xref: /openbmc/qemu/tcg/optimize.c (revision 961b80aecd1a503eedb885c309a1d5267d89c98c)
1 /*
2  * Optimizations for Tiny Code Generator for QEMU
3  *
4  * Copyright (c) 2010 Samsung Electronics.
5  * Contributed by Kirill Batuzov <batuzovk@ispras.ru>
6  *
7  * Permission is hereby granted, free of charge, to any person obtaining a copy
8  * of this software and associated documentation files (the "Software"), to deal
9  * in the Software without restriction, including without limitation the rights
10  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
11  * copies of the Software, and to permit persons to whom the Software is
12  * furnished to do so, subject to the following conditions:
13  *
14  * The above copyright notice and this permission notice shall be included in
15  * all copies or substantial portions of the Software.
16  *
17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
23  * THE SOFTWARE.
24  */
25 
26 #include "qemu/osdep.h"
27 #include "qemu/int128.h"
28 #include "qemu/interval-tree.h"
29 #include "tcg/tcg-op-common.h"
30 #include "tcg-internal.h"
31 #include "tcg-has.h"
32 
33 #define CASE_OP_32_64(x)                        \
34         glue(glue(case INDEX_op_, x), _i32):    \
35         glue(glue(case INDEX_op_, x), _i64)
36 
37 #define CASE_OP_32_64_VEC(x)                    \
38         glue(glue(case INDEX_op_, x), _i32):    \
39         glue(glue(case INDEX_op_, x), _i64):    \
40         glue(glue(case INDEX_op_, x), _vec)
41 
42 typedef struct MemCopyInfo {
43     IntervalTreeNode itree;
44     QSIMPLEQ_ENTRY (MemCopyInfo) next;
45     TCGTemp *ts;
46     TCGType type;
47 } MemCopyInfo;
48 
49 typedef struct TempOptInfo {
50     bool is_const;
51     TCGTemp *prev_copy;
52     TCGTemp *next_copy;
53     QSIMPLEQ_HEAD(, MemCopyInfo) mem_copy;
54     uint64_t val;
55     uint64_t z_mask;  /* mask bit is 0 if and only if value bit is 0 */
56     uint64_t s_mask;  /* mask bit is 1 if value bit matches msb */
57 } TempOptInfo;
58 
59 typedef struct OptContext {
60     TCGContext *tcg;
61     TCGOp *prev_mb;
62     TCGTempSet temps_used;
63 
64     IntervalTreeRoot mem_copy;
65     QSIMPLEQ_HEAD(, MemCopyInfo) mem_free;
66 
67     /* In flight values from optimization. */
68     TCGType type;
69 } OptContext;
70 
71 static inline TempOptInfo *ts_info(TCGTemp *ts)
72 {
73     return ts->state_ptr;
74 }
75 
76 static inline TempOptInfo *arg_info(TCGArg arg)
77 {
78     return ts_info(arg_temp(arg));
79 }
80 
81 static inline bool ti_is_const(TempOptInfo *ti)
82 {
83     return ti->is_const;
84 }
85 
86 static inline uint64_t ti_const_val(TempOptInfo *ti)
87 {
88     return ti->val;
89 }
90 
91 static inline bool ti_is_const_val(TempOptInfo *ti, uint64_t val)
92 {
93     return ti_is_const(ti) && ti_const_val(ti) == val;
94 }
95 
96 static inline bool ts_is_const(TCGTemp *ts)
97 {
98     return ti_is_const(ts_info(ts));
99 }
100 
101 static inline bool ts_is_const_val(TCGTemp *ts, uint64_t val)
102 {
103     return ti_is_const_val(ts_info(ts), val);
104 }
105 
106 static inline bool arg_is_const(TCGArg arg)
107 {
108     return ts_is_const(arg_temp(arg));
109 }
110 
111 static inline bool arg_is_const_val(TCGArg arg, uint64_t val)
112 {
113     return ts_is_const_val(arg_temp(arg), val);
114 }
115 
116 static inline bool ts_is_copy(TCGTemp *ts)
117 {
118     return ts_info(ts)->next_copy != ts;
119 }
120 
121 static TCGTemp *cmp_better_copy(TCGTemp *a, TCGTemp *b)
122 {
123     return a->kind < b->kind ? b : a;
124 }
125 
126 /* Initialize and activate a temporary.  */
127 static void init_ts_info(OptContext *ctx, TCGTemp *ts)
128 {
129     size_t idx = temp_idx(ts);
130     TempOptInfo *ti;
131 
132     if (test_bit(idx, ctx->temps_used.l)) {
133         return;
134     }
135     set_bit(idx, ctx->temps_used.l);
136 
137     ti = ts->state_ptr;
138     if (ti == NULL) {
139         ti = tcg_malloc(sizeof(TempOptInfo));
140         ts->state_ptr = ti;
141     }
142 
143     ti->next_copy = ts;
144     ti->prev_copy = ts;
145     QSIMPLEQ_INIT(&ti->mem_copy);
146     if (ts->kind == TEMP_CONST) {
147         ti->is_const = true;
148         ti->val = ts->val;
149         ti->z_mask = ts->val;
150         ti->s_mask = INT64_MIN >> clrsb64(ts->val);
151     } else {
152         ti->is_const = false;
153         ti->z_mask = -1;
154         ti->s_mask = 0;
155     }
156 }
157 
158 static MemCopyInfo *mem_copy_first(OptContext *ctx, intptr_t s, intptr_t l)
159 {
160     IntervalTreeNode *r = interval_tree_iter_first(&ctx->mem_copy, s, l);
161     return r ? container_of(r, MemCopyInfo, itree) : NULL;
162 }
163 
164 static MemCopyInfo *mem_copy_next(MemCopyInfo *mem, intptr_t s, intptr_t l)
165 {
166     IntervalTreeNode *r = interval_tree_iter_next(&mem->itree, s, l);
167     return r ? container_of(r, MemCopyInfo, itree) : NULL;
168 }
169 
170 static void remove_mem_copy(OptContext *ctx, MemCopyInfo *mc)
171 {
172     TCGTemp *ts = mc->ts;
173     TempOptInfo *ti = ts_info(ts);
174 
175     interval_tree_remove(&mc->itree, &ctx->mem_copy);
176     QSIMPLEQ_REMOVE(&ti->mem_copy, mc, MemCopyInfo, next);
177     QSIMPLEQ_INSERT_TAIL(&ctx->mem_free, mc, next);
178 }
179 
180 static void remove_mem_copy_in(OptContext *ctx, intptr_t s, intptr_t l)
181 {
182     while (true) {
183         MemCopyInfo *mc = mem_copy_first(ctx, s, l);
184         if (!mc) {
185             break;
186         }
187         remove_mem_copy(ctx, mc);
188     }
189 }
190 
191 static void remove_mem_copy_all(OptContext *ctx)
192 {
193     remove_mem_copy_in(ctx, 0, -1);
194     tcg_debug_assert(interval_tree_is_empty(&ctx->mem_copy));
195 }
196 
197 static TCGTemp *find_better_copy(TCGTemp *ts)
198 {
199     TCGTemp *i, *ret;
200 
201     /* If this is already readonly, we can't do better. */
202     if (temp_readonly(ts)) {
203         return ts;
204     }
205 
206     ret = ts;
207     for (i = ts_info(ts)->next_copy; i != ts; i = ts_info(i)->next_copy) {
208         ret = cmp_better_copy(ret, i);
209     }
210     return ret;
211 }
212 
213 static void move_mem_copies(TCGTemp *dst_ts, TCGTemp *src_ts)
214 {
215     TempOptInfo *si = ts_info(src_ts);
216     TempOptInfo *di = ts_info(dst_ts);
217     MemCopyInfo *mc;
218 
219     QSIMPLEQ_FOREACH(mc, &si->mem_copy, next) {
220         tcg_debug_assert(mc->ts == src_ts);
221         mc->ts = dst_ts;
222     }
223     QSIMPLEQ_CONCAT(&di->mem_copy, &si->mem_copy);
224 }
225 
226 /* Reset TEMP's state, possibly removing the temp for the list of copies.  */
227 static void reset_ts(OptContext *ctx, TCGTemp *ts)
228 {
229     TempOptInfo *ti = ts_info(ts);
230     TCGTemp *pts = ti->prev_copy;
231     TCGTemp *nts = ti->next_copy;
232     TempOptInfo *pi = ts_info(pts);
233     TempOptInfo *ni = ts_info(nts);
234 
235     ni->prev_copy = ti->prev_copy;
236     pi->next_copy = ti->next_copy;
237     ti->next_copy = ts;
238     ti->prev_copy = ts;
239     ti->is_const = false;
240     ti->z_mask = -1;
241     ti->s_mask = 0;
242 
243     if (!QSIMPLEQ_EMPTY(&ti->mem_copy)) {
244         if (ts == nts) {
245             /* Last temp copy being removed, the mem copies die. */
246             MemCopyInfo *mc;
247             QSIMPLEQ_FOREACH(mc, &ti->mem_copy, next) {
248                 interval_tree_remove(&mc->itree, &ctx->mem_copy);
249             }
250             QSIMPLEQ_CONCAT(&ctx->mem_free, &ti->mem_copy);
251         } else {
252             move_mem_copies(find_better_copy(nts), ts);
253         }
254     }
255 }
256 
257 static void reset_temp(OptContext *ctx, TCGArg arg)
258 {
259     reset_ts(ctx, arg_temp(arg));
260 }
261 
262 static void record_mem_copy(OptContext *ctx, TCGType type,
263                             TCGTemp *ts, intptr_t start, intptr_t last)
264 {
265     MemCopyInfo *mc;
266     TempOptInfo *ti;
267 
268     mc = QSIMPLEQ_FIRST(&ctx->mem_free);
269     if (mc) {
270         QSIMPLEQ_REMOVE_HEAD(&ctx->mem_free, next);
271     } else {
272         mc = tcg_malloc(sizeof(*mc));
273     }
274 
275     memset(mc, 0, sizeof(*mc));
276     mc->itree.start = start;
277     mc->itree.last = last;
278     mc->type = type;
279     interval_tree_insert(&mc->itree, &ctx->mem_copy);
280 
281     ts = find_better_copy(ts);
282     ti = ts_info(ts);
283     mc->ts = ts;
284     QSIMPLEQ_INSERT_TAIL(&ti->mem_copy, mc, next);
285 }
286 
287 static bool ts_are_copies(TCGTemp *ts1, TCGTemp *ts2)
288 {
289     TCGTemp *i;
290 
291     if (ts1 == ts2) {
292         return true;
293     }
294 
295     if (!ts_is_copy(ts1) || !ts_is_copy(ts2)) {
296         return false;
297     }
298 
299     for (i = ts_info(ts1)->next_copy; i != ts1; i = ts_info(i)->next_copy) {
300         if (i == ts2) {
301             return true;
302         }
303     }
304 
305     return false;
306 }
307 
308 static bool args_are_copies(TCGArg arg1, TCGArg arg2)
309 {
310     return ts_are_copies(arg_temp(arg1), arg_temp(arg2));
311 }
312 
313 static TCGTemp *find_mem_copy_for(OptContext *ctx, TCGType type, intptr_t s)
314 {
315     MemCopyInfo *mc;
316 
317     for (mc = mem_copy_first(ctx, s, s); mc; mc = mem_copy_next(mc, s, s)) {
318         if (mc->itree.start == s && mc->type == type) {
319             return find_better_copy(mc->ts);
320         }
321     }
322     return NULL;
323 }
324 
325 static TCGArg arg_new_constant(OptContext *ctx, uint64_t val)
326 {
327     TCGType type = ctx->type;
328     TCGTemp *ts;
329 
330     if (type == TCG_TYPE_I32) {
331         val = (int32_t)val;
332     }
333 
334     ts = tcg_constant_internal(type, val);
335     init_ts_info(ctx, ts);
336 
337     return temp_arg(ts);
338 }
339 
340 static TCGArg arg_new_temp(OptContext *ctx)
341 {
342     TCGTemp *ts = tcg_temp_new_internal(ctx->type, TEMP_EBB);
343     init_ts_info(ctx, ts);
344     return temp_arg(ts);
345 }
346 
347 static TCGOp *opt_insert_after(OptContext *ctx, TCGOp *op,
348                                TCGOpcode opc, unsigned narg)
349 {
350     return tcg_op_insert_after(ctx->tcg, op, opc, ctx->type, narg);
351 }
352 
353 static TCGOp *opt_insert_before(OptContext *ctx, TCGOp *op,
354                                 TCGOpcode opc, unsigned narg)
355 {
356     return tcg_op_insert_before(ctx->tcg, op, opc, ctx->type, narg);
357 }
358 
359 static bool tcg_opt_gen_mov(OptContext *ctx, TCGOp *op, TCGArg dst, TCGArg src)
360 {
361     TCGTemp *dst_ts = arg_temp(dst);
362     TCGTemp *src_ts = arg_temp(src);
363     TempOptInfo *di;
364     TempOptInfo *si;
365     TCGOpcode new_op;
366 
367     if (ts_are_copies(dst_ts, src_ts)) {
368         tcg_op_remove(ctx->tcg, op);
369         return true;
370     }
371 
372     reset_ts(ctx, dst_ts);
373     di = ts_info(dst_ts);
374     si = ts_info(src_ts);
375 
376     switch (ctx->type) {
377     case TCG_TYPE_I32:
378     case TCG_TYPE_I64:
379         new_op = INDEX_op_mov;
380         break;
381     case TCG_TYPE_V64:
382     case TCG_TYPE_V128:
383     case TCG_TYPE_V256:
384         /* TCGOP_TYPE and TCGOP_VECE remain unchanged.  */
385         new_op = INDEX_op_mov_vec;
386         break;
387     default:
388         g_assert_not_reached();
389     }
390     op->opc = new_op;
391     op->args[0] = dst;
392     op->args[1] = src;
393 
394     di->z_mask = si->z_mask;
395     di->s_mask = si->s_mask;
396 
397     if (src_ts->type == dst_ts->type) {
398         TempOptInfo *ni = ts_info(si->next_copy);
399 
400         di->next_copy = si->next_copy;
401         di->prev_copy = src_ts;
402         ni->prev_copy = dst_ts;
403         si->next_copy = dst_ts;
404         di->is_const = si->is_const;
405         di->val = si->val;
406 
407         if (!QSIMPLEQ_EMPTY(&si->mem_copy)
408             && cmp_better_copy(src_ts, dst_ts) == dst_ts) {
409             move_mem_copies(dst_ts, src_ts);
410         }
411     }
412     return true;
413 }
414 
415 static bool tcg_opt_gen_movi(OptContext *ctx, TCGOp *op,
416                              TCGArg dst, uint64_t val)
417 {
418     /* Convert movi to mov with constant temp. */
419     return tcg_opt_gen_mov(ctx, op, dst, arg_new_constant(ctx, val));
420 }
421 
422 static uint64_t do_constant_folding_2(TCGOpcode op, TCGType type,
423                                       uint64_t x, uint64_t y)
424 {
425     uint64_t l64, h64;
426 
427     switch (op) {
428     case INDEX_op_add:
429         return x + y;
430 
431     case INDEX_op_sub:
432         return x - y;
433 
434     case INDEX_op_mul:
435         return x * y;
436 
437     case INDEX_op_and:
438     case INDEX_op_and_vec:
439         return x & y;
440 
441     case INDEX_op_or:
442     case INDEX_op_or_vec:
443         return x | y;
444 
445     case INDEX_op_xor:
446     case INDEX_op_xor_vec:
447         return x ^ y;
448 
449     case INDEX_op_shl_i32:
450         return (uint32_t)x << (y & 31);
451 
452     case INDEX_op_shl_i64:
453         return (uint64_t)x << (y & 63);
454 
455     case INDEX_op_shr_i32:
456         return (uint32_t)x >> (y & 31);
457 
458     case INDEX_op_shr_i64:
459         return (uint64_t)x >> (y & 63);
460 
461     case INDEX_op_sar_i32:
462         return (int32_t)x >> (y & 31);
463 
464     case INDEX_op_sar_i64:
465         return (int64_t)x >> (y & 63);
466 
467     case INDEX_op_rotr_i32:
468         return ror32(x, y & 31);
469 
470     case INDEX_op_rotr_i64:
471         return ror64(x, y & 63);
472 
473     case INDEX_op_rotl_i32:
474         return rol32(x, y & 31);
475 
476     case INDEX_op_rotl_i64:
477         return rol64(x, y & 63);
478 
479     case INDEX_op_not:
480     case INDEX_op_not_vec:
481         return ~x;
482 
483     case INDEX_op_neg:
484         return -x;
485 
486     case INDEX_op_andc:
487     case INDEX_op_andc_vec:
488         return x & ~y;
489 
490     case INDEX_op_orc:
491     case INDEX_op_orc_vec:
492         return x | ~y;
493 
494     case INDEX_op_eqv:
495     case INDEX_op_eqv_vec:
496         return ~(x ^ y);
497 
498     case INDEX_op_nand:
499     case INDEX_op_nand_vec:
500         return ~(x & y);
501 
502     case INDEX_op_nor:
503     case INDEX_op_nor_vec:
504         return ~(x | y);
505 
506     case INDEX_op_clz_i32:
507         return (uint32_t)x ? clz32(x) : y;
508 
509     case INDEX_op_clz_i64:
510         return x ? clz64(x) : y;
511 
512     case INDEX_op_ctz_i32:
513         return (uint32_t)x ? ctz32(x) : y;
514 
515     case INDEX_op_ctz_i64:
516         return x ? ctz64(x) : y;
517 
518     case INDEX_op_ctpop_i32:
519         return ctpop32(x);
520 
521     case INDEX_op_ctpop_i64:
522         return ctpop64(x);
523 
524     CASE_OP_32_64(bswap16):
525         x = bswap16(x);
526         return y & TCG_BSWAP_OS ? (int16_t)x : x;
527 
528     CASE_OP_32_64(bswap32):
529         x = bswap32(x);
530         return y & TCG_BSWAP_OS ? (int32_t)x : x;
531 
532     case INDEX_op_bswap64_i64:
533         return bswap64(x);
534 
535     case INDEX_op_ext_i32_i64:
536         return (int32_t)x;
537 
538     case INDEX_op_extu_i32_i64:
539     case INDEX_op_extrl_i64_i32:
540         return (uint32_t)x;
541 
542     case INDEX_op_extrh_i64_i32:
543         return (uint64_t)x >> 32;
544 
545     case INDEX_op_muluh:
546         if (type == TCG_TYPE_I32) {
547             return ((uint64_t)(uint32_t)x * (uint32_t)y) >> 32;
548         }
549         mulu64(&l64, &h64, x, y);
550         return h64;
551 
552     case INDEX_op_mulsh:
553         if (type == TCG_TYPE_I32) {
554             return ((int64_t)(int32_t)x * (int32_t)y) >> 32;
555         }
556         muls64(&l64, &h64, x, y);
557         return h64;
558 
559     case INDEX_op_divs:
560         /* Avoid crashing on divide by zero, otherwise undefined.  */
561         if (type == TCG_TYPE_I32) {
562             return (int32_t)x / ((int32_t)y ? : 1);
563         }
564         return (int64_t)x / ((int64_t)y ? : 1);
565 
566     case INDEX_op_divu:
567         if (type == TCG_TYPE_I32) {
568             return (uint32_t)x / ((uint32_t)y ? : 1);
569         }
570         return (uint64_t)x / ((uint64_t)y ? : 1);
571 
572     case INDEX_op_rem_i32:
573         return (int32_t)x % ((int32_t)y ? : 1);
574     case INDEX_op_remu_i32:
575         return (uint32_t)x % ((uint32_t)y ? : 1);
576     case INDEX_op_rem_i64:
577         return (int64_t)x % ((int64_t)y ? : 1);
578     case INDEX_op_remu_i64:
579         return (uint64_t)x % ((uint64_t)y ? : 1);
580 
581     default:
582         g_assert_not_reached();
583     }
584 }
585 
586 static uint64_t do_constant_folding(TCGOpcode op, TCGType type,
587                                     uint64_t x, uint64_t y)
588 {
589     uint64_t res = do_constant_folding_2(op, type, x, y);
590     if (type == TCG_TYPE_I32) {
591         res = (int32_t)res;
592     }
593     return res;
594 }
595 
596 static bool do_constant_folding_cond_32(uint32_t x, uint32_t y, TCGCond c)
597 {
598     switch (c) {
599     case TCG_COND_EQ:
600         return x == y;
601     case TCG_COND_NE:
602         return x != y;
603     case TCG_COND_LT:
604         return (int32_t)x < (int32_t)y;
605     case TCG_COND_GE:
606         return (int32_t)x >= (int32_t)y;
607     case TCG_COND_LE:
608         return (int32_t)x <= (int32_t)y;
609     case TCG_COND_GT:
610         return (int32_t)x > (int32_t)y;
611     case TCG_COND_LTU:
612         return x < y;
613     case TCG_COND_GEU:
614         return x >= y;
615     case TCG_COND_LEU:
616         return x <= y;
617     case TCG_COND_GTU:
618         return x > y;
619     case TCG_COND_TSTEQ:
620         return (x & y) == 0;
621     case TCG_COND_TSTNE:
622         return (x & y) != 0;
623     case TCG_COND_ALWAYS:
624     case TCG_COND_NEVER:
625         break;
626     }
627     g_assert_not_reached();
628 }
629 
630 static bool do_constant_folding_cond_64(uint64_t x, uint64_t y, TCGCond c)
631 {
632     switch (c) {
633     case TCG_COND_EQ:
634         return x == y;
635     case TCG_COND_NE:
636         return x != y;
637     case TCG_COND_LT:
638         return (int64_t)x < (int64_t)y;
639     case TCG_COND_GE:
640         return (int64_t)x >= (int64_t)y;
641     case TCG_COND_LE:
642         return (int64_t)x <= (int64_t)y;
643     case TCG_COND_GT:
644         return (int64_t)x > (int64_t)y;
645     case TCG_COND_LTU:
646         return x < y;
647     case TCG_COND_GEU:
648         return x >= y;
649     case TCG_COND_LEU:
650         return x <= y;
651     case TCG_COND_GTU:
652         return x > y;
653     case TCG_COND_TSTEQ:
654         return (x & y) == 0;
655     case TCG_COND_TSTNE:
656         return (x & y) != 0;
657     case TCG_COND_ALWAYS:
658     case TCG_COND_NEVER:
659         break;
660     }
661     g_assert_not_reached();
662 }
663 
664 static int do_constant_folding_cond_eq(TCGCond c)
665 {
666     switch (c) {
667     case TCG_COND_GT:
668     case TCG_COND_LTU:
669     case TCG_COND_LT:
670     case TCG_COND_GTU:
671     case TCG_COND_NE:
672         return 0;
673     case TCG_COND_GE:
674     case TCG_COND_GEU:
675     case TCG_COND_LE:
676     case TCG_COND_LEU:
677     case TCG_COND_EQ:
678         return 1;
679     case TCG_COND_TSTEQ:
680     case TCG_COND_TSTNE:
681         return -1;
682     case TCG_COND_ALWAYS:
683     case TCG_COND_NEVER:
684         break;
685     }
686     g_assert_not_reached();
687 }
688 
689 /*
690  * Return -1 if the condition can't be simplified,
691  * and the result of the condition (0 or 1) if it can.
692  */
693 static int do_constant_folding_cond(TCGType type, TCGArg x,
694                                     TCGArg y, TCGCond c)
695 {
696     if (arg_is_const(x) && arg_is_const(y)) {
697         uint64_t xv = arg_info(x)->val;
698         uint64_t yv = arg_info(y)->val;
699 
700         switch (type) {
701         case TCG_TYPE_I32:
702             return do_constant_folding_cond_32(xv, yv, c);
703         case TCG_TYPE_I64:
704             return do_constant_folding_cond_64(xv, yv, c);
705         default:
706             /* Only scalar comparisons are optimizable */
707             return -1;
708         }
709     } else if (args_are_copies(x, y)) {
710         return do_constant_folding_cond_eq(c);
711     } else if (arg_is_const_val(y, 0)) {
712         switch (c) {
713         case TCG_COND_LTU:
714         case TCG_COND_TSTNE:
715             return 0;
716         case TCG_COND_GEU:
717         case TCG_COND_TSTEQ:
718             return 1;
719         default:
720             return -1;
721         }
722     }
723     return -1;
724 }
725 
726 /**
727  * swap_commutative:
728  * @dest: TCGArg of the destination argument, or NO_DEST.
729  * @p1: first paired argument
730  * @p2: second paired argument
731  *
732  * If *@p1 is a constant and *@p2 is not, swap.
733  * If *@p2 matches @dest, swap.
734  * Return true if a swap was performed.
735  */
736 
737 #define NO_DEST  temp_arg(NULL)
738 
739 static bool swap_commutative(TCGArg dest, TCGArg *p1, TCGArg *p2)
740 {
741     TCGArg a1 = *p1, a2 = *p2;
742     int sum = 0;
743     sum += arg_is_const(a1);
744     sum -= arg_is_const(a2);
745 
746     /* Prefer the constant in second argument, and then the form
747        op a, a, b, which is better handled on non-RISC hosts. */
748     if (sum > 0 || (sum == 0 && dest == a2)) {
749         *p1 = a2;
750         *p2 = a1;
751         return true;
752     }
753     return false;
754 }
755 
756 static bool swap_commutative2(TCGArg *p1, TCGArg *p2)
757 {
758     int sum = 0;
759     sum += arg_is_const(p1[0]);
760     sum += arg_is_const(p1[1]);
761     sum -= arg_is_const(p2[0]);
762     sum -= arg_is_const(p2[1]);
763     if (sum > 0) {
764         TCGArg t;
765         t = p1[0], p1[0] = p2[0], p2[0] = t;
766         t = p1[1], p1[1] = p2[1], p2[1] = t;
767         return true;
768     }
769     return false;
770 }
771 
772 /*
773  * Return -1 if the condition can't be simplified,
774  * and the result of the condition (0 or 1) if it can.
775  */
776 static int do_constant_folding_cond1(OptContext *ctx, TCGOp *op, TCGArg dest,
777                                      TCGArg *p1, TCGArg *p2, TCGArg *pcond)
778 {
779     TCGCond cond;
780     TempOptInfo *i1;
781     bool swap;
782     int r;
783 
784     swap = swap_commutative(dest, p1, p2);
785     cond = *pcond;
786     if (swap) {
787         *pcond = cond = tcg_swap_cond(cond);
788     }
789 
790     r = do_constant_folding_cond(ctx->type, *p1, *p2, cond);
791     if (r >= 0) {
792         return r;
793     }
794     if (!is_tst_cond(cond)) {
795         return -1;
796     }
797 
798     i1 = arg_info(*p1);
799 
800     /*
801      * TSTNE x,x -> NE x,0
802      * TSTNE x,i -> NE x,0 if i includes all nonzero bits of x
803      */
804     if (args_are_copies(*p1, *p2) ||
805         (arg_is_const(*p2) && (i1->z_mask & ~arg_info(*p2)->val) == 0)) {
806         *p2 = arg_new_constant(ctx, 0);
807         *pcond = tcg_tst_eqne_cond(cond);
808         return -1;
809     }
810 
811     /* TSTNE x,i -> LT x,0 if i only includes sign bit copies */
812     if (arg_is_const(*p2) && (arg_info(*p2)->val & ~i1->s_mask) == 0) {
813         *p2 = arg_new_constant(ctx, 0);
814         *pcond = tcg_tst_ltge_cond(cond);
815         return -1;
816     }
817 
818     /* Expand to AND with a temporary if no backend support. */
819     if (!TCG_TARGET_HAS_tst) {
820         TCGOp *op2 = opt_insert_before(ctx, op, INDEX_op_and, 3);
821         TCGArg tmp = arg_new_temp(ctx);
822 
823         op2->args[0] = tmp;
824         op2->args[1] = *p1;
825         op2->args[2] = *p2;
826 
827         *p1 = tmp;
828         *p2 = arg_new_constant(ctx, 0);
829         *pcond = tcg_tst_eqne_cond(cond);
830     }
831     return -1;
832 }
833 
834 static int do_constant_folding_cond2(OptContext *ctx, TCGOp *op, TCGArg *args)
835 {
836     TCGArg al, ah, bl, bh;
837     TCGCond c;
838     bool swap;
839     int r;
840 
841     swap = swap_commutative2(args, args + 2);
842     c = args[4];
843     if (swap) {
844         args[4] = c = tcg_swap_cond(c);
845     }
846 
847     al = args[0];
848     ah = args[1];
849     bl = args[2];
850     bh = args[3];
851 
852     if (arg_is_const(bl) && arg_is_const(bh)) {
853         tcg_target_ulong blv = arg_info(bl)->val;
854         tcg_target_ulong bhv = arg_info(bh)->val;
855         uint64_t b = deposit64(blv, 32, 32, bhv);
856 
857         if (arg_is_const(al) && arg_is_const(ah)) {
858             tcg_target_ulong alv = arg_info(al)->val;
859             tcg_target_ulong ahv = arg_info(ah)->val;
860             uint64_t a = deposit64(alv, 32, 32, ahv);
861 
862             r = do_constant_folding_cond_64(a, b, c);
863             if (r >= 0) {
864                 return r;
865             }
866         }
867 
868         if (b == 0) {
869             switch (c) {
870             case TCG_COND_LTU:
871             case TCG_COND_TSTNE:
872                 return 0;
873             case TCG_COND_GEU:
874             case TCG_COND_TSTEQ:
875                 return 1;
876             default:
877                 break;
878             }
879         }
880 
881         /* TSTNE x,-1 -> NE x,0 */
882         if (b == -1 && is_tst_cond(c)) {
883             args[3] = args[2] = arg_new_constant(ctx, 0);
884             args[4] = tcg_tst_eqne_cond(c);
885             return -1;
886         }
887 
888         /* TSTNE x,sign -> LT x,0 */
889         if (b == INT64_MIN && is_tst_cond(c)) {
890             /* bl must be 0, so copy that to bh */
891             args[3] = bl;
892             args[4] = tcg_tst_ltge_cond(c);
893             return -1;
894         }
895     }
896 
897     if (args_are_copies(al, bl) && args_are_copies(ah, bh)) {
898         r = do_constant_folding_cond_eq(c);
899         if (r >= 0) {
900             return r;
901         }
902 
903         /* TSTNE x,x -> NE x,0 */
904         if (is_tst_cond(c)) {
905             args[3] = args[2] = arg_new_constant(ctx, 0);
906             args[4] = tcg_tst_eqne_cond(c);
907             return -1;
908         }
909     }
910 
911     /* Expand to AND with a temporary if no backend support. */
912     if (!TCG_TARGET_HAS_tst && is_tst_cond(c)) {
913         TCGOp *op1 = opt_insert_before(ctx, op, INDEX_op_and, 3);
914         TCGOp *op2 = opt_insert_before(ctx, op, INDEX_op_and, 3);
915         TCGArg t1 = arg_new_temp(ctx);
916         TCGArg t2 = arg_new_temp(ctx);
917 
918         op1->args[0] = t1;
919         op1->args[1] = al;
920         op1->args[2] = bl;
921         op2->args[0] = t2;
922         op2->args[1] = ah;
923         op2->args[2] = bh;
924 
925         args[0] = t1;
926         args[1] = t2;
927         args[3] = args[2] = arg_new_constant(ctx, 0);
928         args[4] = tcg_tst_eqne_cond(c);
929     }
930     return -1;
931 }
932 
933 static void init_arguments(OptContext *ctx, TCGOp *op, int nb_args)
934 {
935     for (int i = 0; i < nb_args; i++) {
936         TCGTemp *ts = arg_temp(op->args[i]);
937         init_ts_info(ctx, ts);
938     }
939 }
940 
941 static void copy_propagate(OptContext *ctx, TCGOp *op,
942                            int nb_oargs, int nb_iargs)
943 {
944     for (int i = nb_oargs; i < nb_oargs + nb_iargs; i++) {
945         TCGTemp *ts = arg_temp(op->args[i]);
946         if (ts_is_copy(ts)) {
947             op->args[i] = temp_arg(find_better_copy(ts));
948         }
949     }
950 }
951 
952 static void finish_bb(OptContext *ctx)
953 {
954     /* We only optimize memory barriers across basic blocks. */
955     ctx->prev_mb = NULL;
956 }
957 
958 static void finish_ebb(OptContext *ctx)
959 {
960     finish_bb(ctx);
961     /* We only optimize across extended basic blocks. */
962     memset(&ctx->temps_used, 0, sizeof(ctx->temps_used));
963     remove_mem_copy_all(ctx);
964 }
965 
966 static bool finish_folding(OptContext *ctx, TCGOp *op)
967 {
968     const TCGOpDef *def = &tcg_op_defs[op->opc];
969     int i, nb_oargs;
970 
971     nb_oargs = def->nb_oargs;
972     for (i = 0; i < nb_oargs; i++) {
973         TCGTemp *ts = arg_temp(op->args[i]);
974         reset_ts(ctx, ts);
975     }
976     return true;
977 }
978 
979 /*
980  * The fold_* functions return true when processing is complete,
981  * usually by folding the operation to a constant or to a copy,
982  * and calling tcg_opt_gen_{mov,movi}.  They may do other things,
983  * like collect information about the value produced, for use in
984  * optimizing a subsequent operation.
985  *
986  * These first fold_* functions are all helpers, used by other
987  * folders for more specific operations.
988  */
989 
990 static bool fold_const1(OptContext *ctx, TCGOp *op)
991 {
992     if (arg_is_const(op->args[1])) {
993         uint64_t t;
994 
995         t = arg_info(op->args[1])->val;
996         t = do_constant_folding(op->opc, ctx->type, t, 0);
997         return tcg_opt_gen_movi(ctx, op, op->args[0], t);
998     }
999     return false;
1000 }
1001 
1002 static bool fold_const2(OptContext *ctx, TCGOp *op)
1003 {
1004     if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) {
1005         uint64_t t1 = arg_info(op->args[1])->val;
1006         uint64_t t2 = arg_info(op->args[2])->val;
1007 
1008         t1 = do_constant_folding(op->opc, ctx->type, t1, t2);
1009         return tcg_opt_gen_movi(ctx, op, op->args[0], t1);
1010     }
1011     return false;
1012 }
1013 
1014 static bool fold_commutative(OptContext *ctx, TCGOp *op)
1015 {
1016     swap_commutative(op->args[0], &op->args[1], &op->args[2]);
1017     return false;
1018 }
1019 
1020 static bool fold_const2_commutative(OptContext *ctx, TCGOp *op)
1021 {
1022     swap_commutative(op->args[0], &op->args[1], &op->args[2]);
1023     return fold_const2(ctx, op);
1024 }
1025 
1026 /*
1027  * Record "zero" and "sign" masks for the single output of @op.
1028  * See TempOptInfo definition of z_mask and s_mask.
1029  * If z_mask allows, fold the output to constant zero.
1030  * The passed s_mask may be augmented by z_mask.
1031  */
1032 static bool fold_masks_zs(OptContext *ctx, TCGOp *op,
1033                           uint64_t z_mask, int64_t s_mask)
1034 {
1035     const TCGOpDef *def = &tcg_op_defs[op->opc];
1036     TCGTemp *ts;
1037     TempOptInfo *ti;
1038     int rep;
1039 
1040     /* Only single-output opcodes are supported here. */
1041     tcg_debug_assert(def->nb_oargs == 1);
1042 
1043     /*
1044      * 32-bit ops generate 32-bit results, which for the purpose of
1045      * simplifying tcg are sign-extended.  Certainly that's how we
1046      * represent our constants elsewhere.  Note that the bits will
1047      * be reset properly for a 64-bit value when encountering the
1048      * type changing opcodes.
1049      */
1050     if (ctx->type == TCG_TYPE_I32) {
1051         z_mask = (int32_t)z_mask;
1052         s_mask |= INT32_MIN;
1053     }
1054 
1055     if (z_mask == 0) {
1056         return tcg_opt_gen_movi(ctx, op, op->args[0], 0);
1057     }
1058 
1059     ts = arg_temp(op->args[0]);
1060     reset_ts(ctx, ts);
1061 
1062     ti = ts_info(ts);
1063     ti->z_mask = z_mask;
1064 
1065     /* Canonicalize s_mask and incorporate data from z_mask. */
1066     rep = clz64(~s_mask);
1067     rep = MAX(rep, clz64(z_mask));
1068     rep = MAX(rep - 1, 0);
1069     ti->s_mask = INT64_MIN >> rep;
1070 
1071     return true;
1072 }
1073 
1074 static bool fold_masks_z(OptContext *ctx, TCGOp *op, uint64_t z_mask)
1075 {
1076     return fold_masks_zs(ctx, op, z_mask, 0);
1077 }
1078 
1079 static bool fold_masks_s(OptContext *ctx, TCGOp *op, uint64_t s_mask)
1080 {
1081     return fold_masks_zs(ctx, op, -1, s_mask);
1082 }
1083 
1084 /*
1085  * An "affected" mask bit is 0 if and only if the result is identical
1086  * to the first input.  Thus if the entire mask is 0, the operation
1087  * is equivalent to a copy.
1088  */
1089 static bool fold_affected_mask(OptContext *ctx, TCGOp *op, uint64_t a_mask)
1090 {
1091     if (ctx->type == TCG_TYPE_I32) {
1092         a_mask = (uint32_t)a_mask;
1093     }
1094     if (a_mask == 0) {
1095         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
1096     }
1097     return false;
1098 }
1099 
1100 /*
1101  * Convert @op to NOT, if NOT is supported by the host.
1102  * Return true f the conversion is successful, which will still
1103  * indicate that the processing is complete.
1104  */
1105 static bool fold_not(OptContext *ctx, TCGOp *op);
1106 static bool fold_to_not(OptContext *ctx, TCGOp *op, int idx)
1107 {
1108     TCGOpcode not_op;
1109     bool have_not;
1110 
1111     switch (ctx->type) {
1112     case TCG_TYPE_I32:
1113     case TCG_TYPE_I64:
1114         not_op = INDEX_op_not;
1115         have_not = tcg_op_supported(INDEX_op_not, ctx->type, 0);
1116         break;
1117     case TCG_TYPE_V64:
1118     case TCG_TYPE_V128:
1119     case TCG_TYPE_V256:
1120         not_op = INDEX_op_not_vec;
1121         have_not = TCG_TARGET_HAS_not_vec;
1122         break;
1123     default:
1124         g_assert_not_reached();
1125     }
1126     if (have_not) {
1127         op->opc = not_op;
1128         op->args[1] = op->args[idx];
1129         return fold_not(ctx, op);
1130     }
1131     return false;
1132 }
1133 
1134 /* If the binary operation has first argument @i, fold to @i. */
1135 static bool fold_ix_to_i(OptContext *ctx, TCGOp *op, uint64_t i)
1136 {
1137     if (arg_is_const_val(op->args[1], i)) {
1138         return tcg_opt_gen_movi(ctx, op, op->args[0], i);
1139     }
1140     return false;
1141 }
1142 
1143 /* If the binary operation has first argument @i, fold to NOT. */
1144 static bool fold_ix_to_not(OptContext *ctx, TCGOp *op, uint64_t i)
1145 {
1146     if (arg_is_const_val(op->args[1], i)) {
1147         return fold_to_not(ctx, op, 2);
1148     }
1149     return false;
1150 }
1151 
1152 /* If the binary operation has second argument @i, fold to @i. */
1153 static bool fold_xi_to_i(OptContext *ctx, TCGOp *op, uint64_t i)
1154 {
1155     if (arg_is_const_val(op->args[2], i)) {
1156         return tcg_opt_gen_movi(ctx, op, op->args[0], i);
1157     }
1158     return false;
1159 }
1160 
1161 /* If the binary operation has second argument @i, fold to identity. */
1162 static bool fold_xi_to_x(OptContext *ctx, TCGOp *op, uint64_t i)
1163 {
1164     if (arg_is_const_val(op->args[2], i)) {
1165         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
1166     }
1167     return false;
1168 }
1169 
1170 /* If the binary operation has second argument @i, fold to NOT. */
1171 static bool fold_xi_to_not(OptContext *ctx, TCGOp *op, uint64_t i)
1172 {
1173     if (arg_is_const_val(op->args[2], i)) {
1174         return fold_to_not(ctx, op, 1);
1175     }
1176     return false;
1177 }
1178 
1179 /* If the binary operation has both arguments equal, fold to @i. */
1180 static bool fold_xx_to_i(OptContext *ctx, TCGOp *op, uint64_t i)
1181 {
1182     if (args_are_copies(op->args[1], op->args[2])) {
1183         return tcg_opt_gen_movi(ctx, op, op->args[0], i);
1184     }
1185     return false;
1186 }
1187 
1188 /* If the binary operation has both arguments equal, fold to identity. */
1189 static bool fold_xx_to_x(OptContext *ctx, TCGOp *op)
1190 {
1191     if (args_are_copies(op->args[1], op->args[2])) {
1192         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
1193     }
1194     return false;
1195 }
1196 
1197 /*
1198  * These outermost fold_<op> functions are sorted alphabetically.
1199  *
1200  * The ordering of the transformations should be:
1201  *   1) those that produce a constant
1202  *   2) those that produce a copy
1203  *   3) those that produce information about the result value.
1204  */
1205 
1206 static bool fold_or(OptContext *ctx, TCGOp *op);
1207 static bool fold_orc(OptContext *ctx, TCGOp *op);
1208 static bool fold_xor(OptContext *ctx, TCGOp *op);
1209 
1210 static bool fold_add(OptContext *ctx, TCGOp *op)
1211 {
1212     if (fold_const2_commutative(ctx, op) ||
1213         fold_xi_to_x(ctx, op, 0)) {
1214         return true;
1215     }
1216     return finish_folding(ctx, op);
1217 }
1218 
1219 /* We cannot as yet do_constant_folding with vectors. */
1220 static bool fold_add_vec(OptContext *ctx, TCGOp *op)
1221 {
1222     if (fold_commutative(ctx, op) ||
1223         fold_xi_to_x(ctx, op, 0)) {
1224         return true;
1225     }
1226     return finish_folding(ctx, op);
1227 }
1228 
1229 static bool fold_addsub2(OptContext *ctx, TCGOp *op, bool add)
1230 {
1231     bool a_const = arg_is_const(op->args[2]) && arg_is_const(op->args[3]);
1232     bool b_const = arg_is_const(op->args[4]) && arg_is_const(op->args[5]);
1233 
1234     if (a_const && b_const) {
1235         uint64_t al = arg_info(op->args[2])->val;
1236         uint64_t ah = arg_info(op->args[3])->val;
1237         uint64_t bl = arg_info(op->args[4])->val;
1238         uint64_t bh = arg_info(op->args[5])->val;
1239         TCGArg rl, rh;
1240         TCGOp *op2;
1241 
1242         if (ctx->type == TCG_TYPE_I32) {
1243             uint64_t a = deposit64(al, 32, 32, ah);
1244             uint64_t b = deposit64(bl, 32, 32, bh);
1245 
1246             if (add) {
1247                 a += b;
1248             } else {
1249                 a -= b;
1250             }
1251 
1252             al = sextract64(a, 0, 32);
1253             ah = sextract64(a, 32, 32);
1254         } else {
1255             Int128 a = int128_make128(al, ah);
1256             Int128 b = int128_make128(bl, bh);
1257 
1258             if (add) {
1259                 a = int128_add(a, b);
1260             } else {
1261                 a = int128_sub(a, b);
1262             }
1263 
1264             al = int128_getlo(a);
1265             ah = int128_gethi(a);
1266         }
1267 
1268         rl = op->args[0];
1269         rh = op->args[1];
1270 
1271         /* The proper opcode is supplied by tcg_opt_gen_mov. */
1272         op2 = opt_insert_before(ctx, op, 0, 2);
1273 
1274         tcg_opt_gen_movi(ctx, op, rl, al);
1275         tcg_opt_gen_movi(ctx, op2, rh, ah);
1276         return true;
1277     }
1278 
1279     /* Fold sub2 r,x,i to add2 r,x,-i */
1280     if (!add && b_const) {
1281         uint64_t bl = arg_info(op->args[4])->val;
1282         uint64_t bh = arg_info(op->args[5])->val;
1283 
1284         /* Negate the two parts without assembling and disassembling. */
1285         bl = -bl;
1286         bh = ~bh + !bl;
1287 
1288         op->opc = (ctx->type == TCG_TYPE_I32
1289                    ? INDEX_op_add2_i32 : INDEX_op_add2_i64);
1290         op->args[4] = arg_new_constant(ctx, bl);
1291         op->args[5] = arg_new_constant(ctx, bh);
1292     }
1293     return finish_folding(ctx, op);
1294 }
1295 
1296 static bool fold_add2(OptContext *ctx, TCGOp *op)
1297 {
1298     /* Note that the high and low parts may be independently swapped. */
1299     swap_commutative(op->args[0], &op->args[2], &op->args[4]);
1300     swap_commutative(op->args[1], &op->args[3], &op->args[5]);
1301 
1302     return fold_addsub2(ctx, op, true);
1303 }
1304 
1305 static bool fold_and(OptContext *ctx, TCGOp *op)
1306 {
1307     uint64_t z1, z2, z_mask, s_mask;
1308     TempOptInfo *t1, *t2;
1309 
1310     if (fold_const2_commutative(ctx, op) ||
1311         fold_xi_to_i(ctx, op, 0) ||
1312         fold_xi_to_x(ctx, op, -1) ||
1313         fold_xx_to_x(ctx, op)) {
1314         return true;
1315     }
1316 
1317     t1 = arg_info(op->args[1]);
1318     t2 = arg_info(op->args[2]);
1319     z1 = t1->z_mask;
1320     z2 = t2->z_mask;
1321 
1322     /*
1323      * Known-zeros does not imply known-ones.  Therefore unless
1324      * arg2 is constant, we can't infer affected bits from it.
1325      */
1326     if (ti_is_const(t2) && fold_affected_mask(ctx, op, z1 & ~z2)) {
1327         return true;
1328     }
1329 
1330     z_mask = z1 & z2;
1331 
1332     /*
1333      * Sign repetitions are perforce all identical, whether they are 1 or 0.
1334      * Bitwise operations preserve the relative quantity of the repetitions.
1335      */
1336     s_mask = t1->s_mask & t2->s_mask;
1337 
1338     return fold_masks_zs(ctx, op, z_mask, s_mask);
1339 }
1340 
1341 static bool fold_andc(OptContext *ctx, TCGOp *op)
1342 {
1343     uint64_t z_mask, s_mask;
1344     TempOptInfo *t1, *t2;
1345 
1346     if (fold_const2(ctx, op) ||
1347         fold_xx_to_i(ctx, op, 0) ||
1348         fold_xi_to_x(ctx, op, 0) ||
1349         fold_ix_to_not(ctx, op, -1)) {
1350         return true;
1351     }
1352 
1353     t1 = arg_info(op->args[1]);
1354     t2 = arg_info(op->args[2]);
1355     z_mask = t1->z_mask;
1356 
1357     if (ti_is_const(t2)) {
1358         /* Fold andc r,x,i to and r,x,~i. */
1359         switch (ctx->type) {
1360         case TCG_TYPE_I32:
1361         case TCG_TYPE_I64:
1362             op->opc = INDEX_op_and;
1363             break;
1364         case TCG_TYPE_V64:
1365         case TCG_TYPE_V128:
1366         case TCG_TYPE_V256:
1367             op->opc = INDEX_op_and_vec;
1368             break;
1369         default:
1370             g_assert_not_reached();
1371         }
1372         op->args[2] = arg_new_constant(ctx, ~ti_const_val(t2));
1373         return fold_and(ctx, op);
1374     }
1375 
1376     /*
1377      * Known-zeros does not imply known-ones.  Therefore unless
1378      * arg2 is constant, we can't infer anything from it.
1379      */
1380     if (ti_is_const(t2)) {
1381         uint64_t v2 = ti_const_val(t2);
1382         if (fold_affected_mask(ctx, op, z_mask & v2)) {
1383             return true;
1384         }
1385         z_mask &= ~v2;
1386     }
1387 
1388     s_mask = t1->s_mask & t2->s_mask;
1389     return fold_masks_zs(ctx, op, z_mask, s_mask);
1390 }
1391 
1392 static bool fold_bitsel_vec(OptContext *ctx, TCGOp *op)
1393 {
1394     /* If true and false values are the same, eliminate the cmp. */
1395     if (args_are_copies(op->args[2], op->args[3])) {
1396         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[2]);
1397     }
1398 
1399     if (arg_is_const(op->args[2]) && arg_is_const(op->args[3])) {
1400         uint64_t tv = arg_info(op->args[2])->val;
1401         uint64_t fv = arg_info(op->args[3])->val;
1402 
1403         if (tv == -1 && fv == 0) {
1404             return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
1405         }
1406         if (tv == 0 && fv == -1) {
1407             if (TCG_TARGET_HAS_not_vec) {
1408                 op->opc = INDEX_op_not_vec;
1409                 return fold_not(ctx, op);
1410             } else {
1411                 op->opc = INDEX_op_xor_vec;
1412                 op->args[2] = arg_new_constant(ctx, -1);
1413                 return fold_xor(ctx, op);
1414             }
1415         }
1416     }
1417     if (arg_is_const(op->args[2])) {
1418         uint64_t tv = arg_info(op->args[2])->val;
1419         if (tv == -1) {
1420             op->opc = INDEX_op_or_vec;
1421             op->args[2] = op->args[3];
1422             return fold_or(ctx, op);
1423         }
1424         if (tv == 0 && TCG_TARGET_HAS_andc_vec) {
1425             op->opc = INDEX_op_andc_vec;
1426             op->args[2] = op->args[1];
1427             op->args[1] = op->args[3];
1428             return fold_andc(ctx, op);
1429         }
1430     }
1431     if (arg_is_const(op->args[3])) {
1432         uint64_t fv = arg_info(op->args[3])->val;
1433         if (fv == 0) {
1434             op->opc = INDEX_op_and_vec;
1435             return fold_and(ctx, op);
1436         }
1437         if (fv == -1 && TCG_TARGET_HAS_orc_vec) {
1438             op->opc = INDEX_op_orc_vec;
1439             op->args[2] = op->args[1];
1440             op->args[1] = op->args[3];
1441             return fold_orc(ctx, op);
1442         }
1443     }
1444     return finish_folding(ctx, op);
1445 }
1446 
1447 static bool fold_brcond(OptContext *ctx, TCGOp *op)
1448 {
1449     int i = do_constant_folding_cond1(ctx, op, NO_DEST, &op->args[0],
1450                                       &op->args[1], &op->args[2]);
1451     if (i == 0) {
1452         tcg_op_remove(ctx->tcg, op);
1453         return true;
1454     }
1455     if (i > 0) {
1456         op->opc = INDEX_op_br;
1457         op->args[0] = op->args[3];
1458         finish_ebb(ctx);
1459     } else {
1460         finish_bb(ctx);
1461     }
1462     return true;
1463 }
1464 
1465 static bool fold_brcond2(OptContext *ctx, TCGOp *op)
1466 {
1467     TCGCond cond;
1468     TCGArg label;
1469     int i, inv = 0;
1470 
1471     i = do_constant_folding_cond2(ctx, op, &op->args[0]);
1472     cond = op->args[4];
1473     label = op->args[5];
1474     if (i >= 0) {
1475         goto do_brcond_const;
1476     }
1477 
1478     switch (cond) {
1479     case TCG_COND_LT:
1480     case TCG_COND_GE:
1481         /*
1482          * Simplify LT/GE comparisons vs zero to a single compare
1483          * vs the high word of the input.
1484          */
1485         if (arg_is_const_val(op->args[2], 0) &&
1486             arg_is_const_val(op->args[3], 0)) {
1487             goto do_brcond_high;
1488         }
1489         break;
1490 
1491     case TCG_COND_NE:
1492         inv = 1;
1493         QEMU_FALLTHROUGH;
1494     case TCG_COND_EQ:
1495         /*
1496          * Simplify EQ/NE comparisons where one of the pairs
1497          * can be simplified.
1498          */
1499         i = do_constant_folding_cond(TCG_TYPE_I32, op->args[0],
1500                                      op->args[2], cond);
1501         switch (i ^ inv) {
1502         case 0:
1503             goto do_brcond_const;
1504         case 1:
1505             goto do_brcond_high;
1506         }
1507 
1508         i = do_constant_folding_cond(TCG_TYPE_I32, op->args[1],
1509                                      op->args[3], cond);
1510         switch (i ^ inv) {
1511         case 0:
1512             goto do_brcond_const;
1513         case 1:
1514             goto do_brcond_low;
1515         }
1516         break;
1517 
1518     case TCG_COND_TSTEQ:
1519     case TCG_COND_TSTNE:
1520         if (arg_is_const_val(op->args[2], 0)) {
1521             goto do_brcond_high;
1522         }
1523         if (arg_is_const_val(op->args[3], 0)) {
1524             goto do_brcond_low;
1525         }
1526         break;
1527 
1528     default:
1529         break;
1530 
1531     do_brcond_low:
1532         op->opc = INDEX_op_brcond_i32;
1533         op->args[1] = op->args[2];
1534         op->args[2] = cond;
1535         op->args[3] = label;
1536         return fold_brcond(ctx, op);
1537 
1538     do_brcond_high:
1539         op->opc = INDEX_op_brcond_i32;
1540         op->args[0] = op->args[1];
1541         op->args[1] = op->args[3];
1542         op->args[2] = cond;
1543         op->args[3] = label;
1544         return fold_brcond(ctx, op);
1545 
1546     do_brcond_const:
1547         if (i == 0) {
1548             tcg_op_remove(ctx->tcg, op);
1549             return true;
1550         }
1551         op->opc = INDEX_op_br;
1552         op->args[0] = label;
1553         finish_ebb(ctx);
1554         return true;
1555     }
1556 
1557     finish_bb(ctx);
1558     return true;
1559 }
1560 
1561 static bool fold_bswap(OptContext *ctx, TCGOp *op)
1562 {
1563     uint64_t z_mask, s_mask, sign;
1564     TempOptInfo *t1 = arg_info(op->args[1]);
1565 
1566     if (ti_is_const(t1)) {
1567         return tcg_opt_gen_movi(ctx, op, op->args[0],
1568                                 do_constant_folding(op->opc, ctx->type,
1569                                                     ti_const_val(t1),
1570                                                     op->args[2]));
1571     }
1572 
1573     z_mask = t1->z_mask;
1574     switch (op->opc) {
1575     case INDEX_op_bswap16_i32:
1576     case INDEX_op_bswap16_i64:
1577         z_mask = bswap16(z_mask);
1578         sign = INT16_MIN;
1579         break;
1580     case INDEX_op_bswap32_i32:
1581     case INDEX_op_bswap32_i64:
1582         z_mask = bswap32(z_mask);
1583         sign = INT32_MIN;
1584         break;
1585     case INDEX_op_bswap64_i64:
1586         z_mask = bswap64(z_mask);
1587         sign = INT64_MIN;
1588         break;
1589     default:
1590         g_assert_not_reached();
1591     }
1592 
1593     s_mask = 0;
1594     switch (op->args[2] & (TCG_BSWAP_OZ | TCG_BSWAP_OS)) {
1595     case TCG_BSWAP_OZ:
1596         break;
1597     case TCG_BSWAP_OS:
1598         /* If the sign bit may be 1, force all the bits above to 1. */
1599         if (z_mask & sign) {
1600             z_mask |= sign;
1601         }
1602         /* The value and therefore s_mask is explicitly sign-extended. */
1603         s_mask = sign;
1604         break;
1605     default:
1606         /* The high bits are undefined: force all bits above the sign to 1. */
1607         z_mask |= sign << 1;
1608         break;
1609     }
1610 
1611     return fold_masks_zs(ctx, op, z_mask, s_mask);
1612 }
1613 
1614 static bool fold_call(OptContext *ctx, TCGOp *op)
1615 {
1616     TCGContext *s = ctx->tcg;
1617     int nb_oargs = TCGOP_CALLO(op);
1618     int nb_iargs = TCGOP_CALLI(op);
1619     int flags, i;
1620 
1621     init_arguments(ctx, op, nb_oargs + nb_iargs);
1622     copy_propagate(ctx, op, nb_oargs, nb_iargs);
1623 
1624     /* If the function reads or writes globals, reset temp data. */
1625     flags = tcg_call_flags(op);
1626     if (!(flags & (TCG_CALL_NO_READ_GLOBALS | TCG_CALL_NO_WRITE_GLOBALS))) {
1627         int nb_globals = s->nb_globals;
1628 
1629         for (i = 0; i < nb_globals; i++) {
1630             if (test_bit(i, ctx->temps_used.l)) {
1631                 reset_ts(ctx, &ctx->tcg->temps[i]);
1632             }
1633         }
1634     }
1635 
1636     /* If the function has side effects, reset mem data. */
1637     if (!(flags & TCG_CALL_NO_SIDE_EFFECTS)) {
1638         remove_mem_copy_all(ctx);
1639     }
1640 
1641     /* Reset temp data for outputs. */
1642     for (i = 0; i < nb_oargs; i++) {
1643         reset_temp(ctx, op->args[i]);
1644     }
1645 
1646     /* Stop optimizing MB across calls. */
1647     ctx->prev_mb = NULL;
1648     return true;
1649 }
1650 
1651 static bool fold_cmp_vec(OptContext *ctx, TCGOp *op)
1652 {
1653     /* Canonicalize the comparison to put immediate second. */
1654     if (swap_commutative(NO_DEST, &op->args[1], &op->args[2])) {
1655         op->args[3] = tcg_swap_cond(op->args[3]);
1656     }
1657     return finish_folding(ctx, op);
1658 }
1659 
1660 static bool fold_cmpsel_vec(OptContext *ctx, TCGOp *op)
1661 {
1662     /* If true and false values are the same, eliminate the cmp. */
1663     if (args_are_copies(op->args[3], op->args[4])) {
1664         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[3]);
1665     }
1666 
1667     /* Canonicalize the comparison to put immediate second. */
1668     if (swap_commutative(NO_DEST, &op->args[1], &op->args[2])) {
1669         op->args[5] = tcg_swap_cond(op->args[5]);
1670     }
1671     /*
1672      * Canonicalize the "false" input reg to match the destination,
1673      * so that the tcg backend can implement "move if true".
1674      */
1675     if (swap_commutative(op->args[0], &op->args[4], &op->args[3])) {
1676         op->args[5] = tcg_invert_cond(op->args[5]);
1677     }
1678     return finish_folding(ctx, op);
1679 }
1680 
1681 static bool fold_count_zeros(OptContext *ctx, TCGOp *op)
1682 {
1683     uint64_t z_mask, s_mask;
1684     TempOptInfo *t1 = arg_info(op->args[1]);
1685     TempOptInfo *t2 = arg_info(op->args[2]);
1686 
1687     if (ti_is_const(t1)) {
1688         uint64_t t = ti_const_val(t1);
1689 
1690         if (t != 0) {
1691             t = do_constant_folding(op->opc, ctx->type, t, 0);
1692             return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1693         }
1694         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[2]);
1695     }
1696 
1697     switch (ctx->type) {
1698     case TCG_TYPE_I32:
1699         z_mask = 31;
1700         break;
1701     case TCG_TYPE_I64:
1702         z_mask = 63;
1703         break;
1704     default:
1705         g_assert_not_reached();
1706     }
1707     s_mask = ~z_mask;
1708     z_mask |= t2->z_mask;
1709     s_mask &= t2->s_mask;
1710 
1711     return fold_masks_zs(ctx, op, z_mask, s_mask);
1712 }
1713 
1714 static bool fold_ctpop(OptContext *ctx, TCGOp *op)
1715 {
1716     uint64_t z_mask;
1717 
1718     if (fold_const1(ctx, op)) {
1719         return true;
1720     }
1721 
1722     switch (ctx->type) {
1723     case TCG_TYPE_I32:
1724         z_mask = 32 | 31;
1725         break;
1726     case TCG_TYPE_I64:
1727         z_mask = 64 | 63;
1728         break;
1729     default:
1730         g_assert_not_reached();
1731     }
1732     return fold_masks_z(ctx, op, z_mask);
1733 }
1734 
1735 static bool fold_deposit(OptContext *ctx, TCGOp *op)
1736 {
1737     TempOptInfo *t1 = arg_info(op->args[1]);
1738     TempOptInfo *t2 = arg_info(op->args[2]);
1739     int ofs = op->args[3];
1740     int len = op->args[4];
1741     int width = 8 * tcg_type_size(ctx->type);
1742     uint64_t z_mask, s_mask;
1743 
1744     if (ti_is_const(t1) && ti_is_const(t2)) {
1745         return tcg_opt_gen_movi(ctx, op, op->args[0],
1746                                 deposit64(ti_const_val(t1), ofs, len,
1747                                           ti_const_val(t2)));
1748     }
1749 
1750     /* Inserting a value into zero at offset 0. */
1751     if (ti_is_const_val(t1, 0) && ofs == 0) {
1752         uint64_t mask = MAKE_64BIT_MASK(0, len);
1753 
1754         op->opc = INDEX_op_and;
1755         op->args[1] = op->args[2];
1756         op->args[2] = arg_new_constant(ctx, mask);
1757         return fold_and(ctx, op);
1758     }
1759 
1760     /* Inserting zero into a value. */
1761     if (ti_is_const_val(t2, 0)) {
1762         uint64_t mask = deposit64(-1, ofs, len, 0);
1763 
1764         op->opc = INDEX_op_and;
1765         op->args[2] = arg_new_constant(ctx, mask);
1766         return fold_and(ctx, op);
1767     }
1768 
1769     /* The s_mask from the top portion of the deposit is still valid. */
1770     if (ofs + len == width) {
1771         s_mask = t2->s_mask << ofs;
1772     } else {
1773         s_mask = t1->s_mask & ~MAKE_64BIT_MASK(0, ofs + len);
1774     }
1775 
1776     z_mask = deposit64(t1->z_mask, ofs, len, t2->z_mask);
1777     return fold_masks_zs(ctx, op, z_mask, s_mask);
1778 }
1779 
1780 static bool fold_divide(OptContext *ctx, TCGOp *op)
1781 {
1782     if (fold_const2(ctx, op) ||
1783         fold_xi_to_x(ctx, op, 1)) {
1784         return true;
1785     }
1786     return finish_folding(ctx, op);
1787 }
1788 
1789 static bool fold_dup(OptContext *ctx, TCGOp *op)
1790 {
1791     if (arg_is_const(op->args[1])) {
1792         uint64_t t = arg_info(op->args[1])->val;
1793         t = dup_const(TCGOP_VECE(op), t);
1794         return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1795     }
1796     return finish_folding(ctx, op);
1797 }
1798 
1799 static bool fold_dup2(OptContext *ctx, TCGOp *op)
1800 {
1801     if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) {
1802         uint64_t t = deposit64(arg_info(op->args[1])->val, 32, 32,
1803                                arg_info(op->args[2])->val);
1804         return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1805     }
1806 
1807     if (args_are_copies(op->args[1], op->args[2])) {
1808         op->opc = INDEX_op_dup_vec;
1809         TCGOP_VECE(op) = MO_32;
1810     }
1811     return finish_folding(ctx, op);
1812 }
1813 
1814 static bool fold_eqv(OptContext *ctx, TCGOp *op)
1815 {
1816     uint64_t s_mask;
1817     TempOptInfo *t1, *t2;
1818 
1819     if (fold_const2_commutative(ctx, op) ||
1820         fold_xi_to_x(ctx, op, -1) ||
1821         fold_xi_to_not(ctx, op, 0)) {
1822         return true;
1823     }
1824 
1825     t2 = arg_info(op->args[2]);
1826     if (ti_is_const(t2)) {
1827         /* Fold eqv r,x,i to xor r,x,~i. */
1828         switch (ctx->type) {
1829         case TCG_TYPE_I32:
1830         case TCG_TYPE_I64:
1831             op->opc = INDEX_op_xor;
1832             break;
1833         case TCG_TYPE_V64:
1834         case TCG_TYPE_V128:
1835         case TCG_TYPE_V256:
1836             op->opc = INDEX_op_xor_vec;
1837             break;
1838         default:
1839             g_assert_not_reached();
1840         }
1841         op->args[2] = arg_new_constant(ctx, ~ti_const_val(t2));
1842         return fold_xor(ctx, op);
1843     }
1844 
1845     t1 = arg_info(op->args[1]);
1846     s_mask = t1->s_mask & t2->s_mask;
1847     return fold_masks_s(ctx, op, s_mask);
1848 }
1849 
1850 static bool fold_extract(OptContext *ctx, TCGOp *op)
1851 {
1852     uint64_t z_mask_old, z_mask;
1853     TempOptInfo *t1 = arg_info(op->args[1]);
1854     int pos = op->args[2];
1855     int len = op->args[3];
1856 
1857     if (ti_is_const(t1)) {
1858         return tcg_opt_gen_movi(ctx, op, op->args[0],
1859                                 extract64(ti_const_val(t1), pos, len));
1860     }
1861 
1862     z_mask_old = t1->z_mask;
1863     z_mask = extract64(z_mask_old, pos, len);
1864     if (pos == 0 && fold_affected_mask(ctx, op, z_mask_old ^ z_mask)) {
1865         return true;
1866     }
1867 
1868     return fold_masks_z(ctx, op, z_mask);
1869 }
1870 
1871 static bool fold_extract2(OptContext *ctx, TCGOp *op)
1872 {
1873     if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) {
1874         uint64_t v1 = arg_info(op->args[1])->val;
1875         uint64_t v2 = arg_info(op->args[2])->val;
1876         int shr = op->args[3];
1877 
1878         if (op->opc == INDEX_op_extract2_i64) {
1879             v1 >>= shr;
1880             v2 <<= 64 - shr;
1881         } else {
1882             v1 = (uint32_t)v1 >> shr;
1883             v2 = (uint64_t)((int32_t)v2 << (32 - shr));
1884         }
1885         return tcg_opt_gen_movi(ctx, op, op->args[0], v1 | v2);
1886     }
1887     return finish_folding(ctx, op);
1888 }
1889 
1890 static bool fold_exts(OptContext *ctx, TCGOp *op)
1891 {
1892     uint64_t s_mask, z_mask;
1893     TempOptInfo *t1;
1894 
1895     if (fold_const1(ctx, op)) {
1896         return true;
1897     }
1898 
1899     t1 = arg_info(op->args[1]);
1900     z_mask = t1->z_mask;
1901     s_mask = t1->s_mask;
1902 
1903     switch (op->opc) {
1904     case INDEX_op_ext_i32_i64:
1905         s_mask |= INT32_MIN;
1906         z_mask = (int32_t)z_mask;
1907         break;
1908     default:
1909         g_assert_not_reached();
1910     }
1911     return fold_masks_zs(ctx, op, z_mask, s_mask);
1912 }
1913 
1914 static bool fold_extu(OptContext *ctx, TCGOp *op)
1915 {
1916     uint64_t z_mask;
1917 
1918     if (fold_const1(ctx, op)) {
1919         return true;
1920     }
1921 
1922     z_mask = arg_info(op->args[1])->z_mask;
1923     switch (op->opc) {
1924     case INDEX_op_extrl_i64_i32:
1925     case INDEX_op_extu_i32_i64:
1926         z_mask = (uint32_t)z_mask;
1927         break;
1928     case INDEX_op_extrh_i64_i32:
1929         z_mask >>= 32;
1930         break;
1931     default:
1932         g_assert_not_reached();
1933     }
1934     return fold_masks_z(ctx, op, z_mask);
1935 }
1936 
1937 static bool fold_mb(OptContext *ctx, TCGOp *op)
1938 {
1939     /* Eliminate duplicate and redundant fence instructions.  */
1940     if (ctx->prev_mb) {
1941         /*
1942          * Merge two barriers of the same type into one,
1943          * or a weaker barrier into a stronger one,
1944          * or two weaker barriers into a stronger one.
1945          *   mb X; mb Y => mb X|Y
1946          *   mb; strl => mb; st
1947          *   ldaq; mb => ld; mb
1948          *   ldaq; strl => ld; mb; st
1949          * Other combinations are also merged into a strong
1950          * barrier.  This is stricter than specified but for
1951          * the purposes of TCG is better than not optimizing.
1952          */
1953         ctx->prev_mb->args[0] |= op->args[0];
1954         tcg_op_remove(ctx->tcg, op);
1955     } else {
1956         ctx->prev_mb = op;
1957     }
1958     return true;
1959 }
1960 
1961 static bool fold_mov(OptContext *ctx, TCGOp *op)
1962 {
1963     return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
1964 }
1965 
1966 static bool fold_movcond(OptContext *ctx, TCGOp *op)
1967 {
1968     uint64_t z_mask, s_mask;
1969     TempOptInfo *tt, *ft;
1970     int i;
1971 
1972     /* If true and false values are the same, eliminate the cmp. */
1973     if (args_are_copies(op->args[3], op->args[4])) {
1974         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[3]);
1975     }
1976 
1977     /*
1978      * Canonicalize the "false" input reg to match the destination reg so
1979      * that the tcg backend can implement a "move if true" operation.
1980      */
1981     if (swap_commutative(op->args[0], &op->args[4], &op->args[3])) {
1982         op->args[5] = tcg_invert_cond(op->args[5]);
1983     }
1984 
1985     i = do_constant_folding_cond1(ctx, op, NO_DEST, &op->args[1],
1986                                   &op->args[2], &op->args[5]);
1987     if (i >= 0) {
1988         return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[4 - i]);
1989     }
1990 
1991     tt = arg_info(op->args[3]);
1992     ft = arg_info(op->args[4]);
1993     z_mask = tt->z_mask | ft->z_mask;
1994     s_mask = tt->s_mask & ft->s_mask;
1995 
1996     if (ti_is_const(tt) && ti_is_const(ft)) {
1997         uint64_t tv = ti_const_val(tt);
1998         uint64_t fv = ti_const_val(ft);
1999         TCGOpcode opc, negopc = 0;
2000         TCGCond cond = op->args[5];
2001 
2002         switch (ctx->type) {
2003         case TCG_TYPE_I32:
2004             opc = INDEX_op_setcond_i32;
2005             if (TCG_TARGET_HAS_negsetcond_i32) {
2006                 negopc = INDEX_op_negsetcond_i32;
2007             }
2008             tv = (int32_t)tv;
2009             fv = (int32_t)fv;
2010             break;
2011         case TCG_TYPE_I64:
2012             opc = INDEX_op_setcond_i64;
2013             if (TCG_TARGET_HAS_negsetcond_i64) {
2014                 negopc = INDEX_op_negsetcond_i64;
2015             }
2016             break;
2017         default:
2018             g_assert_not_reached();
2019         }
2020 
2021         if (tv == 1 && fv == 0) {
2022             op->opc = opc;
2023             op->args[3] = cond;
2024         } else if (fv == 1 && tv == 0) {
2025             op->opc = opc;
2026             op->args[3] = tcg_invert_cond(cond);
2027         } else if (negopc) {
2028             if (tv == -1 && fv == 0) {
2029                 op->opc = negopc;
2030                 op->args[3] = cond;
2031             } else if (fv == -1 && tv == 0) {
2032                 op->opc = negopc;
2033                 op->args[3] = tcg_invert_cond(cond);
2034             }
2035         }
2036     }
2037 
2038     return fold_masks_zs(ctx, op, z_mask, s_mask);
2039 }
2040 
2041 static bool fold_mul(OptContext *ctx, TCGOp *op)
2042 {
2043     if (fold_const2(ctx, op) ||
2044         fold_xi_to_i(ctx, op, 0) ||
2045         fold_xi_to_x(ctx, op, 1)) {
2046         return true;
2047     }
2048     return finish_folding(ctx, op);
2049 }
2050 
2051 static bool fold_mul_highpart(OptContext *ctx, TCGOp *op)
2052 {
2053     if (fold_const2_commutative(ctx, op) ||
2054         fold_xi_to_i(ctx, op, 0)) {
2055         return true;
2056     }
2057     return finish_folding(ctx, op);
2058 }
2059 
2060 static bool fold_multiply2(OptContext *ctx, TCGOp *op)
2061 {
2062     swap_commutative(op->args[0], &op->args[2], &op->args[3]);
2063 
2064     if (arg_is_const(op->args[2]) && arg_is_const(op->args[3])) {
2065         uint64_t a = arg_info(op->args[2])->val;
2066         uint64_t b = arg_info(op->args[3])->val;
2067         uint64_t h, l;
2068         TCGArg rl, rh;
2069         TCGOp *op2;
2070 
2071         switch (op->opc) {
2072         case INDEX_op_mulu2_i32:
2073             l = (uint64_t)(uint32_t)a * (uint32_t)b;
2074             h = (int32_t)(l >> 32);
2075             l = (int32_t)l;
2076             break;
2077         case INDEX_op_muls2_i32:
2078             l = (int64_t)(int32_t)a * (int32_t)b;
2079             h = l >> 32;
2080             l = (int32_t)l;
2081             break;
2082         case INDEX_op_mulu2_i64:
2083             mulu64(&l, &h, a, b);
2084             break;
2085         case INDEX_op_muls2_i64:
2086             muls64(&l, &h, a, b);
2087             break;
2088         default:
2089             g_assert_not_reached();
2090         }
2091 
2092         rl = op->args[0];
2093         rh = op->args[1];
2094 
2095         /* The proper opcode is supplied by tcg_opt_gen_mov. */
2096         op2 = opt_insert_before(ctx, op, 0, 2);
2097 
2098         tcg_opt_gen_movi(ctx, op, rl, l);
2099         tcg_opt_gen_movi(ctx, op2, rh, h);
2100         return true;
2101     }
2102     return finish_folding(ctx, op);
2103 }
2104 
2105 static bool fold_nand(OptContext *ctx, TCGOp *op)
2106 {
2107     uint64_t s_mask;
2108 
2109     if (fold_const2_commutative(ctx, op) ||
2110         fold_xi_to_not(ctx, op, -1)) {
2111         return true;
2112     }
2113 
2114     s_mask = arg_info(op->args[1])->s_mask
2115            & arg_info(op->args[2])->s_mask;
2116     return fold_masks_s(ctx, op, s_mask);
2117 }
2118 
2119 static bool fold_neg_no_const(OptContext *ctx, TCGOp *op)
2120 {
2121     /* Set to 1 all bits to the left of the rightmost.  */
2122     uint64_t z_mask = arg_info(op->args[1])->z_mask;
2123     z_mask = -(z_mask & -z_mask);
2124 
2125     return fold_masks_z(ctx, op, z_mask);
2126 }
2127 
2128 static bool fold_neg(OptContext *ctx, TCGOp *op)
2129 {
2130     return fold_const1(ctx, op) || fold_neg_no_const(ctx, op);
2131 }
2132 
2133 static bool fold_nor(OptContext *ctx, TCGOp *op)
2134 {
2135     uint64_t s_mask;
2136 
2137     if (fold_const2_commutative(ctx, op) ||
2138         fold_xi_to_not(ctx, op, 0)) {
2139         return true;
2140     }
2141 
2142     s_mask = arg_info(op->args[1])->s_mask
2143            & arg_info(op->args[2])->s_mask;
2144     return fold_masks_s(ctx, op, s_mask);
2145 }
2146 
2147 static bool fold_not(OptContext *ctx, TCGOp *op)
2148 {
2149     if (fold_const1(ctx, op)) {
2150         return true;
2151     }
2152     return fold_masks_s(ctx, op, arg_info(op->args[1])->s_mask);
2153 }
2154 
2155 static bool fold_or(OptContext *ctx, TCGOp *op)
2156 {
2157     uint64_t z_mask, s_mask;
2158     TempOptInfo *t1, *t2;
2159 
2160     if (fold_const2_commutative(ctx, op) ||
2161         fold_xi_to_x(ctx, op, 0) ||
2162         fold_xx_to_x(ctx, op)) {
2163         return true;
2164     }
2165 
2166     t1 = arg_info(op->args[1]);
2167     t2 = arg_info(op->args[2]);
2168     z_mask = t1->z_mask | t2->z_mask;
2169     s_mask = t1->s_mask & t2->s_mask;
2170     return fold_masks_zs(ctx, op, z_mask, s_mask);
2171 }
2172 
2173 static bool fold_orc(OptContext *ctx, TCGOp *op)
2174 {
2175     uint64_t s_mask;
2176     TempOptInfo *t1, *t2;
2177 
2178     if (fold_const2(ctx, op) ||
2179         fold_xx_to_i(ctx, op, -1) ||
2180         fold_xi_to_x(ctx, op, -1) ||
2181         fold_ix_to_not(ctx, op, 0)) {
2182         return true;
2183     }
2184 
2185     t2 = arg_info(op->args[2]);
2186     if (ti_is_const(t2)) {
2187         /* Fold orc r,x,i to or r,x,~i. */
2188         switch (ctx->type) {
2189         case TCG_TYPE_I32:
2190         case TCG_TYPE_I64:
2191             op->opc = INDEX_op_or;
2192             break;
2193         case TCG_TYPE_V64:
2194         case TCG_TYPE_V128:
2195         case TCG_TYPE_V256:
2196             op->opc = INDEX_op_or_vec;
2197             break;
2198         default:
2199             g_assert_not_reached();
2200         }
2201         op->args[2] = arg_new_constant(ctx, ~ti_const_val(t2));
2202         return fold_or(ctx, op);
2203     }
2204 
2205     t1 = arg_info(op->args[1]);
2206     s_mask = t1->s_mask & t2->s_mask;
2207     return fold_masks_s(ctx, op, s_mask);
2208 }
2209 
2210 static bool fold_qemu_ld_1reg(OptContext *ctx, TCGOp *op)
2211 {
2212     const TCGOpDef *def = &tcg_op_defs[op->opc];
2213     MemOpIdx oi = op->args[def->nb_oargs + def->nb_iargs];
2214     MemOp mop = get_memop(oi);
2215     int width = 8 * memop_size(mop);
2216     uint64_t z_mask = -1, s_mask = 0;
2217 
2218     if (width < 64) {
2219         if (mop & MO_SIGN) {
2220             s_mask = MAKE_64BIT_MASK(width - 1, 64 - (width - 1));
2221         } else {
2222             z_mask = MAKE_64BIT_MASK(0, width);
2223         }
2224     }
2225 
2226     /* Opcodes that touch guest memory stop the mb optimization.  */
2227     ctx->prev_mb = NULL;
2228 
2229     return fold_masks_zs(ctx, op, z_mask, s_mask);
2230 }
2231 
2232 static bool fold_qemu_ld_2reg(OptContext *ctx, TCGOp *op)
2233 {
2234     /* Opcodes that touch guest memory stop the mb optimization.  */
2235     ctx->prev_mb = NULL;
2236     return finish_folding(ctx, op);
2237 }
2238 
2239 static bool fold_qemu_st(OptContext *ctx, TCGOp *op)
2240 {
2241     /* Opcodes that touch guest memory stop the mb optimization.  */
2242     ctx->prev_mb = NULL;
2243     return true;
2244 }
2245 
2246 static bool fold_remainder(OptContext *ctx, TCGOp *op)
2247 {
2248     if (fold_const2(ctx, op) ||
2249         fold_xx_to_i(ctx, op, 0)) {
2250         return true;
2251     }
2252     return finish_folding(ctx, op);
2253 }
2254 
2255 /* Return 1 if finished, -1 if simplified, 0 if unchanged. */
2256 static int fold_setcond_zmask(OptContext *ctx, TCGOp *op, bool neg)
2257 {
2258     uint64_t a_zmask, b_val;
2259     TCGCond cond;
2260 
2261     if (!arg_is_const(op->args[2])) {
2262         return false;
2263     }
2264 
2265     a_zmask = arg_info(op->args[1])->z_mask;
2266     b_val = arg_info(op->args[2])->val;
2267     cond = op->args[3];
2268 
2269     if (ctx->type == TCG_TYPE_I32) {
2270         a_zmask = (uint32_t)a_zmask;
2271         b_val = (uint32_t)b_val;
2272     }
2273 
2274     /*
2275      * A with only low bits set vs B with high bits set means that A < B.
2276      */
2277     if (a_zmask < b_val) {
2278         bool inv = false;
2279 
2280         switch (cond) {
2281         case TCG_COND_NE:
2282         case TCG_COND_LEU:
2283         case TCG_COND_LTU:
2284             inv = true;
2285             /* fall through */
2286         case TCG_COND_GTU:
2287         case TCG_COND_GEU:
2288         case TCG_COND_EQ:
2289             return tcg_opt_gen_movi(ctx, op, op->args[0], neg ? -inv : inv);
2290         default:
2291             break;
2292         }
2293     }
2294 
2295     /*
2296      * A with only lsb set is already boolean.
2297      */
2298     if (a_zmask <= 1) {
2299         bool convert = false;
2300         bool inv = false;
2301 
2302         switch (cond) {
2303         case TCG_COND_EQ:
2304             inv = true;
2305             /* fall through */
2306         case TCG_COND_NE:
2307             convert = (b_val == 0);
2308             break;
2309         case TCG_COND_LTU:
2310         case TCG_COND_TSTEQ:
2311             inv = true;
2312             /* fall through */
2313         case TCG_COND_GEU:
2314         case TCG_COND_TSTNE:
2315             convert = (b_val == 1);
2316             break;
2317         default:
2318             break;
2319         }
2320         if (convert) {
2321             if (!inv && !neg) {
2322                 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
2323             }
2324 
2325             if (!inv) {
2326                 op->opc = INDEX_op_neg;
2327             } else if (neg) {
2328                 op->opc = INDEX_op_add;
2329                 op->args[2] = arg_new_constant(ctx, -1);
2330             } else {
2331                 op->opc = INDEX_op_xor;
2332                 op->args[2] = arg_new_constant(ctx, 1);
2333             }
2334             return -1;
2335         }
2336     }
2337     return 0;
2338 }
2339 
2340 static void fold_setcond_tst_pow2(OptContext *ctx, TCGOp *op, bool neg)
2341 {
2342     TCGOpcode shr_opc;
2343     TCGOpcode uext_opc = 0, sext_opc = 0;
2344     TCGCond cond = op->args[3];
2345     TCGArg ret, src1, src2;
2346     TCGOp *op2;
2347     uint64_t val;
2348     int sh;
2349     bool inv;
2350 
2351     if (!is_tst_cond(cond) || !arg_is_const(op->args[2])) {
2352         return;
2353     }
2354 
2355     src2 = op->args[2];
2356     val = arg_info(src2)->val;
2357     if (!is_power_of_2(val)) {
2358         return;
2359     }
2360     sh = ctz64(val);
2361 
2362     switch (ctx->type) {
2363     case TCG_TYPE_I32:
2364         shr_opc = INDEX_op_shr_i32;
2365         if (TCG_TARGET_extract_valid(TCG_TYPE_I32, sh, 1)) {
2366             uext_opc = INDEX_op_extract_i32;
2367         }
2368         if (TCG_TARGET_sextract_valid(TCG_TYPE_I32, sh, 1)) {
2369             sext_opc = INDEX_op_sextract_i32;
2370         }
2371         break;
2372     case TCG_TYPE_I64:
2373         shr_opc = INDEX_op_shr_i64;
2374         if (TCG_TARGET_extract_valid(TCG_TYPE_I64, sh, 1)) {
2375             uext_opc = INDEX_op_extract_i64;
2376         }
2377         if (TCG_TARGET_sextract_valid(TCG_TYPE_I64, sh, 1)) {
2378             sext_opc = INDEX_op_sextract_i64;
2379         }
2380         break;
2381     default:
2382         g_assert_not_reached();
2383     }
2384 
2385     ret = op->args[0];
2386     src1 = op->args[1];
2387     inv = cond == TCG_COND_TSTEQ;
2388 
2389     if (sh && sext_opc && neg && !inv) {
2390         op->opc = sext_opc;
2391         op->args[1] = src1;
2392         op->args[2] = sh;
2393         op->args[3] = 1;
2394         return;
2395     } else if (sh && uext_opc) {
2396         op->opc = uext_opc;
2397         op->args[1] = src1;
2398         op->args[2] = sh;
2399         op->args[3] = 1;
2400     } else {
2401         if (sh) {
2402             op2 = opt_insert_before(ctx, op, shr_opc, 3);
2403             op2->args[0] = ret;
2404             op2->args[1] = src1;
2405             op2->args[2] = arg_new_constant(ctx, sh);
2406             src1 = ret;
2407         }
2408         op->opc = INDEX_op_and;
2409         op->args[1] = src1;
2410         op->args[2] = arg_new_constant(ctx, 1);
2411     }
2412 
2413     if (neg && inv) {
2414         op2 = opt_insert_after(ctx, op, INDEX_op_add, 3);
2415         op2->args[0] = ret;
2416         op2->args[1] = ret;
2417         op2->args[2] = arg_new_constant(ctx, -1);
2418     } else if (inv) {
2419         op2 = opt_insert_after(ctx, op, INDEX_op_xor, 3);
2420         op2->args[0] = ret;
2421         op2->args[1] = ret;
2422         op2->args[2] = arg_new_constant(ctx, 1);
2423     } else if (neg) {
2424         op2 = opt_insert_after(ctx, op, INDEX_op_neg, 2);
2425         op2->args[0] = ret;
2426         op2->args[1] = ret;
2427     }
2428 }
2429 
2430 static bool fold_setcond(OptContext *ctx, TCGOp *op)
2431 {
2432     int i = do_constant_folding_cond1(ctx, op, op->args[0], &op->args[1],
2433                                       &op->args[2], &op->args[3]);
2434     if (i >= 0) {
2435         return tcg_opt_gen_movi(ctx, op, op->args[0], i);
2436     }
2437 
2438     i = fold_setcond_zmask(ctx, op, false);
2439     if (i > 0) {
2440         return true;
2441     }
2442     if (i == 0) {
2443         fold_setcond_tst_pow2(ctx, op, false);
2444     }
2445 
2446     return fold_masks_z(ctx, op, 1);
2447 }
2448 
2449 static bool fold_negsetcond(OptContext *ctx, TCGOp *op)
2450 {
2451     int i = do_constant_folding_cond1(ctx, op, op->args[0], &op->args[1],
2452                                       &op->args[2], &op->args[3]);
2453     if (i >= 0) {
2454         return tcg_opt_gen_movi(ctx, op, op->args[0], -i);
2455     }
2456 
2457     i = fold_setcond_zmask(ctx, op, true);
2458     if (i > 0) {
2459         return true;
2460     }
2461     if (i == 0) {
2462         fold_setcond_tst_pow2(ctx, op, true);
2463     }
2464 
2465     /* Value is {0,-1} so all bits are repetitions of the sign. */
2466     return fold_masks_s(ctx, op, -1);
2467 }
2468 
2469 static bool fold_setcond2(OptContext *ctx, TCGOp *op)
2470 {
2471     TCGCond cond;
2472     int i, inv = 0;
2473 
2474     i = do_constant_folding_cond2(ctx, op, &op->args[1]);
2475     cond = op->args[5];
2476     if (i >= 0) {
2477         goto do_setcond_const;
2478     }
2479 
2480     switch (cond) {
2481     case TCG_COND_LT:
2482     case TCG_COND_GE:
2483         /*
2484          * Simplify LT/GE comparisons vs zero to a single compare
2485          * vs the high word of the input.
2486          */
2487         if (arg_is_const_val(op->args[3], 0) &&
2488             arg_is_const_val(op->args[4], 0)) {
2489             goto do_setcond_high;
2490         }
2491         break;
2492 
2493     case TCG_COND_NE:
2494         inv = 1;
2495         QEMU_FALLTHROUGH;
2496     case TCG_COND_EQ:
2497         /*
2498          * Simplify EQ/NE comparisons where one of the pairs
2499          * can be simplified.
2500          */
2501         i = do_constant_folding_cond(TCG_TYPE_I32, op->args[1],
2502                                      op->args[3], cond);
2503         switch (i ^ inv) {
2504         case 0:
2505             goto do_setcond_const;
2506         case 1:
2507             goto do_setcond_high;
2508         }
2509 
2510         i = do_constant_folding_cond(TCG_TYPE_I32, op->args[2],
2511                                      op->args[4], cond);
2512         switch (i ^ inv) {
2513         case 0:
2514             goto do_setcond_const;
2515         case 1:
2516             goto do_setcond_low;
2517         }
2518         break;
2519 
2520     case TCG_COND_TSTEQ:
2521     case TCG_COND_TSTNE:
2522         if (arg_is_const_val(op->args[3], 0)) {
2523             goto do_setcond_high;
2524         }
2525         if (arg_is_const_val(op->args[4], 0)) {
2526             goto do_setcond_low;
2527         }
2528         break;
2529 
2530     default:
2531         break;
2532 
2533     do_setcond_low:
2534         op->args[2] = op->args[3];
2535         op->args[3] = cond;
2536         op->opc = INDEX_op_setcond_i32;
2537         return fold_setcond(ctx, op);
2538 
2539     do_setcond_high:
2540         op->args[1] = op->args[2];
2541         op->args[2] = op->args[4];
2542         op->args[3] = cond;
2543         op->opc = INDEX_op_setcond_i32;
2544         return fold_setcond(ctx, op);
2545     }
2546 
2547     return fold_masks_z(ctx, op, 1);
2548 
2549  do_setcond_const:
2550     return tcg_opt_gen_movi(ctx, op, op->args[0], i);
2551 }
2552 
2553 static bool fold_sextract(OptContext *ctx, TCGOp *op)
2554 {
2555     uint64_t z_mask, s_mask, s_mask_old;
2556     TempOptInfo *t1 = arg_info(op->args[1]);
2557     int pos = op->args[2];
2558     int len = op->args[3];
2559 
2560     if (ti_is_const(t1)) {
2561         return tcg_opt_gen_movi(ctx, op, op->args[0],
2562                                 sextract64(ti_const_val(t1), pos, len));
2563     }
2564 
2565     s_mask_old = t1->s_mask;
2566     s_mask = s_mask_old >> pos;
2567     s_mask |= -1ull << (len - 1);
2568 
2569     if (pos == 0 && fold_affected_mask(ctx, op, s_mask & ~s_mask_old)) {
2570         return true;
2571     }
2572 
2573     z_mask = sextract64(t1->z_mask, pos, len);
2574     return fold_masks_zs(ctx, op, z_mask, s_mask);
2575 }
2576 
2577 static bool fold_shift(OptContext *ctx, TCGOp *op)
2578 {
2579     uint64_t s_mask, z_mask;
2580     TempOptInfo *t1, *t2;
2581 
2582     if (fold_const2(ctx, op) ||
2583         fold_ix_to_i(ctx, op, 0) ||
2584         fold_xi_to_x(ctx, op, 0)) {
2585         return true;
2586     }
2587 
2588     t1 = arg_info(op->args[1]);
2589     t2 = arg_info(op->args[2]);
2590     s_mask = t1->s_mask;
2591     z_mask = t1->z_mask;
2592 
2593     if (ti_is_const(t2)) {
2594         int sh = ti_const_val(t2);
2595 
2596         z_mask = do_constant_folding(op->opc, ctx->type, z_mask, sh);
2597         s_mask = do_constant_folding(op->opc, ctx->type, s_mask, sh);
2598 
2599         return fold_masks_zs(ctx, op, z_mask, s_mask);
2600     }
2601 
2602     switch (op->opc) {
2603     CASE_OP_32_64(sar):
2604         /*
2605          * Arithmetic right shift will not reduce the number of
2606          * input sign repetitions.
2607          */
2608         return fold_masks_s(ctx, op, s_mask);
2609     CASE_OP_32_64(shr):
2610         /*
2611          * If the sign bit is known zero, then logical right shift
2612          * will not reduce the number of input sign repetitions.
2613          */
2614         if (~z_mask & -s_mask) {
2615             return fold_masks_s(ctx, op, s_mask);
2616         }
2617         break;
2618     default:
2619         break;
2620     }
2621 
2622     return finish_folding(ctx, op);
2623 }
2624 
2625 static bool fold_sub_to_neg(OptContext *ctx, TCGOp *op)
2626 {
2627     TCGOpcode neg_op;
2628     bool have_neg;
2629 
2630     if (!arg_is_const(op->args[1]) || arg_info(op->args[1])->val != 0) {
2631         return false;
2632     }
2633 
2634     switch (ctx->type) {
2635     case TCG_TYPE_I32:
2636     case TCG_TYPE_I64:
2637         neg_op = INDEX_op_neg;
2638         have_neg = true;
2639         break;
2640     case TCG_TYPE_V64:
2641     case TCG_TYPE_V128:
2642     case TCG_TYPE_V256:
2643         neg_op = INDEX_op_neg_vec;
2644         have_neg = (TCG_TARGET_HAS_neg_vec &&
2645                     tcg_can_emit_vec_op(neg_op, ctx->type, TCGOP_VECE(op)) > 0);
2646         break;
2647     default:
2648         g_assert_not_reached();
2649     }
2650     if (have_neg) {
2651         op->opc = neg_op;
2652         op->args[1] = op->args[2];
2653         return fold_neg_no_const(ctx, op);
2654     }
2655     return false;
2656 }
2657 
2658 /* We cannot as yet do_constant_folding with vectors. */
2659 static bool fold_sub_vec(OptContext *ctx, TCGOp *op)
2660 {
2661     if (fold_xx_to_i(ctx, op, 0) ||
2662         fold_xi_to_x(ctx, op, 0) ||
2663         fold_sub_to_neg(ctx, op)) {
2664         return true;
2665     }
2666     return finish_folding(ctx, op);
2667 }
2668 
2669 static bool fold_sub(OptContext *ctx, TCGOp *op)
2670 {
2671     if (fold_const2(ctx, op) ||
2672         fold_xx_to_i(ctx, op, 0) ||
2673         fold_xi_to_x(ctx, op, 0) ||
2674         fold_sub_to_neg(ctx, op)) {
2675         return true;
2676     }
2677 
2678     /* Fold sub r,x,i to add r,x,-i */
2679     if (arg_is_const(op->args[2])) {
2680         uint64_t val = arg_info(op->args[2])->val;
2681 
2682         op->opc = INDEX_op_add;
2683         op->args[2] = arg_new_constant(ctx, -val);
2684     }
2685     return finish_folding(ctx, op);
2686 }
2687 
2688 static bool fold_sub2(OptContext *ctx, TCGOp *op)
2689 {
2690     return fold_addsub2(ctx, op, false);
2691 }
2692 
2693 static bool fold_tcg_ld(OptContext *ctx, TCGOp *op)
2694 {
2695     uint64_t z_mask = -1, s_mask = 0;
2696 
2697     /* We can't do any folding with a load, but we can record bits. */
2698     switch (op->opc) {
2699     CASE_OP_32_64(ld8s):
2700         s_mask = INT8_MIN;
2701         break;
2702     CASE_OP_32_64(ld8u):
2703         z_mask = MAKE_64BIT_MASK(0, 8);
2704         break;
2705     CASE_OP_32_64(ld16s):
2706         s_mask = INT16_MIN;
2707         break;
2708     CASE_OP_32_64(ld16u):
2709         z_mask = MAKE_64BIT_MASK(0, 16);
2710         break;
2711     case INDEX_op_ld32s_i64:
2712         s_mask = INT32_MIN;
2713         break;
2714     case INDEX_op_ld32u_i64:
2715         z_mask = MAKE_64BIT_MASK(0, 32);
2716         break;
2717     default:
2718         g_assert_not_reached();
2719     }
2720     return fold_masks_zs(ctx, op, z_mask, s_mask);
2721 }
2722 
2723 static bool fold_tcg_ld_memcopy(OptContext *ctx, TCGOp *op)
2724 {
2725     TCGTemp *dst, *src;
2726     intptr_t ofs;
2727     TCGType type;
2728 
2729     if (op->args[1] != tcgv_ptr_arg(tcg_env)) {
2730         return finish_folding(ctx, op);
2731     }
2732 
2733     type = ctx->type;
2734     ofs = op->args[2];
2735     dst = arg_temp(op->args[0]);
2736     src = find_mem_copy_for(ctx, type, ofs);
2737     if (src && src->base_type == type) {
2738         return tcg_opt_gen_mov(ctx, op, temp_arg(dst), temp_arg(src));
2739     }
2740 
2741     reset_ts(ctx, dst);
2742     record_mem_copy(ctx, type, dst, ofs, ofs + tcg_type_size(type) - 1);
2743     return true;
2744 }
2745 
2746 static bool fold_tcg_st(OptContext *ctx, TCGOp *op)
2747 {
2748     intptr_t ofs = op->args[2];
2749     intptr_t lm1;
2750 
2751     if (op->args[1] != tcgv_ptr_arg(tcg_env)) {
2752         remove_mem_copy_all(ctx);
2753         return true;
2754     }
2755 
2756     switch (op->opc) {
2757     CASE_OP_32_64(st8):
2758         lm1 = 0;
2759         break;
2760     CASE_OP_32_64(st16):
2761         lm1 = 1;
2762         break;
2763     case INDEX_op_st32_i64:
2764     case INDEX_op_st_i32:
2765         lm1 = 3;
2766         break;
2767     case INDEX_op_st_i64:
2768         lm1 = 7;
2769         break;
2770     case INDEX_op_st_vec:
2771         lm1 = tcg_type_size(ctx->type) - 1;
2772         break;
2773     default:
2774         g_assert_not_reached();
2775     }
2776     remove_mem_copy_in(ctx, ofs, ofs + lm1);
2777     return true;
2778 }
2779 
2780 static bool fold_tcg_st_memcopy(OptContext *ctx, TCGOp *op)
2781 {
2782     TCGTemp *src;
2783     intptr_t ofs, last;
2784     TCGType type;
2785 
2786     if (op->args[1] != tcgv_ptr_arg(tcg_env)) {
2787         return fold_tcg_st(ctx, op);
2788     }
2789 
2790     src = arg_temp(op->args[0]);
2791     ofs = op->args[2];
2792     type = ctx->type;
2793 
2794     /*
2795      * Eliminate duplicate stores of a constant.
2796      * This happens frequently when the target ISA zero-extends.
2797      */
2798     if (ts_is_const(src)) {
2799         TCGTemp *prev = find_mem_copy_for(ctx, type, ofs);
2800         if (src == prev) {
2801             tcg_op_remove(ctx->tcg, op);
2802             return true;
2803         }
2804     }
2805 
2806     last = ofs + tcg_type_size(type) - 1;
2807     remove_mem_copy_in(ctx, ofs, last);
2808     record_mem_copy(ctx, type, src, ofs, last);
2809     return true;
2810 }
2811 
2812 static bool fold_xor(OptContext *ctx, TCGOp *op)
2813 {
2814     uint64_t z_mask, s_mask;
2815     TempOptInfo *t1, *t2;
2816 
2817     if (fold_const2_commutative(ctx, op) ||
2818         fold_xx_to_i(ctx, op, 0) ||
2819         fold_xi_to_x(ctx, op, 0) ||
2820         fold_xi_to_not(ctx, op, -1)) {
2821         return true;
2822     }
2823 
2824     t1 = arg_info(op->args[1]);
2825     t2 = arg_info(op->args[2]);
2826     z_mask = t1->z_mask | t2->z_mask;
2827     s_mask = t1->s_mask & t2->s_mask;
2828     return fold_masks_zs(ctx, op, z_mask, s_mask);
2829 }
2830 
2831 /* Propagate constants and copies, fold constant expressions. */
2832 void tcg_optimize(TCGContext *s)
2833 {
2834     int nb_temps, i;
2835     TCGOp *op, *op_next;
2836     OptContext ctx = { .tcg = s };
2837 
2838     QSIMPLEQ_INIT(&ctx.mem_free);
2839 
2840     /* Array VALS has an element for each temp.
2841        If this temp holds a constant then its value is kept in VALS' element.
2842        If this temp is a copy of other ones then the other copies are
2843        available through the doubly linked circular list. */
2844 
2845     nb_temps = s->nb_temps;
2846     for (i = 0; i < nb_temps; ++i) {
2847         s->temps[i].state_ptr = NULL;
2848     }
2849 
2850     QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) {
2851         TCGOpcode opc = op->opc;
2852         const TCGOpDef *def;
2853         bool done = false;
2854 
2855         /* Calls are special. */
2856         if (opc == INDEX_op_call) {
2857             fold_call(&ctx, op);
2858             continue;
2859         }
2860 
2861         def = &tcg_op_defs[opc];
2862         init_arguments(&ctx, op, def->nb_oargs + def->nb_iargs);
2863         copy_propagate(&ctx, op, def->nb_oargs, def->nb_iargs);
2864 
2865         /* Pre-compute the type of the operation. */
2866         ctx.type = TCGOP_TYPE(op);
2867 
2868         /*
2869          * Process each opcode.
2870          * Sorted alphabetically by opcode as much as possible.
2871          */
2872         switch (opc) {
2873         case INDEX_op_add:
2874             done = fold_add(&ctx, op);
2875             break;
2876         case INDEX_op_add_vec:
2877             done = fold_add_vec(&ctx, op);
2878             break;
2879         CASE_OP_32_64(add2):
2880             done = fold_add2(&ctx, op);
2881             break;
2882         case INDEX_op_and:
2883         case INDEX_op_and_vec:
2884             done = fold_and(&ctx, op);
2885             break;
2886         case INDEX_op_andc:
2887         case INDEX_op_andc_vec:
2888             done = fold_andc(&ctx, op);
2889             break;
2890         CASE_OP_32_64(brcond):
2891             done = fold_brcond(&ctx, op);
2892             break;
2893         case INDEX_op_brcond2_i32:
2894             done = fold_brcond2(&ctx, op);
2895             break;
2896         CASE_OP_32_64(bswap16):
2897         CASE_OP_32_64(bswap32):
2898         case INDEX_op_bswap64_i64:
2899             done = fold_bswap(&ctx, op);
2900             break;
2901         CASE_OP_32_64(clz):
2902         CASE_OP_32_64(ctz):
2903             done = fold_count_zeros(&ctx, op);
2904             break;
2905         CASE_OP_32_64(ctpop):
2906             done = fold_ctpop(&ctx, op);
2907             break;
2908         CASE_OP_32_64(deposit):
2909             done = fold_deposit(&ctx, op);
2910             break;
2911         case INDEX_op_divs:
2912         case INDEX_op_divu:
2913             done = fold_divide(&ctx, op);
2914             break;
2915         case INDEX_op_dup_vec:
2916             done = fold_dup(&ctx, op);
2917             break;
2918         case INDEX_op_dup2_vec:
2919             done = fold_dup2(&ctx, op);
2920             break;
2921         case INDEX_op_eqv:
2922         case INDEX_op_eqv_vec:
2923             done = fold_eqv(&ctx, op);
2924             break;
2925         CASE_OP_32_64(extract):
2926             done = fold_extract(&ctx, op);
2927             break;
2928         CASE_OP_32_64(extract2):
2929             done = fold_extract2(&ctx, op);
2930             break;
2931         case INDEX_op_ext_i32_i64:
2932             done = fold_exts(&ctx, op);
2933             break;
2934         case INDEX_op_extu_i32_i64:
2935         case INDEX_op_extrl_i64_i32:
2936         case INDEX_op_extrh_i64_i32:
2937             done = fold_extu(&ctx, op);
2938             break;
2939         CASE_OP_32_64(ld8s):
2940         CASE_OP_32_64(ld8u):
2941         CASE_OP_32_64(ld16s):
2942         CASE_OP_32_64(ld16u):
2943         case INDEX_op_ld32s_i64:
2944         case INDEX_op_ld32u_i64:
2945             done = fold_tcg_ld(&ctx, op);
2946             break;
2947         case INDEX_op_ld_i32:
2948         case INDEX_op_ld_i64:
2949         case INDEX_op_ld_vec:
2950             done = fold_tcg_ld_memcopy(&ctx, op);
2951             break;
2952         CASE_OP_32_64(st8):
2953         CASE_OP_32_64(st16):
2954         case INDEX_op_st32_i64:
2955             done = fold_tcg_st(&ctx, op);
2956             break;
2957         case INDEX_op_st_i32:
2958         case INDEX_op_st_i64:
2959         case INDEX_op_st_vec:
2960             done = fold_tcg_st_memcopy(&ctx, op);
2961             break;
2962         case INDEX_op_mb:
2963             done = fold_mb(&ctx, op);
2964             break;
2965         case INDEX_op_mov:
2966         case INDEX_op_mov_vec:
2967             done = fold_mov(&ctx, op);
2968             break;
2969         CASE_OP_32_64(movcond):
2970             done = fold_movcond(&ctx, op);
2971             break;
2972         case INDEX_op_mul:
2973             done = fold_mul(&ctx, op);
2974             break;
2975         case INDEX_op_mulsh:
2976         case INDEX_op_muluh:
2977             done = fold_mul_highpart(&ctx, op);
2978             break;
2979         CASE_OP_32_64(muls2):
2980         CASE_OP_32_64(mulu2):
2981             done = fold_multiply2(&ctx, op);
2982             break;
2983         case INDEX_op_nand:
2984         case INDEX_op_nand_vec:
2985             done = fold_nand(&ctx, op);
2986             break;
2987         case INDEX_op_neg:
2988             done = fold_neg(&ctx, op);
2989             break;
2990         case INDEX_op_nor:
2991         case INDEX_op_nor_vec:
2992             done = fold_nor(&ctx, op);
2993             break;
2994         case INDEX_op_not:
2995         case INDEX_op_not_vec:
2996             done = fold_not(&ctx, op);
2997             break;
2998         case INDEX_op_or:
2999         case INDEX_op_or_vec:
3000             done = fold_or(&ctx, op);
3001             break;
3002         case INDEX_op_orc:
3003         case INDEX_op_orc_vec:
3004             done = fold_orc(&ctx, op);
3005             break;
3006         case INDEX_op_qemu_ld_i32:
3007             done = fold_qemu_ld_1reg(&ctx, op);
3008             break;
3009         case INDEX_op_qemu_ld_i64:
3010             if (TCG_TARGET_REG_BITS == 64) {
3011                 done = fold_qemu_ld_1reg(&ctx, op);
3012                 break;
3013             }
3014             QEMU_FALLTHROUGH;
3015         case INDEX_op_qemu_ld_i128:
3016             done = fold_qemu_ld_2reg(&ctx, op);
3017             break;
3018         case INDEX_op_qemu_st8_i32:
3019         case INDEX_op_qemu_st_i32:
3020         case INDEX_op_qemu_st_i64:
3021         case INDEX_op_qemu_st_i128:
3022             done = fold_qemu_st(&ctx, op);
3023             break;
3024         CASE_OP_32_64(rem):
3025         CASE_OP_32_64(remu):
3026             done = fold_remainder(&ctx, op);
3027             break;
3028         CASE_OP_32_64(rotl):
3029         CASE_OP_32_64(rotr):
3030         CASE_OP_32_64(sar):
3031         CASE_OP_32_64(shl):
3032         CASE_OP_32_64(shr):
3033             done = fold_shift(&ctx, op);
3034             break;
3035         CASE_OP_32_64(setcond):
3036             done = fold_setcond(&ctx, op);
3037             break;
3038         CASE_OP_32_64(negsetcond):
3039             done = fold_negsetcond(&ctx, op);
3040             break;
3041         case INDEX_op_setcond2_i32:
3042             done = fold_setcond2(&ctx, op);
3043             break;
3044         case INDEX_op_cmp_vec:
3045             done = fold_cmp_vec(&ctx, op);
3046             break;
3047         case INDEX_op_cmpsel_vec:
3048             done = fold_cmpsel_vec(&ctx, op);
3049             break;
3050         case INDEX_op_bitsel_vec:
3051             done = fold_bitsel_vec(&ctx, op);
3052             break;
3053         CASE_OP_32_64(sextract):
3054             done = fold_sextract(&ctx, op);
3055             break;
3056         case INDEX_op_sub:
3057             done = fold_sub(&ctx, op);
3058             break;
3059         case INDEX_op_sub_vec:
3060             done = fold_sub_vec(&ctx, op);
3061             break;
3062         CASE_OP_32_64(sub2):
3063             done = fold_sub2(&ctx, op);
3064             break;
3065         case INDEX_op_xor:
3066         case INDEX_op_xor_vec:
3067             done = fold_xor(&ctx, op);
3068             break;
3069         case INDEX_op_set_label:
3070         case INDEX_op_br:
3071         case INDEX_op_exit_tb:
3072         case INDEX_op_goto_tb:
3073         case INDEX_op_goto_ptr:
3074             finish_ebb(&ctx);
3075             done = true;
3076             break;
3077         default:
3078             done = finish_folding(&ctx, op);
3079             break;
3080         }
3081         tcg_debug_assert(done);
3082     }
3083 }
3084