xref: /openbmc/qemu/tcg/optimize.c (revision b8116f4c)
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 
32 #define CASE_OP_32_64(x)                        \
33         glue(glue(case INDEX_op_, x), _i32):    \
34         glue(glue(case INDEX_op_, x), _i64)
35 
36 #define CASE_OP_32_64_VEC(x)                    \
37         glue(glue(case INDEX_op_, x), _i32):    \
38         glue(glue(case INDEX_op_, x), _i64):    \
39         glue(glue(case INDEX_op_, x), _vec)
40 
41 typedef struct MemCopyInfo {
42     IntervalTreeNode itree;
43     QSIMPLEQ_ENTRY (MemCopyInfo) next;
44     TCGTemp *ts;
45     TCGType type;
46 } MemCopyInfo;
47 
48 typedef struct TempOptInfo {
49     bool is_const;
50     TCGTemp *prev_copy;
51     TCGTemp *next_copy;
52     QSIMPLEQ_HEAD(, MemCopyInfo) mem_copy;
53     uint64_t val;
54     uint64_t z_mask;  /* mask bit is 0 if and only if value bit is 0 */
55     uint64_t s_mask;  /* a left-aligned mask of clrsb(value) bits. */
56 } TempOptInfo;
57 
58 typedef struct OptContext {
59     TCGContext *tcg;
60     TCGOp *prev_mb;
61     TCGTempSet temps_used;
62 
63     IntervalTreeRoot mem_copy;
64     QSIMPLEQ_HEAD(, MemCopyInfo) mem_free;
65 
66     /* In flight values from optimization. */
67     uint64_t a_mask;  /* mask bit is 0 iff value identical to first input */
68     uint64_t z_mask;  /* mask bit is 0 iff value bit is 0 */
69     uint64_t s_mask;  /* mask of clrsb(value) bits */
70     TCGType type;
71 } OptContext;
72 
73 /* Calculate the smask for a specific value. */
74 static uint64_t smask_from_value(uint64_t value)
75 {
76     int rep = clrsb64(value);
77     return ~(~0ull >> rep);
78 }
79 
80 /*
81  * Calculate the smask for a given set of known-zeros.
82  * If there are lots of zeros on the left, we can consider the remainder
83  * an unsigned field, and thus the corresponding signed field is one bit
84  * larger.
85  */
86 static uint64_t smask_from_zmask(uint64_t zmask)
87 {
88     /*
89      * Only the 0 bits are significant for zmask, thus the msb itself
90      * must be zero, else we have no sign information.
91      */
92     int rep = clz64(zmask);
93     if (rep == 0) {
94         return 0;
95     }
96     rep -= 1;
97     return ~(~0ull >> rep);
98 }
99 
100 /*
101  * Recreate a properly left-aligned smask after manipulation.
102  * Some bit-shuffling, particularly shifts and rotates, may
103  * retain sign bits on the left, but may scatter disconnected
104  * sign bits on the right.  Retain only what remains to the left.
105  */
106 static uint64_t smask_from_smask(int64_t smask)
107 {
108     /* Only the 1 bits are significant for smask */
109     return smask_from_zmask(~smask);
110 }
111 
112 static inline TempOptInfo *ts_info(TCGTemp *ts)
113 {
114     return ts->state_ptr;
115 }
116 
117 static inline TempOptInfo *arg_info(TCGArg arg)
118 {
119     return ts_info(arg_temp(arg));
120 }
121 
122 static inline bool ts_is_const(TCGTemp *ts)
123 {
124     return ts_info(ts)->is_const;
125 }
126 
127 static inline bool ts_is_const_val(TCGTemp *ts, uint64_t val)
128 {
129     TempOptInfo *ti = ts_info(ts);
130     return ti->is_const && ti->val == val;
131 }
132 
133 static inline bool arg_is_const(TCGArg arg)
134 {
135     return ts_is_const(arg_temp(arg));
136 }
137 
138 static inline bool arg_is_const_val(TCGArg arg, uint64_t val)
139 {
140     return ts_is_const_val(arg_temp(arg), val);
141 }
142 
143 static inline bool ts_is_copy(TCGTemp *ts)
144 {
145     return ts_info(ts)->next_copy != ts;
146 }
147 
148 static TCGTemp *cmp_better_copy(TCGTemp *a, TCGTemp *b)
149 {
150     return a->kind < b->kind ? b : a;
151 }
152 
153 /* Initialize and activate a temporary.  */
154 static void init_ts_info(OptContext *ctx, TCGTemp *ts)
155 {
156     size_t idx = temp_idx(ts);
157     TempOptInfo *ti;
158 
159     if (test_bit(idx, ctx->temps_used.l)) {
160         return;
161     }
162     set_bit(idx, ctx->temps_used.l);
163 
164     ti = ts->state_ptr;
165     if (ti == NULL) {
166         ti = tcg_malloc(sizeof(TempOptInfo));
167         ts->state_ptr = ti;
168     }
169 
170     ti->next_copy = ts;
171     ti->prev_copy = ts;
172     QSIMPLEQ_INIT(&ti->mem_copy);
173     if (ts->kind == TEMP_CONST) {
174         ti->is_const = true;
175         ti->val = ts->val;
176         ti->z_mask = ts->val;
177         ti->s_mask = smask_from_value(ts->val);
178     } else {
179         ti->is_const = false;
180         ti->z_mask = -1;
181         ti->s_mask = 0;
182     }
183 }
184 
185 static MemCopyInfo *mem_copy_first(OptContext *ctx, intptr_t s, intptr_t l)
186 {
187     IntervalTreeNode *r = interval_tree_iter_first(&ctx->mem_copy, s, l);
188     return r ? container_of(r, MemCopyInfo, itree) : NULL;
189 }
190 
191 static MemCopyInfo *mem_copy_next(MemCopyInfo *mem, intptr_t s, intptr_t l)
192 {
193     IntervalTreeNode *r = interval_tree_iter_next(&mem->itree, s, l);
194     return r ? container_of(r, MemCopyInfo, itree) : NULL;
195 }
196 
197 static void remove_mem_copy(OptContext *ctx, MemCopyInfo *mc)
198 {
199     TCGTemp *ts = mc->ts;
200     TempOptInfo *ti = ts_info(ts);
201 
202     interval_tree_remove(&mc->itree, &ctx->mem_copy);
203     QSIMPLEQ_REMOVE(&ti->mem_copy, mc, MemCopyInfo, next);
204     QSIMPLEQ_INSERT_TAIL(&ctx->mem_free, mc, next);
205 }
206 
207 static void remove_mem_copy_in(OptContext *ctx, intptr_t s, intptr_t l)
208 {
209     while (true) {
210         MemCopyInfo *mc = mem_copy_first(ctx, s, l);
211         if (!mc) {
212             break;
213         }
214         remove_mem_copy(ctx, mc);
215     }
216 }
217 
218 static void remove_mem_copy_all(OptContext *ctx)
219 {
220     remove_mem_copy_in(ctx, 0, -1);
221     tcg_debug_assert(interval_tree_is_empty(&ctx->mem_copy));
222 }
223 
224 static TCGTemp *find_better_copy(TCGTemp *ts)
225 {
226     TCGTemp *i, *ret;
227 
228     /* If this is already readonly, we can't do better. */
229     if (temp_readonly(ts)) {
230         return ts;
231     }
232 
233     ret = ts;
234     for (i = ts_info(ts)->next_copy; i != ts; i = ts_info(i)->next_copy) {
235         ret = cmp_better_copy(ret, i);
236     }
237     return ret;
238 }
239 
240 static void move_mem_copies(TCGTemp *dst_ts, TCGTemp *src_ts)
241 {
242     TempOptInfo *si = ts_info(src_ts);
243     TempOptInfo *di = ts_info(dst_ts);
244     MemCopyInfo *mc;
245 
246     QSIMPLEQ_FOREACH(mc, &si->mem_copy, next) {
247         tcg_debug_assert(mc->ts == src_ts);
248         mc->ts = dst_ts;
249     }
250     QSIMPLEQ_CONCAT(&di->mem_copy, &si->mem_copy);
251 }
252 
253 /* Reset TEMP's state, possibly removing the temp for the list of copies.  */
254 static void reset_ts(OptContext *ctx, TCGTemp *ts)
255 {
256     TempOptInfo *ti = ts_info(ts);
257     TCGTemp *pts = ti->prev_copy;
258     TCGTemp *nts = ti->next_copy;
259     TempOptInfo *pi = ts_info(pts);
260     TempOptInfo *ni = ts_info(nts);
261 
262     ni->prev_copy = ti->prev_copy;
263     pi->next_copy = ti->next_copy;
264     ti->next_copy = ts;
265     ti->prev_copy = ts;
266     ti->is_const = false;
267     ti->z_mask = -1;
268     ti->s_mask = 0;
269 
270     if (!QSIMPLEQ_EMPTY(&ti->mem_copy)) {
271         if (ts == nts) {
272             /* Last temp copy being removed, the mem copies die. */
273             MemCopyInfo *mc;
274             QSIMPLEQ_FOREACH(mc, &ti->mem_copy, next) {
275                 interval_tree_remove(&mc->itree, &ctx->mem_copy);
276             }
277             QSIMPLEQ_CONCAT(&ctx->mem_free, &ti->mem_copy);
278         } else {
279             move_mem_copies(find_better_copy(nts), ts);
280         }
281     }
282 }
283 
284 static void reset_temp(OptContext *ctx, TCGArg arg)
285 {
286     reset_ts(ctx, arg_temp(arg));
287 }
288 
289 static void record_mem_copy(OptContext *ctx, TCGType type,
290                             TCGTemp *ts, intptr_t start, intptr_t last)
291 {
292     MemCopyInfo *mc;
293     TempOptInfo *ti;
294 
295     mc = QSIMPLEQ_FIRST(&ctx->mem_free);
296     if (mc) {
297         QSIMPLEQ_REMOVE_HEAD(&ctx->mem_free, next);
298     } else {
299         mc = tcg_malloc(sizeof(*mc));
300     }
301 
302     memset(mc, 0, sizeof(*mc));
303     mc->itree.start = start;
304     mc->itree.last = last;
305     mc->type = type;
306     interval_tree_insert(&mc->itree, &ctx->mem_copy);
307 
308     ts = find_better_copy(ts);
309     ti = ts_info(ts);
310     mc->ts = ts;
311     QSIMPLEQ_INSERT_TAIL(&ti->mem_copy, mc, next);
312 }
313 
314 static bool ts_are_copies(TCGTemp *ts1, TCGTemp *ts2)
315 {
316     TCGTemp *i;
317 
318     if (ts1 == ts2) {
319         return true;
320     }
321 
322     if (!ts_is_copy(ts1) || !ts_is_copy(ts2)) {
323         return false;
324     }
325 
326     for (i = ts_info(ts1)->next_copy; i != ts1; i = ts_info(i)->next_copy) {
327         if (i == ts2) {
328             return true;
329         }
330     }
331 
332     return false;
333 }
334 
335 static bool args_are_copies(TCGArg arg1, TCGArg arg2)
336 {
337     return ts_are_copies(arg_temp(arg1), arg_temp(arg2));
338 }
339 
340 static TCGTemp *find_mem_copy_for(OptContext *ctx, TCGType type, intptr_t s)
341 {
342     MemCopyInfo *mc;
343 
344     for (mc = mem_copy_first(ctx, s, s); mc; mc = mem_copy_next(mc, s, s)) {
345         if (mc->itree.start == s && mc->type == type) {
346             return find_better_copy(mc->ts);
347         }
348     }
349     return NULL;
350 }
351 
352 static TCGArg arg_new_constant(OptContext *ctx, uint64_t val)
353 {
354     TCGType type = ctx->type;
355     TCGTemp *ts;
356 
357     if (type == TCG_TYPE_I32) {
358         val = (int32_t)val;
359     }
360 
361     ts = tcg_constant_internal(type, val);
362     init_ts_info(ctx, ts);
363 
364     return temp_arg(ts);
365 }
366 
367 static TCGArg arg_new_temp(OptContext *ctx)
368 {
369     TCGTemp *ts = tcg_temp_new_internal(ctx->type, TEMP_EBB);
370     init_ts_info(ctx, ts);
371     return temp_arg(ts);
372 }
373 
374 static bool tcg_opt_gen_mov(OptContext *ctx, TCGOp *op, TCGArg dst, TCGArg src)
375 {
376     TCGTemp *dst_ts = arg_temp(dst);
377     TCGTemp *src_ts = arg_temp(src);
378     TempOptInfo *di;
379     TempOptInfo *si;
380     TCGOpcode new_op;
381 
382     if (ts_are_copies(dst_ts, src_ts)) {
383         tcg_op_remove(ctx->tcg, op);
384         return true;
385     }
386 
387     reset_ts(ctx, dst_ts);
388     di = ts_info(dst_ts);
389     si = ts_info(src_ts);
390 
391     switch (ctx->type) {
392     case TCG_TYPE_I32:
393         new_op = INDEX_op_mov_i32;
394         break;
395     case TCG_TYPE_I64:
396         new_op = INDEX_op_mov_i64;
397         break;
398     case TCG_TYPE_V64:
399     case TCG_TYPE_V128:
400     case TCG_TYPE_V256:
401         /* TCGOP_VECL and TCGOP_VECE remain unchanged.  */
402         new_op = INDEX_op_mov_vec;
403         break;
404     default:
405         g_assert_not_reached();
406     }
407     op->opc = new_op;
408     op->args[0] = dst;
409     op->args[1] = src;
410 
411     di->z_mask = si->z_mask;
412     di->s_mask = si->s_mask;
413 
414     if (src_ts->type == dst_ts->type) {
415         TempOptInfo *ni = ts_info(si->next_copy);
416 
417         di->next_copy = si->next_copy;
418         di->prev_copy = src_ts;
419         ni->prev_copy = dst_ts;
420         si->next_copy = dst_ts;
421         di->is_const = si->is_const;
422         di->val = si->val;
423 
424         if (!QSIMPLEQ_EMPTY(&si->mem_copy)
425             && cmp_better_copy(src_ts, dst_ts) == dst_ts) {
426             move_mem_copies(dst_ts, src_ts);
427         }
428     }
429     return true;
430 }
431 
432 static bool tcg_opt_gen_movi(OptContext *ctx, TCGOp *op,
433                              TCGArg dst, uint64_t val)
434 {
435     /* Convert movi to mov with constant temp. */
436     return tcg_opt_gen_mov(ctx, op, dst, arg_new_constant(ctx, val));
437 }
438 
439 static uint64_t do_constant_folding_2(TCGOpcode op, uint64_t x, uint64_t y)
440 {
441     uint64_t l64, h64;
442 
443     switch (op) {
444     CASE_OP_32_64(add):
445         return x + y;
446 
447     CASE_OP_32_64(sub):
448         return x - y;
449 
450     CASE_OP_32_64(mul):
451         return x * y;
452 
453     CASE_OP_32_64_VEC(and):
454         return x & y;
455 
456     CASE_OP_32_64_VEC(or):
457         return x | y;
458 
459     CASE_OP_32_64_VEC(xor):
460         return x ^ y;
461 
462     case INDEX_op_shl_i32:
463         return (uint32_t)x << (y & 31);
464 
465     case INDEX_op_shl_i64:
466         return (uint64_t)x << (y & 63);
467 
468     case INDEX_op_shr_i32:
469         return (uint32_t)x >> (y & 31);
470 
471     case INDEX_op_shr_i64:
472         return (uint64_t)x >> (y & 63);
473 
474     case INDEX_op_sar_i32:
475         return (int32_t)x >> (y & 31);
476 
477     case INDEX_op_sar_i64:
478         return (int64_t)x >> (y & 63);
479 
480     case INDEX_op_rotr_i32:
481         return ror32(x, y & 31);
482 
483     case INDEX_op_rotr_i64:
484         return ror64(x, y & 63);
485 
486     case INDEX_op_rotl_i32:
487         return rol32(x, y & 31);
488 
489     case INDEX_op_rotl_i64:
490         return rol64(x, y & 63);
491 
492     CASE_OP_32_64_VEC(not):
493         return ~x;
494 
495     CASE_OP_32_64(neg):
496         return -x;
497 
498     CASE_OP_32_64_VEC(andc):
499         return x & ~y;
500 
501     CASE_OP_32_64_VEC(orc):
502         return x | ~y;
503 
504     CASE_OP_32_64_VEC(eqv):
505         return ~(x ^ y);
506 
507     CASE_OP_32_64_VEC(nand):
508         return ~(x & y);
509 
510     CASE_OP_32_64_VEC(nor):
511         return ~(x | y);
512 
513     case INDEX_op_clz_i32:
514         return (uint32_t)x ? clz32(x) : y;
515 
516     case INDEX_op_clz_i64:
517         return x ? clz64(x) : y;
518 
519     case INDEX_op_ctz_i32:
520         return (uint32_t)x ? ctz32(x) : y;
521 
522     case INDEX_op_ctz_i64:
523         return x ? ctz64(x) : y;
524 
525     case INDEX_op_ctpop_i32:
526         return ctpop32(x);
527 
528     case INDEX_op_ctpop_i64:
529         return ctpop64(x);
530 
531     CASE_OP_32_64(ext8s):
532         return (int8_t)x;
533 
534     CASE_OP_32_64(ext16s):
535         return (int16_t)x;
536 
537     CASE_OP_32_64(ext8u):
538         return (uint8_t)x;
539 
540     CASE_OP_32_64(ext16u):
541         return (uint16_t)x;
542 
543     CASE_OP_32_64(bswap16):
544         x = bswap16(x);
545         return y & TCG_BSWAP_OS ? (int16_t)x : x;
546 
547     CASE_OP_32_64(bswap32):
548         x = bswap32(x);
549         return y & TCG_BSWAP_OS ? (int32_t)x : x;
550 
551     case INDEX_op_bswap64_i64:
552         return bswap64(x);
553 
554     case INDEX_op_ext_i32_i64:
555     case INDEX_op_ext32s_i64:
556         return (int32_t)x;
557 
558     case INDEX_op_extu_i32_i64:
559     case INDEX_op_extrl_i64_i32:
560     case INDEX_op_ext32u_i64:
561         return (uint32_t)x;
562 
563     case INDEX_op_extrh_i64_i32:
564         return (uint64_t)x >> 32;
565 
566     case INDEX_op_muluh_i32:
567         return ((uint64_t)(uint32_t)x * (uint32_t)y) >> 32;
568     case INDEX_op_mulsh_i32:
569         return ((int64_t)(int32_t)x * (int32_t)y) >> 32;
570 
571     case INDEX_op_muluh_i64:
572         mulu64(&l64, &h64, x, y);
573         return h64;
574     case INDEX_op_mulsh_i64:
575         muls64(&l64, &h64, x, y);
576         return h64;
577 
578     case INDEX_op_div_i32:
579         /* Avoid crashing on divide by zero, otherwise undefined.  */
580         return (int32_t)x / ((int32_t)y ? : 1);
581     case INDEX_op_divu_i32:
582         return (uint32_t)x / ((uint32_t)y ? : 1);
583     case INDEX_op_div_i64:
584         return (int64_t)x / ((int64_t)y ? : 1);
585     case INDEX_op_divu_i64:
586         return (uint64_t)x / ((uint64_t)y ? : 1);
587 
588     case INDEX_op_rem_i32:
589         return (int32_t)x % ((int32_t)y ? : 1);
590     case INDEX_op_remu_i32:
591         return (uint32_t)x % ((uint32_t)y ? : 1);
592     case INDEX_op_rem_i64:
593         return (int64_t)x % ((int64_t)y ? : 1);
594     case INDEX_op_remu_i64:
595         return (uint64_t)x % ((uint64_t)y ? : 1);
596 
597     default:
598         g_assert_not_reached();
599     }
600 }
601 
602 static uint64_t do_constant_folding(TCGOpcode op, TCGType type,
603                                     uint64_t x, uint64_t y)
604 {
605     uint64_t res = do_constant_folding_2(op, x, y);
606     if (type == TCG_TYPE_I32) {
607         res = (int32_t)res;
608     }
609     return res;
610 }
611 
612 static bool do_constant_folding_cond_32(uint32_t x, uint32_t y, TCGCond c)
613 {
614     switch (c) {
615     case TCG_COND_EQ:
616         return x == y;
617     case TCG_COND_NE:
618         return x != y;
619     case TCG_COND_LT:
620         return (int32_t)x < (int32_t)y;
621     case TCG_COND_GE:
622         return (int32_t)x >= (int32_t)y;
623     case TCG_COND_LE:
624         return (int32_t)x <= (int32_t)y;
625     case TCG_COND_GT:
626         return (int32_t)x > (int32_t)y;
627     case TCG_COND_LTU:
628         return x < y;
629     case TCG_COND_GEU:
630         return x >= y;
631     case TCG_COND_LEU:
632         return x <= y;
633     case TCG_COND_GTU:
634         return x > y;
635     case TCG_COND_TSTEQ:
636         return (x & y) == 0;
637     case TCG_COND_TSTNE:
638         return (x & y) != 0;
639     case TCG_COND_ALWAYS:
640     case TCG_COND_NEVER:
641         break;
642     }
643     g_assert_not_reached();
644 }
645 
646 static bool do_constant_folding_cond_64(uint64_t x, uint64_t y, TCGCond c)
647 {
648     switch (c) {
649     case TCG_COND_EQ:
650         return x == y;
651     case TCG_COND_NE:
652         return x != y;
653     case TCG_COND_LT:
654         return (int64_t)x < (int64_t)y;
655     case TCG_COND_GE:
656         return (int64_t)x >= (int64_t)y;
657     case TCG_COND_LE:
658         return (int64_t)x <= (int64_t)y;
659     case TCG_COND_GT:
660         return (int64_t)x > (int64_t)y;
661     case TCG_COND_LTU:
662         return x < y;
663     case TCG_COND_GEU:
664         return x >= y;
665     case TCG_COND_LEU:
666         return x <= y;
667     case TCG_COND_GTU:
668         return x > y;
669     case TCG_COND_TSTEQ:
670         return (x & y) == 0;
671     case TCG_COND_TSTNE:
672         return (x & y) != 0;
673     case TCG_COND_ALWAYS:
674     case TCG_COND_NEVER:
675         break;
676     }
677     g_assert_not_reached();
678 }
679 
680 static int do_constant_folding_cond_eq(TCGCond c)
681 {
682     switch (c) {
683     case TCG_COND_GT:
684     case TCG_COND_LTU:
685     case TCG_COND_LT:
686     case TCG_COND_GTU:
687     case TCG_COND_NE:
688         return 0;
689     case TCG_COND_GE:
690     case TCG_COND_GEU:
691     case TCG_COND_LE:
692     case TCG_COND_LEU:
693     case TCG_COND_EQ:
694         return 1;
695     case TCG_COND_TSTEQ:
696     case TCG_COND_TSTNE:
697         return -1;
698     case TCG_COND_ALWAYS:
699     case TCG_COND_NEVER:
700         break;
701     }
702     g_assert_not_reached();
703 }
704 
705 /*
706  * Return -1 if the condition can't be simplified,
707  * and the result of the condition (0 or 1) if it can.
708  */
709 static int do_constant_folding_cond(TCGType type, TCGArg x,
710                                     TCGArg y, TCGCond c)
711 {
712     if (arg_is_const(x) && arg_is_const(y)) {
713         uint64_t xv = arg_info(x)->val;
714         uint64_t yv = arg_info(y)->val;
715 
716         switch (type) {
717         case TCG_TYPE_I32:
718             return do_constant_folding_cond_32(xv, yv, c);
719         case TCG_TYPE_I64:
720             return do_constant_folding_cond_64(xv, yv, c);
721         default:
722             /* Only scalar comparisons are optimizable */
723             return -1;
724         }
725     } else if (args_are_copies(x, y)) {
726         return do_constant_folding_cond_eq(c);
727     } else if (arg_is_const_val(y, 0)) {
728         switch (c) {
729         case TCG_COND_LTU:
730         case TCG_COND_TSTNE:
731             return 0;
732         case TCG_COND_GEU:
733         case TCG_COND_TSTEQ:
734             return 1;
735         default:
736             return -1;
737         }
738     }
739     return -1;
740 }
741 
742 /**
743  * swap_commutative:
744  * @dest: TCGArg of the destination argument, or NO_DEST.
745  * @p1: first paired argument
746  * @p2: second paired argument
747  *
748  * If *@p1 is a constant and *@p2 is not, swap.
749  * If *@p2 matches @dest, swap.
750  * Return true if a swap was performed.
751  */
752 
753 #define NO_DEST  temp_arg(NULL)
754 
755 static bool swap_commutative(TCGArg dest, TCGArg *p1, TCGArg *p2)
756 {
757     TCGArg a1 = *p1, a2 = *p2;
758     int sum = 0;
759     sum += arg_is_const(a1);
760     sum -= arg_is_const(a2);
761 
762     /* Prefer the constant in second argument, and then the form
763        op a, a, b, which is better handled on non-RISC hosts. */
764     if (sum > 0 || (sum == 0 && dest == a2)) {
765         *p1 = a2;
766         *p2 = a1;
767         return true;
768     }
769     return false;
770 }
771 
772 static bool swap_commutative2(TCGArg *p1, TCGArg *p2)
773 {
774     int sum = 0;
775     sum += arg_is_const(p1[0]);
776     sum += arg_is_const(p1[1]);
777     sum -= arg_is_const(p2[0]);
778     sum -= arg_is_const(p2[1]);
779     if (sum > 0) {
780         TCGArg t;
781         t = p1[0], p1[0] = p2[0], p2[0] = t;
782         t = p1[1], p1[1] = p2[1], p2[1] = t;
783         return true;
784     }
785     return false;
786 }
787 
788 /*
789  * Return -1 if the condition can't be simplified,
790  * and the result of the condition (0 or 1) if it can.
791  */
792 static int do_constant_folding_cond1(OptContext *ctx, TCGOp *op, TCGArg dest,
793                                      TCGArg *p1, TCGArg *p2, TCGArg *pcond)
794 {
795     TCGCond cond;
796     bool swap;
797     int r;
798 
799     swap = swap_commutative(dest, p1, p2);
800     cond = *pcond;
801     if (swap) {
802         *pcond = cond = tcg_swap_cond(cond);
803     }
804 
805     r = do_constant_folding_cond(ctx->type, *p1, *p2, cond);
806     if (r >= 0) {
807         return r;
808     }
809     if (!is_tst_cond(cond)) {
810         return -1;
811     }
812 
813     /*
814      * TSTNE x,x -> NE x,0
815      * TSTNE x,-1 -> NE x,0
816      */
817     if (args_are_copies(*p1, *p2) || arg_is_const_val(*p2, -1)) {
818         *p2 = arg_new_constant(ctx, 0);
819         *pcond = tcg_tst_eqne_cond(cond);
820         return -1;
821     }
822 
823     /* TSTNE x,sign -> LT x,0 */
824     if (arg_is_const_val(*p2, (ctx->type == TCG_TYPE_I32
825                                ? INT32_MIN : INT64_MIN))) {
826         *p2 = arg_new_constant(ctx, 0);
827         *pcond = tcg_tst_ltge_cond(cond);
828         return -1;
829     }
830 
831     /* Expand to AND with a temporary if no backend support. */
832     if (!TCG_TARGET_HAS_tst) {
833         TCGOpcode and_opc = (ctx->type == TCG_TYPE_I32
834                              ? INDEX_op_and_i32 : INDEX_op_and_i64);
835         TCGOp *op2 = tcg_op_insert_before(ctx->tcg, op, and_opc, 3);
836         TCGArg tmp = arg_new_temp(ctx);
837 
838         op2->args[0] = tmp;
839         op2->args[1] = *p1;
840         op2->args[2] = *p2;
841 
842         *p1 = tmp;
843         *p2 = arg_new_constant(ctx, 0);
844         *pcond = tcg_tst_eqne_cond(cond);
845     }
846     return -1;
847 }
848 
849 static int do_constant_folding_cond2(OptContext *ctx, TCGOp *op, TCGArg *args)
850 {
851     TCGArg al, ah, bl, bh;
852     TCGCond c;
853     bool swap;
854     int r;
855 
856     swap = swap_commutative2(args, args + 2);
857     c = args[4];
858     if (swap) {
859         args[4] = c = tcg_swap_cond(c);
860     }
861 
862     al = args[0];
863     ah = args[1];
864     bl = args[2];
865     bh = args[3];
866 
867     if (arg_is_const(bl) && arg_is_const(bh)) {
868         tcg_target_ulong blv = arg_info(bl)->val;
869         tcg_target_ulong bhv = arg_info(bh)->val;
870         uint64_t b = deposit64(blv, 32, 32, bhv);
871 
872         if (arg_is_const(al) && arg_is_const(ah)) {
873             tcg_target_ulong alv = arg_info(al)->val;
874             tcg_target_ulong ahv = arg_info(ah)->val;
875             uint64_t a = deposit64(alv, 32, 32, ahv);
876 
877             r = do_constant_folding_cond_64(a, b, c);
878             if (r >= 0) {
879                 return r;
880             }
881         }
882 
883         if (b == 0) {
884             switch (c) {
885             case TCG_COND_LTU:
886             case TCG_COND_TSTNE:
887                 return 0;
888             case TCG_COND_GEU:
889             case TCG_COND_TSTEQ:
890                 return 1;
891             default:
892                 break;
893             }
894         }
895 
896         /* TSTNE x,-1 -> NE x,0 */
897         if (b == -1 && is_tst_cond(c)) {
898             args[3] = args[2] = arg_new_constant(ctx, 0);
899             args[4] = tcg_tst_eqne_cond(c);
900             return -1;
901         }
902 
903         /* TSTNE x,sign -> LT x,0 */
904         if (b == INT64_MIN && is_tst_cond(c)) {
905             /* bl must be 0, so copy that to bh */
906             args[3] = bl;
907             args[4] = tcg_tst_ltge_cond(c);
908             return -1;
909         }
910     }
911 
912     if (args_are_copies(al, bl) && args_are_copies(ah, bh)) {
913         r = do_constant_folding_cond_eq(c);
914         if (r >= 0) {
915             return r;
916         }
917 
918         /* TSTNE x,x -> NE x,0 */
919         if (is_tst_cond(c)) {
920             args[3] = args[2] = arg_new_constant(ctx, 0);
921             args[4] = tcg_tst_eqne_cond(c);
922             return -1;
923         }
924     }
925 
926     /* Expand to AND with a temporary if no backend support. */
927     if (!TCG_TARGET_HAS_tst && is_tst_cond(c)) {
928         TCGOp *op1 = tcg_op_insert_before(ctx->tcg, op, INDEX_op_and_i32, 3);
929         TCGOp *op2 = tcg_op_insert_before(ctx->tcg, op, INDEX_op_and_i32, 3);
930         TCGArg t1 = arg_new_temp(ctx);
931         TCGArg t2 = arg_new_temp(ctx);
932 
933         op1->args[0] = t1;
934         op1->args[1] = al;
935         op1->args[2] = bl;
936         op2->args[0] = t2;
937         op2->args[1] = ah;
938         op2->args[2] = bh;
939 
940         args[0] = t1;
941         args[1] = t2;
942         args[3] = args[2] = arg_new_constant(ctx, 0);
943         args[4] = tcg_tst_eqne_cond(c);
944     }
945     return -1;
946 }
947 
948 static void init_arguments(OptContext *ctx, TCGOp *op, int nb_args)
949 {
950     for (int i = 0; i < nb_args; i++) {
951         TCGTemp *ts = arg_temp(op->args[i]);
952         init_ts_info(ctx, ts);
953     }
954 }
955 
956 static void copy_propagate(OptContext *ctx, TCGOp *op,
957                            int nb_oargs, int nb_iargs)
958 {
959     for (int i = nb_oargs; i < nb_oargs + nb_iargs; i++) {
960         TCGTemp *ts = arg_temp(op->args[i]);
961         if (ts_is_copy(ts)) {
962             op->args[i] = temp_arg(find_better_copy(ts));
963         }
964     }
965 }
966 
967 static void finish_folding(OptContext *ctx, TCGOp *op)
968 {
969     const TCGOpDef *def = &tcg_op_defs[op->opc];
970     int i, nb_oargs;
971 
972     /*
973      * We only optimize extended basic blocks.  If the opcode ends a BB
974      * and is not a conditional branch, reset all temp data.
975      */
976     if (def->flags & TCG_OPF_BB_END) {
977         ctx->prev_mb = NULL;
978         if (!(def->flags & TCG_OPF_COND_BRANCH)) {
979             memset(&ctx->temps_used, 0, sizeof(ctx->temps_used));
980             remove_mem_copy_all(ctx);
981         }
982         return;
983     }
984 
985     nb_oargs = def->nb_oargs;
986     for (i = 0; i < nb_oargs; i++) {
987         TCGTemp *ts = arg_temp(op->args[i]);
988         reset_ts(ctx, ts);
989         /*
990          * Save the corresponding known-zero/sign bits mask for the
991          * first output argument (only one supported so far).
992          */
993         if (i == 0) {
994             ts_info(ts)->z_mask = ctx->z_mask;
995             ts_info(ts)->s_mask = ctx->s_mask;
996         }
997     }
998 }
999 
1000 /*
1001  * The fold_* functions return true when processing is complete,
1002  * usually by folding the operation to a constant or to a copy,
1003  * and calling tcg_opt_gen_{mov,movi}.  They may do other things,
1004  * like collect information about the value produced, for use in
1005  * optimizing a subsequent operation.
1006  *
1007  * These first fold_* functions are all helpers, used by other
1008  * folders for more specific operations.
1009  */
1010 
1011 static bool fold_const1(OptContext *ctx, TCGOp *op)
1012 {
1013     if (arg_is_const(op->args[1])) {
1014         uint64_t t;
1015 
1016         t = arg_info(op->args[1])->val;
1017         t = do_constant_folding(op->opc, ctx->type, t, 0);
1018         return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1019     }
1020     return false;
1021 }
1022 
1023 static bool fold_const2(OptContext *ctx, TCGOp *op)
1024 {
1025     if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) {
1026         uint64_t t1 = arg_info(op->args[1])->val;
1027         uint64_t t2 = arg_info(op->args[2])->val;
1028 
1029         t1 = do_constant_folding(op->opc, ctx->type, t1, t2);
1030         return tcg_opt_gen_movi(ctx, op, op->args[0], t1);
1031     }
1032     return false;
1033 }
1034 
1035 static bool fold_commutative(OptContext *ctx, TCGOp *op)
1036 {
1037     swap_commutative(op->args[0], &op->args[1], &op->args[2]);
1038     return false;
1039 }
1040 
1041 static bool fold_const2_commutative(OptContext *ctx, TCGOp *op)
1042 {
1043     swap_commutative(op->args[0], &op->args[1], &op->args[2]);
1044     return fold_const2(ctx, op);
1045 }
1046 
1047 static bool fold_masks(OptContext *ctx, TCGOp *op)
1048 {
1049     uint64_t a_mask = ctx->a_mask;
1050     uint64_t z_mask = ctx->z_mask;
1051     uint64_t s_mask = ctx->s_mask;
1052 
1053     /*
1054      * 32-bit ops generate 32-bit results, which for the purpose of
1055      * simplifying tcg are sign-extended.  Certainly that's how we
1056      * represent our constants elsewhere.  Note that the bits will
1057      * be reset properly for a 64-bit value when encountering the
1058      * type changing opcodes.
1059      */
1060     if (ctx->type == TCG_TYPE_I32) {
1061         a_mask = (int32_t)a_mask;
1062         z_mask = (int32_t)z_mask;
1063         s_mask |= MAKE_64BIT_MASK(32, 32);
1064         ctx->z_mask = z_mask;
1065         ctx->s_mask = s_mask;
1066     }
1067 
1068     if (z_mask == 0) {
1069         return tcg_opt_gen_movi(ctx, op, op->args[0], 0);
1070     }
1071     if (a_mask == 0) {
1072         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
1073     }
1074     return false;
1075 }
1076 
1077 /*
1078  * Convert @op to NOT, if NOT is supported by the host.
1079  * Return true f the conversion is successful, which will still
1080  * indicate that the processing is complete.
1081  */
1082 static bool fold_not(OptContext *ctx, TCGOp *op);
1083 static bool fold_to_not(OptContext *ctx, TCGOp *op, int idx)
1084 {
1085     TCGOpcode not_op;
1086     bool have_not;
1087 
1088     switch (ctx->type) {
1089     case TCG_TYPE_I32:
1090         not_op = INDEX_op_not_i32;
1091         have_not = TCG_TARGET_HAS_not_i32;
1092         break;
1093     case TCG_TYPE_I64:
1094         not_op = INDEX_op_not_i64;
1095         have_not = TCG_TARGET_HAS_not_i64;
1096         break;
1097     case TCG_TYPE_V64:
1098     case TCG_TYPE_V128:
1099     case TCG_TYPE_V256:
1100         not_op = INDEX_op_not_vec;
1101         have_not = TCG_TARGET_HAS_not_vec;
1102         break;
1103     default:
1104         g_assert_not_reached();
1105     }
1106     if (have_not) {
1107         op->opc = not_op;
1108         op->args[1] = op->args[idx];
1109         return fold_not(ctx, op);
1110     }
1111     return false;
1112 }
1113 
1114 /* If the binary operation has first argument @i, fold to @i. */
1115 static bool fold_ix_to_i(OptContext *ctx, TCGOp *op, uint64_t i)
1116 {
1117     if (arg_is_const_val(op->args[1], i)) {
1118         return tcg_opt_gen_movi(ctx, op, op->args[0], i);
1119     }
1120     return false;
1121 }
1122 
1123 /* If the binary operation has first argument @i, fold to NOT. */
1124 static bool fold_ix_to_not(OptContext *ctx, TCGOp *op, uint64_t i)
1125 {
1126     if (arg_is_const_val(op->args[1], i)) {
1127         return fold_to_not(ctx, op, 2);
1128     }
1129     return false;
1130 }
1131 
1132 /* If the binary operation has second argument @i, fold to @i. */
1133 static bool fold_xi_to_i(OptContext *ctx, TCGOp *op, uint64_t i)
1134 {
1135     if (arg_is_const_val(op->args[2], i)) {
1136         return tcg_opt_gen_movi(ctx, op, op->args[0], i);
1137     }
1138     return false;
1139 }
1140 
1141 /* If the binary operation has second argument @i, fold to identity. */
1142 static bool fold_xi_to_x(OptContext *ctx, TCGOp *op, uint64_t i)
1143 {
1144     if (arg_is_const_val(op->args[2], i)) {
1145         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
1146     }
1147     return false;
1148 }
1149 
1150 /* If the binary operation has second argument @i, fold to NOT. */
1151 static bool fold_xi_to_not(OptContext *ctx, TCGOp *op, uint64_t i)
1152 {
1153     if (arg_is_const_val(op->args[2], i)) {
1154         return fold_to_not(ctx, op, 1);
1155     }
1156     return false;
1157 }
1158 
1159 /* If the binary operation has both arguments equal, fold to @i. */
1160 static bool fold_xx_to_i(OptContext *ctx, TCGOp *op, uint64_t i)
1161 {
1162     if (args_are_copies(op->args[1], op->args[2])) {
1163         return tcg_opt_gen_movi(ctx, op, op->args[0], i);
1164     }
1165     return false;
1166 }
1167 
1168 /* If the binary operation has both arguments equal, fold to identity. */
1169 static bool fold_xx_to_x(OptContext *ctx, TCGOp *op)
1170 {
1171     if (args_are_copies(op->args[1], op->args[2])) {
1172         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
1173     }
1174     return false;
1175 }
1176 
1177 /*
1178  * These outermost fold_<op> functions are sorted alphabetically.
1179  *
1180  * The ordering of the transformations should be:
1181  *   1) those that produce a constant
1182  *   2) those that produce a copy
1183  *   3) those that produce information about the result value.
1184  */
1185 
1186 static bool fold_add(OptContext *ctx, TCGOp *op)
1187 {
1188     if (fold_const2_commutative(ctx, op) ||
1189         fold_xi_to_x(ctx, op, 0)) {
1190         return true;
1191     }
1192     return false;
1193 }
1194 
1195 /* We cannot as yet do_constant_folding with vectors. */
1196 static bool fold_add_vec(OptContext *ctx, TCGOp *op)
1197 {
1198     if (fold_commutative(ctx, op) ||
1199         fold_xi_to_x(ctx, op, 0)) {
1200         return true;
1201     }
1202     return false;
1203 }
1204 
1205 static bool fold_addsub2(OptContext *ctx, TCGOp *op, bool add)
1206 {
1207     bool a_const = arg_is_const(op->args[2]) && arg_is_const(op->args[3]);
1208     bool b_const = arg_is_const(op->args[4]) && arg_is_const(op->args[5]);
1209 
1210     if (a_const && b_const) {
1211         uint64_t al = arg_info(op->args[2])->val;
1212         uint64_t ah = arg_info(op->args[3])->val;
1213         uint64_t bl = arg_info(op->args[4])->val;
1214         uint64_t bh = arg_info(op->args[5])->val;
1215         TCGArg rl, rh;
1216         TCGOp *op2;
1217 
1218         if (ctx->type == TCG_TYPE_I32) {
1219             uint64_t a = deposit64(al, 32, 32, ah);
1220             uint64_t b = deposit64(bl, 32, 32, bh);
1221 
1222             if (add) {
1223                 a += b;
1224             } else {
1225                 a -= b;
1226             }
1227 
1228             al = sextract64(a, 0, 32);
1229             ah = sextract64(a, 32, 32);
1230         } else {
1231             Int128 a = int128_make128(al, ah);
1232             Int128 b = int128_make128(bl, bh);
1233 
1234             if (add) {
1235                 a = int128_add(a, b);
1236             } else {
1237                 a = int128_sub(a, b);
1238             }
1239 
1240             al = int128_getlo(a);
1241             ah = int128_gethi(a);
1242         }
1243 
1244         rl = op->args[0];
1245         rh = op->args[1];
1246 
1247         /* The proper opcode is supplied by tcg_opt_gen_mov. */
1248         op2 = tcg_op_insert_before(ctx->tcg, op, 0, 2);
1249 
1250         tcg_opt_gen_movi(ctx, op, rl, al);
1251         tcg_opt_gen_movi(ctx, op2, rh, ah);
1252         return true;
1253     }
1254 
1255     /* Fold sub2 r,x,i to add2 r,x,-i */
1256     if (!add && b_const) {
1257         uint64_t bl = arg_info(op->args[4])->val;
1258         uint64_t bh = arg_info(op->args[5])->val;
1259 
1260         /* Negate the two parts without assembling and disassembling. */
1261         bl = -bl;
1262         bh = ~bh + !bl;
1263 
1264         op->opc = (ctx->type == TCG_TYPE_I32
1265                    ? INDEX_op_add2_i32 : INDEX_op_add2_i64);
1266         op->args[4] = arg_new_constant(ctx, bl);
1267         op->args[5] = arg_new_constant(ctx, bh);
1268     }
1269     return false;
1270 }
1271 
1272 static bool fold_add2(OptContext *ctx, TCGOp *op)
1273 {
1274     /* Note that the high and low parts may be independently swapped. */
1275     swap_commutative(op->args[0], &op->args[2], &op->args[4]);
1276     swap_commutative(op->args[1], &op->args[3], &op->args[5]);
1277 
1278     return fold_addsub2(ctx, op, true);
1279 }
1280 
1281 static bool fold_and(OptContext *ctx, TCGOp *op)
1282 {
1283     uint64_t z1, z2;
1284 
1285     if (fold_const2_commutative(ctx, op) ||
1286         fold_xi_to_i(ctx, op, 0) ||
1287         fold_xi_to_x(ctx, op, -1) ||
1288         fold_xx_to_x(ctx, op)) {
1289         return true;
1290     }
1291 
1292     z1 = arg_info(op->args[1])->z_mask;
1293     z2 = arg_info(op->args[2])->z_mask;
1294     ctx->z_mask = z1 & z2;
1295 
1296     /*
1297      * Sign repetitions are perforce all identical, whether they are 1 or 0.
1298      * Bitwise operations preserve the relative quantity of the repetitions.
1299      */
1300     ctx->s_mask = arg_info(op->args[1])->s_mask
1301                 & arg_info(op->args[2])->s_mask;
1302 
1303     /*
1304      * Known-zeros does not imply known-ones.  Therefore unless
1305      * arg2 is constant, we can't infer affected bits from it.
1306      */
1307     if (arg_is_const(op->args[2])) {
1308         ctx->a_mask = z1 & ~z2;
1309     }
1310 
1311     return fold_masks(ctx, op);
1312 }
1313 
1314 static bool fold_andc(OptContext *ctx, TCGOp *op)
1315 {
1316     uint64_t z1;
1317 
1318     if (fold_const2(ctx, op) ||
1319         fold_xx_to_i(ctx, op, 0) ||
1320         fold_xi_to_x(ctx, op, 0) ||
1321         fold_ix_to_not(ctx, op, -1)) {
1322         return true;
1323     }
1324 
1325     z1 = arg_info(op->args[1])->z_mask;
1326 
1327     /*
1328      * Known-zeros does not imply known-ones.  Therefore unless
1329      * arg2 is constant, we can't infer anything from it.
1330      */
1331     if (arg_is_const(op->args[2])) {
1332         uint64_t z2 = ~arg_info(op->args[2])->z_mask;
1333         ctx->a_mask = z1 & ~z2;
1334         z1 &= z2;
1335     }
1336     ctx->z_mask = z1;
1337 
1338     ctx->s_mask = arg_info(op->args[1])->s_mask
1339                 & arg_info(op->args[2])->s_mask;
1340     return fold_masks(ctx, op);
1341 }
1342 
1343 static bool fold_brcond(OptContext *ctx, TCGOp *op)
1344 {
1345     int i = do_constant_folding_cond1(ctx, op, NO_DEST, &op->args[0],
1346                                       &op->args[1], &op->args[2]);
1347     if (i == 0) {
1348         tcg_op_remove(ctx->tcg, op);
1349         return true;
1350     }
1351     if (i > 0) {
1352         op->opc = INDEX_op_br;
1353         op->args[0] = op->args[3];
1354     }
1355     return false;
1356 }
1357 
1358 static bool fold_brcond2(OptContext *ctx, TCGOp *op)
1359 {
1360     TCGCond cond;
1361     TCGArg label;
1362     int i, inv = 0;
1363 
1364     i = do_constant_folding_cond2(ctx, op, &op->args[0]);
1365     cond = op->args[4];
1366     label = op->args[5];
1367     if (i >= 0) {
1368         goto do_brcond_const;
1369     }
1370 
1371     switch (cond) {
1372     case TCG_COND_LT:
1373     case TCG_COND_GE:
1374         /*
1375          * Simplify LT/GE comparisons vs zero to a single compare
1376          * vs the high word of the input.
1377          */
1378         if (arg_is_const_val(op->args[2], 0) &&
1379             arg_is_const_val(op->args[3], 0)) {
1380             goto do_brcond_high;
1381         }
1382         break;
1383 
1384     case TCG_COND_NE:
1385         inv = 1;
1386         QEMU_FALLTHROUGH;
1387     case TCG_COND_EQ:
1388         /*
1389          * Simplify EQ/NE comparisons where one of the pairs
1390          * can be simplified.
1391          */
1392         i = do_constant_folding_cond(TCG_TYPE_I32, op->args[0],
1393                                      op->args[2], cond);
1394         switch (i ^ inv) {
1395         case 0:
1396             goto do_brcond_const;
1397         case 1:
1398             goto do_brcond_high;
1399         }
1400 
1401         i = do_constant_folding_cond(TCG_TYPE_I32, op->args[1],
1402                                      op->args[3], cond);
1403         switch (i ^ inv) {
1404         case 0:
1405             goto do_brcond_const;
1406         case 1:
1407             goto do_brcond_low;
1408         }
1409         break;
1410 
1411     case TCG_COND_TSTEQ:
1412     case TCG_COND_TSTNE:
1413         if (arg_is_const_val(op->args[2], 0)) {
1414             goto do_brcond_high;
1415         }
1416         if (arg_is_const_val(op->args[3], 0)) {
1417             goto do_brcond_low;
1418         }
1419         break;
1420 
1421     default:
1422         break;
1423 
1424     do_brcond_low:
1425         op->opc = INDEX_op_brcond_i32;
1426         op->args[1] = op->args[2];
1427         op->args[2] = cond;
1428         op->args[3] = label;
1429         return fold_brcond(ctx, op);
1430 
1431     do_brcond_high:
1432         op->opc = INDEX_op_brcond_i32;
1433         op->args[0] = op->args[1];
1434         op->args[1] = op->args[3];
1435         op->args[2] = cond;
1436         op->args[3] = label;
1437         return fold_brcond(ctx, op);
1438 
1439     do_brcond_const:
1440         if (i == 0) {
1441             tcg_op_remove(ctx->tcg, op);
1442             return true;
1443         }
1444         op->opc = INDEX_op_br;
1445         op->args[0] = label;
1446         break;
1447     }
1448     return false;
1449 }
1450 
1451 static bool fold_bswap(OptContext *ctx, TCGOp *op)
1452 {
1453     uint64_t z_mask, s_mask, sign;
1454 
1455     if (arg_is_const(op->args[1])) {
1456         uint64_t t = arg_info(op->args[1])->val;
1457 
1458         t = do_constant_folding(op->opc, ctx->type, t, op->args[2]);
1459         return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1460     }
1461 
1462     z_mask = arg_info(op->args[1])->z_mask;
1463 
1464     switch (op->opc) {
1465     case INDEX_op_bswap16_i32:
1466     case INDEX_op_bswap16_i64:
1467         z_mask = bswap16(z_mask);
1468         sign = INT16_MIN;
1469         break;
1470     case INDEX_op_bswap32_i32:
1471     case INDEX_op_bswap32_i64:
1472         z_mask = bswap32(z_mask);
1473         sign = INT32_MIN;
1474         break;
1475     case INDEX_op_bswap64_i64:
1476         z_mask = bswap64(z_mask);
1477         sign = INT64_MIN;
1478         break;
1479     default:
1480         g_assert_not_reached();
1481     }
1482     s_mask = smask_from_zmask(z_mask);
1483 
1484     switch (op->args[2] & (TCG_BSWAP_OZ | TCG_BSWAP_OS)) {
1485     case TCG_BSWAP_OZ:
1486         break;
1487     case TCG_BSWAP_OS:
1488         /* If the sign bit may be 1, force all the bits above to 1. */
1489         if (z_mask & sign) {
1490             z_mask |= sign;
1491             s_mask = sign << 1;
1492         }
1493         break;
1494     default:
1495         /* The high bits are undefined: force all bits above the sign to 1. */
1496         z_mask |= sign << 1;
1497         s_mask = 0;
1498         break;
1499     }
1500     ctx->z_mask = z_mask;
1501     ctx->s_mask = s_mask;
1502 
1503     return fold_masks(ctx, op);
1504 }
1505 
1506 static bool fold_call(OptContext *ctx, TCGOp *op)
1507 {
1508     TCGContext *s = ctx->tcg;
1509     int nb_oargs = TCGOP_CALLO(op);
1510     int nb_iargs = TCGOP_CALLI(op);
1511     int flags, i;
1512 
1513     init_arguments(ctx, op, nb_oargs + nb_iargs);
1514     copy_propagate(ctx, op, nb_oargs, nb_iargs);
1515 
1516     /* If the function reads or writes globals, reset temp data. */
1517     flags = tcg_call_flags(op);
1518     if (!(flags & (TCG_CALL_NO_READ_GLOBALS | TCG_CALL_NO_WRITE_GLOBALS))) {
1519         int nb_globals = s->nb_globals;
1520 
1521         for (i = 0; i < nb_globals; i++) {
1522             if (test_bit(i, ctx->temps_used.l)) {
1523                 reset_ts(ctx, &ctx->tcg->temps[i]);
1524             }
1525         }
1526     }
1527 
1528     /* If the function has side effects, reset mem data. */
1529     if (!(flags & TCG_CALL_NO_SIDE_EFFECTS)) {
1530         remove_mem_copy_all(ctx);
1531     }
1532 
1533     /* Reset temp data for outputs. */
1534     for (i = 0; i < nb_oargs; i++) {
1535         reset_temp(ctx, op->args[i]);
1536     }
1537 
1538     /* Stop optimizing MB across calls. */
1539     ctx->prev_mb = NULL;
1540     return true;
1541 }
1542 
1543 static bool fold_count_zeros(OptContext *ctx, TCGOp *op)
1544 {
1545     uint64_t z_mask;
1546 
1547     if (arg_is_const(op->args[1])) {
1548         uint64_t t = arg_info(op->args[1])->val;
1549 
1550         if (t != 0) {
1551             t = do_constant_folding(op->opc, ctx->type, t, 0);
1552             return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1553         }
1554         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[2]);
1555     }
1556 
1557     switch (ctx->type) {
1558     case TCG_TYPE_I32:
1559         z_mask = 31;
1560         break;
1561     case TCG_TYPE_I64:
1562         z_mask = 63;
1563         break;
1564     default:
1565         g_assert_not_reached();
1566     }
1567     ctx->z_mask = arg_info(op->args[2])->z_mask | z_mask;
1568     ctx->s_mask = smask_from_zmask(ctx->z_mask);
1569     return false;
1570 }
1571 
1572 static bool fold_ctpop(OptContext *ctx, TCGOp *op)
1573 {
1574     if (fold_const1(ctx, op)) {
1575         return true;
1576     }
1577 
1578     switch (ctx->type) {
1579     case TCG_TYPE_I32:
1580         ctx->z_mask = 32 | 31;
1581         break;
1582     case TCG_TYPE_I64:
1583         ctx->z_mask = 64 | 63;
1584         break;
1585     default:
1586         g_assert_not_reached();
1587     }
1588     ctx->s_mask = smask_from_zmask(ctx->z_mask);
1589     return false;
1590 }
1591 
1592 static bool fold_deposit(OptContext *ctx, TCGOp *op)
1593 {
1594     TCGOpcode and_opc;
1595 
1596     if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) {
1597         uint64_t t1 = arg_info(op->args[1])->val;
1598         uint64_t t2 = arg_info(op->args[2])->val;
1599 
1600         t1 = deposit64(t1, op->args[3], op->args[4], t2);
1601         return tcg_opt_gen_movi(ctx, op, op->args[0], t1);
1602     }
1603 
1604     switch (ctx->type) {
1605     case TCG_TYPE_I32:
1606         and_opc = INDEX_op_and_i32;
1607         break;
1608     case TCG_TYPE_I64:
1609         and_opc = INDEX_op_and_i64;
1610         break;
1611     default:
1612         g_assert_not_reached();
1613     }
1614 
1615     /* Inserting a value into zero at offset 0. */
1616     if (arg_is_const_val(op->args[1], 0) && op->args[3] == 0) {
1617         uint64_t mask = MAKE_64BIT_MASK(0, op->args[4]);
1618 
1619         op->opc = and_opc;
1620         op->args[1] = op->args[2];
1621         op->args[2] = arg_new_constant(ctx, mask);
1622         ctx->z_mask = mask & arg_info(op->args[1])->z_mask;
1623         return false;
1624     }
1625 
1626     /* Inserting zero into a value. */
1627     if (arg_is_const_val(op->args[2], 0)) {
1628         uint64_t mask = deposit64(-1, op->args[3], op->args[4], 0);
1629 
1630         op->opc = and_opc;
1631         op->args[2] = arg_new_constant(ctx, mask);
1632         ctx->z_mask = mask & arg_info(op->args[1])->z_mask;
1633         return false;
1634     }
1635 
1636     ctx->z_mask = deposit64(arg_info(op->args[1])->z_mask,
1637                             op->args[3], op->args[4],
1638                             arg_info(op->args[2])->z_mask);
1639     return false;
1640 }
1641 
1642 static bool fold_divide(OptContext *ctx, TCGOp *op)
1643 {
1644     if (fold_const2(ctx, op) ||
1645         fold_xi_to_x(ctx, op, 1)) {
1646         return true;
1647     }
1648     return false;
1649 }
1650 
1651 static bool fold_dup(OptContext *ctx, TCGOp *op)
1652 {
1653     if (arg_is_const(op->args[1])) {
1654         uint64_t t = arg_info(op->args[1])->val;
1655         t = dup_const(TCGOP_VECE(op), t);
1656         return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1657     }
1658     return false;
1659 }
1660 
1661 static bool fold_dup2(OptContext *ctx, TCGOp *op)
1662 {
1663     if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) {
1664         uint64_t t = deposit64(arg_info(op->args[1])->val, 32, 32,
1665                                arg_info(op->args[2])->val);
1666         return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1667     }
1668 
1669     if (args_are_copies(op->args[1], op->args[2])) {
1670         op->opc = INDEX_op_dup_vec;
1671         TCGOP_VECE(op) = MO_32;
1672     }
1673     return false;
1674 }
1675 
1676 static bool fold_eqv(OptContext *ctx, TCGOp *op)
1677 {
1678     if (fold_const2_commutative(ctx, op) ||
1679         fold_xi_to_x(ctx, op, -1) ||
1680         fold_xi_to_not(ctx, op, 0)) {
1681         return true;
1682     }
1683 
1684     ctx->s_mask = arg_info(op->args[1])->s_mask
1685                 & arg_info(op->args[2])->s_mask;
1686     return false;
1687 }
1688 
1689 static bool fold_extract(OptContext *ctx, TCGOp *op)
1690 {
1691     uint64_t z_mask_old, z_mask;
1692     int pos = op->args[2];
1693     int len = op->args[3];
1694 
1695     if (arg_is_const(op->args[1])) {
1696         uint64_t t;
1697 
1698         t = arg_info(op->args[1])->val;
1699         t = extract64(t, pos, len);
1700         return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1701     }
1702 
1703     z_mask_old = arg_info(op->args[1])->z_mask;
1704     z_mask = extract64(z_mask_old, pos, len);
1705     if (pos == 0) {
1706         ctx->a_mask = z_mask_old ^ z_mask;
1707     }
1708     ctx->z_mask = z_mask;
1709     ctx->s_mask = smask_from_zmask(z_mask);
1710 
1711     return fold_masks(ctx, op);
1712 }
1713 
1714 static bool fold_extract2(OptContext *ctx, TCGOp *op)
1715 {
1716     if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) {
1717         uint64_t v1 = arg_info(op->args[1])->val;
1718         uint64_t v2 = arg_info(op->args[2])->val;
1719         int shr = op->args[3];
1720 
1721         if (op->opc == INDEX_op_extract2_i64) {
1722             v1 >>= shr;
1723             v2 <<= 64 - shr;
1724         } else {
1725             v1 = (uint32_t)v1 >> shr;
1726             v2 = (uint64_t)((int32_t)v2 << (32 - shr));
1727         }
1728         return tcg_opt_gen_movi(ctx, op, op->args[0], v1 | v2);
1729     }
1730     return false;
1731 }
1732 
1733 static bool fold_exts(OptContext *ctx, TCGOp *op)
1734 {
1735     uint64_t s_mask_old, s_mask, z_mask, sign;
1736     bool type_change = false;
1737 
1738     if (fold_const1(ctx, op)) {
1739         return true;
1740     }
1741 
1742     z_mask = arg_info(op->args[1])->z_mask;
1743     s_mask = arg_info(op->args[1])->s_mask;
1744     s_mask_old = s_mask;
1745 
1746     switch (op->opc) {
1747     CASE_OP_32_64(ext8s):
1748         sign = INT8_MIN;
1749         z_mask = (uint8_t)z_mask;
1750         break;
1751     CASE_OP_32_64(ext16s):
1752         sign = INT16_MIN;
1753         z_mask = (uint16_t)z_mask;
1754         break;
1755     case INDEX_op_ext_i32_i64:
1756         type_change = true;
1757         QEMU_FALLTHROUGH;
1758     case INDEX_op_ext32s_i64:
1759         sign = INT32_MIN;
1760         z_mask = (uint32_t)z_mask;
1761         break;
1762     default:
1763         g_assert_not_reached();
1764     }
1765 
1766     if (z_mask & sign) {
1767         z_mask |= sign;
1768     }
1769     s_mask |= sign << 1;
1770 
1771     ctx->z_mask = z_mask;
1772     ctx->s_mask = s_mask;
1773     if (!type_change) {
1774         ctx->a_mask = s_mask & ~s_mask_old;
1775     }
1776 
1777     return fold_masks(ctx, op);
1778 }
1779 
1780 static bool fold_extu(OptContext *ctx, TCGOp *op)
1781 {
1782     uint64_t z_mask_old, z_mask;
1783     bool type_change = false;
1784 
1785     if (fold_const1(ctx, op)) {
1786         return true;
1787     }
1788 
1789     z_mask_old = z_mask = arg_info(op->args[1])->z_mask;
1790 
1791     switch (op->opc) {
1792     CASE_OP_32_64(ext8u):
1793         z_mask = (uint8_t)z_mask;
1794         break;
1795     CASE_OP_32_64(ext16u):
1796         z_mask = (uint16_t)z_mask;
1797         break;
1798     case INDEX_op_extrl_i64_i32:
1799     case INDEX_op_extu_i32_i64:
1800         type_change = true;
1801         QEMU_FALLTHROUGH;
1802     case INDEX_op_ext32u_i64:
1803         z_mask = (uint32_t)z_mask;
1804         break;
1805     case INDEX_op_extrh_i64_i32:
1806         type_change = true;
1807         z_mask >>= 32;
1808         break;
1809     default:
1810         g_assert_not_reached();
1811     }
1812 
1813     ctx->z_mask = z_mask;
1814     ctx->s_mask = smask_from_zmask(z_mask);
1815     if (!type_change) {
1816         ctx->a_mask = z_mask_old ^ z_mask;
1817     }
1818     return fold_masks(ctx, op);
1819 }
1820 
1821 static bool fold_mb(OptContext *ctx, TCGOp *op)
1822 {
1823     /* Eliminate duplicate and redundant fence instructions.  */
1824     if (ctx->prev_mb) {
1825         /*
1826          * Merge two barriers of the same type into one,
1827          * or a weaker barrier into a stronger one,
1828          * or two weaker barriers into a stronger one.
1829          *   mb X; mb Y => mb X|Y
1830          *   mb; strl => mb; st
1831          *   ldaq; mb => ld; mb
1832          *   ldaq; strl => ld; mb; st
1833          * Other combinations are also merged into a strong
1834          * barrier.  This is stricter than specified but for
1835          * the purposes of TCG is better than not optimizing.
1836          */
1837         ctx->prev_mb->args[0] |= op->args[0];
1838         tcg_op_remove(ctx->tcg, op);
1839     } else {
1840         ctx->prev_mb = op;
1841     }
1842     return true;
1843 }
1844 
1845 static bool fold_mov(OptContext *ctx, TCGOp *op)
1846 {
1847     return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
1848 }
1849 
1850 static bool fold_movcond(OptContext *ctx, TCGOp *op)
1851 {
1852     int i;
1853 
1854     /*
1855      * Canonicalize the "false" input reg to match the destination reg so
1856      * that the tcg backend can implement a "move if true" operation.
1857      */
1858     if (swap_commutative(op->args[0], &op->args[4], &op->args[3])) {
1859         op->args[5] = tcg_invert_cond(op->args[5]);
1860     }
1861 
1862     i = do_constant_folding_cond1(ctx, op, NO_DEST, &op->args[1],
1863                                   &op->args[2], &op->args[5]);
1864     if (i >= 0) {
1865         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[4 - i]);
1866     }
1867 
1868     ctx->z_mask = arg_info(op->args[3])->z_mask
1869                 | arg_info(op->args[4])->z_mask;
1870     ctx->s_mask = arg_info(op->args[3])->s_mask
1871                 & arg_info(op->args[4])->s_mask;
1872 
1873     if (arg_is_const(op->args[3]) && arg_is_const(op->args[4])) {
1874         uint64_t tv = arg_info(op->args[3])->val;
1875         uint64_t fv = arg_info(op->args[4])->val;
1876         TCGOpcode opc, negopc = 0;
1877         TCGCond cond = op->args[5];
1878 
1879         switch (ctx->type) {
1880         case TCG_TYPE_I32:
1881             opc = INDEX_op_setcond_i32;
1882             if (TCG_TARGET_HAS_negsetcond_i32) {
1883                 negopc = INDEX_op_negsetcond_i32;
1884             }
1885             tv = (int32_t)tv;
1886             fv = (int32_t)fv;
1887             break;
1888         case TCG_TYPE_I64:
1889             opc = INDEX_op_setcond_i64;
1890             if (TCG_TARGET_HAS_negsetcond_i64) {
1891                 negopc = INDEX_op_negsetcond_i64;
1892             }
1893             break;
1894         default:
1895             g_assert_not_reached();
1896         }
1897 
1898         if (tv == 1 && fv == 0) {
1899             op->opc = opc;
1900             op->args[3] = cond;
1901         } else if (fv == 1 && tv == 0) {
1902             op->opc = opc;
1903             op->args[3] = tcg_invert_cond(cond);
1904         } else if (negopc) {
1905             if (tv == -1 && fv == 0) {
1906                 op->opc = negopc;
1907                 op->args[3] = cond;
1908             } else if (fv == -1 && tv == 0) {
1909                 op->opc = negopc;
1910                 op->args[3] = tcg_invert_cond(cond);
1911             }
1912         }
1913     }
1914     return false;
1915 }
1916 
1917 static bool fold_mul(OptContext *ctx, TCGOp *op)
1918 {
1919     if (fold_const2(ctx, op) ||
1920         fold_xi_to_i(ctx, op, 0) ||
1921         fold_xi_to_x(ctx, op, 1)) {
1922         return true;
1923     }
1924     return false;
1925 }
1926 
1927 static bool fold_mul_highpart(OptContext *ctx, TCGOp *op)
1928 {
1929     if (fold_const2_commutative(ctx, op) ||
1930         fold_xi_to_i(ctx, op, 0)) {
1931         return true;
1932     }
1933     return false;
1934 }
1935 
1936 static bool fold_multiply2(OptContext *ctx, TCGOp *op)
1937 {
1938     swap_commutative(op->args[0], &op->args[2], &op->args[3]);
1939 
1940     if (arg_is_const(op->args[2]) && arg_is_const(op->args[3])) {
1941         uint64_t a = arg_info(op->args[2])->val;
1942         uint64_t b = arg_info(op->args[3])->val;
1943         uint64_t h, l;
1944         TCGArg rl, rh;
1945         TCGOp *op2;
1946 
1947         switch (op->opc) {
1948         case INDEX_op_mulu2_i32:
1949             l = (uint64_t)(uint32_t)a * (uint32_t)b;
1950             h = (int32_t)(l >> 32);
1951             l = (int32_t)l;
1952             break;
1953         case INDEX_op_muls2_i32:
1954             l = (int64_t)(int32_t)a * (int32_t)b;
1955             h = l >> 32;
1956             l = (int32_t)l;
1957             break;
1958         case INDEX_op_mulu2_i64:
1959             mulu64(&l, &h, a, b);
1960             break;
1961         case INDEX_op_muls2_i64:
1962             muls64(&l, &h, a, b);
1963             break;
1964         default:
1965             g_assert_not_reached();
1966         }
1967 
1968         rl = op->args[0];
1969         rh = op->args[1];
1970 
1971         /* The proper opcode is supplied by tcg_opt_gen_mov. */
1972         op2 = tcg_op_insert_before(ctx->tcg, op, 0, 2);
1973 
1974         tcg_opt_gen_movi(ctx, op, rl, l);
1975         tcg_opt_gen_movi(ctx, op2, rh, h);
1976         return true;
1977     }
1978     return false;
1979 }
1980 
1981 static bool fold_nand(OptContext *ctx, TCGOp *op)
1982 {
1983     if (fold_const2_commutative(ctx, op) ||
1984         fold_xi_to_not(ctx, op, -1)) {
1985         return true;
1986     }
1987 
1988     ctx->s_mask = arg_info(op->args[1])->s_mask
1989                 & arg_info(op->args[2])->s_mask;
1990     return false;
1991 }
1992 
1993 static bool fold_neg_no_const(OptContext *ctx, TCGOp *op)
1994 {
1995     /* Set to 1 all bits to the left of the rightmost.  */
1996     uint64_t z_mask = arg_info(op->args[1])->z_mask;
1997     ctx->z_mask = -(z_mask & -z_mask);
1998 
1999     /*
2000      * Because of fold_sub_to_neg, we want to always return true,
2001      * via finish_folding.
2002      */
2003     finish_folding(ctx, op);
2004     return true;
2005 }
2006 
2007 static bool fold_neg(OptContext *ctx, TCGOp *op)
2008 {
2009     return fold_const1(ctx, op) || fold_neg_no_const(ctx, op);
2010 }
2011 
2012 static bool fold_nor(OptContext *ctx, TCGOp *op)
2013 {
2014     if (fold_const2_commutative(ctx, op) ||
2015         fold_xi_to_not(ctx, op, 0)) {
2016         return true;
2017     }
2018 
2019     ctx->s_mask = arg_info(op->args[1])->s_mask
2020                 & arg_info(op->args[2])->s_mask;
2021     return false;
2022 }
2023 
2024 static bool fold_not(OptContext *ctx, TCGOp *op)
2025 {
2026     if (fold_const1(ctx, op)) {
2027         return true;
2028     }
2029 
2030     ctx->s_mask = arg_info(op->args[1])->s_mask;
2031 
2032     /* Because of fold_to_not, we want to always return true, via finish. */
2033     finish_folding(ctx, op);
2034     return true;
2035 }
2036 
2037 static bool fold_or(OptContext *ctx, TCGOp *op)
2038 {
2039     if (fold_const2_commutative(ctx, op) ||
2040         fold_xi_to_x(ctx, op, 0) ||
2041         fold_xx_to_x(ctx, op)) {
2042         return true;
2043     }
2044 
2045     ctx->z_mask = arg_info(op->args[1])->z_mask
2046                 | arg_info(op->args[2])->z_mask;
2047     ctx->s_mask = arg_info(op->args[1])->s_mask
2048                 & arg_info(op->args[2])->s_mask;
2049     return fold_masks(ctx, op);
2050 }
2051 
2052 static bool fold_orc(OptContext *ctx, TCGOp *op)
2053 {
2054     if (fold_const2(ctx, op) ||
2055         fold_xx_to_i(ctx, op, -1) ||
2056         fold_xi_to_x(ctx, op, -1) ||
2057         fold_ix_to_not(ctx, op, 0)) {
2058         return true;
2059     }
2060 
2061     ctx->s_mask = arg_info(op->args[1])->s_mask
2062                 & arg_info(op->args[2])->s_mask;
2063     return false;
2064 }
2065 
2066 static bool fold_qemu_ld(OptContext *ctx, TCGOp *op)
2067 {
2068     const TCGOpDef *def = &tcg_op_defs[op->opc];
2069     MemOpIdx oi = op->args[def->nb_oargs + def->nb_iargs];
2070     MemOp mop = get_memop(oi);
2071     int width = 8 * memop_size(mop);
2072 
2073     if (width < 64) {
2074         ctx->s_mask = MAKE_64BIT_MASK(width, 64 - width);
2075         if (!(mop & MO_SIGN)) {
2076             ctx->z_mask = MAKE_64BIT_MASK(0, width);
2077             ctx->s_mask <<= 1;
2078         }
2079     }
2080 
2081     /* Opcodes that touch guest memory stop the mb optimization.  */
2082     ctx->prev_mb = NULL;
2083     return false;
2084 }
2085 
2086 static bool fold_qemu_st(OptContext *ctx, TCGOp *op)
2087 {
2088     /* Opcodes that touch guest memory stop the mb optimization.  */
2089     ctx->prev_mb = NULL;
2090     return false;
2091 }
2092 
2093 static bool fold_remainder(OptContext *ctx, TCGOp *op)
2094 {
2095     if (fold_const2(ctx, op) ||
2096         fold_xx_to_i(ctx, op, 0)) {
2097         return true;
2098     }
2099     return false;
2100 }
2101 
2102 static bool fold_setcond_zmask(OptContext *ctx, TCGOp *op, bool neg)
2103 {
2104     uint64_t a_zmask, b_val;
2105     TCGCond cond;
2106 
2107     if (!arg_is_const(op->args[2])) {
2108         return false;
2109     }
2110 
2111     a_zmask = arg_info(op->args[1])->z_mask;
2112     b_val = arg_info(op->args[2])->val;
2113     cond = op->args[3];
2114 
2115     if (ctx->type == TCG_TYPE_I32) {
2116         a_zmask = (uint32_t)a_zmask;
2117         b_val = (uint32_t)b_val;
2118     }
2119 
2120     /*
2121      * A with only low bits set vs B with high bits set means that A < B.
2122      */
2123     if (a_zmask < b_val) {
2124         bool inv = false;
2125 
2126         switch (cond) {
2127         case TCG_COND_NE:
2128         case TCG_COND_LEU:
2129         case TCG_COND_LTU:
2130             inv = true;
2131             /* fall through */
2132         case TCG_COND_GTU:
2133         case TCG_COND_GEU:
2134         case TCG_COND_EQ:
2135             return tcg_opt_gen_movi(ctx, op, op->args[0], neg ? -inv : inv);
2136         default:
2137             break;
2138         }
2139     }
2140 
2141     /*
2142      * A with only lsb set is already boolean.
2143      */
2144     if (a_zmask <= 1) {
2145         bool convert = false;
2146         bool inv = false;
2147 
2148         switch (cond) {
2149         case TCG_COND_EQ:
2150             inv = true;
2151             /* fall through */
2152         case TCG_COND_NE:
2153             convert = (b_val == 0);
2154             break;
2155         case TCG_COND_LTU:
2156         case TCG_COND_TSTEQ:
2157             inv = true;
2158             /* fall through */
2159         case TCG_COND_GEU:
2160         case TCG_COND_TSTNE:
2161             convert = (b_val == 1);
2162             break;
2163         default:
2164             break;
2165         }
2166         if (convert) {
2167             TCGOpcode add_opc, xor_opc, neg_opc;
2168 
2169             if (!inv && !neg) {
2170                 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
2171             }
2172 
2173             switch (ctx->type) {
2174             case TCG_TYPE_I32:
2175                 add_opc = INDEX_op_add_i32;
2176                 neg_opc = INDEX_op_neg_i32;
2177                 xor_opc = INDEX_op_xor_i32;
2178                 break;
2179             case TCG_TYPE_I64:
2180                 add_opc = INDEX_op_add_i64;
2181                 neg_opc = INDEX_op_neg_i64;
2182                 xor_opc = INDEX_op_xor_i64;
2183                 break;
2184             default:
2185                 g_assert_not_reached();
2186             }
2187 
2188             if (!inv) {
2189                 op->opc = neg_opc;
2190             } else if (neg) {
2191                 op->opc = add_opc;
2192                 op->args[2] = arg_new_constant(ctx, -1);
2193             } else {
2194                 op->opc = xor_opc;
2195                 op->args[2] = arg_new_constant(ctx, 1);
2196             }
2197             return false;
2198         }
2199     }
2200 
2201     return false;
2202 }
2203 
2204 static void fold_setcond_tst_pow2(OptContext *ctx, TCGOp *op, bool neg)
2205 {
2206     TCGOpcode and_opc, sub_opc, xor_opc, neg_opc, shr_opc;
2207     TCGOpcode uext_opc = 0, sext_opc = 0;
2208     TCGCond cond = op->args[3];
2209     TCGArg ret, src1, src2;
2210     TCGOp *op2;
2211     uint64_t val;
2212     int sh;
2213     bool inv;
2214 
2215     if (!is_tst_cond(cond) || !arg_is_const(op->args[2])) {
2216         return;
2217     }
2218 
2219     src2 = op->args[2];
2220     val = arg_info(src2)->val;
2221     if (!is_power_of_2(val)) {
2222         return;
2223     }
2224     sh = ctz64(val);
2225 
2226     switch (ctx->type) {
2227     case TCG_TYPE_I32:
2228         and_opc = INDEX_op_and_i32;
2229         sub_opc = INDEX_op_sub_i32;
2230         xor_opc = INDEX_op_xor_i32;
2231         shr_opc = INDEX_op_shr_i32;
2232         neg_opc = INDEX_op_neg_i32;
2233         if (TCG_TARGET_extract_i32_valid(sh, 1)) {
2234             uext_opc = TCG_TARGET_HAS_extract_i32 ? INDEX_op_extract_i32 : 0;
2235             sext_opc = TCG_TARGET_HAS_sextract_i32 ? INDEX_op_sextract_i32 : 0;
2236         }
2237         break;
2238     case TCG_TYPE_I64:
2239         and_opc = INDEX_op_and_i64;
2240         sub_opc = INDEX_op_sub_i64;
2241         xor_opc = INDEX_op_xor_i64;
2242         shr_opc = INDEX_op_shr_i64;
2243         neg_opc = INDEX_op_neg_i64;
2244         if (TCG_TARGET_extract_i64_valid(sh, 1)) {
2245             uext_opc = TCG_TARGET_HAS_extract_i64 ? INDEX_op_extract_i64 : 0;
2246             sext_opc = TCG_TARGET_HAS_sextract_i64 ? INDEX_op_sextract_i64 : 0;
2247         }
2248         break;
2249     default:
2250         g_assert_not_reached();
2251     }
2252 
2253     ret = op->args[0];
2254     src1 = op->args[1];
2255     inv = cond == TCG_COND_TSTEQ;
2256 
2257     if (sh && sext_opc && neg && !inv) {
2258         op->opc = sext_opc;
2259         op->args[1] = src1;
2260         op->args[2] = sh;
2261         op->args[3] = 1;
2262         return;
2263     } else if (sh && uext_opc) {
2264         op->opc = uext_opc;
2265         op->args[1] = src1;
2266         op->args[2] = sh;
2267         op->args[3] = 1;
2268     } else {
2269         if (sh) {
2270             op2 = tcg_op_insert_before(ctx->tcg, op, shr_opc, 3);
2271             op2->args[0] = ret;
2272             op2->args[1] = src1;
2273             op2->args[2] = arg_new_constant(ctx, sh);
2274             src1 = ret;
2275         }
2276         op->opc = and_opc;
2277         op->args[1] = src1;
2278         op->args[2] = arg_new_constant(ctx, 1);
2279     }
2280 
2281     if (neg && inv) {
2282         op2 = tcg_op_insert_after(ctx->tcg, op, sub_opc, 3);
2283         op2->args[0] = ret;
2284         op2->args[1] = ret;
2285         op2->args[2] = arg_new_constant(ctx, 1);
2286     } else if (inv) {
2287         op2 = tcg_op_insert_after(ctx->tcg, op, xor_opc, 3);
2288         op2->args[0] = ret;
2289         op2->args[1] = ret;
2290         op2->args[2] = arg_new_constant(ctx, 1);
2291     } else if (neg) {
2292         op2 = tcg_op_insert_after(ctx->tcg, op, neg_opc, 2);
2293         op2->args[0] = ret;
2294         op2->args[1] = ret;
2295     }
2296 }
2297 
2298 static bool fold_setcond(OptContext *ctx, TCGOp *op)
2299 {
2300     int i = do_constant_folding_cond1(ctx, op, op->args[0], &op->args[1],
2301                                       &op->args[2], &op->args[3]);
2302     if (i >= 0) {
2303         return tcg_opt_gen_movi(ctx, op, op->args[0], i);
2304     }
2305 
2306     if (fold_setcond_zmask(ctx, op, false)) {
2307         return true;
2308     }
2309     fold_setcond_tst_pow2(ctx, op, false);
2310 
2311     ctx->z_mask = 1;
2312     ctx->s_mask = smask_from_zmask(1);
2313     return false;
2314 }
2315 
2316 static bool fold_negsetcond(OptContext *ctx, TCGOp *op)
2317 {
2318     int i = do_constant_folding_cond1(ctx, op, op->args[0], &op->args[1],
2319                                       &op->args[2], &op->args[3]);
2320     if (i >= 0) {
2321         return tcg_opt_gen_movi(ctx, op, op->args[0], -i);
2322     }
2323 
2324     if (fold_setcond_zmask(ctx, op, true)) {
2325         return true;
2326     }
2327     fold_setcond_tst_pow2(ctx, op, true);
2328 
2329     /* Value is {0,-1} so all bits are repetitions of the sign. */
2330     ctx->s_mask = -1;
2331     return false;
2332 }
2333 
2334 static bool fold_setcond2(OptContext *ctx, TCGOp *op)
2335 {
2336     TCGCond cond;
2337     int i, inv = 0;
2338 
2339     i = do_constant_folding_cond2(ctx, op, &op->args[1]);
2340     cond = op->args[5];
2341     if (i >= 0) {
2342         goto do_setcond_const;
2343     }
2344 
2345     switch (cond) {
2346     case TCG_COND_LT:
2347     case TCG_COND_GE:
2348         /*
2349          * Simplify LT/GE comparisons vs zero to a single compare
2350          * vs the high word of the input.
2351          */
2352         if (arg_is_const_val(op->args[3], 0) &&
2353             arg_is_const_val(op->args[4], 0)) {
2354             goto do_setcond_high;
2355         }
2356         break;
2357 
2358     case TCG_COND_NE:
2359         inv = 1;
2360         QEMU_FALLTHROUGH;
2361     case TCG_COND_EQ:
2362         /*
2363          * Simplify EQ/NE comparisons where one of the pairs
2364          * can be simplified.
2365          */
2366         i = do_constant_folding_cond(TCG_TYPE_I32, op->args[1],
2367                                      op->args[3], cond);
2368         switch (i ^ inv) {
2369         case 0:
2370             goto do_setcond_const;
2371         case 1:
2372             goto do_setcond_high;
2373         }
2374 
2375         i = do_constant_folding_cond(TCG_TYPE_I32, op->args[2],
2376                                      op->args[4], cond);
2377         switch (i ^ inv) {
2378         case 0:
2379             goto do_setcond_const;
2380         case 1:
2381             goto do_setcond_low;
2382         }
2383         break;
2384 
2385     case TCG_COND_TSTEQ:
2386     case TCG_COND_TSTNE:
2387         if (arg_is_const_val(op->args[2], 0)) {
2388             goto do_setcond_high;
2389         }
2390         if (arg_is_const_val(op->args[4], 0)) {
2391             goto do_setcond_low;
2392         }
2393         break;
2394 
2395     default:
2396         break;
2397 
2398     do_setcond_low:
2399         op->args[2] = op->args[3];
2400         op->args[3] = cond;
2401         op->opc = INDEX_op_setcond_i32;
2402         return fold_setcond(ctx, op);
2403 
2404     do_setcond_high:
2405         op->args[1] = op->args[2];
2406         op->args[2] = op->args[4];
2407         op->args[3] = cond;
2408         op->opc = INDEX_op_setcond_i32;
2409         return fold_setcond(ctx, op);
2410     }
2411 
2412     ctx->z_mask = 1;
2413     ctx->s_mask = smask_from_zmask(1);
2414     return false;
2415 
2416  do_setcond_const:
2417     return tcg_opt_gen_movi(ctx, op, op->args[0], i);
2418 }
2419 
2420 static bool fold_sextract(OptContext *ctx, TCGOp *op)
2421 {
2422     uint64_t z_mask, s_mask, s_mask_old;
2423     int pos = op->args[2];
2424     int len = op->args[3];
2425 
2426     if (arg_is_const(op->args[1])) {
2427         uint64_t t;
2428 
2429         t = arg_info(op->args[1])->val;
2430         t = sextract64(t, pos, len);
2431         return tcg_opt_gen_movi(ctx, op, op->args[0], t);
2432     }
2433 
2434     z_mask = arg_info(op->args[1])->z_mask;
2435     z_mask = sextract64(z_mask, pos, len);
2436     ctx->z_mask = z_mask;
2437 
2438     s_mask_old = arg_info(op->args[1])->s_mask;
2439     s_mask = sextract64(s_mask_old, pos, len);
2440     s_mask |= MAKE_64BIT_MASK(len, 64 - len);
2441     ctx->s_mask = s_mask;
2442 
2443     if (pos == 0) {
2444         ctx->a_mask = s_mask & ~s_mask_old;
2445     }
2446 
2447     return fold_masks(ctx, op);
2448 }
2449 
2450 static bool fold_shift(OptContext *ctx, TCGOp *op)
2451 {
2452     uint64_t s_mask, z_mask, sign;
2453 
2454     if (fold_const2(ctx, op) ||
2455         fold_ix_to_i(ctx, op, 0) ||
2456         fold_xi_to_x(ctx, op, 0)) {
2457         return true;
2458     }
2459 
2460     s_mask = arg_info(op->args[1])->s_mask;
2461     z_mask = arg_info(op->args[1])->z_mask;
2462 
2463     if (arg_is_const(op->args[2])) {
2464         int sh = arg_info(op->args[2])->val;
2465 
2466         ctx->z_mask = do_constant_folding(op->opc, ctx->type, z_mask, sh);
2467 
2468         s_mask = do_constant_folding(op->opc, ctx->type, s_mask, sh);
2469         ctx->s_mask = smask_from_smask(s_mask);
2470 
2471         return fold_masks(ctx, op);
2472     }
2473 
2474     switch (op->opc) {
2475     CASE_OP_32_64(sar):
2476         /*
2477          * Arithmetic right shift will not reduce the number of
2478          * input sign repetitions.
2479          */
2480         ctx->s_mask = s_mask;
2481         break;
2482     CASE_OP_32_64(shr):
2483         /*
2484          * If the sign bit is known zero, then logical right shift
2485          * will not reduced the number of input sign repetitions.
2486          */
2487         sign = (s_mask & -s_mask) >> 1;
2488         if (sign && !(z_mask & sign)) {
2489             ctx->s_mask = s_mask;
2490         }
2491         break;
2492     default:
2493         break;
2494     }
2495 
2496     return false;
2497 }
2498 
2499 static bool fold_sub_to_neg(OptContext *ctx, TCGOp *op)
2500 {
2501     TCGOpcode neg_op;
2502     bool have_neg;
2503 
2504     if (!arg_is_const(op->args[1]) || arg_info(op->args[1])->val != 0) {
2505         return false;
2506     }
2507 
2508     switch (ctx->type) {
2509     case TCG_TYPE_I32:
2510         neg_op = INDEX_op_neg_i32;
2511         have_neg = true;
2512         break;
2513     case TCG_TYPE_I64:
2514         neg_op = INDEX_op_neg_i64;
2515         have_neg = true;
2516         break;
2517     case TCG_TYPE_V64:
2518     case TCG_TYPE_V128:
2519     case TCG_TYPE_V256:
2520         neg_op = INDEX_op_neg_vec;
2521         have_neg = (TCG_TARGET_HAS_neg_vec &&
2522                     tcg_can_emit_vec_op(neg_op, ctx->type, TCGOP_VECE(op)) > 0);
2523         break;
2524     default:
2525         g_assert_not_reached();
2526     }
2527     if (have_neg) {
2528         op->opc = neg_op;
2529         op->args[1] = op->args[2];
2530         return fold_neg_no_const(ctx, op);
2531     }
2532     return false;
2533 }
2534 
2535 /* We cannot as yet do_constant_folding with vectors. */
2536 static bool fold_sub_vec(OptContext *ctx, TCGOp *op)
2537 {
2538     if (fold_xx_to_i(ctx, op, 0) ||
2539         fold_xi_to_x(ctx, op, 0) ||
2540         fold_sub_to_neg(ctx, op)) {
2541         return true;
2542     }
2543     return false;
2544 }
2545 
2546 static bool fold_sub(OptContext *ctx, TCGOp *op)
2547 {
2548     if (fold_const2(ctx, op) || fold_sub_vec(ctx, op)) {
2549         return true;
2550     }
2551 
2552     /* Fold sub r,x,i to add r,x,-i */
2553     if (arg_is_const(op->args[2])) {
2554         uint64_t val = arg_info(op->args[2])->val;
2555 
2556         op->opc = (ctx->type == TCG_TYPE_I32
2557                    ? INDEX_op_add_i32 : INDEX_op_add_i64);
2558         op->args[2] = arg_new_constant(ctx, -val);
2559     }
2560     return false;
2561 }
2562 
2563 static bool fold_sub2(OptContext *ctx, TCGOp *op)
2564 {
2565     return fold_addsub2(ctx, op, false);
2566 }
2567 
2568 static bool fold_tcg_ld(OptContext *ctx, TCGOp *op)
2569 {
2570     /* We can't do any folding with a load, but we can record bits. */
2571     switch (op->opc) {
2572     CASE_OP_32_64(ld8s):
2573         ctx->s_mask = MAKE_64BIT_MASK(8, 56);
2574         break;
2575     CASE_OP_32_64(ld8u):
2576         ctx->z_mask = MAKE_64BIT_MASK(0, 8);
2577         ctx->s_mask = MAKE_64BIT_MASK(9, 55);
2578         break;
2579     CASE_OP_32_64(ld16s):
2580         ctx->s_mask = MAKE_64BIT_MASK(16, 48);
2581         break;
2582     CASE_OP_32_64(ld16u):
2583         ctx->z_mask = MAKE_64BIT_MASK(0, 16);
2584         ctx->s_mask = MAKE_64BIT_MASK(17, 47);
2585         break;
2586     case INDEX_op_ld32s_i64:
2587         ctx->s_mask = MAKE_64BIT_MASK(32, 32);
2588         break;
2589     case INDEX_op_ld32u_i64:
2590         ctx->z_mask = MAKE_64BIT_MASK(0, 32);
2591         ctx->s_mask = MAKE_64BIT_MASK(33, 31);
2592         break;
2593     default:
2594         g_assert_not_reached();
2595     }
2596     return false;
2597 }
2598 
2599 static bool fold_tcg_ld_memcopy(OptContext *ctx, TCGOp *op)
2600 {
2601     TCGTemp *dst, *src;
2602     intptr_t ofs;
2603     TCGType type;
2604 
2605     if (op->args[1] != tcgv_ptr_arg(tcg_env)) {
2606         return false;
2607     }
2608 
2609     type = ctx->type;
2610     ofs = op->args[2];
2611     dst = arg_temp(op->args[0]);
2612     src = find_mem_copy_for(ctx, type, ofs);
2613     if (src && src->base_type == type) {
2614         return tcg_opt_gen_mov(ctx, op, temp_arg(dst), temp_arg(src));
2615     }
2616 
2617     reset_ts(ctx, dst);
2618     record_mem_copy(ctx, type, dst, ofs, ofs + tcg_type_size(type) - 1);
2619     return true;
2620 }
2621 
2622 static bool fold_tcg_st(OptContext *ctx, TCGOp *op)
2623 {
2624     intptr_t ofs = op->args[2];
2625     intptr_t lm1;
2626 
2627     if (op->args[1] != tcgv_ptr_arg(tcg_env)) {
2628         remove_mem_copy_all(ctx);
2629         return false;
2630     }
2631 
2632     switch (op->opc) {
2633     CASE_OP_32_64(st8):
2634         lm1 = 0;
2635         break;
2636     CASE_OP_32_64(st16):
2637         lm1 = 1;
2638         break;
2639     case INDEX_op_st32_i64:
2640     case INDEX_op_st_i32:
2641         lm1 = 3;
2642         break;
2643     case INDEX_op_st_i64:
2644         lm1 = 7;
2645         break;
2646     case INDEX_op_st_vec:
2647         lm1 = tcg_type_size(ctx->type) - 1;
2648         break;
2649     default:
2650         g_assert_not_reached();
2651     }
2652     remove_mem_copy_in(ctx, ofs, ofs + lm1);
2653     return false;
2654 }
2655 
2656 static bool fold_tcg_st_memcopy(OptContext *ctx, TCGOp *op)
2657 {
2658     TCGTemp *src;
2659     intptr_t ofs, last;
2660     TCGType type;
2661 
2662     if (op->args[1] != tcgv_ptr_arg(tcg_env)) {
2663         fold_tcg_st(ctx, op);
2664         return false;
2665     }
2666 
2667     src = arg_temp(op->args[0]);
2668     ofs = op->args[2];
2669     type = ctx->type;
2670 
2671     /*
2672      * Eliminate duplicate stores of a constant.
2673      * This happens frequently when the target ISA zero-extends.
2674      */
2675     if (ts_is_const(src)) {
2676         TCGTemp *prev = find_mem_copy_for(ctx, type, ofs);
2677         if (src == prev) {
2678             tcg_op_remove(ctx->tcg, op);
2679             return true;
2680         }
2681     }
2682 
2683     last = ofs + tcg_type_size(type) - 1;
2684     remove_mem_copy_in(ctx, ofs, last);
2685     record_mem_copy(ctx, type, src, ofs, last);
2686     return false;
2687 }
2688 
2689 static bool fold_xor(OptContext *ctx, TCGOp *op)
2690 {
2691     if (fold_const2_commutative(ctx, op) ||
2692         fold_xx_to_i(ctx, op, 0) ||
2693         fold_xi_to_x(ctx, op, 0) ||
2694         fold_xi_to_not(ctx, op, -1)) {
2695         return true;
2696     }
2697 
2698     ctx->z_mask = arg_info(op->args[1])->z_mask
2699                 | arg_info(op->args[2])->z_mask;
2700     ctx->s_mask = arg_info(op->args[1])->s_mask
2701                 & arg_info(op->args[2])->s_mask;
2702     return fold_masks(ctx, op);
2703 }
2704 
2705 /* Propagate constants and copies, fold constant expressions. */
2706 void tcg_optimize(TCGContext *s)
2707 {
2708     int nb_temps, i;
2709     TCGOp *op, *op_next;
2710     OptContext ctx = { .tcg = s };
2711 
2712     QSIMPLEQ_INIT(&ctx.mem_free);
2713 
2714     /* Array VALS has an element for each temp.
2715        If this temp holds a constant then its value is kept in VALS' element.
2716        If this temp is a copy of other ones then the other copies are
2717        available through the doubly linked circular list. */
2718 
2719     nb_temps = s->nb_temps;
2720     for (i = 0; i < nb_temps; ++i) {
2721         s->temps[i].state_ptr = NULL;
2722     }
2723 
2724     QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) {
2725         TCGOpcode opc = op->opc;
2726         const TCGOpDef *def;
2727         bool done = false;
2728 
2729         /* Calls are special. */
2730         if (opc == INDEX_op_call) {
2731             fold_call(&ctx, op);
2732             continue;
2733         }
2734 
2735         def = &tcg_op_defs[opc];
2736         init_arguments(&ctx, op, def->nb_oargs + def->nb_iargs);
2737         copy_propagate(&ctx, op, def->nb_oargs, def->nb_iargs);
2738 
2739         /* Pre-compute the type of the operation. */
2740         if (def->flags & TCG_OPF_VECTOR) {
2741             ctx.type = TCG_TYPE_V64 + TCGOP_VECL(op);
2742         } else if (def->flags & TCG_OPF_64BIT) {
2743             ctx.type = TCG_TYPE_I64;
2744         } else {
2745             ctx.type = TCG_TYPE_I32;
2746         }
2747 
2748         /* Assume all bits affected, no bits known zero, no sign reps. */
2749         ctx.a_mask = -1;
2750         ctx.z_mask = -1;
2751         ctx.s_mask = 0;
2752 
2753         /*
2754          * Process each opcode.
2755          * Sorted alphabetically by opcode as much as possible.
2756          */
2757         switch (opc) {
2758         CASE_OP_32_64(add):
2759             done = fold_add(&ctx, op);
2760             break;
2761         case INDEX_op_add_vec:
2762             done = fold_add_vec(&ctx, op);
2763             break;
2764         CASE_OP_32_64(add2):
2765             done = fold_add2(&ctx, op);
2766             break;
2767         CASE_OP_32_64_VEC(and):
2768             done = fold_and(&ctx, op);
2769             break;
2770         CASE_OP_32_64_VEC(andc):
2771             done = fold_andc(&ctx, op);
2772             break;
2773         CASE_OP_32_64(brcond):
2774             done = fold_brcond(&ctx, op);
2775             break;
2776         case INDEX_op_brcond2_i32:
2777             done = fold_brcond2(&ctx, op);
2778             break;
2779         CASE_OP_32_64(bswap16):
2780         CASE_OP_32_64(bswap32):
2781         case INDEX_op_bswap64_i64:
2782             done = fold_bswap(&ctx, op);
2783             break;
2784         CASE_OP_32_64(clz):
2785         CASE_OP_32_64(ctz):
2786             done = fold_count_zeros(&ctx, op);
2787             break;
2788         CASE_OP_32_64(ctpop):
2789             done = fold_ctpop(&ctx, op);
2790             break;
2791         CASE_OP_32_64(deposit):
2792             done = fold_deposit(&ctx, op);
2793             break;
2794         CASE_OP_32_64(div):
2795         CASE_OP_32_64(divu):
2796             done = fold_divide(&ctx, op);
2797             break;
2798         case INDEX_op_dup_vec:
2799             done = fold_dup(&ctx, op);
2800             break;
2801         case INDEX_op_dup2_vec:
2802             done = fold_dup2(&ctx, op);
2803             break;
2804         CASE_OP_32_64_VEC(eqv):
2805             done = fold_eqv(&ctx, op);
2806             break;
2807         CASE_OP_32_64(extract):
2808             done = fold_extract(&ctx, op);
2809             break;
2810         CASE_OP_32_64(extract2):
2811             done = fold_extract2(&ctx, op);
2812             break;
2813         CASE_OP_32_64(ext8s):
2814         CASE_OP_32_64(ext16s):
2815         case INDEX_op_ext32s_i64:
2816         case INDEX_op_ext_i32_i64:
2817             done = fold_exts(&ctx, op);
2818             break;
2819         CASE_OP_32_64(ext8u):
2820         CASE_OP_32_64(ext16u):
2821         case INDEX_op_ext32u_i64:
2822         case INDEX_op_extu_i32_i64:
2823         case INDEX_op_extrl_i64_i32:
2824         case INDEX_op_extrh_i64_i32:
2825             done = fold_extu(&ctx, op);
2826             break;
2827         CASE_OP_32_64(ld8s):
2828         CASE_OP_32_64(ld8u):
2829         CASE_OP_32_64(ld16s):
2830         CASE_OP_32_64(ld16u):
2831         case INDEX_op_ld32s_i64:
2832         case INDEX_op_ld32u_i64:
2833             done = fold_tcg_ld(&ctx, op);
2834             break;
2835         case INDEX_op_ld_i32:
2836         case INDEX_op_ld_i64:
2837         case INDEX_op_ld_vec:
2838             done = fold_tcg_ld_memcopy(&ctx, op);
2839             break;
2840         CASE_OP_32_64(st8):
2841         CASE_OP_32_64(st16):
2842         case INDEX_op_st32_i64:
2843             done = fold_tcg_st(&ctx, op);
2844             break;
2845         case INDEX_op_st_i32:
2846         case INDEX_op_st_i64:
2847         case INDEX_op_st_vec:
2848             done = fold_tcg_st_memcopy(&ctx, op);
2849             break;
2850         case INDEX_op_mb:
2851             done = fold_mb(&ctx, op);
2852             break;
2853         CASE_OP_32_64_VEC(mov):
2854             done = fold_mov(&ctx, op);
2855             break;
2856         CASE_OP_32_64(movcond):
2857             done = fold_movcond(&ctx, op);
2858             break;
2859         CASE_OP_32_64(mul):
2860             done = fold_mul(&ctx, op);
2861             break;
2862         CASE_OP_32_64(mulsh):
2863         CASE_OP_32_64(muluh):
2864             done = fold_mul_highpart(&ctx, op);
2865             break;
2866         CASE_OP_32_64(muls2):
2867         CASE_OP_32_64(mulu2):
2868             done = fold_multiply2(&ctx, op);
2869             break;
2870         CASE_OP_32_64_VEC(nand):
2871             done = fold_nand(&ctx, op);
2872             break;
2873         CASE_OP_32_64(neg):
2874             done = fold_neg(&ctx, op);
2875             break;
2876         CASE_OP_32_64_VEC(nor):
2877             done = fold_nor(&ctx, op);
2878             break;
2879         CASE_OP_32_64_VEC(not):
2880             done = fold_not(&ctx, op);
2881             break;
2882         CASE_OP_32_64_VEC(or):
2883             done = fold_or(&ctx, op);
2884             break;
2885         CASE_OP_32_64_VEC(orc):
2886             done = fold_orc(&ctx, op);
2887             break;
2888         case INDEX_op_qemu_ld_a32_i32:
2889         case INDEX_op_qemu_ld_a64_i32:
2890         case INDEX_op_qemu_ld_a32_i64:
2891         case INDEX_op_qemu_ld_a64_i64:
2892         case INDEX_op_qemu_ld_a32_i128:
2893         case INDEX_op_qemu_ld_a64_i128:
2894             done = fold_qemu_ld(&ctx, op);
2895             break;
2896         case INDEX_op_qemu_st8_a32_i32:
2897         case INDEX_op_qemu_st8_a64_i32:
2898         case INDEX_op_qemu_st_a32_i32:
2899         case INDEX_op_qemu_st_a64_i32:
2900         case INDEX_op_qemu_st_a32_i64:
2901         case INDEX_op_qemu_st_a64_i64:
2902         case INDEX_op_qemu_st_a32_i128:
2903         case INDEX_op_qemu_st_a64_i128:
2904             done = fold_qemu_st(&ctx, op);
2905             break;
2906         CASE_OP_32_64(rem):
2907         CASE_OP_32_64(remu):
2908             done = fold_remainder(&ctx, op);
2909             break;
2910         CASE_OP_32_64(rotl):
2911         CASE_OP_32_64(rotr):
2912         CASE_OP_32_64(sar):
2913         CASE_OP_32_64(shl):
2914         CASE_OP_32_64(shr):
2915             done = fold_shift(&ctx, op);
2916             break;
2917         CASE_OP_32_64(setcond):
2918             done = fold_setcond(&ctx, op);
2919             break;
2920         CASE_OP_32_64(negsetcond):
2921             done = fold_negsetcond(&ctx, op);
2922             break;
2923         case INDEX_op_setcond2_i32:
2924             done = fold_setcond2(&ctx, op);
2925             break;
2926         CASE_OP_32_64(sextract):
2927             done = fold_sextract(&ctx, op);
2928             break;
2929         CASE_OP_32_64(sub):
2930             done = fold_sub(&ctx, op);
2931             break;
2932         case INDEX_op_sub_vec:
2933             done = fold_sub_vec(&ctx, op);
2934             break;
2935         CASE_OP_32_64(sub2):
2936             done = fold_sub2(&ctx, op);
2937             break;
2938         CASE_OP_32_64_VEC(xor):
2939             done = fold_xor(&ctx, op);
2940             break;
2941         default:
2942             break;
2943         }
2944 
2945         if (!done) {
2946             finish_folding(&ctx, op);
2947         }
2948     }
2949 }
2950