X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbeabi.c;h=d8dafd90a5284baa7b207c227d95e2bf9e3a4ed7;hb=964aebaae5f9bf41d57592e1cfdc878410c6297e;hp=068191bd270a11aac8d0946cd70470c47e579b28;hpb=443730f33bedbae60304c7773d6b14062eceaa6d;p=libfirm diff --git a/ir/be/beabi.c b/ir/be/beabi.c index 068191bd2..d8dafd90a 100644 --- a/ir/be/beabi.c +++ b/ir/be/beabi.c @@ -1,5 +1,5 @@ /* - * Copyright (C) 1995-2007 University of Karlsruhe. All right reserved. + * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved. * * This file is part of libFirm. * @@ -61,16 +61,18 @@ typedef struct _be_abi_call_arg_t { unsigned in_reg : 1; /**< 1: this argument is transmitted in registers. */ unsigned on_stack : 1; /**< 1: this argument is transmitted on the stack. */ - int pos; + int pos; const arch_register_t *reg; - ir_entity *stack_ent; - unsigned alignment; /**< stack alignment */ - unsigned space_before; /**< allocate space before */ - unsigned space_after; /**< allocate space after */ + ir_entity *stack_ent; + ir_mode *load_mode; + unsigned alignment; /**< stack alignment */ + unsigned space_before; /**< allocate space before */ + unsigned space_after; /**< allocate space after */ } be_abi_call_arg_t; struct _be_abi_call_t { - be_abi_call_flags_t flags; + be_abi_call_flags_t flags; + int pop; const be_abi_callbacks_t *cb; ir_type *between_type; set *params; @@ -90,8 +92,6 @@ struct _be_abi_irg_t { ir_node *init_sp; /**< The node representing the stack pointer at the start of the function. */ - ir_node *start_barrier; /**< The barrier of the start block */ - ir_node *reg_params; /**< The reg params node. */ pmap *regs; /**< A map of all callee-save and ignore regs to their Projs to the RegParams node. */ @@ -108,7 +108,7 @@ struct _be_abi_irg_t { arch_register_req_t sp_req; arch_register_req_t sp_cls_req; - DEBUG_ONLY(firm_dbg_module_t *dbg;) /**< The debugging module. */ + DEBUG_ONLY(firm_dbg_module_t *dbg;) /**< The debugging module. */ }; static heights_t *ir_heights; @@ -180,6 +180,11 @@ void be_abi_call_set_flags(be_abi_call_t *call, be_abi_call_flags_t flags, const call->cb = cb; } +void be_abi_call_set_pop(be_abi_call_t *call, int pop) +{ + assert(pop >= 0); + call->pop = pop; +} /* Set register class for call address */ void be_abi_call_set_call_address_reg_class(be_abi_call_t *call, const arch_register_class_t *cls) @@ -188,10 +193,11 @@ void be_abi_call_set_call_address_reg_class(be_abi_call_t *call, const arch_regi } -void be_abi_call_param_stack(be_abi_call_t *call, int arg_pos, unsigned alignment, unsigned space_before, unsigned space_after) +void be_abi_call_param_stack(be_abi_call_t *call, int arg_pos, ir_mode *load_mode, unsigned alignment, unsigned space_before, unsigned space_after) { be_abi_call_arg_t *arg = get_or_set_call_arg(call, 0, arg_pos, 1); arg->on_stack = 1; + arg->load_mode = load_mode; arg->alignment = alignment; arg->space_before = space_before; arg->space_after = space_after; @@ -226,6 +232,7 @@ be_abi_call_flags_t be_abi_call_get_flags(const be_abi_call_t *call) static be_abi_call_t *be_abi_call_new(const arch_register_class_t *cls_addr) { be_abi_call_t *call = xmalloc(sizeof(call[0])); + memset(call, 0, sizeof(call[0])); call->flags.val = 0; call->params = new_set(cmp_call_arg, 16); @@ -263,7 +270,8 @@ static void be_abi_call_free(be_abi_call_t *call) and the spills. */ -static int get_stack_entity_offset(be_stack_layout_t *frame, ir_entity *ent, int bias) +static int get_stack_entity_offset(be_stack_layout_t *frame, ir_entity *ent, + int bias) { ir_type *t = get_entity_owner(ent); int ofs = get_entity_offset(ent); @@ -398,14 +406,14 @@ static INLINE int is_on_stack(be_abi_call_t *call, int pos) * @param curr_sp The stack pointer node to use. * @return The stack pointer after the call. */ -static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp, ir_node *alloca_copy) +static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp) { ir_graph *irg = env->birg->irg; - const arch_env_t *arch_env = env->birg->main_env->arch_env; + const arch_env_t *arch_env = &env->birg->main_env->arch_env; const arch_isa_t *isa = arch_env->isa; - ir_type *mt = get_Call_type(irn); + ir_type *call_tp = get_Call_type(irn); ir_node *call_ptr = get_Call_ptr(irn); - int n_params = get_method_n_params(mt); + int n_params = get_method_n_params(call_tp); ir_node *curr_mem = get_Call_mem(irn); ir_node *bl = get_nodes_block(irn); pset *results = pset_new_ptr(8); @@ -418,9 +426,10 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp, i ir_mode *mach_mode = sp->reg_class->mode; struct obstack *obst = &env->obst; int no_alloc = call->flags.bits.frame_is_setup_on_call; + int n_res = get_method_n_ress(call_tp); + int do_seq = call->flags.bits.store_args_sequential && !no_alloc; ir_node *res_proj = NULL; - int curr_res_proj = pn_Call_max; int n_reg_params = 0; int n_stack_params = 0; int n_ins; @@ -428,6 +437,7 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp, i ir_node *low_call; ir_node **in; ir_node **res_projs; + int n_reg_results = 0; const arch_register_t *reg; const ir_edge_t *edge; int *reg_param_idxs; @@ -435,7 +445,7 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp, i int i, n; /* Let the isa fill out the abi description for that call node. */ - arch_isa_get_call_abi(isa, mt, call); + arch_isa_get_call_abi(isa, call_tp, call); /* Insert code to put the stack arguments on the stack. */ assert(get_Call_n_params(irn) == n_params); @@ -443,7 +453,7 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp, i be_abi_call_arg_t *arg = get_call_arg(call, 0, i); assert(arg); if (arg->on_stack) { - int arg_size = get_type_size_bytes(get_method_param_type(mt, i)); + int arg_size = get_type_size_bytes(get_method_param_type(call_tp, i)); stack_size += round_up2(arg->space_before, arg->alignment); stack_size += round_up2(arg_size, arg->alignment); @@ -464,10 +474,22 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp, i } reg_param_idxs = obstack_finish(obst); + /* + * If the stack is decreasing and we do not want to store sequentially, + * or someone else allocated the call frame + * we allocate as much space on the stack all parameters need, by + * moving the stack pointer along the stack's direction. + * + * Note: we also have to do this for stack_size == 0, because we may have + * to adjust stack alignment for the call. + */ + if (stack_dir < 0 && !do_seq && !no_alloc) { + curr_sp = be_new_IncSP(sp, irg, bl, curr_sp, stack_size, 1); + } + /* If there are some parameters which shall be passed on the stack. */ if (n_stack_params > 0) { int curr_ofs = 0; - int do_seq = call->flags.bits.store_args_sequential && !no_alloc; /* * Reverse list of stack parameters if call arguments are from left to right. @@ -483,25 +505,9 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp, i } } - /* - * If the stack is decreasing and we do not want to store sequentially, - * or someone else allocated the call frame - * we allocate as much space on the stack all parameters need, by - * moving the stack pointer along the stack's direction. - */ - if (stack_dir < 0 && !do_seq && !no_alloc) { - curr_sp = be_new_IncSP(sp, irg, bl, curr_sp, stack_size); - if (alloca_copy) { - add_irn_dep(curr_sp, alloca_copy); - alloca_copy = NULL; - } - } - + curr_mem = get_Call_mem(irn); if (! do_seq) { - obstack_ptr_grow(obst, get_Call_mem(irn)); - curr_mem = new_NoMem(); - } else { - curr_mem = get_Call_mem(irn); + obstack_ptr_grow(obst, curr_mem); } for (i = 0; i < n_stack_params; ++i) { @@ -510,7 +516,7 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp, i ir_node *param = get_Call_param(irn, p); ir_node *addr = curr_sp; ir_node *mem = NULL; - ir_type *param_type = get_method_param_type(mt, p); + ir_type *param_type = get_method_param_type(call_tp, p); int param_size = get_type_size_bytes(param_type) + arg->space_after; /* @@ -520,11 +526,7 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp, i */ if (do_seq) { curr_ofs = 0; - addr = curr_sp = be_new_IncSP(sp, irg, bl, curr_sp, param_size + arg->space_before); - if (alloca_copy) { - add_irn_dep(curr_sp, alloca_copy); - alloca_copy = NULL; - } + addr = curr_sp = be_new_IncSP(sp, irg, bl, curr_sp, param_size + arg->space_before, 0); add_irn_dep(curr_sp, curr_mem); } else { @@ -545,7 +547,8 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp, i /* Insert a store for primitive arguments. */ if (is_atomic_type(param_type)) { ir_node *store; - store = new_r_Store(irg, bl, curr_mem, addr, param); + ir_node *mem_input = do_seq ? curr_mem : new_NoMem(); + store = new_r_Store(irg, bl, mem_input, addr, param); mem = new_r_Proj(irg, bl, store, mode_M, pn_Store_M); } @@ -581,7 +584,7 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp, i /* Collect caller save registers */ for (i = 0, n = arch_isa_get_n_reg_class(isa); i < n; ++i) { - int j; + unsigned j; const arch_register_class_t *cls = arch_isa_get_reg_class(isa, i); for (j = 0; j < cls->n_regs; ++j) { const arch_register_t *reg = arch_register_for_index(cls, j); @@ -597,52 +600,35 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp, i /* search the greatest result proj number */ - /* TODO: what if the result is NOT used? Currently there is - * no way to detect this later, especially there is no way to - * see this in the proj numbers. - * While this is ok for the register allocator, it is bad for - * backends which need to change the be_Call further (x87 simulator - * for instance. However for this particular case the call_type is - * sufficient.). - */ + res_projs = alloca(n_res * sizeof(res_projs[0])); + memset(res_projs, 0, n_res * sizeof(res_projs[0])); + foreach_out_edge(irn, edge) { const ir_edge_t *res_edge; - ir_node *irn = get_edge_src_irn(edge); + ir_node *irn = get_edge_src_irn(edge); - if (is_Proj(irn) && get_Proj_proj(irn) == pn_Call_T_result) { - res_proj = irn; - foreach_out_edge(irn, res_edge) { - int proj; - be_abi_call_arg_t *arg; - ir_node *res = get_edge_src_irn(res_edge); - - assert(is_Proj(res)); - - proj = get_Proj_proj(res); - arg = get_call_arg(call, 1, proj); - - /* - shift the proj number to the right, since we will drop the - unspeakable Proj_T from the Call. Therefore, all real argument - Proj numbers must be increased by pn_be_Call_first_res - */ - proj += pn_be_Call_first_res; - set_Proj_proj(res, proj); - obstack_ptr_grow(obst, res); - - if (proj > curr_res_proj) - curr_res_proj = proj; - if (arg->in_reg) { - pset_remove_ptr(caller_save, arg->reg); - //pmap_insert(arg_regs, arg->reg, INT_TO_PTR(proj + 1)) - } - } + if(!is_Proj(irn) || get_Proj_proj(irn) != pn_Call_T_result) + continue; + + foreach_out_edge(irn, res_edge) { + int proj; + ir_node *res = get_edge_src_irn(res_edge); + + assert(is_Proj(res)); + + proj = get_Proj_proj(res); + assert(proj < n_res); + assert(res_projs[proj] == NULL); + res_projs[proj] = res; } + res_proj = irn; + break; } - curr_res_proj++; - obstack_ptr_grow(obst, NULL); - res_projs = obstack_finish(obst); + /** TODO: this is not correct for cases where return values are passed + * on the stack, but no known ABI does this currentl... + */ + n_reg_results = n_res; /* make the back end call node and set its register requirements. */ for (i = 0; i < n_reg_params; ++i) { @@ -665,18 +651,57 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp, i /* direct call */ low_call = be_new_Call(get_irn_dbg_info(irn), irg, bl, curr_mem, curr_sp, curr_sp, - curr_res_proj + pset_count(caller_save), n_ins, - in, get_Call_type(irn)); + n_reg_results + pn_be_Call_first_res + pset_count(caller_save), + n_ins, in, get_Call_type(irn)); be_Call_set_entity(low_call, get_SymConst_entity(call_ptr)); } else { /* indirect call */ low_call = be_new_Call(get_irn_dbg_info(irn), irg, bl, curr_mem, curr_sp, call_ptr, - curr_res_proj + pset_count(caller_save), + n_reg_results + pn_be_Call_first_res + pset_count(caller_save), n_ins, in, get_Call_type(irn)); } + be_Call_set_pop(low_call, call->pop); ARR_APP1(ir_node *, env->calls, low_call); + /* create new stack pointer */ + curr_sp = new_r_Proj(irg, bl, low_call, get_irn_mode(curr_sp), + pn_be_Call_sp); + be_set_constr_single_reg(low_call, BE_OUT_POS(pn_be_Call_sp), sp); + arch_set_irn_register(arch_env, curr_sp, sp); + be_node_set_flags(low_call, BE_OUT_POS(pn_be_Call_sp), + arch_irn_flags_ignore | arch_irn_flags_modify_sp); + + for(i = 0; i < n_res; ++i) { + int pn; + ir_node *proj = res_projs[i]; + be_abi_call_arg_t *arg = get_call_arg(call, 1, i); + + /* returns values on stack not supported yet */ + assert(arg->in_reg); + + /* + shift the proj number to the right, since we will drop the + unspeakable Proj_T from the Call. Therefore, all real argument + Proj numbers must be increased by pn_be_Call_first_res + */ + pn = i + pn_be_Call_first_res; + + if(proj == NULL) { + ir_type *res_type = get_method_res_type(call_tp, i); + ir_mode *mode = get_type_mode(res_type); + proj = new_r_Proj(irg, bl, low_call, mode, pn); + res_projs[i] = proj; + } else { + set_Proj_pred(proj, low_call); + set_Proj_proj(proj, pn); + } + + if (arg->in_reg) { + pset_remove_ptr(caller_save, arg->reg); + } + } + /* Set the register class of the call address to the backend provided class (default: stack pointer class) @@ -695,62 +720,65 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp, i } /* Set the register constraints of the results. */ - for (i = 0; res_projs[i]; ++i) { - int pn = get_Proj_proj(res_projs[i]); - - /* Correct Proj number since it has been adjusted! (see above) */ - const be_abi_call_arg_t *arg = get_call_arg(call, 1, pn - pn_Call_max); - - /* Matze: we need the information about the real mode for later - * transforms (signed/unsigend compares, stores...), so leave the fixup - * for the backend transform phase... */ -#if 0 - /* correct mode */ - const arch_register_class_t *cls = arch_register_get_class(arg->reg); - ir_mode *mode = arch_register_class_mode(cls); - set_irn_mode(irn, mode); -#endif + for (i = 0; i < n_res; ++i) { + ir_node *proj = res_projs[i]; + const be_abi_call_arg_t *arg = get_call_arg(call, 1, i); + int pn = get_Proj_proj(proj); assert(arg->in_reg); be_set_constr_single_reg(low_call, BE_OUT_POS(pn), arg->reg); - arch_set_irn_register(arch_env, res_projs[i], arg->reg); + arch_set_irn_register(arch_env, proj, arg->reg); } obstack_free(obst, in); exchange(irn, low_call); - /* redirect the result projs to the lowered call instead of the Proj_T */ - for (i = 0; res_projs[i]; ++i) - set_Proj_pred(res_projs[i], low_call); - - /* set the now unnecessary projT to bad */ + /* kill the ProjT node */ if (res_proj != NULL) { be_kill_node(res_proj); } /* Make additional projs for the caller save registers and the Keep node which keeps them alive. */ - if (pset_count(caller_save) > 0) { + if (1 || pset_count(caller_save) + n_reg_results > 0) { const arch_register_t *reg; ir_node **in, *keep; - int i, n; + int i; + int n = 0; + int curr_res_proj + = pn_be_Call_first_res + n_reg_results; - for (reg = pset_first(caller_save), n = 0; reg; reg = pset_next(caller_save), ++n) { - ir_node *proj = new_r_Proj(irg, bl, low_call, reg->reg_class->mode, curr_res_proj); + /* also keep the stack pointer */ + ++n; + set_irn_link(curr_sp, (void*) sp); + obstack_ptr_grow(obst, curr_sp); + + for (reg = pset_first(caller_save); reg; reg = pset_next(caller_save), ++n) { + ir_node *proj = new_r_Proj(irg, bl, low_call, reg->reg_class->mode, + curr_res_proj); /* memorize the register in the link field. we need afterwards to set the register class of the keep correctly. */ be_set_constr_single_reg(low_call, BE_OUT_POS(curr_res_proj), reg); + arch_set_irn_register(arch_env, proj, reg); /* a call can produce ignore registers, in this case set the flag and register for the Proj */ if (arch_register_type_is(reg, ignore)) { - arch_set_irn_register(env->birg->main_env->arch_env, proj, reg); - be_node_set_flags(low_call, BE_OUT_POS(curr_res_proj), arch_irn_flags_ignore); + be_node_set_flags(low_call, BE_OUT_POS(curr_res_proj), + arch_irn_flags_ignore); } - set_irn_link(proj, (void *) reg); + set_irn_link(proj, (void*) reg); obstack_ptr_grow(obst, proj); curr_res_proj++; } + for(i = 0; i < n_reg_results; ++i) { + ir_node *proj = res_projs[i]; + const arch_register_t *reg = arch_get_irn_register(arch_env, proj); + set_irn_link(proj, (void*) reg); + obstack_ptr_grow(obst, proj); + } + n += n_reg_results; + /* create the Keep for the caller save registers */ in = (ir_node **) obstack_finish(obst); keep = be_new_Keep(NULL, irg, bl, n, in); @@ -762,6 +790,9 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp, i } /* Clean up the stack. */ + assert(stack_size >= call->pop); + stack_size -= call->pop; + if (stack_size > 0) { ir_node *mem_proj = NULL; @@ -774,19 +805,13 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp, i } if (! mem_proj) { - mem_proj = new_r_Proj(irg, bl, low_call, mode_M, pn_Call_M); + mem_proj = new_r_Proj(irg, bl, low_call, mode_M, pn_be_Call_M_regular); keep_alive(mem_proj); } - - /* Clean up the stack frame if we allocated it */ - if (! no_alloc) { - curr_sp = be_new_IncSP(sp, irg, bl, curr_sp, -stack_size); - add_irn_dep(curr_sp, mem_proj); - if (alloca_copy) { - add_irn_dep(curr_sp, alloca_copy); - alloca_copy = NULL; - } - } + } + /* Clean up the stack frame or revert alignment fixes if we allocated it */ + if (! no_alloc) { + curr_sp = be_new_IncSP(sp, irg, bl, curr_sp, -stack_size, 0); } be_abi_call_free(call); @@ -813,10 +838,15 @@ static ir_node *adjust_alloc_size(unsigned stack_alignment, ir_node *size, ir_graph *irg, ir_node *block, dbg_info *dbg) { if (stack_alignment > 1) { - ir_mode *mode = get_irn_mode(size); - tarval *tv = new_tarval_from_long(stack_alignment-1, mode); - ir_node *mask = new_r_Const(irg, block, mode, tv); + ir_mode *mode; + tarval *tv; + ir_node *mask; + + assert(is_po2(stack_alignment)); + mode = get_irn_mode(size); + tv = new_tarval_from_long(stack_alignment-1, mode); + mask = new_r_Const(irg, block, mode, tv); size = new_rd_Add(dbg, irg, block, size, mask, mode); tv = new_tarval_from_long(-(long)stack_alignment, mode); @@ -829,7 +859,7 @@ static ir_node *adjust_alloc_size(unsigned stack_alignment, ir_node *size, * Adjust an alloca. * The alloca is transformed into a back end alloca node and connected to the stack nodes. */ -static ir_node *adjust_alloc(be_abi_irg_t *env, ir_node *alloc, ir_node *curr_sp, ir_node **result_copy) +static ir_node *adjust_alloc(be_abi_irg_t *env, ir_node *alloc, ir_node *curr_sp) { ir_node *block; ir_graph *irg; @@ -839,7 +869,7 @@ static ir_node *adjust_alloc(be_abi_irg_t *env, ir_node *alloc, ir_node *curr_sp dbg_info *dbg; const ir_edge_t *edge; - ir_node *new_alloc, *size, *addr, *copy, *ins[2]; + ir_node *new_alloc, *size, *addr, *ins[2]; unsigned stack_alignment; if (get_Alloc_where(alloc) != stack_alloc) { @@ -881,10 +911,11 @@ static ir_node *adjust_alloc(be_abi_irg_t *env, ir_node *alloc, ir_node *curr_sp /* we might need to multiply the size with the element size */ if(type != get_unknown_type() && get_type_size_bytes(type) != 1) { - tarval *tv = new_tarval_from_long(get_type_size_bytes(type), mode_Iu); + tarval *tv = new_tarval_from_long(get_type_size_bytes(type), + mode_Iu); ir_node *cnst = new_rd_Const(dbg, irg, block, mode_Iu, tv); - ir_node *mul = new_rd_Mul(dbg, irg, block, get_Alloc_size(alloc), - cnst, mode_Iu); + ir_node *mul = new_rd_Mul(dbg, irg, block, get_Alloc_size(alloc), + cnst, mode_Iu); size = mul; } else { size = get_Alloc_size(alloc); @@ -894,9 +925,7 @@ static ir_node *adjust_alloc(be_abi_irg_t *env, ir_node *alloc, ir_node *curr_sp We cannot omit it. */ env->call->flags.bits.try_omit_fp = 0; - /* FIXME: size must be here round up for the stack alignment, but - this must be transmitted from the backend. */ - stack_alignment = 4; + stack_alignment = env->isa->stack_alignment; size = adjust_alloc_size(stack_alignment, size, irg, block, dbg); new_alloc = be_new_AddSP(env->isa->sp, irg, block, curr_sp, size); set_irn_dbg_info(new_alloc, dbg); @@ -921,20 +950,9 @@ static ir_node *adjust_alloc(be_abi_irg_t *env, ir_node *alloc, ir_node *curr_sp /* fix projnum of alloca res */ set_Proj_proj(alloc_res, pn_be_AddSP_res); - addr = env->isa->stack_dir < 0 ? alloc_res : curr_sp; - - /* copy the address away, since it could be used after further stack pointer modifications. */ - /* Let it point curr_sp just for the moment, I'll reroute it in a second. */ - *result_copy = copy = be_new_Copy(env->isa->sp->reg_class, irg, block, curr_sp); - set_irn_mode(copy, mode_P); - - /* Let all users of the Alloc() result now point to the copy. */ - edges_reroute(alloc_res, copy, irg); - - /* Rewire the copy appropriately. */ - set_irn_n(copy, be_pos_Copy_op, addr); - - curr_sp = alloc_res; + addr = alloc_res; + curr_sp = new_r_Proj(irg, block, new_alloc, get_irn_mode(curr_sp), + pn_be_AddSP_sp); return curr_sp; } /* adjust_alloc */ @@ -976,10 +994,8 @@ static ir_node *adjust_free(be_abi_irg_t *env, ir_node *free, ir_node *curr_sp) size = get_Free_size(free); } - /* FIXME: size must be here round up for the stack alignment, but - this must be transmitted from the backend. */ - stack_alignment = 4; - size = adjust_alloc_size(stack_alignment, size, irg, block, dbg); + stack_alignment = env->isa->stack_alignment; + size = adjust_alloc_size(stack_alignment, size, irg, block, dbg); /* The stack pointer will be modified in an unknown manner. We cannot omit it. */ @@ -988,7 +1004,7 @@ static ir_node *adjust_free(be_abi_irg_t *env, ir_node *free, ir_node *curr_sp) set_irn_dbg_info(subsp, dbg); mem = new_r_Proj(irg, block, subsp, mode_M, pn_be_SubSP_M); - res = new_r_Proj(irg, block, subsp, sp_mode, pn_be_SubSP_res); + res = new_r_Proj(irg, block, subsp, sp_mode, pn_be_SubSP_sp); /* we need to sync the memory */ in[0] = get_Free_mem(free); @@ -1112,7 +1128,6 @@ static void process_calls_in_block(ir_node *bl, void *data) if(n > 0) { ir_node *keep; ir_node **nodes; - ir_node *copy = NULL; int i; nodes = obstack_finish(&env->obst); @@ -1126,10 +1141,10 @@ static void process_calls_in_block(ir_node *bl, void *data) DBG((env->dbg, LEVEL_3, "\tprocessing call %+F\n", irn)); switch(get_irn_opcode(irn)) { case iro_Call: - curr_sp = adjust_call(env, irn, curr_sp, copy); + curr_sp = adjust_call(env, irn, curr_sp); break; case iro_Alloc: - curr_sp = adjust_alloc(env, irn, curr_sp, ©); + curr_sp = adjust_alloc(env, irn, curr_sp); break; case iro_Free: curr_sp = adjust_free(env, irn, curr_sp); @@ -1142,8 +1157,10 @@ static void process_calls_in_block(ir_node *bl, void *data) obstack_free(&env->obst, nodes); - /* Keep the last stack state in the block by tying it to Keep node */ - if(curr_sp != env->init_sp) { + /* Keep the last stack state in the block by tying it to Keep node, + * the proj from calls is already kept */ + if(curr_sp != env->init_sp + && !(is_Proj(curr_sp) && be_is_Call(get_Proj_pred(curr_sp)))) { nodes[0] = curr_sp; keep = be_new_Keep(env->isa->sp->reg_class, get_irn_irg(bl), bl, 1, nodes); @@ -1169,86 +1186,6 @@ static void process_calls(be_abi_irg_t *env) heights_free(ir_heights); } -#if 0 /* -static ir_node *setup_frame(be_abi_irg_t *env) -{ - const arch_isa_t *isa = env->birg->main_env->arch_env->isa; - const arch_register_t *sp = isa->sp; - const arch_register_t *bp = isa->bp; - be_abi_call_flags_bits_t flags = env->call->flags.bits; - ir_graph *irg = env->birg->irg; - ir_node *bl = get_irg_start_block(irg); - ir_node *no_mem = get_irg_no_mem(irg); - ir_node *old_frame = get_irg_frame(irg); - ir_node *stack = pmap_get(env->regs, (void *) sp); - ir_node *frame = pmap_get(env->regs, (void *) bp); - - int stack_nr = get_Proj_proj(stack); - - if(flags.try_omit_fp) { - stack = be_new_IncSP(sp, irg, bl, stack, no_mem, BE_STACK_FRAME_SIZE_EXPAND); - frame = stack; - } - - else { - frame = be_new_Copy(bp->reg_class, irg, bl, stack); - - be_node_set_flags(frame, -1, arch_irn_flags_dont_spill); - if(!flags.fp_free) { - be_set_constr_single_reg(frame, -1, bp); - be_node_set_flags(frame, -1, arch_irn_flags_ignore); - arch_set_irn_register(env->birg->main_env->arch_env, frame, bp); - } - - stack = be_new_IncSP(sp, irg, bl, stack, frame, BE_STACK_FRAME_SIZE_EXPAND); - } - - be_node_set_flags(env->reg_params, -(stack_nr + 1), arch_irn_flags_ignore); - env->init_sp = stack; - set_irg_frame(irg, frame); - edges_reroute(old_frame, frame, irg); - - return frame; -} - -static void clearup_frame(be_abi_irg_t *env, ir_node *ret, pmap *reg_map, struct obstack *obst) -{ - const arch_isa_t *isa = env->birg->main_env->arch_env->isa; - const arch_register_t *sp = isa->sp; - const arch_register_t *bp = isa->bp; - ir_graph *irg = env->birg->irg; - ir_node *ret_mem = get_Return_mem(ret); - ir_node *frame = get_irg_frame(irg); - ir_node *bl = get_nodes_block(ret); - ir_node *stack = get_irn_link(bl); - - pmap_entry *ent; - - if(env->call->flags.bits.try_omit_fp) { - stack = be_new_IncSP(sp, irg, bl, stack, ret_mem, -BE_STACK_FRAME_SIZE_SHRINK); - } - - else { - stack = be_new_SetSP(sp, irg, bl, stack, frame, ret_mem); - be_set_constr_single_reg(stack, -1, sp); - be_node_set_flags(stack, -1, arch_irn_flags_ignore); - } - - pmap_foreach(env->regs, ent) { - const arch_register_t *reg = ent->key; - ir_node *irn = ent->value; - - if(reg == sp) - obstack_ptr_grow(&env->obst, stack); - else if(reg == bp) - obstack_ptr_grow(&env->obst, frame); - else if(arch_register_type_is(reg, callee_save) || arch_register_type_is(reg, ignore)) - obstack_ptr_grow(obst, irn); - } -} -*/ -#endif - /** * Computes the stack argument layout type. * Changes a possibly allocated value param type by moving @@ -1264,7 +1201,7 @@ static void clearup_frame(be_abi_irg_t *env, ir_node *ret, pmap *reg_map, struct static ir_type *compute_arg_type(be_abi_irg_t *env, be_abi_call_t *call, ir_type *method_type, ir_entity ***param_map) { int dir = env->call->flags.bits.left_to_right ? 1 : -1; - int inc = env->birg->main_env->arch_env->isa->stack_dir * dir; + int inc = env->birg->main_env->arch_env.isa->stack_dir * dir; int n = get_method_n_params(method_type); int curr = inc > 0 ? 0 : n - 1; int ofs = 0; @@ -1377,7 +1314,7 @@ static reg_node_map_t *reg_map_to_arr(struct obstack *obst, pmap *reg_map) int i = 0; reg_node_map_t *res = obstack_alloc(obst, n * sizeof(res[0])); - pmap_foreach(reg_map, ent) { + foreach_pmap(reg_map, ent) { res[i].reg = ent->key; res[i].irn = ent->value; i++; @@ -1425,13 +1362,13 @@ static ir_node *create_barrier(be_abi_irg_t *env, ir_node *bl, ir_node **mem, pm be_set_constr_single_reg(irn, n, reg); be_set_constr_single_reg(irn, pos, reg); be_node_set_reg_class(irn, pos, reg->reg_class); - arch_set_irn_register(env->birg->main_env->arch_env, proj, reg); + arch_set_irn_register(&env->birg->main_env->arch_env, proj, reg); /* if the proj projects a ignore register or a node which is set to ignore, propagate this property. */ - if(arch_register_type_is(reg, ignore) || arch_irn_is(env->birg->main_env->arch_env, in[n], ignore)) + if(arch_register_type_is(reg, ignore) || arch_irn_is(&env->birg->main_env->arch_env, in[n], ignore)) flags |= arch_irn_flags_ignore; - if(arch_irn_is(env->birg->main_env->arch_env, in[n], modify_sp)) + if(arch_irn_is(&env->birg->main_env->arch_env, in[n], modify_sp)) flags |= arch_irn_flags_modify_sp; be_node_set_flags(irn, pos, flags); @@ -1456,15 +1393,18 @@ static ir_node *create_barrier(be_abi_irg_t *env, ir_node *bl, ir_node **mem, pm * @param mem the current memory * @param n_res number of return results */ -static ir_node *create_be_return(be_abi_irg_t *env, ir_node *irn, ir_node *bl, ir_node *mem, int n_res) { - be_abi_call_t *call = env->call; - const arch_isa_t *isa = env->birg->main_env->arch_env->isa; - +static ir_node *create_be_return(be_abi_irg_t *env, ir_node *irn, ir_node *bl, + ir_node *mem, int n_res) +{ + be_abi_call_t *call = env->call; + const arch_isa_t *isa = env->birg->main_env->arch_env.isa; + dbg_info *dbgi; pmap *reg_map = pmap_create(); ir_node *keep = pmap_get(env->keep_map, bl); int in_max; ir_node *ret; int i, n; + unsigned pop; ir_node **in; ir_node *stack; const arch_register_t **regs; @@ -1493,7 +1433,7 @@ static ir_node *create_be_return(be_abi_irg_t *env, ir_node *irn, ir_node *bl, i } /* Add uses of the callee save registers. */ - pmap_foreach(env->regs, ent) { + foreach_pmap(env->regs, ent) { const arch_register_t *reg = ent->key; if(arch_register_type_is(reg, callee_save) || arch_register_type_is(reg, ignore)) pmap_insert(reg_map, ent->key, ent->value); @@ -1533,7 +1473,7 @@ static ir_node *create_be_return(be_abi_irg_t *env, ir_node *irn, ir_node *bl, i } /* grow the rest of the stuff. */ - pmap_foreach(reg_map, ent) { + foreach_pmap(reg_map, ent) { if(ent->value) { in[n] = ent->value; regs[n++] = ent->key; @@ -1541,7 +1481,14 @@ static ir_node *create_be_return(be_abi_irg_t *env, ir_node *irn, ir_node *bl, i } /* The in array for the new back end return is now ready. */ - ret = be_new_Return(irn ? get_irn_dbg_info(irn) : NULL, env->birg->irg, bl, n_res, n, in); + if(irn != NULL) { + dbgi = get_irn_dbg_info(irn); + } else { + dbgi = NULL; + } + /* we have to pop the shadow parameter in in case of struct returns */ + pop = call->pop; + ret = be_new_Return(dbgi, env->birg->irg, bl, n_res, pop, n, in); /* Set the register classes of the return's parameter accordingly. */ for(i = 0; i < n; ++i) @@ -1633,7 +1580,7 @@ static void fix_address_of_parameter_access(be_abi_irg_t *env, ir_entity *value_ ir_node *frame, *imem, *nmem, *store, *mem, *args, *args_bl; const ir_edge_t *edge; optimization_state_t state; - int offset; + unsigned offset; foreach_block_succ(start_bl, edge) { ir_node *succ = get_edge_src_irn(edge); @@ -1686,12 +1633,16 @@ static void fix_address_of_parameter_access(be_abi_irg_t *env, ir_entity *value_ /* move all entities to the frame type */ frame_tp = get_irg_frame_type(irg); offset = get_type_size_bytes(frame_tp); + + /* we will add new entities: set the layout to undefined */ + assert(get_type_state(frame_tp) == layout_fixed); + set_type_state(frame_tp, layout_undefined); for (ent = new_list; ent; ent = get_entity_link(ent)) { - ir_type *tp = get_entity_type(ent); - int align = get_type_alignment_bytes(tp); + ir_type *tp = get_entity_type(ent); + unsigned align = get_type_alignment_bytes(tp); offset += align - 1; - offset &= -align; + offset &= ~(align - 1); set_entity_owner(ent, frame_tp); add_class_member(frame_tp, ent); /* must be automatic to set a fixed layout */ @@ -1700,9 +1651,12 @@ static void fix_address_of_parameter_access(be_abi_irg_t *env, ir_entity *value_ offset += get_type_size_bytes(tp); } set_type_size_bytes(frame_tp, offset); + /* fix the layout again */ + set_type_state(frame_tp, layout_fixed); } } +#if 1 /** * The start block has no jump, instead it has an initial exec Proj. * The backend wants to handle all blocks the same way, so we replace @@ -1733,6 +1687,7 @@ static void fix_start_block(ir_node *block, void *env) { } } } +#endif /** * Modify the irg itself and the frame type. @@ -1740,7 +1695,7 @@ static void fix_start_block(ir_node *block, void *env) { static void modify_irg(be_abi_irg_t *env) { be_abi_call_t *call = env->call; - const arch_isa_t *isa = env->birg->main_env->arch_env->isa; + const arch_isa_t *isa = env->birg->main_env->arch_env.isa; const arch_register_t *sp = arch_isa_sp(isa); ir_graph *irg = env->birg->irg; ir_node *bl = get_irg_start_block(irg); @@ -1752,12 +1707,12 @@ static void modify_irg(be_abi_irg_t *env) pset *dont_save = pset_new_ptr(8); int n_params; - int i, j, n, temp; + int i, n; + unsigned j; reg_node_map_t *rm; const arch_register_t *fp_reg; ir_node *frame_pointer; - ir_node *barrier; ir_node *reg_params_bl; ir_node **args; ir_node *arg_tuple; @@ -1881,7 +1836,7 @@ static void modify_irg(be_abi_irg_t *env) proj = new_r_Proj(irg, reg_params_bl, env->reg_params, mode, nr); pmap_insert(env->regs, (void *) reg, proj); be_set_constr_single_reg(env->reg_params, pos, reg); - arch_set_irn_register(env->birg->main_env->arch_env, proj, reg); + arch_set_irn_register(&env->birg->main_env->arch_env, proj, reg); /* * If the register is an ignore register, @@ -1912,13 +1867,13 @@ static void modify_irg(be_abi_irg_t *env) /* do the stack allocation BEFORE the barrier, or spill code might be added before it */ env->init_sp = be_abi_reg_map_get(env->regs, sp); - env->init_sp = be_new_IncSP(sp, irg, bl, env->init_sp, BE_STACK_FRAME_SIZE_EXPAND); + env->init_sp = be_new_IncSP(sp, irg, bl, env->init_sp, BE_STACK_FRAME_SIZE_EXPAND, 0); be_abi_reg_map_set(env->regs, sp, env->init_sp); - env->start_barrier = barrier = create_barrier(env, bl, &mem, env->regs, 0); + create_barrier(env, bl, &mem, env->regs, 0); env->init_sp = be_abi_reg_map_get(env->regs, sp); - arch_set_irn_register(env->birg->main_env->arch_env, env->init_sp, sp); + arch_set_irn_register(&env->birg->main_env->arch_env, env->init_sp, sp); frame_pointer = be_abi_reg_map_get(env->regs, fp_reg); set_irg_frame(irg, frame_pointer); @@ -1947,21 +1902,26 @@ static void modify_irg(be_abi_irg_t *env) if (arg->in_reg) { repl = pmap_get(env->regs, (void *) arg->reg); - } + } else if(arg->on_stack) { + ir_node *addr = be_new_FrameAddr(sp->reg_class, irg, reg_params_bl, frame_pointer, arg->stack_ent); - else if(arg->on_stack) { - /* For atomic parameters which are actually used, we create a StackParam node. */ + /* For atomic parameters which are actually used, we create a Load node. */ if(is_atomic_type(param_type) && get_irn_n_edges(args[i]) > 0) { - ir_mode *mode = get_type_mode(param_type); - const arch_register_class_t *cls = arch_isa_get_reg_class_for_mode(isa, mode); - repl = be_new_StackParam(cls, isa->bp->reg_class, irg, reg_params_bl, mode, frame_pointer, arg->stack_ent); - } + ir_mode *mode = get_type_mode(param_type); + ir_mode *load_mode = arg->load_mode; - /* The stack parameter is not primitive (it is a struct or array), - we thus will create a node representing the parameter's address - on the stack. */ - else { - repl = be_new_FrameAddr(sp->reg_class, irg, reg_params_bl, frame_pointer, arg->stack_ent); + ir_node *load = new_r_Load(irg, reg_params_bl, new_NoMem(), addr, load_mode); + set_irn_pinned(load, op_pin_state_floats); + repl = new_r_Proj(irg, reg_params_bl, load, load_mode, pn_Load_res); + + if (mode != load_mode) { + repl = new_r_Conv(irg, reg_params_bl, repl, mode); + } + } else { + /* The stack parameter is not primitive (it is a struct or array), + * we thus will create a node representing the parameter's address + * on the stack. */ + repl = addr; } } @@ -1987,7 +1947,9 @@ static void modify_irg(be_abi_irg_t *env) ir_node *irn = get_Block_cfgpred(end, i); if (is_Return(irn)) { - ir_node *ret = create_be_return(env, irn, get_nodes_block(irn), get_Return_mem(irn), get_Return_n_ress(irn)); + ir_node *blk = get_nodes_block(irn); + ir_node *mem = get_Return_mem(irn); + ir_node *ret = create_be_return(env, irn, blk, mem, get_Return_n_ress(irn)); exchange(irn, ret); } } @@ -1998,8 +1960,8 @@ static void modify_irg(be_abi_irg_t *env) obstack_free(&env->obst, args); /* handle start block here (place a jump in the block) */ - temp = 0; - irg_block_walk_graph(irg, fix_start_block, NULL, &temp); + i = 0; + irg_block_walk_graph(irg, fix_start_block, NULL, &i); } /** Fix the state inputs of calls that still hang on unknowns */ @@ -2013,7 +1975,7 @@ void fix_call_state_inputs(be_abi_irg_t *env) /* Collect caller save registers */ n = arch_isa_get_n_reg_class(isa); for(i = 0; i < n; ++i) { - int j; + unsigned j; const arch_register_class_t *cls = arch_isa_get_reg_class(isa, i); for(j = 0; j < cls->n_regs; ++j) { const arch_register_t *reg = arch_register_for_index(cls, j); @@ -2031,7 +1993,7 @@ void fix_call_state_inputs(be_abi_irg_t *env) arity = get_irn_arity(call); - /* the statereg inputs are the last n inputs of the calls */ + /* the state reg inputs are the last n inputs of the calls */ for(s = 0; s < n_states; ++s) { int inp = arity - n_states + s; const arch_register_t *reg = stateregs[s]; @@ -2042,6 +2004,115 @@ void fix_call_state_inputs(be_abi_irg_t *env) } } +/** + * Create a trampoline entity for the given method. + */ +static ir_entity *create_trampoline(be_main_env_t *be, ir_entity *method) +{ + ir_type *type = get_entity_type(method); + ident *old_id = get_entity_ld_ident(method); + ident *id = mangle3("L", old_id, "$stub"); + ir_type *parent = be->pic_trampolines_type; + ir_entity *ent = new_entity(parent, old_id, type); + set_entity_ld_ident(ent, id); + set_entity_visibility(ent, visibility_local); + set_entity_variability(ent, variability_uninitialized); + + return ent; +} + +/** + * Returns the trampoline entity for the given method. + */ +static ir_entity *get_trampoline(be_main_env_t *env, ir_entity *method) +{ + ir_entity *result = pmap_get(env->ent_trampoline_map, method); + if (result == NULL) { + result = create_trampoline(env, method); + pmap_insert(env->ent_trampoline_map, method, result); + } + + return result; +} + +/** + * Returns non-zero if a given entity can be accessed using a relative address. + */ +static int can_address_relative(ir_entity *entity) +{ + return get_entity_variability(entity) == variability_initialized + || get_entity_visibility(entity) == visibility_local; +} + +/** patches SymConsts to work in position independent code */ +static void fix_pic_symconsts(ir_node *node, void *data) +{ + ir_graph *irg; + ir_node *pic_base; + ir_node *add; + ir_node *block; + ir_node *unknown; + ir_mode *mode; + ir_node *load; + ir_node *load_res; + be_abi_irg_t *env = data; + int arity, i; + be_main_env_t *be = env->birg->main_env; + + arity = get_irn_arity(node); + for (i = 0; i < arity; ++i) { + ir_node *pred = get_irn_n(node, i); + ir_entity *entity; + if (!is_SymConst(pred)) + continue; + + entity = get_SymConst_entity(pred); + block = get_nodes_block(pred); + irg = get_irn_irg(pred); + + /* calls can jump to relative addresses, so we can directly jump to + the (relatively) known call address or the trampoline */ + if (is_Call(node) && i == 1) { + dbg_info *dbgi; + ir_entity *trampoline; + ir_node *trampoline_const; + + if (can_address_relative(entity)) + continue; + + dbgi = get_irn_dbg_info(pred); + trampoline = get_trampoline(be, entity); + trampoline_const = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code, trampoline, NULL); + set_irn_n(node, i, trampoline_const); + continue; + } + + /* everything else is accessed relative to EIP */ + mode = get_irn_mode(pred); + unknown = new_r_Unknown(irg, mode); + pic_base = arch_code_generator_get_pic_base(env->birg->cg); + add = new_r_Add(irg, block, pic_base, pred, mode); + + /* make sure the walker doesn't visit this add again */ + mark_irn_visited(add); + + /* all ok now for locally constructed stuff */ + if (can_address_relative(entity)) { + set_irn_n(node, i, add); + continue; + } + + /* we need an extra indirection for global data outside our current + module. The loads are always safe and can therefore float + and need no memory input */ + load = new_r_Load(irg, block, new_NoMem(), add, mode); + load_res = new_r_Proj(irg, block, load, mode, pn_Load_res); + set_irn_pinned(load, op_pin_state_floats); + + set_irn_n(node, i, load_res); + } +} + be_abi_irg_t *be_abi_introduce(be_irg_t *birg) { be_abi_irg_t *env = xmalloc(sizeof(env[0])); @@ -2057,7 +2128,7 @@ be_abi_irg_t *be_abi_introduce(be_irg_t *birg) obstack_init(&env->obst); - env->isa = birg->main_env->arch_env->isa; + env->isa = birg->main_env->arch_env.isa; env->method_type = get_entity_type(get_irg_entity(irg)); env->call = be_abi_call_new(env->isa->sp->reg_class); arch_isa_get_call_abi(env->isa, env->method_type, env->call); @@ -2086,6 +2157,10 @@ be_abi_irg_t *be_abi_introduce(be_irg_t *birg) env->calls = NEW_ARR_F(ir_node*, 0); + if (birg->main_env->options->pic) { + irg_walk_graph(irg, fix_pic_symconsts, NULL, env); + } + /* Lower all call nodes in the IRG. */ process_calls(env); @@ -2093,7 +2168,7 @@ be_abi_irg_t *be_abi_introduce(be_irg_t *birg) Beware: init backend abi call object after processing calls, otherwise some information might be not yet available. */ - env->cb = env->call->cb->init(env->call, birg->main_env->arch_env, irg); + env->cb = env->call->cb->init(env->call, &birg->main_env->arch_env, irg); /* Process the IRG */ modify_irg(env); @@ -2103,9 +2178,11 @@ be_abi_irg_t *be_abi_introduce(be_irg_t *birg) /* We don't need the keep map anymore. */ pmap_destroy(env->keep_map); + env->keep_map = NULL; /* calls array is not needed anymore */ DEL_ARR_F(env->calls); + env->calls = NULL; /* reroute the stack origin of the calls to the true stack origin. */ exchange(dummy, env->init_sp); @@ -2113,7 +2190,7 @@ be_abi_irg_t *be_abi_introduce(be_irg_t *birg) /* Make some important node pointers survive the dead node elimination. */ survive_dce_register_irn(env->dce_survivor, &env->init_sp); - pmap_foreach(env->regs, ent) { + foreach_pmap(env->regs, ent) { survive_dce_register_irn(env->dce_survivor, (ir_node **) &ent->value); } @@ -2187,7 +2264,7 @@ void be_abi_fix_stack_nodes(be_abi_irg_t *env) arch_isa_t *isa; walker_env.sp_nodes = NEW_ARR_F(ir_node*, 0); - walker_env.arch_env = birg->main_env->arch_env; + walker_env.arch_env = &birg->main_env->arch_env; isa = walker_env.arch_env->isa; irg_walk_graph(birg->irg, collect_stack_nodes_walker, NULL, &walker_env); @@ -2231,11 +2308,12 @@ void be_abi_fix_stack_nodes(be_abi_irg_t *env) DEL_ARR_F(walker_env.sp_nodes); } -static int process_stack_bias(be_abi_irg_t *env, ir_node *bl, int bias) +static int process_stack_bias(be_abi_irg_t *env, ir_node *bl, int real_bias) { - const arch_env_t *arch_env = env->birg->main_env->arch_env; - int omit_fp = env->call->flags.bits.try_omit_fp; - ir_node *irn; + const arch_env_t *arch_env = &env->birg->main_env->arch_env; + int omit_fp = env->call->flags.bits.try_omit_fp; + ir_node *irn; + int wanted_bias = real_bias; sched_foreach(bl, irn) { int ofs; @@ -2247,34 +2325,59 @@ static int process_stack_bias(be_abi_irg_t *env, ir_node *bl, int bias) */ ir_entity *ent = arch_get_frame_entity(arch_env, irn); if(ent) { + int bias = omit_fp ? real_bias : 0; int offset = get_stack_entity_offset(env->frame, ent, bias); arch_set_frame_offset(arch_env, irn, offset); - DBG((env->dbg, LEVEL_2, "%F has offset %d (including bias %d)\n", ent, offset, bias)); + DBG((env->dbg, LEVEL_2, "%F has offset %d (including bias %d)\n", + ent, offset, bias)); } - if(omit_fp || be_is_IncSP(irn)) { - /* - * If the node modifies the stack pointer by a constant offset, - * record that in the bias. - */ - ofs = arch_get_sp_bias(arch_env, irn); - - if(be_is_IncSP(irn)) { - if(ofs == BE_STACK_FRAME_SIZE_EXPAND) { - ofs = get_type_size_bytes(get_irg_frame_type(env->birg->irg)); - be_set_IncSP_offset(irn, ofs); - } else if(ofs == BE_STACK_FRAME_SIZE_SHRINK) { - ofs = - get_type_size_bytes(get_irg_frame_type(env->birg->irg)); - be_set_IncSP_offset(irn, ofs); + /* + * If the node modifies the stack pointer by a constant offset, + * record that in the bias. + */ + ofs = arch_get_sp_bias(arch_env, irn); + + if(be_is_IncSP(irn)) { + /* fill in real stack frame size */ + if(ofs == BE_STACK_FRAME_SIZE_EXPAND) { + ir_type *frame_type = get_irg_frame_type(env->birg->irg); + ofs = (int) get_type_size_bytes(frame_type); + be_set_IncSP_offset(irn, ofs); + } else if(ofs == BE_STACK_FRAME_SIZE_SHRINK) { + ir_type *frame_type = get_irg_frame_type(env->birg->irg); + ofs = - (int)get_type_size_bytes(frame_type); + be_set_IncSP_offset(irn, ofs); + } else { + if (be_get_IncSP_align(irn)) { + /* patch IncSP to produce an aligned stack pointer */ + ir_type *between_type = env->frame->between_type; + int between_size = get_type_size_bytes(between_type); + int alignment = env->isa->stack_alignment; + int delta = (real_bias + ofs + between_size) % env->isa->stack_alignment; + assert(ofs >= 0); + if (delta > 0) { + be_set_IncSP_offset(irn, ofs + alignment - delta); + real_bias += alignment - delta; + } + } else { + /* adjust so real_bias corresponds with wanted_bias */ + int delta = wanted_bias - real_bias; + assert(delta <= 0); + if(delta != 0) { + be_set_IncSP_offset(irn, ofs + delta); + real_bias += delta; + } } } - - if(omit_fp) - bias += ofs; } + + real_bias += ofs; + wanted_bias += ofs; } - return bias; + assert(real_bias == wanted_bias); + return real_bias; } /** @@ -2282,8 +2385,9 @@ static int process_stack_bias(be_abi_irg_t *env, ir_node *bl, int bias) */ struct bias_walk { be_abi_irg_t *env; /**< The ABI irg environment. */ - int start_block_bias; /**< The bias at the end of the start block. */ - ir_node *start_block; /**< The start block of the current graph. */ + int start_block_bias; /**< The bias at the end of the start block. */ + int between_size; + ir_node *start_block; /**< The start block of the current graph. */ }; /** @@ -2307,6 +2411,7 @@ void be_abi_fix_stack_bias(be_abi_irg_t *env) /* Determine the stack bias at the end of the start block. */ bw.start_block_bias = process_stack_bias(env, get_irg_start_block(irg), 0); + bw.between_size = get_type_size_bytes(env->frame->between_type); /* fix the bias is all other blocks */ bw.env = env; @@ -2328,11 +2433,6 @@ ir_node *be_abi_get_ignore_irn(be_abi_irg_t *abi, const arch_register_t *reg) return pmap_get(abi->regs, (void *) reg); } -ir_node *be_abi_get_start_barrier(be_abi_irg_t *abi) -{ - return abi->start_barrier; -} - /** * Returns non-zero if the ABI has omitted the frame pointer in * the current graph.