* @file
* @brief Implements several optimizations for IA32.
* @author Matthias Braun, Christian Wuerdig
- * @version $Id$
*/
#include "config.h"
#include "irgmod.h"
#include "irgwalk.h"
#include "heights.h"
-#include "irbitset.h"
#include "irprintf.h"
#include "irdump.h"
#include "error.h"
-#include "../be_t.h"
-#include "../beabi.h"
-#include "../benode.h"
-#include "../besched.h"
-#include "../bepeephole.h"
+#include "be_t.h"
+#include "beabi.h"
+#include "benode.h"
+#include "besched.h"
+#include "bepeephole.h"
#include "ia32_new_nodes.h"
#include "ia32_optimize.h"
typedef enum produces_flag_t {
produces_no_flag,
- produces_flag_zero,
- produces_flag_carry
+ produces_zero_sign,
+ produces_zero_in_carry
} produces_flag_t;
/**
- * Return which usable flag the given node produces
+ * Return which usable flag the given node produces about the result.
+ * That is zero (ZF) and sign(SF).
+ * We do not check for carry (CF) or overflow (OF).
*
* @param node the node to check
* @param pn the projection number of the used result
*/
-static produces_flag_t produces_test_flag(ir_node *node, int pn)
+static produces_flag_t check_produces_zero_sign(ir_node *node, int pn)
{
ir_node *count;
const ia32_immediate_attr_t *imm_attr;
case iro_ia32_Mul:
return pn == pn_ia32_Mul_res_high ?
- produces_flag_carry : produces_no_flag;
+ produces_zero_in_carry : produces_no_flag;
default:
return produces_no_flag;
}
- return pn == pn_ia32_res ?
- produces_flag_zero : produces_no_flag;
+ return pn == pn_ia32_res ? produces_zero_sign : produces_no_flag;
}
/**
}
set_ia32_ls_mode(test, get_ia32_ls_mode(node));
- reg = arch_irn_get_register(node, pn_ia32_Cmp_eflags);
- arch_irn_set_register(test, pn_ia32_Test_eflags, reg);
+ reg = arch_get_irn_register_out(node, pn_ia32_Cmp_eflags);
+ arch_set_irn_register_out(test, pn_ia32_Test_eflags, reg);
foreach_out_edge_safe(node, edge, tmp) {
ir_node *const user = get_edge_src_irn(edge);
if (left == right) { /* we need a test for 0 */
ir_node *block = get_nodes_block(node);
int pn = pn_ia32_res;
+ ir_node *op = left;
ir_node *flags_proj;
ir_mode *flags_mode;
+ ir_mode *op_mode;
ir_node *schedpoint;
- ir_node *op = left;
const ir_edge_t *edge;
+ produces_flag_t produced;
if (get_nodes_block(left) != block)
return;
panic("couldn't find left");
}
- /* make sure only Lg/Eq tests are used */
+ produced = check_produces_zero_sign(op, pn);
+ if (produced == produces_no_flag)
+ return;
+
+ /* make sure users only look at the sign/zero flag */
foreach_out_edge(node, edge) {
ir_node *user = get_edge_src_irn(edge);
ia32_condition_code_t cc = get_ia32_condcode(user);
- if (cc != ia32_cc_equal && cc != ia32_cc_not_equal) {
- return;
+ if (cc == ia32_cc_equal || cc == ia32_cc_not_equal)
+ continue;
+ if (produced == produces_zero_sign
+ && (cc == ia32_cc_sign || cc == ia32_cc_not_sign)) {
+ continue;
}
+ return;
}
- switch (produces_test_flag(op, pn)) {
- case produces_flag_zero:
- break;
+ op_mode = get_ia32_ls_mode(op);
+ if (op_mode == NULL)
+ op_mode = get_irn_mode(op);
- case produces_flag_carry:
- foreach_out_edge(node, edge) {
- ir_node *user = get_edge_src_irn(edge);
- ia32_condition_code_t cc = get_ia32_condcode(user);
-
- switch (cc) {
- case ia32_cc_equal: cc = ia32_cc_above_equal; break; /* CF = 0 */
- case ia32_cc_not_equal: cc = ia32_cc_below; break; /* CF = 1 */
- default: panic("unexpected pn");
- }
- set_ia32_condcode(user, cc);
- }
- break;
+ /* Make sure we operate on the same bit size */
+ if (get_mode_size_bits(op_mode) != get_mode_size_bits(get_ia32_ls_mode(node)))
+ return;
- default:
- return;
+ if (produced == produces_zero_in_carry) {
+ /* patch users to look at the carry instead of the zero flag */
+ foreach_out_edge(node, edge) {
+ ir_node *user = get_edge_src_irn(edge);
+ ia32_condition_code_t cc = get_ia32_condcode(user);
+
+ switch (cc) {
+ case ia32_cc_equal: cc = ia32_cc_above_equal; break;
+ case ia32_cc_not_equal: cc = ia32_cc_below; break;
+ default: panic("unexpected pn");
+ }
+ set_ia32_condcode(user, cc);
+ }
}
if (get_irn_mode(op) != mode_T) {
if (loads[loadslot] != NULL)
break;
- dreg = arch_irn_get_register(node, pn_ia32_Load_res);
+ dreg = arch_get_irn_register_out(node, pn_ia32_Load_res);
if (regmask & (1 << dreg->index)) {
/* this register is already used */
break;
const arch_register_t *reg;
mem = get_irn_n(load, n_ia32_mem);
- reg = arch_irn_get_register(load, pn_ia32_Load_res);
+ reg = arch_get_irn_register_out(load, pn_ia32_Load_res);
pop = new_bd_ia32_Pop(get_irn_dbg_info(load), block, mem, pred_sp);
- arch_irn_set_register(pop, pn_ia32_Load_res, reg);
+ arch_set_irn_register_out(pop, pn_ia32_Load_res, reg);
copy_mark(load, pop);
if (!rbitset_is_set(birg->allocatable_regs, reg->global_index))
continue;
- if (be_peephole_get_value(CLASS_ia32_gp, i) == NULL)
+ if (be_peephole_get_value(reg->global_index) == NULL)
return reg;
}
if (ia32_cg_config.use_mov_0)
return;
/* xor destroys the flags, so no-one must be using them */
- if (be_peephole_get_value(CLASS_ia32_flags, REG_FLAGS_EFLAGS) != NULL)
+ if (be_peephole_get_value(REG_EFLAGS) != NULL)
return;
reg = arch_get_irn_register(node);
assert(is_ia32_Lea(node));
/* we can only do this if it is allowed to clobber the flags */
- if (be_peephole_get_value(CLASS_ia32_flags, REG_FLAGS_EFLAGS) != NULL)
+ if (be_peephole_get_value(REG_EFLAGS) != NULL)
return;
base = get_irn_n(node, n_ia32_Lea_base);
if (get_mode_size_bits(smaller_mode) != 16 ||
!mode_is_signed(smaller_mode) ||
eax != arch_get_irn_register(val) ||
- eax != arch_irn_get_register(node, pn_ia32_Conv_I2I_res))
+ eax != arch_get_irn_register_out(node, pn_ia32_Conv_I2I_res))
return;
dbgi = get_irn_dbg_info(node);
/* Perform peephole-optimizations. */
void ia32_peephole_optimization(ir_graph *irg)
{
- /* register peephole optimisations */
- clear_irp_opcodes_generic_func();
- register_peephole_optimisation(op_ia32_Const, peephole_ia32_Const);
- register_peephole_optimisation(op_be_IncSP, peephole_be_IncSP);
- register_peephole_optimisation(op_ia32_Lea, peephole_ia32_Lea);
+ /* we currently do it in 2 passes because:
+ * Lea -> Add could be usefull as flag producer for Test later
+ */
+
+ /* pass 1 */
+ ir_clear_opcodes_generic_func();
register_peephole_optimisation(op_ia32_Cmp, peephole_ia32_Cmp);
register_peephole_optimisation(op_ia32_Cmp8Bit, peephole_ia32_Cmp);
- register_peephole_optimisation(op_ia32_Test, peephole_ia32_Test);
- register_peephole_optimisation(op_ia32_Test8Bit, 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);
+ register_peephole_optimisation(op_ia32_Lea, peephole_ia32_Lea);
if (ia32_cg_config.use_short_sex_eax)
register_peephole_optimisation(op_ia32_Conv_I2I, peephole_ia32_Conv_I2I);
+ if (ia32_cg_config.use_pxor)
+ register_peephole_optimisation(op_ia32_xZero, peephole_ia32_xZero);
+ if (! ia32_cg_config.use_imul_mem_imm32)
+ register_peephole_optimisation(op_ia32_IMul, peephole_ia32_Imul_split);
+ be_peephole_opt(irg);
+ /* pass 2 */
+ ir_clear_opcodes_generic_func();
+ register_peephole_optimisation(op_ia32_Const, peephole_ia32_Const);
+ register_peephole_optimisation(op_be_IncSP, peephole_be_IncSP);
+ register_peephole_optimisation(op_ia32_Test, peephole_ia32_Test);
+ register_peephole_optimisation(op_ia32_Test8Bit, peephole_ia32_Test);
+ register_peephole_optimisation(op_be_Return, peephole_ia32_Return);
be_peephole_opt(irg);
}
/* Argh:We must change the opcode to 8bit AND copy the register constraints */
if (get_mode_size_bits(conv_mode) == 8) {
+ const arch_register_req_t **reqs = arch_get_irn_register_reqs_in(node);
set_irn_op(pred, op_ia32_Conv_I2I8Bit);
- arch_set_in_register_reqs(pred,
- arch_get_in_register_reqs(node));
+ arch_set_irn_register_reqs_in(pred, reqs);
}
} else {
/* we don't want to end up with 2 loads, so we better do nothing */
/* Argh:We must change the opcode to 8bit AND copy the register constraints */
if (get_mode_size_bits(conv_mode) == 8) {
+ const arch_register_req_t **reqs = arch_get_irn_register_reqs_in(node);
set_irn_op(result_conv, op_ia32_Conv_I2I8Bit);
- arch_set_in_register_reqs(result_conv,
- arch_get_in_register_reqs(node));
+ arch_set_irn_register_reqs_in(result_conv, reqs);
}
}
} else {
projects / libfirm / blobdiff