2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief This file implements functions to finalize the irg for emit.
23 * @author Christian Wuerdig
35 #include "firmstat_t.h"
41 #include "bearch_ia32_t.h"
42 #include "ia32_finish.h"
43 #include "ia32_new_nodes.h"
44 #include "ia32_common_transform.h"
45 #include "ia32_transform.h"
46 #include "ia32_dbg_stat.h"
47 #include "ia32_optimize.h"
48 #include "gen_ia32_regalloc_if.h"
50 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
53 * Transforms a Sub or xSub into Neg--Add iff OUT_REG != SRC1_REG && OUT_REG == SRC2_REG.
54 * THIS FUNCTIONS MUST BE CALLED AFTER REGISTER ALLOCATION.
56 static void ia32_transform_sub_to_neg_add(ir_node *irn)
59 ir_node *in1, *in2, *noreg, *nomem, *res;
60 ir_node *noreg_fp, *block;
62 const arch_register_t *in1_reg, *in2_reg, *out_reg;
64 /* fix_am will solve this for AddressMode variants */
65 if (get_ia32_op_type(irn) != ia32_Normal)
68 irg = get_irn_irg(irn);
69 noreg = ia32_new_NoReg_gp(irg);
70 noreg_fp = ia32_new_NoReg_xmm(irg);
71 nomem = get_irg_no_mem(irg);
72 in1 = get_irn_n(irn, n_ia32_binary_left);
73 in2 = get_irn_n(irn, n_ia32_binary_right);
74 in1_reg = arch_get_irn_register(in1);
75 in2_reg = arch_get_irn_register(in2);
76 out_reg = arch_get_irn_register_out(irn, 0);
78 if (out_reg == in1_reg)
81 /* in case of sub and OUT == SRC2 we can transform the sequence into neg src2 -- add */
82 if (out_reg != in2_reg)
85 block = get_nodes_block(irn);
86 dbgi = get_irn_dbg_info(irn);
88 /* generate the neg src2 */
89 if (is_ia32_xSub(irn)) {
92 ir_mode *op_mode = get_ia32_ls_mode(irn);
94 assert(get_irn_mode(irn) != mode_T);
96 res = new_bd_ia32_xXor(dbgi, block, noreg, noreg, nomem, in2, noreg_fp);
97 size = get_mode_size_bits(op_mode);
98 entity = ia32_gen_fp_known_const(size == 32 ? ia32_SSIGN : ia32_DSIGN);
99 set_ia32_am_sc(res, entity);
100 set_ia32_op_type(res, ia32_AddrModeS);
101 set_ia32_ls_mode(res, op_mode);
103 arch_set_irn_register(res, in2_reg);
105 /* add to schedule */
106 sched_add_before(irn, res);
108 /* generate the add */
109 res = new_bd_ia32_xAdd(dbgi, block, noreg, noreg, nomem, res, in1);
110 set_ia32_ls_mode(res, get_ia32_ls_mode(irn));
112 ir_node *flags_proj = NULL;
115 if (get_irn_mode(irn) == mode_T) {
116 /* collect the Proj uses */
117 foreach_out_edge(irn, edge) {
118 ir_node *proj = get_edge_src_irn(edge);
119 long pn = get_Proj_proj(proj);
120 if (pn == pn_ia32_flags) {
121 assert(flags_proj == NULL);
128 if (is_ia32_Sbb(irn)) {
129 /* Feed borrow (in CF) as carry (via CMC) into NOT+ADC. */
130 carry = get_irn_n(irn, n_ia32_Sbb_eflags);
131 carry = new_bd_ia32_Cmc(dbgi, block, carry);
133 } else if (flags_proj != 0) {
135 * ARG, the above technique does NOT set the flags right.
136 * So, we must produce the following code:
138 * t2 = a + ~b + Carry
141 * a + -b = a + (~b + 1) would set the carry flag wrong IFF both a and b are zero.
148 carry = new_bd_ia32_Stc(dbgi, block);
151 nnot = new_bd_ia32_Not(dbgi, block, in2);
152 arch_set_irn_register(nnot, in2_reg);
153 sched_add_before(irn, nnot);
155 arch_set_irn_register(carry, &ia32_registers[REG_EFLAGS]);
156 sched_add_before(irn, carry);
158 adc = new_bd_ia32_Adc(dbgi, block, noreg, noreg, nomem, nnot, in1, carry);
159 arch_set_irn_register(adc, out_reg);
160 set_ia32_commutative(adc);
162 if (flags_proj != NULL) {
163 set_irn_mode(adc, mode_T);
164 adc_flags = new_r_Proj(adc, mode_Iu, pn_ia32_Adc_flags);
165 arch_set_irn_register(adc_flags, &ia32_registers[REG_EFLAGS]);
167 cmc = new_bd_ia32_Cmc(dbgi, block, adc_flags);
168 arch_set_irn_register(cmc, &ia32_registers[REG_EFLAGS]);
169 sched_add_after(irn, cmc);
170 exchange(flags_proj, cmc);
175 res = new_bd_ia32_Neg(dbgi, block, in2);
176 arch_set_irn_register(res, in2_reg);
178 /* add to schedule */
179 sched_add_before(irn, res);
181 /* generate the add */
182 res = new_bd_ia32_Add(dbgi, block, noreg, noreg, nomem, res, in1);
183 arch_set_irn_register(res, out_reg);
184 set_ia32_commutative(res);
188 /* exchange the add and the sub */
189 edges_reroute(irn, res);
190 sched_replace(irn, res);
192 set_irn_mode(res, get_irn_mode(irn));
194 SET_IA32_ORIG_NODE(res, irn);
196 /* remove the old sub */
199 DBG_OPT_SUB2NEGADD(irn, res);
202 static inline int need_constraint_copy(ir_node *irn)
204 /* TODO this should be determined from the node specification */
205 switch (get_ia32_irn_opcode(irn)) {
206 case iro_ia32_IMul: {
207 /* the 3 operand form of IMul needs no constraint copy */
208 ir_node *right = get_irn_n(irn, n_ia32_IMul_right);
209 return !is_ia32_Immediate(right);
213 case iro_ia32_Conv_I2I:
214 case iro_ia32_Conv_I2I8Bit:
215 case iro_ia32_CMovcc:
216 case iro_ia32_Minus64Bit:
225 * Returns the index of the "same" register.
226 * On the x86, we should have only one.
228 static int get_first_same(const arch_register_req_t* req)
230 const unsigned other = req->other_same;
233 for (i = 0; i < 32; ++i) {
234 if (other & (1U << i)) return i;
236 panic("same position not found");
240 * Insert copies for all ia32 nodes where the should_be_same requirement
242 * Transform Sub into Neg -- Add if IN2 == OUT
244 static void assure_should_be_same_requirements(ir_node *node)
246 const arch_register_t *out_reg, *in_reg;
247 ir_node *in_node, *block;
249 block = get_nodes_block(node);
251 /* check all OUT requirements, if there is a should_be_same */
252 be_foreach_out(node, i) {
255 ir_node *uses_out_reg;
256 const arch_register_req_t *req = arch_get_irn_register_req_out(node, i);
257 int uses_out_reg_pos;
259 if (!arch_register_req_is(req, should_be_same))
262 same_pos = get_first_same(req);
264 /* get in and out register */
265 out_reg = arch_get_irn_register_out(node, i);
266 in_node = get_irn_n(node, same_pos);
267 in_reg = arch_get_irn_register(in_node);
269 /* requirement already fulfilled? */
270 if (in_reg == out_reg)
272 assert(in_reg->reg_class == out_reg->reg_class);
274 /* check if any other input operands uses the out register */
275 arity = get_irn_arity(node);
277 uses_out_reg_pos = -1;
278 for (i2 = 0; i2 < arity; ++i2) {
279 ir_node *in = get_irn_n(node, i2);
280 const arch_register_t *other_in_reg;
282 if (!mode_is_data(get_irn_mode(in)))
285 other_in_reg = arch_get_irn_register(in);
287 if (other_in_reg != out_reg)
290 if (uses_out_reg != NULL && in != uses_out_reg) {
291 panic("invalid register allocation");
294 if (uses_out_reg_pos >= 0)
295 uses_out_reg_pos = -1; /* multiple inputs... */
297 uses_out_reg_pos = i2;
300 /* no-one else is using the out reg, we can simply copy it
301 * (the register can't be live since the operation will override it
303 if (uses_out_reg == NULL) {
304 ir_node *copy = be_new_Copy(block, in_node);
305 DBG_OPT_2ADDRCPY(copy);
307 /* destination is the out register */
308 arch_set_irn_register(copy, out_reg);
310 /* insert copy before the node into the schedule */
311 sched_add_before(node, copy);
314 set_irn_n(node, same_pos, copy);
317 "created copy %+F for should be same argument at input %d of %+F\n",
318 copy, same_pos, node));
322 /* for commutative nodes we can simply swap the left/right */
323 if (uses_out_reg_pos == n_ia32_binary_right && is_ia32_commutative(node)) {
324 ia32_swap_left_right(node);
326 "swapped left/right input of %+F to resolve should be same constraint\n",
331 panic("Unresolved should_be_same constraint");
337 * We have a source address mode node with base or index register equal to
338 * result register and unfulfilled should_be_same requirement. The constraint
339 * handler will insert a copy from the remaining input operand to the result
340 * register -> base or index is broken then.
341 * Solution: Turn back this address mode into explicit Load + Operation.
343 static void fix_am_source(ir_node *irn)
345 /* check only ia32 nodes with source address mode */
346 if (!is_ia32_irn(irn) || get_ia32_op_type(irn) != ia32_AddrModeS)
348 /* only need to fix binary operations */
349 if (get_ia32_am_support(irn) != ia32_am_binary)
352 be_foreach_out(irn, i) {
353 const arch_register_req_t *req = arch_get_irn_register_req_out(irn, i);
354 const arch_register_t *out_reg;
357 const arch_register_t *same_reg;
360 if (!arch_register_req_is(req, should_be_same))
363 /* get in and out register */
364 out_reg = arch_get_irn_register_out(irn, i);
365 same_pos = get_first_same(req);
366 same_node = get_irn_n(irn, same_pos);
367 same_reg = arch_get_irn_register(same_node);
369 /* should_be same constraint is fullfilled, nothing to do */
370 if (out_reg == same_reg)
373 /* we only need to do something if the out reg is the same as base
375 if (out_reg != arch_get_irn_register(get_irn_n(irn, n_ia32_base)) &&
376 out_reg != arch_get_irn_register(get_irn_n(irn, n_ia32_index)))
379 load_res = ia32_turn_back_am(irn);
380 arch_set_irn_register(load_res, out_reg);
383 "irg %+F: build back AM source for node %+F, inserted load %+F\n",
384 get_irn_irg(irn), irn, get_Proj_pred(load_res)));
390 * Block walker: finishes a block
392 static void ia32_finish_irg_walker(ir_node *block, void *env)
397 /* first: turn back AM source if necessary */
398 for (irn = sched_first(block); ! sched_is_end(irn); irn = next) {
399 next = sched_next(irn);
403 for (irn = sched_first(block); ! sched_is_end(irn); irn = next) {
404 next = sched_next(irn);
406 /* check if there is a sub which need to be transformed */
407 if (is_ia32_Sub(irn) || is_ia32_Sbb(irn) || is_ia32_xSub(irn)) {
408 ia32_transform_sub_to_neg_add(irn);
412 /* second: insert copies and finish irg */
413 for (irn = sched_first(block); ! sched_is_end(irn); irn = next) {
414 next = sched_next(irn);
415 if (is_ia32_irn(irn)) {
416 /* some nodes are just a bit less efficient, but need no fixing if the
417 * should be same requirement is not fulfilled */
418 if (need_constraint_copy(irn))
419 assure_should_be_same_requirements(irn);
425 * Block walker: pushes all blocks on a wait queue
427 static void ia32_push_on_queue_walker(ir_node *block, void *env)
429 waitq *wq = (waitq*)env;
430 waitq_put(wq, block);
435 * Add Copy nodes for not fulfilled should_be_equal constraints
437 void ia32_finish_irg(ir_graph *irg)
439 waitq *wq = new_waitq();
441 /* Push the blocks on the waitq because ia32_finish_irg_walker starts more walks ... */
442 irg_block_walk_graph(irg, NULL, ia32_push_on_queue_walker, wq);
444 while (! waitq_empty(wq)) {
445 ir_node *block = (ir_node*)waitq_get(wq);
446 ia32_finish_irg_walker(block, NULL);
451 void ia32_init_finish(void)
453 FIRM_DBG_REGISTER(dbg, "firm.be.ia32.finish");