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 int sregs [MONO_MAX_SRC_REGS];
70 if ((ins->dreg != -1) && (ins->dreg < max)) {
71 defs [ins->dreg] = NULL;
72 #if SIZEOF_REGISTER == 4
73 defs [ins->dreg + 1] = NULL;
77 num_sregs = mono_inst_get_src_registers (ins, sregs);
78 for (i = 0; i < num_sregs; ++i) {
82 #if SIZEOF_REGISTER == 4
83 defs [sreg + 1] = NULL;
91 MONO_BB_FOR_EACH_INS (bb, ins) {
92 const char *spec = INS_INFO (ins->opcode);
93 int regtype, srcindex, sreg;
95 int sregs [MONO_MAX_SRC_REGS];
97 if (ins->opcode == OP_NOP) {
98 MONO_DELETE_INS (bb, ins);
102 g_assert (ins->opcode > MONO_CEE_LAST);
104 /* FIXME: Optimize this */
105 if (ins->opcode == OP_LDADDR) {
106 MonoInst *var = ins->inst_p0;
108 defs [var->dreg] = NULL;
110 if (!MONO_TYPE_ISSTRUCT (var->inst_vtype))
115 if (MONO_IS_STORE_MEMBASE (ins)) {
119 if ((regtype == 'i') && (sreg != -1) && defs [sreg]) {
120 MonoInst *def = defs [sreg];
122 if ((def->opcode == OP_MOVE) && (!defs [def->sreg1] || (def_index [def->sreg1] < def_index [sreg])) && !vreg_is_volatile (cfg, def->sreg1)) {
123 int vreg = def->sreg1;
124 if (cfg->verbose_level > 2) printf ("CCOPY: R%d -> R%d\n", sreg, vreg);
130 num_sregs = mono_inst_get_src_registers (ins, sregs);
131 for (srcindex = 0; srcindex < num_sregs; ++srcindex) {
134 regtype = spec [MONO_INST_SRC1 + srcindex];
135 sreg = sregs [srcindex];
137 if ((regtype == ' ') || (sreg == -1) || (!defs [sreg]))
142 /* Copy propagation */
144 * The first check makes sure the source of the copy did not change since
146 * The second check avoids volatile variables.
147 * The third check avoids copy propagating local vregs through a call,
148 * since the lvreg will be spilled
149 * The fourth check avoids copy propagating a vreg in cases where
150 * it would be eliminated anyway by reverse copy propagation later,
151 * because propagating it would create another use for it, thus making
152 * it impossible to use reverse copy propagation.
154 /* Enabling this for floats trips up the fp stack */
156 * Enabling this for floats on amd64 seems to cause a failure in
157 * basic-math.cs, most likely because it gets rid of some r8->r4
160 if (MONO_IS_MOVE (def) &&
161 (!defs [def->sreg1] || (def_index [def->sreg1] < def_index [sreg])) &&
162 !vreg_is_volatile (cfg, def->sreg1) &&
163 /* This avoids propagating local vregs across calls */
164 ((get_vreg_to_inst (cfg, def->sreg1) || !defs [def->sreg1] || (def_index [def->sreg1] >= last_call_index) || (def->opcode == OP_VMOVE))) &&
165 !(defs [def->sreg1] && defs [def->sreg1]->next == def) &&
166 (!MONO_ARCH_USE_FPSTACK || (def->opcode != OP_FMOVE)) &&
167 (def->opcode != OP_FMOVE)) {
168 int vreg = def->sreg1;
170 if (cfg->verbose_level > 2) printf ("CCOPY/2: R%d -> R%d\n", sreg, vreg);
171 sregs [srcindex] = vreg;
172 mono_inst_set_src_registers (ins, sregs);
174 /* Allow further iterations */
179 /* Constant propagation */
180 /* FIXME: Make is_inst_imm a macro */
181 /* FIXME: Make is_inst_imm take an opcode argument */
182 /* is_inst_imm is only needed for binops */
183 if ((((def->opcode == OP_ICONST) || ((sizeof (gpointer) == 8) && (def->opcode == OP_I8CONST))) &&
184 (((srcindex == 0) && (ins->sreg2 == -1)) || mono_arch_is_inst_imm (def->inst_c0))) ||
185 (!MONO_ARCH_USE_FPSTACK && (def->opcode == OP_R8CONST))) {
188 /* srcindex == 1 -> binop, ins->sreg2 == -1 -> unop */
189 if ((srcindex == 1) && (ins->sreg1 != -1) && defs [ins->sreg1] && (defs [ins->sreg1]->opcode == OP_ICONST) && defs [ins->sreg2]) {
190 /* Both arguments are constants, perform cfold */
191 mono_constant_fold_ins (cfg, ins, defs [ins->sreg1], defs [ins->sreg2], TRUE);
192 } else if ((srcindex == 0) && (ins->sreg2 != -1) && defs [ins->sreg2]) {
193 /* Arg 1 is constant, swap arguments if possible */
194 int opcode = ins->opcode;
195 mono_constant_fold_ins (cfg, ins, defs [ins->sreg1], defs [ins->sreg2], TRUE);
196 if (ins->opcode != opcode) {
197 /* Allow further iterations */
201 } else if ((srcindex == 0) && (ins->sreg2 == -1)) {
202 /* Constant unop, perform cfold */
203 mono_constant_fold_ins (cfg, ins, defs [ins->sreg1], NULL, TRUE);
206 opcode2 = mono_op_to_op_imm (ins->opcode);
207 if ((opcode2 != -1) && mono_arch_is_inst_imm (def->inst_c0) && ((srcindex == 1) || (ins->sreg2 == -1))) {
208 ins->opcode = opcode2;
209 if ((def->opcode == OP_I8CONST) && (sizeof (gpointer) == 4)) {
210 ins->inst_ls_word = def->inst_ls_word;
211 ins->inst_ms_word = def->inst_ms_word;
213 ins->inst_imm = def->inst_c0;
215 sregs [srcindex] = -1;
216 mono_inst_set_src_registers (ins, sregs);
218 if ((opcode2 == OP_VOIDCALL) || (opcode2 == OP_CALL) || (opcode2 == OP_LCALL) || (opcode2 == OP_FCALL))
219 ((MonoCallInst*)ins)->fptr = (gpointer)ins->inst_imm;
221 /* Allow further iterations */
227 #if defined(TARGET_X86) || defined(TARGET_AMD64)
228 if ((ins->opcode == OP_X86_LEA) && (srcindex == 1)) {
229 #if SIZEOF_REGISTER == 8
230 /* FIXME: Use OP_PADD_IMM when the new JIT is done */
231 ins->opcode = OP_LADD_IMM;
233 ins->opcode = OP_ADD_IMM;
235 ins->inst_imm += def->inst_c0 << ins->backend.shift_amount;
239 opcode2 = mono_load_membase_to_load_mem (ins->opcode);
240 if ((srcindex == 0) && (opcode2 != -1) && mono_arch_is_inst_imm (def->inst_c0)) {
241 ins->opcode = opcode2;
242 ins->inst_imm = def->inst_c0 + ins->inst_offset;
247 else if (((def->opcode == OP_ADD_IMM) || (def->opcode == OP_LADD_IMM)) && (MONO_IS_LOAD_MEMBASE (ins) || MONO_ARCH_IS_OP_MEMBASE (ins->opcode))) {
248 /* ADD_IMM is created by spill_global_vars */
250 * We have to guarantee that def->sreg1 haven't changed since def->dreg
251 * was defined. cfg->frame_reg is assumed to remain constant.
253 if ((def->sreg1 == cfg->frame_reg) || ((def->next == ins) && (def->dreg != def->sreg1))) {
254 ins->inst_basereg = def->sreg1;
255 ins->inst_offset += def->inst_imm;
257 } else if ((ins->opcode == OP_ISUB_IMM) && (def->opcode == OP_IADD_IMM) && (def->next == ins) && (def->dreg != def->sreg1)) {
258 ins->sreg1 = def->sreg1;
259 ins->inst_imm -= def->inst_imm;
260 } else if ((ins->opcode == OP_IADD_IMM) && (def->opcode == OP_ISUB_IMM) && (def->next == ins) && (def->dreg != def->sreg1)) {
261 ins->sreg1 = def->sreg1;
262 ins->inst_imm -= def->inst_imm;
263 } else if (ins->opcode == OP_STOREI1_MEMBASE_REG &&
264 (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)) &&
265 (!defs [def->sreg1] || (def_index [def->sreg1] < def_index [sreg]))) {
266 /* Avoid needless sign extension */
267 ins->sreg1 = def->sreg1;
268 } else if (ins->opcode == OP_STOREI2_MEMBASE_REG &&
269 (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)) &&
270 (!defs [def->sreg1] || (def_index [def->sreg1] < def_index [sreg]))) {
271 /* Avoid needless sign extension */
272 ins->sreg1 = def->sreg1;
276 /* Do strength reduction here */
277 /* FIXME: Add long/float */
278 switch (ins->opcode) {
281 if (ins->dreg == ins->sreg1) {
282 MONO_DELETE_INS (bb, ins);
283 spec = INS_INFO (ins->opcode);
290 #if SIZEOF_REGISTER == 8
294 if (ins->inst_imm == 0) {
295 ins->opcode = OP_MOVE;
296 spec = INS_INFO (ins->opcode);
301 #if SIZEOF_REGISTER == 8
304 if (ins->inst_imm == 0) {
305 ins->opcode = (ins->opcode == OP_LMUL_IMM) ? OP_I8CONST : OP_ICONST;
308 } else if (ins->inst_imm == 1) {
309 ins->opcode = OP_MOVE;
310 } else if ((ins->opcode == OP_IMUL_IMM) && (ins->inst_imm == -1)) {
311 ins->opcode = OP_INEG;
312 } else if ((ins->opcode == OP_LMUL_IMM) && (ins->inst_imm == -1)) {
313 ins->opcode = OP_LNEG;
315 int power2 = mono_is_power_of_two (ins->inst_imm);
317 ins->opcode = (ins->opcode == OP_MUL_IMM) ? OP_SHL_IMM : ((ins->opcode == OP_LMUL_IMM) ? OP_LSHL_IMM : OP_ISHL_IMM);
318 ins->inst_imm = power2;
321 spec = INS_INFO (ins->opcode);
324 case OP_IDIV_UN_IMM: {
325 int c = ins->inst_imm;
326 int power2 = mono_is_power_of_two (c);
329 if (ins->opcode == OP_IREM_UN_IMM) {
330 ins->opcode = OP_IAND_IMM;
332 ins->inst_imm = (1 << power2) - 1;
333 } else if (ins->opcode == OP_IDIV_UN_IMM) {
334 ins->opcode = OP_ISHR_UN_IMM;
336 ins->inst_imm = power2;
339 spec = INS_INFO (ins->opcode);
343 int c = ins->inst_imm;
344 int power2 = mono_is_power_of_two (c);
345 MonoInst *tmp1, *tmp2, *tmp3, *tmp4;
347 /* FIXME: Move this elsewhere cause its hard to implement it here */
349 int r1 = mono_alloc_ireg (cfg);
351 NEW_BIALU_IMM (cfg, tmp1, OP_ISHR_UN_IMM, r1, ins->sreg1, 31);
352 mono_bblock_insert_after_ins (bb, ins, tmp1);
353 NEW_BIALU (cfg, tmp2, OP_IADD, r1, r1, ins->sreg1);
354 mono_bblock_insert_after_ins (bb, tmp1, tmp2);
355 NEW_BIALU_IMM (cfg, tmp3, OP_ISHR_IMM, ins->dreg, r1, 1);
356 mono_bblock_insert_after_ins (bb, tmp2, tmp3);
360 // We allocated a new vreg, so need to restart
362 } else if (power2 > 0) {
363 int r1 = mono_alloc_ireg (cfg);
365 NEW_BIALU_IMM (cfg, tmp1, OP_ISHR_IMM, r1, ins->sreg1, 31);
366 mono_bblock_insert_after_ins (bb, ins, tmp1);
367 NEW_BIALU_IMM (cfg, tmp2, OP_ISHR_UN_IMM, r1, r1, (32 - power2));
368 mono_bblock_insert_after_ins (bb, tmp1, tmp2);
369 NEW_BIALU (cfg, tmp3, OP_IADD, r1, r1, ins->sreg1);
370 mono_bblock_insert_after_ins (bb, tmp2, tmp3);
371 NEW_BIALU_IMM (cfg, tmp4, OP_ISHR_IMM, ins->dreg, r1, power2);
372 mono_bblock_insert_after_ins (bb, tmp3, tmp4);
376 // We allocated a new vreg, so need to restart
383 if (spec [MONO_INST_DEST] != ' ') {
384 MonoInst *def = defs [ins->dreg];
386 if (def && (def->opcode == OP_ADD_IMM) && (def->sreg1 == cfg->frame_reg) && (MONO_IS_STORE_MEMBASE (ins))) {
387 /* ADD_IMM is created by spill_global_vars */
388 /* cfg->frame_reg is assumed to remain constant */
389 ins->inst_destbasereg = def->sreg1;
390 ins->inst_offset += def->inst_imm;
394 if ((spec [MONO_INST_DEST] != ' ') && !MONO_IS_STORE_MEMBASE (ins) && !vreg_is_volatile (cfg, ins->dreg)) {
395 defs [ins->dreg] = ins;
396 def_index [ins->dreg] = ins_index;
399 if (MONO_IS_CALL (ins))
400 last_call_index = ins_index;
407 static inline gboolean
408 reg_is_softreg_no_fpstack (int reg, const char spec)
410 return (spec == 'i' && reg >= MONO_MAX_IREGS)
411 || ((spec == 'f' && reg >= MONO_MAX_FREGS) && !MONO_ARCH_USE_FPSTACK)
412 #ifdef MONO_ARCH_SIMD_INTRINSICS
413 || (spec == 'x' && reg >= MONO_MAX_XREGS)
418 static inline gboolean
419 reg_is_softreg (int reg, const char spec)
421 return (spec == 'i' && reg >= MONO_MAX_IREGS)
422 || (spec == 'f' && reg >= MONO_MAX_FREGS)
423 #ifdef MONO_ARCH_SIMD_INTRINSICS
424 || (spec == 'x' && reg >= MONO_MAX_XREGS)
429 static inline gboolean
430 mono_is_simd_accessor (MonoInst *ins)
432 switch (ins->opcode) {
433 #ifdef MONO_ARCH_SIMD_INTRINSICS
441 case OP_INSERTX_U1_SLOW:
442 case OP_INSERTX_I4_SLOW:
443 case OP_INSERTX_R4_SLOW:
444 case OP_INSERTX_R8_SLOW:
445 case OP_INSERTX_I8_SLOW:
456 * Get rid of the dead assignments to local vregs like the ones created by the
460 mono_local_deadce (MonoCompile *cfg)
463 MonoInst *ins, *prev;
464 MonoBitSet *used, *defined;
466 //mono_print_code (cfg, "BEFORE LOCAL-DEADCE");
469 * Assignments to global vregs can't be eliminated so this pass must come
470 * after the handle_global_vregs () pass.
473 used = mono_bitset_mp_new_noinit (cfg->mempool, cfg->next_vreg + 1);
474 defined = mono_bitset_mp_new_noinit (cfg->mempool, cfg->next_vreg + 1);
476 /* First pass: collect liveness info */
477 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
478 /* Manually init the defs entries used by the bblock */
479 MONO_BB_FOR_EACH_INS (bb, ins) {
480 const char *spec = INS_INFO (ins->opcode);
481 int sregs [MONO_MAX_SRC_REGS];
484 if (spec [MONO_INST_DEST] != ' ') {
485 mono_bitset_clear_fast (used, ins->dreg);
486 mono_bitset_clear_fast (defined, ins->dreg);
487 #if SIZEOF_REGISTER == 4
489 mono_bitset_clear_fast (used, ins->dreg + 1);
490 mono_bitset_clear_fast (defined, ins->dreg + 1);
493 num_sregs = mono_inst_get_src_registers (ins, sregs);
494 for (i = 0; i < num_sregs; ++i) {
495 mono_bitset_clear_fast (used, sregs [i]);
496 #if SIZEOF_REGISTER == 4
497 mono_bitset_clear_fast (used, sregs [i] + 1);
503 * Make a reverse pass over the instruction list
505 MONO_BB_FOR_EACH_INS_REVERSE_SAFE (bb, prev, ins) {
506 const char *spec = INS_INFO (ins->opcode);
507 int sregs [MONO_MAX_SRC_REGS];
510 if (ins->opcode == OP_NOP) {
511 MONO_DELETE_INS (bb, ins);
515 g_assert (ins->opcode > MONO_CEE_LAST);
517 if (MONO_IS_NON_FP_MOVE (ins) && ins->prev) {
522 while (def->prev && (def->opcode == OP_NOP))
524 spec2 = INS_INFO (def->opcode);
527 * Perform a limited kind of reverse copy propagation, i.e.
528 * transform B <- FOO; A <- B into A <- FOO
529 * This isn't copyprop, not deadce, but it can only be performed
530 * after handle_global_vregs () has run.
532 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)) {
533 if (cfg->verbose_level > 2) {
534 printf ("\tReverse copyprop in BB%d on ", bb->block_num);
535 mono_print_ins (ins);
538 def->dreg = ins->dreg;
539 MONO_DELETE_INS (bb, ins);
540 spec = INS_INFO (ins->opcode);
544 /* Enabling this on x86 could screw up the fp stack */
545 if (reg_is_softreg_no_fpstack (ins->dreg, spec [MONO_INST_DEST])) {
547 * Assignments to global vregs can only be eliminated if there is another
548 * assignment to the same vreg later in the same bblock.
550 if (!mono_bitset_test_fast (used, ins->dreg) &&
551 (!get_vreg_to_inst (cfg, ins->dreg) || (!bb->extended && !vreg_is_volatile (cfg, ins->dreg) && mono_bitset_test_fast (defined, ins->dreg))) &&
552 MONO_INS_HAS_NO_SIDE_EFFECT (ins)) {
553 /* Happens with CMOV instructions */
554 if (ins->prev && ins->prev->opcode == OP_ICOMPARE_IMM) {
555 MonoInst *prev = ins->prev;
557 * Can't use DELETE_INS since that would interfere with the
562 //printf ("DEADCE: "); mono_print_ins (ins);
563 MONO_DELETE_INS (bb, ins);
564 spec = INS_INFO (ins->opcode);
567 if (spec [MONO_INST_DEST] != ' ')
568 mono_bitset_clear_fast (used, ins->dreg);
571 if (spec [MONO_INST_DEST] != ' ')
572 mono_bitset_set_fast (defined, ins->dreg);
573 num_sregs = mono_inst_get_src_registers (ins, sregs);
574 for (i = 0; i < num_sregs; ++i)
575 mono_bitset_set_fast (used, sregs [i]);
576 if (MONO_IS_STORE_MEMBASE (ins))
577 mono_bitset_set_fast (used, ins->dreg);
579 if (MONO_IS_CALL (ins)) {
580 MonoCallInst *call = (MonoCallInst*)ins;
583 if (call->out_ireg_args) {
584 for (l = call->out_ireg_args; l; l = l->next) {
585 guint32 regpair, reg;
587 regpair = (guint32)(gssize)(l->data);
588 reg = regpair & 0xffffff;
590 mono_bitset_set_fast (used, reg);
594 if (call->out_freg_args) {
595 for (l = call->out_freg_args; l; l = l->next) {
596 guint32 regpair, reg;
598 regpair = (guint32)(gssize)(l->data);
599 reg = regpair & 0xffffff;
601 mono_bitset_set_fast (used, reg);
608 //mono_print_code (cfg, "AFTER LOCAL-DEADCE");