Remove the attribute const arch_env_t *arch_env from struct ia32_code_gen_t. We...
[libfirm] / ir / be / ia32 / ia32_common_transform.c
index a273dec..d91cbf3 100644 (file)
@@ -24,6 +24,7 @@
  * @author      Matthias Braun, Sebastian Buchwald
  * @version     $Id: ia32_common_transform.c 21012 2008-08-06 13:35:17Z beck $
  */
+#include "config.h"
 
 #include "error.h"
 #include "irargs_t.h"
@@ -171,7 +172,7 @@ ir_node *create_Immediate(ir_entity *symconst, int symconst_sign, long val)
        ir_node  *start_block = get_irg_start_block(irg);
        ir_node  *immediate   = new_rd_ia32_Immediate(NULL, irg, start_block,
                                                      symconst, symconst_sign, val);
-       arch_set_irn_register(env_cg->arch_env, immediate, &ia32_gp_regs[REG_GP_NOREG]);
+       arch_set_irn_register(immediate, &ia32_gp_regs[REG_GP_NOREG]);
 
        return immediate;
 }
@@ -204,7 +205,7 @@ const arch_register_t *ia32_get_clobber_register(const char *clobber)
 
 #ifndef NDEBUG
 const char *ia32_get_old_node_name(ia32_code_gen_t *cg, ir_node *irn) {
-       ia32_isa_t *isa = (ia32_isa_t*) cg->arch_env;
+       const ia32_isa_t *isa = cg->isa;
 
        lc_eoprintf(firm_get_arg_env(), isa->name_obst, "%+F", irn);
        obstack_1grow(isa->name_obst, 0);
@@ -223,7 +224,6 @@ int ia32_mode_needs_gp_reg(ir_mode *mode) {
 static void parse_asm_constraints(constraint_t *constraint, const char *c,
                            int is_output)
 {
-       asm_constraint_flags_t       flags              = 0;
        char                         immediate_type     = '\0';
        unsigned                     limited            = 0;
        const arch_register_class_t *cls                = NULL;
@@ -251,15 +251,9 @@ static void parse_asm_constraints(constraint_t *constraint, const char *c,
                case '\n':
                        break;
 
-               case '=':
-                       flags |= ASM_CONSTRAINT_FLAG_MODIFIER_WRITE
-                               | ASM_CONSTRAINT_FLAG_MODIFIER_NO_READ;
-                       break;
-
-               case '+':
-                       flags |= ASM_CONSTRAINT_FLAG_MODIFIER_WRITE
-                               | ASM_CONSTRAINT_FLAG_MODIFIER_READ;
-                       break;
+               /* Skip out/in-out marker */
+               case '=': break;
+               case '+': break;
 
                case '*':
                        ++c;
@@ -457,8 +451,8 @@ static void parse_asm_constraints(constraint_t *constraint, const char *c,
 ir_node *gen_ASM(ir_node *node)
 {
        ir_graph                   *irg       = current_ir_graph;
-       ir_node                    *block     = get_nodes_block(node);
-       ir_node                    *new_block = be_transform_node(block);
+       ir_node                    *block = NULL;
+       ir_node                    *new_block = NULL;
        dbg_info                   *dbgi      = get_irn_dbg_info(node);
        int                         i, arity;
        int                         out_idx;
@@ -480,6 +474,23 @@ ir_node *gen_ASM(ir_node *node)
 
        memset(&clobber_bits, 0, sizeof(clobber_bits));
 
+       switch (be_transformer) {
+       case TRANSFORMER_DEFAULT:
+               block     = get_nodes_block(node);
+               new_block = be_transform_node(block);
+               break;
+
+#ifdef FIRM_GRGEN_BE
+       case TRANSFORMER_PBQP:
+       case TRANSFORMER_RAND:
+               new_block = get_nodes_block(node);
+               break;
+#endif
+
+       default:
+               panic("invalid transformer");
+       }
+
        /* workaround for lots of buggy code out there as most people think volatile
         * asm is enough for everything and forget the flags (linux kernel, etc.)
         */
@@ -493,7 +504,7 @@ ir_node *gen_ASM(ir_node *node)
 
        clobbers   = get_ASM_clobbers(node);
        n_clobbers = 0;
-       for(i = 0; i < get_ASM_n_clobbers(node); ++i) {
+       for (i = 0; i < get_ASM_n_clobbers(node); ++i) {
                const arch_register_req_t *req;
                const char                *c = get_id_str(clobbers[i]);
 
@@ -516,12 +527,12 @@ ir_node *gen_ASM(ir_node *node)
        out_constraints = get_ASM_output_constraints(node);
 
        /* determine size of register_map */
-       for(out_idx = 0; out_idx < n_out_constraints; ++out_idx) {
+       for (out_idx = 0; out_idx < n_out_constraints; ++out_idx) {
                const ir_asm_constraint *constraint = &out_constraints[out_idx];
                if (constraint->pos > reg_map_size)
                        reg_map_size = constraint->pos;
        }
-       for(i = 0; i < arity; ++i) {
+       for (i = 0; i < arity; ++i) {
                const ir_asm_constraint   *constraint = &in_constraints[i];
                if(constraint->pos > reg_map_size)
                        reg_map_size = constraint->pos;
@@ -535,7 +546,7 @@ ir_node *gen_ASM(ir_node *node)
        /* construct output constraints */
        out_reg_reqs = obstack_alloc(obst, out_arity * sizeof(out_reg_reqs[0]));
 
-       for(out_idx = 0; out_idx < n_out_constraints; ++out_idx) {
+       for (out_idx = 0; out_idx < n_out_constraints; ++out_idx) {
                const ir_asm_constraint   *constraint = &out_constraints[out_idx];
                const char                *c       = get_id_str(constraint->constraint);
                unsigned                   pos        = constraint->pos;
@@ -556,7 +567,7 @@ ir_node *gen_ASM(ir_node *node)
 
        /* inputs + input constraints */
        in_reg_reqs = obstack_alloc(obst, arity * sizeof(in_reg_reqs[0]));
-       for(i = 0; i < arity; ++i) {
+       for (i = 0; i < arity; ++i) {
                ir_node                   *pred         = get_irn_n(node, i);
                const ir_asm_constraint   *constraint   = &in_constraints[i];
                ident                     *constr_id    = constraint->constraint;
@@ -592,8 +603,22 @@ ir_node *gen_ASM(ir_node *node)
                }
 
                if (input == NULL) {
-                       ir_node *pred = get_irn_n(node, i);
-                       input         = be_transform_node(pred);
+                       ir_node *pred = NULL;
+                       switch (be_transformer) {
+                       case TRANSFORMER_DEFAULT:
+                               pred  = get_irn_n(node, i);
+                               input = be_transform_node(pred);
+                               break;
+
+#ifdef FIRM_GRGEN_BE
+                       case TRANSFORMER_PBQP:
+                       case TRANSFORMER_RAND:
+                               input = get_irn_n(node, i);
+                               break;
+#endif
+
+                       default: panic("invalid transformer");
+                       }
 
                        if (parsed_constraint.cls == NULL
                                        && parsed_constraint.same_as < 0) {
@@ -612,7 +637,7 @@ ir_node *gen_ASM(ir_node *node)
        }
 
        /* parse clobbers */
-       for(i = 0; i < get_ASM_n_clobbers(node); ++i) {
+       for (i = 0; i < get_ASM_n_clobbers(node); ++i) {
                const char                *c = get_id_str(clobbers[i]);
                const arch_register_req_t *req;
 
@@ -627,6 +652,9 @@ ir_node *gen_ASM(ir_node *node)
        new_node = new_rd_ia32_Asm(dbgi, irg, new_block, arity, in, out_arity,
                                   get_ASM_text(node), register_map);
 
+       if (arity == 0)
+               be_dep_on_frame(new_node);
+
        set_ia32_out_req_all(new_node, out_reg_reqs);
        set_ia32_in_req_all(new_node, in_reg_reqs);
 
@@ -635,6 +663,93 @@ ir_node *gen_ASM(ir_node *node)
        return new_node;
 }
 
+ir_node *gen_CopyB(ir_node *node) {
+       ir_node  *block    = NULL;
+       ir_node  *src      = NULL;
+       ir_node  *new_src  = NULL;
+       ir_node  *dst      = NULL;
+       ir_node  *new_dst  = NULL;
+       ir_node  *mem      = NULL;
+       ir_node  *new_mem  = NULL;
+       ir_node  *res      = NULL;
+       ir_graph *irg      = current_ir_graph;
+       dbg_info *dbgi     = get_irn_dbg_info(node);
+       int      size      = get_type_size_bytes(get_CopyB_type(node));
+       int      rem;
+
+       switch (be_transformer) {
+               case TRANSFORMER_DEFAULT:
+                       block    = be_transform_node(get_nodes_block(node));
+                       src      = get_CopyB_src(node);
+                       new_src  = be_transform_node(src);
+                       dst      = get_CopyB_dst(node);
+                       new_dst  = be_transform_node(dst);
+                       mem      = get_CopyB_mem(node);
+                       new_mem  = be_transform_node(mem);
+                       break;
+
+#ifdef FIRM_GRGEN_BE
+               case TRANSFORMER_PBQP:
+               case TRANSFORMER_RAND:
+                       block    = get_nodes_block(node);
+                       new_src  = get_CopyB_src(node);
+                       new_dst  = get_CopyB_dst(node);
+                       new_mem  = get_CopyB_mem(node);
+                       break;
+#endif
+
+               default: panic("invalid transformer");
+       }
+
+       /* If we have to copy more than 32 bytes, we use REP MOVSx and */
+       /* then we need the size explicitly in ECX.                    */
+       if (size >= 32 * 4) {
+               rem = size & 0x3; /* size % 4 */
+               size >>= 2;
+
+               res = new_rd_ia32_Const(dbgi, irg, block, NULL, 0, size);
+               be_dep_on_frame(res);
+
+               res = new_rd_ia32_CopyB(dbgi, irg, block, new_dst, new_src, res, new_mem, rem);
+       } else {
+               if(size == 0) {
+                       ir_fprintf(stderr, "Optimization warning copyb %+F with size <4\n",
+                                  node);
+               }
+               res = new_rd_ia32_CopyB_i(dbgi, irg, block, new_dst, new_src, new_mem, size);
+       }
+
+       SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
+
+       return res;
+}
+
+ir_node *gen_Proj_tls(ir_node *node) {
+       ir_node  *block = NULL;
+       ir_graph *irg   = current_ir_graph;
+       dbg_info *dbgi  = NULL;
+       ir_node  *res   = NULL;
+
+       switch (be_transformer) {
+               case TRANSFORMER_DEFAULT:
+                       block = be_transform_node(get_nodes_block(node));
+                       break;
+
+#ifdef FIRM_GRGEN_BE
+               case TRANSFORMER_PBQP:
+               case TRANSFORMER_RAND:
+                       block = get_nodes_block(node);
+                       break;
+#endif
+
+               default: panic("invalid transformer");
+       }
+
+       res   = new_rd_ia32_LdTls(dbgi, irg, block, mode_Iu);
+
+       return res;
+}
+
 ir_node *gen_Unknown(ir_node *node)
 {
        ir_mode *mode = get_irn_mode(node);
@@ -649,12 +764,7 @@ ir_node *gen_Unknown(ir_node *node)
                        ir_node  *block = get_irg_start_block(irg);
                        ir_node  *ret   = new_rd_ia32_vfldz(dbgi, irg, block);
 
-                       /* Const Nodes before the initial IncSP are a bad idea, because
-                        * they could be spilled and we have no SP ready at that point yet.
-                        * So add a dependency to the initial frame pointer calculation to
-                        * avoid that situation.
-                        */
-                       add_irn_dep(ret, get_irg_frame(irg));
+                       be_dep_on_frame(ret);
                        return ret;
                }
        } else if (ia32_mode_needs_gp_reg(mode)) {
@@ -678,14 +788,12 @@ const arch_register_req_t *make_register_req(const constraint_t *constraint,
                if (same_as >= n_outs)
                        panic("invalid output number in same_as constraint");
 
-               other_constr         = out_reqs[same_as];
+               other_constr     = out_reqs[same_as];
 
-               req                  = obstack_alloc(obst, sizeof(req[0]));
-               req->cls             = other_constr->cls;
-               req->type            = arch_register_req_type_should_be_same;
-               req->limited         = NULL;
-               req->other_same      = 1U << pos;
-               req->other_different = 0;
+               req              = obstack_alloc(obst, sizeof(req[0]));
+               *req             = *other_constr;
+               req->type       |= arch_register_req_type_should_be_same;
+               req->other_same  = 1U << pos;
 
                /* switch constraints. This is because in firm we have same_as
                 * constraints on the output constraints while in the gcc asm syntax
@@ -728,7 +836,7 @@ const arch_register_req_t *parse_clobber(const char *clobber)
        unsigned              *limited;
 
        if(reg == NULL) {
-               panic("Register '%s' mentioned in asm clobber is unknown\n", clobber);
+               panic("Register '%s' mentioned in asm clobber is unknown", clobber);
        }
 
        assert(reg->index < 32);
@@ -745,6 +853,45 @@ const arch_register_req_t *parse_clobber(const char *clobber)
        return req;
 }
 
+
+int prevents_AM(ir_node *const block, ir_node *const am_candidate,
+                       ir_node *const other)
+{
+       if (get_nodes_block(other) != block)
+               return 0;
+
+       if (is_Sync(other)) {
+               int i;
+
+               for (i = get_Sync_n_preds(other) - 1; i >= 0; --i) {
+                       ir_node *const pred = get_Sync_pred(other, i);
+
+                       if (get_nodes_block(pred) != block)
+                               continue;
+
+                       /* Do not block ourselves from getting eaten */
+                       if (is_Proj(pred) && get_Proj_pred(pred) == am_candidate)
+                               continue;
+
+                       if (!heights_reachable_in_block(heights, pred, am_candidate))
+                               continue;
+
+                       return 1;
+               }
+
+               return 0;
+       } else {
+               /* Do not block ourselves from getting eaten */
+               if (is_Proj(other) && get_Proj_pred(other) == am_candidate)
+                       return 0;
+
+               if (!heights_reachable_in_block(heights, other, am_candidate))
+                       return 0;
+
+               return 1;
+       }
+}
+
 ir_node *try_create_Immediate(ir_node *node, char immediate_constraint_type)
 {
        int          minus         = 0;