USE_SSE2(env_cg)) { - return gen_binop_sse_float(node, op1, op2, new_rd_ia32_xDiv, 0); + if (ia32_cg_config.use_sse2) { + return gen_binop(node, op1, op2, new_rd_ia32_xDiv, match_am); } else { - return gen_binop_x87_float(node, op1, op2, new_rd_ia32_vfdiv, 0); + return gen_binop_x87_float(node, op1, op2, new_rd_ia32_vfdiv, match_am); } } @@ -1422,21 +1584,20 @@ static ir_node *gen_Shl(ir_node *node) { ir_node *left = get_Shl_left(node); ir_node *right = get_Shl_right(node); - left = ia32_skip_downconv(left); - return gen_shift_binop(node, left, right, new_rd_ia32_Shl); + return gen_shift_binop(node, left, right, new_rd_ia32_Shl, + match_mode_neutral | match_immediate); } - - /** * Creates an ia32 Shr. * * @return The created ia32 Shr node */ static ir_node *gen_Shr(ir_node *node) { - assert(get_mode_size_bits(get_irn_mode(node)) == 32); - return gen_shift_binop(node, get_Shr_left(node), - get_Shr_right(node), new_rd_ia32_Shr); + ir_node *left = get_Shr_left(node); + ir_node *right = get_Shr_right(node); + + return gen_shift_binop(node, left, right, new_rd_ia32_Shr, match_immediate); } @@ -1451,8 +1612,6 @@ static ir_node *gen_Shrs(ir_node *node) { ir_node *right = get_Shrs_right(node); ir_mode *mode = get_irn_mode(node); - assert(get_mode_size_bits(mode) == 32); - if(is_Const(right) && mode == mode_Is) { tarval *tv = get_Const_tarval(right); long val = get_tarval_long(tv); @@ -1500,7 +1659,7 @@ static ir_node *gen_Shrs(ir_node *node) { } } - return gen_shift_binop(node, left, right, new_rd_ia32_Sar); + return gen_shift_binop(node, left, right, new_rd_ia32_Sar, match_immediate); } @@ -1512,10 +1671,8 @@ static ir_node *gen_Shrs(ir_node *node) { * @param op2 The second operator * @return The created ia32 RotL node */ -static ir_node *gen_RotL(ir_node *node, - ir_node *op1, ir_node *op2) { - assert(get_mode_size_bits(get_irn_mode(node)) == 32); - return gen_shift_binop(node, op1, op2, new_rd_ia32_Rol); +static ir_node *gen_RotL(ir_node *node, ir_node *op1, ir_node *op2) { + return gen_shift_binop(node, op1, op2, new_rd_ia32_Rol, match_immediate); } @@ -1529,10 +1686,8 @@ static ir_node *gen_RotL(ir_node *node, * @param op2 The second operator * @return The created ia32 RotR node */ -static ir_node *gen_RotR(ir_node *node, ir_node *op1, - ir_node *op2) { - assert(get_mode_size_bits(get_irn_mode(node)) == 32); - return gen_shift_binop(node, op1, op2, new_rd_ia32_Ror); +static ir_node *gen_RotR(ir_node *node, ir_node *op1, ir_node *op2) { + return gen_shift_binop(node, op1, op2, new_rd_ia32_Ror, match_immediate); } @@ -1593,36 +1748,38 @@ static ir_node *gen_Minus(ir_node *node) dbg_info *dbgi = get_irn_dbg_info(node); ir_mode *mode = get_irn_mode(node); ir_entity *ent; - ir_node *res; - int size; + ir_node *new_node; + int size; if (mode_is_float(mode)) { ir_node *new_op = be_transform_node(op); - if (USE_SSE2(env_cg)) { + if (ia32_cg_config.use_sse2) { + /* TODO: non-optimal... if we have many xXors, then we should + * rather create a load for the const and use that instead of + * several AM nodes... */ ir_node *noreg_gp = ia32_new_NoReg_gp(env_cg); ir_node *noreg_xmm = ia32_new_NoReg_xmm(env_cg); ir_node *nomem = new_rd_NoMem(irg); - res = new_rd_ia32_xXor(dbgi, irg, block, noreg_gp, noreg_gp, nomem, - new_op, noreg_xmm); + new_node = new_rd_ia32_xXor(dbgi, irg, block, noreg_gp, noreg_gp, + nomem, new_op, noreg_xmm); size = get_mode_size_bits(mode); ent = ia32_gen_fp_known_const(size == 32 ? ia32_SSIGN : ia32_DSIGN); - set_ia32_am_sc(res, ent); - set_ia32_op_type(res, ia32_AddrModeS); - set_ia32_ls_mode(res, mode); + set_ia32_am_sc(new_node, ent); + set_ia32_op_type(new_node, ia32_AddrModeS); + set_ia32_ls_mode(new_node, mode); } else { - res = new_rd_ia32_vfchs(dbgi, irg, block, new_op); + new_node = new_rd_ia32_vfchs(dbgi, irg, block, new_op); } } else { - op = ia32_skip_downconv(op); - res = gen_unop(node, op, new_rd_ia32_Neg); + new_node = gen_unop(node, op, new_rd_ia32_Neg, match_mode_neutral); } - SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node)); + SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node)); - return res; + return new_node; } /** @@ -1636,8 +1793,7 @@ static ir_node *gen_Not(ir_node *node) { assert(get_irn_mode(node) != mode_b); /* should be lowered already */ assert (! mode_is_float(get_irn_mode(node))); - node = ia32_skip_downconv(node); - return gen_unop(node, op, new_rd_ia32_Not); + return gen_unop(node, op, new_rd_ia32_Not, match_mode_neutral); } @@ -1649,56 +1805,153 @@ static ir_node *gen_Not(ir_node *node) { */ static ir_node *gen_Abs(ir_node *node) { - ir_node *block = be_transform_node(get_nodes_block(node)); - ir_node *op = get_Abs_op(node); - ir_node *new_op = be_transform_node(op); - ir_graph *irg = current_ir_graph; - dbg_info *dbgi = get_irn_dbg_info(node); - ir_mode *mode = get_irn_mode(node); - ir_node *noreg_gp = ia32_new_NoReg_gp(env_cg); - ir_node *noreg_fp = ia32_new_NoReg_fp(env_cg); - ir_node *nomem = new_NoMem(); - ir_node *res; - int size; + ir_node *block = get_nodes_block(node); + ir_node *new_block = be_transform_node(block); + ir_node *op = get_Abs_op(node); + ir_graph *irg = current_ir_graph; + dbg_info *dbgi = get_irn_dbg_info(node); + ir_mode *mode = get_irn_mode(node); + ir_node *noreg_gp = ia32_new_NoReg_gp(env_cg); + ir_node *nomem = new_NoMem(); + ir_node *new_op; + ir_node *new_node; + int size; ir_entity *ent; if (mode_is_float(mode)) { - if (USE_SSE2(env_cg)) { - res = new_rd_ia32_xAnd(dbgi,irg, block, noreg_gp, noreg_gp, nomem, new_op, noreg_fp); + new_op = be_transform_node(op); + + if (ia32_cg_config.use_sse2) { + ir_node *noreg_fp = ia32_new_NoReg_xmm(env_cg); + new_node = new_rd_ia32_xAnd(dbgi,irg, new_block, noreg_gp, noreg_gp, + nomem, new_op, noreg_fp); size = get_mode_size_bits(mode); ent = ia32_gen_fp_known_const(size == 32 ? ia32_SABS : ia32_DABS); - set_ia32_am_sc(res, ent); + set_ia32_am_sc(new_node, ent); - SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node)); + SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node)); - set_ia32_op_type(res, ia32_AddrModeS); - set_ia32_ls_mode(res, mode); + set_ia32_op_type(new_node, ia32_AddrModeS); + set_ia32_ls_mode(new_node, mode); } else { - res = new_rd_ia32_vfabs(dbgi, irg, block, new_op); - SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node)); + new_node = new_rd_ia32_vfabs(dbgi, irg, new_block, new_op); + SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node)); } } else { - ir_node *xor; - ir_node *pval = new_rd_ia32_ProduceVal(dbgi, irg, block); - ir_node *sign_extension = new_rd_ia32_Cltd(dbgi, irg, block, new_op, - pval); + ir_node *xor, *pval, *sign_extension; + + if (get_mode_size_bits(mode) == 32) { + new_op = be_transform_node(op); + } else { + new_op = create_I2I_Conv(mode, mode_Is, dbgi, block, op, node); + } + + pval = new_rd_ia32_ProduceVal(dbgi, irg, new_block); + sign_extension = new_rd_ia32_Cltd(dbgi, irg, new_block, + new_op, pval); add_irn_dep(pval, get_irg_frame(irg)); - SET_IA32_ORIG_NODE(sign_extension, - ia32_get_old_node_name(env_cg, node)); + SET_IA32_ORIG_NODE(sign_extension,ia32_get_old_node_name(env_cg, node)); - xor = new_rd_ia32_Xor(dbgi, irg, block, noreg_gp, noreg_gp, nomem, new_op, - sign_extension); + xor = new_rd_ia32_Xor(dbgi, irg, new_block, noreg_gp, noreg_gp, + nomem, new_op, sign_extension); SET_IA32_ORIG_NODE(xor, ia32_get_old_node_name(env_cg, node)); - res = new_rd_ia32_Sub(dbgi, irg, block, noreg_gp, noreg_gp, nomem, xor, - sign_extension); - SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node)); + new_node = new_rd_ia32_Sub(dbgi, irg, new_block, noreg_gp, noreg_gp, + nomem, xor, sign_extension); + SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node)); } - return res; + return new_node; +} + +/** + * Create a bt instruction for x & (1 << n) and place it into the block of cmp. + */ +static ir_node *gen_bt(ir_node *cmp, ir_node *x, ir_node *n) { + dbg_info *dbgi = get_irn_dbg_info(cmp); + ir_node *block = get_nodes_block(cmp); + ir_node *new_block = be_transform_node(block); + ir_node *op1 = be_transform_node(x); + ir_node *op2 = be_transform_node(n); + + return new_rd_ia32_Bt(dbgi, current_ir_graph, new_block, op1, op2); +} + +/** + * Transform a node returning a "flag" result. + * + * @param node the node to transform + * @param pnc_out the compare mode to use + */ +static ir_node *get_flags_node(ir_node *node, pn_Cmp *pnc_out) +{ + ir_node *flags; + ir_node *new_op; + ir_node *noreg; + ir_node *nomem; + ir_node *new_block; + dbg_info *dbgi; + + /* we have a Cmp as input */ + if (is_Proj(node)) { + ir_node *pred = get_Proj_pred(node); + if (is_Cmp(pred)) { + pn_Cmp pnc = get_Proj_proj(node); + if (ia32_cg_config.use_bt && (pnc == pn_Cmp_Lg || pnc == pn_Cmp_Eq)) { + ir_node *l = get_Cmp_left(pred); + ir_node *r = get_Cmp_right(pred); + if (is_And(l)) { + ir_node *la = get_And_left(l); + ir_node *ra = get_And_right(l); + if (is_Shl(la)) { + ir_node *c = get_Shl_left(la); + if (is_Const_1(c) && (is_Const_0(r) || r == la)) { + /* (1 << n) & ra) */ + ir_node *n = get_Shl_right(la); + flags = gen_bt(pred, ra, n); + /* we must generate a Jc/Jnc jump */ + pnc = pnc == pn_Cmp_Lg ? pn_Cmp_Lt : pn_Cmp_Ge; + if (r == la) + pnc ^= pn_Cmp_Leg; + *pnc_out = ia32_pn_Cmp_unsigned | pnc; + return flags; + } + } + if (is_Shl(ra)) { + ir_node *c = get_Shl_left(ra); + if (is_Const_1(c) && (is_Const_0(r) || r == ra)) { + /* la & (1 << n)) */ + ir_node *n = get_Shl_right(ra); + flags = gen_bt(pred, la, n); + /* we must generate a Jc/Jnc jump */ + pnc = pnc == pn_Cmp_Lg ? pn_Cmp_Lt : pn_Cmp_Ge; + if (r == ra) + pnc ^= pn_Cmp_Leg; + *pnc_out = ia32_pn_Cmp_unsigned | pnc; + return flags; + } + } + } + } + flags = be_transform_node(pred); + *pnc_out = pnc; + return flags; + } + } + + /* a mode_b value, we have to compare it against 0 */ + dbgi = get_irn_dbg_info(node); + new_block = be_transform_node(get_nodes_block(node)); + new_op = be_transform_node(node); + noreg = ia32_new_NoReg_gp(env_cg); + nomem = new_NoMem(); + flags = new_rd_ia32_Test(dbgi, current_ir_graph, new_block, noreg, noreg, nomem, + new_op, new_op, /*is_permuted=*/0, /*cmp_unsigned=*/0); + *pnc_out = pn_Cmp_Lg; + return flags; } /** @@ -1719,7 +1972,7 @@ static ir_node *gen_Load(ir_node *node) { ir_node *noreg = ia32_new_NoReg_gp(env_cg); ir_mode *mode = get_Load_mode(node); ir_mode *res_mode; - ir_node *new_op; + ir_node *new_node; ia32_address_t addr; /* construct load address */ @@ -1741,45 +1994,47 @@ static ir_node *gen_Load(ir_node *node) { } if (mode_is_float(mode)) { - if (USE_SSE2(env_cg)) { - new_op = new_rd_ia32_xLoad(dbgi, irg, block, base, index, new_mem, - mode); + if (ia32_cg_config.use_sse2) { + new_node = new_rd_ia32_xLoad(dbgi, irg, block, base, index, new_mem, + mode); res_mode = mode_xmm; } else { - new_op = new_rd_ia32_vfld(dbgi, irg, block, base, index, new_mem, + new_node = new_rd_ia32_vfld(dbgi, irg, block, base, index, new_mem, mode); res_mode = mode_vfp; } } else { - if(mode == mode_b) - mode = mode_Iu; + assert(mode != mode_b); /* create a conv node with address mode for smaller modes */ if(get_mode_size_bits(mode) < 32) { - new_op = new_rd_ia32_Conv_I2I(dbgi, irg, block, base, index, - new_mem, noreg, mode); + new_node = new_rd_ia32_Conv_I2I(dbgi, irg, block, base, index, + new_mem, noreg, mode); } else { - new_op = new_rd_ia32_Load(dbgi, irg, block, base, index, new_mem); + new_node = new_rd_ia32_Load(dbgi, irg, block, base, index, new_mem); } res_mode = mode_Iu; } - set_irn_pinned(new_op, get_irn_pinned(node)); - set_ia32_op_type(new_op, ia32_AddrModeS); - set_ia32_ls_mode(new_op, mode); - set_address(new_op, &addr); + set_irn_pinned(new_node, get_irn_pinned(node)); + set_ia32_op_type(new_node, ia32_AddrModeS); + set_ia32_ls_mode(new_node, mode); + set_address(new_node, &addr); + + if(get_irn_pinned(node) == op_pin_state_floats) { + add_ia32_flags(new_node, arch_irn_flags_rematerializable); + } /* make sure we are scheduled behind the initial IncSP/Barrier * to avoid spills being placed before it */ if (block == get_irg_start_block(irg)) { - add_irn_dep(new_op, get_irg_frame(irg)); + add_irn_dep(new_node, get_irg_frame(irg)); } - set_ia32_exc_label(new_op, be_get_Proj_for_pn(node, pn_Load_X_except) != NULL); - SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node)); + SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node)); - return new_op; + return new_node; } static int use_dest_am(ir_node *block, ir_node *node, ir_node *mem, @@ -1819,19 +2074,24 @@ static ir_node *dest_am_binop(ir_node *node, ir_node *op1, ir_node *op2, ir_node *mem, ir_node *ptr, ir_mode *mode, construct_binop_dest_func *func, construct_binop_dest_func *func8bit, - int commutative) + match_flags_t flags) { - ir_node *src_block = get_nodes_block(node); - ir_node *block; - ir_node *noreg_gp = ia32_new_NoReg_gp(env_cg); + ir_node *src_block = get_nodes_block(node); + ir_node *block; + ir_node *noreg_gp = ia32_new_NoReg_gp(env_cg); ir_graph *irg = current_ir_graph; dbg_info *dbgi; - ir_node *new_node; - ir_node *new_op; + ir_node *new_node; + ir_node *new_op; + int commutative; ia32_address_mode_t am; - ia32_address_t *addr = &am.addr; + ia32_address_t *addr = &am.addr; memset(&am, 0, sizeof(am)); + assert(flags & match_dest_am); + assert(flags & match_immediate); /* there is no destam node without... */ + commutative = (flags & match_commutative) != 0; + if(use_dest_am(src_block, op1, mem, ptr, op2)) { build_address(&am, op1); new_op = create_immediate_or_transform(op2, 0); @@ -1849,8 +2109,8 @@ static ir_node *dest_am_binop(ir_node *node, ir_node *op1, ir_node *op2, if(addr->mem == NULL) addr->mem = new_NoMem(); - dbgi = get_irn_dbg_info(node); - block = be_transform_node(src_block); + dbgi = get_irn_dbg_info(node); + block = be_transform_node(src_block); if(get_mode_size_bits(mode) == 8) { new_node = func8bit(dbgi, irg, block, addr->base, addr->index, addr->mem, new_op); @@ -1870,10 +2130,9 @@ static ir_node *dest_am_unop(ir_node *node, ir_node *op, ir_node *mem, ir_node *ptr, ir_mode *mode, construct_unop_dest_func *func) { + ir_graph *irg = current_ir_graph; ir_node *src_block = get_nodes_block(node); ir_node *block; - ir_node *noreg_gp = ia32_new_NoReg_gp(env_cg); - ir_graph *irg = current_ir_graph; dbg_info *dbgi; ir_node *new_node; ia32_address_mode_t am; @@ -1885,13 +2144,6 @@ static ir_node *dest_am_unop(ir_node *node, ir_node *op, ir_node *mem, build_address(&am, op); - if(addr->base == NULL) - addr->base = noreg_gp; - if(addr->index == NULL) - addr->index = noreg_gp; - if(addr->mem == NULL) - addr->mem = new_NoMem(); - dbgi = get_irn_dbg_info(node); block = be_transform_node(src_block); new_node = func(dbgi, irg, block, addr->base, addr->index, addr->mem); @@ -1903,11 +2155,58 @@ static ir_node *dest_am_unop(ir_node *node, ir_node *op, ir_node *mem, return new_node; } +static ir_node *try_create_SetMem(ir_node *node, ir_node *ptr, ir_node *mem) { + ir_mode *mode = get_irn_mode(node); + ir_node *psi_true = get_Psi_val(node, 0); + ir_node *psi_default = get_Psi_default(node); + ir_graph *irg; + ir_node *cond; + ir_node *new_mem; + dbg_info *dbgi; + ir_node *block; + ir_node *new_block; + ir_node *flags; + ir_node *new_node; + int negated; + pn_Cmp pnc; + ia32_address_t addr; + + if(get_mode_size_bits(mode) != 8) + return NULL; + + if(is_Const_1(psi_true) && is_Const_0(psi_default)) { + negated = 0; + } else if(is_Const_0(psi_true) && is_Const_1(psi_default)) { + negated = 1; + } else { + return NULL; + } + + build_address_ptr(&addr, ptr, mem); + + irg = current_ir_graph; + dbgi = get_irn_dbg_info(node); + block = get_nodes_block(node); + new_block = be_transform_node(block); + cond = get_Psi_cond(node, 0); + flags = get_flags_node(cond, &pnc); + new_mem = be_transform_node(mem); + new_node = new_rd_ia32_SetMem(dbgi, irg, new_block, addr.base, + addr.index, addr.mem, flags, pnc, negated); + set_address(new_node, &addr); + set_ia32_op_type(new_node, ia32_AddrModeD); + set_ia32_ls_mode(new_node, mode); + SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node)); + + return new_node; +} + static ir_node *try_create_dest_am(ir_node *node) { - ir_node *val = get_Store_value(node); - ir_node *mem = get_Store_mem(node); - ir_node *ptr = get_Store_ptr(node); - ir_mode *mode = get_irn_mode(val); + ir_node *val = get_Store_value(node); + ir_node *mem = get_Store_mem(node); + ir_node *ptr = get_Store_ptr(node); + ir_mode *mode = get_irn_mode(val); + unsigned bits = get_mode_size_bits(mode); ir_node *op1; ir_node *op2; ir_node *new_node; @@ -1916,8 +2215,24 @@ static ir_node *try_create_dest_am(ir_node *node) { if(!mode_needs_gp_reg(mode)) return NULL; - /* store must be the only user of the val node */ - if(get_irn_n_edges(val) > 1) + while(1) { + /* store must be the only user of the val node */ + if(get_irn_n_edges(val) > 1) + return NULL; + /* skip pointless convs */ + if(is_Conv(val)) { + ir_node *conv_op = get_Conv_op(val); + ir_mode *pred_mode = get_irn_mode(conv_op); + if(pred_mode == mode_b || bits <= get_mode_size_bits(pred_mode)) { + val = conv_op; + continue; + } + } + break; + } + + /* value must be in the same block */ + if(get_nodes_block(node) != get_nodes_block(val)) return NULL; switch(get_irn_opcode(val)) { @@ -1934,7 +2249,9 @@ static ir_node *try_create_dest_am(ir_node *node) { break; } new_node = dest_am_binop(val, op1, op2, mem, ptr, mode, - new_rd_ia32_AddMem, new_rd_ia32_AddMem8Bit, 1); + new_rd_ia32_AddMem, new_rd_ia32_AddMem8Bit, + match_dest_am | match_commutative | + match_immediate); break; case iro_Sub: op1 = get_Sub_left(val); @@ -1944,51 +2261,66 @@ static ir_node *try_create_dest_am(ir_node *node) { "found\n"); } new_node = dest_am_binop(val, op1, op2, mem, ptr, mode, - new_rd_ia32_SubMem, new_rd_ia32_SubMem8Bit, 0); + new_rd_ia32_SubMem, new_rd_ia32_SubMem8Bit, + match_dest_am | match_immediate | + match_immediate); break; case iro_And: op1 = get_And_left(val); op2 = get_And_right(val); new_node = dest_am_binop(val, op1, op2, mem, ptr, mode, - new_rd_ia32_AndMem, new_rd_ia32_AndMem8Bit, 1); + new_rd_ia32_AndMem, new_rd_ia32_AndMem8Bit, + match_dest_am | match_commutative | + match_immediate); break; case iro_Or: op1 = get_Or_left(val); op2 = get_Or_right(val); new_node = dest_am_binop(val, op1, op2, mem, ptr, mode, - new_rd_ia32_OrMem, new_rd_ia32_OrMem8Bit, 1); + new_rd_ia32_OrMem, new_rd_ia32_OrMem8Bit, + match_dest_am | match_commutative | + match_immediate); break; case iro_Eor: op1 = get_Eor_left(val); op2 = get_Eor_right(val); new_node = dest_am_binop(val, op1, op2, mem, ptr, mode, - new_rd_ia32_XorMem, new_rd_ia32_XorMem8Bit, 1); + new_rd_ia32_XorMem, new_rd_ia32_XorMem8Bit, + match_dest_am | match_commutative | + match_immediate); break; case iro_Shl: op1 = get_Shl_left(val); op2 = get_Shl_right(val); new_node = dest_am_binop(val, op1, op2, mem, ptr, mode, - new_rd_ia32_ShlMem, new_rd_ia32_ShlMem, 0); + new_rd_ia32_ShlMem, new_rd_ia32_ShlMem, + match_dest_am | match_immediate); break; case iro_Shr: op1 = get_Shr_left(val); op2 = get_Shr_right(val); new_node = dest_am_binop(val, op1, op2, mem, ptr, mode, - new_rd_ia32_ShrMem, new_rd_ia32_ShrMem, 0); + new_rd_ia32_ShrMem, new_rd_ia32_ShrMem, + match_dest_am | match_immediate); break; case iro_Shrs: op1 = get_Shrs_left(val); op2 = get_Shrs_right(val); new_node = dest_am_binop(val, op1, op2, mem, ptr, mode, - new_rd_ia32_SarMem, new_rd_ia32_SarMem, 0); + new_rd_ia32_SarMem, new_rd_ia32_SarMem, + match_dest_am | match_immediate); break; case iro_Rot: op1 = get_Rot_left(val); op2 = get_Rot_right(val); new_node = dest_am_binop(val, op1, op2, mem, ptr, mode, - new_rd_ia32_RolMem, new_rd_ia32_RolMem, 0); + new_rd_ia32_RolMem, new_rd_ia32_RolMem, + match_dest_am | match_immediate); break; /* TODO: match ROR patterns... */ + case iro_Psi: + new_node = try_create_SetMem(val, ptr, mem); + break; case iro_Minus: op1 = get_Minus_op(val); new_node = dest_am_unop(val, op1, mem, ptr, mode, new_rd_ia32_NegMem); @@ -2003,101 +2335,273 @@ static ir_node *try_create_dest_am(ir_node *node) { return NULL; } + if(new_node != NULL) { + if(get_irn_pinned(new_node) != op_pin_state_pinned && + get_irn_pinned(node) == op_pin_state_pinned) { + set_irn_pinned(new_node, op_pin_state_pinned); + } + } + return new_node; } +static int is_float_to_int32_conv(const ir_node *node) +{ + ir_mode *mode = get_irn_mode(node); + ir_node *conv_op; + ir_mode *conv_mode; + + if(get_mode_size_bits(mode) != 32 || !mode_needs_gp_reg(mode)) + return 0; + + if(!is_Conv(node)) + return 0; + conv_op = get_Conv_op(node); + conv_mode = get_irn_mode(conv_op); + + if(!mode_is_float(conv_mode)) + return 0; + + return 1; +} + /** - * Transforms a Store. + * Transform a Store(floatConst). * * @return the created ia32 Store node */ -static ir_node *gen_Store(ir_node *node) { - ir_node *block = be_transform_node(get_nodes_block(node)); - ir_node *ptr = get_Store_ptr(node); - ir_node *base; - ir_node *index; - ir_node *val = get_Store_value(node); - ir_node *new_val; - ir_node *mem = get_Store_mem(node); - ir_node *new_mem = be_transform_node(mem); - ir_graph *irg = current_ir_graph; - dbg_info *dbgi = get_irn_dbg_info(node); - ir_node *noreg = ia32_new_NoReg_gp(env_cg); - ir_mode *mode = get_irn_mode(val); - ir_node *new_op; +static ir_node *gen_float_const_Store(ir_node *node, ir_node *cns) { + ir_mode *mode = get_irn_mode(cns); + int size = get_mode_size_bits(mode); + tarval *tv = get_Const_tarval(cns); + ir_node *block = get_nodes_block(node); + ir_node *new_block = be_transform_node(block); + ir_node *ptr = get_Store_ptr(node); + ir_node *mem = get_Store_mem(node); + ir_graph *irg = current_ir_graph; + dbg_info *dbgi = get_irn_dbg_info(node); + ir_node *noreg = ia32_new_NoReg_gp(env_cg); + int ofs = 4; + ir_node *new_node; + ia32_address_t addr; + + unsigned val = get_tarval_sub_bits(tv, 0) | + (get_tarval_sub_bits(tv, 1) << 8) | + (get_tarval_sub_bits(tv, 2) << 16) | + (get_tarval_sub_bits(tv, 3) << 24); + ir_node *imm = create_Immediate(NULL, 0, val); + + /* construct store address */ + memset(&addr, 0, sizeof(addr)); + ia32_create_address_mode(&addr, ptr, /*force=*/0); + + if (addr.base == NULL) { + addr.base = noreg; + } else { + addr.base = be_transform_node(addr.base); + } + + if (addr.index == NULL) { + addr.index = noreg; + } else { + addr.index = be_transform_node(addr.index); + } + addr.mem = be_transform_node(mem); + + new_node = new_rd_ia32_Store(dbgi, irg, new_block, addr.base, + addr.index, addr.mem, imm); + + set_irn_pinned(new_node, get_irn_pinned(node)); + set_ia32_op_type(new_node, ia32_AddrModeD); + set_ia32_ls_mode(new_node, mode_Iu); + + set_address(new_node, &addr); + + /** add more stores if needed */ + while (size > 32) { + unsigned val = get_tarval_sub_bits(tv, ofs) | + (get_tarval_sub_bits(tv, ofs + 1) << 8) | + (get_tarval_sub_bits(tv, ofs + 2) << 16) | + (get_tarval_sub_bits(tv, ofs + 3) << 24); + ir_node *imm = create_Immediate(NULL, 0, val); + + addr.offset += 4; + addr.mem = new_node; + + new_node = new_rd_ia32_Store(dbgi, irg, new_block, addr.base, + addr.index, addr.mem, imm); + + set_irn_pinned(new_node, get_irn_pinned(node)); + set_ia32_op_type(new_node, ia32_AddrModeD); + set_ia32_ls_mode(new_node, mode_Iu); + + set_address(new_node, &addr); + size -= 32; + ofs += 4; + } + + SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node)); + return new_node; +} + +/** + * Generate a vfist or vfisttp instruction. + */ +static ir_node *gen_vfist(dbg_info *dbgi, ir_graph *irg, ir_node *block, ir_node *base, ir_node *index, + ir_node *mem, ir_node *val, ir_node **fist) +{ + ir_node *new_node; + + if (ia32_cg_config.use_fisttp) { + /* Note: fisttp ALWAYS pop the tos. We have to ensure here that the value is copied + if other users exists */ + const arch_register_class_t *reg_class = &ia32_reg_classes[CLASS_ia32_vfp]; + ir_node *vfisttp = new_rd_ia32_vfisttp(dbgi, irg, block, base, index, mem, val); + ir_node *value = new_r_Proj(irg, block, vfisttp, mode_E, pn_ia32_vfisttp_res); + be_new_Keep(reg_class, irg, block, 1, &value); + + new_node = new_r_Proj(irg, block, vfisttp, mode_M, pn_ia32_vfisttp_M); + *fist = vfisttp; + } else { + ir_node *trunc_mode = ia32_new_Fpu_truncate(env_cg); + + /* do a fist */ + new_node = new_rd_ia32_vfist(dbgi, irg, block, base, index, mem, val, trunc_mode); + *fist = new_node; + } + return new_node; +} +/** + * Transforms a normal Store. + * + * @return the created ia32 Store node + */ +static ir_node *gen_normal_Store(ir_node *node) +{ + ir_node *val = get_Store_value(node); + ir_mode *mode = get_irn_mode(val); + ir_node *block = get_nodes_block(node); + ir_node *new_block = be_transform_node(block); + ir_node *ptr = get_Store_ptr(node); + ir_node *mem = get_Store_mem(node); + ir_graph *irg = current_ir_graph; + dbg_info *dbgi = get_irn_dbg_info(node); + ir_node *noreg = ia32_new_NoReg_gp(env_cg); + ir_node *new_val, *new_node, *store; ia32_address_t addr; /* check for destination address mode */ - new_op = try_create_dest_am(node); - if(new_op != NULL) - return new_op; + new_node = try_create_dest_am(node); + if (new_node != NULL) + return new_node; /* construct store address */ memset(&addr, 0, sizeof(addr)); ia32_create_address_mode(&addr, ptr, /*force=*/0); - base = addr.base; - index = addr.index; - if(base == NULL) { - base = noreg; + if (addr.base == NULL) { + addr.base = noreg; } else { - base = be_transform_node(base); + addr.base = be_transform_node(addr.base); } - if(index == NULL) { - index = noreg; + if (addr.index == NULL) { + addr.index = noreg; } else { - index = be_transform_node(index); + addr.index = be_transform_node(addr.index); } + addr.mem = be_transform_node(mem); if (mode_is_float(mode)) { /* convs (and strict-convs) before stores are unnecessary if the mode is the same */ - while(is_Conv(val) && mode == get_irn_mode(get_Conv_op(val))) { + while (is_Conv(val) && mode == get_irn_mode(get_Conv_op(val))) { val = get_Conv_op(val); } new_val = be_transform_node(val); - if (USE_SSE2(env_cg)) { - new_op = new_rd_ia32_xStore(dbgi, irg, block, base, index, new_mem, - new_val); + if (ia32_cg_config.use_sse2) { + new_node = new_rd_ia32_xStore(dbgi, irg, new_block, addr.base, + addr.index, addr.mem, new_val); } else { - new_op = new_rd_ia32_vfst(dbgi, irg, block, base, index, new_mem, new_val, - mode); + new_node = new_rd_ia32_vfst(dbgi, irg, new_block, addr.base, + addr.index, addr.mem, new_val, mode); } + store = new_node; + } else if (is_float_to_int32_conv(val)) { + val = get_Conv_op(val); + + /* convs (and strict-convs) before stores are unnecessary if the mode + is the same */ + while(is_Conv(val) && mode == get_irn_mode(get_Conv_op(val))) { + val = get_Conv_op(val); + } + new_val = be_transform_node(val); + new_node = gen_vfist(dbgi, irg, new_block, addr.base, addr.index, addr.mem, new_val, &store); } else { new_val = create_immediate_or_transform(val, 0); - if(mode == mode_b) - mode = mode_Iu; + assert(mode != mode_b); if (get_mode_size_bits(mode) == 8) { - new_op = new_rd_ia32_Store8Bit(dbgi, irg, block, base, index, new_mem, - new_val); + new_node = new_rd_ia32_Store8Bit(dbgi, irg, new_block, addr.base, + addr.index, addr.mem, new_val); } else { - new_op = new_rd_ia32_Store(dbgi, irg, block, base, index, new_mem, - new_val); + new_node = new_rd_ia32_Store(dbgi, irg, new_block, addr.base, + addr.index, addr.mem, new_val); } + store = new_node; } - set_irn_pinned(new_op, get_irn_pinned(node)); - set_ia32_op_type(new_op, ia32_AddrModeD); - set_ia32_ls_mode(new_op, mode); + set_irn_pinned(store, get_irn_pinned(node)); + set_ia32_op_type(store, ia32_AddrModeD); + set_ia32_ls_mode(store, mode); - set_ia32_exc_label(new_op, be_get_Proj_for_pn(node, pn_Store_X_except) != NULL); - set_address(new_op, &addr); - SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node)); + set_address(store, &addr); + SET_IA32_ORIG_NODE(store, ia32_get_old_node_name(env_cg, node)); - return new_op; + return new_node; +} + +/** + * Transforms a Store. + * + * @return the created ia32 Store node + */ +static ir_node *gen_Store(ir_node *node) +{ + ir_node *val = get_Store_value(node); + ir_mode *mode = get_irn_mode(val); + + if (mode_is_float(mode) && is_Const(val)) { + int transform = 1; + + /* we are storing a floating point constant */ + if (ia32_cg_config.use_sse2) { + transform = !is_simple_sse_Const(val); + } else { + transform = !is_simple_x87_Const(val); + } + if (transform) + return gen_float_const_Store(node, val); + } + return gen_normal_Store(node); } +/** + * Transforms a Switch. + * + * @return the created ia32 SwitchJmp node + */ static ir_node *create_Switch(ir_node *node) { - ir_graph *irg = current_ir_graph; - dbg_info *dbgi = get_irn_dbg_info(node); - ir_node *block = be_transform_node(get_nodes_block(node)); - ir_node *sel = get_Cond_selector(node); - ir_node *new_sel = be_transform_node(sel); - ir_node *res; - int switch_min = INT_MAX; + ir_graph *irg = current_ir_graph; + dbg_info *dbgi = get_irn_dbg_info(node); + ir_node *block = be_transform_node(get_nodes_block(node)); + ir_node *sel = get_Cond_selector(node); + ir_node *new_sel = be_transform_node(sel); + int switch_min = INT_MAX; + int switch_max = INT_MIN; + long default_pn = get_Cond_defaultProj(node); + ir_node *new_node; const ir_edge_t *edge; assert(get_mode_size_bits(get_irn_mode(sel)) == 32); @@ -2105,9 +2609,18 @@ static ir_node *create_Switch(ir_node *node) /* determine the smallest switch case value */ foreach_out_edge(node, edge) { ir_node *proj = get_edge_src_irn(edge); - int pn = get_Proj_proj(proj); + long pn = get_Proj_proj(proj); + if(pn == default_pn) + continue; + if(pn < switch_min) switch_min = pn; + if(pn > switch_max) + switch_max = pn; + } + + if((unsigned) (switch_max - switch_min) > 256000) { + panic("Size of switch %+F bigger than 256000", node); } if (switch_min != 0) { @@ -2121,46 +2634,15 @@ static ir_node *create_Switch(ir_node *node) SET_IA32_ORIG_NODE(new_sel, ia32_get_old_node_name(env_cg, node)); } - res = new_rd_ia32_SwitchJmp(dbgi, irg, block, new_sel); - set_ia32_pncode(res, get_Cond_defaultProj(node)); - - SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node)); - - return res; -} - -static ir_node *get_flags_node(ir_node *node, pn_Cmp *pnc_out) -{ - ir_graph *irg = current_ir_graph; - ir_node *flags; - ir_node *new_op; - ir_node *noreg; - ir_node *nomem; - ir_node *new_block; - dbg_info *dbgi; - - /* we have a Cmp as input */ - if(is_Proj(node)) { - ir_node *pred = get_Proj_pred(node); - if(is_Cmp(pred)) { - flags = be_transform_node(pred); - *pnc_out = get_Proj_proj(node); - return flags; - } - } + new_node = new_rd_ia32_SwitchJmp(dbgi, irg, block, new_sel, default_pn); + SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node)); - /* a mode_b value, we have to compare it against 0 */ - dbgi = get_irn_dbg_info(node); - new_block = be_transform_node(get_nodes_block(node)); - new_op = be_transform_node(node); - noreg = ia32_new_NoReg_gp(env_cg); - nomem = new_NoMem(); - flags = new_rd_ia32_Test(dbgi, irg, new_block, noreg, noreg, nomem, - new_op, new_op, 0, 0); - *pnc_out = pn_Cmp_Lg; - return flags; + return new_node; } +/** + * Transform a Cond node. + */ static ir_node *gen_Cond(ir_node *node) { ir_node *block = get_nodes_block(node); ir_node *new_block = be_transform_node(block); @@ -2168,25 +2650,23 @@ static ir_node *gen_Cond(ir_node *node) { dbg_info *dbgi = get_irn_dbg_info(node); ir_node *sel = get_Cond_selector(node); ir_mode *sel_mode = get_irn_mode(sel); - ir_node *res; ir_node *flags = NULL; + ir_node *new_node; pn_Cmp pnc; if (sel_mode != mode_b) { return create_Switch(node); } - /* we get flags from a cmp */ + /* we get flags from a Cmp */ flags = get_flags_node(sel, &pnc); - res = new_rd_ia32_Jcc(dbgi, irg, new_block, flags, pnc); - SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node)); + new_node = new_rd_ia32_Jcc(dbgi, irg, new_block, flags, pnc); + SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node)); - return res; + return new_node; } - - /** * Transforms a CopyB node. * @@ -2215,16 +2695,13 @@ static ir_node *gen_CopyB(ir_node *node) { res = new_rd_ia32_Const(dbgi, irg, block, NULL, 0, size); add_irn_dep(res, get_irg_frame(irg)); - res = new_rd_ia32_CopyB(dbgi, irg, block, new_dst, new_src, res, new_mem); - /* we misuse the pncode field for the copyb size */ - set_ia32_pncode(res, rem); + res = new_rd_ia32_CopyB(dbgi, irg, block, new_dst, new_src, res, new_mem, rem); } else { if(size == 0) { ir_fprintf(stderr, "Optimisation warning copyb %+F with size <4\n", node); } - res = new_rd_ia32_CopyB_i(dbgi, irg, block, new_dst, new_src, new_mem); - set_ia32_pncode(res, size); + res = new_rd_ia32_CopyB_i(dbgi, irg, block, new_dst, new_src, new_mem, size); } SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node)); @@ -2234,92 +2711,14 @@ static ir_node *gen_CopyB(ir_node *node) { static ir_node *gen_be_Copy(ir_node *node) { - ir_node *result = be_duplicate_node(node); - ir_mode *mode = get_irn_mode(result); + ir_node *new_node = be_duplicate_node(node); + ir_mode *mode = get_irn_mode(new_node); if (mode_needs_gp_reg(mode)) { - set_irn_mode(result, mode_Iu); - } - - return result; -} - -/** - * helper function: checks wether all Cmp projs are Lg or Eq which is needed - * to fold an and into a test node - */ -static int can_fold_test_and(ir_node *node) -{ - const ir_edge_t *edge; - - /** we can only have eq and lg projs */ - foreach_out_edge(node, edge) { - ir_node *proj = get_edge_src_irn(edge); - pn_Cmp pnc = get_Proj_proj(proj); - if(pnc != pn_Cmp_Eq && pnc != pn_Cmp_Lg) - return 0; - } - - return 1; -} - -static ir_node *try_create_Test(ir_node *node) -{ - ir_graph *irg = current_ir_graph; - dbg_info *dbgi = get_irn_dbg_info(node); - ir_node *block = get_nodes_block(node); - ir_node *new_block = be_transform_node(block); - ir_node *cmp_left = get_Cmp_left(node); - ir_node *cmp_right = get_Cmp_right(node); - ir_mode *mode; - ir_node *left; - ir_node *right; - ir_node *res; - ia32_address_mode_t am; - ia32_address_t *addr = &am.addr; - int cmp_unsigned; - - /* can we use a test instruction? */ - if(!is_Const_0(cmp_right)) - return NULL; - - if(is_And(cmp_left) && get_irn_n_edges(cmp_left) == 1 && - can_fold_test_and(node)) { - ir_node *and_left = get_And_left(cmp_left); - ir_node *and_right = get_And_right(cmp_left); - - mode = get_irn_mode(and_left); - left = and_left; - right = and_right; - } else { - mode = get_irn_mode(cmp_left); - left = cmp_left; - right = cmp_left; - } - - assert(get_mode_size_bits(mode) <= 32); - - match_arguments(&am, block, left, right, match_commutative | - match_8_bit_am | match_16_bit_am | match_am_and_immediates); - - cmp_unsigned = !mode_is_signed(mode); - if(get_mode_size_bits(mode) == 8) { - res = new_rd_ia32_Test8Bit(dbgi, irg, new_block, addr->base, - addr->index, addr->mem, am.new_op1, - am.new_op2, am.ins_permuted, cmp_unsigned); - } else { - res = new_rd_ia32_Test(dbgi, irg, new_block, addr->base, addr->index, - addr->mem, am.new_op1, am.new_op2, - am.ins_permuted, cmp_unsigned); + set_irn_mode(new_node, mode_Iu); } - set_am_attributes(res, &am); - assert(mode != NULL); - set_ia32_ls_mode(res, mode); - SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node)); - - res = fix_mem_proj(res, &am); - return res; + return new_node; } static ir_node *create_Fucom(ir_node *node) @@ -2332,31 +2731,33 @@ static ir_node *create_Fucom(ir_node *node) ir_node *new_left = be_transform_node(left); ir_node *right = get_Cmp_right(node); ir_node *new_right; - ir_node *res; + ir_node *new_node; - if(transform_config.use_fucomi) { + if(ia32_cg_config.use_fucomi) { new_right = be_transform_node(right); - res = new_rd_ia32_vFucomi(dbgi, irg, new_block, new_left, new_right, 0); - set_ia32_commutative(res); - SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node)); + new_node = new_rd_ia32_vFucomi(dbgi, irg, new_block, new_left, + new_right, 0); + set_ia32_commutative(new_node); + SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node)); } else { - if(transform_config.use_ftst && is_Const_null(right)) { - res = new_rd_ia32_vFtstFnstsw(dbgi, irg, new_block, new_left, 0); + if(ia32_cg_config.use_ftst && is_Const_0(right)) { + new_node = new_rd_ia32_vFtstFnstsw(dbgi, irg, new_block, new_left, + 0); } else { new_right = be_transform_node(right); - res = new_rd_ia32_vFucomFnstsw(dbgi, irg, new_block, new_left, + new_node = new_rd_ia32_vFucomFnstsw(dbgi, irg, new_block, new_left, new_right, 0); } - set_ia32_commutative(res); + set_ia32_commutative(new_node); - SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node)); + SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node)); - res = new_rd_ia32_Sahf(dbgi, irg, new_block, res); - SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node)); + new_node = new_rd_ia32_Sahf(dbgi, irg, new_block, new_node); + SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node)); } - return res; + return new_node; } static ir_node *create_Ucomi(ir_node *node) @@ -2371,7 +2772,8 @@ static ir_node *create_Ucomi(ir_node *node) ia32_address_mode_t am; ia32_address_t *addr = &am.addr; - match_arguments(&am, src_block, left, right, match_commutative); + match_arguments(&am, src_block, left, right, NULL, + match_commutative | match_am); new_node = new_rd_ia32_Ucomi(dbgi, irg, new_block, addr->base, addr->index, addr->mem, am.new_op1, am.new_op2, @@ -2385,6 +2787,28 @@ static ir_node *create_Ucomi(ir_node *node) return new_node; } +/** + * helper function: checks wether all Cmp projs are Lg or Eq which is needed + * to fold an and into a test node + */ +static int can_fold_test_and(ir_node *node) +{ + const ir_edge_t *edge; + + /** we can only have eq and lg projs */ + foreach_out_edge(node, edge) { + ir_node *proj = get_edge_src_irn(edge); + pn_Cmp pnc = get_Proj_proj(proj); + if(pnc != pn_Cmp_Eq && pnc != pn_Cmp_Lg) + return 0; + } + + return 1; +} + +/** + * Generate code for a Cmp. + */ static ir_node *gen_Cmp(ir_node *node) { ir_graph *irg = current_ir_graph; @@ -2394,13 +2818,13 @@ static ir_node *gen_Cmp(ir_node *node) ir_node *left = get_Cmp_left(node); ir_node *right = get_Cmp_right(node); ir_mode *cmp_mode = get_irn_mode(left); - ir_node *res; + ir_node *new_node; ia32_address_mode_t am; ia32_address_t *addr = &am.addr; int cmp_unsigned; if(mode_is_float(cmp_mode)) { - if (USE_SSE2(env_cg)) { + if (ia32_cg_config.use_sse2) { return create_Ucomi(node); } else { return create_Fucom(node); @@ -2411,39 +2835,93 @@ static ir_node *gen_Cmp(ir_node *node) /* we prefer the Test instruction where possible except cases where * we can use SourceAM */ - if(!ia32_use_source_address_mode(block, left, right) && - !ia32_use_source_address_mode(block, right, left)) { - res = try_create_Test(node); - if(res != NULL) - return res; - } - - match_arguments(&am, block, left, right, - match_commutative | match_8_bit_am | match_16_bit_am | - match_am_and_immediates); - - cmp_unsigned = !mode_is_signed(get_irn_mode(left)); - if(get_mode_size_bits(cmp_mode) == 8) { - res = new_rd_ia32_Cmp8Bit(dbgi, irg, new_block, addr->base, addr->index, - addr->mem, am.new_op1, am.new_op2, - am.ins_permuted, cmp_unsigned); + cmp_unsigned = !mode_is_signed(cmp_mode); + if (is_Const_0(right)) { + if (is_And(left) && + get_irn_n_edges(left) == 1 && + can_fold_test_and(node)) { + /* Test(and_left, and_right) */ + ir_node *and_left = get_And_left(left); + ir_node *and_right = get_And_right(left); + ir_mode *mode = get_irn_mode(and_left); + + match_arguments(&am, block, and_left, and_right, NULL, + match_commutative | + match_am | match_8bit_am | match_16bit_am | + match_am_and_immediates | match_immediate | + match_8bit | match_16bit); + if (get_mode_size_bits(mode) == 8) { + new_node = new_rd_ia32_Test8Bit(dbgi, irg, new_block, addr->base, + addr->index, addr->mem, am.new_op1, + am.new_op2, am.ins_permuted, + cmp_unsigned); + } else { + new_node = new_rd_ia32_Test(dbgi, irg, new_block, addr->base, + addr->index, addr->mem, am.new_op1, + am.new_op2, am.ins_permuted, cmp_unsigned); + } + } else { + match_arguments(&am, block, NULL, left, NULL, + match_am | match_8bit_am | match_16bit_am | + match_8bit | match_16bit); + if (am.op_type == ia32_AddrModeS) { + /* Cmp(AM, 0) */ + ir_node *imm_zero = try_create_Immediate(right, 0); + if (get_mode_size_bits(cmp_mode) == 8) { + new_node = new_rd_ia32_Cmp8Bit(dbgi, irg, new_block, addr->base, + addr->index, addr->mem, am.new_op2, + imm_zero, am.ins_permuted, + cmp_unsigned); + } else { + new_node = new_rd_ia32_Cmp(dbgi, irg, new_block, addr->base, + addr->index, addr->mem, am.new_op2, + imm_zero, am.ins_permuted, cmp_unsigned); + } + } else { + /* Test(left, left) */ + if (get_mode_size_bits(cmp_mode) == 8) { + new_node = new_rd_ia32_Test8Bit(dbgi, irg, new_block, addr->base, + addr->index, addr->mem, am.new_op2, + am.new_op2, am.ins_permuted, + cmp_unsigned); + } else { + new_node = new_rd_ia32_Test(dbgi, irg, new_block, addr->base, + addr->index, addr->mem, am.new_op2, + am.new_op2, am.ins_permuted, + cmp_unsigned); + } + } + } } else { - res = new_rd_ia32_Cmp(dbgi, irg, new_block, addr->base, addr->index, - addr->mem, am.new_op1, am.new_op2, - am.ins_permuted, cmp_unsigned); + /* Cmp(left, right) */ + match_arguments(&am, block, left, right, NULL, + match_commutative | match_am | match_8bit_am | + match_16bit_am | match_am_and_immediates | + match_immediate | match_8bit | match_16bit); + if (get_mode_size_bits(cmp_mode) == 8) { + new_node = new_rd_ia32_Cmp8Bit(dbgi, irg, new_block, addr->base, + addr->index, addr->mem, am.new_op1, + am.new_op2, am.ins_permuted, + cmp_unsigned); + } else { + new_node = new_rd_ia32_Cmp(dbgi, irg, new_block, addr->base, + addr->index, addr->mem, am.new_op1, + am.new_op2, am.ins_permuted, cmp_unsigned); + } } - set_am_attributes(res, &am); + set_am_attributes(new_node, &am); assert(cmp_mode != NULL); - set_ia32_ls_mode(res, cmp_mode); + set_ia32_ls_mode(new_node, cmp_mode); - SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node)); + SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node)); - res = fix_mem_proj(res, &am); + new_node = fix_mem_proj(new_node, &am); - return res; + return new_node; } -static ir_node *create_CMov(ir_node *node, ir_node *new_flags, pn_Cmp pnc) +static ir_node *create_CMov(ir_node *node, ir_node *flags, ir_node *new_flags, + pn_Cmp pnc) { ir_graph *irg = current_ir_graph; dbg_info *dbgi = get_irn_dbg_info(node); @@ -2456,15 +2934,15 @@ static ir_node *create_CMov(ir_node *node, ir_node *new_flags, pn_Cmp pnc) ia32_address_mode_t am; ia32_address_t *addr; - assert(transform_config.use_cmov); + assert(ia32_cg_config.use_cmov); assert(mode_needs_gp_reg(get_irn_mode(val_true))); addr = &am.addr; - match_flags = match_commutative | match_no_immediate | match_16_bit_am - | match_force_32bit_op; + match_flags = match_commutative | match_am | match_16bit_am | + match_mode_neutral; - match_arguments(&am, block, val_false, val_true, match_flags); + match_arguments(&am, block, val_false, val_true, flags, match_flags); new_node = new_rd_ia32_CMov(dbgi, irg, new_block, addr->base, addr->index, addr->mem, am.new_op1, am.new_op2, new_flags, @@ -2478,8 +2956,9 @@ static ir_node *create_CMov(ir_node *node, ir_node *new_flags, pn_Cmp pnc) return new_node; } - - +/** + * Creates a ia32 Setcc instruction. + */ static ir_node *create_set_32bit(dbg_info *dbgi, ir_node *new_block, ir_node *flags, pn_Cmp pnc, ir_node *orig_node, int ins_permuted) @@ -2487,16 +2966,55 @@ static ir_node *create_set_32bit(dbg_info *dbgi, ir_node *new_block, ir_graph *irg = current_ir_graph; ir_node *noreg = ia32_new_NoReg_gp(env_cg); ir_node *nomem = new_NoMem(); - ir_node *res; + ir_mode *mode = get_irn_mode(orig_node); + ir_node *new_node; + + new_node = new_rd_ia32_Set(dbgi, irg, new_block, flags, pnc, ins_permuted); + SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, orig_node)); + + /* we might need to conv the result up */ + if (get_mode_size_bits(mode) > 8) { + new_node = new_rd_ia32_Conv_I2I8Bit(dbgi, irg, new_block, noreg, noreg, + nomem, new_node, mode_Bu); + SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, orig_node)); + } - res = new_rd_ia32_Set(dbgi, irg, new_block, flags, pnc, ins_permuted); - SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, orig_node)); - res = new_rd_ia32_Conv_I2I8Bit(dbgi, irg, new_block, noreg, noreg, - nomem, res, mode_Bu); - SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, orig_node)); - (void) orig_node; + return new_node; +} + +/** + * Create instruction for an unsigned Difference or Zero. + */ +static ir_node *create_Doz(ir_node *psi, ir_node *a, ir_node *b) { + ir_graph *irg = current_ir_graph; + ir_mode *mode = get_irn_mode(psi); + ir_node *new_node, *sub, *sbb, *eflags, *block, *noreg, *tmpreg, *nomem; + dbg_info *dbgi; + + new_node = gen_binop(psi, a, b, new_rd_ia32_Sub, + match_mode_neutral | match_am | match_immediate | match_two_users); + + block = get_nodes_block(new_node); + + if (is_Proj(new_node)) { + sub = get_Proj_pred(new_node); + assert(is_ia32_Sub(sub)); + } else { + sub = new_node; + set_irn_mode(sub, mode_T); + new_node = new_rd_Proj(NULL, irg, block, sub, mode, pn_ia32_res); + } + eflags = new_rd_Proj(NULL, irg, block, sub, mode_Iu, pn_ia32_Sub_flags); + + dbgi = get_irn_dbg_info(psi); + noreg = ia32_new_NoReg_gp(env_cg); + tmpreg = new_rd_ia32_ProduceVal(dbgi, irg, block); + nomem = new_NoMem(); + sbb = new_rd_ia32_Sbb(dbgi, irg, block, noreg, noreg, nomem, tmpreg, tmpreg, eflags); - return res; + new_node = new_rd_ia32_And(dbgi, irg, block, noreg, noreg, nomem, new_node, sbb); + set_ia32_commutative(new_node); + return new_node; } /** @@ -2512,24 +3030,79 @@ static ir_node *gen_Psi(ir_node *node) ir_node *psi_true = get_Psi_val(node, 0); ir_node *psi_default = get_Psi_default(node); ir_node *cond = get_Psi_cond(node, 0); - ir_node *flags = NULL; - ir_node *res; - pn_Cmp pnc; + ir_mode *mode = get_irn_mode(node); + ir_node *cmp = get_Proj_pred(cond); + ir_node *cmp_left = get_Cmp_left(cmp); + ir_node *cmp_right = get_Cmp_right(cmp); + pn_Cmp pnc = get_Proj_proj(cond); assert(get_Psi_n_conds(node) == 1); assert(get_irn_mode(cond) == mode_b); - assert(mode_needs_gp_reg(get_irn_mode(node))); - flags = get_flags_node(cond, &pnc); + /* Note: a Psi node uses a Load two times IFF it's used in the compare AND in the result */ + if (mode_is_float(mode)) { + if (ia32_cg_config.use_sse2) { + if (pnc == pn_Cmp_Lt || pnc == pn_Cmp_Le) { + if (cmp_left == psi_true && cmp_right == psi_default) { + /* psi(a <= b, a, b) => MIN */ + return gen_binop(node, cmp_left, cmp_right, new_rd_ia32_xMin, + match_commutative | match_am | match_two_users); + } else if (cmp_left == psi_default && cmp_right == psi_true) { + /* psi(a <= b, b, a) => MAX */ + return gen_binop(node, cmp_left, cmp_right, new_rd_ia32_xMax, + match_commutative | match_am | match_two_users); + } + } else if (pnc == pn_Cmp_Gt || pnc == pn_Cmp_Ge) { + if (cmp_left == psi_true && cmp_right == psi_default) { + /* psi(a >= b, a, b) => MAX */ + return gen_binop(node, cmp_left, cmp_right, new_rd_ia32_xMax, + match_commutative | match_am | match_two_users); + } else if (cmp_left == psi_default && cmp_right == psi_true) { + /* psi(a >= b, b, a) => MIN */ + return gen_binop(node, cmp_left, cmp_right, new_rd_ia32_xMin, + match_commutative | match_am | match_two_users); + } + } + } + panic("cannot transform floating point Psi"); - if(is_Const_1(psi_true) && is_Const_0(psi_default)) { - res = create_set_32bit(dbgi, new_block, flags, pnc, node, 0); - } else if(is_Const_0(psi_true) && is_Const_1(psi_default)) { - res = create_set_32bit(dbgi, new_block, flags, pnc, node, 1); } else { - res = create_CMov(node, flags, pnc); + ir_node *flags; + ir_node *new_node; + + assert(mode_needs_gp_reg(mode)); + + /* check for unsigned Doz first */ + if ((pnc & pn_Cmp_Gt) && !mode_is_signed(mode) && + is_Const_0(psi_default) && is_Sub(psi_true) && + get_Sub_left(psi_true) == cmp_left && get_Sub_right(psi_true) == cmp_right) { + /* Psi(a >=u b, a - b, 0) unsigned Doz */ + return create_Doz(node, cmp_left, cmp_right); + } else if ((pnc & pn_Cmp_Lt) && !mode_is_signed(mode) && + is_Const_0(psi_true) && is_Sub(psi_default) && + get_Sub_left(psi_default) == cmp_left && get_Sub_right(psi_default) == cmp_right) { + /* Psi(a <=u b, 0, a - b) unsigned Doz */ + return create_Doz(node, cmp_left, cmp_right); + } + + flags = get_flags_node(cond, &pnc); + + if (is_Const(psi_true) && is_Const(psi_default)) { + /* both are const, good */ + if (is_Const_1(psi_true) && is_Const_0(psi_default)) { + new_node = create_set_32bit(dbgi, new_block, flags, pnc, node, /*is_premuted=*/0); + } else if (is_Const_0(psi_true) && is_Const_1(psi_default)) { + new_node = create_set_32bit(dbgi, new_block, flags, pnc, node, /*is_premuted=*/1); + } else { + /* Not that simple. */ + goto need_cmov; + } + } else { +need_cmov: + new_node = create_CMov(node, cond, flags, pnc); + } + return new_node; } - return res; } @@ -2544,14 +3117,10 @@ static ir_node *gen_x87_fp_to_gp(ir_node *node) { ir_graph *irg = current_ir_graph; dbg_info *dbgi = get_irn_dbg_info(node); ir_node *noreg = ia32_new_NoReg_gp(cg); - ir_node *trunc_mode = ia32_new_Fpu_truncate(cg); ir_mode *mode = get_irn_mode(node); - ir_node *fist, *load; - - /* do a fist */ - fist = new_rd_ia32_vfist(dbgi, irg, block, get_irg_frame(irg), noreg, - new_NoMem(), new_op, trunc_mode); + ir_node *fist, *load, *mem; + mem = gen_vfist(dbgi, irg, block, get_irg_frame(irg), noreg, new_NoMem(), new_op, &fist); set_irn_pinned(fist, op_pin_state_floats); set_ia32_use_frame(fist); set_ia32_op_type(fist, ia32_AddrModeD); @@ -2567,7 +3136,7 @@ static ir_node *gen_x87_fp_to_gp(ir_node *node) { SET_IA32_ORIG_NODE(fist, ia32_get_old_node_name(cg, node)); /* do a Load */ - load = new_rd_ia32_Load(dbgi, irg, block, get_irg_frame(irg), noreg, fist); + load = new_rd_ia32_Load(dbgi, irg, block, get_irg_frame(irg), noreg, mem); set_irn_pinned(load, op_pin_state_floats); set_ia32_use_frame(load); @@ -2597,7 +3166,7 @@ static ir_node *gen_x87_strict_conv(ir_mode *tgt_mode, ir_node *node) ir_node *nomem = new_NoMem(); ir_node *frame = get_irg_frame(irg); ir_node *store, *load; - ir_node *res; + ir_node *new_node; store = new_rd_ia32_vfst(dbgi, irg, block, frame, noreg, nomem, node, tgt_mode); @@ -2611,60 +3180,52 @@ static ir_node *gen_x87_strict_conv(ir_mode *tgt_mode, ir_node *node) set_ia32_op_type(load, ia32_AddrModeS); SET_IA32_ORIG_NODE(load, ia32_get_old_node_name(env_cg, node)); - res = new_r_Proj(irg, block, load, mode_E, pn_ia32_vfld_res); - return res; -} - -static ir_node *create_Immediate(ir_entity *symconst, int symconst_sign, long 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, - symconst, symconst_sign, val); - arch_set_irn_register(env_cg->arch_env, immediate, &ia32_gp_regs[REG_GP_NOREG]); - - return immediate; + new_node = new_r_Proj(irg, block, load, mode_E, pn_ia32_vfld_res); + return new_node; } /** * Create a conversion from general purpose to x87 register */ static ir_node *gen_x87_gp_to_fp(ir_node *node, ir_mode *src_mode) { - ir_node *src_block = get_nodes_block(node); - ir_node *block = be_transform_node(src_block); - ir_graph *irg = current_ir_graph; - dbg_info *dbgi = get_irn_dbg_info(node); - ir_node *op = get_Conv_op(node); - ir_node *new_op; + ir_node *src_block = get_nodes_block(node); + ir_node *block = be_transform_node(src_block); + ir_graph *irg = current_ir_graph; + dbg_info *dbgi = get_irn_dbg_info(node); + ir_node *op = get_Conv_op(node); + ir_node *new_op = NULL; ir_node *noreg; ir_node *nomem; ir_mode *mode; ir_mode *store_mode; ir_node *fild; ir_node *store; - ir_node *res; + ir_node *new_node; int src_bits; /* fild can use source AM if the operand is a signed 32bit integer */ if (src_mode == mode_Is) { ia32_address_mode_t am; - match_arguments(&am, src_block, NULL, op, match_no_immediate); + match_arguments(&am, src_block, NULL, op, NULL, + match_am | match_try_am); if (am.op_type == ia32_AddrModeS) { ia32_address_t *addr = &am.addr; - fild = new_rd_ia32_vfild(dbgi, irg, block, addr->base, addr->index, addr->mem); - res = new_r_Proj(irg, block, fild, mode_vfp, pn_ia32_vfild_res); + fild = new_rd_ia32_vfild(dbgi, irg, block, addr->base, + addr->index, addr->mem); + new_node = new_r_Proj(irg, block, fild, mode_vfp, + pn_ia32_vfild_res); set_am_attributes(fild, &am); SET_IA32_ORIG_NODE(fild, ia32_get_old_node_name(env_cg, node)); fix_mem_proj(fild, &am); - return res; + return new_node; } - new_op = am.new_op2; - } else { + } + if(new_op == NULL) { new_op = be_transform_node(op); } @@ -2727,9 +3288,9 @@ static ir_node *gen_x87_gp_to_fp(ir_node *node, ir_mode *src_mode) { set_ia32_op_type(fild, ia32_AddrModeS); set_ia32_ls_mode(fild, store_mode); - res = new_r_Proj(irg, block, fild, mode_vfp, pn_ia32_vfild_res); + new_node = new_r_Proj(irg, block, fild, mode_vfp, pn_ia32_vfild_res); - return res; + return new_node; } /** @@ -2743,12 +3304,13 @@ static ir_node *create_I2I_Conv(ir_mode *src_mode, ir_mode *tgt_mode, int src_bits = get_mode_size_bits(src_mode); int tgt_bits = get_mode_size_bits(tgt_mode); ir_node *new_block = be_transform_node(block); - ir_node *res; + ir_node *new_node; ir_mode *smaller_mode; int smaller_bits; ia32_address_mode_t am; ia32_address_t *addr = &am.addr; + (void) node; if (src_bits < tgt_bits) { smaller_mode = src_mode; smaller_bits = src_bits; @@ -2757,21 +3319,32 @@ static ir_node *create_I2I_Conv(ir_mode *src_mode, ir_mode *tgt_mode, smaller_bits = tgt_bits; } - match_arguments(&am, block, NULL, op, match_8_bit_am | match_16_bit_am); - if (smaller_bits == 8 && am.op_type == ia32_Normal) { - res = new_rd_ia32_Conv_I2I8Bit(dbgi, irg, new_block, addr->base, - addr->index, addr->mem, am.new_op2, - smaller_mode); +#ifdef DEBUG_libfirm + if(is_Const(op)) { + ir_fprintf(stderr, "Optimisation warning: conv after constant %+F\n", + op); + } +#endif + + match_arguments(&am, block, NULL, op, NULL, + match_8bit | match_16bit | + match_am | match_8bit_am | match_16bit_am); + if (smaller_bits == 8) { + new_node = new_rd_ia32_Conv_I2I8Bit(dbgi, irg, new_block, addr->base, + addr->index, addr->mem, am.new_op2, + smaller_mode); } else { - res = new_rd_ia32_Conv_I2I(dbgi, irg, new_block, addr->base, - addr->index, addr->mem, am.new_op2, - smaller_mode); + new_node = new_rd_ia32_Conv_I2I(dbgi, irg, new_block, addr->base, + addr->index, addr->mem, am.new_op2, + smaller_mode); } - set_am_attributes(res, &am); - set_ia32_ls_mode(res, smaller_mode); - SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node)); - res = fix_mem_proj(res, &am); - return res; + set_am_attributes(new_node, &am); + /* match_arguments assume that out-mode = in-mode, this isn't true here + * so fix it */ + set_ia32_ls_mode(new_node, smaller_mode); + SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node)); + new_node = fix_mem_proj(new_node, &am); + return new_node; } /** @@ -2795,14 +3368,14 @@ static ir_node *gen_Conv(ir_node *node) { ir_node *res = NULL; if (src_mode == mode_b) { - assert(mode_is_int(tgt_mode)); + assert(mode_is_int(tgt_mode) || mode_is_reference(tgt_mode)); /* nothing to do, we already model bools as 0/1 ints */ return be_transform_node(op); } if (src_mode == tgt_mode) { if (get_Conv_strict(node)) { - if (USE_SSE2(env_cg)) { + if (ia32_cg_config.use_sse2) { /* when we are in SSE mode, we can kill all strict no-op conversion */ return be_transform_node(op); } @@ -2825,7 +3398,7 @@ static ir_node *gen_Conv(ir_node *node) { } /* ... to float */ - if (USE_SSE2(env_cg)) { + if (ia32_cg_config.use_sse2) { DB((dbg, LEVEL_1, "create Conv(float, float) ...")); res = new_rd_ia32_Conv_FP2FP(dbgi, irg, new_block, noreg, noreg, nomem, new_op); @@ -2842,7 +3415,7 @@ static ir_node *gen_Conv(ir_node *node) { } else { /* ... to int */ DB((dbg, LEVEL_1, "create Conv(float, int) ...")); - if (USE_SSE2(env_cg)) { + if (ia32_cg_config.use_sse2) { res = new_rd_ia32_Conv_FP2I(dbgi, irg, new_block, noreg, noreg, nomem, new_op); set_ia32_ls_mode(res, src_mode); @@ -2855,7 +3428,7 @@ static ir_node *gen_Conv(ir_node *node) { if (mode_is_float(tgt_mode)) { /* ... to float */ DB((dbg, LEVEL_1, "create Conv(int, float) ...")); - if (USE_SSE2(env_cg)) { + if (ia32_cg_config.use_sse2) { new_op = be_transform_node(op); res = new_rd_ia32_Conv_I2FP(dbgi, irg, new_block, noreg, noreg, nomem, new_op); @@ -2928,7 +3501,7 @@ static ir_node *try_create_Immediate(ir_node *node, ir_mode *mode; ir_node *cnst = NULL; ir_node *symconst = NULL; - ir_node *res; + ir_node *new_node; mode = get_irn_mode(node); if(!mode_is_int(mode) && !mode_is_reference(mode)) { @@ -3014,9 +3587,9 @@ static ir_node *try_create_Immediate(ir_node *node, offset = tarval_neg(offset); } - res = create_Immediate(symconst_ent, symconst_sign, val); + new_node = create_Immediate(symconst_ent, symconst_sign, val); - return res; + return new_node; } static ir_node *create_immediate_or_transform(ir_node *node, @@ -3033,8 +3606,8 @@ static const arch_register_req_t no_register_req = { arch_register_req_type_none, NULL, /* regclass */ NULL, /* limit bitset */ - { -1, -1 }, /* same pos */ - -1 /* different pos */ + 0, /* same pos */ + 0 /* different pos */ }; /** @@ -3060,7 +3633,7 @@ static void parse_asm_constraint(int pos, constraint_t *constraint, const char * ir_graph *irg = current_ir_graph; struct obstack *obst = get_irg_obstack(irg); arch_register_req_t *req; - unsigned *limited_ptr; + unsigned *limited_ptr = NULL; int p; int same_as = -1; @@ -3070,7 +3643,7 @@ static void parse_asm_constraint(int pos, constraint_t *constraint, const char * /* a memory constraint: no need to do anything in backend about it * (the dependencies are already respected by the memory edge of * the node) */ - constraint->req = &no_register_req; + constraint->req = &no_register_req; return; } @@ -3252,9 +3825,8 @@ static void parse_asm_constraint(int pos, constraint_t *constraint, const char * req->cls = other_constr->cls; req->type = arch_register_req_type_should_be_same; req->limited = NULL; - req->other_same[0] = pos; - req->other_same[1] = -1; - req->other_different = -1; + req->other_same = 1U << pos; + req->other_different = 0; /* switch constraints. This is because in firm we have same_as * constraints on the output constraints while in the gcc asm syntax @@ -3300,13 +3872,53 @@ static void parse_asm_constraint(int pos, constraint_t *constraint, const char * } static void parse_clobber(ir_node *node, int pos, constraint_t *constraint, - const char *c) + const char *clobber) { - (void) node; + ir_graph *irg = get_irn_irg(node); + struct obstack *obst = get_irg_obstack(irg); + const arch_register_t *reg = NULL; + int c; + size_t r; + arch_register_req_t *req; + const arch_register_class_t *cls; + unsigned *limited; + (void) pos; - (void) constraint; - (void) c; - panic("Clobbers not supported yet"); + + /* TODO: construct a hashmap instead of doing linear search for clobber + * register */ + for(c = 0; c < N_CLASSES; ++c) { + cls = & ia32_reg_classes[c]; + for(r = 0; r < cls->n_regs; ++r) { + const arch_register_t *temp_reg = arch_register_for_index(cls, r); + if(strcmp(temp_reg->name, clobber) == 0 + || (c == CLASS_ia32_gp && strcmp(temp_reg->name+1, clobber) == 0)) { + reg = temp_reg; + break; + } + } + if(reg != NULL) + break; + } + if(reg == NULL) { + panic("Register '%s' mentioned in asm clobber is unknown\n", clobber); + return; + } + + assert(reg->index < 32); + + limited = obstack_alloc(obst, sizeof(limited[0])); + *limited = 1 << reg->index; + + req = obstack_alloc(obst, sizeof(req[0])); + memset(req, 0, sizeof(req[0])); + req->type = arch_register_req_type_limited; + req->cls = cls; + req->limited = limited; + + constraint->req = req; + constraint->immediate_possible = 0; + constraint->immediate_type = 0; } static int is_memory_op(const ir_asm_constraint *constraint) @@ -3334,7 +3946,7 @@ static ir_node *gen_ASM(ir_node *node) ir_node *new_block = be_transform_node(block); dbg_info *dbgi = get_irn_dbg_info(node); ir_node **in; - ir_node *res; + ir_node *new_node; int out_arity; int n_out_constraints; int n_clobbers; @@ -3355,6 +3967,9 @@ static ir_node *gen_ASM(ir_node *node) n_out_constraints = get_ASM_n_output_constraints(node); n_clobbers = get_ASM_n_clobbers(node); out_arity = n_out_constraints + n_clobbers; + /* hack to keep space for mem proj */ + if(n_clobbers > 0) + out_arity += 1; in_constraints = get_ASM_input_constraints(node); out_constraints = get_ASM_output_constraints(node); @@ -3377,14 +3992,19 @@ static ir_node *gen_ASM(ir_node *node) if(constraint->pos > reg_map_size) reg_map_size = constraint->pos; - } else { + + out_reg_reqs[i] = parsed_constraint.req; + } else if(i < out_arity - 1) { ident *glob_id = clobbers [i - n_out_constraints]; + assert(glob_id != NULL); c = get_id_str(glob_id); parse_clobber(node, i, &parsed_constraint, c); - } - out_reg_reqs[i] = parsed_constraint.req; + out_reg_reqs[i+1] = parsed_constraint.req; + } } + if(n_clobbers > 1) + out_reg_reqs[n_out_constraints] = &no_register_req; /* construct input constraints */ in_reg_reqs = obstack_alloc(obst, arity * sizeof(in_reg_reqs[0])); @@ -3448,27 +4068,17 @@ static ir_node *gen_ASM(ir_node *node) in[i] = transformed; } - res = new_rd_ia32_Asm(dbgi, irg, new_block, arity, in, out_arity, - get_ASM_text(node), register_map); + new_node = new_rd_ia32_Asm(dbgi, irg, new_block, arity, in, out_arity, + get_ASM_text(node), register_map); - set_ia32_out_req_all(res, out_reg_reqs); - set_ia32_in_req_all(res, in_reg_reqs); + set_ia32_out_req_all(new_node, out_reg_reqs); + set_ia32_in_req_all(new_node, in_reg_reqs); - SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node)); + SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node)); - return res; + return new_node; } -/******************************************** - * _ _ - * | | | | - * | |__ ___ _ __ ___ __| | ___ ___ - * | '_ \ / _ \ '_ \ / _ \ / _` |/ _ \/ __| - * | |_) | __/ | | | (_) | (_| | __/\__ \ - * |_.__/ \___|_| |_|\___/ \__,_|\___||___/ - * - ********************************************/ - /** * Transforms a FrameAddr into an ia32 Add. */ @@ -3479,15 +4089,15 @@ static ir_node *gen_be_FrameAddr(ir_node *node) { ir_graph *irg = current_ir_graph; dbg_info *dbgi = get_irn_dbg_info(node); ir_node *noreg = ia32_new_NoReg_gp(env_cg); - ir_node *res; + ir_node *new_node; - res = new_rd_ia32_Lea(dbgi, irg, block, new_op, noreg); - set_ia32_frame_ent(res, arch_get_frame_entity(env_cg->arch_env, node)); - set_ia32_use_frame(res); + new_node = new_rd_ia32_Lea(dbgi, irg, block, new_op, noreg); + set_ia32_frame_ent(new_node, arch_get_frame_entity(env_cg->arch_env, node)); + set_ia32_use_frame(new_node); - SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node)); + SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node)); - return res; + return new_node; } /** @@ -3510,7 +4120,7 @@ static ir_node *gen_be_Return(ir_node *node) { int pn_ret_val, pn_ret_mem, arity, i; assert(ret_val != NULL); - if (be_Return_get_n_rets(node) < 1 || ! USE_SSE2(env_cg)) { + if (be_Return_get_n_rets(node) < 1 || ! ia32_cg_config.use_sse2) { return be_duplicate_node(node); } @@ -3597,30 +4207,10 @@ static ir_node *gen_be_Return(ir_node *node) { */ static ir_node *gen_be_AddSP(ir_node *node) { - ir_node *src_block = get_nodes_block(node); - ir_node *new_block = be_transform_node(src_block); - ir_node *sz = get_irn_n(node, be_pos_AddSP_size); - ir_node *sp = get_irn_n(node, be_pos_AddSP_old_sp); - ir_graph *irg = current_ir_graph; - dbg_info *dbgi = get_irn_dbg_info(node); - ir_node *new_node; - ia32_address_mode_t am; - ia32_address_t *addr = &am.addr; - match_flags_t flags = 0; - - match_arguments(&am, src_block, sp, sz, flags); - - new_node = new_rd_ia32_SubSP(dbgi, irg, new_block, addr->base, addr->index, - addr->mem, am.new_op1, am.new_op2); - set_am_attributes(new_node, &am); - /* we can't use source address mode anymore when using immediates */ - if(is_ia32_Immediate(am.new_op1) || is_ia32_Immediate(am.new_op2)) - set_ia32_am_support(new_node, ia32_am_None, ia32_am_arity_none); - SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node)); - - new_node = fix_mem_proj(new_node, &am); + ir_node *sz = get_irn_n(node, be_pos_AddSP_size); + ir_node *sp = get_irn_n(node, be_pos_AddSP_old_sp); - return new_node; + return gen_binop(node, sp, sz, new_rd_ia32_SubSP, match_am); } /** @@ -3628,30 +4218,10 @@ static ir_node *gen_be_AddSP(ir_node *node) */ static ir_node *gen_be_SubSP(ir_node *node) { - ir_node *src_block = get_nodes_block(node); - ir_node *new_block = be_transform_node(src_block); - ir_node *sz = get_irn_n(node, be_pos_SubSP_size); - ir_node *sp = get_irn_n(node, be_pos_SubSP_old_sp); - ir_graph *irg = current_ir_graph; - dbg_info *dbgi = get_irn_dbg_info(node); - ir_node *new_node; - ia32_address_mode_t am; - ia32_address_t *addr = &am.addr; - match_flags_t flags = 0; - - match_arguments(&am, src_block, sp, sz, flags); - - new_node = new_rd_ia32_AddSP(dbgi, irg, new_block, addr->base, addr->index, - addr->mem, am.new_op1, am.new_op2); - set_am_attributes(new_node, &am); - /* we can't use source address mode anymore when using immediates */ - if(is_ia32_Immediate(am.new_op1) || is_ia32_Immediate(am.new_op2)) - set_ia32_am_support(new_node, ia32_am_None, ia32_am_arity_none); - SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node)); - - new_node = fix_mem_proj(new_node, &am); + ir_node *sz = get_irn_n(node, be_pos_SubSP_size); + ir_node *sp = get_irn_n(node, be_pos_SubSP_old_sp); - return new_node; + return gen_binop(node, sp, sz, new_rd_ia32_AddSP, match_am); } /** @@ -3663,21 +4233,28 @@ static ir_node *gen_Unknown(ir_node *node) { ir_mode *mode = get_irn_mode(node); if (mode_is_float(mode)) { - if (USE_SSE2(env_cg)) { + if (ia32_cg_config.use_sse2) { return ia32_new_Unknown_xmm(env_cg); } else { - /* Unknown nodes are buggy in x87 sim, use zero for now... */ + /* Unknown nodes are buggy in x87 simulator, use zero for now... */ ir_graph *irg = current_ir_graph; dbg_info *dbgi = get_irn_dbg_info(node); ir_node *block = get_irg_start_block(irg); - return new_rd_ia32_vfldz(dbgi, irg, block); + ir_node *ret = new_rd_ia32_vfldz(dbgi, irg, block); + + /* Const Nodes before the initial IncSP are a bad idea, because + * they could be spilled and we have no SP ready at that point yet. + * So add a dependency to the initial frame pointer calculation to + * avoid that situation. + */ + add_irn_dep(ret, get_irg_frame(irg)); + return ret; } } else if (mode_needs_gp_reg(mode)) { return ia32_new_Unknown_gp(env_cg); } else { - assert(0 && "unsupported Unknown-Mode"); + panic("unsupported Unknown-Mode"); } - return NULL; } @@ -3697,7 +4274,7 @@ static ir_node *gen_Phi(ir_node *node) { /* all integer operations are on 32bit registers now */ mode = mode_Iu; } else if(mode_is_float(mode)) { - if (USE_SSE2(env_cg)) { + if (ia32_cg_config.use_sse2) { mode = mode_xmm; } else { mode = mode_vfp; @@ -3730,11 +4307,12 @@ static ir_node *gen_IJmp(ir_node *node) ir_node *new_node; ia32_address_mode_t am; ia32_address_t *addr = &am.addr; - match_flags_t flags; - flags = match_force_32bit_op | match_no_immediate; + assert(get_irn_mode(op) == mode_P); - match_arguments(&am, block, NULL, op, flags); + match_arguments(&am, block, NULL, op, NULL, + match_am | match_8bit_am | match_16bit_am | + match_immediate | match_8bit | match_16bit); new_node = new_rd_ia32_IJmp(dbgi, irg, new_block, addr->base, addr->index, addr->mem, am.new_op2); @@ -3746,19 +4324,6 @@ static ir_node *gen_IJmp(ir_node *node) return new_node; } - -/********************************************************************** - * _ _ _ - * | | | | | | - * | | _____ _____ _ __ ___ __| | _ __ ___ __| | ___ ___ - * | |/ _ \ \ /\ / / _ \ '__/ _ \/ _` | | '_ \ / _ \ / _` |/ _ \/ __| - * | | (_) \ V V / __/ | | __/ (_| | | | | | (_) | (_| | __/\__ \ - * |_|\___/ \_/\_/ \___|_| \___|\__,_| |_| |_|\___/ \__,_|\___||___/ - * - **********************************************************************/ - -/* These nodes are created in intrinsic lowering (64bit -> 32bit) */ - typedef ir_node *construct_load_func(dbg_info *db, ir_graph *irg, ir_node *block, ir_node *base, ir_node *index, \ ir_node *mem); @@ -3834,27 +4399,37 @@ static ir_node *gen_lowered_Store(ir_node *node, construct_store_func func) return new_op; } +static ir_node *gen_ia32_l_ShlDep(ir_node *node) +{ + ir_node *left = get_irn_n(node, n_ia32_l_ShlDep_val); + ir_node *right = get_irn_n(node, n_ia32_l_ShlDep_count); -/** - * Transforms an ia32_l_XXX into a "real" XXX node - * - * @param node The node to transform - * @return the created ia32 XXX node - */ -#define GEN_LOWERED_SHIFT_OP(l_op, op) \ - static ir_node *gen_ia32_##l_op(ir_node *node) { \ - return gen_shift_binop(node, get_irn_n(node, 0), \ - get_irn_n(node, 1), new_rd_ia32_##op); \ - } + return gen_shift_binop(node, left, right, new_rd_ia32_Shl, + match_immediate | match_mode_neutral); +} + +static ir_node *gen_ia32_l_ShrDep(ir_node *node) +{ + ir_node *left = get_irn_n(node, n_ia32_l_ShrDep_val); + ir_node *right = get_irn_n(node, n_ia32_l_ShrDep_count); + return gen_shift_binop(node, left, right, new_rd_ia32_Shr, + match_immediate); +} -GEN_LOWERED_SHIFT_OP(l_ShlDep, Shl) -GEN_LOWERED_SHIFT_OP(l_ShrDep, Shr) -GEN_LOWERED_SHIFT_OP(l_SarDep, Sar) +static ir_node *gen_ia32_l_SarDep(ir_node *node) +{ + ir_node *left = get_irn_n(node, n_ia32_l_SarDep_val); + ir_node *right = get_irn_n(node, n_ia32_l_SarDep_count); + return gen_shift_binop(node, left, right, new_rd_ia32_Sar, + match_immediate); +} static ir_node *gen_ia32_l_Add(ir_node *node) { ir_node *left = get_irn_n(node, n_ia32_l_Add_left); ir_node *right = get_irn_n(node, n_ia32_l_Add_right); - ir_node *lowered = gen_binop(node, left, right, new_rd_ia32_Add, match_commutative); + ir_node *lowered = gen_binop(node, left, right, new_rd_ia32_Add, + match_commutative | match_am | match_immediate | + match_mode_neutral); if(is_Proj(lowered)) { lowered = get_Proj_pred(lowered); @@ -3868,17 +4443,9 @@ static ir_node *gen_ia32_l_Add(ir_node *node) { static ir_node *gen_ia32_l_Adc(ir_node *node) { - return gen_binop_flags(node, new_rd_ia32_Adc, match_commutative); -} - -/** - * Transforms an ia32_l_Neg into a "real" ia32_Neg node - * - * @param node The node to transform - * @return the created ia32 Neg node - */ -static ir_node *gen_ia32_l_Neg(ir_node *node) { - return gen_unop(node, get_unop_op(node), new_rd_ia32_Neg); + return gen_binop_flags(node, new_rd_ia32_Adc, + match_commutative | match_am | match_immediate | + match_mode_neutral); } /** @@ -3929,24 +4496,21 @@ static ir_node *gen_ia32_l_vfist(ir_node *node) { dbg_info *dbgi = get_irn_dbg_info(node); ir_node *noreg = ia32_new_NoReg_gp(env_cg); ir_mode *mode = get_ia32_ls_mode(node); - ir_node *trunc_mode = ia32_new_Fpu_truncate(env_cg); - ir_node *new_op; + ir_node *memres, *fist; long am_offs; - new_op = new_rd_ia32_vfist(dbgi, irg, block, new_ptr, noreg, new_mem, - new_val, trunc_mode); - + memres = gen_vfist(dbgi, irg, block, new_ptr, noreg, new_mem, new_val, &fist); am_offs = get_ia32_am_offs_int(node); - add_ia32_am_offs_int(new_op, am_offs); + add_ia32_am_offs_int(fist, am_offs); - set_ia32_op_type(new_op, ia32_AddrModeD); - set_ia32_ls_mode(new_op, mode); - set_ia32_frame_ent(new_op, get_ia32_frame_ent(node)); - set_ia32_use_frame(new_op); + set_ia32_op_type(fist, ia32_AddrModeD); + set_ia32_ls_mode(fist, mode); + set_ia32_frame_ent(fist, get_ia32_frame_ent(node)); + set_ia32_use_frame(fist); - SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node)); + SET_IA32_ORIG_NODE(fist, ia32_get_old_node_name(env_cg, node)); - return new_op; + return memres; } /** @@ -3959,7 +4523,7 @@ static ir_node *gen_ia32_l_Mul(ir_node *node) { ir_node *right = get_binop_right(node); return gen_binop(node, left, right, new_rd_ia32_Mul, - match_commutative | match_no_immediate); + match_commutative | match_am | match_mode_neutral); } /** @@ -3968,17 +4532,18 @@ static ir_node *gen_ia32_l_Mul(ir_node *node) { * @return the created ia32 IMul1OP node */ static ir_node *gen_ia32_l_IMul(ir_node *node) { - ir_node *left = get_binop_left(node); - ir_node *right = get_binop_right(node); + ir_node *left = get_binop_left(node); + ir_node *right = get_binop_right(node); return gen_binop(node, left, right, new_rd_ia32_IMul1OP, - match_commutative | match_no_immediate); + match_commutative | match_am | match_mode_neutral); } static ir_node *gen_ia32_l_Sub(ir_node *node) { - ir_node *left = get_irn_n(node, n_ia32_l_Sub_left); - ir_node *right = get_irn_n(node, n_ia32_l_Sub_right); - ir_node *lowered = gen_binop(node, left, right, new_rd_ia32_Sub, 0); + ir_node *left = get_irn_n(node, n_ia32_l_Sub_minuend); + ir_node *right = get_irn_n(node, n_ia32_l_Sub_subtrahend); + ir_node *lowered = gen_binop(node, left, right, new_rd_ia32_Sub, + match_am | match_immediate | match_mode_neutral); if(is_Proj(lowered)) { lowered = get_Proj_pred(lowered); @@ -3991,7 +4556,8 @@ static ir_node *gen_ia32_l_Sub(ir_node *node) { } static ir_node *gen_ia32_l_Sbb(ir_node *node) { - return gen_binop_flags(node, new_rd_ia32_Sbb, 0); + return gen_binop_flags(node, new_rd_ia32_Sbb, + match_am | match_immediate | match_mode_neutral); } /** @@ -4001,150 +4567,128 @@ static ir_node *gen_ia32_l_Sbb(ir_node *node) { * op3 - shift count * Only op3 can be an immediate. */ -static ir_node *gen_lowered_64bit_shifts(ir_node *node, ir_node *op1, - ir_node *op2, ir_node *count) +static ir_node *gen_lowered_64bit_shifts(ir_node *node, ir_node *high, + ir_node *low, ir_node *count) { - ir_node *block = be_transform_node(get_nodes_block(node)); - ir_node *new_op = NULL; + ir_node *block = get_nodes_block(node); + ir_node *new_block = be_transform_node(block); ir_graph *irg = current_ir_graph; dbg_info *dbgi = get_irn_dbg_info(node); - ir_node *new_op1 = be_transform_node(op1); - ir_node *new_op2 = be_transform_node(op2); - ir_node *new_count = create_immediate_or_transform(count, 'I'); - - /* TODO proper AM support */ + ir_node *new_high = be_transform_node(high); + ir_node *new_low = be_transform_node(low); + ir_node *new_count; + ir_node *new_node; - if (is_ia32_l_ShlD(node)) - new_op = new_rd_ia32_ShlD(dbgi, irg, block, new_op1, new_op2, new_count); - else - new_op = new_rd_ia32_ShrD(dbgi, irg, block, new_op1, new_op2, new_count); + /* the shift amount can be any mode that is bigger than 5 bits, since all + * other bits are ignored anyway */ + while (is_Conv(count) && get_irn_n_edges(count) == 1) { + assert(get_mode_size_bits(get_irn_mode(count)) >= 5); + count = get_Conv_op(count); + } + new_count = create_immediate_or_transform(count, 0); - SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node)); + if (is_ia32_l_ShlD(node)) { + new_node = new_rd_ia32_ShlD(dbgi, irg, new_block, new_high, new_low, + new_count); + } else { + new_node = new_rd_ia32_ShrD(dbgi, irg, new_block, new_high, new_low, + new_count); + } + SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node)); - return new_op; + return new_node; } -static ir_node *gen_ia32_l_ShlD(ir_node *node) { - return gen_lowered_64bit_shifts(node, get_irn_n(node, 0), - get_irn_n(node, 1), get_irn_n(node, 2)); +static ir_node *gen_ia32_l_ShlD(ir_node *node) +{ + ir_node *high = get_irn_n(node, n_ia32_l_ShlD_val_high); + ir_node *low = get_irn_n(node, n_ia32_l_ShlD_val_low); + ir_node *count = get_irn_n(node, n_ia32_l_ShlD_count); + return gen_lowered_64bit_shifts(node, high, low, count); } -static ir_node *gen_ia32_l_ShrD(ir_node *node) { - return gen_lowered_64bit_shifts(node, get_irn_n(node, 0), - get_irn_n(node, 1), get_irn_n(node, 2)); +static ir_node *gen_ia32_l_ShrD(ir_node *node) +{ + ir_node *high = get_irn_n(node, n_ia32_l_ShrD_val_high); + ir_node *low = get_irn_n(node, n_ia32_l_ShrD_val_low); + ir_node *count = get_irn_n(node, n_ia32_l_ShrD_count); + return gen_lowered_64bit_shifts(node, high, low, count); } -/** - * In case SSE Unit is used, the node is transformed into a vfst + xLoad. - */ -static ir_node *gen_ia32_l_X87toSSE(ir_node *node) { - ir_node *block = be_transform_node(get_nodes_block(node)); - ir_node *val = get_irn_n(node, 1); - ir_node *new_val = be_transform_node(val); - ia32_code_gen_t *cg = env_cg; - ir_node *res = NULL; - ir_graph *irg = current_ir_graph; - dbg_info *dbgi; - ir_node *noreg, *new_ptr, *new_mem; - ir_node *ptr, *mem; - - if (USE_SSE2(cg)) { - return new_val; - } - - mem = get_irn_n(node, 2); - new_mem = be_transform_node(mem); - ptr = get_irn_n(node, 0); - new_ptr = be_transform_node(ptr); - noreg = ia32_new_NoReg_gp(cg); - dbgi = get_irn_dbg_info(node); - - /* Store x87 -> MEM */ - res = new_rd_ia32_vfst(dbgi, irg, block, new_ptr, noreg, new_mem, new_val, - get_ia32_ls_mode(node)); - set_ia32_frame_ent(res, get_ia32_frame_ent(node)); - set_ia32_use_frame(res); - set_ia32_ls_mode(res, get_ia32_ls_mode(node)); - set_ia32_op_type(res, ia32_AddrModeD); - - /* Load MEM -> SSE */ - res = new_rd_ia32_xLoad(dbgi, irg, block, new_ptr, noreg, res, - get_ia32_ls_mode(node)); - set_ia32_frame_ent(res, get_ia32_frame_ent(node)); - set_ia32_use_frame(res); - set_ia32_op_type(res, ia32_AddrModeS); - res = new_rd_Proj(dbgi, irg, block, res, mode_xmm, pn_ia32_xLoad_res); +static ir_node *gen_ia32_l_LLtoFloat(ir_node *node) { + ir_node *src_block = get_nodes_block(node); + ir_node *block = be_transform_node(src_block); + ir_graph *irg = current_ir_graph; + dbg_info *dbgi = get_irn_dbg_info(node); + ir_node *frame = get_irg_frame(irg); + ir_node *noreg = ia32_new_NoReg_gp(env_cg); + ir_node *nomem = new_NoMem(); + ir_node *val_low = get_irn_n(node, n_ia32_l_LLtoFloat_val_low); + ir_node *val_high = get_irn_n(node, n_ia32_l_LLtoFloat_val_high); + ir_node *new_val_low = be_transform_node(val_low); + ir_node *new_val_high = be_transform_node(val_high); + ir_node *in[2]; + ir_node *sync; + ir_node *fild; + ir_node *store_low; + ir_node *store_high; - return res; -} + if(!mode_is_signed(get_irn_mode(val_high))) { + panic("unsigned long long -> float not supported yet (%+F)", node); + } -/** - * In case SSE Unit is used, the node is transformed into a xStore + vfld. - */ -static ir_node *gen_ia32_l_SSEtoX87(ir_node *node) { - ir_node *block = be_transform_node(get_nodes_block(node)); - ir_node *val = get_irn_n(node, 1); - ir_node *new_val = be_transform_node(val); - ia32_code_gen_t *cg = env_cg; - ir_graph *irg = current_ir_graph; - ir_node *res = NULL; - ir_entity *fent = get_ia32_frame_ent(node); - ir_mode *lsmode = get_ia32_ls_mode(node); - int offs = 0; - ir_node *noreg, *new_ptr, *new_mem; - ir_node *ptr, *mem; - dbg_info *dbgi; - - if (! USE_SSE2(cg)) { - /* SSE unit is not used -> skip this node. */ - return new_val; - } - - ptr = get_irn_n(node, 0); - new_ptr = be_transform_node(ptr); - mem = get_irn_n(node, 2); - new_mem = be_transform_node(mem); - noreg = ia32_new_NoReg_gp(cg); - dbgi = get_irn_dbg_info(node); - - /* Store SSE -> MEM */ - if (is_ia32_xLoad(skip_Proj(new_val))) { - ir_node *ld = skip_Proj(new_val); - - /* we can vfld the value directly into the fpu */ - fent = get_ia32_frame_ent(ld); - ptr = get_irn_n(ld, 0); - offs = get_ia32_am_offs_int(ld); - } else { - res = new_rd_ia32_xStore(dbgi, irg, block, new_ptr, noreg, new_mem, - new_val); - set_ia32_frame_ent(res, fent); - set_ia32_use_frame(res); - set_ia32_ls_mode(res, lsmode); - set_ia32_op_type(res, ia32_AddrModeD); - mem = res; - } - - /* Load MEM -> x87 */ - res = new_rd_ia32_vfld(dbgi, irg, block, new_ptr, noreg, new_mem, lsmode); - set_ia32_frame_ent(res, fent); - set_ia32_use_frame(res); - add_ia32_am_offs_int(res, offs); - set_ia32_op_type(res, ia32_AddrModeS); - res = new_rd_Proj(dbgi, irg, block, res, mode_vfp, pn_ia32_vfld_res); + /* do a store */ + store_low = new_rd_ia32_Store(dbgi, irg, block, frame, noreg, nomem, + new_val_low); + store_high = new_rd_ia32_Store(dbgi, irg, block, frame, noreg, nomem, + new_val_high); + SET_IA32_ORIG_NODE(store_low, ia32_get_old_node_name(env_cg, node)); + SET_IA32_ORIG_NODE(store_high, ia32_get_old_node_name(env_cg, node)); + + set_ia32_use_frame(store_low); + set_ia32_use_frame(store_high); + set_ia32_op_type(store_low, ia32_AddrModeD); + set_ia32_op_type(store_high, ia32_AddrModeD); + set_ia32_ls_mode(store_low, mode_Iu); + set_ia32_ls_mode(store_high, mode_Is); + add_ia32_am_offs_int(store_high, 4); + + in[0] = store_low; + in[1] = store_high; + sync = new_rd_Sync(dbgi, irg, block, 2, in); - return res; + /* do a fild */ + fild = new_rd_ia32_vfild(dbgi, irg, block, frame, noreg, sync); + + set_ia32_use_frame(fild); + set_ia32_op_type(fild, ia32_AddrModeS); + set_ia32_ls_mode(fild, mode_Ls); + + SET_IA32_ORIG_NODE(fild, ia32_get_old_node_name(env_cg, node)); + + return new_r_Proj(irg, block, fild, mode_vfp, pn_ia32_vfild_res); } -/********************************************************* - * _ _ _ - * (_) | | (_) - * _ __ ___ __ _ _ _ __ __| |_ __ ___ _____ _ __ - * | '_ ` _ \ / _` | | '_ \ / _` | '__| \ \ / / _ \ '__| - * | | | | | | (_| | | | | | | (_| | | | |\ V / __/ | - * |_| |_| |_|\__,_|_|_| |_| \__,_|_| |_| \_/ \___|_| - * - *********************************************************/ +static ir_node *gen_ia32_l_FloattoLL(ir_node *node) { + ir_node *src_block = get_nodes_block(node); + ir_node *block = be_transform_node(src_block); + ir_graph *irg = current_ir_graph; + dbg_info *dbgi = get_irn_dbg_info(node); + ir_node *frame = get_irg_frame(irg); + ir_node *noreg = ia32_new_NoReg_gp(env_cg); + ir_node *nomem = new_NoMem(); + ir_node *val = get_irn_n(node, n_ia32_l_FloattoLL_val); + ir_node *new_val = be_transform_node(val); + ir_node *fist, *mem; + + mem = gen_vfist(dbgi, irg, block, frame, noreg, nomem, new_val, &fist); + SET_IA32_ORIG_NODE(fist, ia32_get_old_node_name(env_cg, node)); + set_ia32_use_frame(fist); + set_ia32_op_type(fist, ia32_AddrModeD); + set_ia32_ls_mode(fist, mode_Ls); + + return mem; +} /** * the BAD transformer. @@ -4154,6 +4698,40 @@ static ir_node *bad_transform(ir_node *node) { return NULL; } +static ir_node *gen_Proj_l_FloattoLL(ir_node *node) { + ir_graph *irg = current_ir_graph; + ir_node *block = be_transform_node(get_nodes_block(node)); + ir_node *pred = get_Proj_pred(node); + ir_node *new_pred = be_transform_node(pred); + ir_node *frame = get_irg_frame(irg); + ir_node *noreg = ia32_new_NoReg_gp(env_cg); + dbg_info *dbgi = get_irn_dbg_info(node); + long pn = get_Proj_proj(node); + ir_node *load; + ir_node *proj; + ia32_attr_t *attr; + + load = new_rd_ia32_Load(dbgi, irg, block, frame, noreg, new_pred); + SET_IA32_ORIG_NODE(load, ia32_get_old_node_name(env_cg, node)); + set_ia32_use_frame(load); + set_ia32_op_type(load, ia32_AddrModeS); + set_ia32_ls_mode(load, mode_Iu); + /* we need a 64bit stackslot (fist stores 64bit) even though we only load + * 32 bit from it with this particular load */ + attr = get_ia32_attr(load); + attr->data.need_64bit_stackent = 1; + + if (pn == pn_ia32_l_FloattoLL_res_high) { + add_ia32_am_offs_int(load, 4); + } else { + assert(pn == pn_ia32_l_FloattoLL_res_low); + } + + proj = new_r_Proj(irg, block, load, mode_Iu, pn_ia32_Load_res); + + return proj; +} + /** * Transform the Projs of an AddSP. */ @@ -4238,15 +4816,22 @@ static ir_node *gen_Proj_Load(ir_node *node) { /* renumber the proj */ new_pred = be_transform_node(pred); if (is_ia32_Load(new_pred)) { - if (proj == pn_Load_res) { - return new_rd_Proj(dbgi, irg, block, new_pred, mode_Iu, - pn_ia32_Load_res); - } else if (proj == pn_Load_M) { - return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, - pn_ia32_Load_M); + switch (proj) { + case pn_Load_res: + return new_rd_Proj(dbgi, irg, block, new_pred, mode_Iu, pn_ia32_Load_res); + case pn_Load_M: + return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_Load_M); + case pn_Load_X_regular: + return new_rd_Jmp(dbgi, irg, block); + case pn_Load_X_except: + /* This Load might raise an exception. Mark it. */ + set_ia32_exc_label(new_pred, 1); + return new_rd_Proj(dbgi, irg, block, new_pred, mode_X, pn_ia32_Load_X_exc); + default: + break; } - } else if(is_ia32_Conv_I2I(new_pred) - || is_ia32_Conv_I2I8Bit(new_pred)) { + } else if (is_ia32_Conv_I2I(new_pred) || + is_ia32_Conv_I2I8Bit(new_pred)) { set_irn_mode(new_pred, mode_T); if (proj == pn_Load_res) { return new_rd_Proj(dbgi, irg, block, new_pred, mode_Iu, pn_ia32_res); @@ -4254,20 +4839,34 @@ static ir_node *gen_Proj_Load(ir_node *node) { return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_mem); } } else if (is_ia32_xLoad(new_pred)) { - if (proj == pn_Load_res) { - return new_rd_Proj(dbgi, irg, block, new_pred, mode_xmm, - pn_ia32_xLoad_res); - } else if (proj == pn_Load_M) { - return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, - pn_ia32_xLoad_M); + switch (proj) { + case pn_Load_res: + return new_rd_Proj(dbgi, irg, block, new_pred, mode_xmm, pn_ia32_xLoad_res); + case pn_Load_M: + return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_xLoad_M); + case pn_Load_X_regular: + return new_rd_Jmp(dbgi, irg, block); + case pn_Load_X_except: + /* This Load might raise an exception. Mark it. */ + set_ia32_exc_label(new_pred, 1); + return new_rd_Proj(dbgi, irg, block, new_pred, mode_X, pn_ia32_xLoad_X_exc); + default: + break; } } else if (is_ia32_vfld(new_pred)) { - if (proj == pn_Load_res) { - return new_rd_Proj(dbgi, irg, block, new_pred, mode_vfp, - pn_ia32_vfld_res); - } else if (proj == pn_Load_M) { - return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, - pn_ia32_vfld_M); + switch (proj) { + case pn_Load_res: + return new_rd_Proj(dbgi, irg, block, new_pred, mode_vfp, pn_ia32_vfld_res); + case pn_Load_M: + return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_vfld_M); + case pn_Load_X_regular: + return new_rd_Jmp(dbgi, irg, block); + case pn_Load_X_except: + /* This Load might raise an exception. Mark it. */ + set_ia32_exc_label(new_pred, 1); + return new_rd_Proj(dbgi, irg, block, new_pred, mode_X, pn_ia32_xLoad_X_exc); + default: + break; } } else { /* can happen for ProJMs when source address mode happened for the @@ -4276,7 +4875,7 @@ static ir_node *gen_Proj_Load(ir_node *node) { /* however it should not be the result proj, as that would mean the load had multiple users and should not have been used for SourceAM */ - if(proj != pn_Load_M) { + if (proj != pn_Load_M) { panic("internal error: transformed node not a Load"); } return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, 1); @@ -4307,6 +4906,11 @@ static ir_node *gen_Proj_DivMod(ir_node *node) { return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_Div_M); case pn_Div_res: return new_rd_Proj(dbgi, irg, block, new_pred, mode_Iu, pn_ia32_Div_div_res); + case pn_Div_X_regular: + return new_rd_Jmp(dbgi, irg, block); + case pn_Div_X_except: + set_ia32_exc_label(new_pred, 1); + return new_rd_Proj(dbgi, irg, block, new_pred, mode_X, pn_ia32_Div_X_exc); default: break; } @@ -4317,6 +4921,9 @@ static ir_node *gen_Proj_DivMod(ir_node *node) { return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_Div_M); case pn_Mod_res: return new_rd_Proj(dbgi, irg, block, new_pred, mode_Iu, pn_ia32_Div_mod_res); + case pn_Mod_X_except: + set_ia32_exc_label(new_pred, 1); + return new_rd_Proj(dbgi, irg, block, new_pred, mode_X, pn_ia32_Div_X_exc); default: break; } @@ -4329,6 +4936,11 @@ static ir_node *gen_Proj_DivMod(ir_node *node) { return new_rd_Proj(dbgi, irg, block, new_pred, mode_Iu, pn_ia32_Div_div_res); case pn_DivMod_res_mod: return new_rd_Proj(dbgi, irg, block, new_pred, mode_Iu, pn_ia32_Div_mod_res); + case pn_DivMod_X_regular: + return new_rd_Jmp(dbgi, irg, block); + case pn_DivMod_X_except: + set_ia32_exc_label(new_pred, 1); + return new_rd_Proj(dbgi, irg, block, new_pred, mode_X, pn_ia32_Div_X_exc); default: break; } @@ -4396,6 +5008,8 @@ static ir_node *gen_Proj_Quot(ir_node *node) { return new_rd_Proj(dbgi, irg, block, new_pred, mode_vfp, pn_ia32_vfdiv_res); } break; + case pn_Quot_X_regular: + case pn_Quot_X_except: default: break; } @@ -4472,9 +5086,8 @@ static ir_node *gen_Proj_be_Call(ir_node *node) { pn_ia32_xLoad_M); } } - if (USE_SSE2(env_cg) && proj >= pn_be_Call_first_res - && proj < (pn_be_Call_first_res + n_res) && mode_is_float(mode) - && USE_SSE2(env_cg)) { + if (ia32_cg_config.use_sse2 && proj >= pn_be_Call_first_res + && proj < (pn_be_Call_first_res + n_res) && mode_is_float(mode)) { ir_node *fstp; ir_node *frame = get_irg_frame(irg); ir_node *noreg = ia32_new_NoReg_gp(env_cg); @@ -4521,36 +5134,23 @@ static ir_node *gen_Proj_be_Call(ir_node *node) { */ static ir_node *gen_Proj_Cmp(ir_node *node) { - /* normally Cmps are processed when looking at Cond nodes, but this case - * can happen in complicated Psi conditions */ - dbg_info *dbgi = get_irn_dbg_info(node); - ir_node *block = get_nodes_block(node); - ir_node *new_block = be_transform_node(block); - ir_node *cmp = get_Proj_pred(node); - ir_node *new_cmp = be_transform_node(cmp); - long pnc = get_Proj_proj(node); - ir_node *res; - - res = create_set_32bit(dbgi, new_block, new_cmp, pnc, node, 0); - - return res; + /* this probably means not all mode_b nodes were lowered... */ + panic("trying to directly transform Proj_Cmp %+F (mode_b not lowered?)", + node); } /** * Transform and potentially renumber Proj nodes. */ static ir_node *gen_Proj(ir_node *node) { - ir_graph *irg = current_ir_graph; - dbg_info *dbgi = get_irn_dbg_info(node); ir_node *pred = get_Proj_pred(node); - long proj = get_Proj_proj(node); - if (is_Store(pred)) { + long proj = get_Proj_proj(node); if (proj == pn_Store_M) { return be_transform_node(pred); } else { assert(0); - return new_r_Bad(irg); + return new_r_Bad(current_ir_graph); } } else if (is_Load(pred)) { return gen_Proj_Load(node); @@ -4569,29 +5169,33 @@ static ir_node *gen_Proj(ir_node *node) { } else if (is_Cmp(pred)) { return gen_Proj_Cmp(node); } else if (get_irn_op(pred) == op_Start) { + long proj = get_Proj_proj(node); if (proj == pn_Start_X_initial_exec) { ir_node *block = get_nodes_block(pred); + dbg_info *dbgi = get_irn_dbg_info(node); ir_node *jump; /* we exchange the ProjX with a jump */ block = be_transform_node(block); - jump = new_rd_Jmp(dbgi, irg, block); + jump = new_rd_Jmp(dbgi, current_ir_graph, block); return jump; } if (node == be_get_old_anchor(anchor_tls)) { return gen_Proj_tls(node); } + } else if (is_ia32_l_FloattoLL(pred)) { + return gen_Proj_l_FloattoLL(node); #ifdef FIRM_EXT_GRS } else if(!is_ia32_irn(pred)) { // Quick hack for SIMD optimization #else } else { #endif - ir_node *new_pred = be_transform_node(pred); - ir_node *block = be_transform_node(get_nodes_block(node)); - ir_mode *mode = get_irn_mode(node); + ir_mode *mode = get_irn_mode(node); if (mode_needs_gp_reg(mode)) { - ir_node *new_proj = new_r_Proj(irg, block, new_pred, mode_Iu, - get_Proj_proj(node)); + ir_node *new_pred = be_transform_node(pred); + ir_node *block = be_transform_node(get_nodes_block(node)); + ir_node *new_proj = new_r_Proj(current_ir_graph, block, new_pred, + mode_Iu, get_Proj_proj(node)); #ifdef DEBUG_libfirm new_proj->node_nr = node->node_nr; #endif @@ -4654,7 +5258,6 @@ static void register_transformers(void) /* transform ops from intrinsic lowering */ GEN(ia32_l_Add); GEN(ia32_l_Adc); - GEN(ia32_l_Neg); GEN(ia32_l_Mul); GEN(ia32_l_IMul); GEN(ia32_l_ShlDep); @@ -4668,8 +5271,8 @@ static void register_transformers(void) GEN(ia32_l_Load); GEN(ia32_l_vfist); GEN(ia32_l_Store); - GEN(ia32_l_X87toSSE); - GEN(ia32_l_SSEtoX87); + GEN(ia32_l_LLtoFloat); + GEN(ia32_l_FloattoLL); GEN(Const); GEN(SymConst); @@ -4805,21 +5408,14 @@ void ia32_transform_graph(ia32_code_gen_t *cg) { int cse_last; ir_graph *irg = cg->irg; - /* TODO: look at cpu and fill transform config in with that... */ - transform_config.use_incdec = 1; - transform_config.use_sse2 = 0; - transform_config.use_ffreep = 0; - transform_config.use_ftst = 0; - transform_config.use_femms = 0; - transform_config.use_fucomi = 1; - transform_config.use_cmov = 1; - register_transformers(); env_cg = cg; initial_fpcw = NULL; +BE_TIMER_PUSH(t_heights); heights = heights_new(irg); - calculate_non_address_mode_nodes(irg); +BE_TIMER_POP(t_heights); + ia32_calculate_non_address_mode_nodes(cg->birg); /* the transform phase is not safe for CSE (yet) because several nodes get * attributes set after their creation */ @@ -4830,7 +5426,7 @@ void ia32_transform_graph(ia32_code_gen_t *cg) { set_opt_cse(cse_last); - free_non_address_mode_nodes(); + ia32_free_non_address_mode_nodes(); heights_free(heights); heights = NULL; }