block = get_nodes_block(node);
- if (get_Block_n_cfgpreds(block) == 1) {
- ir_node *pred = get_Block_cfgpred(block, 0);
-
- if (is_Jmp(pred)) {
- /* The block of the return has only one predecessor,
- which jumps directly to this block.
- This jump will be encoded as a fall through, so we
- ignore it here.
- However, the predecessor might be empty, so it must be
- ensured that empty blocks are gone away ... */
- return;
- }
- }
-
/* check if this return is the first on the block */
sched_foreach_reverse_from(node, irn) {
switch (get_irn_opcode(irn)) {
return;
}
}
- /* yep, return is the first real instruction in this block */
-#if 0
- {
- /* add an rep prefix to the return */
- ir_node *rep = new_rd_ia32_RepPrefix(get_irn_dbg_info(node), current_ir_graph, block);
- keep_alive(rep);
- sched_add_before(node, rep);
- }
-#else
- /* ensure, that the 3 byte return is generated */
+
+ /* ensure, that the 3 byte return is generated
+ * actually the emitter tests again if the block beginning has a label and
+ * isn't just a fallthrough */
be_Return_set_emit_pop(node, 1);
-#endif
}
/* only optimize up to 48 stores behind IncSPs */
be_set_IncSP_pred(irn, curr_sp);
}
+/**
+ * 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;
+ if (get_mode_size_bits(mode) > 32)
+ return 0;
+ return mode_is_int(mode) || mode_is_reference(mode) || mode == mode_b;
+}
+
/**
* Tries to create Pops from Load, IncSP combinations.
* The Loads are replaced by Pops, the IncSP is modified
static void peephole_Load_IncSP_to_pop(ir_node *irn)
{
const arch_register_t *esp = &ia32_gp_regs[REG_ESP];
- int i, maxslot, inc_ofs;
+ int i, maxslot, inc_ofs, ofs;
ir_node *node, *pred_sp, *block;
ir_node *loads[MAXPUSH_OPTIMIZE];
ir_graph *irg;
unsigned regmask = 0;
+ unsigned copymask = ~0;
memset(loads, 0, sizeof(loads));
assert(be_is_IncSP(irn));
ir_node *mem;
int offset;
int loadslot;
- const arch_register_t *dreg;
+ const arch_register_t *sreg, *dreg;
/* it has to be a Load */
if (!is_ia32_Load(node)) {
if (be_is_Copy(node)) {
- if (get_irn_mode(node) != mode_Iu) {
+ if (!mode_needs_gp_reg(get_irn_mode(node))) {
/* not a GP copy, ignore */
continue;
}
dreg = arch_get_irn_register(arch_env, node);
- if (regmask & (1 << dreg->index)) {
+ sreg = arch_get_irn_register(arch_env, be_get_Copy_op(node));
+ if (regmask & copymask & (1 << sreg->index)) {
break;
}
- /* we CAN skip Copies if the destination is not in our regmask, ie
- none of our future Pop will overwrite it */
- regmask |= (1 << dreg->index);
+ if (regmask & copymask & (1 << dreg->index)) {
+ break;
+ }
+ /* we CAN skip Copies if neither the destination nor the source
+ * is not in our regmask, ie none of our future Pop will overwrite it */
+ regmask |= (1 << dreg->index) | (1 << sreg->index);
+ copymask &= ~((1 << dreg->index) | (1 << sreg->index));
continue;
}
break;
}
/* we can handle only GP loads */
- if (get_ia32_ls_mode(node) != mode_Iu)
+ if (!mode_needs_gp_reg(get_ia32_ls_mode(node)))
continue;
/* it has to use our predecessor sp value */
- if (get_irn_n(node, n_ia32_base) != pred_sp)
- continue;
+ if (get_irn_n(node, n_ia32_base) != pred_sp) {
+ /* it would be ok if this load does not use a Pop result,
+ * but we do not check this */
+ break;
+ }
/* Load has to be attached to Spill-Mem */
mem = skip_Proj(get_irn_n(node, n_ia32_mem));
if (!is_Phi(mem) && !is_ia32_Store(mem) && !is_ia32_Push(mem))
- continue;
+ break;
/* should have NO index */
if (get_ia32_am_scale(node) > 0 || !is_ia32_NoReg_GP(get_irn_n(node, n_ia32_index)))
if (offset < 0 || offset >= MAXPUSH_OPTIMIZE * 4)
continue;
+ /* ignore those outside the possible windows */
+ if (offset > inc_ofs - 4)
+ continue;
loadslot = offset >> 2;
/* loading from the same slot twice is bad (and shouldn't happen...) */
if (maxslot < 0)
return;
- /* walk through the Loads and create Pops for them */
+ /* find the first slot */
for (i = maxslot; i >= 0; --i) {
ir_node *load = loads[i];
if (load == NULL)
break;
- inc_ofs -= 4;
}
+ ofs = inc_ofs - (maxslot + 1) * 4;
+ inc_ofs = (i+1) * 4;
+
/* create a new IncSP if needed */
block = get_nodes_block(irn);
irg = cg->irg;
- if (inc_ofs != 0) {
- assert(inc_ofs > 0);
+ if (inc_ofs > 0) {
pred_sp = be_new_IncSP(esp, irg, block, pred_sp, -inc_ofs, be_get_IncSP_align(irn));
sched_add_before(irn, pred_sp);
}
+ /* walk through the Loads and create Pops for them */
for (++i; i <= maxslot; ++i) {
ir_node *load = loads[i];
ir_node *mem, *pop;
/* we can remove the Load now */
sched_remove(load);
kill_node(load);
-
}
- be_set_IncSP_offset(irn, 0);
+ be_set_IncSP_offset(irn, -ofs);
be_set_IncSP_pred(irn, pred_sp);
}