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
21 #include <mono/metadata/debug-helpers.h>
22 #include <mono/metadata/mempool.h>
23 #include <mono/metadata/opcodes.h>
27 #ifndef MONO_ARCH_IS_OP_MEMBASE
28 #define MONO_ARCH_IS_OP_MEMBASE(opcode) FALSE
31 static inline MonoBitSet*
32 mono_bitset_mp_new_noinit (MonoMemPool *mp, guint32 max_size)
34 int size = mono_bitset_alloc_size (max_size, 0);
37 mem = mono_mempool_alloc (mp, size);
38 return mono_bitset_mem_new (mem, max_size, MONO_BITSET_DONT_FREE);
44 * A combined local copy and constant propagation pass.
47 mono_local_cprop (MonoCompile *cfg)
57 defs = mono_mempool_alloc (cfg->mempool, sizeof (MonoInst*) * (cfg->next_vreg + 1));
58 def_index = mono_mempool_alloc (cfg->mempool, sizeof (guint32) * (cfg->next_vreg + 1));
60 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
65 /* Manually init the defs entries used by the bblock */
66 MONO_BB_FOR_EACH_INS (bb, ins) {
67 if ((ins->dreg != -1) && (ins->dreg < max)) {
68 defs [ins->dreg] = NULL;
69 #if SIZEOF_REGISTER == 4
70 defs [ins->dreg + 1] = NULL;
73 if ((ins->sreg1 != -1) && (ins->sreg1 < max)) {
74 defs [ins->sreg1] = NULL;
75 #if SIZEOF_REGISTER == 4
76 defs [ins->sreg1 + 1] = NULL;
79 if ((ins->sreg2 != -1) && (ins->sreg2 < max)) {
80 defs [ins->sreg2] = NULL;
81 #if SIZEOF_REGISTER == 4
82 defs [ins->sreg2 + 1] = NULL;
89 MONO_BB_FOR_EACH_INS (bb, ins) {
90 const char *spec = INS_INFO (ins->opcode);
91 int regtype, srcindex, sreg;
93 if (ins->opcode == OP_NOP) {
94 MONO_DELETE_INS (bb, ins);
98 g_assert (ins->opcode > MONO_CEE_LAST);
100 /* FIXME: Optimize this */
101 if (ins->opcode == OP_LDADDR) {
102 MonoInst *var = ins->inst_p0;
104 defs [var->dreg] = NULL;
106 if (!MONO_TYPE_ISSTRUCT (var->inst_vtype))
111 if (MONO_IS_STORE_MEMBASE (ins)) {
115 if ((regtype == 'i') && (sreg != -1) && defs [sreg]) {
116 MonoInst *def = defs [sreg];
118 if ((def->opcode == OP_MOVE) && (!defs [def->sreg1] || (def_index [def->sreg1] < def_index [sreg])) && !vreg_is_volatile (cfg, def->sreg1)) {
119 int vreg = def->sreg1;
120 //printf ("CCOPY: R%d -> R%d\n", sreg, vreg);
126 for (srcindex = 0; srcindex < 2; ++srcindex) {
129 regtype = srcindex == 0 ? spec [MONO_INST_SRC1] : spec [MONO_INST_SRC2];
130 sreg = srcindex == 0 ? ins->sreg1 : ins->sreg2;
132 if ((regtype == ' ') || (sreg == -1) || (!defs [sreg]))
137 /* Copy propagation */
139 * The first check makes sure the source of the copy did not change since
141 * The second check avoids volatile variables.
142 * The third check avoids copy propagating local vregs through a call,
143 * since the lvreg will be spilled
144 * The fourth check avoids copy propagating a vreg in cases where
145 * it would be eliminated anyway by reverse copy propagation later,
146 * because propagating it would create another use for it, thus making
147 * it impossible to use reverse copy propagation.
149 /* Enabling this for floats trips up the fp stack */
151 * Enabling this for floats on amd64 seems to cause a failure in
152 * basic-math.cs, most likely because it gets rid of some r8->r4
155 if (MONO_IS_MOVE (def) &&
156 (!defs [def->sreg1] || (def_index [def->sreg1] < def_index [sreg])) &&
157 !vreg_is_volatile (cfg, def->sreg1) &&
158 /* This avoids propagating local vregs across calls */
159 ((get_vreg_to_inst (cfg, def->sreg1) || !defs [def->sreg1] || (def_index [def->sreg1] >= last_call_index) || (def->opcode == OP_VMOVE))) &&
160 !(defs [def->sreg1] && defs [def->sreg1]->next == def) &&
161 (!MONO_ARCH_USE_FPSTACK || (def->opcode != OP_FMOVE)) &&
162 (def->opcode != OP_FMOVE)) {
163 int vreg = def->sreg1;
165 //printf ("CCOPY: R%d -> R%d\n", sreg, vreg);
171 /* Allow further iterations */
176 /* Constant propagation */
177 /* FIXME: Make is_inst_imm a macro */
178 /* FIXME: Make is_inst_imm take an opcode argument */
179 /* is_inst_imm is only needed for binops */
180 if ((((def->opcode == OP_ICONST) || ((sizeof (gpointer) == 8) && (def->opcode == OP_I8CONST))) &&
181 (((srcindex == 0) && (ins->sreg2 == -1)) || mono_arch_is_inst_imm (def->inst_c0))) ||
182 (!MONO_ARCH_USE_FPSTACK && (def->opcode == OP_R8CONST))) {
185 /* srcindex == 1 -> binop, ins->sreg2 == -1 -> unop */
186 if ((srcindex == 1) && (ins->sreg1 != -1) && defs [ins->sreg1] && (defs [ins->sreg1]->opcode == OP_ICONST) && defs [ins->sreg2]) {
187 /* Both arguments are constants, perform cfold */
188 mono_constant_fold_ins (cfg, ins, defs [ins->sreg1], defs [ins->sreg2], TRUE);
189 } else if ((srcindex == 0) && (ins->sreg2 != -1) && defs [ins->sreg2]) {
190 /* Arg 1 is constant, swap arguments if possible */
191 int opcode = ins->opcode;
192 mono_constant_fold_ins (cfg, ins, defs [ins->sreg1], defs [ins->sreg2], TRUE);
193 if (ins->opcode != opcode) {
194 /* Allow further iterations */
198 } else if ((srcindex == 0) && (ins->sreg2 == -1)) {
199 /* Constant unop, perform cfold */
200 mono_constant_fold_ins (cfg, ins, defs [ins->sreg1], NULL, TRUE);
203 opcode2 = mono_op_to_op_imm (ins->opcode);
204 if ((opcode2 != -1) && mono_arch_is_inst_imm (def->inst_c0) && ((srcindex == 1) || (ins->sreg2 == -1))) {
205 ins->opcode = opcode2;
206 if ((def->opcode == OP_I8CONST) && (sizeof (gpointer) == 4)) {
207 ins->inst_ls_word = def->inst_ls_word;
208 ins->inst_ms_word = def->inst_ms_word;
210 ins->inst_imm = def->inst_c0;
217 if ((opcode2 == OP_VOIDCALL) || (opcode2 == OP_CALL) || (opcode2 == OP_LCALL) || (opcode2 == OP_FCALL))
218 ((MonoCallInst*)ins)->fptr = (gpointer)ins->inst_imm;
220 /* Allow further iterations */
226 #if defined(__i386__) || defined(__x86_64__)
227 if ((ins->opcode == OP_X86_LEA) && (srcindex == 1)) {
228 #if SIZEOF_REGISTER == 8
229 /* FIXME: Use OP_PADD_IMM when the new JIT is done */
230 ins->opcode = OP_LADD_IMM;
232 ins->opcode = OP_ADD_IMM;
234 ins->inst_imm += def->inst_c0 << ins->backend.shift_amount;
238 opcode2 = mono_load_membase_to_load_mem (ins->opcode);
239 if ((srcindex == 0) && (opcode2 != -1) && mono_arch_is_inst_imm (def->inst_c0)) {
240 ins->opcode = opcode2;
241 ins->inst_imm = def->inst_c0 + ins->inst_offset;
246 else if (((def->opcode == OP_ADD_IMM) || (def->opcode == OP_LADD_IMM)) && (MONO_IS_LOAD_MEMBASE (ins) || MONO_ARCH_IS_OP_MEMBASE (ins->opcode))) {
247 /* ADD_IMM is created by spill_global_vars */
249 * We have to guarantee that def->sreg1 haven't changed since def->dreg
250 * was defined. cfg->frame_reg is assumed to remain constant.
252 if ((def->sreg1 == cfg->frame_reg) || ((def->next == ins) && (def->dreg != def->sreg1))) {
253 ins->inst_basereg = def->sreg1;
254 ins->inst_offset += def->inst_imm;
256 } else if ((ins->opcode == OP_ISUB_IMM) && (def->opcode == OP_IADD_IMM) && (def->next == ins)) {
257 ins->sreg1 = def->sreg1;
258 ins->inst_imm -= def->inst_imm;
259 } else if ((ins->opcode == OP_IADD_IMM) && (def->opcode == OP_ISUB_IMM) && (def->next == ins)) {
260 ins->sreg1 = def->sreg1;
261 ins->inst_imm -= def->inst_imm;
262 } else if (ins->opcode == OP_STOREI1_MEMBASE_REG &&
263 (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)) &&
264 (!defs [def->sreg1] || (def_index [def->sreg1] < def_index [sreg]))) {
265 /* Avoid needless sign extension */
266 ins->sreg1 = def->sreg1;
267 } else if (ins->opcode == OP_STOREI2_MEMBASE_REG &&
268 (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)) &&
269 (!defs [def->sreg1] || (def_index [def->sreg1] < def_index [sreg]))) {
270 /* Avoid needless sign extension */
271 ins->sreg1 = def->sreg1;
275 /* Do strength reduction here */
276 /* FIXME: Add long/float */
277 switch (ins->opcode) {
280 if (ins->dreg == ins->sreg1) {
281 MONO_DELETE_INS (bb, ins);
282 spec = INS_INFO (ins->opcode);
289 #if SIZEOF_REGISTER == 8
293 if (ins->inst_imm == 0) {
294 ins->opcode = OP_MOVE;
295 spec = INS_INFO (ins->opcode);
300 #if SIZEOF_REGISTER == 8
303 if (ins->inst_imm == 0) {
304 ins->opcode = (ins->opcode == OP_LMUL_IMM) ? OP_I8CONST : OP_ICONST;
307 } else if (ins->inst_imm == 1) {
308 ins->opcode = OP_MOVE;
309 } else if ((ins->opcode == OP_IMUL_IMM) && (ins->inst_imm == -1)) {
310 ins->opcode = OP_INEG;
311 } else if ((ins->opcode == OP_LMUL_IMM) && (ins->inst_imm == -1)) {
312 ins->opcode = OP_LNEG;
314 int power2 = mono_is_power_of_two (ins->inst_imm);
316 ins->opcode = (ins->opcode == OP_MUL_IMM) ? OP_SHL_IMM : ((ins->opcode == OP_LMUL_IMM) ? OP_LSHL_IMM : OP_ISHL_IMM);
317 ins->inst_imm = power2;
320 spec = INS_INFO (ins->opcode);
323 case OP_IDIV_UN_IMM: {
324 int c = ins->inst_imm;
325 int power2 = mono_is_power_of_two (c);
328 if (ins->opcode == OP_IREM_UN_IMM) {
329 ins->opcode = OP_IAND_IMM;
331 ins->inst_imm = (1 << power2) - 1;
332 } else if (ins->opcode == OP_IDIV_UN_IMM) {
333 ins->opcode = OP_ISHR_UN_IMM;
335 ins->inst_imm = power2;
338 spec = INS_INFO (ins->opcode);
342 int c = ins->inst_imm;
343 int power2 = mono_is_power_of_two (c);
344 MonoInst *tmp1, *tmp2, *tmp3, *tmp4;
346 /* FIXME: Move this elsewhere cause its hard to implement it here */
348 int r1 = mono_alloc_ireg (cfg);
350 NEW_BIALU_IMM (cfg, tmp1, OP_ISHR_UN_IMM, r1, ins->sreg1, 31);
351 mono_bblock_insert_after_ins (bb, ins, tmp1);
352 NEW_BIALU (cfg, tmp2, OP_IADD, r1, r1, ins->sreg1);
353 mono_bblock_insert_after_ins (bb, tmp1, tmp2);
354 NEW_BIALU_IMM (cfg, tmp3, OP_ISHR_IMM, ins->dreg, r1, 1);
355 mono_bblock_insert_after_ins (bb, tmp2, tmp3);
359 // We allocated a new vreg, so need to restart
361 } else if (power2 > 0) {
362 int r1 = mono_alloc_ireg (cfg);
364 NEW_BIALU_IMM (cfg, tmp1, OP_ISHR_IMM, r1, ins->sreg1, 31);
365 mono_bblock_insert_after_ins (bb, ins, tmp1);
366 NEW_BIALU_IMM (cfg, tmp2, OP_ISHR_UN_IMM, r1, r1, (32 - power2));
367 mono_bblock_insert_after_ins (bb, tmp1, tmp2);
368 NEW_BIALU (cfg, tmp3, OP_IADD, r1, r1, ins->sreg1);
369 mono_bblock_insert_after_ins (bb, tmp2, tmp3);
370 NEW_BIALU_IMM (cfg, tmp4, OP_ISHR_IMM, ins->dreg, r1, power2);
371 mono_bblock_insert_after_ins (bb, tmp3, tmp4);
375 // We allocated a new vreg, so need to restart
382 if (spec [MONO_INST_DEST] != ' ') {
383 MonoInst *def = defs [ins->dreg];
385 if (def && (def->opcode == OP_ADD_IMM) && (def->sreg1 == cfg->frame_reg) && (MONO_IS_STORE_MEMBASE (ins))) {
386 /* ADD_IMM is created by spill_global_vars */
387 /* cfg->frame_reg is assumed to remain constant */
388 ins->inst_destbasereg = def->sreg1;
389 ins->inst_offset += def->inst_imm;
393 if ((spec [MONO_INST_DEST] != ' ') && !MONO_IS_STORE_MEMBASE (ins) && !vreg_is_volatile (cfg, ins->dreg)) {
394 defs [ins->dreg] = ins;
395 def_index [ins->dreg] = ins_index;
398 if (MONO_IS_CALL (ins))
399 last_call_index = ins_index;
406 static inline gboolean
407 reg_is_softreg_no_fpstack (int reg, const char spec)
409 return (spec == 'i' && reg > MONO_MAX_IREGS)
410 || ((spec == 'f' && reg > MONO_MAX_FREGS) && !MONO_ARCH_USE_FPSTACK)
411 #ifdef MONO_ARCH_SIMD_INTRINSICS
412 || (spec == 'x' && reg > MONO_MAX_XREGS)
417 static inline gboolean
418 reg_is_softreg (int reg, const char spec)
420 return (spec == 'i' && reg > MONO_MAX_IREGS)
421 || (spec == 'f' && reg > MONO_MAX_FREGS)
422 #ifdef MONO_ARCH_SIMD_INTRINSICS
423 || (spec == 'x' && reg > MONO_MAX_XREGS)
431 * Get rid of the dead assignments to local vregs like the ones created by the
435 mono_local_deadce (MonoCompile *cfg)
438 MonoInst *ins, *prev;
439 MonoBitSet *used, *defined;
441 //mono_print_code (cfg, "BEFORE LOCAL-DEADCE");
444 * Assignments to global vregs can't be eliminated so this pass must come
445 * after the handle_global_vregs () pass.
448 used = mono_bitset_mp_new_noinit (cfg->mempool, cfg->next_vreg + 1);
449 defined = mono_bitset_mp_new_noinit (cfg->mempool, cfg->next_vreg + 1);
451 /* First pass: collect liveness info */
452 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
453 /* Manually init the defs entries used by the bblock */
454 MONO_BB_FOR_EACH_INS (bb, ins) {
455 const char *spec = INS_INFO (ins->opcode);
457 if (spec [MONO_INST_DEST] != ' ') {
458 mono_bitset_clear_fast (used, ins->dreg);
459 mono_bitset_clear_fast (defined, ins->dreg);
460 #if SIZEOF_REGISTER == 4
462 mono_bitset_clear_fast (used, ins->dreg + 1);
463 mono_bitset_clear_fast (defined, ins->dreg + 1);
466 if (spec [MONO_INST_SRC1] != ' ') {
467 mono_bitset_clear_fast (used, ins->sreg1);
468 #if SIZEOF_REGISTER == 4
469 mono_bitset_clear_fast (used, ins->sreg1 + 1);
472 if (spec [MONO_INST_SRC2] != ' ') {
473 mono_bitset_clear_fast (used, ins->sreg2);
474 #if SIZEOF_REGISTER == 4
475 mono_bitset_clear_fast (used, ins->sreg2 + 1);
481 * Make a reverse pass over the instruction list
483 MONO_BB_FOR_EACH_INS_REVERSE_SAFE (bb, prev, ins) {
484 const char *spec = INS_INFO (ins->opcode);
486 if (ins->opcode == OP_NOP) {
487 MONO_DELETE_INS (bb, ins);
491 g_assert (ins->opcode > MONO_CEE_LAST);
493 if (MONO_IS_NON_FP_MOVE (ins) && ins->prev) {
498 while (def->prev && (def->opcode == OP_NOP))
500 spec2 = INS_INFO (def->opcode);
503 * Perform a limited kind of reverse copy propagation, i.e.
504 * transform B <- FOO; A <- B into A <- FOO
505 * This isn't copyprop, not deadce, but it can only be performed
506 * after handle_global_vregs () has run.
508 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])) {
509 if (cfg->verbose_level > 2) {
510 printf ("\tReverse copyprop in BB%d on ", bb->block_num);
511 mono_print_ins (ins);
514 def->dreg = ins->dreg;
515 MONO_DELETE_INS (bb, ins);
516 spec = INS_INFO (ins->opcode);
520 /* Enabling this on x86 could screw up the fp stack */
521 if (reg_is_softreg_no_fpstack (ins->dreg, spec [MONO_INST_DEST])) {
523 * Assignments to global vregs can only be eliminated if there is another
524 * assignment to the same vreg later in the same bblock.
526 if (!mono_bitset_test_fast (used, ins->dreg) &&
527 (!get_vreg_to_inst (cfg, ins->dreg) || (!bb->extended && !vreg_is_volatile (cfg, ins->dreg) && mono_bitset_test_fast (defined, ins->dreg))) &&
528 MONO_INS_HAS_NO_SIDE_EFFECT (ins)) {
529 /* Happens with CMOV instructions */
530 if (ins->prev && ins->prev->opcode == OP_ICOMPARE_IMM) {
531 MonoInst *prev = ins->prev;
533 * Can't use DELETE_INS since that would interfere with the
538 //printf ("DEADCE: "); mono_print_ins (ins);
539 MONO_DELETE_INS (bb, ins);
540 spec = INS_INFO (ins->opcode);
543 if (spec [MONO_INST_DEST] != ' ')
544 mono_bitset_clear_fast (used, ins->dreg);
547 if (spec [MONO_INST_DEST] != ' ')
548 mono_bitset_set_fast (defined, ins->dreg);
549 if (spec [MONO_INST_SRC1] != ' ')
550 mono_bitset_set_fast (used, ins->sreg1);
551 if (spec [MONO_INST_SRC2] != ' ')
552 mono_bitset_set_fast (used, ins->sreg2);
553 if (MONO_IS_STORE_MEMBASE (ins))
554 mono_bitset_set_fast (used, ins->dreg);
556 if (MONO_IS_CALL (ins)) {
557 MonoCallInst *call = (MonoCallInst*)ins;
560 if (call->out_ireg_args) {
561 for (l = call->out_ireg_args; l; l = l->next) {
562 guint32 regpair, reg;
564 regpair = (guint32)(gssize)(l->data);
565 reg = regpair & 0xffffff;
567 mono_bitset_set_fast (used, reg);
571 if (call->out_freg_args) {
572 for (l = call->out_freg_args; l; l = l->next) {
573 guint32 regpair, reg;
575 regpair = (guint32)(gssize)(l->data);
576 reg = regpair & 0xffffff;
578 mono_bitset_set_fast (used, reg);
585 //mono_print_code (cfg, "AFTER LOCAL-DEADCE");