X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fia32%2Fia32_emitter.c;h=94769152e3e2a042fd00ad9979baaf9925bf5b85;hb=752769d9f2f8b8bb4711db0942cff982a1bd1029;hp=088ef5141602a6de9f12b0f8176c50fad939485a;hpb=ff4acf5218ebcb851bb04d2cc4181b456833d5d8;p=libfirm diff --git a/ir/be/ia32/ia32_emitter.c b/ir/be/ia32/ia32_emitter.c index 088ef5141..94769152e 100644 --- a/ir/be/ia32/ia32_emitter.c +++ b/ir/be/ia32/ia32_emitter.c @@ -23,9 +23,7 @@ * @author Christian Wuerdig, Matthias Braun * @version $Id$ */ -#ifdef HAVE_CONFIG_H #include "config.h" -#endif #include @@ -39,18 +37,20 @@ #include "irargs_t.h" #include "irprog_t.h" #include "iredges_t.h" +#include "irtools.h" #include "execfreq.h" #include "error.h" #include "raw_bitset.h" #include "dbginfo.h" +#include "lc_opts.h" -#include "../besched_t.h" -#include "../benode_t.h" +#include "../besched.h" +#include "../benode.h" #include "../beabi.h" #include "../be_dbgout.h" #include "../beemitter.h" #include "../begnuas.h" -#include "../beirg_t.h" +#include "../beirg.h" #include "../be_dbgout.h" #include "ia32_emitter.h" @@ -68,13 +68,12 @@ DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;) #define SNPRINTF_BUF_LEN 128 -static const arch_env_t *arch_env; static const ia32_isa_t *isa; static ia32_code_gen_t *cg; -static int do_pic; static char pic_base_label[128]; static ir_label_t exc_label_id; static int mark_spill_reload = 0; +static int do_pic; /** Return the next block in Block schedule */ static ir_node *get_prev_block_sched(const ir_node *block) @@ -82,6 +81,7 @@ static ir_node *get_prev_block_sched(const ir_node *block) return get_irn_link(block); } +/** Checks if the current block is a fall-through target. */ static int is_fallthrough(const ir_node *cfgpred) { ir_node *pred; @@ -95,11 +95,18 @@ static int is_fallthrough(const ir_node *cfgpred) return 1; } +/** + * returns non-zero if the given block needs a label + * because of being a jump-target (and not a fall-through) + */ static int block_needs_label(const ir_node *block) { int need_label = 1; int n_cfgpreds = get_Block_n_cfgpreds(block); + if (has_Block_entity(block)) + return 1; + if (n_cfgpreds == 0) { need_label = 0; } else if (n_cfgpreds == 1) { @@ -129,7 +136,7 @@ static const arch_register_t *get_in_reg(const ir_node *irn, int pos) in register we need. */ op = get_irn_n(irn, pos); - reg = arch_get_irn_register(arch_env, op); + reg = arch_get_irn_register(op); assert(reg && "no in register found"); @@ -138,10 +145,7 @@ static const arch_register_t *get_in_reg(const ir_node *irn, int pos) /* in case of unknown register: just return a valid register */ if (reg == &ia32_gp_regs[REG_GP_UKNWN]) { - const arch_register_req_t *req; - - /* ask for the requirements */ - req = arch_get_register_req(arch_env, irn, pos); + const arch_register_req_t *req = arch_get_register_req(irn, pos); if (arch_register_req_is(req, limited)) { /* in case of limited requirements: get the first allowed register */ @@ -171,9 +175,9 @@ static const arch_register_t *get_out_reg(const ir_node *irn, int pos) if (get_irn_mode(irn) != mode_T) { assert(pos == 0); - reg = arch_get_irn_register(arch_env, irn); + reg = arch_get_irn_register(irn); } else if (is_ia32_irn(irn)) { - reg = get_ia32_out_reg(irn, pos); + reg = arch_irn_get_register(irn, pos); } else { const ir_edge_t *edge; @@ -181,7 +185,7 @@ static const arch_register_t *get_out_reg(const ir_node *irn, int pos) proj = get_edge_src_irn(edge); assert(is_Proj(proj) && "non-Proj from mode_T node"); if (get_Proj_proj(proj) == pos) { - reg = arch_get_irn_register(arch_env, proj); + reg = arch_get_irn_register(proj); break; } } @@ -212,6 +216,9 @@ static char *get_unique_label(char *buf, size_t buflen, const char *prefix) * |_| |_| *************************************************************/ +/** + * Emit the name of the 8bit low register + */ static void emit_8bit_register(const arch_register_t *reg) { const char *reg_name = arch_register_get_name(reg); @@ -221,6 +228,18 @@ static void emit_8bit_register(const arch_register_t *reg) be_emit_char('l'); } +/** + * Emit the name of the 8bit high register + */ +static void emit_8bit_register_high(const arch_register_t *reg) +{ + const char *reg_name = arch_register_get_name(reg); + + be_emit_char('%'); + be_emit_char(reg_name[1]); + be_emit_char('h'); +} + static void emit_16bit_register(const arch_register_t *reg) { const char *reg_name = ia32_get_mapped_reg_name(isa->regs_16bit, reg); @@ -229,6 +248,12 @@ static void emit_16bit_register(const arch_register_t *reg) be_emit_string(reg_name); } +/** + * emit a register, possible shortened by a mode + * + * @param reg the register + * @param mode the mode of the register or NULL for full register + */ static void emit_register(const arch_register_t *reg, const ir_mode *mode) { const char *reg_name; @@ -255,12 +280,53 @@ void ia32_emit_source_register(const ir_node *node, int pos) emit_register(reg, NULL); } -static void emit_ia32_Immediate(const ir_node *node); +static void ia32_emit_entity(ir_entity *entity, int no_pic_adjust) +{ + set_entity_backend_marked(entity, 1); + be_gas_emit_entity(entity); + + if (get_entity_owner(entity) == get_tls_type()) { + if (get_entity_visibility(entity) == visibility_external_allocated) { + be_emit_cstring("@INDNTPOFF"); + } else { + be_emit_cstring("@NTPOFF"); + } + } + + if (do_pic && !no_pic_adjust) { + be_emit_char('-'); + be_emit_string(pic_base_label); + } +} + +static void emit_ia32_Immediate_no_prefix(const ir_node *node) +{ + const ia32_immediate_attr_t *attr = get_ia32_immediate_attr_const(node); + + if (attr->symconst != NULL) { + if (attr->sc_sign) + be_emit_char('-'); + ia32_emit_entity(attr->symconst, attr->no_pic_adjust); + } + if (attr->symconst == NULL || attr->offset != 0) { + if (attr->symconst != NULL) { + be_emit_irprintf("%+d", attr->offset); + } else { + be_emit_irprintf("0x%X", attr->offset); + } + } +} + +static void emit_ia32_Immediate(const ir_node *node) +{ + be_emit_char('$'); + emit_ia32_Immediate_no_prefix(node); +} void ia32_emit_8bit_source_register_or_immediate(const ir_node *node, int pos) { const arch_register_t *reg; - ir_node *in = get_irn_n(node, pos); + const ir_node *in = get_irn_n(node, pos); if (is_ia32_Immediate(in)) { emit_ia32_Immediate(in); return; @@ -270,6 +336,25 @@ void ia32_emit_8bit_source_register_or_immediate(const ir_node *node, int pos) emit_8bit_register(reg); } +void ia32_emit_8bit_high_source_register(const ir_node *node, int pos) +{ + const arch_register_t *reg = get_in_reg(node, pos); + emit_8bit_register_high(reg); +} + +void ia32_emit_16bit_source_register_or_immediate(const ir_node *node, int pos) +{ + const arch_register_t *reg; + const ir_node *in = get_irn_n(node, pos); + if (is_ia32_Immediate(in)) { + emit_ia32_Immediate(in); + return; + } + + reg = get_in_reg(node, pos); + emit_16bit_register(reg); +} + void ia32_emit_dest_register(const ir_node *node, int pos) { const arch_register_t *reg = get_out_reg(node, pos); @@ -277,6 +362,13 @@ void ia32_emit_dest_register(const ir_node *node, int pos) emit_register(reg, NULL); } +void ia32_emit_dest_register_size(const ir_node *node, int pos) +{ + const arch_register_t *reg = get_out_reg(node, pos); + + emit_register(reg, get_ia32_ls_mode(node)); +} + void ia32_emit_8bit_dest_register(const ir_node *node, int pos) { const arch_register_t *reg = get_out_reg(node, pos); @@ -372,11 +464,13 @@ void ia32_emit_xmm_mode_suffix_s(const ir_node *node) be_emit_char(get_xmm_mode_suffix(mode)); } -void ia32_emit_extend_suffix(const ir_mode *mode) +void ia32_emit_extend_suffix(const ir_node *node) { + ir_mode *mode = get_ia32_ls_mode(node); if (get_mode_size_bits(mode) == 32) return; be_emit_char(mode_is_signed(mode) ? 's' : 'z'); + ia32_emit_mode_suffix_mode(mode); } void ia32_emit_source_register_or_immediate(const ir_node *node, int pos) @@ -405,9 +499,9 @@ static ir_node *get_cfop_target_block(const ir_node *irn) */ static void ia32_emit_block_name(const ir_node *block) { - if (has_Block_label(block)) { - be_emit_string(be_gas_block_label_prefix()); - be_emit_irprintf("%lu", get_Block_label(block)); + if (has_Block_entity(block)) { + ir_entity *entity = get_Block_entity(block); + be_gas_emit_entity(entity); } else { be_emit_cstring(BLOCK_PREFIX); be_emit_irprintf("%ld", get_irn_node_nr(block)); @@ -423,59 +517,117 @@ static void ia32_emit_cfop_target(const ir_node *node) ia32_emit_block_name(block); } -/* - * coding of conditions - */ -struct cmp2conditon_t { - const char *name; - int num; -}; - /* * positive conditions for signed compares */ -static const struct cmp2conditon_t cmp2condition_s[] = { - { NULL, pn_Cmp_False }, /* always false */ - { "e", pn_Cmp_Eq }, /* == */ - { "l", pn_Cmp_Lt }, /* < */ - { "le", pn_Cmp_Le }, /* <= */ - { "g", pn_Cmp_Gt }, /* > */ - { "ge", pn_Cmp_Ge }, /* >= */ - { "ne", pn_Cmp_Lg }, /* != */ - { NULL, pn_Cmp_Leg}, /* always true */ +static const char *const cmp2condition_s[] = { + NULL, /* always false */ + "e", /* == */ + "l", /* < */ + "le", /* <= */ + "g", /* > */ + "ge", /* >= */ + "ne", /* != */ + NULL /* always true */ }; /* * positive conditions for unsigned compares */ -static const struct cmp2conditon_t cmp2condition_u[] = { - { NULL, pn_Cmp_False }, /* always false */ - { "e", pn_Cmp_Eq }, /* == */ - { "b", pn_Cmp_Lt }, /* < */ - { "be", pn_Cmp_Le }, /* <= */ - { "a", pn_Cmp_Gt }, /* > */ - { "ae", pn_Cmp_Ge }, /* >= */ - { "ne", pn_Cmp_Lg }, /* != */ - { NULL, pn_Cmp_Leg }, /* always true */ +static const char *const cmp2condition_u[] = { + NULL, /* always false */ + "e", /* == */ + "b", /* < */ + "be", /* <= */ + "a", /* > */ + "ae", /* >= */ + "ne", /* != */ + NULL /* always true */ }; +/** + * Emit the suffix for a compare instruction. + */ static void ia32_emit_cmp_suffix(int pnc) { const char *str; - if ((pnc & ia32_pn_Cmp_float) || (pnc & ia32_pn_Cmp_unsigned)) { - pnc = pnc & 7; - assert(cmp2condition_u[pnc].num == pnc); - str = cmp2condition_u[pnc].name; + if (pnc == ia32_pn_Cmp_parity) { + be_emit_char('p'); + return; + } + if (pnc & ia32_pn_Cmp_float || pnc & ia32_pn_Cmp_unsigned) { + str = cmp2condition_u[pnc & 7]; } else { - pnc = pnc & 7; - assert(cmp2condition_s[pnc].num == pnc); - str = cmp2condition_s[pnc].name; + str = cmp2condition_s[pnc & 7]; } be_emit_string(str); } +typedef enum ia32_emit_mod_t { + EMIT_RESPECT_LS = 1U << 0, + EMIT_ALTERNATE_AM = 1U << 1, + EMIT_LONG = 1U << 2 +} ia32_emit_mod_t; + +/** + * Emits address mode. + */ +void ia32_emit_am(const ir_node *node) +{ + ir_entity *ent = get_ia32_am_sc(node); + int offs = get_ia32_am_offs_int(node); + ir_node *base = get_irn_n(node, n_ia32_base); + int has_base = !is_ia32_NoReg_GP(base); + ir_node *index = get_irn_n(node, n_ia32_index); + int has_index = !is_ia32_NoReg_GP(index); + + /* just to be sure... */ + assert(!is_ia32_use_frame(node) || get_ia32_frame_ent(node) != NULL); + + /* emit offset */ + if (ent != NULL) { + const ia32_attr_t *attr = get_ia32_attr_const(node); + if (is_ia32_am_sc_sign(node)) + be_emit_char('-'); + ia32_emit_entity(ent, attr->data.am_sc_no_pic_adjust); + } + + /* also handle special case if nothing is set */ + if (offs != 0 || (ent == NULL && !has_base && !has_index)) { + if (ent != NULL) { + be_emit_irprintf("%+d", offs); + } else { + be_emit_irprintf("%d", offs); + } + } + + if (has_base || has_index) { + be_emit_char('('); + + /* emit base */ + if (has_base) { + const arch_register_t *reg = get_in_reg(node, n_ia32_base); + emit_register(reg, NULL); + } + + /* emit index + scale */ + if (has_index) { + const arch_register_t *reg = get_in_reg(node, n_ia32_index); + int scale; + be_emit_char(','); + emit_register(reg, NULL); + + scale = get_ia32_am_scale(node); + if (scale > 0) { + be_emit_irprintf(",%d", 1 << scale); + } + } + be_emit_char(')'); + } +} + /** * fmt parameter output * ---- ---------------------- --------------------------------------------- @@ -492,9 +644,12 @@ static void ia32_emit_cmp_suffix(int pnc) * %Sx source register x * %s const char* string * %u unsigned int unsigned int + * %d signed int signed int * * x starts at 0 * # modifier for %ASx, %D and %S uses ls mode of node to alter register width + * * modifier does not prefix immediates with $, but AM with * + * l modifier for %lu and %ld */ static void ia32_emitf(const ir_node *node, const char *fmt, ...) { @@ -502,8 +657,8 @@ static void ia32_emitf(const ir_node *node, const char *fmt, ...) va_start(ap, fmt); for (;;) { - const char *start = fmt; - const ir_mode *mode = NULL; + const char *start = fmt; + ia32_emit_mod_t mod = 0; while (*fmt != '%' && *fmt != '\n' && *fmt != '\0') ++fmt; @@ -523,8 +678,18 @@ static void ia32_emitf(const ir_node *node, const char *fmt, ...) break; ++fmt; + if (*fmt == '*') { + mod |= EMIT_ALTERNATE_AM; + ++fmt; + } + if (*fmt == '#') { - mode = get_ia32_ls_mode(node); + mod |= EMIT_RESPECT_LS; + ++fmt; + } + + if (*fmt == 'l') { + mod |= EMIT_LONG; ++fmt; } @@ -536,11 +701,16 @@ static void ia32_emitf(const ir_node *node, const char *fmt, ...) case 'A': { switch (*fmt++) { case 'M': + if (mod & EMIT_ALTERNATE_AM) + be_emit_char('*'); + ia32_emit_am(node); break; case 'R': { const arch_register_t *reg = va_arg(ap, const arch_register_t*); + if (mod & EMIT_ALTERNATE_AM) + be_emit_char('*'); if (get_ia32_op_type(node) == ia32_AddrModeS) { ia32_emit_am(node); } else { @@ -551,6 +721,8 @@ static void ia32_emitf(const ir_node *node, const char *fmt, ...) case 'S': if (get_ia32_op_type(node) == ia32_AddrModeS) { + if (mod & EMIT_ALTERNATE_AM) + be_emit_char('*'); ia32_emit_am(node); ++fmt; } else { @@ -573,12 +745,14 @@ static void ia32_emitf(const ir_node *node, const char *fmt, ...) pos = *fmt++ - '0'; reg = get_out_reg(node, pos); - emit_register(reg, mode); + emit_register(reg, mod & EMIT_RESPECT_LS ? get_ia32_ls_mode(node) : NULL); break; } case 'I': - emit_ia32_Immediate(node); + if (!(mod & EMIT_ALTERNATE_AM)) + be_emit_char('$'); + emit_ia32_Immediate_no_prefix(node); break; case 'L': @@ -613,10 +787,16 @@ emit_S: pos = *fmt++ - '0'; in = get_irn_n(node, pos); if (is_ia32_Immediate(in)) { - emit_ia32_Immediate(in); + if (!(mod & EMIT_ALTERNATE_AM)) + be_emit_char('$'); + emit_ia32_Immediate_no_prefix(in); } else { - const arch_register_t *reg = get_in_reg(node, pos); - emit_register(reg, mode); + const arch_register_t *reg; + + if (mod & EMIT_ALTERNATE_AM) + be_emit_char('*'); + reg = get_in_reg(node, pos); + emit_register(reg, mod & EMIT_RESPECT_LS ? get_ia32_ls_mode(node) : NULL); } break; } @@ -627,15 +807,29 @@ emit_S: break; } - case 'u': { - unsigned num = va_arg(ap, unsigned); - be_emit_irprintf("%u", num); + case 'u': + if (mod & EMIT_LONG) { + unsigned long num = va_arg(ap, unsigned long); + be_emit_irprintf("%lu", num); + } else { + unsigned num = va_arg(ap, unsigned); + be_emit_irprintf("%u", num); + } + break; + + case 'd': + if (mod & EMIT_LONG) { + long num = va_arg(ap, long); + be_emit_irprintf("%ld", num); + } else { + int num = va_arg(ap, int); + be_emit_irprintf("%d", num); + } break; - } default: unknown: - panic("unknown conversion"); + panic("unknown format conversion in ia32_emitf()"); } } @@ -698,85 +892,6 @@ void ia32_emit_unop(const ir_node *node, int pos) ia32_emitf(node, fmt); } -static void ia32_emit_entity(ir_entity *entity, int no_pic_adjust) -{ - ident *id; - - set_entity_backend_marked(entity, 1); - id = get_entity_ld_ident(entity); - be_emit_ident(id); - - if (get_entity_owner(entity) == get_tls_type()) { - if (get_entity_visibility(entity) == visibility_external_allocated) { - be_emit_cstring("@INDNTPOFF"); - } else { - be_emit_cstring("@NTPOFF"); - } - } - - if (!no_pic_adjust && do_pic) { - /* TODO: only do this when necessary */ - be_emit_char('-'); - be_emit_string(pic_base_label); - } -} - -/** - * Emits address mode. - */ -void ia32_emit_am(const ir_node *node) -{ - ir_entity *ent = get_ia32_am_sc(node); - int offs = get_ia32_am_offs_int(node); - ir_node *base = get_irn_n(node, n_ia32_base); - int has_base = !is_ia32_NoReg_GP(base); - ir_node *index = get_irn_n(node, n_ia32_index); - int has_index = !is_ia32_NoReg_GP(index); - - /* just to be sure... */ - assert(!is_ia32_use_frame(node) || get_ia32_frame_ent(node) != NULL); - - /* emit offset */ - if (ent != NULL) { - if (is_ia32_am_sc_sign(node)) - be_emit_char('-'); - ia32_emit_entity(ent, 0); - } - - /* also handle special case if nothing is set */ - if (offs != 0 || (ent == NULL && !has_base && !has_index)) { - if (ent != NULL) { - be_emit_irprintf("%+d", offs); - } else { - be_emit_irprintf("%d", offs); - } - } - - if (has_base || has_index) { - be_emit_char('('); - - /* emit base */ - if (has_base) { - const arch_register_t *reg = get_in_reg(node, n_ia32_base); - emit_register(reg, NULL); - } - - /* emit index + scale */ - if (has_index) { - const arch_register_t *reg = get_in_reg(node, n_ia32_index); - int scale; - be_emit_char(','); - emit_register(reg, NULL); - - scale = get_ia32_am_scale(node); - if (scale > 0) { - be_emit_irprintf(",%d", 1 << scale); - } - } - be_emit_char(')'); - } -} - static void emit_ia32_IMul(const ir_node *node) { ir_node *left = get_irn_n(node, n_ia32_IMul_left); @@ -875,6 +990,12 @@ static int determine_final_pnc(const ir_node *node, int flags_pos, return pnc; } +static pn_Cmp ia32_get_negated_pnc(pn_Cmp pnc) +{ + ir_mode *mode = pnc & ia32_pn_Cmp_float ? mode_F : mode_Iu; + return get_negated_pnc(pnc, mode); +} + void ia32_emit_cmp_suffix_node(const ir_node *node, int flags_pos) { @@ -883,13 +1004,8 @@ void ia32_emit_cmp_suffix_node(const ir_node *node, pn_Cmp pnc = get_ia32_condcode(node); pnc = determine_final_pnc(node, flags_pos, pnc); - if (attr->data.ins_permuted) { - if (pnc & ia32_pn_Cmp_float) { - pnc = get_negated_pnc(pnc, mode_F); - } else { - pnc = get_negated_pnc(pnc, mode_Iu); - } - } + if (attr->data.ins_permuted) + pnc = ia32_get_negated_pnc(pnc); ia32_emit_cmp_suffix(pnc); } @@ -961,11 +1077,7 @@ static void emit_ia32_Jcc(const ir_node *node) proj_true = proj_false; proj_false = t; - if (pnc & ia32_pn_Cmp_float) { - pnc = get_negated_pnc(pnc, mode_F); - } else { - pnc = get_negated_pnc(pnc, mode_Iu); - } + pnc = ia32_get_negated_pnc(pnc); } if (pnc & ia32_pn_Cmp_float) { @@ -1024,17 +1136,15 @@ static void emit_ia32_CMov(const ir_node *node) { const ia32_attr_t *attr = get_ia32_attr_const(node); int ins_permuted = attr->data.ins_permuted; - const arch_register_t *out = arch_get_irn_register(arch_env, node); + const arch_register_t *out = arch_irn_get_register(node, pn_ia32_res); pn_Cmp pnc = get_ia32_condcode(node); const arch_register_t *in_true; const arch_register_t *in_false; pnc = determine_final_pnc(node, n_ia32_CMov_eflags, pnc); - in_true = arch_get_irn_register(arch_env, - get_irn_n(node, n_ia32_CMov_val_true)); - in_false = arch_get_irn_register(arch_env, - get_irn_n(node, n_ia32_CMov_val_false)); + in_true = arch_get_irn_register(get_irn_n(node, n_ia32_CMov_val_true)); + in_false = arch_get_irn_register(get_irn_n(node, n_ia32_CMov_val_false)); /* should be same constraint fullfilled? */ if (out == in_false) { @@ -1054,17 +1164,12 @@ static void emit_ia32_CMov(const ir_node *node) ia32_emitf(node, "\tmovl %R, %R\n", in_false, out); } - if (ins_permuted) { - if (pnc & ia32_pn_Cmp_float) { - pnc = get_negated_pnc(pnc, mode_F); - } else { - pnc = get_negated_pnc(pnc, mode_Iu); - } - } + if (ins_permuted) + pnc = ia32_get_negated_pnc(pnc); /* TODO: handling of Nans isn't correct yet */ - ia32_emitf(node, "\tcmov%P %AR, %#R\n", pnc, in_true, out); + ia32_emitf(node, "\tcmov%P %#AR, %#R\n", pnc, in_true, out); } /********************************************************* @@ -1200,7 +1305,7 @@ static void emit_ia32_SwitchJmp(const ir_node *node) /** * Emits code for a unconditional jump. */ -static void emit_Jmp(const ir_node *node) +static void emit_ia32_Jmp(const ir_node *node) { ir_node *block; @@ -1215,34 +1320,15 @@ static void emit_Jmp(const ir_node *node) } } -static void emit_ia32_Immediate(const ir_node *node) -{ - const ia32_immediate_attr_t *attr = get_ia32_immediate_attr_const(node); - - be_emit_char('$'); - if (attr->symconst != NULL) { - if (attr->sc_sign) - be_emit_char('-'); - ia32_emit_entity(attr->symconst, 0); - } - if (attr->symconst == NULL || attr->offset != 0) { - if (attr->symconst != NULL) { - be_emit_irprintf("%+d", attr->offset); - } else { - be_emit_irprintf("0x%X", attr->offset); - } - } -} - -/** - * Emit an inline assembler operand. - * - * @param node the ia32_ASM node - * @param s points to the operand (a %c) - * - * @return pointer to the first char in s NOT in the current operand - */ -static const char* emit_asm_operand(const ir_node *node, const char *s) +/** + * Emit an inline assembler operand. + * + * @param node the ia32_ASM node + * @param s points to the operand (a %c) + * + * @return pointer to the first char in s NOT in the current operand + */ +static const char* emit_asm_operand(const ir_node *node, const char *s) { const ia32_attr_t *ia32_attr = get_ia32_attr_const(node); const ia32_asm_attr_t *attr = CONST_CAST_IA32_ATTR(ia32_asm_attr_t, @@ -1262,7 +1348,7 @@ static const char* emit_asm_operand(const ir_node *node, const char *s) /* parse modifiers */ switch(c) { case 0: - ir_fprintf(stderr, "Warning: asm text (%+F) ends with %\n", node); + ir_fprintf(stderr, "Warning: asm text (%+F) ends with %%\n", node); be_emit_char('%'); return s + 1; case '%': @@ -1286,8 +1372,9 @@ static const char* emit_asm_operand(const ir_node *node, const char *s) case '9': break; default: - ir_fprintf(stderr, "Warning: asm text (%+F) contains unknown modifier " - "'%c' for asm op\n", node, c); + ir_fprintf(stderr, + "Warning: asm text (%+F) contains unknown modifier '%c' for asm op\n", + node, c); ++s; break; } @@ -1302,9 +1389,10 @@ static const char* emit_asm_operand(const ir_node *node, const char *s) s += p; } - if (num < 0 || num >= ARR_LEN(asm_regs)) { - ir_fprintf(stderr, "Error: Custom assembler references invalid " - "input/output (%+F)\n", node); + if (num < 0 || ARR_LEN(asm_regs) <= num) { + ir_fprintf(stderr, + "Error: Custom assembler references invalid input/output (%+F)\n", + node); return s; } asm_reg = & asm_regs[num]; @@ -1324,8 +1412,9 @@ static const char* emit_asm_operand(const ir_node *node, const char *s) reg = get_in_reg(node, asm_reg->inout_pos); } if (reg == NULL) { - ir_fprintf(stderr, "Warning: no register assigned for %d asm op " - "(%+F)\n", num, node); + ir_fprintf(stderr, + "Warning: no register assigned for %d asm op (%+F)\n", + num, node); return s; } @@ -1451,49 +1540,29 @@ static void emit_ia32_CopyB_i(const ir_node *node) /** * Emit code for conversions (I, FP), (FP, I) and (FP, FP). */ -static void emit_ia32_Conv_with_FP(const ir_node *node) +static void emit_ia32_Conv_with_FP(const ir_node *node, const char* conv_f, + const char* conv_d) { ir_mode *ls_mode = get_ia32_ls_mode(node); int ls_bits = get_mode_size_bits(ls_mode); - const char *conv; - - if (is_ia32_Conv_I2FP(node)) { - if (ls_bits == 32) { - conv = "si2ss"; - } else { - conv = "si2sd"; - } - } else if (is_ia32_Conv_FP2I(node)) { - if (ls_bits == 32) { - conv = "ss2si"; - } else { - conv = "sd2si"; - } - } else { - assert(is_ia32_Conv_FP2FP(node)); - if (ls_bits == 32) { - conv = "sd2ss"; - } else { - conv = "ss2sd"; - } - } + const char *conv = ls_bits == 32 ? conv_f : conv_d; ia32_emitf(node, "\tcvt%s %AS3, %D0\n", conv); } static void emit_ia32_Conv_I2FP(const ir_node *node) { - emit_ia32_Conv_with_FP(node); + emit_ia32_Conv_with_FP(node, "si2ss", "si2sd"); } static void emit_ia32_Conv_FP2I(const ir_node *node) { - emit_ia32_Conv_with_FP(node); + emit_ia32_Conv_with_FP(node, "ss2si", "sd2si"); } static void emit_ia32_Conv_FP2FP(const ir_node *node) { - emit_ia32_Conv_with_FP(node); + emit_ia32_Conv_with_FP(node, "sd2ss", "ss2sd"); } /** @@ -1502,23 +1571,23 @@ static void emit_ia32_Conv_FP2FP(const ir_node *node) static void emit_ia32_Conv_I2I(const ir_node *node) { ir_mode *smaller_mode = get_ia32_ls_mode(node); - int smaller_bits = get_mode_size_bits(smaller_mode); int signed_mode = mode_is_signed(smaller_mode); + const char *sign_suffix; assert(!mode_is_float(smaller_mode)); - assert(smaller_bits == 8 || smaller_bits == 16); - - if (signed_mode && - smaller_bits == 16 && - &ia32_gp_regs[REG_EAX] == get_out_reg(node, 0) && - &ia32_gp_regs[REG_EAX] == arch_get_irn_register(arch_env, get_irn_n(node, n_ia32_unary_op))) { - /* argument and result are both in EAX and signedness is ok: use the - * smaller cwtl opcode */ - ia32_emitf(node, "\tcwtl\n"); - } else { - const char *sign_suffix = signed_mode ? "s" : "z"; - ia32_emitf(node, "\tmov%s%Ml %#AS3, %D0\n", sign_suffix); - } + + sign_suffix = signed_mode ? "s" : "z"; + ia32_emitf(node, "\tmov%s%Ml %#AS3, %D0\n", sign_suffix); +} + +/** + * Emits a call + */ +static void emit_ia32_Call(const ir_node *node) +{ + /* Special case: Call must not have its immediates prefixed by $, instead + * address mode is prefixed by *. */ + ia32_emitf(node, "\tcall %*AS3\n"); } @@ -1532,24 +1601,6 @@ static void emit_ia32_Conv_I2I(const ir_node *node) * *******************************************/ -/** - * Emits a backend call - */ -static void emit_be_Call(const ir_node *node) -{ - ir_entity *ent = be_Call_get_entity(node); - - be_emit_cstring("\tcall "); - if (ent) { - ia32_emit_entity(ent, 1); - } else { - const arch_register_t *reg = get_in_reg(node, be_pos_Call_ptr); - be_emit_char('*'); - emit_register(reg, NULL); - } - be_emit_finish_line_gas(node); -} - /** * Emits code to increase stack pointer. */ @@ -1567,13 +1618,23 @@ static void emit_be_IncSP(const ir_node *node) } } +static inline bool is_unknown_reg(const arch_register_t *reg) +{ + if(reg == &ia32_gp_regs[REG_GP_UKNWN] + || reg == &ia32_xmm_regs[REG_XMM_UKNWN] + || reg == &ia32_vfp_regs[REG_VFP_UKNWN]) + return true; + + return false; +} + /** * Emits code for Copy/CopyKeep. */ static void Copy_emitter(const ir_node *node, const ir_node *op) { - const arch_register_t *in = arch_get_irn_register(arch_env, op); - const arch_register_t *out = arch_get_irn_register(arch_env, node); + const arch_register_t *in = arch_get_irn_register(op); + const arch_register_t *out = arch_get_irn_register(node); if (in == out) { return; @@ -1609,8 +1670,8 @@ static void emit_be_Perm(const ir_node *node) const arch_register_t *in0, *in1; const arch_register_class_t *cls0, *cls1; - in0 = arch_get_irn_register(arch_env, get_irn_n(node, 0)); - in1 = arch_get_irn_register(arch_env, get_irn_n(node, 1)); + in0 = arch_get_irn_register(get_irn_n(node, 0)); + in1 = arch_get_irn_register(get_irn_n(node, 1)); cls0 = arch_register_get_class(in0); cls1 = arch_register_get_class(in1); @@ -1745,6 +1806,18 @@ static void emit_ia32_GetEIP(const ir_node *node) ia32_emitf(node, "\tpopl %D0\n"); } +static void emit_ia32_ClimbFrame(const ir_node *node) +{ + const ia32_climbframe_attr_t *attr = get_ia32_climbframe_attr_const(node); + + ia32_emitf(node, "\tmovl %S0, %D0\n"); + ia32_emitf(node, "\tmovl $%u, %S1\n", attr->count); + ia32_emitf(NULL, BLOCK_PREFIX "%ld:\n", get_irn_node_nr(node)); + ia32_emitf(node, "\tmovl (%D0), %D0\n"); + ia32_emitf(node, "\tdec %S1\n"); + ia32_emitf(node, "\tjnz " BLOCK_PREFIX "%ld\n", get_irn_node_nr(node)); +} + static void emit_be_Return(const ir_node *node) { unsigned pop = be_Return_get_pop(node); @@ -1792,40 +1865,39 @@ static void ia32_register_emitters(void) ia32_register_spec_emitters(); /* other ia32 emitter functions */ + IA32_EMIT2(Conv_I2I8Bit, Conv_I2I); IA32_EMIT(Asm); IA32_EMIT(CMov); - IA32_EMIT(IMul); - IA32_EMIT(SwitchJmp); - IA32_EMIT(CopyB); - IA32_EMIT(CopyB_i); - IA32_EMIT(Conv_I2FP); - IA32_EMIT(Conv_FP2I); + IA32_EMIT(Call); + IA32_EMIT(Const); IA32_EMIT(Conv_FP2FP); + IA32_EMIT(Conv_FP2I); + IA32_EMIT(Conv_I2FP); IA32_EMIT(Conv_I2I); - IA32_EMIT2(Conv_I2I8Bit, Conv_I2I); - IA32_EMIT(Const); + IA32_EMIT(CopyB); + IA32_EMIT(CopyB_i); + IA32_EMIT(GetEIP); + IA32_EMIT(IMul); + IA32_EMIT(Jcc); IA32_EMIT(LdTls); IA32_EMIT(Minus64Bit); - IA32_EMIT(Jcc); - IA32_EMIT(GetEIP); + IA32_EMIT(SwitchJmp); + IA32_EMIT(ClimbFrame); + IA32_EMIT(Jmp); /* benode emitter */ - BE_EMIT(Call); - BE_EMIT(IncSP); BE_EMIT(Copy); BE_EMIT(CopyKeep); + BE_EMIT(IncSP); BE_EMIT(Perm); BE_EMIT(Return); - BE_IGN(RegParams); BE_IGN(Barrier); BE_IGN(Keep); + BE_IGN(Start); /* firm emitter */ - EMIT(Jmp); - IGN(Proj); IGN(Phi); - IGN(Start); #undef BE_EMIT #undef EMIT @@ -1968,7 +2040,7 @@ static void ia32_emit_block_header(ir_node *block) int i, arity; ir_exec_freq *exec_freq = cg->birg->exec_freq; - if (block == get_irg_end_block(irg) || block == get_irg_start_block(irg)) + if (block == get_irg_end_block(irg)) return; if (ia32_cg_config.label_alignment > 0) { @@ -1997,7 +2069,7 @@ static void ia32_emit_block_header(ir_node *block) } } - if (need_label || has_Block_label(block)) { + if (need_label) { ia32_emit_block_name(block); be_emit_char(':'); @@ -2013,12 +2085,16 @@ static void ia32_emit_block_header(ir_node *block) /* emit list of pred blocks in comment */ arity = get_irn_arity(block); - for (i = 0; i < arity; ++i) { - ir_node *predblock = get_Block_cfgpred_block(block, i); - be_emit_irprintf(" %d", get_irn_node_nr(predblock)); + if (arity <= 0) { + be_emit_cstring(" none"); + } else { + for (i = 0; i < arity; ++i) { + ir_node *predblock = get_Block_cfgpred_block(block, i); + be_emit_irprintf(" %d", get_irn_node_nr(predblock)); + } } if (exec_freq != NULL) { - be_emit_irprintf(" freq: %f", + be_emit_irprintf(", freq: %f", get_block_execfreq(exec_freq, block)); } be_emit_cstring(" */\n"); @@ -2097,8 +2173,7 @@ void ia32_gen_routine(ia32_code_gen_t *ia32_cg, ir_graph *irg) int i, n; cg = ia32_cg; - isa = (const ia32_isa_t*) cg->arch_env; - arch_env = cg->arch_env; + isa = cg->isa; do_pic = cg->birg->main_env->options->pic; ia32_register_emitters(); @@ -2157,6 +2232,986 @@ static const lc_opt_table_entry_t ia32_emitter_options[] = { LC_OPT_LAST }; +/* ==== Experimental binary emitter ==== */ + +static unsigned char reg_gp_map[N_ia32_gp_REGS]; +static unsigned char reg_mmx_map[N_ia32_mmx_REGS]; +static unsigned char reg_sse_map[N_ia32_xmm_REGS]; +static unsigned char pnc_map_signed[8]; +static unsigned char pnc_map_unsigned[8]; + +static void build_reg_map(void) +{ + reg_gp_map[REG_EAX] = 0x0; + reg_gp_map[REG_ECX] = 0x1; + reg_gp_map[REG_EDX] = 0x2; + reg_gp_map[REG_EBX] = 0x3; + reg_gp_map[REG_ESP] = 0x4; + reg_gp_map[REG_EBP] = 0x5; + reg_gp_map[REG_ESI] = 0x6; + reg_gp_map[REG_EDI] = 0x7; + + pnc_map_signed[pn_Cmp_False] = 0xFF; + pnc_map_signed[pn_Cmp_Eq] = 0x04; + pnc_map_signed[pn_Cmp_Lt] = 0x0C; + pnc_map_signed[pn_Cmp_Le] = 0x0E; + pnc_map_signed[pn_Cmp_Gt] = 0x0F; + pnc_map_signed[pn_Cmp_Ge] = 0x0D; + pnc_map_signed[pn_Cmp_Lg] = 0x05; + + pnc_map_unsigned[pn_Cmp_False] = 0xFF; + pnc_map_unsigned[pn_Cmp_Eq] = 0x04; + pnc_map_unsigned[pn_Cmp_Lt] = 0x02; + pnc_map_unsigned[pn_Cmp_Le] = 0x06; + pnc_map_unsigned[pn_Cmp_Gt] = 0x07; + pnc_map_unsigned[pn_Cmp_Ge] = 0x03; + pnc_map_unsigned[pn_Cmp_Lg] = 0x05; +} + +#define GET_MODE(code) ((code) & 0xC0) + +/** Sign extension bit values for binops */ +enum SignExt { + UNSIGNED_IMM = 0, /**< unsigned immediate */ + SIGNEXT_IMM = 2, /**< sign extended immediate */ +}; + +/** The mod encoding of the ModR/M */ +enum Mod { + MOD_IND = 0x00, /**< [reg1] */ + MOD_IND_BYTE_OFS = 0x40, /**< [reg1 + byte ofs] */ + MOD_IND_WORD_OFS = 0x80, /**< [reg1 + word ofs] */ + MOD_REG = 0xC0 /**< reg1 */ +}; + +/** create R/M encoding for ModR/M */ +#define ENC_RM(x) (x) +/** create REG encoding for ModR/M */ +#define ENC_REG(x) ((x) << 3) + +/** create encoding for a SIB byte */ +#define ENC_SIB(scale, index, base) ((scale) << 6 | (index) << 3 | (base)) + +/* Node: The following routines are supposed to append bytes, words, dwords + to the output stream. + Currently the implementation is stupid in that it still creates output + for an "assembler" in the form of .byte, .long + We will change this when enough infrastructure is there to create complete + machine code in memory/object files */ + +static void bemit8(const unsigned char byte) +{ + be_emit_irprintf("\t.byte 0x%x\n", byte); + be_emit_write_line(); +} + +static void bemit16(const unsigned u16) +{ + be_emit_irprintf("\t.word 0x%x\n", u16); + be_emit_write_line(); +} + +static void bemit32(const unsigned u32) +{ + be_emit_irprintf("\t.long 0x%x\n", u32); + be_emit_write_line(); +} + +/** + * Emit address of an entity. If @p is_relative is true then a relative + * offset from behind the address to the entity is created. + */ +static void bemit_entity(ir_entity *entity, bool entity_sign, int offset, + bool is_relative) +{ + if (entity == NULL) { + bemit32(offset); + return; + } + + if (is_relative) { + offset -= 4; + } + + /* the final version should remember the position in the bytestream + and patch it with the correct address at linktime... */ + be_emit_cstring("\t.long "); + if (entity_sign) + be_emit_char('-'); + set_entity_backend_marked(entity, 1); + be_gas_emit_entity(entity); + + if (is_relative) { + be_emit_cstring("-."); + } + + if (offset != 0) { + be_emit_irprintf("%+d", offset); + } + be_emit_char('\n'); + be_emit_write_line(); +} + +static void bemit_jmp_destination(const ir_node *dest_block) +{ + be_emit_cstring("\t.long "); + ia32_emit_block_name(dest_block); + be_emit_cstring(" - . - 4\n"); + be_emit_write_line(); +} + +/* end emit routines, all emitters following here should only use the functions + above. */ + +/** Create a ModR/M byte for src1,src2 registers */ +static void bemit_modrr(const arch_register_t *src1, + const arch_register_t *src2) +{ + unsigned char modrm = MOD_REG; + modrm |= ENC_RM(reg_gp_map[src1->index]); + modrm |= ENC_REG(reg_gp_map[src2->index]); + bemit8(modrm); +} + +/** Create a ModR/M byte for one register and extension */ +static void bemit_modru(const arch_register_t *reg, unsigned ext) +{ + unsigned char modrm = MOD_REG; + assert(ext <= 7); + modrm |= ENC_RM(reg_gp_map[reg->index]); + modrm |= ENC_REG(ext); + bemit8(modrm); +} + +/** + * Calculate the size of an (unsigned) immediate in bytes. + * + * @param offset an offset + */ +static unsigned get_unsigned_imm_size(unsigned offset) +{ + if (offset < 256) { + return 1; + } else if (offset < 65536) { + return 2; + } else { + return 4; + } +} + +/** + * Calculate the size of an signed immediate in bytes. + * + * @param offset an offset + */ +static unsigned get_signed_imm_size(int offset) +{ + if (-128 <= offset && offset < 128) { + return 1; + } else if (-32768 <= offset && offset < 32768) { + return 2; + } else { + return 4; + } +} + +/** + * Emit an address mode. + * + * @param reg content of the reg field: either a register index or an opcode extension + * @param node the node + */ +static void bemit_mod_am(unsigned reg, const ir_node *node) +{ + ir_entity *ent = get_ia32_am_sc(node); + int offs = get_ia32_am_offs_int(node); + ir_node *base = get_irn_n(node, n_ia32_base); + int has_base = !is_ia32_NoReg_GP(base); + ir_node *index = get_irn_n(node, n_ia32_index); + int has_index = !is_ia32_NoReg_GP(index); + unsigned modrm = 0; + unsigned sib = 0; + unsigned emitoffs = 0; + bool emitsib = false; + unsigned base_enc; + + /* set the mod part depending on displacement */ + if (ent != NULL) { + modrm |= MOD_IND_WORD_OFS; + emitoffs = 32; + } else if (offs == 0) { + modrm |= MOD_IND; + emitoffs = 0; + } else if (-128 <= offs && offs < 128) { + modrm |= MOD_IND_BYTE_OFS; + emitoffs = 8; + } else { + modrm |= MOD_IND_WORD_OFS; + emitoffs = 32; + } + + if (has_base) { + const arch_register_t *base_reg = arch_get_irn_register(base); + base_enc = reg_gp_map[base_reg->index]; + } else { + /* Use the EBP encoding + MOD_IND if NO base register. There is + * always a 32bit offset present in this case. */ + modrm = MOD_IND; + base_enc = 0x05; + emitoffs = 32; + } + + /* Determine if we need a SIB byte. */ + if (has_index) { + const arch_register_t *reg_index = arch_get_irn_register(index); + int scale = get_ia32_am_scale(node); + assert(scale < 4); + /* R/M set to ESP means SIB in 32bit mode. */ + modrm |= ENC_RM(0x04); + sib = ENC_SIB(scale, reg_gp_map[reg_index->index], base_enc); + emitsib = true; + } else if (base_enc == 0x04) { + /* for the above reason we are forced to emit a SIB when base is ESP. + * Only the base is used, index must be ESP too, which means no index. + */ + modrm |= ENC_RM(0x04); + sib = ENC_SIB(0, 0x04, 0x04); + emitsib = true; + } else { + modrm |= ENC_RM(base_enc); + /* We are forced to emit an 8bit offset as EBP base without offset is a + * special case for SIB without base register. */ + if (base_enc == 0x05 && emitoffs == 0) { + modrm |= MOD_IND_BYTE_OFS; + emitoffs = 8; + } + } + + modrm |= ENC_REG(reg); + + bemit8(modrm); + if (emitsib) + bemit8(sib); + + /* emit displacement */ + if (emitoffs == 8) { + bemit8((unsigned) offs); + } else if (emitoffs == 32) { + bemit_entity(ent, is_ia32_am_sc_sign(node), offs, false); + } +} + +/** + * Emit a binop with a immediate operand. + * + * @param node the node to emit + * @param opcode_eax the opcode for the op eax, imm variant + * @param opcode the opcode for the reg, imm variant + * @param ruval the opcode extension for opcode + */ +static void bemit_binop_with_imm( + const ir_node *node, + unsigned char opcode_ax, + unsigned char opcode, unsigned char ruval) +{ + /* Use in-reg, because some instructions (cmp, test) have no out-reg. */ + const ir_node *op = get_irn_n(node, n_ia32_binary_right); + const ia32_immediate_attr_t *attr = get_ia32_immediate_attr_const(op); + unsigned size; + + /* Some instructions (test) have no short form with 32bit value + 8bit + * immediate. */ + if (attr->symconst != NULL || opcode & SIGNEXT_IMM) { + size = 4; + } else { + /* check for sign extension */ + size = get_signed_imm_size(attr->offset); + } + + switch (size) { + case 1: + bemit8(opcode | SIGNEXT_IMM); + /* cmp has this special mode */ + if (get_ia32_op_type(node) == ia32_AddrModeS) { + bemit_mod_am(ruval, node); + } else { + const arch_register_t *reg = get_in_reg(node, n_ia32_binary_left); + bemit_modru(reg, ruval); + } + bemit8((unsigned char)attr->offset); + return; + case 2: + case 4: + /* check for eax variant: this variant is shorter for 32bit immediates only */ + if (get_ia32_op_type(node) == ia32_AddrModeS) { + bemit8(opcode); + bemit_mod_am(ruval, node); + } else { + const arch_register_t *reg = get_in_reg(node, n_ia32_binary_left); + if (reg->index == REG_EAX) { + bemit8(opcode_ax); + } else { + bemit8(opcode); + bemit_modru(reg, ruval); + } + } + bemit_entity(attr->symconst, attr->sc_sign, attr->offset, false); + return; + } + panic("invalid imm size?!?"); +} + +/** + * Emits a binop. + */ +static void bemit_binop_2(const ir_node *node, unsigned code) +{ + const arch_register_t *out = get_in_reg(node, n_ia32_binary_left); + ia32_op_type_t am_type = get_ia32_op_type(node); + unsigned char d = 0; + const arch_register_t *op2; + + switch (am_type) { + case ia32_AddrModeS: + d = 2; + /* FALLTHROUGH */ + case ia32_AddrModeD: + bemit8(code | d); + bemit_mod_am(reg_gp_map[out->index], node); + return; + case ia32_Normal: + bemit8(code); + op2 = get_in_reg(node, n_ia32_binary_right); + bemit_modrr(out, op2); + return; + } + panic("invalid address mode"); +} + +/** + * Emit a binop. + */ +static void bemit_binop(const ir_node *node, const unsigned char opcodes[4]) +{ + ir_node *right = get_irn_n(node, n_ia32_binary_right); + if (is_ia32_Immediate(right)) { + bemit_binop_with_imm(node, opcodes[1], opcodes[2], opcodes[3]); + } else { + bemit_binop_2(node, opcodes[0]); + } +} + +/** + * Emit an unop. + */ +static void bemit_unop(const ir_node *node, unsigned char code, unsigned char ext, int input) +{ + bemit8(code); + if (get_ia32_op_type(node) == ia32_Normal) { + const arch_register_t *in = get_in_reg(node, input); + bemit_modru(in, ext); + } else { + bemit_mod_am(ext, node); + } +} + +static void bemit_unop_reg(const ir_node *node, unsigned char code, int input) +{ + const arch_register_t *out = get_out_reg(node, 0); + bemit_unop(node, code, reg_gp_map[out->index], input); +} + +static void bemit_immediate(const ir_node *node, bool relative) +{ + const ia32_immediate_attr_t *attr = get_ia32_immediate_attr_const(node); + bemit_entity(attr->symconst, attr->sc_sign, attr->offset, relative); +} + +static void bemit_copy(const ir_node *copy) +{ + const arch_register_t *in = get_in_reg(copy, 0); + const arch_register_t *out = get_out_reg(copy, 0); + + if (in == out || is_unknown_reg(in)) + return; + /* copies of vf nodes aren't real... */ + if (arch_register_get_class(in) == &ia32_reg_classes[CLASS_ia32_vfp]) + return; + + if (get_irn_mode(copy) == mode_E) { + panic("NIY"); + } else { + assert(arch_register_get_class(in) == &ia32_reg_classes[CLASS_ia32_gp]); + bemit8(0x89); + bemit_modrr(out, in); + } +} + +static void bemit_xor0(const ir_node *node) +{ + const arch_register_t *out = get_out_reg(node, 0); + bemit8(0x31); + bemit_modrr(out, out); +} + +static void bemit_mov_const(const ir_node *node) +{ + const arch_register_t *out = get_out_reg(node, 0); + bemit8(0xB8 + reg_gp_map[out->index]); + bemit_immediate(node, false); +} + +/** + * Creates a function for a Binop with 3 possible encodings. + */ +#define BINOP(op, op0, op1, op2, op2_ext) \ +static void bemit_ ## op(const ir_node *node) { \ + static const unsigned char op ## _codes[] = {op0, op1, op2, op2_ext}; \ + bemit_binop(node, op ## _codes); \ +} + +/* insn def eax,imm imm */ +BINOP(add, 0x01, 0x05, 0x81, 0) +BINOP(or, 0x09, 0x0D, 0x81, 1) +BINOP(adc, 0x11, 0x15, 0x81, 2) +BINOP(sbb, 0x19, 0x1D, 0x81, 3) +BINOP(and, 0x21, 0x25, 0x81, 4) +BINOP(sub, 0x29, 0x2D, 0x81, 5) +BINOP(xor, 0x31, 0x35, 0x81, 6) +BINOP(cmp, 0x39, 0x3D, 0x81, 7) +BINOP(test, 0x85, 0xA9, 0xF7, 0) + +/** + * Creates a function for an Unop with code /ext encoding. + */ +#define UNOP(op, code, ext, input) \ +static void bemit_ ## op(const ir_node *node) { \ + bemit_unop(node, code, ext, input); \ +} + +UNOP(not, 0xF7, 2, n_ia32_Not_val) +UNOP(neg, 0xF7, 3, n_ia32_Neg_val) +UNOP(mul, 0xF7, 4, n_ia32_Mul_right) +UNOP(imul1op, 0xF7, 5, n_ia32_IMul1OP_right) +UNOP(div, 0xF7, 6, n_ia32_Div_divisor) +UNOP(idiv, 0xF7, 7, n_ia32_IDiv_divisor) + +/* TODO: am support for IJmp */ +UNOP(ijmp, 0xFF, 4, n_ia32_IJmp_target) + +#define SHIFT(op, ext) \ +static void bemit_##op(const ir_node *node) \ +{ \ + const arch_register_t *out = get_out_reg(node, 0); \ + ir_node *count = get_irn_n(node, 1); \ + if (is_ia32_Immediate(count)) { \ + int offset = get_ia32_immediate_attr_const(count)->offset; \ + if (offset == 1) { \ + bemit8(0xD1); \ + bemit_modru(out, ext); \ + } else { \ + bemit8(0xC1); \ + bemit_modru(out, ext); \ + bemit8(offset); \ + } \ + } else { \ + bemit8(0xD3); \ + bemit_modru(out, ext); \ + } \ +} + +SHIFT(rol, 0) +SHIFT(ror, 1) +SHIFT(shl, 4) +SHIFT(shr, 5) +SHIFT(sar, 7) + +static void bemit_dec(const ir_node *node) +{ + const arch_register_t *out = get_out_reg(node, pn_ia32_Dec_res); + bemit8(0x48 + reg_gp_map[out->index]); +} + +static void bemit_inc(const ir_node *node) +{ + const arch_register_t *out = get_out_reg(node, pn_ia32_Inc_res); + bemit8(0x40 + reg_gp_map[out->index]); +} + +/** + * Emit a Lea. + */ +static void bemit_lea(const ir_node *node) +{ + const arch_register_t *out = get_out_reg(node, 0); + bemit8(0x8D); + bemit_mod_am(reg_gp_map[out->index], node); +} + +/** + * Emit a single optcode. + */ +#define EMIT_SINGLEOP(op, code) \ +static void bemit_ ## op(const ir_node *node) { \ + (void) node; \ + bemit8(code); \ +} + +//EMIT_SINGLEOP(daa, 0x27) +//EMIT_SINGLEOP(das, 0x2F) +//EMIT_SINGLEOP(aaa, 0x37) +//EMIT_SINGLEOP(aas, 0x3F) +//EMIT_SINGLEOP(nop, 0x90) +//EMIT_SINGLEOP(cwde, 0x98) +EMIT_SINGLEOP(cltd, 0x99) +//EMIT_SINGLEOP(fwait, 0x9B) +EMIT_SINGLEOP(sahf, 0x9E) +//EMIT_SINGLEOP(popf, 0x9D) +EMIT_SINGLEOP(int3, 0xCC) +//EMIT_SINGLEOP(iret, 0xCF) +//EMIT_SINGLEOP(xlat, 0xD7) +//EMIT_SINGLEOP(lock, 0xF0) +EMIT_SINGLEOP(rep, 0xF3) +//EMIT_SINGLEOP(halt, 0xF4) +EMIT_SINGLEOP(cmc, 0xF5) +EMIT_SINGLEOP(stc, 0xF9) +//EMIT_SINGLEOP(cli, 0xFA) +//EMIT_SINGLEOP(sti, 0xFB) +//EMIT_SINGLEOP(std, 0xFD) + +/** + * Emits a MOV out, [MEM]. + */ +static void bemit_load(const ir_node *node) +{ + const arch_register_t *out = get_out_reg(node, 0); + + if (out->index == REG_EAX) { + ir_entity *ent = get_ia32_am_sc(node); + int offs = get_ia32_am_offs_int(node); + ir_node *base = get_irn_n(node, n_ia32_base); + int has_base = !is_ia32_NoReg_GP(base); + ir_node *index = get_irn_n(node, n_ia32_index); + int has_index = !is_ia32_NoReg_GP(index); + + if (ent == NULL && !has_base && !has_index) { + /* load from constant address to EAX can be encoded + as 0xA1 [offset] */ + bemit8(0xA1); + bemit_entity(NULL, 0, offs, false); + return; + } + } + bemit8(0x8B); + bemit_mod_am(reg_gp_map[out->index], node); +} + +/** + * Emits a MOV [mem], in. + */ +static void bemit_store(const ir_node *node) +{ + const ir_node *value = get_irn_n(node, n_ia32_Store_val); + + if (is_ia32_Immediate(value)) { + bemit8(0xC7); + bemit_mod_am(0, node); + bemit_immediate(value, false); + } else { + const arch_register_t *in = get_in_reg(node, n_ia32_Store_val); + + if (in->index == REG_EAX) { + ir_entity *ent = get_ia32_am_sc(node); + int offs = get_ia32_am_offs_int(node); + ir_node *base = get_irn_n(node, n_ia32_base); + int has_base = !is_ia32_NoReg_GP(base); + ir_node *index = get_irn_n(node, n_ia32_index); + int has_index = !is_ia32_NoReg_GP(index); + + if (ent == NULL && !has_base && !has_index) { + /* store to constant address from EAX can be encoded as + 0xA3 [offset]*/ + bemit8(0xA3); + bemit_entity(NULL, 0, offs, false); + return; + } + } + bemit8(0x89); + bemit_mod_am(reg_gp_map[in->index], node); + } +} + +static void bemit_conv_i2i(const ir_node *node) +{ + ir_mode *smaller_mode = get_ia32_ls_mode(node); + unsigned opcode; + + bemit8(0x0F); + /* 8 16 bit source + * movzx B6 B7 + * movsx BE BF + */ + opcode = 0xB6; + if (mode_is_signed(smaller_mode)) opcode |= 0x08; + if (get_mode_size_bits(smaller_mode) == 16) opcode |= 0x01; + bemit_unop_reg(node, opcode, n_ia32_Conv_I2I_val); +} + +/** + * Emit a Push. + */ +static void bemit_push(const ir_node *node) +{ + const ir_node *value = get_irn_n(node, n_ia32_Push_val); + + if (is_ia32_Immediate(value)) { + const ia32_immediate_attr_t *attr + = get_ia32_immediate_attr_const(value); + unsigned size = get_unsigned_imm_size(attr->offset); + if (attr->symconst) + size = 4; + switch (size) { + case 1: + bemit8(0x6A); + bemit8((unsigned char)attr->offset); + break; + case 2: + case 4: + bemit8(0x68); + bemit_immediate(value, false); + break; + } + } else if (is_ia32_NoReg_GP(value)) { + bemit8(0xFF); + bemit_mod_am(6, node); + } else { + const arch_register_t *reg = get_in_reg(node, n_ia32_Push_val); + bemit8(0x50 + reg_gp_map[reg->index]); + } +} + +/** + * Emit a Pop. + */ +static void bemit_pop(const ir_node *node) +{ + const arch_register_t *reg = get_out_reg(node, pn_ia32_Pop_res); + bemit8(0x58 + reg_gp_map[reg->index]); +} + +static void bemit_popmem(const ir_node *node) +{ + bemit8(0x8F); + bemit_mod_am(0, node); +} + +static void bemit_call(const ir_node *node) +{ + ir_node *proc = get_irn_n(node, n_ia32_Call_addr); + + if (is_ia32_Immediate(proc)) { + bemit8(0xE8); + bemit_immediate(proc, true); + } else { + bemit_unop(node, 0xFF, 2, n_ia32_Call_addr); + } +} + +static void bemit_jmp(const ir_node *dest_block) +{ + bemit8(0xE9); + bemit_jmp_destination(dest_block); +} + +static void bemit_jump(const ir_node *node) +{ + if (can_be_fallthrough(node)) + return; + + bemit_jmp(get_cfop_target_block(node)); +} + +static void bemit_jcc(int pnc, const ir_node *dest_block) +{ + unsigned char cc; + + if (pnc == ia32_pn_Cmp_parity) { + cc = 0x0A; + } else { + if (pnc & ia32_pn_Cmp_float || pnc & ia32_pn_Cmp_unsigned) { + cc = pnc_map_unsigned[pnc & 0x07]; + } else { + cc = pnc_map_signed[pnc & 0x07]; + } + } + assert(cc != 0xFF); + + bemit8(0x0F); + bemit8(0x80 + cc); + bemit_jmp_destination(dest_block); +} + +static void bemit_ia32_jcc(const ir_node *node) +{ + int pnc = get_ia32_condcode(node); + int need_parity_label = 0; + const ir_node *proj_true; + const ir_node *proj_false; + const ir_node *dest_true; + const ir_node *dest_false; + const ir_node *block; + + pnc = determine_final_pnc(node, 0, pnc); + + /* get both Projs */ + proj_true = get_proj(node, pn_ia32_Jcc_true); + assert(proj_true && "Jcc without true Proj"); + + proj_false = get_proj(node, pn_ia32_Jcc_false); + assert(proj_false && "Jcc without false Proj"); + + block = get_nodes_block(node); + + if (can_be_fallthrough(proj_true)) { + /* exchange both proj's so the second one can be omitted */ + const ir_node *t = proj_true; + + proj_true = proj_false; + proj_false = t; + pnc = ia32_get_negated_pnc(pnc); + } + + dest_true = get_cfop_target_block(proj_true); + dest_false = get_cfop_target_block(proj_false); + + if (pnc & ia32_pn_Cmp_float) { + panic("Float jump NIY"); + /* Some floating point comparisons require a test of the parity flag, + * which indicates that the result is unordered */ + switch (pnc & 15) { + case pn_Cmp_Uo: { + ia32_emitf(proj_true, "\tjp %L\n"); + break; + } + + case pn_Cmp_Leg: + ia32_emitf(proj_true, "\tjnp %L\n"); + break; + + case pn_Cmp_Eq: + case pn_Cmp_Lt: + case pn_Cmp_Le: + /* we need a local label if the false proj is a fallthrough + * as the falseblock might have no label emitted then */ + if (can_be_fallthrough(proj_false)) { + need_parity_label = 1; + ia32_emitf(proj_false, "\tjp 1f\n"); + } else { + ia32_emitf(proj_false, "\tjp %L\n"); + } + goto emit_jcc; + + case pn_Cmp_Ug: + case pn_Cmp_Uge: + case pn_Cmp_Ne: + ia32_emitf(proj_true, "\tjp %L\n"); + goto emit_jcc; + + default: + goto emit_jcc; + } + } else { +emit_jcc: + bemit_jcc(pnc, dest_true); + } + + if (need_parity_label) { + panic("parity label NIY"); + } + + /* the second Proj might be a fallthrough */ + if (can_be_fallthrough(proj_false)) { + /* it's a fallthrough */ + } else { + bemit_jmp(dest_false); + } +} + +/** + * Emits a return. + */ +static void bemit_return(const ir_node *node) +{ + unsigned pop = be_Return_get_pop(node); + if (pop > 0 || be_Return_get_emit_pop(node)) { + bemit8(0xC2); + assert(pop <= 0xffff); + bemit16(pop); + } else { + bemit8(0xC3); + } +} + +static void bemit_incsp(const ir_node *node) +{ + int offs; + const arch_register_t *reg; + unsigned size; + unsigned ext; + + offs = be_get_IncSP_offset(node); + if (offs == 0) + return; + + if (offs > 0) { + ext = 5; /* sub */ + } else { + ext = 0; /* add */ + offs = -offs; + } + + size = get_signed_imm_size(offs); + bemit8(size == 1 ? 0x83 : 0x81); + + reg = get_out_reg(node, 0); + bemit_modru(reg, ext); + + if (size == 1) { + bemit8(offs); + } else { + bemit32(offs); + } +} + +/** + * The type of a emitter function. + */ +typedef void (*emit_func) (const ir_node *); + +/** + * Set a node emitter. Make it a bit more type safe. + */ +static void register_emitter(ir_op *op, emit_func func) +{ + op->ops.generic = (op_func) func; +} + +static void ia32_register_binary_emitters(void) +{ + /* first clear the generic function pointer for all ops */ + clear_irp_opcodes_generic_func(); + + /* benode emitter */ + register_emitter(op_be_Copy, bemit_copy); + register_emitter(op_be_CopyKeep, bemit_copy); + register_emitter(op_be_IncSP, bemit_incsp); + register_emitter(op_be_Return, bemit_return); + register_emitter(op_ia32_Adc, bemit_adc); + register_emitter(op_ia32_Add, bemit_add); + register_emitter(op_ia32_And, bemit_and); + register_emitter(op_ia32_Breakpoint, bemit_int3); + register_emitter(op_ia32_Call, bemit_call); + register_emitter(op_ia32_Cltd, bemit_cltd); + register_emitter(op_ia32_Cmc, bemit_cmc); + register_emitter(op_ia32_Cmp, bemit_cmp); + register_emitter(op_ia32_Const, bemit_mov_const); + register_emitter(op_ia32_Conv_I2I, bemit_conv_i2i); + register_emitter(op_ia32_Conv_I2I8Bit, bemit_conv_i2i); + register_emitter(op_ia32_Dec, bemit_dec); + register_emitter(op_ia32_Div, bemit_div); + register_emitter(op_ia32_IDiv, bemit_idiv); + register_emitter(op_ia32_IJmp, bemit_ijmp); + register_emitter(op_ia32_IMul1OP, bemit_imul1op); + register_emitter(op_ia32_Inc, bemit_inc); + register_emitter(op_ia32_Jcc, bemit_ia32_jcc); + register_emitter(op_ia32_Jmp, bemit_jump); + register_emitter(op_ia32_Lea, bemit_lea); + register_emitter(op_ia32_Load, bemit_load); + register_emitter(op_ia32_Mul, bemit_mul); + register_emitter(op_ia32_Neg, bemit_neg); + register_emitter(op_ia32_Not, bemit_not); + register_emitter(op_ia32_Or, bemit_or); + register_emitter(op_ia32_Pop, bemit_pop); + register_emitter(op_ia32_PopEbp, bemit_pop); + register_emitter(op_ia32_PopMem, bemit_popmem); + register_emitter(op_ia32_Push, bemit_push); + register_emitter(op_ia32_RepPrefix, bemit_rep); + register_emitter(op_ia32_Rol, bemit_rol); + register_emitter(op_ia32_Ror, bemit_ror); + register_emitter(op_ia32_Sahf, bemit_sahf); + register_emitter(op_ia32_Sar, bemit_sar); + register_emitter(op_ia32_Sbb, bemit_sbb); + register_emitter(op_ia32_Shl, bemit_shl); + register_emitter(op_ia32_Shr, bemit_shr); + register_emitter(op_ia32_Stc, bemit_stc); + register_emitter(op_ia32_Store, bemit_store); + register_emitter(op_ia32_Sub, bemit_sub); + register_emitter(op_ia32_Test, bemit_test); + register_emitter(op_ia32_Xor, bemit_xor); + register_emitter(op_ia32_Xor0, bemit_xor0); + + /* ignore the following nodes */ + register_emitter(op_ia32_ProduceVal, emit_Nothing); + register_emitter(op_be_Barrier, emit_Nothing); + register_emitter(op_be_Keep, emit_Nothing); + register_emitter(op_be_Start, emit_Nothing); + register_emitter(op_Phi, emit_Nothing); + register_emitter(op_Start, emit_Nothing); +} + +static void gen_binary_block(ir_node *block) +{ + ir_node *node; + + ia32_emit_block_header(block); + + /* emit the contents of the block */ + sched_foreach(block, node) { + ia32_emit_node(node); + } +} + +void ia32_gen_binary_routine(ia32_code_gen_t *ia32_cg, ir_graph *irg) +{ + ir_entity *entity = get_irg_entity(irg); + int i, n; + + cg = ia32_cg; + isa = cg->isa; + + ia32_register_binary_emitters(); + + be_gas_emit_function_prolog(entity, ia32_cg_config.function_alignment); + + /* we use links to point to target blocks */ + ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK); + irg_block_walk_graph(irg, ia32_gen_labels, NULL, NULL); + + /* initialize next block links */ + n = ARR_LEN(cg->blk_sched); + for (i = 0; i < n; ++i) { + ir_node *block = cg->blk_sched[i]; + ir_node *prev = i > 0 ? cg->blk_sched[i-1] : NULL; + + set_irn_link(block, prev); + } + + for (i = 0; i < n; ++i) { + ir_node *block = cg->blk_sched[i]; + gen_binary_block(block); + } + + be_gas_emit_function_epilog(entity); + be_dbg_method_end(); + be_emit_char('\n'); + be_emit_write_line(); + + ir_free_resources(irg, IR_RESOURCE_IRN_LINK); +} + + + + void ia32_init_emitter(void) { lc_opt_entry_t *be_grp; @@ -2167,5 +3222,7 @@ void ia32_init_emitter(void) lc_opt_add_table(ia32_grp, ia32_emitter_options); + build_reg_map(); + FIRM_DBG_REGISTER(dbg, "firm.be.ia32.emitter"); }