void be_default_lower_va_arg(ir_node *const node, bool const compound_is_ptr,
unsigned const stack_param_align)
{
ir_node *block = get_nodes_block(node);
dbg_info *dbgi = get_irn_dbg_info(node);
ir_graph *irg = get_irn_irg(node);
ir_type *aptype = get_method_res_type(get_Builtin_type(node), 0);
ir_node *const ap = get_irn_n(node, 1);
ir_node *const node_mem = get_Builtin_mem(node);
ir_mode *apmode = get_type_mode(aptype);
ir_node *res;
ir_node *new_mem;
if (apmode) {
goto load;
} else if (compound_is_ptr) {
apmode = mode_P;
aptype = get_type_for_mode(apmode);
load:;
ir_node *const load = new_rd_Load(dbgi, block, node_mem, ap, apmode, aptype, cons_none);
res = new_r_Proj(load, apmode, pn_Load_res);
new_mem = new_r_Proj(load, mode_M,pn_Load_M);
} else {
/* aptype has no associated mode, so it is represented as a pointer. */
res = ap;
new_mem = node_mem;
}
unsigned const round_up = round_up2(get_type_size(aptype),
stack_param_align);
ir_mode *const offset_mode = get_reference_offset_mode(mode_P);
ir_node *const offset = new_r_Const_long(irg, offset_mode, round_up);
ir_node *const new_ap = new_rd_Add(dbgi, block, ap, offset);
ir_node *const in[] = { new_mem, res, new_ap };
turn_into_tuple(node, ARRAY_SIZE(in), in);
}
calling_convention_t *sparc_decide_calling_convention(ir_type *function_type,
ir_graph *irg)
{
bool omit_fp = false;
if (irg != NULL) {
omit_fp = be_options.omit_fp;
/* our current vaarg handling needs the standard space to store the
* args 0-5 in it */
if (is_method_variadic(function_type))
omit_fp = false;
/* The pointer to the aggregate return value belongs to the 92 magic bytes.
* Thus, if the called functions increases the stack size,
* it must copy the value to the appropriate location.
* This is not implemented yet, so we forbid to omit the frame pointer.
*/
if (get_method_calling_convention(function_type) & cc_compound_ret)
omit_fp = false;
if (omit_fp)
irg_walk_graph(irg, check_omit_fp, NULL, &omit_fp);
sparc_get_irg_data(irg)->omit_fp = omit_fp;
}
mtp_additional_properties mtp
= get_method_additional_properties(function_type);
unsigned *caller_saves = rbitset_malloc(N_SPARC_REGISTERS);
if (mtp & mtp_property_returns_twice) {
rbitset_copy(caller_saves, default_returns_twice_saves,
N_SPARC_REGISTERS);
} else {
rbitset_copy(caller_saves, default_caller_saves, N_SPARC_REGISTERS);
}
/* determine how parameters are passed */
int n_params = get_method_n_params(function_type);
int regnum = 0;
reg_or_stackslot_t *params = XMALLOCNZ(reg_or_stackslot_t, n_params);
int n_param_regs = ARRAY_SIZE(param_regs);
unsigned stack_offset = !omit_fp ? SPARC_MIN_STACKSIZE : 0;
for (int i = 0; i < n_params; ++i) {
ir_type *param_type = get_method_param_type(function_type,i);
ir_mode *mode;
int bits;
reg_or_stackslot_t *param;
if (is_compound_type(param_type))
panic("compound arguments not supported yet");
mode = get_type_mode(param_type);
bits = get_mode_size_bits(mode);
param = ¶ms[i];
if (i == 0 &&
(get_method_calling_convention(function_type) & cc_compound_ret)) {
assert(mode_is_reference(mode) && bits == 32);
/* special case, we have reserved space for this on the between
* type */
param->type = param_type;
param->offset = SPARC_AGGREGATE_RETURN_OFFSET;
param->already_stored = true;
continue;
}
if (regnum < n_param_regs) {
param->offset = SPARC_PARAMS_SPILL_OFFSET
+ regnum * SPARC_REGISTER_SIZE;
param->type = param_type;
arch_register_t const *reg = param_regs[regnum++];
if (irg == NULL || omit_fp)
reg = map_i_to_o_reg(reg);
param->reg0 = reg;
param->req0 = reg->single_req;
} else {
param->type = param_type;
param->offset = stack_offset;
param->already_stored = true;
/* increase offset by at least SPARC_REGISTER_SIZE bytes so
* everything is aligned */
stack_offset += MAX(bits / 8, SPARC_REGISTER_SIZE);
continue;
}
/* we might need a 2nd 32bit component (for 64bit or double values) */
if (bits > 32) {
if (bits > 64)
panic("only 32 and 64bit modes supported");
if (regnum < n_param_regs) {
param->offset = SPARC_PARAMS_SPILL_OFFSET
+ regnum * SPARC_REGISTER_SIZE;
arch_register_t const *reg = param_regs[regnum++];
if (irg == NULL || omit_fp)
reg = map_i_to_o_reg(reg);
param->reg1 = reg;
param->req1 = reg->single_req;
} else {
ir_mode *regmode = param_regs[0]->cls->mode;
ir_type *type = get_type_for_mode(regmode);
param->type = type;
param->offset = stack_offset;
assert(get_mode_size_bits(regmode) == 32);
stack_offset += SPARC_REGISTER_SIZE;
}
}
}
unsigned n_param_regs_used = regnum;
/* determine how results are passed */
int n_results = get_method_n_ress(function_type);
unsigned float_regnum = 0;
unsigned n_reg_results = 0;
unsigned n_float_result_regs = ARRAY_SIZE(float_result_regs);
reg_or_stackslot_t *results = XMALLOCNZ(reg_or_stackslot_t, n_results);
regnum = 0;
for (int i = 0; i < n_results; ++i) {
ir_type *result_type = get_method_res_type(function_type, i);
ir_mode *result_mode = get_type_mode(result_type);
reg_or_stackslot_t *result = &results[i];
if (mode_is_float(result_mode)) {
unsigned n_regs = determine_n_float_regs(result_mode);
unsigned next_reg = round_up2(float_regnum, n_regs);
if (next_reg >= n_float_result_regs) {
panic("too many float results");
} else {
const arch_register_t *reg = float_result_regs[next_reg];
rbitset_clear(caller_saves, reg->global_index);
if (n_regs == 1) {
result->req0 = reg->single_req;
} else if (n_regs == 2) {
result->req0 = &float_result_reqs_double[next_reg];
rbitset_clear(caller_saves, reg->global_index+1);
} else if (n_regs == 4) {
result->req0 = &float_result_reqs_quad[next_reg];
rbitset_clear(caller_saves, reg->global_index+1);
rbitset_clear(caller_saves, reg->global_index+2);
rbitset_clear(caller_saves, reg->global_index+3);
} else {
panic("invalid number of registers in result");
}
float_regnum = next_reg + n_regs;
++n_reg_results;
}
} else {
if (get_mode_size_bits(result_mode) > 32) {
panic("results with more than 32bits not supported yet");
}
if (regnum >= n_param_regs) {
panic("too many results");
} else {
const arch_register_t *reg = param_regs[regnum++];
if (irg == NULL || omit_fp)
reg = map_i_to_o_reg(reg);
result->req0 = reg->single_req;
rbitset_clear(caller_saves, reg->global_index);
++n_reg_results;
}
}
}
calling_convention_t *cconv = XMALLOCZ(calling_convention_t);
cconv->n_parameters = n_params;
cconv->parameters = params;
cconv->param_stack_size = stack_offset - SPARC_MIN_STACKSIZE;
cconv->n_param_regs = n_param_regs_used;
cconv->results = results;
cconv->omit_fp = omit_fp;
cconv->caller_saves = caller_saves;
cconv->n_reg_results = n_reg_results;
/* setup ignore register array */
if (irg != NULL) {
be_irg_t *birg = be_birg_from_irg(irg);
birg->allocatable_regs = be_cconv_alloc_all_regs(&birg->obst, N_SPARC_REGISTERS);
be_cconv_rem_regs(birg->allocatable_regs, ignore_regs, ARRAY_SIZE(ignore_regs));
}
return cconv;
}
void layout_compound(compound_t *const compound)
{
bool const is_union = compound->base.kind == ENTITY_UNION;
bool const is_packed = compound->packed;
il_alignment_t alignment = compound->alignment;
size_t bit_offset = 0;
il_size_t size = 0;
bool need_pad = false;
for (entity_t *entry = compound->members.entities; entry; entry = entry->base.next) {
if (entry->kind != ENTITY_COMPOUND_MEMBER)
continue;
compound_member_t *const member = &entry->compound_member;
type_t *const m_type = skip_typeref(member->base.type);
if (!is_type_valid(m_type))
continue;
if (is_packed) {
/* GCC: Specifying this attribute for `struct' and `union' types is
* equivalent to specifying the `packed' attribute on each of the
* structure or union members. */
member->base.alignment = 1;
}
il_alignment_t const m_alignment = member->base.alignment;
alignment = MAX(alignment, m_alignment);
unsigned const m_size = get_type_size(m_type);
if (is_union) {
size = MAX(size, m_size);
} else if (member->bitfield) {
il_alignment_t const alignment_mask = m_alignment - 1;
size_t const base_size = m_size * BITS_PER_BYTE;
size_t const bit_size = member->bit_size;
if (!is_packed) {
bit_offset += (size & alignment_mask) * BITS_PER_BYTE;
size &= ~alignment_mask;
if (bit_offset + bit_size > base_size || bit_size == 0) {
size += (bit_offset + BITS_PER_BYTE - 1) / BITS_PER_BYTE;
size = round_up2(size, m_alignment);
bit_offset = 0;
}
}
if (target.byte_order_big_endian) {
member->offset = size & ~alignment_mask;
member->bit_offset = base_size - bit_offset - bit_size;
} else {
member->offset = size;
member->bit_offset = bit_offset;
}
bit_offset += bit_size;
size += bit_offset / BITS_PER_BYTE;
bit_offset %= BITS_PER_BYTE;
} else {
if (bit_offset != 0) {
bit_offset = 0;
size += 1;
}
il_size_t const new_size = round_up2(size, m_alignment);
if (new_size > size) {
need_pad = true;
size = new_size;
}
member->offset = size;
size += m_size;
}
}
if (bit_offset != 0)
size += 1;
il_size_t const new_size = round_up2(size, alignment);
if (new_size > size) {
need_pad = true;
size = new_size;
}
position_t const *const pos = &compound->base.pos;
if (need_pad) {
warningf(WARN_PADDED, pos, "'%N' needs padding", compound);
} else if (is_packed) {
warningf(WARN_PACKED, pos, "superfluous packed attribute on '%N'", compound);
}
compound->size = size;
compound->alignment = alignment;
}
void layout_compound(compound_t *const compound)
{
bool const is_union = compound->base.kind == ENTITY_UNION;
unsigned alignment = compound->alignment;
size_t bit_offset = 0;
unsigned size = 0;
bool need_pad = false;
for (entity_t *entry = compound->members.first_entity; entry;
entry = entry->base.next) {
if (entry->kind != ENTITY_COMPOUND_MEMBER)
continue;
compound_member_t *const member = &entry->compound_member;
type_t *const m_type = skip_typeref(member->base.type);
if (!is_type_valid(m_type))
continue;
unsigned m_alignment = get_declaration_alignment(&member->base);
alignment = MAX(alignment, m_alignment);
unsigned const m_size = get_ctype_size(m_type);
if (is_union) {
size = MAX(size, m_size);
} else if (member->bitfield) {
unsigned const alignment_mask = m_alignment - 1;
size_t const base_size = m_size * BITS_PER_BYTE;
size_t const bit_size = member->bit_size;
bit_offset += (size & alignment_mask) * BITS_PER_BYTE;
size &= ~alignment_mask;
if (bit_offset + bit_size > base_size
|| (bit_size == 0 && !(member->base.modifiers & DM_PACKED))) {
size += (bit_offset + BITS_PER_BYTE - 1) / BITS_PER_BYTE;
size = round_up2(size, m_alignment);
bit_offset = 0;
}
if (target.byte_order_big_endian) {
member->offset = size & ~alignment_mask;
member->bit_offset = base_size - bit_offset - bit_size;
} else {
member->offset = size;
member->bit_offset = bit_offset;
}
bit_offset += bit_size;
size += bit_offset / BITS_PER_BYTE;
bit_offset %= BITS_PER_BYTE;
} else {
if (bit_offset != 0) {
bit_offset = 0;
size += 1;
}
unsigned const new_size = round_up2(size, m_alignment);
if (new_size > size) {
need_pad = true;
size = new_size;
}
member->offset = size;
size += m_size;
}
}
if (bit_offset != 0)
size += 1;
unsigned const new_size = round_up2(size, alignment);
if (new_size > size) {
need_pad = true;
size = new_size;
}
position_t const *const pos = &compound->base.pos;
if (need_pad) {
warningf(WARN_PADDED, pos, "%N needs padding", compound);
} else if (compound->packed) {
warningf(WARN_PACKED, pos, "superfluous packed attribute on %N",
compound);
}
compound->size = size;
compound->alignment = alignment;
}
x86_cconv_t *amd64_decide_calling_convention(ir_type *function_type,
ir_graph *irg)
{
bool omit_fp = false;
if (irg != NULL) {
omit_fp = be_options.omit_fp;
if (omit_fp)
irg_walk_graph(irg, check_omit_fp, NULL, &omit_fp);
amd64_get_irg_data(irg)->omit_fp = omit_fp;
}
unsigned *caller_saves = rbitset_malloc(N_AMD64_REGISTERS);
unsigned *callee_saves = rbitset_malloc(N_AMD64_REGISTERS);
rbitset_copy(caller_saves, default_caller_saves, N_AMD64_REGISTERS);
rbitset_copy(callee_saves, default_callee_saves, N_AMD64_REGISTERS);
/* determine how parameters are passed */
size_t n_params = get_method_n_params(function_type);
size_t param_regnum = 0;
size_t float_param_regnum = 0;
reg_or_stackslot_t *params = XMALLOCNZ(reg_or_stackslot_t,
n_params);
/* x64 always reserves space to spill the first 4 arguments to have it
* easy in case of variadic functions. */
unsigned stack_offset = amd64_use_x64_abi ? 32 : 0;
for (size_t i = 0; i < n_params; ++i) {
ir_type *param_type = get_method_param_type(function_type,i);
if (is_compound_type(param_type))
panic("compound arguments NIY");
ir_mode *mode = get_type_mode(param_type);
int bits = get_mode_size_bits(mode);
reg_or_stackslot_t *param = ¶ms[i];
if (mode_is_float(mode) && float_param_regnum < n_float_param_regs
&& mode != x86_mode_E) {
param->reg = float_param_regs[float_param_regnum++];
if (amd64_use_x64_abi) {
++param_regnum;
}
} else if (!mode_is_float(mode) && param_regnum < n_param_regs) {
param->reg = param_regs[param_regnum++];
if (amd64_use_x64_abi) {
++float_param_regnum;
}
} else {
param->type = param_type;
param->offset = stack_offset;
/* increase offset by at least AMD64_REGISTER_SIZE bytes so
* everything is aligned */
stack_offset += round_up2(bits / 8, AMD64_REGISTER_SIZE);
}
}
/* If the function is variadic, we add all unused parameter
* passing registers to the end of the params array, first GP,
* then XMM. */
if (irg && is_method_variadic(function_type)) {
if (amd64_use_x64_abi) {
panic("Variadic functions on Windows ABI not supported");
}
int params_remaining = (n_param_regs - param_regnum) +
(n_float_param_regs - float_param_regnum);
params = XREALLOC(params, reg_or_stackslot_t, n_params + params_remaining);
size_t i = n_params;
for (; param_regnum < n_param_regs; param_regnum++, i++) {
params[i].reg = param_regs[param_regnum];
}
for (; float_param_regnum < n_float_param_regs; float_param_regnum++, i++) {
params[i].reg = float_param_regs[float_param_regnum];
}
}
unsigned n_param_regs_used
= amd64_use_x64_abi ? param_regnum : param_regnum + float_param_regnum;
/* determine how results are passed */
size_t n_results = get_method_n_ress(function_type);
unsigned n_reg_results = 0;
reg_or_stackslot_t *results = XMALLOCNZ(reg_or_stackslot_t, n_results);
unsigned res_regnum = 0;
unsigned res_float_regnum = 0;
unsigned res_x87_regnum = 0;
size_t n_result_regs = ARRAY_SIZE(result_regs);
size_t n_float_result_regs = ARRAY_SIZE(float_result_regs);
size_t n_x87_result_regs = ARRAY_SIZE(x87_result_regs);
for (size_t i = 0; i < n_results; ++i) {
ir_type *result_type = get_method_res_type(function_type, i);
ir_mode *result_mode = get_type_mode(result_type);
reg_or_stackslot_t *result = &results[i];
const arch_register_t *reg;
if (result_mode == x86_mode_E) {
if (res_x87_regnum >= n_x87_result_regs)
panic("too manu x87 floating point results");
reg = x87_result_regs[res_x87_regnum++];
} else if (mode_is_float(result_mode)) {
if (res_float_regnum >= n_float_result_regs) {
panic("too many floating points results");
}
reg = float_result_regs[res_float_regnum++];
} else {
if (res_regnum >= n_result_regs) {
panic("too many results");
}
reg = result_regs[res_regnum++];
}
result->reg = reg;
rbitset_clear(caller_saves, reg->global_index);
++n_reg_results;
}
x86_cconv_t *cconv = XMALLOCZ(x86_cconv_t);
cconv->parameters = params;
cconv->n_parameters = n_params;
cconv->param_stacksize = stack_offset;
cconv->n_param_regs = n_param_regs_used;
cconv->n_xmm_regs = float_param_regnum;
cconv->results = results;
cconv->omit_fp = omit_fp;
cconv->caller_saves = caller_saves;
cconv->callee_saves = callee_saves;
cconv->n_reg_results = n_reg_results;
if (irg != NULL) {
be_irg_t *birg = be_birg_from_irg(irg);
birg->allocatable_regs = be_cconv_alloc_all_regs(&birg->obst, N_AMD64_REGISTERS);
be_cconv_rem_regs(birg->allocatable_regs, ignore_regs, ARRAY_SIZE(ignore_regs));
if (!omit_fp)
rbitset_clear(birg->allocatable_regs, REG_RBP);
}
return cconv;
}