X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fia32%2Fbearch_ia32.c;h=3715368079ec3780f3d762d6b61e594ea210400c;hb=39e27d1fe336c1238bfe4a7b614d103762e39ba2;hp=9b7a5526295acf8f688edf71a2ab6e16e681d194;hpb=2f10d96357f98a6d313d35aee8c49ca3caa4752a;p=libfirm diff --git a/ir/be/ia32/bearch_ia32.c b/ir/be/ia32/bearch_ia32.c index 9b7a55262..371536807 100644 --- a/ir/be/ia32/bearch_ia32.c +++ b/ir/be/ia32/bearch_ia32.c @@ -1,6 +1,6 @@ /** * This is the main ia32 firm backend driver. - * + * @author Christian Wuerdig * $Id$ */ @@ -8,12 +8,21 @@ #include "config.h" #endif -#ifdef _WIN32 +#ifdef HAVE_MALLOC_H #include -#else +#endif + +#ifdef HAVE_ALLOCA_H #include #endif +#ifdef WITH_LIBCORE +#include +#include +#endif /* WITH_LIBCORE */ + +#include + #include "pseudo_irg.h" #include "irgwalk.h" #include "irprog.h" @@ -22,8 +31,8 @@ #include "ircons.h" #include "irgmod.h" #include "irgopt.h" - -#include "bitset.h" +#include "irbitset.h" +#include "pdeq.h" #include "debug.h" #include "../beabi.h" /* the general register allocator interface */ @@ -31,6 +40,7 @@ #include "../belower.h" #include "../besched_t.h" #include "../be.h" +#include "../be_t.h" #include "bearch_ia32_t.h" #include "ia32_new_nodes.h" /* ia32 nodes interface */ @@ -40,6 +50,10 @@ #include "ia32_emitter.h" #include "ia32_map_regs.h" #include "ia32_optimize.h" +#include "ia32_x87.h" +#include "ia32_dbg_stat.h" +#include "ia32_finish.h" +#include "ia32_util.h" #define DEBUG_MODULE "firm.be.ia32.isa" @@ -49,12 +63,15 @@ static set *cur_reg_set = NULL; #undef is_Start #define is_Start(irn) (get_irn_opcode(irn) == iro_Start) +/* Creates the unique per irg GP NoReg node. */ ir_node *ia32_new_NoReg_gp(ia32_code_gen_t *cg) { - return be_abi_get_callee_save_irn(cg->birg->abi, &ia32_gp_regs[REG_XXX]); + return be_abi_get_callee_save_irn(cg->birg->abi, &ia32_gp_regs[REG_GP_NOREG]); } +/* Creates the unique per irg FP NoReg node. */ ir_node *ia32_new_NoReg_fp(ia32_code_gen_t *cg) { - return be_abi_get_callee_save_irn(cg->birg->abi, &ia32_fp_regs[REG_XXXX]); + return be_abi_get_callee_save_irn(cg->birg->abi, + USE_SSE2(cg) ? &ia32_xmm_regs[REG_XMM_NOREG] : &ia32_vfp_regs[REG_VFP_NOREG]); } /************************************************** @@ -81,10 +98,11 @@ static ir_node *my_skip_proj(const ir_node *n) { * will be asked for this information. */ static const arch_register_req_t *ia32_get_irn_reg_req(const void *self, arch_register_req_t *req, const ir_node *irn, int pos) { + const ia32_irn_ops_t *ops = self; const ia32_register_req_t *irn_req; long node_pos = pos == -1 ? 0 : pos; ir_mode *mode = is_Block(irn) ? NULL : get_irn_mode(irn); - firm_dbg_module_t *mod = firm_dbg_register(DEBUG_MODULE); + FIRM_DBG_REGISTER(firm_dbg_module_t *mod, DEBUG_MODULE); if (is_Block(irn) || mode == mode_M || mode == mode_X) { DBG((mod, LEVEL_1, "ignoring Block, mode_M, mode_X node %+F\n", irn)); @@ -98,7 +116,6 @@ static const arch_register_req_t *ia32_get_irn_reg_req(const void *self, arch_re DBG((mod, LEVEL_1, "get requirements at pos %d for %+F ... ", pos, irn)); - if (is_Proj(irn)) { if (pos == -1) { node_pos = ia32_translate_proj_pos(irn); @@ -135,19 +152,23 @@ static const arch_register_req_t *ia32_get_irn_reg_req(const void *self, arch_re } } else { - /* treat Phi like Const with default requirements */ - if (is_Phi(irn)) { - DB((mod, LEVEL_1, "returning standard reqs for %+F\n", irn)); - if (mode_is_float(mode)) - memcpy(req, &(ia32_default_req_ia32_fp.req), sizeof(*req)); + /* treat Unknowns like Const with default requirements */ + if (is_Unknown(irn)) { + DB((mod, LEVEL_1, "returning UKNWN reqs for %+F\n", irn)); + if (mode_is_float(mode)) { + if (USE_SSE2(ops->cg)) + memcpy(req, &(ia32_default_req_ia32_xmm_xmm_UKNWN), sizeof(*req)); + else + memcpy(req, &(ia32_default_req_ia32_vfp_vfp_UKNWN), sizeof(*req)); + } else if (mode_is_int(mode) || mode_is_reference(mode)) - memcpy(req, &(ia32_default_req_ia32_gp.req), sizeof(*req)); + memcpy(req, &(ia32_default_req_ia32_gp_gp_UKNWN), sizeof(*req)); else if (mode == mode_T || mode == mode_M) { - DBG((mod, LEVEL_1, "ignoring Phi node %+F\n", irn)); + DBG((mod, LEVEL_1, "ignoring Unknown node %+F\n", irn)); return NULL; } else - assert(0 && "unsupported Phi-Mode"); + assert(0 && "unsupported Unknown-Mode"); } else { DB((mod, LEVEL_1, "returning NULL for %+F (not ia32)\n", irn)); @@ -211,20 +232,49 @@ static const arch_register_t *ia32_get_irn_reg(const void *self, const ir_node * } static arch_irn_class_t ia32_classify(const void *self, const ir_node *irn) { + arch_irn_class_t classification = arch_irn_class_normal; + irn = my_skip_proj(irn); + if (is_cfop(irn)) - return arch_irn_class_branch; - else if (is_ia32_irn(irn)) - return arch_irn_class_normal; - else - return 0; + classification |= arch_irn_class_branch; + + if (! is_ia32_irn(irn)) + return classification & ~arch_irn_class_normal; + + if (is_ia32_Cnst(irn)) + classification |= arch_irn_class_const; + + if (is_ia32_Ld(irn)) + classification |= arch_irn_class_load; + + if (is_ia32_St(irn) || is_ia32_Store8Bit(irn)) + classification |= arch_irn_class_store; + + if (is_ia32_got_reload(irn)) + classification |= arch_irn_class_reload; + + return classification; } static arch_irn_flags_t ia32_get_flags(const void *self, const ir_node *irn) { + + if(is_Proj(irn)) { + ir_node *pred = get_Proj_pred(irn); + if(is_ia32_Push(pred) && get_Proj_proj(irn) == 0) { + return arch_irn_flags_modify_sp; + } + if(is_ia32_Pop(pred) && get_Proj_proj(irn) == 1) { + return arch_irn_flags_modify_sp; + } + } + irn = my_skip_proj(irn); if (is_ia32_irn(irn)) return get_ia32_flags(irn); else { + if (is_Unknown(irn)) + return arch_irn_flags_ignore; return 0; } } @@ -233,24 +283,57 @@ static entity *ia32_get_frame_entity(const void *self, const ir_node *irn) { return is_ia32_irn(irn) ? get_ia32_frame_ent(irn) : NULL; } -static void ia32_set_stack_bias(const void *self, ir_node *irn, int bias) { +static void ia32_set_frame_entity(const void *self, ir_node *irn, entity *ent) { + set_ia32_frame_ent(irn, ent); +} + +static void ia32_set_frame_offset(const void *self, ir_node *irn, int bias) { char buf[64]; const ia32_irn_ops_t *ops = self; - if (is_ia32_use_frame(irn) && bias != 0) { + if (get_ia32_frame_ent(irn)) { ia32_am_flavour_t am_flav = get_ia32_am_flavour(irn); + /* Pop nodes modify the stack pointer before reading the destination + * address, so fix this here + */ + if(is_ia32_Pop(irn)) { + bias -= 4; + } + DBG((ops->cg->mod, LEVEL_1, "stack biased %+F with %d\n", irn, bias)); + snprintf(buf, sizeof(buf), "%d", bias); - add_ia32_am_offs(irn, buf); - am_flav |= ia32_O; - set_ia32_am_flavour(irn, am_flav); + + if (get_ia32_op_type(irn) == ia32_Normal) { + set_ia32_cnst(irn, buf); + } + else { + add_ia32_am_offs(irn, buf); + am_flav |= ia32_O; + set_ia32_am_flavour(irn, am_flav); + } + } +} + +static int ia32_get_sp_bias(const void *self, const ir_node *irn) { + if(is_Proj(irn)) { + int proj = get_Proj_proj(irn); + ir_node *pred = get_Proj_pred(irn); + + if(is_ia32_Push(pred) && proj == 0) + return 4; + else if(is_ia32_Pop(pred) && proj == 1) + return -4; } + + return 0; } typedef struct { be_abi_call_flags_bits_t flags; const arch_isa_t *isa; + const arch_env_t *aenv; ir_graph *irg; } ia32_abi_env_t; @@ -260,10 +343,17 @@ static void *ia32_abi_init(const be_abi_call_t *call, const arch_env_t *aenv, ir be_abi_call_flags_t fl = be_abi_call_get_flags(call); env->flags = fl.bits; env->irg = irg; + env->aenv = aenv; env->isa = aenv->isa; return env; } +/** + * Put all registers which are saved by the prologue/epilogue in a set. + * + * @param self The callback object. + * @param s The result set. + */ static void ia32_abi_dont_save_regs(void *self, pset *s) { ia32_abi_env_t *env = self; @@ -271,59 +361,107 @@ static void ia32_abi_dont_save_regs(void *self, pset *s) pset_insert_ptr(s, env->isa->bp); } +/** + * Generate the routine prologue. + * + * @param self The callback object. + * @param mem A pointer to the mem node. Update this if you define new memory. + * @param reg_map A map mapping all callee_save/ignore/parameter registers to their defining nodes. + * + * @return The register which shall be used as a stack frame base. + * + * All nodes which define registers in @p reg_map must keep @p reg_map current. + */ static const arch_register_t *ia32_abi_prologue(void *self, ir_node **mem, pmap *reg_map) { - ia32_abi_env_t *env = self; - const arch_register_t *frame_reg = env->isa->sp; - - if(!env->flags.try_omit_fp) { - int reg_size = get_mode_size_bytes(env->isa->bp->reg_class->mode); - ir_node *bl = get_irg_start_block(env->irg); - ir_node *curr_sp = be_abi_reg_map_get(reg_map, env->isa->sp); - ir_node *curr_bp = be_abi_reg_map_get(reg_map, env->isa->bp); - ir_node *curr_no_reg = be_abi_reg_map_get(reg_map, &ia32_gp_regs[REG_XXX]); - ir_node *store_bp; - - curr_sp = be_new_IncSP(env->isa->sp, env->irg, bl, curr_sp, *mem, reg_size, be_stack_dir_along); - store_bp = new_rd_ia32_Store(NULL, env->irg, bl, curr_sp, curr_no_reg, curr_bp, *mem, mode_T); - set_ia32_am_support(store_bp, ia32_am_Dest); - set_ia32_am_flavour(store_bp, ia32_B); - set_ia32_op_type(store_bp, ia32_AddrModeD); - *mem = new_r_Proj(env->irg, bl, store_bp, mode_M, 0); + ia32_abi_env_t *env = self; + + if (! env->flags.try_omit_fp) { + ir_node *bl = get_irg_start_block(env->irg); + ir_node *curr_sp = be_abi_reg_map_get(reg_map, env->isa->sp); + ir_node *curr_bp = be_abi_reg_map_get(reg_map, env->isa->bp); + ir_node *push; + + /* push ebp */ + push = new_rd_ia32_Push(NULL, env->irg, bl, curr_sp, curr_bp, *mem); + curr_sp = new_r_Proj(env->irg, bl, push, get_irn_mode(curr_sp), pn_ia32_Push_stack); + *mem = new_r_Proj(env->irg, bl, push, mode_M, pn_ia32_Push_M); + + /* the push must have SP out register */ + arch_set_irn_register(env->aenv, curr_sp, env->isa->sp); + set_ia32_flags(push, arch_irn_flags_ignore); + + /* move esp to ebp */ curr_bp = be_new_Copy(env->isa->bp->reg_class, env->irg, bl, curr_sp); be_set_constr_single_reg(curr_bp, BE_OUT_POS(0), env->isa->bp); + arch_set_irn_register(env->aenv, curr_bp, env->isa->bp); be_node_set_flags(curr_bp, BE_OUT_POS(0), arch_irn_flags_ignore); + /* beware: the copy must be done before any other sp use */ + curr_sp = be_new_CopyKeep_single(env->isa->sp->reg_class, env->irg, bl, curr_sp, curr_bp, get_irn_mode(curr_sp)); + be_set_constr_single_reg(curr_sp, BE_OUT_POS(0), env->isa->sp); + arch_set_irn_register(env->aenv, curr_sp, env->isa->sp); + be_node_set_flags(curr_sp, BE_OUT_POS(0), arch_irn_flags_ignore); + be_abi_reg_map_set(reg_map, env->isa->sp, curr_sp); be_abi_reg_map_set(reg_map, env->isa->bp, curr_bp); + + return env->isa->bp; } - return frame_reg; + return env->isa->sp; } +/** + * Generate the routine epilogue. + * @param self The callback object. + * @param bl The block for the epilog + * @param mem A pointer to the mem node. Update this if you define new memory. + * @param reg_map A map mapping all callee_save/ignore/parameter registers to their defining nodes. + * @return The register which shall be used as a stack frame base. + * + * All nodes which define registers in @p reg_map must keep @p reg_map current. + */ static void ia32_abi_epilogue(void *self, ir_node *bl, ir_node **mem, pmap *reg_map) { - ia32_abi_env_t *env = self; - ir_node *curr_sp = be_abi_reg_map_get(reg_map, env->isa->sp); - ir_node *curr_bp = be_abi_reg_map_get(reg_map, env->isa->bp); - ir_node *curr_no_reg = be_abi_reg_map_get(reg_map, &ia32_gp_regs[REG_XXX]); + ia32_abi_env_t *env = self; + ir_node *curr_sp = be_abi_reg_map_get(reg_map, env->isa->sp); + ir_node *curr_bp = be_abi_reg_map_get(reg_map, env->isa->bp); - if(env->flags.try_omit_fp) { - curr_sp = be_new_IncSP(env->isa->sp, env->irg, bl, curr_sp, *mem, BE_STACK_FRAME_SIZE, be_stack_dir_against); + if (env->flags.try_omit_fp) { + /* simply remove the stack frame here */ + curr_sp = be_new_IncSP(env->isa->sp, env->irg, bl, curr_sp, *mem, BE_STACK_FRAME_SIZE_SHRINK); } - else { - ir_node *load_bp; - ir_mode *mode_bp = env->isa->bp->reg_class->mode; + const ia32_isa_t *isa = (ia32_isa_t *)env->isa; + ir_mode *mode_bp = env->isa->bp->reg_class->mode; + + /* gcc always emits a leave at the end of a routine */ + if (1 || ARCH_AMD(isa->opt_arch)) { + ir_node *leave; + + /* leave */ + leave = new_rd_ia32_Leave(NULL, env->irg, bl, curr_sp, *mem); + set_ia32_flags(leave, arch_irn_flags_ignore); + curr_bp = new_r_Proj(current_ir_graph, bl, leave, mode_bp, pn_ia32_Leave_frame); + curr_sp = new_r_Proj(current_ir_graph, bl, leave, get_irn_mode(curr_sp), pn_ia32_Leave_stack); + *mem = new_r_Proj(current_ir_graph, bl, leave, mode_M, pn_ia32_Leave_M); + } + else { + ir_node *pop; + + /* copy ebp to esp */ + curr_sp = be_new_SetSP(env->isa->sp, env->irg, bl, curr_sp, curr_bp, *mem); - curr_sp = be_new_SetSP(env->isa->sp, env->irg, bl, curr_sp, curr_bp, *mem); - load_bp = new_rd_ia32_Load(NULL, env->irg, bl, curr_sp, curr_no_reg, *mem, mode_T); - set_ia32_am_support(load_bp, ia32_am_Source); - set_ia32_am_flavour(load_bp, ia32_B); - set_ia32_op_type(load_bp, ia32_AddrModeS); - set_ia32_ls_mode(load_bp, mode_bp); - curr_bp = new_r_Proj(env->irg, bl, load_bp, mode_bp, 0); - *mem = new_r_Proj(env->irg, bl, load_bp, mode_M, 1); + /* pop ebp */ + pop = new_rd_ia32_Pop(NULL, env->irg, bl, curr_sp, *mem); + set_ia32_flags(pop, arch_irn_flags_ignore); + curr_bp = new_r_Proj(current_ir_graph, bl, pop, mode_bp, pn_ia32_Pop_res); + curr_sp = new_r_Proj(current_ir_graph, bl, pop, get_irn_mode(curr_sp), pn_ia32_Pop_stack); + *mem = new_r_Proj(current_ir_graph, bl, pop, mode_M, pn_ia32_Pop_M); + } + arch_set_irn_register(env->aenv, curr_sp, env->isa->sp); + arch_set_irn_register(env->aenv, curr_bp, env->isa->bp); } be_abi_reg_map_set(reg_map, env->isa->sp, curr_sp); @@ -337,35 +475,270 @@ static void ia32_abi_epilogue(void *self, ir_node *bl, ir_node **mem, pmap *reg_ */ static ir_type *ia32_abi_get_between_type(void *self) { +#define IDENT(s) new_id_from_chars(s, sizeof(s)-1) static ir_type *omit_fp_between_type = NULL; static ir_type *between_type = NULL; ia32_abi_env_t *env = self; - if(!between_type) { + if ( !between_type) { entity *old_bp_ent; entity *ret_addr_ent; entity *omit_fp_ret_addr_ent; - ir_type *old_bp_type = new_type_primitive(new_id_from_str("bp"), mode_P); - ir_type *ret_addr_type = new_type_primitive(new_id_from_str("return_addr"), mode_P); + ir_type *old_bp_type = new_type_primitive(IDENT("bp"), mode_P); + ir_type *ret_addr_type = new_type_primitive(IDENT("return_addr"), mode_P); - between_type = new_type_class(new_id_from_str("ia32_between_type")); - old_bp_ent = new_entity(between_type, new_id_from_str("old_bp"), old_bp_type); - ret_addr_ent = new_entity(between_type, new_id_from_str("ret_addr"), ret_addr_type); + between_type = new_type_struct(IDENT("ia32_between_type")); + old_bp_ent = new_entity(between_type, IDENT("old_bp"), old_bp_type); + ret_addr_ent = new_entity(between_type, IDENT("ret_addr"), ret_addr_type); set_entity_offset_bytes(old_bp_ent, 0); set_entity_offset_bytes(ret_addr_ent, get_type_size_bytes(old_bp_type)); set_type_size_bytes(between_type, get_type_size_bytes(old_bp_type) + get_type_size_bytes(ret_addr_type)); + set_type_state(between_type, layout_fixed); - omit_fp_between_type = new_type_class(new_id_from_str("ia32_between_type_omit_fp")); - omit_fp_ret_addr_ent = new_entity(omit_fp_between_type, new_id_from_str("ret_addr"), ret_addr_type); + omit_fp_between_type = new_type_struct(IDENT("ia32_between_type_omit_fp")); + omit_fp_ret_addr_ent = new_entity(omit_fp_between_type, IDENT("ret_addr"), ret_addr_type); set_entity_offset_bytes(omit_fp_ret_addr_ent, 0); set_type_size_bytes(omit_fp_between_type, get_type_size_bytes(ret_addr_type)); + set_type_state(omit_fp_between_type, layout_fixed); } return env->flags.try_omit_fp ? omit_fp_between_type : between_type; +#undef IDENT +} + +/** + * Get the estimated cycle count for @p irn. + * + * @param self The this pointer. + * @param irn The node. + * + * @return The estimated cycle count for this operation + */ +static int ia32_get_op_estimated_cost(const void *self, const ir_node *irn) +{ + int cost; + ia32_op_type_t op_tp; + const ia32_irn_ops_t *ops = self; + + if (is_Proj(irn)) + return 0; + + assert(is_ia32_irn(irn)); + + cost = get_ia32_latency(irn); + op_tp = get_ia32_op_type(irn); + + if (is_ia32_CopyB(irn)) { + cost = 250; + if (ARCH_INTEL(ops->cg->arch)) + cost += 150; + } + else if (is_ia32_CopyB_i(irn)) { + int size = get_tarval_long(get_ia32_Immop_tarval(irn)); + cost = 20 + (int)ceil((4/3) * size); + if (ARCH_INTEL(ops->cg->arch)) + cost += 150; + } + /* in case of address mode operations add additional cycles */ + else if (op_tp == ia32_AddrModeD || op_tp == ia32_AddrModeS) { + /* + In case of stack access add 5 cycles (we assume stack is in cache), + other memory operations cost 20 cycles. + */ + cost += is_ia32_use_frame(irn) ? 5 : 20; + } + + return cost; +} + +/** + * Returns the inverse operation if @p irn, recalculating the argument at position @p i. + * + * @param irn The original operation + * @param i Index of the argument we want the inverse operation to yield + * @param inverse struct to be filled with the resulting inverse op + * @param obstack The obstack to use for allocation of the returned nodes array + * @return The inverse operation or NULL if operation invertible + */ +static arch_inverse_t *ia32_get_inverse(const void *self, const ir_node *irn, int i, arch_inverse_t *inverse, struct obstack *obst) { + ir_graph *irg; + ir_mode *mode; + ir_node *block, *noreg, *nomem; + int pnc; + + /* we cannot invert non-ia32 irns */ + if (! is_ia32_irn(irn)) + return NULL; + + /* operand must always be a real operand (not base, index or mem) */ + if (i != 2 && i != 3) + return NULL; + + /* we don't invert address mode operations */ + if (get_ia32_op_type(irn) != ia32_Normal) + return NULL; + + irg = get_irn_irg(irn); + block = get_nodes_block(irn); + mode = get_ia32_res_mode(irn); + noreg = get_irn_n(irn, 0); + nomem = new_r_NoMem(irg); + + /* initialize structure */ + inverse->nodes = obstack_alloc(obst, 2 * sizeof(inverse->nodes[0])); + inverse->costs = 0; + inverse->n = 2; + + switch (get_ia32_irn_opcode(irn)) { + case iro_ia32_Add: + if (get_ia32_immop_type(irn) == ia32_ImmConst) { + /* we have an add with a const here */ + /* invers == add with negated const */ + inverse->nodes[0] = new_rd_ia32_Add(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), noreg, nomem); + pnc = pn_ia32_Add_res; + inverse->costs += 1; + copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn); + set_ia32_Immop_tarval(inverse->nodes[0], tarval_neg(get_ia32_Immop_tarval(irn))); + set_ia32_commutative(inverse->nodes[0]); + } + else if (get_ia32_immop_type(irn) == ia32_ImmSymConst) { + /* we have an add with a symconst here */ + /* invers == sub with const */ + inverse->nodes[0] = new_rd_ia32_Sub(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), noreg, nomem); + pnc = pn_ia32_Sub_res; + inverse->costs += 2; + copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn); + } + else { + /* normal add: inverse == sub */ + ir_node *proj = ia32_get_res_proj(irn); + assert(proj); + + inverse->nodes[0] = new_rd_ia32_Sub(NULL, irg, block, noreg, noreg, proj, get_irn_n(irn, i ^ 1), nomem); + pnc = pn_ia32_Sub_res; + inverse->costs += 2; + } + break; + case iro_ia32_Sub: + if (get_ia32_immop_type(irn) != ia32_ImmNone) { + /* we have a sub with a const/symconst here */ + /* invers == add with this const */ + inverse->nodes[0] = new_rd_ia32_Add(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), noreg, nomem); + pnc = pn_ia32_Add_res; + inverse->costs += (get_ia32_immop_type(irn) == ia32_ImmSymConst) ? 5 : 1; + copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn); + } + else { + /* normal sub */ + ir_node *proj = ia32_get_res_proj(irn); + assert(proj); + + if (i == 2) { + inverse->nodes[0] = new_rd_ia32_Add(NULL, irg, block, noreg, noreg, proj, get_irn_n(irn, 3), nomem); + } + else { + inverse->nodes[0] = new_rd_ia32_Sub(NULL, irg, block, noreg, noreg, get_irn_n(irn, 2), proj, nomem); + } + pnc = pn_ia32_Sub_res; + inverse->costs += 1; + } + break; + case iro_ia32_Eor: + if (get_ia32_immop_type(irn) != ia32_ImmNone) { + /* xor with const: inverse = xor */ + inverse->nodes[0] = new_rd_ia32_Eor(NULL, irg, block, noreg, noreg, get_irn_n(irn, i), noreg, nomem); + pnc = pn_ia32_Eor_res; + inverse->costs += (get_ia32_immop_type(irn) == ia32_ImmSymConst) ? 5 : 1; + copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn); + } + else { + /* normal xor */ + inverse->nodes[0] = new_rd_ia32_Eor(NULL, irg, block, noreg, noreg, (ir_node *)irn, get_irn_n(irn, i), nomem); + pnc = pn_ia32_Eor_res; + inverse->costs += 1; + } + break; + case iro_ia32_Not: { + ir_node *proj = ia32_get_res_proj(irn); + assert(proj); + + inverse->nodes[0] = new_rd_ia32_Not(NULL, irg, block, noreg, noreg, proj, nomem); + pnc = pn_ia32_Not_res; + inverse->costs += 1; + break; + } + case iro_ia32_Minus: { + ir_node *proj = ia32_get_res_proj(irn); + assert(proj); + + inverse->nodes[0] = new_rd_ia32_Minus(NULL, irg, block, noreg, noreg, proj, nomem); + pnc = pn_ia32_Minus_res; + inverse->costs += 1; + break; + } + default: + /* inverse operation not supported */ + return NULL; + } + + set_ia32_res_mode(inverse->nodes[0], mode); + inverse->nodes[1] = new_r_Proj(irg, block, inverse->nodes[0], mode, pnc); + + return inverse; +} + +/** + * Check if irn can load it's operand at position i from memory (source addressmode). + * @param self Pointer to irn ops itself + * @param irn The irn to be checked + * @param i The operands position + * @return Non-Zero if operand can be loaded + */ +static int ia32_possible_memory_operand(const void *self, const ir_node *irn, unsigned int i) { + if (! is_ia32_irn(irn) || /* must be an ia32 irn */ + get_irn_arity(irn) != 5 || /* must be a binary operation */ + get_ia32_op_type(irn) != ia32_Normal || /* must not already be a addressmode irn */ + ! (get_ia32_am_support(irn) & ia32_am_Source) || /* must be capable of source addressmode */ + (i != 2 && i != 3) || /* a "real" operand position must be requested */ + (i == 2 && ! is_ia32_commutative(irn)) || /* if first operand requested irn must be commutative */ + is_ia32_use_frame(irn)) /* must not already use frame */ + return 0; + + return 1; +} + +static void ia32_perform_memory_operand(const void *self, ir_node *irn, ir_node *spill, unsigned int i) { + assert(ia32_possible_memory_operand(self, irn, i) && "Cannot perform memory operand change"); + + if (i == 2) { + ir_node *tmp = get_irn_n(irn, 3); + set_irn_n(irn, 3, get_irn_n(irn, 2)); + set_irn_n(irn, 2, tmp); + } + + set_ia32_am_support(irn, ia32_am_Source); + set_ia32_op_type(irn, ia32_AddrModeS); + set_ia32_am_flavour(irn, ia32_B); + set_ia32_ls_mode(irn, get_irn_mode(get_irn_n(irn, i))); + set_ia32_frame_ent(irn, be_get_frame_entity(spill)); + set_ia32_use_frame(irn); + set_ia32_got_reload(irn); + + set_irn_n(irn, 0, be_get_Spill_frame(spill)); + set_irn_n(irn, 4, spill); + + /* + Input at position one is index register, which is NoReg. + We would need cg object to get a real noreg, but we cannot + access it from here. + */ + set_irn_n(irn, 3, get_irn_n(irn, 1)); + + //FIXME DBG_OPT_AM_S(reload, irn); } static const be_abi_callbacks_t ia32_abi_callbacks = { @@ -386,7 +759,13 @@ static const arch_irn_ops_if_t ia32_irn_ops_if = { ia32_classify, ia32_get_flags, ia32_get_frame_entity, - ia32_set_stack_bias + ia32_set_frame_entity, + ia32_set_frame_offset, + ia32_get_sp_bias, + ia32_get_inverse, + ia32_get_op_estimated_cost, + ia32_possible_memory_operand, + ia32_perform_memory_operand, }; ia32_irn_ops_t ia32_irn_ops = { @@ -407,111 +786,137 @@ ia32_irn_ops_t ia32_irn_ops = { * |___/ **************************************************/ +static void ia32_kill_convs(ia32_code_gen_t *cg) { + ir_node *irn; + + /* BEWARE: the Projs are inserted in the set */ + foreach_nodeset(cg->kill_conv, irn) { + ir_node *in = get_irn_n(get_Proj_pred(irn), 2); + edges_reroute(irn, in, cg->birg->irg); + } +} + /** * Transforms the standard firm graph into * an ia32 firm graph */ static void ia32_prepare_graph(void *self) { ia32_code_gen_t *cg = self; - firm_dbg_module_t *old_mod = cg->mod; + dom_front_info_t *dom; + DEBUG_ONLY(firm_dbg_module_t *old_mod = cg->mod;) + + FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.transform"); + + /* 1st: transform constants and psi condition trees */ + ia32_pre_transform_phase(cg); + + /* 2nd: transform all remaining nodes */ + ia32_register_transformers(); + dom = be_compute_dominance_frontiers(cg->irg); + + cg->kill_conv = new_nodeset(5); + irg_walk_blkwise_graph(cg->irg, NULL, ia32_transform_node, cg); + ia32_kill_convs(cg); + del_nodeset(cg->kill_conv); - cg->mod = firm_dbg_register("firm.be.ia32.transform"); - irg_walk_blkwise_graph(cg->irg, ia32_place_consts_set_modes, ia32_transform_node, cg); - be_dump(cg->irg, "-transformed", dump_ir_block_graph_sched); + be_free_dominance_frontiers(dom); - cg->mod = old_mod; + if (cg->dump) + be_dump(cg->irg, "-transformed", dump_ir_block_graph_sched); - if (cg->opt.doam) { - edges_deactivate(cg->irg); - //dead_node_elimination(cg->irg); - edges_activate(cg->irg); + /* 3rd: optimize address mode */ + FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.am"); + ia32_optimize_addressmode(cg); - irg_walk_blkwise_graph(cg->irg, NULL, ia32_optimize_am, cg); + if (cg->dump) be_dump(cg->irg, "-am", dump_ir_block_graph_sched); - } -} + DEBUG_ONLY(cg->mod = old_mod;) +} /** - * Insert copies for all ia32 nodes where the should_be_same requirement - * is not fulfilled. - * Transform Sub into Neg -- Add if IN2 == OUT + * Dummy functions for hooks we don't need but which must be filled. */ -static void ia32_finish_irg_walker(ir_node *irn, void *env) { - ia32_code_gen_t *cg = env; - const ia32_register_req_t **reqs; - const arch_register_t *out_reg, *in_reg; - int n_res, i; - ir_node *copy, *in_node, *block; +static void ia32_before_sched(void *self) { +} - if (! is_ia32_irn(irn)) - return; +static void remove_unused_nodes(ir_node *irn, bitset_t *already_visited) { + int i; + ir_mode *mode; + ir_node *mem_proj; - /* nodes with destination address mode don't produce values */ - if (get_ia32_op_type(irn) == ia32_AddrModeD) + if (is_Block(irn)) return; - reqs = get_ia32_out_req_all(irn); - n_res = get_ia32_n_res(irn); - block = get_nodes_block(irn); + mode = get_irn_mode(irn); - /* check all OUT requirements, if there is a should_be_same */ - for (i = 0; i < n_res; i++) { - if (arch_register_req_is(&(reqs[i]->req), should_be_same)) { - /* get in and out register */ - out_reg = get_ia32_out_reg(irn, i); - in_node = get_irn_n(irn, reqs[i]->same_pos); - in_reg = arch_get_irn_register(cg->arch_env, in_node); + /* check if we already saw this node or the node has more than one user */ + if (bitset_contains_irn(already_visited, irn) || get_irn_n_edges(irn) > 1) + return; - /* check if in and out register are equal */ - if (arch_register_get_index(out_reg) != arch_register_get_index(in_reg)) { - DBG((cg->mod, LEVEL_1, "inserting copy for %+F in_pos %d\n", irn, reqs[i]->same_pos)); + /* mark irn visited */ + bitset_add_irn(already_visited, irn); - /* create copy from in register */ - copy = be_new_Copy(arch_register_get_class(in_reg), cg->irg, block, in_node); + /* non-Tuple nodes with one user: ok, return */ + if (get_irn_n_edges(irn) >= 1 && mode != mode_T) + return; - /* destination is the out register */ - arch_set_irn_register(cg->arch_env, copy, out_reg); + /* tuple node has one user which is not the mem proj-> ok */ + if (mode == mode_T && get_irn_n_edges(irn) == 1) { + mem_proj = ia32_get_proj_for_mode(irn, mode_M); + if (! mem_proj) + return; + } - /* insert copy before the node into the schedule */ - sched_add_before(irn, copy); + for (i = get_irn_arity(irn) - 1; i >= 0; i--) { + ir_node *pred = get_irn_n(irn, i); - /* set copy as in */ - set_irn_n(irn, reqs[i]->same_pos, copy); - } - } - } + /* do not follow memory edges or we will accidentally remove stores */ + if (is_Proj(pred) && get_irn_mode(pred) == mode_M) + continue; - /* check if there is a sub which need to be transformed */ - ia32_transform_sub_to_neg_add(irn, cg); + set_irn_n(irn, i, new_Bad()); - /* transform a LEA into an Add if possible */ - ia32_transform_lea_to_add(irn, cg); + /* + The current node is about to be removed: if the predecessor + has only this node as user, it need to be removed as well. + */ + if (get_irn_n_edges(pred) <= 1) + remove_unused_nodes(pred, already_visited); + } - /* check for peephole optimization */ - ia32_peephole_optimization(irn, cg); + if (sched_is_scheduled(irn)) + sched_remove(irn); } -/** - * Add Copy nodes for not fulfilled should_be_equal constraints - */ -static void ia32_finish_irg(ir_graph *irg, ia32_code_gen_t *cg) { - irg_walk_blkwise_graph(irg, NULL, ia32_finish_irg_walker, cg); +static void remove_unused_loads_walker(ir_node *irn, void *env) { + bitset_t *already_visited = env; + if (is_ia32_Ld(irn) && ! bitset_contains_irn(already_visited, irn)) + remove_unused_nodes(irn, env); } - - /** - * Dummy functions for hooks we don't need but which must be filled. + * Called before the register allocator. + * Calculate a block schedule here. We need it for the x87 + * simulator and the emitter. */ -static void ia32_before_sched(void *self) { -} - static void ia32_before_ra(void *self) { + ia32_code_gen_t *cg = self; + bitset_t *already_visited = bitset_irg_malloc(cg->irg); + + cg->blk_sched = sched_create_block_schedule(cg->irg); + + /* + Handle special case: + There are sometimes unused loads, only pinned by memory. + We need to remove those Loads and all other nodes which won't be used + after removing the Load from schedule. + */ + irg_walk_graph(cg->irg, remove_unused_loads_walker, NULL, already_visited); + bitset_free(already_visited); } - /** * Transforms a be node into a Load. */ @@ -532,10 +937,13 @@ static void transform_to_Load(ia32_transform_env_t *env) { } if (mode_is_float(mode)) { - new_op = new_rd_ia32_fLoad(env->dbg, env->irg, env->block, ptr, noreg, mem, mode_T); + if (USE_SSE2(env->cg)) + new_op = new_rd_ia32_xLoad(env->dbg, env->irg, env->block, ptr, noreg, mem); + else + new_op = new_rd_ia32_vfld(env->dbg, env->irg, env->block, ptr, noreg, mem); } else { - new_op = new_rd_ia32_Load(env->dbg, env->irg, env->block, ptr, noreg, mem, mode_T); + new_op = new_rd_ia32_Load(env->dbg, env->irg, env->block, ptr, noreg, mem); } set_ia32_am_support(new_op, ia32_am_Source); @@ -545,6 +953,8 @@ static void transform_to_Load(ia32_transform_env_t *env) { set_ia32_frame_ent(new_op, ent); set_ia32_use_frame(new_op); + DBG_OPT_RELOAD2LD(irn, new_op); + proj = new_rd_Proj(env->dbg, env->irg, env->block, new_op, mode, pn_Load_res); if (sched_point) { @@ -558,8 +968,9 @@ static void transform_to_Load(ia32_transform_env_t *env) { reg = arch_get_irn_register(env->cg->arch_env, irn); arch_set_irn_register(env->cg->arch_env, new_op, reg); - exchange(irn, proj); + SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, irn)); + exchange(irn, proj); } /** @@ -581,71 +992,240 @@ static void transform_to_Store(ia32_transform_env_t *env) { } if (mode_is_float(mode)) { - new_op = new_rd_ia32_fStore(env->dbg, env->irg, env->block, ptr, noreg, val, nomem, mode_T); + if (USE_SSE2(env->cg)) + new_op = new_rd_ia32_xStore(env->dbg, env->irg, env->block, ptr, noreg, val, nomem); + else + new_op = new_rd_ia32_vfst(env->dbg, env->irg, env->block, ptr, noreg, val, nomem); } else if (get_mode_size_bits(mode) == 8) { - new_op = new_rd_ia32_Store8Bit(env->dbg, env->irg, env->block, ptr, noreg, val, nomem, mode_T); + new_op = new_rd_ia32_Store8Bit(env->dbg, env->irg, env->block, ptr, noreg, val, nomem); } else { - new_op = new_rd_ia32_Store(env->dbg, env->irg, env->block, ptr, noreg, val, nomem, mode_T); + new_op = new_rd_ia32_Store(env->dbg, env->irg, env->block, ptr, noreg, val, nomem); } set_ia32_am_support(new_op, ia32_am_Dest); set_ia32_op_type(new_op, ia32_AddrModeD); set_ia32_am_flavour(new_op, ia32_B); - set_ia32_ls_mode(new_op, get_irn_mode(val)); + set_ia32_ls_mode(new_op, mode); set_ia32_frame_ent(new_op, ent); set_ia32_use_frame(new_op); - proj = new_rd_Proj(env->dbg, env->irg, env->block, new_op, mode, 0); + DBG_OPT_SPILL2ST(irn, new_op); + + proj = new_rd_Proj(env->dbg, env->irg, env->block, new_op, mode_M, pn_ia32_Store_M); if (sched_point) { sched_add_after(sched_point, new_op); - sched_add_after(new_op, proj); - sched_remove(irn); } + SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, irn)); + exchange(irn, proj); +} + +static ir_node *create_push(ia32_transform_env_t *env, ir_node *schedpoint, ir_node *sp, ir_node *mem, entity *ent, const char *offset) { + ir_node *noreg = ia32_new_NoReg_gp(env->cg); + + ir_node *push = new_rd_ia32_Push(env->dbg, env->irg, env->block, sp, noreg, mem); + + set_ia32_frame_ent(push, ent); + set_ia32_use_frame(push); + set_ia32_op_type(push, ia32_AddrModeS); + set_ia32_am_flavour(push, ia32_B); + set_ia32_ls_mode(push, mode_Is); + if(offset != NULL) + add_ia32_am_offs(push, offset); + + sched_add_before(schedpoint, push); + return push; +} + +static ir_node *create_pop(ia32_transform_env_t *env, ir_node *schedpoint, ir_node *sp, entity *ent, const char *offset) { + ir_node *pop = new_rd_ia32_Pop(env->dbg, env->irg, env->block, sp, new_NoMem()); + + set_ia32_frame_ent(pop, ent); + set_ia32_use_frame(pop); + set_ia32_op_type(pop, ia32_AddrModeD); + set_ia32_am_flavour(pop, ia32_B); + set_ia32_ls_mode(pop, mode_Is); + if(offset != NULL) + add_ia32_am_offs(pop, offset); + + sched_add_before(schedpoint, pop); + + return pop; +} + +static ir_node* create_spproj(ia32_transform_env_t *env, ir_node *pred, int pos, ir_node *schedpoint, const ir_node *oldsp) { + ir_mode *spmode = get_irn_mode(oldsp); + const arch_register_t *spreg = arch_get_irn_register(env->cg->arch_env, oldsp); + ir_node *sp; + + sp = new_rd_Proj(env->dbg, env->irg, env->block, pred, spmode, pos); + arch_set_irn_register(env->cg->arch_env, sp, spreg); + sched_add_before(schedpoint, sp); + + return sp; +} + +static void transform_MemPerm(ia32_transform_env_t *env) { + /* + * Transform memperm, currently we do this the ugly way and produce + * push/pop into/from memory cascades. This is possible without using + * any registers. + */ + ir_node *node = env->irn; + int i, arity; + ir_node *sp = get_irn_n(node, 0); + const ir_edge_t *edge; + const ir_edge_t *next; + ir_node **pops; + + arity = be_get_MemPerm_entity_arity(node); + pops = alloca(arity * sizeof(pops[0])); + + // create pushs + for(i = 0; i < arity; ++i) { + entity *ent = be_get_MemPerm_in_entity(node, i); + ir_type *enttype = get_entity_type(ent); + int entbits = get_type_size_bits(enttype); + ir_node *mem = get_irn_n(node, i + 1); + ir_node *push; + + assert( (entbits == 32 || entbits == 64) && "spillslot on x86 should be 32 or 64 bit"); + + push = create_push(env, node, sp, mem, ent, NULL); + sp = create_spproj(env, push, 0, node, sp); + if(entbits == 64) { + // add another push after the first one + push = create_push(env, node, sp, mem, ent, "4"); + sp = create_spproj(env, push, 0, node, sp); + } + + set_irn_n(node, i, new_Bad()); + } + + // create pops + for(i = arity - 1; i >= 0; --i) { + entity *ent = be_get_MemPerm_out_entity(node, i); + ir_type *enttype = get_entity_type(ent); + int entbits = get_type_size_bits(enttype); + ir_node *pop; + + assert( (entbits == 32 || entbits == 64) && "spillslot on x86 should be 32 or 64 bit"); + + pop = create_pop(env, node, sp, ent, NULL); + if(entbits == 64) { + // add another pop after the first one + sp = create_spproj(env, pop, 1, node, sp); + pop = create_pop(env, node, sp, ent, "4"); + } + //if(i != 0) { + sp = create_spproj(env, pop, 1, node, sp); + //} + + pops[i] = pop; + } + + // exchange memprojs + foreach_out_edge_safe(node, edge, next) { + ir_node *proj = get_edge_src_irn(edge); + int p = get_Proj_proj(proj); + + assert(p < arity); + + set_Proj_pred(proj, pops[p]); + set_Proj_proj(proj, 3); + } + + sched_remove(node); } /** - * Calls the transform functions for StackParam, Spill and Reload. + * Fix the mode of Spill/Reload */ -static void ia32_after_ra_walker(ir_node *node, void *env) { +static ir_mode *fix_spill_mode(ia32_code_gen_t *cg, ir_mode *mode) +{ + if (mode_is_float(mode)) { + if (USE_SSE2(cg)) + mode = mode_D; + else + mode = mode_E; + } + else + mode = mode_Is; + return mode; +} + +/** + * Block-Walker: Calls the transform functions Spill and Reload. + */ +static void ia32_after_ra_walker(ir_node *block, void *env) { + ir_node *node, *prev; ia32_code_gen_t *cg = env; ia32_transform_env_t tenv; - if (is_Block(node)) - return; - - tenv.block = get_nodes_block(node); - tenv.dbg = get_irn_dbg_info(node); + tenv.block = block; tenv.irg = current_ir_graph; - tenv.irn = node; - tenv.mod = cg->mod; - tenv.mode = get_irn_mode(node); tenv.cg = cg; - - /* be_is_StackParam(node) || */ - if (be_is_Reload(node)) { - transform_to_Load(&tenv); - } - else if (be_is_Spill(node)) { - transform_to_Store(&tenv); + DEBUG_ONLY(tenv.mod = cg->mod;) + + /* beware: the schedule is changed here */ + for (node = sched_last(block); !sched_is_begin(node); node = prev) { + prev = sched_prev(node); + if (be_is_Reload(node)) { + /* we always reload the whole register */ + tenv.dbg = get_irn_dbg_info(node); + tenv.irn = node; + tenv.mode = fix_spill_mode(cg, get_irn_mode(node)); + transform_to_Load(&tenv); + } + else if (be_is_Spill(node)) { + ir_node *spillval = get_irn_n(node, be_pos_Spill_val); + /* we always spill the whole register */ + tenv.dbg = get_irn_dbg_info(node); + tenv.irn = node; + tenv.mode = fix_spill_mode(cg, get_irn_mode(spillval)); + transform_to_Store(&tenv); + } + else if(be_is_MemPerm(node)) { + tenv.dbg = get_irn_dbg_info(node); + tenv.irn = node; + transform_MemPerm(&tenv); + } } } /** - * We transform StackParam, Spill and Reload here. This needs to be done before + * We transform Spill and Reload here. This needs to be done before * stack biasing otherwise we would miss the corrected offset for these nodes. + * + * If x87 instruction should be emitted, run the x87 simulator and patch + * the virtual instructions. This must obviously be done after register allocation. */ static void ia32_after_ra(void *self) { ia32_code_gen_t *cg = self; - irg_walk_blkwise_graph(cg->irg, NULL, ia32_after_ra_walker, self); + + irg_block_walk_graph(cg->irg, NULL, ia32_after_ra_walker, self); + + /* if we do x87 code generation, rewrite all the virtual instructions and registers */ + if (cg->used_fp == fp_x87 || cg->force_sim) { + x87_simulate_graph(cg->arch_env, cg->irg, cg->blk_sched); + } } +/** + * Last touchups for the graph before emit + */ +static void ia32_finish(void *self) { + ia32_code_gen_t *cg = self; + ir_graph *irg = cg->irg; + + ia32_finish_irg(irg, cg); +} /** * Emits the code, closes the output file and frees @@ -654,28 +1234,20 @@ static void ia32_after_ra(void *self) { static void ia32_codegen(void *self) { ia32_code_gen_t *cg = self; ir_graph *irg = cg->irg; - FILE *out = cg->out; - if (cg->emit_decls) { - ia32_gen_decls(cg->out); - cg->emit_decls = 0; - } - - ia32_finish_irg(irg, cg); - be_dump(irg, "-finished", dump_ir_block_graph_sched); - ia32_gen_routine(out, irg, cg); + ia32_gen_routine(cg->isa->out, irg, cg); cur_reg_set = NULL; - pmap_destroy(cg->tv_ent); - pmap_destroy(cg->types); + /* remove it from the isa */ + cg->isa->cg = NULL; /* de-allocate code generator */ del_set(cg->reg_set); free(self); } -static void *ia32_cg_init(FILE *F, const be_irg_t *birg); +static void *ia32_cg_init(const be_irg_t *birg); static const arch_code_generator_if_t ia32_code_gen_if = { ia32_cg_init, @@ -684,32 +1256,39 @@ static const arch_code_generator_if_t ia32_code_gen_if = { ia32_before_sched, /* before scheduling hook */ ia32_before_ra, /* before register allocation hook */ ia32_after_ra, /* after register allocation hook */ + ia32_finish, /* called before codegen */ ia32_codegen /* emit && done */ }; /** - * Initializes the code generator. + * Initializes a IA32 code generator. */ -static void *ia32_cg_init(FILE *F, const be_irg_t *birg) { +static void *ia32_cg_init(const be_irg_t *birg) { ia32_isa_t *isa = (ia32_isa_t *)birg->main_env->arch_env->isa; ia32_code_gen_t *cg = xcalloc(1, sizeof(*cg)); - cg->impl = &ia32_code_gen_if; - cg->irg = birg->irg; - cg->reg_set = new_set(ia32_cmp_irn_reg_assoc, 1024); - cg->mod = firm_dbg_register("firm.be.ia32.cg"); - cg->out = F; - cg->arch_env = birg->main_env->arch_env; - cg->types = pmap_create(); - cg->tv_ent = pmap_create(); - cg->birg = birg; - - /* set optimizations */ - cg->opt.incdec = 0; - cg->opt.doam = 1; - cg->opt.placecnst = 1; - cg->opt.immops = 1; - cg->opt.extbb = 1; + cg->impl = &ia32_code_gen_if; + cg->irg = birg->irg; + cg->reg_set = new_set(ia32_cmp_irn_reg_assoc, 1024); + cg->arch_env = birg->main_env->arch_env; + cg->isa = isa; + cg->birg = birg; + cg->blk_sched = NULL; + cg->fp_to_gp = NULL; + cg->gp_to_fp = NULL; + cg->fp_kind = isa->fp_kind; + cg->used_fp = fp_none; + cg->dump = (birg->main_env->options->dump_flags & DUMP_BE) ? 1 : 0; + + FIRM_DBG_REGISTER(cg->mod, "firm.be.ia32.cg"); + + /* copy optimizations from isa for easier access */ + cg->opt = isa->opt; + cg->arch = isa->arch; + cg->opt_arch = isa->opt_arch; + + /* enter it */ + isa->cg = cg; #ifndef NDEBUG if (isa->name_obst_size) { @@ -720,13 +1299,6 @@ static void *ia32_cg_init(FILE *F, const be_irg_t *birg) { } #endif /* NDEBUG */ - isa->num_codegens++; - - if (isa->num_codegens > 1) - cg->emit_decls = 0; - else - cg->emit_decls = 1; - cur_reg_set = cg->reg_set; ia32_irn_ops.cg = cg; @@ -746,14 +1318,61 @@ static void *ia32_cg_init(FILE *F, const be_irg_t *birg) { * *****************************************************************/ +/** + * Set output modes for GCC + */ +static const tarval_mode_info mo_integer = { + TVO_DECIMAL, + NULL, + NULL, +}; + +/* +* set the tarval output mode to C-semantics +*/ +static void set_tarval_output_modes(void) +{ + set_tarval_mode_output_option(get_modeLs(), &mo_integer); + set_tarval_mode_output_option(get_modeLu(), &mo_integer); + set_tarval_mode_output_option(get_modeIs(), &mo_integer); + set_tarval_mode_output_option(get_modeIu(), &mo_integer); + set_tarval_mode_output_option(get_modeHs(), &mo_integer); + set_tarval_mode_output_option(get_modeHu(), &mo_integer); + set_tarval_mode_output_option(get_modeBs(), &mo_integer); + set_tarval_mode_output_option(get_modeBu(), &mo_integer); + set_tarval_mode_output_option(get_modeC(), &mo_integer); + set_tarval_mode_output_option(get_modeU(), &mo_integer); + set_tarval_mode_output_option(get_modeIu(), &mo_integer); +} + + +/** + * The template that generates a new ISA object. + * Note that this template can be changed by command line + * arguments. + */ static ia32_isa_t ia32_isa_template = { - &ia32_isa_if, /* isa interface implementation */ - &ia32_gp_regs[REG_ESP], /* stack pointer register */ - &ia32_gp_regs[REG_EBP], /* base pointer register */ - -1, /* stack direction */ - 0, /* number of code generator objects so far */ + { + &ia32_isa_if, /* isa interface implementation */ + &ia32_gp_regs[REG_ESP], /* stack pointer register */ + &ia32_gp_regs[REG_EBP], /* base pointer register */ + -1, /* stack direction */ + }, NULL, /* 16bit register names */ NULL, /* 8bit register names */ + NULL, /* types */ + NULL, /* tv_ents */ + (0 | + IA32_OPT_INCDEC | /* optimize add 1, sub 1 into inc/dec default: on */ + IA32_OPT_DOAM | /* optimize address mode default: on */ + IA32_OPT_LEA | /* optimize for LEAs default: on */ + IA32_OPT_PLACECNST | /* place constants immediately before instructions, default: on */ + IA32_OPT_IMMOPS | /* operations can use immediates, default: on */ + IA32_OPT_EXTBB), /* use extended basic block scheduling, default: on */ + arch_pentium_4, /* instruction architecture */ + arch_pentium_4, /* optimize for architecture */ + fp_sse2, /* use sse2 unit */ + NULL, /* current code generator */ #ifndef NDEBUG NULL, /* name obstack */ 0 /* name obst size */ @@ -763,32 +1382,62 @@ static ia32_isa_t ia32_isa_template = { /** * Initializes the backend ISA. */ -static void *ia32_init(void) { +static void *ia32_init(FILE *file_handle) { static int inited = 0; ia32_isa_t *isa; - if(inited) + if (inited) return NULL; - isa = xcalloc(1, sizeof(*isa)); + set_tarval_output_modes(); + + isa = xmalloc(sizeof(*isa)); memcpy(isa, &ia32_isa_template, sizeof(*isa)); ia32_register_init(isa); ia32_create_opcodes(); - ia32_register_copy_attr_func(); + + if ((ARCH_INTEL(isa->arch) && isa->arch < arch_pentium_4) || + (ARCH_AMD(isa->arch) && isa->arch < arch_athlon)) + /* no SSE2 for these cpu's */ + isa->fp_kind = fp_x87; + + if (ARCH_INTEL(isa->opt_arch) && isa->opt_arch >= arch_pentium_4) { + /* Pentium 4 don't like inc and dec instructions */ + isa->opt &= ~IA32_OPT_INCDEC; + } isa->regs_16bit = pmap_create(); isa->regs_8bit = pmap_create(); + isa->types = pmap_create(); + isa->tv_ent = pmap_create(); + isa->out = file_handle; ia32_build_16bit_reg_map(isa->regs_16bit); ia32_build_8bit_reg_map(isa->regs_8bit); + /* patch register names of x87 registers */ + if (USE_x87(isa)) { + ia32_st_regs[0].name = "st"; + ia32_st_regs[1].name = "st(1)"; + ia32_st_regs[2].name = "st(2)"; + ia32_st_regs[3].name = "st(3)"; + ia32_st_regs[4].name = "st(4)"; + ia32_st_regs[5].name = "st(5)"; + ia32_st_regs[6].name = "st(6)"; + ia32_st_regs[7].name = "st(7)"; + } + #ifndef NDEBUG - isa->name_obst = xcalloc(1, sizeof(*(isa->name_obst))); + isa->name_obst = xmalloc(sizeof(*isa->name_obst)); obstack_init(isa->name_obst); isa->name_obst_size = 0; #endif /* NDEBUG */ + ia32_handle_intrinsics(); + ia32_switch_section(NULL, NO_SECTION); + fprintf(isa->out, "\t.intel_syntax\n"); + inited = 1; return isa; @@ -802,8 +1451,13 @@ static void *ia32_init(void) { static void ia32_done(void *self) { ia32_isa_t *isa = self; + /* emit now all global declarations */ + ia32_gen_decls(isa->out); + pmap_destroy(isa->regs_16bit); pmap_destroy(isa->regs_8bit); + pmap_destroy(isa->tv_ent); + pmap_destroy(isa->types); #ifndef NDEBUG //printf("name obst size = %d bytes\n", isa->name_obst_size); @@ -814,14 +1468,28 @@ static void ia32_done(void *self) { } - +/** + * Return the number of register classes for this architecture. + * We report always these: + * - the general purpose registers + * - the SSE floating point register set + * - the virtual floating point registers + */ static int ia32_get_n_reg_class(const void *self) { - return N_CLASSES; + return 3; } +/** + * Return the register class for index i. + */ static const arch_register_class_t *ia32_get_reg_class(const void *self, int i) { - assert(i >= 0 && i < N_CLASSES && "Invalid ia32 register class requested."); - return &ia32_reg_classes[i]; + assert(i >= 0 && i < 3 && "Invalid ia32 register class requested."); + if (i == 0) + return &ia32_reg_classes[CLASS_ia32_gp]; + else if (i == 1) + return &ia32_reg_classes[CLASS_ia32_xmm]; + else + return &ia32_reg_classes[CLASS_ia32_vfp]; } /** @@ -831,8 +1499,10 @@ static const arch_register_class_t *ia32_get_reg_class(const void *self, int i) * @return A register class which can hold values of the given mode. */ const arch_register_class_t *ia32_get_reg_class_for_mode(const void *self, const ir_mode *mode) { - if (mode_is_float(mode)) - return &ia32_reg_classes[CLASS_ia32_fp]; + const ia32_isa_t *isa = self; + if (mode_is_float(mode)) { + return USE_SSE2(isa) ? &ia32_reg_classes[CLASS_ia32_xmm] : &ia32_reg_classes[CLASS_ia32_vfp]; + } else return &ia32_reg_classes[CLASS_ia32_gp]; } @@ -843,7 +1513,8 @@ const arch_register_class_t *ia32_get_reg_class_for_mode(const void *self, const * @param method_type The type of the method (procedure) in question. * @param abi The abi object to be modified */ -void ia32_get_call_abi(const void *self, ir_type *method_type, be_abi_call_t *abi) { +static void ia32_get_call_abi(const void *self, ir_type *method_type, be_abi_call_t *abi) { + const ia32_isa_t *isa = self; ir_type *tp; ir_mode *mode; unsigned cc = get_method_calling_convention(method_type); @@ -853,14 +1524,16 @@ void ia32_get_call_abi(const void *self, ir_type *method_type, be_abi_call_t *ab int i, ignore_1, ignore_2; ir_mode **modes; const arch_register_t *reg; - be_abi_call_flags_t call_flags; + be_abi_call_flags_t call_flags = be_abi_call_get_flags(abi); + + unsigned use_push = !IS_P6_ARCH(isa->opt_arch); /* set abi flags for calls */ - call_flags.bits.left_to_right = 0; - call_flags.bits.store_args_sequential = 0; - call_flags.bits.try_omit_fp = 1; - call_flags.bits.fp_free = 0; - call_flags.bits.call_has_imm = 1; + call_flags.bits.left_to_right = 0; /* always last arg first on stack */ + call_flags.bits.store_args_sequential = use_push; + /* call_flags.bits.try_omit_fp not changed: can handle both settings */ + call_flags.bits.fp_free = 0; /* the frame pointer is fixed in IA32 */ + call_flags.bits.call_has_imm = 1; /* IA32 calls can have immediate address */ /* set stack parameter passing style */ be_abi_call_set_flags(abi, call_flags, &ia32_abi_callbacks); @@ -916,11 +1589,17 @@ void ia32_get_call_abi(const void *self, ir_type *method_type, be_abi_call_t *ab be_abi_call_res_reg(abi, 1, &ia32_gp_regs[REG_EDX]); } else if (n == 1) { + const arch_register_t *reg; + tp = get_method_res_type(method_type, 0); assert(is_atomic_type(tp)); mode = get_type_mode(tp); - be_abi_call_res_reg(abi, 0, mode_is_float(mode) ? &ia32_fp_regs[REG_XMM0] : &ia32_gp_regs[REG_EAX]); + reg = mode_is_float(mode) ? + (USE_SSE2(isa) ? &ia32_xmm_regs[REG_XMM0] : &ia32_vfp_regs[REG_VF0]) : + &ia32_gp_regs[REG_EAX]; + + be_abi_call_res_reg(abi, 0, reg); } } @@ -938,7 +1617,7 @@ const arch_irn_handler_t *ia32_get_irn_handler(const void *self) { } int ia32_to_appear_in_schedule(void *block_env, const ir_node *irn) { - return is_ia32_irn(irn); + return is_ia32_irn(irn) ? 1 : -1; } /** @@ -948,27 +1627,157 @@ static const arch_code_generator_if_t *ia32_get_code_generator_if(void *self) { return &ia32_code_gen_if; } +/** + * Returns the estimated execution time of an ia32 irn. + */ +static sched_timestep_t ia32_sched_exectime(void *env, const ir_node *irn) { + const arch_env_t *arch_env = env; + return is_ia32_irn(irn) ? ia32_get_op_estimated_cost(arch_get_irn_ops(arch_env, irn), irn) : 1; +} + list_sched_selector_t ia32_sched_selector; /** * Returns the reg_pressure scheduler with to_appear_in_schedule() overloaded */ static const list_sched_selector_t *ia32_get_list_sched_selector(const void *self) { - memcpy(&ia32_sched_selector, reg_pressure_selector, sizeof(list_sched_selector_t)); +// memcpy(&ia32_sched_selector, reg_pressure_selector, sizeof(list_sched_selector_t)); + memcpy(&ia32_sched_selector, trivial_selector, sizeof(list_sched_selector_t)); + ia32_sched_selector.exectime = ia32_sched_exectime; ia32_sched_selector.to_appear_in_schedule = ia32_to_appear_in_schedule; return &ia32_sched_selector; } +/** + * Returns the necessary byte alignment for storing a register of given class. + */ +static int ia32_get_reg_class_alignment(const void *self, const arch_register_class_t *cls) { + ir_mode *mode = arch_register_class_mode(cls); + int bytes = get_mode_size_bytes(mode); + + if (mode_is_float(mode) && bytes > 8) + return 16; + return bytes; +} + +static ia32_intrinsic_env_t intrinsic_env = { NULL, NULL }; + +/** + * Returns the libFirm configuration parameter for this backend. + */ +static const backend_params *ia32_get_libfirm_params(void) { + static const arch_dep_params_t ad = { + 1, /* also use subs */ + 4, /* maximum shifts */ + 31, /* maximum shift amount */ + + 1, /* allow Mulhs */ + 1, /* allow Mulus */ + 32 /* Mulh allowed up to 32 bit */ + }; + static backend_params p = { + NULL, /* no additional opcodes */ + NULL, /* will be set later */ + 1, /* need dword lowering */ + ia32_create_intrinsic_fkt, + &intrinsic_env, /* context for ia32_create_intrinsic_fkt */ + }; + + p.dep_param = &ad; + return &p; +} #ifdef WITH_LIBCORE + +/* instruction set architectures. */ +static const lc_opt_enum_int_items_t arch_items[] = { + { "386", arch_i386, }, + { "486", arch_i486, }, + { "pentium", arch_pentium, }, + { "586", arch_pentium, }, + { "pentiumpro", arch_pentium_pro, }, + { "686", arch_pentium_pro, }, + { "pentiummmx", arch_pentium_mmx, }, + { "pentium2", arch_pentium_2, }, + { "p2", arch_pentium_2, }, + { "pentium3", arch_pentium_3, }, + { "p3", arch_pentium_3, }, + { "pentium4", arch_pentium_4, }, + { "p4", arch_pentium_4, }, + { "pentiumm", arch_pentium_m, }, + { "pm", arch_pentium_m, }, + { "core", arch_core, }, + { "k6", arch_k6, }, + { "athlon", arch_athlon, }, + { "athlon64", arch_athlon_64, }, + { "opteron", arch_opteron, }, + { NULL, 0 } +}; + +static lc_opt_enum_int_var_t arch_var = { + &ia32_isa_template.arch, arch_items +}; + +static lc_opt_enum_int_var_t opt_arch_var = { + &ia32_isa_template.opt_arch, arch_items +}; + +static const lc_opt_enum_int_items_t fp_unit_items[] = { + { "x87" , fp_x87 }, + { "sse2", fp_sse2 }, + { NULL, 0 } +}; + +static lc_opt_enum_int_var_t fp_unit_var = { + &ia32_isa_template.fp_kind, fp_unit_items +}; + +static const lc_opt_enum_int_items_t gas_items[] = { + { "linux", ASM_LINUX_GAS }, + { "mingw", ASM_MINGW_GAS }, + { NULL, 0 } +}; + +static lc_opt_enum_int_var_t gas_var = { + (int *)&asm_flavour, gas_items +}; + +static const lc_opt_table_entry_t ia32_options[] = { + LC_OPT_ENT_ENUM_INT("arch", "select the instruction architecture", &arch_var), + LC_OPT_ENT_ENUM_INT("opt", "optimize for instruction architecture", &opt_arch_var), + LC_OPT_ENT_ENUM_INT("fpunit", "select the floating point unit", &fp_unit_var), + LC_OPT_ENT_NEGBIT("noaddrmode", "do not use address mode", &ia32_isa_template.opt, IA32_OPT_DOAM), + LC_OPT_ENT_NEGBIT("nolea", "do not optimize for LEAs", &ia32_isa_template.opt, IA32_OPT_LEA), + LC_OPT_ENT_NEGBIT("noplacecnst", "do not place constants", &ia32_isa_template.opt, IA32_OPT_PLACECNST), + LC_OPT_ENT_NEGBIT("noimmop", "no operations with immediates", &ia32_isa_template.opt, IA32_OPT_IMMOPS), + LC_OPT_ENT_NEGBIT("noextbb", "do not use extended basic block scheduling", &ia32_isa_template.opt, IA32_OPT_EXTBB), + LC_OPT_ENT_ENUM_INT("gasmode", "set the GAS compatibility mode", &gas_var), + { NULL } +}; + +/** + * Register command line options for the ia32 backend. + * + * Options so far: + * + * ia32-arch=arch create instruction for arch + * ia32-opt=arch optimize for run on arch + * ia32-fpunit=unit select floating point unit (x87 or SSE2) + * ia32-incdec optimize for inc/dec + * ia32-noaddrmode do not use address mode + * ia32-nolea do not optimize for LEAs + * ia32-noplacecnst do not place constants, + * ia32-noimmop no operations with immediates + * ia32-noextbb do not use extended basic block scheduling + * ia32-gasmode set the GAS compatibility mode + */ static void ia32_register_options(lc_opt_entry_t *ent) { + lc_opt_entry_t *be_grp_ia32 = lc_opt_get_grp(ent, "ia32"); + lc_opt_add_table(be_grp_ia32, ia32_options); } #endif /* WITH_LIBCORE */ const arch_isa_if_t ia32_isa_if = { -#ifdef WITH_LIBCORE - ia32_register_options, -#endif ia32_init, ia32_done, ia32_get_n_reg_class, @@ -977,5 +1786,10 @@ const arch_isa_if_t ia32_isa_if = { ia32_get_call_abi, ia32_get_irn_handler, ia32_get_code_generator_if, - ia32_get_list_sched_selector + ia32_get_list_sched_selector, + ia32_get_reg_class_alignment, + ia32_get_libfirm_params, +#ifdef WITH_LIBCORE + ia32_register_options +#endif };