2 * local-propagation.c: Local constant, copy and tree propagation.
4 * To make some sense of the tree mover, read mono/docs/tree-mover.txt
7 * Paolo Molaro (lupus@ximian.com)
8 * Dietmar Maurer (dietmar@ximian.com)
9 * Massimiliano Mantione (massi@ximian.com)
11 * (C) 2006 Novell, Inc. http://www.novell.com
12 * Copyright 2011 Xamarin, Inc (http://www.xamarin.com)
13 * Licensed under the MIT license. See LICENSE file in the project root for full license information.
25 #include <mono/metadata/debug-helpers.h>
26 #include <mono/metadata/mempool.h>
27 #include <mono/metadata/opcodes.h>
31 #ifndef MONO_ARCH_IS_OP_MEMBASE
32 #define MONO_ARCH_IS_OP_MEMBASE(opcode) FALSE
35 static inline MonoBitSet*
36 mono_bitset_mp_new_noinit (MonoMemPool *mp, guint32 max_size)
38 int size = mono_bitset_alloc_size (max_size, 0);
41 mem = mono_mempool_alloc (mp, size);
42 return mono_bitset_mem_new (mem, max_size, MONO_BITSET_DONT_FREE);
45 struct magic_unsigned {
56 /* http://www.hackersdelight.org/hdcodetxt/magicu.c.txt */
57 static struct magic_unsigned
58 compute_magic_unsigned (guint32 divisor) {
59 guint32 nc, delta, q1, r1, q2, r2;
60 struct magic_unsigned magu;
65 nc = -1 - (-divisor) % divisor;
68 r1 = 0x80000000 - q1 * nc;
69 q2 = 0x7FFFFFFF / divisor;
70 r2 = 0x7FFFFFFF - q2 * divisor;
82 if (r2 + 1 >= divisor - r2) {
86 r2 = 2 * r2 + 1 - divisor;
93 delta = divisor - 1 - r2;
94 } while (!gt && (q1 < delta || (q1 == delta && r1 == 0)));
96 magu.magic_number = q2 + 1;
101 /* http://www.hackersdelight.org/hdcodetxt/magic.c.txt */
102 static struct magic_signed
103 compute_magic_signed (gint32 divisor) {
105 guint32 ad, anc, delta, q1, r1, q2, r2, t;
106 const guint32 two31 = 0x80000000;
107 struct magic_signed mag;
110 t = two31 + ((unsigned)divisor >> 31);
111 anc = t - 1 - t % ad;
114 r1 = two31 - q1 * anc;
116 r2 = two31 - q2 * ad;
135 } while (q1 < delta || (q1 == delta && r1 == 0));
137 mag.magic_number = q2 + 1;
139 mag.magic_number = -mag.magic_number;
145 mono_strength_reduction_division (MonoCompile *cfg, MonoInst *ins)
147 gboolean allocated_vregs = FALSE;
149 * We don't use it on 32bit systems because on those
150 * platforms we emulate long multiplication, driving the
151 * performance back down.
153 switch (ins->opcode) {
154 case OP_IDIV_UN_IMM: {
156 #if SIZEOF_REGISTER == 8
157 guint32 dividend_reg;
161 struct magic_unsigned mag;
162 int power2 = mono_is_power_of_two (ins->inst_imm);
164 /* The decomposition doesn't handle exception throwing */
165 if (ins->inst_imm == 0)
169 ins->opcode = OP_ISHR_UN_IMM;
171 ins->inst_imm = power2;
174 allocated_vregs = TRUE;
176 * Replacement of unsigned division with multiplication,
177 * shifts and additions Hacker's Delight, chapter 10-10.
179 mag = compute_magic_unsigned (ins->inst_imm);
180 tmp_regl = alloc_lreg (cfg);
181 #if SIZEOF_REGISTER == 8
182 dividend_reg = alloc_lreg (cfg);
183 MONO_EMIT_NEW_I8CONST (cfg, tmp_regl, mag.magic_number);
184 MONO_EMIT_NEW_UNALU (cfg, OP_ZEXT_I4, dividend_reg, ins->sreg1);
185 MONO_EMIT_NEW_BIALU (cfg, OP_LMUL, tmp_regl, dividend_reg, tmp_regl);
187 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_LSHR_UN_IMM, tmp_regl, tmp_regl, 32);
188 MONO_EMIT_NEW_BIALU (cfg, OP_LADD, tmp_regl, tmp_regl, dividend_reg);
189 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_LSHR_UN_IMM, ins->dreg, tmp_regl, mag.shift);
191 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_LSHR_UN_IMM, ins->dreg, tmp_regl, 32 + mag.shift);
194 tmp_regi = alloc_ireg (cfg);
195 MONO_EMIT_NEW_ICONST (cfg, tmp_regi, mag.magic_number);
196 MONO_EMIT_NEW_BIALU (cfg, OP_BIGMUL_UN, tmp_regl, ins->sreg1, tmp_regi);
197 /* Long shifts below will be decomposed during cprop */
199 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_LSHR_UN_IMM, tmp_regl, tmp_regl, 32);
200 MONO_EMIT_NEW_BIALU (cfg, OP_IADDCC, MONO_LVREG_LS (tmp_regl), MONO_LVREG_LS (tmp_regl), ins->sreg1);
201 /* MONO_LVREG_MS (tmp_reg) is 0, save in it the carry */
202 MONO_EMIT_NEW_BIALU (cfg, OP_IADC, MONO_LVREG_MS (tmp_regl), MONO_LVREG_MS (tmp_regl), MONO_LVREG_MS (tmp_regl));
203 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_LSHR_UN_IMM, tmp_regl, tmp_regl, mag.shift);
205 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_LSHR_UN_IMM, tmp_regl, tmp_regl, 32 + mag.shift);
207 MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, ins->dreg, MONO_LVREG_LS (tmp_regl));
209 mono_jit_stats.optimized_divisions++;
214 #if SIZEOF_REGISTER == 8
215 guint32 dividend_reg;
219 struct magic_signed mag;
220 int power2 = mono_is_power_of_two (ins->inst_imm);
221 /* The decomposition doesn't handle exception throwing */
222 if (ins->inst_imm == 0 || ins->inst_imm == -1)
224 allocated_vregs = TRUE;
226 guint32 r1 = alloc_ireg (cfg);
227 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_ISHR_UN_IMM, r1, ins->sreg1, 31);
228 MONO_EMIT_NEW_BIALU (cfg, OP_IADD, r1, r1, ins->sreg1);
229 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_ISHR_IMM, ins->dreg, r1, 1);
231 } else if (power2 > 0 && power2 < 31) {
232 guint32 r1 = alloc_ireg (cfg);
233 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_ISHR_IMM, r1, ins->sreg1, 31);
234 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_ISHR_UN_IMM, r1, r1, (32 - power2));
235 MONO_EMIT_NEW_BIALU (cfg, OP_IADD, r1, r1, ins->sreg1);
236 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_ISHR_IMM, ins->dreg, r1, power2);
241 * Replacement of signed division with multiplication,
242 * shifts and additions Hacker's Delight, chapter 10-6.
244 mag = compute_magic_signed (ins->inst_imm);
245 tmp_regl = alloc_lreg (cfg);
246 #if SIZEOF_REGISTER == 8
247 dividend_reg = alloc_lreg (cfg);
248 MONO_EMIT_NEW_I8CONST (cfg, tmp_regl, mag.magic_number);
249 MONO_EMIT_NEW_UNALU (cfg, OP_SEXT_I4, dividend_reg, ins->sreg1);
250 MONO_EMIT_NEW_BIALU (cfg, OP_LMUL, tmp_regl, dividend_reg, tmp_regl);
251 if ((ins->inst_imm > 0 && mag.magic_number < 0) || (ins->inst_imm < 0 && mag.magic_number > 0)) {
252 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_LSHR_IMM, tmp_regl, tmp_regl, 32);
253 if (ins->inst_imm > 0 && mag.magic_number < 0) {
254 MONO_EMIT_NEW_BIALU (cfg, OP_LADD, tmp_regl, tmp_regl, dividend_reg);
255 } else if (ins->inst_imm < 0 && mag.magic_number > 0) {
256 MONO_EMIT_NEW_BIALU (cfg, OP_LSUB, tmp_regl, tmp_regl, dividend_reg);
258 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_LSHR_IMM, tmp_regl, tmp_regl, mag.shift);
260 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_LSHR_IMM, tmp_regl, tmp_regl, 32 + mag.shift);
262 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_LSHR_UN_IMM, ins->dreg, tmp_regl, SIZEOF_REGISTER * 8 - 1);
263 MONO_EMIT_NEW_BIALU (cfg, OP_LADD, ins->dreg, ins->dreg, tmp_regl);
265 tmp_regi = alloc_ireg (cfg);
266 MONO_EMIT_NEW_ICONST (cfg, tmp_regi, mag.magic_number);
267 MONO_EMIT_NEW_BIALU (cfg, OP_BIGMUL, tmp_regl, ins->sreg1, tmp_regi);
268 if ((ins->inst_imm > 0 && mag.magic_number < 0) || (ins->inst_imm < 0 && mag.magic_number > 0)) {
269 if (ins->inst_imm > 0 && mag.magic_number < 0) {
270 /* Opposite sign, cannot overflow */
271 MONO_EMIT_NEW_BIALU (cfg, OP_IADD, tmp_regi, MONO_LVREG_MS (tmp_regl), ins->sreg1);
272 } else if (ins->inst_imm < 0 && mag.magic_number > 0) {
273 /* Same sign, cannot overflow */
274 MONO_EMIT_NEW_BIALU (cfg, OP_ISUB, tmp_regi, MONO_LVREG_MS (tmp_regl), ins->sreg1);
276 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_ISHR_IMM, tmp_regi, tmp_regi, mag.shift);
278 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_ISHR_IMM, tmp_regi, MONO_LVREG_MS (tmp_regl), mag.shift);
280 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_ISHR_UN_IMM, ins->dreg, tmp_regi, SIZEOF_REGISTER * 8 - 1);
281 MONO_EMIT_NEW_BIALU (cfg, OP_IADD, ins->dreg, ins->dreg, tmp_regi);
283 mono_jit_stats.optimized_divisions++;
287 return allocated_vregs;
291 * Replaces ins with optimized opcodes.
293 * We can emit to cbb the equivalent instructions which will be used as
294 * replacement for ins, or simply change the fields of ins. Spec needs to
295 * be updated if we silently change the opcode of ins.
297 * Returns TRUE if additional vregs were allocated.
300 mono_strength_reduction_ins (MonoCompile *cfg, MonoInst *ins, const char **spec)
302 gboolean allocated_vregs = FALSE;
304 /* FIXME: Add long/float */
305 switch (ins->opcode) {
308 if (ins->dreg == ins->sreg1) {
316 #if SIZEOF_REGISTER == 8
320 if (ins->inst_imm == 0) {
321 ins->opcode = OP_MOVE;
326 #if SIZEOF_REGISTER == 8
329 if (ins->inst_imm == 0) {
330 ins->opcode = (ins->opcode == OP_LMUL_IMM) ? OP_I8CONST : OP_ICONST;
333 } else if (ins->inst_imm == 1) {
334 ins->opcode = OP_MOVE;
335 } else if ((ins->opcode == OP_IMUL_IMM) && (ins->inst_imm == -1)) {
336 ins->opcode = OP_INEG;
337 } else if ((ins->opcode == OP_LMUL_IMM) && (ins->inst_imm == -1)) {
338 ins->opcode = OP_LNEG;
340 int power2 = mono_is_power_of_two (ins->inst_imm);
342 ins->opcode = (ins->opcode == OP_MUL_IMM) ? OP_SHL_IMM : ((ins->opcode == OP_LMUL_IMM) ? OP_LSHL_IMM : OP_ISHL_IMM);
343 ins->inst_imm = power2;
347 case OP_IREM_UN_IMM: {
348 int power2 = mono_is_power_of_two (ins->inst_imm);
351 ins->opcode = OP_IAND_IMM;
353 ins->inst_imm = (1 << power2) - 1;
359 allocated_vregs = mono_strength_reduction_division (cfg, ins);
362 #if SIZEOF_REGISTER == 8
366 int power = mono_is_power_of_two (ins->inst_imm);
367 if (ins->inst_imm == 1) {
368 ins->opcode = OP_ICONST;
369 MONO_INST_NULLIFY_SREGS (ins);
374 } else if ((ins->inst_imm > 0) && (ins->inst_imm < (1LL << 32)) && (power != -1)) {
375 gboolean is_long = ins->opcode == OP_LREM_IMM;
376 int compensator_reg = alloc_ireg (cfg);
377 int intermediate_reg;
379 /* Based on gcc code */
381 /* Add compensation for negative numerators */
384 intermediate_reg = compensator_reg;
385 MONO_EMIT_NEW_BIALU_IMM (cfg, is_long ? OP_LSHR_IMM : OP_ISHR_IMM, intermediate_reg, ins->sreg1, is_long ? 63 : 31);
387 intermediate_reg = ins->sreg1;
390 MONO_EMIT_NEW_BIALU_IMM (cfg, is_long ? OP_LSHR_UN_IMM : OP_ISHR_UN_IMM, compensator_reg, intermediate_reg, (is_long ? 64 : 32) - power);
391 MONO_EMIT_NEW_BIALU (cfg, is_long ? OP_LADD : OP_IADD, ins->dreg, ins->sreg1, compensator_reg);
392 /* Compute remainder */
393 MONO_EMIT_NEW_BIALU_IMM (cfg, is_long ? OP_LAND_IMM : OP_AND_IMM, ins->dreg, ins->dreg, (1 << power) - 1);
394 /* Remove compensation */
395 MONO_EMIT_NEW_BIALU (cfg, is_long ? OP_LSUB : OP_ISUB, ins->dreg, ins->dreg, compensator_reg);
397 allocated_vregs = TRUE;
402 #if SIZEOF_REGISTER == 4
404 if (COMPILE_LLVM (cfg))
406 if (ins->inst_c1 == 32) {
407 MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, MONO_LVREG_LS (ins->dreg), MONO_LVREG_MS (ins->sreg1));
408 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_ISHR_IMM, MONO_LVREG_MS (ins->dreg), MONO_LVREG_MS (ins->sreg1), 31);
409 } else if (ins->inst_c1 == 0) {
410 MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, MONO_LVREG_LS (ins->dreg), MONO_LVREG_LS (ins->sreg1));
411 MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, MONO_LVREG_MS (ins->dreg), MONO_LVREG_MS (ins->sreg1));
412 } else if (ins->inst_c1 > 32) {
413 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_ISHR_IMM, MONO_LVREG_LS (ins->dreg), MONO_LVREG_MS (ins->sreg1), ins->inst_c1 - 32);
414 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_ISHR_IMM, MONO_LVREG_MS (ins->dreg), MONO_LVREG_MS (ins->sreg1), 31);
416 guint32 tmpreg = alloc_ireg (cfg);
417 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_ISHL_IMM, tmpreg, MONO_LVREG_MS (ins->sreg1), 32 - ins->inst_c1);
418 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_ISHR_IMM, MONO_LVREG_MS (ins->dreg), MONO_LVREG_MS (ins->sreg1), ins->inst_c1);
419 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_ISHR_UN_IMM, MONO_LVREG_LS (ins->dreg), MONO_LVREG_LS (ins->sreg1), ins->inst_c1);
420 MONO_EMIT_NEW_BIALU (cfg, OP_IOR, MONO_LVREG_LS (ins->dreg), MONO_LVREG_LS (ins->dreg), tmpreg);
421 allocated_vregs = TRUE;
425 case OP_LSHR_UN_IMM: {
426 if (COMPILE_LLVM (cfg))
428 if (ins->inst_c1 == 32) {
429 MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, MONO_LVREG_LS (ins->dreg), MONO_LVREG_MS (ins->sreg1));
430 MONO_EMIT_NEW_ICONST (cfg, MONO_LVREG_MS (ins->dreg), 0);
431 } else if (ins->inst_c1 == 0) {
432 MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, MONO_LVREG_LS (ins->dreg), MONO_LVREG_LS (ins->sreg1));
433 MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, MONO_LVREG_MS (ins->dreg), MONO_LVREG_MS (ins->sreg1));
434 } else if (ins->inst_c1 > 32) {
435 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_ISHR_UN_IMM, MONO_LVREG_LS (ins->dreg), MONO_LVREG_MS (ins->sreg1), ins->inst_c1 - 32);
436 MONO_EMIT_NEW_ICONST (cfg, MONO_LVREG_MS (ins->dreg), 0);
438 guint32 tmpreg = alloc_ireg (cfg);
439 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_ISHL_IMM, tmpreg, MONO_LVREG_MS (ins->sreg1), 32 - ins->inst_c1);
440 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_ISHR_UN_IMM, MONO_LVREG_MS (ins->dreg), MONO_LVREG_MS (ins->sreg1), ins->inst_c1);
441 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_ISHR_UN_IMM, MONO_LVREG_LS (ins->dreg), MONO_LVREG_LS (ins->sreg1), ins->inst_c1);
442 MONO_EMIT_NEW_BIALU (cfg, OP_IOR, MONO_LVREG_LS (ins->dreg), MONO_LVREG_LS (ins->dreg), tmpreg);
443 allocated_vregs = TRUE;
448 if (COMPILE_LLVM (cfg))
450 if (ins->inst_c1 == 32) {
451 /* just move the lower half to the upper and zero the lower word */
452 MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, MONO_LVREG_MS (ins->dreg), MONO_LVREG_LS (ins->sreg1));
453 MONO_EMIT_NEW_ICONST (cfg, MONO_LVREG_LS (ins->dreg), 0);
454 } else if (ins->inst_c1 == 0) {
455 MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, MONO_LVREG_LS (ins->dreg), MONO_LVREG_LS (ins->sreg1));
456 MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, MONO_LVREG_MS (ins->dreg), MONO_LVREG_MS (ins->sreg1));
457 } else if (ins->inst_c1 > 32) {
458 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_ISHL_IMM, MONO_LVREG_MS (ins->dreg), MONO_LVREG_LS (ins->sreg1), ins->inst_c1 - 32);
459 MONO_EMIT_NEW_ICONST (cfg, MONO_LVREG_LS (ins->dreg), 0);
461 guint32 tmpreg = alloc_ireg (cfg);
462 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_ISHR_UN_IMM, tmpreg, MONO_LVREG_LS (ins->sreg1), 32 - ins->inst_c1);
463 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_ISHL_IMM, MONO_LVREG_MS (ins->dreg), MONO_LVREG_MS (ins->sreg1), ins->inst_c1);
464 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_ISHL_IMM, MONO_LVREG_LS (ins->dreg), MONO_LVREG_LS (ins->sreg1), ins->inst_c1);
465 MONO_EMIT_NEW_BIALU (cfg, OP_IOR, MONO_LVREG_MS (ins->dreg), MONO_LVREG_MS (ins->dreg), tmpreg);
466 allocated_vregs = TRUE;
476 *spec = INS_INFO (ins->opcode);
477 return allocated_vregs;
483 * A combined local copy and constant propagation pass.
486 mono_local_cprop (MonoCompile *cfg)
488 MonoBasicBlock *bb, *bb_opt;
492 int filter = FILTER_IL_SEQ_POINT;
493 int initial_max_vregs = cfg->next_vreg;
495 max = cfg->next_vreg;
496 defs = (MonoInst **)mono_mempool_alloc (cfg->mempool, sizeof (MonoInst*) * cfg->next_vreg);
497 def_index = (gint32 *)mono_mempool_alloc (cfg->mempool, sizeof (guint32) * cfg->next_vreg);
498 cfg->cbb = bb_opt = mono_mempool_alloc0 ((cfg)->mempool, sizeof (MonoBasicBlock));
500 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
505 /* Manually init the defs entries used by the bblock */
506 MONO_BB_FOR_EACH_INS (bb, ins) {
507 int sregs [MONO_MAX_SRC_REGS];
510 if (ins->dreg != -1) {
511 #if SIZEOF_REGISTER == 4
512 const char *spec = INS_INFO (ins->opcode);
513 if (spec [MONO_INST_DEST] == 'l') {
514 defs [ins->dreg + 1] = NULL;
515 defs [ins->dreg + 2] = NULL;
518 defs [ins->dreg] = NULL;
521 num_sregs = mono_inst_get_src_registers (ins, sregs);
522 for (i = 0; i < num_sregs; ++i) {
523 int sreg = sregs [i];
524 #if SIZEOF_REGISTER == 4
525 const char *spec = INS_INFO (ins->opcode);
526 if (spec [MONO_INST_SRC1 + i] == 'l') {
527 defs [sreg + 1] = NULL;
528 defs [sreg + 2] = NULL;
536 last_call_index = -1;
537 MONO_BB_FOR_EACH_INS (bb, ins) {
538 const char *spec = INS_INFO (ins->opcode);
539 int regtype, srcindex, sreg;
541 int sregs [MONO_MAX_SRC_REGS];
543 if (ins->opcode == OP_NOP) {
544 MONO_DELETE_INS (bb, ins);
548 g_assert (ins->opcode > MONO_CEE_LAST);
550 /* FIXME: Optimize this */
551 if (ins->opcode == OP_LDADDR) {
552 MonoInst *var = (MonoInst *)ins->inst_p0;
554 defs [var->dreg] = NULL;
556 if (!MONO_TYPE_ISSTRUCT (var->inst_vtype))
561 if (MONO_IS_STORE_MEMBASE (ins)) {
565 if ((regtype == 'i') && (sreg != -1) && defs [sreg]) {
566 MonoInst *def = defs [sreg];
568 if ((def->opcode == OP_MOVE) && (!defs [def->sreg1] || (def_index [def->sreg1] < def_index [sreg])) && !vreg_is_volatile (cfg, def->sreg1)) {
569 int vreg = def->sreg1;
570 if (cfg->verbose_level > 2) printf ("CCOPY: R%d -> R%d\n", sreg, vreg);
576 num_sregs = mono_inst_get_src_registers (ins, sregs);
577 for (srcindex = 0; srcindex < num_sregs; ++srcindex) {
580 mono_inst_get_src_registers (ins, sregs);
582 regtype = spec [MONO_INST_SRC1 + srcindex];
583 sreg = sregs [srcindex];
585 if ((regtype == ' ') || (sreg == -1) || (!defs [sreg]))
590 /* Copy propagation */
592 * The first check makes sure the source of the copy did not change since
594 * The second check avoids volatile variables.
595 * The third check avoids copy propagating local vregs through a call,
596 * since the lvreg will be spilled
597 * The fourth check avoids copy propagating a vreg in cases where
598 * it would be eliminated anyway by reverse copy propagation later,
599 * because propagating it would create another use for it, thus making
600 * it impossible to use reverse copy propagation.
602 /* Enabling this for floats trips up the fp stack */
604 * Enabling this for floats on amd64 seems to cause a failure in
605 * basic-math.cs, most likely because it gets rid of some r8->r4
608 if (MONO_IS_MOVE (def) &&
609 (!defs [def->sreg1] || (def_index [def->sreg1] < def_index [sreg])) &&
610 !vreg_is_volatile (cfg, def->sreg1) &&
611 /* This avoids propagating local vregs across calls */
612 ((get_vreg_to_inst (cfg, def->sreg1) || !defs [def->sreg1] || (def_index [def->sreg1] >= last_call_index) || (def->opcode == OP_VMOVE))) &&
613 !(defs [def->sreg1] && mono_inst_next (defs [def->sreg1], filter) == def) &&
614 (!MONO_ARCH_USE_FPSTACK || (def->opcode != OP_FMOVE)) &&
615 (def->opcode != OP_FMOVE)) {
616 int vreg = def->sreg1;
618 if (cfg->verbose_level > 2) printf ("CCOPY/2: R%d -> R%d\n", sreg, vreg);
619 sregs [srcindex] = vreg;
620 mono_inst_set_src_registers (ins, sregs);
622 /* Allow further iterations */
627 /* Constant propagation */
628 /* FIXME: Make is_inst_imm a macro */
629 /* FIXME: Make is_inst_imm take an opcode argument */
630 /* is_inst_imm is only needed for binops */
631 if ((((def->opcode == OP_ICONST) || ((sizeof (gpointer) == 8) && (def->opcode == OP_I8CONST))) &&
632 (((srcindex == 0) && (ins->sreg2 == -1)) || mono_arch_is_inst_imm (def->inst_c0))) ||
633 (!MONO_ARCH_USE_FPSTACK && (def->opcode == OP_R8CONST))) {
636 /* srcindex == 1 -> binop, ins->sreg2 == -1 -> unop */
637 if ((srcindex == 1) && (ins->sreg1 != -1) && defs [ins->sreg1] && (defs [ins->sreg1]->opcode == OP_ICONST) && defs [ins->sreg2]) {
638 /* Both arguments are constants, perform cfold */
639 mono_constant_fold_ins (cfg, ins, defs [ins->sreg1], defs [ins->sreg2], TRUE);
640 } else if ((srcindex == 0) && (ins->sreg2 != -1) && defs [ins->sreg2]) {
641 /* Arg 1 is constant, swap arguments if possible */
642 int opcode = ins->opcode;
643 mono_constant_fold_ins (cfg, ins, defs [ins->sreg1], defs [ins->sreg2], TRUE);
644 if (ins->opcode != opcode) {
645 /* Allow further iterations */
649 } else if ((srcindex == 0) && (ins->sreg2 == -1)) {
650 /* Constant unop, perform cfold */
651 mono_constant_fold_ins (cfg, ins, defs [ins->sreg1], NULL, TRUE);
654 opcode2 = mono_op_to_op_imm (ins->opcode);
655 if ((opcode2 != -1) && mono_arch_is_inst_imm (def->inst_c0) && ((srcindex == 1) || (ins->sreg2 == -1))) {
656 ins->opcode = opcode2;
657 if ((def->opcode == OP_I8CONST) && (sizeof (gpointer) == 4)) {
658 ins->inst_ls_word = def->inst_ls_word;
659 ins->inst_ms_word = def->inst_ms_word;
661 ins->inst_imm = def->inst_c0;
663 sregs [srcindex] = -1;
664 mono_inst_set_src_registers (ins, sregs);
666 if ((opcode2 == OP_VOIDCALL) || (opcode2 == OP_CALL) || (opcode2 == OP_LCALL) || (opcode2 == OP_FCALL))
667 ((MonoCallInst*)ins)->fptr = (gpointer)ins->inst_imm;
669 /* Allow further iterations */
675 #if defined(TARGET_X86) || defined(TARGET_AMD64)
676 if ((ins->opcode == OP_X86_LEA) && (srcindex == 1)) {
677 #if SIZEOF_REGISTER == 8
678 /* FIXME: Use OP_PADD_IMM when the new JIT is done */
679 ins->opcode = OP_LADD_IMM;
681 ins->opcode = OP_ADD_IMM;
683 ins->inst_imm += def->inst_c0 << ins->backend.shift_amount;
687 opcode2 = mono_load_membase_to_load_mem (ins->opcode);
688 if ((srcindex == 0) && (opcode2 != -1) && mono_arch_is_inst_imm (def->inst_c0)) {
689 ins->opcode = opcode2;
690 ins->inst_imm = def->inst_c0 + ins->inst_offset;
695 else if (((def->opcode == OP_ADD_IMM) || (def->opcode == OP_LADD_IMM)) && (MONO_IS_LOAD_MEMBASE (ins) || MONO_ARCH_IS_OP_MEMBASE (ins->opcode))) {
696 /* ADD_IMM is created by spill_global_vars */
698 * We have to guarantee that def->sreg1 haven't changed since def->dreg
699 * was defined. cfg->frame_reg is assumed to remain constant.
701 if ((def->sreg1 == cfg->frame_reg) || ((mono_inst_next (def, filter) == ins) && (def->dreg != def->sreg1))) {
702 ins->inst_basereg = def->sreg1;
703 ins->inst_offset += def->inst_imm;
705 } else if ((ins->opcode == OP_ISUB_IMM) && (def->opcode == OP_IADD_IMM) && (mono_inst_next (def, filter) == ins) && (def->dreg != def->sreg1)) {
706 ins->sreg1 = def->sreg1;
707 ins->inst_imm -= def->inst_imm;
708 } else if ((ins->opcode == OP_IADD_IMM) && (def->opcode == OP_ISUB_IMM) && (mono_inst_next (def, filter) == ins) && (def->dreg != def->sreg1)) {
709 ins->sreg1 = def->sreg1;
710 ins->inst_imm -= def->inst_imm;
711 } else if (ins->opcode == OP_STOREI1_MEMBASE_REG &&
712 (def->opcode == OP_ICONV_TO_U1 || def->opcode == OP_ICONV_TO_I1 || def->opcode == OP_SEXT_I4 || (SIZEOF_REGISTER == 8 && def->opcode == OP_LCONV_TO_U1)) &&
713 (!defs [def->sreg1] || (def_index [def->sreg1] < def_index [sreg]))) {
714 /* Avoid needless sign extension */
715 ins->sreg1 = def->sreg1;
716 } else if (ins->opcode == OP_STOREI2_MEMBASE_REG &&
717 (def->opcode == OP_ICONV_TO_U2 || def->opcode == OP_ICONV_TO_I2 || def->opcode == OP_SEXT_I4 || (SIZEOF_REGISTER == 8 && def->opcode == OP_LCONV_TO_I2)) &&
718 (!defs [def->sreg1] || (def_index [def->sreg1] < def_index [sreg]))) {
719 /* Avoid needless sign extension */
720 ins->sreg1 = def->sreg1;
721 } else if (ins->opcode == OP_COMPARE_IMM && def->opcode == OP_LDADDR && ins->inst_imm == 0) {
724 memset (&dummy_arg1, 0, sizeof (MonoInst));
725 dummy_arg1.opcode = OP_ICONST;
726 dummy_arg1.inst_c0 = 1;
728 mono_constant_fold_ins (cfg, ins, &dummy_arg1, NULL, TRUE);
732 g_assert (cfg->cbb == bb_opt);
733 g_assert (!bb_opt->code);
734 /* Do strength reduction here */
735 if (mono_strength_reduction_ins (cfg, ins, &spec) && max < cfg->next_vreg) {
736 MonoInst **defs_prev = defs;
737 gint32 *def_index_prev = def_index;
738 guint32 prev_max = max;
739 guint32 additional_vregs = cfg->next_vreg - initial_max_vregs;
741 /* We have more vregs so we need to reallocate defs and def_index arrays */
742 max = initial_max_vregs + additional_vregs * 2;
743 defs = (MonoInst **)mono_mempool_alloc (cfg->mempool, sizeof (MonoInst*) * max);
744 def_index = (gint32 *)mono_mempool_alloc (cfg->mempool, sizeof (guint32) * max);
746 /* Keep the entries for the previous vregs, zero the rest */
747 memcpy (defs, defs_prev, sizeof (MonoInst*) * prev_max);
748 memset (defs + prev_max, 0, sizeof (MonoInst*) * (max - prev_max));
749 memcpy (def_index, def_index_prev, sizeof (guint32) * prev_max);
750 memset (def_index + prev_max, 0, sizeof (guint32) * (max - prev_max));
753 if (cfg->cbb->code || (cfg->cbb != bb_opt)) {
754 MonoInst *saved_prev = ins->prev;
756 /* If we have code in cbb, we need to replace ins with the decomposition */
757 mono_replace_ins (cfg, bb, ins, &ins->prev, bb_opt, cfg->cbb);
758 bb_opt->code = bb_opt->last_ins = NULL;
759 bb_opt->in_count = bb_opt->out_count = 0;
762 /* ins is hanging, continue scanning the emitted code */
767 if (spec [MONO_INST_DEST] != ' ') {
768 MonoInst *def = defs [ins->dreg];
770 if (def && (def->opcode == OP_ADD_IMM) && (def->sreg1 == cfg->frame_reg) && (MONO_IS_STORE_MEMBASE (ins))) {
771 /* ADD_IMM is created by spill_global_vars */
772 /* cfg->frame_reg is assumed to remain constant */
773 ins->inst_destbasereg = def->sreg1;
774 ins->inst_offset += def->inst_imm;
777 if (!MONO_IS_STORE_MEMBASE (ins) && !vreg_is_volatile (cfg, ins->dreg)) {
778 defs [ins->dreg] = ins;
779 def_index [ins->dreg] = ins_index;
783 if (MONO_IS_CALL (ins))
784 last_call_index = ins_index;
791 static inline gboolean
792 reg_is_softreg_no_fpstack (int reg, const char spec)
794 return (spec == 'i' && reg >= MONO_MAX_IREGS)
795 || ((spec == 'f' && reg >= MONO_MAX_FREGS) && !MONO_ARCH_USE_FPSTACK)
796 #ifdef MONO_ARCH_SIMD_INTRINSICS
797 || (spec == 'x' && reg >= MONO_MAX_XREGS)
802 static inline gboolean
803 reg_is_softreg (int reg, const char spec)
805 return (spec == 'i' && reg >= MONO_MAX_IREGS)
806 || (spec == 'f' && reg >= MONO_MAX_FREGS)
807 #ifdef MONO_ARCH_SIMD_INTRINSICS
808 || (spec == 'x' && reg >= MONO_MAX_XREGS)
813 static inline gboolean
814 mono_is_simd_accessor (MonoInst *ins)
816 switch (ins->opcode) {
817 #ifdef MONO_ARCH_SIMD_INTRINSICS
825 case OP_INSERTX_U1_SLOW:
826 case OP_INSERTX_I4_SLOW:
827 case OP_INSERTX_R4_SLOW:
828 case OP_INSERTX_R8_SLOW:
829 case OP_INSERTX_I8_SLOW:
840 * Get rid of the dead assignments to local vregs like the ones created by the
844 mono_local_deadce (MonoCompile *cfg)
847 MonoInst *ins, *prev;
848 MonoBitSet *used, *defined;
850 //mono_print_code (cfg, "BEFORE LOCAL-DEADCE");
853 * Assignments to global vregs can't be eliminated so this pass must come
854 * after the handle_global_vregs () pass.
857 used = mono_bitset_mp_new_noinit (cfg->mempool, cfg->next_vreg + 1);
858 defined = mono_bitset_mp_new_noinit (cfg->mempool, cfg->next_vreg + 1);
860 /* First pass: collect liveness info */
861 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
862 /* Manually init the defs entries used by the bblock */
863 MONO_BB_FOR_EACH_INS (bb, ins) {
864 const char *spec = INS_INFO (ins->opcode);
865 int sregs [MONO_MAX_SRC_REGS];
868 if (spec [MONO_INST_DEST] != ' ') {
869 mono_bitset_clear_fast (used, ins->dreg);
870 mono_bitset_clear_fast (defined, ins->dreg);
871 #if SIZEOF_REGISTER == 4
873 mono_bitset_clear_fast (used, ins->dreg + 1);
874 mono_bitset_clear_fast (defined, ins->dreg + 1);
877 num_sregs = mono_inst_get_src_registers (ins, sregs);
878 for (i = 0; i < num_sregs; ++i) {
879 mono_bitset_clear_fast (used, sregs [i]);
880 #if SIZEOF_REGISTER == 4
881 mono_bitset_clear_fast (used, sregs [i] + 1);
887 * Make a reverse pass over the instruction list
889 MONO_BB_FOR_EACH_INS_REVERSE_SAFE (bb, prev, ins) {
890 const char *spec = INS_INFO (ins->opcode);
891 int sregs [MONO_MAX_SRC_REGS];
893 MonoInst *prev_f = mono_inst_prev (ins, FILTER_NOP | FILTER_IL_SEQ_POINT);
895 if (ins->opcode == OP_NOP) {
896 MONO_DELETE_INS (bb, ins);
900 g_assert (ins->opcode > MONO_CEE_LAST);
902 if (MONO_IS_NON_FP_MOVE (ins) && prev_f) {
907 spec2 = INS_INFO (def->opcode);
910 * Perform a limited kind of reverse copy propagation, i.e.
911 * transform B <- FOO; A <- B into A <- FOO
912 * This isn't copyprop, not deadce, but it can only be performed
913 * after handle_global_vregs () has run.
915 if (!get_vreg_to_inst (cfg, ins->sreg1) && (spec2 [MONO_INST_DEST] != ' ') && (def->dreg == ins->sreg1) && !mono_bitset_test_fast (used, ins->sreg1) && !MONO_IS_STORE_MEMBASE (def) && reg_is_softreg (ins->sreg1, spec [MONO_INST_DEST]) && !mono_is_simd_accessor (def)) {
916 if (cfg->verbose_level > 2) {
917 printf ("\tReverse copyprop in BB%d on ", bb->block_num);
918 mono_print_ins (ins);
921 def->dreg = ins->dreg;
922 MONO_DELETE_INS (bb, ins);
923 spec = INS_INFO (ins->opcode);
927 /* Enabling this on x86 could screw up the fp stack */
928 if (reg_is_softreg_no_fpstack (ins->dreg, spec [MONO_INST_DEST])) {
930 * Assignments to global vregs can only be eliminated if there is another
931 * assignment to the same vreg later in the same bblock.
933 if (!mono_bitset_test_fast (used, ins->dreg) &&
934 (!get_vreg_to_inst (cfg, ins->dreg) || (!bb->extended && !vreg_is_volatile (cfg, ins->dreg) && mono_bitset_test_fast (defined, ins->dreg))) &&
935 MONO_INS_HAS_NO_SIDE_EFFECT (ins)) {
936 /* Happens with CMOV instructions */
937 if (prev_f && prev_f->opcode == OP_ICOMPARE_IMM) {
938 MonoInst *prev = prev_f;
940 * Can't use DELETE_INS since that would interfere with the
945 //printf ("DEADCE: "); mono_print_ins (ins);
946 MONO_DELETE_INS (bb, ins);
947 spec = INS_INFO (ins->opcode);
950 if (spec [MONO_INST_DEST] != ' ')
951 mono_bitset_clear_fast (used, ins->dreg);
954 if (spec [MONO_INST_DEST] != ' ')
955 mono_bitset_set_fast (defined, ins->dreg);
956 num_sregs = mono_inst_get_src_registers (ins, sregs);
957 for (i = 0; i < num_sregs; ++i)
958 mono_bitset_set_fast (used, sregs [i]);
959 if (MONO_IS_STORE_MEMBASE (ins))
960 mono_bitset_set_fast (used, ins->dreg);
962 if (MONO_IS_CALL (ins)) {
963 MonoCallInst *call = (MonoCallInst*)ins;
966 if (call->out_ireg_args) {
967 for (l = call->out_ireg_args; l; l = l->next) {
968 guint32 regpair, reg;
970 regpair = (guint32)(gssize)(l->data);
971 reg = regpair & 0xffffff;
973 mono_bitset_set_fast (used, reg);
977 if (call->out_freg_args) {
978 for (l = call->out_freg_args; l; l = l->next) {
979 guint32 regpair, reg;
981 regpair = (guint32)(gssize)(l->data);
982 reg = regpair & 0xffffff;
984 mono_bitset_set_fast (used, reg);
991 //mono_print_code (cfg, "AFTER LOCAL-DEADCE");
994 #endif /* DISABLE_JIT */