xref: /openbmc/qemu/tcg/optimize.c (revision 59379a45af1f4d62fc8c1ae0ddee988f47075787)
1 /*
2  * Optimizations for Tiny Code Generator for QEMU
3  *
4  * Copyright (c) 2010 Samsung Electronics.
5  * Contributed by Kirill Batuzov <batuzovk@ispras.ru>
6  *
7  * Permission is hereby granted, free of charge, to any person obtaining a copy
8  * of this software and associated documentation files (the "Software"), to deal
9  * in the Software without restriction, including without limitation the rights
10  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
11  * copies of the Software, and to permit persons to whom the Software is
12  * furnished to do so, subject to the following conditions:
13  *
14  * The above copyright notice and this permission notice shall be included in
15  * all copies or substantial portions of the Software.
16  *
17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
23  * THE SOFTWARE.
24  */
25 
26 #include "qemu/osdep.h"
27 #include "qemu/int128.h"
28 #include "qemu/interval-tree.h"
29 #include "tcg/tcg-op-common.h"
30 #include "tcg-internal.h"
31 #include "tcg-has.h"
32 
33 #define CASE_OP_32_64(x)                        \
34         glue(glue(case INDEX_op_, x), _i32):    \
35         glue(glue(case INDEX_op_, x), _i64)
36 
37 #define CASE_OP_32_64_VEC(x)                    \
38         glue(glue(case INDEX_op_, x), _i32):    \
39         glue(glue(case INDEX_op_, x), _i64):    \
40         glue(glue(case INDEX_op_, x), _vec)
41 
42 typedef struct MemCopyInfo {
43     IntervalTreeNode itree;
44     QSIMPLEQ_ENTRY (MemCopyInfo) next;
45     TCGTemp *ts;
46     TCGType type;
47 } MemCopyInfo;
48 
49 typedef struct TempOptInfo {
50     bool is_const;
51     TCGTemp *prev_copy;
52     TCGTemp *next_copy;
53     QSIMPLEQ_HEAD(, MemCopyInfo) mem_copy;
54     uint64_t val;
55     uint64_t z_mask;  /* mask bit is 0 if and only if value bit is 0 */
56     uint64_t s_mask;  /* mask bit is 1 if value bit matches msb */
57 } TempOptInfo;
58 
59 typedef struct OptContext {
60     TCGContext *tcg;
61     TCGOp *prev_mb;
62     TCGTempSet temps_used;
63 
64     IntervalTreeRoot mem_copy;
65     QSIMPLEQ_HEAD(, MemCopyInfo) mem_free;
66 
67     /* In flight values from optimization. */
68     TCGType type;
69 } OptContext;
70 
71 static inline TempOptInfo *ts_info(TCGTemp *ts)
72 {
73     return ts->state_ptr;
74 }
75 
76 static inline TempOptInfo *arg_info(TCGArg arg)
77 {
78     return ts_info(arg_temp(arg));
79 }
80 
81 static inline bool ti_is_const(TempOptInfo *ti)
82 {
83     return ti->is_const;
84 }
85 
86 static inline uint64_t ti_const_val(TempOptInfo *ti)
87 {
88     return ti->val;
89 }
90 
91 static inline bool ti_is_const_val(TempOptInfo *ti, uint64_t val)
92 {
93     return ti_is_const(ti) && ti_const_val(ti) == val;
94 }
95 
96 static inline bool ts_is_const(TCGTemp *ts)
97 {
98     return ti_is_const(ts_info(ts));
99 }
100 
101 static inline bool ts_is_const_val(TCGTemp *ts, uint64_t val)
102 {
103     return ti_is_const_val(ts_info(ts), val);
104 }
105 
106 static inline bool arg_is_const(TCGArg arg)
107 {
108     return ts_is_const(arg_temp(arg));
109 }
110 
111 static inline bool arg_is_const_val(TCGArg arg, uint64_t val)
112 {
113     return ts_is_const_val(arg_temp(arg), val);
114 }
115 
116 static inline bool ts_is_copy(TCGTemp *ts)
117 {
118     return ts_info(ts)->next_copy != ts;
119 }
120 
121 static TCGTemp *cmp_better_copy(TCGTemp *a, TCGTemp *b)
122 {
123     return a->kind < b->kind ? b : a;
124 }
125 
126 /* Initialize and activate a temporary.  */
127 static void init_ts_info(OptContext *ctx, TCGTemp *ts)
128 {
129     size_t idx = temp_idx(ts);
130     TempOptInfo *ti;
131 
132     if (test_bit(idx, ctx->temps_used.l)) {
133         return;
134     }
135     set_bit(idx, ctx->temps_used.l);
136 
137     ti = ts->state_ptr;
138     if (ti == NULL) {
139         ti = tcg_malloc(sizeof(TempOptInfo));
140         ts->state_ptr = ti;
141     }
142 
143     ti->next_copy = ts;
144     ti->prev_copy = ts;
145     QSIMPLEQ_INIT(&ti->mem_copy);
146     if (ts->kind == TEMP_CONST) {
147         ti->is_const = true;
148         ti->val = ts->val;
149         ti->z_mask = ts->val;
150         ti->s_mask = INT64_MIN >> clrsb64(ts->val);
151     } else {
152         ti->is_const = false;
153         ti->z_mask = -1;
154         ti->s_mask = 0;
155     }
156 }
157 
158 static MemCopyInfo *mem_copy_first(OptContext *ctx, intptr_t s, intptr_t l)
159 {
160     IntervalTreeNode *r = interval_tree_iter_first(&ctx->mem_copy, s, l);
161     return r ? container_of(r, MemCopyInfo, itree) : NULL;
162 }
163 
164 static MemCopyInfo *mem_copy_next(MemCopyInfo *mem, intptr_t s, intptr_t l)
165 {
166     IntervalTreeNode *r = interval_tree_iter_next(&mem->itree, s, l);
167     return r ? container_of(r, MemCopyInfo, itree) : NULL;
168 }
169 
170 static void remove_mem_copy(OptContext *ctx, MemCopyInfo *mc)
171 {
172     TCGTemp *ts = mc->ts;
173     TempOptInfo *ti = ts_info(ts);
174 
175     interval_tree_remove(&mc->itree, &ctx->mem_copy);
176     QSIMPLEQ_REMOVE(&ti->mem_copy, mc, MemCopyInfo, next);
177     QSIMPLEQ_INSERT_TAIL(&ctx->mem_free, mc, next);
178 }
179 
180 static void remove_mem_copy_in(OptContext *ctx, intptr_t s, intptr_t l)
181 {
182     while (true) {
183         MemCopyInfo *mc = mem_copy_first(ctx, s, l);
184         if (!mc) {
185             break;
186         }
187         remove_mem_copy(ctx, mc);
188     }
189 }
190 
191 static void remove_mem_copy_all(OptContext *ctx)
192 {
193     remove_mem_copy_in(ctx, 0, -1);
194     tcg_debug_assert(interval_tree_is_empty(&ctx->mem_copy));
195 }
196 
197 static TCGTemp *find_better_copy(TCGTemp *ts)
198 {
199     TCGTemp *i, *ret;
200 
201     /* If this is already readonly, we can't do better. */
202     if (temp_readonly(ts)) {
203         return ts;
204     }
205 
206     ret = ts;
207     for (i = ts_info(ts)->next_copy; i != ts; i = ts_info(i)->next_copy) {
208         ret = cmp_better_copy(ret, i);
209     }
210     return ret;
211 }
212 
213 static void move_mem_copies(TCGTemp *dst_ts, TCGTemp *src_ts)
214 {
215     TempOptInfo *si = ts_info(src_ts);
216     TempOptInfo *di = ts_info(dst_ts);
217     MemCopyInfo *mc;
218 
219     QSIMPLEQ_FOREACH(mc, &si->mem_copy, next) {
220         tcg_debug_assert(mc->ts == src_ts);
221         mc->ts = dst_ts;
222     }
223     QSIMPLEQ_CONCAT(&di->mem_copy, &si->mem_copy);
224 }
225 
226 /* Reset TEMP's state, possibly removing the temp for the list of copies.  */
227 static void reset_ts(OptContext *ctx, TCGTemp *ts)
228 {
229     TempOptInfo *ti = ts_info(ts);
230     TCGTemp *pts = ti->prev_copy;
231     TCGTemp *nts = ti->next_copy;
232     TempOptInfo *pi = ts_info(pts);
233     TempOptInfo *ni = ts_info(nts);
234 
235     ni->prev_copy = ti->prev_copy;
236     pi->next_copy = ti->next_copy;
237     ti->next_copy = ts;
238     ti->prev_copy = ts;
239     ti->is_const = false;
240     ti->z_mask = -1;
241     ti->s_mask = 0;
242 
243     if (!QSIMPLEQ_EMPTY(&ti->mem_copy)) {
244         if (ts == nts) {
245             /* Last temp copy being removed, the mem copies die. */
246             MemCopyInfo *mc;
247             QSIMPLEQ_FOREACH(mc, &ti->mem_copy, next) {
248                 interval_tree_remove(&mc->itree, &ctx->mem_copy);
249             }
250             QSIMPLEQ_CONCAT(&ctx->mem_free, &ti->mem_copy);
251         } else {
252             move_mem_copies(find_better_copy(nts), ts);
253         }
254     }
255 }
256 
257 static void reset_temp(OptContext *ctx, TCGArg arg)
258 {
259     reset_ts(ctx, arg_temp(arg));
260 }
261 
262 static void record_mem_copy(OptContext *ctx, TCGType type,
263                             TCGTemp *ts, intptr_t start, intptr_t last)
264 {
265     MemCopyInfo *mc;
266     TempOptInfo *ti;
267 
268     mc = QSIMPLEQ_FIRST(&ctx->mem_free);
269     if (mc) {
270         QSIMPLEQ_REMOVE_HEAD(&ctx->mem_free, next);
271     } else {
272         mc = tcg_malloc(sizeof(*mc));
273     }
274 
275     memset(mc, 0, sizeof(*mc));
276     mc->itree.start = start;
277     mc->itree.last = last;
278     mc->type = type;
279     interval_tree_insert(&mc->itree, &ctx->mem_copy);
280 
281     ts = find_better_copy(ts);
282     ti = ts_info(ts);
283     mc->ts = ts;
284     QSIMPLEQ_INSERT_TAIL(&ti->mem_copy, mc, next);
285 }
286 
287 static bool ts_are_copies(TCGTemp *ts1, TCGTemp *ts2)
288 {
289     TCGTemp *i;
290 
291     if (ts1 == ts2) {
292         return true;
293     }
294 
295     if (!ts_is_copy(ts1) || !ts_is_copy(ts2)) {
296         return false;
297     }
298 
299     for (i = ts_info(ts1)->next_copy; i != ts1; i = ts_info(i)->next_copy) {
300         if (i == ts2) {
301             return true;
302         }
303     }
304 
305     return false;
306 }
307 
308 static bool args_are_copies(TCGArg arg1, TCGArg arg2)
309 {
310     return ts_are_copies(arg_temp(arg1), arg_temp(arg2));
311 }
312 
313 static TCGTemp *find_mem_copy_for(OptContext *ctx, TCGType type, intptr_t s)
314 {
315     MemCopyInfo *mc;
316 
317     for (mc = mem_copy_first(ctx, s, s); mc; mc = mem_copy_next(mc, s, s)) {
318         if (mc->itree.start == s && mc->type == type) {
319             return find_better_copy(mc->ts);
320         }
321     }
322     return NULL;
323 }
324 
325 static TCGArg arg_new_constant(OptContext *ctx, uint64_t val)
326 {
327     TCGType type = ctx->type;
328     TCGTemp *ts;
329 
330     if (type == TCG_TYPE_I32) {
331         val = (int32_t)val;
332     }
333 
334     ts = tcg_constant_internal(type, val);
335     init_ts_info(ctx, ts);
336 
337     return temp_arg(ts);
338 }
339 
340 static TCGArg arg_new_temp(OptContext *ctx)
341 {
342     TCGTemp *ts = tcg_temp_new_internal(ctx->type, TEMP_EBB);
343     init_ts_info(ctx, ts);
344     return temp_arg(ts);
345 }
346 
347 static TCGOp *opt_insert_after(OptContext *ctx, TCGOp *op,
348                                TCGOpcode opc, unsigned narg)
349 {
350     return tcg_op_insert_after(ctx->tcg, op, opc, ctx->type, narg);
351 }
352 
353 static TCGOp *opt_insert_before(OptContext *ctx, TCGOp *op,
354                                 TCGOpcode opc, unsigned narg)
355 {
356     return tcg_op_insert_before(ctx->tcg, op, opc, ctx->type, narg);
357 }
358 
359 static bool tcg_opt_gen_mov(OptContext *ctx, TCGOp *op, TCGArg dst, TCGArg src)
360 {
361     TCGTemp *dst_ts = arg_temp(dst);
362     TCGTemp *src_ts = arg_temp(src);
363     TempOptInfo *di;
364     TempOptInfo *si;
365     TCGOpcode new_op;
366 
367     if (ts_are_copies(dst_ts, src_ts)) {
368         tcg_op_remove(ctx->tcg, op);
369         return true;
370     }
371 
372     reset_ts(ctx, dst_ts);
373     di = ts_info(dst_ts);
374     si = ts_info(src_ts);
375 
376     switch (ctx->type) {
377     case TCG_TYPE_I32:
378     case TCG_TYPE_I64:
379         new_op = INDEX_op_mov;
380         break;
381     case TCG_TYPE_V64:
382     case TCG_TYPE_V128:
383     case TCG_TYPE_V256:
384         /* TCGOP_TYPE and TCGOP_VECE remain unchanged.  */
385         new_op = INDEX_op_mov_vec;
386         break;
387     default:
388         g_assert_not_reached();
389     }
390     op->opc = new_op;
391     op->args[0] = dst;
392     op->args[1] = src;
393 
394     di->z_mask = si->z_mask;
395     di->s_mask = si->s_mask;
396 
397     if (src_ts->type == dst_ts->type) {
398         TempOptInfo *ni = ts_info(si->next_copy);
399 
400         di->next_copy = si->next_copy;
401         di->prev_copy = src_ts;
402         ni->prev_copy = dst_ts;
403         si->next_copy = dst_ts;
404         di->is_const = si->is_const;
405         di->val = si->val;
406 
407         if (!QSIMPLEQ_EMPTY(&si->mem_copy)
408             && cmp_better_copy(src_ts, dst_ts) == dst_ts) {
409             move_mem_copies(dst_ts, src_ts);
410         }
411     }
412     return true;
413 }
414 
415 static bool tcg_opt_gen_movi(OptContext *ctx, TCGOp *op,
416                              TCGArg dst, uint64_t val)
417 {
418     /* Convert movi to mov with constant temp. */
419     return tcg_opt_gen_mov(ctx, op, dst, arg_new_constant(ctx, val));
420 }
421 
422 static uint64_t do_constant_folding_2(TCGOpcode op, uint64_t x, uint64_t y)
423 {
424     uint64_t l64, h64;
425 
426     switch (op) {
427     case INDEX_op_add:
428         return x + y;
429 
430     CASE_OP_32_64(sub):
431         return x - y;
432 
433     CASE_OP_32_64(mul):
434         return x * y;
435 
436     case INDEX_op_and:
437     case INDEX_op_and_vec:
438         return x & y;
439 
440     case INDEX_op_or:
441     case INDEX_op_or_vec:
442         return x | y;
443 
444     case INDEX_op_xor:
445     case INDEX_op_xor_vec:
446         return x ^ y;
447 
448     case INDEX_op_shl_i32:
449         return (uint32_t)x << (y & 31);
450 
451     case INDEX_op_shl_i64:
452         return (uint64_t)x << (y & 63);
453 
454     case INDEX_op_shr_i32:
455         return (uint32_t)x >> (y & 31);
456 
457     case INDEX_op_shr_i64:
458         return (uint64_t)x >> (y & 63);
459 
460     case INDEX_op_sar_i32:
461         return (int32_t)x >> (y & 31);
462 
463     case INDEX_op_sar_i64:
464         return (int64_t)x >> (y & 63);
465 
466     case INDEX_op_rotr_i32:
467         return ror32(x, y & 31);
468 
469     case INDEX_op_rotr_i64:
470         return ror64(x, y & 63);
471 
472     case INDEX_op_rotl_i32:
473         return rol32(x, y & 31);
474 
475     case INDEX_op_rotl_i64:
476         return rol64(x, y & 63);
477 
478     CASE_OP_32_64_VEC(not):
479         return ~x;
480 
481     CASE_OP_32_64(neg):
482         return -x;
483 
484     case INDEX_op_andc:
485     case INDEX_op_andc_vec:
486         return x & ~y;
487 
488     case INDEX_op_orc:
489     case INDEX_op_orc_vec:
490         return x | ~y;
491 
492     case INDEX_op_eqv:
493     case INDEX_op_eqv_vec:
494         return ~(x ^ y);
495 
496     case INDEX_op_nand:
497     case INDEX_op_nand_vec:
498         return ~(x & y);
499 
500     CASE_OP_32_64_VEC(nor):
501         return ~(x | y);
502 
503     case INDEX_op_clz_i32:
504         return (uint32_t)x ? clz32(x) : y;
505 
506     case INDEX_op_clz_i64:
507         return x ? clz64(x) : y;
508 
509     case INDEX_op_ctz_i32:
510         return (uint32_t)x ? ctz32(x) : y;
511 
512     case INDEX_op_ctz_i64:
513         return x ? ctz64(x) : y;
514 
515     case INDEX_op_ctpop_i32:
516         return ctpop32(x);
517 
518     case INDEX_op_ctpop_i64:
519         return ctpop64(x);
520 
521     CASE_OP_32_64(bswap16):
522         x = bswap16(x);
523         return y & TCG_BSWAP_OS ? (int16_t)x : x;
524 
525     CASE_OP_32_64(bswap32):
526         x = bswap32(x);
527         return y & TCG_BSWAP_OS ? (int32_t)x : x;
528 
529     case INDEX_op_bswap64_i64:
530         return bswap64(x);
531 
532     case INDEX_op_ext_i32_i64:
533         return (int32_t)x;
534 
535     case INDEX_op_extu_i32_i64:
536     case INDEX_op_extrl_i64_i32:
537         return (uint32_t)x;
538 
539     case INDEX_op_extrh_i64_i32:
540         return (uint64_t)x >> 32;
541 
542     case INDEX_op_muluh_i32:
543         return ((uint64_t)(uint32_t)x * (uint32_t)y) >> 32;
544     case INDEX_op_mulsh_i32:
545         return ((int64_t)(int32_t)x * (int32_t)y) >> 32;
546 
547     case INDEX_op_muluh_i64:
548         mulu64(&l64, &h64, x, y);
549         return h64;
550     case INDEX_op_mulsh_i64:
551         muls64(&l64, &h64, x, y);
552         return h64;
553 
554     case INDEX_op_div_i32:
555         /* Avoid crashing on divide by zero, otherwise undefined.  */
556         return (int32_t)x / ((int32_t)y ? : 1);
557     case INDEX_op_divu_i32:
558         return (uint32_t)x / ((uint32_t)y ? : 1);
559     case INDEX_op_div_i64:
560         return (int64_t)x / ((int64_t)y ? : 1);
561     case INDEX_op_divu_i64:
562         return (uint64_t)x / ((uint64_t)y ? : 1);
563 
564     case INDEX_op_rem_i32:
565         return (int32_t)x % ((int32_t)y ? : 1);
566     case INDEX_op_remu_i32:
567         return (uint32_t)x % ((uint32_t)y ? : 1);
568     case INDEX_op_rem_i64:
569         return (int64_t)x % ((int64_t)y ? : 1);
570     case INDEX_op_remu_i64:
571         return (uint64_t)x % ((uint64_t)y ? : 1);
572 
573     default:
574         g_assert_not_reached();
575     }
576 }
577 
578 static uint64_t do_constant_folding(TCGOpcode op, TCGType type,
579                                     uint64_t x, uint64_t y)
580 {
581     uint64_t res = do_constant_folding_2(op, x, y);
582     if (type == TCG_TYPE_I32) {
583         res = (int32_t)res;
584     }
585     return res;
586 }
587 
588 static bool do_constant_folding_cond_32(uint32_t x, uint32_t y, TCGCond c)
589 {
590     switch (c) {
591     case TCG_COND_EQ:
592         return x == y;
593     case TCG_COND_NE:
594         return x != y;
595     case TCG_COND_LT:
596         return (int32_t)x < (int32_t)y;
597     case TCG_COND_GE:
598         return (int32_t)x >= (int32_t)y;
599     case TCG_COND_LE:
600         return (int32_t)x <= (int32_t)y;
601     case TCG_COND_GT:
602         return (int32_t)x > (int32_t)y;
603     case TCG_COND_LTU:
604         return x < y;
605     case TCG_COND_GEU:
606         return x >= y;
607     case TCG_COND_LEU:
608         return x <= y;
609     case TCG_COND_GTU:
610         return x > y;
611     case TCG_COND_TSTEQ:
612         return (x & y) == 0;
613     case TCG_COND_TSTNE:
614         return (x & y) != 0;
615     case TCG_COND_ALWAYS:
616     case TCG_COND_NEVER:
617         break;
618     }
619     g_assert_not_reached();
620 }
621 
622 static bool do_constant_folding_cond_64(uint64_t x, uint64_t y, TCGCond c)
623 {
624     switch (c) {
625     case TCG_COND_EQ:
626         return x == y;
627     case TCG_COND_NE:
628         return x != y;
629     case TCG_COND_LT:
630         return (int64_t)x < (int64_t)y;
631     case TCG_COND_GE:
632         return (int64_t)x >= (int64_t)y;
633     case TCG_COND_LE:
634         return (int64_t)x <= (int64_t)y;
635     case TCG_COND_GT:
636         return (int64_t)x > (int64_t)y;
637     case TCG_COND_LTU:
638         return x < y;
639     case TCG_COND_GEU:
640         return x >= y;
641     case TCG_COND_LEU:
642         return x <= y;
643     case TCG_COND_GTU:
644         return x > y;
645     case TCG_COND_TSTEQ:
646         return (x & y) == 0;
647     case TCG_COND_TSTNE:
648         return (x & y) != 0;
649     case TCG_COND_ALWAYS:
650     case TCG_COND_NEVER:
651         break;
652     }
653     g_assert_not_reached();
654 }
655 
656 static int do_constant_folding_cond_eq(TCGCond c)
657 {
658     switch (c) {
659     case TCG_COND_GT:
660     case TCG_COND_LTU:
661     case TCG_COND_LT:
662     case TCG_COND_GTU:
663     case TCG_COND_NE:
664         return 0;
665     case TCG_COND_GE:
666     case TCG_COND_GEU:
667     case TCG_COND_LE:
668     case TCG_COND_LEU:
669     case TCG_COND_EQ:
670         return 1;
671     case TCG_COND_TSTEQ:
672     case TCG_COND_TSTNE:
673         return -1;
674     case TCG_COND_ALWAYS:
675     case TCG_COND_NEVER:
676         break;
677     }
678     g_assert_not_reached();
679 }
680 
681 /*
682  * Return -1 if the condition can't be simplified,
683  * and the result of the condition (0 or 1) if it can.
684  */
685 static int do_constant_folding_cond(TCGType type, TCGArg x,
686                                     TCGArg y, TCGCond c)
687 {
688     if (arg_is_const(x) && arg_is_const(y)) {
689         uint64_t xv = arg_info(x)->val;
690         uint64_t yv = arg_info(y)->val;
691 
692         switch (type) {
693         case TCG_TYPE_I32:
694             return do_constant_folding_cond_32(xv, yv, c);
695         case TCG_TYPE_I64:
696             return do_constant_folding_cond_64(xv, yv, c);
697         default:
698             /* Only scalar comparisons are optimizable */
699             return -1;
700         }
701     } else if (args_are_copies(x, y)) {
702         return do_constant_folding_cond_eq(c);
703     } else if (arg_is_const_val(y, 0)) {
704         switch (c) {
705         case TCG_COND_LTU:
706         case TCG_COND_TSTNE:
707             return 0;
708         case TCG_COND_GEU:
709         case TCG_COND_TSTEQ:
710             return 1;
711         default:
712             return -1;
713         }
714     }
715     return -1;
716 }
717 
718 /**
719  * swap_commutative:
720  * @dest: TCGArg of the destination argument, or NO_DEST.
721  * @p1: first paired argument
722  * @p2: second paired argument
723  *
724  * If *@p1 is a constant and *@p2 is not, swap.
725  * If *@p2 matches @dest, swap.
726  * Return true if a swap was performed.
727  */
728 
729 #define NO_DEST  temp_arg(NULL)
730 
731 static bool swap_commutative(TCGArg dest, TCGArg *p1, TCGArg *p2)
732 {
733     TCGArg a1 = *p1, a2 = *p2;
734     int sum = 0;
735     sum += arg_is_const(a1);
736     sum -= arg_is_const(a2);
737 
738     /* Prefer the constant in second argument, and then the form
739        op a, a, b, which is better handled on non-RISC hosts. */
740     if (sum > 0 || (sum == 0 && dest == a2)) {
741         *p1 = a2;
742         *p2 = a1;
743         return true;
744     }
745     return false;
746 }
747 
748 static bool swap_commutative2(TCGArg *p1, TCGArg *p2)
749 {
750     int sum = 0;
751     sum += arg_is_const(p1[0]);
752     sum += arg_is_const(p1[1]);
753     sum -= arg_is_const(p2[0]);
754     sum -= arg_is_const(p2[1]);
755     if (sum > 0) {
756         TCGArg t;
757         t = p1[0], p1[0] = p2[0], p2[0] = t;
758         t = p1[1], p1[1] = p2[1], p2[1] = t;
759         return true;
760     }
761     return false;
762 }
763 
764 /*
765  * Return -1 if the condition can't be simplified,
766  * and the result of the condition (0 or 1) if it can.
767  */
768 static int do_constant_folding_cond1(OptContext *ctx, TCGOp *op, TCGArg dest,
769                                      TCGArg *p1, TCGArg *p2, TCGArg *pcond)
770 {
771     TCGCond cond;
772     TempOptInfo *i1;
773     bool swap;
774     int r;
775 
776     swap = swap_commutative(dest, p1, p2);
777     cond = *pcond;
778     if (swap) {
779         *pcond = cond = tcg_swap_cond(cond);
780     }
781 
782     r = do_constant_folding_cond(ctx->type, *p1, *p2, cond);
783     if (r >= 0) {
784         return r;
785     }
786     if (!is_tst_cond(cond)) {
787         return -1;
788     }
789 
790     i1 = arg_info(*p1);
791 
792     /*
793      * TSTNE x,x -> NE x,0
794      * TSTNE x,i -> NE x,0 if i includes all nonzero bits of x
795      */
796     if (args_are_copies(*p1, *p2) ||
797         (arg_is_const(*p2) && (i1->z_mask & ~arg_info(*p2)->val) == 0)) {
798         *p2 = arg_new_constant(ctx, 0);
799         *pcond = tcg_tst_eqne_cond(cond);
800         return -1;
801     }
802 
803     /* TSTNE x,i -> LT x,0 if i only includes sign bit copies */
804     if (arg_is_const(*p2) && (arg_info(*p2)->val & ~i1->s_mask) == 0) {
805         *p2 = arg_new_constant(ctx, 0);
806         *pcond = tcg_tst_ltge_cond(cond);
807         return -1;
808     }
809 
810     /* Expand to AND with a temporary if no backend support. */
811     if (!TCG_TARGET_HAS_tst) {
812         TCGOp *op2 = opt_insert_before(ctx, op, INDEX_op_and, 3);
813         TCGArg tmp = arg_new_temp(ctx);
814 
815         op2->args[0] = tmp;
816         op2->args[1] = *p1;
817         op2->args[2] = *p2;
818 
819         *p1 = tmp;
820         *p2 = arg_new_constant(ctx, 0);
821         *pcond = tcg_tst_eqne_cond(cond);
822     }
823     return -1;
824 }
825 
826 static int do_constant_folding_cond2(OptContext *ctx, TCGOp *op, TCGArg *args)
827 {
828     TCGArg al, ah, bl, bh;
829     TCGCond c;
830     bool swap;
831     int r;
832 
833     swap = swap_commutative2(args, args + 2);
834     c = args[4];
835     if (swap) {
836         args[4] = c = tcg_swap_cond(c);
837     }
838 
839     al = args[0];
840     ah = args[1];
841     bl = args[2];
842     bh = args[3];
843 
844     if (arg_is_const(bl) && arg_is_const(bh)) {
845         tcg_target_ulong blv = arg_info(bl)->val;
846         tcg_target_ulong bhv = arg_info(bh)->val;
847         uint64_t b = deposit64(blv, 32, 32, bhv);
848 
849         if (arg_is_const(al) && arg_is_const(ah)) {
850             tcg_target_ulong alv = arg_info(al)->val;
851             tcg_target_ulong ahv = arg_info(ah)->val;
852             uint64_t a = deposit64(alv, 32, 32, ahv);
853 
854             r = do_constant_folding_cond_64(a, b, c);
855             if (r >= 0) {
856                 return r;
857             }
858         }
859 
860         if (b == 0) {
861             switch (c) {
862             case TCG_COND_LTU:
863             case TCG_COND_TSTNE:
864                 return 0;
865             case TCG_COND_GEU:
866             case TCG_COND_TSTEQ:
867                 return 1;
868             default:
869                 break;
870             }
871         }
872 
873         /* TSTNE x,-1 -> NE x,0 */
874         if (b == -1 && is_tst_cond(c)) {
875             args[3] = args[2] = arg_new_constant(ctx, 0);
876             args[4] = tcg_tst_eqne_cond(c);
877             return -1;
878         }
879 
880         /* TSTNE x,sign -> LT x,0 */
881         if (b == INT64_MIN && is_tst_cond(c)) {
882             /* bl must be 0, so copy that to bh */
883             args[3] = bl;
884             args[4] = tcg_tst_ltge_cond(c);
885             return -1;
886         }
887     }
888 
889     if (args_are_copies(al, bl) && args_are_copies(ah, bh)) {
890         r = do_constant_folding_cond_eq(c);
891         if (r >= 0) {
892             return r;
893         }
894 
895         /* TSTNE x,x -> NE x,0 */
896         if (is_tst_cond(c)) {
897             args[3] = args[2] = arg_new_constant(ctx, 0);
898             args[4] = tcg_tst_eqne_cond(c);
899             return -1;
900         }
901     }
902 
903     /* Expand to AND with a temporary if no backend support. */
904     if (!TCG_TARGET_HAS_tst && is_tst_cond(c)) {
905         TCGOp *op1 = opt_insert_before(ctx, op, INDEX_op_and, 3);
906         TCGOp *op2 = opt_insert_before(ctx, op, INDEX_op_and, 3);
907         TCGArg t1 = arg_new_temp(ctx);
908         TCGArg t2 = arg_new_temp(ctx);
909 
910         op1->args[0] = t1;
911         op1->args[1] = al;
912         op1->args[2] = bl;
913         op2->args[0] = t2;
914         op2->args[1] = ah;
915         op2->args[2] = bh;
916 
917         args[0] = t1;
918         args[1] = t2;
919         args[3] = args[2] = arg_new_constant(ctx, 0);
920         args[4] = tcg_tst_eqne_cond(c);
921     }
922     return -1;
923 }
924 
925 static void init_arguments(OptContext *ctx, TCGOp *op, int nb_args)
926 {
927     for (int i = 0; i < nb_args; i++) {
928         TCGTemp *ts = arg_temp(op->args[i]);
929         init_ts_info(ctx, ts);
930     }
931 }
932 
933 static void copy_propagate(OptContext *ctx, TCGOp *op,
934                            int nb_oargs, int nb_iargs)
935 {
936     for (int i = nb_oargs; i < nb_oargs + nb_iargs; i++) {
937         TCGTemp *ts = arg_temp(op->args[i]);
938         if (ts_is_copy(ts)) {
939             op->args[i] = temp_arg(find_better_copy(ts));
940         }
941     }
942 }
943 
944 static void finish_bb(OptContext *ctx)
945 {
946     /* We only optimize memory barriers across basic blocks. */
947     ctx->prev_mb = NULL;
948 }
949 
950 static void finish_ebb(OptContext *ctx)
951 {
952     finish_bb(ctx);
953     /* We only optimize across extended basic blocks. */
954     memset(&ctx->temps_used, 0, sizeof(ctx->temps_used));
955     remove_mem_copy_all(ctx);
956 }
957 
958 static bool finish_folding(OptContext *ctx, TCGOp *op)
959 {
960     const TCGOpDef *def = &tcg_op_defs[op->opc];
961     int i, nb_oargs;
962 
963     nb_oargs = def->nb_oargs;
964     for (i = 0; i < nb_oargs; i++) {
965         TCGTemp *ts = arg_temp(op->args[i]);
966         reset_ts(ctx, ts);
967     }
968     return true;
969 }
970 
971 /*
972  * The fold_* functions return true when processing is complete,
973  * usually by folding the operation to a constant or to a copy,
974  * and calling tcg_opt_gen_{mov,movi}.  They may do other things,
975  * like collect information about the value produced, for use in
976  * optimizing a subsequent operation.
977  *
978  * These first fold_* functions are all helpers, used by other
979  * folders for more specific operations.
980  */
981 
982 static bool fold_const1(OptContext *ctx, TCGOp *op)
983 {
984     if (arg_is_const(op->args[1])) {
985         uint64_t t;
986 
987         t = arg_info(op->args[1])->val;
988         t = do_constant_folding(op->opc, ctx->type, t, 0);
989         return tcg_opt_gen_movi(ctx, op, op->args[0], t);
990     }
991     return false;
992 }
993 
994 static bool fold_const2(OptContext *ctx, TCGOp *op)
995 {
996     if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) {
997         uint64_t t1 = arg_info(op->args[1])->val;
998         uint64_t t2 = arg_info(op->args[2])->val;
999 
1000         t1 = do_constant_folding(op->opc, ctx->type, t1, t2);
1001         return tcg_opt_gen_movi(ctx, op, op->args[0], t1);
1002     }
1003     return false;
1004 }
1005 
1006 static bool fold_commutative(OptContext *ctx, TCGOp *op)
1007 {
1008     swap_commutative(op->args[0], &op->args[1], &op->args[2]);
1009     return false;
1010 }
1011 
1012 static bool fold_const2_commutative(OptContext *ctx, TCGOp *op)
1013 {
1014     swap_commutative(op->args[0], &op->args[1], &op->args[2]);
1015     return fold_const2(ctx, op);
1016 }
1017 
1018 /*
1019  * Record "zero" and "sign" masks for the single output of @op.
1020  * See TempOptInfo definition of z_mask and s_mask.
1021  * If z_mask allows, fold the output to constant zero.
1022  * The passed s_mask may be augmented by z_mask.
1023  */
1024 static bool fold_masks_zs(OptContext *ctx, TCGOp *op,
1025                           uint64_t z_mask, int64_t s_mask)
1026 {
1027     const TCGOpDef *def = &tcg_op_defs[op->opc];
1028     TCGTemp *ts;
1029     TempOptInfo *ti;
1030     int rep;
1031 
1032     /* Only single-output opcodes are supported here. */
1033     tcg_debug_assert(def->nb_oargs == 1);
1034 
1035     /*
1036      * 32-bit ops generate 32-bit results, which for the purpose of
1037      * simplifying tcg are sign-extended.  Certainly that's how we
1038      * represent our constants elsewhere.  Note that the bits will
1039      * be reset properly for a 64-bit value when encountering the
1040      * type changing opcodes.
1041      */
1042     if (ctx->type == TCG_TYPE_I32) {
1043         z_mask = (int32_t)z_mask;
1044         s_mask |= INT32_MIN;
1045     }
1046 
1047     if (z_mask == 0) {
1048         return tcg_opt_gen_movi(ctx, op, op->args[0], 0);
1049     }
1050 
1051     ts = arg_temp(op->args[0]);
1052     reset_ts(ctx, ts);
1053 
1054     ti = ts_info(ts);
1055     ti->z_mask = z_mask;
1056 
1057     /* Canonicalize s_mask and incorporate data from z_mask. */
1058     rep = clz64(~s_mask);
1059     rep = MAX(rep, clz64(z_mask));
1060     rep = MAX(rep - 1, 0);
1061     ti->s_mask = INT64_MIN >> rep;
1062 
1063     return true;
1064 }
1065 
1066 static bool fold_masks_z(OptContext *ctx, TCGOp *op, uint64_t z_mask)
1067 {
1068     return fold_masks_zs(ctx, op, z_mask, 0);
1069 }
1070 
1071 static bool fold_masks_s(OptContext *ctx, TCGOp *op, uint64_t s_mask)
1072 {
1073     return fold_masks_zs(ctx, op, -1, s_mask);
1074 }
1075 
1076 /*
1077  * An "affected" mask bit is 0 if and only if the result is identical
1078  * to the first input.  Thus if the entire mask is 0, the operation
1079  * is equivalent to a copy.
1080  */
1081 static bool fold_affected_mask(OptContext *ctx, TCGOp *op, uint64_t a_mask)
1082 {
1083     if (ctx->type == TCG_TYPE_I32) {
1084         a_mask = (uint32_t)a_mask;
1085     }
1086     if (a_mask == 0) {
1087         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
1088     }
1089     return false;
1090 }
1091 
1092 /*
1093  * Convert @op to NOT, if NOT is supported by the host.
1094  * Return true f the conversion is successful, which will still
1095  * indicate that the processing is complete.
1096  */
1097 static bool fold_not(OptContext *ctx, TCGOp *op);
1098 static bool fold_to_not(OptContext *ctx, TCGOp *op, int idx)
1099 {
1100     TCGOpcode not_op;
1101     bool have_not;
1102 
1103     switch (ctx->type) {
1104     case TCG_TYPE_I32:
1105         not_op = INDEX_op_not_i32;
1106         have_not = TCG_TARGET_HAS_not_i32;
1107         break;
1108     case TCG_TYPE_I64:
1109         not_op = INDEX_op_not_i64;
1110         have_not = TCG_TARGET_HAS_not_i64;
1111         break;
1112     case TCG_TYPE_V64:
1113     case TCG_TYPE_V128:
1114     case TCG_TYPE_V256:
1115         not_op = INDEX_op_not_vec;
1116         have_not = TCG_TARGET_HAS_not_vec;
1117         break;
1118     default:
1119         g_assert_not_reached();
1120     }
1121     if (have_not) {
1122         op->opc = not_op;
1123         op->args[1] = op->args[idx];
1124         return fold_not(ctx, op);
1125     }
1126     return false;
1127 }
1128 
1129 /* If the binary operation has first argument @i, fold to @i. */
1130 static bool fold_ix_to_i(OptContext *ctx, TCGOp *op, uint64_t i)
1131 {
1132     if (arg_is_const_val(op->args[1], i)) {
1133         return tcg_opt_gen_movi(ctx, op, op->args[0], i);
1134     }
1135     return false;
1136 }
1137 
1138 /* If the binary operation has first argument @i, fold to NOT. */
1139 static bool fold_ix_to_not(OptContext *ctx, TCGOp *op, uint64_t i)
1140 {
1141     if (arg_is_const_val(op->args[1], i)) {
1142         return fold_to_not(ctx, op, 2);
1143     }
1144     return false;
1145 }
1146 
1147 /* If the binary operation has second argument @i, fold to @i. */
1148 static bool fold_xi_to_i(OptContext *ctx, TCGOp *op, uint64_t i)
1149 {
1150     if (arg_is_const_val(op->args[2], i)) {
1151         return tcg_opt_gen_movi(ctx, op, op->args[0], i);
1152     }
1153     return false;
1154 }
1155 
1156 /* If the binary operation has second argument @i, fold to identity. */
1157 static bool fold_xi_to_x(OptContext *ctx, TCGOp *op, uint64_t i)
1158 {
1159     if (arg_is_const_val(op->args[2], i)) {
1160         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
1161     }
1162     return false;
1163 }
1164 
1165 /* If the binary operation has second argument @i, fold to NOT. */
1166 static bool fold_xi_to_not(OptContext *ctx, TCGOp *op, uint64_t i)
1167 {
1168     if (arg_is_const_val(op->args[2], i)) {
1169         return fold_to_not(ctx, op, 1);
1170     }
1171     return false;
1172 }
1173 
1174 /* If the binary operation has both arguments equal, fold to @i. */
1175 static bool fold_xx_to_i(OptContext *ctx, TCGOp *op, uint64_t i)
1176 {
1177     if (args_are_copies(op->args[1], op->args[2])) {
1178         return tcg_opt_gen_movi(ctx, op, op->args[0], i);
1179     }
1180     return false;
1181 }
1182 
1183 /* If the binary operation has both arguments equal, fold to identity. */
1184 static bool fold_xx_to_x(OptContext *ctx, TCGOp *op)
1185 {
1186     if (args_are_copies(op->args[1], op->args[2])) {
1187         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
1188     }
1189     return false;
1190 }
1191 
1192 /*
1193  * These outermost fold_<op> functions are sorted alphabetically.
1194  *
1195  * The ordering of the transformations should be:
1196  *   1) those that produce a constant
1197  *   2) those that produce a copy
1198  *   3) those that produce information about the result value.
1199  */
1200 
1201 static bool fold_or(OptContext *ctx, TCGOp *op);
1202 static bool fold_orc(OptContext *ctx, TCGOp *op);
1203 static bool fold_xor(OptContext *ctx, TCGOp *op);
1204 
1205 static bool fold_add(OptContext *ctx, TCGOp *op)
1206 {
1207     if (fold_const2_commutative(ctx, op) ||
1208         fold_xi_to_x(ctx, op, 0)) {
1209         return true;
1210     }
1211     return finish_folding(ctx, op);
1212 }
1213 
1214 /* We cannot as yet do_constant_folding with vectors. */
1215 static bool fold_add_vec(OptContext *ctx, TCGOp *op)
1216 {
1217     if (fold_commutative(ctx, op) ||
1218         fold_xi_to_x(ctx, op, 0)) {
1219         return true;
1220     }
1221     return finish_folding(ctx, op);
1222 }
1223 
1224 static bool fold_addsub2(OptContext *ctx, TCGOp *op, bool add)
1225 {
1226     bool a_const = arg_is_const(op->args[2]) && arg_is_const(op->args[3]);
1227     bool b_const = arg_is_const(op->args[4]) && arg_is_const(op->args[5]);
1228 
1229     if (a_const && b_const) {
1230         uint64_t al = arg_info(op->args[2])->val;
1231         uint64_t ah = arg_info(op->args[3])->val;
1232         uint64_t bl = arg_info(op->args[4])->val;
1233         uint64_t bh = arg_info(op->args[5])->val;
1234         TCGArg rl, rh;
1235         TCGOp *op2;
1236 
1237         if (ctx->type == TCG_TYPE_I32) {
1238             uint64_t a = deposit64(al, 32, 32, ah);
1239             uint64_t b = deposit64(bl, 32, 32, bh);
1240 
1241             if (add) {
1242                 a += b;
1243             } else {
1244                 a -= b;
1245             }
1246 
1247             al = sextract64(a, 0, 32);
1248             ah = sextract64(a, 32, 32);
1249         } else {
1250             Int128 a = int128_make128(al, ah);
1251             Int128 b = int128_make128(bl, bh);
1252 
1253             if (add) {
1254                 a = int128_add(a, b);
1255             } else {
1256                 a = int128_sub(a, b);
1257             }
1258 
1259             al = int128_getlo(a);
1260             ah = int128_gethi(a);
1261         }
1262 
1263         rl = op->args[0];
1264         rh = op->args[1];
1265 
1266         /* The proper opcode is supplied by tcg_opt_gen_mov. */
1267         op2 = opt_insert_before(ctx, op, 0, 2);
1268 
1269         tcg_opt_gen_movi(ctx, op, rl, al);
1270         tcg_opt_gen_movi(ctx, op2, rh, ah);
1271         return true;
1272     }
1273 
1274     /* Fold sub2 r,x,i to add2 r,x,-i */
1275     if (!add && b_const) {
1276         uint64_t bl = arg_info(op->args[4])->val;
1277         uint64_t bh = arg_info(op->args[5])->val;
1278 
1279         /* Negate the two parts without assembling and disassembling. */
1280         bl = -bl;
1281         bh = ~bh + !bl;
1282 
1283         op->opc = (ctx->type == TCG_TYPE_I32
1284                    ? INDEX_op_add2_i32 : INDEX_op_add2_i64);
1285         op->args[4] = arg_new_constant(ctx, bl);
1286         op->args[5] = arg_new_constant(ctx, bh);
1287     }
1288     return finish_folding(ctx, op);
1289 }
1290 
1291 static bool fold_add2(OptContext *ctx, TCGOp *op)
1292 {
1293     /* Note that the high and low parts may be independently swapped. */
1294     swap_commutative(op->args[0], &op->args[2], &op->args[4]);
1295     swap_commutative(op->args[1], &op->args[3], &op->args[5]);
1296 
1297     return fold_addsub2(ctx, op, true);
1298 }
1299 
1300 static bool fold_and(OptContext *ctx, TCGOp *op)
1301 {
1302     uint64_t z1, z2, z_mask, s_mask;
1303     TempOptInfo *t1, *t2;
1304 
1305     if (fold_const2_commutative(ctx, op) ||
1306         fold_xi_to_i(ctx, op, 0) ||
1307         fold_xi_to_x(ctx, op, -1) ||
1308         fold_xx_to_x(ctx, op)) {
1309         return true;
1310     }
1311 
1312     t1 = arg_info(op->args[1]);
1313     t2 = arg_info(op->args[2]);
1314     z1 = t1->z_mask;
1315     z2 = t2->z_mask;
1316 
1317     /*
1318      * Known-zeros does not imply known-ones.  Therefore unless
1319      * arg2 is constant, we can't infer affected bits from it.
1320      */
1321     if (ti_is_const(t2) && fold_affected_mask(ctx, op, z1 & ~z2)) {
1322         return true;
1323     }
1324 
1325     z_mask = z1 & z2;
1326 
1327     /*
1328      * Sign repetitions are perforce all identical, whether they are 1 or 0.
1329      * Bitwise operations preserve the relative quantity of the repetitions.
1330      */
1331     s_mask = t1->s_mask & t2->s_mask;
1332 
1333     return fold_masks_zs(ctx, op, z_mask, s_mask);
1334 }
1335 
1336 static bool fold_andc(OptContext *ctx, TCGOp *op)
1337 {
1338     uint64_t z_mask, s_mask;
1339     TempOptInfo *t1, *t2;
1340 
1341     if (fold_const2(ctx, op) ||
1342         fold_xx_to_i(ctx, op, 0) ||
1343         fold_xi_to_x(ctx, op, 0) ||
1344         fold_ix_to_not(ctx, op, -1)) {
1345         return true;
1346     }
1347 
1348     t1 = arg_info(op->args[1]);
1349     t2 = arg_info(op->args[2]);
1350     z_mask = t1->z_mask;
1351 
1352     if (ti_is_const(t2)) {
1353         /* Fold andc r,x,i to and r,x,~i. */
1354         switch (ctx->type) {
1355         case TCG_TYPE_I32:
1356         case TCG_TYPE_I64:
1357             op->opc = INDEX_op_and;
1358             break;
1359         case TCG_TYPE_V64:
1360         case TCG_TYPE_V128:
1361         case TCG_TYPE_V256:
1362             op->opc = INDEX_op_and_vec;
1363             break;
1364         default:
1365             g_assert_not_reached();
1366         }
1367         op->args[2] = arg_new_constant(ctx, ~ti_const_val(t2));
1368         return fold_and(ctx, op);
1369     }
1370 
1371     /*
1372      * Known-zeros does not imply known-ones.  Therefore unless
1373      * arg2 is constant, we can't infer anything from it.
1374      */
1375     if (ti_is_const(t2)) {
1376         uint64_t v2 = ti_const_val(t2);
1377         if (fold_affected_mask(ctx, op, z_mask & v2)) {
1378             return true;
1379         }
1380         z_mask &= ~v2;
1381     }
1382 
1383     s_mask = t1->s_mask & t2->s_mask;
1384     return fold_masks_zs(ctx, op, z_mask, s_mask);
1385 }
1386 
1387 static bool fold_bitsel_vec(OptContext *ctx, TCGOp *op)
1388 {
1389     /* If true and false values are the same, eliminate the cmp. */
1390     if (args_are_copies(op->args[2], op->args[3])) {
1391         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[2]);
1392     }
1393 
1394     if (arg_is_const(op->args[2]) && arg_is_const(op->args[3])) {
1395         uint64_t tv = arg_info(op->args[2])->val;
1396         uint64_t fv = arg_info(op->args[3])->val;
1397 
1398         if (tv == -1 && fv == 0) {
1399             return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
1400         }
1401         if (tv == 0 && fv == -1) {
1402             if (TCG_TARGET_HAS_not_vec) {
1403                 op->opc = INDEX_op_not_vec;
1404                 return fold_not(ctx, op);
1405             } else {
1406                 op->opc = INDEX_op_xor_vec;
1407                 op->args[2] = arg_new_constant(ctx, -1);
1408                 return fold_xor(ctx, op);
1409             }
1410         }
1411     }
1412     if (arg_is_const(op->args[2])) {
1413         uint64_t tv = arg_info(op->args[2])->val;
1414         if (tv == -1) {
1415             op->opc = INDEX_op_or_vec;
1416             op->args[2] = op->args[3];
1417             return fold_or(ctx, op);
1418         }
1419         if (tv == 0 && TCG_TARGET_HAS_andc_vec) {
1420             op->opc = INDEX_op_andc_vec;
1421             op->args[2] = op->args[1];
1422             op->args[1] = op->args[3];
1423             return fold_andc(ctx, op);
1424         }
1425     }
1426     if (arg_is_const(op->args[3])) {
1427         uint64_t fv = arg_info(op->args[3])->val;
1428         if (fv == 0) {
1429             op->opc = INDEX_op_and_vec;
1430             return fold_and(ctx, op);
1431         }
1432         if (fv == -1 && TCG_TARGET_HAS_orc_vec) {
1433             op->opc = INDEX_op_orc_vec;
1434             op->args[2] = op->args[1];
1435             op->args[1] = op->args[3];
1436             return fold_orc(ctx, op);
1437         }
1438     }
1439     return finish_folding(ctx, op);
1440 }
1441 
1442 static bool fold_brcond(OptContext *ctx, TCGOp *op)
1443 {
1444     int i = do_constant_folding_cond1(ctx, op, NO_DEST, &op->args[0],
1445                                       &op->args[1], &op->args[2]);
1446     if (i == 0) {
1447         tcg_op_remove(ctx->tcg, op);
1448         return true;
1449     }
1450     if (i > 0) {
1451         op->opc = INDEX_op_br;
1452         op->args[0] = op->args[3];
1453         finish_ebb(ctx);
1454     } else {
1455         finish_bb(ctx);
1456     }
1457     return true;
1458 }
1459 
1460 static bool fold_brcond2(OptContext *ctx, TCGOp *op)
1461 {
1462     TCGCond cond;
1463     TCGArg label;
1464     int i, inv = 0;
1465 
1466     i = do_constant_folding_cond2(ctx, op, &op->args[0]);
1467     cond = op->args[4];
1468     label = op->args[5];
1469     if (i >= 0) {
1470         goto do_brcond_const;
1471     }
1472 
1473     switch (cond) {
1474     case TCG_COND_LT:
1475     case TCG_COND_GE:
1476         /*
1477          * Simplify LT/GE comparisons vs zero to a single compare
1478          * vs the high word of the input.
1479          */
1480         if (arg_is_const_val(op->args[2], 0) &&
1481             arg_is_const_val(op->args[3], 0)) {
1482             goto do_brcond_high;
1483         }
1484         break;
1485 
1486     case TCG_COND_NE:
1487         inv = 1;
1488         QEMU_FALLTHROUGH;
1489     case TCG_COND_EQ:
1490         /*
1491          * Simplify EQ/NE comparisons where one of the pairs
1492          * can be simplified.
1493          */
1494         i = do_constant_folding_cond(TCG_TYPE_I32, op->args[0],
1495                                      op->args[2], cond);
1496         switch (i ^ inv) {
1497         case 0:
1498             goto do_brcond_const;
1499         case 1:
1500             goto do_brcond_high;
1501         }
1502 
1503         i = do_constant_folding_cond(TCG_TYPE_I32, op->args[1],
1504                                      op->args[3], cond);
1505         switch (i ^ inv) {
1506         case 0:
1507             goto do_brcond_const;
1508         case 1:
1509             goto do_brcond_low;
1510         }
1511         break;
1512 
1513     case TCG_COND_TSTEQ:
1514     case TCG_COND_TSTNE:
1515         if (arg_is_const_val(op->args[2], 0)) {
1516             goto do_brcond_high;
1517         }
1518         if (arg_is_const_val(op->args[3], 0)) {
1519             goto do_brcond_low;
1520         }
1521         break;
1522 
1523     default:
1524         break;
1525 
1526     do_brcond_low:
1527         op->opc = INDEX_op_brcond_i32;
1528         op->args[1] = op->args[2];
1529         op->args[2] = cond;
1530         op->args[3] = label;
1531         return fold_brcond(ctx, op);
1532 
1533     do_brcond_high:
1534         op->opc = INDEX_op_brcond_i32;
1535         op->args[0] = op->args[1];
1536         op->args[1] = op->args[3];
1537         op->args[2] = cond;
1538         op->args[3] = label;
1539         return fold_brcond(ctx, op);
1540 
1541     do_brcond_const:
1542         if (i == 0) {
1543             tcg_op_remove(ctx->tcg, op);
1544             return true;
1545         }
1546         op->opc = INDEX_op_br;
1547         op->args[0] = label;
1548         finish_ebb(ctx);
1549         return true;
1550     }
1551 
1552     finish_bb(ctx);
1553     return true;
1554 }
1555 
1556 static bool fold_bswap(OptContext *ctx, TCGOp *op)
1557 {
1558     uint64_t z_mask, s_mask, sign;
1559     TempOptInfo *t1 = arg_info(op->args[1]);
1560 
1561     if (ti_is_const(t1)) {
1562         return tcg_opt_gen_movi(ctx, op, op->args[0],
1563                                 do_constant_folding(op->opc, ctx->type,
1564                                                     ti_const_val(t1),
1565                                                     op->args[2]));
1566     }
1567 
1568     z_mask = t1->z_mask;
1569     switch (op->opc) {
1570     case INDEX_op_bswap16_i32:
1571     case INDEX_op_bswap16_i64:
1572         z_mask = bswap16(z_mask);
1573         sign = INT16_MIN;
1574         break;
1575     case INDEX_op_bswap32_i32:
1576     case INDEX_op_bswap32_i64:
1577         z_mask = bswap32(z_mask);
1578         sign = INT32_MIN;
1579         break;
1580     case INDEX_op_bswap64_i64:
1581         z_mask = bswap64(z_mask);
1582         sign = INT64_MIN;
1583         break;
1584     default:
1585         g_assert_not_reached();
1586     }
1587 
1588     s_mask = 0;
1589     switch (op->args[2] & (TCG_BSWAP_OZ | TCG_BSWAP_OS)) {
1590     case TCG_BSWAP_OZ:
1591         break;
1592     case TCG_BSWAP_OS:
1593         /* If the sign bit may be 1, force all the bits above to 1. */
1594         if (z_mask & sign) {
1595             z_mask |= sign;
1596         }
1597         /* The value and therefore s_mask is explicitly sign-extended. */
1598         s_mask = sign;
1599         break;
1600     default:
1601         /* The high bits are undefined: force all bits above the sign to 1. */
1602         z_mask |= sign << 1;
1603         break;
1604     }
1605 
1606     return fold_masks_zs(ctx, op, z_mask, s_mask);
1607 }
1608 
1609 static bool fold_call(OptContext *ctx, TCGOp *op)
1610 {
1611     TCGContext *s = ctx->tcg;
1612     int nb_oargs = TCGOP_CALLO(op);
1613     int nb_iargs = TCGOP_CALLI(op);
1614     int flags, i;
1615 
1616     init_arguments(ctx, op, nb_oargs + nb_iargs);
1617     copy_propagate(ctx, op, nb_oargs, nb_iargs);
1618 
1619     /* If the function reads or writes globals, reset temp data. */
1620     flags = tcg_call_flags(op);
1621     if (!(flags & (TCG_CALL_NO_READ_GLOBALS | TCG_CALL_NO_WRITE_GLOBALS))) {
1622         int nb_globals = s->nb_globals;
1623 
1624         for (i = 0; i < nb_globals; i++) {
1625             if (test_bit(i, ctx->temps_used.l)) {
1626                 reset_ts(ctx, &ctx->tcg->temps[i]);
1627             }
1628         }
1629     }
1630 
1631     /* If the function has side effects, reset mem data. */
1632     if (!(flags & TCG_CALL_NO_SIDE_EFFECTS)) {
1633         remove_mem_copy_all(ctx);
1634     }
1635 
1636     /* Reset temp data for outputs. */
1637     for (i = 0; i < nb_oargs; i++) {
1638         reset_temp(ctx, op->args[i]);
1639     }
1640 
1641     /* Stop optimizing MB across calls. */
1642     ctx->prev_mb = NULL;
1643     return true;
1644 }
1645 
1646 static bool fold_cmp_vec(OptContext *ctx, TCGOp *op)
1647 {
1648     /* Canonicalize the comparison to put immediate second. */
1649     if (swap_commutative(NO_DEST, &op->args[1], &op->args[2])) {
1650         op->args[3] = tcg_swap_cond(op->args[3]);
1651     }
1652     return finish_folding(ctx, op);
1653 }
1654 
1655 static bool fold_cmpsel_vec(OptContext *ctx, TCGOp *op)
1656 {
1657     /* If true and false values are the same, eliminate the cmp. */
1658     if (args_are_copies(op->args[3], op->args[4])) {
1659         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[3]);
1660     }
1661 
1662     /* Canonicalize the comparison to put immediate second. */
1663     if (swap_commutative(NO_DEST, &op->args[1], &op->args[2])) {
1664         op->args[5] = tcg_swap_cond(op->args[5]);
1665     }
1666     /*
1667      * Canonicalize the "false" input reg to match the destination,
1668      * so that the tcg backend can implement "move if true".
1669      */
1670     if (swap_commutative(op->args[0], &op->args[4], &op->args[3])) {
1671         op->args[5] = tcg_invert_cond(op->args[5]);
1672     }
1673     return finish_folding(ctx, op);
1674 }
1675 
1676 static bool fold_count_zeros(OptContext *ctx, TCGOp *op)
1677 {
1678     uint64_t z_mask, s_mask;
1679     TempOptInfo *t1 = arg_info(op->args[1]);
1680     TempOptInfo *t2 = arg_info(op->args[2]);
1681 
1682     if (ti_is_const(t1)) {
1683         uint64_t t = ti_const_val(t1);
1684 
1685         if (t != 0) {
1686             t = do_constant_folding(op->opc, ctx->type, t, 0);
1687             return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1688         }
1689         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[2]);
1690     }
1691 
1692     switch (ctx->type) {
1693     case TCG_TYPE_I32:
1694         z_mask = 31;
1695         break;
1696     case TCG_TYPE_I64:
1697         z_mask = 63;
1698         break;
1699     default:
1700         g_assert_not_reached();
1701     }
1702     s_mask = ~z_mask;
1703     z_mask |= t2->z_mask;
1704     s_mask &= t2->s_mask;
1705 
1706     return fold_masks_zs(ctx, op, z_mask, s_mask);
1707 }
1708 
1709 static bool fold_ctpop(OptContext *ctx, TCGOp *op)
1710 {
1711     uint64_t z_mask;
1712 
1713     if (fold_const1(ctx, op)) {
1714         return true;
1715     }
1716 
1717     switch (ctx->type) {
1718     case TCG_TYPE_I32:
1719         z_mask = 32 | 31;
1720         break;
1721     case TCG_TYPE_I64:
1722         z_mask = 64 | 63;
1723         break;
1724     default:
1725         g_assert_not_reached();
1726     }
1727     return fold_masks_z(ctx, op, z_mask);
1728 }
1729 
1730 static bool fold_deposit(OptContext *ctx, TCGOp *op)
1731 {
1732     TempOptInfo *t1 = arg_info(op->args[1]);
1733     TempOptInfo *t2 = arg_info(op->args[2]);
1734     int ofs = op->args[3];
1735     int len = op->args[4];
1736     int width = 8 * tcg_type_size(ctx->type);
1737     uint64_t z_mask, s_mask;
1738 
1739     if (ti_is_const(t1) && ti_is_const(t2)) {
1740         return tcg_opt_gen_movi(ctx, op, op->args[0],
1741                                 deposit64(ti_const_val(t1), ofs, len,
1742                                           ti_const_val(t2)));
1743     }
1744 
1745     /* Inserting a value into zero at offset 0. */
1746     if (ti_is_const_val(t1, 0) && ofs == 0) {
1747         uint64_t mask = MAKE_64BIT_MASK(0, len);
1748 
1749         op->opc = INDEX_op_and;
1750         op->args[1] = op->args[2];
1751         op->args[2] = arg_new_constant(ctx, mask);
1752         return fold_and(ctx, op);
1753     }
1754 
1755     /* Inserting zero into a value. */
1756     if (ti_is_const_val(t2, 0)) {
1757         uint64_t mask = deposit64(-1, ofs, len, 0);
1758 
1759         op->opc = INDEX_op_and;
1760         op->args[2] = arg_new_constant(ctx, mask);
1761         return fold_and(ctx, op);
1762     }
1763 
1764     /* The s_mask from the top portion of the deposit is still valid. */
1765     if (ofs + len == width) {
1766         s_mask = t2->s_mask << ofs;
1767     } else {
1768         s_mask = t1->s_mask & ~MAKE_64BIT_MASK(0, ofs + len);
1769     }
1770 
1771     z_mask = deposit64(t1->z_mask, ofs, len, t2->z_mask);
1772     return fold_masks_zs(ctx, op, z_mask, s_mask);
1773 }
1774 
1775 static bool fold_divide(OptContext *ctx, TCGOp *op)
1776 {
1777     if (fold_const2(ctx, op) ||
1778         fold_xi_to_x(ctx, op, 1)) {
1779         return true;
1780     }
1781     return finish_folding(ctx, op);
1782 }
1783 
1784 static bool fold_dup(OptContext *ctx, TCGOp *op)
1785 {
1786     if (arg_is_const(op->args[1])) {
1787         uint64_t t = arg_info(op->args[1])->val;
1788         t = dup_const(TCGOP_VECE(op), t);
1789         return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1790     }
1791     return finish_folding(ctx, op);
1792 }
1793 
1794 static bool fold_dup2(OptContext *ctx, TCGOp *op)
1795 {
1796     if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) {
1797         uint64_t t = deposit64(arg_info(op->args[1])->val, 32, 32,
1798                                arg_info(op->args[2])->val);
1799         return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1800     }
1801 
1802     if (args_are_copies(op->args[1], op->args[2])) {
1803         op->opc = INDEX_op_dup_vec;
1804         TCGOP_VECE(op) = MO_32;
1805     }
1806     return finish_folding(ctx, op);
1807 }
1808 
1809 static bool fold_eqv(OptContext *ctx, TCGOp *op)
1810 {
1811     uint64_t s_mask;
1812     TempOptInfo *t1, *t2;
1813 
1814     if (fold_const2_commutative(ctx, op) ||
1815         fold_xi_to_x(ctx, op, -1) ||
1816         fold_xi_to_not(ctx, op, 0)) {
1817         return true;
1818     }
1819 
1820     t2 = arg_info(op->args[2]);
1821     if (ti_is_const(t2)) {
1822         /* Fold eqv r,x,i to xor r,x,~i. */
1823         switch (ctx->type) {
1824         case TCG_TYPE_I32:
1825         case TCG_TYPE_I64:
1826             op->opc = INDEX_op_xor;
1827             break;
1828         case TCG_TYPE_V64:
1829         case TCG_TYPE_V128:
1830         case TCG_TYPE_V256:
1831             op->opc = INDEX_op_xor_vec;
1832             break;
1833         default:
1834             g_assert_not_reached();
1835         }
1836         op->args[2] = arg_new_constant(ctx, ~ti_const_val(t2));
1837         return fold_xor(ctx, op);
1838     }
1839 
1840     t1 = arg_info(op->args[1]);
1841     s_mask = t1->s_mask & t2->s_mask;
1842     return fold_masks_s(ctx, op, s_mask);
1843 }
1844 
1845 static bool fold_extract(OptContext *ctx, TCGOp *op)
1846 {
1847     uint64_t z_mask_old, z_mask;
1848     TempOptInfo *t1 = arg_info(op->args[1]);
1849     int pos = op->args[2];
1850     int len = op->args[3];
1851 
1852     if (ti_is_const(t1)) {
1853         return tcg_opt_gen_movi(ctx, op, op->args[0],
1854                                 extract64(ti_const_val(t1), pos, len));
1855     }
1856 
1857     z_mask_old = t1->z_mask;
1858     z_mask = extract64(z_mask_old, pos, len);
1859     if (pos == 0 && fold_affected_mask(ctx, op, z_mask_old ^ z_mask)) {
1860         return true;
1861     }
1862 
1863     return fold_masks_z(ctx, op, z_mask);
1864 }
1865 
1866 static bool fold_extract2(OptContext *ctx, TCGOp *op)
1867 {
1868     if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) {
1869         uint64_t v1 = arg_info(op->args[1])->val;
1870         uint64_t v2 = arg_info(op->args[2])->val;
1871         int shr = op->args[3];
1872 
1873         if (op->opc == INDEX_op_extract2_i64) {
1874             v1 >>= shr;
1875             v2 <<= 64 - shr;
1876         } else {
1877             v1 = (uint32_t)v1 >> shr;
1878             v2 = (uint64_t)((int32_t)v2 << (32 - shr));
1879         }
1880         return tcg_opt_gen_movi(ctx, op, op->args[0], v1 | v2);
1881     }
1882     return finish_folding(ctx, op);
1883 }
1884 
1885 static bool fold_exts(OptContext *ctx, TCGOp *op)
1886 {
1887     uint64_t s_mask, z_mask;
1888     TempOptInfo *t1;
1889 
1890     if (fold_const1(ctx, op)) {
1891         return true;
1892     }
1893 
1894     t1 = arg_info(op->args[1]);
1895     z_mask = t1->z_mask;
1896     s_mask = t1->s_mask;
1897 
1898     switch (op->opc) {
1899     case INDEX_op_ext_i32_i64:
1900         s_mask |= INT32_MIN;
1901         z_mask = (int32_t)z_mask;
1902         break;
1903     default:
1904         g_assert_not_reached();
1905     }
1906     return fold_masks_zs(ctx, op, z_mask, s_mask);
1907 }
1908 
1909 static bool fold_extu(OptContext *ctx, TCGOp *op)
1910 {
1911     uint64_t z_mask;
1912 
1913     if (fold_const1(ctx, op)) {
1914         return true;
1915     }
1916 
1917     z_mask = arg_info(op->args[1])->z_mask;
1918     switch (op->opc) {
1919     case INDEX_op_extrl_i64_i32:
1920     case INDEX_op_extu_i32_i64:
1921         z_mask = (uint32_t)z_mask;
1922         break;
1923     case INDEX_op_extrh_i64_i32:
1924         z_mask >>= 32;
1925         break;
1926     default:
1927         g_assert_not_reached();
1928     }
1929     return fold_masks_z(ctx, op, z_mask);
1930 }
1931 
1932 static bool fold_mb(OptContext *ctx, TCGOp *op)
1933 {
1934     /* Eliminate duplicate and redundant fence instructions.  */
1935     if (ctx->prev_mb) {
1936         /*
1937          * Merge two barriers of the same type into one,
1938          * or a weaker barrier into a stronger one,
1939          * or two weaker barriers into a stronger one.
1940          *   mb X; mb Y => mb X|Y
1941          *   mb; strl => mb; st
1942          *   ldaq; mb => ld; mb
1943          *   ldaq; strl => ld; mb; st
1944          * Other combinations are also merged into a strong
1945          * barrier.  This is stricter than specified but for
1946          * the purposes of TCG is better than not optimizing.
1947          */
1948         ctx->prev_mb->args[0] |= op->args[0];
1949         tcg_op_remove(ctx->tcg, op);
1950     } else {
1951         ctx->prev_mb = op;
1952     }
1953     return true;
1954 }
1955 
1956 static bool fold_mov(OptContext *ctx, TCGOp *op)
1957 {
1958     return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
1959 }
1960 
1961 static bool fold_movcond(OptContext *ctx, TCGOp *op)
1962 {
1963     uint64_t z_mask, s_mask;
1964     TempOptInfo *tt, *ft;
1965     int i;
1966 
1967     /* If true and false values are the same, eliminate the cmp. */
1968     if (args_are_copies(op->args[3], op->args[4])) {
1969         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[3]);
1970     }
1971 
1972     /*
1973      * Canonicalize the "false" input reg to match the destination reg so
1974      * that the tcg backend can implement a "move if true" operation.
1975      */
1976     if (swap_commutative(op->args[0], &op->args[4], &op->args[3])) {
1977         op->args[5] = tcg_invert_cond(op->args[5]);
1978     }
1979 
1980     i = do_constant_folding_cond1(ctx, op, NO_DEST, &op->args[1],
1981                                   &op->args[2], &op->args[5]);
1982     if (i >= 0) {
1983         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[4 - i]);
1984     }
1985 
1986     tt = arg_info(op->args[3]);
1987     ft = arg_info(op->args[4]);
1988     z_mask = tt->z_mask | ft->z_mask;
1989     s_mask = tt->s_mask & ft->s_mask;
1990 
1991     if (ti_is_const(tt) && ti_is_const(ft)) {
1992         uint64_t tv = ti_const_val(tt);
1993         uint64_t fv = ti_const_val(ft);
1994         TCGOpcode opc, negopc = 0;
1995         TCGCond cond = op->args[5];
1996 
1997         switch (ctx->type) {
1998         case TCG_TYPE_I32:
1999             opc = INDEX_op_setcond_i32;
2000             if (TCG_TARGET_HAS_negsetcond_i32) {
2001                 negopc = INDEX_op_negsetcond_i32;
2002             }
2003             tv = (int32_t)tv;
2004             fv = (int32_t)fv;
2005             break;
2006         case TCG_TYPE_I64:
2007             opc = INDEX_op_setcond_i64;
2008             if (TCG_TARGET_HAS_negsetcond_i64) {
2009                 negopc = INDEX_op_negsetcond_i64;
2010             }
2011             break;
2012         default:
2013             g_assert_not_reached();
2014         }
2015 
2016         if (tv == 1 && fv == 0) {
2017             op->opc = opc;
2018             op->args[3] = cond;
2019         } else if (fv == 1 && tv == 0) {
2020             op->opc = opc;
2021             op->args[3] = tcg_invert_cond(cond);
2022         } else if (negopc) {
2023             if (tv == -1 && fv == 0) {
2024                 op->opc = negopc;
2025                 op->args[3] = cond;
2026             } else if (fv == -1 && tv == 0) {
2027                 op->opc = negopc;
2028                 op->args[3] = tcg_invert_cond(cond);
2029             }
2030         }
2031     }
2032 
2033     return fold_masks_zs(ctx, op, z_mask, s_mask);
2034 }
2035 
2036 static bool fold_mul(OptContext *ctx, TCGOp *op)
2037 {
2038     if (fold_const2(ctx, op) ||
2039         fold_xi_to_i(ctx, op, 0) ||
2040         fold_xi_to_x(ctx, op, 1)) {
2041         return true;
2042     }
2043     return finish_folding(ctx, op);
2044 }
2045 
2046 static bool fold_mul_highpart(OptContext *ctx, TCGOp *op)
2047 {
2048     if (fold_const2_commutative(ctx, op) ||
2049         fold_xi_to_i(ctx, op, 0)) {
2050         return true;
2051     }
2052     return finish_folding(ctx, op);
2053 }
2054 
2055 static bool fold_multiply2(OptContext *ctx, TCGOp *op)
2056 {
2057     swap_commutative(op->args[0], &op->args[2], &op->args[3]);
2058 
2059     if (arg_is_const(op->args[2]) && arg_is_const(op->args[3])) {
2060         uint64_t a = arg_info(op->args[2])->val;
2061         uint64_t b = arg_info(op->args[3])->val;
2062         uint64_t h, l;
2063         TCGArg rl, rh;
2064         TCGOp *op2;
2065 
2066         switch (op->opc) {
2067         case INDEX_op_mulu2_i32:
2068             l = (uint64_t)(uint32_t)a * (uint32_t)b;
2069             h = (int32_t)(l >> 32);
2070             l = (int32_t)l;
2071             break;
2072         case INDEX_op_muls2_i32:
2073             l = (int64_t)(int32_t)a * (int32_t)b;
2074             h = l >> 32;
2075             l = (int32_t)l;
2076             break;
2077         case INDEX_op_mulu2_i64:
2078             mulu64(&l, &h, a, b);
2079             break;
2080         case INDEX_op_muls2_i64:
2081             muls64(&l, &h, a, b);
2082             break;
2083         default:
2084             g_assert_not_reached();
2085         }
2086 
2087         rl = op->args[0];
2088         rh = op->args[1];
2089 
2090         /* The proper opcode is supplied by tcg_opt_gen_mov. */
2091         op2 = opt_insert_before(ctx, op, 0, 2);
2092 
2093         tcg_opt_gen_movi(ctx, op, rl, l);
2094         tcg_opt_gen_movi(ctx, op2, rh, h);
2095         return true;
2096     }
2097     return finish_folding(ctx, op);
2098 }
2099 
2100 static bool fold_nand(OptContext *ctx, TCGOp *op)
2101 {
2102     uint64_t s_mask;
2103 
2104     if (fold_const2_commutative(ctx, op) ||
2105         fold_xi_to_not(ctx, op, -1)) {
2106         return true;
2107     }
2108 
2109     s_mask = arg_info(op->args[1])->s_mask
2110            & arg_info(op->args[2])->s_mask;
2111     return fold_masks_s(ctx, op, s_mask);
2112 }
2113 
2114 static bool fold_neg_no_const(OptContext *ctx, TCGOp *op)
2115 {
2116     /* Set to 1 all bits to the left of the rightmost.  */
2117     uint64_t z_mask = arg_info(op->args[1])->z_mask;
2118     z_mask = -(z_mask & -z_mask);
2119 
2120     return fold_masks_z(ctx, op, z_mask);
2121 }
2122 
2123 static bool fold_neg(OptContext *ctx, TCGOp *op)
2124 {
2125     return fold_const1(ctx, op) || fold_neg_no_const(ctx, op);
2126 }
2127 
2128 static bool fold_nor(OptContext *ctx, TCGOp *op)
2129 {
2130     uint64_t s_mask;
2131 
2132     if (fold_const2_commutative(ctx, op) ||
2133         fold_xi_to_not(ctx, op, 0)) {
2134         return true;
2135     }
2136 
2137     s_mask = arg_info(op->args[1])->s_mask
2138            & arg_info(op->args[2])->s_mask;
2139     return fold_masks_s(ctx, op, s_mask);
2140 }
2141 
2142 static bool fold_not(OptContext *ctx, TCGOp *op)
2143 {
2144     if (fold_const1(ctx, op)) {
2145         return true;
2146     }
2147     return fold_masks_s(ctx, op, arg_info(op->args[1])->s_mask);
2148 }
2149 
2150 static bool fold_or(OptContext *ctx, TCGOp *op)
2151 {
2152     uint64_t z_mask, s_mask;
2153     TempOptInfo *t1, *t2;
2154 
2155     if (fold_const2_commutative(ctx, op) ||
2156         fold_xi_to_x(ctx, op, 0) ||
2157         fold_xx_to_x(ctx, op)) {
2158         return true;
2159     }
2160 
2161     t1 = arg_info(op->args[1]);
2162     t2 = arg_info(op->args[2]);
2163     z_mask = t1->z_mask | t2->z_mask;
2164     s_mask = t1->s_mask & t2->s_mask;
2165     return fold_masks_zs(ctx, op, z_mask, s_mask);
2166 }
2167 
2168 static bool fold_orc(OptContext *ctx, TCGOp *op)
2169 {
2170     uint64_t s_mask;
2171     TempOptInfo *t1, *t2;
2172 
2173     if (fold_const2(ctx, op) ||
2174         fold_xx_to_i(ctx, op, -1) ||
2175         fold_xi_to_x(ctx, op, -1) ||
2176         fold_ix_to_not(ctx, op, 0)) {
2177         return true;
2178     }
2179 
2180     t2 = arg_info(op->args[2]);
2181     if (ti_is_const(t2)) {
2182         /* Fold orc r,x,i to or r,x,~i. */
2183         switch (ctx->type) {
2184         case TCG_TYPE_I32:
2185         case TCG_TYPE_I64:
2186             op->opc = INDEX_op_or;
2187             break;
2188         case TCG_TYPE_V64:
2189         case TCG_TYPE_V128:
2190         case TCG_TYPE_V256:
2191             op->opc = INDEX_op_or_vec;
2192             break;
2193         default:
2194             g_assert_not_reached();
2195         }
2196         op->args[2] = arg_new_constant(ctx, ~ti_const_val(t2));
2197         return fold_or(ctx, op);
2198     }
2199 
2200     t1 = arg_info(op->args[1]);
2201     s_mask = t1->s_mask & t2->s_mask;
2202     return fold_masks_s(ctx, op, s_mask);
2203 }
2204 
2205 static bool fold_qemu_ld_1reg(OptContext *ctx, TCGOp *op)
2206 {
2207     const TCGOpDef *def = &tcg_op_defs[op->opc];
2208     MemOpIdx oi = op->args[def->nb_oargs + def->nb_iargs];
2209     MemOp mop = get_memop(oi);
2210     int width = 8 * memop_size(mop);
2211     uint64_t z_mask = -1, s_mask = 0;
2212 
2213     if (width < 64) {
2214         if (mop & MO_SIGN) {
2215             s_mask = MAKE_64BIT_MASK(width - 1, 64 - (width - 1));
2216         } else {
2217             z_mask = MAKE_64BIT_MASK(0, width);
2218         }
2219     }
2220 
2221     /* Opcodes that touch guest memory stop the mb optimization.  */
2222     ctx->prev_mb = NULL;
2223 
2224     return fold_masks_zs(ctx, op, z_mask, s_mask);
2225 }
2226 
2227 static bool fold_qemu_ld_2reg(OptContext *ctx, TCGOp *op)
2228 {
2229     /* Opcodes that touch guest memory stop the mb optimization.  */
2230     ctx->prev_mb = NULL;
2231     return finish_folding(ctx, op);
2232 }
2233 
2234 static bool fold_qemu_st(OptContext *ctx, TCGOp *op)
2235 {
2236     /* Opcodes that touch guest memory stop the mb optimization.  */
2237     ctx->prev_mb = NULL;
2238     return true;
2239 }
2240 
2241 static bool fold_remainder(OptContext *ctx, TCGOp *op)
2242 {
2243     if (fold_const2(ctx, op) ||
2244         fold_xx_to_i(ctx, op, 0)) {
2245         return true;
2246     }
2247     return finish_folding(ctx, op);
2248 }
2249 
2250 /* Return 1 if finished, -1 if simplified, 0 if unchanged. */
2251 static int fold_setcond_zmask(OptContext *ctx, TCGOp *op, bool neg)
2252 {
2253     uint64_t a_zmask, b_val;
2254     TCGCond cond;
2255 
2256     if (!arg_is_const(op->args[2])) {
2257         return false;
2258     }
2259 
2260     a_zmask = arg_info(op->args[1])->z_mask;
2261     b_val = arg_info(op->args[2])->val;
2262     cond = op->args[3];
2263 
2264     if (ctx->type == TCG_TYPE_I32) {
2265         a_zmask = (uint32_t)a_zmask;
2266         b_val = (uint32_t)b_val;
2267     }
2268 
2269     /*
2270      * A with only low bits set vs B with high bits set means that A < B.
2271      */
2272     if (a_zmask < b_val) {
2273         bool inv = false;
2274 
2275         switch (cond) {
2276         case TCG_COND_NE:
2277         case TCG_COND_LEU:
2278         case TCG_COND_LTU:
2279             inv = true;
2280             /* fall through */
2281         case TCG_COND_GTU:
2282         case TCG_COND_GEU:
2283         case TCG_COND_EQ:
2284             return tcg_opt_gen_movi(ctx, op, op->args[0], neg ? -inv : inv);
2285         default:
2286             break;
2287         }
2288     }
2289 
2290     /*
2291      * A with only lsb set is already boolean.
2292      */
2293     if (a_zmask <= 1) {
2294         bool convert = false;
2295         bool inv = false;
2296 
2297         switch (cond) {
2298         case TCG_COND_EQ:
2299             inv = true;
2300             /* fall through */
2301         case TCG_COND_NE:
2302             convert = (b_val == 0);
2303             break;
2304         case TCG_COND_LTU:
2305         case TCG_COND_TSTEQ:
2306             inv = true;
2307             /* fall through */
2308         case TCG_COND_GEU:
2309         case TCG_COND_TSTNE:
2310             convert = (b_val == 1);
2311             break;
2312         default:
2313             break;
2314         }
2315         if (convert) {
2316             TCGOpcode neg_opc;
2317 
2318             if (!inv && !neg) {
2319                 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
2320             }
2321 
2322             switch (ctx->type) {
2323             case TCG_TYPE_I32:
2324                 neg_opc = INDEX_op_neg_i32;
2325                 break;
2326             case TCG_TYPE_I64:
2327                 neg_opc = INDEX_op_neg_i64;
2328                 break;
2329             default:
2330                 g_assert_not_reached();
2331             }
2332 
2333             if (!inv) {
2334                 op->opc = neg_opc;
2335             } else if (neg) {
2336                 op->opc = INDEX_op_add;
2337                 op->args[2] = arg_new_constant(ctx, -1);
2338             } else {
2339                 op->opc = INDEX_op_xor;
2340                 op->args[2] = arg_new_constant(ctx, 1);
2341             }
2342             return -1;
2343         }
2344     }
2345     return 0;
2346 }
2347 
2348 static void fold_setcond_tst_pow2(OptContext *ctx, TCGOp *op, bool neg)
2349 {
2350     TCGOpcode neg_opc, shr_opc;
2351     TCGOpcode uext_opc = 0, sext_opc = 0;
2352     TCGCond cond = op->args[3];
2353     TCGArg ret, src1, src2;
2354     TCGOp *op2;
2355     uint64_t val;
2356     int sh;
2357     bool inv;
2358 
2359     if (!is_tst_cond(cond) || !arg_is_const(op->args[2])) {
2360         return;
2361     }
2362 
2363     src2 = op->args[2];
2364     val = arg_info(src2)->val;
2365     if (!is_power_of_2(val)) {
2366         return;
2367     }
2368     sh = ctz64(val);
2369 
2370     switch (ctx->type) {
2371     case TCG_TYPE_I32:
2372         shr_opc = INDEX_op_shr_i32;
2373         neg_opc = INDEX_op_neg_i32;
2374         if (TCG_TARGET_extract_valid(TCG_TYPE_I32, sh, 1)) {
2375             uext_opc = INDEX_op_extract_i32;
2376         }
2377         if (TCG_TARGET_sextract_valid(TCG_TYPE_I32, sh, 1)) {
2378             sext_opc = INDEX_op_sextract_i32;
2379         }
2380         break;
2381     case TCG_TYPE_I64:
2382         shr_opc = INDEX_op_shr_i64;
2383         neg_opc = INDEX_op_neg_i64;
2384         if (TCG_TARGET_extract_valid(TCG_TYPE_I64, sh, 1)) {
2385             uext_opc = INDEX_op_extract_i64;
2386         }
2387         if (TCG_TARGET_sextract_valid(TCG_TYPE_I64, sh, 1)) {
2388             sext_opc = INDEX_op_sextract_i64;
2389         }
2390         break;
2391     default:
2392         g_assert_not_reached();
2393     }
2394 
2395     ret = op->args[0];
2396     src1 = op->args[1];
2397     inv = cond == TCG_COND_TSTEQ;
2398 
2399     if (sh && sext_opc && neg && !inv) {
2400         op->opc = sext_opc;
2401         op->args[1] = src1;
2402         op->args[2] = sh;
2403         op->args[3] = 1;
2404         return;
2405     } else if (sh && uext_opc) {
2406         op->opc = uext_opc;
2407         op->args[1] = src1;
2408         op->args[2] = sh;
2409         op->args[3] = 1;
2410     } else {
2411         if (sh) {
2412             op2 = opt_insert_before(ctx, op, shr_opc, 3);
2413             op2->args[0] = ret;
2414             op2->args[1] = src1;
2415             op2->args[2] = arg_new_constant(ctx, sh);
2416             src1 = ret;
2417         }
2418         op->opc = INDEX_op_and;
2419         op->args[1] = src1;
2420         op->args[2] = arg_new_constant(ctx, 1);
2421     }
2422 
2423     if (neg && inv) {
2424         op2 = opt_insert_after(ctx, op, INDEX_op_add, 3);
2425         op2->args[0] = ret;
2426         op2->args[1] = ret;
2427         op2->args[2] = arg_new_constant(ctx, -1);
2428     } else if (inv) {
2429         op2 = opt_insert_after(ctx, op, INDEX_op_xor, 3);
2430         op2->args[0] = ret;
2431         op2->args[1] = ret;
2432         op2->args[2] = arg_new_constant(ctx, 1);
2433     } else if (neg) {
2434         op2 = opt_insert_after(ctx, op, neg_opc, 2);
2435         op2->args[0] = ret;
2436         op2->args[1] = ret;
2437     }
2438 }
2439 
2440 static bool fold_setcond(OptContext *ctx, TCGOp *op)
2441 {
2442     int i = do_constant_folding_cond1(ctx, op, op->args[0], &op->args[1],
2443                                       &op->args[2], &op->args[3]);
2444     if (i >= 0) {
2445         return tcg_opt_gen_movi(ctx, op, op->args[0], i);
2446     }
2447 
2448     i = fold_setcond_zmask(ctx, op, false);
2449     if (i > 0) {
2450         return true;
2451     }
2452     if (i == 0) {
2453         fold_setcond_tst_pow2(ctx, op, false);
2454     }
2455 
2456     return fold_masks_z(ctx, op, 1);
2457 }
2458 
2459 static bool fold_negsetcond(OptContext *ctx, TCGOp *op)
2460 {
2461     int i = do_constant_folding_cond1(ctx, op, op->args[0], &op->args[1],
2462                                       &op->args[2], &op->args[3]);
2463     if (i >= 0) {
2464         return tcg_opt_gen_movi(ctx, op, op->args[0], -i);
2465     }
2466 
2467     i = fold_setcond_zmask(ctx, op, true);
2468     if (i > 0) {
2469         return true;
2470     }
2471     if (i == 0) {
2472         fold_setcond_tst_pow2(ctx, op, true);
2473     }
2474 
2475     /* Value is {0,-1} so all bits are repetitions of the sign. */
2476     return fold_masks_s(ctx, op, -1);
2477 }
2478 
2479 static bool fold_setcond2(OptContext *ctx, TCGOp *op)
2480 {
2481     TCGCond cond;
2482     int i, inv = 0;
2483 
2484     i = do_constant_folding_cond2(ctx, op, &op->args[1]);
2485     cond = op->args[5];
2486     if (i >= 0) {
2487         goto do_setcond_const;
2488     }
2489 
2490     switch (cond) {
2491     case TCG_COND_LT:
2492     case TCG_COND_GE:
2493         /*
2494          * Simplify LT/GE comparisons vs zero to a single compare
2495          * vs the high word of the input.
2496          */
2497         if (arg_is_const_val(op->args[3], 0) &&
2498             arg_is_const_val(op->args[4], 0)) {
2499             goto do_setcond_high;
2500         }
2501         break;
2502 
2503     case TCG_COND_NE:
2504         inv = 1;
2505         QEMU_FALLTHROUGH;
2506     case TCG_COND_EQ:
2507         /*
2508          * Simplify EQ/NE comparisons where one of the pairs
2509          * can be simplified.
2510          */
2511         i = do_constant_folding_cond(TCG_TYPE_I32, op->args[1],
2512                                      op->args[3], cond);
2513         switch (i ^ inv) {
2514         case 0:
2515             goto do_setcond_const;
2516         case 1:
2517             goto do_setcond_high;
2518         }
2519 
2520         i = do_constant_folding_cond(TCG_TYPE_I32, op->args[2],
2521                                      op->args[4], cond);
2522         switch (i ^ inv) {
2523         case 0:
2524             goto do_setcond_const;
2525         case 1:
2526             goto do_setcond_low;
2527         }
2528         break;
2529 
2530     case TCG_COND_TSTEQ:
2531     case TCG_COND_TSTNE:
2532         if (arg_is_const_val(op->args[3], 0)) {
2533             goto do_setcond_high;
2534         }
2535         if (arg_is_const_val(op->args[4], 0)) {
2536             goto do_setcond_low;
2537         }
2538         break;
2539 
2540     default:
2541         break;
2542 
2543     do_setcond_low:
2544         op->args[2] = op->args[3];
2545         op->args[3] = cond;
2546         op->opc = INDEX_op_setcond_i32;
2547         return fold_setcond(ctx, op);
2548 
2549     do_setcond_high:
2550         op->args[1] = op->args[2];
2551         op->args[2] = op->args[4];
2552         op->args[3] = cond;
2553         op->opc = INDEX_op_setcond_i32;
2554         return fold_setcond(ctx, op);
2555     }
2556 
2557     return fold_masks_z(ctx, op, 1);
2558 
2559  do_setcond_const:
2560     return tcg_opt_gen_movi(ctx, op, op->args[0], i);
2561 }
2562 
2563 static bool fold_sextract(OptContext *ctx, TCGOp *op)
2564 {
2565     uint64_t z_mask, s_mask, s_mask_old;
2566     TempOptInfo *t1 = arg_info(op->args[1]);
2567     int pos = op->args[2];
2568     int len = op->args[3];
2569 
2570     if (ti_is_const(t1)) {
2571         return tcg_opt_gen_movi(ctx, op, op->args[0],
2572                                 sextract64(ti_const_val(t1), pos, len));
2573     }
2574 
2575     s_mask_old = t1->s_mask;
2576     s_mask = s_mask_old >> pos;
2577     s_mask |= -1ull << (len - 1);
2578 
2579     if (pos == 0 && fold_affected_mask(ctx, op, s_mask & ~s_mask_old)) {
2580         return true;
2581     }
2582 
2583     z_mask = sextract64(t1->z_mask, pos, len);
2584     return fold_masks_zs(ctx, op, z_mask, s_mask);
2585 }
2586 
2587 static bool fold_shift(OptContext *ctx, TCGOp *op)
2588 {
2589     uint64_t s_mask, z_mask;
2590     TempOptInfo *t1, *t2;
2591 
2592     if (fold_const2(ctx, op) ||
2593         fold_ix_to_i(ctx, op, 0) ||
2594         fold_xi_to_x(ctx, op, 0)) {
2595         return true;
2596     }
2597 
2598     t1 = arg_info(op->args[1]);
2599     t2 = arg_info(op->args[2]);
2600     s_mask = t1->s_mask;
2601     z_mask = t1->z_mask;
2602 
2603     if (ti_is_const(t2)) {
2604         int sh = ti_const_val(t2);
2605 
2606         z_mask = do_constant_folding(op->opc, ctx->type, z_mask, sh);
2607         s_mask = do_constant_folding(op->opc, ctx->type, s_mask, sh);
2608 
2609         return fold_masks_zs(ctx, op, z_mask, s_mask);
2610     }
2611 
2612     switch (op->opc) {
2613     CASE_OP_32_64(sar):
2614         /*
2615          * Arithmetic right shift will not reduce the number of
2616          * input sign repetitions.
2617          */
2618         return fold_masks_s(ctx, op, s_mask);
2619     CASE_OP_32_64(shr):
2620         /*
2621          * If the sign bit is known zero, then logical right shift
2622          * will not reduce the number of input sign repetitions.
2623          */
2624         if (~z_mask & -s_mask) {
2625             return fold_masks_s(ctx, op, s_mask);
2626         }
2627         break;
2628     default:
2629         break;
2630     }
2631 
2632     return finish_folding(ctx, op);
2633 }
2634 
2635 static bool fold_sub_to_neg(OptContext *ctx, TCGOp *op)
2636 {
2637     TCGOpcode neg_op;
2638     bool have_neg;
2639 
2640     if (!arg_is_const(op->args[1]) || arg_info(op->args[1])->val != 0) {
2641         return false;
2642     }
2643 
2644     switch (ctx->type) {
2645     case TCG_TYPE_I32:
2646         neg_op = INDEX_op_neg_i32;
2647         have_neg = true;
2648         break;
2649     case TCG_TYPE_I64:
2650         neg_op = INDEX_op_neg_i64;
2651         have_neg = true;
2652         break;
2653     case TCG_TYPE_V64:
2654     case TCG_TYPE_V128:
2655     case TCG_TYPE_V256:
2656         neg_op = INDEX_op_neg_vec;
2657         have_neg = (TCG_TARGET_HAS_neg_vec &&
2658                     tcg_can_emit_vec_op(neg_op, ctx->type, TCGOP_VECE(op)) > 0);
2659         break;
2660     default:
2661         g_assert_not_reached();
2662     }
2663     if (have_neg) {
2664         op->opc = neg_op;
2665         op->args[1] = op->args[2];
2666         return fold_neg_no_const(ctx, op);
2667     }
2668     return false;
2669 }
2670 
2671 /* We cannot as yet do_constant_folding with vectors. */
2672 static bool fold_sub_vec(OptContext *ctx, TCGOp *op)
2673 {
2674     if (fold_xx_to_i(ctx, op, 0) ||
2675         fold_xi_to_x(ctx, op, 0) ||
2676         fold_sub_to_neg(ctx, op)) {
2677         return true;
2678     }
2679     return finish_folding(ctx, op);
2680 }
2681 
2682 static bool fold_sub(OptContext *ctx, TCGOp *op)
2683 {
2684     if (fold_const2(ctx, op) ||
2685         fold_xx_to_i(ctx, op, 0) ||
2686         fold_xi_to_x(ctx, op, 0) ||
2687         fold_sub_to_neg(ctx, op)) {
2688         return true;
2689     }
2690 
2691     /* Fold sub r,x,i to add r,x,-i */
2692     if (arg_is_const(op->args[2])) {
2693         uint64_t val = arg_info(op->args[2])->val;
2694 
2695         op->opc = INDEX_op_add;
2696         op->args[2] = arg_new_constant(ctx, -val);
2697     }
2698     return finish_folding(ctx, op);
2699 }
2700 
2701 static bool fold_sub2(OptContext *ctx, TCGOp *op)
2702 {
2703     return fold_addsub2(ctx, op, false);
2704 }
2705 
2706 static bool fold_tcg_ld(OptContext *ctx, TCGOp *op)
2707 {
2708     uint64_t z_mask = -1, s_mask = 0;
2709 
2710     /* We can't do any folding with a load, but we can record bits. */
2711     switch (op->opc) {
2712     CASE_OP_32_64(ld8s):
2713         s_mask = INT8_MIN;
2714         break;
2715     CASE_OP_32_64(ld8u):
2716         z_mask = MAKE_64BIT_MASK(0, 8);
2717         break;
2718     CASE_OP_32_64(ld16s):
2719         s_mask = INT16_MIN;
2720         break;
2721     CASE_OP_32_64(ld16u):
2722         z_mask = MAKE_64BIT_MASK(0, 16);
2723         break;
2724     case INDEX_op_ld32s_i64:
2725         s_mask = INT32_MIN;
2726         break;
2727     case INDEX_op_ld32u_i64:
2728         z_mask = MAKE_64BIT_MASK(0, 32);
2729         break;
2730     default:
2731         g_assert_not_reached();
2732     }
2733     return fold_masks_zs(ctx, op, z_mask, s_mask);
2734 }
2735 
2736 static bool fold_tcg_ld_memcopy(OptContext *ctx, TCGOp *op)
2737 {
2738     TCGTemp *dst, *src;
2739     intptr_t ofs;
2740     TCGType type;
2741 
2742     if (op->args[1] != tcgv_ptr_arg(tcg_env)) {
2743         return finish_folding(ctx, op);
2744     }
2745 
2746     type = ctx->type;
2747     ofs = op->args[2];
2748     dst = arg_temp(op->args[0]);
2749     src = find_mem_copy_for(ctx, type, ofs);
2750     if (src && src->base_type == type) {
2751         return tcg_opt_gen_mov(ctx, op, temp_arg(dst), temp_arg(src));
2752     }
2753 
2754     reset_ts(ctx, dst);
2755     record_mem_copy(ctx, type, dst, ofs, ofs + tcg_type_size(type) - 1);
2756     return true;
2757 }
2758 
2759 static bool fold_tcg_st(OptContext *ctx, TCGOp *op)
2760 {
2761     intptr_t ofs = op->args[2];
2762     intptr_t lm1;
2763 
2764     if (op->args[1] != tcgv_ptr_arg(tcg_env)) {
2765         remove_mem_copy_all(ctx);
2766         return true;
2767     }
2768 
2769     switch (op->opc) {
2770     CASE_OP_32_64(st8):
2771         lm1 = 0;
2772         break;
2773     CASE_OP_32_64(st16):
2774         lm1 = 1;
2775         break;
2776     case INDEX_op_st32_i64:
2777     case INDEX_op_st_i32:
2778         lm1 = 3;
2779         break;
2780     case INDEX_op_st_i64:
2781         lm1 = 7;
2782         break;
2783     case INDEX_op_st_vec:
2784         lm1 = tcg_type_size(ctx->type) - 1;
2785         break;
2786     default:
2787         g_assert_not_reached();
2788     }
2789     remove_mem_copy_in(ctx, ofs, ofs + lm1);
2790     return true;
2791 }
2792 
2793 static bool fold_tcg_st_memcopy(OptContext *ctx, TCGOp *op)
2794 {
2795     TCGTemp *src;
2796     intptr_t ofs, last;
2797     TCGType type;
2798 
2799     if (op->args[1] != tcgv_ptr_arg(tcg_env)) {
2800         return fold_tcg_st(ctx, op);
2801     }
2802 
2803     src = arg_temp(op->args[0]);
2804     ofs = op->args[2];
2805     type = ctx->type;
2806 
2807     /*
2808      * Eliminate duplicate stores of a constant.
2809      * This happens frequently when the target ISA zero-extends.
2810      */
2811     if (ts_is_const(src)) {
2812         TCGTemp *prev = find_mem_copy_for(ctx, type, ofs);
2813         if (src == prev) {
2814             tcg_op_remove(ctx->tcg, op);
2815             return true;
2816         }
2817     }
2818 
2819     last = ofs + tcg_type_size(type) - 1;
2820     remove_mem_copy_in(ctx, ofs, last);
2821     record_mem_copy(ctx, type, src, ofs, last);
2822     return true;
2823 }
2824 
2825 static bool fold_xor(OptContext *ctx, TCGOp *op)
2826 {
2827     uint64_t z_mask, s_mask;
2828     TempOptInfo *t1, *t2;
2829 
2830     if (fold_const2_commutative(ctx, op) ||
2831         fold_xx_to_i(ctx, op, 0) ||
2832         fold_xi_to_x(ctx, op, 0) ||
2833         fold_xi_to_not(ctx, op, -1)) {
2834         return true;
2835     }
2836 
2837     t1 = arg_info(op->args[1]);
2838     t2 = arg_info(op->args[2]);
2839     z_mask = t1->z_mask | t2->z_mask;
2840     s_mask = t1->s_mask & t2->s_mask;
2841     return fold_masks_zs(ctx, op, z_mask, s_mask);
2842 }
2843 
2844 /* Propagate constants and copies, fold constant expressions. */
2845 void tcg_optimize(TCGContext *s)
2846 {
2847     int nb_temps, i;
2848     TCGOp *op, *op_next;
2849     OptContext ctx = { .tcg = s };
2850 
2851     QSIMPLEQ_INIT(&ctx.mem_free);
2852 
2853     /* Array VALS has an element for each temp.
2854        If this temp holds a constant then its value is kept in VALS' element.
2855        If this temp is a copy of other ones then the other copies are
2856        available through the doubly linked circular list. */
2857 
2858     nb_temps = s->nb_temps;
2859     for (i = 0; i < nb_temps; ++i) {
2860         s->temps[i].state_ptr = NULL;
2861     }
2862 
2863     QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) {
2864         TCGOpcode opc = op->opc;
2865         const TCGOpDef *def;
2866         bool done = false;
2867 
2868         /* Calls are special. */
2869         if (opc == INDEX_op_call) {
2870             fold_call(&ctx, op);
2871             continue;
2872         }
2873 
2874         def = &tcg_op_defs[opc];
2875         init_arguments(&ctx, op, def->nb_oargs + def->nb_iargs);
2876         copy_propagate(&ctx, op, def->nb_oargs, def->nb_iargs);
2877 
2878         /* Pre-compute the type of the operation. */
2879         ctx.type = TCGOP_TYPE(op);
2880 
2881         /*
2882          * Process each opcode.
2883          * Sorted alphabetically by opcode as much as possible.
2884          */
2885         switch (opc) {
2886         case INDEX_op_add:
2887             done = fold_add(&ctx, op);
2888             break;
2889         case INDEX_op_add_vec:
2890             done = fold_add_vec(&ctx, op);
2891             break;
2892         CASE_OP_32_64(add2):
2893             done = fold_add2(&ctx, op);
2894             break;
2895         case INDEX_op_and:
2896         case INDEX_op_and_vec:
2897             done = fold_and(&ctx, op);
2898             break;
2899         case INDEX_op_andc:
2900         case INDEX_op_andc_vec:
2901             done = fold_andc(&ctx, op);
2902             break;
2903         CASE_OP_32_64(brcond):
2904             done = fold_brcond(&ctx, op);
2905             break;
2906         case INDEX_op_brcond2_i32:
2907             done = fold_brcond2(&ctx, op);
2908             break;
2909         CASE_OP_32_64(bswap16):
2910         CASE_OP_32_64(bswap32):
2911         case INDEX_op_bswap64_i64:
2912             done = fold_bswap(&ctx, op);
2913             break;
2914         CASE_OP_32_64(clz):
2915         CASE_OP_32_64(ctz):
2916             done = fold_count_zeros(&ctx, op);
2917             break;
2918         CASE_OP_32_64(ctpop):
2919             done = fold_ctpop(&ctx, op);
2920             break;
2921         CASE_OP_32_64(deposit):
2922             done = fold_deposit(&ctx, op);
2923             break;
2924         CASE_OP_32_64(div):
2925         CASE_OP_32_64(divu):
2926             done = fold_divide(&ctx, op);
2927             break;
2928         case INDEX_op_dup_vec:
2929             done = fold_dup(&ctx, op);
2930             break;
2931         case INDEX_op_dup2_vec:
2932             done = fold_dup2(&ctx, op);
2933             break;
2934         case INDEX_op_eqv:
2935         case INDEX_op_eqv_vec:
2936             done = fold_eqv(&ctx, op);
2937             break;
2938         CASE_OP_32_64(extract):
2939             done = fold_extract(&ctx, op);
2940             break;
2941         CASE_OP_32_64(extract2):
2942             done = fold_extract2(&ctx, op);
2943             break;
2944         case INDEX_op_ext_i32_i64:
2945             done = fold_exts(&ctx, op);
2946             break;
2947         case INDEX_op_extu_i32_i64:
2948         case INDEX_op_extrl_i64_i32:
2949         case INDEX_op_extrh_i64_i32:
2950             done = fold_extu(&ctx, op);
2951             break;
2952         CASE_OP_32_64(ld8s):
2953         CASE_OP_32_64(ld8u):
2954         CASE_OP_32_64(ld16s):
2955         CASE_OP_32_64(ld16u):
2956         case INDEX_op_ld32s_i64:
2957         case INDEX_op_ld32u_i64:
2958             done = fold_tcg_ld(&ctx, op);
2959             break;
2960         case INDEX_op_ld_i32:
2961         case INDEX_op_ld_i64:
2962         case INDEX_op_ld_vec:
2963             done = fold_tcg_ld_memcopy(&ctx, op);
2964             break;
2965         CASE_OP_32_64(st8):
2966         CASE_OP_32_64(st16):
2967         case INDEX_op_st32_i64:
2968             done = fold_tcg_st(&ctx, op);
2969             break;
2970         case INDEX_op_st_i32:
2971         case INDEX_op_st_i64:
2972         case INDEX_op_st_vec:
2973             done = fold_tcg_st_memcopy(&ctx, op);
2974             break;
2975         case INDEX_op_mb:
2976             done = fold_mb(&ctx, op);
2977             break;
2978         case INDEX_op_mov:
2979         case INDEX_op_mov_vec:
2980             done = fold_mov(&ctx, op);
2981             break;
2982         CASE_OP_32_64(movcond):
2983             done = fold_movcond(&ctx, op);
2984             break;
2985         CASE_OP_32_64(mul):
2986             done = fold_mul(&ctx, op);
2987             break;
2988         CASE_OP_32_64(mulsh):
2989         CASE_OP_32_64(muluh):
2990             done = fold_mul_highpart(&ctx, op);
2991             break;
2992         CASE_OP_32_64(muls2):
2993         CASE_OP_32_64(mulu2):
2994             done = fold_multiply2(&ctx, op);
2995             break;
2996         case INDEX_op_nand:
2997         case INDEX_op_nand_vec:
2998             done = fold_nand(&ctx, op);
2999             break;
3000         CASE_OP_32_64(neg):
3001             done = fold_neg(&ctx, op);
3002             break;
3003         CASE_OP_32_64_VEC(nor):
3004             done = fold_nor(&ctx, op);
3005             break;
3006         CASE_OP_32_64_VEC(not):
3007             done = fold_not(&ctx, op);
3008             break;
3009         case INDEX_op_or:
3010         case INDEX_op_or_vec:
3011             done = fold_or(&ctx, op);
3012             break;
3013         case INDEX_op_orc:
3014         case INDEX_op_orc_vec:
3015             done = fold_orc(&ctx, op);
3016             break;
3017         case INDEX_op_qemu_ld_i32:
3018             done = fold_qemu_ld_1reg(&ctx, op);
3019             break;
3020         case INDEX_op_qemu_ld_i64:
3021             if (TCG_TARGET_REG_BITS == 64) {
3022                 done = fold_qemu_ld_1reg(&ctx, op);
3023                 break;
3024             }
3025             QEMU_FALLTHROUGH;
3026         case INDEX_op_qemu_ld_i128:
3027             done = fold_qemu_ld_2reg(&ctx, op);
3028             break;
3029         case INDEX_op_qemu_st8_i32:
3030         case INDEX_op_qemu_st_i32:
3031         case INDEX_op_qemu_st_i64:
3032         case INDEX_op_qemu_st_i128:
3033             done = fold_qemu_st(&ctx, op);
3034             break;
3035         CASE_OP_32_64(rem):
3036         CASE_OP_32_64(remu):
3037             done = fold_remainder(&ctx, op);
3038             break;
3039         CASE_OP_32_64(rotl):
3040         CASE_OP_32_64(rotr):
3041         CASE_OP_32_64(sar):
3042         CASE_OP_32_64(shl):
3043         CASE_OP_32_64(shr):
3044             done = fold_shift(&ctx, op);
3045             break;
3046         CASE_OP_32_64(setcond):
3047             done = fold_setcond(&ctx, op);
3048             break;
3049         CASE_OP_32_64(negsetcond):
3050             done = fold_negsetcond(&ctx, op);
3051             break;
3052         case INDEX_op_setcond2_i32:
3053             done = fold_setcond2(&ctx, op);
3054             break;
3055         case INDEX_op_cmp_vec:
3056             done = fold_cmp_vec(&ctx, op);
3057             break;
3058         case INDEX_op_cmpsel_vec:
3059             done = fold_cmpsel_vec(&ctx, op);
3060             break;
3061         case INDEX_op_bitsel_vec:
3062             done = fold_bitsel_vec(&ctx, op);
3063             break;
3064         CASE_OP_32_64(sextract):
3065             done = fold_sextract(&ctx, op);
3066             break;
3067         CASE_OP_32_64(sub):
3068             done = fold_sub(&ctx, op);
3069             break;
3070         case INDEX_op_sub_vec:
3071             done = fold_sub_vec(&ctx, op);
3072             break;
3073         CASE_OP_32_64(sub2):
3074             done = fold_sub2(&ctx, op);
3075             break;
3076         case INDEX_op_xor:
3077         case INDEX_op_xor_vec:
3078             done = fold_xor(&ctx, op);
3079             break;
3080         case INDEX_op_set_label:
3081         case INDEX_op_br:
3082         case INDEX_op_exit_tb:
3083         case INDEX_op_goto_tb:
3084         case INDEX_op_goto_ptr:
3085             finish_ebb(&ctx);
3086             done = true;
3087             break;
3088         default:
3089             done = finish_folding(&ctx, op);
3090             break;
3091         }
3092         tcg_debug_assert(done);
3093     }
3094 }
3095