}
/**
- * Get an atomic entity that is initialized with a tarval
+ * Creates an immediate.
+ *
+ * @param symconst if set, create a SymConst immediate
+ * @param symconst_sign sign for the symconst
+ * @param val integer value for the immediate
+ */
+static ir_node *create_Immediate(ir_entity *symconst, int symconst_sign, long val)
+{
+ ir_graph *irg = current_ir_graph;
+ 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]);
+
+ return immediate;
+}
+
+/**
+ * Get an atomic entity that is initialized with a tarval forming
+ * a given constant.
+ *
+ * @param cnst the node representing the constant
*/
static ir_entity *create_float_const_entity(ir_node *cnst)
{
ia32_isa_t *isa = env_cg->isa;
- tarval *tv = get_Const_tarval(cnst);
- pmap_entry *e = pmap_find(isa->tv_ent, tv);
+ tarval *key = get_Const_tarval(cnst);
+ pmap_entry *e = pmap_find(isa->tv_ent, key);
ir_entity *res;
ir_graph *rem;
- if (! e) {
- ir_mode *mode = get_irn_mode(cnst);
- ir_type *tp = get_Const_type(cnst);
- if (tp == firm_unknown_type)
+ if (e == NULL) {
+ tarval *tv = key;
+ ir_mode *mode = get_tarval_mode(tv);
+ ir_type *tp;
+
+ if (! ia32_cg_config.use_sse2) {
+ /* try to reduce the mode to produce smaller sized entities */
+ if (mode != mode_F) {
+ if (tarval_ieee754_can_conv_lossless(tv, mode_F)) {
+ mode = mode_F;
+ tv = tarval_convert_to(tv, mode);
+ } else if (mode != mode_D) {
+ if (tarval_ieee754_can_conv_lossless(tv, mode_D)) {
+ mode = mode_D;
+ tv = tarval_convert_to(tv, mode);
+ }
+ }
+ }
+ }
+
+ if (mode == get_irn_mode(cnst)) {
+ /* mode was not changed */
+ tp = get_Const_type(cnst);
+ if (tp == firm_unknown_type)
+ tp = get_prim_type(isa->types, mode);
+ } else
tp = get_prim_type(isa->types, mode);
res = new_entity(get_glob_type(), unique_id(".LC%u"), tp);
set_atomic_ent_value(res, new_Const_type(tv, tp));
current_ir_graph = rem;
- pmap_insert(isa->tv_ent, tv, res);
+ pmap_insert(isa->tv_ent, key, res);
} else {
res = e->value;
}
static int is_simple_x87_Const(ir_node *node)
{
tarval *tv = get_Const_tarval(node);
+ if (tarval_is_null(tv) || tarval_is_one(tv))
+ return 1;
+
+ /* TODO: match all the other float constants */
+ return 0;
+}
+
+/**
+ * returns true if constant can be created with a simple float command
+ */
+static int is_simple_sse_Const(ir_node *node)
+{
+ tarval *tv = get_Const_tarval(node);
+ ir_mode *mode = get_tarval_mode(tv);
+
+ if (mode == mode_F)
+ return 1;
- if(tarval_is_null(tv) || tarval_is_one(tv))
+ if (tarval_is_null(tv) || tarval_is_one(tv))
return 1;
+ if (mode == mode_D) {
+ unsigned val = get_tarval_sub_bits(tv, 0) |
+ (get_tarval_sub_bits(tv, 1) << 8) |
+ (get_tarval_sub_bits(tv, 2) << 16) |
+ (get_tarval_sub_bits(tv, 3) << 24);
+ if (val == 0)
+ /* lower 32bit are zero, really a 32bit constant */
+ return 1;
+ }
+
/* TODO: match all the other float constants */
return 0;
}
ir_entity *floatent;
if (ia32_cg_config.use_sse2) {
- if (is_Const_null(node)) {
+ tarval *tv = get_Const_tarval(node);
+ if (tarval_is_null(tv)) {
load = new_rd_ia32_xZero(dbgi, irg, block);
set_ia32_ls_mode(load, mode);
res = load;
+ } else if (tarval_is_one(tv)) {
+ int cnst = mode == mode_F ? 26 : 55;
+ ir_node *imm1 = create_Immediate(NULL, 0, cnst);
+ ir_node *imm2 = create_Immediate(NULL, 0, 2);
+ ir_node *pslld, *psrld;
+
+ load = new_rd_ia32_xAllOnes(dbgi, irg, block);
+ set_ia32_ls_mode(load, mode);
+ pslld = new_rd_ia32_xPslld(dbgi, irg, block, load, imm1);
+ set_ia32_ls_mode(pslld, mode);
+ psrld = new_rd_ia32_xPsrld(dbgi, irg, block, pslld, imm2);
+ set_ia32_ls_mode(psrld, mode);
+ res = psrld;
+ } else if (mode == mode_F) {
+ /* we can place any 32bit constant by using a movd gp, sse */
+ unsigned val = get_tarval_sub_bits(tv, 0) |
+ (get_tarval_sub_bits(tv, 1) << 8) |
+ (get_tarval_sub_bits(tv, 2) << 16) |
+ (get_tarval_sub_bits(tv, 3) << 24);
+ ir_node *cnst = new_rd_ia32_Const(dbgi, irg, block, NULL, 0, val);
+ load = new_rd_ia32_xMovd(dbgi, irg, block, cnst);
+ set_ia32_ls_mode(load, mode);
+ res = load;
} else {
+ if (mode == mode_D) {
+ unsigned val = get_tarval_sub_bits(tv, 0) |
+ (get_tarval_sub_bits(tv, 1) << 8) |
+ (get_tarval_sub_bits(tv, 2) << 16) |
+ (get_tarval_sub_bits(tv, 3) << 24);
+ if (val == 0) {
+ ir_node *imm32 = create_Immediate(NULL, 0, 32);
+ ir_node *cnst, *psllq;
+
+ /* fine, lower 32bit are zero, produce 32bit value */
+ val = get_tarval_sub_bits(tv, 4) |
+ (get_tarval_sub_bits(tv, 5) << 8) |
+ (get_tarval_sub_bits(tv, 6) << 16) |
+ (get_tarval_sub_bits(tv, 7) << 24);
+ cnst = new_rd_ia32_Const(dbgi, irg, block, NULL, 0, val);
+ load = new_rd_ia32_xMovd(dbgi, irg, block, cnst);
+ set_ia32_ls_mode(load, mode);
+ psllq = new_rd_ia32_xPsllq(dbgi, irg, block, load, imm32);
+ set_ia32_ls_mode(psllq, mode);
+ res = psllq;
+ goto end;
+ }
+ }
floatent = create_float_const_entity(node);
load = new_rd_ia32_xLoad(dbgi, irg, block, noreg, noreg, nomem,
- mode);
+ mode);
set_ia32_op_type(load, ia32_AddrModeS);
set_ia32_am_sc(load, floatent);
set_ia32_flags(load, get_ia32_flags(load) | arch_irn_flags_rematerializable);
if (is_Const_null(node)) {
load = new_rd_ia32_vfldz(dbgi, irg, block);
res = load;
+ set_ia32_ls_mode(load, mode);
} else if (is_Const_one(node)) {
load = new_rd_ia32_vfld1(dbgi, irg, block);
res = load;
+ set_ia32_ls_mode(load, mode);
} else {
floatent = create_float_const_entity(node);
set_ia32_am_sc(load, floatent);
set_ia32_flags(load, get_ia32_flags(load) | arch_irn_flags_rematerializable);
res = new_r_Proj(irg, block, load, mode_vfp, pn_ia32_vfld_res);
+ /* take the mode from the entity */
+ set_ia32_ls_mode(load, get_type_mode(get_entity_type(floatent)));
}
- set_ia32_ls_mode(load, mode);
}
-
- SET_IA32_ORIG_NODE(load, ia32_get_old_node_name(env_cg, node));
-
+end:
/* 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
SET_IA32_ORIG_NODE(load, ia32_get_old_node_name(env_cg, node));
return res;
- } else {
+ } else { /* non-float mode */
ir_node *cnst;
tarval *tv = get_Const_tarval(node);
long val;
tv = tarval_convert_to(tv, mode_Iu);
- if(tv == get_tarval_bad() || tv == get_tarval_undefined()
- || tv == NULL) {
+ if (tv == get_tarval_bad() || tv == get_tarval_undefined() ||
+ tv == NULL) {
panic("couldn't convert constant tarval (%+F)", node);
}
val = get_tarval_long(tv);
* input here, for unary operations use NULL).
*/
static int ia32_use_source_address_mode(ir_node *block, ir_node *node,
- ir_node *other, ir_node *other2)
+ ir_node *other, ir_node *other2, match_flags_t flags)
{
- ir_mode *mode = get_irn_mode(node);
ir_node *load;
long pn;
/* float constants are always available */
- if(is_Const(node) && mode_is_float(mode)) {
- if(!is_simple_x87_Const(node))
- return 0;
- if(get_irn_n_edges(node) > 1)
- return 0;
- return 1;
+ if (is_Const(node)) {
+ ir_mode *mode = get_irn_mode(node);
+ if (mode_is_float(mode)) {
+ if (ia32_cg_config.use_sse2) {
+ if (is_simple_sse_Const(node))
+ return 0;
+ } else {
+ if (is_simple_x87_Const(node))
+ return 0;
+ }
+ if (get_irn_n_edges(node) > 1)
+ return 0;
+ return 1;
+ }
}
- if(!is_Proj(node))
+ if (!is_Proj(node))
return 0;
load = get_Proj_pred(node);
pn = get_Proj_proj(node);
- if(!is_Load(load) || pn != pn_Load_res)
+ if (!is_Load(load) || pn != pn_Load_res)
return 0;
- if(get_nodes_block(load) != block)
+ if (get_nodes_block(load) != block)
return 0;
/* we only use address mode if we're the only user of the load */
- if(get_irn_n_edges(node) > 1)
+ if (get_irn_n_edges(node) != (flags & match_two_users ? 2 : 1))
return 0;
/* in some edge cases with address mode we might reach the load normally
* and through some AM sequence, if it is already materialized then we
* can't create an AM node from it */
- if(be_is_transformed(node))
+ if (be_is_transformed(node))
return 0;
/* don't do AM if other node inputs depend on the load (via mem-proj) */
- if(other != NULL && get_nodes_block(other) == block
- && heights_reachable_in_block(heights, other, load))
+ if (other != NULL && get_nodes_block(other) == block &&
+ heights_reachable_in_block(heights, other, load))
return 0;
- if(other2 != NULL && get_nodes_block(other2) == block
- && heights_reachable_in_block(heights, other2, load))
+ if (other2 != NULL && get_nodes_block(other2) == block &&
+ heights_reachable_in_block(heights, other2, load))
return 0;
return 1;
static void build_address_ptr(ia32_address_t *addr, ir_node *ptr, ir_node *mem)
{
- ir_node *noreg_gp = ia32_new_NoReg_gp(env_cg);
+ ir_node *noreg_gp;
/* construct load address */
memset(addr, 0, sizeof(addr[0]));
ia32_create_address_mode(addr, ptr, /*force=*/0);
+ noreg_gp = ia32_new_NoReg_gp(env_cg);
addr->base = addr->base ? be_transform_node(addr->base) : noreg_gp;
addr->index = addr->index ? be_transform_node(addr->index) : noreg_gp;
addr->mem = be_transform_node(mem);
ir_node *mem;
ir_node *new_mem;
- if(is_Const(node)) {
+ if (is_Const(node)) {
ir_entity *entity = create_float_const_entity(node);
addr->base = noreg_gp;
addr->index = noreg_gp;
addr->mem = new_NoMem();
addr->symconst_ent = entity;
addr->use_frame = 1;
- am->ls_mode = get_irn_mode(node);
+ am->ls_mode = get_type_mode(get_entity_type(entity));
am->pinned = op_pin_state_floats;
return;
}
set_ia32_frame_ent(node, addr->frame_entity);
}
+/**
+ * Apply attributes of a given address mode to a node.
+ */
static void set_am_attributes(ir_node *node, const ia32_address_mode_t *am)
{
set_address(node, &am->addr);
set_ia32_op_type(node, am->op_type);
set_ia32_ls_mode(node, am->ls_mode);
- if(am->pinned == op_pin_state_pinned && get_irn_pinned(node) != op_pin_state_pinned) {
+ if (am->pinned == op_pin_state_pinned) {
set_irn_pinned(node, am->pinned);
}
- if(am->commutative)
+ if (am->commutative)
set_ia32_commutative(node);
}
ir_node *op1, ir_node *op2, ir_node *other_op,
match_flags_t flags)
{
- ia32_address_t *addr = &am->addr;
- ir_node *noreg_gp = ia32_new_NoReg_gp(env_cg);
- ir_node *new_op1;
- ir_node *new_op2;
- ir_mode *mode = get_irn_mode(op2);
+ ia32_address_t *addr = &am->addr;
+ ir_mode *mode = get_irn_mode(op2);
+ int mode_bits = get_mode_size_bits(mode);
+ ir_node *noreg_gp, *new_op1, *new_op2;
int use_am;
unsigned commutative;
int use_am_and_immediates;
int use_immediate;
- int mode_bits = get_mode_size_bits(mode);
memset(am, 0, sizeof(am[0]));
assert(use_am || !(flags & match_8bit_am));
assert(use_am || !(flags & match_16bit_am));
- if(mode_bits == 8) {
- if (! (flags & match_8bit_am))
+ if (mode_bits == 8) {
+ if (!(flags & match_8bit_am))
use_am = 0;
/* we don't automatically add upconvs yet */
assert((flags & match_mode_neutral) || (flags & match_8bit));
- } else if(mode_bits == 16) {
- if(! (flags & match_16bit_am))
+ } else if (mode_bits == 16) {
+ if (!(flags & match_16bit_am))
use_am = 0;
/* we don't automatically add upconvs yet */
assert((flags & match_mode_neutral) || (flags & match_16bit));
/* we can simply skip downconvs for mode neutral nodes: the upper bits
* can be random for these operations */
- if(flags & match_mode_neutral) {
+ if (flags & match_mode_neutral) {
op2 = ia32_skip_downconv(op2);
- if(op1 != NULL) {
+ if (op1 != NULL) {
op1 = ia32_skip_downconv(op1);
}
}
/* match immediates. firm nodes are normalized: constants are always on the
* op2 input */
new_op2 = NULL;
- if(! (flags & match_try_am) && use_immediate) {
+ if (!(flags & match_try_am) && use_immediate) {
new_op2 = try_create_Immediate(op2, 0);
}
- if(new_op2 == NULL
- && use_am && ia32_use_source_address_mode(block, op2, op1, other_op)) {
+ noreg_gp = ia32_new_NoReg_gp(env_cg);
+ if (new_op2 == NULL &&
+ use_am && ia32_use_source_address_mode(block, op2, op1, other_op, flags)) {
build_address(am, op2);
new_op1 = (op1 == NULL ? NULL : be_transform_node(op1));
- if(mode_is_float(mode)) {
+ if (mode_is_float(mode)) {
new_op2 = ia32_new_NoReg_vfp(env_cg);
} else {
new_op2 = noreg_gp;
}
am->op_type = ia32_AddrModeS;
- } else if(commutative && (new_op2 == NULL || use_am_and_immediates) &&
- use_am
- && ia32_use_source_address_mode(block, op1, op2, other_op)) {
+ } else if (commutative && (new_op2 == NULL || use_am_and_immediates) &&
+ use_am &&
+ ia32_use_source_address_mode(block, op1, op2, other_op, flags)) {
ir_node *noreg;
build_address(am, op1);
- if(mode_is_float(mode)) {
+ if (mode_is_float(mode)) {
noreg = ia32_new_NoReg_vfp(env_cg);
} else {
noreg = noreg_gp;
}
- if(new_op2 != NULL) {
+ if (new_op2 != NULL) {
new_op1 = noreg;
} else {
new_op1 = be_transform_node(op2);
}
am->op_type = ia32_AddrModeS;
} else {
- if(flags & match_try_am) {
+ if (flags & match_try_am) {
am->new_op1 = NULL;
am->new_op2 = NULL;
am->op_type = ia32_Normal;
}
new_op1 = (op1 == NULL ? NULL : be_transform_node(op1));
- if(new_op2 == NULL)
+ if (new_op2 == NULL)
new_op2 = be_transform_node(op2);
am->op_type = ia32_Normal;
am->ls_mode = get_irn_mode(op2);
- if(flags & match_mode_neutral)
+ if (flags & match_mode_neutral)
am->ls_mode = mode_Iu;
}
- if(addr->base == NULL)
+ if (addr->base == NULL)
addr->base = noreg_gp;
- if(addr->index == NULL)
+ if (addr->index == NULL)
addr->index = noreg_gp;
- if(addr->mem == NULL)
+ if (addr->mem == NULL)
addr->mem = new_NoMem();
am->new_op1 = new_op1;
static ir_node *fix_mem_proj(ir_node *node, ia32_address_mode_t *am)
{
- ir_graph *irg = current_ir_graph;
ir_mode *mode;
ir_node *load;
- if(am->mem_proj == NULL)
+ if (am->mem_proj == NULL)
return node;
/* we have to create a mode_T so the old MemProj can attach to us */
mark_irn_visited(load);
be_set_transformed_node(load, node);
- if(mode != mode_T) {
+ if (mode != mode_T) {
set_irn_mode(node, mode_T);
- return new_rd_Proj(NULL, irg, get_nodes_block(node), node, mode, pn_ia32_res);
+ return new_rd_Proj(NULL, current_ir_graph, get_nodes_block(node), node, mode, pn_ia32_res);
} else {
return node;
}
/**
* Construct a standard binary operation, set AM and immediate if required.
*
+ * @param node The original node for which the binop is created
* @param op1 The first operand
* @param op2 The second operand
* @param func The node constructor function
static ir_node *gen_binop(ir_node *node, ir_node *op1, ir_node *op2,
construct_binop_func *func, match_flags_t flags)
{
- ir_node *block = get_nodes_block(node);
- ir_node *new_block = be_transform_node(block);
- ir_graph *irg = current_ir_graph;
- dbg_info *dbgi = get_irn_dbg_info(node);
- ir_node *new_node;
+ dbg_info *dbgi;
+ ir_node *block, *new_block, *new_node;
ia32_address_mode_t am;
ia32_address_t *addr = &am.addr;
+ block = get_nodes_block(node);
match_arguments(&am, block, op1, op2, NULL, flags);
- new_node = func(dbgi, irg, new_block, addr->base, addr->index, addr->mem,
- am.new_op1, am.new_op2);
+ dbgi = get_irn_dbg_info(node);
+ new_block = be_transform_node(block);
+ new_node = func(dbgi, current_ir_graph, new_block,
+ addr->base, addr->index, addr->mem,
+ am.new_op1, am.new_op2);
set_am_attributes(new_node, &am);
/* we can't use source address mode anymore when using immediates */
- if(is_ia32_Immediate(am.new_op1) || is_ia32_Immediate(am.new_op2))
+ if (is_ia32_Immediate(am.new_op1) || is_ia32_Immediate(am.new_op2))
set_ia32_am_support(new_node, ia32_am_None, ia32_am_arity_none);
SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node));
n_ia32_l_binop_right,
n_ia32_l_binop_eflags
};
-COMPILETIME_ASSERT(n_ia32_l_binop_left == n_ia32_l_Adc_left, n_Adc_left)
-COMPILETIME_ASSERT(n_ia32_l_binop_right == n_ia32_l_Adc_right, n_Adc_right)
-COMPILETIME_ASSERT(n_ia32_l_binop_eflags == n_ia32_l_Adc_eflags, n_Adc_eflags)
-COMPILETIME_ASSERT(n_ia32_l_binop_left == n_ia32_l_Sbb_left, n_Sbb_left)
-COMPILETIME_ASSERT(n_ia32_l_binop_right == n_ia32_l_Sbb_right, n_Sbb_right)
-COMPILETIME_ASSERT(n_ia32_l_binop_eflags == n_ia32_l_Sbb_eflags, n_Sbb_eflags)
+COMPILETIME_ASSERT(n_ia32_l_binop_left == n_ia32_l_Adc_left, n_Adc_left)
+COMPILETIME_ASSERT(n_ia32_l_binop_right == n_ia32_l_Adc_right, n_Adc_right)
+COMPILETIME_ASSERT(n_ia32_l_binop_eflags == n_ia32_l_Adc_eflags, n_Adc_eflags)
+COMPILETIME_ASSERT(n_ia32_l_binop_left == n_ia32_l_Sbb_minuend, n_Sbb_minuend)
+COMPILETIME_ASSERT(n_ia32_l_binop_right == n_ia32_l_Sbb_subtrahend, n_Sbb_subtrahend)
+COMPILETIME_ASSERT(n_ia32_l_binop_eflags == n_ia32_l_Sbb_eflags, n_Sbb_eflags)
/**
* Construct a binary operation which also consumes the eflags.
match_flags_t flags)
{
ir_node *src_block = get_nodes_block(node);
- ir_node *block = be_transform_node(src_block);
ir_node *op1 = get_irn_n(node, n_ia32_l_binop_left);
ir_node *op2 = get_irn_n(node, n_ia32_l_binop_right);
- ir_node *eflags = get_irn_n(node, n_ia32_l_binop_eflags);
- ir_node *new_eflags = be_transform_node(eflags);
- ir_graph *irg = current_ir_graph;
- dbg_info *dbgi = get_irn_dbg_info(node);
- ir_node *new_node;
+ dbg_info *dbgi;
+ ir_node *block, *new_node, *eflags, *new_eflags;
ia32_address_mode_t am;
ia32_address_t *addr = &am.addr;
match_arguments(&am, src_block, op1, op2, NULL, flags);
- new_node = func(dbgi, irg, block, addr->base, addr->index,
- addr->mem, am.new_op1, am.new_op2, new_eflags);
+ dbgi = get_irn_dbg_info(node);
+ block = be_transform_node(src_block);
+ eflags = get_irn_n(node, n_ia32_l_binop_eflags);
+ new_eflags = be_transform_node(eflags);
+ new_node = func(dbgi, current_ir_graph, block, addr->base, addr->index,
+ addr->mem, am.new_op1, am.new_op2, new_eflags);
set_am_attributes(new_node, &am);
/* we can't use source address mode anymore when using immediates */
if(is_ia32_Immediate(am.new_op1) || is_ia32_Immediate(am.new_op2))
static ir_node *get_fpcw(void)
{
ir_node *fpcw;
- if(initial_fpcw != NULL)
+ if (initial_fpcw != NULL)
return initial_fpcw;
fpcw = be_abi_get_ignore_irn(env_cg->birg->abi,
construct_binop_float_func *func,
match_flags_t flags)
{
- ir_graph *irg = current_ir_graph;
- dbg_info *dbgi = get_irn_dbg_info(node);
- ir_node *block = get_nodes_block(node);
- ir_node *new_block = be_transform_node(block);
- ir_mode *mode = get_irn_mode(node);
- ir_node *new_node;
+ ir_mode *mode = get_irn_mode(node);
+ dbg_info *dbgi;
+ ir_node *block, *new_block, *new_node;
ia32_address_mode_t am;
ia32_address_t *addr = &am.addr;
- /* cannot use addresmode with long double on x87 */
+ /* cannot use address mode with long double on x87 */
if (get_mode_size_bits(mode) > 64)
flags &= ~match_am;
+ block = get_nodes_block(node);
match_arguments(&am, block, op1, op2, NULL, flags);
- new_node = func(dbgi, irg, new_block, addr->base, addr->index, addr->mem,
- am.new_op1, am.new_op2, get_fpcw());
+ dbgi = get_irn_dbg_info(node);
+ new_block = be_transform_node(block);
+ new_node = func(dbgi, current_ir_graph, new_block,
+ addr->base, addr->index, addr->mem,
+ am.new_op1, am.new_op2, get_fpcw());
set_am_attributes(new_node, &am);
SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node));
construct_shift_func *func,
match_flags_t flags)
{
- dbg_info *dbgi = get_irn_dbg_info(node);
- ir_graph *irg = current_ir_graph;
- ir_node *block = get_nodes_block(node);
- ir_node *new_block = be_transform_node(block);
- ir_node *new_op1;
- ir_node *new_op2;
- ir_node *new_node;
+ dbg_info *dbgi;
+ ir_node *block, *new_block, *new_op1, *new_op2, *new_node;
assert(! mode_is_float(get_irn_mode(node)));
assert(flags & match_immediate);
assert((flags & ~(match_mode_neutral | match_immediate)) == 0);
- if(flags & match_mode_neutral) {
+ if (flags & match_mode_neutral) {
op1 = ia32_skip_downconv(op1);
+ } else if (get_mode_size_bits(get_irn_mode(node)) != 32) {
+ panic("right shifting of non-32bit values not supported, yet");
}
new_op1 = be_transform_node(op1);
}
new_op2 = create_immediate_or_transform(op2, 0);
- new_node = func(dbgi, irg, new_block, new_op1, new_op2);
+ dbgi = get_irn_dbg_info(node);
+ block = get_nodes_block(node);
+ new_block = be_transform_node(block);
+ new_node = func(dbgi, current_ir_graph, new_block, new_op1, new_op2);
SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node));
/* lowered shift instruction may have a dependency operand, handle it here */
static ir_node *gen_unop(ir_node *node, ir_node *op, construct_unop_func *func,
match_flags_t flags)
{
- ir_graph *irg = current_ir_graph;
- dbg_info *dbgi = get_irn_dbg_info(node);
- ir_node *block = get_nodes_block(node);
- ir_node *new_block = be_transform_node(block);
- ir_node *new_op;
- ir_node *new_node;
+ dbg_info *dbgi;
+ ir_node *block, *new_block, *new_op, *new_node;
assert(flags == 0 || flags == match_mode_neutral);
- if(flags & match_mode_neutral) {
+ if (flags & match_mode_neutral) {
op = ia32_skip_downconv(op);
}
- new_op = be_transform_node(op);
- new_node = func(dbgi, irg, new_block, new_op);
+ new_op = be_transform_node(op);
+ dbgi = get_irn_dbg_info(node);
+ block = get_nodes_block(node);
+ new_block = be_transform_node(block);
+ new_node = func(dbgi, current_ir_graph, new_block, new_op);
SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node));
static ir_node *create_lea_from_address(dbg_info *dbgi, ir_node *block,
ia32_address_t *addr)
{
- ir_graph *irg = current_ir_graph;
- ir_node *base = addr->base;
- ir_node *index = addr->index;
- ir_node *res;
+ ir_node *base, *index, *res;
- if(base == NULL) {
+ base = addr->base;
+ if (base == NULL) {
base = ia32_new_NoReg_gp(env_cg);
} else {
base = be_transform_node(base);
}
- if(index == NULL) {
+ index = addr->index;
+ if (index == NULL) {
index = ia32_new_NoReg_gp(env_cg);
} else {
index = be_transform_node(index);
}
- res = new_rd_ia32_Lea(dbgi, irg, block, base, index);
+ res = new_rd_ia32_Lea(dbgi, current_ir_graph, block, base, index);
set_address(res, addr);
return res;
}
+/**
+ * Returns non-zero if a given address mode has a symbolic or
+ * numerical offset != 0.
+ */
static int am_has_immediates(const ia32_address_t *addr)
{
return addr->offset != 0 || addr->symconst_ent != NULL
* @return the created ia32 Add node
*/
static ir_node *gen_Add(ir_node *node) {
- ir_graph *irg = current_ir_graph;
- dbg_info *dbgi = get_irn_dbg_info(node);
- ir_node *block = get_nodes_block(node);
- ir_node *new_block = be_transform_node(block);
- ir_node *op1 = get_Add_left(node);
- ir_node *op2 = get_Add_right(node);
- ir_mode *mode = get_irn_mode(node);
- ir_node *new_node;
- ir_node *add_immediate_op;
+ ir_mode *mode = get_irn_mode(node);
+ ir_node *op1 = get_Add_left(node);
+ ir_node *op2 = get_Add_right(node);
+ dbg_info *dbgi;
+ ir_node *block, *new_block, *new_node, *add_immediate_op;
ia32_address_t addr;
ia32_address_mode_t am;
memset(&addr, 0, sizeof(addr));
ia32_create_address_mode(&addr, node, /*force=*/1);
add_immediate_op = NULL;
+
+ dbgi = get_irn_dbg_info(node);
+ block = get_nodes_block(node);
+ new_block = be_transform_node(block);
+
/* a constant? */
if(addr.base == NULL && addr.index == NULL) {
+ ir_graph *irg = current_ir_graph;
new_node = new_rd_ia32_Const(dbgi, irg, new_block, addr.symconst_ent,
addr.symconst_sign, addr.offset);
add_irn_dep(new_node, get_irg_frame(irg));
/* construct an Add with source address mode */
if (am.op_type == ia32_AddrModeS) {
+ ir_graph *irg = current_ir_graph;
ia32_address_t *am_addr = &am.addr;
new_node = new_rd_ia32_Add(dbgi, irg, new_block, am_addr->base,
am_addr->index, am_addr->mem, am.new_op1,
return gen_binop_x87_float(node, op1, op2, new_rd_ia32_vfmul,
match_commutative | match_am);
}
-
- /* for the lower 32bit of the result it doesn't matter whether we use
- * signed or unsigned multiplication so we use IMul as it has fewer
- * constraints */
return gen_binop(node, op1, op2, new_rd_ia32_IMul,
match_commutative | match_am | match_mode_neutral |
match_immediate | match_am_and_immediates);
return res;
}
}
-
return gen_binop(node, op1, op2, new_rd_ia32_And,
match_commutative | match_mode_neutral | match_am
| match_immediate);
match_am);
}
- if(is_Const(op2)) {
+ if (is_Const(op2)) {
ir_fprintf(stderr, "Optimisation warning: found sub with const (%+F)\n",
node);
}
match_arguments(&am, block, op1, op2, NULL, match_am);
- if(!is_NoMem(mem)) {
+ /* Beware: We don't need a Sync, if the memory predecessor of the Div node
+ is the memory of the consumed address. We can have only the second op as address
+ in Div nodes, so check only op2. */
+ if(!is_NoMem(mem) && skip_Proj(mem) != skip_Proj(op2)) {
new_mem = be_transform_node(mem);
if(!is_NoMem(addr->mem)) {
ir_node *in[2];
produceval);
new_node = new_rd_ia32_IDiv(dbgi, irg, new_block, addr->base,
- addr->index, new_mem, am.new_op1,
- sign_extension, am.new_op2);
+ addr->index, new_mem, am.new_op2,
+ am.new_op1, sign_extension);
} else {
sign_extension = new_rd_ia32_Const(dbgi, irg, new_block, NULL, 0, 0);
add_irn_dep(sign_extension, get_irg_frame(irg));
new_node = new_rd_ia32_Div(dbgi, irg, new_block, addr->base,
- addr->index, new_mem, am.new_op1,
- sign_extension, am.new_op2);
+ addr->index, new_mem, am.new_op2,
+ am.new_op1, sign_extension);
}
set_irn_pinned(new_node, get_irn_pinned(node));
*/
static ir_node *gen_Quot(ir_node *node)
{
- ir_node *op1 = get_Quot_left(node);
- ir_node *op2 = get_Quot_right(node);
+ ir_node *op1 = get_Quot_left(node);
+ ir_node *op2 = get_Quot_right(node);
if (ia32_cg_config.use_sse2) {
return gen_binop(node, op1, op2, new_rd_ia32_xDiv, match_am);
return new_node;
}
+/**
+ * Create a bt instruction for x & (1 << n) and place it into the block of cmp.
+ */
+static ir_node *gen_bt(ir_node *cmp, ir_node *x, ir_node *n) {
+ dbg_info *dbgi = get_irn_dbg_info(cmp);
+ ir_node *block = get_nodes_block(cmp);
+ ir_node *new_block = be_transform_node(block);
+ ir_node *op1 = be_transform_node(x);
+ ir_node *op2 = be_transform_node(n);
+
+ return new_rd_ia32_Bt(dbgi, current_ir_graph, new_block, op1, op2);
+}
+
+/**
+ * Transform a node returning a "flag" result.
+ *
+ * @param node the node to transform
+ * @param pnc_out the compare mode to use
+ */
static ir_node *get_flags_node(ir_node *node, pn_Cmp *pnc_out)
{
- ir_graph *irg = current_ir_graph;
ir_node *flags;
ir_node *new_op;
ir_node *noreg;
dbg_info *dbgi;
/* we have a Cmp as input */
- if(is_Proj(node)) {
+ if (is_Proj(node)) {
ir_node *pred = get_Proj_pred(node);
- if(is_Cmp(pred)) {
+ if (is_Cmp(pred)) {
+ pn_Cmp pnc = get_Proj_proj(node);
+ if (ia32_cg_config.use_bt && (pnc == pn_Cmp_Lg || pnc == pn_Cmp_Eq)) {
+ ir_node *l = get_Cmp_left(pred);
+ ir_node *r = get_Cmp_right(pred);
+ if (is_And(l)) {
+ ir_node *la = get_And_left(l);
+ ir_node *ra = get_And_right(l);
+ if (is_Shl(la)) {
+ ir_node *c = get_Shl_left(la);
+ if (is_Const_1(c) && (is_Const_0(r) || r == la)) {
+ /* (1 << n) & ra) */
+ ir_node *n = get_Shl_right(la);
+ flags = gen_bt(pred, ra, n);
+ /* we must generate a Jc/Jnc jump */
+ pnc = pnc == pn_Cmp_Lg ? pn_Cmp_Lt : pn_Cmp_Ge;
+ if (r == la)
+ pnc ^= pn_Cmp_Leg;
+ *pnc_out = ia32_pn_Cmp_unsigned | pnc;
+ return flags;
+ }
+ }
+ if (is_Shl(ra)) {
+ ir_node *c = get_Shl_left(ra);
+ if (is_Const_1(c) && (is_Const_0(r) || r == ra)) {
+ /* la & (1 << n)) */
+ ir_node *n = get_Shl_right(ra);
+ flags = gen_bt(pred, la, n);
+ /* we must generate a Jc/Jnc jump */
+ pnc = pnc == pn_Cmp_Lg ? pn_Cmp_Lt : pn_Cmp_Ge;
+ if (r == ra)
+ pnc ^= pn_Cmp_Leg;
+ *pnc_out = ia32_pn_Cmp_unsigned | pnc;
+ return flags;
+ }
+ }
+ }
+ }
flags = be_transform_node(pred);
- *pnc_out = get_Proj_proj(node);
+ *pnc_out = pnc;
return flags;
}
}
new_op = be_transform_node(node);
noreg = ia32_new_NoReg_gp(env_cg);
nomem = new_NoMem();
- flags = new_rd_ia32_Test(dbgi, irg, new_block, noreg, noreg, nomem,
- new_op, new_op, 0, 0);
+ flags = new_rd_ia32_Test(dbgi, current_ir_graph, new_block, noreg, noreg, nomem,
+ new_op, new_op, /*is_permuted=*/0, /*cmp_unsigned=*/0);
*pnc_out = pn_Cmp_Lg;
return flags;
}
}
/**
- * Transforms a Store.
+ * Transform a Store(floatConst).
*
* @return the created ia32 Store node
*/
-static ir_node *gen_Store(ir_node *node)
-{
+static ir_node *gen_float_const_Store(ir_node *node, ir_node *cns) {
+ ir_mode *mode = get_irn_mode(cns);
+ int size = get_mode_size_bits(mode);
+ tarval *tv = get_Const_tarval(cns);
ir_node *block = get_nodes_block(node);
ir_node *new_block = be_transform_node(block);
ir_node *ptr = get_Store_ptr(node);
- ir_node *val = get_Store_value(node);
ir_node *mem = get_Store_mem(node);
ir_graph *irg = current_ir_graph;
dbg_info *dbgi = get_irn_dbg_info(node);
ir_node *noreg = ia32_new_NoReg_gp(env_cg);
- ir_mode *mode = get_irn_mode(val);
- ir_node *new_val;
+ int ofs = 4;
ir_node *new_node;
ia32_address_t addr;
+ unsigned val = get_tarval_sub_bits(tv, 0) |
+ (get_tarval_sub_bits(tv, 1) << 8) |
+ (get_tarval_sub_bits(tv, 2) << 16) |
+ (get_tarval_sub_bits(tv, 3) << 24);
+ ir_node *imm = create_Immediate(NULL, 0, val);
+
+ /* construct store address */
+ memset(&addr, 0, sizeof(addr));
+ ia32_create_address_mode(&addr, ptr, /*force=*/0);
+
+ if (addr.base == NULL) {
+ addr.base = noreg;
+ } else {
+ addr.base = be_transform_node(addr.base);
+ }
+
+ if (addr.index == NULL) {
+ addr.index = noreg;
+ } else {
+ addr.index = be_transform_node(addr.index);
+ }
+ addr.mem = be_transform_node(mem);
+
+ new_node = new_rd_ia32_Store(dbgi, irg, new_block, addr.base,
+ addr.index, addr.mem, imm);
+
+ set_irn_pinned(new_node, get_irn_pinned(node));
+ set_ia32_op_type(new_node, ia32_AddrModeD);
+ set_ia32_ls_mode(new_node, mode_Iu);
+
+ set_address(new_node, &addr);
+
+ /** add more stores if needed */
+ while (size > 32) {
+ unsigned val = get_tarval_sub_bits(tv, ofs) |
+ (get_tarval_sub_bits(tv, ofs + 1) << 8) |
+ (get_tarval_sub_bits(tv, ofs + 2) << 16) |
+ (get_tarval_sub_bits(tv, ofs + 3) << 24);
+ ir_node *imm = create_Immediate(NULL, 0, val);
+
+ addr.offset += 4;
+ addr.mem = new_node;
+
+ new_node = new_rd_ia32_Store(dbgi, irg, new_block, addr.base,
+ addr.index, addr.mem, imm);
+
+ set_irn_pinned(new_node, get_irn_pinned(node));
+ set_ia32_op_type(new_node, ia32_AddrModeD);
+ set_ia32_ls_mode(new_node, mode_Iu);
+
+ set_address(new_node, &addr);
+ size -= 32;
+ ofs += 4;
+ }
+
+ SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node));
+ return new_node;
+}
+
+/**
+ * Generate a vfist or vfisttp instruction.
+ */
+static ir_node *gen_vfist(dbg_info *dbgi, ir_graph *irg, ir_node *block, ir_node *base, ir_node *index,
+ ir_node *mem, ir_node *val, ir_node **fist)
+{
+ ir_node *new_node;
+
+ if (ia32_cg_config.use_fisttp) {
+ /* Note: fisttp ALWAYS pop the tos. We have to ensure here that the value is copied
+ if other users exists */
+ const arch_register_class_t *reg_class = &ia32_reg_classes[CLASS_ia32_vfp];
+ ir_node *vfisttp = new_rd_ia32_vfisttp(dbgi, irg, block, base, index, mem, val);
+ ir_node *value = new_r_Proj(irg, block, vfisttp, mode_E, pn_ia32_vfisttp_res);
+ be_new_Keep(reg_class, irg, block, 1, &value);
+
+ new_node = new_r_Proj(irg, block, vfisttp, mode_M, pn_ia32_vfisttp_M);
+ *fist = vfisttp;
+ } else {
+ ir_node *trunc_mode = ia32_new_Fpu_truncate(env_cg);
+
+ /* do a fist */
+ new_node = new_rd_ia32_vfist(dbgi, irg, block, base, index, mem, val, trunc_mode);
+ *fist = new_node;
+ }
+ return new_node;
+}
+/**
+ * Transforms a normal Store.
+ *
+ * @return the created ia32 Store node
+ */
+static ir_node *gen_normal_Store(ir_node *node)
+{
+ ir_node *val = get_Store_value(node);
+ ir_mode *mode = get_irn_mode(val);
+ ir_node *block = get_nodes_block(node);
+ ir_node *new_block = be_transform_node(block);
+ ir_node *ptr = get_Store_ptr(node);
+ ir_node *mem = get_Store_mem(node);
+ ir_graph *irg = current_ir_graph;
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ ir_node *noreg = ia32_new_NoReg_gp(env_cg);
+ ir_node *new_val, *new_node, *store;
+ ia32_address_t addr;
+
/* check for destination address mode */
new_node = try_create_dest_am(node);
- if(new_node != NULL)
+ if (new_node != NULL)
return new_node;
/* construct store address */
memset(&addr, 0, sizeof(addr));
ia32_create_address_mode(&addr, ptr, /*force=*/0);
- if(addr.base == NULL) {
+ if (addr.base == NULL) {
addr.base = noreg;
} else {
addr.base = be_transform_node(addr.base);
}
- if(addr.index == NULL) {
+ if (addr.index == NULL) {
addr.index = noreg;
} else {
addr.index = be_transform_node(addr.index);
if (mode_is_float(mode)) {
/* convs (and strict-convs) before stores are unnecessary if the mode
is the same */
- while(is_Conv(val) && mode == get_irn_mode(get_Conv_op(val))) {
+ while (is_Conv(val) && mode == get_irn_mode(get_Conv_op(val))) {
val = get_Conv_op(val);
}
new_val = be_transform_node(val);
new_node = new_rd_ia32_vfst(dbgi, irg, new_block, addr.base,
addr.index, addr.mem, new_val, mode);
}
- } else if(is_float_to_int32_conv(val)) {
- ir_node *trunc_mode = ia32_new_Fpu_truncate(env_cg);
+ store = new_node;
+ } else if (is_float_to_int32_conv(val)) {
val = get_Conv_op(val);
/* convs (and strict-convs) before stores are unnecessary if the mode
while(is_Conv(val) && mode == get_irn_mode(get_Conv_op(val))) {
val = get_Conv_op(val);
}
- new_val = be_transform_node(val);
-
- new_node = new_rd_ia32_vfist(dbgi, irg, new_block, addr.base,
- addr.index, addr.mem, new_val, trunc_mode);
+ new_val = be_transform_node(val);
+ new_node = gen_vfist(dbgi, irg, new_block, addr.base, addr.index, addr.mem, new_val, &store);
} else {
new_val = create_immediate_or_transform(val, 0);
assert(mode != mode_b);
new_node = new_rd_ia32_Store(dbgi, irg, new_block, addr.base,
addr.index, addr.mem, new_val);
}
+ store = new_node;
}
- set_irn_pinned(new_node, get_irn_pinned(node));
- set_ia32_op_type(new_node, ia32_AddrModeD);
- set_ia32_ls_mode(new_node, mode);
+ set_irn_pinned(store, get_irn_pinned(node));
+ set_ia32_op_type(store, ia32_AddrModeD);
+ set_ia32_ls_mode(store, mode);
- set_address(new_node, &addr);
- SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node));
+ set_address(store, &addr);
+ SET_IA32_ORIG_NODE(store, ia32_get_old_node_name(env_cg, node));
return new_node;
}
+/**
+ * Transforms a Store.
+ *
+ * @return the created ia32 Store node
+ */
+static ir_node *gen_Store(ir_node *node)
+{
+ ir_node *val = get_Store_value(node);
+ ir_mode *mode = get_irn_mode(val);
+
+ if (mode_is_float(mode) && is_Const(val)) {
+ int transform = 1;
+
+ /* we are storing a floating point constant */
+ if (ia32_cg_config.use_sse2) {
+ transform = !is_simple_sse_Const(val);
+ } else {
+ transform = !is_simple_x87_Const(val);
+ }
+ if (transform)
+ return gen_float_const_Store(node, val);
+ }
+ return gen_normal_Store(node);
+}
+
+/**
+ * Transforms a Switch.
+ *
+ * @return the created ia32 SwitchJmp node
+ */
static ir_node *create_Switch(ir_node *node)
{
ir_graph *irg = current_ir_graph;
return new_node;
}
+/**
+ * Transform a Cond node.
+ */
static ir_node *gen_Cond(ir_node *node) {
ir_node *block = get_nodes_block(node);
ir_node *new_block = be_transform_node(block);
return create_Switch(node);
}
- /* we get flags from a cmp */
+ /* we get flags from a Cmp */
flags = get_flags_node(sel, &pnc);
new_node = new_rd_ia32_Jcc(dbgi, irg, new_block, flags, pnc);
return new_node;
}
-
-
/**
* Transforms a CopyB node.
*
return 1;
}
+/**
+ * Generate code for a Cmp.
+ */
static ir_node *gen_Cmp(ir_node *node)
{
ir_graph *irg = current_ir_graph;
return new_node;
}
-
-
+/**
+ * Creates a ia32 Setcc instruction.
+ */
static ir_node *create_set_32bit(dbg_info *dbgi, ir_node *new_block,
ir_node *flags, pn_Cmp pnc, ir_node *orig_node,
int ins_permuted)
SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, orig_node));
/* we might need to conv the result up */
- if(get_mode_size_bits(mode) > 8) {
+ if (get_mode_size_bits(mode) > 8) {
new_node = new_rd_ia32_Conv_I2I8Bit(dbgi, irg, new_block, noreg, noreg,
nomem, new_node, mode_Bu);
SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, orig_node));
return new_node;
}
+/**
+ * Create instruction for an unsigned Difference or Zero.
+ */
+static ir_node *create_Doz(ir_node *psi, ir_node *a, ir_node *b) {
+ ir_graph *irg = current_ir_graph;
+ ir_mode *mode = get_irn_mode(psi);
+ ir_node *new_node, *sub, *sbb, *eflags, *block, *noreg, *tmpreg, *nomem;
+ dbg_info *dbgi;
+
+ new_node = gen_binop(psi, a, b, new_rd_ia32_Sub,
+ match_mode_neutral | match_am | match_immediate | match_two_users);
+
+ block = get_nodes_block(new_node);
+
+ if (is_Proj(new_node)) {
+ sub = get_Proj_pred(new_node);
+ assert(is_ia32_Sub(sub));
+ } else {
+ sub = new_node;
+ set_irn_mode(sub, mode_T);
+ new_node = new_rd_Proj(NULL, irg, block, sub, mode, pn_ia32_res);
+ }
+ eflags = new_rd_Proj(NULL, irg, block, sub, mode_Iu, pn_ia32_Sub_flags);
+
+ dbgi = get_irn_dbg_info(psi);
+ noreg = ia32_new_NoReg_gp(env_cg);
+ tmpreg = new_rd_ia32_ProduceVal(dbgi, irg, block);
+ nomem = new_NoMem();
+ sbb = new_rd_ia32_Sbb(dbgi, irg, block, noreg, noreg, nomem, tmpreg, tmpreg, eflags);
+
+ new_node = new_rd_ia32_And(dbgi, irg, block, noreg, noreg, nomem, new_node, sbb);
+ set_ia32_commutative(new_node);
+ return new_node;
+}
+
/**
* Transforms a Psi node into CMov.
*
ir_node *psi_true = get_Psi_val(node, 0);
ir_node *psi_default = get_Psi_default(node);
ir_node *cond = get_Psi_cond(node, 0);
- ir_node *flags = NULL;
- ir_node *new_node;
- pn_Cmp pnc;
+ ir_mode *mode = get_irn_mode(node);
+ ir_node *cmp = get_Proj_pred(cond);
+ ir_node *cmp_left = get_Cmp_left(cmp);
+ ir_node *cmp_right = get_Cmp_right(cmp);
+ pn_Cmp pnc = get_Proj_proj(cond);
assert(get_Psi_n_conds(node) == 1);
assert(get_irn_mode(cond) == mode_b);
- assert(mode_needs_gp_reg(get_irn_mode(node)));
- flags = get_flags_node(cond, &pnc);
+ /* Note: a Psi node uses a Load two times IFF it's used in the compare AND in the result */
+ if (mode_is_float(mode)) {
+ if (ia32_cg_config.use_sse2) {
+ if (pnc == pn_Cmp_Lt || pnc == pn_Cmp_Le) {
+ if (cmp_left == psi_true && cmp_right == psi_default) {
+ /* psi(a <= b, a, b) => MIN */
+ return gen_binop(node, cmp_left, cmp_right, new_rd_ia32_xMin,
+ match_commutative | match_am | match_two_users);
+ } else if (cmp_left == psi_default && cmp_right == psi_true) {
+ /* psi(a <= b, b, a) => MAX */
+ return gen_binop(node, cmp_left, cmp_right, new_rd_ia32_xMax,
+ match_commutative | match_am | match_two_users);
+ }
+ } else if (pnc == pn_Cmp_Gt || pnc == pn_Cmp_Ge) {
+ if (cmp_left == psi_true && cmp_right == psi_default) {
+ /* psi(a >= b, a, b) => MAX */
+ return gen_binop(node, cmp_left, cmp_right, new_rd_ia32_xMax,
+ match_commutative | match_am | match_two_users);
+ } else if (cmp_left == psi_default && cmp_right == psi_true) {
+ /* psi(a >= b, b, a) => MIN */
+ return gen_binop(node, cmp_left, cmp_right, new_rd_ia32_xMin,
+ match_commutative | match_am | match_two_users);
+ }
+ }
+ }
+ panic("cannot transform floating point Psi");
- if(is_Const_1(psi_true) && is_Const_0(psi_default)) {
- new_node = create_set_32bit(dbgi, new_block, flags, pnc, node, 0);
- } else if(is_Const_0(psi_true) && is_Const_1(psi_default)) {
- new_node = create_set_32bit(dbgi, new_block, flags, pnc, node, 1);
} else {
- new_node = create_CMov(node, cond, flags, pnc);
+ ir_node *flags;
+ ir_node *new_node;
+
+ assert(mode_needs_gp_reg(mode));
+
+ /* check for unsigned Doz first */
+ if ((pnc & pn_Cmp_Gt) && !mode_is_signed(mode) &&
+ is_Const_0(psi_default) && is_Sub(psi_true) &&
+ get_Sub_left(psi_true) == cmp_left && get_Sub_right(psi_true) == cmp_right) {
+ /* Psi(a >=u b, a - b, 0) unsigned Doz */
+ return create_Doz(node, cmp_left, cmp_right);
+ } else if ((pnc & pn_Cmp_Lt) && !mode_is_signed(mode) &&
+ is_Const_0(psi_true) && is_Sub(psi_default) &&
+ get_Sub_left(psi_default) == cmp_left && get_Sub_right(psi_default) == cmp_right) {
+ /* Psi(a <=u b, 0, a - b) unsigned Doz */
+ return create_Doz(node, cmp_left, cmp_right);
+ }
+
+ flags = get_flags_node(cond, &pnc);
+
+ if (is_Const(psi_true) && is_Const(psi_default)) {
+ /* both are const, good */
+ if (is_Const_1(psi_true) && is_Const_0(psi_default)) {
+ new_node = create_set_32bit(dbgi, new_block, flags, pnc, node, /*is_premuted=*/0);
+ } else if (is_Const_0(psi_true) && is_Const_1(psi_default)) {
+ new_node = create_set_32bit(dbgi, new_block, flags, pnc, node, /*is_premuted=*/1);
+ } else {
+ /* Not that simple. */
+ goto need_cmov;
+ }
+ } else {
+need_cmov:
+ new_node = create_CMov(node, cond, flags, pnc);
+ }
+ return new_node;
}
- return new_node;
}
ir_graph *irg = current_ir_graph;
dbg_info *dbgi = get_irn_dbg_info(node);
ir_node *noreg = ia32_new_NoReg_gp(cg);
- ir_node *trunc_mode = ia32_new_Fpu_truncate(cg);
ir_mode *mode = get_irn_mode(node);
- ir_node *fist, *load;
-
- /* do a fist */
- fist = new_rd_ia32_vfist(dbgi, irg, block, get_irg_frame(irg), noreg,
- new_NoMem(), new_op, trunc_mode);
+ ir_node *fist, *load, *mem;
+ mem = gen_vfist(dbgi, irg, block, get_irg_frame(irg), noreg, new_NoMem(), new_op, &fist);
set_irn_pinned(fist, op_pin_state_floats);
set_ia32_use_frame(fist);
set_ia32_op_type(fist, ia32_AddrModeD);
SET_IA32_ORIG_NODE(fist, ia32_get_old_node_name(cg, node));
/* do a Load */
- load = new_rd_ia32_Load(dbgi, irg, block, get_irg_frame(irg), noreg, fist);
+ load = new_rd_ia32_Load(dbgi, irg, block, get_irg_frame(irg), noreg, mem);
set_irn_pinned(load, op_pin_state_floats);
set_ia32_use_frame(load);
return new_node;
}
-static ir_node *create_Immediate(ir_entity *symconst, int symconst_sign, long val)
-{
- ir_graph *irg = current_ir_graph;
- 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]);
-
- return immediate;
-}
-
/**
* Create a conversion from general purpose to x87 register
*/
ir_node *res = NULL;
if (src_mode == mode_b) {
- assert(mode_is_int(tgt_mode));
+ assert(mode_is_int(tgt_mode) || mode_is_reference(tgt_mode));
/* nothing to do, we already model bools as 0/1 ints */
return be_transform_node(op);
}
/* a memory constraint: no need to do anything in backend about it
* (the dependencies are already respected by the memory edge of
* the node) */
- constraint->req = &no_register_req;
+ constraint->req = &no_register_req;
return;
}
}
static void parse_clobber(ir_node *node, int pos, constraint_t *constraint,
- const char *c)
+ const char *clobber)
{
- (void) node;
+ ir_graph *irg = get_irn_irg(node);
+ struct obstack *obst = get_irg_obstack(irg);
+ const arch_register_t *reg = NULL;
+ int c;
+ size_t r;
+ arch_register_req_t *req;
+ const arch_register_class_t *cls;
+ unsigned *limited;
+
(void) pos;
- (void) constraint;
- (void) c;
- panic("Clobbers not supported yet");
+
+ /* TODO: construct a hashmap instead of doing linear search for clobber
+ * register */
+ for(c = 0; c < N_CLASSES; ++c) {
+ cls = & ia32_reg_classes[c];
+ for(r = 0; r < cls->n_regs; ++r) {
+ const arch_register_t *temp_reg = arch_register_for_index(cls, r);
+ if(strcmp(temp_reg->name, clobber) == 0
+ || (c == CLASS_ia32_gp && strcmp(temp_reg->name+1, clobber) == 0)) {
+ reg = temp_reg;
+ break;
+ }
+ }
+ if(reg != NULL)
+ break;
+ }
+ if(reg == NULL) {
+ panic("Register '%s' mentioned in asm clobber is unknown\n", clobber);
+ return;
+ }
+
+ assert(reg->index < 32);
+
+ limited = obstack_alloc(obst, sizeof(limited[0]));
+ *limited = 1 << reg->index;
+
+ req = obstack_alloc(obst, sizeof(req[0]));
+ memset(req, 0, sizeof(req[0]));
+ req->type = arch_register_req_type_limited;
+ req->cls = cls;
+ req->limited = limited;
+
+ constraint->req = req;
+ constraint->immediate_possible = 0;
+ constraint->immediate_type = 0;
}
static int is_memory_op(const ir_asm_constraint *constraint)
n_out_constraints = get_ASM_n_output_constraints(node);
n_clobbers = get_ASM_n_clobbers(node);
out_arity = n_out_constraints + n_clobbers;
+ /* hack to keep space for mem proj */
+ if(n_clobbers > 0)
+ out_arity += 1;
in_constraints = get_ASM_input_constraints(node);
out_constraints = get_ASM_output_constraints(node);
if(constraint->pos > reg_map_size)
reg_map_size = constraint->pos;
- } else {
+
+ out_reg_reqs[i] = parsed_constraint.req;
+ } else if(i < out_arity - 1) {
ident *glob_id = clobbers [i - n_out_constraints];
+ assert(glob_id != NULL);
c = get_id_str(glob_id);
parse_clobber(node, i, &parsed_constraint, c);
- }
- out_reg_reqs[i] = parsed_constraint.req;
+ out_reg_reqs[i+1] = parsed_constraint.req;
+ }
}
+ if(n_clobbers > 1)
+ out_reg_reqs[n_out_constraints] = &no_register_req;
/* construct input constraints */
in_reg_reqs = obstack_alloc(obst, arity * sizeof(in_reg_reqs[0]));
if (ia32_cg_config.use_sse2) {
return ia32_new_Unknown_xmm(env_cg);
} else {
- /* Unknown nodes are buggy in x87 sim, use zero for now... */
+ /* Unknown nodes are buggy in x87 simulator, use zero for now... */
ir_graph *irg = current_ir_graph;
dbg_info *dbgi = get_irn_dbg_info(node);
ir_node *block = get_irg_start_block(irg);
- return new_rd_ia32_vfldz(dbgi, irg, block);
+ 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));
+ return ret;
}
} else if (mode_needs_gp_reg(mode)) {
return ia32_new_Unknown_gp(env_cg);
dbg_info *dbgi = get_irn_dbg_info(node);
ir_node *noreg = ia32_new_NoReg_gp(env_cg);
ir_mode *mode = get_ia32_ls_mode(node);
- ir_node *trunc_mode = ia32_new_Fpu_truncate(env_cg);
- ir_node *new_op;
+ ir_node *memres, *fist;
long am_offs;
- new_op = new_rd_ia32_vfist(dbgi, irg, block, new_ptr, noreg, new_mem,
- new_val, trunc_mode);
-
+ memres = gen_vfist(dbgi, irg, block, new_ptr, noreg, new_mem, new_val, &fist);
am_offs = get_ia32_am_offs_int(node);
- add_ia32_am_offs_int(new_op, am_offs);
+ add_ia32_am_offs_int(fist, am_offs);
- set_ia32_op_type(new_op, ia32_AddrModeD);
- set_ia32_ls_mode(new_op, mode);
- set_ia32_frame_ent(new_op, get_ia32_frame_ent(node));
- set_ia32_use_frame(new_op);
+ set_ia32_op_type(fist, ia32_AddrModeD);
+ set_ia32_ls_mode(fist, mode);
+ set_ia32_frame_ent(fist, get_ia32_frame_ent(node));
+ set_ia32_use_frame(fist);
- SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
+ SET_IA32_ORIG_NODE(fist, ia32_get_old_node_name(env_cg, node));
- return new_op;
+ return memres;
}
/**
* @return the created ia32 IMul1OP node
*/
static ir_node *gen_ia32_l_IMul(ir_node *node) {
- ir_node *left = get_binop_left(node);
- ir_node *right = get_binop_right(node);
+ ir_node *left = get_binop_left(node);
+ ir_node *right = get_binop_right(node);
return gen_binop(node, left, right, new_rd_ia32_IMul1OP,
match_commutative | match_am | match_mode_neutral);
}
static ir_node *gen_ia32_l_Sub(ir_node *node) {
- ir_node *left = get_irn_n(node, n_ia32_l_Sub_left);
- ir_node *right = get_irn_n(node, n_ia32_l_Sub_right);
+ ir_node *left = get_irn_n(node, n_ia32_l_Sub_minuend);
+ ir_node *right = get_irn_n(node, n_ia32_l_Sub_subtrahend);
ir_node *lowered = gen_binop(node, left, right, new_rd_ia32_Sub,
match_am | match_immediate | match_mode_neutral);
}
static ir_node *gen_ia32_l_FloattoLL(ir_node *node) {
- (void) node;
- panic("LLtoFloat NIY");
+ ir_node *src_block = get_nodes_block(node);
+ ir_node *block = be_transform_node(src_block);
+ ir_graph *irg = current_ir_graph;
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ ir_node *frame = get_irg_frame(irg);
+ ir_node *noreg = ia32_new_NoReg_gp(env_cg);
+ ir_node *nomem = new_NoMem();
+ ir_node *val = get_irn_n(node, n_ia32_l_FloattoLL_val);
+ ir_node *new_val = be_transform_node(val);
+ ir_node *fist, *mem;
+
+ mem = gen_vfist(dbgi, irg, block, frame, noreg, nomem, new_val, &fist);
+ SET_IA32_ORIG_NODE(fist, ia32_get_old_node_name(env_cg, node));
+ set_ia32_use_frame(fist);
+ set_ia32_op_type(fist, ia32_AddrModeD);
+ set_ia32_ls_mode(fist, mode_Ls);
+
+ return mem;
}
/**
return NULL;
}
+static ir_node *gen_Proj_l_FloattoLL(ir_node *node) {
+ ir_graph *irg = current_ir_graph;
+ ir_node *block = be_transform_node(get_nodes_block(node));
+ ir_node *pred = get_Proj_pred(node);
+ ir_node *new_pred = be_transform_node(pred);
+ ir_node *frame = get_irg_frame(irg);
+ ir_node *noreg = ia32_new_NoReg_gp(env_cg);
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ long pn = get_Proj_proj(node);
+ ir_node *load;
+ ir_node *proj;
+ ia32_attr_t *attr;
+
+ load = new_rd_ia32_Load(dbgi, irg, block, frame, noreg, new_pred);
+ SET_IA32_ORIG_NODE(load, ia32_get_old_node_name(env_cg, node));
+ set_ia32_use_frame(load);
+ set_ia32_op_type(load, ia32_AddrModeS);
+ set_ia32_ls_mode(load, mode_Iu);
+ /* we need a 64bit stackslot (fist stores 64bit) even though we only load
+ * 32 bit from it with this particular load */
+ attr = get_ia32_attr(load);
+ attr->data.need_64bit_stackent = 1;
+
+ if (pn == pn_ia32_l_FloattoLL_res_high) {
+ add_ia32_am_offs_int(load, 4);
+ } else {
+ assert(pn == pn_ia32_l_FloattoLL_res_low);
+ }
+
+ proj = new_r_Proj(irg, block, load, mode_Iu, pn_ia32_Load_res);
+
+ return proj;
+}
+
/**
* Transform the Projs of an AddSP.
*/
return new_rd_Proj(dbgi, irg, block, new_pred, mode_vfp, pn_ia32_vfdiv_res);
}
break;
+ case pn_Quot_X_regular:
+ case pn_Quot_X_except:
default:
break;
}
* Transform and potentially renumber Proj nodes.
*/
static ir_node *gen_Proj(ir_node *node) {
- ir_graph *irg = current_ir_graph;
- dbg_info *dbgi = get_irn_dbg_info(node);
ir_node *pred = get_Proj_pred(node);
- long proj = get_Proj_proj(node);
-
if (is_Store(pred)) {
+ long proj = get_Proj_proj(node);
if (proj == pn_Store_M) {
return be_transform_node(pred);
} else {
assert(0);
- return new_r_Bad(irg);
+ return new_r_Bad(current_ir_graph);
}
} else if (is_Load(pred)) {
return gen_Proj_Load(node);
} else if (is_Cmp(pred)) {
return gen_Proj_Cmp(node);
} else if (get_irn_op(pred) == op_Start) {
+ long proj = get_Proj_proj(node);
if (proj == pn_Start_X_initial_exec) {
ir_node *block = get_nodes_block(pred);
+ dbg_info *dbgi = get_irn_dbg_info(node);
ir_node *jump;
/* we exchange the ProjX with a jump */
block = be_transform_node(block);
- jump = new_rd_Jmp(dbgi, irg, block);
+ jump = new_rd_Jmp(dbgi, current_ir_graph, block);
return jump;
}
if (node == be_get_old_anchor(anchor_tls)) {
return gen_Proj_tls(node);
}
+ } else if (is_ia32_l_FloattoLL(pred)) {
+ return gen_Proj_l_FloattoLL(node);
#ifdef FIRM_EXT_GRS
} else if(!is_ia32_irn(pred)) { // Quick hack for SIMD optimization
#else
} else {
#endif
- ir_node *new_pred = be_transform_node(pred);
- ir_node *block = be_transform_node(get_nodes_block(node));
- ir_mode *mode = get_irn_mode(node);
+ ir_mode *mode = get_irn_mode(node);
if (mode_needs_gp_reg(mode)) {
- ir_node *new_proj = new_r_Proj(irg, block, new_pred, mode_Iu,
- get_Proj_proj(node));
+ ir_node *new_pred = be_transform_node(pred);
+ ir_node *block = be_transform_node(get_nodes_block(node));
+ ir_node *new_proj = new_r_Proj(current_ir_graph, block, new_pred,
+ mode_Iu, get_Proj_proj(node));
#ifdef DEBUG_libfirm
new_proj->node_nr = node->node_nr;
#endif