/*
- * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
+ * Copyright (C) 1995-2011 University of Karlsruhe. All right reserved.
*
* This file is part of libFirm.
*
* @brief This file implements the IR transformation from firm into
* ia32-Firm.
* @author Christian Wuerdig, Matthias Braun
- * @version $Id$
*/
#include "config.h"
#include "irprog_t.h"
#include "iredges_t.h"
#include "irgmod.h"
-#include "irvrfy.h"
#include "ircons.h"
#include "irgwalk.h"
#include "irprintf.h"
#include "debug.h"
#include "irdom.h"
+#include "iropt.h"
#include "error.h"
#include "array_t.h"
-#include "height.h"
+#include "heights.h"
-#include "../benode.h"
-#include "../besched.h"
-#include "../beabi.h"
-#include "../beutil.h"
-#include "../beirg.h"
-#include "../betranshlp.h"
-#include "../be_t.h"
+#include "benode.h"
+#include "besched.h"
+#include "beabi.h"
+#include "beutil.h"
+#include "beirg.h"
+#include "betranshlp.h"
+#include "be_t.h"
#include "bearch_ia32_t.h"
#include "ia32_common_transform.h"
#include "ia32_nodes_attr.h"
#include "ia32_transform.h"
#include "ia32_new_nodes.h"
-#include "ia32_map_regs.h"
#include "ia32_dbg_stat.h"
#include "ia32_optimize.h"
-#include "ia32_util.h"
#include "ia32_address_mode.h"
#include "ia32_architecture.h"
/* define this to construct SSE constants instead of load them */
#undef CONSTRUCT_SSE_CONST
-
-#define SFP_SIGN "0x80000000"
-#define DFP_SIGN "0x8000000000000000"
-#define SFP_ABS "0x7FFFFFFF"
-#define DFP_ABS "0x7FFFFFFFFFFFFFFF"
-#define DFP_INTMAX "9223372036854775807"
-#define ULL_BIAS "18446744073709551616"
-
-#define ENT_SFP_SIGN "C_ia32_sfp_sign"
-#define ENT_DFP_SIGN "C_ia32_dfp_sign"
-#define ENT_SFP_ABS "C_ia32_sfp_abs"
-#define ENT_DFP_ABS "C_ia32_dfp_abs"
-#define ENT_ULL_BIAS "C_ia32_ull_bias"
-
-#define mode_vfp (ia32_reg_classes[CLASS_ia32_vfp].mode)
+#define mode_vfp (ia32_reg_classes[CLASS_ia32_vfp].mode)
#define mode_xmm (ia32_reg_classes[CLASS_ia32_xmm].mode)
DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
+static ir_node *old_initial_fpcw = NULL;
static ir_node *initial_fpcw = NULL;
-int no_pic_adjust;
+int ia32_no_pic_adjust;
typedef ir_node *construct_binop_func(dbg_info *db, ir_node *block,
ir_node *base, ir_node *index, ir_node *mem, ir_node *op1,
*/
static bool is_simple_x87_Const(ir_node *node)
{
- tarval *tv = get_Const_tarval(node);
+ ir_tarval *tv = get_Const_tarval(node);
if (tarval_is_null(tv) || tarval_is_one(tv))
return true;
*/
static bool is_simple_sse_Const(ir_node *node)
{
- tarval *tv = get_Const_tarval(node);
- ir_mode *mode = get_tarval_mode(tv);
+ ir_tarval *tv = get_Const_tarval(node);
+ ir_mode *mode = get_tarval_mode(tv);
if (mode == mode_F)
return true;
*/
static ir_node *get_symconst_base(void)
{
- if (be_get_irg_options(env_cg->irg)->pic) {
- return arch_code_generator_get_pic_base(env_cg);
+ ir_graph *irg = current_ir_graph;
+
+ if (be_options.pic) {
+ const arch_env_t *arch_env = be_get_irg_arch_env(irg);
+ return arch_env->impl->get_pic_base(irg);
}
return noreg_GP;
*/
static ir_node *gen_Const(ir_node *node)
{
- ir_node *old_block = get_nodes_block(node);
- ir_node *block = be_transform_node(old_block);
- dbg_info *dbgi = get_irn_dbg_info(node);
- ir_mode *mode = get_irn_mode(node);
+ ir_node *old_block = get_nodes_block(node);
+ ir_node *block = be_transform_node(old_block);
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ ir_mode *mode = get_irn_mode(node);
+ ir_tarval *tv = get_Const_tarval(node);
assert(is_Const(node));
if (mode_is_float(mode)) {
- ir_node *res = NULL;
- ir_node *load;
- ir_node *base;
- ir_entity *floatent;
+ ir_graph *irg = get_irn_irg(node);
+ const arch_env_t *arch_env = be_get_irg_arch_env(irg);
+ ia32_isa_t *isa = (ia32_isa_t*) arch_env;
+ ir_node *res = NULL;
+ ir_node *load;
+ ir_entity *floatent;
if (ia32_cg_config.use_sse2) {
- tarval *tv = get_Const_tarval(node);
if (tarval_is_null(tv)) {
load = new_bd_ia32_xZero(dbgi, block);
set_ia32_ls_mode(load, mode);
set_ia32_ls_mode(load, mode);
res = load;
} else {
+ ir_node *base;
#ifdef CONSTRUCT_SSE_CONST
if (mode == mode_D) {
unsigned val = get_tarval_sub_bits(tv, 0) |
}
}
#endif /* CONSTRUCT_SSE_CONST */
- floatent = create_float_const_entity(node);
+ floatent = ia32_create_float_const_entity(isa, tv, NULL);
base = get_symconst_base();
load = new_bd_ia32_xLoad(dbgi, block, base, noreg_GP, nomem,
mode);
set_ia32_op_type(load, ia32_AddrModeS);
set_ia32_am_sc(load, floatent);
- arch_irn_add_flags(load, arch_irn_flags_rematerializable);
+ arch_add_irn_flags(load, arch_irn_flags_rematerializable);
res = new_r_Proj(load, mode_xmm, pn_ia32_xLoad_res);
}
} else {
- if (is_Const_null(node)) {
+ if (tarval_is_null(tv)) {
load = new_bd_ia32_vfldz(dbgi, block);
res = load;
set_ia32_ls_mode(load, mode);
- } else if (is_Const_one(node)) {
+ } else if (tarval_is_one(tv)) {
load = new_bd_ia32_vfld1(dbgi, block);
res = load;
set_ia32_ls_mode(load, mode);
ir_mode *ls_mode;
ir_node *base;
- floatent = create_float_const_entity(node);
+ floatent = ia32_create_float_const_entity(isa, tv, NULL);
/* create_float_const_ent is smart and sometimes creates
smaller entities */
ls_mode = get_type_mode(get_entity_type(floatent));
ls_mode);
set_ia32_op_type(load, ia32_AddrModeS);
set_ia32_am_sc(load, floatent);
- arch_irn_add_flags(load, arch_irn_flags_rematerializable);
+ arch_add_irn_flags(load, arch_irn_flags_rematerializable);
res = new_r_Proj(load, mode_vfp, pn_ia32_vfld_res);
}
}
end:
#endif /* CONSTRUCT_SSE_CONST */
SET_IA32_ORIG_NODE(load, node);
-
- be_dep_on_frame(load);
return res;
} else { /* non-float mode */
ir_node *cnst;
- tarval *tv = get_Const_tarval(node);
long val;
tv = tarval_convert_to(tv, mode_Iu);
cnst = new_bd_ia32_Const(dbgi, block, NULL, 0, 0, val);
SET_IA32_ORIG_NODE(cnst, node);
- be_dep_on_frame(cnst);
return cnst;
}
}
if (mode_is_float(mode)) {
if (ia32_cg_config.use_sse2)
- cnst = new_bd_ia32_xLoad(dbgi, block, noreg_GP, noreg_GP, nomem, mode_E);
+ cnst = new_bd_ia32_xLoad(dbgi, block, noreg_GP, noreg_GP, nomem, mode_D);
else
- cnst = new_bd_ia32_vfld(dbgi, block, noreg_GP, noreg_GP, nomem, mode_E);
+ cnst = new_bd_ia32_vfld(dbgi, block, noreg_GP, noreg_GP, nomem, ia32_mode_E);
set_ia32_am_sc(cnst, get_SymConst_entity(node));
set_ia32_use_frame(cnst);
} else {
panic("backend only support symconst_addr_ent (at %+F)", node);
}
entity = get_SymConst_entity(node);
- cnst = new_bd_ia32_Const(dbgi, block, entity, 0, 0, 0);
+ if (get_entity_owner(entity) == get_tls_type()) {
+ ir_node *tls_base = new_bd_ia32_LdTls(NULL, block);
+ ir_node *lea = new_bd_ia32_Lea(dbgi, block, tls_base, noreg_GP);
+ set_ia32_am_sc(lea, entity);
+ cnst = lea;
+ } else {
+ cnst = new_bd_ia32_Const(dbgi, block, entity, 0, 0, 0);
+ }
}
SET_IA32_ORIG_NODE(cnst, node);
- be_dep_on_frame(cnst);
return cnst;
}
-/**
- * Create a float type for the given mode and cache it.
- *
- * @param mode the mode for the float type (might be integer mode for SSE2 types)
- * @param align alignment
- */
-static ir_type *ia32_create_float_type(ir_mode *mode, unsigned align)
+static ir_type *make_array_type(ir_type *tp)
{
- ir_type *tp;
-
- assert(align <= 16);
-
- if (mode == mode_Iu) {
- static ir_type *int_Iu[16] = {NULL, };
-
- if (int_Iu[align] == NULL) {
- int_Iu[align] = tp = new_type_primitive(mode);
- /* set the specified alignment */
- set_type_alignment_bytes(tp, align);
- }
- return int_Iu[align];
- } else if (mode == mode_Lu) {
- static ir_type *int_Lu[16] = {NULL, };
-
- if (int_Lu[align] == NULL) {
- int_Lu[align] = tp = new_type_primitive(mode);
- /* set the specified alignment */
- set_type_alignment_bytes(tp, align);
- }
- return int_Lu[align];
- } else if (mode == mode_F) {
- static ir_type *float_F[16] = {NULL, };
-
- if (float_F[align] == NULL) {
- float_F[align] = tp = new_type_primitive(mode);
- /* set the specified alignment */
- set_type_alignment_bytes(tp, align);
- }
- return float_F[align];
- } else if (mode == mode_D) {
- static ir_type *float_D[16] = {NULL, };
-
- if (float_D[align] == NULL) {
- float_D[align] = tp = new_type_primitive(mode);
- /* set the specified alignment */
- set_type_alignment_bytes(tp, align);
- }
- return float_D[align];
- } else {
- static ir_type *float_E[16] = {NULL, };
-
- if (float_E[align] == NULL) {
- float_E[align] = tp = new_type_primitive(mode);
- /* set the specified alignment */
- set_type_alignment_bytes(tp, align);
- }
- return float_E[align];
- }
+ unsigned alignment = get_type_alignment_bytes(tp);
+ unsigned size = get_type_size_bytes(tp);
+ ir_type *res = new_type_array(1, tp);
+ set_type_alignment_bytes(res, alignment);
+ set_array_bounds_int(res, 0, 0, 2);
+ if (alignment > size)
+ size = alignment;
+ set_type_size_bytes(res, 2 * size);
+ set_type_state(res, layout_fixed);
+ return res;
}
/**
static ir_type *ia32_create_float_array(ir_type *tp)
{
ir_mode *mode = get_type_mode(tp);
- unsigned align = get_type_alignment_bytes(tp);
ir_type *arr;
- assert(align <= 16);
-
if (mode == mode_F) {
- static ir_type *float_F[16] = {NULL, };
+ static ir_type *float_F;
- if (float_F[align] != NULL)
- return float_F[align];
- arr = float_F[align] = new_type_array(1, tp);
+ arr = float_F;
+ if (arr == NULL)
+ arr = float_F = make_array_type(tp);
} else if (mode == mode_D) {
- static ir_type *float_D[16] = {NULL, };
+ static ir_type *float_D;
- if (float_D[align] != NULL)
- return float_D[align];
- arr = float_D[align] = new_type_array(1, tp);
+ arr = float_D;
+ if (arr == NULL)
+ arr = float_D = make_array_type(tp);
} else {
- static ir_type *float_E[16] = {NULL, };
+ static ir_type *float_E;
- if (float_E[align] != NULL)
- return float_E[align];
- arr = float_E[align] = new_type_array(1, tp);
+ arr = float_E;
+ if (arr == NULL)
+ arr = float_E = make_array_type(tp);
}
- set_type_alignment_bytes(arr, align);
- set_type_size_bytes(arr, 2 * get_type_size_bytes(tp));
- set_type_state(arr, layout_fixed);
return arr;
}
ir_entity *ia32_gen_fp_known_const(ia32_known_const_t kct)
{
static const struct {
- const char *ent_name;
+ const char *name;
const char *cnst_str;
- char mode;
- unsigned char align;
+ char mode;
} names [ia32_known_const_max] = {
- { ENT_SFP_SIGN, SFP_SIGN, 0, 16 }, /* ia32_SSIGN */
- { ENT_DFP_SIGN, DFP_SIGN, 1, 16 }, /* ia32_DSIGN */
- { ENT_SFP_ABS, SFP_ABS, 0, 16 }, /* ia32_SABS */
- { ENT_DFP_ABS, DFP_ABS, 1, 16 }, /* ia32_DABS */
- { ENT_ULL_BIAS, ULL_BIAS, 2, 4 } /* ia32_ULLBIAS */
+ { "C_sfp_sign", "0x80000000", 0 },
+ { "C_dfp_sign", "0x8000000000000000", 1 },
+ { "C_sfp_abs", "0x7FFFFFFF", 0 },
+ { "C_dfp_abs", "0x7FFFFFFFFFFFFFFF", 1 },
+ { "C_ull_bias", "0x10000000000000000", 2 }
};
static ir_entity *ent_cache[ia32_known_const_max];
- const char *ent_name, *cnst_str;
- ir_type *tp;
- ir_entity *ent;
- tarval *tv;
- ir_mode *mode;
-
- ent_name = names[kct].ent_name;
- if (! ent_cache[kct]) {
- cnst_str = names[kct].cnst_str;
+ ir_entity *ent = ent_cache[kct];
+ if (ent == NULL) {
+ ir_graph *irg = current_ir_graph;
+ const arch_env_t *arch_env = be_get_irg_arch_env(irg);
+ ia32_isa_t *isa = (ia32_isa_t*) arch_env;
+ const char *cnst_str = names[kct].cnst_str;
+ ident *name = new_id_from_str(names[kct].name);
+ ir_mode *mode;
+ ir_tarval *tv;
switch (names[kct].mode) {
case 0: mode = mode_Iu; break;
case 1: mode = mode_Lu; break;
- default: mode = mode_F; break;
+ case 2: mode = mode_F; break;
+ default: panic("internal compiler error (ia32_gen_fp_known_const)");
}
- tv = new_tarval_from_str(cnst_str, strlen(cnst_str), mode);
- tp = ia32_create_float_type(mode, names[kct].align);
+ tv = new_tarval_from_str(cnst_str, strlen(cnst_str), mode);
- if (kct == ia32_ULLBIAS)
- tp = ia32_create_float_array(tp);
- ent = new_entity(get_glob_type(), new_id_from_str(ent_name), tp);
+ if (kct == ia32_ULLBIAS) {
+ ir_type *type = ia32_get_prim_type(mode_F);
+ ir_type *atype = ia32_create_float_array(type);
+ ir_initializer_t *initializer;
- set_entity_ld_ident(ent, get_entity_ident(ent));
- add_entity_linkage(ent, IR_LINKAGE_CONSTANT);
- set_entity_visibility(ent, ir_visibility_private);
+ ent = new_entity(get_glob_type(), name, atype);
- if (kct == ia32_ULLBIAS) {
- ir_initializer_t *initializer = create_initializer_compound(2);
+ set_entity_ld_ident(ent, name);
+ set_entity_visibility(ent, ir_visibility_private);
+ add_entity_linkage(ent, IR_LINKAGE_CONSTANT);
+ initializer = create_initializer_compound(2);
set_initializer_compound_value(initializer, 0,
create_initializer_tarval(get_mode_null(mode)));
set_initializer_compound_value(initializer, 1,
create_initializer_tarval(tv));
-
set_entity_initializer(ent, initializer);
} else {
- set_entity_initializer(ent, create_initializer_tarval(tv));
+ ent = ia32_create_float_const_entity(isa, tv, name);
}
-
/* cache the entry */
ent_cache[kct] = ent;
}
* 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, match_flags_t flags)
+ ir_node *other, ir_node *other2,
+ match_flags_t flags)
{
ir_node *load;
+ ir_mode *mode;
long pn;
/* float constants are always available */
if (is_Const(node)) {
- ir_mode *mode = get_irn_mode(node);
+ mode = get_irn_mode(node);
if (mode_is_float(mode)) {
+ ir_tarval *tv = get_Const_tarval(node);
+ if (!tarval_ieee754_can_conv_lossless(tv, mode_D))
+ return 0;
if (ia32_cg_config.use_sse2) {
if (is_simple_sse_Const(node))
return 0;
return 0;
return 1;
}
+ return 0;
}
if (!is_Proj(node))
return 0;
if (get_nodes_block(load) != block)
return 0;
+ mode = get_irn_mode(node);
+ /* we can't fold mode_E AM */
+ if (mode == ia32_mode_E)
+ return 0;
/* we only use address mode if we're the only user of the load */
if (get_irn_n_edges(node) != (flags & match_two_users ? 2 : 1))
return 0;
return 0;
/* don't do AM if other node inputs depend on the load (via mem-proj) */
- if (other != NULL && prevents_AM(block, load, other))
+ if (other != NULL && ia32_prevents_AM(block, load, other))
return 0;
- if (other2 != NULL && prevents_AM(block, load, other2))
+ if (other2 != NULL && ia32_prevents_AM(block, load, other2))
return 0;
return 1;
{
/* construct load address */
memset(addr, 0, sizeof(addr[0]));
- ia32_create_address_mode(addr, ptr, 0);
+ ia32_create_address_mode(addr, ptr, ia32_create_am_normal);
addr->base = addr->base ? be_transform_node(addr->base) : noreg_GP;
addr->index = addr->index ? be_transform_node(addr->index) : noreg_GP;
/* floating point immediates */
if (is_Const(node)) {
- ir_entity *entity = create_float_const_entity(node);
+ ir_graph *irg = get_irn_irg(node);
+ const arch_env_t *arch_env = be_get_irg_arch_env(irg);
+ ia32_isa_t *isa = (ia32_isa_t*) arch_env;
+ ir_tarval *tv = get_Const_tarval(node);
+ ir_entity *entity = ia32_create_float_const_entity(isa, tv, NULL);
addr->base = get_symconst_base();
addr->index = noreg_GP;
addr->mem = nomem;
addr->symconst_ent = entity;
+ addr->tls_segment = false;
addr->use_frame = 1;
am->ls_mode = get_type_mode(get_entity_type(entity));
am->pinned = op_pin_state_floats;
set_ia32_am_scale(node, addr->scale);
set_ia32_am_sc(node, addr->symconst_ent);
set_ia32_am_offs_int(node, addr->offset);
+ set_ia32_am_tls_segment(node, addr->tls_segment);
if (addr->symconst_sign)
set_ia32_am_sc_sign(node);
if (addr->use_frame)
* op2 input */
new_op2 = NULL;
if (!(flags & match_try_am) && use_immediate) {
- new_op2 = try_create_Immediate(op2, 0);
+ new_op2 = ia32_try_create_Immediate(op2, 0);
}
if (new_op2 == NULL &&
use_am && ia32_use_source_address_mode(block, op2, op1, other_op, flags)) {
- build_address(am, op2, 0);
+ build_address(am, op2, ia32_create_am_normal);
new_op1 = (op1 == NULL ? NULL : be_transform_node(op1));
if (mode_is_float(mode)) {
- new_op2 = ia32_new_NoReg_vfp(env_cg);
+ new_op2 = ia32_new_NoReg_vfp(current_ir_graph);
} else {
new_op2 = noreg_GP;
}
use_am &&
ia32_use_source_address_mode(block, op1, op2, other_op, flags)) {
ir_node *noreg;
- build_address(am, op1, 0);
+ build_address(am, op1, ia32_create_am_normal);
if (mode_is_float(mode)) {
- noreg = ia32_new_NoReg_vfp(env_cg);
+ noreg = ia32_new_NoReg_vfp(current_ir_graph);
} else {
noreg = noreg_GP;
}
} else {
new_op1 = be_transform_node(op2);
new_op2 = noreg;
- am->ins_permuted = 1;
+ am->ins_permuted = true;
}
am->op_type = ia32_AddrModeS;
} else {
- ir_mode *mode;
am->op_type = ia32_Normal;
if (flags & match_try_am) {
n_ia32_l_binop_right, /**< ia32 right input */
n_ia32_l_binop_eflags /**< ia32 eflags input */
};
-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)
+COMPILETIME_ASSERT((int)n_ia32_l_binop_left == (int)n_ia32_l_Adc_left, n_Adc_left)
+COMPILETIME_ASSERT((int)n_ia32_l_binop_right == (int)n_ia32_l_Adc_right, n_Adc_right)
+COMPILETIME_ASSERT((int)n_ia32_l_binop_eflags == (int)n_ia32_l_Adc_eflags, n_Adc_eflags)
+COMPILETIME_ASSERT((int)n_ia32_l_binop_left == (int)n_ia32_l_Sbb_minuend, n_Sbb_minuend)
+COMPILETIME_ASSERT((int)n_ia32_l_binop_right == (int)n_ia32_l_Sbb_subtrahend, n_Sbb_subtrahend)
+COMPILETIME_ASSERT((int)n_ia32_l_binop_eflags == (int)n_ia32_l_Sbb_eflags, n_Sbb_eflags)
/**
* Construct a binary operation which also consumes the eflags.
static ir_node *get_fpcw(void)
{
- ir_node *fpcw;
if (initial_fpcw != NULL)
return initial_fpcw;
- fpcw = be_abi_get_ignore_irn(be_get_irg_abi(env_cg->irg),
- &ia32_fp_cw_regs[REG_FPCW]);
- initial_fpcw = be_transform_node(fpcw);
-
+ initial_fpcw = be_transform_node(old_initial_fpcw);
return initial_fpcw;
}
+static ir_node *skip_float_upconv(ir_node *node)
+{
+ ir_mode *mode = get_irn_mode(node);
+ assert(mode_is_float(mode));
+
+ while (is_Conv(node)) {
+ ir_node *pred = get_Conv_op(node);
+ ir_mode *pred_mode = get_irn_mode(pred);
+
+ /**
+ * suboptimal, but without this check the address mode matcher
+ * can incorrectly think that something has only 1 user
+ */
+ if (get_irn_n_edges(node) > 1)
+ break;
+
+ if (!mode_is_float(pred_mode)
+ || get_mode_size_bits(pred_mode) > get_mode_size_bits(mode))
+ break;
+ node = pred;
+ mode = pred_mode;
+ }
+ return node;
+}
+
/**
* Construct a standard binary operation, set AM and immediate if required.
*
static ir_node *gen_binop_x87_float(ir_node *node, ir_node *op1, ir_node *op2,
construct_binop_float_func *func)
{
- ir_mode *mode = get_irn_mode(node);
dbg_info *dbgi;
- ir_node *block, *new_block, *new_node;
+ ir_node *block;
+ ir_node *new_block;
+ ir_node *new_node;
ia32_address_mode_t am;
ia32_address_t *addr = &am.addr;
ia32_x87_attr_t *attr;
/* All operations are considered commutative, because there are reverse
* variants */
- match_flags_t flags = match_commutative;
-
- /* happens for div nodes... */
- if (mode == mode_T)
- mode = get_divop_resmod(node);
+ match_flags_t flags = match_commutative | match_am;
- /* cannot use address mode with long double on x87 */
- if (get_mode_size_bits(mode) <= 64)
- flags |= match_am;
+ op1 = skip_float_upconv(op1);
+ op2 = skip_float_upconv(op2);
block = get_nodes_block(node);
match_arguments(&am, block, op1, op2, NULL, flags);
{
dbg_info *dbgi;
ir_node *block, *new_block, *new_op1, *new_op2, *new_node;
+ ir_mode *mode = get_irn_mode(node);
- assert(! mode_is_float(get_irn_mode(node)));
+ assert(! mode_is_float(mode));
assert(flags & match_immediate);
assert((flags & ~(match_mode_neutral | match_immediate)) == 0);
+ if (get_mode_modulo_shift(mode) != 32)
+ panic("modulo shift!=32 not supported by ia32 backend");
+
if (flags & match_mode_neutral) {
op1 = ia32_skip_downconv(op1);
new_op1 = be_transform_node(op1);
- } else if (get_mode_size_bits(get_irn_mode(node)) != 32) {
+ } else if (get_mode_size_bits(mode) != 32) {
new_op1 = create_upconv(op1, node);
} else {
new_op1 = be_transform_node(op1);
/* lowered shift instruction may have a dependency operand, handle it here */
if (get_irn_arity(node) == 3) {
/* we have a dependency */
- ir_node *new_dep = be_transform_node(get_irn_n(node, 2));
- add_irn_dep(new_node, new_dep);
+ ir_node* dep = get_irn_n(node, 2);
+ if (get_irn_n_edges(dep) > 1) {
+ /* ... which has at least one user other than 'node' */
+ ir_node *new_dep = be_transform_node(dep);
+ add_irn_dep(new_node, new_dep);
+ }
}
return new_node;
return new_node;
}
-static ir_node *create_lea_from_address(dbg_info *dbgi, ir_node *block,
+static ir_node *create_lea_from_address(dbg_info *dbgi, ir_node *block,
ia32_address_t *addr)
{
- ir_node *base, *index, *res;
+ ir_node *base;
+ ir_node *idx;
+ ir_node *res;
base = addr->base;
if (base == NULL) {
base = be_transform_node(base);
}
- index = addr->index;
- if (index == NULL) {
- index = noreg_GP;
+ idx = addr->index;
+ if (idx == NULL) {
+ idx = noreg_GP;
} else {
- index = be_transform_node(index);
+ idx = be_transform_node(idx);
+ }
+
+ /* segment overrides are ineffective for Leas :-( so we have to patch
+ * around... */
+ if (addr->tls_segment) {
+ ir_node *tls_base = new_bd_ia32_LdTls(NULL, block);
+ assert(addr->symconst_ent != NULL);
+ if (base == noreg_GP)
+ base = tls_base;
+ else
+ base = new_bd_ia32_Lea(dbgi, block, tls_base, base);
+ addr->tls_segment = false;
}
- res = new_bd_ia32_Lea(dbgi, block, base, index);
+ res = new_bd_ia32_Lea(dbgi, block, base, idx);
set_address(res, addr);
return res;
|| addr->frame_entity || addr->use_frame;
}
+typedef ir_node* (*new_shiftd_func)(dbg_info *dbgi, ir_node *block,
+ ir_node *high, ir_node *low,
+ ir_node *count);
+
+/**
+ * Transforms a l_ShlD/l_ShrD into a ShlD/ShrD. Those nodes have 3 data inputs:
+ * op1 - target to be shifted
+ * op2 - contains bits to be shifted into target
+ * op3 - shift count
+ * Only op3 can be an immediate.
+ */
+static ir_node *gen_64bit_shifts(dbg_info *dbgi, ir_node *block,
+ ir_node *high, ir_node *low, ir_node *count,
+ new_shiftd_func func)
+{
+ ir_node *new_block = be_transform_node(block);
+ ir_node *new_high = be_transform_node(high);
+ ir_node *new_low = be_transform_node(low);
+ ir_node *new_count;
+ ir_node *new_node;
+
+ /* the shift amount can be any mode that is bigger than 5 bits, since all
+ * other bits are ignored anyway */
+ while (is_Conv(count) &&
+ get_irn_n_edges(count) == 1 &&
+ mode_is_int(get_irn_mode(count))) {
+ assert(get_mode_size_bits(get_irn_mode(count)) >= 5);
+ count = get_Conv_op(count);
+ }
+ new_count = create_immediate_or_transform(count, 0);
+
+ new_node = func(dbgi, new_block, new_high, new_low, new_count);
+ return new_node;
+}
+
+/**
+ * test wether 2 values result in 'x' and '32-x' when interpreted as a shift
+ * value.
+ */
+static bool is_complementary_shifts(ir_node *value1, ir_node *value2)
+{
+ if (is_Const(value1) && is_Const(value2)) {
+ ir_tarval *tv1 = get_Const_tarval(value1);
+ ir_tarval *tv2 = get_Const_tarval(value2);
+ if (tarval_is_long(tv1) && tarval_is_long(tv2)) {
+ long v1 = get_tarval_long(tv1);
+ long v2 = get_tarval_long(tv2);
+ return v1 <= v2 && v2 == 32-v1;
+ }
+ }
+ return false;
+}
+
+static ir_node *match_64bit_shift(ir_node *node)
+{
+ ir_node *op1 = get_binop_left(node);
+ ir_node *op2 = get_binop_right(node);
+ assert(is_Or(node) || is_Add(node));
+
+ if (is_Shr(op1)) {
+ ir_node *tmp = op1;
+ op1 = op2;
+ op2 = tmp;
+ }
+
+ /* match ShlD operation */
+ if (is_Shl(op1) && is_Shr(op2)) {
+ ir_node *shl_right = get_Shl_right(op1);
+ ir_node *shl_left = get_Shl_left(op1);
+ ir_node *shr_right = get_Shr_right(op2);
+ ir_node *shr_left = get_Shr_left(op2);
+ /* constant ShlD operation */
+ if (is_complementary_shifts(shl_right, shr_right)) {
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ ir_node *block = get_nodes_block(node);
+ return gen_64bit_shifts(dbgi, block, shl_left, shr_left, shl_right,
+ new_bd_ia32_ShlD);
+ }
+ /* constant ShrD operation */
+ if (is_complementary_shifts(shr_right, shl_right)) {
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ ir_node *block = get_nodes_block(node);
+ return gen_64bit_shifts(dbgi, block, shr_left, shl_left, shr_right,
+ new_bd_ia32_ShrD);
+ }
+ /* lower_dw produces the following for ShlD:
+ * Or(Shr(Shr(high,1),Not(c)),Shl(low,c)) */
+ if (is_Shr(shr_left) && is_Not(shr_right)
+ && is_Const_1(get_Shr_right(shr_left))
+ && get_Not_op(shr_right) == shl_right) {
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ ir_node *block = get_nodes_block(node);
+ ir_node *val_h = get_Shr_left(shr_left);
+ return gen_64bit_shifts(dbgi, block, shl_left, val_h, shl_right,
+ new_bd_ia32_ShlD);
+ }
+ /* lower_dw produces the following for ShrD:
+ * Or(Shl(Shl(high,1),Not(c)), Shr(low,c)) */
+ if (is_Shl(shl_left) && is_Not(shl_right)
+ && is_Const_1(get_Shl_right(shl_left))
+ && get_Not_op(shl_right) == shr_right) {
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ ir_node *block = get_nodes_block(node);
+ ir_node *val_h = get_Shl_left(shl_left);
+ return gen_64bit_shifts(dbgi, block, shr_left, val_h, shr_right,
+ new_bd_ia32_ShrD);
+ }
+ }
+
+ return NULL;
+}
+
/**
* Creates an ia32 Add.
*
ia32_address_t addr;
ia32_address_mode_t am;
+ new_node = match_64bit_shift(node);
+ if (new_node != NULL)
+ return new_node;
+
if (mode_is_float(mode)) {
if (ia32_cg_config.use_sse2)
return gen_binop(node, op1, op2, new_bd_ia32_xAdd,
ia32_mark_non_am(node);
- op2 = ia32_skip_downconv(op2);
- op1 = ia32_skip_downconv(op1);
-
/**
* Rules for an Add:
* 0. Immediate Trees (example Add(Symconst, Const) -> Const)
if (addr.base == NULL && addr.index == NULL) {
new_node = new_bd_ia32_Const(dbgi, new_block, addr.symconst_ent,
addr.symconst_sign, 0, addr.offset);
- be_dep_on_frame(new_node);
SET_IA32_ORIG_NODE(new_node, node);
return new_node;
}
/* is it a zero extension? */
if (is_Const(op2)) {
- tarval *tv = get_Const_tarval(op2);
- long v = get_tarval_long(tv);
+ ir_tarval *tv = get_Const_tarval(op2);
+ long v = get_tarval_long(tv);
if (v == 0xFF || v == 0xFFFF) {
dbg_info *dbgi = get_irn_dbg_info(node);
match_commutative | match_mode_neutral | match_am | match_immediate);
}
-
-
/**
* Creates an ia32 Or.
*
{
ir_node *op1 = get_Or_left(node);
ir_node *op2 = get_Or_right(node);
+ ir_node *res;
+
+ res = match_64bit_shift(node);
+ if (res != NULL)
+ return res;
assert (! mode_is_float(get_irn_mode(node)));
return gen_binop(node, op1, op2, new_bd_ia32_Or, match_commutative
ins[n++] = be_transform_node(pred);
}
- ins[n++] = am_mem;
+ if (n==1 && ins[0] == am_mem) {
+ return am_mem;
+ /* creating a new Sync and relying on CSE may fail,
+ * if am_mem is a ProjM, which does not yet verify. */
+ }
+ ins[n++] = am_mem;
return new_r_Sync(block, n, ins);
} else {
ir_node *ins[2];
(void)orig;
if (ia32_cg_config.use_short_sex_eax) {
ir_node *pval = new_bd_ia32_ProduceVal(dbgi, block);
- be_dep_on_frame(pval);
res = new_bd_ia32_Cltd(dbgi, block, val, pval);
} else {
ir_node *imm31 = ia32_create_Immediate(NULL, 0, 31);
}
/**
- * Generates an ia32 DivMod with additional infrastructure for the
+ * Generates an ia32 Div with additional infrastructure for the
* register allocator if needed.
*/
static ir_node *create_Div(ir_node *node)
{
- dbg_info *dbgi = get_irn_dbg_info(node);
- ir_node *block = get_nodes_block(node);
- ir_node *new_block = be_transform_node(block);
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ ir_node *block = get_nodes_block(node);
+ ir_node *new_block = be_transform_node(block);
+ int throws_exception = ir_throws_exception(node);
ir_node *mem;
ir_node *new_mem;
ir_node *op1;
mem = get_Mod_mem(node);
mode = get_Mod_resmode(node);
break;
- case iro_DivMod:
- op1 = get_DivMod_left(node);
- op2 = get_DivMod_right(node);
- mem = get_DivMod_mem(node);
- mode = get_DivMod_resmode(node);
- break;
default:
panic("invalid divmod node %+F", node);
}
addr->index, new_mem, am.new_op2, am.new_op1, sign_extension);
} else {
sign_extension = new_bd_ia32_Const(dbgi, new_block, NULL, 0, 0, 0);
- be_dep_on_frame(sign_extension);
new_node = new_bd_ia32_Div(dbgi, new_block, addr->base,
addr->index, new_mem, am.new_op2,
am.new_op1, sign_extension);
}
+ ir_set_throws_exception(new_node, throws_exception);
set_irn_pinned(new_node, get_irn_pinned(node));
*/
static ir_node *gen_Div(ir_node *node)
{
- return create_Div(node);
-}
-
-/**
- * Generates an ia32 DivMod.
- */
-static ir_node *gen_DivMod(ir_node *node)
-{
- return create_Div(node);
-}
-
-
-
-/**
- * Creates an ia32 floating Div.
- *
- * @return The created ia32 xDiv node
- */
-static ir_node *gen_Quot(ir_node *node)
-{
- ir_node *op1 = get_Quot_left(node);
- ir_node *op2 = get_Quot_right(node);
+ ir_mode *mode = get_Div_resmode(node);
+ if (mode_is_float(mode)) {
+ ir_node *op1 = get_Div_left(node);
+ ir_node *op2 = get_Div_right(node);
- if (ia32_cg_config.use_sse2) {
- return gen_binop(node, op1, op2, new_bd_ia32_xDiv, match_am);
- } else {
- return gen_binop_x87_float(node, op1, op2, new_bd_ia32_vfdiv);
+ if (ia32_cg_config.use_sse2) {
+ return gen_binop(node, op1, op2, new_bd_ia32_xDiv, match_am);
+ } else {
+ return gen_binop_x87_float(node, op1, op2, new_bd_ia32_vfdiv);
+ }
}
-}
+ return create_Div(node);
+}
/**
* Creates an ia32 Shl.
ir_node *right = get_Shrs_right(node);
if (is_Const(right)) {
- tarval *tv = get_Const_tarval(right);
- long val = get_tarval_long(tv);
+ ir_tarval *tv = get_Const_tarval(right);
+ long val = get_tarval_long(tv);
if (val == 31) {
/* this is a sign extension */
dbg_info *dbgi = get_irn_dbg_info(node);
ir_node *shl_left = get_Shl_left(left);
ir_node *shl_right = get_Shl_right(left);
if (is_Const(shl_right)) {
- tarval *tv1 = get_Const_tarval(right);
- tarval *tv2 = get_Const_tarval(shl_right);
+ ir_tarval *tv1 = get_Const_tarval(right);
+ ir_tarval *tv2 = get_Const_tarval(shl_right);
if (tv1 == tv2 && tarval_is_long(tv1)) {
long val = get_tarval_long(tv1);
if (val == 16 || val == 24) {
*/
static ir_node *gen_Rotl(ir_node *node)
{
- ir_node *rotate = NULL;
ir_node *op1 = get_Rotl_left(node);
ir_node *op2 = get_Rotl_right(node);
- /* Firm has only RotL, so we are looking for a right (op2)
- operand "-e+mode_size_bits" (it's an already modified "mode_size_bits-e",
- that means we can create a RotR instead of an Add and a RotL */
-
- if (is_Add(op2)) {
- ir_node *add = op2;
- ir_node *left = get_Add_left(add);
- ir_node *right = get_Add_right(add);
- if (is_Const(right)) {
- tarval *tv = get_Const_tarval(right);
- ir_mode *mode = get_irn_mode(node);
- long bits = get_mode_size_bits(mode);
-
- if (is_Minus(left) &&
- tarval_is_long(tv) &&
- get_tarval_long(tv) == bits &&
- bits == 32)
- {
- DB((dbg, LEVEL_1, "RotL into RotR ... "));
- rotate = gen_Ror(node, op1, get_Minus_op(left));
- }
- }
- }
-
- if (rotate == NULL) {
- rotate = gen_Rol(node, op1, op2);
+ if (is_Minus(op2)) {
+ return gen_Ror(node, op1, get_Minus_op(op2));
}
- return rotate;
+ return gen_Rol(node, op1, op2);
}
/* TODO: non-optimal... if we have many xXors, then we should
* rather create a load for the const and use that instead of
* several AM nodes... */
- ir_node *noreg_xmm = ia32_new_NoReg_xmm(env_cg);
+ ir_node *noreg_xmm = ia32_new_NoReg_xmm(current_ir_graph);
new_node = new_bd_ia32_xXor(dbgi, block, get_symconst_base(),
noreg_GP, nomem, new_op, noreg_xmm);
return gen_unop(node, op, new_bd_ia32_Not, match_mode_neutral);
}
-
-
-/**
- * Transforms an Abs node.
- *
- * @return The created ia32 Abs node
- */
-static ir_node *gen_Abs(ir_node *node)
+static ir_node *create_float_abs(dbg_info *dbgi, ir_node *block, ir_node *op,
+ bool negate, ir_node *node)
{
- ir_node *block = get_nodes_block(node);
ir_node *new_block = be_transform_node(block);
- ir_node *op = get_Abs_op(node);
- dbg_info *dbgi = get_irn_dbg_info(node);
- ir_mode *mode = get_irn_mode(node);
- ir_node *new_op;
+ ir_mode *mode = get_irn_mode(op);
+ ir_node *new_op = be_transform_node(op);
ir_node *new_node;
int size;
ir_entity *ent;
- if (mode_is_float(mode)) {
- new_op = be_transform_node(op);
-
- if (ia32_cg_config.use_sse2) {
- ir_node *noreg_fp = ia32_new_NoReg_xmm(env_cg);
- new_node = new_bd_ia32_xAnd(dbgi, new_block, get_symconst_base(),
- noreg_GP, nomem, new_op, noreg_fp);
-
- size = get_mode_size_bits(mode);
- ent = ia32_gen_fp_known_const(size == 32 ? ia32_SABS : ia32_DABS);
-
- set_ia32_am_sc(new_node, ent);
+ assert(mode_is_float(mode));
- SET_IA32_ORIG_NODE(new_node, node);
+ if (ia32_cg_config.use_sse2) {
+ ir_node *noreg_fp = ia32_new_NoReg_xmm(current_ir_graph);
+ new_node = new_bd_ia32_xAnd(dbgi, new_block, get_symconst_base(),
+ noreg_GP, nomem, new_op, noreg_fp);
- set_ia32_op_type(new_node, ia32_AddrModeS);
- set_ia32_ls_mode(new_node, mode);
- } else {
- new_node = new_bd_ia32_vfabs(dbgi, new_block, new_op);
- SET_IA32_ORIG_NODE(new_node, node);
- }
- } else {
- ir_node *xor, *sign_extension;
+ size = get_mode_size_bits(mode);
+ ent = ia32_gen_fp_known_const(size == 32 ? ia32_SABS : ia32_DABS);
- if (get_mode_size_bits(mode) == 32) {
- new_op = be_transform_node(op);
- } else {
- new_op = create_I2I_Conv(mode, mode_Is, dbgi, block, op, node);
- }
+ set_ia32_am_sc(new_node, ent);
- sign_extension = create_sex_32_64(dbgi, new_block, new_op, node);
+ SET_IA32_ORIG_NODE(new_node, node);
- xor = new_bd_ia32_Xor(dbgi, new_block, noreg_GP, noreg_GP,
- nomem, new_op, sign_extension);
- SET_IA32_ORIG_NODE(xor, node);
+ set_ia32_op_type(new_node, ia32_AddrModeS);
+ set_ia32_ls_mode(new_node, mode);
- new_node = new_bd_ia32_Sub(dbgi, new_block, noreg_GP, noreg_GP,
- nomem, xor, sign_extension);
+ /* TODO, implement -Abs case */
+ assert(!negate);
+ } else {
+ new_node = new_bd_ia32_vfabs(dbgi, new_block, new_op);
SET_IA32_ORIG_NODE(new_node, node);
+ if (negate) {
+ new_node = new_bd_ia32_vfchs(dbgi, new_block, new_node);
+ SET_IA32_ORIG_NODE(new_node, node);
+ }
}
return new_node;
return new_bd_ia32_Bt(dbgi, 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)
+static ia32_condition_code_t relation_to_condition_code(ir_relation relation,
+ ir_mode *mode,
+ bool overflow_possible)
{
- ir_node *flags;
- ir_node *new_op;
- ir_node *new_block;
- dbg_info *dbgi;
-
- /* we have a Cmp as input */
- if (is_Proj(node)) {
- ir_node *pred = get_Proj_pred(node);
- 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;
- }
- }
- }
- }
- /* add ia32 compare flags */
- {
- ir_node *l = get_Cmp_left(pred);
- ir_mode *mode = get_irn_mode(l);
- if (mode_is_float(mode))
- pnc |= ia32_pn_Cmp_float;
- else if (! mode_is_signed(mode))
- pnc |= ia32_pn_Cmp_unsigned;
+ if (mode_is_float(mode)) {
+ switch (relation) {
+ case ir_relation_equal: return ia32_cc_float_equal;
+ case ir_relation_less: return ia32_cc_float_below;
+ case ir_relation_less_equal: return ia32_cc_float_below_equal;
+ case ir_relation_greater: return ia32_cc_float_above;
+ case ir_relation_greater_equal: return ia32_cc_float_above_equal;
+ case ir_relation_less_greater: return ia32_cc_not_equal;
+ case ir_relation_less_equal_greater: return ia32_cc_not_parity;
+ case ir_relation_unordered: return ia32_cc_parity;
+ case ir_relation_unordered_equal: return ia32_cc_equal;
+ case ir_relation_unordered_less: return ia32_cc_float_unordered_below;
+ case ir_relation_unordered_less_equal:
+ return ia32_cc_float_unordered_below_equal;
+ case ir_relation_unordered_greater:
+ return ia32_cc_float_unordered_above;
+ case ir_relation_unordered_greater_equal:
+ return ia32_cc_float_unordered_above_equal;
+ case ir_relation_unordered_less_greater:
+ return ia32_cc_float_not_equal;
+ case ir_relation_false:
+ case ir_relation_true:
+ /* should we introduce a jump always/jump never? */
+ break;
+ }
+ panic("Unexpected float pnc");
+ } else if (mode_is_signed(mode)) {
+ switch (relation) {
+ case ir_relation_unordered_equal:
+ case ir_relation_equal: return ia32_cc_equal;
+ case ir_relation_unordered_less:
+ case ir_relation_less:
+ return overflow_possible ? ia32_cc_less : ia32_cc_sign;
+ case ir_relation_unordered_less_equal:
+ case ir_relation_less_equal: return ia32_cc_less_equal;
+ case ir_relation_unordered_greater:
+ case ir_relation_greater: return ia32_cc_greater;
+ case ir_relation_unordered_greater_equal:
+ case ir_relation_greater_equal:
+ return overflow_possible ? ia32_cc_greater_equal : ia32_cc_not_sign;
+ case ir_relation_unordered_less_greater:
+ case ir_relation_less_greater: return ia32_cc_not_equal;
+ case ir_relation_less_equal_greater:
+ case ir_relation_unordered:
+ case ir_relation_false:
+ case ir_relation_true:
+ /* introduce jump always/jump never? */
+ break;
+ }
+ panic("Unexpected pnc");
+ } else {
+ switch (relation) {
+ case ir_relation_unordered_equal:
+ case ir_relation_equal: return ia32_cc_equal;
+ case ir_relation_unordered_less:
+ case ir_relation_less: return ia32_cc_below;
+ case ir_relation_unordered_less_equal:
+ case ir_relation_less_equal: return ia32_cc_below_equal;
+ case ir_relation_unordered_greater:
+ case ir_relation_greater: return ia32_cc_above;
+ case ir_relation_unordered_greater_equal:
+ case ir_relation_greater_equal: return ia32_cc_above_equal;
+ case ir_relation_unordered_less_greater:
+ case ir_relation_less_greater: return ia32_cc_not_equal;
+ case ir_relation_less_equal_greater:
+ case ir_relation_unordered:
+ case ir_relation_false:
+ case ir_relation_true:
+ /* introduce jump always/jump never? */
+ break;
+ }
+ panic("Unexpected pnc");
+ }
+}
+
+static ir_node *get_flags_node(ir_node *cmp, ia32_condition_code_t *cc_out)
+{
+ /* must have a Cmp as input */
+ ir_relation relation = get_Cmp_relation(cmp);
+ ir_node *l = get_Cmp_left(cmp);
+ ir_node *r = get_Cmp_right(cmp);
+ ir_mode *mode = get_irn_mode(l);
+ bool overflow_possible;
+ ir_node *flags;
+
+ /* check for bit-test */
+ if (ia32_cg_config.use_bt
+ && (relation == ir_relation_equal
+ || (mode_is_signed(mode) && relation == ir_relation_less_greater)
+ || (!mode_is_signed(mode) && ((relation & ir_relation_greater_equal) == ir_relation_greater)))
+ && is_And(l)) {
+ ir_node *la = get_And_left(l);
+ ir_node *ra = get_And_right(l);
+ if (is_Shl(ra)) {
+ ir_node *tmp = la;
+ la = ra;
+ ra = tmp;
+ }
+ if (is_Shl(la)) {
+ ir_node *c = get_Shl_left(la);
+ if (is_Const_1(c) && is_Const_0(r)) {
+ /* (1 << n) & ra) */
+ ir_node *n = get_Shl_right(la);
+ flags = gen_bt(cmp, ra, n);
+ /* the bit is copied into the CF flag */
+ if (relation & ir_relation_equal)
+ *cc_out = ia32_cc_above_equal; /* test for CF=0 */
+ else
+ *cc_out = ia32_cc_below; /* test for CF=1 */
+ return flags;
}
- *pnc_out = pnc;
- flags = be_transform_node(pred);
- return flags;
}
}
- /* a mode_b value, we have to compare it against 0 */
- dbgi = get_irn_dbg_info(node);
- new_block = be_transform_node(get_nodes_block(node));
- new_op = be_transform_node(node);
- flags = new_bd_ia32_Test(dbgi, new_block, noreg_GP, noreg_GP, nomem, new_op,
- new_op, /*is_permuted=*/0, /*cmp_unsigned=*/0);
- *pnc_out = pn_Cmp_Lg;
+ /* the middle-end tries to eliminate impossible relations, so a ptr <> 0
+ * test becomes ptr > 0. But for x86 an equal comparison is preferable to
+ * a >0 (we can sometimes eliminate the cmp in favor of flags produced by
+ * a predecessor node). So add the < bit.
+ * (Note that we do not want to produce <=> (which can happen for
+ * unoptimized code), because no x86 flag can represent that */
+ if (!(relation & ir_relation_equal) && relation & ir_relation_less_greater)
+ relation |= get_negated_relation(ir_get_possible_cmp_relations(l, r)) & ir_relation_less_greater;
+
+ overflow_possible = true;
+ if (is_Const(r) && is_Const_null(r))
+ overflow_possible = false;
+
+ /* just do a normal transformation of the Cmp */
+ *cc_out = relation_to_condition_code(relation, mode, overflow_possible);
+ flags = be_transform_node(cmp);
return flags;
}
static ir_node *gen_Load(ir_node *node)
{
ir_node *old_block = get_nodes_block(node);
- ir_node *block = be_transform_node(old_block);
- ir_node *ptr = get_Load_ptr(node);
- ir_node *mem = get_Load_mem(node);
- ir_node *new_mem = be_transform_node(mem);
+ ir_node *block = be_transform_node(old_block);
+ ir_node *ptr = get_Load_ptr(node);
+ ir_node *mem = get_Load_mem(node);
+ ir_node *new_mem = be_transform_node(mem);
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ ir_mode *mode = get_Load_mode(node);
+ int throws_exception = ir_throws_exception(node);
ir_node *base;
- ir_node *index;
- dbg_info *dbgi = get_irn_dbg_info(node);
- ir_mode *mode = get_Load_mode(node);
+ ir_node *idx;
ir_node *new_node;
ia32_address_t addr;
/* construct load address */
memset(&addr, 0, sizeof(addr));
- ia32_create_address_mode(&addr, ptr, 0);
- base = addr.base;
- index = addr.index;
+ ia32_create_address_mode(&addr, ptr, ia32_create_am_normal);
+ base = addr.base;
+ idx = addr.index;
if (base == NULL) {
base = noreg_GP;
base = be_transform_node(base);
}
- if (index == NULL) {
- index = noreg_GP;
+ if (idx == NULL) {
+ idx = noreg_GP;
} else {
- index = be_transform_node(index);
+ idx = be_transform_node(idx);
}
if (mode_is_float(mode)) {
if (ia32_cg_config.use_sse2) {
- new_node = new_bd_ia32_xLoad(dbgi, block, base, index, new_mem,
+ new_node = new_bd_ia32_xLoad(dbgi, block, base, idx, new_mem,
mode);
} else {
- new_node = new_bd_ia32_vfld(dbgi, block, base, index, new_mem,
+ new_node = new_bd_ia32_vfld(dbgi, block, base, idx, new_mem,
mode);
}
} else {
/* create a conv node with address mode for smaller modes */
if (get_mode_size_bits(mode) < 32) {
- new_node = new_bd_ia32_Conv_I2I(dbgi, block, base, index,
+ new_node = new_bd_ia32_Conv_I2I(dbgi, block, base, idx,
new_mem, noreg_GP, mode);
} else {
- new_node = new_bd_ia32_Load(dbgi, block, base, index, new_mem);
+ new_node = new_bd_ia32_Load(dbgi, block, base, idx, new_mem);
}
}
+ ir_set_throws_exception(new_node, throws_exception);
set_irn_pinned(new_node, get_irn_pinned(node));
set_ia32_op_type(new_node, ia32_AddrModeS);
set_address(new_node, &addr);
if (get_irn_pinned(node) == op_pin_state_floats) {
- assert(pn_ia32_xLoad_res == pn_ia32_vfld_res
- && pn_ia32_vfld_res == pn_ia32_Load_res
- && pn_ia32_Load_res == pn_ia32_res);
- arch_irn_add_flags(new_node, arch_irn_flags_rematerializable);
+ assert((int)pn_ia32_xLoad_res == (int)pn_ia32_vfld_res
+ && (int)pn_ia32_vfld_res == (int)pn_ia32_Load_res
+ && (int)pn_ia32_Load_res == (int)pn_ia32_res);
+ arch_add_irn_flags(new_node, arch_irn_flags_rematerializable);
}
SET_IA32_ORIG_NODE(new_node, node);
- be_dep_on_frame(new_node);
return new_node;
}
/* 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)) {
+ heights_reachable_in_block(ia32_heights, other, load)) {
return 0;
}
- if (prevents_AM(block, load, mem))
+ if (ia32_prevents_AM(block, load, mem))
return 0;
/* Store should be attached to the load via mem */
- assert(heights_reachable_in_block(heights, mem, load));
+ assert(heights_reachable_in_block(ia32_heights, mem, load));
return 1;
}
be_set_transformed_node(get_Proj_pred(am.mem_proj), new_node);
mem_proj = be_transform_node(am.mem_proj);
- be_set_transformed_node(mem_proj ? mem_proj : am.mem_proj, new_node);
+ be_set_transformed_node(am.mem_proj, new_node);
+ be_set_transformed_node(mem_proj, new_node);
return new_node;
}
be_set_transformed_node(get_Proj_pred(am.mem_proj), new_node);
mem_proj = be_transform_node(am.mem_proj);
- be_set_transformed_node(mem_proj ? mem_proj : am.mem_proj, new_node);
+ be_set_transformed_node(am.mem_proj, new_node);
+ be_set_transformed_node(mem_proj, new_node);
return new_node;
}
-static pn_Cmp ia32_get_negated_pnc(pn_Cmp pnc)
-{
- ir_mode *mode = pnc & ia32_pn_Cmp_float ? mode_F : mode_Iu;
- return get_negated_pnc(pnc, mode);
-}
-
static ir_node *try_create_SetMem(ir_node *node, ir_node *ptr, ir_node *mem)
{
- ir_mode *mode = get_irn_mode(node);
- ir_node *mux_true = get_Mux_true(node);
- ir_node *mux_false = get_Mux_false(node);
- ir_node *cond;
- dbg_info *dbgi;
- ir_node *block;
- ir_node *new_block;
- ir_node *flags;
- ir_node *new_node;
- bool negated;
- pn_Cmp pnc;
- ia32_address_t addr;
+ ir_mode *mode = get_irn_mode(node);
+ ir_node *mux_true = get_Mux_true(node);
+ ir_node *mux_false = get_Mux_false(node);
+ ir_node *cond;
+ dbg_info *dbgi;
+ ir_node *block;
+ ir_node *new_block;
+ ir_node *flags;
+ ir_node *new_node;
+ bool negated;
+ ia32_condition_code_t cc;
+ ia32_address_t addr;
if (get_mode_size_bits(mode) != 8)
return NULL;
}
cond = get_Mux_sel(node);
- flags = get_flags_node(cond, &pnc);
+ flags = get_flags_node(cond, &cc);
/* we can't handle the float special cases with SetM */
- if (pnc & ia32_pn_Cmp_float)
+ if (cc & ia32_cc_additional_float_cases)
return NULL;
if (negated)
- pnc = ia32_get_negated_pnc(pnc);
+ cc = ia32_negate_condition_code(cc);
build_address_ptr(&addr, ptr, mem);
block = get_nodes_block(node);
new_block = be_transform_node(block);
new_node = new_bd_ia32_SetccMem(dbgi, new_block, addr.base,
- addr.index, addr.mem, flags, pnc);
+ addr.index, addr.mem, flags, cc);
set_address(new_node, &addr);
set_ia32_op_type(new_node, ia32_AddrModeD);
set_ia32_ls_mode(new_node, mode);
{
ir_mode *mode = get_irn_mode(cns);
unsigned size = get_mode_size_bytes(mode);
- tarval *tv = get_Const_tarval(cns);
+ ir_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 *mem = get_Store_mem(node);
dbg_info *dbgi = get_irn_dbg_info(node);
int ofs = 0;
- size_t i = 0;
+ int i = 0;
+ int throws_exception = ir_throws_exception(node);
ir_node *ins[4];
ia32_address_t addr;
- assert(size % 4 == 0);
- assert(size <= 16);
-
build_address_ptr(&addr, ptr, mem);
do {
- 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);
+ unsigned val;
+ unsigned delta;
+ ir_mode *mode;
+ if (size >= 4) {
+ 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);
+ delta = 4;
+ mode = mode_Iu;
+ } else if (size >= 2) {
+ val= get_tarval_sub_bits(tv, ofs) |
+ (get_tarval_sub_bits(tv, ofs + 1) << 8);
+ delta = 2;
+ mode = mode_Hu;
+ } else {
+ panic("invalid size of Store float to mem (%+F)", node);
+ }
ir_node *imm = ia32_create_Immediate(NULL, 0, val);
ir_node *new_node = new_bd_ia32_Store(dbgi, new_block, addr.base,
addr.index, addr.mem, imm);
+ ir_node *new_mem = new_r_Proj(new_node, mode_M, pn_ia32_Store_M);
+ ir_set_throws_exception(new_node, throws_exception);
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_ia32_ls_mode(new_node, mode);
set_address(new_node, &addr);
SET_IA32_ORIG_NODE(new_node, node);
assert(i < 4);
- ins[i++] = new_node;
+ ins[i++] = new_mem;
- size -= 4;
- ofs += 4;
- addr.offset += 4;
+ size -= delta;
+ ofs += delta;
+ addr.offset += delta;
} while (size != 0);
if (i > 1) {
return new_rd_Sync(dbgi, new_block, i, ins);
} else {
- return ins[0];
+ return get_Proj_pred(ins[0]);
}
}
/**
* Generate a vfist or vfisttp instruction.
*/
-static ir_node *gen_vfist(dbg_info *dbgi, ir_node *block, ir_node *base, ir_node *index,
- ir_node *mem, ir_node *val, ir_node **fist)
+static ir_node *gen_vfist(dbg_info *dbgi, ir_node *block, ir_node *base,
+ ir_node *index, ir_node *mem, ir_node *val)
{
- 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 */
ir_node *vfisttp = new_bd_ia32_vfisttp(dbgi, block, base, index, mem, val);
- ir_node *value = new_r_Proj(vfisttp, mode_E, pn_ia32_vfisttp_res);
+ ir_node *value = new_r_Proj(vfisttp, ia32_mode_E, pn_ia32_vfisttp_res);
be_new_Keep(block, 1, &value);
- new_node = new_r_Proj(vfisttp, mode_M, pn_ia32_vfisttp_M);
- *fist = vfisttp;
+ return vfisttp;
} else {
- ir_node *trunc_mode = ia32_new_Fpu_truncate(env_cg);
+ ir_node *trunc_mode = ia32_new_Fpu_truncate(current_ir_graph);
/* do a fist */
- new_node = new_bd_ia32_vfist(dbgi, block, base, index, mem, val, trunc_mode);
- *fist = new_node;
+ ir_node *vfist = new_bd_ia32_vfist(dbgi, block, base, index, mem, val, trunc_mode);
+ return vfist;
}
- return new_node;
}
+
/**
* Transforms a general (no special case) Store.
*
ir_node *ptr = get_Store_ptr(node);
ir_node *mem = get_Store_mem(node);
dbg_info *dbgi = get_irn_dbg_info(node);
- ir_node *new_val, *new_node, *store;
+ int throws_exception = ir_throws_exception(node);
+ ir_node *new_val;
+ ir_node *new_node;
ia32_address_t addr;
/* check for destination address mode */
/* construct store address */
memset(&addr, 0, sizeof(addr));
- ia32_create_address_mode(&addr, ptr, 0);
+ ia32_create_address_mode(&addr, ptr, ia32_create_am_normal);
if (addr.base == NULL) {
addr.base = noreg_GP;
new_node = new_bd_ia32_vfst(dbgi, new_block, addr.base,
addr.index, addr.mem, new_val, mode);
}
- store = new_node;
} else if (!ia32_cg_config.use_sse2 && is_float_to_int_conv(val)) {
val = get_Conv_op(val);
/* TODO: is this optimisation still necessary at all (middleend)? */
- /* We can skip ALL float->float up-Convs (and strict-up-Convs) before stores. */
+ /* We can skip ALL float->float up-Convs (and strict-up-Convs) before
+ * stores. */
while (is_Conv(val)) {
ir_node *op = get_Conv_op(val);
if (!mode_is_float(get_irn_mode(op)))
val = op;
}
new_val = be_transform_node(val);
- new_node = gen_vfist(dbgi, new_block, addr.base, addr.index, addr.mem, new_val, &store);
+ new_node = gen_vfist(dbgi, new_block, addr.base, addr.index, addr.mem, new_val);
} else {
new_val = create_immediate_or_transform(val, 0);
assert(mode != mode_b);
new_node = new_bd_ia32_Store(dbgi, new_block, addr.base,
addr.index, addr.mem, new_val);
}
- store = new_node;
}
+ ir_set_throws_exception(new_node, throws_exception);
- set_irn_pinned(store, get_irn_pinned(node));
- set_ia32_op_type(store, ia32_AddrModeD);
- set_ia32_ls_mode(store, mode);
+ 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_address(store, &addr);
- SET_IA32_ORIG_NODE(store, node);
+ set_address(new_node, &addr);
+ SET_IA32_ORIG_NODE(new_node, node);
return new_node;
}
*/
static ir_node *gen_Store(ir_node *node)
{
- ir_node *val = get_Store_value(node);
- ir_mode *mode = get_irn_mode(val);
+ ir_node *val = get_Store_value(node);
+ ir_mode *mode = get_irn_mode(val);
if (mode_is_float(mode) && is_Const(val)) {
/* We can transform every floating const store
*
* @return the created ia32 SwitchJmp node
*/
-static ir_node *create_Switch(ir_node *node)
-{
- dbg_info *dbgi = get_irn_dbg_info(node);
- ir_node *block = be_transform_node(get_nodes_block(node));
- ir_node *sel = get_Cond_selector(node);
- ir_node *new_sel = be_transform_node(sel);
- long switch_min = LONG_MAX;
- long switch_max = LONG_MIN;
- long default_pn = get_Cond_default_proj(node);
- ir_node *new_node;
- const ir_edge_t *edge;
-
- assert(get_mode_size_bits(get_irn_mode(sel)) == 32);
-
- /* determine the smallest switch case value */
- foreach_out_edge(node, edge) {
- ir_node *proj = get_edge_src_irn(edge);
- long pn = get_Proj_proj(proj);
- if (pn == default_pn)
- continue;
-
- if (pn < switch_min)
- switch_min = pn;
- if (pn > switch_max)
- switch_max = pn;
- }
-
- if ((unsigned long) (switch_max - switch_min) > 128000) {
- panic("Size of switch %+F bigger than 128000", node);
- }
-
- if (switch_min != 0) {
- /* if smallest switch case is not 0 we need an additional sub */
- new_sel = new_bd_ia32_Lea(dbgi, block, new_sel, noreg_GP);
- add_ia32_am_offs_int(new_sel, -switch_min);
- set_ia32_op_type(new_sel, ia32_AddrModeS);
-
- SET_IA32_ORIG_NODE(new_sel, node);
- }
-
- new_node = new_bd_ia32_SwitchJmp(dbgi, block, new_sel, default_pn);
+static ir_node *gen_Switch(ir_node *node)
+{
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ ir_graph *irg = get_irn_irg(node);
+ ir_node *block = be_transform_node(get_nodes_block(node));
+ ir_node *sel = get_Switch_selector(node);
+ ir_node *new_sel = be_transform_node(sel);
+ ir_mode *sel_mode = get_irn_mode(sel);
+ const ir_switch_table *table = get_Switch_table(node);
+ unsigned n_outs = get_Switch_n_outs(node);
+ ir_node *new_node;
+ ir_entity *entity;
+
+ assert(get_mode_size_bits(get_irn_mode(sel)) <= 32);
+ if (get_mode_size_bits(sel_mode) != 32)
+ new_sel = create_upconv(new_sel, sel);
+
+ entity = new_entity(NULL, id_unique("TBL%u"), get_unknown_type());
+ set_entity_visibility(entity, ir_visibility_private);
+ add_entity_linkage(entity, IR_LINKAGE_CONSTANT);
+
+ table = ir_switch_table_duplicate(irg, table);
+
+ new_node = new_bd_ia32_SwitchJmp(dbgi, block, noreg_GP, new_sel, n_outs, table);
+ set_ia32_am_scale(new_node, 2);
+ set_ia32_am_sc(new_node, entity);
+ set_ia32_op_type(new_node, ia32_AddrModeS);
+ set_ia32_ls_mode(new_node, mode_Iu);
SET_IA32_ORIG_NODE(new_node, node);
+ // FIXME This seems wrong. GCC uses PIC for switch on OS X.
+ get_ia32_attr(new_node)->data.am_sc_no_pic_adjust = true;
return new_node;
}
*/
static ir_node *gen_Cond(ir_node *node)
{
- ir_node *block = get_nodes_block(node);
- ir_node *new_block = be_transform_node(block);
- dbg_info *dbgi = get_irn_dbg_info(node);
- ir_node *sel = get_Cond_selector(node);
- ir_mode *sel_mode = get_irn_mode(sel);
- ir_node *flags = NULL;
- ir_node *new_node;
- pn_Cmp pnc;
-
- if (sel_mode != mode_b) {
- return create_Switch(node);
- }
+ ir_node *block = get_nodes_block(node);
+ ir_node *new_block = be_transform_node(block);
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ ir_node *sel = get_Cond_selector(node);
+ ir_node *flags = NULL;
+ ir_node *new_node;
+ ia32_condition_code_t cc;
/* we get flags from a Cmp */
- flags = get_flags_node(sel, &pnc);
+ flags = get_flags_node(sel, &cc);
- new_node = new_bd_ia32_Jcc(dbgi, new_block, flags, pnc);
+ new_node = new_bd_ia32_Jcc(dbgi, new_block, flags, cc);
SET_IA32_ORIG_NODE(new_node, node);
return new_node;
set_ia32_commutative(new_node);
SET_IA32_ORIG_NODE(new_node, node);
} else {
- if (ia32_cg_config.use_ftst && is_Const_0(right)) {
+ if (is_Const_0(right)) {
new_node = new_bd_ia32_vFtstFnstsw(dbgi, new_block, new_left, 0);
} else {
new_right = be_transform_node(right);
new_node = new_bd_ia32_vFucomFnstsw(dbgi, new_block, new_left, new_right, 0);
+ set_ia32_commutative(new_node);
}
- set_ia32_commutative(new_node);
-
SET_IA32_ORIG_NODE(new_node, node);
new_node = new_bd_ia32_Sahf(dbgi, new_block, new_node);
return new_node;
}
-/**
- * helper function: checks whether all Cmp projs are Lg or Eq which is needed
- * to fold an and into a test node
- */
-static bool can_fold_test_and(ir_node *node)
-{
- const ir_edge_t *edge;
-
- /** we can only have eq and lg projs */
- foreach_out_edge(node, edge) {
- ir_node *proj = get_edge_src_irn(edge);
- pn_Cmp pnc = get_Proj_proj(proj);
- if (pnc != pn_Cmp_Eq && pnc != pn_Cmp_Lg)
- return false;
- }
-
- return true;
-}
-
/**
* returns true if it is assured, that the upper bits of a node are "clean"
* which means for a 16 or 8 bit value, that the upper bits in the register
return shifted == 0 || shifted == -1;
} else {
unsigned long shifted = (unsigned long)attr->offset;
- shifted >>= get_mode_size_bits(mode);
+ shifted >>= get_mode_size_bits(mode)-1;
+ shifted >>= 1;
return shifted == 0;
}
}
ir_node *new_node;
ia32_address_mode_t am;
ia32_address_t *addr = &am.addr;
- int cmp_unsigned;
if (mode_is_float(cmp_mode)) {
if (ia32_cg_config.use_sse2) {
assert(ia32_mode_needs_gp_reg(cmp_mode));
/* Prefer the Test instruction, when encountering (x & y) ==/!= 0 */
- cmp_unsigned = !mode_is_signed(cmp_mode);
if (is_Const_0(right) &&
is_And(left) &&
- get_irn_n_edges(left) == 1 &&
- can_fold_test_and(node)) {
+ get_irn_n_edges(left) == 1) {
/* Test(and_left, and_right) */
ir_node *and_left = get_And_left(left);
ir_node *and_right = get_And_right(left);
assert(get_irn_mode(and_left) == cmp_mode);
match_arguments(&am, block, and_left, and_right, NULL,
- match_commutative |
- match_am | match_8bit_am | match_16bit_am |
- match_am_and_immediates | match_immediate);
+ match_commutative |
+ match_am | match_8bit_am | match_16bit_am |
+ match_am_and_immediates | match_immediate);
/* use 32bit compare mode if possible since the opcode is smaller */
if (upper_bits_clean(am.new_op1, cmp_mode) &&
if (get_mode_size_bits(cmp_mode) == 8) {
new_node = new_bd_ia32_Test8Bit(dbgi, new_block, addr->base,
- addr->index, addr->mem, am.new_op1, am.new_op2, am.ins_permuted,
- cmp_unsigned);
+ addr->index, addr->mem,
+ am.new_op1, am.new_op2,
+ am.ins_permuted);
} else {
- new_node = new_bd_ia32_Test(dbgi, new_block, addr->base, addr->index,
- addr->mem, am.new_op1, am.new_op2, am.ins_permuted, cmp_unsigned);
+ new_node = new_bd_ia32_Test(dbgi, new_block, addr->base,
+ addr->index, addr->mem, am.new_op1,
+ am.new_op2, am.ins_permuted);
}
} else {
/* Cmp(left, right) */
if (get_mode_size_bits(cmp_mode) == 8) {
new_node = new_bd_ia32_Cmp8Bit(dbgi, new_block, addr->base,
addr->index, addr->mem, am.new_op1,
- am.new_op2, am.ins_permuted,
- cmp_unsigned);
+ am.new_op2, am.ins_permuted);
} else {
new_node = new_bd_ia32_Cmp(dbgi, new_block, addr->base, addr->index,
- addr->mem, am.new_op1, am.new_op2, am.ins_permuted, cmp_unsigned);
+ addr->mem, am.new_op1, am.new_op2,
+ am.ins_permuted);
}
}
set_am_attributes(new_node, &am);
}
static ir_node *create_CMov(ir_node *node, ir_node *flags, ir_node *new_flags,
- pn_Cmp pnc)
+ ia32_condition_code_t cc)
{
dbg_info *dbgi = get_irn_dbg_info(node);
ir_node *block = get_nodes_block(node);
match_commutative | match_am | match_16bit_am | match_mode_neutral);
if (am.ins_permuted)
- pnc = ia32_get_negated_pnc(pnc);
+ cc = ia32_negate_condition_code(cc);
new_node = new_bd_ia32_CMovcc(dbgi, new_block, addr->base, addr->index,
addr->mem, am.new_op1, am.new_op2, new_flags,
- pnc);
+ cc);
set_am_attributes(new_node, &am);
SET_IA32_ORIG_NODE(new_node, 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 *flags, ia32_condition_code_t cc,
ir_node *orig_node)
{
ir_mode *mode = get_irn_mode(orig_node);
ir_node *new_node;
- new_node = new_bd_ia32_Setcc(dbgi, new_block, flags, pnc);
+ new_node = new_bd_ia32_Setcc(dbgi, new_block, flags, cc);
SET_IA32_ORIG_NODE(new_node, orig_node);
/* we might need to conv the result up */
ir_node *new_node;
ir_node *sub;
ir_node *sbb;
- ir_node *not;
+ ir_node *notn;
ir_node *eflags;
ir_node *block;
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, sub, mode, pn_ia32_res);
}
+ assert(is_ia32_Sub(sub));
eflags = new_rd_Proj(NULL, sub, mode_Iu, pn_ia32_Sub_flags);
- dbgi = get_irn_dbg_info(psi);
- sbb = new_bd_ia32_Sbb0(dbgi, block, eflags);
- not = new_bd_ia32_Not(dbgi, block, sbb);
+ dbgi = get_irn_dbg_info(psi);
+ sbb = new_bd_ia32_Sbb0(dbgi, block, eflags);
+ set_ia32_ls_mode(sbb, mode_Iu);
+ notn = new_bd_ia32_Not(dbgi, block, sbb);
- new_node = new_bd_ia32_And(dbgi, block, noreg_GP, noreg_GP, nomem, new_node, not);
+ new_node = new_bd_ia32_And(dbgi, block, noreg_GP, noreg_GP, nomem, new_node, notn);
+ set_ia32_ls_mode(new_node, mode_Iu);
set_ia32_commutative(new_node);
return new_node;
}
ir_mode *mode = *new_mode;
ir_type *tp;
ir_initializer_t *initializer;
- tarval *tv0 = get_Const_tarval(c0);
- tarval *tv1 = get_Const_tarval(c1);
+ ir_tarval *tv0 = get_Const_tarval(c0);
+ ir_tarval *tv1 = get_Const_tarval(c1);
if (mode == NULL) {
/* detect the best mode for the constants */
}
- tp = ia32_create_float_type(mode, 4);
+ tp = ia32_get_prim_type(mode);
tp = ia32_create_float_array(tp);
ent = new_entity(get_glob_type(), id_unique("C%u"), tp);
};
typedef struct setcc_transform {
- unsigned num_steps;
- unsigned permutate_cmp_ins;
- pn_Cmp pnc;
+ unsigned num_steps;
+ ia32_condition_code_t cc;
struct {
enum setcc_transform_insn transform;
long val;
* Find a transformation that creates 0 and 1 from
* tv_t and tv_f.
*/
-static void find_const_transform(pn_Cmp pnc, tarval *t, tarval *f,
+static void find_const_transform(ia32_condition_code_t cc,
+ ir_tarval *t, ir_tarval *f,
setcc_transform_t *res)
{
unsigned step = 0;
res->num_steps = 0;
- res->permutate_cmp_ins = 0;
if (tarval_is_null(t)) {
- tarval *tmp = t;
+ ir_tarval *tmp = t;
t = f;
f = tmp;
- pnc = ia32_get_negated_pnc(pnc);
- } else if (tarval_cmp(t, f) == pn_Cmp_Lt) {
+ cc = ia32_negate_condition_code(cc);
+ } else if (tarval_cmp(t, f) == ir_relation_less) {
// now, t is the bigger one
- tarval *tmp = t;
+ ir_tarval *tmp = t;
t = f;
f = tmp;
- pnc = ia32_get_negated_pnc(pnc);
+ cc = ia32_negate_condition_code(cc);
}
- res->pnc = pnc;
+ res->cc = cc;
if (! tarval_is_null(f)) {
- tarval *t_sub = tarval_sub(t, f, NULL);
+ ir_tarval *t_sub = tarval_sub(t, f, NULL);
t = t_sub;
res->steps[step].transform = SETCC_TR_ADD;
++step;
res->steps[step].transform = SETCC_TR_NEG;
} else {
- int v = get_tarval_lowest_bit(t);
- assert(v >= 0);
+ int val = get_tarval_lowest_bit(t);
+ assert(val >= 0);
res->steps[step].transform = SETCC_TR_SHL;
- res->steps[step].scale = v;
+ res->steps[step].scale = val;
}
}
++step;
*/
static ir_node *gen_Mux(ir_node *node)
{
- 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 *mux_true = get_Mux_true(node);
- ir_node *mux_false = get_Mux_false(node);
- ir_node *cond = get_Mux_sel(node);
- ir_mode *mode = get_irn_mode(node);
- ir_node *flags;
- ir_node *new_node;
- pn_Cmp pnc;
-
- assert(get_irn_mode(cond) == mode_b);
+ 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 *mux_true = get_Mux_true(node);
+ ir_node *mux_false = get_Mux_false(node);
+ ir_node *sel = get_Mux_sel(node);
+ ir_mode *mode = get_irn_mode(node);
+ ir_node *flags;
+ ir_node *new_node;
+ int is_abs;
+ ia32_condition_code_t cc;
+
+ assert(get_irn_mode(sel) == mode_b);
+
+ is_abs = ir_mux_is_abs(sel, mux_false, mux_true);
+ if (is_abs != 0) {
+ if (ia32_mode_needs_gp_reg(mode)) {
+ ir_fprintf(stderr, "Optimisation warning: Integer abs %+F not transformed\n",
+ node);
+ } else {
+ ir_node *op = ir_get_abs_op(sel, mux_false, mux_true);
+ return create_float_abs(dbgi, block, op, is_abs < 0, node);
+ }
+ }
/* Note: a Mux node uses a Load two times IFF it's used in the compare AND in the result */
if (mode_is_float(mode)) {
- 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);
+ ir_node *cmp_left = get_Cmp_left(sel);
+ ir_node *cmp_right = get_Cmp_right(sel);
+ ir_relation relation = get_Cmp_relation(sel);
if (ia32_cg_config.use_sse2) {
- if (pnc == pn_Cmp_Lt || pnc == pn_Cmp_Le) {
+ if (relation == ir_relation_less || relation == ir_relation_less_equal) {
if (cmp_left == mux_true && cmp_right == mux_false) {
/* Mux(a <= b, a, b) => MIN */
return gen_binop(node, cmp_left, cmp_right, new_bd_ia32_xMin,
return gen_binop(node, cmp_left, cmp_right, new_bd_ia32_xMax,
match_commutative | match_am | match_two_users);
}
- } else if (pnc == pn_Cmp_Gt || pnc == pn_Cmp_Ge) {
+ } else if (relation == ir_relation_greater || relation == ir_relation_greater_equal) {
if (cmp_left == mux_true && cmp_right == mux_false) {
/* Mux(a >= b, a, b) => MAX */
return gen_binop(node, cmp_left, cmp_right, new_bd_ia32_xMax,
ir_mode *new_mode;
unsigned scale;
- flags = get_flags_node(cond, &pnc);
- new_node = create_set_32bit(dbgi, new_block, flags, pnc, node);
+ flags = get_flags_node(sel, &cc);
+ new_node = create_set_32bit(dbgi, new_block, flags, cc, node);
if (ia32_cg_config.use_sse2) {
/* cannot load from different mode on SSE */
am.addr.symconst_ent = ia32_create_const_array(mux_false, mux_true, &new_mode);
- switch (get_mode_size_bytes(new_mode)) {
- case 4:
+ if (new_mode == mode_F) {
scale = 2;
- break;
- case 8:
+ } else if (new_mode == mode_D) {
scale = 3;
- break;
- case 10:
- /* use 2 * 5 */
- scale = 1;
- new_node = new_bd_ia32_Lea(dbgi, new_block, new_node, new_node);
- set_ia32_am_scale(new_node, 2);
- break;
- case 12:
- /* use 4 * 3 */
- scale = 2;
- new_node = new_bd_ia32_Lea(dbgi, new_block, new_node, new_node);
- set_ia32_am_scale(new_node, 1);
- break;
- case 16:
+ } else if (new_mode == ia32_mode_E) {
/* arg, shift 16 NOT supported */
scale = 3;
- new_node = new_bd_ia32_Add(dbgi, new_block, noreg_GP, noreg_GP, nomem, new_node, new_node);
- break;
- default:
+ new_node = new_bd_ia32_Lea(dbgi, new_block, new_node, new_node);
+ } else {
panic("Unsupported constant size");
}
am.addr.offset = 0;
am.addr.scale = scale;
am.addr.use_frame = 0;
+ am.addr.tls_segment = false;
am.addr.frame_entity = NULL;
am.addr.symconst_sign = 0;
am.mem_proj = am.addr.mem;
am.new_op2 = NULL;
am.pinned = op_pin_state_floats;
am.commutative = 1;
- am.ins_permuted = 0;
+ am.ins_permuted = false;
if (ia32_cg_config.use_sse2)
load = new_bd_ia32_xLoad(dbgi, block, am.addr.base, am.addr.index, am.addr.mem, new_mode);
} else {
assert(ia32_mode_needs_gp_reg(mode));
- if (is_Proj(cond)) {
- ir_node *cmp = get_Proj_pred(cond);
- if (is_Cmp(cmp)) {
- ir_node *cmp_left = get_Cmp_left(cmp);
- ir_node *cmp_right = get_Cmp_right(cmp);
- pn_Cmp pnc = get_Proj_proj(cond);
-
- /* check for unsigned Doz first */
- if ((pnc & pn_Cmp_Gt) && !mode_is_signed(mode) &&
- is_Const_0(mux_false) && is_Sub(mux_true) &&
- get_Sub_left(mux_true) == cmp_left && get_Sub_right(mux_true) == cmp_right) {
- /* Mux(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(mux_true) && is_Sub(mux_false) &&
- get_Sub_left(mux_false) == cmp_left && get_Sub_right(mux_false) == cmp_right) {
- /* Mux(a <=u b, 0, a - b) unsigned Doz */
+ if (is_Cmp(sel)) {
+ ir_node *cmp_left = get_Cmp_left(sel);
+ ir_node *cmp_right = get_Cmp_right(sel);
+ ir_relation relation = get_Cmp_relation(sel);
+ ir_node *val_true = mux_true;
+ ir_node *val_false = mux_false;
+
+ if (is_Const(val_true) && is_Const_null(val_true)) {
+ ir_node *tmp = val_false;
+ val_false = val_true;
+ val_true = tmp;
+ relation = get_negated_relation(relation);
+ }
+ if (is_Const_0(val_false) && is_Sub(val_true)) {
+ if ((relation & ir_relation_greater)
+ && get_Sub_left(val_true) == cmp_left
+ && get_Sub_right(val_true) == cmp_right) {
return create_doz(node, cmp_left, cmp_right);
}
+ if ((relation & ir_relation_less)
+ && get_Sub_left(val_true) == cmp_right
+ && get_Sub_right(val_true) == cmp_left) {
+ return create_doz(node, cmp_right, cmp_left);
+ }
}
}
- flags = get_flags_node(cond, &pnc);
+ flags = get_flags_node(sel, &cc);
if (is_Const(mux_true) && is_Const(mux_false)) {
/* both are const, good */
- tarval *tv_true = get_Const_tarval(mux_true);
- tarval *tv_false = get_Const_tarval(mux_false);
+ ir_tarval *tv_true = get_Const_tarval(mux_true);
+ ir_tarval *tv_false = get_Const_tarval(mux_false);
setcc_transform_t res;
int step;
- find_const_transform(pnc, tv_true, tv_false, &res);
+ find_const_transform(cc, tv_true, tv_false, &res);
new_node = node;
- if (res.permutate_cmp_ins) {
- ia32_attr_t *attr = get_ia32_attr(flags);
- attr->data.ins_permuted ^= 1;
- }
for (step = (int)res.num_steps - 1; step >= 0; --step) {
ir_node *imm;
switch (res.steps[step].transform) {
case SETCC_TR_ADD:
- imm = ia32_immediate_from_long(res.steps[step].val);
- new_node = new_bd_ia32_Add(dbgi, new_block, noreg_GP, noreg_GP, nomem, new_node, imm);
+ new_node = new_bd_ia32_Lea(dbgi, new_block, new_node, noreg_GP);
+ add_ia32_am_offs_int(new_node, res.steps[step].val);
break;
case SETCC_TR_ADDxx:
new_node = new_bd_ia32_Lea(dbgi, new_block, new_node, new_node);
new_node = new_bd_ia32_And(dbgi, new_block, noreg_GP, noreg_GP, nomem, new_node, imm);
break;
case SETCC_TR_SET:
- new_node = create_set_32bit(dbgi, new_block, flags, res.pnc, new_node);
+ new_node = create_set_32bit(dbgi, new_block, flags, res.cc, node);
break;
case SETCC_TR_SBB:
new_node = new_bd_ia32_Sbb0(dbgi, new_block, flags);
}
}
} else {
- new_node = create_CMov(node, cond, flags, pnc);
+ new_node = create_CMov(node, sel, flags, cc);
}
return new_node;
}
}
-
/**
* Create a conversion from x87 state register to general purpose.
*/
ir_graph *irg = current_ir_graph;
dbg_info *dbgi = get_irn_dbg_info(node);
ir_mode *mode = get_irn_mode(node);
+ ir_node *frame = get_irg_frame(irg);
ir_node *fist, *load, *mem;
- mem = gen_vfist(dbgi, block, get_irg_frame(irg), noreg_GP, nomem, new_op, &fist);
+ fist = gen_vfist(dbgi, block, frame, noreg_GP, nomem, new_op);
set_irn_pinned(fist, op_pin_state_floats);
set_ia32_use_frame(fist);
set_ia32_op_type(fist, ia32_AddrModeD);
+ assert((long)pn_ia32_vfist_M == (long) pn_ia32_vfisttp_M);
+ mem = new_r_Proj(fist, mode_M, pn_ia32_vfist_M);
+
assert(get_mode_size_bits(mode) <= 32);
/* exception we can only store signed 32 bit integers, so for unsigned
we store a 64bit (signed) integer and load the lower bits */
ir_graph *irg = get_Block_irg(block);
dbg_info *dbgi = get_irn_dbg_info(node);
ir_node *frame = get_irg_frame(irg);
+ ir_node *store_mem;
ir_node *store, *load;
ir_node *new_node;
set_ia32_op_type(store, ia32_AddrModeD);
SET_IA32_ORIG_NODE(store, node);
- load = new_bd_ia32_vfld(dbgi, block, frame, noreg_GP, store, tgt_mode);
+ store_mem = new_r_Proj(store, mode_M, pn_ia32_vfst_M);
+
+ load = new_bd_ia32_vfld(dbgi, block, frame, noreg_GP, store_mem, tgt_mode);
set_ia32_use_frame(load);
set_ia32_op_type(load, ia32_AddrModeS);
SET_IA32_ORIG_NODE(load, node);
- new_node = new_r_Proj(load, mode_E, pn_ia32_vfld_res);
+ new_node = new_r_Proj(load, ia32_mode_E, pn_ia32_vfld_res);
return new_node;
}
ir_mode *store_mode;
ir_node *fild;
ir_node *store;
+ ir_node *store_mem;
ir_node *new_node;
/* fild can use source AM if the operand is a signed 16bit or 32bit integer */
set_ia32_op_type(store, ia32_AddrModeD);
set_ia32_ls_mode(store, mode_Iu);
+ store_mem = new_r_Proj(store, mode_M, pn_ia32_Store_M);
+
/* exception for 32bit unsigned, do a 64bit spill+load */
if (!mode_is_signed(mode)) {
ir_node *in[2];
ir_node *zero_store = new_bd_ia32_Store(dbgi, block, get_irg_frame(irg),
noreg_GP, nomem, zero_const);
+ ir_node *zero_store_mem = new_r_Proj(zero_store, mode_M, pn_ia32_Store_M);
set_ia32_use_frame(zero_store);
set_ia32_op_type(zero_store, ia32_AddrModeD);
add_ia32_am_offs_int(zero_store, 4);
set_ia32_ls_mode(zero_store, mode_Iu);
- in[0] = zero_store;
- in[1] = store;
+ in[0] = zero_store_mem;
+ in[1] = store_mem;
- store = new_rd_Sync(dbgi, block, 2, in);
+ store_mem = new_rd_Sync(dbgi, block, 2, in);
store_mode = mode_Ls;
} else {
store_mode = mode_Is;
}
/* do a fild */
- fild = new_bd_ia32_vfild(dbgi, block, get_irg_frame(irg), noreg_GP, store);
+ fild = new_bd_ia32_vfild(dbgi, block, get_irg_frame(irg), noreg_GP, store_mem);
set_ia32_use_frame(fild);
set_ia32_op_type(fild, ia32_AddrModeS);
}
} else {
/* this should be optimized already, but who knows... */
- DEBUG_ONLY(ir_fprintf(stderr, "Debug warning: conv %+F is pointless\n", node));
+ DEBUG_ONLY(ir_fprintf(stderr, "Debug warning: conv %+F is pointless\n", node);)
DB((dbg, LEVEL_1, "killed Conv(mode, mode) ..."));
return be_transform_node(op);
}
/* if fp_no_float_fold is not set then we assume that we
* don't have any float operations in a non
* mode_float_arithmetic mode and can skip strict upconvs */
- if (src_bits < tgt_bits
- && !(get_irg_fp_model(current_ir_graph) & fp_no_float_fold)) {
+ if (src_bits < tgt_bits) {
DB((dbg, LEVEL_1, "killed Conv(float, float) ..."));
return new_op;
} else {
set_ia32_ls_mode(res, tgt_mode);
} else {
unsigned int_mantissa = get_mode_size_bits(src_mode) - (mode_is_signed(src_mode) ? 1 : 0);
- unsigned float_mantissa = tarval_ieee754_get_mantissa_size(tgt_mode);
+ unsigned float_mantissa = get_mode_mantissa_size(tgt_mode);
res = gen_x87_gp_to_fp(node, src_mode);
/* we need a strict-Conv, if the int mode has more bits than the
static ir_node *create_immediate_or_transform(ir_node *node,
char immediate_constraint_type)
{
- ir_node *new_node = try_create_Immediate(node, immediate_constraint_type);
+ ir_node *new_node = ia32_try_create_Immediate(node, immediate_constraint_type);
if (new_node == NULL) {
new_node = be_transform_node(node);
}
*/
static ir_node *gen_be_Return(ir_node *node)
{
- ir_graph *irg = current_ir_graph;
- ir_node *ret_val = get_irn_n(node, be_pos_Return_val);
- ir_node *ret_mem = get_irn_n(node, be_pos_Return_mem);
- ir_entity *ent = get_irg_entity(irg);
- ir_type *tp = get_entity_type(ent);
- dbg_info *dbgi;
- ir_node *block;
+ ir_graph *irg = current_ir_graph;
+ ir_node *ret_val = get_irn_n(node, n_be_Return_val);
+ ir_node *ret_mem = get_irn_n(node, n_be_Return_mem);
+ ir_node *new_ret_val = be_transform_node(ret_val);
+ ir_node *new_ret_mem = be_transform_node(ret_mem);
+ ir_entity *ent = get_irg_entity(irg);
+ ir_type *tp = get_entity_type(ent);
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ ir_node *block = be_transform_node(get_nodes_block(node));
ir_type *res_type;
ir_mode *mode;
- ir_node *frame, *sse_store, *fld, *mproj, *barrier;
- ir_node *new_barrier, *new_ret_val, *new_ret_mem;
- ir_node **in;
- int pn_ret_val, pn_ret_mem, arity, i;
+ ir_node *frame;
+ ir_node *sse_store;
+ ir_node *store_mem;
+ ir_node *fld;
+ ir_node *mproj;
+ int i;
+ int arity;
+ unsigned pop;
+ ir_node **in;
+ ir_node *new_node;
assert(ret_val != NULL);
if (be_Return_get_n_rets(node) < 1 || ! ia32_cg_config.use_sse2) {
assert(get_method_n_ress(tp) == 1);
- pn_ret_val = get_Proj_proj(ret_val);
- pn_ret_mem = get_Proj_proj(ret_mem);
-
- /* get the Barrier */
- barrier = get_Proj_pred(ret_val);
-
- /* get result input of the Barrier */
- ret_val = get_irn_n(barrier, pn_ret_val);
- new_ret_val = be_transform_node(ret_val);
-
- /* get memory input of the Barrier */
- ret_mem = get_irn_n(barrier, pn_ret_mem);
- new_ret_mem = be_transform_node(ret_mem);
-
frame = get_irg_frame(irg);
- dbgi = get_irn_dbg_info(barrier);
- block = be_transform_node(get_nodes_block(barrier));
-
/* store xmm0 onto stack */
sse_store = new_bd_ia32_xStoreSimple(dbgi, block, frame, noreg_GP,
new_ret_mem, new_ret_val);
set_ia32_ls_mode(sse_store, mode);
set_ia32_op_type(sse_store, ia32_AddrModeD);
set_ia32_use_frame(sse_store);
+ store_mem = new_r_Proj(sse_store, mode_M, pn_ia32_xStoreSimple_M);
/* load into x87 register */
- fld = new_bd_ia32_vfld(dbgi, block, frame, noreg_GP, sse_store, mode);
+ fld = new_bd_ia32_vfld(dbgi, block, frame, noreg_GP, store_mem, mode);
set_ia32_op_type(fld, ia32_AddrModeS);
set_ia32_use_frame(fld);
mproj = new_r_Proj(fld, mode_M, pn_ia32_vfld_M);
fld = new_r_Proj(fld, mode_vfp, pn_ia32_vfld_res);
- /* create a new barrier */
- arity = get_irn_arity(barrier);
+ /* create a new return */
+ arity = get_irn_arity(node);
in = ALLOCAN(ir_node*, arity);
+ pop = be_Return_get_pop(node);
for (i = 0; i < arity; ++i) {
- ir_node *new_in;
-
- if (i == pn_ret_val) {
- new_in = fld;
- } else if (i == pn_ret_mem) {
- new_in = mproj;
+ ir_node *op = get_irn_n(node, i);
+ if (op == ret_val) {
+ in[i] = fld;
+ } else if (op == ret_mem) {
+ in[i] = mproj;
} else {
- ir_node *in = get_irn_n(barrier, i);
- new_in = be_transform_node(in);
+ in[i] = be_transform_node(op);
}
- in[i] = new_in;
}
+ new_node = be_new_Return(dbgi, irg, block, arity, pop, arity, in);
+ copy_node_attr(irg, node, new_node);
- new_barrier = new_ir_node(dbgi, irg, block,
- get_irn_op(barrier), get_irn_mode(barrier),
- arity, in);
- copy_node_attr(irg, barrier, new_barrier);
- be_duplicate_deps(barrier, new_barrier);
- be_set_transformed_node(barrier, new_barrier);
-
- /* transform normally */
- return be_duplicate_node(node);
+ return new_node;
}
/**
*/
static ir_node *gen_be_AddSP(ir_node *node)
{
- ir_node *sz = get_irn_n(node, be_pos_AddSP_size);
- ir_node *sp = get_irn_n(node, be_pos_AddSP_old_sp);
+ ir_node *sz = get_irn_n(node, n_be_AddSP_size);
+ ir_node *sp = get_irn_n(node, n_be_AddSP_old_sp);
- return gen_binop(node, sp, sz, new_bd_ia32_SubSP,
- match_am | match_immediate);
+ ir_node *new_node = gen_binop(node, sp, sz, new_bd_ia32_SubSP,
+ match_am | match_immediate);
+ assert(is_ia32_SubSP(new_node));
+ arch_set_irn_register_out(new_node, pn_ia32_SubSP_stack,
+ &ia32_registers[REG_ESP]);
+ return new_node;
}
/**
*/
static ir_node *gen_be_SubSP(ir_node *node)
{
- ir_node *sz = get_irn_n(node, be_pos_SubSP_size);
- ir_node *sp = get_irn_n(node, be_pos_SubSP_old_sp);
+ ir_node *sz = get_irn_n(node, n_be_SubSP_size);
+ ir_node *sp = get_irn_n(node, n_be_SubSP_old_sp);
- return gen_binop(node, sp, sz, new_bd_ia32_AddSP,
- match_am | match_immediate);
+ ir_node *new_node = gen_binop(node, sp, sz, new_bd_ia32_AddSP,
+ match_am | match_immediate);
+ assert(is_ia32_AddSP(new_node));
+ arch_set_irn_register_out(new_node, pn_ia32_AddSP_stack,
+ &ia32_registers[REG_ESP]);
+ return new_node;
}
/**
copy_node_attr(irg, node, phi);
be_duplicate_deps(node, phi);
- arch_set_out_register_req(phi, 0, req);
+ arch_set_irn_register_req_out(phi, 0, req);
be_enqueue_preds(node);
return new_node;
}
-/**
- * Transform a Bound node.
- */
-static ir_node *gen_Bound(ir_node *node)
-{
- ir_node *new_node;
- ir_node *lower = get_Bound_lower(node);
- dbg_info *dbgi = get_irn_dbg_info(node);
-
- if (is_Const_0(lower)) {
- /* typical case for Java */
- ir_node *sub, *res, *flags, *block;
-
- res = gen_binop(node, get_Bound_index(node), get_Bound_upper(node),
- new_bd_ia32_Sub, match_mode_neutral | match_am | match_immediate);
-
- block = get_nodes_block(res);
- if (! is_Proj(res)) {
- sub = res;
- set_irn_mode(sub, mode_T);
- res = new_rd_Proj(NULL, sub, mode_Iu, pn_ia32_res);
- } else {
- sub = get_Proj_pred(res);
- }
- flags = new_rd_Proj(NULL, sub, mode_Iu, pn_ia32_Sub_flags);
- new_node = new_bd_ia32_Jcc(dbgi, block, flags, pn_Cmp_Lt | ia32_pn_Cmp_unsigned);
- SET_IA32_ORIG_NODE(new_node, node);
- } else {
- panic("generic Bound not supported in ia32 Backend");
- }
- return new_node;
-}
-
-
-static ir_node *gen_ia32_l_ShlDep(ir_node *node)
-{
- ir_node *left = get_irn_n(node, n_ia32_l_ShlDep_val);
- ir_node *right = get_irn_n(node, n_ia32_l_ShlDep_count);
-
- return gen_shift_binop(node, left, right, new_bd_ia32_Shl,
- match_immediate | match_mode_neutral);
-}
-
-static ir_node *gen_ia32_l_ShrDep(ir_node *node)
-{
- ir_node *left = get_irn_n(node, n_ia32_l_ShrDep_val);
- ir_node *right = get_irn_n(node, n_ia32_l_ShrDep_count);
- return gen_shift_binop(node, left, right, new_bd_ia32_Shr,
- match_immediate);
-}
-
-static ir_node *gen_ia32_l_SarDep(ir_node *node)
-{
- ir_node *left = get_irn_n(node, n_ia32_l_SarDep_val);
- ir_node *right = get_irn_n(node, n_ia32_l_SarDep_count);
- return gen_shift_binop(node, left, right, new_bd_ia32_Sar,
- match_immediate);
-}
-
static ir_node *gen_ia32_l_Add(ir_node *node)
{
ir_node *left = get_irn_n(node, n_ia32_l_Add_left);
match_mode_neutral);
if (is_Proj(lowered)) {
- lowered = get_Proj_pred(lowered);
+ lowered = get_Proj_pred(lowered);
} else {
assert(is_ia32_Add(lowered));
set_irn_mode(lowered, mode_T);
match_am | match_immediate | match_mode_neutral);
if (is_Proj(lowered)) {
- lowered = get_Proj_pred(lowered);
+ lowered = get_Proj_pred(lowered);
} else {
assert(is_ia32_Sub(lowered));
set_irn_mode(lowered, mode_T);
match_am | match_immediate | match_mode_neutral);
}
-/**
- * Transforms a l_ShlD/l_ShrD into a ShlD/ShrD. Those nodes have 3 data inputs:
- * op1 - target to be shifted
- * op2 - contains bits to be shifted into target
- * op3 - shift count
- * Only op3 can be an immediate.
- */
-static ir_node *gen_lowered_64bit_shifts(ir_node *node, ir_node *high,
- ir_node *low, ir_node *count)
-{
- ir_node *block = get_nodes_block(node);
- ir_node *new_block = be_transform_node(block);
- dbg_info *dbgi = get_irn_dbg_info(node);
- ir_node *new_high = be_transform_node(high);
- ir_node *new_low = be_transform_node(low);
- ir_node *new_count;
- ir_node *new_node;
-
- /* the shift amount can be any mode that is bigger than 5 bits, since all
- * other bits are ignored anyway */
- while (is_Conv(count) &&
- get_irn_n_edges(count) == 1 &&
- mode_is_int(get_irn_mode(count))) {
- assert(get_mode_size_bits(get_irn_mode(count)) >= 5);
- count = get_Conv_op(count);
- }
- new_count = create_immediate_or_transform(count, 0);
-
- if (is_ia32_l_ShlD(node)) {
- new_node = new_bd_ia32_ShlD(dbgi, new_block, new_high, new_low,
- new_count);
- } else {
- new_node = new_bd_ia32_ShrD(dbgi, new_block, new_high, new_low,
- new_count);
- }
- SET_IA32_ORIG_NODE(new_node, node);
-
- return new_node;
-}
-
-static ir_node *gen_ia32_l_ShlD(ir_node *node)
-{
- ir_node *high = get_irn_n(node, n_ia32_l_ShlD_val_high);
- ir_node *low = get_irn_n(node, n_ia32_l_ShlD_val_low);
- ir_node *count = get_irn_n(node, n_ia32_l_ShlD_count);
- return gen_lowered_64bit_shifts(node, high, low, count);
-}
-
-static ir_node *gen_ia32_l_ShrD(ir_node *node)
-{
- ir_node *high = get_irn_n(node, n_ia32_l_ShrD_val_high);
- ir_node *low = get_irn_n(node, n_ia32_l_ShrD_val_low);
- ir_node *count = get_irn_n(node, n_ia32_l_ShrD_count);
- return gen_lowered_64bit_shifts(node, high, low, count);
-}
-
static ir_node *gen_ia32_l_LLtoFloat(ir_node *node)
{
ir_node *src_block = get_nodes_block(node);
ir_node *new_val_high = be_transform_node(val_high);
ir_node *in[2];
ir_node *sync, *fild, *res;
- ir_node *store_low, *store_high;
+ ir_node *store_low;
+ ir_node *store_high;
+ ir_node *mem_low;
+ ir_node *mem_high;
if (ia32_cg_config.use_sse2) {
panic("ia32_l_LLtoFloat not implemented for SSE2");
SET_IA32_ORIG_NODE(store_low, node);
SET_IA32_ORIG_NODE(store_high, node);
+ mem_low = new_r_Proj(store_low, mode_M, pn_ia32_Store_M);
+ mem_high = new_r_Proj(store_high, mode_M, pn_ia32_Store_M);
+
set_ia32_use_frame(store_low);
set_ia32_use_frame(store_high);
set_ia32_op_type(store_low, ia32_AddrModeD);
set_ia32_ls_mode(store_high, mode_Is);
add_ia32_am_offs_int(store_high, 4);
- in[0] = store_low;
- in[1] = store_high;
+ in[0] = mem_low;
+ in[1] = mem_high;
sync = new_rd_Sync(dbgi, block, 2, in);
/* do a fild */
am.addr.offset = 0;
am.addr.scale = 2;
am.addr.symconst_ent = ia32_gen_fp_known_const(ia32_ULLBIAS);
+ am.addr.tls_segment = false;
am.addr.use_frame = 0;
am.addr.frame_entity = NULL;
am.addr.symconst_sign = 0;
am.mem_proj = nomem;
am.op_type = ia32_AddrModeS;
am.new_op1 = res;
- am.new_op2 = ia32_new_NoReg_vfp(env_cg);
+ am.new_op2 = ia32_new_NoReg_vfp(current_ir_graph);
am.pinned = op_pin_state_floats;
am.commutative = 1;
- am.ins_permuted = 0;
+ am.ins_permuted = false;
fadd = new_bd_ia32_vfadd(dbgi, block, am.addr.base, am.addr.index, am.addr.mem,
am.new_op1, am.new_op2, get_fpcw());
ir_node *frame = get_irg_frame(irg);
ir_node *val = get_irn_n(node, n_ia32_l_FloattoLL_val);
ir_node *new_val = be_transform_node(val);
- ir_node *fist, *mem;
+ ir_node *fist;
- mem = gen_vfist(dbgi, block, frame, noreg_GP, nomem, new_val, &fist);
+ fist = gen_vfist(dbgi, block, frame, noreg_GP, nomem, new_val);
SET_IA32_ORIG_NODE(fist, node);
set_ia32_use_frame(fist);
set_ia32_op_type(fist, ia32_AddrModeD);
set_ia32_ls_mode(fist, mode_Ls);
- return mem;
+ assert((long)pn_ia32_vfist_M == (long) pn_ia32_vfisttp_M);
+ return new_r_Proj(fist, mode_M, pn_ia32_vfist_M);
}
static ir_node *gen_Proj_l_FloattoLL(ir_node *node)
if (proj == pn_be_AddSP_sp) {
ir_node *res = new_rd_Proj(dbgi, new_pred, mode_Iu,
pn_ia32_SubSP_stack);
- arch_set_irn_register(res, &ia32_gp_regs[REG_ESP]);
+ arch_set_irn_register(res, &ia32_registers[REG_ESP]);
return res;
} else if (proj == pn_be_AddSP_res) {
return new_rd_Proj(dbgi, new_pred, mode_Iu,
if (proj == pn_be_SubSP_sp) {
ir_node *res = new_rd_Proj(dbgi, new_pred, mode_Iu,
pn_ia32_AddSP_stack);
- arch_set_irn_register(res, &ia32_gp_regs[REG_ESP]);
+ arch_set_irn_register(res, &ia32_registers[REG_ESP]);
return res;
} else if (proj == pn_be_SubSP_M) {
return new_rd_Proj(dbgi, new_pred, mode_M, pn_ia32_AddSP_M);
static ir_node *gen_Proj_Load(ir_node *node)
{
ir_node *new_pred;
- ir_node *block = be_transform_node(get_nodes_block(node));
- ir_node *pred = get_Proj_pred(node);
- dbg_info *dbgi = get_irn_dbg_info(node);
- long proj = get_Proj_proj(node);
+ ir_node *pred = get_Proj_pred(node);
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ long proj = get_Proj_proj(node);
/* loads might be part of source address mode matches, so we don't
* transform the ProjMs yet (with the exception of loads whose result is
/* renumber the proj */
new_pred = be_transform_node(pred);
if (is_ia32_Load(new_pred)) {
- switch (proj) {
+ switch ((pn_Load)proj) {
case pn_Load_res:
return new_rd_Proj(dbgi, new_pred, mode_Iu, pn_ia32_Load_res);
case pn_Load_M:
return new_rd_Proj(dbgi, new_pred, mode_M, pn_ia32_Load_M);
- case pn_Load_X_regular:
- return new_rd_Jmp(dbgi, block);
case pn_Load_X_except:
/* This Load might raise an exception. Mark it. */
set_ia32_exc_label(new_pred, 1);
- return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_Load_X_exc);
- default:
- break;
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_Load_X_except);
+ case pn_Load_X_regular:
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_Load_X_regular);
}
} else if (is_ia32_Conv_I2I(new_pred) ||
is_ia32_Conv_I2I8Bit(new_pred)) {
set_irn_mode(new_pred, mode_T);
- if (proj == pn_Load_res) {
+ switch ((pn_Load)proj) {
+ case pn_Load_res:
return new_rd_Proj(dbgi, new_pred, mode_Iu, pn_ia32_res);
- } else if (proj == pn_Load_M) {
+ case pn_Load_M:
return new_rd_Proj(dbgi, new_pred, mode_M, pn_ia32_mem);
+ case pn_Load_X_except:
+ /* This Load might raise an exception. Mark it. */
+ set_ia32_exc_label(new_pred, 1);
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_Conv_I2I_X_except);
+ case pn_Load_X_regular:
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_Conv_I2I_X_regular);
}
} else if (is_ia32_xLoad(new_pred)) {
- switch (proj) {
+ switch ((pn_Load)proj) {
case pn_Load_res:
return new_rd_Proj(dbgi, new_pred, mode_xmm, pn_ia32_xLoad_res);
case pn_Load_M:
return new_rd_Proj(dbgi, new_pred, mode_M, pn_ia32_xLoad_M);
- case pn_Load_X_regular:
- return new_rd_Jmp(dbgi, block);
case pn_Load_X_except:
/* This Load might raise an exception. Mark it. */
set_ia32_exc_label(new_pred, 1);
- return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_xLoad_X_exc);
- default:
- break;
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_xLoad_X_except);
+ case pn_Load_X_regular:
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_xLoad_X_regular);
}
} else if (is_ia32_vfld(new_pred)) {
- switch (proj) {
+ switch ((pn_Load)proj) {
case pn_Load_res:
return new_rd_Proj(dbgi, new_pred, mode_vfp, pn_ia32_vfld_res);
case pn_Load_M:
return new_rd_Proj(dbgi, new_pred, mode_M, pn_ia32_vfld_M);
- case pn_Load_X_regular:
- return new_rd_Jmp(dbgi, block);
case pn_Load_X_except:
/* This Load might raise an exception. Mark it. */
set_ia32_exc_label(new_pred, 1);
- return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_vfld_X_exc);
- default:
- break;
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_vfld_X_except);
+ case pn_Load_X_regular:
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_vfld_X_regular);
}
} else {
/* can happen for ProJMs when source address mode happened for the
return new_rd_Proj(dbgi, new_pred, mode_M, 1);
}
- panic("No idea how to transform proj");
+ panic("No idea how to transform Proj(Load) %+F", node);
+}
+
+static ir_node *gen_Proj_Store(ir_node *node)
+{
+ ir_node *pred = get_Proj_pred(node);
+ ir_node *new_pred = be_transform_node(pred);
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ long pn = get_Proj_proj(node);
+
+ if (is_ia32_Store(new_pred) || is_ia32_Store8Bit(new_pred)) {
+ switch ((pn_Store)pn) {
+ case pn_Store_M:
+ return new_rd_Proj(dbgi, new_pred, mode_M, pn_ia32_Store_M);
+ case pn_Store_X_except:
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_Store_X_except);
+ case pn_Store_X_regular:
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_Store_X_regular);
+ }
+ } else if (is_ia32_vfist(new_pred)) {
+ switch ((pn_Store)pn) {
+ case pn_Store_M:
+ return new_rd_Proj(dbgi, new_pred, mode_M, pn_ia32_vfist_M);
+ case pn_Store_X_except:
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_vfist_X_except);
+ case pn_Store_X_regular:
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_vfist_X_regular);
+ }
+ } else if (is_ia32_vfisttp(new_pred)) {
+ switch ((pn_Store)pn) {
+ case pn_Store_M:
+ return new_rd_Proj(dbgi, new_pred, mode_M, pn_ia32_vfisttp_M);
+ case pn_Store_X_except:
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_vfisttp_X_except);
+ case pn_Store_X_regular:
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_vfisttp_X_regular);
+ }
+ } else if (is_ia32_vfst(new_pred)) {
+ switch ((pn_Store)pn) {
+ case pn_Store_M:
+ return new_rd_Proj(dbgi, new_pred, mode_M, pn_ia32_vfst_M);
+ case pn_Store_X_except:
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_vfst_X_except);
+ case pn_Store_X_regular:
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_vfst_X_regular);
+ }
+ } else if (is_ia32_xStore(new_pred)) {
+ switch ((pn_Store)pn) {
+ case pn_Store_M:
+ return new_rd_Proj(dbgi, new_pred, mode_M, pn_ia32_xStore_M);
+ case pn_Store_X_except:
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_xStore_X_except);
+ case pn_Store_X_regular:
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_xStore_X_regular);
+ }
+ } else if (is_Sync(new_pred)) {
+ /* hack for the case that gen_float_const_Store produced a Sync */
+ if (pn == pn_Store_M) {
+ return new_pred;
+ }
+ panic("exception control flow for gen_float_const_Store not implemented yet");
+ } else if (get_ia32_op_type(new_pred) == ia32_AddrModeD) {
+ /* destination address mode */
+ if (pn == pn_Store_M) {
+ return new_pred;
+ }
+ panic("exception control flow for destination AM not implemented yet");
+ }
+
+ panic("No idea how to transform Proj(Store) %+F", node);
}
/**
- * Transform and renumber the Projs from a DivMod like instruction.
+ * Transform and renumber the Projs from a Div or Mod instruction.
*/
-static ir_node *gen_Proj_DivMod(ir_node *node)
+static ir_node *gen_Proj_Div(ir_node *node)
{
- 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);
dbg_info *dbgi = get_irn_dbg_info(node);
- long proj = get_Proj_proj(node);
+ long proj = get_Proj_proj(node);
- assert(is_ia32_Div(new_pred) || is_ia32_IDiv(new_pred));
+ assert((long)pn_ia32_Div_M == (long)pn_ia32_IDiv_M);
+ assert((long)pn_ia32_Div_div_res == (long)pn_ia32_IDiv_div_res);
- switch (get_irn_opcode(pred)) {
- case iro_Div:
- switch (proj) {
- case pn_Div_M:
+ switch ((pn_Div)proj) {
+ case pn_Div_M:
+ if (is_ia32_Div(new_pred) || is_ia32_IDiv(new_pred)) {
return new_rd_Proj(dbgi, new_pred, mode_M, pn_ia32_Div_M);
- case pn_Div_res:
- return new_rd_Proj(dbgi, new_pred, mode_Iu, pn_ia32_Div_div_res);
- case pn_Div_X_regular:
- return new_rd_Jmp(dbgi, block);
- case pn_Div_X_except:
- set_ia32_exc_label(new_pred, 1);
- return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_Div_X_exc);
- default:
- break;
- }
- break;
- case iro_Mod:
- switch (proj) {
- case pn_Mod_M:
- return new_rd_Proj(dbgi, new_pred, mode_M, pn_ia32_Div_M);
- case pn_Mod_res:
- return new_rd_Proj(dbgi, new_pred, mode_Iu, pn_ia32_Div_mod_res);
- case pn_Mod_X_except:
- set_ia32_exc_label(new_pred, 1);
- return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_Div_X_exc);
- default:
- break;
+ } else if (is_ia32_xDiv(new_pred)) {
+ return new_rd_Proj(dbgi, new_pred, mode_M, pn_ia32_xDiv_M);
+ } else if (is_ia32_vfdiv(new_pred)) {
+ return new_rd_Proj(dbgi, new_pred, mode_M, pn_ia32_vfdiv_M);
+ } else {
+ panic("Div transformed to unexpected thing %+F", new_pred);
}
- break;
- case iro_DivMod:
- switch (proj) {
- case pn_DivMod_M:
- return new_rd_Proj(dbgi, new_pred, mode_M, pn_ia32_Div_M);
- case pn_DivMod_res_div:
+ case pn_Div_res:
+ if (is_ia32_Div(new_pred) || is_ia32_IDiv(new_pred)) {
return new_rd_Proj(dbgi, new_pred, mode_Iu, pn_ia32_Div_div_res);
- case pn_DivMod_res_mod:
- return new_rd_Proj(dbgi, new_pred, mode_Iu, pn_ia32_Div_mod_res);
- case pn_DivMod_X_regular:
- return new_rd_Jmp(dbgi, block);
- case pn_DivMod_X_except:
- set_ia32_exc_label(new_pred, 1);
- return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_Div_X_exc);
- default:
- break;
+ } else if (is_ia32_xDiv(new_pred)) {
+ return new_rd_Proj(dbgi, new_pred, mode_xmm, pn_ia32_xDiv_res);
+ } else if (is_ia32_vfdiv(new_pred)) {
+ return new_rd_Proj(dbgi, new_pred, mode_vfp, pn_ia32_vfdiv_res);
+ } else {
+ panic("Div transformed to unexpected thing %+F", new_pred);
}
- break;
- default:
- break;
+ case pn_Div_X_except:
+ set_ia32_exc_label(new_pred, 1);
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_Div_X_except);
+ case pn_Div_X_regular:
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_Div_X_regular);
}
- panic("No idea how to transform proj->DivMod");
+ panic("No idea how to transform proj->Div");
}
/**
- * Transform and renumber the Projs from a CopyB.
+ * Transform and renumber the Projs from a Div or Mod instruction.
*/
-static ir_node *gen_Proj_CopyB(ir_node *node)
+static ir_node *gen_Proj_Mod(ir_node *node)
{
ir_node *pred = get_Proj_pred(node);
ir_node *new_pred = be_transform_node(pred);
dbg_info *dbgi = get_irn_dbg_info(node);
long proj = get_Proj_proj(node);
- switch (proj) {
- case pn_CopyB_M_regular:
- if (is_ia32_CopyB_i(new_pred)) {
- return new_rd_Proj(dbgi, new_pred, mode_M, pn_ia32_CopyB_i_M);
- } else if (is_ia32_CopyB(new_pred)) {
- return new_rd_Proj(dbgi, new_pred, mode_M, pn_ia32_CopyB_M);
- }
- break;
- default:
- break;
- }
+ assert(is_ia32_Div(new_pred) || is_ia32_IDiv(new_pred));
+ assert((long)pn_ia32_Div_M == (long)pn_ia32_IDiv_M);
+ assert((long)pn_ia32_Div_mod_res == (long)pn_ia32_IDiv_mod_res);
- panic("No idea how to transform proj->CopyB");
+ switch ((pn_Mod)proj) {
+ case pn_Mod_M:
+ return new_rd_Proj(dbgi, new_pred, mode_M, pn_ia32_Div_M);
+ case pn_Mod_res:
+ return new_rd_Proj(dbgi, new_pred, mode_Iu, pn_ia32_Div_mod_res);
+ case pn_Mod_X_except:
+ set_ia32_exc_label(new_pred, 1);
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_Div_X_except);
+ case pn_Mod_X_regular:
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_Div_X_regular);
+ }
+ panic("No idea how to transform proj->Mod");
}
/**
- * Transform and renumber the Projs from a Quot.
+ * Transform and renumber the Projs from a CopyB.
*/
-static ir_node *gen_Proj_Quot(ir_node *node)
+static ir_node *gen_Proj_CopyB(ir_node *node)
{
ir_node *pred = get_Proj_pred(node);
ir_node *new_pred = be_transform_node(pred);
dbg_info *dbgi = get_irn_dbg_info(node);
long proj = get_Proj_proj(node);
- switch (proj) {
- case pn_Quot_M:
- if (is_ia32_xDiv(new_pred)) {
- return new_rd_Proj(dbgi, new_pred, mode_M, pn_ia32_xDiv_M);
- } else if (is_ia32_vfdiv(new_pred)) {
- return new_rd_Proj(dbgi, new_pred, mode_M, pn_ia32_vfdiv_M);
+ switch ((pn_CopyB)proj) {
+ case pn_CopyB_M:
+ if (is_ia32_CopyB_i(new_pred)) {
+ return new_rd_Proj(dbgi, new_pred, mode_M, pn_ia32_CopyB_i_M);
+ } else if (is_ia32_CopyB(new_pred)) {
+ return new_rd_Proj(dbgi, new_pred, mode_M, pn_ia32_CopyB_M);
}
break;
- case pn_Quot_res:
- if (is_ia32_xDiv(new_pred)) {
- return new_rd_Proj(dbgi, new_pred, mode_xmm, pn_ia32_xDiv_res);
- } else if (is_ia32_vfdiv(new_pred)) {
- return new_rd_Proj(dbgi, new_pred, mode_vfp, pn_ia32_vfdiv_res);
+ case pn_CopyB_X_regular:
+ if (is_ia32_CopyB_i(new_pred)) {
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_CopyB_i_X_regular);
+ } else if (is_ia32_CopyB(new_pred)) {
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_CopyB_X_regular);
}
break;
- case pn_Quot_X_regular:
- case pn_Quot_X_except:
- default:
+ case pn_CopyB_X_except:
+ if (is_ia32_CopyB_i(new_pred)) {
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_CopyB_i_X_except);
+ } else if (is_ia32_CopyB(new_pred)) {
+ return new_rd_Proj(dbgi, new_pred, mode_X, pn_ia32_CopyB_X_except);
+ }
break;
}
- panic("No idea how to transform proj->Quot");
+ panic("No idea how to transform proj->CopyB");
}
static ir_node *gen_be_Call(ir_node *node)
dbg_info *const dbgi = get_irn_dbg_info(node);
ir_node *const src_block = get_nodes_block(node);
ir_node *const block = be_transform_node(src_block);
- ir_node *const src_mem = get_irn_n(node, be_pos_Call_mem);
- ir_node *const src_sp = get_irn_n(node, be_pos_Call_sp);
+ ir_node *const src_mem = get_irn_n(node, n_be_Call_mem);
+ ir_node *const src_sp = get_irn_n(node, n_be_Call_sp);
ir_node *const sp = be_transform_node(src_sp);
- ir_node *const src_ptr = get_irn_n(node, be_pos_Call_ptr);
+ ir_node *const src_ptr = get_irn_n(node, n_be_Call_ptr);
ia32_address_mode_t am;
ia32_address_t *const addr = &am.addr;
ir_node * mem;
unsigned const pop = be_Call_get_pop(node);
ir_type *const call_tp = be_Call_get_type(node);
int old_no_pic_adjust;
+ int throws_exception = ir_throws_exception(node);
/* Run the x87 simulator if the call returns a float value */
if (get_method_n_ress(call_tp) > 0) {
ir_mode *const res_mode = get_type_mode(res_type);
if (res_mode != NULL && mode_is_float(res_mode)) {
- env_cg->do_x87_sim = 1;
+ ir_graph *irg = current_ir_graph;
+ ia32_irg_data_t *irg_data = ia32_get_irg_data(irg);
+ irg_data->do_x87_sim = 1;
}
}
assert(be_Call_get_entity(node) == NULL);
/* special case for PIC trampoline calls */
- old_no_pic_adjust = no_pic_adjust;
- no_pic_adjust = be_get_irg_options(env_cg->irg)->pic;
+ old_no_pic_adjust = ia32_no_pic_adjust;
+ ia32_no_pic_adjust = be_options.pic;
match_arguments(&am, src_block, NULL, src_ptr, src_mem,
match_am | match_immediate);
- no_pic_adjust = old_no_pic_adjust;
+ ia32_no_pic_adjust = old_no_pic_adjust;
i = get_irn_arity(node) - 1;
fpcw = be_transform_node(get_irn_n(node, i--));
- for (; i >= be_pos_Call_first_arg; --i) {
- arch_register_req_t const *const req = arch_get_register_req(node, i);
+ for (; i >= n_be_Call_first_arg; --i) {
+ arch_register_req_t const *const req
+ = arch_get_irn_register_req_in(node, i);
ir_node *const reg_parm = be_transform_node(get_irn_n(node, i));
assert(req->type == arch_register_req_type_limited);
assert(req->cls == &ia32_reg_classes[CLASS_ia32_gp]);
switch (*req->limited) {
- case 1 << REG_EAX: assert(eax == noreg_GP); eax = reg_parm; break;
- case 1 << REG_ECX: assert(ecx == noreg_GP); ecx = reg_parm; break;
- case 1 << REG_EDX: assert(edx == noreg_GP); edx = reg_parm; break;
+ case 1 << REG_GP_EAX: assert(eax == noreg_GP); eax = reg_parm; break;
+ case 1 << REG_GP_ECX: assert(ecx == noreg_GP); ecx = reg_parm; break;
+ case 1 << REG_GP_EDX: assert(edx == noreg_GP); edx = reg_parm; break;
default: panic("Invalid GP register for register parameter");
}
}
mem = transform_AM_mem(block, src_ptr, src_mem, addr->mem);
call = new_bd_ia32_Call(dbgi, block, addr->base, addr->index, mem,
am.new_op2, sp, fpcw, eax, ecx, edx, pop, call_tp);
+ ir_set_throws_exception(call, throws_exception);
set_am_attributes(call, &am);
call = fix_mem_proj(call, &am);
*/
static ir_node *gen_return_address(ir_node *node)
{
- ir_node *param = get_Builtin_param(node, 0);
- ir_node *frame = get_Builtin_param(node, 1);
- dbg_info *dbgi = get_irn_dbg_info(node);
- tarval *tv = get_Const_tarval(param);
+ ir_node *param = get_Builtin_param(node, 0);
+ ir_node *frame = get_Builtin_param(node, 1);
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ ir_tarval *tv = get_Const_tarval(param);
+ ir_graph *irg = get_irn_irg(node);
unsigned long value = get_tarval_long(tv);
ir_node *block = be_transform_node(get_nodes_block(node));
set_ia32_am_offs_int(load, 0);
set_ia32_use_frame(load);
- set_ia32_frame_ent(load, ia32_get_return_address_entity());
+ set_ia32_frame_ent(load, ia32_get_return_address_entity(irg));
if (get_irn_pinned(node) == op_pin_state_floats) {
- assert(pn_ia32_xLoad_res == pn_ia32_vfld_res
- && pn_ia32_vfld_res == pn_ia32_Load_res
- && pn_ia32_Load_res == pn_ia32_res);
- arch_irn_add_flags(load, arch_irn_flags_rematerializable);
+ assert((int)pn_ia32_xLoad_res == (int)pn_ia32_vfld_res
+ && (int)pn_ia32_vfld_res == (int)pn_ia32_Load_res
+ && (int)pn_ia32_Load_res == (int)pn_ia32_res);
+ arch_add_irn_flags(load, arch_irn_flags_rematerializable);
}
SET_IA32_ORIG_NODE(load, node);
*/
static ir_node *gen_frame_address(ir_node *node)
{
- ir_node *param = get_Builtin_param(node, 0);
- ir_node *frame = get_Builtin_param(node, 1);
- dbg_info *dbgi = get_irn_dbg_info(node);
- tarval *tv = get_Const_tarval(param);
+ ir_node *param = get_Builtin_param(node, 0);
+ ir_node *frame = get_Builtin_param(node, 1);
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ ir_tarval *tv = get_Const_tarval(param);
+ ir_graph *irg = get_irn_irg(node);
unsigned long value = get_tarval_long(tv);
ir_node *block = be_transform_node(get_nodes_block(node));
set_ia32_op_type(load, ia32_AddrModeS);
set_ia32_ls_mode(load, mode_Iu);
- ent = ia32_get_frame_address_entity();
+ ent = ia32_get_frame_address_entity(irg);
if (ent != NULL) {
set_ia32_am_offs_int(load, 0);
set_ia32_use_frame(load);
}
if (get_irn_pinned(node) == op_pin_state_floats) {
- assert(pn_ia32_xLoad_res == pn_ia32_vfld_res
- && pn_ia32_vfld_res == pn_ia32_Load_res
- && pn_ia32_Load_res == pn_ia32_res);
- arch_irn_add_flags(load, arch_irn_flags_rematerializable);
+ assert((int)pn_ia32_xLoad_res == (int)pn_ia32_vfld_res
+ && (int)pn_ia32_vfld_res == (int)pn_ia32_Load_res
+ && (int)pn_ia32_Load_res == (int)pn_ia32_res);
+ arch_add_irn_flags(load, arch_irn_flags_rematerializable);
}
SET_IA32_ORIG_NODE(load, node);
static ir_node *gen_prefetch(ir_node *node)
{
dbg_info *dbgi;
- ir_node *ptr, *block, *mem, *base, *index;
+ ir_node *ptr, *block, *mem, *base, *idx;
ir_node *param, *new_node;
long rw, locality;
- tarval *tv;
+ ir_tarval *tv;
ia32_address_t addr;
if (!ia32_cg_config.use_sse_prefetch && !ia32_cg_config.use_3dnow_prefetch) {
/* construct load address */
memset(&addr, 0, sizeof(addr));
ptr = get_Builtin_param(node, 0);
- ia32_create_address_mode(&addr, ptr, 0);
- base = addr.base;
- index = addr.index;
+ ia32_create_address_mode(&addr, ptr, ia32_create_am_normal);
+ base = addr.base;
+ idx = addr.index;
if (base == NULL) {
base = noreg_GP;
base = be_transform_node(base);
}
- if (index == NULL) {
- index = noreg_GP;
+ if (idx == NULL) {
+ idx = noreg_GP;
} else {
- index = be_transform_node(index);
+ idx = be_transform_node(idx);
}
dbgi = get_irn_dbg_info(node);
if (rw == 1 && ia32_cg_config.use_3dnow_prefetch) {
/* we have 3DNow!, this was already checked above */
- new_node = new_bd_ia32_PrefetchW(dbgi, block, base, index, mem);
+ new_node = new_bd_ia32_PrefetchW(dbgi, block, base, idx, mem);
} else if (ia32_cg_config.use_sse_prefetch) {
/* note: rw == 1 is IGNORED in that case */
param = get_Builtin_param(node, 2);
/* SSE style prefetch */
switch (locality) {
case 0:
- new_node = new_bd_ia32_PrefetchNTA(dbgi, block, base, index, mem);
+ new_node = new_bd_ia32_PrefetchNTA(dbgi, block, base, idx, mem);
break;
case 1:
- new_node = new_bd_ia32_Prefetch2(dbgi, block, base, index, mem);
+ new_node = new_bd_ia32_Prefetch2(dbgi, block, base, idx, mem);
break;
case 2:
- new_node = new_bd_ia32_Prefetch1(dbgi, block, base, index, mem);
+ new_node = new_bd_ia32_Prefetch1(dbgi, block, base, idx, mem);
break;
default:
- new_node = new_bd_ia32_Prefetch0(dbgi, block, base, index, mem);
+ new_node = new_bd_ia32_Prefetch0(dbgi, block, base, idx, mem);
break;
}
} else {
assert(ia32_cg_config.use_3dnow_prefetch);
/* 3DNow! style prefetch */
- new_node = new_bd_ia32_Prefetch(dbgi, block, base, index, mem);
+ new_node = new_bd_ia32_Prefetch(dbgi, block, base, idx, mem);
}
set_irn_pinned(new_node, get_irn_pinned(node));
SET_IA32_ORIG_NODE(new_node, node);
- be_dep_on_frame(new_node);
return new_r_Proj(new_node, mode_M, pn_ia32_Prefetch_M);
}
ir_node *real = skip_Proj(bsf);
dbg_info *dbgi = get_irn_dbg_info(real);
ir_node *block = get_nodes_block(real);
- ir_node *flag, *set, *conv, *neg, *or;
+ ir_node *flag, *set, *conv, *neg, *orn, *add;
/* bsf x */
if (get_irn_mode(real) != mode_T) {
flag = new_r_Proj(real, mode_b, pn_ia32_flags);
/* sete */
- set = new_bd_ia32_Setcc(dbgi, block, flag, pn_Cmp_Eq);
+ set = new_bd_ia32_Setcc(dbgi, block, flag, ia32_cc_equal);
SET_IA32_ORIG_NODE(set, node);
/* conv to 32bit */
neg = new_bd_ia32_Neg(dbgi, block, conv);
/* or */
- or = new_bd_ia32_Or(dbgi, block, noreg_GP, noreg_GP, nomem, bsf, neg);
- set_ia32_commutative(or);
+ orn = new_bd_ia32_Or(dbgi, block, noreg_GP, noreg_GP, nomem, bsf, neg);
+ set_ia32_ls_mode(orn, mode_Iu);
+ set_ia32_commutative(orn);
/* add 1 */
- return new_bd_ia32_Add(dbgi, block, noreg_GP, noreg_GP, nomem, or, ia32_create_Immediate(NULL, 0, 1));
+ add = new_bd_ia32_Lea(dbgi, block, orn, noreg_GP);
+ add_ia32_am_offs_int(add, 1);
+ return add;
}
/**
*/
static ir_node *gen_parity(ir_node *node)
{
- ir_node *param = get_Builtin_param(node, 0);
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 *param = get_Builtin_param(node, 0);
+ ir_node *new_param = be_transform_node(param);
+ ir_node *new_node;
- ir_node *block = get_nodes_block(node);
-
- ir_node *new_block = be_transform_node(block);
- ir_node *imm, *cmp, *new_node;
-
- ia32_address_mode_t am;
- ia32_address_t *addr = &am.addr;
-
-
- /* cmp param, 0 */
- match_arguments(&am, block, NULL, param, NULL, match_am);
- imm = ia32_create_Immediate(NULL, 0, 0);
- cmp = new_bd_ia32_Cmp(dbgi, new_block, addr->base, addr->index,
- addr->mem, imm, am.new_op2, am.ins_permuted, 0);
- set_am_attributes(cmp, &am);
- set_ia32_ls_mode(cmp, mode_Iu);
+ /* the x86 parity bit is stupid: it only looks at the lowest byte,
+ * so we have to do complicated xoring first.
+ * (we should also better lower this before the backend so we still have a
+ * chance for CSE, constant folding and other goodies for some of these
+ * operations)
+ */
+ ir_node *count = ia32_create_Immediate(NULL, 0, 16);
+ ir_node *shr = new_bd_ia32_Shr(dbgi, new_block, new_param, count);
+ ir_node *xor = new_bd_ia32_Xor(dbgi, new_block, noreg_GP, noreg_GP, nomem,
+ shr, new_param);
+ ir_node *xor2 = new_bd_ia32_XorHighLow(dbgi, new_block, xor);
+ ir_node *flags;
- SET_IA32_ORIG_NODE(cmp, node);
+ set_ia32_ls_mode(xor, mode_Iu);
+ set_ia32_commutative(xor);
- cmp = fix_mem_proj(cmp, &am);
+ set_irn_mode(xor2, mode_T);
+ flags = new_r_Proj(xor2, mode_Iu, pn_ia32_XorHighLow_flags);
/* setp */
- new_node = new_bd_ia32_Setcc(dbgi, new_block, cmp, ia32_pn_Cmp_parity);
+ new_node = new_bd_ia32_Setcc(dbgi, new_block, flags, ia32_cc_not_parity);
SET_IA32_ORIG_NODE(new_node, node);
/* conv to 32bit */
new_param = be_transform_node(param);
/* do the standard popcount algo */
+ /* TODO: This is stupid, we should transform this before the backend,
+ * to get CSE, localopts, etc. for the operations
+ * TODO: This is also not the optimal algorithm (it is just the starting
+ * example in hackers delight, they optimize it more on the following page)
+ * But I'm too lazy to fix this now, as the code should get lowered before
+ * the backend anyway.
+ */
/* m1 = x & 0x55555555 */
imm = ia32_create_Immediate(NULL, 0, 0x55555555);
/* s1 = x >> 1 */
simm = ia32_create_Immediate(NULL, 0, 1);
- s1 = new_bd_ia32_Shl(dbgi, new_block, new_param, simm);
+ s1 = new_bd_ia32_Shr(dbgi, new_block, new_param, simm);
/* m2 = s1 & 0x55555555 */
m2 = new_bd_ia32_And(dbgi, new_block, noreg_GP, noreg_GP, nomem, s1, imm);
/* s2 = m3 >> 2 */
simm = ia32_create_Immediate(NULL, 0, 2);
- s2 = new_bd_ia32_Shl(dbgi, new_block, m3, simm);
+ s2 = new_bd_ia32_Shr(dbgi, new_block, m3, simm);
/* m5 = s2 & 0x33333333 */
m5 = new_bd_ia32_And(dbgi, new_block, noreg_GP, noreg_GP, nomem, s2, imm);
/* s3 = m6 >> 4 */
simm = ia32_create_Immediate(NULL, 0, 4);
- s3 = new_bd_ia32_Shl(dbgi, new_block, m6, simm);
+ s3 = new_bd_ia32_Shr(dbgi, new_block, m6, simm);
/* m8 = s3 & 0x0F0F0F0F */
m8 = new_bd_ia32_And(dbgi, new_block, noreg_GP, noreg_GP, nomem, s3, imm);
/* s4 = m9 >> 8 */
simm = ia32_create_Immediate(NULL, 0, 8);
- s4 = new_bd_ia32_Shl(dbgi, new_block, m9, simm);
+ s4 = new_bd_ia32_Shr(dbgi, new_block, m9, simm);
/* m11 = s4 & 0x00FF00FF */
m11 = new_bd_ia32_And(dbgi, new_block, noreg_GP, noreg_GP, nomem, s4, imm);
/* s5 = m12 >> 16 */
simm = ia32_create_Immediate(NULL, 0, 16);
- s5 = new_bd_ia32_Shl(dbgi, new_block, m12, simm);
+ s5 = new_bd_ia32_Shr(dbgi, new_block, m12, simm);
/* res = m13 + s5 */
return new_bd_ia32_Lea(dbgi, new_block, m13, s5);
ir_node *new_block = be_transform_node(block);
ir_mode *mode = get_irn_mode(param);
unsigned size = get_mode_size_bits(mode);
- ir_node *m1, *m2, *m3, *m4, *s1, *s2, *s3, *s4;
switch (size) {
case 32:
- if (ia32_cg_config.use_i486) {
+ if (ia32_cg_config.use_bswap) {
/* swap available */
return new_bd_ia32_Bswap(dbgi, new_block, param);
+ } else {
+ ir_node *i8 = ia32_create_Immediate(NULL, 0, 8);
+ ir_node *rol1 = new_bd_ia32_Rol(dbgi, new_block, param, i8);
+ ir_node *i16 = ia32_create_Immediate(NULL, 0, 16);
+ ir_node *rol2 = new_bd_ia32_Rol(dbgi, new_block, rol1, i16);
+ ir_node *rol3 = new_bd_ia32_Rol(dbgi, new_block, rol2, i8);
+ set_ia32_ls_mode(rol1, mode_Hu);
+ set_ia32_ls_mode(rol2, mode_Iu);
+ set_ia32_ls_mode(rol3, mode_Hu);
+ return rol3;
}
- s1 = new_bd_ia32_Shl(dbgi, new_block, param, ia32_create_Immediate(NULL, 0, 24));
- s2 = new_bd_ia32_Shl(dbgi, new_block, param, ia32_create_Immediate(NULL, 0, 8));
-
- m1 = new_bd_ia32_And(dbgi, new_block, noreg_GP, noreg_GP, nomem, s2, ia32_create_Immediate(NULL, 0, 0xFF00));
- m2 = new_bd_ia32_Lea(dbgi, new_block, s1, m1);
-
- s3 = new_bd_ia32_Shr(dbgi, new_block, param, ia32_create_Immediate(NULL, 0, 8));
-
- m3 = new_bd_ia32_And(dbgi, new_block, noreg_GP, noreg_GP, nomem, s3, ia32_create_Immediate(NULL, 0, 0xFF0000));
- m4 = new_bd_ia32_Lea(dbgi, new_block, m2, m3);
-
- s4 = new_bd_ia32_Shr(dbgi, new_block, param, ia32_create_Immediate(NULL, 0, 24));
- return new_bd_ia32_Lea(dbgi, new_block, m4, s4);
case 16:
/* swap16 always available */
/* construct store address */
memset(&addr, 0, sizeof(addr));
- ia32_create_address_mode(&addr, ptr, 0);
+ ia32_create_address_mode(&addr, ptr, ia32_create_am_normal);
if (addr.base == NULL) {
addr.base = noreg_GP;
if (is_SymConst(callee)) {
rel = new_bd_ia32_Const(dbgi, new_block, get_SymConst_entity(callee), 0, 0, -10);
} else {
- rel = new_bd_ia32_Lea(dbgi, new_block, be_transform_node(callee), ia32_create_Immediate(NULL, 0, -10));
+ rel = new_bd_ia32_Lea(dbgi, new_block, be_transform_node(callee), noreg_GP);
+ add_ia32_am_offs_int(rel, -10);
}
rel = new_bd_ia32_Sub(dbgi, new_block, noreg_GP, noreg_GP, nomem, rel, trampoline);
case ir_bk_parity:
case ir_bk_popcount:
case ir_bk_bswap:
- assert(get_Proj_proj(proj) == pn_Builtin_1_result);
+ assert(get_Proj_proj(proj) == pn_Builtin_max+1);
return new_node;
case ir_bk_trap:
case ir_bk_debugbreak:
assert(get_Proj_proj(proj) == pn_Builtin_M);
return new_node;
case ir_bk_inport:
- if (get_Proj_proj(proj) == pn_Builtin_1_result) {
+ if (get_Proj_proj(proj) == pn_Builtin_max+1) {
return new_r_Proj(new_node, get_irn_mode(proj), pn_ia32_Inport_res);
} else {
assert(get_Proj_proj(proj) == pn_Builtin_M);
return new_r_Proj(new_node, mode_M, pn_ia32_Inport_M);
}
case ir_bk_inner_trampoline:
- if (get_Proj_proj(proj) == pn_Builtin_1_result) {
+ if (get_Proj_proj(proj) == pn_Builtin_max+1) {
return get_Tuple_pred(new_node, 1);
} else {
assert(get_Proj_proj(proj) == pn_Builtin_M);
static ir_node *gen_be_IncSP(ir_node *node)
{
ir_node *res = be_duplicate_node(node);
- arch_irn_add_flags(res, arch_irn_flags_modify_flags);
+ arch_add_irn_flags(res, arch_irn_flags_modify_flags);
return res;
}
ir_mode *mode = get_irn_mode(node);
ir_node *res;
- if (proj == pn_be_Call_M_regular) {
+ if (proj == pn_be_Call_M) {
return new_rd_Proj(dbgi, new_call, mode_M, n_ia32_Call_mem);
}
/* transform call modes */
if (mode_is_data(mode)) {
- const arch_register_class_t *cls = arch_get_irn_reg_class_out(node);
+ const arch_register_class_t *cls = arch_get_irn_reg_class(node);
mode = cls->mode;
}
/* Map from be_Call to ia32_Call proj number */
if (proj == pn_be_Call_sp) {
proj = pn_ia32_Call_stack;
- } else if (proj == pn_be_Call_M_regular) {
+ } else if (proj == pn_be_Call_M) {
proj = pn_ia32_Call_M;
+ } else if (proj == pn_be_Call_X_except) {
+ proj = pn_ia32_Call_X_except;
+ } else if (proj == pn_be_Call_X_regular) {
+ proj = pn_ia32_Call_X_regular;
} else {
- arch_register_req_t const *const req = arch_get_register_req_out(node);
- int const n_outs = arch_irn_get_n_outs(new_call);
+ arch_register_req_t const *const req = arch_get_irn_register_req(node);
+ int const n_outs = arch_get_irn_n_outs(new_call);
int i;
assert(proj >= pn_be_Call_first_res);
for (i = 0; i < n_outs; ++i) {
arch_register_req_t const *const new_req
- = arch_get_out_register_req(new_call, i);
+ = arch_get_irn_register_req_out(new_call, i);
if (!(new_req->type & arch_register_req_type_limited) ||
new_req->cls != req->cls ||
/* TODO arch_set_irn_register() only operates on Projs, need variant with index */
switch (proj) {
- case pn_ia32_Call_stack:
- arch_set_irn_register(res, &ia32_gp_regs[REG_ESP]);
- break;
+ case pn_ia32_Call_stack:
+ arch_set_irn_register(res, &ia32_registers[REG_ESP]);
+ break;
- case pn_ia32_Call_fpcw:
- arch_set_irn_register(res, &ia32_fp_cw_regs[REG_FPCW]);
- break;
+ case pn_ia32_Call_fpcw:
+ arch_set_irn_register(res, &ia32_registers[REG_FPCW]);
+ break;
}
return res;
node);
}
-/**
- * Transform the Projs from a Bound.
- */
-static ir_node *gen_Proj_Bound(ir_node *node)
-{
- ir_node *new_node;
- ir_node *pred = get_Proj_pred(node);
-
- switch (get_Proj_proj(node)) {
- case pn_Bound_M:
- return be_transform_node(get_Bound_mem(pred));
- case pn_Bound_X_regular:
- new_node = be_transform_node(pred);
- return new_r_Proj(new_node, mode_X, pn_ia32_Jcc_true);
- case pn_Bound_X_except:
- new_node = be_transform_node(pred);
- return new_r_Proj(new_node, mode_X, pn_ia32_Jcc_false);
- case pn_Bound_res:
- return be_transform_node(get_Bound_index(pred));
- default:
- panic("unsupported Proj from Bound");
- }
-}
-
static ir_node *gen_Proj_ASM(ir_node *node)
{
ir_mode *mode = get_irn_mode(node);
long pos = get_Proj_proj(node);
if (mode == mode_M) {
- pos = arch_irn_get_n_outs(new_pred)-1;
+ pos = arch_get_irn_n_outs(new_pred)-1;
} else if (mode_is_int(mode) || mode_is_reference(mode)) {
mode = mode_Iu;
} else if (mode_is_float(mode)) {
- mode = mode_E;
+ mode = ia32_mode_E;
} else {
panic("unexpected proj mode at ASM");
}
long proj;
switch (get_irn_opcode(pred)) {
- case iro_Store:
- proj = get_Proj_proj(node);
- if (proj == pn_Store_M) {
- return be_transform_node(pred);
- } else {
- panic("No idea how to transform proj->Store");
- }
case iro_Load:
return gen_Proj_Load(node);
+ case iro_Store:
+ return gen_Proj_Store(node);
case iro_ASM:
return gen_Proj_ASM(node);
case iro_Builtin:
return gen_Proj_Builtin(node);
case iro_Div:
+ return gen_Proj_Div(node);
case iro_Mod:
- case iro_DivMod:
- return gen_Proj_DivMod(node);
+ return gen_Proj_Mod(node);
case iro_CopyB:
return gen_Proj_CopyB(node);
- case iro_Quot:
- return gen_Proj_Quot(node);
case beo_SubSP:
return gen_Proj_be_SubSP(node);
case beo_AddSP:
return gen_Proj_be_Call(node);
case iro_Cmp:
return gen_Proj_Cmp(node);
- case iro_Bound:
- return gen_Proj_Bound(node);
case iro_Start:
proj = get_Proj_proj(node);
switch (proj) {
return jump;
}
-
- case pn_Start_P_tls:
- return gen_Proj_tls(node);
}
break;
/* first clear the generic function pointer for all ops */
be_start_transform_setup();
- be_set_transform_function(op_Abs, gen_Abs);
be_set_transform_function(op_Add, gen_Add);
be_set_transform_function(op_And, gen_And);
- be_set_transform_function(op_ASM, gen_ASM);
+ be_set_transform_function(op_ASM, ia32_gen_ASM);
be_set_transform_function(op_be_AddSP, gen_be_AddSP);
be_set_transform_function(op_be_Call, gen_be_Call);
be_set_transform_function(op_be_Copy, gen_be_Copy);
be_set_transform_function(op_be_IncSP, gen_be_IncSP);
be_set_transform_function(op_be_Return, gen_be_Return);
be_set_transform_function(op_be_SubSP, gen_be_SubSP);
- be_set_transform_function(op_Bound, gen_Bound);
be_set_transform_function(op_Builtin, gen_Builtin);
be_set_transform_function(op_Cmp, gen_Cmp);
be_set_transform_function(op_Cond, gen_Cond);
be_set_transform_function(op_Const, gen_Const);
be_set_transform_function(op_Conv, gen_Conv);
- be_set_transform_function(op_CopyB, gen_CopyB);
+ be_set_transform_function(op_CopyB, ia32_gen_CopyB);
be_set_transform_function(op_Div, gen_Div);
- be_set_transform_function(op_DivMod, gen_DivMod);
be_set_transform_function(op_Eor, gen_Eor);
be_set_transform_function(op_ia32_l_Adc, gen_ia32_l_Adc);
be_set_transform_function(op_ia32_l_Add, gen_ia32_l_Add);
be_set_transform_function(op_ia32_l_IMul, gen_ia32_l_IMul);
be_set_transform_function(op_ia32_l_LLtoFloat, gen_ia32_l_LLtoFloat);
be_set_transform_function(op_ia32_l_Mul, gen_ia32_l_Mul);
- be_set_transform_function(op_ia32_l_SarDep, gen_ia32_l_SarDep);
be_set_transform_function(op_ia32_l_Sbb, gen_ia32_l_Sbb);
- be_set_transform_function(op_ia32_l_ShlDep, gen_ia32_l_ShlDep);
- be_set_transform_function(op_ia32_l_ShlD, gen_ia32_l_ShlD);
- be_set_transform_function(op_ia32_l_ShrDep, gen_ia32_l_ShrDep);
- be_set_transform_function(op_ia32_l_ShrD, gen_ia32_l_ShrD);
be_set_transform_function(op_ia32_l_Sub, gen_ia32_l_Sub);
+ be_set_transform_function(op_ia32_GetEIP, be_duplicate_node);
be_set_transform_function(op_ia32_Minus64Bit, be_duplicate_node);
be_set_transform_function(op_ia32_NoReg_GP, be_duplicate_node);
be_set_transform_function(op_ia32_NoReg_VFP, be_duplicate_node);
be_set_transform_function(op_Or, gen_Or);
be_set_transform_function(op_Phi, gen_Phi);
be_set_transform_function(op_Proj, gen_Proj);
- be_set_transform_function(op_Quot, gen_Quot);
be_set_transform_function(op_Rotl, gen_Rotl);
be_set_transform_function(op_Shl, gen_Shl);
be_set_transform_function(op_Shr, gen_Shr);
be_set_transform_function(op_Shrs, gen_Shrs);
be_set_transform_function(op_Store, gen_Store);
be_set_transform_function(op_Sub, gen_Sub);
+ be_set_transform_function(op_Switch, gen_Switch);
be_set_transform_function(op_SymConst, gen_SymConst);
- be_set_transform_function(op_Unknown, gen_Unknown);
+ be_set_transform_function(op_Unknown, ia32_gen_Unknown);
}
/**
*/
static void ia32_pretransform_node(void)
{
- ia32_code_gen_t *cg = env_cg;
+ ir_graph *irg = current_ir_graph;
+ ia32_irg_data_t *irg_data = ia32_get_irg_data(current_ir_graph);
- cg->noreg_gp = be_pre_transform_node(cg->noreg_gp);
- cg->noreg_vfp = be_pre_transform_node(cg->noreg_vfp);
- cg->noreg_xmm = be_pre_transform_node(cg->noreg_xmm);
+ irg_data->noreg_gp = be_pre_transform_node(irg_data->noreg_gp);
+ irg_data->noreg_vfp = be_pre_transform_node(irg_data->noreg_vfp);
+ irg_data->noreg_xmm = be_pre_transform_node(irg_data->noreg_xmm);
+ irg_data->get_eip = be_pre_transform_node(irg_data->get_eip);
+ irg_data->fpu_trunc_mode = be_pre_transform_node(irg_data->fpu_trunc_mode);
- nomem = get_irg_no_mem(current_ir_graph);
- noreg_GP = ia32_new_NoReg_gp(cg);
-
- get_fpcw();
-}
-
-/**
- * Walker, checks if all ia32 nodes producing more than one result have their
- * Projs, otherwise creates new Projs and keeps them using a be_Keep node.
- */
-static void add_missing_keep_walker(ir_node *node, void *data)
-{
- int n_outs, i;
- unsigned found_projs = 0;
- const ir_edge_t *edge;
- ir_mode *mode = get_irn_mode(node);
- ir_node *last_keep;
- (void) data;
- if (mode != mode_T)
- return;
- if (!is_ia32_irn(node))
- return;
-
- n_outs = arch_irn_get_n_outs(node);
- if (n_outs <= 0)
- return;
- if (is_ia32_SwitchJmp(node))
- return;
-
- assert(n_outs < (int) sizeof(unsigned) * 8);
- foreach_out_edge(node, edge) {
- ir_node *proj = get_edge_src_irn(edge);
- int pn;
-
- /* The node could be kept */
- if (is_End(proj))
- continue;
-
- if (get_irn_mode(proj) == mode_M)
- continue;
-
- pn = get_Proj_proj(proj);
- assert(pn < n_outs);
- found_projs |= 1 << pn;
- }
-
-
- /* are keeps missing? */
- last_keep = NULL;
- for (i = 0; i < n_outs; ++i) {
- ir_node *block;
- ir_node *in[1];
- const arch_register_req_t *req;
- const arch_register_class_t *cls;
-
- if (found_projs & (1 << i)) {
- continue;
- }
-
- req = arch_get_out_register_req(node, i);
- cls = req->cls;
- if (cls == NULL) {
- continue;
- }
- if (cls == &ia32_reg_classes[CLASS_ia32_flags]) {
- continue;
- }
-
- block = get_nodes_block(node);
- in[0] = new_r_Proj(node, arch_register_class_mode(cls), i);
- if (last_keep != NULL) {
- be_Keep_add_node(last_keep, cls, in[0]);
- } else {
- last_keep = be_new_Keep(block, 1, in);
- if (sched_is_scheduled(node)) {
- sched_add_after(node, last_keep);
- }
- }
- }
-}
-
-/**
- * Adds missing keeps to nodes. Adds missing Proj nodes for unused outputs
- * and keeps them.
- */
-void ia32_add_missing_keeps(ia32_code_gen_t *cg)
-{
- irg_walk_graph(cg->irg, add_missing_keep_walker, NULL, NULL);
+ nomem = get_irg_no_mem(irg);
+ noreg_GP = ia32_new_NoReg_gp(irg);
}
/**
*/
static void postprocess_fp_call_results(void)
{
- int i;
+ size_t i, n;
- for (i = ARR_LEN(call_list) - 1; i >= 0; --i) {
+ for (i = 0, n = ARR_LEN(call_list); i < n; ++i) {
ir_node *call = call_list[i];
ir_type *mtp = call_types[i];
int j;
ir_type *res_tp = get_method_res_type(mtp, j);
ir_node *res, *new_res;
const ir_edge_t *edge, *next;
- ir_mode *mode;
+ ir_mode *res_mode;
if (! is_atomic_type(res_tp)) {
/* no floating point return */
continue;
}
- mode = get_type_mode(res_tp);
- if (! mode_is_float(mode)) {
+ res_mode = get_type_mode(res_tp);
+ if (! mode_is_float(res_mode)) {
/* no floating point return */
continue;
}
dbg_info *db = get_irn_dbg_info(succ);
ir_node *block = get_nodes_block(succ);
ir_node *base = get_irn_n(succ, n_ia32_xStore_base);
- ir_node *index = get_irn_n(succ, n_ia32_xStore_index);
+ ir_node *idx = get_irn_n(succ, n_ia32_xStore_index);
ir_node *mem = get_irn_n(succ, n_ia32_xStore_mem);
ir_node *value = get_irn_n(succ, n_ia32_xStore_val);
ir_mode *mode = get_ia32_ls_mode(succ);
- ir_node *st = new_bd_ia32_vfst(db, block, base, index, mem, value, mode);
+ ir_node *st = new_bd_ia32_vfst(db, block, base, idx, mem, value, mode);
+ //ir_node *mem = new_r_Proj(st, mode_M, pn_ia32_vfst_M);
set_ia32_am_offs_int(st, get_ia32_am_offs_int(succ));
if (is_ia32_use_frame(succ))
set_ia32_use_frame(st);
set_irn_pinned(st, get_irn_pinned(succ));
set_ia32_op_type(st, ia32_AddrModeD);
+ assert((long)pn_ia32_xStore_M == (long)pn_ia32_vfst_M);
+ assert((long)pn_ia32_xStore_X_regular == (long)pn_ia32_vfst_X_regular);
+ assert((long)pn_ia32_xStore_X_except == (long)pn_ia32_vfst_X_except);
+
exchange(succ, st);
- } else {
- if (new_res == NULL) {
- dbg_info *db = get_irn_dbg_info(call);
- ir_node *block = get_nodes_block(call);
- ir_node *frame = get_irg_frame(current_ir_graph);
- ir_node *old_mem = be_get_Proj_for_pn(call, pn_ia32_Call_M);
- ir_node *call_mem = new_r_Proj(call, mode_M, pn_ia32_Call_M);
- ir_node *vfst, *xld, *new_mem;
-
- /* store st(0) on stack */
- vfst = new_bd_ia32_vfst(db, block, frame, noreg_GP, call_mem, res, mode);
- set_ia32_op_type(vfst, ia32_AddrModeD);
- set_ia32_use_frame(vfst);
-
- /* load into SSE register */
- xld = new_bd_ia32_xLoad(db, block, frame, noreg_GP, vfst, mode);
- set_ia32_op_type(xld, ia32_AddrModeS);
- set_ia32_use_frame(xld);
-
- new_res = new_r_Proj(xld, mode, pn_ia32_xLoad_res);
- new_mem = new_r_Proj(xld, mode_M, pn_ia32_xLoad_M);
-
- if (old_mem != NULL) {
- edges_reroute(old_mem, new_mem, current_ir_graph);
- kill_node(old_mem);
- }
+
+ continue;
+ }
+
+ if (new_res == NULL) {
+ dbg_info *db = get_irn_dbg_info(call);
+ ir_node *block = get_nodes_block(call);
+ ir_node *frame = get_irg_frame(current_ir_graph);
+ ir_node *old_mem = be_get_Proj_for_pn(call, pn_ia32_Call_M);
+ ir_node *call_mem = new_r_Proj(call, mode_M, pn_ia32_Call_M);
+ ir_node *vfst, *xld, *new_mem;
+ ir_node *vfst_mem;
+
+ /* store st(0) on stack */
+ vfst = new_bd_ia32_vfst(db, block, frame, noreg_GP, call_mem,
+ res, res_mode);
+ set_ia32_op_type(vfst, ia32_AddrModeD);
+ set_ia32_use_frame(vfst);
+
+ vfst_mem = new_r_Proj(vfst, mode_M, pn_ia32_vfst_M);
+
+ /* load into SSE register */
+ xld = new_bd_ia32_xLoad(db, block, frame, noreg_GP, vfst_mem,
+ res_mode);
+ set_ia32_op_type(xld, ia32_AddrModeS);
+ set_ia32_use_frame(xld);
+
+ new_res = new_r_Proj(xld, res_mode, pn_ia32_xLoad_res);
+ new_mem = new_r_Proj(xld, mode_M, pn_ia32_xLoad_M);
+
+ if (old_mem != NULL) {
+ edges_reroute(old_mem, new_mem);
+ kill_node(old_mem);
}
- set_irn_n(succ, get_edge_src_pos(edge), new_res);
}
+ set_irn_n(succ, get_edge_src_pos(edge), new_res);
}
}
}
}
/* do the transformation */
-void ia32_transform_graph(ia32_code_gen_t *cg)
+void ia32_transform_graph(ir_graph *irg)
{
int cse_last;
register_transformers();
- env_cg = cg;
- initial_fpcw = NULL;
- no_pic_adjust = 0;
+ initial_fpcw = NULL;
+ ia32_no_pic_adjust = 0;
+
+ old_initial_fpcw = be_get_initial_reg_value(irg, &ia32_registers[REG_FPCW]);
be_timer_push(T_HEIGHTS);
- heights = heights_new(cg->irg);
+ ia32_heights = heights_new(irg);
be_timer_pop(T_HEIGHTS);
- ia32_calculate_non_address_mode_nodes(cg->irg);
+ ia32_calculate_non_address_mode_nodes(irg);
/* the transform phase is not safe for CSE (yet) because several nodes get
* attributes set after their creation */
call_list = NEW_ARR_F(ir_node *, 0);
call_types = NEW_ARR_F(ir_type *, 0);
- be_transform_graph(cg->irg, ia32_pretransform_node);
+ be_transform_graph(irg, ia32_pretransform_node);
if (ia32_cg_config.use_sse2)
postprocess_fp_call_results();
set_opt_cse(cse_last);
ia32_free_non_address_mode_nodes();
- heights_free(heights);
- heights = NULL;
+ heights_free(ia32_heights);
+ ia32_heights = NULL;
}
void ia32_init_transform(void)