/**
* @file
- * @brief This file implements the IR transformation from firm into ia32-Firm.
+ * @brief This file implements the IR transformation from firm into
+ * ia32-Firm.
* @author Christian Wuerdig, Matthias Braun
* @version $Id$
*/
#include "irvrfy.h"
#include "ircons.h"
#include "irgwalk.h"
-#include "dbginfo.h"
#include "irprintf.h"
#include "debug.h"
#include "irdom.h"
#include "archop.h"
#include "error.h"
+#include "height.h"
#include "../benode_t.h"
#include "../besched.h"
#include "ia32_dbg_stat.h"
#include "ia32_optimize.h"
#include "ia32_util.h"
+#include "ia32_address_mode.h"
#include "gen_ia32_regalloc_if.h"
DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
/** hold the current code generator during transformation */
-static ia32_code_gen_t *env_cg = NULL;
+static ia32_code_gen_t *env_cg = NULL;
+static ir_node *initial_fpcw = NULL;
+static heights_t *heights = NULL;
extern ir_op *get_op_Mulh(void);
ir_node *block, ir_node *base, ir_node *index, ir_node *op1,
ir_node *op2, ir_node *mem);
+typedef ir_node *construct_shift_func(dbg_info *db, ir_graph *irg,
+ ir_node *block, ir_node *op1, ir_node *op2);
+
+typedef ir_node *construct_binop_dest_func(dbg_info *db, ir_graph *irg,
+ ir_node *block, ir_node *base, ir_node *index, ir_node *op,
+ ir_node *mem);
+
+typedef ir_node *construct_unop_dest_func(dbg_info *db, ir_graph *irg,
+ ir_node *block, ir_node *base, ir_node *index, ir_node *mem);
+
typedef ir_node *construct_binop_float_func(dbg_info *db, ir_graph *irg,
ir_node *block, ir_node *base, ir_node *index, ir_node *op1,
ir_node *op2, ir_node *mem, ir_node *fpcw);
typedef ir_node *construct_unop_func(dbg_info *db, ir_graph *irg,
- ir_node *block, ir_node *base, ir_node *index, ir_node *op,
- ir_node *mem);
+ ir_node *block, ir_node *op);
/****************************************************************************************************
* _ _ __ _ _
static ir_node *create_immediate_or_transform(ir_node *node,
char immediate_constraint_type);
+static ir_node *create_I2I_Conv(ir_mode *src_mode, ir_mode *tgt_mode,
+ dbg_info *dbgi, ir_node *new_block,
+ ir_node *new_op);
+
/**
* Return true if a mode can be stored in the GP register set
*/
static INLINE int mode_needs_gp_reg(ir_mode *mode) {
if(mode == mode_fpcw)
return 0;
- return mode_is_int(mode) || mode_is_character(mode) || mode_is_reference(mode);
-}
-
-/**
- * Returns 1 if irn is a Const representing 0, 0 otherwise
- */
-static INLINE int is_ia32_Const_0(ir_node *irn) {
- return is_ia32_irn(irn) && is_ia32_Const(irn) && get_ia32_immop_type(irn) == ia32_ImmConst
- && tarval_is_null(get_ia32_Immop_tarval(irn));
-}
-
-/**
- * Returns 1 if irn is a Const representing 1, 0 otherwise
- */
-static INLINE int is_ia32_Const_1(ir_node *irn) {
- return is_ia32_irn(irn) && is_ia32_Const(irn) && get_ia32_immop_type(irn) == ia32_ImmConst
- && tarval_is_one(get_ia32_Immop_tarval(irn));
-}
-
-/**
- * Collects all Projs of a node into the node array. Index is the projnum.
- * BEWARE: The caller has to assure the appropriate array size!
- */
-static void ia32_collect_Projs(ir_node *irn, ir_node **projs, int size) {
- const ir_edge_t *edge;
- assert(get_irn_mode(irn) == mode_T && "need mode_T");
-
- memset(projs, 0, size * sizeof(projs[0]));
-
- foreach_out_edge(irn, edge) {
- ir_node *proj = get_edge_src_irn(edge);
- int proj_proj = get_Proj_proj(proj);
- assert(proj_proj < size);
- projs[proj_proj] = proj;
- }
-}
-
-/**
- * Renumbers the proj having pn_old in the array tp pn_new
- * and removes the proj from the array.
- */
-static INLINE void ia32_renumber_Proj(ir_node **projs, long pn_old, long pn_new) {
- fprintf(stderr, "Warning: renumber_Proj used!\n");
- if (projs[pn_old]) {
- set_Proj_proj(projs[pn_old], pn_new);
- projs[pn_old] = NULL;
- }
+ return mode_is_int(mode) || mode_is_reference(mode) || mode == mode_b;
}
/**
}
/**
- * Get an entity that is initialized with a tarval
+ * Get an atomic entity that is initialized with a tarval
*/
-static ir_entity *get_entity_for_tv(ia32_code_gen_t *cg, ir_node *cnst)
+static ir_entity *ia32_get_entity_for_tv(ia32_isa_t *isa, ir_node *cnst)
{
tarval *tv = get_Const_tarval(cnst);
- pmap_entry *e = pmap_find(cg->isa->tv_ent, tv);
+ pmap_entry *e = pmap_find(isa->tv_ent, tv);
ir_entity *res;
ir_graph *rem;
ir_mode *mode = get_irn_mode(cnst);
ir_type *tp = get_Const_type(cnst);
if (tp == firm_unknown_type)
- tp = get_prim_type(cg->isa->types, mode);
+ tp = get_prim_type(isa->types, mode);
res = new_entity(get_glob_type(), unique_id(".LC%u"), tp);
set_atomic_ent_value(res, new_Const_type(tv, tp));
current_ir_graph = rem;
- pmap_insert(cg->isa->tv_ent, tv, res);
+ pmap_insert(isa->tv_ent, tv, res);
} else {
res = e->value;
}
return classify_Const(node) == CNST_ONE;
}
+static int is_Const_Minus_1(ir_node *node) {
+ tarval *tv;
+ ir_mode *mode;
+ if(!is_Const(node))
+ return 0;
+
+ mode = get_irn_mode(node);
+ if(!mode_is_signed(mode))
+ return 0;
+
+ tv = get_Const_tarval(node);
+ tv = tarval_neg(tv);
+
+ return classify_tarval(tv) == CNST_ONE;
+}
+
/**
* Transforms a Const.
*/
static ir_node *gen_Const(ir_node *node) {
ir_graph *irg = current_ir_graph;
- ir_node *block = be_transform_node(get_nodes_block(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 *nomem = new_NoMem();
ir_node *load;
ir_entity *floatent;
+ cnst_classify_t clss = classify_Const(node);
- FP_USED(env_cg);
- if (! USE_SSE2(env_cg)) {
- cnst_classify_t clss = classify_Const(node);
+ if (USE_SSE2(env_cg)) {
+ if (clss == CNST_NULL) {
+ load = new_rd_ia32_xZero(dbgi, irg, block);
+ set_ia32_ls_mode(load, mode);
+ res = load;
+ } else {
+ floatent = ia32_get_entity_for_tv(env_cg->isa, node);
+ load = new_rd_ia32_xLoad(dbgi, irg, block, noreg, noreg, nomem,
+ mode);
+ set_ia32_op_type(load, ia32_AddrModeS);
+ set_ia32_am_sc(load, floatent);
+ set_ia32_flags(load, get_ia32_flags(load) | arch_irn_flags_rematerializable);
+ res = new_r_Proj(irg, block, load, mode_xmm, pn_ia32_xLoad_res);
+ }
+ } else {
if (clss == CNST_NULL) {
load = new_rd_ia32_vfldz(dbgi, irg, block);
res = load;
load = new_rd_ia32_vfld1(dbgi, irg, block);
res = load;
} else {
- floatent = get_entity_for_tv(env_cg, node);
+ floatent = ia32_get_entity_for_tv(env_cg->isa, node);
load = new_rd_ia32_vfld(dbgi, irg, block, noreg, noreg, nomem, mode);
set_ia32_op_type(load, ia32_AddrModeS);
- set_ia32_am_flavour(load, ia32_am_N);
set_ia32_am_sc(load, floatent);
set_ia32_flags(load, get_ia32_flags(load) | arch_irn_flags_rematerializable);
res = new_r_Proj(irg, block, load, mode_vfp, pn_ia32_vfld_res);
}
set_ia32_ls_mode(load, mode);
- } else {
- floatent = get_entity_for_tv(env_cg, node);
-
- load = new_rd_ia32_xLoad(dbgi, irg, block, noreg, noreg, nomem);
- set_ia32_op_type(load, ia32_AddrModeS);
- set_ia32_am_flavour(load, ia32_am_N);
- set_ia32_am_sc(load, floatent);
- set_ia32_ls_mode(load, mode);
- set_ia32_flags(load, get_ia32_flags(load) | arch_irn_flags_rematerializable);
-
- res = new_r_Proj(irg, block, load, mode_xmm, pn_ia32_xLoad_res);
}
SET_IA32_ORIG_NODE(load, ia32_get_old_node_name(env_cg, node));
SET_IA32_ORIG_NODE(load, ia32_get_old_node_name(env_cg, node));
return res;
} else {
- ir_node *cnst = new_rd_ia32_Const(dbgi, irg, block);
+ ir_node *cnst;
+ tarval *tv = get_Const_tarval(node);
+ long val;
+
+ tv = tarval_convert_to(tv, mode_Iu);
+
+ if(tv == get_tarval_bad() || tv == get_tarval_undefined()
+ || tv == NULL) {
+ panic("couldn't convert constant tarval (%+F)", node);
+ }
+ val = get_tarval_long(tv);
+
+ cnst = new_rd_ia32_Const(dbgi, irg, block, NULL, 0, val);
+ SET_IA32_ORIG_NODE(cnst, ia32_get_old_node_name(env_cg, node));
/* see above */
if (get_irg_start_block(irg) == block) {
add_irn_dep(cnst, get_irg_frame(irg));
}
- set_ia32_Const_attr(cnst, node);
- SET_IA32_ORIG_NODE(cnst, ia32_get_old_node_name(env_cg, node));
return cnst;
}
-
- assert(0);
- return new_r_Bad(irg);
}
/**
*/
static ir_node *gen_SymConst(ir_node *node) {
ir_graph *irg = current_ir_graph;
- ir_node *block = be_transform_node(get_nodes_block(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 *cnst;
if (mode_is_float(mode)) {
- FP_USED(env_cg);
+ ir_node *noreg = ia32_new_NoReg_gp(env_cg);
+ ir_node *nomem = new_NoMem();
+
if (USE_SSE2(env_cg))
- cnst = new_rd_ia32_xConst(dbgi, irg, block);
+ cnst = new_rd_ia32_xLoad(dbgi, irg, block, noreg, noreg, nomem, mode_E);
else
- cnst = new_rd_ia32_vfConst(dbgi, irg, block);
- //set_ia32_ls_mode(cnst, mode);
- set_ia32_ls_mode(cnst, mode_E);
+ cnst = new_rd_ia32_vfld(dbgi, irg, block, noreg, noreg, nomem, mode_E);
+ set_ia32_am_sc(cnst, get_SymConst_entity(node));
} else {
- cnst = new_rd_ia32_Const(dbgi, irg, block);
+ ir_entity *entity;
+
+ if(get_SymConst_kind(node) != symconst_addr_ent) {
+ panic("backend only support symconst_addr_ent (at %+F)", node);
+ }
+ entity = get_SymConst_entity(node);
+ cnst = new_rd_ia32_Const(dbgi, irg, block, entity, 0, 0);
}
/* Const Nodes before the initial IncSP are a bad idea, because
add_irn_dep(cnst, get_irg_frame(irg));
}
- set_ia32_Const_attr(cnst, node);
SET_IA32_ORIG_NODE(cnst, ia32_get_old_node_name(env_cg, node));
return cnst;
//mode = mode_xmm;
tv = new_tarval_from_str(cnst_str, strlen(cnst_str), mode);
tp = new_type_primitive(new_id_from_str(tp_name), mode);
+ /* these constants are loaded as part of an instruction, so they must be aligned
+ to 128 bit. */
+ set_type_alignment_bytes(tp, 16);
ent = new_entity(get_glob_type(), new_id_from_str(ent_name), tp);
set_entity_ld_ident(ent, get_entity_ident(ent));
}
#endif /* NDEBUG */
-/* determine if one operator is an Imm */
-static ir_node *get_immediate_op(ir_node *op1, ir_node *op2) {
- if (op1) {
- return is_ia32_Cnst(op1) ? op1 : (is_ia32_Cnst(op2) ? op2 : NULL);
+static int use_source_address_mode(ir_node *block, ir_node *node,
+ ir_node *other)
+{
+ ir_mode *mode;
+ ir_node *load;
+ long pn;
+
+ if(!is_Proj(node))
+ return 0;
+ load = get_Proj_pred(node);
+ pn = get_Proj_proj(node);
+ if(!is_Load(load) || pn != pn_Load_res)
+ return 0;
+ if(get_nodes_block(load) != block)
+ return 0;
+ /* we only use address mode if we're the only user of the load */
+ if(get_irn_n_edges(node) > 1)
+ return 0;
+
+ mode = get_irn_mode(node);
+ if(!mode_needs_gp_reg(mode))
+ return 0;
+ if(get_mode_size_bits(mode) != 32)
+ return 0;
+
+ /* don't do AM if other node inputs depend on the load (via mem-proj) */
+ if(other != NULL && get_nodes_block(other) == block
+ && heights_reachable_in_block(heights, other, load))
+ return 0;
+
+ return 1;
+}
+
+typedef struct ia32_address_mode_t ia32_address_mode_t;
+struct ia32_address_mode_t {
+ ia32_address_t addr;
+ ir_mode *ls_mode;
+ ir_node *mem_proj;
+ ia32_op_type_t op_type;
+ ir_node *new_op1;
+ ir_node *new_op2;
+ int commutative;
+ int flipped;
+};
+
+static void build_address(ia32_address_mode_t *am, ir_node *node)
+{
+ ia32_address_t *addr = &am->addr;
+ ir_node *load = get_Proj_pred(node);
+ ir_node *ptr = get_Load_ptr(load);
+ ir_node *mem = get_Load_mem(load);
+ ir_node *new_mem = be_transform_node(mem);
+ ir_node *base;
+ ir_node *index;
+
+ am->ls_mode = get_Load_mode(load);
+ am->mem_proj = be_get_Proj_for_pn(load, pn_Load_M);
+
+ /* construct load address */
+ ia32_create_address_mode(addr, ptr, 0);
+ base = addr->base;
+ index = addr->index;
+
+ if(base == NULL) {
+ base = ia32_new_NoReg_gp(env_cg);
+ } else {
+ base = be_transform_node(base);
+ }
+
+ if(index == NULL) {
+ index = ia32_new_NoReg_gp(env_cg);
} else {
- return is_ia32_Cnst(op2) ? op2 : NULL;
+ index = be_transform_node(index);
}
+
+ addr->base = base;
+ addr->index = index;
+ addr->mem = new_mem;
}
-/* determine if one operator is not an Imm */
-static ir_node *get_expr_op(ir_node *op1, ir_node *op2) {
- return !is_ia32_Cnst(op1) ? op1 : (!is_ia32_Cnst(op2) ? op2 : NULL);
+static void set_address(ir_node *node, ia32_address_t *addr)
+{
+ set_ia32_am_scale(node, addr->scale);
+ set_ia32_am_sc(node, addr->symconst_ent);
+ set_ia32_am_offs_int(node, addr->offset);
+ if(addr->symconst_sign)
+ set_ia32_am_sc_sign(node);
+ if(addr->use_frame)
+ set_ia32_use_frame(node);
+ set_ia32_frame_ent(node, addr->frame_entity);
}
-static void fold_immediate(ir_node *node, int in1, int in2) {
- ir_node *left;
- ir_node *right;
+static void set_am_attributes(ir_node *node, ia32_address_mode_t *am)
+{
+ set_address(node, &am->addr);
- if (!(env_cg->opt & IA32_OPT_IMMOPS))
- return;
+ set_ia32_op_type(node, am->op_type);
+ set_ia32_ls_mode(node, am->ls_mode);
+ if(am->commutative)
+ set_ia32_commutative(node);
+}
- left = get_irn_n(node, in1);
- right = get_irn_n(node, in2);
- if (! is_ia32_Cnst(right) && is_ia32_Cnst(left)) {
- /* we can only set right operand to immediate */
- if(!is_ia32_commutative(node))
- return;
- /* exchange left/right */
- set_irn_n(node, in1, right);
- set_irn_n(node, in2, ia32_get_admissible_noreg(env_cg, node, in2));
- copy_ia32_Immop_attr(node, left);
- } else if(is_ia32_Cnst(right)) {
- set_irn_n(node, in2, ia32_get_admissible_noreg(env_cg, node, in2));
- copy_ia32_Immop_attr(node, right);
+static void match_arguments(ia32_address_mode_t *am, ir_node *block,
+ ir_node *op1, ir_node *op2, int commutative,
+ int use_am_and_immediates)
+{
+ ia32_address_t *addr = &am->addr;
+ ir_node *noreg_gp = ia32_new_NoReg_gp(env_cg);
+ ir_node *new_op1;
+ ir_node *new_op2;
+
+ memset(am, 0, sizeof(am[0]));
+
+ new_op2 = try_create_Immediate(op2, 0);
+ if(new_op2 == NULL && use_source_address_mode(block, op2, op1)) {
+ build_address(am, op2);
+ new_op1 = be_transform_node(op1);
+ new_op2 = noreg_gp;
+ am->op_type = ia32_AddrModeS;
+ } else if(commutative && (new_op2 == NULL || use_am_and_immediates) &&
+ use_source_address_mode(block, op1, op2)) {
+ build_address(am, op1);
+ if(new_op2 != NULL) {
+ new_op1 = noreg_gp;
+ } else {
+ new_op1 = be_transform_node(op2);
+ new_op2 = noreg_gp;
+ am->flipped = 1;
+ }
+ am->op_type = ia32_AddrModeS;
} else {
- return;
+ new_op1 = be_transform_node(op1);
+ if(new_op2 == NULL)
+ new_op2 = be_transform_node(op2);
+ am->op_type = ia32_Normal;
}
+ if(addr->base == NULL)
+ addr->base = noreg_gp;
+ if(addr->index == NULL)
+ addr->index = noreg_gp;
+ if(addr->mem == NULL)
+ addr->mem = new_NoMem();
+
+ am->new_op1 = new_op1;
+ am->new_op2 = new_op2;
+ am->commutative = commutative;
+}
- clear_ia32_commutative(node);
- set_ia32_am_support(node, get_ia32_am_support(node) & ~ia32_am_Source,
- get_ia32_am_arity(node));
+static ir_node *fix_mem_proj(ir_node *node, ia32_address_mode_t *am)
+{
+ ir_graph *irg = current_ir_graph;
+ ir_mode *mode;
+ ir_node *load;
+
+ if(am->mem_proj == NULL)
+ return node;
+
+ /* we have to create a mode_T so the old MemProj can attach to us */
+ mode = get_irn_mode(node);
+ load = get_Proj_pred(am->mem_proj);
+
+ mark_irn_visited(load);
+ be_set_transformed_node(load, node);
+
+ if(mode != mode_T) {
+ set_irn_mode(node, mode_T);
+ return new_rd_Proj(NULL, irg, get_nodes_block(node), node, mode, 0);
+ } else {
+ return node;
+ }
}
/**
static ir_node *gen_binop(ir_node *node, ir_node *op1, ir_node *op2,
construct_binop_func *func, int commutative)
{
- ir_node *block = be_transform_node(get_nodes_block(node));
- ir_graph *irg = current_ir_graph;
- dbg_info *dbgi = get_irn_dbg_info(node);
- ir_node *noreg_gp = ia32_new_NoReg_gp(env_cg);
- ir_node *nomem = new_NoMem();
+ ir_node *src_block = get_nodes_block(node);
+ ir_node *block = be_transform_node(src_block);
+ ir_graph *irg = current_ir_graph;
+ dbg_info *dbgi = get_irn_dbg_info(node);
ir_node *new_node;
+ ia32_address_mode_t am;
+ ia32_address_t *addr = &am.addr;
- ir_node *new_op1 = be_transform_node(op1);
- ir_node *new_op2 = create_immediate_or_transform(op2, 0);
- if (is_ia32_Immediate(new_op2)) {
- commutative = 0;
- }
-
- new_node = func(dbgi, irg, block, noreg_gp, noreg_gp, new_op1, new_op2, nomem);
- if (func == new_rd_ia32_IMul) {
- set_ia32_am_support(new_node, ia32_am_Source, ia32_am_binary);
- } else {
- set_ia32_am_support(new_node, ia32_am_Full, ia32_am_binary);
- }
+ match_arguments(&am, src_block, op1, op2, commutative, 0);
+ new_node = func(dbgi, irg, block, addr->base, addr->index, am.new_op1,
+ am.new_op2, addr->mem);
+ set_am_attributes(new_node, &am);
+ /* we can't use source address mode anymore when using immediates */
+ if(is_ia32_Immediate(am.new_op1) || is_ia32_Immediate(am.new_op2))
+ set_ia32_am_support(new_node, ia32_am_None, ia32_am_arity_none);
SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node));
- if (commutative) {
- set_ia32_commutative(new_node);
- }
+
+ new_node = fix_mem_proj(new_node, &am);
return new_node;
}
new_node = func(dbgi, irg, block, noreg_gp, noreg_gp, new_op1, new_op2,
nomem);
- set_ia32_am_support(new_node, ia32_am_Source, ia32_am_binary);
if (is_op_commutative(get_irn_op(node))) {
set_ia32_commutative(new_node);
}
- if (USE_SSE2(env_cg)) {
- set_ia32_ls_mode(new_node, mode);
- }
+ set_ia32_ls_mode(new_node, mode);
SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node));
return new_node;
}
+static ir_node *get_fpcw(void)
+{
+ ir_node *fpcw;
+ if(initial_fpcw != NULL)
+ return initial_fpcw;
+
+ fpcw = be_abi_get_ignore_irn(env_cg->birg->abi,
+ &ia32_fp_cw_regs[REG_FPCW]);
+ initial_fpcw = be_transform_node(fpcw);
+
+ return initial_fpcw;
+}
+
/**
* Construct a standard binary operation, set AM and immediate if required.
*
ir_node *new_node = NULL;
dbg_info *dbgi = get_irn_dbg_info(node);
ir_graph *irg = current_ir_graph;
- ir_mode *mode = get_irn_mode(node);
ir_node *noreg_gp = ia32_new_NoReg_gp(env_cg);
ir_node *nomem = new_NoMem();
- ir_node *fpcw = be_abi_get_ignore_irn(env_cg->birg->abi,
- &ia32_fp_cw_regs[REG_FPCW]);
new_node = func(dbgi, irg, block, noreg_gp, noreg_gp, new_op1, new_op2,
- nomem, fpcw);
- set_ia32_am_support(new_node, ia32_am_Source, ia32_am_binary);
+ nomem, get_fpcw());
if (is_op_commutative(get_irn_op(node))) {
set_ia32_commutative(new_node);
}
- if (USE_SSE2(env_cg)) {
- set_ia32_ls_mode(new_node, mode);
- }
SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node));
* @return The constructed ia32 node.
*/
static ir_node *gen_shift_binop(ir_node *node, ir_node *op1, ir_node *op2,
- construct_binop_func *func)
+ construct_shift_func *func)
{
- ir_node *block = be_transform_node(get_nodes_block(node));
- ir_node *new_op1 = be_transform_node(op1);
- ir_node *new_op2;
- ir_node *new_op = NULL;
- dbg_info *dbgi = get_irn_dbg_info(node);
- ir_graph *irg = current_ir_graph;
- ir_node *noreg = ia32_new_NoReg_gp(env_cg);
- ir_node *nomem = new_NoMem();
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ ir_graph *irg = current_ir_graph;
+ ir_node *block = get_nodes_block(node);
+ ir_node *new_block = be_transform_node(block);
+ ir_node *new_op1 = be_transform_node(op1);
+ ir_node *new_op2 = create_immediate_or_transform(op2, 0);
+ ir_node *res;
assert(! mode_is_float(get_irn_mode(node))
&& "Shift/Rotate with float not supported");
- new_op2 = create_immediate_or_transform(op2, 'N');
-
- new_op = func(dbgi, irg, block, noreg, noreg, new_op1, new_op2, nomem);
-
- /* set AM support */
- set_ia32_am_support(new_op, ia32_am_Dest, ia32_am_binary);
-
- SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
+ res = func(dbgi, irg, new_block, new_op1, new_op2);
+ SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
- set_ia32_emit_cl(new_op);
+ /* 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(res, new_dep);
+ }
- return new_op;
+ return res;
}
ir_node *new_node = NULL;
ir_graph *irg = current_ir_graph;
dbg_info *dbgi = get_irn_dbg_info(node);
- ir_node *noreg = ia32_new_NoReg_gp(env_cg);
- ir_node *nomem = new_NoMem();
- new_node = func(dbgi, irg, block, noreg, noreg, new_op, nomem);
- DB((dbg, LEVEL_1, "INT unop ..."));
- set_ia32_am_support(new_node, ia32_am_Dest, ia32_am_unary);
+ new_node = func(dbgi, irg, block, new_op);
SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node));
return new_node;
}
+static ir_node *create_lea_from_address(dbg_info *dbgi, ir_node *block,
+ ia32_address_t *addr)
+{
+ ir_graph *irg = current_ir_graph;
+ ir_node *base = addr->base;
+ ir_node *index = addr->index;
+ ir_node *res;
+
+ if(base == NULL) {
+ base = ia32_new_NoReg_gp(env_cg);
+ } else {
+ base = be_transform_node(base);
+ }
+
+ if(index == NULL) {
+ index = ia32_new_NoReg_gp(env_cg);
+ } else {
+ index = be_transform_node(index);
+ }
+
+ res = new_rd_ia32_Lea(dbgi, irg, block, base, index);
+ set_address(res, addr);
+
+ return res;
+}
+
+static int am_has_immediates(const ia32_address_t *addr)
+{
+ return addr->offset != 0 || addr->symconst_ent != NULL
+ || addr->frame_entity || addr->use_frame;
+}
+
/**
* Creates an ia32 Add.
*
static ir_node *gen_Add(ir_node *node) {
ir_node *block = be_transform_node(get_nodes_block(node));
ir_node *op1 = get_Add_left(node);
- ir_node *new_op1 = be_transform_node(op1);
ir_node *op2 = get_Add_right(node);
- ir_node *new_op2 = be_transform_node(op2);
- ir_node *new_op = NULL;
+ ir_node *new_op;
+ ir_node *new_op1;
ir_graph *irg = current_ir_graph;
dbg_info *dbgi = get_irn_dbg_info(node);
ir_mode *mode = get_irn_mode(node);
ir_node *noreg = ia32_new_NoReg_gp(env_cg);
- ir_node *nomem = new_NoMem();
- ir_node *expr_op, *imm_op;
-
- /* Check if immediate optimization is on and */
- /* if it's an operation with immediate. */
- imm_op = (env_cg->opt & IA32_OPT_IMMOPS) ? get_immediate_op(new_op1, new_op2) : NULL;
- expr_op = get_expr_op(new_op1, new_op2);
-
- assert((expr_op || imm_op) && "invalid operands");
+ ir_node *src_block = get_nodes_block(node);
+ ir_node *add_immediate_op;
+ ia32_address_t addr;
+ ia32_address_mode_t am;
if (mode_is_float(mode)) {
- FP_USED(env_cg);
if (USE_SSE2(env_cg))
return gen_binop_sse_float(node, op1, op2, new_rd_ia32_xAdd);
else
return gen_binop_x87_float(node, op1, op2, new_rd_ia32_vfadd);
}
- /* integer ADD */
- if (! expr_op) {
- ia32_immop_type_t tp1 = get_ia32_immop_type(new_op1);
- ia32_immop_type_t tp2 = get_ia32_immop_type(new_op2);
-
- /* No expr_op means, that we have two const - one symconst and */
- /* one tarval or another symconst - because this case is not */
- /* covered by constant folding */
- /* We need to check for: */
- /* 1) symconst + const -> becomes a LEA */
- /* 2) symconst + symconst -> becomes a const + LEA as the elf */
- /* linker doesn't support two symconsts */
-
- if (tp1 == ia32_ImmSymConst && tp2 == ia32_ImmSymConst) {
- /* this is the 2nd case */
- new_op = new_rd_ia32_Lea(dbgi, irg, block, new_op1, noreg);
- set_ia32_am_sc(new_op, get_ia32_Immop_symconst(new_op2));
- set_ia32_am_flavour(new_op, ia32_am_B);
- set_ia32_op_type(new_op, ia32_AddrModeS);
-
- DBG_OPT_LEA3(new_op1, new_op2, node, new_op);
- } else if (tp1 == ia32_ImmSymConst) {
- tarval *tv = get_ia32_Immop_tarval(new_op2);
- long offs = get_tarval_long(tv);
-
- new_op = new_rd_ia32_Lea(dbgi, irg, block, noreg, noreg);
- add_irn_dep(new_op, get_irg_frame(irg));
- DBG_OPT_LEA3(new_op1, new_op2, node, new_op);
-
- set_ia32_am_sc(new_op, get_ia32_Immop_symconst(new_op1));
- add_ia32_am_offs_int(new_op, offs);
- set_ia32_am_flavour(new_op, ia32_am_OB);
- set_ia32_op_type(new_op, ia32_AddrModeS);
- } else if (tp2 == ia32_ImmSymConst) {
- tarval *tv = get_ia32_Immop_tarval(new_op1);
- long offs = get_tarval_long(tv);
-
- new_op = new_rd_ia32_Lea(dbgi, irg, block, noreg, noreg);
- add_irn_dep(new_op, get_irg_frame(irg));
- DBG_OPT_LEA3(new_op1, new_op2, node, new_op);
-
- add_ia32_am_offs_int(new_op, offs);
- set_ia32_am_sc(new_op, get_ia32_Immop_symconst(new_op2));
- set_ia32_am_flavour(new_op, ia32_am_OB);
- set_ia32_op_type(new_op, ia32_AddrModeS);
- } else {
- tarval *tv1 = get_ia32_Immop_tarval(new_op1);
- tarval *tv2 = get_ia32_Immop_tarval(new_op2);
- tarval *restv = tarval_add(tv1, tv2);
-
- DEBUG_ONLY(ir_fprintf(stderr, "Warning: add with 2 consts not folded: %+F\n", node));
+ /**
+ * Rules for an Add:
+ * 0. Immediate Trees (example Add(Symconst, Const) -> Const)
+ * 1. Add with immediate -> Lea
+ * 2. Add with possible source address mode -> Add
+ * 3. Otherwise -> Lea
+ */
+ memset(&addr, 0, sizeof(addr));
+ ia32_create_address_mode(&addr, node, 1);
+ add_immediate_op = NULL;
+ /* a constant? */
+ if(addr.base == NULL && addr.index == NULL) {
+ new_op = new_rd_ia32_Const(dbgi, irg, block, addr.symconst_ent,
+ addr.symconst_sign, addr.offset);
+ add_irn_dep(new_op, get_irg_frame(irg));
+ SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
+ return new_op;
+ }
+ /* add with immediate? */
+ if(addr.index == NULL) {
+ add_immediate_op = addr.base;
+ } else if(addr.base == NULL && addr.scale == 0) {
+ add_immediate_op = addr.index;
+ }
- new_op = new_rd_ia32_Const(dbgi, irg, block);
- set_ia32_Const_tarval(new_op, restv);
- DBG_OPT_LEA3(new_op1, new_op2, node, new_op);
+ if(add_immediate_op != NULL) {
+ if(!am_has_immediates(&addr)) {
+#ifdef DEBUG_libfirm
+ ir_fprintf(stderr, "Optimisation warning Add x,0 (%+F) found\n",
+ node);
+#endif
+ return be_transform_node(add_immediate_op);
}
+ new_op = create_lea_from_address(dbgi, block, &addr);
SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
return new_op;
- } else if (imm_op) {
- if ((env_cg->opt & IA32_OPT_INCDEC) && get_ia32_immop_type(imm_op) == ia32_ImmConst) {
- tarval_classification_t class_tv, class_negtv;
- tarval *tv = get_ia32_Immop_tarval(imm_op);
-
- /* optimize tarvals */
- class_tv = classify_tarval(tv);
- class_negtv = classify_tarval(tarval_neg(tv));
-
- if (class_tv == TV_CLASSIFY_ONE) { /* + 1 == INC */
- DB((dbg, LEVEL_2, "Add(1) to Inc ... "));
- new_op = new_rd_ia32_Inc(dbgi, irg, block, noreg, noreg, expr_op, nomem);
- SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
- return new_op;
- } else if (class_tv == TV_CLASSIFY_ALL_ONE || class_negtv == TV_CLASSIFY_ONE) { /* + (-1) == DEC */
- DB((dbg, LEVEL_2, "Add(-1) to Dec ... "));
- new_op = new_rd_ia32_Dec(dbgi, irg, block, noreg, noreg, expr_op, nomem);
- SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
- return new_op;
- }
- }
}
- /* This is a normal add */
- new_op = new_rd_ia32_Add(dbgi, irg, block, noreg, noreg, new_op1, new_op2, nomem);
+ /* test if we can use source address mode */
+ memset(&am, 0, sizeof(am));
+ new_op1 = NULL;
+ if(use_source_address_mode(src_block, op2, op1)) {
+ build_address(&am, op2);
+ new_op1 = be_transform_node(op1);
+ } else if(use_source_address_mode(src_block, op1, op2)) {
+ build_address(&am, op1);
+ new_op1 = be_transform_node(op2);
+ }
+ /* construct an Add with source address mode */
+ if(new_op1 != NULL) {
+ ia32_address_t *am_addr = &am.addr;
+ new_op = new_rd_ia32_Add(dbgi, irg, block, am_addr->base,
+ am_addr->index, new_op1, noreg, am_addr->mem);
+ set_address(new_op, am_addr);
+ set_ia32_op_type(new_op, ia32_AddrModeS);
+ set_ia32_ls_mode(new_op, am.ls_mode);
+ set_ia32_commutative(new_op);
+ SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
- /* set AM support */
- set_ia32_am_support(new_op, ia32_am_Full, ia32_am_binary);
- set_ia32_commutative(new_op);
+ new_op = fix_mem_proj(new_op, &am);
- fold_immediate(new_op, 2, 3);
+ return new_op;
+ }
+ /* otherwise construct a lea */
+ new_op = create_lea_from_address(dbgi, block, &addr);
SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
-
return new_op;
}
ir_mode *mode = get_irn_mode(node);
if (mode_is_float(mode)) {
- FP_USED(env_cg);
if (USE_SSE2(env_cg))
return gen_binop_sse_float(node, op1, op2, new_rd_ia32_xMul);
else
dbg_info *dbgi = get_irn_dbg_info(node);
ir_node *noreg = ia32_new_NoReg_gp(env_cg);
ir_mode *mode = get_irn_mode(node);
- ir_node *proj_EAX, *proj_EDX, *res;
- ir_node *in[1];
+ ir_node *proj_EDX, *res;
assert(!mode_is_float(mode) && "Mulh with float not supported");
if (mode_is_signed(mode)) {
- res = new_rd_ia32_IMul1OP(dbgi, irg, block, noreg, noreg, new_op1, new_op2, new_NoMem());
+ res = new_rd_ia32_IMul1OP(dbgi, irg, block, noreg, noreg, new_op1,
+ new_op2, new_NoMem());
} else {
- res = new_rd_ia32_Mul(dbgi, irg, block, noreg, noreg, new_op1, new_op2, new_NoMem());
+ res = new_rd_ia32_Mul(dbgi, irg, block, noreg, noreg, new_op1, new_op2,
+ new_NoMem());
}
set_ia32_commutative(res);
- set_ia32_am_support(res, ia32_am_Source, ia32_am_binary);
- proj_EAX = new_rd_Proj(dbgi, irg, block, res, mode_Iu, pn_EAX);
proj_EDX = new_rd_Proj(dbgi, irg, block, res, mode_Iu, pn_EDX);
- /* keep EAX */
- in[0] = proj_EAX;
- be_new_Keep(&ia32_reg_classes[CLASS_ia32_gp], irg, block, 1, in);
-
return proj_EDX;
}
static ir_node *gen_And(ir_node *node) {
ir_node *op1 = get_And_left(node);
ir_node *op2 = get_And_right(node);
+ assert(! mode_is_float(get_irn_mode(node)));
+
+ /* is it a zero extension? */
+ if (is_Const(op2)) {
+ 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);
+ ir_node *block = be_transform_node(get_nodes_block(node));
+ ir_node *new_op = be_transform_node(op1);
+ ir_mode *src_mode;
+ ir_node *res;
+
+ if(v == 0xFF) {
+ src_mode = mode_Bu;
+ } else {
+ assert(v == 0xFFFF);
+ src_mode = mode_Hu;
+ }
+ res = create_I2I_Conv(src_mode, mode_Iu, dbgi, block, new_op);
+ SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
+
+ return res;
+ }
+ }
- assert (! mode_is_float(get_irn_mode(node)));
return gen_binop(node, op1, op2, new_rd_ia32_And, 1);
}
}
-
-/**
- * Creates an ia32 Max.
- *
- * @return the created ia32 Max node
- */
-static ir_node *gen_Max(ir_node *node) {
- ir_node *block = be_transform_node(get_nodes_block(node));
- ir_node *op1 = get_irn_n(node, 0);
- ir_node *new_op1 = be_transform_node(op1);
- ir_node *op2 = get_irn_n(node, 1);
- ir_node *new_op2 = be_transform_node(op2);
- ir_graph *irg = current_ir_graph;
- ir_mode *mode = get_irn_mode(node);
- dbg_info *dbgi = get_irn_dbg_info(node);
- ir_mode *op_mode = get_irn_mode(op1);
- ir_node *new_op;
-
- assert(get_mode_size_bits(mode) == 32);
-
- if (mode_is_float(mode)) {
- FP_USED(env_cg);
- if (USE_SSE2(env_cg)) {
- new_op = gen_binop_sse_float(node, new_op1, new_op2, new_rd_ia32_xMax);
- } else {
- panic("Can't create Max node");
- }
- } else {
- long pnc = pn_Cmp_Gt;
- if (! mode_is_signed(op_mode)) {
- pnc |= ia32_pn_Cmp_Unsigned;
- }
- new_op = new_rd_ia32_CmpCMov(dbgi, irg, block, new_op1, new_op2,
- new_op1, new_op2, pnc);
- }
- SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
-
- return new_op;
-}
-
-/**
- * Creates an ia32 Min.
- *
- * @return the created ia32 Min node
- */
-static ir_node *gen_Min(ir_node *node) {
- ir_node *block = be_transform_node(get_nodes_block(node));
- ir_node *op1 = get_irn_n(node, 0);
- ir_node *new_op1 = be_transform_node(op1);
- ir_node *op2 = get_irn_n(node, 1);
- ir_node *new_op2 = be_transform_node(op2);
- ir_graph *irg = current_ir_graph;
- ir_mode *mode = get_irn_mode(node);
- dbg_info *dbgi = get_irn_dbg_info(node);
- ir_mode *op_mode = get_irn_mode(op1);
- ir_node *new_op;
-
- assert(get_mode_size_bits(mode) == 32);
-
- if (mode_is_float(mode)) {
- FP_USED(env_cg);
- if (USE_SSE2(env_cg)) {
- new_op = gen_binop_sse_float(node, op1, op2, new_rd_ia32_xMin);
- } else {
- panic("can't create Min node");
- }
- } else {
- long pnc = pn_Cmp_Lt;
- if (! mode_is_signed(op_mode)) {
- pnc |= ia32_pn_Cmp_Unsigned;
- }
- new_op = new_rd_ia32_CmpCMov(dbgi, irg, block, new_op1, new_op2,
- new_op1, new_op2, pnc);
- }
- SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
-
- return new_op;
-}
-
-
/**
* Creates an ia32 Sub.
*
* @return The created ia32 Sub node
*/
static ir_node *gen_Sub(ir_node *node) {
- ir_node *block = be_transform_node(get_nodes_block(node));
- ir_node *op1 = get_Sub_left(node);
- ir_node *new_op1 = be_transform_node(op1);
- ir_node *op2 = get_Sub_right(node);
- ir_node *new_op2 = be_transform_node(op2);
- ir_node *new_op = NULL;
- ir_graph *irg = current_ir_graph;
- dbg_info *dbgi = get_irn_dbg_info(node);
- ir_mode *mode = get_irn_mode(node);
- ir_node *noreg = ia32_new_NoReg_gp(env_cg);
- ir_node *nomem = new_NoMem();
- ir_node *expr_op, *imm_op;
-
- /* Check if immediate optimization is on and */
- /* if it's an operation with immediate. */
- imm_op = (env_cg->opt & IA32_OPT_IMMOPS) ? get_immediate_op(NULL, new_op2) : NULL;
- expr_op = get_expr_op(new_op1, new_op2);
-
- assert((expr_op || imm_op) && "invalid operands");
+ ir_node *op1 = get_Sub_left(node);
+ ir_node *op2 = get_Sub_right(node);
+ ir_mode *mode = get_irn_mode(node);
if (mode_is_float(mode)) {
- FP_USED(env_cg);
if (USE_SSE2(env_cg))
return gen_binop_sse_float(node, op1, op2, new_rd_ia32_xSub);
else
return gen_binop_x87_float(node, op1, op2, new_rd_ia32_vfsub);
}
- /* integer SUB */
- if (! expr_op) {
- ia32_immop_type_t tp1 = get_ia32_immop_type(new_op1);
- ia32_immop_type_t tp2 = get_ia32_immop_type(new_op2);
-
- /* No expr_op means, that we have two const - one symconst and */
- /* one tarval or another symconst - because this case is not */
- /* covered by constant folding */
- /* We need to check for: */
- /* 1) symconst - const -> becomes a LEA */
- /* 2) symconst - symconst -> becomes a const - LEA as the elf */
- /* linker doesn't support two symconsts */
- if (tp1 == ia32_ImmSymConst && tp2 == ia32_ImmSymConst) {
- /* this is the 2nd case */
- new_op = new_rd_ia32_Lea(dbgi, irg, block, new_op1, noreg);
- set_ia32_am_sc(new_op, get_ia32_Immop_symconst(op2));
- set_ia32_am_sc_sign(new_op);
- set_ia32_am_flavour(new_op, ia32_am_B);
-
- DBG_OPT_LEA3(op1, op2, node, new_op);
- } else if (tp1 == ia32_ImmSymConst) {
- tarval *tv = get_ia32_Immop_tarval(new_op2);
- long offs = get_tarval_long(tv);
-
- new_op = new_rd_ia32_Lea(dbgi, irg, block, noreg, noreg);
- add_irn_dep(new_op, get_irg_frame(irg));
- DBG_OPT_LEA3(op1, op2, node, new_op);
-
- set_ia32_am_sc(new_op, get_ia32_Immop_symconst(new_op1));
- add_ia32_am_offs_int(new_op, -offs);
- set_ia32_am_flavour(new_op, ia32_am_OB);
- set_ia32_op_type(new_op, ia32_AddrModeS);
- } else if (tp2 == ia32_ImmSymConst) {
- tarval *tv = get_ia32_Immop_tarval(new_op1);
- long offs = get_tarval_long(tv);
-
- new_op = new_rd_ia32_Lea(dbgi, irg, block, noreg, noreg);
- add_irn_dep(new_op, get_irg_frame(irg));
- DBG_OPT_LEA3(op1, op2, node, new_op);
-
- add_ia32_am_offs_int(new_op, offs);
- set_ia32_am_sc(new_op, get_ia32_Immop_symconst(new_op2));
- set_ia32_am_sc_sign(new_op);
- set_ia32_am_flavour(new_op, ia32_am_OB);
- set_ia32_op_type(new_op, ia32_AddrModeS);
- } else {
- tarval *tv1 = get_ia32_Immop_tarval(new_op1);
- tarval *tv2 = get_ia32_Immop_tarval(new_op2);
- tarval *restv = tarval_sub(tv1, tv2);
-
- DEBUG_ONLY(ir_fprintf(stderr, "Warning: sub with 2 consts not folded: %+F\n", node));
-
- new_op = new_rd_ia32_Const(dbgi, irg, block);
- set_ia32_Const_tarval(new_op, restv);
- DBG_OPT_LEA3(new_op1, new_op2, node, new_op);
- }
-
- SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
- return new_op;
- } else if (imm_op) {
- if ((env_cg->opt & IA32_OPT_INCDEC) && get_ia32_immop_type(imm_op) == ia32_ImmConst) {
- tarval_classification_t class_tv, class_negtv;
- tarval *tv = get_ia32_Immop_tarval(imm_op);
-
- /* optimize tarvals */
- class_tv = classify_tarval(tv);
- class_negtv = classify_tarval(tarval_neg(tv));
-
- if (class_tv == TV_CLASSIFY_ONE) {
- DB((dbg, LEVEL_2, "Sub(1) to Dec ... "));
- new_op = new_rd_ia32_Dec(dbgi, irg, block, noreg, noreg, expr_op, nomem);
- SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
- return new_op;
- } else if (class_tv == TV_CLASSIFY_ALL_ONE || class_negtv == TV_CLASSIFY_ONE) {
- DB((dbg, LEVEL_2, "Sub(-1) to Inc ... "));
- new_op = new_rd_ia32_Inc(dbgi, irg, block, noreg, noreg, expr_op, nomem);
- SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
- return new_op;
- }
- }
+ if(is_Const(op2)) {
+ ir_fprintf(stderr, "Optimisation warning: found sub with const (%+F)\n",
+ node);
}
- /* This is a normal sub */
- new_op = new_rd_ia32_Sub(dbgi, irg, block, noreg, noreg, new_op1, new_op2, nomem);
-
- /* set AM support */
- set_ia32_am_support(new_op, ia32_am_Full, ia32_am_binary);
-
- fold_immediate(new_op, 2, 3);
-
- SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
-
- return new_op;
+ return gen_binop(node, op1, op2, new_rd_ia32_Sub, 0);
}
ir_node *noreg = ia32_new_NoReg_gp(env_cg);
ir_node *res, *proj_div, *proj_mod;
ir_node *sign_extension;
- ir_node *in_keep[2];
ir_node *mem, *new_mem;
- ir_node *projs[pn_DivMod_max];
- int i, has_exc;
-
- ia32_collect_Projs(node, projs, pn_DivMod_max);
+ int has_exc;
proj_div = proj_mod = NULL;
has_exc = 0;
if (mode_is_signed(mode)) {
/* in signed mode, we need to sign extend the dividend */
- sign_extension = new_rd_ia32_Cltd(dbgi, irg, block, new_dividend);
+ ir_node *produceval = new_rd_ia32_ProduceVal(dbgi, irg, block);
+ add_irn_dep(produceval, get_irg_frame(irg));
+ sign_extension = new_rd_ia32_Cltd(dbgi, irg, block, new_dividend,
+ produceval);
} else {
- sign_extension = new_rd_ia32_Const(dbgi, irg, block);
- set_ia32_Immop_tarval(sign_extension, get_tarval_null(mode_Iu));
-
+ sign_extension = new_rd_ia32_Const(dbgi, irg, block, NULL, 0, 0);
add_irn_dep(sign_extension, get_irg_frame(irg));
}
set_ia32_exc_label(res, has_exc);
set_irn_pinned(res, get_irn_pinned(node));
- /* Matze: code can't handle this at the moment... */
-#if 0
- /* set AM support */
- set_ia32_am_support(res, ia32_am_Source, ia32_am_binary);
-#endif
-
- /* check, which Proj-Keep, we need to add */
- i = 0;
- if (proj_div == NULL) {
- /* We have only mod result: add div res Proj-Keep */
- in_keep[i] = new_rd_Proj(dbgi, irg, block, res, mode_Iu, pn_ia32_Div_div_res);
- ++i;
- }
- if (proj_mod == NULL) {
- /* We have only div result: add mod res Proj-Keep */
- in_keep[i] = new_rd_Proj(dbgi, irg, block, res, mode_Iu, pn_ia32_Div_mod_res);
- ++i;
- }
- if(i > 0)
- be_new_Keep(&ia32_reg_classes[CLASS_ia32_gp], irg, block, i, in_keep);
-
SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
return res;
ir_node *nomem = new_rd_NoMem(current_ir_graph);
ir_node *new_op;
- FP_USED(env_cg);
if (USE_SSE2(env_cg)) {
ir_mode *mode = get_irn_mode(op1);
- if (is_ia32_xConst(new_op2)) {
- new_op = new_rd_ia32_xDiv(dbgi, irg, block, noreg, noreg, new_op1, noreg, nomem);
- set_ia32_am_support(new_op, ia32_am_None, ia32_am_arity_none);
- copy_ia32_Immop_attr(new_op, new_op2);
- } else {
- new_op = new_rd_ia32_xDiv(dbgi, irg, block, noreg, noreg, new_op1, new_op2, nomem);
- // Matze: disabled for now, spillslot coalescer fails
- //set_ia32_am_support(new_op, ia32_am_Source | ia32_am_binary);
- }
+ new_op = new_rd_ia32_xDiv(dbgi, irg, block, noreg, noreg, new_op1,
+ new_op2, nomem);
set_ia32_ls_mode(new_op, mode);
} else {
- ir_node *fpcw = be_abi_get_ignore_irn(env_cg->birg->abi,
- &ia32_fp_cw_regs[REG_FPCW]);
new_op = new_rd_ia32_vfdiv(dbgi, irg, block, noreg, noreg, new_op1,
- new_op2, nomem, fpcw);
- // Matze: disabled for now (spillslot coalescer fails)
- //set_ia32_am_support(new_op, ia32_am_Source | ia32_am_binary);
+ new_op2, nomem, get_fpcw());
}
SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
return new_op;
* @return The created ia32 Shl node
*/
static ir_node *gen_Shl(ir_node *node) {
+ ir_node *right = get_Shl_right(node);
+
+ /* test whether we can build a lea */
+ if(is_Const(right)) {
+ tarval *tv = get_Const_tarval(right);
+ if(tarval_is_long(tv)) {
+ long val = get_tarval_long(tv);
+ if(val >= 0 && val <= 3) {
+ ir_graph *irg = current_ir_graph;
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ ir_node *block = be_transform_node(get_nodes_block(node));
+ ir_node *base = ia32_new_NoReg_gp(env_cg);
+ ir_node *index = be_transform_node(get_Shl_left(node));
+
+ ir_node *res
+ = new_rd_ia32_Lea(dbgi, irg, block, base, index);
+ set_ia32_am_scale(res, val);
+ SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
+ return res;
+ }
+ }
+ }
+
return gen_shift_binop(node, get_Shl_left(node), get_Shl_right(node),
new_rd_ia32_Shl);
}
static ir_node *gen_Shrs(ir_node *node) {
ir_node *left = get_Shrs_left(node);
ir_node *right = get_Shrs_right(node);
- if(is_Const(right) && get_irn_mode(left) == mode_Is) {
+ ir_mode *mode = get_irn_mode(node);
+ if(is_Const(right) && mode == mode_Is) {
tarval *tv = get_Const_tarval(right);
long val = get_tarval_long(tv);
if(val == 31) {
ir_node *block = be_transform_node(get_nodes_block(node));
ir_node *op = left;
ir_node *new_op = be_transform_node(op);
+ ir_node *pval = new_rd_ia32_ProduceVal(dbgi, irg, block);
+ add_irn_dep(pval, get_irg_frame(irg));
- return new_rd_ia32_Cltd(dbgi, irg, block, new_op);
+ return new_rd_ia32_Cltd(dbgi, irg, block, new_op, pval);
+ }
+ }
+
+ /* 8 or 16 bit sign extension? */
+ if(is_Const(right) && is_Shl(left) && mode == mode_Is) {
+ 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);
+ if(tv1 == tv2 && tarval_is_long(tv1)) {
+ long val = get_tarval_long(tv1);
+ if(val == 16 || val == 24) {
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ ir_node *block = be_transform_node(get_nodes_block(node));
+ ir_node *new_op = be_transform_node(shl_left);
+ ir_mode *src_mode;
+ ir_node *res;
+
+ if(val == 24) {
+ src_mode = mode_Bs;
+ } else {
+ assert(val == 16);
+ src_mode = mode_Hs;
+ }
+ res = create_I2I_Conv(src_mode, mode_Is, dbgi, block,
+ new_op);
+ SET_IA32_ORIG_NODE(res,
+ ia32_get_old_node_name(env_cg, node));
+
+ return res;
+ }
+ }
}
}
if (mode_is_float(mode)) {
ir_node *new_op = be_transform_node(op);
- FP_USED(env_cg);
if (USE_SSE2(env_cg)) {
ir_node *noreg_gp = ia32_new_NoReg_gp(env_cg);
ir_node *noreg_fp = ia32_new_NoReg_fp(env_cg);
return gen_Minus_ex(node, get_Minus_op(node));
}
+static ir_node *create_Immediate_from_int(int val)
+{
+ ir_graph *irg = current_ir_graph;
+ ir_node *start_block = get_irg_start_block(irg);
+ ir_node *immediate = new_rd_ia32_Immediate(NULL, irg, start_block, NULL, 0, val);
+ arch_set_irn_register(env_cg->arch_env, immediate, &ia32_gp_regs[REG_GP_NOREG]);
+
+ return immediate;
+}
+
+static ir_node *gen_bin_Not(ir_node *node)
+{
+ ir_graph *irg = current_ir_graph;
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ ir_node *block = be_transform_node(get_nodes_block(node));
+ ir_node *op = get_Not_op(node);
+ ir_node *new_op = be_transform_node(op);
+ ir_node *noreg = ia32_new_NoReg_gp(env_cg);
+ ir_node *nomem = new_NoMem();
+ ir_node *one = create_Immediate_from_int(1);
+
+ return new_rd_ia32_Xor(dbgi, irg, block, noreg, noreg, new_op, one, nomem);
+}
/**
* Transforms a Not node.
* @return The created ia32 Not node
*/
static ir_node *gen_Not(ir_node *node) {
- ir_node *op = get_Not_op(node);
+ ir_node *op = get_Not_op(node);
+ ir_mode *mode = get_irn_mode(node);
+
+ if(mode == mode_b) {
+ return gen_bin_Not(node);
+ }
assert (! mode_is_float(get_irn_mode(node)));
return gen_unop(node, op, new_rd_ia32_Not);
ir_entity *ent;
if (mode_is_float(mode)) {
- FP_USED(env_cg);
if (USE_SSE2(env_cg)) {
res = new_rd_ia32_xAnd(dbgi,irg, block, noreg_gp, noreg_gp, new_op, noreg_fp, nomem);
}
} else {
ir_node *xor;
- ir_node *sign_extension = new_rd_ia32_Cltd(dbgi, irg, block, new_op);
+ ir_node *pval = new_rd_ia32_ProduceVal(dbgi, irg, block);
+ ir_node *sign_extension = new_rd_ia32_Cltd(dbgi, irg, block, new_op,
+ pval);
+
+ add_irn_dep(pval, get_irg_frame(irg));
SET_IA32_ORIG_NODE(sign_extension,
ia32_get_old_node_name(env_cg, node));
return res;
}
-
-
/**
* Transforms a Load.
*
* @return the created ia32 Load node
*/
static ir_node *gen_Load(ir_node *node) {
- ir_node *block = be_transform_node(get_nodes_block(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 *new_ptr = be_transform_node(ptr);
ir_node *mem = get_Load_mem(node);
ir_node *new_mem = be_transform_node(mem);
+ ir_node *base;
+ ir_node *index;
ir_graph *irg = current_ir_graph;
dbg_info *dbgi = get_irn_dbg_info(node);
ir_node *noreg = ia32_new_NoReg_gp(env_cg);
ir_mode *mode = get_Load_mode(node);
ir_mode *res_mode;
- ir_node *lptr = new_ptr;
- int is_imm = 0;
ir_node *new_op;
- ir_node *projs[pn_Load_max];
- ia32_am_flavour_t am_flav = ia32_am_B;
+ 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_collect_Projs(node, projs, pn_Load_max);
+ if(base == NULL) {
+ base = noreg;
+ } else {
+ base = be_transform_node(base);
+ }
- /* address might be a constant (symconst or absolute address) */
- if (is_ia32_Const(new_ptr)) {
- lptr = noreg;
- is_imm = 1;
+ if(index == NULL) {
+ index = noreg;
+ } else {
+ index = be_transform_node(index);
}
if (mode_is_float(mode)) {
- FP_USED(env_cg);
if (USE_SSE2(env_cg)) {
- new_op = new_rd_ia32_xLoad(dbgi, irg, block, lptr, noreg, new_mem);
+ new_op = new_rd_ia32_xLoad(dbgi, irg, block, base, index, new_mem,
+ mode);
res_mode = mode_xmm;
} else {
- new_op = new_rd_ia32_vfld(dbgi, irg, block, lptr, noreg, new_mem, mode);
+ new_op = new_rd_ia32_vfld(dbgi, irg, block, base, index, new_mem,
+ mode);
res_mode = mode_vfp;
}
} else {
- new_op = new_rd_ia32_Load(dbgi, irg, block, lptr, noreg, new_mem);
- res_mode = mode_Iu;
- }
-
- /*
- check for special case: the loaded value might not be used
- */
- if (be_get_Proj_for_pn(node, pn_Load_res) == NULL) {
- /* add a result proj and a Keep to produce a pseudo use */
- ir_node *proj = new_r_Proj(irg, block, new_op, mode_Iu,
- pn_ia32_Load_res);
- be_new_Keep(arch_get_irn_reg_class(env_cg->arch_env, proj, -1), irg, block, 1, &proj);
- }
+ if(mode == mode_b)
+ mode = mode_Iu;
- /* base is a constant address */
- if (is_imm) {
- if (get_ia32_immop_type(new_ptr) == ia32_ImmSymConst) {
- set_ia32_am_sc(new_op, get_ia32_Immop_symconst(new_ptr));
- am_flav = ia32_am_N;
+ /* create a conv node with address mode for smaller modes */
+ if(get_mode_size_bits(mode) < 32) {
+ new_op = new_rd_ia32_Conv_I2I(dbgi, irg, block, base, index, noreg,
+ new_mem, mode);
} else {
- tarval *tv = get_ia32_Immop_tarval(new_ptr);
- long offs = get_tarval_long(tv);
-
- add_ia32_am_offs_int(new_op, offs);
- am_flav = ia32_am_O;
+ new_op = new_rd_ia32_Load(dbgi, irg, block, base, index, new_mem);
}
+ res_mode = mode_Iu;
}
set_irn_pinned(new_op, get_irn_pinned(node));
set_ia32_op_type(new_op, ia32_AddrModeS);
- set_ia32_am_flavour(new_op, am_flav);
set_ia32_ls_mode(new_op, mode);
+ set_address(new_op, &addr);
/* make sure we are scheduled behind the initial IncSP/Barrier
* to avoid spills being placed before it
set_ia32_exc_label(new_op, be_get_Proj_for_pn(node, pn_Load_X_except) != NULL);
SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
- return new_op;
-}
+ return new_op;
+}
+
+static int use_dest_am(ir_node *block, ir_node *node, ir_node *mem,
+ ir_node *ptr, ir_mode *mode, ir_node *other)
+{
+ ir_node *load;
+
+ if(!is_Proj(node))
+ return 0;
+
+ /* we only use address mode if we're the only user of the load */
+ if(get_irn_n_edges(node) > 1)
+ return 0;
+
+ load = get_Proj_pred(node);
+ if(!is_Load(load))
+ return 0;
+ if(get_nodes_block(load) != block)
+ return 0;
+
+ /* Store should be attached to the load */
+ if(!is_Proj(mem) || get_Proj_pred(mem) != load)
+ return 0;
+ /* store should have the same pointer as the load */
+ if(get_Load_ptr(load) != ptr)
+ return 0;
+
+ /* don't do AM if other node inputs depend on the load (via mem-proj) */
+ if(other != NULL && get_nodes_block(other) == block
+ && heights_reachable_in_block(heights, other, load))
+ return 0;
+
+ assert(get_Load_mode(load) == mode);
+
+ return 1;
+}
+
+static ir_node *dest_am_binop(ir_node *node, ir_node *op1, ir_node *op2,
+ ir_node *mem, ir_node *ptr, ir_mode *mode,
+ construct_binop_dest_func *func, int commutative)
+{
+ ir_node *src_block = get_nodes_block(node);
+ ir_node *block;
+ ir_node *noreg_gp = ia32_new_NoReg_gp(env_cg);
+ ir_graph *irg = current_ir_graph;
+ dbg_info *dbgi;
+ ir_node *new_node;
+ ir_node *new_op;
+ ia32_address_mode_t am;
+ ia32_address_t *addr = &am.addr;
+ memset(&am, 0, sizeof(am));
+
+ if(use_dest_am(src_block, op1, mem, ptr, mode, op2)) {
+ build_address(&am, op1);
+ new_op = create_immediate_or_transform(op2, 0);
+ } else if(commutative && use_dest_am(src_block, op2, mem, ptr, mode, op1)) {
+ build_address(&am, op2);
+ new_op = create_immediate_or_transform(op1, 0);
+ } else {
+ return NULL;
+ }
+
+ if(addr->base == NULL)
+ addr->base = noreg_gp;
+ if(addr->index == NULL)
+ addr->index = noreg_gp;
+ if(addr->mem == NULL)
+ addr->mem = new_NoMem();
+
+ dbgi = get_irn_dbg_info(node);
+ block = be_transform_node(src_block);
+ new_node = func(dbgi, irg, block, addr->base, addr->index, new_op,
+ addr->mem);
+ set_address(new_node, addr);
+ set_ia32_op_type(new_node, ia32_AddrModeD);
+ set_ia32_ls_mode(new_node, mode);
+ SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node));
+
+ return new_node;
+}
+
+static ir_node *dest_am_unop(ir_node *node, ir_node *op, ir_node *mem,
+ ir_node *ptr, ir_mode *mode,
+ construct_unop_dest_func *func)
+{
+ ir_node *src_block = get_nodes_block(node);
+ ir_node *block;
+ ir_node *noreg_gp = ia32_new_NoReg_gp(env_cg);
+ ir_graph *irg = current_ir_graph;
+ dbg_info *dbgi;
+ ir_node *new_node;
+ ia32_address_mode_t am;
+ ia32_address_t *addr = &am.addr;
+ memset(&am, 0, sizeof(am));
+
+ if(!use_dest_am(src_block, op, mem, ptr, mode, NULL))
+ return NULL;
+
+ build_address(&am, op);
+
+ if(addr->base == NULL)
+ addr->base = noreg_gp;
+ if(addr->index == NULL)
+ addr->index = noreg_gp;
+ if(addr->mem == NULL)
+ addr->mem = new_NoMem();
+
+ dbgi = get_irn_dbg_info(node);
+ block = be_transform_node(src_block);
+ new_node = func(dbgi, irg, block, addr->base, addr->index, addr->mem);
+ set_address(new_node, addr);
+ set_ia32_op_type(new_node, ia32_AddrModeD);
+ set_ia32_ls_mode(new_node, mode);
+ SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node));
+
+ return new_node;
+}
+
+static ir_node *try_create_dest_am(ir_node *node) {
+ ir_node *val = get_Store_value(node);
+ ir_node *mem = get_Store_mem(node);
+ ir_node *ptr = get_Store_ptr(node);
+ ir_mode *mode = get_irn_mode(val);
+ ir_node *op1;
+ ir_node *op2;
+ ir_node *new_node;
+
+ /* handle only GP modes for now... */
+ if(!mode_needs_gp_reg(mode))
+ return NULL;
+ if(get_mode_size_bits(mode) != 32)
+ return NULL;
+
+ /* store must be the only user of the val node */
+ if(get_irn_n_edges(val) > 1)
+ return NULL;
+ switch(get_irn_opcode(val)) {
+ case iro_Add:
+ op1 = get_Add_left(val);
+ op2 = get_Add_right(val);
+ if(is_Const_1(op2)) {
+ new_node = dest_am_unop(val, op1, mem, ptr, mode,
+ new_rd_ia32_IncMem);
+ break;
+ } else if(is_Const_Minus_1(op2)) {
+ new_node = dest_am_unop(val, op1, mem, ptr, mode,
+ new_rd_ia32_DecMem);
+ break;
+ }
+ new_node = dest_am_binop(val, op1, op2, mem, ptr, mode,
+ new_rd_ia32_AddMem, 1);
+ break;
+ case iro_Sub:
+ op1 = get_Sub_left(val);
+ op2 = get_Sub_right(val);
+ new_node = dest_am_binop(val, op1, op2, mem, ptr, mode,
+ new_rd_ia32_SubMem, 0);
+ break;
+ case iro_And:
+ op1 = get_And_left(val);
+ op2 = get_And_right(val);
+ new_node = dest_am_binop(val, op1, op2, mem, ptr, mode,
+ new_rd_ia32_AndMem, 1);
+ break;
+ case iro_Or:
+ op1 = get_Or_left(val);
+ op2 = get_Or_right(val);
+ new_node = dest_am_binop(val, op1, op2, mem, ptr, mode,
+ new_rd_ia32_OrMem, 1);
+ break;
+ case iro_Eor:
+ op1 = get_Eor_left(val);
+ op2 = get_Eor_right(val);
+ new_node = dest_am_binop(val, op1, op2, mem, ptr, mode,
+ new_rd_ia32_XorMem, 1);
+ break;
+ case iro_Shl:
+ op1 = get_Shl_left(val);
+ op2 = get_Shl_right(val);
+ new_node = dest_am_binop(val, op1, op2, mem, ptr, mode,
+ new_rd_ia32_ShlMem, 0);
+ break;
+ case iro_Shr:
+ op1 = get_Shr_left(val);
+ op2 = get_Shr_right(val);
+ new_node = dest_am_binop(val, op1, op2, mem, ptr, mode,
+ new_rd_ia32_ShrMem, 0);
+ break;
+ case iro_Shrs:
+ op1 = get_Shrs_left(val);
+ op2 = get_Shrs_right(val);
+ new_node = dest_am_binop(val, op1, op2, mem, ptr, mode,
+ new_rd_ia32_SarMem, 0);
+ break;
+ case iro_Rot:
+ op1 = get_Rot_left(val);
+ op2 = get_Rot_right(val);
+ new_node = dest_am_binop(val, op1, op2, mem, ptr, mode,
+ new_rd_ia32_RolMem, 0);
+ break;
+ /* TODO: match ROR patterns... */
+ case iro_Minus:
+ op1 = get_Minus_op(val);
+ new_node = dest_am_unop(val, op1, mem, ptr, mode, new_rd_ia32_NegMem);
+ break;
+ case iro_Not:
+ /* TODO this would be ^ 1 with DestAM */
+ if(mode == mode_b)
+ return NULL;
+ op1 = get_Not_op(val);
+ new_node = dest_am_unop(val, op1, mem, ptr, mode, new_rd_ia32_NotMem);
+ break;
+ default:
+ return NULL;
+ }
+ return new_node;
+}
/**
* Transforms a Store.
static ir_node *gen_Store(ir_node *node) {
ir_node *block = be_transform_node(get_nodes_block(node));
ir_node *ptr = get_Store_ptr(node);
- ir_node *new_ptr = be_transform_node(ptr);
+ ir_node *base;
+ ir_node *index;
ir_node *val = get_Store_value(node);
ir_node *new_val;
ir_node *mem = get_Store_mem(node);
ir_graph *irg = current_ir_graph;
dbg_info *dbgi = get_irn_dbg_info(node);
ir_node *noreg = ia32_new_NoReg_gp(env_cg);
- ir_node *sptr = new_ptr;
ir_mode *mode = get_irn_mode(val);
- int is_imm = 0;
ir_node *new_op;
- ia32_am_flavour_t am_flav = ia32_am_B;
+ ia32_address_t addr;
+
+ /* check for destination address mode */
+ new_op = try_create_dest_am(node);
+ if(new_op != NULL)
+ return new_op;
- /* address might be a constant (symconst or absolute address) */
- if (is_ia32_Const(new_ptr)) {
- sptr = noreg;
- is_imm = 1;
+ /* construct load address */
+ memset(&addr, 0, sizeof(addr));
+ ia32_create_address_mode(&addr, ptr, 0);
+ base = addr.base;
+ index = addr.index;
+
+ if(base == NULL) {
+ base = noreg;
+ } else {
+ base = be_transform_node(base);
}
- if (mode_is_float(mode)) {
- FP_USED(env_cg);
+ if(index == NULL) {
+ index = noreg;
+ } else {
+ index = be_transform_node(index);
+ }
+ if (mode_is_float(mode)) {
new_val = be_transform_node(val);
if (USE_SSE2(env_cg)) {
- new_op = new_rd_ia32_xStore(dbgi, irg, block, sptr, noreg, new_val,
+ new_op = new_rd_ia32_xStore(dbgi, irg, block, base, index, new_val,
new_mem);
} else {
- new_op = new_rd_ia32_vfst(dbgi, irg, block, sptr, noreg, new_val,
+ new_op = new_rd_ia32_vfst(dbgi, irg, block, base, index, new_val,
new_mem, mode);
}
} else {
new_val = create_immediate_or_transform(val, 0);
+ if(mode == mode_b)
+ mode = mode_Iu;
if (get_mode_size_bits(mode) == 8) {
- new_op = new_rd_ia32_Store8Bit(dbgi, irg, block, sptr, noreg,
+ new_op = new_rd_ia32_Store8Bit(dbgi, irg, block, base, index,
new_val, new_mem);
} else {
- new_op = new_rd_ia32_Store(dbgi, irg, block, sptr, noreg, new_val,
+ new_op = new_rd_ia32_Store(dbgi, irg, block, base, index, new_val,
new_mem);
}
}
- /* base is an constant address */
- if (is_imm) {
- if (get_ia32_immop_type(new_ptr) == ia32_ImmSymConst) {
- set_ia32_am_sc(new_op, get_ia32_Immop_symconst(new_ptr));
- am_flav = ia32_am_N;
- } else {
- tarval *tv = get_ia32_Immop_tarval(new_ptr);
- long offs = get_tarval_long(tv);
-
- add_ia32_am_offs_int(new_op, offs);
- am_flav = ia32_am_O;
- }
- }
-
set_irn_pinned(new_op, get_irn_pinned(node));
set_ia32_op_type(new_op, ia32_AddrModeD);
- set_ia32_am_flavour(new_op, am_flav);
set_ia32_ls_mode(new_op, mode);
set_ia32_exc_label(new_op, be_get_Proj_for_pn(node, pn_Store_X_except) != NULL);
+ set_address(new_op, &addr);
SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
return new_op;
}
-static ir_node *try_create_TestJmp(ir_node *block, ir_node *node, long pnc)
+static ir_node *try_create_TestJmp(ir_node *block, dbg_info *dbgi, long pnc,
+ ir_node *cmp_left, ir_node *cmp_right)
{
- ir_node *cmp_a = get_Cmp_left(node);
- ir_node *new_cmp_a;
- ir_node *cmp_b = get_Cmp_right(node);
- ir_node *new_cmp_b;
- ir_node *and_left;
- ir_node *and_right;
+ ir_node *arg_left;
+ ir_node *arg_right;
ir_node *res;
- ir_node *noreg;
- ir_node *nomem;
- dbg_info *dbgi;
- tarval *tv;
+ ir_mode *mode;
+ long pure_pnc = pnc & ~ia32_pn_Cmp_Unsigned;
+ ia32_address_mode_t am;
+ ia32_address_t *addr = &am.addr;
- if(!is_Const(cmp_b))
+ if(cmp_right != NULL && !is_Const_0(cmp_right))
return NULL;
- tv = get_Const_tarval(cmp_b);
- if(!tarval_is_null(tv))
- return NULL;
+ if(is_And(cmp_left) && (pure_pnc == pn_Cmp_Eq || pure_pnc == pn_Cmp_Lg)) {
+ mode = get_irn_mode(cmp_left);
+ arg_left = get_And_left(cmp_left);
+ arg_right = get_And_right(cmp_left);
+ } else {
+ mode = get_irn_mode(cmp_left);
+ arg_left = cmp_left;
+ arg_right = cmp_left;
+ }
+ if(mode == mode_b)
+ mode = mode_Iu;
- if(is_And(cmp_a) && (pnc == pn_Cmp_Eq || pnc == pn_Cmp_Lg)) {
- and_left = get_And_left(cmp_a);
- and_right = get_And_right(cmp_a);
+ assert(get_mode_size_bits(mode) <= 32);
+ match_arguments(&am, block, arg_left, arg_right, 1, 1);
+ if(am.flipped)
+ pnc = get_inversed_pnc(pnc);
- new_cmp_a = be_transform_node(and_left);
- new_cmp_b = create_immediate_or_transform(and_right, 0);
+ if(get_mode_size_bits(mode) == 8) {
+ res = new_rd_ia32_TestJmp8Bit(dbgi, current_ir_graph, block, addr->base,
+ addr->index, am.new_op1, am.new_op2,
+ addr->mem, pnc);
} else {
- new_cmp_a = be_transform_node(cmp_a);
- new_cmp_b = be_transform_node(cmp_a);
+ res = new_rd_ia32_TestJmp(dbgi, current_ir_graph, block, addr->base,
+ addr->index, am.new_op1, am.new_op2,
+ addr->mem, pnc);
}
+ set_am_attributes(res, &am);
+ set_ia32_ls_mode(res, mode);
- dbgi = get_irn_dbg_info(node);
- noreg = ia32_new_NoReg_gp(env_cg);
- nomem = new_NoMem();
-
- res = new_rd_ia32_TestJmp(dbgi, current_ir_graph, block, noreg, noreg,
- new_cmp_a, new_cmp_b, nomem, pnc);
- set_ia32_am_support(res, ia32_am_Source, ia32_am_binary);
- SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
+ res = fix_mem_proj(res, &am);
return res;
}
int switch_min = INT_MAX;
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);
}
if (switch_min != 0) {
- ir_node *noreg = ia32_new_NoReg_gp(env_cg);
+ ir_node *noreg = ia32_new_NoReg_gp(env_cg);
/* if smallest switch case is not 0 we need an additional sub */
new_sel = new_rd_ia32_Lea(dbgi, irg, block, new_sel, noreg);
add_ia32_am_offs_int(new_sel, -switch_min);
- set_ia32_am_flavour(new_sel, ia32_am_OB);
set_ia32_op_type(new_sel, ia32_AddrModeS);
SET_IA32_ORIG_NODE(new_sel, ia32_get_old_node_name(env_cg, node));
* @return The transformed node.
*/
static ir_node *gen_Cond(ir_node *node) {
- ir_node *block = be_transform_node(get_nodes_block(node));
- ir_graph *irg = current_ir_graph;
- 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 *res = NULL;
- ir_node *noreg = ia32_new_NoReg_gp(env_cg);
+ ir_node *src_block = get_nodes_block(node);
+ ir_node *block = be_transform_node(src_block);
+ ir_graph *irg = current_ir_graph;
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ ir_node *sel = get_Cond_selector(node);
+ ir_mode *sel_mode = get_irn_mode(sel);
+ ir_node *res = NULL;
+ ir_node *noreg = ia32_new_NoReg_gp(env_cg);
+ ir_node *nomem = new_NoMem();
ir_node *cmp;
ir_node *cmp_a;
ir_node *cmp_b;
ir_node *new_cmp_a;
ir_node *new_cmp_b;
ir_mode *cmp_mode;
- ir_node *nomem = new_NoMem();
long pnc;
if (sel_mode != mode_b) {
return create_Switch(node);
}
+ if(!is_Proj(sel) || !is_Cmp(get_Proj_pred(sel))) {
+ /* it's some mode_b value but not a direct comparison -> create a
+ * testjmp */
+ res = try_create_TestJmp(block, dbgi, pn_Cmp_Lg, sel, NULL);
+ SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
+ return res;
+ }
+
cmp = get_Proj_pred(sel);
cmp_a = get_Cmp_left(cmp);
cmp_b = get_Cmp_right(cmp);
cmp_mode = get_irn_mode(cmp_a);
- pnc = get_Proj_proj(sel);
+ pnc = get_Proj_proj(sel);
if(mode_is_float(cmp_mode) || !mode_is_signed(cmp_mode)) {
pnc |= ia32_pn_Cmp_Unsigned;
}
if(mode_needs_gp_reg(cmp_mode)) {
- res = try_create_TestJmp(block, cmp, pnc);
- if(res != NULL)
+ res = try_create_TestJmp(block, dbgi, pnc, cmp_a, cmp_b);
+ if(res != NULL) {
+ SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
return res;
+ }
}
- new_cmp_a = be_transform_node(cmp_a);
- new_cmp_b = create_immediate_or_transform(cmp_b, 0);
-
if (mode_is_float(cmp_mode)) {
- FP_USED(env_cg);
+ new_cmp_a = be_transform_node(cmp_a);
+ new_cmp_b = create_immediate_or_transform(cmp_b, 0);
if (USE_SSE2(env_cg)) {
- res = new_rd_ia32_xCondJmp(dbgi, irg, block, noreg, noreg, cmp_a,
- cmp_b, nomem, pnc);
+ res = new_rd_ia32_xCmpJmp(dbgi, irg, block, noreg, noreg, cmp_a,
+ cmp_b, nomem, pnc);
set_ia32_commutative(res);
- set_ia32_am_support(res, ia32_am_Source, ia32_am_binary);
set_ia32_ls_mode(res, cmp_mode);
} else {
- ir_node *proj_eax;
- res = new_rd_ia32_vfCondJmp(dbgi, irg, block, cmp_a, cmp_b, pnc);
+ res = new_rd_ia32_vfCmpJmp(dbgi, irg, block, cmp_a, cmp_b, pnc);
set_ia32_commutative(res);
- proj_eax = new_r_Proj(irg, block, res, mode_Iu,
- pn_ia32_vfCondJmp_temp_reg_eax);
- be_new_Keep(&ia32_reg_classes[CLASS_ia32_gp], irg, block, 1,
- &proj_eax);
}
} else {
- assert(get_mode_size_bits(cmp_mode) == 32);
- res = new_rd_ia32_CondJmp(dbgi, irg, block, noreg, noreg,
- new_cmp_a, new_cmp_b, nomem, pnc);
- set_ia32_commutative(res);
- set_ia32_am_support(res, ia32_am_Source, ia32_am_binary);
+ ia32_address_mode_t am;
+ ia32_address_t *addr = &am.addr;
+ match_arguments(&am, src_block, cmp_a, cmp_b, 1, 1);
+ if(am.flipped)
+ pnc = get_inversed_pnc(pnc);
+
+ if(get_mode_size_bits(cmp_mode) == 8) {
+ res = new_rd_ia32_CmpJmp8Bit(dbgi, irg, block, addr->base,
+ addr->index, am.new_op1, am.new_op2,
+ addr->mem, pnc);
+ } else {
+ res = new_rd_ia32_CmpJmp(dbgi, irg, block, addr->base, addr->index,
+ am.new_op1, am.new_op2, addr->mem, pnc);
+ }
+ set_am_attributes(res, &am);
+ assert(cmp_mode != NULL);
+ set_ia32_ls_mode(res, cmp_mode);
+
+ res = fix_mem_proj(res, &am);
}
SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
ir_graph *irg = current_ir_graph;
dbg_info *dbgi = get_irn_dbg_info(node);
int size = get_type_size_bytes(get_CopyB_type(node));
- ir_mode *dst_mode = get_irn_mode(dst);
- ir_mode *src_mode = get_irn_mode(src);
int rem;
- ir_node *in[3];
/* If we have to copy more than 32 bytes, we use REP MOVSx and */
/* then we need the size explicitly in ECX. */
rem = size & 0x3; /* size % 4 */
size >>= 2;
- res = new_rd_ia32_Const(dbgi, irg, block);
+ res = new_rd_ia32_Const(dbgi, irg, block, NULL, 0, size);
add_irn_dep(res, be_abi_get_start_barrier(env_cg->birg->abi));
- set_ia32_Immop_tarval(res, new_tarval_from_long(size, mode_Is));
res = new_rd_ia32_CopyB(dbgi, irg, block, new_dst, new_src, res, new_mem);
- set_ia32_Immop_tarval(res, new_tarval_from_long(rem, mode_Is));
-
- /* ok: now attach Proj's because rep movsd will destroy esi, edi and ecx */
- in[0] = new_r_Proj(irg, block, res, dst_mode, pn_ia32_CopyB_DST);
- in[1] = new_r_Proj(irg, block, res, src_mode, pn_ia32_CopyB_SRC);
- in[2] = new_r_Proj(irg, block, res, mode_Iu, pn_ia32_CopyB_CNT);
- be_new_Keep(&ia32_reg_classes[CLASS_ia32_gp], irg, block, 3, in);
- }
- else {
+ /* we misuse the pncode field for the copyb size */
+ set_ia32_pncode(res, rem);
+ } else {
res = new_rd_ia32_CopyB_i(dbgi, irg, block, new_dst, new_src, new_mem);
- set_ia32_Immop_tarval(res, new_tarval_from_long(size, mode_Is));
-
- /* ok: now attach Proj's because movsd will destroy esi and edi */
- in[0] = new_r_Proj(irg, block, res, dst_mode, pn_ia32_CopyB_i_DST);
- in[1] = new_r_Proj(irg, block, res, src_mode, pn_ia32_CopyB_i_SRC);
- be_new_Keep(&ia32_reg_classes[CLASS_ia32_gp], irg, block, 2, in);
+ set_ia32_pncode(res, size);
}
SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
}
-#if 0
-/**
- * Transforms a Mux node into CMov.
- *
- * @return The transformed node.
- */
-static ir_node *gen_Mux(ir_node *node) {
- ir_node *new_op = new_rd_ia32_CMov(env.dbgi, current_ir_graph, env.block, \
- get_Mux_sel(node), get_Mux_false(node), get_Mux_true(node), env.mode);
+static ir_node *create_set(long pnc, ir_node *cmp_left, ir_node *cmp_right,
+ dbg_info *dbgi, ir_node *block)
+{
+ ir_graph *irg = current_ir_graph;
+ ir_node *new_block = be_transform_node(block);
+ ir_node *noreg = ia32_new_NoReg_gp(env_cg);
+ ir_node *nomem = new_rd_NoMem(irg);
+ ir_mode *mode;
+ ir_node *arg_left;
+ ir_node *arg_right;
+ ir_node *res;
+ ia32_address_mode_t am;
+ ia32_address_t *addr = &am.addr;
+
+ /* can we use a test instruction? */
+ if(cmp_right == NULL || is_Const_0(cmp_right)) {
+ long pure_pnc = pnc & ~ia32_pn_Cmp_Unsigned;
+ if(is_And(cmp_left) &&
+ (pure_pnc == pn_Cmp_Eq || pure_pnc == pn_Cmp_Lg)) {
+ ir_node *and_left = get_And_left(cmp_left);
+ ir_node *and_right = get_And_right(cmp_left);
+
+ mode = get_irn_mode(and_left);
+ arg_left = and_left;
+ arg_right = and_right;
+ } else {
+ mode = get_irn_mode(cmp_left);
+ arg_left = cmp_left;
+ arg_right = cmp_left;
+ }
- SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
+ assert(get_mode_size_bits(mode) <= 32);
- return new_op;
+ match_arguments(&am, block, arg_left, arg_right, 1, 1);
+ if(am.flipped)
+ pnc = get_inversed_pnc(pnc);
+
+ if(get_mode_size_bits(mode) == 8) {
+ res = new_rd_ia32_TestSet8Bit(dbgi, irg, new_block, addr->base,
+ addr->index, am.new_op1, am.new_op2,
+ addr->mem, pnc);
+ } else {
+ res = new_rd_ia32_TestSet(dbgi, irg, new_block, addr->base,
+ addr->index, am.new_op1, am.new_op2,
+ addr->mem, pnc);
+ }
+ set_am_attributes(res, &am);
+ set_ia32_ls_mode(res, mode);
+
+ res = fix_mem_proj(res, &am);
+
+ res = new_rd_ia32_Conv_I2I8Bit(dbgi, irg, new_block, noreg, noreg, res,
+ nomem, mode_Bu);
+
+ return res;
+ }
+
+ mode = get_irn_mode(cmp_left);
+ assert(get_mode_size_bits(mode) <= 32);
+
+ match_arguments(&am, block, cmp_left, cmp_right, 1, 1);
+ if(am.flipped)
+ pnc = get_inversed_pnc(pnc);
+
+ if(get_mode_size_bits(mode) == 8) {
+ res = new_rd_ia32_CmpSet8Bit(dbgi, irg, new_block, addr->base,
+ addr->index, am.new_op1, am.new_op2,
+ addr->mem, pnc);
+ } else {
+ res = new_rd_ia32_CmpSet(dbgi, irg, new_block, addr->base, addr->index,
+ am.new_op1, am.new_op2, addr->mem, pnc);
+ }
+ set_am_attributes(res, &am);
+ set_ia32_ls_mode(res, mode);
+
+ res = fix_mem_proj(res, &am);
+
+ res = new_rd_ia32_Conv_I2I8Bit(dbgi, irg, new_block, noreg, noreg, res,
+ nomem, mode_Bu);
+
+ return res;
}
+
+static ir_node *create_cmov(long pnc, ir_node *cmp_left, ir_node *cmp_right,
+ ir_node *val_true, ir_node *val_false,
+ dbg_info *dbgi, ir_node *block)
+{
+ ir_graph *irg = current_ir_graph;
+ ir_node *new_block = be_transform_node(block);
+ ir_node *new_val_true = be_transform_node(val_true);
+ ir_node *new_val_false = be_transform_node(val_false);
+ ir_node *noreg = ia32_new_NoReg_gp(env_cg);
+ ir_node *nomem = new_NoMem();
+ ir_node *new_cmp_left;
+ ir_node *new_cmp_right;
+ ir_node *res;
+ ir_mode *mode;
+
+ /* cmovs with unknowns are pointless... */
+ if(is_Unknown(val_true)) {
+#ifdef DEBUG_libfirm
+ ir_fprintf(stderr, "Optimisation warning: psi with unknown operand\n");
#endif
+ return new_val_false;
+ }
+ if(is_Unknown(val_false)) {
+#ifdef DEBUG_libfirm
+ ir_fprintf(stderr, "Optimisation warning: psi with unknown operand\n");
+#endif
+ return new_val_true;
+ }
+
+ /* can we use a test instruction? */
+ if(is_Const_0(cmp_right)) {
+ long pure_pnc = pnc & ~ia32_pn_Cmp_Unsigned;
+ if(is_And(cmp_left) &&
+ (pure_pnc == pn_Cmp_Eq || pure_pnc == pn_Cmp_Lg)) {
+ ir_node *and_left = get_And_left(cmp_left);
+ ir_node *and_right = get_And_right(cmp_left);
+
+ mode = get_irn_mode(and_left);
+ new_cmp_left = be_transform_node(and_left);
+ new_cmp_right = create_immediate_or_transform(and_right, 0);
+ } else {
+ mode = get_irn_mode(cmp_left);
+ new_cmp_left = be_transform_node(cmp_left);
+ new_cmp_right = be_transform_node(cmp_left);
+ }
+
+ assert(get_mode_size_bits(mode) <= 32);
+
+ if(get_mode_size_bits(mode) == 8) {
+ res = new_rd_ia32_TestCMov8Bit(dbgi, current_ir_graph, new_block,
+ noreg, noreg, new_cmp_left,
+ new_cmp_right, nomem, new_val_true,
+ new_val_false, pnc);
+ } else {
+ res = new_rd_ia32_TestCMov(dbgi, current_ir_graph, new_block, noreg,
+ noreg, new_cmp_left, new_cmp_right,
+ nomem, new_val_true, new_val_false, pnc);
+ }
+ set_ia32_ls_mode(res, mode);
+
+ return res;
+ }
+
+ mode = get_irn_mode(cmp_left);
+ new_cmp_left = be_transform_node(cmp_left);
+ new_cmp_right = create_immediate_or_transform(cmp_right, 0);
+
+ /* no support for 8,16 bit modes yet */
+ assert(get_mode_size_bits(mode) <= 32);
+
+ if(get_mode_size_bits(mode) == 8) {
+ res = new_rd_ia32_CmpCMov8Bit(dbgi, irg, new_block, noreg, noreg,
+ new_cmp_left, new_cmp_right, nomem,
+ new_val_true, new_val_false, pnc);
+ } else {
+ res = new_rd_ia32_CmpCMov(dbgi, irg, new_block, noreg, noreg,
+ new_cmp_left, new_cmp_right, nomem,
+ new_val_true, new_val_false, pnc);
+ }
+ set_ia32_ls_mode(res, mode);
+
+ return res;
+}
-typedef ir_node *cmov_func_t(dbg_info *db, ir_graph *irg, ir_node *block,
- ir_node *cmp_a, ir_node *cmp_b, ir_node *psi_true,
- ir_node *psi_default);
/**
* Transforms a Psi node into CMov.
* @return The transformed node.
*/
static ir_node *gen_Psi(ir_node *node) {
- ir_node *block = be_transform_node(get_nodes_block(node));
- ir_node *psi_true = get_Psi_val(node, 0);
- ir_node *psi_default = get_Psi_default(node);
- ia32_code_gen_t *cg = env_cg;
- ir_graph *irg = current_ir_graph;
- dbg_info *dbgi = get_irn_dbg_info(node);
- ir_node *cond = get_Psi_cond(node, 0);
- ir_node *noreg = ia32_new_NoReg_gp(env_cg);
- ir_node *nomem = new_NoMem();
+ ir_node *psi_true = get_Psi_val(node, 0);
+ ir_node *psi_default = get_Psi_default(node);
+ ia32_code_gen_t *cg = env_cg;
+ ir_node *cond = get_Psi_cond(node, 0);
+ ir_node *block = get_nodes_block(node);
+ dbg_info *dbgi = get_irn_dbg_info(node);
ir_node *new_op;
- ir_node *cmp, *cmp_a, *cmp_b;
- ir_node *new_cmp_a, *new_cmp_b;
+ ir_node *cmp_left;
+ ir_node *cmp_right;
ir_mode *cmp_mode;
- int pnc;
+ long pnc;
assert(get_Psi_n_conds(node) == 1);
assert(get_irn_mode(cond) == mode_b);
+ assert(mode_needs_gp_reg(get_irn_mode(node)));
- if(is_And(cond) || is_Or(cond)) {
- ir_node *new_cond = be_transform_node(cond);
- ir_node *zero = new_rd_ia32_Immediate(NULL, irg, block, NULL, 0, 0);
- arch_set_irn_register(env_cg->arch_env, zero,
- &ia32_gp_regs[REG_GP_NOREG]);
-
- /* we have to compare the result against zero */
- new_cmp_a = new_cond;
- new_cmp_b = zero;
- cmp_mode = mode_Iu;
+ if(!is_Proj(cond) || !is_Cmp(get_Proj_pred(cond))) {
+ /* a mode_b value, we have to compare it against 0 */
+ cmp_left = cond;
+ cmp_right = new_Const_long(mode_Iu, 0);
pnc = pn_Cmp_Lg;
+ cmp_mode = mode_Iu;
} else {
- cmp = get_Proj_pred(cond);
- cmp_a = get_Cmp_left(cmp);
- cmp_b = get_Cmp_right(cmp);
- cmp_mode = get_irn_mode(cmp_a);
+ ir_node *cmp = get_Proj_pred(cond);
+
+ cmp_left = get_Cmp_left(cmp);
+ cmp_right = get_Cmp_right(cmp);
+ cmp_mode = get_irn_mode(cmp_left);
pnc = get_Proj_proj(cond);
- new_cmp_a = be_transform_node(cmp_a);
- new_cmp_b = create_immediate_or_transform(cmp_b, 0);
+ assert(!mode_is_float(cmp_mode));
if (!mode_is_signed(cmp_mode)) {
pnc |= ia32_pn_Cmp_Unsigned;
}
if(is_Const_1(psi_true) && is_Const_0(psi_default)) {
- new_op = new_rd_ia32_CmpSet(dbgi, irg, block, noreg, noreg,
- new_cmp_a, new_cmp_b, nomem, pnc);
+ new_op = create_set(pnc, cmp_left, cmp_right, dbgi, block);
} else if(is_Const_0(psi_true) && is_Const_1(psi_default)) {
pnc = get_negated_pnc(pnc, cmp_mode);
- new_op = new_rd_ia32_CmpSet(dbgi, irg, block, noreg, noreg,
- new_cmp_a, new_cmp_b, nomem, pnc);
+ new_op = create_set(pnc, cmp_left, cmp_right, dbgi, block);
} else {
- ir_node *new_psi_true = be_transform_node(psi_true);
- ir_node *new_psi_default = be_transform_node(psi_default);
- new_op = new_rd_ia32_CmpCMov(dbgi, irg, block, new_cmp_a, new_cmp_b,
- new_psi_true, new_psi_default, pnc);
+ new_op = create_cmov(pnc, cmp_left, cmp_right, psi_true, psi_default,
+ dbgi, block);
}
SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(cg, node));
return new_op;
}
-/**
- * Following conversion rules apply:
- *
- * INT -> INT
- * ============
- * 1) n bit -> m bit n > m (downscale)
- * always ignored
- * 2) n bit -> m bit n == m (sign change)
- * always ignored
- * 3) n bit -> m bit n < m (upscale)
- * a) source is signed: movsx
- * b) source is unsigned: and with lower bits sets
- *
- * INT -> FLOAT
- * ==============
- * SSE(1/2) convert to float or double (cvtsi2ss/sd)
- *
- * FLOAT -> INT
- * ==============
- * SSE(1/2) convert from float or double to 32bit int (cvtss/sd2si)
- *
- * FLOAT -> FLOAT
- * ================
- * SSE(1/2) convert from float or double to double or float (cvtss/sd2sd/ss)
- * x87 is mode_E internally, conversions happen only at load and store
- * in non-strict semantic
- */
-
/**
* Create a conversion from x87 state register to general purpose.
*/
dbg_info *dbgi = get_irn_dbg_info(node);
ir_node *noreg = ia32_new_NoReg_gp(cg);
ir_node *trunc_mode = ia32_new_Fpu_truncate(cg);
+ ir_mode *mode = get_irn_mode(node);
ir_node *fist, *load;
/* do a fist */
set_irn_pinned(fist, op_pin_state_floats);
set_ia32_use_frame(fist);
set_ia32_op_type(fist, ia32_AddrModeD);
- set_ia32_am_flavour(fist, ia32_am_B);
- set_ia32_ls_mode(fist, mode_Iu);
+
+ 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 */
+ if(get_mode_size_bits(mode) == 32 && !mode_is_signed(mode)) {
+ set_ia32_ls_mode(fist, mode_Ls);
+ } else {
+ set_ia32_ls_mode(fist, mode_Is);
+ }
SET_IA32_ORIG_NODE(fist, ia32_get_old_node_name(cg, node));
/* do a Load */
set_irn_pinned(load, op_pin_state_floats);
set_ia32_use_frame(load);
set_ia32_op_type(load, ia32_AddrModeS);
- set_ia32_am_flavour(load, ia32_am_B);
- set_ia32_ls_mode(load, mode_Iu);
+ set_ia32_ls_mode(load, mode_Is);
+ if(get_ia32_ls_mode(fist) == mode_Ls) {
+ ia32_attr_t *attr = get_ia32_attr(load);
+ attr->data.need_64bit_stackent = 1;
+ } else {
+ ia32_attr_t *attr = get_ia32_attr(load);
+ attr->data.need_32bit_stackent = 1;
+ }
SET_IA32_ORIG_NODE(load, ia32_get_old_node_name(cg, node));
return new_r_Proj(irg, block, load, mode_Iu, pn_ia32_Load_res);
}
+/**
+ * Creates a x87 strict Conv by placing a Sore and a Load
+ */
+static ir_node *gen_x87_strict_conv(ir_mode *tgt_mode, ir_node *node)
+{
+ ir_node *block = get_nodes_block(node);
+ ir_graph *irg = current_ir_graph;
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ ir_node *noreg = ia32_new_NoReg_gp(env_cg);
+ ir_node *nomem = new_NoMem();
+ ir_node *frame = get_irg_frame(irg);
+ ir_node *store, *load;
+ ir_node *res;
+
+ store = new_rd_ia32_vfst(dbgi, irg, block, frame, noreg, node, nomem,
+ tgt_mode);
+ set_ia32_use_frame(store);
+ set_ia32_op_type(store, ia32_AddrModeD);
+ SET_IA32_ORIG_NODE(store, ia32_get_old_node_name(env_cg, node));
+
+ load = new_rd_ia32_vfld(dbgi, irg, block, frame, noreg, store,
+ tgt_mode);
+ set_ia32_use_frame(load);
+ set_ia32_op_type(load, ia32_AddrModeS);
+ SET_IA32_ORIG_NODE(load, ia32_get_old_node_name(env_cg, node));
+
+ res = new_r_Proj(irg, block, load, mode_E, pn_ia32_vfld_res);
+ return res;
+}
+
/**
* Create a conversion from general purpose to x87 register
*/
dbg_info *dbgi = get_irn_dbg_info(node);
ir_node *noreg = ia32_new_NoReg_gp(env_cg);
ir_node *nomem = new_NoMem();
+ ir_mode *mode = get_irn_mode(op);
+ ir_mode *store_mode;
ir_node *fild, *store;
- int src_bits;
+ ir_node *res;
+ int src_bits;
- /* first convert to 32 bit if necessary */
+ /* first convert to 32 bit signed if necessary */
src_bits = get_mode_size_bits(src_mode);
if (src_bits == 8) {
- new_op = new_rd_ia32_Conv_I2I8Bit(dbgi, irg, block, noreg, noreg, new_op, nomem);
- set_ia32_am_support(new_op, ia32_am_Source, ia32_am_unary);
- set_ia32_ls_mode(new_op, src_mode);
+ new_op = new_rd_ia32_Conv_I2I8Bit(dbgi, irg, block, noreg, noreg, new_op, nomem,
+ src_mode);
SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
+ mode = mode_Is;
} else if (src_bits < 32) {
- new_op = new_rd_ia32_Conv_I2I(dbgi, irg, block, noreg, noreg, new_op, nomem);
- set_ia32_am_support(new_op, ia32_am_Source, ia32_am_unary);
- set_ia32_ls_mode(new_op, src_mode);
+ new_op = new_rd_ia32_Conv_I2I(dbgi, irg, block, noreg, noreg, new_op, nomem, src_mode);
SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
+ mode = mode_Is;
}
+ assert(get_mode_size_bits(mode) == 32);
+
/* do a store */
store = new_rd_ia32_Store(dbgi, irg, block, get_irg_frame(irg), noreg, new_op, nomem);
set_ia32_use_frame(store);
set_ia32_op_type(store, ia32_AddrModeD);
- set_ia32_am_flavour(store, ia32_am_OB);
set_ia32_ls_mode(store, mode_Iu);
+ /* exception for 32bit unsigned, do a 64bit spill+load */
+ if(!mode_is_signed(mode)) {
+ ir_node *in[2];
+ /* store a zero */
+ ir_node *zero_const = create_Immediate_from_int(0);
+
+ ir_node *zero_store = new_rd_ia32_Store(dbgi, irg, block, get_irg_frame(irg), noreg,
+ zero_const, nomem);
+
+ 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;
+
+ store = new_rd_Sync(dbgi, irg, block, 2, in);
+ store_mode = mode_Ls;
+ } else {
+ store_mode = mode_Is;
+ }
+
/* do a fild */
fild = new_rd_ia32_vfild(dbgi, irg, block, get_irg_frame(irg), noreg, store);
set_ia32_use_frame(fild);
set_ia32_op_type(fild, ia32_AddrModeS);
- set_ia32_am_flavour(fild, ia32_am_OB);
- set_ia32_ls_mode(fild, mode_Iu);
+ set_ia32_ls_mode(fild, store_mode);
+
+ res = new_r_Proj(irg, block, fild, mode_vfp, pn_ia32_vfild_res);
- return new_r_Proj(irg, block, fild, mode_vfp, pn_ia32_vfild_res);
+ return res;
}
-static ir_node *create_Strict_conv(ir_mode *src_mode, ir_mode *tgt_mode,
- ir_node *node)
+/**
+ * Crete a conversion from one integer mode into another one
+ */
+static ir_node *create_I2I_Conv(ir_mode *src_mode, ir_mode *tgt_mode,
+ dbg_info *dbgi, ir_node *new_block,
+ ir_node *new_op)
{
- ir_node *block = get_nodes_block(node);
ir_graph *irg = current_ir_graph;
- dbg_info *dbgi = get_irn_dbg_info(node);
- ir_node *noreg = ia32_new_NoReg_gp(env_cg);
- ir_node *nomem = new_NoMem();
int src_bits = get_mode_size_bits(src_mode);
int tgt_bits = get_mode_size_bits(tgt_mode);
- ir_node *frame = get_irg_frame(irg);
- ir_mode *smaller_mode;
- ir_node *store, *load;
+ ir_node *noreg = ia32_new_NoReg_gp(env_cg);
+ ir_node *nomem = new_rd_NoMem(irg);
ir_node *res;
+ ir_mode *smaller_mode;
+ int smaller_bits;
- if(src_bits <= tgt_bits)
+ if (src_bits < tgt_bits) {
smaller_mode = src_mode;
- else
+ smaller_bits = src_bits;
+ } else {
smaller_mode = tgt_mode;
+ smaller_bits = tgt_bits;
+ }
- store = new_rd_ia32_vfst(dbgi, irg, block, frame, noreg, node, nomem,
- smaller_mode);
- set_ia32_use_frame(store);
- set_ia32_op_type(store, ia32_AddrModeD);
- set_ia32_am_flavour(store, ia32_am_OB);
-
- load = new_rd_ia32_vfld(dbgi, irg, block, frame, noreg, store,
- smaller_mode);
- set_ia32_use_frame(load);
- set_ia32_op_type(load, ia32_AddrModeS);
- set_ia32_am_flavour(load, ia32_am_OB);
+ DB((dbg, LEVEL_1, "create Conv(int, int) ...", src_mode, tgt_mode));
+ if (smaller_bits == 8) {
+ res = new_rd_ia32_Conv_I2I8Bit(dbgi, irg, new_block, noreg, noreg,
+ new_op, nomem, smaller_mode);
+ } else {
+ res = new_rd_ia32_Conv_I2I(dbgi, irg, new_block, noreg, noreg, new_op,
+ nomem, smaller_mode);
+ }
- res = new_r_Proj(irg, block, load, mode_E, pn_ia32_vfld_res);
return res;
}
dbg_info *dbgi = get_irn_dbg_info(node);
ir_mode *src_mode = get_irn_mode(op);
ir_mode *tgt_mode = get_irn_mode(node);
- int src_bits = get_mode_size_bits(src_mode);
- int tgt_bits = get_mode_size_bits(tgt_mode);
+ int src_bits = get_mode_size_bits(src_mode);
+ int tgt_bits = get_mode_size_bits(tgt_mode);
ir_node *noreg = ia32_new_NoReg_gp(env_cg);
ir_node *nomem = new_rd_NoMem(irg);
ir_node *res;
+ if (src_mode == mode_b) {
+ assert(mode_is_int(tgt_mode));
+ /* nothing to do, we already model bools as 0/1 ints */
+ return new_op;
+ }
+
if (src_mode == tgt_mode) {
if (get_Conv_strict(node)) {
if (USE_SSE2(env_cg)) {
res = new_rd_ia32_Conv_FP2FP(dbgi, irg, block, noreg, noreg, new_op, nomem);
set_ia32_ls_mode(res, tgt_mode);
} else {
- // Matze: TODO what about strict convs?
if(get_Conv_strict(node)) {
- res = create_Strict_conv(src_mode, tgt_mode, new_op);
+ res = gen_x87_strict_conv(tgt_mode, new_op);
SET_IA32_ORIG_NODE(get_Proj_pred(res), ia32_get_old_node_name(env_cg, node));
return res;
}
} else {
/* we convert from int ... */
if (mode_is_float(tgt_mode)) {
- FP_USED(env_cg);
/* ... to float */
DB((dbg, LEVEL_1, "create Conv(int, float) ..."));
if (USE_SSE2(env_cg)) {
res = new_rd_ia32_Conv_I2FP(dbgi, irg, block, noreg, noreg, new_op, nomem);
set_ia32_ls_mode(res, tgt_mode);
- if(src_bits == 32) {
- set_ia32_am_support(res, ia32_am_Source, ia32_am_unary);
- }
} else {
- return gen_x87_gp_to_fp(node, src_mode);
+ res = gen_x87_gp_to_fp(node, src_mode);
+ if(get_Conv_strict(node)) {
+ res = gen_x87_strict_conv(tgt_mode, res);
+ SET_IA32_ORIG_NODE(get_Proj_pred(res),
+ ia32_get_old_node_name(env_cg, node));
+ }
+ return res;
}
+ } else if(tgt_mode == mode_b) {
+ /* mode_b lowering already took care that we only have 0/1 values */
+ DB((dbg, LEVEL_1, "omitting unnecessary Conv(%+F, %+F) ...",
+ src_mode, tgt_mode));
+ return new_op;
} else {
/* to int */
- ir_mode *smaller_mode;
- int smaller_bits;
-
if (src_bits == tgt_bits) {
- DB((dbg, LEVEL_1, "omitting unnecessary Conv(%+F, %+F) ...", src_mode, tgt_mode));
+ DB((dbg, LEVEL_1, "omitting unnecessary Conv(%+F, %+F) ...",
+ src_mode, tgt_mode));
return new_op;
}
- if (src_bits < tgt_bits) {
- smaller_mode = src_mode;
- smaller_bits = src_bits;
- } else {
- smaller_mode = tgt_mode;
- smaller_bits = tgt_bits;
- }
-
- DB((dbg, LEVEL_1, "create Conv(int, int) ...", src_mode, tgt_mode));
- if (smaller_bits == 8) {
- res = new_rd_ia32_Conv_I2I8Bit(dbgi, irg, block, noreg, noreg, new_op, nomem);
- set_ia32_ls_mode(res, smaller_mode);
- } else {
- res = new_rd_ia32_Conv_I2I(dbgi, irg, block, noreg, noreg, new_op, nomem);
- set_ia32_ls_mode(res, smaller_mode);
- }
- set_ia32_am_support(res, ia32_am_Source, ia32_am_unary);
+ res = create_I2I_Conv(src_mode, tgt_mode, dbgi, block, new_op);
}
}
ir_node *block;
mode = get_irn_mode(node);
- if(!mode_is_int(mode) && !mode_is_character(mode) &&
- !mode_is_reference(mode)) {
+ if(!mode_is_int(mode) && !mode_is_reference(mode)) {
return NULL;
}
irg = current_ir_graph;
dbgi = get_irn_dbg_info(node);
block = get_irg_start_block(irg);
- res = new_rd_ia32_Immediate(dbgi, irg, block, symconst_ent, symconst_sign,
- val);
+ res = new_rd_ia32_Immediate(dbgi, irg, block, symconst_ent,
+ symconst_sign, val);
arch_set_irn_register(env_cg->arch_env, res, &ia32_gp_regs[REG_GP_NOREG]);
- /* make sure we don't schedule stuff before the barrier */
- add_irn_dep(res, get_irg_frame(irg));
-
return res;
}
char immediate_type = 0;
unsigned limited = 0;
const arch_register_class_t *cls = NULL;
- ir_graph *irg;
- struct obstack *obst;
+ ir_graph *irg = current_ir_graph;
+ struct obstack *obst = get_irg_obstack(irg);
arch_register_req_t *req;
unsigned *limited_ptr;
int p;
assert(constraint->is_in
&& "imeediates make no sense for output constraints");
}
- /* todo: check types (no float input on 'r' constrainted in and such... */
-
- irg = current_ir_graph;
- obst = get_irg_obstack(irg);
+ /* todo: check types (no float input on 'r' constrained in and such... */
if(limited != 0) {
req = obstack_alloc(obst, sizeof(req[0]) + sizeof(unsigned));
panic("Clobbers not supported yet");
}
-ir_node *gen_ASM(ir_node *node)
+/**
+ * generates code for a ASM node
+ */
+static ir_node *gen_ASM(ir_node *node)
{
int i, arity;
ir_graph *irg = current_ir_graph;
struct obstack *obst;
constraint_t parsed_constraint;
- /* assembler could contain float statements */
- FP_USED(env_cg);
-
/* transform inputs */
arity = get_irn_arity(node);
in = alloca(arity * sizeof(in[0]));
* | |_) | __/ | | | (_) | (_| | __/\__ \
* |_.__/ \___|_| |_|\___/ \__,_|\___||___/
*
- ********************************************/
-
-static ir_node *gen_be_StackParam(ir_node *node) {
- ir_node *block = be_transform_node(get_nodes_block(node));
- ir_node *ptr = get_irn_n(node, be_pos_StackParam_ptr);
- ir_node *new_ptr = be_transform_node(ptr);
- ir_node *new_op = NULL;
- ir_graph *irg = current_ir_graph;
- dbg_info *dbgi = get_irn_dbg_info(node);
- ir_node *nomem = new_rd_NoMem(current_ir_graph);
- ir_entity *ent = arch_get_frame_entity(env_cg->arch_env, node);
- ir_mode *load_mode = get_irn_mode(node);
- ir_node *noreg = ia32_new_NoReg_gp(env_cg);
- ir_mode *proj_mode;
- long pn_res;
-
- if (mode_is_float(load_mode)) {
- FP_USED(env_cg);
- if (USE_SSE2(env_cg)) {
- new_op = new_rd_ia32_xLoad(dbgi, irg, block, new_ptr, noreg, nomem);
- pn_res = pn_ia32_xLoad_res;
- proj_mode = mode_xmm;
- } else {
- new_op = new_rd_ia32_vfld(dbgi, irg, block, new_ptr, noreg, nomem, load_mode);
- pn_res = pn_ia32_vfld_res;
- proj_mode = mode_vfp;
- }
- } else {
- new_op = new_rd_ia32_Load(dbgi, irg, block, new_ptr, noreg, nomem);
- proj_mode = mode_Iu;
- pn_res = pn_ia32_Load_res;
- }
-
- set_irn_pinned(new_op, op_pin_state_floats);
- set_ia32_frame_ent(new_op, ent);
- set_ia32_use_frame(new_op);
-
- set_ia32_op_type(new_op, ia32_AddrModeS);
- set_ia32_am_flavour(new_op, ia32_am_B);
- set_ia32_ls_mode(new_op, load_mode);
- set_ia32_flags(new_op, get_ia32_flags(new_op) | arch_irn_flags_rematerializable);
-
- SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
-
- return new_rd_Proj(dbgi, irg, block, new_op, proj_mode, pn_res);
-}
+ ********************************************/
/**
* Transforms a FrameAddr into an ia32 Add.
res = new_rd_ia32_Lea(dbgi, irg, block, new_op, noreg);
set_ia32_frame_ent(res, arch_get_frame_entity(env_cg->arch_env, node));
set_ia32_use_frame(res);
- set_ia32_am_flavour(res, ia32_am_OB);
SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
return res;
}
-/**
- * Transforms a FrameLoad into an ia32 Load.
- */
-static ir_node *gen_be_FrameLoad(ir_node *node) {
- ir_node *block = be_transform_node(get_nodes_block(node));
- ir_node *mem = get_irn_n(node, be_pos_FrameLoad_mem);
- ir_node *new_mem = be_transform_node(mem);
- ir_node *ptr = get_irn_n(node, be_pos_FrameLoad_ptr);
- ir_node *new_ptr = be_transform_node(ptr);
- ir_node *new_op = NULL;
- ir_graph *irg = current_ir_graph;
- dbg_info *dbgi = get_irn_dbg_info(node);
- ir_node *noreg = ia32_new_NoReg_gp(env_cg);
- ir_entity *ent = arch_get_frame_entity(env_cg->arch_env, node);
- ir_mode *mode = get_type_mode(get_entity_type(ent));
- ir_node *projs[pn_Load_max];
-
- ia32_collect_Projs(node, projs, pn_Load_max);
-
- if (mode_is_float(mode)) {
- FP_USED(env_cg);
- if (USE_SSE2(env_cg)) {
- new_op = new_rd_ia32_xLoad(dbgi, irg, block, new_ptr, noreg, new_mem);
- }
- else {
- new_op = new_rd_ia32_vfld(dbgi, irg, block, new_ptr, noreg, new_mem, mode);
- }
- }
- else {
- new_op = new_rd_ia32_Load(dbgi, irg, block, new_ptr, noreg, new_mem);
- }
-
- set_irn_pinned(new_op, op_pin_state_floats);
- set_ia32_frame_ent(new_op, ent);
- set_ia32_use_frame(new_op);
-
- set_ia32_op_type(new_op, ia32_AddrModeS);
- set_ia32_am_flavour(new_op, ia32_am_B);
- set_ia32_ls_mode(new_op, mode);
- set_ia32_flags(new_op, get_ia32_flags(new_op) | arch_irn_flags_rematerializable);
-
- SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
-
- return new_op;
-}
-
-
-/**
- * Transforms a FrameStore into an ia32 Store.
- */
-static ir_node *gen_be_FrameStore(ir_node *node) {
- ir_node *block = be_transform_node(get_nodes_block(node));
- ir_node *mem = get_irn_n(node, be_pos_FrameStore_mem);
- ir_node *new_mem = be_transform_node(mem);
- ir_node *ptr = get_irn_n(node, be_pos_FrameStore_ptr);
- ir_node *new_ptr = be_transform_node(ptr);
- ir_node *val = get_irn_n(node, be_pos_FrameStore_val);
- ir_node *new_val = be_transform_node(val);
- ir_node *new_op = NULL;
- ir_graph *irg = current_ir_graph;
- dbg_info *dbgi = get_irn_dbg_info(node);
- ir_node *noreg = ia32_new_NoReg_gp(env_cg);
- ir_entity *ent = arch_get_frame_entity(env_cg->arch_env, node);
- ir_mode *mode = get_irn_mode(val);
-
- if (mode_is_float(mode)) {
- FP_USED(env_cg);
- if (USE_SSE2(env_cg)) {
- new_op = new_rd_ia32_xStore(dbgi, irg, block, new_ptr, noreg, new_val, new_mem);
- } else {
- new_op = new_rd_ia32_vfst(dbgi, irg, block, new_ptr, noreg, new_val, new_mem, mode);
- }
- } else if (get_mode_size_bits(mode) == 8) {
- new_op = new_rd_ia32_Store8Bit(dbgi, irg, block, new_ptr, noreg, new_val, new_mem);
- } else {
- new_op = new_rd_ia32_Store(dbgi, irg, block, new_ptr, noreg, new_val, new_mem);
- }
-
- set_ia32_frame_ent(new_op, ent);
- set_ia32_use_frame(new_op);
-
- set_ia32_op_type(new_op, ia32_AddrModeD);
- set_ia32_am_flavour(new_op, ia32_am_B);
- set_ia32_ls_mode(new_op, mode);
-
- SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
-
- return new_op;
-}
-
/**
* In case SSE is used we need to copy the result from XMM0 to FPU TOS before return.
*/
noreg = ia32_new_NoReg_gp(env_cg);
/* store xmm0 onto stack */
- sse_store = new_rd_ia32_xStoreSimple(dbgi, irg, block, frame, noreg, new_ret_val, new_ret_mem);
+ sse_store = new_rd_ia32_xStoreSimple(dbgi, irg, block, frame, noreg,
+ new_ret_val, new_ret_mem);
set_ia32_ls_mode(sse_store, mode);
set_ia32_op_type(sse_store, ia32_AddrModeD);
set_ia32_use_frame(sse_store);
- set_ia32_am_flavour(sse_store, ia32_am_B);
- /* load into st0 */
- fld = new_rd_ia32_SetST0(dbgi, irg, block, frame, noreg, sse_store);
- set_ia32_ls_mode(fld, mode);
+ /* load into x87 register */
+ fld = new_rd_ia32_vfld(dbgi, irg, block, frame, noreg, sse_store, mode);
set_ia32_op_type(fld, ia32_AddrModeS);
set_ia32_use_frame(fld);
- set_ia32_am_flavour(fld, ia32_am_B);
- mproj = new_r_Proj(irg, block, fld, mode_M, pn_ia32_SetST0_M);
- fld = new_r_Proj(irg, block, fld, mode_vfp, pn_ia32_SetST0_res);
- arch_set_irn_register(env_cg->arch_env, fld, &ia32_vfp_regs[REG_VF0]);
+ mproj = new_r_Proj(irg, block, fld, mode_M, pn_ia32_vfld_M);
+ fld = new_r_Proj(irg, block, fld, mode_vfp, pn_ia32_vfld_res);
/* create a new barrier */
arity = get_irn_arity(barrier);
/* ia32 stack grows in reverse direction, make a SubSP */
new_op = new_rd_ia32_SubSP(dbgi, irg, block, noreg, noreg, new_sp, new_sz,
nomem);
- set_ia32_am_support(new_op, ia32_am_Source, ia32_am_binary);
SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
return new_op;
/* ia32 stack grows in reverse direction, make an AddSP */
new_op = new_rd_ia32_AddSP(dbgi, irg, block, noreg, noreg, new_sp, new_sz, nomem);
- set_ia32_am_support(new_op, ia32_am_Source, ia32_am_binary);
SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
return new_op;
ir_mode *mode = get_irn_mode(node);
if (mode_is_float(mode)) {
- if (USE_SSE2(env_cg))
+ if (USE_SSE2(env_cg)) {
return ia32_new_Unknown_xmm(env_cg);
- else
- return ia32_new_Unknown_vfp(env_cg);
+ } else {
+ /* Unknown nodes are buggy in x87 sim, use zero for now... */
+ ir_graph *irg = current_ir_graph;
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ ir_node *block = get_irg_start_block(irg);
+ return new_rd_ia32_vfldz(dbgi, irg, block);
+ }
} else if (mode_needs_gp_reg(mode)) {
return ia32_new_Unknown_gp(env_cg);
} else {
/* phi nodes allow loops, so we use the old arguments for now
* and fix this later */
- phi = new_ir_node(dbgi, irg, block, op_Phi, mode, get_irn_arity(node), get_irn_in(node) + 1);
+ phi = new_ir_node(dbgi, irg, block, op_Phi, mode, get_irn_arity(node),
+ get_irn_in(node) + 1);
copy_node_attr(node, phi);
be_duplicate_deps(node, phi);
return phi;
}
+/**
+ * Transform IJmp
+ */
+static ir_node *gen_IJmp(ir_node *node) {
+ /* TODO: support AM */
+ return gen_unop(node, get_IJmp_target(node), new_rd_ia32_IJmp);
+}
+
+
/**********************************************************************
* _ _ _
* | | | | | |
/**
* Transforms a lowered Load into a "real" one.
*/
-static ir_node *gen_lowered_Load(ir_node *node, construct_load_func func, char fp_unit) {
+static ir_node *gen_lowered_Load(ir_node *node, construct_load_func func)
+{
ir_node *block = be_transform_node(get_nodes_block(node));
ir_node *ptr = get_irn_n(node, 0);
ir_node *new_ptr = be_transform_node(ptr);
ir_node *noreg = ia32_new_NoReg_gp(env_cg);
ir_node *new_op;
- /*
- Could be that we have SSE2 unit, but due to 64Bit Div/Conv
- lowering we have x87 nodes, so we need to enforce simulation.
- */
- if (mode_is_float(mode)) {
- FP_USED(env_cg);
- if (fp_unit == fp_x87)
- FORCE_x87(env_cg);
- }
-
new_op = func(dbgi, irg, block, new_ptr, noreg, new_mem);
set_ia32_op_type(new_op, ia32_AddrModeS);
- set_ia32_am_flavour(new_op, ia32_am_OB);
- set_ia32_am_offs_int(new_op, 0);
- set_ia32_am_scale(new_op, 1);
+ set_ia32_am_offs_int(new_op, get_ia32_am_offs_int(node));
+ set_ia32_am_scale(new_op, get_ia32_am_scale(node));
set_ia32_am_sc(new_op, get_ia32_am_sc(node));
if (is_ia32_am_sc_sign(node))
set_ia32_am_sc_sign(new_op);
- set_ia32_ls_mode(new_op, get_ia32_ls_mode(node));
+ set_ia32_ls_mode(new_op, mode);
if (is_ia32_use_frame(node)) {
set_ia32_frame_ent(new_op, get_ia32_frame_ent(node));
set_ia32_use_frame(new_op);
}
/**
-* Transforms a lowered Store into a "real" one.
-*/
-static ir_node *gen_lowered_Store(ir_node *node, construct_store_func func, char fp_unit) {
+ * Transforms a lowered Store into a "real" one.
+ */
+static ir_node *gen_lowered_Store(ir_node *node, construct_store_func func)
+{
ir_node *block = be_transform_node(get_nodes_block(node));
ir_node *ptr = get_irn_n(node, 0);
ir_node *new_ptr = be_transform_node(ptr);
ir_mode *mode = get_ia32_ls_mode(node);
ir_node *new_op;
long am_offs;
- ia32_am_flavour_t am_flav = ia32_B;
-
- /*
- Could be that we have SSE2 unit, but due to 64Bit Div/Conv
- lowering we have x87 nodes, so we need to enforce simulation.
- */
- if (mode_is_float(mode)) {
- FP_USED(env_cg);
- if (fp_unit == fp_x87)
- FORCE_x87(env_cg);
- }
new_op = func(dbgi, irg, block, new_ptr, noreg, new_val, new_mem);
- if ((am_offs = get_ia32_am_offs_int(node)) != 0) {
- am_flav |= ia32_O;
- add_ia32_am_offs_int(new_op, am_offs);
- }
+ am_offs = get_ia32_am_offs_int(node);
+ add_ia32_am_offs_int(new_op, am_offs);
set_ia32_op_type(new_op, ia32_AddrModeD);
- set_ia32_am_flavour(new_op, am_flav);
set_ia32_ls_mode(new_op, mode);
set_ia32_frame_ent(new_op, get_ia32_frame_ent(node));
set_ia32_use_frame(new_op);
/**
* Transforms an ia32_l_XXX into a "real" XXX node
*
- * @param env The transformation environment
+ * @param node The node to transform
* @return the created ia32 XXX node
*/
#define GEN_LOWERED_OP(op) \
static ir_node *gen_ia32_l_##op(ir_node *node) { \
- ir_mode *mode = get_irn_mode(node); \
- if (mode_is_float(mode)) \
- FP_USED(env_cg); \
return gen_binop(node, get_binop_left(node), \
get_binop_right(node), new_rd_ia32_##op,0); \
}
#define GEN_LOWERED_x87_OP(op) \
static ir_node *gen_ia32_l_##op(ir_node *node) { \
ir_node *new_op; \
- FORCE_x87(env_cg); \
new_op = gen_binop_x87_float(node, get_binop_left(node), \
get_binop_right(node), new_rd_ia32_##op); \
return new_op; \
}
-#define GEN_LOWERED_UNOP(op) \
- static ir_node *gen_ia32_l_##op(ir_node *node) {\
- return gen_unop(node, get_unop_op(node), new_rd_ia32_##op); \
- }
-
-#define GEN_LOWERED_SHIFT_OP(op) \
- static ir_node *gen_ia32_l_##op(ir_node *node) {\
- return gen_shift_binop(node, get_binop_left(node), \
- get_binop_right(node), new_rd_ia32_##op); \
- }
-
-#define GEN_LOWERED_LOAD(op, fp_unit) \
- static ir_node *gen_ia32_l_##op(ir_node *node) {\
- return gen_lowered_Load(node, new_rd_ia32_##op, fp_unit); \
- }
-
-#define GEN_LOWERED_STORE(op, fp_unit) \
- static ir_node *gen_ia32_l_##op(ir_node *node) {\
- return gen_lowered_Store(node, new_rd_ia32_##op, fp_unit); \
+#define GEN_LOWERED_SHIFT_OP(l_op, op) \
+ static ir_node *gen_ia32_##l_op(ir_node *node) { \
+ return gen_shift_binop(node, get_irn_n(node, 0), \
+ get_irn_n(node, 1), new_rd_ia32_##op); \
}
GEN_LOWERED_OP(Adc)
GEN_LOWERED_OP(Add)
GEN_LOWERED_OP(Sbb)
GEN_LOWERED_OP(Sub)
-GEN_LOWERED_OP(IMul)
GEN_LOWERED_OP(Xor)
GEN_LOWERED_x87_OP(vfprem)
GEN_LOWERED_x87_OP(vfmul)
GEN_LOWERED_x87_OP(vfsub)
+GEN_LOWERED_SHIFT_OP(l_ShlDep, Shl)
+GEN_LOWERED_SHIFT_OP(l_ShrDep, Shr)
+GEN_LOWERED_SHIFT_OP(l_Sar, Sar)
+GEN_LOWERED_SHIFT_OP(l_SarDep, Sar)
+
+
+/**
+ * Transforms an ia32_l_Neg into a "real" ia32_Neg node
+ *
+ * @param node The node to transform
+ * @return the created ia32 Neg node
+ */
+static ir_node *gen_ia32_l_Neg(ir_node *node) {
+ return gen_unop(node, get_unop_op(node), new_rd_ia32_Neg);
+}
+
+/**
+ * Transforms an ia32_l_vfild into a "real" ia32_vfild node
+ *
+ * @param node The node to transform
+ * @return the created ia32 vfild node
+ */
+static ir_node *gen_ia32_l_vfild(ir_node *node) {
+ return gen_lowered_Load(node, new_rd_ia32_vfild);
+}
+
+/**
+ * Transforms an ia32_l_Load into a "real" ia32_Load node
+ *
+ * @param node The node to transform
+ * @return the created ia32 Load node
+ */
+static ir_node *gen_ia32_l_Load(ir_node *node) {
+ return gen_lowered_Load(node, new_rd_ia32_Load);
+}
-GEN_LOWERED_UNOP(Neg)
+/**
+ * Transforms an ia32_l_Store into a "real" ia32_Store node
+ *
+ * @param node The node to transform
+ * @return the created ia32 Store node
+ */
+static ir_node *gen_ia32_l_Store(ir_node *node) {
+ return gen_lowered_Store(node, new_rd_ia32_Store);
+}
-GEN_LOWERED_LOAD(vfild, fp_x87)
-GEN_LOWERED_LOAD(Load, fp_none)
-/*GEN_LOWERED_STORE(vfist, fp_x87)
- *TODO
+/**
+ * Transforms a l_vfist into a "real" vfist node.
+ *
+ * @param node The node to transform
+ * @return the created ia32 vfist node
*/
-GEN_LOWERED_STORE(Store, fp_none)
+static ir_node *gen_ia32_l_vfist(ir_node *node) {
+ ir_node *block = be_transform_node(get_nodes_block(node));
+ ir_node *ptr = get_irn_n(node, 0);
+ ir_node *new_ptr = be_transform_node(ptr);
+ ir_node *val = get_irn_n(node, 1);
+ ir_node *new_val = be_transform_node(val);
+ ir_node *mem = get_irn_n(node, 2);
+ ir_node *new_mem = be_transform_node(mem);
+ ir_graph *irg = current_ir_graph;
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ ir_node *noreg = ia32_new_NoReg_gp(env_cg);
+ ir_mode *mode = get_ia32_ls_mode(node);
+ ir_node *trunc_mode = ia32_new_Fpu_truncate(env_cg);
+ ir_node *new_op;
+ long am_offs;
+
+ new_op = new_rd_ia32_vfist(dbgi, irg, block, new_ptr, noreg, new_val,
+ trunc_mode, new_mem);
+
+ am_offs = get_ia32_am_offs_int(node);
+ add_ia32_am_offs_int(new_op, am_offs);
+
+ set_ia32_op_type(new_op, ia32_AddrModeD);
+ set_ia32_ls_mode(new_op, mode);
+ set_ia32_frame_ent(new_op, get_ia32_frame_ent(node));
+ set_ia32_use_frame(new_op);
+ SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
+
+ return new_op;
+}
+
+/**
+ * Transforms a l_vfdiv into a "real" vfdiv node.
+ *
+ * @param env The transformation environment
+ * @return the created ia32 vfdiv node
+ */
static ir_node *gen_ia32_l_vfdiv(ir_node *node) {
ir_node *block = be_transform_node(get_nodes_block(node));
ir_node *left = get_binop_left(node);
ir_node *noreg = ia32_new_NoReg_gp(env_cg);
ir_graph *irg = current_ir_graph;
dbg_info *dbgi = get_irn_dbg_info(node);
- ir_node *fpcw = be_abi_get_ignore_irn(env_cg->birg->abi,
- &ia32_fp_cw_regs[REG_FPCW]);
+ ir_node *fpcw = get_fpcw();
ir_node *vfdiv;
vfdiv = new_rd_ia32_vfdiv(dbgi, irg, block, noreg, noreg, new_left,
new_right, new_NoMem(), fpcw);
clear_ia32_commutative(vfdiv);
- set_ia32_am_support(vfdiv, ia32_am_Source, ia32_am_binary);
SET_IA32_ORIG_NODE(vfdiv, ia32_get_old_node_name(env_cg, node));
- FORCE_x87(env_cg);
-
return vfdiv;
}
ir_node *noreg = ia32_new_NoReg_gp(env_cg);
ir_graph *irg = current_ir_graph;
dbg_info *dbgi = get_irn_dbg_info(node);
- ir_node *in[2];
/* l_Mul is already a mode_T node, so we create the Mul in the normal way */
/* and then skip the result Proj, because all needed Projs are already there. */
ir_node *muls = new_rd_ia32_Mul(dbgi, irg, block, noreg, noreg, new_left,
new_right, new_NoMem());
clear_ia32_commutative(muls);
- set_ia32_am_support(muls, ia32_am_Source, ia32_am_binary);
-
- /* check if EAX and EDX proj exist, add missing one */
- in[0] = new_rd_Proj(dbgi, irg, block, muls, mode_Iu, pn_EAX);
- in[1] = new_rd_Proj(dbgi, irg, block, muls, mode_Iu, pn_EDX);
- be_new_Keep(&ia32_reg_classes[CLASS_ia32_gp], irg, block, 2, in);
SET_IA32_ORIG_NODE(muls, ia32_get_old_node_name(env_cg, node));
return muls;
}
-GEN_LOWERED_SHIFT_OP(Shl)
-GEN_LOWERED_SHIFT_OP(Shr)
-GEN_LOWERED_SHIFT_OP(Sar)
+/**
+ * Transforms a l_IMulS into a "real" IMul1OPS node.
+ *
+ * @param env The transformation environment
+ * @return the created ia32 IMul1OP node
+ */
+static ir_node *gen_ia32_l_IMul(ir_node *node) {
+ ir_node *block = be_transform_node(get_nodes_block(node));
+ ir_node *left = get_binop_left(node);
+ ir_node *new_left = be_transform_node(left);
+ ir_node *right = get_binop_right(node);
+ ir_node *new_right = be_transform_node(right);
+ ir_node *noreg = ia32_new_NoReg_gp(env_cg);
+ ir_graph *irg = current_ir_graph;
+ dbg_info *dbgi = get_irn_dbg_info(node);
+
+ /* l_IMul is already a mode_T node, so we create the IMul1OP in the normal way */
+ /* and then skip the result Proj, because all needed Projs are already there. */
+ ir_node *muls = new_rd_ia32_IMul1OP(dbgi, irg, block, noreg, noreg, new_left,
+ new_right, new_NoMem());
+ clear_ia32_commutative(muls);
+ set_ia32_am_support(muls, ia32_am_Source, ia32_am_binary);
+
+ SET_IA32_ORIG_NODE(muls, ia32_get_old_node_name(env_cg, node));
+
+ return muls;
+}
/**
* Transforms a l_ShlD/l_ShrD into a ShlD/ShrD. Those nodes have 3 data inputs:
ir_node *op2, ir_node *count)
{
ir_node *block = be_transform_node(get_nodes_block(node));
- ir_node *new_op1 = be_transform_node(op1);
- ir_node *new_op2 = be_transform_node(op2);
- ir_node *new_count = be_transform_node(count);
ir_node *new_op = NULL;
ir_graph *irg = current_ir_graph;
dbg_info *dbgi = get_irn_dbg_info(node);
- ir_node *noreg = ia32_new_NoReg_gp(env_cg);
- ir_node *nomem = new_NoMem();
- ir_node *imm_op;
- tarval *tv;
-
- assert(! mode_is_float(get_irn_mode(node)) && "Shift/Rotate with float not supported");
-
- /* Check if immediate optimization is on and */
- /* if it's an operation with immediate. */
- imm_op = (env_cg->opt & IA32_OPT_IMMOPS) ? get_immediate_op(NULL, new_count) : NULL;
-
- /* Limit imm_op within range imm8 */
- if (imm_op) {
- tv = get_ia32_Immop_tarval(imm_op);
-
- if (tv) {
- tv = tarval_mod(tv, new_tarval_from_long(32, get_tarval_mode(tv)));
- set_ia32_Immop_tarval(imm_op, tv);
- }
- else {
- imm_op = NULL;
- }
- }
-
- /* integer operations */
- if (imm_op) {
- /* This is ShiftD with const */
- DB((dbg, LEVEL_1, "ShiftD with immediate ..."));
+ ir_node *new_op1 = be_transform_node(op1);
+ ir_node *new_op2 = create_immediate_or_transform(op2, 'I');
+ ir_node *new_count = be_transform_node(count);
- if (is_ia32_l_ShlD(node))
- new_op = new_rd_ia32_ShlD(dbgi, irg, block, noreg, noreg,
- new_op1, new_op2, noreg, nomem);
- else
- new_op = new_rd_ia32_ShrD(dbgi, irg, block, noreg, noreg,
- new_op1, new_op2, noreg, nomem);
- copy_ia32_Immop_attr(new_op, imm_op);
- }
- else {
- /* This is a normal ShiftD */
- DB((dbg, LEVEL_1, "ShiftD binop ..."));
- if (is_ia32_l_ShlD(node))
- new_op = new_rd_ia32_ShlD(dbgi, irg, block, noreg, noreg,
- new_op1, new_op2, new_count, nomem);
- else
- new_op = new_rd_ia32_ShrD(dbgi, irg, block, noreg, noreg,
- new_op1, new_op2, new_count, nomem);
- }
+ /* TODO proper AM support */
- /* set AM support */
- set_ia32_am_support(new_op, ia32_am_Dest, ia32_am_binary);
+ if (is_ia32_l_ShlD(node))
+ new_op = new_rd_ia32_ShlD(dbgi, irg, block, new_op1, new_op2, new_count);
+ else
+ new_op = new_rd_ia32_ShrD(dbgi, irg, block, new_op1, new_op2, new_count);
SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, node));
- set_ia32_emit_cl(new_op);
-
return new_op;
}
set_ia32_frame_ent(res, get_ia32_frame_ent(node));
set_ia32_use_frame(res);
set_ia32_ls_mode(res, get_ia32_ls_mode(node));
- set_ia32_am_flavour(res, ia32_B);
set_ia32_op_type(res, ia32_AddrModeD);
/* Load MEM -> SSE */
- res = new_rd_ia32_xLoad(dbgi, irg, block, new_ptr, noreg, res);
+ res = new_rd_ia32_xLoad(dbgi, irg, block, new_ptr, noreg, res,
+ get_ia32_ls_mode(node));
set_ia32_frame_ent(res, get_ia32_frame_ent(node));
set_ia32_use_frame(res);
- set_ia32_ls_mode(res, get_ia32_ls_mode(node));
- set_ia32_am_flavour(res, ia32_B);
set_ia32_op_type(res, ia32_AddrModeS);
res = new_rd_Proj(dbgi, irg, block, res, mode_xmm, pn_ia32_xLoad_res);
set_ia32_frame_ent(res, fent);
set_ia32_use_frame(res);
set_ia32_ls_mode(res, lsmode);
- set_ia32_am_flavour(res, ia32_B);
set_ia32_op_type(res, ia32_AddrModeD);
mem = res;
}
set_ia32_frame_ent(res, fent);
set_ia32_use_frame(res);
add_ia32_am_offs_int(res, offs);
- set_ia32_am_flavour(res, ia32_B);
set_ia32_op_type(res, ia32_AddrModeS);
res = new_rd_Proj(dbgi, irg, block, res, mode_vfp, pn_ia32_vfld_res);
dbg_info *dbgi = get_irn_dbg_info(node);
long proj = get_Proj_proj(node);
- if (proj == pn_be_AddSP_res) {
- ir_node *res = new_rd_Proj(dbgi, irg, block, new_pred, mode_Iu, pn_ia32_AddSP_stack);
+ if (proj == pn_be_AddSP_sp) {
+ ir_node *res = new_rd_Proj(dbgi, irg, block, new_pred, mode_Iu,
+ pn_ia32_SubSP_stack);
arch_set_irn_register(env_cg->arch_env, res, &ia32_gp_regs[REG_ESP]);
return res;
+ } else if(proj == pn_be_AddSP_res) {
+ return new_rd_Proj(dbgi, irg, block, new_pred, mode_Iu,
+ pn_ia32_SubSP_addr);
} else if (proj == pn_be_AddSP_M) {
- return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_AddSP_M);
+ return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_SubSP_M);
}
assert(0);
dbg_info *dbgi = get_irn_dbg_info(node);
long proj = get_Proj_proj(node);
- if (proj == pn_be_SubSP_res) {
- ir_node *res = new_rd_Proj(dbgi, irg, block, new_pred, mode_Iu, pn_ia32_SubSP_stack);
+ if (proj == pn_be_SubSP_sp) {
+ ir_node *res = new_rd_Proj(dbgi, irg, block, new_pred, mode_Iu,
+ pn_ia32_AddSP_stack);
arch_set_irn_register(env_cg->arch_env, res, &ia32_gp_regs[REG_ESP]);
return res;
} else if (proj == pn_be_SubSP_M) {
- return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_SubSP_M);
+ return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_AddSP_M);
}
assert(0);
* Transform and renumber the Projs from a Load.
*/
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);
- ir_node *new_pred = be_transform_node(pred);
ir_graph *irg = current_ir_graph;
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
+ not used)
+ */
+ if (is_Load(pred) && proj == pn_Load_M && get_irn_n_edges(pred) > 1) {
+ ir_node *res;
+
+ assert(pn_ia32_Load_M == 1); /* convention: mem-result of Source-AM
+ nodes is 1 */
+ /* this is needed, because sometimes we have loops that are only
+ reachable through the ProjM */
+ be_enqueue_preds(node);
+ /* do it in 2 steps, to silence firm verifier */
+ res = new_rd_Proj(dbgi, irg, block, pred, mode_M, pn_Load_M);
+ set_Proj_proj(res, pn_ia32_Load_M);
+ return res;
+ }
+
/* renumber the proj */
+ new_pred = be_transform_node(pred);
if (is_ia32_Load(new_pred)) {
if (proj == pn_Load_res) {
- return new_rd_Proj(dbgi, irg, block, new_pred, mode_Iu, pn_ia32_Load_res);
+ return new_rd_Proj(dbgi, irg, block, new_pred, mode_Iu,
+ pn_ia32_Load_res);
+ } else if (proj == pn_Load_M) {
+ return new_rd_Proj(dbgi, irg, block, new_pred, mode_M,
+ pn_ia32_Load_M);
+ }
+ } else if(is_ia32_Conv_I2I(new_pred)) {
+ set_irn_mode(new_pred, mode_T);
+ if (proj == pn_Load_res) {
+ return new_rd_Proj(dbgi, irg, block, new_pred, mode_Iu, 0);
} else if (proj == pn_Load_M) {
- return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_Load_M);
+ return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, 1);
}
} else if (is_ia32_xLoad(new_pred)) {
if (proj == pn_Load_res) {
- return new_rd_Proj(dbgi, irg, block, new_pred, mode_xmm, pn_ia32_xLoad_res);
+ return new_rd_Proj(dbgi, irg, block, new_pred, mode_xmm,
+ pn_ia32_xLoad_res);
} else if (proj == pn_Load_M) {
- return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_xLoad_M);
+ return new_rd_Proj(dbgi, irg, block, new_pred, mode_M,
+ pn_ia32_xLoad_M);
}
} else if (is_ia32_vfld(new_pred)) {
if (proj == pn_Load_res) {
- return new_rd_Proj(dbgi, irg, block, new_pred, mode_vfp, pn_ia32_vfld_res);
+ return new_rd_Proj(dbgi, irg, block, new_pred, mode_vfp,
+ pn_ia32_vfld_res);
} else if (proj == pn_Load_M) {
- return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, pn_ia32_vfld_M);
+ return new_rd_Proj(dbgi, irg, block, new_pred, mode_M,
+ pn_ia32_vfld_M);
+ }
+ } else {
+ /* can happen for ProJMs when source address mode happened for the
+ node */
+
+ /* however it should not be the result proj, as that would mean the
+ load had multiple users and should not have been used for
+ SourceAM */
+ if(proj != pn_Load_M) {
+ panic("internal error: transformed node not a Load");
}
+ return new_rd_Proj(dbgi, irg, block, new_pred, mode_M, 1);
}
assert(0);
* Transform the Projs from a be_Call.
*/
static ir_node *gen_Proj_be_Call(ir_node *node) {
- ir_node *block = be_transform_node(get_nodes_block(node));
- ir_node *call = get_Proj_pred(node);
- ir_node *new_call = be_transform_node(call);
- ir_graph *irg = current_ir_graph;
- dbg_info *dbgi = get_irn_dbg_info(node);
- long proj = get_Proj_proj(node);
- ir_mode *mode = get_irn_mode(node);
+ ir_node *block = be_transform_node(get_nodes_block(node));
+ ir_node *call = get_Proj_pred(node);
+ ir_node *new_call = be_transform_node(call);
+ ir_graph *irg = current_ir_graph;
+ dbg_info *dbgi = get_irn_dbg_info(node);
+ ir_type *method_type = be_Call_get_type(call);
+ int n_res = get_method_n_ress(method_type);
+ long proj = get_Proj_proj(node);
+ ir_mode *mode = get_irn_mode(node);
ir_node *sse_load;
const arch_register_class_t *cls;
}
if (call_res_pred == NULL || be_is_Call(call_res_pred)) {
- return new_rd_Proj(dbgi, irg, block, new_call, mode_M, pn_be_Call_M_regular);
+ return new_rd_Proj(dbgi, irg, block, new_call, mode_M,
+ pn_be_Call_M_regular);
} else {
assert(is_ia32_xLoad(call_res_pred));
- return new_rd_Proj(dbgi, irg, block, call_res_pred, mode_M, pn_ia32_xLoad_M);
+ return new_rd_Proj(dbgi, irg, block, call_res_pred, mode_M,
+ pn_ia32_xLoad_M);
}
}
- if (proj == pn_be_Call_first_res && mode_is_float(mode) && USE_SSE2(env_cg)) {
+ if (USE_SSE2(env_cg) && proj >= pn_be_Call_first_res
+ && proj < (pn_be_Call_first_res + n_res) && mode_is_float(mode)
+ && USE_SSE2(env_cg)) {
ir_node *fstp;
ir_node *frame = get_irg_frame(irg);
ir_node *noreg = ia32_new_NoReg_gp(env_cg);
- ir_node *p;
+ //ir_node *p;
ir_node *call_mem = be_get_Proj_for_pn(call, pn_be_Call_M_regular);
- ir_node *keepin[1];
- const arch_register_class_t *cls;
+ ir_node *call_res;
- /* in case there is no memory output: create one to serialize the copy FPU -> SSE */
- call_mem = new_rd_Proj(dbgi, irg, block, new_call, mode_M, pn_be_Call_M_regular);
+ /* in case there is no memory output: create one to serialize the copy
+ FPU -> SSE */
+ call_mem = new_rd_Proj(dbgi, irg, block, new_call, mode_M,
+ pn_be_Call_M_regular);
+ call_res = new_rd_Proj(dbgi, irg, block, new_call, mode,
+ pn_be_Call_first_res);
/* store st(0) onto stack */
- fstp = new_rd_ia32_GetST0(dbgi, irg, block, frame, noreg, call_mem);
-
- set_ia32_ls_mode(fstp, mode);
+ fstp = new_rd_ia32_vfst(dbgi, irg, block, frame, noreg, call_res,
+ call_mem, mode);
set_ia32_op_type(fstp, ia32_AddrModeD);
set_ia32_use_frame(fstp);
- set_ia32_am_flavour(fstp, ia32_am_B);
/* load into SSE register */
- sse_load = new_rd_ia32_xLoad(dbgi, irg, block, frame, noreg, fstp);
- set_ia32_ls_mode(sse_load, mode);
+ sse_load = new_rd_ia32_xLoad(dbgi, irg, block, frame, noreg, fstp,
+ mode);
set_ia32_op_type(sse_load, ia32_AddrModeS);
set_ia32_use_frame(sse_load);
- set_ia32_am_flavour(sse_load, ia32_am_B);
-
- sse_load = new_rd_Proj(dbgi, irg, block, sse_load, mode_xmm, pn_ia32_xLoad_res);
-
- /* now: create new Keep whith all former ins and one additional in - the result Proj */
-
- /* get a Proj representing a caller save register */
- p = be_get_Proj_for_pn(call, pn_be_Call_first_res + 1);
- assert(is_Proj(p) && "Proj expected.");
- /* user of the the proj is the Keep */
- p = get_edge_src_irn(get_irn_out_edge_first(p));
- assert(be_is_Keep(p) && "Keep expected.");
-
- /* keep the result */
- cls = arch_get_irn_reg_class(env_cg->arch_env, sse_load, -1);
- keepin[0] = sse_load;
- be_new_Keep(cls, irg, block, 1, keepin);
+ sse_load = new_rd_Proj(dbgi, irg, block, sse_load, mode_xmm,
+ pn_ia32_xLoad_res);
return sse_load;
}
/* transform call modes */
if (mode_is_data(mode)) {
- cls = arch_get_irn_reg_class(env_cg->arch_env, node, -1);
+ cls = arch_get_irn_reg_class(env_cg->arch_env, node, -1);
mode = cls->mode;
}
/* normally Cmps are processed when looking at Cond nodes, but this case
* can happen in complicated Psi conditions */
- ir_graph *irg = current_ir_graph;
- dbg_info *dbgi = get_irn_dbg_info(node);
- ir_node *block = be_transform_node(get_nodes_block(node));
- ir_node *cmp = get_Proj_pred(node);
- long pnc = get_Proj_proj(node);
- ir_node *cmp_left = get_Cmp_left(cmp);
- ir_node *cmp_right = get_Cmp_right(cmp);
- ir_node *new_cmp_left;
- ir_node *new_cmp_right;
- ir_node *noreg = ia32_new_NoReg_gp(env_cg);
- ir_node *nomem = new_rd_NoMem(irg);
- ir_mode *cmp_mode = get_irn_mode(cmp_left);
- ir_node *new_op;
+ ir_node *cmp = get_Proj_pred(node);
+ long pnc = get_Proj_proj(node);
+ ir_node *cmp_left = get_Cmp_left(cmp);
+ ir_node *cmp_right = get_Cmp_right(cmp);
+ ir_mode *cmp_mode = get_irn_mode(cmp_left);
+ dbg_info *dbgi = get_irn_dbg_info(cmp);
+ ir_node *block = get_nodes_block(node);
+ ir_node *res;
assert(!mode_is_float(cmp_mode));
- /* (a != b) -> (a ^ b) */
- if(pnc == pn_Cmp_Lg) {
- if(is_Const_0(cmp_left)) {
- new_op = be_transform_node(cmp_right);
- } else if(is_Const_0(cmp_right)) {
- new_op = be_transform_node(cmp_left);
- } else {
- new_op = gen_binop(cmp, cmp_left, cmp_right, new_rd_ia32_Xor, 1);
- }
-
- return new_op;
- }
- /* TODO:
- * (a == b) -> !(a ^ b)
- * (a < 0) -> (a & 0x80000000) oder a >> 31
- * (a >= 0) -> (a >> 31) ^ 1
- */
-
if(!mode_is_signed(cmp_mode)) {
pnc |= ia32_pn_Cmp_Unsigned;
}
- new_cmp_left = be_transform_node(cmp_left);
- new_cmp_right = create_immediate_or_transform(cmp_right, 0);
-
- new_op = new_rd_ia32_CmpSet(dbgi, irg, block, noreg, noreg, new_cmp_left,
- new_cmp_right, nomem, pnc);
- SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env_cg, cmp));
+ res = create_set(pnc, cmp_left, cmp_right, dbgi, block);
+ SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, cmp));
- return new_op;
+ return res;
}
/**
ir_node *pred = get_Proj_pred(node);
long proj = get_Proj_proj(node);
- if (is_Store(pred) || be_is_FrameStore(pred)) {
+ if (is_Store(pred)) {
if (proj == pn_Store_M) {
return be_transform_node(pred);
} else {
assert(0);
return new_r_Bad(irg);
}
- } else if (is_Load(pred) || be_is_FrameLoad(pred)) {
+ } else if (is_Load(pred)) {
return gen_Proj_Load(node);
} else if (is_Div(pred) || is_Mod(pred) || is_DivMod(pred)) {
return gen_Proj_DivMod(node);
if (node == be_get_old_anchor(anchor_tls)) {
return gen_Proj_tls(node);
}
+#ifdef FIRM_EXT_GRS
+ } else if(!is_ia32_irn(pred)) { // Quick hack for SIMD optimization
+#else
} else {
+#endif
ir_node *new_pred = be_transform_node(pred);
ir_node *block = be_transform_node(get_nodes_block(node));
ir_mode *mode = get_irn_mode(node);
/**
* Enters all transform functions into the generic pointer
*/
-static void register_transformers(void) {
- ir_op *op_Max, *op_Min, *op_Mulh;
+static void register_transformers(void)
+{
+ ir_op *op_Mulh;
/* first clear the generic function pointer for all ops */
clear_irp_opcodes_generic_func();
GEN(ASM);
GEN(CopyB);
- //GEN(Mux);
BAD(Mux);
GEN(Psi);
GEN(Proj);
GEN(Phi);
+ GEN(IJmp);
/* transform ops from intrinsic lowering */
GEN(ia32_l_Add);
GEN(ia32_l_Sbb);
GEN(ia32_l_Neg);
GEN(ia32_l_Mul);
- GEN(ia32_l_Xor);
GEN(ia32_l_IMul);
- GEN(ia32_l_Shl);
- GEN(ia32_l_Shr);
+ GEN(ia32_l_Xor);
+ GEN(ia32_l_ShlDep);
+ GEN(ia32_l_ShrDep);
GEN(ia32_l_Sar);
+ GEN(ia32_l_SarDep);
GEN(ia32_l_ShlD);
GEN(ia32_l_ShrD);
GEN(ia32_l_vfdiv);
GEN(ia32_l_vfsub);
GEN(ia32_l_vfild);
GEN(ia32_l_Load);
- /* GEN(ia32_l_vfist); TODO */
+ GEN(ia32_l_vfist);
GEN(ia32_l_Store);
GEN(ia32_l_X87toSSE);
GEN(ia32_l_SSEtoX87);
GEN(Const);
GEN(SymConst);
+ GEN(Unknown);
/* we should never see these nodes */
BAD(Raise);
GEN(be_FrameAddr);
//GEN(be_Call);
GEN(be_Return);
- GEN(be_FrameLoad);
- GEN(be_FrameStore);
- GEN(be_StackParam);
GEN(be_AddSP);
GEN(be_SubSP);
GEN(be_Copy);
- /* set the register for all Unknown nodes */
- GEN(Unknown);
-
- op_Max = get_op_Max();
- if (op_Max)
- GEN(Max);
- op_Min = get_op_Min();
- if (op_Min)
- GEN(Min);
op_Mulh = get_op_Mulh();
if (op_Mulh)
GEN(Mulh);
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);
+ get_fpcw();
+}
+
+/**
+ * Walker, checks if all ia32 nodes producing more than one result have
+ * its Projs, other wise creates new projs and keep 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 = get_ia32_n_res(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 = get_Proj_proj(proj);
+
+ assert(get_irn_mode(proj) == mode_M || 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 *class;
+
+ if(found_projs & (1 << i)) {
+ continue;
+ }
+
+ req = get_ia32_out_req(node, i);
+ class = req->cls;
+ if(class == NULL) {
+ continue;
+ }
+
+ block = get_nodes_block(node);
+ in[0] = new_r_Proj(current_ir_graph, block, node,
+ arch_register_class_mode(class), i);
+ if(last_keep != NULL) {
+ be_Keep_add_node(last_keep, class, in[0]);
+ } else {
+ last_keep = be_new_Keep(class, current_ir_graph, block, 1, in);
+ }
+ }
+}
+
+/**
+ * Adds missing keeps to nodes. Adds missing Proj nodes for unused outputs
+ * and keeps them.
+ */
+static
+void add_missing_keeps(ia32_code_gen_t *cg)
+{
+ ir_graph *irg = be_get_birg_irg(cg->birg);
+ irg_walk_graph(irg, add_missing_keep_walker, NULL, NULL);
}
/* do the transformation */
void ia32_transform_graph(ia32_code_gen_t *cg) {
register_transformers();
- env_cg = cg;
+ env_cg = cg;
+ initial_fpcw = NULL;
+
+ heights = heights_new(cg->irg);
+
be_transform_graph(cg->birg, ia32_pretransform_node, cg);
+
+ heights_free(heights);
+ heights = NULL;
+ add_missing_keeps(cg);
}
void ia32_init_transform(void)