xref: /openbmc/qemu/tcg/optimize.c (revision 46f96bff163512f9f8f9959de4a18c0799001422)
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_OP_32_64_VEC(or):
441         return x | y;
442 
443     CASE_OP_32_64_VEC(xor):
444         return x ^ y;
445 
446     case INDEX_op_shl_i32:
447         return (uint32_t)x << (y & 31);
448 
449     case INDEX_op_shl_i64:
450         return (uint64_t)x << (y & 63);
451 
452     case INDEX_op_shr_i32:
453         return (uint32_t)x >> (y & 31);
454 
455     case INDEX_op_shr_i64:
456         return (uint64_t)x >> (y & 63);
457 
458     case INDEX_op_sar_i32:
459         return (int32_t)x >> (y & 31);
460 
461     case INDEX_op_sar_i64:
462         return (int64_t)x >> (y & 63);
463 
464     case INDEX_op_rotr_i32:
465         return ror32(x, y & 31);
466 
467     case INDEX_op_rotr_i64:
468         return ror64(x, y & 63);
469 
470     case INDEX_op_rotl_i32:
471         return rol32(x, y & 31);
472 
473     case INDEX_op_rotl_i64:
474         return rol64(x, y & 63);
475 
476     CASE_OP_32_64_VEC(not):
477         return ~x;
478 
479     CASE_OP_32_64(neg):
480         return -x;
481 
482     case INDEX_op_andc:
483     case INDEX_op_andc_vec:
484         return x & ~y;
485 
486     CASE_OP_32_64_VEC(orc):
487         return x | ~y;
488 
489     CASE_OP_32_64_VEC(eqv):
490         return ~(x ^ y);
491 
492     CASE_OP_32_64_VEC(nand):
493         return ~(x & y);
494 
495     CASE_OP_32_64_VEC(nor):
496         return ~(x | y);
497 
498     case INDEX_op_clz_i32:
499         return (uint32_t)x ? clz32(x) : y;
500 
501     case INDEX_op_clz_i64:
502         return x ? clz64(x) : y;
503 
504     case INDEX_op_ctz_i32:
505         return (uint32_t)x ? ctz32(x) : y;
506 
507     case INDEX_op_ctz_i64:
508         return x ? ctz64(x) : y;
509 
510     case INDEX_op_ctpop_i32:
511         return ctpop32(x);
512 
513     case INDEX_op_ctpop_i64:
514         return ctpop64(x);
515 
516     CASE_OP_32_64(bswap16):
517         x = bswap16(x);
518         return y & TCG_BSWAP_OS ? (int16_t)x : x;
519 
520     CASE_OP_32_64(bswap32):
521         x = bswap32(x);
522         return y & TCG_BSWAP_OS ? (int32_t)x : x;
523 
524     case INDEX_op_bswap64_i64:
525         return bswap64(x);
526 
527     case INDEX_op_ext_i32_i64:
528         return (int32_t)x;
529 
530     case INDEX_op_extu_i32_i64:
531     case INDEX_op_extrl_i64_i32:
532         return (uint32_t)x;
533 
534     case INDEX_op_extrh_i64_i32:
535         return (uint64_t)x >> 32;
536 
537     case INDEX_op_muluh_i32:
538         return ((uint64_t)(uint32_t)x * (uint32_t)y) >> 32;
539     case INDEX_op_mulsh_i32:
540         return ((int64_t)(int32_t)x * (int32_t)y) >> 32;
541 
542     case INDEX_op_muluh_i64:
543         mulu64(&l64, &h64, x, y);
544         return h64;
545     case INDEX_op_mulsh_i64:
546         muls64(&l64, &h64, x, y);
547         return h64;
548 
549     case INDEX_op_div_i32:
550         /* Avoid crashing on divide by zero, otherwise undefined.  */
551         return (int32_t)x / ((int32_t)y ? : 1);
552     case INDEX_op_divu_i32:
553         return (uint32_t)x / ((uint32_t)y ? : 1);
554     case INDEX_op_div_i64:
555         return (int64_t)x / ((int64_t)y ? : 1);
556     case INDEX_op_divu_i64:
557         return (uint64_t)x / ((uint64_t)y ? : 1);
558 
559     case INDEX_op_rem_i32:
560         return (int32_t)x % ((int32_t)y ? : 1);
561     case INDEX_op_remu_i32:
562         return (uint32_t)x % ((uint32_t)y ? : 1);
563     case INDEX_op_rem_i64:
564         return (int64_t)x % ((int64_t)y ? : 1);
565     case INDEX_op_remu_i64:
566         return (uint64_t)x % ((uint64_t)y ? : 1);
567 
568     default:
569         g_assert_not_reached();
570     }
571 }
572 
573 static uint64_t do_constant_folding(TCGOpcode op, TCGType type,
574                                     uint64_t x, uint64_t y)
575 {
576     uint64_t res = do_constant_folding_2(op, x, y);
577     if (type == TCG_TYPE_I32) {
578         res = (int32_t)res;
579     }
580     return res;
581 }
582 
583 static bool do_constant_folding_cond_32(uint32_t x, uint32_t y, TCGCond c)
584 {
585     switch (c) {
586     case TCG_COND_EQ:
587         return x == y;
588     case TCG_COND_NE:
589         return x != y;
590     case TCG_COND_LT:
591         return (int32_t)x < (int32_t)y;
592     case TCG_COND_GE:
593         return (int32_t)x >= (int32_t)y;
594     case TCG_COND_LE:
595         return (int32_t)x <= (int32_t)y;
596     case TCG_COND_GT:
597         return (int32_t)x > (int32_t)y;
598     case TCG_COND_LTU:
599         return x < y;
600     case TCG_COND_GEU:
601         return x >= y;
602     case TCG_COND_LEU:
603         return x <= y;
604     case TCG_COND_GTU:
605         return x > y;
606     case TCG_COND_TSTEQ:
607         return (x & y) == 0;
608     case TCG_COND_TSTNE:
609         return (x & y) != 0;
610     case TCG_COND_ALWAYS:
611     case TCG_COND_NEVER:
612         break;
613     }
614     g_assert_not_reached();
615 }
616 
617 static bool do_constant_folding_cond_64(uint64_t x, uint64_t y, TCGCond c)
618 {
619     switch (c) {
620     case TCG_COND_EQ:
621         return x == y;
622     case TCG_COND_NE:
623         return x != y;
624     case TCG_COND_LT:
625         return (int64_t)x < (int64_t)y;
626     case TCG_COND_GE:
627         return (int64_t)x >= (int64_t)y;
628     case TCG_COND_LE:
629         return (int64_t)x <= (int64_t)y;
630     case TCG_COND_GT:
631         return (int64_t)x > (int64_t)y;
632     case TCG_COND_LTU:
633         return x < y;
634     case TCG_COND_GEU:
635         return x >= y;
636     case TCG_COND_LEU:
637         return x <= y;
638     case TCG_COND_GTU:
639         return x > y;
640     case TCG_COND_TSTEQ:
641         return (x & y) == 0;
642     case TCG_COND_TSTNE:
643         return (x & y) != 0;
644     case TCG_COND_ALWAYS:
645     case TCG_COND_NEVER:
646         break;
647     }
648     g_assert_not_reached();
649 }
650 
651 static int do_constant_folding_cond_eq(TCGCond c)
652 {
653     switch (c) {
654     case TCG_COND_GT:
655     case TCG_COND_LTU:
656     case TCG_COND_LT:
657     case TCG_COND_GTU:
658     case TCG_COND_NE:
659         return 0;
660     case TCG_COND_GE:
661     case TCG_COND_GEU:
662     case TCG_COND_LE:
663     case TCG_COND_LEU:
664     case TCG_COND_EQ:
665         return 1;
666     case TCG_COND_TSTEQ:
667     case TCG_COND_TSTNE:
668         return -1;
669     case TCG_COND_ALWAYS:
670     case TCG_COND_NEVER:
671         break;
672     }
673     g_assert_not_reached();
674 }
675 
676 /*
677  * Return -1 if the condition can't be simplified,
678  * and the result of the condition (0 or 1) if it can.
679  */
680 static int do_constant_folding_cond(TCGType type, TCGArg x,
681                                     TCGArg y, TCGCond c)
682 {
683     if (arg_is_const(x) && arg_is_const(y)) {
684         uint64_t xv = arg_info(x)->val;
685         uint64_t yv = arg_info(y)->val;
686 
687         switch (type) {
688         case TCG_TYPE_I32:
689             return do_constant_folding_cond_32(xv, yv, c);
690         case TCG_TYPE_I64:
691             return do_constant_folding_cond_64(xv, yv, c);
692         default:
693             /* Only scalar comparisons are optimizable */
694             return -1;
695         }
696     } else if (args_are_copies(x, y)) {
697         return do_constant_folding_cond_eq(c);
698     } else if (arg_is_const_val(y, 0)) {
699         switch (c) {
700         case TCG_COND_LTU:
701         case TCG_COND_TSTNE:
702             return 0;
703         case TCG_COND_GEU:
704         case TCG_COND_TSTEQ:
705             return 1;
706         default:
707             return -1;
708         }
709     }
710     return -1;
711 }
712 
713 /**
714  * swap_commutative:
715  * @dest: TCGArg of the destination argument, or NO_DEST.
716  * @p1: first paired argument
717  * @p2: second paired argument
718  *
719  * If *@p1 is a constant and *@p2 is not, swap.
720  * If *@p2 matches @dest, swap.
721  * Return true if a swap was performed.
722  */
723 
724 #define NO_DEST  temp_arg(NULL)
725 
726 static bool swap_commutative(TCGArg dest, TCGArg *p1, TCGArg *p2)
727 {
728     TCGArg a1 = *p1, a2 = *p2;
729     int sum = 0;
730     sum += arg_is_const(a1);
731     sum -= arg_is_const(a2);
732 
733     /* Prefer the constant in second argument, and then the form
734        op a, a, b, which is better handled on non-RISC hosts. */
735     if (sum > 0 || (sum == 0 && dest == a2)) {
736         *p1 = a2;
737         *p2 = a1;
738         return true;
739     }
740     return false;
741 }
742 
743 static bool swap_commutative2(TCGArg *p1, TCGArg *p2)
744 {
745     int sum = 0;
746     sum += arg_is_const(p1[0]);
747     sum += arg_is_const(p1[1]);
748     sum -= arg_is_const(p2[0]);
749     sum -= arg_is_const(p2[1]);
750     if (sum > 0) {
751         TCGArg t;
752         t = p1[0], p1[0] = p2[0], p2[0] = t;
753         t = p1[1], p1[1] = p2[1], p2[1] = t;
754         return true;
755     }
756     return false;
757 }
758 
759 /*
760  * Return -1 if the condition can't be simplified,
761  * and the result of the condition (0 or 1) if it can.
762  */
763 static int do_constant_folding_cond1(OptContext *ctx, TCGOp *op, TCGArg dest,
764                                      TCGArg *p1, TCGArg *p2, TCGArg *pcond)
765 {
766     TCGCond cond;
767     TempOptInfo *i1;
768     bool swap;
769     int r;
770 
771     swap = swap_commutative(dest, p1, p2);
772     cond = *pcond;
773     if (swap) {
774         *pcond = cond = tcg_swap_cond(cond);
775     }
776 
777     r = do_constant_folding_cond(ctx->type, *p1, *p2, cond);
778     if (r >= 0) {
779         return r;
780     }
781     if (!is_tst_cond(cond)) {
782         return -1;
783     }
784 
785     i1 = arg_info(*p1);
786 
787     /*
788      * TSTNE x,x -> NE x,0
789      * TSTNE x,i -> NE x,0 if i includes all nonzero bits of x
790      */
791     if (args_are_copies(*p1, *p2) ||
792         (arg_is_const(*p2) && (i1->z_mask & ~arg_info(*p2)->val) == 0)) {
793         *p2 = arg_new_constant(ctx, 0);
794         *pcond = tcg_tst_eqne_cond(cond);
795         return -1;
796     }
797 
798     /* TSTNE x,i -> LT x,0 if i only includes sign bit copies */
799     if (arg_is_const(*p2) && (arg_info(*p2)->val & ~i1->s_mask) == 0) {
800         *p2 = arg_new_constant(ctx, 0);
801         *pcond = tcg_tst_ltge_cond(cond);
802         return -1;
803     }
804 
805     /* Expand to AND with a temporary if no backend support. */
806     if (!TCG_TARGET_HAS_tst) {
807         TCGOp *op2 = opt_insert_before(ctx, op, INDEX_op_and, 3);
808         TCGArg tmp = arg_new_temp(ctx);
809 
810         op2->args[0] = tmp;
811         op2->args[1] = *p1;
812         op2->args[2] = *p2;
813 
814         *p1 = tmp;
815         *p2 = arg_new_constant(ctx, 0);
816         *pcond = tcg_tst_eqne_cond(cond);
817     }
818     return -1;
819 }
820 
821 static int do_constant_folding_cond2(OptContext *ctx, TCGOp *op, TCGArg *args)
822 {
823     TCGArg al, ah, bl, bh;
824     TCGCond c;
825     bool swap;
826     int r;
827 
828     swap = swap_commutative2(args, args + 2);
829     c = args[4];
830     if (swap) {
831         args[4] = c = tcg_swap_cond(c);
832     }
833 
834     al = args[0];
835     ah = args[1];
836     bl = args[2];
837     bh = args[3];
838 
839     if (arg_is_const(bl) && arg_is_const(bh)) {
840         tcg_target_ulong blv = arg_info(bl)->val;
841         tcg_target_ulong bhv = arg_info(bh)->val;
842         uint64_t b = deposit64(blv, 32, 32, bhv);
843 
844         if (arg_is_const(al) && arg_is_const(ah)) {
845             tcg_target_ulong alv = arg_info(al)->val;
846             tcg_target_ulong ahv = arg_info(ah)->val;
847             uint64_t a = deposit64(alv, 32, 32, ahv);
848 
849             r = do_constant_folding_cond_64(a, b, c);
850             if (r >= 0) {
851                 return r;
852             }
853         }
854 
855         if (b == 0) {
856             switch (c) {
857             case TCG_COND_LTU:
858             case TCG_COND_TSTNE:
859                 return 0;
860             case TCG_COND_GEU:
861             case TCG_COND_TSTEQ:
862                 return 1;
863             default:
864                 break;
865             }
866         }
867 
868         /* TSTNE x,-1 -> NE x,0 */
869         if (b == -1 && is_tst_cond(c)) {
870             args[3] = args[2] = arg_new_constant(ctx, 0);
871             args[4] = tcg_tst_eqne_cond(c);
872             return -1;
873         }
874 
875         /* TSTNE x,sign -> LT x,0 */
876         if (b == INT64_MIN && is_tst_cond(c)) {
877             /* bl must be 0, so copy that to bh */
878             args[3] = bl;
879             args[4] = tcg_tst_ltge_cond(c);
880             return -1;
881         }
882     }
883 
884     if (args_are_copies(al, bl) && args_are_copies(ah, bh)) {
885         r = do_constant_folding_cond_eq(c);
886         if (r >= 0) {
887             return r;
888         }
889 
890         /* TSTNE x,x -> NE x,0 */
891         if (is_tst_cond(c)) {
892             args[3] = args[2] = arg_new_constant(ctx, 0);
893             args[4] = tcg_tst_eqne_cond(c);
894             return -1;
895         }
896     }
897 
898     /* Expand to AND with a temporary if no backend support. */
899     if (!TCG_TARGET_HAS_tst && is_tst_cond(c)) {
900         TCGOp *op1 = opt_insert_before(ctx, op, INDEX_op_and, 3);
901         TCGOp *op2 = opt_insert_before(ctx, op, INDEX_op_and, 3);
902         TCGArg t1 = arg_new_temp(ctx);
903         TCGArg t2 = arg_new_temp(ctx);
904 
905         op1->args[0] = t1;
906         op1->args[1] = al;
907         op1->args[2] = bl;
908         op2->args[0] = t2;
909         op2->args[1] = ah;
910         op2->args[2] = bh;
911 
912         args[0] = t1;
913         args[1] = t2;
914         args[3] = args[2] = arg_new_constant(ctx, 0);
915         args[4] = tcg_tst_eqne_cond(c);
916     }
917     return -1;
918 }
919 
920 static void init_arguments(OptContext *ctx, TCGOp *op, int nb_args)
921 {
922     for (int i = 0; i < nb_args; i++) {
923         TCGTemp *ts = arg_temp(op->args[i]);
924         init_ts_info(ctx, ts);
925     }
926 }
927 
928 static void copy_propagate(OptContext *ctx, TCGOp *op,
929                            int nb_oargs, int nb_iargs)
930 {
931     for (int i = nb_oargs; i < nb_oargs + nb_iargs; i++) {
932         TCGTemp *ts = arg_temp(op->args[i]);
933         if (ts_is_copy(ts)) {
934             op->args[i] = temp_arg(find_better_copy(ts));
935         }
936     }
937 }
938 
939 static void finish_bb(OptContext *ctx)
940 {
941     /* We only optimize memory barriers across basic blocks. */
942     ctx->prev_mb = NULL;
943 }
944 
945 static void finish_ebb(OptContext *ctx)
946 {
947     finish_bb(ctx);
948     /* We only optimize across extended basic blocks. */
949     memset(&ctx->temps_used, 0, sizeof(ctx->temps_used));
950     remove_mem_copy_all(ctx);
951 }
952 
953 static bool finish_folding(OptContext *ctx, TCGOp *op)
954 {
955     const TCGOpDef *def = &tcg_op_defs[op->opc];
956     int i, nb_oargs;
957 
958     nb_oargs = def->nb_oargs;
959     for (i = 0; i < nb_oargs; i++) {
960         TCGTemp *ts = arg_temp(op->args[i]);
961         reset_ts(ctx, ts);
962     }
963     return true;
964 }
965 
966 /*
967  * The fold_* functions return true when processing is complete,
968  * usually by folding the operation to a constant or to a copy,
969  * and calling tcg_opt_gen_{mov,movi}.  They may do other things,
970  * like collect information about the value produced, for use in
971  * optimizing a subsequent operation.
972  *
973  * These first fold_* functions are all helpers, used by other
974  * folders for more specific operations.
975  */
976 
977 static bool fold_const1(OptContext *ctx, TCGOp *op)
978 {
979     if (arg_is_const(op->args[1])) {
980         uint64_t t;
981 
982         t = arg_info(op->args[1])->val;
983         t = do_constant_folding(op->opc, ctx->type, t, 0);
984         return tcg_opt_gen_movi(ctx, op, op->args[0], t);
985     }
986     return false;
987 }
988 
989 static bool fold_const2(OptContext *ctx, TCGOp *op)
990 {
991     if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) {
992         uint64_t t1 = arg_info(op->args[1])->val;
993         uint64_t t2 = arg_info(op->args[2])->val;
994 
995         t1 = do_constant_folding(op->opc, ctx->type, t1, t2);
996         return tcg_opt_gen_movi(ctx, op, op->args[0], t1);
997     }
998     return false;
999 }
1000 
1001 static bool fold_commutative(OptContext *ctx, TCGOp *op)
1002 {
1003     swap_commutative(op->args[0], &op->args[1], &op->args[2]);
1004     return false;
1005 }
1006 
1007 static bool fold_const2_commutative(OptContext *ctx, TCGOp *op)
1008 {
1009     swap_commutative(op->args[0], &op->args[1], &op->args[2]);
1010     return fold_const2(ctx, op);
1011 }
1012 
1013 /*
1014  * Record "zero" and "sign" masks for the single output of @op.
1015  * See TempOptInfo definition of z_mask and s_mask.
1016  * If z_mask allows, fold the output to constant zero.
1017  * The passed s_mask may be augmented by z_mask.
1018  */
1019 static bool fold_masks_zs(OptContext *ctx, TCGOp *op,
1020                           uint64_t z_mask, int64_t s_mask)
1021 {
1022     const TCGOpDef *def = &tcg_op_defs[op->opc];
1023     TCGTemp *ts;
1024     TempOptInfo *ti;
1025     int rep;
1026 
1027     /* Only single-output opcodes are supported here. */
1028     tcg_debug_assert(def->nb_oargs == 1);
1029 
1030     /*
1031      * 32-bit ops generate 32-bit results, which for the purpose of
1032      * simplifying tcg are sign-extended.  Certainly that's how we
1033      * represent our constants elsewhere.  Note that the bits will
1034      * be reset properly for a 64-bit value when encountering the
1035      * type changing opcodes.
1036      */
1037     if (ctx->type == TCG_TYPE_I32) {
1038         z_mask = (int32_t)z_mask;
1039         s_mask |= INT32_MIN;
1040     }
1041 
1042     if (z_mask == 0) {
1043         return tcg_opt_gen_movi(ctx, op, op->args[0], 0);
1044     }
1045 
1046     ts = arg_temp(op->args[0]);
1047     reset_ts(ctx, ts);
1048 
1049     ti = ts_info(ts);
1050     ti->z_mask = z_mask;
1051 
1052     /* Canonicalize s_mask and incorporate data from z_mask. */
1053     rep = clz64(~s_mask);
1054     rep = MAX(rep, clz64(z_mask));
1055     rep = MAX(rep - 1, 0);
1056     ti->s_mask = INT64_MIN >> rep;
1057 
1058     return true;
1059 }
1060 
1061 static bool fold_masks_z(OptContext *ctx, TCGOp *op, uint64_t z_mask)
1062 {
1063     return fold_masks_zs(ctx, op, z_mask, 0);
1064 }
1065 
1066 static bool fold_masks_s(OptContext *ctx, TCGOp *op, uint64_t s_mask)
1067 {
1068     return fold_masks_zs(ctx, op, -1, s_mask);
1069 }
1070 
1071 /*
1072  * An "affected" mask bit is 0 if and only if the result is identical
1073  * to the first input.  Thus if the entire mask is 0, the operation
1074  * is equivalent to a copy.
1075  */
1076 static bool fold_affected_mask(OptContext *ctx, TCGOp *op, uint64_t a_mask)
1077 {
1078     if (ctx->type == TCG_TYPE_I32) {
1079         a_mask = (uint32_t)a_mask;
1080     }
1081     if (a_mask == 0) {
1082         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
1083     }
1084     return false;
1085 }
1086 
1087 /*
1088  * Convert @op to NOT, if NOT is supported by the host.
1089  * Return true f the conversion is successful, which will still
1090  * indicate that the processing is complete.
1091  */
1092 static bool fold_not(OptContext *ctx, TCGOp *op);
1093 static bool fold_to_not(OptContext *ctx, TCGOp *op, int idx)
1094 {
1095     TCGOpcode not_op;
1096     bool have_not;
1097 
1098     switch (ctx->type) {
1099     case TCG_TYPE_I32:
1100         not_op = INDEX_op_not_i32;
1101         have_not = TCG_TARGET_HAS_not_i32;
1102         break;
1103     case TCG_TYPE_I64:
1104         not_op = INDEX_op_not_i64;
1105         have_not = TCG_TARGET_HAS_not_i64;
1106         break;
1107     case TCG_TYPE_V64:
1108     case TCG_TYPE_V128:
1109     case TCG_TYPE_V256:
1110         not_op = INDEX_op_not_vec;
1111         have_not = TCG_TARGET_HAS_not_vec;
1112         break;
1113     default:
1114         g_assert_not_reached();
1115     }
1116     if (have_not) {
1117         op->opc = not_op;
1118         op->args[1] = op->args[idx];
1119         return fold_not(ctx, op);
1120     }
1121     return false;
1122 }
1123 
1124 /* If the binary operation has first argument @i, fold to @i. */
1125 static bool fold_ix_to_i(OptContext *ctx, TCGOp *op, uint64_t i)
1126 {
1127     if (arg_is_const_val(op->args[1], i)) {
1128         return tcg_opt_gen_movi(ctx, op, op->args[0], i);
1129     }
1130     return false;
1131 }
1132 
1133 /* If the binary operation has first argument @i, fold to NOT. */
1134 static bool fold_ix_to_not(OptContext *ctx, TCGOp *op, uint64_t i)
1135 {
1136     if (arg_is_const_val(op->args[1], i)) {
1137         return fold_to_not(ctx, op, 2);
1138     }
1139     return false;
1140 }
1141 
1142 /* If the binary operation has second argument @i, fold to @i. */
1143 static bool fold_xi_to_i(OptContext *ctx, TCGOp *op, uint64_t i)
1144 {
1145     if (arg_is_const_val(op->args[2], i)) {
1146         return tcg_opt_gen_movi(ctx, op, op->args[0], i);
1147     }
1148     return false;
1149 }
1150 
1151 /* If the binary operation has second argument @i, fold to identity. */
1152 static bool fold_xi_to_x(OptContext *ctx, TCGOp *op, uint64_t i)
1153 {
1154     if (arg_is_const_val(op->args[2], i)) {
1155         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
1156     }
1157     return false;
1158 }
1159 
1160 /* If the binary operation has second argument @i, fold to NOT. */
1161 static bool fold_xi_to_not(OptContext *ctx, TCGOp *op, uint64_t i)
1162 {
1163     if (arg_is_const_val(op->args[2], i)) {
1164         return fold_to_not(ctx, op, 1);
1165     }
1166     return false;
1167 }
1168 
1169 /* If the binary operation has both arguments equal, fold to @i. */
1170 static bool fold_xx_to_i(OptContext *ctx, TCGOp *op, uint64_t i)
1171 {
1172     if (args_are_copies(op->args[1], op->args[2])) {
1173         return tcg_opt_gen_movi(ctx, op, op->args[0], i);
1174     }
1175     return false;
1176 }
1177 
1178 /* If the binary operation has both arguments equal, fold to identity. */
1179 static bool fold_xx_to_x(OptContext *ctx, TCGOp *op)
1180 {
1181     if (args_are_copies(op->args[1], op->args[2])) {
1182         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
1183     }
1184     return false;
1185 }
1186 
1187 /*
1188  * These outermost fold_<op> functions are sorted alphabetically.
1189  *
1190  * The ordering of the transformations should be:
1191  *   1) those that produce a constant
1192  *   2) those that produce a copy
1193  *   3) those that produce information about the result value.
1194  */
1195 
1196 static bool fold_or(OptContext *ctx, TCGOp *op);
1197 static bool fold_orc(OptContext *ctx, TCGOp *op);
1198 static bool fold_xor(OptContext *ctx, TCGOp *op);
1199 
1200 static bool fold_add(OptContext *ctx, TCGOp *op)
1201 {
1202     if (fold_const2_commutative(ctx, op) ||
1203         fold_xi_to_x(ctx, op, 0)) {
1204         return true;
1205     }
1206     return finish_folding(ctx, op);
1207 }
1208 
1209 /* We cannot as yet do_constant_folding with vectors. */
1210 static bool fold_add_vec(OptContext *ctx, TCGOp *op)
1211 {
1212     if (fold_commutative(ctx, op) ||
1213         fold_xi_to_x(ctx, op, 0)) {
1214         return true;
1215     }
1216     return finish_folding(ctx, op);
1217 }
1218 
1219 static bool fold_addsub2(OptContext *ctx, TCGOp *op, bool add)
1220 {
1221     bool a_const = arg_is_const(op->args[2]) && arg_is_const(op->args[3]);
1222     bool b_const = arg_is_const(op->args[4]) && arg_is_const(op->args[5]);
1223 
1224     if (a_const && b_const) {
1225         uint64_t al = arg_info(op->args[2])->val;
1226         uint64_t ah = arg_info(op->args[3])->val;
1227         uint64_t bl = arg_info(op->args[4])->val;
1228         uint64_t bh = arg_info(op->args[5])->val;
1229         TCGArg rl, rh;
1230         TCGOp *op2;
1231 
1232         if (ctx->type == TCG_TYPE_I32) {
1233             uint64_t a = deposit64(al, 32, 32, ah);
1234             uint64_t b = deposit64(bl, 32, 32, bh);
1235 
1236             if (add) {
1237                 a += b;
1238             } else {
1239                 a -= b;
1240             }
1241 
1242             al = sextract64(a, 0, 32);
1243             ah = sextract64(a, 32, 32);
1244         } else {
1245             Int128 a = int128_make128(al, ah);
1246             Int128 b = int128_make128(bl, bh);
1247 
1248             if (add) {
1249                 a = int128_add(a, b);
1250             } else {
1251                 a = int128_sub(a, b);
1252             }
1253 
1254             al = int128_getlo(a);
1255             ah = int128_gethi(a);
1256         }
1257 
1258         rl = op->args[0];
1259         rh = op->args[1];
1260 
1261         /* The proper opcode is supplied by tcg_opt_gen_mov. */
1262         op2 = opt_insert_before(ctx, op, 0, 2);
1263 
1264         tcg_opt_gen_movi(ctx, op, rl, al);
1265         tcg_opt_gen_movi(ctx, op2, rh, ah);
1266         return true;
1267     }
1268 
1269     /* Fold sub2 r,x,i to add2 r,x,-i */
1270     if (!add && b_const) {
1271         uint64_t bl = arg_info(op->args[4])->val;
1272         uint64_t bh = arg_info(op->args[5])->val;
1273 
1274         /* Negate the two parts without assembling and disassembling. */
1275         bl = -bl;
1276         bh = ~bh + !bl;
1277 
1278         op->opc = (ctx->type == TCG_TYPE_I32
1279                    ? INDEX_op_add2_i32 : INDEX_op_add2_i64);
1280         op->args[4] = arg_new_constant(ctx, bl);
1281         op->args[5] = arg_new_constant(ctx, bh);
1282     }
1283     return finish_folding(ctx, op);
1284 }
1285 
1286 static bool fold_add2(OptContext *ctx, TCGOp *op)
1287 {
1288     /* Note that the high and low parts may be independently swapped. */
1289     swap_commutative(op->args[0], &op->args[2], &op->args[4]);
1290     swap_commutative(op->args[1], &op->args[3], &op->args[5]);
1291 
1292     return fold_addsub2(ctx, op, true);
1293 }
1294 
1295 static bool fold_and(OptContext *ctx, TCGOp *op)
1296 {
1297     uint64_t z1, z2, z_mask, s_mask;
1298     TempOptInfo *t1, *t2;
1299 
1300     if (fold_const2_commutative(ctx, op) ||
1301         fold_xi_to_i(ctx, op, 0) ||
1302         fold_xi_to_x(ctx, op, -1) ||
1303         fold_xx_to_x(ctx, op)) {
1304         return true;
1305     }
1306 
1307     t1 = arg_info(op->args[1]);
1308     t2 = arg_info(op->args[2]);
1309     z1 = t1->z_mask;
1310     z2 = t2->z_mask;
1311 
1312     /*
1313      * Known-zeros does not imply known-ones.  Therefore unless
1314      * arg2 is constant, we can't infer affected bits from it.
1315      */
1316     if (ti_is_const(t2) && fold_affected_mask(ctx, op, z1 & ~z2)) {
1317         return true;
1318     }
1319 
1320     z_mask = z1 & z2;
1321 
1322     /*
1323      * Sign repetitions are perforce all identical, whether they are 1 or 0.
1324      * Bitwise operations preserve the relative quantity of the repetitions.
1325      */
1326     s_mask = t1->s_mask & t2->s_mask;
1327 
1328     return fold_masks_zs(ctx, op, z_mask, s_mask);
1329 }
1330 
1331 static bool fold_andc(OptContext *ctx, TCGOp *op)
1332 {
1333     uint64_t z_mask, s_mask;
1334     TempOptInfo *t1, *t2;
1335 
1336     if (fold_const2(ctx, op) ||
1337         fold_xx_to_i(ctx, op, 0) ||
1338         fold_xi_to_x(ctx, op, 0) ||
1339         fold_ix_to_not(ctx, op, -1)) {
1340         return true;
1341     }
1342 
1343     t1 = arg_info(op->args[1]);
1344     t2 = arg_info(op->args[2]);
1345     z_mask = t1->z_mask;
1346 
1347     if (ti_is_const(t2)) {
1348         /* Fold andc r,x,i to and r,x,~i. */
1349         switch (ctx->type) {
1350         case TCG_TYPE_I32:
1351         case TCG_TYPE_I64:
1352             op->opc = INDEX_op_and;
1353             break;
1354         case TCG_TYPE_V64:
1355         case TCG_TYPE_V128:
1356         case TCG_TYPE_V256:
1357             op->opc = INDEX_op_and_vec;
1358             break;
1359         default:
1360             g_assert_not_reached();
1361         }
1362         op->args[2] = arg_new_constant(ctx, ~ti_const_val(t2));
1363         return fold_and(ctx, op);
1364     }
1365 
1366     /*
1367      * Known-zeros does not imply known-ones.  Therefore unless
1368      * arg2 is constant, we can't infer anything from it.
1369      */
1370     if (ti_is_const(t2)) {
1371         uint64_t v2 = ti_const_val(t2);
1372         if (fold_affected_mask(ctx, op, z_mask & v2)) {
1373             return true;
1374         }
1375         z_mask &= ~v2;
1376     }
1377 
1378     s_mask = t1->s_mask & t2->s_mask;
1379     return fold_masks_zs(ctx, op, z_mask, s_mask);
1380 }
1381 
1382 static bool fold_bitsel_vec(OptContext *ctx, TCGOp *op)
1383 {
1384     /* If true and false values are the same, eliminate the cmp. */
1385     if (args_are_copies(op->args[2], op->args[3])) {
1386         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[2]);
1387     }
1388 
1389     if (arg_is_const(op->args[2]) && arg_is_const(op->args[3])) {
1390         uint64_t tv = arg_info(op->args[2])->val;
1391         uint64_t fv = arg_info(op->args[3])->val;
1392 
1393         if (tv == -1 && fv == 0) {
1394             return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
1395         }
1396         if (tv == 0 && fv == -1) {
1397             if (TCG_TARGET_HAS_not_vec) {
1398                 op->opc = INDEX_op_not_vec;
1399                 return fold_not(ctx, op);
1400             } else {
1401                 op->opc = INDEX_op_xor_vec;
1402                 op->args[2] = arg_new_constant(ctx, -1);
1403                 return fold_xor(ctx, op);
1404             }
1405         }
1406     }
1407     if (arg_is_const(op->args[2])) {
1408         uint64_t tv = arg_info(op->args[2])->val;
1409         if (tv == -1) {
1410             op->opc = INDEX_op_or_vec;
1411             op->args[2] = op->args[3];
1412             return fold_or(ctx, op);
1413         }
1414         if (tv == 0 && TCG_TARGET_HAS_andc_vec) {
1415             op->opc = INDEX_op_andc_vec;
1416             op->args[2] = op->args[1];
1417             op->args[1] = op->args[3];
1418             return fold_andc(ctx, op);
1419         }
1420     }
1421     if (arg_is_const(op->args[3])) {
1422         uint64_t fv = arg_info(op->args[3])->val;
1423         if (fv == 0) {
1424             op->opc = INDEX_op_and_vec;
1425             return fold_and(ctx, op);
1426         }
1427         if (fv == -1 && TCG_TARGET_HAS_orc_vec) {
1428             op->opc = INDEX_op_orc_vec;
1429             op->args[2] = op->args[1];
1430             op->args[1] = op->args[3];
1431             return fold_orc(ctx, op);
1432         }
1433     }
1434     return finish_folding(ctx, op);
1435 }
1436 
1437 static bool fold_brcond(OptContext *ctx, TCGOp *op)
1438 {
1439     int i = do_constant_folding_cond1(ctx, op, NO_DEST, &op->args[0],
1440                                       &op->args[1], &op->args[2]);
1441     if (i == 0) {
1442         tcg_op_remove(ctx->tcg, op);
1443         return true;
1444     }
1445     if (i > 0) {
1446         op->opc = INDEX_op_br;
1447         op->args[0] = op->args[3];
1448         finish_ebb(ctx);
1449     } else {
1450         finish_bb(ctx);
1451     }
1452     return true;
1453 }
1454 
1455 static bool fold_brcond2(OptContext *ctx, TCGOp *op)
1456 {
1457     TCGCond cond;
1458     TCGArg label;
1459     int i, inv = 0;
1460 
1461     i = do_constant_folding_cond2(ctx, op, &op->args[0]);
1462     cond = op->args[4];
1463     label = op->args[5];
1464     if (i >= 0) {
1465         goto do_brcond_const;
1466     }
1467 
1468     switch (cond) {
1469     case TCG_COND_LT:
1470     case TCG_COND_GE:
1471         /*
1472          * Simplify LT/GE comparisons vs zero to a single compare
1473          * vs the high word of the input.
1474          */
1475         if (arg_is_const_val(op->args[2], 0) &&
1476             arg_is_const_val(op->args[3], 0)) {
1477             goto do_brcond_high;
1478         }
1479         break;
1480 
1481     case TCG_COND_NE:
1482         inv = 1;
1483         QEMU_FALLTHROUGH;
1484     case TCG_COND_EQ:
1485         /*
1486          * Simplify EQ/NE comparisons where one of the pairs
1487          * can be simplified.
1488          */
1489         i = do_constant_folding_cond(TCG_TYPE_I32, op->args[0],
1490                                      op->args[2], cond);
1491         switch (i ^ inv) {
1492         case 0:
1493             goto do_brcond_const;
1494         case 1:
1495             goto do_brcond_high;
1496         }
1497 
1498         i = do_constant_folding_cond(TCG_TYPE_I32, op->args[1],
1499                                      op->args[3], cond);
1500         switch (i ^ inv) {
1501         case 0:
1502             goto do_brcond_const;
1503         case 1:
1504             goto do_brcond_low;
1505         }
1506         break;
1507 
1508     case TCG_COND_TSTEQ:
1509     case TCG_COND_TSTNE:
1510         if (arg_is_const_val(op->args[2], 0)) {
1511             goto do_brcond_high;
1512         }
1513         if (arg_is_const_val(op->args[3], 0)) {
1514             goto do_brcond_low;
1515         }
1516         break;
1517 
1518     default:
1519         break;
1520 
1521     do_brcond_low:
1522         op->opc = INDEX_op_brcond_i32;
1523         op->args[1] = op->args[2];
1524         op->args[2] = cond;
1525         op->args[3] = label;
1526         return fold_brcond(ctx, op);
1527 
1528     do_brcond_high:
1529         op->opc = INDEX_op_brcond_i32;
1530         op->args[0] = op->args[1];
1531         op->args[1] = op->args[3];
1532         op->args[2] = cond;
1533         op->args[3] = label;
1534         return fold_brcond(ctx, op);
1535 
1536     do_brcond_const:
1537         if (i == 0) {
1538             tcg_op_remove(ctx->tcg, op);
1539             return true;
1540         }
1541         op->opc = INDEX_op_br;
1542         op->args[0] = label;
1543         finish_ebb(ctx);
1544         return true;
1545     }
1546 
1547     finish_bb(ctx);
1548     return true;
1549 }
1550 
1551 static bool fold_bswap(OptContext *ctx, TCGOp *op)
1552 {
1553     uint64_t z_mask, s_mask, sign;
1554     TempOptInfo *t1 = arg_info(op->args[1]);
1555 
1556     if (ti_is_const(t1)) {
1557         return tcg_opt_gen_movi(ctx, op, op->args[0],
1558                                 do_constant_folding(op->opc, ctx->type,
1559                                                     ti_const_val(t1),
1560                                                     op->args[2]));
1561     }
1562 
1563     z_mask = t1->z_mask;
1564     switch (op->opc) {
1565     case INDEX_op_bswap16_i32:
1566     case INDEX_op_bswap16_i64:
1567         z_mask = bswap16(z_mask);
1568         sign = INT16_MIN;
1569         break;
1570     case INDEX_op_bswap32_i32:
1571     case INDEX_op_bswap32_i64:
1572         z_mask = bswap32(z_mask);
1573         sign = INT32_MIN;
1574         break;
1575     case INDEX_op_bswap64_i64:
1576         z_mask = bswap64(z_mask);
1577         sign = INT64_MIN;
1578         break;
1579     default:
1580         g_assert_not_reached();
1581     }
1582 
1583     s_mask = 0;
1584     switch (op->args[2] & (TCG_BSWAP_OZ | TCG_BSWAP_OS)) {
1585     case TCG_BSWAP_OZ:
1586         break;
1587     case TCG_BSWAP_OS:
1588         /* If the sign bit may be 1, force all the bits above to 1. */
1589         if (z_mask & sign) {
1590             z_mask |= sign;
1591         }
1592         /* The value and therefore s_mask is explicitly sign-extended. */
1593         s_mask = sign;
1594         break;
1595     default:
1596         /* The high bits are undefined: force all bits above the sign to 1. */
1597         z_mask |= sign << 1;
1598         break;
1599     }
1600 
1601     return fold_masks_zs(ctx, op, z_mask, s_mask);
1602 }
1603 
1604 static bool fold_call(OptContext *ctx, TCGOp *op)
1605 {
1606     TCGContext *s = ctx->tcg;
1607     int nb_oargs = TCGOP_CALLO(op);
1608     int nb_iargs = TCGOP_CALLI(op);
1609     int flags, i;
1610 
1611     init_arguments(ctx, op, nb_oargs + nb_iargs);
1612     copy_propagate(ctx, op, nb_oargs, nb_iargs);
1613 
1614     /* If the function reads or writes globals, reset temp data. */
1615     flags = tcg_call_flags(op);
1616     if (!(flags & (TCG_CALL_NO_READ_GLOBALS | TCG_CALL_NO_WRITE_GLOBALS))) {
1617         int nb_globals = s->nb_globals;
1618 
1619         for (i = 0; i < nb_globals; i++) {
1620             if (test_bit(i, ctx->temps_used.l)) {
1621                 reset_ts(ctx, &ctx->tcg->temps[i]);
1622             }
1623         }
1624     }
1625 
1626     /* If the function has side effects, reset mem data. */
1627     if (!(flags & TCG_CALL_NO_SIDE_EFFECTS)) {
1628         remove_mem_copy_all(ctx);
1629     }
1630 
1631     /* Reset temp data for outputs. */
1632     for (i = 0; i < nb_oargs; i++) {
1633         reset_temp(ctx, op->args[i]);
1634     }
1635 
1636     /* Stop optimizing MB across calls. */
1637     ctx->prev_mb = NULL;
1638     return true;
1639 }
1640 
1641 static bool fold_cmp_vec(OptContext *ctx, TCGOp *op)
1642 {
1643     /* Canonicalize the comparison to put immediate second. */
1644     if (swap_commutative(NO_DEST, &op->args[1], &op->args[2])) {
1645         op->args[3] = tcg_swap_cond(op->args[3]);
1646     }
1647     return finish_folding(ctx, op);
1648 }
1649 
1650 static bool fold_cmpsel_vec(OptContext *ctx, TCGOp *op)
1651 {
1652     /* If true and false values are the same, eliminate the cmp. */
1653     if (args_are_copies(op->args[3], op->args[4])) {
1654         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[3]);
1655     }
1656 
1657     /* Canonicalize the comparison to put immediate second. */
1658     if (swap_commutative(NO_DEST, &op->args[1], &op->args[2])) {
1659         op->args[5] = tcg_swap_cond(op->args[5]);
1660     }
1661     /*
1662      * Canonicalize the "false" input reg to match the destination,
1663      * so that the tcg backend can implement "move if true".
1664      */
1665     if (swap_commutative(op->args[0], &op->args[4], &op->args[3])) {
1666         op->args[5] = tcg_invert_cond(op->args[5]);
1667     }
1668     return finish_folding(ctx, op);
1669 }
1670 
1671 static bool fold_count_zeros(OptContext *ctx, TCGOp *op)
1672 {
1673     uint64_t z_mask, s_mask;
1674     TempOptInfo *t1 = arg_info(op->args[1]);
1675     TempOptInfo *t2 = arg_info(op->args[2]);
1676 
1677     if (ti_is_const(t1)) {
1678         uint64_t t = ti_const_val(t1);
1679 
1680         if (t != 0) {
1681             t = do_constant_folding(op->opc, ctx->type, t, 0);
1682             return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1683         }
1684         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[2]);
1685     }
1686 
1687     switch (ctx->type) {
1688     case TCG_TYPE_I32:
1689         z_mask = 31;
1690         break;
1691     case TCG_TYPE_I64:
1692         z_mask = 63;
1693         break;
1694     default:
1695         g_assert_not_reached();
1696     }
1697     s_mask = ~z_mask;
1698     z_mask |= t2->z_mask;
1699     s_mask &= t2->s_mask;
1700 
1701     return fold_masks_zs(ctx, op, z_mask, s_mask);
1702 }
1703 
1704 static bool fold_ctpop(OptContext *ctx, TCGOp *op)
1705 {
1706     uint64_t z_mask;
1707 
1708     if (fold_const1(ctx, op)) {
1709         return true;
1710     }
1711 
1712     switch (ctx->type) {
1713     case TCG_TYPE_I32:
1714         z_mask = 32 | 31;
1715         break;
1716     case TCG_TYPE_I64:
1717         z_mask = 64 | 63;
1718         break;
1719     default:
1720         g_assert_not_reached();
1721     }
1722     return fold_masks_z(ctx, op, z_mask);
1723 }
1724 
1725 static bool fold_deposit(OptContext *ctx, TCGOp *op)
1726 {
1727     TempOptInfo *t1 = arg_info(op->args[1]);
1728     TempOptInfo *t2 = arg_info(op->args[2]);
1729     int ofs = op->args[3];
1730     int len = op->args[4];
1731     int width = 8 * tcg_type_size(ctx->type);
1732     uint64_t z_mask, s_mask;
1733 
1734     if (ti_is_const(t1) && ti_is_const(t2)) {
1735         return tcg_opt_gen_movi(ctx, op, op->args[0],
1736                                 deposit64(ti_const_val(t1), ofs, len,
1737                                           ti_const_val(t2)));
1738     }
1739 
1740     /* Inserting a value into zero at offset 0. */
1741     if (ti_is_const_val(t1, 0) && ofs == 0) {
1742         uint64_t mask = MAKE_64BIT_MASK(0, len);
1743 
1744         op->opc = INDEX_op_and;
1745         op->args[1] = op->args[2];
1746         op->args[2] = arg_new_constant(ctx, mask);
1747         return fold_and(ctx, op);
1748     }
1749 
1750     /* Inserting zero into a value. */
1751     if (ti_is_const_val(t2, 0)) {
1752         uint64_t mask = deposit64(-1, ofs, len, 0);
1753 
1754         op->opc = INDEX_op_and;
1755         op->args[2] = arg_new_constant(ctx, mask);
1756         return fold_and(ctx, op);
1757     }
1758 
1759     /* The s_mask from the top portion of the deposit is still valid. */
1760     if (ofs + len == width) {
1761         s_mask = t2->s_mask << ofs;
1762     } else {
1763         s_mask = t1->s_mask & ~MAKE_64BIT_MASK(0, ofs + len);
1764     }
1765 
1766     z_mask = deposit64(t1->z_mask, ofs, len, t2->z_mask);
1767     return fold_masks_zs(ctx, op, z_mask, s_mask);
1768 }
1769 
1770 static bool fold_divide(OptContext *ctx, TCGOp *op)
1771 {
1772     if (fold_const2(ctx, op) ||
1773         fold_xi_to_x(ctx, op, 1)) {
1774         return true;
1775     }
1776     return finish_folding(ctx, op);
1777 }
1778 
1779 static bool fold_dup(OptContext *ctx, TCGOp *op)
1780 {
1781     if (arg_is_const(op->args[1])) {
1782         uint64_t t = arg_info(op->args[1])->val;
1783         t = dup_const(TCGOP_VECE(op), t);
1784         return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1785     }
1786     return finish_folding(ctx, op);
1787 }
1788 
1789 static bool fold_dup2(OptContext *ctx, TCGOp *op)
1790 {
1791     if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) {
1792         uint64_t t = deposit64(arg_info(op->args[1])->val, 32, 32,
1793                                arg_info(op->args[2])->val);
1794         return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1795     }
1796 
1797     if (args_are_copies(op->args[1], op->args[2])) {
1798         op->opc = INDEX_op_dup_vec;
1799         TCGOP_VECE(op) = MO_32;
1800     }
1801     return finish_folding(ctx, op);
1802 }
1803 
1804 static bool fold_eqv(OptContext *ctx, TCGOp *op)
1805 {
1806     uint64_t s_mask;
1807 
1808     if (fold_const2_commutative(ctx, op) ||
1809         fold_xi_to_x(ctx, op, -1) ||
1810         fold_xi_to_not(ctx, op, 0)) {
1811         return true;
1812     }
1813 
1814     s_mask = arg_info(op->args[1])->s_mask
1815            & arg_info(op->args[2])->s_mask;
1816     return fold_masks_s(ctx, op, s_mask);
1817 }
1818 
1819 static bool fold_extract(OptContext *ctx, TCGOp *op)
1820 {
1821     uint64_t z_mask_old, z_mask;
1822     TempOptInfo *t1 = arg_info(op->args[1]);
1823     int pos = op->args[2];
1824     int len = op->args[3];
1825 
1826     if (ti_is_const(t1)) {
1827         return tcg_opt_gen_movi(ctx, op, op->args[0],
1828                                 extract64(ti_const_val(t1), pos, len));
1829     }
1830 
1831     z_mask_old = t1->z_mask;
1832     z_mask = extract64(z_mask_old, pos, len);
1833     if (pos == 0 && fold_affected_mask(ctx, op, z_mask_old ^ z_mask)) {
1834         return true;
1835     }
1836 
1837     return fold_masks_z(ctx, op, z_mask);
1838 }
1839 
1840 static bool fold_extract2(OptContext *ctx, TCGOp *op)
1841 {
1842     if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) {
1843         uint64_t v1 = arg_info(op->args[1])->val;
1844         uint64_t v2 = arg_info(op->args[2])->val;
1845         int shr = op->args[3];
1846 
1847         if (op->opc == INDEX_op_extract2_i64) {
1848             v1 >>= shr;
1849             v2 <<= 64 - shr;
1850         } else {
1851             v1 = (uint32_t)v1 >> shr;
1852             v2 = (uint64_t)((int32_t)v2 << (32 - shr));
1853         }
1854         return tcg_opt_gen_movi(ctx, op, op->args[0], v1 | v2);
1855     }
1856     return finish_folding(ctx, op);
1857 }
1858 
1859 static bool fold_exts(OptContext *ctx, TCGOp *op)
1860 {
1861     uint64_t s_mask, z_mask;
1862     TempOptInfo *t1;
1863 
1864     if (fold_const1(ctx, op)) {
1865         return true;
1866     }
1867 
1868     t1 = arg_info(op->args[1]);
1869     z_mask = t1->z_mask;
1870     s_mask = t1->s_mask;
1871 
1872     switch (op->opc) {
1873     case INDEX_op_ext_i32_i64:
1874         s_mask |= INT32_MIN;
1875         z_mask = (int32_t)z_mask;
1876         break;
1877     default:
1878         g_assert_not_reached();
1879     }
1880     return fold_masks_zs(ctx, op, z_mask, s_mask);
1881 }
1882 
1883 static bool fold_extu(OptContext *ctx, TCGOp *op)
1884 {
1885     uint64_t z_mask;
1886 
1887     if (fold_const1(ctx, op)) {
1888         return true;
1889     }
1890 
1891     z_mask = arg_info(op->args[1])->z_mask;
1892     switch (op->opc) {
1893     case INDEX_op_extrl_i64_i32:
1894     case INDEX_op_extu_i32_i64:
1895         z_mask = (uint32_t)z_mask;
1896         break;
1897     case INDEX_op_extrh_i64_i32:
1898         z_mask >>= 32;
1899         break;
1900     default:
1901         g_assert_not_reached();
1902     }
1903     return fold_masks_z(ctx, op, z_mask);
1904 }
1905 
1906 static bool fold_mb(OptContext *ctx, TCGOp *op)
1907 {
1908     /* Eliminate duplicate and redundant fence instructions.  */
1909     if (ctx->prev_mb) {
1910         /*
1911          * Merge two barriers of the same type into one,
1912          * or a weaker barrier into a stronger one,
1913          * or two weaker barriers into a stronger one.
1914          *   mb X; mb Y => mb X|Y
1915          *   mb; strl => mb; st
1916          *   ldaq; mb => ld; mb
1917          *   ldaq; strl => ld; mb; st
1918          * Other combinations are also merged into a strong
1919          * barrier.  This is stricter than specified but for
1920          * the purposes of TCG is better than not optimizing.
1921          */
1922         ctx->prev_mb->args[0] |= op->args[0];
1923         tcg_op_remove(ctx->tcg, op);
1924     } else {
1925         ctx->prev_mb = op;
1926     }
1927     return true;
1928 }
1929 
1930 static bool fold_mov(OptContext *ctx, TCGOp *op)
1931 {
1932     return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
1933 }
1934 
1935 static bool fold_movcond(OptContext *ctx, TCGOp *op)
1936 {
1937     uint64_t z_mask, s_mask;
1938     TempOptInfo *tt, *ft;
1939     int i;
1940 
1941     /* If true and false values are the same, eliminate the cmp. */
1942     if (args_are_copies(op->args[3], op->args[4])) {
1943         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[3]);
1944     }
1945 
1946     /*
1947      * Canonicalize the "false" input reg to match the destination reg so
1948      * that the tcg backend can implement a "move if true" operation.
1949      */
1950     if (swap_commutative(op->args[0], &op->args[4], &op->args[3])) {
1951         op->args[5] = tcg_invert_cond(op->args[5]);
1952     }
1953 
1954     i = do_constant_folding_cond1(ctx, op, NO_DEST, &op->args[1],
1955                                   &op->args[2], &op->args[5]);
1956     if (i >= 0) {
1957         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[4 - i]);
1958     }
1959 
1960     tt = arg_info(op->args[3]);
1961     ft = arg_info(op->args[4]);
1962     z_mask = tt->z_mask | ft->z_mask;
1963     s_mask = tt->s_mask & ft->s_mask;
1964 
1965     if (ti_is_const(tt) && ti_is_const(ft)) {
1966         uint64_t tv = ti_const_val(tt);
1967         uint64_t fv = ti_const_val(ft);
1968         TCGOpcode opc, negopc = 0;
1969         TCGCond cond = op->args[5];
1970 
1971         switch (ctx->type) {
1972         case TCG_TYPE_I32:
1973             opc = INDEX_op_setcond_i32;
1974             if (TCG_TARGET_HAS_negsetcond_i32) {
1975                 negopc = INDEX_op_negsetcond_i32;
1976             }
1977             tv = (int32_t)tv;
1978             fv = (int32_t)fv;
1979             break;
1980         case TCG_TYPE_I64:
1981             opc = INDEX_op_setcond_i64;
1982             if (TCG_TARGET_HAS_negsetcond_i64) {
1983                 negopc = INDEX_op_negsetcond_i64;
1984             }
1985             break;
1986         default:
1987             g_assert_not_reached();
1988         }
1989 
1990         if (tv == 1 && fv == 0) {
1991             op->opc = opc;
1992             op->args[3] = cond;
1993         } else if (fv == 1 && tv == 0) {
1994             op->opc = opc;
1995             op->args[3] = tcg_invert_cond(cond);
1996         } else if (negopc) {
1997             if (tv == -1 && fv == 0) {
1998                 op->opc = negopc;
1999                 op->args[3] = cond;
2000             } else if (fv == -1 && tv == 0) {
2001                 op->opc = negopc;
2002                 op->args[3] = tcg_invert_cond(cond);
2003             }
2004         }
2005     }
2006 
2007     return fold_masks_zs(ctx, op, z_mask, s_mask);
2008 }
2009 
2010 static bool fold_mul(OptContext *ctx, TCGOp *op)
2011 {
2012     if (fold_const2(ctx, op) ||
2013         fold_xi_to_i(ctx, op, 0) ||
2014         fold_xi_to_x(ctx, op, 1)) {
2015         return true;
2016     }
2017     return finish_folding(ctx, op);
2018 }
2019 
2020 static bool fold_mul_highpart(OptContext *ctx, TCGOp *op)
2021 {
2022     if (fold_const2_commutative(ctx, op) ||
2023         fold_xi_to_i(ctx, op, 0)) {
2024         return true;
2025     }
2026     return finish_folding(ctx, op);
2027 }
2028 
2029 static bool fold_multiply2(OptContext *ctx, TCGOp *op)
2030 {
2031     swap_commutative(op->args[0], &op->args[2], &op->args[3]);
2032 
2033     if (arg_is_const(op->args[2]) && arg_is_const(op->args[3])) {
2034         uint64_t a = arg_info(op->args[2])->val;
2035         uint64_t b = arg_info(op->args[3])->val;
2036         uint64_t h, l;
2037         TCGArg rl, rh;
2038         TCGOp *op2;
2039 
2040         switch (op->opc) {
2041         case INDEX_op_mulu2_i32:
2042             l = (uint64_t)(uint32_t)a * (uint32_t)b;
2043             h = (int32_t)(l >> 32);
2044             l = (int32_t)l;
2045             break;
2046         case INDEX_op_muls2_i32:
2047             l = (int64_t)(int32_t)a * (int32_t)b;
2048             h = l >> 32;
2049             l = (int32_t)l;
2050             break;
2051         case INDEX_op_mulu2_i64:
2052             mulu64(&l, &h, a, b);
2053             break;
2054         case INDEX_op_muls2_i64:
2055             muls64(&l, &h, a, b);
2056             break;
2057         default:
2058             g_assert_not_reached();
2059         }
2060 
2061         rl = op->args[0];
2062         rh = op->args[1];
2063 
2064         /* The proper opcode is supplied by tcg_opt_gen_mov. */
2065         op2 = opt_insert_before(ctx, op, 0, 2);
2066 
2067         tcg_opt_gen_movi(ctx, op, rl, l);
2068         tcg_opt_gen_movi(ctx, op2, rh, h);
2069         return true;
2070     }
2071     return finish_folding(ctx, op);
2072 }
2073 
2074 static bool fold_nand(OptContext *ctx, TCGOp *op)
2075 {
2076     uint64_t s_mask;
2077 
2078     if (fold_const2_commutative(ctx, op) ||
2079         fold_xi_to_not(ctx, op, -1)) {
2080         return true;
2081     }
2082 
2083     s_mask = arg_info(op->args[1])->s_mask
2084            & arg_info(op->args[2])->s_mask;
2085     return fold_masks_s(ctx, op, s_mask);
2086 }
2087 
2088 static bool fold_neg_no_const(OptContext *ctx, TCGOp *op)
2089 {
2090     /* Set to 1 all bits to the left of the rightmost.  */
2091     uint64_t z_mask = arg_info(op->args[1])->z_mask;
2092     z_mask = -(z_mask & -z_mask);
2093 
2094     return fold_masks_z(ctx, op, z_mask);
2095 }
2096 
2097 static bool fold_neg(OptContext *ctx, TCGOp *op)
2098 {
2099     return fold_const1(ctx, op) || fold_neg_no_const(ctx, op);
2100 }
2101 
2102 static bool fold_nor(OptContext *ctx, TCGOp *op)
2103 {
2104     uint64_t s_mask;
2105 
2106     if (fold_const2_commutative(ctx, op) ||
2107         fold_xi_to_not(ctx, op, 0)) {
2108         return true;
2109     }
2110 
2111     s_mask = arg_info(op->args[1])->s_mask
2112            & arg_info(op->args[2])->s_mask;
2113     return fold_masks_s(ctx, op, s_mask);
2114 }
2115 
2116 static bool fold_not(OptContext *ctx, TCGOp *op)
2117 {
2118     if (fold_const1(ctx, op)) {
2119         return true;
2120     }
2121     return fold_masks_s(ctx, op, arg_info(op->args[1])->s_mask);
2122 }
2123 
2124 static bool fold_or(OptContext *ctx, TCGOp *op)
2125 {
2126     uint64_t z_mask, s_mask;
2127     TempOptInfo *t1, *t2;
2128 
2129     if (fold_const2_commutative(ctx, op) ||
2130         fold_xi_to_x(ctx, op, 0) ||
2131         fold_xx_to_x(ctx, op)) {
2132         return true;
2133     }
2134 
2135     t1 = arg_info(op->args[1]);
2136     t2 = arg_info(op->args[2]);
2137     z_mask = t1->z_mask | t2->z_mask;
2138     s_mask = t1->s_mask & t2->s_mask;
2139     return fold_masks_zs(ctx, op, z_mask, s_mask);
2140 }
2141 
2142 static bool fold_orc(OptContext *ctx, TCGOp *op)
2143 {
2144     uint64_t s_mask;
2145 
2146     if (fold_const2(ctx, op) ||
2147         fold_xx_to_i(ctx, op, -1) ||
2148         fold_xi_to_x(ctx, op, -1) ||
2149         fold_ix_to_not(ctx, op, 0)) {
2150         return true;
2151     }
2152 
2153     s_mask = arg_info(op->args[1])->s_mask
2154            & arg_info(op->args[2])->s_mask;
2155     return fold_masks_s(ctx, op, s_mask);
2156 }
2157 
2158 static bool fold_qemu_ld_1reg(OptContext *ctx, TCGOp *op)
2159 {
2160     const TCGOpDef *def = &tcg_op_defs[op->opc];
2161     MemOpIdx oi = op->args[def->nb_oargs + def->nb_iargs];
2162     MemOp mop = get_memop(oi);
2163     int width = 8 * memop_size(mop);
2164     uint64_t z_mask = -1, s_mask = 0;
2165 
2166     if (width < 64) {
2167         if (mop & MO_SIGN) {
2168             s_mask = MAKE_64BIT_MASK(width - 1, 64 - (width - 1));
2169         } else {
2170             z_mask = MAKE_64BIT_MASK(0, width);
2171         }
2172     }
2173 
2174     /* Opcodes that touch guest memory stop the mb optimization.  */
2175     ctx->prev_mb = NULL;
2176 
2177     return fold_masks_zs(ctx, op, z_mask, s_mask);
2178 }
2179 
2180 static bool fold_qemu_ld_2reg(OptContext *ctx, TCGOp *op)
2181 {
2182     /* Opcodes that touch guest memory stop the mb optimization.  */
2183     ctx->prev_mb = NULL;
2184     return finish_folding(ctx, op);
2185 }
2186 
2187 static bool fold_qemu_st(OptContext *ctx, TCGOp *op)
2188 {
2189     /* Opcodes that touch guest memory stop the mb optimization.  */
2190     ctx->prev_mb = NULL;
2191     return true;
2192 }
2193 
2194 static bool fold_remainder(OptContext *ctx, TCGOp *op)
2195 {
2196     if (fold_const2(ctx, op) ||
2197         fold_xx_to_i(ctx, op, 0)) {
2198         return true;
2199     }
2200     return finish_folding(ctx, op);
2201 }
2202 
2203 /* Return 1 if finished, -1 if simplified, 0 if unchanged. */
2204 static int fold_setcond_zmask(OptContext *ctx, TCGOp *op, bool neg)
2205 {
2206     uint64_t a_zmask, b_val;
2207     TCGCond cond;
2208 
2209     if (!arg_is_const(op->args[2])) {
2210         return false;
2211     }
2212 
2213     a_zmask = arg_info(op->args[1])->z_mask;
2214     b_val = arg_info(op->args[2])->val;
2215     cond = op->args[3];
2216 
2217     if (ctx->type == TCG_TYPE_I32) {
2218         a_zmask = (uint32_t)a_zmask;
2219         b_val = (uint32_t)b_val;
2220     }
2221 
2222     /*
2223      * A with only low bits set vs B with high bits set means that A < B.
2224      */
2225     if (a_zmask < b_val) {
2226         bool inv = false;
2227 
2228         switch (cond) {
2229         case TCG_COND_NE:
2230         case TCG_COND_LEU:
2231         case TCG_COND_LTU:
2232             inv = true;
2233             /* fall through */
2234         case TCG_COND_GTU:
2235         case TCG_COND_GEU:
2236         case TCG_COND_EQ:
2237             return tcg_opt_gen_movi(ctx, op, op->args[0], neg ? -inv : inv);
2238         default:
2239             break;
2240         }
2241     }
2242 
2243     /*
2244      * A with only lsb set is already boolean.
2245      */
2246     if (a_zmask <= 1) {
2247         bool convert = false;
2248         bool inv = false;
2249 
2250         switch (cond) {
2251         case TCG_COND_EQ:
2252             inv = true;
2253             /* fall through */
2254         case TCG_COND_NE:
2255             convert = (b_val == 0);
2256             break;
2257         case TCG_COND_LTU:
2258         case TCG_COND_TSTEQ:
2259             inv = true;
2260             /* fall through */
2261         case TCG_COND_GEU:
2262         case TCG_COND_TSTNE:
2263             convert = (b_val == 1);
2264             break;
2265         default:
2266             break;
2267         }
2268         if (convert) {
2269             TCGOpcode xor_opc, neg_opc;
2270 
2271             if (!inv && !neg) {
2272                 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
2273             }
2274 
2275             switch (ctx->type) {
2276             case TCG_TYPE_I32:
2277                 neg_opc = INDEX_op_neg_i32;
2278                 xor_opc = INDEX_op_xor_i32;
2279                 break;
2280             case TCG_TYPE_I64:
2281                 neg_opc = INDEX_op_neg_i64;
2282                 xor_opc = INDEX_op_xor_i64;
2283                 break;
2284             default:
2285                 g_assert_not_reached();
2286             }
2287 
2288             if (!inv) {
2289                 op->opc = neg_opc;
2290             } else if (neg) {
2291                 op->opc = INDEX_op_add;
2292                 op->args[2] = arg_new_constant(ctx, -1);
2293             } else {
2294                 op->opc = xor_opc;
2295                 op->args[2] = arg_new_constant(ctx, 1);
2296             }
2297             return -1;
2298         }
2299     }
2300     return 0;
2301 }
2302 
2303 static void fold_setcond_tst_pow2(OptContext *ctx, TCGOp *op, bool neg)
2304 {
2305     TCGOpcode xor_opc, neg_opc, shr_opc;
2306     TCGOpcode uext_opc = 0, sext_opc = 0;
2307     TCGCond cond = op->args[3];
2308     TCGArg ret, src1, src2;
2309     TCGOp *op2;
2310     uint64_t val;
2311     int sh;
2312     bool inv;
2313 
2314     if (!is_tst_cond(cond) || !arg_is_const(op->args[2])) {
2315         return;
2316     }
2317 
2318     src2 = op->args[2];
2319     val = arg_info(src2)->val;
2320     if (!is_power_of_2(val)) {
2321         return;
2322     }
2323     sh = ctz64(val);
2324 
2325     switch (ctx->type) {
2326     case TCG_TYPE_I32:
2327         xor_opc = INDEX_op_xor_i32;
2328         shr_opc = INDEX_op_shr_i32;
2329         neg_opc = INDEX_op_neg_i32;
2330         if (TCG_TARGET_extract_valid(TCG_TYPE_I32, sh, 1)) {
2331             uext_opc = INDEX_op_extract_i32;
2332         }
2333         if (TCG_TARGET_sextract_valid(TCG_TYPE_I32, sh, 1)) {
2334             sext_opc = INDEX_op_sextract_i32;
2335         }
2336         break;
2337     case TCG_TYPE_I64:
2338         xor_opc = INDEX_op_xor_i64;
2339         shr_opc = INDEX_op_shr_i64;
2340         neg_opc = INDEX_op_neg_i64;
2341         if (TCG_TARGET_extract_valid(TCG_TYPE_I64, sh, 1)) {
2342             uext_opc = INDEX_op_extract_i64;
2343         }
2344         if (TCG_TARGET_sextract_valid(TCG_TYPE_I64, sh, 1)) {
2345             sext_opc = INDEX_op_sextract_i64;
2346         }
2347         break;
2348     default:
2349         g_assert_not_reached();
2350     }
2351 
2352     ret = op->args[0];
2353     src1 = op->args[1];
2354     inv = cond == TCG_COND_TSTEQ;
2355 
2356     if (sh && sext_opc && neg && !inv) {
2357         op->opc = sext_opc;
2358         op->args[1] = src1;
2359         op->args[2] = sh;
2360         op->args[3] = 1;
2361         return;
2362     } else if (sh && uext_opc) {
2363         op->opc = uext_opc;
2364         op->args[1] = src1;
2365         op->args[2] = sh;
2366         op->args[3] = 1;
2367     } else {
2368         if (sh) {
2369             op2 = opt_insert_before(ctx, op, shr_opc, 3);
2370             op2->args[0] = ret;
2371             op2->args[1] = src1;
2372             op2->args[2] = arg_new_constant(ctx, sh);
2373             src1 = ret;
2374         }
2375         op->opc = INDEX_op_and;
2376         op->args[1] = src1;
2377         op->args[2] = arg_new_constant(ctx, 1);
2378     }
2379 
2380     if (neg && inv) {
2381         op2 = opt_insert_after(ctx, op, INDEX_op_add, 3);
2382         op2->args[0] = ret;
2383         op2->args[1] = ret;
2384         op2->args[2] = arg_new_constant(ctx, -1);
2385     } else if (inv) {
2386         op2 = opt_insert_after(ctx, op, xor_opc, 3);
2387         op2->args[0] = ret;
2388         op2->args[1] = ret;
2389         op2->args[2] = arg_new_constant(ctx, 1);
2390     } else if (neg) {
2391         op2 = opt_insert_after(ctx, op, neg_opc, 2);
2392         op2->args[0] = ret;
2393         op2->args[1] = ret;
2394     }
2395 }
2396 
2397 static bool fold_setcond(OptContext *ctx, TCGOp *op)
2398 {
2399     int i = do_constant_folding_cond1(ctx, op, op->args[0], &op->args[1],
2400                                       &op->args[2], &op->args[3]);
2401     if (i >= 0) {
2402         return tcg_opt_gen_movi(ctx, op, op->args[0], i);
2403     }
2404 
2405     i = fold_setcond_zmask(ctx, op, false);
2406     if (i > 0) {
2407         return true;
2408     }
2409     if (i == 0) {
2410         fold_setcond_tst_pow2(ctx, op, false);
2411     }
2412 
2413     return fold_masks_z(ctx, op, 1);
2414 }
2415 
2416 static bool fold_negsetcond(OptContext *ctx, TCGOp *op)
2417 {
2418     int i = do_constant_folding_cond1(ctx, op, op->args[0], &op->args[1],
2419                                       &op->args[2], &op->args[3]);
2420     if (i >= 0) {
2421         return tcg_opt_gen_movi(ctx, op, op->args[0], -i);
2422     }
2423 
2424     i = fold_setcond_zmask(ctx, op, true);
2425     if (i > 0) {
2426         return true;
2427     }
2428     if (i == 0) {
2429         fold_setcond_tst_pow2(ctx, op, true);
2430     }
2431 
2432     /* Value is {0,-1} so all bits are repetitions of the sign. */
2433     return fold_masks_s(ctx, op, -1);
2434 }
2435 
2436 static bool fold_setcond2(OptContext *ctx, TCGOp *op)
2437 {
2438     TCGCond cond;
2439     int i, inv = 0;
2440 
2441     i = do_constant_folding_cond2(ctx, op, &op->args[1]);
2442     cond = op->args[5];
2443     if (i >= 0) {
2444         goto do_setcond_const;
2445     }
2446 
2447     switch (cond) {
2448     case TCG_COND_LT:
2449     case TCG_COND_GE:
2450         /*
2451          * Simplify LT/GE comparisons vs zero to a single compare
2452          * vs the high word of the input.
2453          */
2454         if (arg_is_const_val(op->args[3], 0) &&
2455             arg_is_const_val(op->args[4], 0)) {
2456             goto do_setcond_high;
2457         }
2458         break;
2459 
2460     case TCG_COND_NE:
2461         inv = 1;
2462         QEMU_FALLTHROUGH;
2463     case TCG_COND_EQ:
2464         /*
2465          * Simplify EQ/NE comparisons where one of the pairs
2466          * can be simplified.
2467          */
2468         i = do_constant_folding_cond(TCG_TYPE_I32, op->args[1],
2469                                      op->args[3], cond);
2470         switch (i ^ inv) {
2471         case 0:
2472             goto do_setcond_const;
2473         case 1:
2474             goto do_setcond_high;
2475         }
2476 
2477         i = do_constant_folding_cond(TCG_TYPE_I32, op->args[2],
2478                                      op->args[4], cond);
2479         switch (i ^ inv) {
2480         case 0:
2481             goto do_setcond_const;
2482         case 1:
2483             goto do_setcond_low;
2484         }
2485         break;
2486 
2487     case TCG_COND_TSTEQ:
2488     case TCG_COND_TSTNE:
2489         if (arg_is_const_val(op->args[3], 0)) {
2490             goto do_setcond_high;
2491         }
2492         if (arg_is_const_val(op->args[4], 0)) {
2493             goto do_setcond_low;
2494         }
2495         break;
2496 
2497     default:
2498         break;
2499 
2500     do_setcond_low:
2501         op->args[2] = op->args[3];
2502         op->args[3] = cond;
2503         op->opc = INDEX_op_setcond_i32;
2504         return fold_setcond(ctx, op);
2505 
2506     do_setcond_high:
2507         op->args[1] = op->args[2];
2508         op->args[2] = op->args[4];
2509         op->args[3] = cond;
2510         op->opc = INDEX_op_setcond_i32;
2511         return fold_setcond(ctx, op);
2512     }
2513 
2514     return fold_masks_z(ctx, op, 1);
2515 
2516  do_setcond_const:
2517     return tcg_opt_gen_movi(ctx, op, op->args[0], i);
2518 }
2519 
2520 static bool fold_sextract(OptContext *ctx, TCGOp *op)
2521 {
2522     uint64_t z_mask, s_mask, s_mask_old;
2523     TempOptInfo *t1 = arg_info(op->args[1]);
2524     int pos = op->args[2];
2525     int len = op->args[3];
2526 
2527     if (ti_is_const(t1)) {
2528         return tcg_opt_gen_movi(ctx, op, op->args[0],
2529                                 sextract64(ti_const_val(t1), pos, len));
2530     }
2531 
2532     s_mask_old = t1->s_mask;
2533     s_mask = s_mask_old >> pos;
2534     s_mask |= -1ull << (len - 1);
2535 
2536     if (pos == 0 && fold_affected_mask(ctx, op, s_mask & ~s_mask_old)) {
2537         return true;
2538     }
2539 
2540     z_mask = sextract64(t1->z_mask, pos, len);
2541     return fold_masks_zs(ctx, op, z_mask, s_mask);
2542 }
2543 
2544 static bool fold_shift(OptContext *ctx, TCGOp *op)
2545 {
2546     uint64_t s_mask, z_mask;
2547     TempOptInfo *t1, *t2;
2548 
2549     if (fold_const2(ctx, op) ||
2550         fold_ix_to_i(ctx, op, 0) ||
2551         fold_xi_to_x(ctx, op, 0)) {
2552         return true;
2553     }
2554 
2555     t1 = arg_info(op->args[1]);
2556     t2 = arg_info(op->args[2]);
2557     s_mask = t1->s_mask;
2558     z_mask = t1->z_mask;
2559 
2560     if (ti_is_const(t2)) {
2561         int sh = ti_const_val(t2);
2562 
2563         z_mask = do_constant_folding(op->opc, ctx->type, z_mask, sh);
2564         s_mask = do_constant_folding(op->opc, ctx->type, s_mask, sh);
2565 
2566         return fold_masks_zs(ctx, op, z_mask, s_mask);
2567     }
2568 
2569     switch (op->opc) {
2570     CASE_OP_32_64(sar):
2571         /*
2572          * Arithmetic right shift will not reduce the number of
2573          * input sign repetitions.
2574          */
2575         return fold_masks_s(ctx, op, s_mask);
2576     CASE_OP_32_64(shr):
2577         /*
2578          * If the sign bit is known zero, then logical right shift
2579          * will not reduce the number of input sign repetitions.
2580          */
2581         if (~z_mask & -s_mask) {
2582             return fold_masks_s(ctx, op, s_mask);
2583         }
2584         break;
2585     default:
2586         break;
2587     }
2588 
2589     return finish_folding(ctx, op);
2590 }
2591 
2592 static bool fold_sub_to_neg(OptContext *ctx, TCGOp *op)
2593 {
2594     TCGOpcode neg_op;
2595     bool have_neg;
2596 
2597     if (!arg_is_const(op->args[1]) || arg_info(op->args[1])->val != 0) {
2598         return false;
2599     }
2600 
2601     switch (ctx->type) {
2602     case TCG_TYPE_I32:
2603         neg_op = INDEX_op_neg_i32;
2604         have_neg = true;
2605         break;
2606     case TCG_TYPE_I64:
2607         neg_op = INDEX_op_neg_i64;
2608         have_neg = true;
2609         break;
2610     case TCG_TYPE_V64:
2611     case TCG_TYPE_V128:
2612     case TCG_TYPE_V256:
2613         neg_op = INDEX_op_neg_vec;
2614         have_neg = (TCG_TARGET_HAS_neg_vec &&
2615                     tcg_can_emit_vec_op(neg_op, ctx->type, TCGOP_VECE(op)) > 0);
2616         break;
2617     default:
2618         g_assert_not_reached();
2619     }
2620     if (have_neg) {
2621         op->opc = neg_op;
2622         op->args[1] = op->args[2];
2623         return fold_neg_no_const(ctx, op);
2624     }
2625     return false;
2626 }
2627 
2628 /* We cannot as yet do_constant_folding with vectors. */
2629 static bool fold_sub_vec(OptContext *ctx, TCGOp *op)
2630 {
2631     if (fold_xx_to_i(ctx, op, 0) ||
2632         fold_xi_to_x(ctx, op, 0) ||
2633         fold_sub_to_neg(ctx, op)) {
2634         return true;
2635     }
2636     return finish_folding(ctx, op);
2637 }
2638 
2639 static bool fold_sub(OptContext *ctx, TCGOp *op)
2640 {
2641     if (fold_const2(ctx, op) ||
2642         fold_xx_to_i(ctx, op, 0) ||
2643         fold_xi_to_x(ctx, op, 0) ||
2644         fold_sub_to_neg(ctx, op)) {
2645         return true;
2646     }
2647 
2648     /* Fold sub r,x,i to add r,x,-i */
2649     if (arg_is_const(op->args[2])) {
2650         uint64_t val = arg_info(op->args[2])->val;
2651 
2652         op->opc = INDEX_op_add;
2653         op->args[2] = arg_new_constant(ctx, -val);
2654     }
2655     return finish_folding(ctx, op);
2656 }
2657 
2658 static bool fold_sub2(OptContext *ctx, TCGOp *op)
2659 {
2660     return fold_addsub2(ctx, op, false);
2661 }
2662 
2663 static bool fold_tcg_ld(OptContext *ctx, TCGOp *op)
2664 {
2665     uint64_t z_mask = -1, s_mask = 0;
2666 
2667     /* We can't do any folding with a load, but we can record bits. */
2668     switch (op->opc) {
2669     CASE_OP_32_64(ld8s):
2670         s_mask = INT8_MIN;
2671         break;
2672     CASE_OP_32_64(ld8u):
2673         z_mask = MAKE_64BIT_MASK(0, 8);
2674         break;
2675     CASE_OP_32_64(ld16s):
2676         s_mask = INT16_MIN;
2677         break;
2678     CASE_OP_32_64(ld16u):
2679         z_mask = MAKE_64BIT_MASK(0, 16);
2680         break;
2681     case INDEX_op_ld32s_i64:
2682         s_mask = INT32_MIN;
2683         break;
2684     case INDEX_op_ld32u_i64:
2685         z_mask = MAKE_64BIT_MASK(0, 32);
2686         break;
2687     default:
2688         g_assert_not_reached();
2689     }
2690     return fold_masks_zs(ctx, op, z_mask, s_mask);
2691 }
2692 
2693 static bool fold_tcg_ld_memcopy(OptContext *ctx, TCGOp *op)
2694 {
2695     TCGTemp *dst, *src;
2696     intptr_t ofs;
2697     TCGType type;
2698 
2699     if (op->args[1] != tcgv_ptr_arg(tcg_env)) {
2700         return finish_folding(ctx, op);
2701     }
2702 
2703     type = ctx->type;
2704     ofs = op->args[2];
2705     dst = arg_temp(op->args[0]);
2706     src = find_mem_copy_for(ctx, type, ofs);
2707     if (src && src->base_type == type) {
2708         return tcg_opt_gen_mov(ctx, op, temp_arg(dst), temp_arg(src));
2709     }
2710 
2711     reset_ts(ctx, dst);
2712     record_mem_copy(ctx, type, dst, ofs, ofs + tcg_type_size(type) - 1);
2713     return true;
2714 }
2715 
2716 static bool fold_tcg_st(OptContext *ctx, TCGOp *op)
2717 {
2718     intptr_t ofs = op->args[2];
2719     intptr_t lm1;
2720 
2721     if (op->args[1] != tcgv_ptr_arg(tcg_env)) {
2722         remove_mem_copy_all(ctx);
2723         return true;
2724     }
2725 
2726     switch (op->opc) {
2727     CASE_OP_32_64(st8):
2728         lm1 = 0;
2729         break;
2730     CASE_OP_32_64(st16):
2731         lm1 = 1;
2732         break;
2733     case INDEX_op_st32_i64:
2734     case INDEX_op_st_i32:
2735         lm1 = 3;
2736         break;
2737     case INDEX_op_st_i64:
2738         lm1 = 7;
2739         break;
2740     case INDEX_op_st_vec:
2741         lm1 = tcg_type_size(ctx->type) - 1;
2742         break;
2743     default:
2744         g_assert_not_reached();
2745     }
2746     remove_mem_copy_in(ctx, ofs, ofs + lm1);
2747     return true;
2748 }
2749 
2750 static bool fold_tcg_st_memcopy(OptContext *ctx, TCGOp *op)
2751 {
2752     TCGTemp *src;
2753     intptr_t ofs, last;
2754     TCGType type;
2755 
2756     if (op->args[1] != tcgv_ptr_arg(tcg_env)) {
2757         return fold_tcg_st(ctx, op);
2758     }
2759 
2760     src = arg_temp(op->args[0]);
2761     ofs = op->args[2];
2762     type = ctx->type;
2763 
2764     /*
2765      * Eliminate duplicate stores of a constant.
2766      * This happens frequently when the target ISA zero-extends.
2767      */
2768     if (ts_is_const(src)) {
2769         TCGTemp *prev = find_mem_copy_for(ctx, type, ofs);
2770         if (src == prev) {
2771             tcg_op_remove(ctx->tcg, op);
2772             return true;
2773         }
2774     }
2775 
2776     last = ofs + tcg_type_size(type) - 1;
2777     remove_mem_copy_in(ctx, ofs, last);
2778     record_mem_copy(ctx, type, src, ofs, last);
2779     return true;
2780 }
2781 
2782 static bool fold_xor(OptContext *ctx, TCGOp *op)
2783 {
2784     uint64_t z_mask, s_mask;
2785     TempOptInfo *t1, *t2;
2786 
2787     if (fold_const2_commutative(ctx, op) ||
2788         fold_xx_to_i(ctx, op, 0) ||
2789         fold_xi_to_x(ctx, op, 0) ||
2790         fold_xi_to_not(ctx, op, -1)) {
2791         return true;
2792     }
2793 
2794     t1 = arg_info(op->args[1]);
2795     t2 = arg_info(op->args[2]);
2796     z_mask = t1->z_mask | t2->z_mask;
2797     s_mask = t1->s_mask & t2->s_mask;
2798     return fold_masks_zs(ctx, op, z_mask, s_mask);
2799 }
2800 
2801 /* Propagate constants and copies, fold constant expressions. */
2802 void tcg_optimize(TCGContext *s)
2803 {
2804     int nb_temps, i;
2805     TCGOp *op, *op_next;
2806     OptContext ctx = { .tcg = s };
2807 
2808     QSIMPLEQ_INIT(&ctx.mem_free);
2809 
2810     /* Array VALS has an element for each temp.
2811        If this temp holds a constant then its value is kept in VALS' element.
2812        If this temp is a copy of other ones then the other copies are
2813        available through the doubly linked circular list. */
2814 
2815     nb_temps = s->nb_temps;
2816     for (i = 0; i < nb_temps; ++i) {
2817         s->temps[i].state_ptr = NULL;
2818     }
2819 
2820     QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) {
2821         TCGOpcode opc = op->opc;
2822         const TCGOpDef *def;
2823         bool done = false;
2824 
2825         /* Calls are special. */
2826         if (opc == INDEX_op_call) {
2827             fold_call(&ctx, op);
2828             continue;
2829         }
2830 
2831         def = &tcg_op_defs[opc];
2832         init_arguments(&ctx, op, def->nb_oargs + def->nb_iargs);
2833         copy_propagate(&ctx, op, def->nb_oargs, def->nb_iargs);
2834 
2835         /* Pre-compute the type of the operation. */
2836         ctx.type = TCGOP_TYPE(op);
2837 
2838         /*
2839          * Process each opcode.
2840          * Sorted alphabetically by opcode as much as possible.
2841          */
2842         switch (opc) {
2843         case INDEX_op_add:
2844             done = fold_add(&ctx, op);
2845             break;
2846         case INDEX_op_add_vec:
2847             done = fold_add_vec(&ctx, op);
2848             break;
2849         CASE_OP_32_64(add2):
2850             done = fold_add2(&ctx, op);
2851             break;
2852         case INDEX_op_and:
2853         case INDEX_op_and_vec:
2854             done = fold_and(&ctx, op);
2855             break;
2856         case INDEX_op_andc:
2857         case INDEX_op_andc_vec:
2858             done = fold_andc(&ctx, op);
2859             break;
2860         CASE_OP_32_64(brcond):
2861             done = fold_brcond(&ctx, op);
2862             break;
2863         case INDEX_op_brcond2_i32:
2864             done = fold_brcond2(&ctx, op);
2865             break;
2866         CASE_OP_32_64(bswap16):
2867         CASE_OP_32_64(bswap32):
2868         case INDEX_op_bswap64_i64:
2869             done = fold_bswap(&ctx, op);
2870             break;
2871         CASE_OP_32_64(clz):
2872         CASE_OP_32_64(ctz):
2873             done = fold_count_zeros(&ctx, op);
2874             break;
2875         CASE_OP_32_64(ctpop):
2876             done = fold_ctpop(&ctx, op);
2877             break;
2878         CASE_OP_32_64(deposit):
2879             done = fold_deposit(&ctx, op);
2880             break;
2881         CASE_OP_32_64(div):
2882         CASE_OP_32_64(divu):
2883             done = fold_divide(&ctx, op);
2884             break;
2885         case INDEX_op_dup_vec:
2886             done = fold_dup(&ctx, op);
2887             break;
2888         case INDEX_op_dup2_vec:
2889             done = fold_dup2(&ctx, op);
2890             break;
2891         CASE_OP_32_64_VEC(eqv):
2892             done = fold_eqv(&ctx, op);
2893             break;
2894         CASE_OP_32_64(extract):
2895             done = fold_extract(&ctx, op);
2896             break;
2897         CASE_OP_32_64(extract2):
2898             done = fold_extract2(&ctx, op);
2899             break;
2900         case INDEX_op_ext_i32_i64:
2901             done = fold_exts(&ctx, op);
2902             break;
2903         case INDEX_op_extu_i32_i64:
2904         case INDEX_op_extrl_i64_i32:
2905         case INDEX_op_extrh_i64_i32:
2906             done = fold_extu(&ctx, op);
2907             break;
2908         CASE_OP_32_64(ld8s):
2909         CASE_OP_32_64(ld8u):
2910         CASE_OP_32_64(ld16s):
2911         CASE_OP_32_64(ld16u):
2912         case INDEX_op_ld32s_i64:
2913         case INDEX_op_ld32u_i64:
2914             done = fold_tcg_ld(&ctx, op);
2915             break;
2916         case INDEX_op_ld_i32:
2917         case INDEX_op_ld_i64:
2918         case INDEX_op_ld_vec:
2919             done = fold_tcg_ld_memcopy(&ctx, op);
2920             break;
2921         CASE_OP_32_64(st8):
2922         CASE_OP_32_64(st16):
2923         case INDEX_op_st32_i64:
2924             done = fold_tcg_st(&ctx, op);
2925             break;
2926         case INDEX_op_st_i32:
2927         case INDEX_op_st_i64:
2928         case INDEX_op_st_vec:
2929             done = fold_tcg_st_memcopy(&ctx, op);
2930             break;
2931         case INDEX_op_mb:
2932             done = fold_mb(&ctx, op);
2933             break;
2934         case INDEX_op_mov:
2935         case INDEX_op_mov_vec:
2936             done = fold_mov(&ctx, op);
2937             break;
2938         CASE_OP_32_64(movcond):
2939             done = fold_movcond(&ctx, op);
2940             break;
2941         CASE_OP_32_64(mul):
2942             done = fold_mul(&ctx, op);
2943             break;
2944         CASE_OP_32_64(mulsh):
2945         CASE_OP_32_64(muluh):
2946             done = fold_mul_highpart(&ctx, op);
2947             break;
2948         CASE_OP_32_64(muls2):
2949         CASE_OP_32_64(mulu2):
2950             done = fold_multiply2(&ctx, op);
2951             break;
2952         CASE_OP_32_64_VEC(nand):
2953             done = fold_nand(&ctx, op);
2954             break;
2955         CASE_OP_32_64(neg):
2956             done = fold_neg(&ctx, op);
2957             break;
2958         CASE_OP_32_64_VEC(nor):
2959             done = fold_nor(&ctx, op);
2960             break;
2961         CASE_OP_32_64_VEC(not):
2962             done = fold_not(&ctx, op);
2963             break;
2964         CASE_OP_32_64_VEC(or):
2965             done = fold_or(&ctx, op);
2966             break;
2967         CASE_OP_32_64_VEC(orc):
2968             done = fold_orc(&ctx, op);
2969             break;
2970         case INDEX_op_qemu_ld_i32:
2971             done = fold_qemu_ld_1reg(&ctx, op);
2972             break;
2973         case INDEX_op_qemu_ld_i64:
2974             if (TCG_TARGET_REG_BITS == 64) {
2975                 done = fold_qemu_ld_1reg(&ctx, op);
2976                 break;
2977             }
2978             QEMU_FALLTHROUGH;
2979         case INDEX_op_qemu_ld_i128:
2980             done = fold_qemu_ld_2reg(&ctx, op);
2981             break;
2982         case INDEX_op_qemu_st8_i32:
2983         case INDEX_op_qemu_st_i32:
2984         case INDEX_op_qemu_st_i64:
2985         case INDEX_op_qemu_st_i128:
2986             done = fold_qemu_st(&ctx, op);
2987             break;
2988         CASE_OP_32_64(rem):
2989         CASE_OP_32_64(remu):
2990             done = fold_remainder(&ctx, op);
2991             break;
2992         CASE_OP_32_64(rotl):
2993         CASE_OP_32_64(rotr):
2994         CASE_OP_32_64(sar):
2995         CASE_OP_32_64(shl):
2996         CASE_OP_32_64(shr):
2997             done = fold_shift(&ctx, op);
2998             break;
2999         CASE_OP_32_64(setcond):
3000             done = fold_setcond(&ctx, op);
3001             break;
3002         CASE_OP_32_64(negsetcond):
3003             done = fold_negsetcond(&ctx, op);
3004             break;
3005         case INDEX_op_setcond2_i32:
3006             done = fold_setcond2(&ctx, op);
3007             break;
3008         case INDEX_op_cmp_vec:
3009             done = fold_cmp_vec(&ctx, op);
3010             break;
3011         case INDEX_op_cmpsel_vec:
3012             done = fold_cmpsel_vec(&ctx, op);
3013             break;
3014         case INDEX_op_bitsel_vec:
3015             done = fold_bitsel_vec(&ctx, op);
3016             break;
3017         CASE_OP_32_64(sextract):
3018             done = fold_sextract(&ctx, op);
3019             break;
3020         CASE_OP_32_64(sub):
3021             done = fold_sub(&ctx, op);
3022             break;
3023         case INDEX_op_sub_vec:
3024             done = fold_sub_vec(&ctx, op);
3025             break;
3026         CASE_OP_32_64(sub2):
3027             done = fold_sub2(&ctx, op);
3028             break;
3029         CASE_OP_32_64_VEC(xor):
3030             done = fold_xor(&ctx, op);
3031             break;
3032         case INDEX_op_set_label:
3033         case INDEX_op_br:
3034         case INDEX_op_exit_tb:
3035         case INDEX_op_goto_tb:
3036         case INDEX_op_goto_ptr:
3037             finish_ebb(&ctx);
3038             done = true;
3039             break;
3040         default:
3041             done = finish_folding(&ctx, op);
3042             break;
3043         }
3044         tcg_debug_assert(done);
3045     }
3046 }
3047