X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fia32%2Fia32_emitter.c;h=391d73c459b04f93ce7d5abfdaef03aa19aa10bb;hb=9b6418bd4144ea51c2ef10d0e929c6c9449ca45d;hp=0ba657a187baa7be78322a5a3f046229ea4af414;hpb=081ef4f7ca73620500e5494f882ee6095c8bb983;p=libfirm diff --git a/ir/be/ia32/ia32_emitter.c b/ir/be/ia32/ia32_emitter.c index 0ba657a18..391d73c45 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" @@ -68,11 +68,59 @@ 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) +{ + 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; + + if (!is_Proj(cfgpred)) + return 1; + pred = get_Proj_pred(cfgpred); + if (is_ia32_SwitchJmp(pred)) + return 0; + + 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) { + ir_node *cfgpred = get_Block_cfgpred(block, 0); + ir_node *cfgpred_block = get_nodes_block(cfgpred); + + if (get_prev_block_sched(block) == cfgpred_block + && is_fallthrough(cfgpred)) { + need_label = 0; + } + } + + return need_label; +} /** * Returns the register at in position pos. @@ -88,19 +136,16 @@ 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"); - if(reg == &ia32_gp_regs[REG_GP_NOREG]) + if (reg == &ia32_gp_regs[REG_GP_NOREG]) panic("trying to emit noreg for %+F input %d", irn, 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 */ @@ -130,9 +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; @@ -140,7 +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; } } @@ -171,6 +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); @@ -180,6 +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); @@ -188,21 +248,23 @@ 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; - if(mode != NULL) { + if (mode != NULL) { int size = get_mode_size_bits(mode); - if(size == 8) { - emit_8bit_register(reg); - return; - } else if(size == 16) { - emit_16bit_register(reg); - return; - } else { - assert(mode_is_float(mode) || size == 32); + switch (size) { + case 8: emit_8bit_register(reg); return; + case 16: emit_16bit_register(reg); return; } + assert(mode_is_float(mode) || size == 32); } reg_name = arch_register_get_name(reg); @@ -213,18 +275,59 @@ static void emit_register(const arch_register_t *reg, const ir_mode *mode) void ia32_emit_source_register(const ir_node *node, int pos) { - const arch_register_t *reg = get_in_reg(node, pos); + const arch_register_t *reg = get_in_reg(node, 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); - if(is_ia32_Immediate(in)) { + const ir_node *in = get_irn_n(node, pos); + if (is_ia32_Immediate(in)) { emit_ia32_Immediate(in); return; } @@ -233,6 +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); @@ -240,6 +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); @@ -258,31 +387,22 @@ void ia32_emit_x87_register(const ir_node *node, int pos) static void ia32_emit_mode_suffix_mode(const ir_mode *mode) { - if(mode_is_float(mode)) { - switch(get_mode_size_bits(mode)) { - case 32: be_emit_char('s'); return; - case 64: be_emit_char('l'); return; - case 80: - case 96: be_emit_char('t'); return; - } - } else { - assert(mode_is_int(mode) || mode_is_reference(mode)); - switch(get_mode_size_bits(mode)) { + assert(mode_is_int(mode) || mode_is_reference(mode)); + switch (get_mode_size_bits(mode)) { + case 8: be_emit_char('b'); return; + case 16: be_emit_char('w'); return; + case 32: be_emit_char('l'); return; + /* gas docu says q is the suffix but gcc, objdump and icc use ll + * apparently */ case 64: be_emit_cstring("ll"); return; - /* gas docu says q is the suffix but gcc, objdump and icc use - ll apparently */ - case 32: be_emit_char('l'); return; - case 16: be_emit_char('w'); return; - case 8: be_emit_char('b'); return; - } } - panic("Can't output mode_suffix for %+F\n", mode); + panic("Can't output mode_suffix for %+F", mode); } void ia32_emit_mode_suffix(const ir_node *node) { ir_mode *mode = get_ia32_ls_mode(node); - if(mode == NULL) + if (mode == NULL) mode = mode_Iu; ia32_emit_mode_suffix_mode(mode); @@ -290,25 +410,43 @@ void ia32_emit_mode_suffix(const ir_node *node) void ia32_emit_x87_mode_suffix(const ir_node *node) { - ir_mode *mode = get_ia32_ls_mode(node); - assert(mode != NULL); + ir_mode *mode; + /* we only need to emit the mode on address mode */ - if(get_ia32_op_type(node) != ia32_Normal) - ia32_emit_mode_suffix_mode(mode); + if (get_ia32_op_type(node) == ia32_Normal) + return; + + mode = get_ia32_ls_mode(node); + assert(mode != NULL); + + if (mode_is_float(mode)) { + switch (get_mode_size_bits(mode)) { + case 32: be_emit_char('s'); return; + case 64: be_emit_char('l'); return; + case 80: + case 96: be_emit_char('t'); return; + } + } else { + assert(mode_is_int(mode)); + switch (get_mode_size_bits(mode)) { + case 16: be_emit_char('s'); return; + case 32: be_emit_char('l'); return; + /* gas docu says q is the suffix but gcc, objdump and icc use ll + * apparently */ + case 64: be_emit_cstring("ll"); return; + } + } + panic("Can't output mode_suffix for %+F", mode); } static char get_xmm_mode_suffix(ir_mode *mode) { assert(mode_is_float(mode)); switch(get_mode_size_bits(mode)) { - case 32: - return 's'; - case 64: - return 'd'; - default: - assert(0); + case 32: return 's'; + case 64: return 'd'; + default: panic("Invalid XMM mode"); } - return '%'; } void ia32_emit_xmm_mode_suffix(const ir_node *node) @@ -326,21 +464,19 @@ 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) { - if(get_mode_size_bits(mode) == 32) + ir_mode *mode = get_ia32_ls_mode(node); + if (get_mode_size_bits(mode) == 32) return; - if(mode_is_signed(mode)) { - be_emit_char('s'); - } else { - be_emit_char('z'); - } + 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) { ir_node *in = get_irn_n(node, pos); - if(is_ia32_Immediate(in)) { + if (is_ia32_Immediate(in)) { emit_ia32_Immediate(in); } else { const ir_mode *mode = get_ia32_ls_mode(node); @@ -350,136 +486,101 @@ void ia32_emit_source_register_or_immediate(const ir_node *node, int pos) } /** - * Emits registers and/or address mode of a binary operation. + * Returns the target block for a control flow node. */ -void ia32_emit_binop(const ir_node *node) { - const ir_node *right_op = get_irn_n(node, n_ia32_binary_right); - const ir_mode *mode = get_ia32_ls_mode(node); - const arch_register_t *reg_left; - - switch(get_ia32_op_type(node)) { - case ia32_Normal: - reg_left = get_in_reg(node, n_ia32_binary_left); - if(is_ia32_Immediate(right_op)) { - emit_ia32_Immediate(right_op); - be_emit_cstring(", "); - emit_register(reg_left, mode); - break; - } else { - const arch_register_t *reg_right - = get_in_reg(node, n_ia32_binary_right); - emit_register(reg_right, mode); - be_emit_cstring(", "); - emit_register(reg_left, mode); - } - break; - case ia32_AddrModeS: - if(is_ia32_Immediate(right_op)) { - emit_ia32_Immediate(right_op); - be_emit_cstring(", "); - ia32_emit_am(node); - } else { - reg_left = get_in_reg(node, n_ia32_binary_left); - ia32_emit_am(node); - be_emit_cstring(", "); - emit_register(reg_left, mode); - } - break; - case ia32_AddrModeD: - panic("DestMode can't be output by %%binop anymore"); - break; - default: - assert(0 && "unsupported op type"); - } +static ir_node *get_cfop_target_block(const ir_node *irn) +{ + assert(get_irn_mode(irn) == mode_X); + return get_irn_link(irn); } /** - * Emits registers and/or address mode of a binary operation. + * Emits a block label for the given block. */ -void ia32_emit_x87_binop(const ir_node *node) { - switch(get_ia32_op_type(node)) { - case ia32_Normal: - { - const ia32_x87_attr_t *x87_attr = get_ia32_x87_attr_const(node); - const arch_register_t *in1 = x87_attr->x87[0]; - const arch_register_t *in2 = x87_attr->x87[1]; - const arch_register_t *out = x87_attr->x87[2]; - const arch_register_t *in; - - in = out ? ((out == in2) ? in1 : in2) : in2; - out = out ? out : in1; - - be_emit_char('%'); - be_emit_string(arch_register_get_name(in)); - be_emit_cstring(", %"); - be_emit_string(arch_register_get_name(out)); - } - break; - case ia32_AddrModeS: - ia32_emit_am(node); - break; - case ia32_AddrModeD: - default: - assert(0 && "unsupported op type"); +static void ia32_emit_block_name(const ir_node *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)); } } /** - * Emits registers and/or address mode of a unary operation. + * Emits the target label for a control flow node. */ -void ia32_emit_unop(const ir_node *node, int pos) { - const ir_node *op; - - switch(get_ia32_op_type(node)) { - case ia32_Normal: - op = get_irn_n(node, pos); - if (is_ia32_Immediate(op)) { - emit_ia32_Immediate(op); - } else { - ia32_emit_source_register(node, pos); - } - break; - case ia32_AddrModeS: - case ia32_AddrModeD: - ia32_emit_am(node); - break; - default: - assert(0 && "unsupported op type"); - } +static void ia32_emit_cfop_target(const ir_node *node) +{ + ir_node *block = get_cfop_target_block(node); + ia32_emit_block_name(block); } -static void ia32_emit_entity(ir_entity *entity, int no_pic_adjust) -{ - ident *id; +/* + * positive conditions for signed compares + */ +static const char *const cmp2condition_s[] = { + NULL, /* always false */ + "e", /* == */ + "l", /* < */ + "le", /* <= */ + "g", /* > */ + "ge", /* >= */ + "ne", /* != */ + NULL /* always true */ +}; - set_entity_backend_marked(entity, 1); - id = get_entity_ld_ident(entity); - be_emit_ident(id); +/* + * positive conditions for unsigned compares + */ +static const char *const cmp2condition_u[] = { + NULL, /* always false */ + "e", /* == */ + "b", /* < */ + "be", /* <= */ + "a", /* > */ + "ae", /* >= */ + "ne", /* != */ + NULL /* always true */ +}; - 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"); - } - } +/** + * Emit the suffix for a compare instruction. + */ +static void ia32_emit_cmp_suffix(int pnc) +{ + const char *str; - if (!no_pic_adjust && do_pic) { - /* TODO: only do this when necessary */ - be_emit_char('-'); - be_emit_string(pic_base_label); + 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 { + 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) { +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, 0); + 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, 1); + ir_node *index = get_irn_n(node, n_ia32_index); int has_index = !is_ia32_NoReg_GP(index); /* just to be sure... */ @@ -487,13 +588,15 @@ void ia32_emit_am(const ir_node *node) { /* 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, 0); + ia32_emit_entity(ent, attr->data.am_sc_no_pic_adjust); } - if(offs != 0) { - if(ent != NULL) { + /* 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); @@ -518,193 +621,379 @@ void ia32_emit_am(const ir_node *node) { scale = get_ia32_am_scale(node); if (scale > 0) { - be_emit_irprintf(",%d", 1 << get_ia32_am_scale(node)); + be_emit_irprintf(",%d", 1 << scale); } } be_emit_char(')'); } - - /* special case if nothing is set */ - if(ent == NULL && offs == 0 && !has_base && !has_index) { - be_emit_char('0'); - } } -static void emit_ia32_IMul(const ir_node *node) +/** + * fmt parameter output + * ---- ---------------------- --------------------------------------------- + * %% % + * %AM address mode of the node + * %AR const arch_register_t* address mode of the node or register + * %ASx address mode of the node or source register x + * %Dx destination register x + * %I immediate of the node + * %L control flow target of the node + * %M mode suffix of the node + * %P int condition code + * %R const arch_register_t* register + * %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, ...) { - ir_node *left = get_irn_n(node, n_ia32_IMul_left); - const arch_register_t *out_reg = get_out_reg(node, pn_ia32_IMul_res); + va_list ap; + va_start(ap, fmt); - be_emit_cstring("\timul"); - ia32_emit_mode_suffix(node); - be_emit_char(' '); + for (;;) { + const char *start = fmt; + ia32_emit_mod_t mod = 0; - ia32_emit_binop(node); - - /* do we need the 3-address form? */ - if(is_ia32_NoReg_GP(left) || - get_in_reg(node, n_ia32_IMul_left) != out_reg) { - be_emit_cstring(", "); - emit_register(out_reg, get_ia32_ls_mode(node)); - } - be_emit_finish_line_gas(node); -} + while (*fmt != '%' && *fmt != '\n' && *fmt != '\0') + ++fmt; + if (fmt != start) { + be_emit_string_len(start, fmt - start); + } -/************************************************* - * _ _ _ - * (_) | | | - * ___ _ __ ___ _| |_ ___ ___ _ __ __| | - * / _ \ '_ ` _ \| | __| / __/ _ \| '_ \ / _` | - * | __/ | | | | | | |_ | (_| (_) | | | | (_| | - * \___|_| |_| |_|_|\__| \___\___/|_| |_|\__,_| - * - *************************************************/ + if (*fmt == '\n') { + be_emit_finish_line_gas(node); + ++fmt; + if (*fmt == '\0') + break; + continue; + } -#undef IA32_DO_EMIT -#define IA32_DO_EMIT(irn) ia32_fprintf_format(F, irn, cmd_buf, cmnt_buf) + if (*fmt == '\0') + break; -/* - * coding of conditions - */ -struct cmp2conditon_t { - const char *name; - int num; -}; + ++fmt; + if (*fmt == '*') { + mod |= EMIT_ALTERNATE_AM; + ++fmt; + } -/* - * 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 */ -}; + if (*fmt == '#') { + mod |= EMIT_RESPECT_LS; + ++fmt; + } -/* - * 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 */ -}; + if (*fmt == 'l') { + mod |= EMIT_LONG; + ++fmt; + } -enum { - ia32_pn_Cmp_unsigned = 0x1000, - ia32_pn_Cmp_float = 0x2000, -}; + switch (*fmt++) { + case '%': + be_emit_char('%'); + break; -/** - * walks up a tree of copies/perms/spills/reloads to find the original value - * that is moved around - */ -static ir_node *find_original_value(ir_node *node) -{ - inc_irg_visited(current_ir_graph); - while(1) { - mark_irn_visited(node); - if(be_is_Copy(node)) { - node = be_get_Copy_op(node); - } else if(be_is_CopyKeep(node)) { - node = be_get_CopyKeep_op(node); - } else if(is_Proj(node)) { - ir_node *pred = get_Proj_pred(node); - if(be_is_Perm(pred)) { - node = get_irn_n(pred, get_Proj_proj(node)); - } else if(be_is_MemPerm(pred)) { - node = get_irn_n(pred, get_Proj_proj(node) + 1); - } else if(is_ia32_Load(pred)) { - node = get_irn_n(pred, n_ia32_Load_mem); - } else { - return node; - } - } else if(is_ia32_Store(node)) { - node = get_irn_n(node, n_ia32_Store_val); - } else if(is_Phi(node)) { - int i, arity; - arity = get_irn_arity(node); - for(i = 0; i < arity; ++i) { - ir_node *in = get_irn_n(node, i); - if(irn_visited(in)) - continue; - node = in; + 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 { + emit_register(reg, NULL); + } + break; + } + + case 'S': + if (get_ia32_op_type(node) == ia32_AddrModeS) { + if (mod & EMIT_ALTERNATE_AM) + be_emit_char('*'); + ia32_emit_am(node); + ++fmt; + } else { + assert(get_ia32_op_type(node) == ia32_Normal); + goto emit_S; + } + break; + + default: goto unknown; + } break; } - assert(i < arity); - } else { - return node; - } - } -} -static int determine_final_pnc(const ir_node *node, int flags_pos, + case 'D': { + unsigned pos; + const arch_register_t *reg; + + if (*fmt < '0' || '9' <= *fmt) + goto unknown; + + pos = *fmt++ - '0'; + reg = get_out_reg(node, pos); + emit_register(reg, mod & EMIT_RESPECT_LS ? get_ia32_ls_mode(node) : NULL); + break; + } + + case 'I': + if (!(mod & EMIT_ALTERNATE_AM)) + be_emit_char('$'); + emit_ia32_Immediate_no_prefix(node); + break; + + case 'L': + ia32_emit_cfop_target(node); + break; + + case 'M': { + ia32_emit_mode_suffix_mode(get_ia32_ls_mode(node)); + break; + } + + case 'P': { + int pnc = va_arg(ap, int); + ia32_emit_cmp_suffix(pnc); + break; + } + + case 'R': { + const arch_register_t *reg = va_arg(ap, const arch_register_t*); + emit_register(reg, NULL); + break; + } + +emit_S: + case 'S': { + unsigned pos; + const ir_node *in; + + if (*fmt < '0' || '9' <= *fmt) + goto unknown; + + pos = *fmt++ - '0'; + in = get_irn_n(node, pos); + if (is_ia32_Immediate(in)) { + if (!(mod & EMIT_ALTERNATE_AM)) + be_emit_char('$'); + emit_ia32_Immediate_no_prefix(in); + } else { + 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; + } + + case 's': { + const char *str = va_arg(ap, const char*); + be_emit_string(str); + break; + } + + 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 format conversion in ia32_emitf()"); + } + } + + va_end(ap); +} + +/** + * Emits registers and/or address mode of a binary operation. + */ +void ia32_emit_binop(const ir_node *node) +{ + if (is_ia32_Immediate(get_irn_n(node, n_ia32_binary_right))) { + ia32_emitf(node, "%#S4, %#AS3"); + } else { + ia32_emitf(node, "%#AS4, %#S3"); + } +} + +/** + * Emits registers and/or address mode of a binary operation. + */ +void ia32_emit_x87_binop(const ir_node *node) +{ + switch(get_ia32_op_type(node)) { + case ia32_Normal: + { + const ia32_x87_attr_t *x87_attr = get_ia32_x87_attr_const(node); + const arch_register_t *in1 = x87_attr->x87[0]; + const arch_register_t *in = x87_attr->x87[1]; + const arch_register_t *out = x87_attr->x87[2]; + + if (out == NULL) { + out = in1; + } else if (out == in) { + in = in1; + } + + be_emit_char('%'); + be_emit_string(arch_register_get_name(in)); + be_emit_cstring(", %"); + be_emit_string(arch_register_get_name(out)); + } + break; + case ia32_AddrModeS: + ia32_emit_am(node); + break; + case ia32_AddrModeD: + default: + assert(0 && "unsupported op type"); + } +} + +/** + * Emits registers and/or address mode of a unary operation. + */ +void ia32_emit_unop(const ir_node *node, int pos) +{ + char fmt[] = "%ASx"; + fmt[3] = '0' + pos; + ia32_emitf(node, fmt); +} + +static void emit_ia32_IMul(const ir_node *node) +{ + ir_node *left = get_irn_n(node, n_ia32_IMul_left); + const arch_register_t *out_reg = get_out_reg(node, pn_ia32_IMul_res); + + /* do we need the 3-address form? */ + if (is_ia32_NoReg_GP(left) || + get_in_reg(node, n_ia32_IMul_left) != out_reg) { + ia32_emitf(node, "\timul%M %#S4, %#AS3, %#D0\n"); + } else { + ia32_emitf(node, "\timul%M %#AS4, %#S3\n"); + } +} + +/** + * walks up a tree of copies/perms/spills/reloads to find the original value + * that is moved around + */ +static ir_node *find_original_value(ir_node *node) +{ + if (irn_visited(node)) + return NULL; + + mark_irn_visited(node); + if (be_is_Copy(node)) { + return find_original_value(be_get_Copy_op(node)); + } else if (be_is_CopyKeep(node)) { + return find_original_value(be_get_CopyKeep_op(node)); + } else if (is_Proj(node)) { + ir_node *pred = get_Proj_pred(node); + if (be_is_Perm(pred)) { + return find_original_value(get_irn_n(pred, get_Proj_proj(node))); + } else if (be_is_MemPerm(pred)) { + return find_original_value(get_irn_n(pred, get_Proj_proj(node) + 1)); + } else if (is_ia32_Load(pred)) { + return find_original_value(get_irn_n(pred, n_ia32_Load_mem)); + } else { + return node; + } + } else if (is_ia32_Store(node)) { + return find_original_value(get_irn_n(node, n_ia32_Store_val)); + } else if (is_Phi(node)) { + int i, arity; + arity = get_irn_arity(node); + for (i = 0; i < arity; ++i) { + ir_node *in = get_irn_n(node, i); + ir_node *res = find_original_value(in); + + if (res != NULL) + return res; + } + return NULL; + } else { + return node; + } +} + +static int determine_final_pnc(const ir_node *node, int flags_pos, int pnc) { ir_node *flags = get_irn_n(node, flags_pos); const ia32_attr_t *flags_attr; flags = skip_Proj(flags); - if(is_ia32_Sahf(flags)) { + if (is_ia32_Sahf(flags)) { ir_node *cmp = get_irn_n(flags, n_ia32_Sahf_val); - if(!(is_ia32_FucomFnstsw(cmp) || is_ia32_FucompFnstsw(cmp) + if (!(is_ia32_FucomFnstsw(cmp) || is_ia32_FucompFnstsw(cmp) || is_ia32_FucomppFnstsw(cmp) || is_ia32_FtstFnstsw(cmp))) { + inc_irg_visited(current_ir_graph); cmp = find_original_value(cmp); + assert(cmp != NULL); assert(is_ia32_FucomFnstsw(cmp) || is_ia32_FucompFnstsw(cmp) || is_ia32_FucomppFnstsw(cmp) || is_ia32_FtstFnstsw(cmp)); } flags_attr = get_ia32_attr_const(cmp); - if(flags_attr->data.ins_permuted) + if (flags_attr->data.ins_permuted) pnc = get_mirrored_pnc(pnc); pnc |= ia32_pn_Cmp_float; - } else if(is_ia32_Ucomi(flags) || is_ia32_Fucomi(flags) + } else if (is_ia32_Ucomi(flags) || is_ia32_Fucomi(flags) || is_ia32_Fucompi(flags)) { flags_attr = get_ia32_attr_const(flags); - if(flags_attr->data.ins_permuted) + if (flags_attr->data.ins_permuted) pnc = get_mirrored_pnc(pnc); pnc |= ia32_pn_Cmp_float; } else { -#if 0 - assert(is_ia32_Cmp(flags) || is_ia32_Test(flags) - || is_ia32_Cmp8Bit(flags) || is_ia32_Test8Bit(flags)); -#endif flags_attr = get_ia32_attr_const(flags); - if(flags_attr->data.ins_permuted) + if (flags_attr->data.ins_permuted) pnc = get_mirrored_pnc(pnc); - if(flags_attr->data.cmp_unsigned) + if (flags_attr->data.cmp_unsigned) pnc |= ia32_pn_Cmp_unsigned; } return pnc; } -static void ia32_emit_cmp_suffix(int pnc) +static pn_Cmp ia32_get_negated_pnc(pn_Cmp 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; - } else { - pnc = pnc & 7; - assert(cmp2condition_s[pnc].num == pnc); - str = cmp2condition_s[pnc].name; - } - - be_emit_string(str); + 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, @@ -715,58 +1004,26 @@ 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); } /** - * Returns the target block for a control flow node. - */ -static ir_node *get_cfop_target_block(const ir_node *irn) { - return get_irn_link(irn); -} - -/** - * Emits a block label for the given block. - */ -static void ia32_emit_block_name(const ir_node *block) -{ - if (has_Block_label(block)) { - be_emit_string(be_gas_label_prefix()); - be_emit_irprintf("%u", (unsigned)get_Block_label(block)); - } else { - be_emit_cstring(BLOCK_PREFIX); - be_emit_irprintf("%d", get_irn_node_nr(block)); - } -} - -/** - * Emits the target label for a control flow node. + * Emits an exception label for a given node. */ -static void ia32_emit_cfop_target(const ir_node *node) -{ - ir_node *block = get_cfop_target_block(node); - - ia32_emit_block_name(block); -} - -/** Return the next block in Block schedule */ -static ir_node *next_blk_sched(const ir_node *block) +static void ia32_emit_exc_label(const ir_node *node) { - return get_irn_link(block); + be_emit_string(be_gas_insn_label_prefix()); + be_emit_irprintf("%lu", get_ia32_exc_label_id(node)); } /** * Returns the Proj with projection number proj and NOT mode_M */ -static ir_node *get_proj(const ir_node *node, long proj) { +static ir_node *get_proj(const ir_node *node, long proj) +{ const ir_edge_t *edge; ir_node *src; @@ -785,6 +1042,13 @@ static ir_node *get_proj(const ir_node *node, long proj) { return NULL; } +static int can_be_fallthrough(const ir_node *node) +{ + ir_node *target_block = get_cfop_target_block(node); + ir_node *block = get_nodes_block(node); + return get_prev_block_sched(target_block) == block; +} + /** * Emits the jump sequence for a conditional jump (cmp + jmp_true + jmp_false) */ @@ -794,7 +1058,6 @@ static void emit_ia32_Jcc(const ir_node *node) const ir_node *proj_true; const ir_node *proj_false; const ir_node *block; - const ir_node *next_block; pn_Cmp pnc = get_ia32_condcode(node); pnc = determine_final_pnc(node, 0, pnc); @@ -807,19 +1070,14 @@ static void emit_ia32_Jcc(const ir_node *node) assert(proj_false && "Jcc without false Proj"); block = get_nodes_block(node); - next_block = next_blk_sched(block); - if (get_cfop_target_block(proj_true) == next_block) { + 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; - 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) { @@ -827,16 +1085,12 @@ static void emit_ia32_Jcc(const ir_node *node) * which indicates that the result is unordered */ switch (pnc & 15) { case pn_Cmp_Uo: { - be_emit_cstring("\tjp "); - ia32_emit_cfop_target(proj_true); - be_emit_finish_line_gas(proj_true); + ia32_emitf(proj_true, "\tjp %L\n"); break; } case pn_Cmp_Leg: - be_emit_cstring("\tjnp "); - ia32_emit_cfop_target(proj_true); - be_emit_finish_line_gas(proj_true); + ia32_emitf(proj_true, "\tjnp %L\n"); break; case pn_Cmp_Eq: @@ -844,22 +1098,18 @@ static void emit_ia32_Jcc(const ir_node *node) 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 (get_cfop_target_block(proj_false) == next_block) { + if (can_be_fallthrough(proj_false)) { need_parity_label = 1; - be_emit_cstring("\tjp 1f"); + ia32_emitf(proj_false, "\tjp 1f\n"); } else { - be_emit_cstring("\tjp "); - ia32_emit_cfop_target(proj_false); + ia32_emitf(proj_false, "\tjp %L\n"); } - be_emit_finish_line_gas(proj_false); goto emit_jcc; case pn_Cmp_Ug: case pn_Cmp_Uge: case pn_Cmp_Ne: - be_emit_cstring("\tjp "); - ia32_emit_cfop_target(proj_true); - be_emit_finish_line_gas(proj_true); + ia32_emitf(proj_true, "\tjp %L\n"); goto emit_jcc; default: @@ -867,28 +1117,18 @@ static void emit_ia32_Jcc(const ir_node *node) } } else { emit_jcc: - be_emit_cstring("\tj"); - ia32_emit_cmp_suffix(pnc); - be_emit_char(' '); - ia32_emit_cfop_target(proj_true); - be_emit_finish_line_gas(proj_true); + ia32_emitf(proj_true, "\tj%P %L\n", pnc); } - if(need_parity_label) { - be_emit_cstring("1:"); - be_emit_write_line(); + if (need_parity_label) { + ia32_emitf(NULL, "1:\n"); } /* the second Proj might be a fallthrough */ - if (get_cfop_target_block(proj_false) != next_block) { - be_emit_cstring("\tjmp "); - ia32_emit_cfop_target(proj_false); - be_emit_finish_line_gas(proj_false); + if (can_be_fallthrough(proj_false)) { + ia32_emitf(proj_false, "\t/* fallthrough to %L */\n"); } else { - be_emit_cstring("\t/* fallthrough to "); - ia32_emit_cfop_target(proj_false); - be_emit_cstring(" */"); - be_emit_finish_line_gas(proj_false); + ia32_emitf(proj_false, "\tjmp %L\n"); } } @@ -896,22 +1136,20 @@ 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) { + if (out == in_false) { /* yes -> nothing to do */ - } else if(out == in_true) { + } else if (out == in_true) { const arch_register_t *tmp; assert(get_ia32_op_type(node) == ia32_Normal); @@ -923,34 +1161,15 @@ static void emit_ia32_CMov(const ir_node *node) in_false = tmp; } else { /* we need a mov */ - be_emit_cstring("\tmovl "); - emit_register(in_false, NULL); - be_emit_cstring(", "); - emit_register(out, NULL); - be_emit_finish_line_gas(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 */ - be_emit_cstring("\tcmov"); - ia32_emit_cmp_suffix(pnc); - be_emit_char(' '); - if(get_ia32_op_type(node) == ia32_AddrModeS) { - ia32_emit_am(node); - } else { - emit_register(in_true, get_ia32_ls_mode(node)); - } - be_emit_cstring(", "); - emit_register(out, get_ia32_ls_mode(node)); - be_emit_finish_line_gas(node); + ia32_emitf(node, "\tcmov%P %#AR, %#R\n", pnc, in_true, out); } /********************************************************* @@ -983,7 +1202,8 @@ typedef struct _jmp_tbl_t { /** * Compare two variables of type branch_t. Used to sort all switch cases */ -static int ia32_cmp_branch_t(const void *a, const void *b) { +static int ia32_cmp_branch_t(const void *a, const void *b) +{ branch_t *b1 = (branch_t *)a; branch_t *b2 = (branch_t *)b; @@ -1009,11 +1229,11 @@ static void emit_ia32_SwitchJmp(const ir_node *node) const ir_edge_t *edge; /* fill the table structure */ - tbl.label = xmalloc(SNPRINTF_BUF_LEN); + tbl.label = XMALLOCN(char, SNPRINTF_BUF_LEN); tbl.label = get_unique_label(tbl.label, SNPRINTF_BUF_LEN, ".TBL_"); tbl.defProj = NULL; tbl.num_branches = get_irn_n_edges(node) - 1; - tbl.branches = xcalloc(tbl.num_branches, sizeof(tbl.branches[0])); + tbl.branches = XMALLOCNZ(branch_t, tbl.num_branches); tbl.min_value = INT_MAX; tbl.max_value = INT_MIN; @@ -1050,53 +1270,30 @@ static void emit_ia32_SwitchJmp(const ir_node *node) interval = tbl.max_value - tbl.min_value; /* emit the table */ - be_emit_cstring("\tcmpl $"); - be_emit_irprintf("%u, ", interval); - ia32_emit_source_register(node, 0); - be_emit_finish_line_gas(node); - - be_emit_cstring("\tja "); - ia32_emit_cfop_target(tbl.defProj); - be_emit_finish_line_gas(node); + ia32_emitf(node, "\tcmpl $%u, %S0\n", interval); + ia32_emitf(tbl.defProj, "\tja %L\n"); if (tbl.num_branches > 1) { /* create table */ - be_emit_cstring("\tjmp *"); - be_emit_string(tbl.label); - be_emit_cstring("(,"); - ia32_emit_source_register(node, 0); - be_emit_cstring(",4)"); - be_emit_finish_line_gas(node); + ia32_emitf(node, "\tjmp *%s(,%S0,4)\n", tbl.label); be_gas_emit_switch_section(GAS_SECTION_RODATA); - be_emit_cstring("\t.align 4\n"); - be_emit_write_line(); - - be_emit_string(tbl.label); - be_emit_cstring(":\n"); - be_emit_write_line(); - - be_emit_cstring(".long "); - ia32_emit_cfop_target(tbl.branches[0].target); - be_emit_finish_line_gas(NULL); + ia32_emitf(NULL, "\t.align 4\n"); + ia32_emitf(NULL, "%s:\n", tbl.label); last_value = tbl.branches[0].value; - for (i = 1; i < tbl.num_branches; ++i) { - while (++last_value < tbl.branches[i].value) { - be_emit_cstring(".long "); - ia32_emit_cfop_target(tbl.defProj); - be_emit_finish_line_gas(NULL); + for (i = 0; i != tbl.num_branches; ++i) { + while (last_value != tbl.branches[i].value) { + ia32_emitf(tbl.defProj, ".long %L\n"); + ++last_value; } - be_emit_cstring(".long "); - ia32_emit_cfop_target(tbl.branches[i].target); - be_emit_finish_line_gas(NULL); + ia32_emitf(tbl.branches[i].target, ".long %L\n"); + ++last_value; } be_gas_emit_switch_section(GAS_SECTION_TEXT); } else { /* one jump is enough */ - be_emit_cstring("\tjmp "); - ia32_emit_cfop_target(tbl.branches[0].target); - be_emit_finish_line_gas(node); + ia32_emitf(tbl.branches[0].target, "\tjmp %L\n"); } if (tbl.label) @@ -1108,42 +1305,18 @@ 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, *next_block; + ir_node *block; /* for now, the code works for scheduled and non-schedules blocks */ block = get_nodes_block(node); /* we have a block schedule */ - next_block = next_blk_sched(block); - if (get_cfop_target_block(node) != next_block) { - be_emit_cstring("\tjmp "); - ia32_emit_cfop_target(node); + if (can_be_fallthrough(node)) { + ia32_emitf(node, "\t/* fallthrough to %L */\n"); } else { - be_emit_cstring("\t/* fallthrough to "); - ia32_emit_cfop_target(node); - be_emit_cstring(" */"); - } - be_emit_finish_line_gas(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); - } + ia32_emitf(node, "\tjmp %L\n"); } } @@ -1175,7 +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 '%': @@ -1199,15 +1372,16 @@ 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; } /* parse number */ sscanf(s, "%d%n", &num, &p); - if(num < 0) { + if (num < 0) { ir_fprintf(stderr, "Warning: Couldn't parse assembler operand (%+F)\n", node); return s; @@ -1215,39 +1389,41 @@ 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]; assert(asm_reg->valid); /* get register */ - if(asm_reg->use_input == 0) { + if (asm_reg->use_input == 0) { reg = get_out_reg(node, asm_reg->inout_pos); } else { ir_node *pred = get_irn_n(node, asm_reg->inout_pos); /* might be an immediate value */ - if(is_ia32_Immediate(pred)) { + if (is_ia32_Immediate(pred)) { emit_ia32_Immediate(pred); return 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); + if (reg == NULL) { + ir_fprintf(stderr, + "Warning: no register assigned for %d asm op (%+F)\n", + num, node); return s; } - if(asm_reg->memory) { + if (asm_reg->memory) { be_emit_char('('); } /* emit it */ - if(modifier != 0) { + if (modifier != 0) { be_emit_char('%'); switch(modifier) { case 'b': @@ -1267,7 +1443,7 @@ static const char* emit_asm_operand(const ir_node *node, const char *s) emit_register(reg, asm_reg->mode); } - if(asm_reg->memory) { + if (asm_reg->memory) { be_emit_char(')'); } @@ -1285,27 +1461,20 @@ static void emit_ia32_Asm(const ir_node *node) ident *asm_text = attr->asm_text; const char *s = get_id_str(asm_text); - be_emit_cstring("# Begin ASM \t"); - be_emit_finish_line_gas(node); + ia32_emitf(node, "#APP\t\n"); if (s[0] != '\t') be_emit_char('\t'); while(*s != 0) { - if(*s == '%') { + if (*s == '%') { s = emit_asm_operand(node, s); - continue; } else { - be_emit_char(*s); + be_emit_char(*s++); } - ++s; } - be_emit_char('\n'); - be_emit_write_line(); - - be_emit_cstring("# End ASM\n"); - be_emit_write_line(); + ia32_emitf(NULL, "\n#NO_APP\n"); } /********************************** @@ -1322,26 +1491,12 @@ static void emit_ia32_Asm(const ir_node *node) /** * Emit movsb/w instructions to make mov count divideable by 4 */ -static void emit_CopyB_prolog(unsigned size) { - be_emit_cstring("\tcld"); - be_emit_finish_line_gas(NULL); - - switch (size) { - case 1: - be_emit_cstring("\tmovsb"); - be_emit_finish_line_gas(NULL); - break; - case 2: - be_emit_cstring("\tmovsw"); - be_emit_finish_line_gas(NULL); - break; - case 3: - be_emit_cstring("\tmovsb"); - be_emit_finish_line_gas(NULL); - be_emit_cstring("\tmovsw"); - be_emit_finish_line_gas(NULL); - break; - } +static void emit_CopyB_prolog(unsigned size) +{ + if (size & 1) + ia32_emitf(NULL, "\tmovsb\n"); + if (size & 2) + ia32_emitf(NULL, "\tmovsw\n"); } /** @@ -1352,9 +1507,7 @@ static void emit_ia32_CopyB(const ir_node *node) unsigned size = get_ia32_copyb_size(node); emit_CopyB_prolog(size); - - be_emit_cstring("\trep movsd"); - be_emit_finish_line_gas(node); + ia32_emitf(node, "\trep movsd\n"); } /** @@ -1364,12 +1517,11 @@ static void emit_ia32_CopyB_i(const ir_node *node) { unsigned size = get_ia32_copyb_size(node); - emit_CopyB_prolog(size & 0x3); + emit_CopyB_prolog(size); size >>= 2; while (size--) { - be_emit_cstring("\tmovsd"); - be_emit_finish_line_gas(NULL); + ia32_emitf(NULL, "\tmovsd\n"); } } @@ -1388,63 +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 = ls_bits == 32 ? conv_f : conv_d; - be_emit_cstring("\tcvt"); - - if(is_ia32_Conv_I2FP(node)) { - if(ls_bits == 32) { - be_emit_cstring("si2ss"); - } else { - be_emit_cstring("si2sd"); - } - } else if(is_ia32_Conv_FP2I(node)) { - if(ls_bits == 32) { - be_emit_cstring("ss2si"); - } else { - be_emit_cstring("sd2si"); - } - } else { - assert(is_ia32_Conv_FP2FP(node)); - if(ls_bits == 32) { - be_emit_cstring("sd2ss"); - } else { - be_emit_cstring("ss2sd"); - } - } - be_emit_char(' '); - - switch(get_ia32_op_type(node)) { - case ia32_Normal: - ia32_emit_source_register(node, n_ia32_unary_op); - break; - case ia32_AddrModeS: - ia32_emit_am(node); - break; - default: - assert(0 && "unsupported op type for Conv"); - } - be_emit_cstring(", "); - ia32_emit_dest_register(node, 0); - be_emit_finish_line_gas(node); + 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"); } /** @@ -1452,62 +1570,24 @@ static void emit_ia32_Conv_FP2FP(const ir_node *node) */ static void emit_ia32_Conv_I2I(const ir_node *node) { - const char *sign_suffix; - ir_mode *smaller_mode = get_ia32_ls_mode(node); - int smaller_bits = get_mode_size_bits(smaller_mode); - int signed_mode; - const arch_register_t *in_reg, *out_reg; + ir_mode *smaller_mode = get_ia32_ls_mode(node); + 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 || smaller_bits == 32); - - signed_mode = mode_is_signed(smaller_mode); - if(smaller_bits == 32) { - // this should not happen as it's no convert - assert(0); - sign_suffix = ""; - } else { - sign_suffix = signed_mode ? "s" : "z"; - } - out_reg = get_out_reg(node, 0); - - switch(get_ia32_op_type(node)) { - case ia32_Normal: - in_reg = get_in_reg(node, n_ia32_unary_op); + sign_suffix = signed_mode ? "s" : "z"; + ia32_emitf(node, "\tmov%s%Ml %#AS3, %D0\n", sign_suffix); +} - if (in_reg == &ia32_gp_regs[REG_EAX] && - out_reg == &ia32_gp_regs[REG_EAX] && - signed_mode && - smaller_bits == 16) - { - /* argument and result are both in EAX and */ - /* signedness is ok: -> use the smaller cwtl opcode */ - be_emit_cstring("\tcwtl"); - } else { - be_emit_cstring("\tmov"); - be_emit_string(sign_suffix); - ia32_emit_mode_suffix_mode(smaller_mode); - be_emit_cstring("l "); - emit_register(in_reg, smaller_mode); - be_emit_cstring(", "); - emit_register(out_reg, NULL); - } - break; - case ia32_AddrModeS: { - be_emit_cstring("\tmov"); - be_emit_string(sign_suffix); - ia32_emit_mode_suffix_mode(smaller_mode); - be_emit_cstring("l "); - ia32_emit_am(node); - be_emit_cstring(", "); - emit_register(out_reg, NULL); - break; - } - default: - assert(0 && "unsupported op type for Conv"); - } - be_emit_finish_line_gas(node); +/** + * 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"); } @@ -1521,45 +1601,31 @@ 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. */ static void emit_be_IncSP(const ir_node *node) { - int offs = be_get_IncSP_offset(node); - const arch_register_t *reg = arch_get_irn_register(arch_env, node); + int offs = be_get_IncSP_offset(node); if (offs == 0) return; if (offs > 0) { - be_emit_cstring("\tsubl $"); - be_emit_irprintf("%u, ", offs); - emit_register(reg, NULL); + ia32_emitf(node, "\tsubl $%u, %D0\n", offs); } else { - be_emit_cstring("\taddl $"); - be_emit_irprintf("%u, ", -offs); - emit_register(reg, NULL); + ia32_emitf(node, "\taddl $%u, %D0\n", -offs); } - be_emit_finish_line_gas(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; } /** @@ -1567,32 +1633,23 @@ static void emit_be_IncSP(const ir_node *node) */ 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); - ir_mode *mode; + const arch_register_t *in = arch_get_irn_register(op); + const arch_register_t *out = arch_get_irn_register(node); - if(in == out) { + if (in == out) { return; } - if(is_unknown_reg(in)) + if (is_unknown_reg(in)) return; /* copies of vf nodes aren't real... */ - if(arch_register_get_class(in) == &ia32_reg_classes[CLASS_ia32_vfp]) + if (arch_register_get_class(in) == &ia32_reg_classes[CLASS_ia32_vfp]) return; - mode = get_irn_mode(node); - if (mode == mode_E) { - be_emit_cstring("\tmovsd "); - emit_register(in, NULL); - be_emit_cstring(", "); - emit_register(out, NULL); + if (get_irn_mode(node) == mode_E) { + ia32_emitf(node, "\tmovsd %R, %R\n", in, out); } else { - be_emit_cstring("\tmovl "); - emit_register(in, NULL); - be_emit_cstring(", "); - emit_register(out, NULL); + ia32_emitf(node, "\tmovl %R, %R\n", in, out); } - be_emit_finish_line_gas(node); } static void emit_be_Copy(const ir_node *node) @@ -1613,8 +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); @@ -1622,35 +1679,17 @@ static void emit_be_Perm(const ir_node *node) assert(cls0 == cls1 && "Register class mismatch at Perm"); if (cls0 == &ia32_reg_classes[CLASS_ia32_gp]) { - be_emit_cstring("\txchg "); - emit_register(in1, NULL); - be_emit_cstring(", "); - emit_register(in0, NULL); - be_emit_finish_line_gas(node); + ia32_emitf(node, "\txchg %R, %R\n", in1, in0); } else if (cls0 == &ia32_reg_classes[CLASS_ia32_xmm]) { - be_emit_cstring("\txorpd "); - emit_register(in1, NULL); - be_emit_cstring(", "); - emit_register(in0, NULL); - be_emit_finish_line_gas(NULL); - - be_emit_cstring("\txorpd "); - emit_register(in0, NULL); - be_emit_cstring(", "); - emit_register(in1, NULL); - be_emit_finish_line_gas(NULL); - - be_emit_cstring("\txorpd "); - emit_register(in1, NULL); - be_emit_cstring(", "); - emit_register(in0, NULL); - be_emit_finish_line_gas(node); + ia32_emitf(NULL, "\txorpd %R, %R\n", in1, in0); + ia32_emitf(NULL, "\txorpd %R, %R\n", in0, in1); + ia32_emitf(node, "\txorpd %R, %R\n", in1, in0); } else if (cls0 == &ia32_reg_classes[CLASS_ia32_vfp]) { /* is a NOP */ } else if (cls0 == &ia32_reg_classes[CLASS_ia32_st]) { /* is a NOP */ } else { - panic("unexpected register class in be_Perm (%+F)\n", node); + panic("unexpected register class in be_Perm (%+F)", node); } } @@ -1659,12 +1698,7 @@ static void emit_be_Perm(const ir_node *node) */ static void emit_ia32_Const(const ir_node *node) { - be_emit_cstring("\tmovl "); - emit_ia32_Immediate(node); - be_emit_cstring(", "); - ia32_emit_dest_register(node, 0); - - be_emit_finish_line_gas(node); + ia32_emitf(node, "\tmovl %I, %D0\n"); } /** @@ -1672,65 +1706,43 @@ static void emit_ia32_Const(const ir_node *node) */ static void emit_ia32_LdTls(const ir_node *node) { - be_emit_cstring("\tmovl %gs:0, "); - ia32_emit_dest_register(node, 0); - be_emit_finish_line_gas(node); + ia32_emitf(node, "\tmovl %%gs:0, %D0\n"); } /* helper function for emit_ia32_Minus64Bit */ static void emit_mov(const ir_node* node, const arch_register_t *src, const arch_register_t *dst) { - be_emit_cstring("\tmovl "); - emit_register(src, NULL); - be_emit_cstring(", "); - emit_register(dst, NULL); - be_emit_finish_line_gas(node); + ia32_emitf(node, "\tmovl %R, %R\n", src, dst); } /* helper function for emit_ia32_Minus64Bit */ static void emit_neg(const ir_node* node, const arch_register_t *reg) { - be_emit_cstring("\tnegl "); - emit_register(reg, NULL); - be_emit_finish_line_gas(node); + ia32_emitf(node, "\tnegl %R\n", reg); } /* helper function for emit_ia32_Minus64Bit */ static void emit_sbb0(const ir_node* node, const arch_register_t *reg) { - be_emit_cstring("\tsbbl $0, "); - emit_register(reg, NULL); - be_emit_finish_line_gas(node); + ia32_emitf(node, "\tsbbl $0, %R\n", reg); } /* helper function for emit_ia32_Minus64Bit */ static void emit_sbb(const ir_node* node, const arch_register_t *src, const arch_register_t *dst) { - be_emit_cstring("\tsbbl "); - emit_register(src, NULL); - be_emit_cstring(", "); - emit_register(dst, NULL); - be_emit_finish_line_gas(node); + ia32_emitf(node, "\tsbbl %R, %R\n", src, dst); } /* helper function for emit_ia32_Minus64Bit */ static void emit_xchg(const ir_node* node, const arch_register_t *src, const arch_register_t *dst) { - be_emit_cstring("\txchgl "); - emit_register(src, NULL); - be_emit_cstring(", "); - emit_register(dst, NULL); - be_emit_finish_line_gas(node); + ia32_emitf(node, "\txchgl %R, %R\n", src, dst); } /* helper function for emit_ia32_Minus64Bit */ static void emit_zero(const ir_node* node, const arch_register_t *reg) { - be_emit_cstring("\txorl "); - emit_register(reg, NULL); - be_emit_cstring(", "); - emit_register(reg, NULL); - be_emit_finish_line_gas(node); + ia32_emitf(node, "\txorl %R, %R\n", reg, reg); } static void emit_ia32_Minus64Bit(const ir_node *node) @@ -1789,30 +1801,32 @@ zero_neg: static void emit_ia32_GetEIP(const ir_node *node) { - be_emit_cstring("\tcall "); - be_emit_string(pic_base_label); - be_emit_finish_line_gas(node); + ia32_emitf(node, "\tcall %s\n", pic_base_label); + ia32_emitf(NULL, "%s:\n", pic_base_label); + ia32_emitf(node, "\tpopl %D0\n"); +} - be_emit_string(pic_base_label); - be_emit_cstring(":\n"); - be_emit_write_line(); +static void emit_ia32_ClimbFrame(const ir_node *node) +{ + const ia32_climbframe_attr_t *attr = get_ia32_climbframe_attr_const(node); - be_emit_cstring("\tpopl "); - ia32_emit_dest_register(node, 0); - be_emit_char('\n'); - be_emit_write_line(); + 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_emit_cstring("\tret"); + unsigned pop = be_Return_get_pop(node); - pop = be_Return_get_pop(node); - if(pop > 0) { - be_emit_irprintf(" $%d", pop); + if (pop > 0 || be_Return_get_emit_pop(node)) { + ia32_emitf(node, "\tret $%u\n", pop); + } else { + ia32_emitf(node, "\tret\n"); } - be_emit_finish_line_gas(node); } static void emit_Nothing(const ir_node *node) @@ -1835,8 +1849,8 @@ static void emit_Nothing(const ir_node *node) * Enters the emitter functions for handled nodes into the generic * pointer of an opcode. */ -static void ia32_register_emitters(void) { - +static void ia32_register_emitters(void) +{ #define IA32_EMIT2(a,b) op_ia32_##a->ops.generic = (op_func)emit_ia32_##b #define IA32_EMIT(a) IA32_EMIT2(a,a) #define EMIT(a) op_##a->ops.generic = (op_func)emit_##a @@ -1851,40 +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 @@ -1895,15 +1908,49 @@ static void ia32_register_emitters(void) { typedef void (*emit_func_ptr) (const ir_node *); +/** + * Assign and emit an exception label if the current instruction can fail. + */ +static void ia32_assign_exc_label(ir_node *node) +{ + /* assign a new ID to the instruction */ + set_ia32_exc_label_id(node, ++exc_label_id); + /* print it */ + ia32_emit_exc_label(node); + be_emit_char(':'); + be_emit_pad_comment(); + be_emit_cstring("/* exception to Block "); + ia32_emit_cfop_target(node); + be_emit_cstring(" */\n"); + be_emit_write_line(); +} + /** * Emits code for a node. */ -static void ia32_emit_node(const ir_node *node) +static void ia32_emit_node(ir_node *node) { ir_op *op = get_irn_op(node); DBG((dbg, LEVEL_1, "emitting code for %+F\n", node)); + if (is_ia32_irn(node)) { + if (get_ia32_exc_label(node)) { + /* emit the exception label of this instruction */ + ia32_assign_exc_label(node); + } + if (mark_spill_reload) { + if (is_ia32_is_spill(node)) { + ia32_emitf(NULL, "\txchg %ebx, %ebx /* spill mark */\n"); + } + if (is_ia32_is_reload(node)) { + ia32_emitf(NULL, "\txchg %edx, %edx /* reload mark */\n"); + } + if (is_ia32_is_remat(node)) { + ia32_emitf(NULL, "\txchg %ecx, %ecx /* remat mark */\n"); + } + } + } if (op->ops.generic) { emit_func_ptr func = (emit_func_ptr) op->ops.generic; @@ -1922,9 +1969,7 @@ static void ia32_emit_node(const ir_node *node) */ static void ia32_emit_alignment(unsigned align, unsigned skip) { - be_emit_cstring("\t.p2align "); - be_emit_irprintf("%u,,%u\n", align, skip); - be_emit_write_line(); + ia32_emitf(NULL, "\t.p2align %u,,%u\n", align, skip); } /** @@ -1943,37 +1988,38 @@ static void ia32_emit_align_label(void) * 16 bytes. However we should only do that if the alignment nops before the * label aren't executed more often than we have jumps to the label. */ -static int should_align_block(ir_node *block, ir_node *prev) +static int should_align_block(const ir_node *block) { static const double DELTA = .0001; ir_exec_freq *exec_freq = cg->birg->exec_freq; + ir_node *prev = get_prev_block_sched(block); double block_freq; double prev_freq = 0; /**< execfreq of the fallthrough block */ double jmp_freq = 0; /**< execfreq of all non-fallthrough blocks */ int i, n_cfgpreds; - if(exec_freq == NULL) + if (exec_freq == NULL) return 0; - if(ia32_cg_config.label_alignment_factor <= 0) + if (ia32_cg_config.label_alignment_factor <= 0) return 0; block_freq = get_block_execfreq(exec_freq, block); - if(block_freq < DELTA) + if (block_freq < DELTA) return 0; n_cfgpreds = get_Block_n_cfgpreds(block); for(i = 0; i < n_cfgpreds; ++i) { - ir_node *pred = get_Block_cfgpred_block(block, i); - double pred_freq = get_block_execfreq(exec_freq, pred); + const ir_node *pred = get_Block_cfgpred_block(block, i); + double pred_freq = get_block_execfreq(exec_freq, pred); - if(pred == prev) { + if (pred == prev) { prev_freq += pred_freq; } else { jmp_freq += pred_freq; } } - if(prev_freq < DELTA && !(jmp_freq < DELTA)) + if (prev_freq < DELTA && !(jmp_freq < DELTA)) return 1; jmp_freq /= prev_freq; @@ -1981,72 +2027,44 @@ static int should_align_block(ir_node *block, ir_node *prev) return jmp_freq > ia32_cg_config.label_alignment_factor; } -/** - * Return non-zero, if a instruction in a fall-through. - */ -static int is_fallthrough(ir_node *cfgpred) -{ - ir_node *pred; - - if(!is_Proj(cfgpred)) - return 1; - pred = get_Proj_pred(cfgpred); - if(is_ia32_SwitchJmp(pred)) - return 0; - - return 1; -} - /** * Emit the block header for a block. * * @param block the block * @param prev_block the previous block */ -static void ia32_emit_block_header(ir_node *block, ir_node *prev_block) +static void ia32_emit_block_header(ir_node *block) { ir_graph *irg = current_ir_graph; - int n_cfgpreds; - int need_label = 1; + int need_label = block_needs_label(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; - n_cfgpreds = get_Block_n_cfgpreds(block); - - if(n_cfgpreds == 0) { - need_label = 0; - } else if(n_cfgpreds == 1) { - ir_node *cfgpred = get_Block_cfgpred(block, 0); - if(get_nodes_block(cfgpred) == prev_block && is_fallthrough(cfgpred)) { - need_label = 0; - } - } - if (ia32_cg_config.label_alignment > 0) { /* align the current block if: * a) if should be aligned due to its execution frequency * b) there is no fall-through here */ - if (should_align_block(block, prev_block)) { + if (should_align_block(block)) { ia32_emit_align_label(); } else { /* if the predecessor block has no fall-through, we can always align the label. */ int i; - ir_node *check_node = NULL; + int has_fallthrough = 0; - for (i = n_cfgpreds - 1; i >= 0; --i) { + for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) { ir_node *cfg_pred = get_Block_cfgpred(block, i); - - if (get_nodes_block(skip_Proj(cfg_pred)) == prev_block) { - check_node = cfg_pred; + if (can_be_fallthrough(cfg_pred)) { + has_fallthrough = 1; break; } } - if (check_node == NULL || !is_fallthrough(check_node)) + + if (!has_fallthrough) ia32_emit_align_label(); } } @@ -2067,12 +2085,16 @@ static void ia32_emit_block_header(ir_node *block, ir_node *prev_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"); @@ -2083,11 +2105,11 @@ static void ia32_emit_block_header(ir_node *block, ir_node *prev_block) * Walks over the nodes in a block connected by scheduling edges * and emits code for each node. */ -static void ia32_gen_block(ir_node *block, ir_node *last_block) +static void ia32_gen_block(ir_node *block) { - const ir_node *node; + ir_node *node; - ia32_emit_block_header(block, last_block); + ia32_emit_block_header(block); /* emit the contents of the block */ be_dbg_set_dbg_info(get_irn_dbg_info(block)); @@ -2096,31 +2118,49 @@ static void ia32_gen_block(ir_node *block, ir_node *last_block) } } +typedef struct exc_entry { + ir_node *exc_instr; /** The instruction that can issue an exception. */ + ir_node *block; /** The block to call then. */ +} exc_entry; + /** * Block-walker: - * Sets labels for control flow nodes (jump target) + * Sets labels for control flow nodes (jump target). + * Links control predecessors to there destination blocks. */ static void ia32_gen_labels(ir_node *block, void *data) { + exc_entry **exc_list = data; ir_node *pred; - int n = get_Block_n_cfgpreds(block); - (void) data; + int n; - for (n--; n >= 0; n--) { + for (n = get_Block_n_cfgpreds(block) - 1; n >= 0; --n) { pred = get_Block_cfgpred(block, n); set_irn_link(pred, block); + + pred = skip_Proj(pred); + if (is_ia32_irn(pred) && get_ia32_exc_label(pred)) { + exc_entry e; + + e.exc_instr = pred; + e.block = block; + ARR_APP1(exc_entry, *exc_list, e); + set_irn_link(pred, block); + } } } /** - * Emit an exception label if the current instruction can fail. + * Compare two exception_entries. */ -void ia32_emit_exc_label(const ir_node *node) +static int cmp_exc_entry(const void *a, const void *b) { - if (get_ia32_exc_label(node)) { - be_emit_irprintf(".EXL%u\n", 0); - be_emit_write_line(); - } + const exc_entry *ea = a; + const exc_entry *eb = b; + + if (get_ia32_exc_label_id(ea->exc_instr) < get_ia32_exc_label_id(eb->exc_instr)) + return -1; + return +1; } /** @@ -2128,14 +2168,12 @@ void ia32_emit_exc_label(const ir_node *node) */ void ia32_gen_routine(ia32_code_gen_t *ia32_cg, ir_graph *irg) { - ir_node *block; - ir_node *last_block = NULL; ir_entity *entity = get_irg_entity(irg); + exc_entry *exc_list = NEW_ARR_F(exc_entry, 0); int i, n; cg = ia32_cg; - isa = (const ia32_isa_t*) cg->arch_env->isa; - arch_env = cg->arch_env; + isa = cg->isa; do_pic = cg->birg->main_env->options->pic; ia32_register_emitters(); @@ -2145,29 +2183,591 @@ void ia32_gen_routine(ia32_code_gen_t *ia32_cg, ir_graph *irg) be_dbg_method_begin(entity, be_abi_get_stack_layout(cg->birg->abi)); 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, &exc_list); + + /* 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]; + + ia32_gen_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); + + /* Sort the exception table using the exception label id's. + Those are ascending with ascending addresses. */ + qsort(exc_list, ARR_LEN(exc_list), sizeof(exc_list[0]), cmp_exc_entry); + { + int i; + + for (i = 0; i < ARR_LEN(exc_list); ++i) { + be_emit_cstring("\t.long "); + ia32_emit_exc_label(exc_list[i].exc_instr); + be_emit_char('\n'); + be_emit_cstring("\t.long "); + ia32_emit_block_name(exc_list[i].block); + be_emit_char('\n'); + } + } + DEL_ARR_F(exc_list); +} + +static const lc_opt_table_entry_t ia32_emitter_options[] = { + LC_OPT_ENT_BOOL("mark_spill_reload", "mark spills and reloads with ud opcodes", &mark_spill_reload), + LC_OPT_LAST +}; + +/* ==== Experimental binary emitter ==== */ + +static unsigned char reg_map[N_ia32_gp_REGS]; + +static void build_reg_map(void) +{ + reg_map[REG_EAX] = 0x0; + reg_map[REG_ECX] = 0x1; + reg_map[REG_EDX] = 0x2; + reg_map[REG_EBX] = 0x3; + reg_map[REG_ESP] = 0x4; + reg_map[REG_EBP] = 0x5; + reg_map[REG_ESI] = 0x6; + reg_map[REG_EDI] = 0x7; +} + +/* 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(); +} + +static void bemit_entity(ir_entity *entity, bool entity_sign, int offset, + bool is_relative) +{ + if (entity == NULL) { + bemit32(offset); + return; + } + + /* 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(); +} + +/* end emit routines, all emitters following here should only use the functions + above. */ + +static void bemit_modrr(const arch_register_t *op1_dest, + const arch_register_t *op2) +{ + unsigned char modrm = 0xC0; + modrm |= reg_map[op1_dest->index]; + modrm |= reg_map[op2->index] << 3; + bemit8(modrm); +} + +static void bemit_modru(const arch_register_t *dest, unsigned val) +{ + unsigned char modrm = 0xC0; + assert(val <= 7); + modrm |= reg_map[dest->index]; + modrm |= val << 3; + bemit8(modrm); +} + +static unsigned get_imm_size(ir_entity *entity, int offset) +{ + if (entity != NULL) + return 32; + if (offset >= -127 && offset < 128) { + return 8; + } else if (offset >= -32768 && offset < 32767) { + return 16; + } else { + return 32; + } +} + +static void bemit_binop_with_imm(const ir_node *node, unsigned opimm8, + unsigned opimm32, unsigned ruval) +{ + const arch_register_t *reg = get_out_reg(node, 0); + 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 = get_imm_size(attr->symconst, attr->offset); + + switch (size) { + case 8: + bemit8(opimm8); + bemit_modru(reg, ruval); + bemit8(attr->offset); + return; + case 16: + case 32: + bemit8(opimm32); + bemit_modru(reg, ruval); + bemit_entity(attr->symconst, attr->sc_sign, attr->offset, false); + return; + } + panic("invalid imm size?!?"); +} + +static void bemit_modsourceam(unsigned dest_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; + + /* set the mod part depending on displacement */ + if (ent != NULL) { + modrm |= 0x80; + emitoffs = 32; + } else if (offs == 0) { + emitoffs = 0; + } else if (offs >= -127 && offs <= 128) { + modrm |= 0x40; + emitoffs = 8; + } else { + modrm |= 0x80; + emitoffs = 32; + } + + /* determine if we need a SIB byte */ + if (has_index) { + int scale; + const arch_register_t *reg_index = arch_get_irn_register(index); + assert(reg_index->index != REG_ESP); + sib |= reg_map[reg_index->index] << 3; + + if (has_base) { + const arch_register_t *reg = arch_get_irn_register(base); + sib |= reg_map[reg->index]; + } else { + sib |= 0x05; + } + + scale = get_ia32_am_scale(node); + assert(scale < 4); + sib |= scale << 6; + emitsib = true; + } + + /* determine modrm byte */ + if (emitsib) { + modrm |= 0x04; + } else if (has_base) { + const arch_register_t *reg = arch_get_irn_register(base); + /* we are forced to emit a sib when base is ESP */ + if (reg->index == REG_ESP) { + sib = 0x24; + emitsib = true; + + /* we are forced to emit a 32bit offset as EBP base without + offset is a special case for displacement without base */ + } else if (reg->index == REG_EBP && emitoffs == 0) { + assert( (modrm & 0xC0) == 0); + emitoffs = 8; + modrm |= 0x40; + } + modrm |= reg_map[reg->index]; + } else { + modrm = 0x05; + emitoffs = 32; + } + + modrm |= dest_reg << 3; + + 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); + } +} + +static void bemit_binop(const ir_node *node, unsigned modrr, unsigned am) +{ + const arch_register_t *out = get_in_reg(node, n_ia32_binary_left); + if (get_ia32_op_type(node) == ia32_AddrModeS) { + bemit8(am); + bemit_modsourceam(reg_map[out->index], node); + } else { + const arch_register_t *op2 = get_in_reg(node, n_ia32_binary_right); + assert(get_ia32_op_type(node) == ia32_Normal); + bemit8(modrr); + bemit_modrr(out, op2); + } +} + +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 ir_node *op = be_get_Copy_op(copy); + const arch_register_t *in = arch_get_irn_register(op); + const arch_register_t *out = arch_get_irn_register(copy); + + 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_const(const ir_node *node) +{ + const arch_register_t *out = get_out_reg(node, 0); + bemit8(0xB8 + reg_map[out->index]); + bemit_immediate(node, false); +} + +static void bemit_add(const ir_node *node) +{ + ir_node *right = get_irn_n(node, n_ia32_binary_right); + if (is_ia32_Immediate(right)) { + /* TODO: there's a shorter variant with DEST=EAX */ + bemit_binop_with_imm(node, 0x83, 0x81, 0); + } else { + bemit_binop(node, 0x01, 0x03); + } +} + +static void bemit_sub(const ir_node *node) +{ + ir_node *right = get_irn_n(node, n_ia32_binary_right); + if (is_ia32_Immediate(right)) { + /* TODO: there's a shorter variant with DEST=EAX */ + bemit_binop_with_imm(node, 0x83, 0x81, 5); + } else { + bemit_binop(node, 0x29, 0x2B); + } +} + +static void bemit_xor(const ir_node *node) +{ + ir_node *right = get_irn_n(node, n_ia32_binary_right); + if (is_ia32_Immediate(right)) { + /* TODO: there's a shorter variant with DEST=EAX */ + bemit_binop_with_imm(node, 0x83, 0x81, 6); + } else { + bemit_binop(node, 0x31, 0x33); + } +} + +static void bemit_not(const ir_node *node) +{ + const arch_register_t *reg = get_out_reg(node, 0); + bemit8(0xF7); + bemit_modru(reg, 2); +} + +static void bemit_lea(const ir_node *node) +{ + const arch_register_t *out = get_out_reg(node, 0); + bemit8(0x8D); + bemit_modsourceam(reg_map[out->index], node); +} + +static void bemit_cltd(const ir_node *node) +{ + (void) node; + bemit8(0x99); +} + +static void bemit_load(const ir_node *node) +{ + const arch_register_t *out = get_out_reg(node, 0); + + /* TODO: load from constant address to EAX can be encoded + as 0xA1 [offset] */ + bemit8(0x8B); + bemit_modsourceam(reg_map[out->index], node); +} + +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_modsourceam(0, node); + bemit_immediate(value, false); + } else { + /* TODO: store to constant address from EAX can be encoded as + 0xA3 [offset]*/ + const arch_register_t *in = get_in_reg(node, n_ia32_Store_val); + bemit8(0x89); + bemit_modsourceam(reg_map[in->index], node); + } +} + +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_imm_size(attr->symconst, attr->offset); + /* TODO: check for bitsizes different from 32... */ + switch (size) { + case 8: + bemit8(0x6A); + bemit8(attr->offset); + break; + case 16: + case 32: + bemit8(0x68); + bemit_immediate(value, false); + break; + } + } else { + bemit8(0xFF); + bemit_modsourceam(6, node); + } +} + +static void bemit_pop(const ir_node *node) +{ + const arch_register_t *reg = get_out_reg(node, pn_ia32_Pop_res); + /* TODO: check for AM pop */ + bemit8(0x58 + reg_map[reg->index]); +} + +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 { + panic("indirect call NIY"); + } +} + +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) +{ + const arch_register_t *reg = get_out_reg(node, 0); + int offs = be_get_IncSP_offset(node); + unsigned size = get_imm_size(NULL, offs); + + if (offs > 0) { + bemit8(size == 8 ? 0x83 : 0x81); + bemit_modru(reg, 5); /* sub */ + if (size == 8) { + bemit8(offs); + } else { + bemit32(offs); + } + } else if (offs < 0) { + bemit8(size == 8 ? 0x83 : 0x81); + bemit_modru(reg, 0); /* add */ + if (size == 8) { + 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_Return, bemit_return); + register_emitter(op_be_IncSP, bemit_incsp); + register_emitter(op_ia32_Add, bemit_add); + register_emitter(op_ia32_Call, bemit_call); + register_emitter(op_ia32_Cltd, bemit_cltd); + register_emitter(op_ia32_Sub, bemit_sub); + register_emitter(op_ia32_Xor0, bemit_xor0); + register_emitter(op_ia32_Xor, bemit_xor); + register_emitter(op_ia32_Const, bemit_const); + register_emitter(op_ia32_Lea, bemit_lea); + register_emitter(op_ia32_Load, bemit_load); + register_emitter(op_ia32_Not, bemit_not); + register_emitter(op_ia32_Push, bemit_push); + register_emitter(op_ia32_Pop, bemit_pop); + register_emitter(op_ia32_Store, bemit_store); + + /* 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;) { - ir_node *next_bl; + for (i = 0; i < n; ++i) { + ir_node *block = cg->blk_sched[i]; + ir_node *prev = i > 0 ? cg->blk_sched[i-1] : NULL; - block = cg->blk_sched[i]; - ++i; - next_bl = i < n ? cg->blk_sched[i] : NULL; + set_irn_link(block, prev); + } - /* set here the link. the emitter expects to find the next block here */ - set_irn_link(block, next_bl); - ia32_gen_block(block, last_block); - last_block = block; + 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; + lc_opt_entry_t *ia32_grp; + + be_grp = lc_opt_get_grp(firm_opt_get_root(), "be"); + ia32_grp = lc_opt_get_grp(be_grp, "ia32"); + + lc_opt_add_table(ia32_grp, ia32_emitter_options); + + build_reg_map(); + FIRM_DBG_REGISTER(dbg, "firm.be.ia32.emitter"); }