xref: /openbmc/qemu/tcg/optimize.c (revision ad66b5cb)
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 "tcg/tcg-op.h"
29 #include "tcg-internal.h"
30 
31 #define CASE_OP_32_64(x)                        \
32         glue(glue(case INDEX_op_, x), _i32):    \
33         glue(glue(case INDEX_op_, x), _i64)
34 
35 #define CASE_OP_32_64_VEC(x)                    \
36         glue(glue(case INDEX_op_, x), _i32):    \
37         glue(glue(case INDEX_op_, x), _i64):    \
38         glue(glue(case INDEX_op_, x), _vec)
39 
40 typedef struct TempOptInfo {
41     bool is_const;
42     TCGTemp *prev_copy;
43     TCGTemp *next_copy;
44     uint64_t val;
45     uint64_t z_mask;  /* mask bit is 0 if and only if value bit is 0 */
46     uint64_t s_mask;  /* a left-aligned mask of clrsb(value) bits. */
47 } TempOptInfo;
48 
49 typedef struct OptContext {
50     TCGContext *tcg;
51     TCGOp *prev_mb;
52     TCGTempSet temps_used;
53 
54     /* In flight values from optimization. */
55     uint64_t a_mask;  /* mask bit is 0 iff value identical to first input */
56     uint64_t z_mask;  /* mask bit is 0 iff value bit is 0 */
57     uint64_t s_mask;  /* mask of clrsb(value) bits */
58     TCGType type;
59 } OptContext;
60 
61 /* Calculate the smask for a specific value. */
62 static uint64_t smask_from_value(uint64_t value)
63 {
64     int rep = clrsb64(value);
65     return ~(~0ull >> rep);
66 }
67 
68 /*
69  * Calculate the smask for a given set of known-zeros.
70  * If there are lots of zeros on the left, we can consider the remainder
71  * an unsigned field, and thus the corresponding signed field is one bit
72  * larger.
73  */
74 static uint64_t smask_from_zmask(uint64_t zmask)
75 {
76     /*
77      * Only the 0 bits are significant for zmask, thus the msb itself
78      * must be zero, else we have no sign information.
79      */
80     int rep = clz64(zmask);
81     if (rep == 0) {
82         return 0;
83     }
84     rep -= 1;
85     return ~(~0ull >> rep);
86 }
87 
88 /*
89  * Recreate a properly left-aligned smask after manipulation.
90  * Some bit-shuffling, particularly shifts and rotates, may
91  * retain sign bits on the left, but may scatter disconnected
92  * sign bits on the right.  Retain only what remains to the left.
93  */
94 static uint64_t smask_from_smask(int64_t smask)
95 {
96     /* Only the 1 bits are significant for smask */
97     return smask_from_zmask(~smask);
98 }
99 
100 static inline TempOptInfo *ts_info(TCGTemp *ts)
101 {
102     return ts->state_ptr;
103 }
104 
105 static inline TempOptInfo *arg_info(TCGArg arg)
106 {
107     return ts_info(arg_temp(arg));
108 }
109 
110 static inline bool ts_is_const(TCGTemp *ts)
111 {
112     return ts_info(ts)->is_const;
113 }
114 
115 static inline bool arg_is_const(TCGArg arg)
116 {
117     return ts_is_const(arg_temp(arg));
118 }
119 
120 static inline bool ts_is_copy(TCGTemp *ts)
121 {
122     return ts_info(ts)->next_copy != ts;
123 }
124 
125 /* Reset TEMP's state, possibly removing the temp for the list of copies.  */
126 static void reset_ts(TCGTemp *ts)
127 {
128     TempOptInfo *ti = ts_info(ts);
129     TempOptInfo *pi = ts_info(ti->prev_copy);
130     TempOptInfo *ni = ts_info(ti->next_copy);
131 
132     ni->prev_copy = ti->prev_copy;
133     pi->next_copy = ti->next_copy;
134     ti->next_copy = ts;
135     ti->prev_copy = ts;
136     ti->is_const = false;
137     ti->z_mask = -1;
138     ti->s_mask = 0;
139 }
140 
141 static void reset_temp(TCGArg arg)
142 {
143     reset_ts(arg_temp(arg));
144 }
145 
146 /* Initialize and activate a temporary.  */
147 static void init_ts_info(OptContext *ctx, TCGTemp *ts)
148 {
149     size_t idx = temp_idx(ts);
150     TempOptInfo *ti;
151 
152     if (test_bit(idx, ctx->temps_used.l)) {
153         return;
154     }
155     set_bit(idx, ctx->temps_used.l);
156 
157     ti = ts->state_ptr;
158     if (ti == NULL) {
159         ti = tcg_malloc(sizeof(TempOptInfo));
160         ts->state_ptr = ti;
161     }
162 
163     ti->next_copy = ts;
164     ti->prev_copy = ts;
165     if (ts->kind == TEMP_CONST) {
166         ti->is_const = true;
167         ti->val = ts->val;
168         ti->z_mask = ts->val;
169         ti->s_mask = smask_from_value(ts->val);
170     } else {
171         ti->is_const = false;
172         ti->z_mask = -1;
173         ti->s_mask = 0;
174     }
175 }
176 
177 static TCGTemp *find_better_copy(TCGContext *s, TCGTemp *ts)
178 {
179     TCGTemp *i, *g, *l;
180 
181     /* If this is already readonly, we can't do better. */
182     if (temp_readonly(ts)) {
183         return ts;
184     }
185 
186     g = l = NULL;
187     for (i = ts_info(ts)->next_copy; i != ts; i = ts_info(i)->next_copy) {
188         if (temp_readonly(i)) {
189             return i;
190         } else if (i->kind > ts->kind) {
191             if (i->kind == TEMP_GLOBAL) {
192                 g = i;
193             } else if (i->kind == TEMP_TB) {
194                 l = i;
195             }
196         }
197     }
198 
199     /* If we didn't find a better representation, return the same temp. */
200     return g ? g : l ? l : ts;
201 }
202 
203 static bool ts_are_copies(TCGTemp *ts1, TCGTemp *ts2)
204 {
205     TCGTemp *i;
206 
207     if (ts1 == ts2) {
208         return true;
209     }
210 
211     if (!ts_is_copy(ts1) || !ts_is_copy(ts2)) {
212         return false;
213     }
214 
215     for (i = ts_info(ts1)->next_copy; i != ts1; i = ts_info(i)->next_copy) {
216         if (i == ts2) {
217             return true;
218         }
219     }
220 
221     return false;
222 }
223 
224 static bool args_are_copies(TCGArg arg1, TCGArg arg2)
225 {
226     return ts_are_copies(arg_temp(arg1), arg_temp(arg2));
227 }
228 
229 static bool tcg_opt_gen_mov(OptContext *ctx, TCGOp *op, TCGArg dst, TCGArg src)
230 {
231     TCGTemp *dst_ts = arg_temp(dst);
232     TCGTemp *src_ts = arg_temp(src);
233     TempOptInfo *di;
234     TempOptInfo *si;
235     TCGOpcode new_op;
236 
237     if (ts_are_copies(dst_ts, src_ts)) {
238         tcg_op_remove(ctx->tcg, op);
239         return true;
240     }
241 
242     reset_ts(dst_ts);
243     di = ts_info(dst_ts);
244     si = ts_info(src_ts);
245 
246     switch (ctx->type) {
247     case TCG_TYPE_I32:
248         new_op = INDEX_op_mov_i32;
249         break;
250     case TCG_TYPE_I64:
251         new_op = INDEX_op_mov_i64;
252         break;
253     case TCG_TYPE_V64:
254     case TCG_TYPE_V128:
255     case TCG_TYPE_V256:
256         /* TCGOP_VECL and TCGOP_VECE remain unchanged.  */
257         new_op = INDEX_op_mov_vec;
258         break;
259     default:
260         g_assert_not_reached();
261     }
262     op->opc = new_op;
263     op->args[0] = dst;
264     op->args[1] = src;
265 
266     di->z_mask = si->z_mask;
267     di->s_mask = si->s_mask;
268 
269     if (src_ts->type == dst_ts->type) {
270         TempOptInfo *ni = ts_info(si->next_copy);
271 
272         di->next_copy = si->next_copy;
273         di->prev_copy = src_ts;
274         ni->prev_copy = dst_ts;
275         si->next_copy = dst_ts;
276         di->is_const = si->is_const;
277         di->val = si->val;
278     }
279     return true;
280 }
281 
282 static bool tcg_opt_gen_movi(OptContext *ctx, TCGOp *op,
283                              TCGArg dst, uint64_t val)
284 {
285     TCGTemp *tv;
286 
287     if (ctx->type == TCG_TYPE_I32) {
288         val = (int32_t)val;
289     }
290 
291     /* Convert movi to mov with constant temp. */
292     tv = tcg_constant_internal(ctx->type, val);
293     init_ts_info(ctx, tv);
294     return tcg_opt_gen_mov(ctx, op, dst, temp_arg(tv));
295 }
296 
297 static uint64_t do_constant_folding_2(TCGOpcode op, uint64_t x, uint64_t y)
298 {
299     uint64_t l64, h64;
300 
301     switch (op) {
302     CASE_OP_32_64(add):
303         return x + y;
304 
305     CASE_OP_32_64(sub):
306         return x - y;
307 
308     CASE_OP_32_64(mul):
309         return x * y;
310 
311     CASE_OP_32_64_VEC(and):
312         return x & y;
313 
314     CASE_OP_32_64_VEC(or):
315         return x | y;
316 
317     CASE_OP_32_64_VEC(xor):
318         return x ^ y;
319 
320     case INDEX_op_shl_i32:
321         return (uint32_t)x << (y & 31);
322 
323     case INDEX_op_shl_i64:
324         return (uint64_t)x << (y & 63);
325 
326     case INDEX_op_shr_i32:
327         return (uint32_t)x >> (y & 31);
328 
329     case INDEX_op_shr_i64:
330         return (uint64_t)x >> (y & 63);
331 
332     case INDEX_op_sar_i32:
333         return (int32_t)x >> (y & 31);
334 
335     case INDEX_op_sar_i64:
336         return (int64_t)x >> (y & 63);
337 
338     case INDEX_op_rotr_i32:
339         return ror32(x, y & 31);
340 
341     case INDEX_op_rotr_i64:
342         return ror64(x, y & 63);
343 
344     case INDEX_op_rotl_i32:
345         return rol32(x, y & 31);
346 
347     case INDEX_op_rotl_i64:
348         return rol64(x, y & 63);
349 
350     CASE_OP_32_64_VEC(not):
351         return ~x;
352 
353     CASE_OP_32_64(neg):
354         return -x;
355 
356     CASE_OP_32_64_VEC(andc):
357         return x & ~y;
358 
359     CASE_OP_32_64_VEC(orc):
360         return x | ~y;
361 
362     CASE_OP_32_64_VEC(eqv):
363         return ~(x ^ y);
364 
365     CASE_OP_32_64_VEC(nand):
366         return ~(x & y);
367 
368     CASE_OP_32_64_VEC(nor):
369         return ~(x | y);
370 
371     case INDEX_op_clz_i32:
372         return (uint32_t)x ? clz32(x) : y;
373 
374     case INDEX_op_clz_i64:
375         return x ? clz64(x) : y;
376 
377     case INDEX_op_ctz_i32:
378         return (uint32_t)x ? ctz32(x) : y;
379 
380     case INDEX_op_ctz_i64:
381         return x ? ctz64(x) : y;
382 
383     case INDEX_op_ctpop_i32:
384         return ctpop32(x);
385 
386     case INDEX_op_ctpop_i64:
387         return ctpop64(x);
388 
389     CASE_OP_32_64(ext8s):
390         return (int8_t)x;
391 
392     CASE_OP_32_64(ext16s):
393         return (int16_t)x;
394 
395     CASE_OP_32_64(ext8u):
396         return (uint8_t)x;
397 
398     CASE_OP_32_64(ext16u):
399         return (uint16_t)x;
400 
401     CASE_OP_32_64(bswap16):
402         x = bswap16(x);
403         return y & TCG_BSWAP_OS ? (int16_t)x : x;
404 
405     CASE_OP_32_64(bswap32):
406         x = bswap32(x);
407         return y & TCG_BSWAP_OS ? (int32_t)x : x;
408 
409     case INDEX_op_bswap64_i64:
410         return bswap64(x);
411 
412     case INDEX_op_ext_i32_i64:
413     case INDEX_op_ext32s_i64:
414         return (int32_t)x;
415 
416     case INDEX_op_extu_i32_i64:
417     case INDEX_op_extrl_i64_i32:
418     case INDEX_op_ext32u_i64:
419         return (uint32_t)x;
420 
421     case INDEX_op_extrh_i64_i32:
422         return (uint64_t)x >> 32;
423 
424     case INDEX_op_muluh_i32:
425         return ((uint64_t)(uint32_t)x * (uint32_t)y) >> 32;
426     case INDEX_op_mulsh_i32:
427         return ((int64_t)(int32_t)x * (int32_t)y) >> 32;
428 
429     case INDEX_op_muluh_i64:
430         mulu64(&l64, &h64, x, y);
431         return h64;
432     case INDEX_op_mulsh_i64:
433         muls64(&l64, &h64, x, y);
434         return h64;
435 
436     case INDEX_op_div_i32:
437         /* Avoid crashing on divide by zero, otherwise undefined.  */
438         return (int32_t)x / ((int32_t)y ? : 1);
439     case INDEX_op_divu_i32:
440         return (uint32_t)x / ((uint32_t)y ? : 1);
441     case INDEX_op_div_i64:
442         return (int64_t)x / ((int64_t)y ? : 1);
443     case INDEX_op_divu_i64:
444         return (uint64_t)x / ((uint64_t)y ? : 1);
445 
446     case INDEX_op_rem_i32:
447         return (int32_t)x % ((int32_t)y ? : 1);
448     case INDEX_op_remu_i32:
449         return (uint32_t)x % ((uint32_t)y ? : 1);
450     case INDEX_op_rem_i64:
451         return (int64_t)x % ((int64_t)y ? : 1);
452     case INDEX_op_remu_i64:
453         return (uint64_t)x % ((uint64_t)y ? : 1);
454 
455     default:
456         fprintf(stderr,
457                 "Unrecognized operation %d in do_constant_folding.\n", op);
458         tcg_abort();
459     }
460 }
461 
462 static uint64_t do_constant_folding(TCGOpcode op, TCGType type,
463                                     uint64_t x, uint64_t y)
464 {
465     uint64_t res = do_constant_folding_2(op, x, y);
466     if (type == TCG_TYPE_I32) {
467         res = (int32_t)res;
468     }
469     return res;
470 }
471 
472 static bool do_constant_folding_cond_32(uint32_t x, uint32_t y, TCGCond c)
473 {
474     switch (c) {
475     case TCG_COND_EQ:
476         return x == y;
477     case TCG_COND_NE:
478         return x != y;
479     case TCG_COND_LT:
480         return (int32_t)x < (int32_t)y;
481     case TCG_COND_GE:
482         return (int32_t)x >= (int32_t)y;
483     case TCG_COND_LE:
484         return (int32_t)x <= (int32_t)y;
485     case TCG_COND_GT:
486         return (int32_t)x > (int32_t)y;
487     case TCG_COND_LTU:
488         return x < y;
489     case TCG_COND_GEU:
490         return x >= y;
491     case TCG_COND_LEU:
492         return x <= y;
493     case TCG_COND_GTU:
494         return x > y;
495     default:
496         tcg_abort();
497     }
498 }
499 
500 static bool do_constant_folding_cond_64(uint64_t x, uint64_t y, TCGCond c)
501 {
502     switch (c) {
503     case TCG_COND_EQ:
504         return x == y;
505     case TCG_COND_NE:
506         return x != y;
507     case TCG_COND_LT:
508         return (int64_t)x < (int64_t)y;
509     case TCG_COND_GE:
510         return (int64_t)x >= (int64_t)y;
511     case TCG_COND_LE:
512         return (int64_t)x <= (int64_t)y;
513     case TCG_COND_GT:
514         return (int64_t)x > (int64_t)y;
515     case TCG_COND_LTU:
516         return x < y;
517     case TCG_COND_GEU:
518         return x >= y;
519     case TCG_COND_LEU:
520         return x <= y;
521     case TCG_COND_GTU:
522         return x > y;
523     default:
524         tcg_abort();
525     }
526 }
527 
528 static bool do_constant_folding_cond_eq(TCGCond c)
529 {
530     switch (c) {
531     case TCG_COND_GT:
532     case TCG_COND_LTU:
533     case TCG_COND_LT:
534     case TCG_COND_GTU:
535     case TCG_COND_NE:
536         return 0;
537     case TCG_COND_GE:
538     case TCG_COND_GEU:
539     case TCG_COND_LE:
540     case TCG_COND_LEU:
541     case TCG_COND_EQ:
542         return 1;
543     default:
544         tcg_abort();
545     }
546 }
547 
548 /*
549  * Return -1 if the condition can't be simplified,
550  * and the result of the condition (0 or 1) if it can.
551  */
552 static int do_constant_folding_cond(TCGType type, TCGArg x,
553                                     TCGArg y, TCGCond c)
554 {
555     if (arg_is_const(x) && arg_is_const(y)) {
556         uint64_t xv = arg_info(x)->val;
557         uint64_t yv = arg_info(y)->val;
558 
559         switch (type) {
560         case TCG_TYPE_I32:
561             return do_constant_folding_cond_32(xv, yv, c);
562         case TCG_TYPE_I64:
563             return do_constant_folding_cond_64(xv, yv, c);
564         default:
565             /* Only scalar comparisons are optimizable */
566             return -1;
567         }
568     } else if (args_are_copies(x, y)) {
569         return do_constant_folding_cond_eq(c);
570     } else if (arg_is_const(y) && arg_info(y)->val == 0) {
571         switch (c) {
572         case TCG_COND_LTU:
573             return 0;
574         case TCG_COND_GEU:
575             return 1;
576         default:
577             return -1;
578         }
579     }
580     return -1;
581 }
582 
583 /*
584  * Return -1 if the condition can't be simplified,
585  * and the result of the condition (0 or 1) if it can.
586  */
587 static int do_constant_folding_cond2(TCGArg *p1, TCGArg *p2, TCGCond c)
588 {
589     TCGArg al = p1[0], ah = p1[1];
590     TCGArg bl = p2[0], bh = p2[1];
591 
592     if (arg_is_const(bl) && arg_is_const(bh)) {
593         tcg_target_ulong blv = arg_info(bl)->val;
594         tcg_target_ulong bhv = arg_info(bh)->val;
595         uint64_t b = deposit64(blv, 32, 32, bhv);
596 
597         if (arg_is_const(al) && arg_is_const(ah)) {
598             tcg_target_ulong alv = arg_info(al)->val;
599             tcg_target_ulong ahv = arg_info(ah)->val;
600             uint64_t a = deposit64(alv, 32, 32, ahv);
601             return do_constant_folding_cond_64(a, b, c);
602         }
603         if (b == 0) {
604             switch (c) {
605             case TCG_COND_LTU:
606                 return 0;
607             case TCG_COND_GEU:
608                 return 1;
609             default:
610                 break;
611             }
612         }
613     }
614     if (args_are_copies(al, bl) && args_are_copies(ah, bh)) {
615         return do_constant_folding_cond_eq(c);
616     }
617     return -1;
618 }
619 
620 /**
621  * swap_commutative:
622  * @dest: TCGArg of the destination argument, or NO_DEST.
623  * @p1: first paired argument
624  * @p2: second paired argument
625  *
626  * If *@p1 is a constant and *@p2 is not, swap.
627  * If *@p2 matches @dest, swap.
628  * Return true if a swap was performed.
629  */
630 
631 #define NO_DEST  temp_arg(NULL)
632 
633 static bool swap_commutative(TCGArg dest, TCGArg *p1, TCGArg *p2)
634 {
635     TCGArg a1 = *p1, a2 = *p2;
636     int sum = 0;
637     sum += arg_is_const(a1);
638     sum -= arg_is_const(a2);
639 
640     /* Prefer the constant in second argument, and then the form
641        op a, a, b, which is better handled on non-RISC hosts. */
642     if (sum > 0 || (sum == 0 && dest == a2)) {
643         *p1 = a2;
644         *p2 = a1;
645         return true;
646     }
647     return false;
648 }
649 
650 static bool swap_commutative2(TCGArg *p1, TCGArg *p2)
651 {
652     int sum = 0;
653     sum += arg_is_const(p1[0]);
654     sum += arg_is_const(p1[1]);
655     sum -= arg_is_const(p2[0]);
656     sum -= arg_is_const(p2[1]);
657     if (sum > 0) {
658         TCGArg t;
659         t = p1[0], p1[0] = p2[0], p2[0] = t;
660         t = p1[1], p1[1] = p2[1], p2[1] = t;
661         return true;
662     }
663     return false;
664 }
665 
666 static void init_arguments(OptContext *ctx, TCGOp *op, int nb_args)
667 {
668     for (int i = 0; i < nb_args; i++) {
669         TCGTemp *ts = arg_temp(op->args[i]);
670         init_ts_info(ctx, ts);
671     }
672 }
673 
674 static void copy_propagate(OptContext *ctx, TCGOp *op,
675                            int nb_oargs, int nb_iargs)
676 {
677     TCGContext *s = ctx->tcg;
678 
679     for (int i = nb_oargs; i < nb_oargs + nb_iargs; i++) {
680         TCGTemp *ts = arg_temp(op->args[i]);
681         if (ts_is_copy(ts)) {
682             op->args[i] = temp_arg(find_better_copy(s, ts));
683         }
684     }
685 }
686 
687 static void finish_folding(OptContext *ctx, TCGOp *op)
688 {
689     const TCGOpDef *def = &tcg_op_defs[op->opc];
690     int i, nb_oargs;
691 
692     /*
693      * For an opcode that ends a BB, reset all temp data.
694      * We do no cross-BB optimization.
695      */
696     if (def->flags & TCG_OPF_BB_END) {
697         memset(&ctx->temps_used, 0, sizeof(ctx->temps_used));
698         ctx->prev_mb = NULL;
699         return;
700     }
701 
702     nb_oargs = def->nb_oargs;
703     for (i = 0; i < nb_oargs; i++) {
704         TCGTemp *ts = arg_temp(op->args[i]);
705         reset_ts(ts);
706         /*
707          * Save the corresponding known-zero/sign bits mask for the
708          * first output argument (only one supported so far).
709          */
710         if (i == 0) {
711             ts_info(ts)->z_mask = ctx->z_mask;
712             ts_info(ts)->s_mask = ctx->s_mask;
713         }
714     }
715 }
716 
717 /*
718  * The fold_* functions return true when processing is complete,
719  * usually by folding the operation to a constant or to a copy,
720  * and calling tcg_opt_gen_{mov,movi}.  They may do other things,
721  * like collect information about the value produced, for use in
722  * optimizing a subsequent operation.
723  *
724  * These first fold_* functions are all helpers, used by other
725  * folders for more specific operations.
726  */
727 
728 static bool fold_const1(OptContext *ctx, TCGOp *op)
729 {
730     if (arg_is_const(op->args[1])) {
731         uint64_t t;
732 
733         t = arg_info(op->args[1])->val;
734         t = do_constant_folding(op->opc, ctx->type, t, 0);
735         return tcg_opt_gen_movi(ctx, op, op->args[0], t);
736     }
737     return false;
738 }
739 
740 static bool fold_const2(OptContext *ctx, TCGOp *op)
741 {
742     if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) {
743         uint64_t t1 = arg_info(op->args[1])->val;
744         uint64_t t2 = arg_info(op->args[2])->val;
745 
746         t1 = do_constant_folding(op->opc, ctx->type, t1, t2);
747         return tcg_opt_gen_movi(ctx, op, op->args[0], t1);
748     }
749     return false;
750 }
751 
752 static bool fold_commutative(OptContext *ctx, TCGOp *op)
753 {
754     swap_commutative(op->args[0], &op->args[1], &op->args[2]);
755     return false;
756 }
757 
758 static bool fold_const2_commutative(OptContext *ctx, TCGOp *op)
759 {
760     swap_commutative(op->args[0], &op->args[1], &op->args[2]);
761     return fold_const2(ctx, op);
762 }
763 
764 static bool fold_masks(OptContext *ctx, TCGOp *op)
765 {
766     uint64_t a_mask = ctx->a_mask;
767     uint64_t z_mask = ctx->z_mask;
768     uint64_t s_mask = ctx->s_mask;
769 
770     /*
771      * 32-bit ops generate 32-bit results, which for the purpose of
772      * simplifying tcg are sign-extended.  Certainly that's how we
773      * represent our constants elsewhere.  Note that the bits will
774      * be reset properly for a 64-bit value when encountering the
775      * type changing opcodes.
776      */
777     if (ctx->type == TCG_TYPE_I32) {
778         a_mask = (int32_t)a_mask;
779         z_mask = (int32_t)z_mask;
780         s_mask |= MAKE_64BIT_MASK(32, 32);
781         ctx->z_mask = z_mask;
782         ctx->s_mask = s_mask;
783     }
784 
785     if (z_mask == 0) {
786         return tcg_opt_gen_movi(ctx, op, op->args[0], 0);
787     }
788     if (a_mask == 0) {
789         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
790     }
791     return false;
792 }
793 
794 /*
795  * Convert @op to NOT, if NOT is supported by the host.
796  * Return true f the conversion is successful, which will still
797  * indicate that the processing is complete.
798  */
799 static bool fold_not(OptContext *ctx, TCGOp *op);
800 static bool fold_to_not(OptContext *ctx, TCGOp *op, int idx)
801 {
802     TCGOpcode not_op;
803     bool have_not;
804 
805     switch (ctx->type) {
806     case TCG_TYPE_I32:
807         not_op = INDEX_op_not_i32;
808         have_not = TCG_TARGET_HAS_not_i32;
809         break;
810     case TCG_TYPE_I64:
811         not_op = INDEX_op_not_i64;
812         have_not = TCG_TARGET_HAS_not_i64;
813         break;
814     case TCG_TYPE_V64:
815     case TCG_TYPE_V128:
816     case TCG_TYPE_V256:
817         not_op = INDEX_op_not_vec;
818         have_not = TCG_TARGET_HAS_not_vec;
819         break;
820     default:
821         g_assert_not_reached();
822     }
823     if (have_not) {
824         op->opc = not_op;
825         op->args[1] = op->args[idx];
826         return fold_not(ctx, op);
827     }
828     return false;
829 }
830 
831 /* If the binary operation has first argument @i, fold to @i. */
832 static bool fold_ix_to_i(OptContext *ctx, TCGOp *op, uint64_t i)
833 {
834     if (arg_is_const(op->args[1]) && arg_info(op->args[1])->val == i) {
835         return tcg_opt_gen_movi(ctx, op, op->args[0], i);
836     }
837     return false;
838 }
839 
840 /* If the binary operation has first argument @i, fold to NOT. */
841 static bool fold_ix_to_not(OptContext *ctx, TCGOp *op, uint64_t i)
842 {
843     if (arg_is_const(op->args[1]) && arg_info(op->args[1])->val == i) {
844         return fold_to_not(ctx, op, 2);
845     }
846     return false;
847 }
848 
849 /* If the binary operation has second argument @i, fold to @i. */
850 static bool fold_xi_to_i(OptContext *ctx, TCGOp *op, uint64_t i)
851 {
852     if (arg_is_const(op->args[2]) && arg_info(op->args[2])->val == i) {
853         return tcg_opt_gen_movi(ctx, op, op->args[0], i);
854     }
855     return false;
856 }
857 
858 /* If the binary operation has second argument @i, fold to identity. */
859 static bool fold_xi_to_x(OptContext *ctx, TCGOp *op, uint64_t i)
860 {
861     if (arg_is_const(op->args[2]) && arg_info(op->args[2])->val == i) {
862         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
863     }
864     return false;
865 }
866 
867 /* If the binary operation has second argument @i, fold to NOT. */
868 static bool fold_xi_to_not(OptContext *ctx, TCGOp *op, uint64_t i)
869 {
870     if (arg_is_const(op->args[2]) && arg_info(op->args[2])->val == i) {
871         return fold_to_not(ctx, op, 1);
872     }
873     return false;
874 }
875 
876 /* If the binary operation has both arguments equal, fold to @i. */
877 static bool fold_xx_to_i(OptContext *ctx, TCGOp *op, uint64_t i)
878 {
879     if (args_are_copies(op->args[1], op->args[2])) {
880         return tcg_opt_gen_movi(ctx, op, op->args[0], i);
881     }
882     return false;
883 }
884 
885 /* If the binary operation has both arguments equal, fold to identity. */
886 static bool fold_xx_to_x(OptContext *ctx, TCGOp *op)
887 {
888     if (args_are_copies(op->args[1], op->args[2])) {
889         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
890     }
891     return false;
892 }
893 
894 /*
895  * These outermost fold_<op> functions are sorted alphabetically.
896  *
897  * The ordering of the transformations should be:
898  *   1) those that produce a constant
899  *   2) those that produce a copy
900  *   3) those that produce information about the result value.
901  */
902 
903 static bool fold_add(OptContext *ctx, TCGOp *op)
904 {
905     if (fold_const2_commutative(ctx, op) ||
906         fold_xi_to_x(ctx, op, 0)) {
907         return true;
908     }
909     return false;
910 }
911 
912 /* We cannot as yet do_constant_folding with vectors. */
913 static bool fold_add_vec(OptContext *ctx, TCGOp *op)
914 {
915     if (fold_commutative(ctx, op) ||
916         fold_xi_to_x(ctx, op, 0)) {
917         return true;
918     }
919     return false;
920 }
921 
922 static bool fold_addsub2(OptContext *ctx, TCGOp *op, bool add)
923 {
924     if (arg_is_const(op->args[2]) && arg_is_const(op->args[3]) &&
925         arg_is_const(op->args[4]) && arg_is_const(op->args[5])) {
926         uint64_t al = arg_info(op->args[2])->val;
927         uint64_t ah = arg_info(op->args[3])->val;
928         uint64_t bl = arg_info(op->args[4])->val;
929         uint64_t bh = arg_info(op->args[5])->val;
930         TCGArg rl, rh;
931         TCGOp *op2;
932 
933         if (ctx->type == TCG_TYPE_I32) {
934             uint64_t a = deposit64(al, 32, 32, ah);
935             uint64_t b = deposit64(bl, 32, 32, bh);
936 
937             if (add) {
938                 a += b;
939             } else {
940                 a -= b;
941             }
942 
943             al = sextract64(a, 0, 32);
944             ah = sextract64(a, 32, 32);
945         } else {
946             Int128 a = int128_make128(al, ah);
947             Int128 b = int128_make128(bl, bh);
948 
949             if (add) {
950                 a = int128_add(a, b);
951             } else {
952                 a = int128_sub(a, b);
953             }
954 
955             al = int128_getlo(a);
956             ah = int128_gethi(a);
957         }
958 
959         rl = op->args[0];
960         rh = op->args[1];
961 
962         /* The proper opcode is supplied by tcg_opt_gen_mov. */
963         op2 = tcg_op_insert_before(ctx->tcg, op, 0, 2);
964 
965         tcg_opt_gen_movi(ctx, op, rl, al);
966         tcg_opt_gen_movi(ctx, op2, rh, ah);
967         return true;
968     }
969     return false;
970 }
971 
972 static bool fold_add2(OptContext *ctx, TCGOp *op)
973 {
974     /* Note that the high and low parts may be independently swapped. */
975     swap_commutative(op->args[0], &op->args[2], &op->args[4]);
976     swap_commutative(op->args[1], &op->args[3], &op->args[5]);
977 
978     return fold_addsub2(ctx, op, true);
979 }
980 
981 static bool fold_and(OptContext *ctx, TCGOp *op)
982 {
983     uint64_t z1, z2;
984 
985     if (fold_const2_commutative(ctx, op) ||
986         fold_xi_to_i(ctx, op, 0) ||
987         fold_xi_to_x(ctx, op, -1) ||
988         fold_xx_to_x(ctx, op)) {
989         return true;
990     }
991 
992     z1 = arg_info(op->args[1])->z_mask;
993     z2 = arg_info(op->args[2])->z_mask;
994     ctx->z_mask = z1 & z2;
995 
996     /*
997      * Sign repetitions are perforce all identical, whether they are 1 or 0.
998      * Bitwise operations preserve the relative quantity of the repetitions.
999      */
1000     ctx->s_mask = arg_info(op->args[1])->s_mask
1001                 & arg_info(op->args[2])->s_mask;
1002 
1003     /*
1004      * Known-zeros does not imply known-ones.  Therefore unless
1005      * arg2 is constant, we can't infer affected bits from it.
1006      */
1007     if (arg_is_const(op->args[2])) {
1008         ctx->a_mask = z1 & ~z2;
1009     }
1010 
1011     return fold_masks(ctx, op);
1012 }
1013 
1014 static bool fold_andc(OptContext *ctx, TCGOp *op)
1015 {
1016     uint64_t z1;
1017 
1018     if (fold_const2(ctx, op) ||
1019         fold_xx_to_i(ctx, op, 0) ||
1020         fold_xi_to_x(ctx, op, 0) ||
1021         fold_ix_to_not(ctx, op, -1)) {
1022         return true;
1023     }
1024 
1025     z1 = arg_info(op->args[1])->z_mask;
1026 
1027     /*
1028      * Known-zeros does not imply known-ones.  Therefore unless
1029      * arg2 is constant, we can't infer anything from it.
1030      */
1031     if (arg_is_const(op->args[2])) {
1032         uint64_t z2 = ~arg_info(op->args[2])->z_mask;
1033         ctx->a_mask = z1 & ~z2;
1034         z1 &= z2;
1035     }
1036     ctx->z_mask = z1;
1037 
1038     ctx->s_mask = arg_info(op->args[1])->s_mask
1039                 & arg_info(op->args[2])->s_mask;
1040     return fold_masks(ctx, op);
1041 }
1042 
1043 static bool fold_brcond(OptContext *ctx, TCGOp *op)
1044 {
1045     TCGCond cond = op->args[2];
1046     int i;
1047 
1048     if (swap_commutative(NO_DEST, &op->args[0], &op->args[1])) {
1049         op->args[2] = cond = tcg_swap_cond(cond);
1050     }
1051 
1052     i = do_constant_folding_cond(ctx->type, op->args[0], op->args[1], cond);
1053     if (i == 0) {
1054         tcg_op_remove(ctx->tcg, op);
1055         return true;
1056     }
1057     if (i > 0) {
1058         op->opc = INDEX_op_br;
1059         op->args[0] = op->args[3];
1060     }
1061     return false;
1062 }
1063 
1064 static bool fold_brcond2(OptContext *ctx, TCGOp *op)
1065 {
1066     TCGCond cond = op->args[4];
1067     TCGArg label = op->args[5];
1068     int i, inv = 0;
1069 
1070     if (swap_commutative2(&op->args[0], &op->args[2])) {
1071         op->args[4] = cond = tcg_swap_cond(cond);
1072     }
1073 
1074     i = do_constant_folding_cond2(&op->args[0], &op->args[2], cond);
1075     if (i >= 0) {
1076         goto do_brcond_const;
1077     }
1078 
1079     switch (cond) {
1080     case TCG_COND_LT:
1081     case TCG_COND_GE:
1082         /*
1083          * Simplify LT/GE comparisons vs zero to a single compare
1084          * vs the high word of the input.
1085          */
1086         if (arg_is_const(op->args[2]) && arg_info(op->args[2])->val == 0 &&
1087             arg_is_const(op->args[3]) && arg_info(op->args[3])->val == 0) {
1088             goto do_brcond_high;
1089         }
1090         break;
1091 
1092     case TCG_COND_NE:
1093         inv = 1;
1094         QEMU_FALLTHROUGH;
1095     case TCG_COND_EQ:
1096         /*
1097          * Simplify EQ/NE comparisons where one of the pairs
1098          * can be simplified.
1099          */
1100         i = do_constant_folding_cond(TCG_TYPE_I32, op->args[0],
1101                                      op->args[2], cond);
1102         switch (i ^ inv) {
1103         case 0:
1104             goto do_brcond_const;
1105         case 1:
1106             goto do_brcond_high;
1107         }
1108 
1109         i = do_constant_folding_cond(TCG_TYPE_I32, op->args[1],
1110                                      op->args[3], cond);
1111         switch (i ^ inv) {
1112         case 0:
1113             goto do_brcond_const;
1114         case 1:
1115             op->opc = INDEX_op_brcond_i32;
1116             op->args[1] = op->args[2];
1117             op->args[2] = cond;
1118             op->args[3] = label;
1119             break;
1120         }
1121         break;
1122 
1123     default:
1124         break;
1125 
1126     do_brcond_high:
1127         op->opc = INDEX_op_brcond_i32;
1128         op->args[0] = op->args[1];
1129         op->args[1] = op->args[3];
1130         op->args[2] = cond;
1131         op->args[3] = label;
1132         break;
1133 
1134     do_brcond_const:
1135         if (i == 0) {
1136             tcg_op_remove(ctx->tcg, op);
1137             return true;
1138         }
1139         op->opc = INDEX_op_br;
1140         op->args[0] = label;
1141         break;
1142     }
1143     return false;
1144 }
1145 
1146 static bool fold_bswap(OptContext *ctx, TCGOp *op)
1147 {
1148     uint64_t z_mask, s_mask, sign;
1149 
1150     if (arg_is_const(op->args[1])) {
1151         uint64_t t = arg_info(op->args[1])->val;
1152 
1153         t = do_constant_folding(op->opc, ctx->type, t, op->args[2]);
1154         return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1155     }
1156 
1157     z_mask = arg_info(op->args[1])->z_mask;
1158 
1159     switch (op->opc) {
1160     case INDEX_op_bswap16_i32:
1161     case INDEX_op_bswap16_i64:
1162         z_mask = bswap16(z_mask);
1163         sign = INT16_MIN;
1164         break;
1165     case INDEX_op_bswap32_i32:
1166     case INDEX_op_bswap32_i64:
1167         z_mask = bswap32(z_mask);
1168         sign = INT32_MIN;
1169         break;
1170     case INDEX_op_bswap64_i64:
1171         z_mask = bswap64(z_mask);
1172         sign = INT64_MIN;
1173         break;
1174     default:
1175         g_assert_not_reached();
1176     }
1177     s_mask = smask_from_zmask(z_mask);
1178 
1179     switch (op->args[2] & (TCG_BSWAP_OZ | TCG_BSWAP_OS)) {
1180     case TCG_BSWAP_OZ:
1181         break;
1182     case TCG_BSWAP_OS:
1183         /* If the sign bit may be 1, force all the bits above to 1. */
1184         if (z_mask & sign) {
1185             z_mask |= sign;
1186             s_mask = sign << 1;
1187         }
1188         break;
1189     default:
1190         /* The high bits are undefined: force all bits above the sign to 1. */
1191         z_mask |= sign << 1;
1192         s_mask = 0;
1193         break;
1194     }
1195     ctx->z_mask = z_mask;
1196     ctx->s_mask = s_mask;
1197 
1198     return fold_masks(ctx, op);
1199 }
1200 
1201 static bool fold_call(OptContext *ctx, TCGOp *op)
1202 {
1203     TCGContext *s = ctx->tcg;
1204     int nb_oargs = TCGOP_CALLO(op);
1205     int nb_iargs = TCGOP_CALLI(op);
1206     int flags, i;
1207 
1208     init_arguments(ctx, op, nb_oargs + nb_iargs);
1209     copy_propagate(ctx, op, nb_oargs, nb_iargs);
1210 
1211     /* If the function reads or writes globals, reset temp data. */
1212     flags = tcg_call_flags(op);
1213     if (!(flags & (TCG_CALL_NO_READ_GLOBALS | TCG_CALL_NO_WRITE_GLOBALS))) {
1214         int nb_globals = s->nb_globals;
1215 
1216         for (i = 0; i < nb_globals; i++) {
1217             if (test_bit(i, ctx->temps_used.l)) {
1218                 reset_ts(&ctx->tcg->temps[i]);
1219             }
1220         }
1221     }
1222 
1223     /* Reset temp data for outputs. */
1224     for (i = 0; i < nb_oargs; i++) {
1225         reset_temp(op->args[i]);
1226     }
1227 
1228     /* Stop optimizing MB across calls. */
1229     ctx->prev_mb = NULL;
1230     return true;
1231 }
1232 
1233 static bool fold_count_zeros(OptContext *ctx, TCGOp *op)
1234 {
1235     uint64_t z_mask;
1236 
1237     if (arg_is_const(op->args[1])) {
1238         uint64_t t = arg_info(op->args[1])->val;
1239 
1240         if (t != 0) {
1241             t = do_constant_folding(op->opc, ctx->type, t, 0);
1242             return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1243         }
1244         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[2]);
1245     }
1246 
1247     switch (ctx->type) {
1248     case TCG_TYPE_I32:
1249         z_mask = 31;
1250         break;
1251     case TCG_TYPE_I64:
1252         z_mask = 63;
1253         break;
1254     default:
1255         g_assert_not_reached();
1256     }
1257     ctx->z_mask = arg_info(op->args[2])->z_mask | z_mask;
1258     ctx->s_mask = smask_from_zmask(ctx->z_mask);
1259     return false;
1260 }
1261 
1262 static bool fold_ctpop(OptContext *ctx, TCGOp *op)
1263 {
1264     if (fold_const1(ctx, op)) {
1265         return true;
1266     }
1267 
1268     switch (ctx->type) {
1269     case TCG_TYPE_I32:
1270         ctx->z_mask = 32 | 31;
1271         break;
1272     case TCG_TYPE_I64:
1273         ctx->z_mask = 64 | 63;
1274         break;
1275     default:
1276         g_assert_not_reached();
1277     }
1278     ctx->s_mask = smask_from_zmask(ctx->z_mask);
1279     return false;
1280 }
1281 
1282 static bool fold_deposit(OptContext *ctx, TCGOp *op)
1283 {
1284     if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) {
1285         uint64_t t1 = arg_info(op->args[1])->val;
1286         uint64_t t2 = arg_info(op->args[2])->val;
1287 
1288         t1 = deposit64(t1, op->args[3], op->args[4], t2);
1289         return tcg_opt_gen_movi(ctx, op, op->args[0], t1);
1290     }
1291 
1292     ctx->z_mask = deposit64(arg_info(op->args[1])->z_mask,
1293                             op->args[3], op->args[4],
1294                             arg_info(op->args[2])->z_mask);
1295     return false;
1296 }
1297 
1298 static bool fold_divide(OptContext *ctx, TCGOp *op)
1299 {
1300     if (fold_const2(ctx, op) ||
1301         fold_xi_to_x(ctx, op, 1)) {
1302         return true;
1303     }
1304     return false;
1305 }
1306 
1307 static bool fold_dup(OptContext *ctx, TCGOp *op)
1308 {
1309     if (arg_is_const(op->args[1])) {
1310         uint64_t t = arg_info(op->args[1])->val;
1311         t = dup_const(TCGOP_VECE(op), t);
1312         return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1313     }
1314     return false;
1315 }
1316 
1317 static bool fold_dup2(OptContext *ctx, TCGOp *op)
1318 {
1319     if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) {
1320         uint64_t t = deposit64(arg_info(op->args[1])->val, 32, 32,
1321                                arg_info(op->args[2])->val);
1322         return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1323     }
1324 
1325     if (args_are_copies(op->args[1], op->args[2])) {
1326         op->opc = INDEX_op_dup_vec;
1327         TCGOP_VECE(op) = MO_32;
1328     }
1329     return false;
1330 }
1331 
1332 static bool fold_eqv(OptContext *ctx, TCGOp *op)
1333 {
1334     if (fold_const2_commutative(ctx, op) ||
1335         fold_xi_to_x(ctx, op, -1) ||
1336         fold_xi_to_not(ctx, op, 0)) {
1337         return true;
1338     }
1339 
1340     ctx->s_mask = arg_info(op->args[1])->s_mask
1341                 & arg_info(op->args[2])->s_mask;
1342     return false;
1343 }
1344 
1345 static bool fold_extract(OptContext *ctx, TCGOp *op)
1346 {
1347     uint64_t z_mask_old, z_mask;
1348     int pos = op->args[2];
1349     int len = op->args[3];
1350 
1351     if (arg_is_const(op->args[1])) {
1352         uint64_t t;
1353 
1354         t = arg_info(op->args[1])->val;
1355         t = extract64(t, pos, len);
1356         return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1357     }
1358 
1359     z_mask_old = arg_info(op->args[1])->z_mask;
1360     z_mask = extract64(z_mask_old, pos, len);
1361     if (pos == 0) {
1362         ctx->a_mask = z_mask_old ^ z_mask;
1363     }
1364     ctx->z_mask = z_mask;
1365     ctx->s_mask = smask_from_zmask(z_mask);
1366 
1367     return fold_masks(ctx, op);
1368 }
1369 
1370 static bool fold_extract2(OptContext *ctx, TCGOp *op)
1371 {
1372     if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) {
1373         uint64_t v1 = arg_info(op->args[1])->val;
1374         uint64_t v2 = arg_info(op->args[2])->val;
1375         int shr = op->args[3];
1376 
1377         if (op->opc == INDEX_op_extract2_i64) {
1378             v1 >>= shr;
1379             v2 <<= 64 - shr;
1380         } else {
1381             v1 = (uint32_t)v1 >> shr;
1382             v2 = (uint64_t)((int32_t)v2 << (32 - shr));
1383         }
1384         return tcg_opt_gen_movi(ctx, op, op->args[0], v1 | v2);
1385     }
1386     return false;
1387 }
1388 
1389 static bool fold_exts(OptContext *ctx, TCGOp *op)
1390 {
1391     uint64_t s_mask_old, s_mask, z_mask, sign;
1392     bool type_change = false;
1393 
1394     if (fold_const1(ctx, op)) {
1395         return true;
1396     }
1397 
1398     z_mask = arg_info(op->args[1])->z_mask;
1399     s_mask = arg_info(op->args[1])->s_mask;
1400     s_mask_old = s_mask;
1401 
1402     switch (op->opc) {
1403     CASE_OP_32_64(ext8s):
1404         sign = INT8_MIN;
1405         z_mask = (uint8_t)z_mask;
1406         break;
1407     CASE_OP_32_64(ext16s):
1408         sign = INT16_MIN;
1409         z_mask = (uint16_t)z_mask;
1410         break;
1411     case INDEX_op_ext_i32_i64:
1412         type_change = true;
1413         QEMU_FALLTHROUGH;
1414     case INDEX_op_ext32s_i64:
1415         sign = INT32_MIN;
1416         z_mask = (uint32_t)z_mask;
1417         break;
1418     default:
1419         g_assert_not_reached();
1420     }
1421 
1422     if (z_mask & sign) {
1423         z_mask |= sign;
1424     }
1425     s_mask |= sign << 1;
1426 
1427     ctx->z_mask = z_mask;
1428     ctx->s_mask = s_mask;
1429     if (!type_change) {
1430         ctx->a_mask = s_mask & ~s_mask_old;
1431     }
1432 
1433     return fold_masks(ctx, op);
1434 }
1435 
1436 static bool fold_extu(OptContext *ctx, TCGOp *op)
1437 {
1438     uint64_t z_mask_old, z_mask;
1439     bool type_change = false;
1440 
1441     if (fold_const1(ctx, op)) {
1442         return true;
1443     }
1444 
1445     z_mask_old = z_mask = arg_info(op->args[1])->z_mask;
1446 
1447     switch (op->opc) {
1448     CASE_OP_32_64(ext8u):
1449         z_mask = (uint8_t)z_mask;
1450         break;
1451     CASE_OP_32_64(ext16u):
1452         z_mask = (uint16_t)z_mask;
1453         break;
1454     case INDEX_op_extrl_i64_i32:
1455     case INDEX_op_extu_i32_i64:
1456         type_change = true;
1457         QEMU_FALLTHROUGH;
1458     case INDEX_op_ext32u_i64:
1459         z_mask = (uint32_t)z_mask;
1460         break;
1461     case INDEX_op_extrh_i64_i32:
1462         type_change = true;
1463         z_mask >>= 32;
1464         break;
1465     default:
1466         g_assert_not_reached();
1467     }
1468 
1469     ctx->z_mask = z_mask;
1470     ctx->s_mask = smask_from_zmask(z_mask);
1471     if (!type_change) {
1472         ctx->a_mask = z_mask_old ^ z_mask;
1473     }
1474     return fold_masks(ctx, op);
1475 }
1476 
1477 static bool fold_mb(OptContext *ctx, TCGOp *op)
1478 {
1479     /* Eliminate duplicate and redundant fence instructions.  */
1480     if (ctx->prev_mb) {
1481         /*
1482          * Merge two barriers of the same type into one,
1483          * or a weaker barrier into a stronger one,
1484          * or two weaker barriers into a stronger one.
1485          *   mb X; mb Y => mb X|Y
1486          *   mb; strl => mb; st
1487          *   ldaq; mb => ld; mb
1488          *   ldaq; strl => ld; mb; st
1489          * Other combinations are also merged into a strong
1490          * barrier.  This is stricter than specified but for
1491          * the purposes of TCG is better than not optimizing.
1492          */
1493         ctx->prev_mb->args[0] |= op->args[0];
1494         tcg_op_remove(ctx->tcg, op);
1495     } else {
1496         ctx->prev_mb = op;
1497     }
1498     return true;
1499 }
1500 
1501 static bool fold_mov(OptContext *ctx, TCGOp *op)
1502 {
1503     return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
1504 }
1505 
1506 static bool fold_movcond(OptContext *ctx, TCGOp *op)
1507 {
1508     TCGCond cond = op->args[5];
1509     int i;
1510 
1511     if (swap_commutative(NO_DEST, &op->args[1], &op->args[2])) {
1512         op->args[5] = cond = tcg_swap_cond(cond);
1513     }
1514     /*
1515      * Canonicalize the "false" input reg to match the destination reg so
1516      * that the tcg backend can implement a "move if true" operation.
1517      */
1518     if (swap_commutative(op->args[0], &op->args[4], &op->args[3])) {
1519         op->args[5] = cond = tcg_invert_cond(cond);
1520     }
1521 
1522     i = do_constant_folding_cond(ctx->type, op->args[1], op->args[2], cond);
1523     if (i >= 0) {
1524         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[4 - i]);
1525     }
1526 
1527     ctx->z_mask = arg_info(op->args[3])->z_mask
1528                 | arg_info(op->args[4])->z_mask;
1529     ctx->s_mask = arg_info(op->args[3])->s_mask
1530                 & arg_info(op->args[4])->s_mask;
1531 
1532     if (arg_is_const(op->args[3]) && arg_is_const(op->args[4])) {
1533         uint64_t tv = arg_info(op->args[3])->val;
1534         uint64_t fv = arg_info(op->args[4])->val;
1535         TCGOpcode opc;
1536 
1537         switch (ctx->type) {
1538         case TCG_TYPE_I32:
1539             opc = INDEX_op_setcond_i32;
1540             break;
1541         case TCG_TYPE_I64:
1542             opc = INDEX_op_setcond_i64;
1543             break;
1544         default:
1545             g_assert_not_reached();
1546         }
1547 
1548         if (tv == 1 && fv == 0) {
1549             op->opc = opc;
1550             op->args[3] = cond;
1551         } else if (fv == 1 && tv == 0) {
1552             op->opc = opc;
1553             op->args[3] = tcg_invert_cond(cond);
1554         }
1555     }
1556     return false;
1557 }
1558 
1559 static bool fold_mul(OptContext *ctx, TCGOp *op)
1560 {
1561     if (fold_const2(ctx, op) ||
1562         fold_xi_to_i(ctx, op, 0) ||
1563         fold_xi_to_x(ctx, op, 1)) {
1564         return true;
1565     }
1566     return false;
1567 }
1568 
1569 static bool fold_mul_highpart(OptContext *ctx, TCGOp *op)
1570 {
1571     if (fold_const2_commutative(ctx, op) ||
1572         fold_xi_to_i(ctx, op, 0)) {
1573         return true;
1574     }
1575     return false;
1576 }
1577 
1578 static bool fold_multiply2(OptContext *ctx, TCGOp *op)
1579 {
1580     swap_commutative(op->args[0], &op->args[2], &op->args[3]);
1581 
1582     if (arg_is_const(op->args[2]) && arg_is_const(op->args[3])) {
1583         uint64_t a = arg_info(op->args[2])->val;
1584         uint64_t b = arg_info(op->args[3])->val;
1585         uint64_t h, l;
1586         TCGArg rl, rh;
1587         TCGOp *op2;
1588 
1589         switch (op->opc) {
1590         case INDEX_op_mulu2_i32:
1591             l = (uint64_t)(uint32_t)a * (uint32_t)b;
1592             h = (int32_t)(l >> 32);
1593             l = (int32_t)l;
1594             break;
1595         case INDEX_op_muls2_i32:
1596             l = (int64_t)(int32_t)a * (int32_t)b;
1597             h = l >> 32;
1598             l = (int32_t)l;
1599             break;
1600         case INDEX_op_mulu2_i64:
1601             mulu64(&l, &h, a, b);
1602             break;
1603         case INDEX_op_muls2_i64:
1604             muls64(&l, &h, a, b);
1605             break;
1606         default:
1607             g_assert_not_reached();
1608         }
1609 
1610         rl = op->args[0];
1611         rh = op->args[1];
1612 
1613         /* The proper opcode is supplied by tcg_opt_gen_mov. */
1614         op2 = tcg_op_insert_before(ctx->tcg, op, 0, 2);
1615 
1616         tcg_opt_gen_movi(ctx, op, rl, l);
1617         tcg_opt_gen_movi(ctx, op2, rh, h);
1618         return true;
1619     }
1620     return false;
1621 }
1622 
1623 static bool fold_nand(OptContext *ctx, TCGOp *op)
1624 {
1625     if (fold_const2_commutative(ctx, op) ||
1626         fold_xi_to_not(ctx, op, -1)) {
1627         return true;
1628     }
1629 
1630     ctx->s_mask = arg_info(op->args[1])->s_mask
1631                 & arg_info(op->args[2])->s_mask;
1632     return false;
1633 }
1634 
1635 static bool fold_neg(OptContext *ctx, TCGOp *op)
1636 {
1637     uint64_t z_mask;
1638 
1639     if (fold_const1(ctx, op)) {
1640         return true;
1641     }
1642 
1643     /* Set to 1 all bits to the left of the rightmost.  */
1644     z_mask = arg_info(op->args[1])->z_mask;
1645     ctx->z_mask = -(z_mask & -z_mask);
1646 
1647     /*
1648      * Because of fold_sub_to_neg, we want to always return true,
1649      * via finish_folding.
1650      */
1651     finish_folding(ctx, op);
1652     return true;
1653 }
1654 
1655 static bool fold_nor(OptContext *ctx, TCGOp *op)
1656 {
1657     if (fold_const2_commutative(ctx, op) ||
1658         fold_xi_to_not(ctx, op, 0)) {
1659         return true;
1660     }
1661 
1662     ctx->s_mask = arg_info(op->args[1])->s_mask
1663                 & arg_info(op->args[2])->s_mask;
1664     return false;
1665 }
1666 
1667 static bool fold_not(OptContext *ctx, TCGOp *op)
1668 {
1669     if (fold_const1(ctx, op)) {
1670         return true;
1671     }
1672 
1673     ctx->s_mask = arg_info(op->args[1])->s_mask;
1674 
1675     /* Because of fold_to_not, we want to always return true, via finish. */
1676     finish_folding(ctx, op);
1677     return true;
1678 }
1679 
1680 static bool fold_or(OptContext *ctx, TCGOp *op)
1681 {
1682     if (fold_const2_commutative(ctx, op) ||
1683         fold_xi_to_x(ctx, op, 0) ||
1684         fold_xx_to_x(ctx, op)) {
1685         return true;
1686     }
1687 
1688     ctx->z_mask = arg_info(op->args[1])->z_mask
1689                 | arg_info(op->args[2])->z_mask;
1690     ctx->s_mask = arg_info(op->args[1])->s_mask
1691                 & arg_info(op->args[2])->s_mask;
1692     return fold_masks(ctx, op);
1693 }
1694 
1695 static bool fold_orc(OptContext *ctx, TCGOp *op)
1696 {
1697     if (fold_const2(ctx, op) ||
1698         fold_xx_to_i(ctx, op, -1) ||
1699         fold_xi_to_x(ctx, op, -1) ||
1700         fold_ix_to_not(ctx, op, 0)) {
1701         return true;
1702     }
1703 
1704     ctx->s_mask = arg_info(op->args[1])->s_mask
1705                 & arg_info(op->args[2])->s_mask;
1706     return false;
1707 }
1708 
1709 static bool fold_qemu_ld(OptContext *ctx, TCGOp *op)
1710 {
1711     const TCGOpDef *def = &tcg_op_defs[op->opc];
1712     MemOpIdx oi = op->args[def->nb_oargs + def->nb_iargs];
1713     MemOp mop = get_memop(oi);
1714     int width = 8 * memop_size(mop);
1715 
1716     if (width < 64) {
1717         ctx->s_mask = MAKE_64BIT_MASK(width, 64 - width);
1718         if (!(mop & MO_SIGN)) {
1719             ctx->z_mask = MAKE_64BIT_MASK(0, width);
1720             ctx->s_mask <<= 1;
1721         }
1722     }
1723 
1724     /* Opcodes that touch guest memory stop the mb optimization.  */
1725     ctx->prev_mb = NULL;
1726     return false;
1727 }
1728 
1729 static bool fold_qemu_st(OptContext *ctx, TCGOp *op)
1730 {
1731     /* Opcodes that touch guest memory stop the mb optimization.  */
1732     ctx->prev_mb = NULL;
1733     return false;
1734 }
1735 
1736 static bool fold_remainder(OptContext *ctx, TCGOp *op)
1737 {
1738     if (fold_const2(ctx, op) ||
1739         fold_xx_to_i(ctx, op, 0)) {
1740         return true;
1741     }
1742     return false;
1743 }
1744 
1745 static bool fold_setcond(OptContext *ctx, TCGOp *op)
1746 {
1747     TCGCond cond = op->args[3];
1748     int i;
1749 
1750     if (swap_commutative(op->args[0], &op->args[1], &op->args[2])) {
1751         op->args[3] = cond = tcg_swap_cond(cond);
1752     }
1753 
1754     i = do_constant_folding_cond(ctx->type, op->args[1], op->args[2], cond);
1755     if (i >= 0) {
1756         return tcg_opt_gen_movi(ctx, op, op->args[0], i);
1757     }
1758 
1759     ctx->z_mask = 1;
1760     ctx->s_mask = smask_from_zmask(1);
1761     return false;
1762 }
1763 
1764 static bool fold_setcond2(OptContext *ctx, TCGOp *op)
1765 {
1766     TCGCond cond = op->args[5];
1767     int i, inv = 0;
1768 
1769     if (swap_commutative2(&op->args[1], &op->args[3])) {
1770         op->args[5] = cond = tcg_swap_cond(cond);
1771     }
1772 
1773     i = do_constant_folding_cond2(&op->args[1], &op->args[3], cond);
1774     if (i >= 0) {
1775         goto do_setcond_const;
1776     }
1777 
1778     switch (cond) {
1779     case TCG_COND_LT:
1780     case TCG_COND_GE:
1781         /*
1782          * Simplify LT/GE comparisons vs zero to a single compare
1783          * vs the high word of the input.
1784          */
1785         if (arg_is_const(op->args[3]) && arg_info(op->args[3])->val == 0 &&
1786             arg_is_const(op->args[4]) && arg_info(op->args[4])->val == 0) {
1787             goto do_setcond_high;
1788         }
1789         break;
1790 
1791     case TCG_COND_NE:
1792         inv = 1;
1793         QEMU_FALLTHROUGH;
1794     case TCG_COND_EQ:
1795         /*
1796          * Simplify EQ/NE comparisons where one of the pairs
1797          * can be simplified.
1798          */
1799         i = do_constant_folding_cond(TCG_TYPE_I32, op->args[1],
1800                                      op->args[3], cond);
1801         switch (i ^ inv) {
1802         case 0:
1803             goto do_setcond_const;
1804         case 1:
1805             goto do_setcond_high;
1806         }
1807 
1808         i = do_constant_folding_cond(TCG_TYPE_I32, op->args[2],
1809                                      op->args[4], cond);
1810         switch (i ^ inv) {
1811         case 0:
1812             goto do_setcond_const;
1813         case 1:
1814             op->args[2] = op->args[3];
1815             op->args[3] = cond;
1816             op->opc = INDEX_op_setcond_i32;
1817             break;
1818         }
1819         break;
1820 
1821     default:
1822         break;
1823 
1824     do_setcond_high:
1825         op->args[1] = op->args[2];
1826         op->args[2] = op->args[4];
1827         op->args[3] = cond;
1828         op->opc = INDEX_op_setcond_i32;
1829         break;
1830     }
1831 
1832     ctx->z_mask = 1;
1833     ctx->s_mask = smask_from_zmask(1);
1834     return false;
1835 
1836  do_setcond_const:
1837     return tcg_opt_gen_movi(ctx, op, op->args[0], i);
1838 }
1839 
1840 static bool fold_sextract(OptContext *ctx, TCGOp *op)
1841 {
1842     uint64_t z_mask, s_mask, s_mask_old;
1843     int pos = op->args[2];
1844     int len = op->args[3];
1845 
1846     if (arg_is_const(op->args[1])) {
1847         uint64_t t;
1848 
1849         t = arg_info(op->args[1])->val;
1850         t = sextract64(t, pos, len);
1851         return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1852     }
1853 
1854     z_mask = arg_info(op->args[1])->z_mask;
1855     z_mask = sextract64(z_mask, pos, len);
1856     ctx->z_mask = z_mask;
1857 
1858     s_mask_old = arg_info(op->args[1])->s_mask;
1859     s_mask = sextract64(s_mask_old, pos, len);
1860     s_mask |= MAKE_64BIT_MASK(len, 64 - len);
1861     ctx->s_mask = s_mask;
1862 
1863     if (pos == 0) {
1864         ctx->a_mask = s_mask & ~s_mask_old;
1865     }
1866 
1867     return fold_masks(ctx, op);
1868 }
1869 
1870 static bool fold_shift(OptContext *ctx, TCGOp *op)
1871 {
1872     uint64_t s_mask, z_mask, sign;
1873 
1874     if (fold_const2(ctx, op) ||
1875         fold_ix_to_i(ctx, op, 0) ||
1876         fold_xi_to_x(ctx, op, 0)) {
1877         return true;
1878     }
1879 
1880     s_mask = arg_info(op->args[1])->s_mask;
1881     z_mask = arg_info(op->args[1])->z_mask;
1882 
1883     if (arg_is_const(op->args[2])) {
1884         int sh = arg_info(op->args[2])->val;
1885 
1886         ctx->z_mask = do_constant_folding(op->opc, ctx->type, z_mask, sh);
1887 
1888         s_mask = do_constant_folding(op->opc, ctx->type, s_mask, sh);
1889         ctx->s_mask = smask_from_smask(s_mask);
1890 
1891         return fold_masks(ctx, op);
1892     }
1893 
1894     switch (op->opc) {
1895     CASE_OP_32_64(sar):
1896         /*
1897          * Arithmetic right shift will not reduce the number of
1898          * input sign repetitions.
1899          */
1900         ctx->s_mask = s_mask;
1901         break;
1902     CASE_OP_32_64(shr):
1903         /*
1904          * If the sign bit is known zero, then logical right shift
1905          * will not reduced the number of input sign repetitions.
1906          */
1907         sign = (s_mask & -s_mask) >> 1;
1908         if (!(z_mask & sign)) {
1909             ctx->s_mask = s_mask;
1910         }
1911         break;
1912     default:
1913         break;
1914     }
1915 
1916     return false;
1917 }
1918 
1919 static bool fold_sub_to_neg(OptContext *ctx, TCGOp *op)
1920 {
1921     TCGOpcode neg_op;
1922     bool have_neg;
1923 
1924     if (!arg_is_const(op->args[1]) || arg_info(op->args[1])->val != 0) {
1925         return false;
1926     }
1927 
1928     switch (ctx->type) {
1929     case TCG_TYPE_I32:
1930         neg_op = INDEX_op_neg_i32;
1931         have_neg = TCG_TARGET_HAS_neg_i32;
1932         break;
1933     case TCG_TYPE_I64:
1934         neg_op = INDEX_op_neg_i64;
1935         have_neg = TCG_TARGET_HAS_neg_i64;
1936         break;
1937     case TCG_TYPE_V64:
1938     case TCG_TYPE_V128:
1939     case TCG_TYPE_V256:
1940         neg_op = INDEX_op_neg_vec;
1941         have_neg = (TCG_TARGET_HAS_neg_vec &&
1942                     tcg_can_emit_vec_op(neg_op, ctx->type, TCGOP_VECE(op)) > 0);
1943         break;
1944     default:
1945         g_assert_not_reached();
1946     }
1947     if (have_neg) {
1948         op->opc = neg_op;
1949         op->args[1] = op->args[2];
1950         return fold_neg(ctx, op);
1951     }
1952     return false;
1953 }
1954 
1955 /* We cannot as yet do_constant_folding with vectors. */
1956 static bool fold_sub_vec(OptContext *ctx, TCGOp *op)
1957 {
1958     if (fold_xx_to_i(ctx, op, 0) ||
1959         fold_xi_to_x(ctx, op, 0) ||
1960         fold_sub_to_neg(ctx, op)) {
1961         return true;
1962     }
1963     return false;
1964 }
1965 
1966 static bool fold_sub(OptContext *ctx, TCGOp *op)
1967 {
1968     return fold_const2(ctx, op) || fold_sub_vec(ctx, op);
1969 }
1970 
1971 static bool fold_sub2(OptContext *ctx, TCGOp *op)
1972 {
1973     return fold_addsub2(ctx, op, false);
1974 }
1975 
1976 static bool fold_tcg_ld(OptContext *ctx, TCGOp *op)
1977 {
1978     /* We can't do any folding with a load, but we can record bits. */
1979     switch (op->opc) {
1980     CASE_OP_32_64(ld8s):
1981         ctx->s_mask = MAKE_64BIT_MASK(8, 56);
1982         break;
1983     CASE_OP_32_64(ld8u):
1984         ctx->z_mask = MAKE_64BIT_MASK(0, 8);
1985         ctx->s_mask = MAKE_64BIT_MASK(9, 55);
1986         break;
1987     CASE_OP_32_64(ld16s):
1988         ctx->s_mask = MAKE_64BIT_MASK(16, 48);
1989         break;
1990     CASE_OP_32_64(ld16u):
1991         ctx->z_mask = MAKE_64BIT_MASK(0, 16);
1992         ctx->s_mask = MAKE_64BIT_MASK(17, 47);
1993         break;
1994     case INDEX_op_ld32s_i64:
1995         ctx->s_mask = MAKE_64BIT_MASK(32, 32);
1996         break;
1997     case INDEX_op_ld32u_i64:
1998         ctx->z_mask = MAKE_64BIT_MASK(0, 32);
1999         ctx->s_mask = MAKE_64BIT_MASK(33, 31);
2000         break;
2001     default:
2002         g_assert_not_reached();
2003     }
2004     return false;
2005 }
2006 
2007 static bool fold_xor(OptContext *ctx, TCGOp *op)
2008 {
2009     if (fold_const2_commutative(ctx, op) ||
2010         fold_xx_to_i(ctx, op, 0) ||
2011         fold_xi_to_x(ctx, op, 0) ||
2012         fold_xi_to_not(ctx, op, -1)) {
2013         return true;
2014     }
2015 
2016     ctx->z_mask = arg_info(op->args[1])->z_mask
2017                 | arg_info(op->args[2])->z_mask;
2018     ctx->s_mask = arg_info(op->args[1])->s_mask
2019                 & arg_info(op->args[2])->s_mask;
2020     return fold_masks(ctx, op);
2021 }
2022 
2023 /* Propagate constants and copies, fold constant expressions. */
2024 void tcg_optimize(TCGContext *s)
2025 {
2026     int nb_temps, i;
2027     TCGOp *op, *op_next;
2028     OptContext ctx = { .tcg = s };
2029 
2030     /* Array VALS has an element for each temp.
2031        If this temp holds a constant then its value is kept in VALS' element.
2032        If this temp is a copy of other ones then the other copies are
2033        available through the doubly linked circular list. */
2034 
2035     nb_temps = s->nb_temps;
2036     for (i = 0; i < nb_temps; ++i) {
2037         s->temps[i].state_ptr = NULL;
2038     }
2039 
2040     QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) {
2041         TCGOpcode opc = op->opc;
2042         const TCGOpDef *def;
2043         bool done = false;
2044 
2045         /* Calls are special. */
2046         if (opc == INDEX_op_call) {
2047             fold_call(&ctx, op);
2048             continue;
2049         }
2050 
2051         def = &tcg_op_defs[opc];
2052         init_arguments(&ctx, op, def->nb_oargs + def->nb_iargs);
2053         copy_propagate(&ctx, op, def->nb_oargs, def->nb_iargs);
2054 
2055         /* Pre-compute the type of the operation. */
2056         if (def->flags & TCG_OPF_VECTOR) {
2057             ctx.type = TCG_TYPE_V64 + TCGOP_VECL(op);
2058         } else if (def->flags & TCG_OPF_64BIT) {
2059             ctx.type = TCG_TYPE_I64;
2060         } else {
2061             ctx.type = TCG_TYPE_I32;
2062         }
2063 
2064         /* Assume all bits affected, no bits known zero, no sign reps. */
2065         ctx.a_mask = -1;
2066         ctx.z_mask = -1;
2067         ctx.s_mask = 0;
2068 
2069         /*
2070          * Process each opcode.
2071          * Sorted alphabetically by opcode as much as possible.
2072          */
2073         switch (opc) {
2074         CASE_OP_32_64(add):
2075             done = fold_add(&ctx, op);
2076             break;
2077         case INDEX_op_add_vec:
2078             done = fold_add_vec(&ctx, op);
2079             break;
2080         CASE_OP_32_64(add2):
2081             done = fold_add2(&ctx, op);
2082             break;
2083         CASE_OP_32_64_VEC(and):
2084             done = fold_and(&ctx, op);
2085             break;
2086         CASE_OP_32_64_VEC(andc):
2087             done = fold_andc(&ctx, op);
2088             break;
2089         CASE_OP_32_64(brcond):
2090             done = fold_brcond(&ctx, op);
2091             break;
2092         case INDEX_op_brcond2_i32:
2093             done = fold_brcond2(&ctx, op);
2094             break;
2095         CASE_OP_32_64(bswap16):
2096         CASE_OP_32_64(bswap32):
2097         case INDEX_op_bswap64_i64:
2098             done = fold_bswap(&ctx, op);
2099             break;
2100         CASE_OP_32_64(clz):
2101         CASE_OP_32_64(ctz):
2102             done = fold_count_zeros(&ctx, op);
2103             break;
2104         CASE_OP_32_64(ctpop):
2105             done = fold_ctpop(&ctx, op);
2106             break;
2107         CASE_OP_32_64(deposit):
2108             done = fold_deposit(&ctx, op);
2109             break;
2110         CASE_OP_32_64(div):
2111         CASE_OP_32_64(divu):
2112             done = fold_divide(&ctx, op);
2113             break;
2114         case INDEX_op_dup_vec:
2115             done = fold_dup(&ctx, op);
2116             break;
2117         case INDEX_op_dup2_vec:
2118             done = fold_dup2(&ctx, op);
2119             break;
2120         CASE_OP_32_64_VEC(eqv):
2121             done = fold_eqv(&ctx, op);
2122             break;
2123         CASE_OP_32_64(extract):
2124             done = fold_extract(&ctx, op);
2125             break;
2126         CASE_OP_32_64(extract2):
2127             done = fold_extract2(&ctx, op);
2128             break;
2129         CASE_OP_32_64(ext8s):
2130         CASE_OP_32_64(ext16s):
2131         case INDEX_op_ext32s_i64:
2132         case INDEX_op_ext_i32_i64:
2133             done = fold_exts(&ctx, op);
2134             break;
2135         CASE_OP_32_64(ext8u):
2136         CASE_OP_32_64(ext16u):
2137         case INDEX_op_ext32u_i64:
2138         case INDEX_op_extu_i32_i64:
2139         case INDEX_op_extrl_i64_i32:
2140         case INDEX_op_extrh_i64_i32:
2141             done = fold_extu(&ctx, op);
2142             break;
2143         CASE_OP_32_64(ld8s):
2144         CASE_OP_32_64(ld8u):
2145         CASE_OP_32_64(ld16s):
2146         CASE_OP_32_64(ld16u):
2147         case INDEX_op_ld32s_i64:
2148         case INDEX_op_ld32u_i64:
2149             done = fold_tcg_ld(&ctx, op);
2150             break;
2151         case INDEX_op_mb:
2152             done = fold_mb(&ctx, op);
2153             break;
2154         CASE_OP_32_64_VEC(mov):
2155             done = fold_mov(&ctx, op);
2156             break;
2157         CASE_OP_32_64(movcond):
2158             done = fold_movcond(&ctx, op);
2159             break;
2160         CASE_OP_32_64(mul):
2161             done = fold_mul(&ctx, op);
2162             break;
2163         CASE_OP_32_64(mulsh):
2164         CASE_OP_32_64(muluh):
2165             done = fold_mul_highpart(&ctx, op);
2166             break;
2167         CASE_OP_32_64(muls2):
2168         CASE_OP_32_64(mulu2):
2169             done = fold_multiply2(&ctx, op);
2170             break;
2171         CASE_OP_32_64_VEC(nand):
2172             done = fold_nand(&ctx, op);
2173             break;
2174         CASE_OP_32_64(neg):
2175             done = fold_neg(&ctx, op);
2176             break;
2177         CASE_OP_32_64_VEC(nor):
2178             done = fold_nor(&ctx, op);
2179             break;
2180         CASE_OP_32_64_VEC(not):
2181             done = fold_not(&ctx, op);
2182             break;
2183         CASE_OP_32_64_VEC(or):
2184             done = fold_or(&ctx, op);
2185             break;
2186         CASE_OP_32_64_VEC(orc):
2187             done = fold_orc(&ctx, op);
2188             break;
2189         case INDEX_op_qemu_ld_i32:
2190         case INDEX_op_qemu_ld_i64:
2191             done = fold_qemu_ld(&ctx, op);
2192             break;
2193         case INDEX_op_qemu_st_i32:
2194         case INDEX_op_qemu_st8_i32:
2195         case INDEX_op_qemu_st_i64:
2196             done = fold_qemu_st(&ctx, op);
2197             break;
2198         CASE_OP_32_64(rem):
2199         CASE_OP_32_64(remu):
2200             done = fold_remainder(&ctx, op);
2201             break;
2202         CASE_OP_32_64(rotl):
2203         CASE_OP_32_64(rotr):
2204         CASE_OP_32_64(sar):
2205         CASE_OP_32_64(shl):
2206         CASE_OP_32_64(shr):
2207             done = fold_shift(&ctx, op);
2208             break;
2209         CASE_OP_32_64(setcond):
2210             done = fold_setcond(&ctx, op);
2211             break;
2212         case INDEX_op_setcond2_i32:
2213             done = fold_setcond2(&ctx, op);
2214             break;
2215         CASE_OP_32_64(sextract):
2216             done = fold_sextract(&ctx, op);
2217             break;
2218         CASE_OP_32_64(sub):
2219             done = fold_sub(&ctx, op);
2220             break;
2221         case INDEX_op_sub_vec:
2222             done = fold_sub_vec(&ctx, op);
2223             break;
2224         CASE_OP_32_64(sub2):
2225             done = fold_sub2(&ctx, op);
2226             break;
2227         CASE_OP_32_64_VEC(xor):
2228             done = fold_xor(&ctx, op);
2229             break;
2230         default:
2231             break;
2232         }
2233 
2234         if (!done) {
2235             finish_folding(&ctx, op);
2236         }
2237     }
2238 }
2239