static const arch_env_t *arch_env;
static ia32_code_gen_t *cg;
-static void peephole_IncSP_IncSP(ir_node *node);
-
-#if 0
-static void peephole_ia32_Store_IncSP_to_push(ir_node *node)
-{
- ir_node *base = get_irn_n(node, n_ia32_Store_base);
- ir_node *index = get_irn_n(node, n_ia32_Store_index);
- ir_node *mem = get_irn_n(node, n_ia32_Store_mem);
- ir_node *incsp = base;
- ir_node *val;
- ir_node *noreg;
- ir_graph *irg;
- ir_node *block;
- dbg_info *dbgi;
- ir_mode *mode;
- ir_node *push;
- ir_node *proj;
- int offset;
- int node_offset;
-
- /* nomem inidicates the store doesn't alias with anything else */
- if(!is_NoMem(mem))
- return;
-
- /* find an IncSP in front of us, we might have to skip barriers for this */
- while(is_Proj(incsp)) {
- ir_node *proj_pred = get_Proj_pred(incsp);
- if(!be_is_Barrier(proj_pred))
- return;
- incsp = get_irn_n(proj_pred, get_Proj_proj(incsp));
- }
- if(!be_is_IncSP(incsp))
- return;
-
- peephole_IncSP_IncSP(incsp);
-
- /* must be in the same block */
- if(get_nodes_block(incsp) != get_nodes_block(node))
- return;
-
- if(!is_ia32_NoReg_GP(index) || get_ia32_am_sc(node) != NULL) {
- panic("Invalid storeAM found (%+F)", node);
- }
-
- /* we should be the store to the end of the stackspace */
- offset = be_get_IncSP_offset(incsp);
- mode = get_ia32_ls_mode(node);
- node_offset = get_ia32_am_offs_int(node);
- if(node_offset != offset - get_mode_size_bytes(mode))
- return;
-
- /* we can use a push instead of the store */
- irg = current_ir_graph;
- block = get_nodes_block(node);
- dbgi = get_irn_dbg_info(node);
- noreg = ia32_new_NoReg_gp(cg);
- base = be_get_IncSP_pred(incsp);
- val = get_irn_n(node, n_ia32_Store_val);
- push = new_rd_ia32_Push(dbgi, irg, block, noreg, noreg, mem, val, base);
-
- proj = new_r_Proj(irg, block, push, mode_M, pn_ia32_Push_M);
-
- be_set_IncSP_offset(incsp, offset - get_mode_size_bytes(mode));
-
- sched_add_before(node, push);
- sched_remove(node);
-
- be_peephole_before_exchange(node, proj);
- exchange(node, proj);
- be_peephole_after_exchange(proj);
-}
-
-static void peephole_ia32_Store(ir_node *node)
-{
- peephole_ia32_Store_IncSP_to_push(node);
-}
-#endif
-
+/**
+ * Returns non-zero if the given node produces
+ * a zero flag.
+ *
+ * @param node the node to check
+ * @param pn if >= 0, the projection number of the used result
+ */
static int produces_zero_flag(ir_node *node, int pn)
{
ir_node *count;
const ia32_immediate_attr_t *imm_attr;
- if(!is_ia32_irn(node))
+ if (!is_ia32_irn(node))
return 0;
- if(pn >= 0) {
- if(pn != pn_ia32_res)
+ if (pn >= 0) {
+ if (pn != pn_ia32_res)
return 0;
}
- switch(get_ia32_irn_opcode(node)) {
+ switch (get_ia32_irn_opcode(node)) {
case iro_ia32_Add:
case iro_ia32_Adc:
case iro_ia32_And:
assert(n_ia32_ShlD_count == n_ia32_ShrD_count);
assert(n_ia32_Shl_count == n_ia32_Shr_count
&& n_ia32_Shl_count == n_ia32_Sar_count);
- if(is_ia32_ShlD(node) || is_ia32_ShrD(node)) {
+ if (is_ia32_ShlD(node) || is_ia32_ShrD(node)) {
count = get_irn_n(node, n_ia32_ShlD_count);
} else {
count = get_irn_n(node, n_ia32_Shl_count);
}
/* when shift count is zero the flags are not affected, so we can only
* do this for constants != 0 */
- if(!is_ia32_Immediate(count))
+ if (!is_ia32_Immediate(count))
return 0;
imm_attr = get_ia32_immediate_attr_const(count);
- if(imm_attr->symconst != NULL)
+ if (imm_attr->symconst != NULL)
return 0;
- if((imm_attr->offset & 0x1f) == 0)
+ if ((imm_attr->offset & 0x1f) == 0)
return 0;
return 1;
return 0;
}
+/**
+ * If the given node has not mode_T, creates a mode_T version (with a result Proj).
+ *
+ * @param node the node to change
+ *
+ * @return the new mode_T node (if the mode was changed) or node itself
+ */
static ir_node *turn_into_mode_t(ir_node *node)
{
ir_node *block;
return new_node;
}
+/**
+ * Peephole optimization for Test instructions.
+ * We can remove the Test, if a zero flags was produced which is still
+ * live.
+ */
static void peephole_ia32_Test(ir_node *node)
{
- ir_node *left = get_irn_n(node, n_ia32_Test_left);
- ir_node *right = get_irn_n(node, n_ia32_Test_right);
- ir_node *flags_proj;
- ir_node *block;
- ir_mode *flags_mode;
- int pn = -1;
- ir_node *schedpoint;
- const ir_edge_t *edge;
+ ir_node *left = get_irn_n(node, n_ia32_Test_left);
+ ir_node *right = get_irn_n(node, n_ia32_Test_right);
+ ir_node *flags_proj;
+ ir_node *block;
+ ir_mode *flags_mode;
+ int pn = -1;
+ ir_node *schedpoint;
+ const ir_edge_t *edge;
assert(n_ia32_Test_left == n_ia32_Test8Bit_left
&& n_ia32_Test_right == n_ia32_Test8Bit_right);
* Can be avoided by placing a Rep prefix before the return.
*/
static void peephole_ia32_Return(ir_node *node) {
- ir_node *block, *irn, *rep;
+ ir_node *block, *irn;
if (!ia32_cg_config.use_pad_return)
return;
/* check if this return is the first on the block */
sched_foreach_reverse_from(node, irn) {
- switch (be_get_irn_opcode(irn)) {
+ switch (get_irn_opcode(irn)) {
case beo_Return:
/* the return node itself, ignore */
continue;
if (be_get_IncSP_offset(irn) == 0)
continue;
return;
+ case iro_Phi:
+ continue;
default:
- if (is_Phi(irn))
- continue;
return;
}
}
/* yep, return is the first real instruction in this block */
#if 0
- /* add an rep prefix to the return */
- rep = new_rd_ia32_RepPrefix(get_irn_dbg_info(node), current_ir_graph, block);
- keep_alive(rep);
- sched_add_before(node, rep);
+ {
+ /* 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 */
be_Return_set_emit_pop(node, 1);
#define MAXPUSH_OPTIMIZE 48
/**
- * Tries to create pushs from IncSP,Store combinations.
+ * Tries to create Push's from IncSP, Store combinations.
* The Stores are replaced by Push's, the IncSP is modified
* (possibly into IncSP 0, but not removed).
*/
static void peephole_IncSP_Store_to_push(ir_node *irn)
{
- int i;
- int offset;
+ int i, maxslot, inc_ofs;
ir_node *node;
ir_node *stores[MAXPUSH_OPTIMIZE];
ir_node *block = get_nodes_block(irn);
assert(be_is_IncSP(irn));
- offset = be_get_IncSP_offset(irn);
- if (offset < 4)
+ inc_ofs = be_get_IncSP_offset(irn);
+ if (inc_ofs < 4)
return;
/*
* We save them into the stores array which is sorted by the frame offset/4
* attached to the node
*/
- for(node = sched_next(irn); !sched_is_end(node); node = sched_next(node)) {
+ maxslot = -1;
+ for (node = sched_next(irn); !sched_is_end(node); node = sched_next(node)) {
ir_node *mem;
int offset;
int storeslot;
- // it has to be a store
- if(!is_ia32_Store(node))
+ /* it has to be a Store */
+ if (!is_ia32_Store(node))
break;
- // it has to use our sp value
- if(get_irn_n(node, n_ia32_base) != irn)
+ /* it has to use our sp value */
+ if (get_irn_n(node, n_ia32_base) != irn)
continue;
- // store has to be attached to NoMem
+ /* Store has to be attached to NoMem */
mem = get_irn_n(node, n_ia32_mem);
- if(!is_NoMem(mem)) {
+ if (!is_NoMem(mem))
continue;
- }
/* unfortunately we can't support the full AMs possible for push at the
* moment. TODO: fix this */
- if(get_ia32_am_scale(node) > 0 || !is_ia32_NoReg_GP(get_irn_n(node, n_ia32_index)))
+ if (get_ia32_am_scale(node) > 0 || !is_ia32_NoReg_GP(get_irn_n(node, n_ia32_index)))
break;
offset = get_ia32_am_offs_int(node);
+ /* we should NEVER access uninitialized stack BELOW the current SP */
+ assert(offset >= 0);
- storeslot = offset / 4;
- if(storeslot >= MAXPUSH_OPTIMIZE)
- continue;
+ offset = inc_ofs - 4 - offset;
- // storing into the same slot twice is bad (and shouldn't happen...)
- if(stores[storeslot] != NULL)
+ /* storing at half-slots is bad */
+ if ((offset & 3) != 0)
break;
- // storing at half-slots is bad
- if(offset % 4 != 0)
+ if (offset < 0 || offset >= MAXPUSH_OPTIMIZE * 4)
+ continue;
+ storeslot = offset >> 2;
+
+ /* storing into the same slot twice is bad (and shouldn't happen...) */
+ if (stores[storeslot] != NULL)
break;
stores[storeslot] = node;
+ if (storeslot > maxslot)
+ maxslot = storeslot;
}
curr_sp = be_get_IncSP_pred(irn);
- // walk the stores in inverse order and create pushs for them
- i = (offset / 4) - 1;
- if(i >= MAXPUSH_OPTIMIZE) {
- i = MAXPUSH_OPTIMIZE - 1;
- }
-
- for( ; i >= 0; --i) {
+ /* walk through the stores and create Pushs for them */
+ for (i = 0; i <= maxslot; ++i) {
const arch_register_t *spreg;
ir_node *push;
ir_node *val, *mem, *mem_proj;
ir_node *store = stores[i];
ir_node *noreg = ia32_new_NoReg_gp(cg);
- if(store == NULL || is_Bad(store))
+ if (store == NULL)
break;
val = get_irn_n(store, n_ia32_unary_op);
sched_add_before(irn, push);
- // create stackpointer proj
+ /* create stackpointer Proj */
curr_sp = new_r_Proj(irg, block, push, spmode, pn_ia32_Push_stack);
arch_set_irn_register(cg->arch_env, curr_sp, spreg);
- // create memory proj
+ /* create memory Proj */
mem_proj = new_r_Proj(irg, block, push, mode_M, pn_ia32_Push_M);
- // use the memproj now
+ /* use the memproj now */
exchange(store, mem_proj);
- // we can remove the store now
+ /* we can remove the store now */
sched_remove(store);
- offset -= 4;
+ inc_ofs -= 4;
}
- be_set_IncSP_offset(irn, offset);
+ be_set_IncSP_offset(irn, inc_ofs);
be_set_IncSP_pred(irn, curr_sp);
}
/**
- * Tries to optimize two following IncSP.
+ * Find a free GP register if possible, else return NULL.
*/
-static void peephole_IncSP_IncSP(ir_node *node)
+static const arch_register_t *get_free_gp_reg(void)
{
- int pred_offs;
- int curr_offs;
- int offs;
- ir_node *pred = be_get_IncSP_pred(node);
- ir_node *predpred;
+ int i;
- if(!be_is_IncSP(pred))
- return;
+ for(i = 0; i < N_ia32_gp_REGS; ++i) {
+ const arch_register_t *reg = &ia32_gp_regs[i];
+ if(arch_register_type_is(reg, ignore))
+ continue;
- if(get_irn_n_edges(pred) > 1)
- return;
+ if(be_peephole_get_value(CLASS_ia32_gp, i) == NULL)
+ return &ia32_gp_regs[i];
+ }
- pred_offs = be_get_IncSP_offset(pred);
- curr_offs = be_get_IncSP_offset(node);
+ return NULL;
+}
- if(pred_offs == BE_STACK_FRAME_SIZE_EXPAND) {
- if(curr_offs != BE_STACK_FRAME_SIZE_SHRINK) {
- return;
- }
- offs = 0;
- } else if(pred_offs == BE_STACK_FRAME_SIZE_SHRINK) {
- if(curr_offs != BE_STACK_FRAME_SIZE_EXPAND) {
- return;
- }
- offs = 0;
- } else if(curr_offs == BE_STACK_FRAME_SIZE_EXPAND
- || curr_offs == BE_STACK_FRAME_SIZE_SHRINK) {
- return;
- } else {
- offs = curr_offs + pred_offs;
- }
+/**
+ * Creates a Pop instruction before the given schedule point.
+ *
+ * @param dbgi debug info
+ * @param irg the graph
+ * @param block the block
+ * @param stack the previous stack value
+ * @param schedpoint the new node is added before this node
+ * @param reg the register to pop
+ *
+ * @return the new stack value
+ */
+static ir_node *create_pop(dbg_info *dbgi, ir_graph *irg, ir_node *block,
+ ir_node *stack, ir_node *schedpoint,
+ const arch_register_t *reg)
+{
+ const arch_register_t *esp = &ia32_gp_regs[REG_ESP];
+ ir_node *pop;
+ ir_node *keep;
+ ir_node *val;
+ ir_node *in[1];
- /* add pred offset to ours and remove pred IncSP */
- be_set_IncSP_offset(node, offs);
+ pop = new_rd_ia32_Pop(dbgi, irg, block, new_NoMem(), stack);
- predpred = be_get_IncSP_pred(pred);
- be_peephole_before_exchange(pred, predpred);
+ stack = new_r_Proj(irg, block, pop, mode_Iu, pn_ia32_Pop_stack);
+ arch_set_irn_register(arch_env, stack, esp);
+ val = new_r_Proj(irg, block, pop, mode_Iu, pn_ia32_Pop_res);
+ arch_set_irn_register(arch_env, val, reg);
- /* rewire dependency edges */
- edges_reroute_kind(pred, predpred, EDGE_KIND_DEP, current_ir_graph);
- be_set_IncSP_pred(node, predpred);
- sched_remove(pred);
- be_kill_node(pred);
+ sched_add_before(schedpoint, pop);
- be_peephole_after_exchange(predpred);
+ in[0] = val;
+ keep = be_new_Keep(&ia32_reg_classes[CLASS_ia32_gp], irg, block, 1, in);
+ sched_add_before(schedpoint, keep);
+
+ return stack;
}
/**
- * Find a free GP register if possible, else return NULL.
+ * Creates a Push instruction before the given schedule point.
+ *
+ * @param dbgi debug info
+ * @param irg the graph
+ * @param block the block
+ * @param stack the previous stack value
+ * @param schedpoint the new node is added before this node
+ * @param reg the register to pop
+ *
+ * @return the new stack value
*/
-static const arch_register_t *get_free_gp_reg(void)
+static ir_node *create_push(dbg_info *dbgi, ir_graph *irg, ir_node *block,
+ ir_node *stack, ir_node *schedpoint,
+ const arch_register_t *reg)
{
- int i;
+ const arch_register_t *esp = &ia32_gp_regs[REG_ESP];
+ ir_node *noreg, *nomem, *push, *val;
- for(i = 0; i < N_ia32_gp_REGS; ++i) {
- const arch_register_t *reg = &ia32_gp_regs[i];
- if(arch_register_type_is(reg, ignore))
- continue;
+ val = new_rd_ia32_ProduceVal(NULL, irg, block);
+ arch_set_irn_register(arch_env, val, reg);
+ sched_add_before(schedpoint, val);
- if(be_peephole_get_value(CLASS_ia32_gp, i) == NULL)
- return &ia32_gp_regs[i];
- }
+ noreg = ia32_new_NoReg_gp(cg);
+ nomem = get_irg_no_mem(irg);
+ push = new_rd_ia32_Push(dbgi, irg, block, noreg, noreg, nomem, val, stack);
+ sched_add_before(schedpoint, push);
- return NULL;
+ stack = new_r_Proj(irg, block, push, mode_Iu, pn_ia32_Push_stack);
+ arch_set_irn_register(arch_env, stack, esp);
+
+ return stack;
}
+/**
+ * Optimize an IncSp by replacing it with Push/Pop.
+ */
static void peephole_be_IncSP(ir_node *node)
{
const arch_register_t *esp = &ia32_gp_regs[REG_ESP];
const arch_register_t *reg;
- ir_graph *irg;
+ ir_graph *irg = current_ir_graph;
dbg_info *dbgi;
ir_node *block;
- ir_node *keep;
- ir_node *val;
- ir_node *pop, *pop2;
ir_node *stack;
int offset;
/* first optimize incsp->incsp combinations */
- peephole_IncSP_IncSP(node);
+ be_peephole_IncSP_IncSP(node);
/* transform IncSP->Store combinations to Push where possible */
peephole_IncSP_Store_to_push(node);
/* replace IncSP -4 by Pop freereg when possible */
offset = be_get_IncSP_offset(node);
- if ((offset != -8 || !ia32_cg_config.use_add_esp_8) &&
- (offset != -4 || !ia32_cg_config.use_add_esp_4) &&
- (offset != +4 || !ia32_cg_config.use_sub_esp_4) &&
- (offset != +8 || !ia32_cg_config.use_sub_esp_8))
+ if ((offset != -8 || ia32_cg_config.use_add_esp_8) &&
+ (offset != -4 || ia32_cg_config.use_add_esp_4) &&
+ (offset != +4 || ia32_cg_config.use_sub_esp_4) &&
+ (offset != +8 || ia32_cg_config.use_sub_esp_8))
return;
if (offset < 0) {
/* we need a free register for pop */
reg = get_free_gp_reg();
- if(reg == NULL)
+ if (reg == NULL)
return;
- irg = current_ir_graph;
dbgi = get_irn_dbg_info(node);
block = get_nodes_block(node);
stack = be_get_IncSP_pred(node);
- pop = new_rd_ia32_Pop(dbgi, irg, block, new_NoMem(), stack);
-
- stack = new_r_Proj(irg, block, pop, mode_Iu, pn_ia32_Pop_stack);
- arch_set_irn_register(arch_env, stack, esp);
- val = new_r_Proj(irg, block, pop, mode_Iu, pn_ia32_Pop_res);
- arch_set_irn_register(arch_env, val, reg);
- sched_add_before(node, pop);
-
- keep = sched_next(node);
- if (!be_is_Keep(keep)) {
- ir_node *in[1];
- in[0] = val;
- keep = be_new_Keep(&ia32_reg_classes[CLASS_ia32_gp], irg, block, 1, in);
- sched_add_before(node, keep);
- } else {
- be_Keep_add_node(keep, &ia32_reg_classes[CLASS_ia32_gp], val);
- }
+ stack = create_pop(dbgi, irg, block, stack, node, reg);
if (offset == -8) {
- pop2 = new_rd_ia32_Pop(dbgi, irg, block, new_NoMem(), stack);
+ stack = create_pop(dbgi, irg, block, stack, node, reg);
+ }
+ } else {
+ dbgi = get_irn_dbg_info(node);
+ block = get_nodes_block(node);
+ stack = be_get_IncSP_pred(node);
+ reg = &ia32_gp_regs[REG_EAX];
- stack = new_r_Proj(irg, block, pop2, mode_Iu, pn_ia32_Pop_stack);
- arch_set_irn_register(arch_env, stack, esp);
- val = new_r_Proj(irg, block, pop2, mode_Iu, pn_ia32_Pop_res);
- arch_set_irn_register(arch_env, val, reg);
+ stack = create_push(dbgi, irg, block, stack, node, reg);
- sched_add_after(pop, pop2);
- be_Keep_add_node(keep, &ia32_reg_classes[CLASS_ia32_gp], val);
+ if (offset == +8) {
+ stack = create_push(dbgi, irg, block, stack, node, reg);
}
- } else {
- /* NIY */
- return;
}
be_peephole_before_exchange(node, stack);
set_ia32_op_type(imul, ia32_Normal);
}
+/**
+ * Replace xorps r,r and xorpd r,r by pxor r,r
+ */
+static void peephole_ia32_xZero(ir_node *xor) {
+ set_irn_op(xor, op_ia32_xPzero);
+}
+
/**
* Register a peephole optimisation function.
*/
/* register peephole optimisations */
clear_irp_opcodes_generic_func();
register_peephole_optimisation(op_ia32_Const, peephole_ia32_Const);
- //register_peephole_optimisation(op_ia32_Store, peephole_ia32_Store);
register_peephole_optimisation(op_be_IncSP, peephole_be_IncSP);
register_peephole_optimisation(op_ia32_Lea, peephole_ia32_Lea);
register_peephole_optimisation(op_ia32_Test, peephole_ia32_Test);
register_peephole_optimisation(op_be_Return, peephole_ia32_Return);
if (! ia32_cg_config.use_imul_mem_imm32)
register_peephole_optimisation(op_ia32_IMul, peephole_ia32_Imul_split);
+ if (ia32_cg_config.use_pxor)
+ register_peephole_optimisation(op_ia32_xZero, peephole_ia32_xZero);
be_peephole_opt(cg->birg);
}