xref: /openbmc/qemu/target/hexagon/translate.c (revision c85cad81)
1 /*
2  *  Copyright(c) 2019-2023 Qualcomm Innovation Center, Inc. All Rights Reserved.
3  *
4  *  This program is free software; you can redistribute it and/or modify
5  *  it under the terms of the GNU General Public License as published by
6  *  the Free Software Foundation; either version 2 of the License, or
7  *  (at your option) any later version.
8  *
9  *  This program is distributed in the hope that it will be useful,
10  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
11  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  *  GNU General Public License for more details.
13  *
14  *  You should have received a copy of the GNU General Public License
15  *  along with this program; if not, see <http://www.gnu.org/licenses/>.
16  */
17 
18 #define QEMU_GENERATE
19 #include "qemu/osdep.h"
20 #include "cpu.h"
21 #include "tcg/tcg-op.h"
22 #include "tcg/tcg-op-gvec.h"
23 #include "exec/helper-gen.h"
24 #include "exec/helper-proto.h"
25 #include "exec/translation-block.h"
26 #include "exec/cpu_ldst.h"
27 #include "exec/log.h"
28 #include "internal.h"
29 #include "attribs.h"
30 #include "insn.h"
31 #include "decode.h"
32 #include "translate.h"
33 #include "genptr.h"
34 #include "printinsn.h"
35 
36 #define HELPER_H "helper.h"
37 #include "exec/helper-info.c.inc"
38 #undef  HELPER_H
39 
40 #include "analyze_funcs_generated.c.inc"
41 
42 typedef void (*AnalyzeInsn)(DisasContext *ctx);
43 static const AnalyzeInsn opcode_analyze[XX_LAST_OPCODE] = {
44 #define OPCODE(X)    [X] = analyze_##X
45 #include "opcodes_def_generated.h.inc"
46 #undef OPCODE
47 };
48 
49 TCGv hex_gpr[TOTAL_PER_THREAD_REGS];
50 TCGv hex_pred[NUM_PREGS];
51 TCGv hex_slot_cancelled;
52 TCGv hex_new_value_usr;
53 TCGv hex_reg_written[TOTAL_PER_THREAD_REGS];
54 TCGv hex_store_addr[STORES_MAX];
55 TCGv hex_store_width[STORES_MAX];
56 TCGv hex_store_val32[STORES_MAX];
57 TCGv_i64 hex_store_val64[STORES_MAX];
58 TCGv hex_llsc_addr;
59 TCGv hex_llsc_val;
60 TCGv_i64 hex_llsc_val_i64;
61 TCGv hex_vstore_addr[VSTORES_MAX];
62 TCGv hex_vstore_size[VSTORES_MAX];
63 TCGv hex_vstore_pending[VSTORES_MAX];
64 
65 static const char * const hexagon_prednames[] = {
66   "p0", "p1", "p2", "p3"
67 };
68 
69 intptr_t ctx_future_vreg_off(DisasContext *ctx, int regnum,
70                           int num, bool alloc_ok)
71 {
72     intptr_t offset;
73 
74     if (!ctx->need_commit) {
75         return offsetof(CPUHexagonState, VRegs[regnum]);
76     }
77 
78     /* See if it is already allocated */
79     for (int i = 0; i < ctx->future_vregs_idx; i++) {
80         if (ctx->future_vregs_num[i] == regnum) {
81             return offsetof(CPUHexagonState, future_VRegs[i]);
82         }
83     }
84 
85     g_assert(alloc_ok);
86     offset = offsetof(CPUHexagonState, future_VRegs[ctx->future_vregs_idx]);
87     for (int i = 0; i < num; i++) {
88         ctx->future_vregs_num[ctx->future_vregs_idx + i] = regnum++;
89     }
90     ctx->future_vregs_idx += num;
91     g_assert(ctx->future_vregs_idx <= VECTOR_TEMPS_MAX);
92     return offset;
93 }
94 
95 intptr_t ctx_tmp_vreg_off(DisasContext *ctx, int regnum,
96                           int num, bool alloc_ok)
97 {
98     intptr_t offset;
99 
100     /* See if it is already allocated */
101     for (int i = 0; i < ctx->tmp_vregs_idx; i++) {
102         if (ctx->tmp_vregs_num[i] == regnum) {
103             return offsetof(CPUHexagonState, tmp_VRegs[i]);
104         }
105     }
106 
107     g_assert(alloc_ok);
108     offset = offsetof(CPUHexagonState, tmp_VRegs[ctx->tmp_vregs_idx]);
109     for (int i = 0; i < num; i++) {
110         ctx->tmp_vregs_num[ctx->tmp_vregs_idx + i] = regnum++;
111     }
112     ctx->tmp_vregs_idx += num;
113     g_assert(ctx->tmp_vregs_idx <= VECTOR_TEMPS_MAX);
114     return offset;
115 }
116 
117 static void gen_exception_raw(int excp)
118 {
119     gen_helper_raise_exception(cpu_env, tcg_constant_i32(excp));
120 }
121 
122 static void gen_exec_counters(DisasContext *ctx)
123 {
124     tcg_gen_addi_tl(hex_gpr[HEX_REG_QEMU_PKT_CNT],
125                     hex_gpr[HEX_REG_QEMU_PKT_CNT], ctx->num_packets);
126     tcg_gen_addi_tl(hex_gpr[HEX_REG_QEMU_INSN_CNT],
127                     hex_gpr[HEX_REG_QEMU_INSN_CNT], ctx->num_insns);
128     tcg_gen_addi_tl(hex_gpr[HEX_REG_QEMU_HVX_CNT],
129                     hex_gpr[HEX_REG_QEMU_HVX_CNT], ctx->num_hvx_insns);
130 }
131 
132 static bool use_goto_tb(DisasContext *ctx, target_ulong dest)
133 {
134     return translator_use_goto_tb(&ctx->base, dest);
135 }
136 
137 static void gen_goto_tb(DisasContext *ctx, int idx, target_ulong dest, bool
138                         move_to_pc)
139 {
140     if (use_goto_tb(ctx, dest)) {
141         tcg_gen_goto_tb(idx);
142         if (move_to_pc) {
143             tcg_gen_movi_tl(hex_gpr[HEX_REG_PC], dest);
144         }
145         tcg_gen_exit_tb(ctx->base.tb, idx);
146     } else {
147         if (move_to_pc) {
148             tcg_gen_movi_tl(hex_gpr[HEX_REG_PC], dest);
149         }
150         tcg_gen_lookup_and_goto_ptr();
151     }
152 }
153 
154 static void gen_end_tb(DisasContext *ctx)
155 {
156     Packet *pkt = ctx->pkt;
157 
158     gen_exec_counters(ctx);
159 
160     if (ctx->branch_cond != TCG_COND_NEVER) {
161         if (ctx->branch_cond != TCG_COND_ALWAYS) {
162             TCGLabel *skip = gen_new_label();
163             tcg_gen_brcondi_tl(ctx->branch_cond, ctx->branch_taken, 0, skip);
164             gen_goto_tb(ctx, 0, ctx->branch_dest, true);
165             gen_set_label(skip);
166             gen_goto_tb(ctx, 1, ctx->next_PC, false);
167         } else {
168             gen_goto_tb(ctx, 0, ctx->branch_dest, true);
169         }
170     } else if (ctx->is_tight_loop &&
171                pkt->insn[pkt->num_insns - 1].opcode == J2_endloop0) {
172         /*
173          * When we're in a tight loop, we defer the endloop0 processing
174          * to take advantage of direct block chaining
175          */
176         TCGLabel *skip = gen_new_label();
177         tcg_gen_brcondi_tl(TCG_COND_LEU, hex_gpr[HEX_REG_LC0], 1, skip);
178         tcg_gen_subi_tl(hex_gpr[HEX_REG_LC0], hex_gpr[HEX_REG_LC0], 1);
179         gen_goto_tb(ctx, 0, ctx->base.tb->pc, true);
180         gen_set_label(skip);
181         gen_goto_tb(ctx, 1, ctx->next_PC, false);
182     } else {
183         tcg_gen_lookup_and_goto_ptr();
184     }
185 
186     ctx->base.is_jmp = DISAS_NORETURN;
187 }
188 
189 static void gen_exception_end_tb(DisasContext *ctx, int excp)
190 {
191     gen_exec_counters(ctx);
192     tcg_gen_movi_tl(hex_gpr[HEX_REG_PC], ctx->next_PC);
193     gen_exception_raw(excp);
194     ctx->base.is_jmp = DISAS_NORETURN;
195 
196 }
197 
198 #define PACKET_BUFFER_LEN              1028
199 static void print_pkt(Packet *pkt)
200 {
201     GString *buf = g_string_sized_new(PACKET_BUFFER_LEN);
202     snprint_a_pkt_debug(buf, pkt);
203     HEX_DEBUG_LOG("%s", buf->str);
204     g_string_free(buf, true);
205 }
206 #define HEX_DEBUG_PRINT_PKT(pkt) \
207     do { \
208         if (HEX_DEBUG) { \
209             print_pkt(pkt); \
210         } \
211     } while (0)
212 
213 static int read_packet_words(CPUHexagonState *env, DisasContext *ctx,
214                              uint32_t words[])
215 {
216     bool found_end = false;
217     int nwords, max_words;
218 
219     memset(words, 0, PACKET_WORDS_MAX * sizeof(uint32_t));
220     for (nwords = 0; !found_end && nwords < PACKET_WORDS_MAX; nwords++) {
221         words[nwords] =
222             translator_ldl(env, &ctx->base,
223                            ctx->base.pc_next + nwords * sizeof(uint32_t));
224         found_end = is_packet_end(words[nwords]);
225     }
226     if (!found_end) {
227         /* Read too many words without finding the end */
228         return 0;
229     }
230 
231     /* Check for page boundary crossing */
232     max_words = -(ctx->base.pc_next | TARGET_PAGE_MASK) / sizeof(uint32_t);
233     if (nwords > max_words) {
234         /* We can only cross a page boundary at the beginning of a TB */
235         g_assert(ctx->base.num_insns == 1);
236     }
237 
238     HEX_DEBUG_LOG("decode_packet: pc = 0x%x\n", ctx->base.pc_next);
239     HEX_DEBUG_LOG("    words = { ");
240     for (int i = 0; i < nwords; i++) {
241         HEX_DEBUG_LOG("0x%x, ", words[i]);
242     }
243     HEX_DEBUG_LOG("}\n");
244 
245     return nwords;
246 }
247 
248 static bool check_for_attrib(Packet *pkt, int attrib)
249 {
250     for (int i = 0; i < pkt->num_insns; i++) {
251         if (GET_ATTRIB(pkt->insn[i].opcode, attrib)) {
252             return true;
253         }
254     }
255     return false;
256 }
257 
258 static bool need_slot_cancelled(Packet *pkt)
259 {
260     /* We only need slot_cancelled for conditional store instructions */
261     for (int i = 0; i < pkt->num_insns; i++) {
262         uint16_t opcode = pkt->insn[i].opcode;
263         if (GET_ATTRIB(opcode, A_CONDEXEC) &&
264             GET_ATTRIB(opcode, A_SCALAR_STORE)) {
265             return true;
266         }
267     }
268     return false;
269 }
270 
271 static bool need_next_PC(DisasContext *ctx)
272 {
273     Packet *pkt = ctx->pkt;
274 
275     /* Check for conditional control flow or HW loop end */
276     for (int i = 0; i < pkt->num_insns; i++) {
277         uint16_t opcode = pkt->insn[i].opcode;
278         if (GET_ATTRIB(opcode, A_CONDEXEC) && GET_ATTRIB(opcode, A_COF)) {
279             return true;
280         }
281         if (GET_ATTRIB(opcode, A_HWLOOP0_END) ||
282             GET_ATTRIB(opcode, A_HWLOOP1_END)) {
283             return true;
284         }
285     }
286     return false;
287 }
288 
289 /*
290  * The opcode_analyze functions mark most of the writes in a packet
291  * However, there are some implicit writes marked as attributes
292  * of the applicable instructions.
293  */
294 static void mark_implicit_reg_write(DisasContext *ctx, int attrib, int rnum)
295 {
296     uint16_t opcode = ctx->insn->opcode;
297     if (GET_ATTRIB(opcode, attrib)) {
298         /*
299          * USR is used to set overflow and FP exceptions,
300          * so treat it as conditional
301          */
302         bool is_predicated = GET_ATTRIB(opcode, A_CONDEXEC) ||
303                              rnum == HEX_REG_USR;
304 
305         /* LC0/LC1 is conditionally written by endloop instructions */
306         if ((rnum == HEX_REG_LC0 || rnum == HEX_REG_LC1) &&
307             (opcode == J2_endloop0 ||
308              opcode == J2_endloop1 ||
309              opcode == J2_endloop01)) {
310             is_predicated = true;
311         }
312 
313         ctx_log_reg_write(ctx, rnum, is_predicated);
314     }
315 }
316 
317 static void mark_implicit_reg_writes(DisasContext *ctx)
318 {
319     mark_implicit_reg_write(ctx, A_IMPLICIT_WRITES_FP,  HEX_REG_FP);
320     mark_implicit_reg_write(ctx, A_IMPLICIT_WRITES_SP,  HEX_REG_SP);
321     mark_implicit_reg_write(ctx, A_IMPLICIT_WRITES_LR,  HEX_REG_LR);
322     mark_implicit_reg_write(ctx, A_IMPLICIT_WRITES_LC0, HEX_REG_LC0);
323     mark_implicit_reg_write(ctx, A_IMPLICIT_WRITES_SA0, HEX_REG_SA0);
324     mark_implicit_reg_write(ctx, A_IMPLICIT_WRITES_LC1, HEX_REG_LC1);
325     mark_implicit_reg_write(ctx, A_IMPLICIT_WRITES_SA1, HEX_REG_SA1);
326     mark_implicit_reg_write(ctx, A_IMPLICIT_WRITES_USR, HEX_REG_USR);
327     mark_implicit_reg_write(ctx, A_FPOP, HEX_REG_USR);
328 }
329 
330 static void mark_implicit_pred_write(DisasContext *ctx, int attrib, int pnum)
331 {
332     if (GET_ATTRIB(ctx->insn->opcode, attrib)) {
333         ctx_log_pred_write(ctx, pnum);
334     }
335 }
336 
337 static void mark_implicit_pred_writes(DisasContext *ctx)
338 {
339     mark_implicit_pred_write(ctx, A_IMPLICIT_WRITES_P0, 0);
340     mark_implicit_pred_write(ctx, A_IMPLICIT_WRITES_P1, 1);
341     mark_implicit_pred_write(ctx, A_IMPLICIT_WRITES_P2, 2);
342     mark_implicit_pred_write(ctx, A_IMPLICIT_WRITES_P3, 3);
343 }
344 
345 static bool pkt_raises_exception(Packet *pkt)
346 {
347     if (check_for_attrib(pkt, A_LOAD) ||
348         check_for_attrib(pkt, A_STORE)) {
349         return true;
350     }
351     return false;
352 }
353 
354 static bool need_commit(DisasContext *ctx)
355 {
356     Packet *pkt = ctx->pkt;
357 
358     /*
359      * If the short-circuit property is set to false, we'll always do the commit
360      */
361     if (!ctx->short_circuit) {
362         return true;
363     }
364 
365     if (pkt_raises_exception(pkt)) {
366         return true;
367     }
368 
369     /* Registers with immutability flags require new_value */
370     for (int i = 0; i < ctx->reg_log_idx; i++) {
371         int rnum = ctx->reg_log[i];
372         if (reg_immut_masks[rnum]) {
373             return true;
374         }
375     }
376 
377     /* Floating point instructions are hard-coded to use new_value */
378     if (check_for_attrib(pkt, A_FPOP)) {
379         return true;
380     }
381 
382     if (pkt->num_insns == 1) {
383         if (pkt->pkt_has_hvx) {
384             /*
385              * The HVX instructions with generated helpers use
386              * pass-by-reference, so they need the read/write overlap
387              * check below.
388              * The HVX instructions with overrides are OK.
389              */
390             if (!ctx->has_hvx_helper) {
391                 return false;
392             }
393         } else {
394             return false;
395         }
396     }
397 
398     /* Check for overlap between register reads and writes */
399     for (int i = 0; i < ctx->reg_log_idx; i++) {
400         int rnum = ctx->reg_log[i];
401         if (test_bit(rnum, ctx->regs_read)) {
402             return true;
403         }
404     }
405 
406     /* Check for overlap between predicate reads and writes */
407     for (int i = 0; i < ctx->preg_log_idx; i++) {
408         int pnum = ctx->preg_log[i];
409         if (test_bit(pnum, ctx->pregs_read)) {
410             return true;
411         }
412     }
413 
414     /* Check for overlap between HVX reads and writes */
415     for (int i = 0; i < ctx->vreg_log_idx; i++) {
416         int vnum = ctx->vreg_log[i];
417         if (test_bit(vnum, ctx->vregs_read)) {
418             return true;
419         }
420     }
421     if (!bitmap_empty(ctx->vregs_updated_tmp, NUM_VREGS)) {
422         int i = find_first_bit(ctx->vregs_updated_tmp, NUM_VREGS);
423         while (i < NUM_VREGS) {
424             if (test_bit(i, ctx->vregs_read)) {
425                 return true;
426             }
427             i = find_next_bit(ctx->vregs_updated_tmp, NUM_VREGS, i + 1);
428         }
429     }
430     if (!bitmap_empty(ctx->vregs_select, NUM_VREGS)) {
431         int i = find_first_bit(ctx->vregs_select, NUM_VREGS);
432         while (i < NUM_VREGS) {
433             if (test_bit(i, ctx->vregs_read)) {
434                 return true;
435             }
436             i = find_next_bit(ctx->vregs_select, NUM_VREGS, i + 1);
437         }
438     }
439 
440     /* Check for overlap between HVX predicate reads and writes */
441     for (int i = 0; i < ctx->qreg_log_idx; i++) {
442         int qnum = ctx->qreg_log[i];
443         if (test_bit(qnum, ctx->qregs_read)) {
444             return true;
445         }
446     }
447 
448     return false;
449 }
450 
451 static void mark_implicit_pred_read(DisasContext *ctx, int attrib, int pnum)
452 {
453     if (GET_ATTRIB(ctx->insn->opcode, attrib)) {
454         ctx_log_pred_read(ctx, pnum);
455     }
456 }
457 
458 static void mark_implicit_pred_reads(DisasContext *ctx)
459 {
460     mark_implicit_pred_read(ctx, A_IMPLICIT_READS_P0, 0);
461     mark_implicit_pred_read(ctx, A_IMPLICIT_READS_P1, 1);
462     mark_implicit_pred_read(ctx, A_IMPLICIT_READS_P3, 2);
463     mark_implicit_pred_read(ctx, A_IMPLICIT_READS_P3, 3);
464 }
465 
466 static void analyze_packet(DisasContext *ctx)
467 {
468     Packet *pkt = ctx->pkt;
469     ctx->has_hvx_helper = false;
470     for (int i = 0; i < pkt->num_insns; i++) {
471         Insn *insn = &pkt->insn[i];
472         ctx->insn = insn;
473         if (opcode_analyze[insn->opcode]) {
474             opcode_analyze[insn->opcode](ctx);
475         }
476         mark_implicit_reg_writes(ctx);
477         mark_implicit_pred_writes(ctx);
478         mark_implicit_pred_reads(ctx);
479     }
480 
481     ctx->need_commit = need_commit(ctx);
482 }
483 
484 static void gen_start_packet(DisasContext *ctx)
485 {
486     Packet *pkt = ctx->pkt;
487     target_ulong next_PC = ctx->base.pc_next + pkt->encod_pkt_size_in_bytes;
488     int i;
489 
490     /* Clear out the disassembly context */
491     ctx->next_PC = next_PC;
492     ctx->reg_log_idx = 0;
493     bitmap_zero(ctx->regs_written, TOTAL_PER_THREAD_REGS);
494     bitmap_zero(ctx->regs_read, TOTAL_PER_THREAD_REGS);
495     bitmap_zero(ctx->predicated_regs, TOTAL_PER_THREAD_REGS);
496     ctx->preg_log_idx = 0;
497     bitmap_zero(ctx->pregs_written, NUM_PREGS);
498     bitmap_zero(ctx->pregs_read, NUM_PREGS);
499     ctx->future_vregs_idx = 0;
500     ctx->tmp_vregs_idx = 0;
501     ctx->vreg_log_idx = 0;
502     bitmap_zero(ctx->vregs_updated_tmp, NUM_VREGS);
503     bitmap_zero(ctx->vregs_updated, NUM_VREGS);
504     bitmap_zero(ctx->vregs_select, NUM_VREGS);
505     bitmap_zero(ctx->predicated_future_vregs, NUM_VREGS);
506     bitmap_zero(ctx->predicated_tmp_vregs, NUM_VREGS);
507     bitmap_zero(ctx->vregs_read, NUM_VREGS);
508     bitmap_zero(ctx->qregs_read, NUM_QREGS);
509     ctx->qreg_log_idx = 0;
510     for (i = 0; i < STORES_MAX; i++) {
511         ctx->store_width[i] = 0;
512     }
513     ctx->s1_store_processed = false;
514     ctx->pre_commit = true;
515     for (i = 0; i < TOTAL_PER_THREAD_REGS; i++) {
516         ctx->new_value[i] = NULL;
517     }
518     for (i = 0; i < NUM_PREGS; i++) {
519         ctx->new_pred_value[i] = NULL;
520     }
521 
522     analyze_packet(ctx);
523 
524     /*
525      * pregs_written is used both in the analyze phase as well as the code
526      * gen phase, so clear it again.
527      */
528     bitmap_zero(ctx->pregs_written, NUM_PREGS);
529 
530     if (HEX_DEBUG) {
531         /* Handy place to set a breakpoint before the packet executes */
532         gen_helper_debug_start_packet(cpu_env);
533     }
534 
535     /* Initialize the runtime state for packet semantics */
536     if (need_slot_cancelled(pkt)) {
537         tcg_gen_movi_tl(hex_slot_cancelled, 0);
538     }
539     ctx->branch_taken = NULL;
540     if (pkt->pkt_has_cof) {
541         ctx->branch_taken = tcg_temp_new();
542         if (pkt->pkt_has_multi_cof) {
543             tcg_gen_movi_tl(ctx->branch_taken, 0);
544         }
545         if (need_next_PC(ctx)) {
546             tcg_gen_movi_tl(hex_gpr[HEX_REG_PC], next_PC);
547         }
548     }
549     if (HEX_DEBUG) {
550         ctx->pred_written = tcg_temp_new();
551         tcg_gen_movi_tl(ctx->pred_written, 0);
552     }
553 
554     /* Preload the predicated registers into get_result_gpr(ctx, i) */
555     if (ctx->need_commit &&
556         !bitmap_empty(ctx->predicated_regs, TOTAL_PER_THREAD_REGS)) {
557         int i = find_first_bit(ctx->predicated_regs, TOTAL_PER_THREAD_REGS);
558         while (i < TOTAL_PER_THREAD_REGS) {
559             tcg_gen_mov_tl(get_result_gpr(ctx, i), hex_gpr[i]);
560             i = find_next_bit(ctx->predicated_regs, TOTAL_PER_THREAD_REGS,
561                               i + 1);
562         }
563     }
564 
565     /*
566      * Preload the predicated pred registers into ctx->new_pred_value[pred_num]
567      * Only endloop instructions conditionally write to pred registers
568      */
569     if (ctx->need_commit && pkt->pkt_has_endloop) {
570         for (int i = 0; i < ctx->preg_log_idx; i++) {
571             int pred_num = ctx->preg_log[i];
572             ctx->new_pred_value[pred_num] = tcg_temp_new();
573             tcg_gen_mov_tl(ctx->new_pred_value[pred_num], hex_pred[pred_num]);
574         }
575     }
576 
577     /* Preload the predicated HVX registers into future_VRegs and tmp_VRegs */
578     if (!bitmap_empty(ctx->predicated_future_vregs, NUM_VREGS)) {
579         int i = find_first_bit(ctx->predicated_future_vregs, NUM_VREGS);
580         while (i < NUM_VREGS) {
581             const intptr_t VdV_off =
582                 ctx_future_vreg_off(ctx, i, 1, true);
583             intptr_t src_off = offsetof(CPUHexagonState, VRegs[i]);
584             tcg_gen_gvec_mov(MO_64, VdV_off,
585                              src_off,
586                              sizeof(MMVector),
587                              sizeof(MMVector));
588             i = find_next_bit(ctx->predicated_future_vregs, NUM_VREGS, i + 1);
589         }
590     }
591     if (!bitmap_empty(ctx->predicated_tmp_vregs, NUM_VREGS)) {
592         int i = find_first_bit(ctx->predicated_tmp_vregs, NUM_VREGS);
593         while (i < NUM_VREGS) {
594             const intptr_t VdV_off =
595                 ctx_tmp_vreg_off(ctx, i, 1, true);
596             intptr_t src_off = offsetof(CPUHexagonState, VRegs[i]);
597             tcg_gen_gvec_mov(MO_64, VdV_off,
598                              src_off,
599                              sizeof(MMVector),
600                              sizeof(MMVector));
601             i = find_next_bit(ctx->predicated_tmp_vregs, NUM_VREGS, i + 1);
602         }
603     }
604 }
605 
606 bool is_gather_store_insn(DisasContext *ctx)
607 {
608     Packet *pkt = ctx->pkt;
609     Insn *insn = ctx->insn;
610     if (GET_ATTRIB(insn->opcode, A_CVI_NEW) &&
611         insn->new_value_producer_slot == 1) {
612         /* Look for gather instruction */
613         for (int i = 0; i < pkt->num_insns; i++) {
614             Insn *in = &pkt->insn[i];
615             if (GET_ATTRIB(in->opcode, A_CVI_GATHER) && in->slot == 1) {
616                 return true;
617             }
618         }
619     }
620     return false;
621 }
622 
623 static void mark_store_width(DisasContext *ctx)
624 {
625     uint16_t opcode = ctx->insn->opcode;
626     uint32_t slot = ctx->insn->slot;
627     uint8_t width = 0;
628 
629     if (GET_ATTRIB(opcode, A_SCALAR_STORE)) {
630         if (GET_ATTRIB(opcode, A_MEMSIZE_0B)) {
631             return;
632         }
633         if (GET_ATTRIB(opcode, A_MEMSIZE_1B)) {
634             width |= 1;
635         }
636         if (GET_ATTRIB(opcode, A_MEMSIZE_2B)) {
637             width |= 2;
638         }
639         if (GET_ATTRIB(opcode, A_MEMSIZE_4B)) {
640             width |= 4;
641         }
642         if (GET_ATTRIB(opcode, A_MEMSIZE_8B)) {
643             width |= 8;
644         }
645         tcg_debug_assert(is_power_of_2(width));
646         ctx->store_width[slot] = width;
647     }
648 }
649 
650 static void gen_insn(DisasContext *ctx)
651 {
652     if (ctx->insn->generate) {
653         ctx->insn->generate(ctx);
654         mark_store_width(ctx);
655     } else {
656         gen_exception_end_tb(ctx, HEX_EXCP_INVALID_OPCODE);
657     }
658 }
659 
660 /*
661  * Helpers for generating the packet commit
662  */
663 static void gen_reg_writes(DisasContext *ctx)
664 {
665     int i;
666 
667     /* Early exit if not needed */
668     if (!ctx->need_commit) {
669         return;
670     }
671 
672     for (i = 0; i < ctx->reg_log_idx; i++) {
673         int reg_num = ctx->reg_log[i];
674 
675         tcg_gen_mov_tl(hex_gpr[reg_num], get_result_gpr(ctx, reg_num));
676 
677         /*
678          * ctx->is_tight_loop is set when SA0 points to the beginning of the TB.
679          * If we write to SA0, we have to turn off tight loop handling.
680          */
681         if (reg_num == HEX_REG_SA0) {
682             ctx->is_tight_loop = false;
683         }
684     }
685 }
686 
687 static void gen_pred_writes(DisasContext *ctx)
688 {
689     /* Early exit if not needed or the log is empty */
690     if (!ctx->need_commit || !ctx->preg_log_idx) {
691         return;
692     }
693 
694     for (int i = 0; i < ctx->preg_log_idx; i++) {
695         int pred_num = ctx->preg_log[i];
696         tcg_gen_mov_tl(hex_pred[pred_num], ctx->new_pred_value[pred_num]);
697     }
698 }
699 
700 static void gen_check_store_width(DisasContext *ctx, int slot_num)
701 {
702     if (HEX_DEBUG) {
703         TCGv slot = tcg_constant_tl(slot_num);
704         TCGv check = tcg_constant_tl(ctx->store_width[slot_num]);
705         gen_helper_debug_check_store_width(cpu_env, slot, check);
706     }
707 }
708 
709 static bool slot_is_predicated(Packet *pkt, int slot_num)
710 {
711     for (int i = 0; i < pkt->num_insns; i++) {
712         if (pkt->insn[i].slot == slot_num) {
713             return GET_ATTRIB(pkt->insn[i].opcode, A_CONDEXEC);
714         }
715     }
716     /* If we get to here, we didn't find an instruction in the requested slot */
717     g_assert_not_reached();
718 }
719 
720 void process_store(DisasContext *ctx, int slot_num)
721 {
722     bool is_predicated = slot_is_predicated(ctx->pkt, slot_num);
723     TCGLabel *label_end = NULL;
724 
725     /*
726      * We may have already processed this store
727      * See CHECK_NOSHUF in macros.h
728      */
729     if (slot_num == 1 && ctx->s1_store_processed) {
730         return;
731     }
732     ctx->s1_store_processed = true;
733 
734     if (is_predicated) {
735         TCGv cancelled = tcg_temp_new();
736         label_end = gen_new_label();
737 
738         /* Don't do anything if the slot was cancelled */
739         tcg_gen_extract_tl(cancelled, hex_slot_cancelled, slot_num, 1);
740         tcg_gen_brcondi_tl(TCG_COND_NE, cancelled, 0, label_end);
741     }
742     {
743         TCGv address = tcg_temp_new();
744         tcg_gen_mov_tl(address, hex_store_addr[slot_num]);
745 
746         /*
747          * If we know the width from the DisasContext, we can
748          * generate much cleaner code.
749          * Unfortunately, not all instructions execute the fSTORE
750          * macro during code generation.  Anything that uses the
751          * generic helper will have this problem.  Instructions
752          * that use fWRAP to generate proper TCG code will be OK.
753          */
754         switch (ctx->store_width[slot_num]) {
755         case 1:
756             gen_check_store_width(ctx, slot_num);
757             tcg_gen_qemu_st_tl(hex_store_val32[slot_num],
758                                hex_store_addr[slot_num],
759                                ctx->mem_idx, MO_UB);
760             break;
761         case 2:
762             gen_check_store_width(ctx, slot_num);
763             tcg_gen_qemu_st_tl(hex_store_val32[slot_num],
764                                hex_store_addr[slot_num],
765                                ctx->mem_idx, MO_TEUW);
766             break;
767         case 4:
768             gen_check_store_width(ctx, slot_num);
769             tcg_gen_qemu_st_tl(hex_store_val32[slot_num],
770                                hex_store_addr[slot_num],
771                                ctx->mem_idx, MO_TEUL);
772             break;
773         case 8:
774             gen_check_store_width(ctx, slot_num);
775             tcg_gen_qemu_st_i64(hex_store_val64[slot_num],
776                                 hex_store_addr[slot_num],
777                                 ctx->mem_idx, MO_TEUQ);
778             break;
779         default:
780             {
781                 /*
782                  * If we get to here, we don't know the width at
783                  * TCG generation time, we'll use a helper to
784                  * avoid branching based on the width at runtime.
785                  */
786                 TCGv slot = tcg_constant_tl(slot_num);
787                 gen_helper_commit_store(cpu_env, slot);
788             }
789         }
790     }
791     if (is_predicated) {
792         gen_set_label(label_end);
793     }
794 }
795 
796 static void process_store_log(DisasContext *ctx)
797 {
798     /*
799      *  When a packet has two stores, the hardware processes
800      *  slot 1 and then slot 0.  This will be important when
801      *  the memory accesses overlap.
802      */
803     Packet *pkt = ctx->pkt;
804     if (pkt->pkt_has_store_s1) {
805         g_assert(!pkt->pkt_has_dczeroa);
806         process_store(ctx, 1);
807     }
808     if (pkt->pkt_has_store_s0) {
809         g_assert(!pkt->pkt_has_dczeroa);
810         process_store(ctx, 0);
811     }
812 }
813 
814 /* Zero out a 32-bit cache line */
815 static void process_dczeroa(DisasContext *ctx)
816 {
817     if (ctx->pkt->pkt_has_dczeroa) {
818         /* Store 32 bytes of zero starting at (addr & ~0x1f) */
819         TCGv addr = tcg_temp_new();
820         TCGv_i64 zero = tcg_constant_i64(0);
821 
822         tcg_gen_andi_tl(addr, ctx->dczero_addr, ~0x1f);
823         tcg_gen_qemu_st_i64(zero, addr, ctx->mem_idx, MO_UQ);
824         tcg_gen_addi_tl(addr, addr, 8);
825         tcg_gen_qemu_st_i64(zero, addr, ctx->mem_idx, MO_UQ);
826         tcg_gen_addi_tl(addr, addr, 8);
827         tcg_gen_qemu_st_i64(zero, addr, ctx->mem_idx, MO_UQ);
828         tcg_gen_addi_tl(addr, addr, 8);
829         tcg_gen_qemu_st_i64(zero, addr, ctx->mem_idx, MO_UQ);
830     }
831 }
832 
833 static bool pkt_has_hvx_store(Packet *pkt)
834 {
835     int i;
836     for (i = 0; i < pkt->num_insns; i++) {
837         int opcode = pkt->insn[i].opcode;
838         if (GET_ATTRIB(opcode, A_CVI) && GET_ATTRIB(opcode, A_STORE)) {
839             return true;
840         }
841     }
842     return false;
843 }
844 
845 static void gen_commit_hvx(DisasContext *ctx)
846 {
847     int i;
848 
849     /* Early exit if not needed */
850     if (!ctx->need_commit) {
851         g_assert(!pkt_has_hvx_store(ctx->pkt));
852         return;
853     }
854 
855     /*
856      *    for (i = 0; i < ctx->vreg_log_idx; i++) {
857      *        int rnum = ctx->vreg_log[i];
858      *        env->VRegs[rnum] = env->future_VRegs[rnum];
859      *    }
860      */
861     for (i = 0; i < ctx->vreg_log_idx; i++) {
862         int rnum = ctx->vreg_log[i];
863         intptr_t dstoff = offsetof(CPUHexagonState, VRegs[rnum]);
864         intptr_t srcoff = ctx_future_vreg_off(ctx, rnum, 1, false);
865         size_t size = sizeof(MMVector);
866 
867         tcg_gen_gvec_mov(MO_64, dstoff, srcoff, size, size);
868     }
869 
870     /*
871      *    for (i = 0; i < ctx->qreg_log_idx; i++) {
872      *        int rnum = ctx->qreg_log[i];
873      *        env->QRegs[rnum] = env->future_QRegs[rnum];
874      *    }
875      */
876     for (i = 0; i < ctx->qreg_log_idx; i++) {
877         int rnum = ctx->qreg_log[i];
878         intptr_t dstoff = offsetof(CPUHexagonState, QRegs[rnum]);
879         intptr_t srcoff = offsetof(CPUHexagonState, future_QRegs[rnum]);
880         size_t size = sizeof(MMQReg);
881 
882         tcg_gen_gvec_mov(MO_64, dstoff, srcoff, size, size);
883     }
884 
885     if (pkt_has_hvx_store(ctx->pkt)) {
886         gen_helper_commit_hvx_stores(cpu_env);
887     }
888 }
889 
890 static void update_exec_counters(DisasContext *ctx)
891 {
892     Packet *pkt = ctx->pkt;
893     int num_insns = pkt->num_insns;
894     int num_real_insns = 0;
895     int num_hvx_insns = 0;
896 
897     for (int i = 0; i < num_insns; i++) {
898         if (!pkt->insn[i].is_endloop &&
899             !pkt->insn[i].part1 &&
900             !GET_ATTRIB(pkt->insn[i].opcode, A_IT_NOP)) {
901             num_real_insns++;
902         }
903         if (GET_ATTRIB(pkt->insn[i].opcode, A_CVI)) {
904             num_hvx_insns++;
905         }
906     }
907 
908     ctx->num_packets++;
909     ctx->num_insns += num_real_insns;
910     ctx->num_hvx_insns += num_hvx_insns;
911 }
912 
913 static void gen_commit_packet(DisasContext *ctx)
914 {
915     /*
916      * If there is more than one store in a packet, make sure they are all OK
917      * before proceeding with the rest of the packet commit.
918      *
919      * dczeroa has to be the only store operation in the packet, so we go
920      * ahead and process that first.
921      *
922      * When there is an HVX store, there can also be a scalar store in either
923      * slot 0 or slot1, so we create a mask for the helper to indicate what
924      * work to do.
925      *
926      * When there are two scalar stores, we probe the one in slot 0.
927      *
928      * Note that we don't call the probe helper for packets with only one
929      * store.  Therefore, we call process_store_log before anything else
930      * involved in committing the packet.
931      */
932     Packet *pkt = ctx->pkt;
933     bool has_store_s0 = pkt->pkt_has_store_s0;
934     bool has_store_s1 = (pkt->pkt_has_store_s1 && !ctx->s1_store_processed);
935     bool has_hvx_store = pkt_has_hvx_store(pkt);
936     if (pkt->pkt_has_dczeroa) {
937         /*
938          * The dczeroa will be the store in slot 0, check that we don't have
939          * a store in slot 1 or an HVX store.
940          */
941         g_assert(!has_store_s1 && !has_hvx_store);
942         process_dczeroa(ctx);
943     } else if (has_hvx_store) {
944         if (!has_store_s0 && !has_store_s1) {
945             TCGv mem_idx = tcg_constant_tl(ctx->mem_idx);
946             gen_helper_probe_hvx_stores(cpu_env, mem_idx);
947         } else {
948             int mask = 0;
949 
950             if (has_store_s0) {
951                 mask =
952                     FIELD_DP32(mask, PROBE_PKT_SCALAR_HVX_STORES, HAS_ST0, 1);
953             }
954             if (has_store_s1) {
955                 mask =
956                     FIELD_DP32(mask, PROBE_PKT_SCALAR_HVX_STORES, HAS_ST1, 1);
957             }
958             if (has_hvx_store) {
959                 mask =
960                     FIELD_DP32(mask, PROBE_PKT_SCALAR_HVX_STORES,
961                                HAS_HVX_STORES, 1);
962             }
963             if (has_store_s0 && slot_is_predicated(pkt, 0)) {
964                 mask =
965                     FIELD_DP32(mask, PROBE_PKT_SCALAR_HVX_STORES,
966                                S0_IS_PRED, 1);
967             }
968             if (has_store_s1 && slot_is_predicated(pkt, 1)) {
969                 mask =
970                     FIELD_DP32(mask, PROBE_PKT_SCALAR_HVX_STORES,
971                                S1_IS_PRED, 1);
972             }
973             mask = FIELD_DP32(mask, PROBE_PKT_SCALAR_HVX_STORES, MMU_IDX,
974                               ctx->mem_idx);
975             gen_helper_probe_pkt_scalar_hvx_stores(cpu_env,
976                                                    tcg_constant_tl(mask));
977         }
978     } else if (has_store_s0 && has_store_s1) {
979         /*
980          * process_store_log will execute the slot 1 store first,
981          * so we only have to probe the store in slot 0
982          */
983         int args = 0;
984         args =
985             FIELD_DP32(args, PROBE_PKT_SCALAR_STORE_S0, MMU_IDX, ctx->mem_idx);
986         if (slot_is_predicated(pkt, 0)) {
987             args =
988                 FIELD_DP32(args, PROBE_PKT_SCALAR_STORE_S0, IS_PREDICATED, 1);
989         }
990         TCGv args_tcgv = tcg_constant_tl(args);
991         gen_helper_probe_pkt_scalar_store_s0(cpu_env, args_tcgv);
992     }
993 
994     process_store_log(ctx);
995 
996     gen_reg_writes(ctx);
997     gen_pred_writes(ctx);
998     if (pkt->pkt_has_hvx) {
999         gen_commit_hvx(ctx);
1000     }
1001     update_exec_counters(ctx);
1002     if (HEX_DEBUG) {
1003         TCGv has_st0 =
1004             tcg_constant_tl(pkt->pkt_has_store_s0 && !pkt->pkt_has_dczeroa);
1005         TCGv has_st1 =
1006             tcg_constant_tl(pkt->pkt_has_store_s1 && !pkt->pkt_has_dczeroa);
1007 
1008         /* Handy place to set a breakpoint at the end of execution */
1009         gen_helper_debug_commit_end(cpu_env, tcg_constant_tl(ctx->pkt->pc),
1010                                     ctx->pred_written, has_st0, has_st1);
1011     }
1012 
1013     if (pkt->vhist_insn != NULL) {
1014         ctx->pre_commit = false;
1015         ctx->insn = pkt->vhist_insn;
1016         pkt->vhist_insn->generate(ctx);
1017     }
1018 
1019     if (pkt->pkt_has_cof) {
1020         gen_end_tb(ctx);
1021     }
1022 }
1023 
1024 static void decode_and_translate_packet(CPUHexagonState *env, DisasContext *ctx)
1025 {
1026     uint32_t words[PACKET_WORDS_MAX];
1027     int nwords;
1028     Packet pkt;
1029     int i;
1030 
1031     nwords = read_packet_words(env, ctx, words);
1032     if (!nwords) {
1033         gen_exception_end_tb(ctx, HEX_EXCP_INVALID_PACKET);
1034         return;
1035     }
1036 
1037     if (decode_packet(nwords, words, &pkt, false) > 0) {
1038         pkt.pc = ctx->base.pc_next;
1039         HEX_DEBUG_PRINT_PKT(&pkt);
1040         ctx->pkt = &pkt;
1041         gen_start_packet(ctx);
1042         for (i = 0; i < pkt.num_insns; i++) {
1043             ctx->insn = &pkt.insn[i];
1044             gen_insn(ctx);
1045         }
1046         gen_commit_packet(ctx);
1047         ctx->base.pc_next += pkt.encod_pkt_size_in_bytes;
1048     } else {
1049         gen_exception_end_tb(ctx, HEX_EXCP_INVALID_PACKET);
1050     }
1051 }
1052 
1053 static void hexagon_tr_init_disas_context(DisasContextBase *dcbase,
1054                                           CPUState *cs)
1055 {
1056     DisasContext *ctx = container_of(dcbase, DisasContext, base);
1057     HexagonCPU *hex_cpu = env_archcpu(cs->env_ptr);
1058     uint32_t hex_flags = dcbase->tb->flags;
1059 
1060     ctx->mem_idx = MMU_USER_IDX;
1061     ctx->num_packets = 0;
1062     ctx->num_insns = 0;
1063     ctx->num_hvx_insns = 0;
1064     ctx->branch_cond = TCG_COND_NEVER;
1065     ctx->is_tight_loop = FIELD_EX32(hex_flags, TB_FLAGS, IS_TIGHT_LOOP);
1066     ctx->short_circuit = hex_cpu->short_circuit;
1067 }
1068 
1069 static void hexagon_tr_tb_start(DisasContextBase *db, CPUState *cpu)
1070 {
1071 }
1072 
1073 static void hexagon_tr_insn_start(DisasContextBase *dcbase, CPUState *cpu)
1074 {
1075     DisasContext *ctx = container_of(dcbase, DisasContext, base);
1076 
1077     tcg_gen_insn_start(ctx->base.pc_next);
1078 }
1079 
1080 static bool pkt_crosses_page(CPUHexagonState *env, DisasContext *ctx)
1081 {
1082     target_ulong page_start = ctx->base.pc_first & TARGET_PAGE_MASK;
1083     bool found_end = false;
1084     int nwords;
1085 
1086     for (nwords = 0; !found_end && nwords < PACKET_WORDS_MAX; nwords++) {
1087         uint32_t word = cpu_ldl_code(env,
1088                             ctx->base.pc_next + nwords * sizeof(uint32_t));
1089         found_end = is_packet_end(word);
1090     }
1091     uint32_t next_ptr =  ctx->base.pc_next + nwords * sizeof(uint32_t);
1092     return found_end && next_ptr - page_start >= TARGET_PAGE_SIZE;
1093 }
1094 
1095 static void hexagon_tr_translate_packet(DisasContextBase *dcbase, CPUState *cpu)
1096 {
1097     DisasContext *ctx = container_of(dcbase, DisasContext, base);
1098     CPUHexagonState *env = cpu->env_ptr;
1099 
1100     decode_and_translate_packet(env, ctx);
1101 
1102     if (ctx->base.is_jmp == DISAS_NEXT) {
1103         target_ulong page_start = ctx->base.pc_first & TARGET_PAGE_MASK;
1104         target_ulong bytes_max = PACKET_WORDS_MAX * sizeof(target_ulong);
1105 
1106         if (ctx->base.pc_next - page_start >= TARGET_PAGE_SIZE ||
1107             (ctx->base.pc_next - page_start >= TARGET_PAGE_SIZE - bytes_max &&
1108              pkt_crosses_page(env, ctx))) {
1109             ctx->base.is_jmp = DISAS_TOO_MANY;
1110         }
1111 
1112         /*
1113          * The CPU log is used to compare against LLDB single stepping,
1114          * so end the TLB after every packet.
1115          */
1116         HexagonCPU *hex_cpu = env_archcpu(env);
1117         if (hex_cpu->lldb_compat && qemu_loglevel_mask(CPU_LOG_TB_CPU)) {
1118             ctx->base.is_jmp = DISAS_TOO_MANY;
1119         }
1120     }
1121 }
1122 
1123 static void hexagon_tr_tb_stop(DisasContextBase *dcbase, CPUState *cpu)
1124 {
1125     DisasContext *ctx = container_of(dcbase, DisasContext, base);
1126 
1127     switch (ctx->base.is_jmp) {
1128     case DISAS_TOO_MANY:
1129         gen_exec_counters(ctx);
1130         tcg_gen_movi_tl(hex_gpr[HEX_REG_PC], ctx->base.pc_next);
1131         tcg_gen_exit_tb(NULL, 0);
1132         break;
1133     case DISAS_NORETURN:
1134         break;
1135     default:
1136         g_assert_not_reached();
1137     }
1138 }
1139 
1140 static void hexagon_tr_disas_log(const DisasContextBase *dcbase,
1141                                  CPUState *cpu, FILE *logfile)
1142 {
1143     fprintf(logfile, "IN: %s\n", lookup_symbol(dcbase->pc_first));
1144     target_disas(logfile, cpu, dcbase->pc_first, dcbase->tb->size);
1145 }
1146 
1147 
1148 static const TranslatorOps hexagon_tr_ops = {
1149     .init_disas_context = hexagon_tr_init_disas_context,
1150     .tb_start           = hexagon_tr_tb_start,
1151     .insn_start         = hexagon_tr_insn_start,
1152     .translate_insn     = hexagon_tr_translate_packet,
1153     .tb_stop            = hexagon_tr_tb_stop,
1154     .disas_log          = hexagon_tr_disas_log,
1155 };
1156 
1157 void gen_intermediate_code(CPUState *cs, TranslationBlock *tb, int *max_insns,
1158                            target_ulong pc, void *host_pc)
1159 {
1160     DisasContext ctx;
1161 
1162     translator_loop(cs, tb, max_insns, pc, host_pc,
1163                     &hexagon_tr_ops, &ctx.base);
1164 }
1165 
1166 #define NAME_LEN               64
1167 static char reg_written_names[TOTAL_PER_THREAD_REGS][NAME_LEN];
1168 static char store_addr_names[STORES_MAX][NAME_LEN];
1169 static char store_width_names[STORES_MAX][NAME_LEN];
1170 static char store_val32_names[STORES_MAX][NAME_LEN];
1171 static char store_val64_names[STORES_MAX][NAME_LEN];
1172 static char vstore_addr_names[VSTORES_MAX][NAME_LEN];
1173 static char vstore_size_names[VSTORES_MAX][NAME_LEN];
1174 static char vstore_pending_names[VSTORES_MAX][NAME_LEN];
1175 
1176 void hexagon_translate_init(void)
1177 {
1178     int i;
1179 
1180     opcode_init();
1181 
1182     for (i = 0; i < TOTAL_PER_THREAD_REGS; i++) {
1183         hex_gpr[i] = tcg_global_mem_new(cpu_env,
1184             offsetof(CPUHexagonState, gpr[i]),
1185             hexagon_regnames[i]);
1186 
1187         if (HEX_DEBUG) {
1188             snprintf(reg_written_names[i], NAME_LEN, "reg_written_%s",
1189                      hexagon_regnames[i]);
1190             hex_reg_written[i] = tcg_global_mem_new(cpu_env,
1191                 offsetof(CPUHexagonState, reg_written[i]),
1192                 reg_written_names[i]);
1193         }
1194     }
1195     hex_new_value_usr = tcg_global_mem_new(cpu_env,
1196         offsetof(CPUHexagonState, new_value_usr), "new_value_usr");
1197 
1198     for (i = 0; i < NUM_PREGS; i++) {
1199         hex_pred[i] = tcg_global_mem_new(cpu_env,
1200             offsetof(CPUHexagonState, pred[i]),
1201             hexagon_prednames[i]);
1202     }
1203     hex_slot_cancelled = tcg_global_mem_new(cpu_env,
1204         offsetof(CPUHexagonState, slot_cancelled), "slot_cancelled");
1205     hex_llsc_addr = tcg_global_mem_new(cpu_env,
1206         offsetof(CPUHexagonState, llsc_addr), "llsc_addr");
1207     hex_llsc_val = tcg_global_mem_new(cpu_env,
1208         offsetof(CPUHexagonState, llsc_val), "llsc_val");
1209     hex_llsc_val_i64 = tcg_global_mem_new_i64(cpu_env,
1210         offsetof(CPUHexagonState, llsc_val_i64), "llsc_val_i64");
1211     for (i = 0; i < STORES_MAX; i++) {
1212         snprintf(store_addr_names[i], NAME_LEN, "store_addr_%d", i);
1213         hex_store_addr[i] = tcg_global_mem_new(cpu_env,
1214             offsetof(CPUHexagonState, mem_log_stores[i].va),
1215             store_addr_names[i]);
1216 
1217         snprintf(store_width_names[i], NAME_LEN, "store_width_%d", i);
1218         hex_store_width[i] = tcg_global_mem_new(cpu_env,
1219             offsetof(CPUHexagonState, mem_log_stores[i].width),
1220             store_width_names[i]);
1221 
1222         snprintf(store_val32_names[i], NAME_LEN, "store_val32_%d", i);
1223         hex_store_val32[i] = tcg_global_mem_new(cpu_env,
1224             offsetof(CPUHexagonState, mem_log_stores[i].data32),
1225             store_val32_names[i]);
1226 
1227         snprintf(store_val64_names[i], NAME_LEN, "store_val64_%d", i);
1228         hex_store_val64[i] = tcg_global_mem_new_i64(cpu_env,
1229             offsetof(CPUHexagonState, mem_log_stores[i].data64),
1230             store_val64_names[i]);
1231     }
1232     for (int i = 0; i < VSTORES_MAX; i++) {
1233         snprintf(vstore_addr_names[i], NAME_LEN, "vstore_addr_%d", i);
1234         hex_vstore_addr[i] = tcg_global_mem_new(cpu_env,
1235             offsetof(CPUHexagonState, vstore[i].va),
1236             vstore_addr_names[i]);
1237 
1238         snprintf(vstore_size_names[i], NAME_LEN, "vstore_size_%d", i);
1239         hex_vstore_size[i] = tcg_global_mem_new(cpu_env,
1240             offsetof(CPUHexagonState, vstore[i].size),
1241             vstore_size_names[i]);
1242 
1243         snprintf(vstore_pending_names[i], NAME_LEN, "vstore_pending_%d", i);
1244         hex_vstore_pending[i] = tcg_global_mem_new(cpu_env,
1245             offsetof(CPUHexagonState, vstore_pending[i]),
1246             vstore_pending_names[i]);
1247     }
1248 }
1249