return new_rd_Proj(dbg, irg, block, conv, tgt_mode, pn_ia32_Conv_I2FP_res);
}
+/**
+* SSE convert of an float node into a double node.
+*/
+static ir_node *gen_sse_conv_f2d(ia32_code_gen_t *cg, dbg_info *dbg, ir_graph *irg, ir_node *block,
+ ir_node *in, ir_node *old_node)
+{
+ ir_node *noreg = ia32_new_NoReg_gp(cg);
+ ir_node *nomem = new_rd_NoMem(irg);
+
+ ir_node *conv = new_rd_ia32_Conv_FP2FP(dbg, irg, block, noreg, noreg, in, nomem);
+ set_ia32_src_mode(conv, mode_F);
+ set_ia32_tgt_mode(conv, mode_D);
+ set_ia32_am_support(conv, ia32_am_Source);
+ SET_IA32_ORIG_NODE(conv, ia32_get_old_node_name(cg, old_node));
+
+ return new_rd_Proj(dbg, irg, block, conv, mode_D, pn_ia32_Conv_FP2FP_res);
+}
+
/* Generates an entity for a known FP const (used for FP Neg + Abs) */
static ident *gen_fp_known_const(ir_mode *mode, ia32_known_const_t kct) {
static const struct {
ir_node *ptr = get_Store_ptr(node);
ir_node *sptr = ptr;
ir_node *mem = get_Store_mem(node);
- ir_mode *mode = get_irn_link(node);
+ ir_mode *mode = get_irn_mode(val);
ir_node *sval = val;
int is_imm = 0;
ir_node *new_op;
int rem;
ir_node *in[3], *tmp;
- /* If we have to copy more than 16 bytes, we use REP MOVSx and */
+ /* If we have to copy more than 32 bytes, we use REP MOVSx and */
/* then we need the size explicitly in ECX. */
- if (size >= 16 * 4) {
+ if (size >= 32 * 4) {
rem = size & 0x3; /* size % 4 */
size >>= 2;
int src_bits = get_mode_size_bits(src_mode);
int tgt_bits = get_mode_size_bits(tgt_mode);
int pn = -1;
+ int kill = 0;
ir_node *block = env->block;
ir_node *new_op = NULL;
ir_node *noreg = ia32_new_NoReg_gp(env->cg);
}
else {
DB((mod, LEVEL_1, "killed Conv(float, float) ..."));
- edges_reroute(env->irn, op, irg);
+ /*
+ remark: we create a intermediate conv here, so modes will be spread correctly
+ these convs will be killed later
+ */
+ new_op = new_rd_ia32_Conv_FP2FP(dbg, irg, block, noreg, noreg, op, nomem);
+ pn = pn_ia32_Conv_FP2FP_res;
+ kill = 1;
}
}
else {
/* ... to int */
if (get_mode_size_bits(src_mode) == tgt_bits) {
DB((mod, LEVEL_1, "omitting equal size Conv(%+F, %+F) ...", src_mode, tgt_mode));
- edges_reroute(env->irn, op, irg);
+ /*
+ remark: we create a intermediate conv here, so modes will be spread correctly
+ these convs will be killed later
+ */
+ new_op = new_rd_ia32_Conv_I2I(dbg, irg, block, noreg, noreg, op, nomem);
+ pn = pn_ia32_Conv_I2I_res;
+ kill = 1;
}
else {
DB((mod, LEVEL_1, "create Conv(int, int) ...", src_mode, tgt_mode));
set_ia32_am_support(new_op, ia32_am_Source);
new_op = new_rd_Proj(dbg, irg, block, new_op, tgt_mode, pn);
+
+ if (kill)
+ nodeset_insert(env->cg->kill_conv, new_op);
}
return new_op;
ir_node *noreg = ia32_new_NoReg_gp(env->cg);
ir_node *mem = new_rd_NoMem(env->irg);
ir_node *ptr = get_irn_n(node, 0);
- entity *ent = be_get_frame_entity(node);
+ entity *ent = arch_get_frame_entity(env->cg->arch_env, node);
ir_mode *mode = env->mode;
/* choose the block where to place the load */
set_ia32_op_type(new_op, ia32_AddrModeS);
set_ia32_am_flavour(new_op, ia32_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, env->irn));
ir_node *nomem = new_rd_NoMem(env->irg);
new_op = new_rd_ia32_Add(env->dbg, env->irg, env->block, noreg, noreg, op, noreg, nomem);
- set_ia32_frame_ent(new_op, be_get_frame_entity(node));
+ set_ia32_frame_ent(new_op, arch_get_frame_entity(env->cg->arch_env, node));
set_ia32_am_support(new_op, ia32_am_Full);
set_ia32_use_frame(new_op);
set_ia32_immop_type(new_op, ia32_ImmConst);
ir_node *noreg = ia32_new_NoReg_gp(env->cg);
ir_node *mem = get_irn_n(node, 0);
ir_node *ptr = get_irn_n(node, 1);
- entity *ent = be_get_frame_entity(node);
+ entity *ent = arch_get_frame_entity(env->cg->arch_env, node);
ir_mode *mode = get_type_mode(get_entity_type(ent));
if (mode_is_float(mode)) {
ir_node *mem = get_irn_n(node, 0);
ir_node *ptr = get_irn_n(node, 1);
ir_node *val = get_irn_n(node, 2);
- entity *ent = be_get_frame_entity(node);
+ entity *ent = arch_get_frame_entity(env->cg->arch_env, node);
ir_mode *mode = get_irn_mode(val);
if (mode_is_float(mode)) {
mode = get_irn_mode(call_res);
+ /* in case there is no memory output: create one to serialize the copy FPU -> SSE */
if (! call_mem)
- call_mem = get_irg_no_mem(env->irg);
+ call_mem = new_r_Proj(env->irg, env->block, env->irn, mode_M, pn_be_Call_M_regular);
if (mode_is_float(mode)) {
/* store st(0) onto stack */
return NULL;
}
+/**
+ * Transform a be_AddSP into an ia32_AddSP. Eat up const sizes.
+ */
+static ir_node *gen_be_AddSP(ia32_transform_env_t *env) {
+ ir_node *new_op;
+ const ir_edge_t *edge;
+ ir_node *sz = get_irn_n(env->irn, be_pos_AddSP_size);
+ ir_node *sp = get_irn_n(env->irn, be_pos_AddSP_old_sp);
+
+ new_op = new_rd_ia32_AddSP(env->dbg, env->irg, env->block, sp, sz);
+
+ if (is_ia32_Const(sz)) {
+ set_ia32_Immop_attr(new_op, sz);
+ set_irn_n(new_op, 1, ia32_new_NoReg_gp(env->cg));
+ }
+ else if (is_ia32_Load(sz) && get_ia32_am_flavour(sz) == ia32_O) {
+ set_ia32_immop_type(new_op, ia32_ImmSymConst);
+ set_ia32_op_type(new_op, ia32_AddrModeS);
+ set_ia32_am_sc(new_op, get_ia32_am_sc(sz));
+ add_ia32_am_offs(new_op, get_ia32_am_offs(sz));
+ set_irn_n(new_op, 1, ia32_new_NoReg_gp(env->cg));
+ }
+
+ /* fix proj nums */
+ foreach_out_edge(env->irn, edge) {
+ ir_node *proj = get_edge_src_irn(edge);
+
+ assert(is_Proj(proj));
+
+ if (get_Proj_proj(proj) == pn_be_AddSP_res) {
+ /* the node is not yet exchanged: we need to set the register manually */
+ ia32_attr_t *attr = get_ia32_attr(new_op);
+ attr->slots[pn_ia32_AddSP_stack] = &ia32_gp_regs[REG_ESP];
+ set_Proj_proj(proj, pn_ia32_AddSP_stack);
+ }
+ else if (get_Proj_proj(proj) == pn_be_AddSP_M) {
+ set_Proj_proj(proj, pn_ia32_AddSP_M);
+ }
+ else {
+ assert(0);
+ }
+ }
+
+ SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(env->cg, env->irn));
+
+ return new_op;
+}
/**
* This function just sets the register for the Unknown node
GEN(be_FrameLoad);
GEN(be_FrameStore);
GEN(be_StackParam);
+ GEN(be_AddSP);
/* set the register for all Unknown nodes */
GEN(Unknown);
ia32_transform_env_t tenv;
transform_func *transform = (transform_func *)op->ops.generic;
- tenv.block = get_nodes_block(node);
- tenv.dbg = get_irn_dbg_info(node);
- tenv.irg = current_ir_graph;
- tenv.irn = node;
- tenv.mode = get_irn_mode(node);
- tenv.cg = cg;
+ tenv.block = get_nodes_block(node);
+ tenv.dbg = get_irn_dbg_info(node);
+ tenv.irg = current_ir_graph;
+ tenv.irn = node;
+ tenv.mode = get_irn_mode(node);
+ tenv.cg = cg;
DEBUG_ONLY(tenv.mod = cg->mod;)
asm_node = (*transform)(&tenv);
/* Psi is float, we need a floating point compare */
if (USE_SSE2(cg)) {
+ ir_mode *m = get_irn_mode(cmp_a);
/* SSE FPU */
- if (! mode_is_float(get_irn_mode(cmp_a))) {
+ if (! mode_is_float(m)) {
cmp_a = gen_sse_conv_int2float(cg, dbg, irg, block, cmp_a, cmp_a, mode);
cmp_b = gen_sse_conv_int2float(cg, dbg, irg, block, cmp_b, cmp_b, mode);
- pnc |= 8;
+ }
+ else if (m == mode_F) {
+ /* we convert cmp values always to double, to get correct bitmask with cmpsd */
+ cmp_a = gen_sse_conv_f2d(cg, dbg, irg, block, cmp_a, cmp_a);
+ cmp_b = gen_sse_conv_f2d(cg, dbg, irg, block, cmp_b, cmp_b);
}
new_op = new_rd_ia32_xCmp(dbg, irg, block, noreg, noreg, cmp_a, cmp_b, nomem);