removed firm proj num translation function, existing projs are renumbered instead
[libfirm] / ir / be / ia32 / ia32_emitter.c
index 997996f..fbd3761 100644 (file)
@@ -128,13 +128,6 @@ static void ia32_dump_function_size(FILE *F, const char *name)
  * |_|                                       |_|
  *************************************************************/
 
-static INLINE int be_is_unknown_reg(const arch_register_t *reg) {
-       return \
-               REGS_ARE_EQUAL(reg, &ia32_gp_regs[REG_GP_UKNWN])   || \
-               REGS_ARE_EQUAL(reg, &ia32_xmm_regs[REG_XMM_UKNWN]) || \
-               REGS_ARE_EQUAL(reg, &ia32_vfp_regs[REG_VFP_UKNWN]);
-}
-
 /**
  * returns true if a node has x87 registers
  */
@@ -165,13 +158,29 @@ static const arch_register_t *get_in_reg(const ir_node *irn, int pos) {
 
        assert(reg && "no in register found");
 
-       /* in case of unknown: just return a register */
-       if (REGS_ARE_EQUAL(reg, &ia32_gp_regs[REG_GP_UKNWN]))
-               reg = &ia32_gp_regs[REG_EAX];
-       else if (REGS_ARE_EQUAL(reg, &ia32_xmm_regs[REG_XMM_UKNWN]))
-               reg = &ia32_xmm_regs[REG_XMM0];
-       else if (REGS_ARE_EQUAL(reg, &ia32_vfp_regs[REG_VFP_UKNWN]))
-               reg = &ia32_vfp_regs[REG_VF0];
+       /* in case of a joker register: just return a valid register */
+       if (arch_register_type_is(reg, joker)) {
+               arch_register_req_t       req;
+               const arch_register_req_t *p_req;
+
+               /* ask for the requirements */
+               p_req = arch_get_register_req(arch_env, &req, irn, pos);
+
+               if (arch_register_req_is(p_req, limited)) {
+                       /* in case of limited requirements: get the first allowed register */
+
+                       bitset_t *bs = bitset_alloca(arch_register_class_n_regs(p_req->cls));
+                       int      idx;
+
+                       p_req->limited(p_req->limited_env, bs);
+                       idx = bitset_next_set(bs, 0);
+                       reg = arch_register_for_index(p_req->cls, idx);
+               }
+               else {
+                       /* otherwise get first register in class */
+                       reg = arch_register_for_index(p_req->cls, 0);
+               }
+       }
 
        return reg;
 }
@@ -653,17 +662,16 @@ const char *ia32_emit_am(const ir_node *n, ia32_emit_env_t *env) {
        }
 
        if (am_flav & ia32_O) {
-               s = get_ia32_am_offs(n);
+               int offs = get_ia32_am_offs_int(n);
 
-               if (s) {
+               if (offs != 0) {
                        /* omit explicit + if there was no base or index */
                        if (! had_output) {
-                               obstack_printf(obst, "[");
-                               if (s[0] == '+')
-                                       s++;
+                               obstack_printf(obst, "[%d", offs);
+                       } else {
+                               obstack_printf(obst, "%+d", offs);
                        }
 
-                       obstack_printf(obst, s);
                        had_output = 1;
                }
        }
@@ -1592,6 +1600,7 @@ static void emit_ia32_Conv_I2I(const ir_node *irn, ia32_emit_env_t *emit_env) {
        char               *conv_cmd = NULL;
        ir_mode            *src_mode = get_ia32_src_mode(irn);
        ir_mode            *tgt_mode = get_ia32_tgt_mode(irn);
+       int                signed_mode;
        int n, m;
        char cmd_buf[SNPRINTF_BUF_LEN], cmnt_buf[SNPRINTF_BUF_LEN];
        const arch_register_t *in_reg, *out_reg;
@@ -1599,33 +1608,34 @@ static void emit_ia32_Conv_I2I(const ir_node *irn, ia32_emit_env_t *emit_env) {
        n = get_mode_size_bits(src_mode);
        m = get_mode_size_bits(tgt_mode);
 
-       if (mode_is_signed(n < m ? src_mode : tgt_mode)) {
+       assert(n == 8 || n == 16 || n == 32);
+       assert(m == 8 || m == 16 || m == 32);
+       assert(n != m);
+
+       signed_mode = mode_is_signed(n < m ? src_mode : tgt_mode);
+       if(signed_mode) {
                move_cmd = "movsx";
-               if (n == 8 || m == 8)
-                       conv_cmd = "cbw";
-               else if (n == 16 || m == 16)
-                       conv_cmd = "cwde";
-               else {
-                       printf("%d -> %d unsupported\n", n, m);
-                       assert(0 && "unsupported Conv_I2I");
-               }
        }
 
-        switch(get_ia32_op_type(irn)) {
+       switch(get_ia32_op_type(irn)) {
                case ia32_Normal:
                        in_reg  = get_in_reg(irn, 2);
                        out_reg = get_out_reg(irn, 0);
 
                        if (REGS_ARE_EQUAL(in_reg, &ia32_gp_regs[REG_EAX]) &&
                                REGS_ARE_EQUAL(out_reg, in_reg)                &&
-                               mode_is_signed(n < m ? src_mode : tgt_mode))
+                               signed_mode)
                        {
+                               if (n == 8 || m == 8)
+                                       conv_cmd = "cbw";
+                               else if (n == 16 || m == 16)
+                                       conv_cmd = "cwde";
+
                                /* argument and result are both in EAX and */
                                /* signedness is ok: -> use converts       */
                                lc_esnprintf(env, cmd_buf, SNPRINTF_BUF_LEN, "%s", conv_cmd);
                        }
-                       else if (REGS_ARE_EQUAL(out_reg, in_reg) &&
-                               ! mode_is_signed(n < m ? src_mode : tgt_mode))
+                       else if (REGS_ARE_EQUAL(out_reg, in_reg) &&     ! signed_mode)
                        {
                                /* argument and result are in the same register */
                                /* and signedness is ok: -> use and with mask   */
@@ -1735,7 +1745,7 @@ static void Copy_emitter(const ir_node *irn, ir_node *op, ia32_emit_env_t *emit_
        char cmd_buf[SNPRINTF_BUF_LEN], cmnt_buf[SNPRINTF_BUF_LEN];
 
        if (REGS_ARE_EQUAL(arch_get_irn_register(aenv, irn), arch_get_irn_register(aenv, op)) ||
-               be_is_unknown_reg(arch_get_irn_register(aenv, op)))
+               arch_register_type_is(arch_get_irn_register(aenv, op), joker))
                return;
 
        if (mode_is_float(get_irn_mode(irn)))
@@ -2112,7 +2122,7 @@ static void ia32_emit_align_label(FILE *F, cpu_support cpu) {
 }
 
 static int is_first_loop_block(ir_node *block, ir_node *prev_block, ia32_emit_env_t *env) {
-       exec_freq_t *execfreqs = env->cg->birg->execfreqs;
+       ir_exec_freq *execfreqs = env->cg->birg->execfreqs;
        double block_freq, prev_freq;
        static const double DELTA = .0001;
        cpu_support cpu = env->isa->opt_arch;
@@ -2134,7 +2144,7 @@ static int is_first_loop_block(ir_node *block, ir_node *prev_block, ia32_emit_en
                case arch_athlon:
                case arch_athlon_64:
                case arch_k6:
-                       return block_freq > 5;
+                       return block_freq > 3;
                default:
                        break;
        }
@@ -2146,28 +2156,53 @@ static int is_first_loop_block(ir_node *block, ir_node *prev_block, ia32_emit_en
  * 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, void *env) {
-       ia32_emit_env_t *emit_env = env;
+static void ia32_gen_block(ir_node *block, ir_node *last_block, ia32_emit_env_t *env) {
+       ir_graph *irg = get_irn_irg(block);
+       ir_node *start_block = get_irg_start_block(irg);
        const ir_node *irn;
-       int need_label = block != get_irg_start_block(get_irn_irg(block));
-       FILE *F = emit_env->out;
+       int need_label = 1;
+       FILE *F = env->out;
 
-       if (! is_Block(block))
-               return;
+       assert(is_Block(block));
+
+       if(block == start_block)
+               need_label = 0;
 
-       if (need_label && (emit_env->cg->opt & IA32_OPT_EXTBB)) {
-               /* if the extended block scheduler is used, only leader blocks need
-                  labels. */
-               need_label = (block == get_extbb_leader(get_nodes_extbb(block)));
+       if (need_label && get_irn_arity(block) == 1) {
+               ir_node *pred_block = get_Block_cfgpred_block(block, 0);
+
+               if(pred_block == last_block && get_irn_n_edges_kind(pred_block, EDGE_KIND_BLOCK) <= 2)
+                       need_label = 0;
+       }
+
+       // special case because the start block contains no jump instruction
+       if(last_block == start_block) {
+               const ir_edge_t *edge;
+               ir_node *startsucc = NULL;
+
+               foreach_block_succ(start_block, edge) {
+                       startsucc = get_edge_src_irn(edge);
+                       if(startsucc != start_block)
+                               break;
+               }
+               assert(startsucc != NULL);
+
+               if(startsucc != block) {
+                       char buf[SNPRINTF_BUF_LEN];
+                       ir_snprintf(buf, sizeof(buf), BLOCK_PREFIX("%d"),
+                                   get_irn_node_nr(startsucc));
+                       ir_fprintf(F, "\tjmp %s\n", buf);
+               }
        }
 
        if (need_label) {
                char cmd_buf[SNPRINTF_BUF_LEN];
                int i, arity;
                int align = 1;
+               ir_exec_freq *execfreqs = env->cg->birg->execfreqs;
 
                // align the loop headers
-               if(!is_first_loop_block(block, last_block, emit_env)) {
+               if(!is_first_loop_block(block, last_block, env)) {
 
                        // align blocks where the previous block has no fallthrough
                        arity = get_irn_arity(block);
@@ -2181,7 +2216,7 @@ static void ia32_gen_block(ir_node *block, ir_node *last_block, void *env) {
                }
 
                if(align)
-                       ia32_emit_align_label(emit_env->out, emit_env->isa->opt_arch);
+                       ia32_emit_align_label(env->out, env->isa->opt_arch);
 
                ir_snprintf(cmd_buf, sizeof(cmd_buf), BLOCK_PREFIX("%d:"),
                            get_irn_node_nr(block));
@@ -2195,6 +2230,10 @@ static void ia32_gen_block(ir_node *block, ir_node *last_block, void *env) {
                        ir_node *predblock = get_Block_cfgpred_block(block, i);
                        fprintf(F, " %ld", get_irn_node_nr(predblock));
                }
+               if(execfreqs != NULL) {
+                       fprintf(F, " freq: %f", get_block_execfreq(execfreqs, block));
+               }
+
                fprintf(F, " */\n");
        }