1/*
2 * Tiny Code Generator for QEMU
3 *
4 * Copyright (c) 2021 WANG Xuerui <git@xen0n.name>
5 *
6 * Based on tcg/riscv/tcg-target.c.inc
7 *
8 * Copyright (c) 2018 SiFive, Inc
9 * Copyright (c) 2008-2009 Arnaud Patard <arnaud.patard@rtp-net.org>
10 * Copyright (c) 2009 Aurelien Jarno <aurelien@aurel32.net>
11 * Copyright (c) 2008 Fabrice Bellard
12 *
13 * Permission is hereby granted, free of charge, to any person obtaining a copy
14 * of this software and associated documentation files (the "Software"), to deal
15 * in the Software without restriction, including without limitation the rights
16 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
17 * copies of the Software, and to permit persons to whom the Software is
18 * furnished to do so, subject to the following conditions:
19 *
20 * The above copyright notice and this permission notice shall be included in
21 * all copies or substantial portions of the Software.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
24 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
25 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
26 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
27 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
28 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
29 * THE SOFTWARE.
30 */
31
32#include "../tcg-ldst.c.inc"
33#include <asm/hwcap.h>
34
35#ifdef CONFIG_DEBUG_TCG
36static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
37    "zero",
38    "ra",
39    "tp",
40    "sp",
41    "a0",
42    "a1",
43    "a2",
44    "a3",
45    "a4",
46    "a5",
47    "a6",
48    "a7",
49    "t0",
50    "t1",
51    "t2",
52    "t3",
53    "t4",
54    "t5",
55    "t6",
56    "t7",
57    "t8",
58    "r21", /* reserved in the LP64* ABI, hence no ABI name */
59    "s9",
60    "s0",
61    "s1",
62    "s2",
63    "s3",
64    "s4",
65    "s5",
66    "s6",
67    "s7",
68    "s8",
69    "vr0",
70    "vr1",
71    "vr2",
72    "vr3",
73    "vr4",
74    "vr5",
75    "vr6",
76    "vr7",
77    "vr8",
78    "vr9",
79    "vr10",
80    "vr11",
81    "vr12",
82    "vr13",
83    "vr14",
84    "vr15",
85    "vr16",
86    "vr17",
87    "vr18",
88    "vr19",
89    "vr20",
90    "vr21",
91    "vr22",
92    "vr23",
93    "vr24",
94    "vr25",
95    "vr26",
96    "vr27",
97    "vr28",
98    "vr29",
99    "vr30",
100    "vr31",
101};
102#endif
103
104static const int tcg_target_reg_alloc_order[] = {
105    /* Registers preserved across calls */
106    /* TCG_REG_S0 reserved for TCG_AREG0 */
107    TCG_REG_S1,
108    TCG_REG_S2,
109    TCG_REG_S3,
110    TCG_REG_S4,
111    TCG_REG_S5,
112    TCG_REG_S6,
113    TCG_REG_S7,
114    TCG_REG_S8,
115    TCG_REG_S9,
116
117    /* Registers (potentially) clobbered across calls */
118    TCG_REG_T0,
119    TCG_REG_T1,
120    TCG_REG_T2,
121    TCG_REG_T3,
122    TCG_REG_T4,
123    TCG_REG_T5,
124    TCG_REG_T6,
125    TCG_REG_T7,
126    TCG_REG_T8,
127
128    /* Argument registers, opposite order of allocation.  */
129    TCG_REG_A7,
130    TCG_REG_A6,
131    TCG_REG_A5,
132    TCG_REG_A4,
133    TCG_REG_A3,
134    TCG_REG_A2,
135    TCG_REG_A1,
136    TCG_REG_A0,
137
138    /* Vector registers */
139    TCG_REG_V0, TCG_REG_V1, TCG_REG_V2, TCG_REG_V3,
140    TCG_REG_V4, TCG_REG_V5, TCG_REG_V6, TCG_REG_V7,
141    TCG_REG_V8, TCG_REG_V9, TCG_REG_V10, TCG_REG_V11,
142    TCG_REG_V12, TCG_REG_V13, TCG_REG_V14, TCG_REG_V15,
143    TCG_REG_V16, TCG_REG_V17, TCG_REG_V18, TCG_REG_V19,
144    TCG_REG_V20, TCG_REG_V21, TCG_REG_V22, TCG_REG_V23,
145    /* V24 - V31 are caller-saved, and skipped.  */
146};
147
148static const int tcg_target_call_iarg_regs[] = {
149    TCG_REG_A0,
150    TCG_REG_A1,
151    TCG_REG_A2,
152    TCG_REG_A3,
153    TCG_REG_A4,
154    TCG_REG_A5,
155    TCG_REG_A6,
156    TCG_REG_A7,
157};
158
159static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
160{
161    tcg_debug_assert(kind == TCG_CALL_RET_NORMAL);
162    tcg_debug_assert(slot >= 0 && slot <= 1);
163    return TCG_REG_A0 + slot;
164}
165
166#define TCG_GUEST_BASE_REG TCG_REG_S1
167
168#define TCG_CT_CONST_ZERO  0x100
169#define TCG_CT_CONST_S12   0x200
170#define TCG_CT_CONST_S32   0x400
171#define TCG_CT_CONST_U12   0x800
172#define TCG_CT_CONST_C12   0x1000
173#define TCG_CT_CONST_WSZ   0x2000
174#define TCG_CT_CONST_VCMP  0x4000
175#define TCG_CT_CONST_VADD  0x8000
176
177#define ALL_GENERAL_REGS   MAKE_64BIT_MASK(0, 32)
178#define ALL_VECTOR_REGS    MAKE_64BIT_MASK(32, 32)
179
180static inline tcg_target_long sextreg(tcg_target_long val, int pos, int len)
181{
182    return sextract64(val, pos, len);
183}
184
185/* test if a constant matches the constraint */
186static bool tcg_target_const_match(int64_t val, int ct,
187                                   TCGType type, TCGCond cond, int vece)
188{
189    if (ct & TCG_CT_CONST) {
190        return true;
191    }
192    if ((ct & TCG_CT_CONST_ZERO) && val == 0) {
193        return true;
194    }
195    if ((ct & TCG_CT_CONST_S12) && val == sextreg(val, 0, 12)) {
196        return true;
197    }
198    if ((ct & TCG_CT_CONST_S32) && val == (int32_t)val) {
199        return true;
200    }
201    if ((ct & TCG_CT_CONST_U12) && val >= 0 && val <= 0xfff) {
202        return true;
203    }
204    if ((ct & TCG_CT_CONST_C12) && ~val >= 0 && ~val <= 0xfff) {
205        return true;
206    }
207    if ((ct & TCG_CT_CONST_WSZ) && val == (type == TCG_TYPE_I32 ? 32 : 64)) {
208        return true;
209    }
210    int64_t vec_val = sextract64(val, 0, 8 << vece);
211    if ((ct & TCG_CT_CONST_VCMP) && -0x10 <= vec_val && vec_val <= 0x1f) {
212        return true;
213    }
214    if ((ct & TCG_CT_CONST_VADD) && -0x1f <= vec_val && vec_val <= 0x1f) {
215        return true;
216    }
217    return false;
218}
219
220/*
221 * Relocations
222 */
223
224/*
225 * Relocation records defined in LoongArch ELF psABI v1.00 is way too
226 * complicated; a whopping stack machine is needed to stuff the fields, at
227 * the very least one SOP_PUSH and one SOP_POP (of the correct format) are
228 * needed.
229 *
230 * Hence, define our own simpler relocation types. Numbers are chosen as to
231 * not collide with potential future additions to the true ELF relocation
232 * type enum.
233 */
234
235/* Field Sk16, shifted right by 2; suitable for conditional jumps */
236#define R_LOONGARCH_BR_SK16     256
237/* Field Sd10k16, shifted right by 2; suitable for B and BL */
238#define R_LOONGARCH_BR_SD10K16  257
239
240static bool reloc_br_sk16(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
241{
242    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
243    intptr_t offset = (intptr_t)target - (intptr_t)src_rx;
244
245    tcg_debug_assert((offset & 3) == 0);
246    offset >>= 2;
247    if (offset == sextreg(offset, 0, 16)) {
248        *src_rw = deposit64(*src_rw, 10, 16, offset);
249        return true;
250    }
251
252    return false;
253}
254
255static bool reloc_br_sd10k16(tcg_insn_unit *src_rw,
256                             const tcg_insn_unit *target)
257{
258    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
259    intptr_t offset = (intptr_t)target - (intptr_t)src_rx;
260
261    tcg_debug_assert((offset & 3) == 0);
262    offset >>= 2;
263    if (offset == sextreg(offset, 0, 26)) {
264        *src_rw = deposit64(*src_rw, 0, 10, offset >> 16); /* slot d10 */
265        *src_rw = deposit64(*src_rw, 10, 16, offset); /* slot k16 */
266        return true;
267    }
268
269    return false;
270}
271
272static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
273                        intptr_t value, intptr_t addend)
274{
275    tcg_debug_assert(addend == 0);
276    switch (type) {
277    case R_LOONGARCH_BR_SK16:
278        return reloc_br_sk16(code_ptr, (tcg_insn_unit *)value);
279    case R_LOONGARCH_BR_SD10K16:
280        return reloc_br_sd10k16(code_ptr, (tcg_insn_unit *)value);
281    default:
282        g_assert_not_reached();
283    }
284}
285
286#include "tcg-insn-defs.c.inc"
287
288/*
289 * TCG intrinsics
290 */
291
292static void tcg_out_mb(TCGContext *s, TCGArg a0)
293{
294    /* Baseline LoongArch only has the full barrier, unfortunately.  */
295    tcg_out_opc_dbar(s, 0);
296}
297
298static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
299{
300    if (ret == arg) {
301        return true;
302    }
303    switch (type) {
304    case TCG_TYPE_I32:
305    case TCG_TYPE_I64:
306        if (ret < TCG_REG_V0) {
307            if (arg < TCG_REG_V0) {
308                /*
309                 * Conventional register-register move used in LoongArch is
310                 * `or dst, src, zero`.
311                 */
312                tcg_out_opc_or(s, ret, arg, TCG_REG_ZERO);
313            } else {
314                tcg_out_opc_movfr2gr_d(s, ret, arg);
315            }
316        } else {
317            if (arg < TCG_REG_V0) {
318                tcg_out_opc_movgr2fr_d(s, ret, arg);
319            } else {
320                tcg_out_opc_fmov_d(s, ret, arg);
321            }
322        }
323        break;
324    case TCG_TYPE_V64:
325    case TCG_TYPE_V128:
326        tcg_out_opc_vori_b(s, ret, arg, 0);
327        break;
328    default:
329        g_assert_not_reached();
330    }
331    return true;
332}
333
334/* Loads a 32-bit immediate into rd, sign-extended.  */
335static void tcg_out_movi_i32(TCGContext *s, TCGReg rd, int32_t val)
336{
337    tcg_target_long lo = sextreg(val, 0, 12);
338    tcg_target_long hi12 = sextreg(val, 12, 20);
339
340    /* Single-instruction cases.  */
341    if (hi12 == 0) {
342        /* val fits in uimm12: ori rd, zero, val */
343        tcg_out_opc_ori(s, rd, TCG_REG_ZERO, val);
344        return;
345    }
346    if (hi12 == sextreg(lo, 12, 20)) {
347        /* val fits in simm12: addi.w rd, zero, val */
348        tcg_out_opc_addi_w(s, rd, TCG_REG_ZERO, val);
349        return;
350    }
351
352    /* High bits must be set; load with lu12i.w + optional ori.  */
353    tcg_out_opc_lu12i_w(s, rd, hi12);
354    if (lo != 0) {
355        tcg_out_opc_ori(s, rd, rd, lo & 0xfff);
356    }
357}
358
359static void tcg_out_movi(TCGContext *s, TCGType type, TCGReg rd,
360                         tcg_target_long val)
361{
362    /*
363     * LoongArch conventionally loads 64-bit immediates in at most 4 steps,
364     * with dedicated instructions for filling the respective bitfields
365     * below:
366     *
367     *        6                   5                   4               3
368     *  3 2 1 0 9 8 7 6 5 4 3 2 1 0 9 8 7 6 5 4 3 2 1 0 9 8 7 6 5 4 3 2
369     * +-----------------------+---------------------------------------+...
370     * |          hi52         |                  hi32                 |
371     * +-----------------------+---------------------------------------+...
372     *       3                   2                   1
373     *     1 0 9 8 7 6 5 4 3 2 1 0 9 8 7 6 5 4 3 2 1 0 9 8 7 6 5 4 3 2 1 0
374     * ...+-------------------------------------+-------------------------+
375     *    |                 hi12                |            lo           |
376     * ...+-------------------------------------+-------------------------+
377     *
378     * Check if val belong to one of the several fast cases, before falling
379     * back to the slow path.
380     */
381
382    intptr_t pc_offset;
383    tcg_target_long val_lo, val_hi, pc_hi, offset_hi;
384    tcg_target_long hi12, hi32, hi52;
385
386    /* Value fits in signed i32.  */
387    if (type == TCG_TYPE_I32 || val == (int32_t)val) {
388        tcg_out_movi_i32(s, rd, val);
389        return;
390    }
391
392    /* PC-relative cases.  */
393    pc_offset = tcg_pcrel_diff(s, (void *)val);
394    if (pc_offset == sextreg(pc_offset, 0, 22) && (pc_offset & 3) == 0) {
395        /* Single pcaddu2i.  */
396        tcg_out_opc_pcaddu2i(s, rd, pc_offset >> 2);
397        return;
398    }
399
400    if (pc_offset == (int32_t)pc_offset) {
401        /* Offset within 32 bits; load with pcalau12i + ori.  */
402        val_lo = sextreg(val, 0, 12);
403        val_hi = val >> 12;
404        pc_hi = (val - pc_offset) >> 12;
405        offset_hi = val_hi - pc_hi;
406
407        tcg_debug_assert(offset_hi == sextreg(offset_hi, 0, 20));
408        tcg_out_opc_pcalau12i(s, rd, offset_hi);
409        if (val_lo != 0) {
410            tcg_out_opc_ori(s, rd, rd, val_lo & 0xfff);
411        }
412        return;
413    }
414
415    hi12 = sextreg(val, 12, 20);
416    hi32 = sextreg(val, 32, 20);
417    hi52 = sextreg(val, 52, 12);
418
419    /* Single cu52i.d case.  */
420    if ((hi52 != 0) && (ctz64(val) >= 52)) {
421        tcg_out_opc_cu52i_d(s, rd, TCG_REG_ZERO, hi52);
422        return;
423    }
424
425    /* Slow path.  Initialize the low 32 bits, then concat high bits.  */
426    tcg_out_movi_i32(s, rd, val);
427
428    /* Load hi32 and hi52 explicitly when they are unexpected values. */
429    if (hi32 != sextreg(hi12, 20, 20)) {
430        tcg_out_opc_cu32i_d(s, rd, hi32);
431    }
432
433    if (hi52 != sextreg(hi32, 20, 12)) {
434        tcg_out_opc_cu52i_d(s, rd, rd, hi52);
435    }
436}
437
438static void tcg_out_addi(TCGContext *s, TCGType type, TCGReg rd,
439                         TCGReg rs, tcg_target_long imm)
440{
441    tcg_target_long lo12 = sextreg(imm, 0, 12);
442    tcg_target_long hi16 = sextreg(imm - lo12, 16, 16);
443
444    /*
445     * Note that there's a hole in between hi16 and lo12:
446     *
447     *       3                   2                   1                   0
448     *     1 0 9 8 7 6 5 4 3 2 1 0 9 8 7 6 5 4 3 2 1 0 9 8 7 6 5 4 3 2 1 0
449     * ...+-------------------------------+-------+-----------------------+
450     *    |             hi16              |       |          lo12         |
451     * ...+-------------------------------+-------+-----------------------+
452     *
453     * For bits within that hole, it's more efficient to use LU12I and ADD.
454     */
455    if (imm == (hi16 << 16) + lo12) {
456        if (hi16) {
457            tcg_out_opc_addu16i_d(s, rd, rs, hi16);
458            rs = rd;
459        }
460        if (type == TCG_TYPE_I32) {
461            tcg_out_opc_addi_w(s, rd, rs, lo12);
462        } else if (lo12) {
463            tcg_out_opc_addi_d(s, rd, rs, lo12);
464        } else {
465            tcg_out_mov(s, type, rd, rs);
466        }
467    } else {
468        tcg_out_movi(s, type, TCG_REG_TMP0, imm);
469        if (type == TCG_TYPE_I32) {
470            tcg_out_opc_add_w(s, rd, rs, TCG_REG_TMP0);
471        } else {
472            tcg_out_opc_add_d(s, rd, rs, TCG_REG_TMP0);
473        }
474    }
475}
476
477static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
478{
479    return false;
480}
481
482static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
483                             tcg_target_long imm)
484{
485    /* This function is only used for passing structs by reference. */
486    g_assert_not_reached();
487}
488
489static void tcg_out_ext8u(TCGContext *s, TCGReg ret, TCGReg arg)
490{
491    tcg_out_opc_andi(s, ret, arg, 0xff);
492}
493
494static void tcg_out_ext16u(TCGContext *s, TCGReg ret, TCGReg arg)
495{
496    tcg_out_opc_bstrpick_w(s, ret, arg, 0, 15);
497}
498
499static void tcg_out_ext32u(TCGContext *s, TCGReg ret, TCGReg arg)
500{
501    tcg_out_opc_bstrpick_d(s, ret, arg, 0, 31);
502}
503
504static void tcg_out_ext8s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
505{
506    tcg_out_opc_sext_b(s, ret, arg);
507}
508
509static void tcg_out_ext16s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
510{
511    tcg_out_opc_sext_h(s, ret, arg);
512}
513
514static void tcg_out_ext32s(TCGContext *s, TCGReg ret, TCGReg arg)
515{
516    tcg_out_opc_addi_w(s, ret, arg, 0);
517}
518
519static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg)
520{
521    if (ret != arg) {
522        tcg_out_ext32s(s, ret, arg);
523    }
524}
525
526static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg)
527{
528    tcg_out_ext32u(s, ret, arg);
529}
530
531static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg ret, TCGReg arg)
532{
533    tcg_out_ext32s(s, ret, arg);
534}
535
536static void tcg_out_clzctz(TCGContext *s, LoongArchInsn opc,
537                           TCGReg a0, TCGReg a1, TCGReg a2,
538                           bool c2, bool is_32bit)
539{
540    if (c2) {
541        /*
542         * Fast path: semantics already satisfied due to constraint and
543         * insn behavior, single instruction is enough.
544         */
545        tcg_debug_assert(a2 == (is_32bit ? 32 : 64));
546        /* all clz/ctz insns belong to DJ-format */
547        tcg_out32(s, encode_dj_insn(opc, a0, a1));
548        return;
549    }
550
551    tcg_out32(s, encode_dj_insn(opc, TCG_REG_TMP0, a1));
552    /* a0 = a1 ? REG_TMP0 : a2 */
553    tcg_out_opc_maskeqz(s, TCG_REG_TMP0, TCG_REG_TMP0, a1);
554    tcg_out_opc_masknez(s, a0, a2, a1);
555    tcg_out_opc_or(s, a0, TCG_REG_TMP0, a0);
556}
557
558#define SETCOND_INV    TCG_TARGET_NB_REGS
559#define SETCOND_NEZ    (SETCOND_INV << 1)
560#define SETCOND_FLAGS  (SETCOND_INV | SETCOND_NEZ)
561
562static int tcg_out_setcond_int(TCGContext *s, TCGCond cond, TCGReg ret,
563                               TCGReg arg1, tcg_target_long arg2, bool c2)
564{
565    int flags = 0;
566
567    switch (cond) {
568    case TCG_COND_EQ:    /* -> NE  */
569    case TCG_COND_GE:    /* -> LT  */
570    case TCG_COND_GEU:   /* -> LTU */
571    case TCG_COND_GT:    /* -> LE  */
572    case TCG_COND_GTU:   /* -> LEU */
573        cond = tcg_invert_cond(cond);
574        flags ^= SETCOND_INV;
575        break;
576    default:
577        break;
578    }
579
580    switch (cond) {
581    case TCG_COND_LE:
582    case TCG_COND_LEU:
583        /*
584         * If we have a constant input, the most efficient way to implement
585         * LE is by adding 1 and using LT.  Watch out for wrap around for LEU.
586         * We don't need to care for this for LE because the constant input
587         * is still constrained to int32_t, and INT32_MAX+1 is representable
588         * in the 64-bit temporary register.
589         */
590        if (c2) {
591            if (cond == TCG_COND_LEU) {
592                /* unsigned <= -1 is true */
593                if (arg2 == -1) {
594                    tcg_out_movi(s, TCG_TYPE_REG, ret, !(flags & SETCOND_INV));
595                    return ret;
596                }
597                cond = TCG_COND_LTU;
598            } else {
599                cond = TCG_COND_LT;
600            }
601            arg2 += 1;
602        } else {
603            TCGReg tmp = arg2;
604            arg2 = arg1;
605            arg1 = tmp;
606            cond = tcg_swap_cond(cond);    /* LE -> GE */
607            cond = tcg_invert_cond(cond);  /* GE -> LT */
608            flags ^= SETCOND_INV;
609        }
610        break;
611    default:
612        break;
613    }
614
615    switch (cond) {
616    case TCG_COND_NE:
617        flags |= SETCOND_NEZ;
618        if (!c2) {
619            tcg_out_opc_xor(s, ret, arg1, arg2);
620        } else if (arg2 == 0) {
621            ret = arg1;
622        } else if (arg2 >= 0 && arg2 <= 0xfff) {
623            tcg_out_opc_xori(s, ret, arg1, arg2);
624        } else {
625            tcg_out_addi(s, TCG_TYPE_REG, ret, arg1, -arg2);
626        }
627        break;
628
629    case TCG_COND_LT:
630    case TCG_COND_LTU:
631        if (c2) {
632            if (arg2 >= -0x800 && arg2 <= 0x7ff) {
633                if (cond == TCG_COND_LT) {
634                    tcg_out_opc_slti(s, ret, arg1, arg2);
635                } else {
636                    tcg_out_opc_sltui(s, ret, arg1, arg2);
637                }
638                break;
639            }
640            tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_TMP0, arg2);
641            arg2 = TCG_REG_TMP0;
642        }
643        if (cond == TCG_COND_LT) {
644            tcg_out_opc_slt(s, ret, arg1, arg2);
645        } else {
646            tcg_out_opc_sltu(s, ret, arg1, arg2);
647        }
648        break;
649
650    default:
651        g_assert_not_reached();
652        break;
653    }
654
655    return ret | flags;
656}
657
658static void tcg_out_setcond(TCGContext *s, TCGCond cond, TCGReg ret,
659                            TCGReg arg1, tcg_target_long arg2, bool c2)
660{
661    int tmpflags = tcg_out_setcond_int(s, cond, ret, arg1, arg2, c2);
662
663    if (tmpflags != ret) {
664        TCGReg tmp = tmpflags & ~SETCOND_FLAGS;
665
666        switch (tmpflags & SETCOND_FLAGS) {
667        case SETCOND_INV:
668            /* Intermediate result is boolean: simply invert. */
669            tcg_out_opc_xori(s, ret, tmp, 1);
670            break;
671        case SETCOND_NEZ:
672            /* Intermediate result is zero/non-zero: test != 0. */
673            tcg_out_opc_sltu(s, ret, TCG_REG_ZERO, tmp);
674            break;
675        case SETCOND_NEZ | SETCOND_INV:
676            /* Intermediate result is zero/non-zero: test == 0. */
677            tcg_out_opc_sltui(s, ret, tmp, 1);
678            break;
679        default:
680            g_assert_not_reached();
681        }
682    }
683}
684
685static void tcg_out_movcond(TCGContext *s, TCGCond cond, TCGReg ret,
686                            TCGReg c1, tcg_target_long c2, bool const2,
687                            TCGReg v1, TCGReg v2)
688{
689    int tmpflags = tcg_out_setcond_int(s, cond, TCG_REG_TMP0, c1, c2, const2);
690    TCGReg t;
691
692    /* Standardize the test below to t != 0. */
693    if (tmpflags & SETCOND_INV) {
694        t = v1, v1 = v2, v2 = t;
695    }
696
697    t = tmpflags & ~SETCOND_FLAGS;
698    if (v1 == TCG_REG_ZERO) {
699        tcg_out_opc_masknez(s, ret, v2, t);
700    } else if (v2 == TCG_REG_ZERO) {
701        tcg_out_opc_maskeqz(s, ret, v1, t);
702    } else {
703        tcg_out_opc_masknez(s, TCG_REG_TMP2, v2, t); /* t ? 0 : v2 */
704        tcg_out_opc_maskeqz(s, TCG_REG_TMP1, v1, t); /* t ? v1 : 0 */
705        tcg_out_opc_or(s, ret, TCG_REG_TMP1, TCG_REG_TMP2);
706    }
707}
708
709/*
710 * Branch helpers
711 */
712
713static const struct {
714    LoongArchInsn op;
715    bool swap;
716} tcg_brcond_to_loongarch[] = {
717    [TCG_COND_EQ] =  { OPC_BEQ,  false },
718    [TCG_COND_NE] =  { OPC_BNE,  false },
719    [TCG_COND_LT] =  { OPC_BGT,  true  },
720    [TCG_COND_GE] =  { OPC_BLE,  true  },
721    [TCG_COND_LE] =  { OPC_BLE,  false },
722    [TCG_COND_GT] =  { OPC_BGT,  false },
723    [TCG_COND_LTU] = { OPC_BGTU, true  },
724    [TCG_COND_GEU] = { OPC_BLEU, true  },
725    [TCG_COND_LEU] = { OPC_BLEU, false },
726    [TCG_COND_GTU] = { OPC_BGTU, false }
727};
728
729static void tcg_out_brcond(TCGContext *s, TCGCond cond, TCGReg arg1,
730                           TCGReg arg2, TCGLabel *l)
731{
732    LoongArchInsn op = tcg_brcond_to_loongarch[cond].op;
733
734    tcg_debug_assert(op != 0);
735
736    if (tcg_brcond_to_loongarch[cond].swap) {
737        TCGReg t = arg1;
738        arg1 = arg2;
739        arg2 = t;
740    }
741
742    /* all conditional branch insns belong to DJSk16-format */
743    tcg_out_reloc(s, s->code_ptr, R_LOONGARCH_BR_SK16, l, 0);
744    tcg_out32(s, encode_djsk16_insn(op, arg1, arg2, 0));
745}
746
747static void tcg_out_call_int(TCGContext *s, const tcg_insn_unit *arg, bool tail)
748{
749    TCGReg link = tail ? TCG_REG_ZERO : TCG_REG_RA;
750    ptrdiff_t offset = tcg_pcrel_diff(s, arg);
751
752    tcg_debug_assert((offset & 3) == 0);
753    if (offset == sextreg(offset, 0, 28)) {
754        /* short jump: +/- 256MiB */
755        if (tail) {
756            tcg_out_opc_b(s, offset >> 2);
757        } else {
758            tcg_out_opc_bl(s, offset >> 2);
759        }
760    } else if (offset == sextreg(offset, 0, 38)) {
761        /* long jump: +/- 256GiB */
762        tcg_target_long lo = sextreg(offset, 0, 18);
763        tcg_target_long hi = offset - lo;
764        tcg_out_opc_pcaddu18i(s, TCG_REG_TMP0, hi >> 18);
765        tcg_out_opc_jirl(s, link, TCG_REG_TMP0, lo >> 2);
766    } else {
767        /* far jump: 64-bit */
768        tcg_target_long lo = sextreg((tcg_target_long)arg, 0, 18);
769        tcg_target_long hi = (tcg_target_long)arg - lo;
770        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, hi);
771        tcg_out_opc_jirl(s, link, TCG_REG_TMP0, lo >> 2);
772    }
773}
774
775static void tcg_out_call(TCGContext *s, const tcg_insn_unit *arg,
776                         const TCGHelperInfo *info)
777{
778    tcg_out_call_int(s, arg, false);
779}
780
781/*
782 * Load/store helpers
783 */
784
785static void tcg_out_ldst(TCGContext *s, LoongArchInsn opc, TCGReg data,
786                         TCGReg addr, intptr_t offset)
787{
788    intptr_t imm12 = sextreg(offset, 0, 12);
789
790    if (offset != imm12) {
791        intptr_t diff = tcg_pcrel_diff(s, (void *)offset);
792
793        if (addr == TCG_REG_ZERO && diff == (int32_t)diff) {
794            imm12 = sextreg(diff, 0, 12);
795            tcg_out_opc_pcaddu12i(s, TCG_REG_TMP2, (diff - imm12) >> 12);
796        } else {
797            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP2, offset - imm12);
798            if (addr != TCG_REG_ZERO) {
799                tcg_out_opc_add_d(s, TCG_REG_TMP2, TCG_REG_TMP2, addr);
800            }
801        }
802        addr = TCG_REG_TMP2;
803    }
804
805    switch (opc) {
806    case OPC_LD_B:
807    case OPC_LD_BU:
808    case OPC_LD_H:
809    case OPC_LD_HU:
810    case OPC_LD_W:
811    case OPC_LD_WU:
812    case OPC_LD_D:
813    case OPC_ST_B:
814    case OPC_ST_H:
815    case OPC_ST_W:
816    case OPC_ST_D:
817        tcg_out32(s, encode_djsk12_insn(opc, data, addr, imm12));
818        break;
819    case OPC_FLD_S:
820    case OPC_FLD_D:
821    case OPC_FST_S:
822    case OPC_FST_D:
823        tcg_out32(s, encode_fdjsk12_insn(opc, data, addr, imm12));
824        break;
825    default:
826        g_assert_not_reached();
827    }
828}
829
830static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg dest,
831                       TCGReg base, intptr_t offset)
832{
833    switch (type) {
834    case TCG_TYPE_I32:
835        if (dest < TCG_REG_V0) {
836            tcg_out_ldst(s, OPC_LD_W, dest, base, offset);
837        } else {
838            tcg_out_ldst(s, OPC_FLD_S, dest, base, offset);
839        }
840        break;
841    case TCG_TYPE_I64:
842    case TCG_TYPE_V64:
843        if (dest < TCG_REG_V0) {
844            tcg_out_ldst(s, OPC_LD_D, dest, base, offset);
845        } else {
846            tcg_out_ldst(s, OPC_FLD_D, dest, base, offset);
847        }
848        break;
849    case TCG_TYPE_V128:
850        if (-0x800 <= offset && offset <= 0x7ff) {
851            tcg_out_opc_vld(s, dest, base, offset);
852        } else {
853            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, offset);
854            tcg_out_opc_vldx(s, dest, base, TCG_REG_TMP0);
855        }
856        break;
857    default:
858        g_assert_not_reached();
859    }
860}
861
862static void tcg_out_st(TCGContext *s, TCGType type, TCGReg src,
863                       TCGReg base, intptr_t offset)
864{
865    switch (type) {
866    case TCG_TYPE_I32:
867        if (src < TCG_REG_V0) {
868            tcg_out_ldst(s, OPC_ST_W, src, base, offset);
869        } else {
870            tcg_out_ldst(s, OPC_FST_S, src, base, offset);
871        }
872        break;
873    case TCG_TYPE_I64:
874    case TCG_TYPE_V64:
875        if (src < TCG_REG_V0) {
876            tcg_out_ldst(s, OPC_ST_D, src, base, offset);
877        } else {
878            tcg_out_ldst(s, OPC_FST_D, src, base, offset);
879        }
880        break;
881    case TCG_TYPE_V128:
882        if (-0x800 <= offset && offset <= 0x7ff) {
883            tcg_out_opc_vst(s, src, base, offset);
884        } else {
885            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, offset);
886            tcg_out_opc_vstx(s, src, base, TCG_REG_TMP0);
887        }
888        break;
889    default:
890        g_assert_not_reached();
891    }
892}
893
894static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
895                        TCGReg base, intptr_t ofs)
896{
897    if (val == 0) {
898        tcg_out_st(s, type, TCG_REG_ZERO, base, ofs);
899        return true;
900    }
901    return false;
902}
903
904/*
905 * Load/store helpers for SoftMMU, and qemu_ld/st implementations
906 */
907
908static bool tcg_out_goto(TCGContext *s, const tcg_insn_unit *target)
909{
910    tcg_out_opc_b(s, 0);
911    return reloc_br_sd10k16(s->code_ptr - 1, target);
912}
913
914static const TCGLdstHelperParam ldst_helper_param = {
915    .ntmp = 1, .tmp = { TCG_REG_TMP0 }
916};
917
918static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
919{
920    MemOp opc = get_memop(l->oi);
921
922    /* resolve label address */
923    if (!reloc_br_sk16(l->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
924        return false;
925    }
926
927    tcg_out_ld_helper_args(s, l, &ldst_helper_param);
928    tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SIZE], false);
929    tcg_out_ld_helper_ret(s, l, false, &ldst_helper_param);
930    return tcg_out_goto(s, l->raddr);
931}
932
933static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
934{
935    MemOp opc = get_memop(l->oi);
936
937    /* resolve label address */
938    if (!reloc_br_sk16(l->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
939        return false;
940    }
941
942    tcg_out_st_helper_args(s, l, &ldst_helper_param);
943    tcg_out_call_int(s, qemu_st_helpers[opc & MO_SIZE], false);
944    return tcg_out_goto(s, l->raddr);
945}
946
947typedef struct {
948    TCGReg base;
949    TCGReg index;
950    TCGAtomAlign aa;
951} HostAddress;
952
953bool tcg_target_has_memory_bswap(MemOp memop)
954{
955    return false;
956}
957
958/* We expect to use a 12-bit negative offset from ENV.  */
959#define MIN_TLB_MASK_TABLE_OFS  -(1 << 11)
960
961/*
962 * For system-mode, perform the TLB load and compare.
963 * For user-mode, perform any required alignment tests.
964 * In both cases, return a TCGLabelQemuLdst structure if the slow path
965 * is required and fill in @h with the host address for the fast path.
966 */
967static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
968                                           TCGReg addr_reg, MemOpIdx oi,
969                                           bool is_ld)
970{
971    TCGType addr_type = s->addr_type;
972    TCGLabelQemuLdst *ldst = NULL;
973    MemOp opc = get_memop(oi);
974    MemOp a_bits;
975
976    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, false);
977    a_bits = h->aa.align;
978
979    if (tcg_use_softmmu) {
980        unsigned s_bits = opc & MO_SIZE;
981        int mem_index = get_mmuidx(oi);
982        int fast_ofs = tlb_mask_table_ofs(s, mem_index);
983        int mask_ofs = fast_ofs + offsetof(CPUTLBDescFast, mask);
984        int table_ofs = fast_ofs + offsetof(CPUTLBDescFast, table);
985
986        ldst = new_ldst_label(s);
987        ldst->is_ld = is_ld;
988        ldst->oi = oi;
989        ldst->addrlo_reg = addr_reg;
990
991        tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP0, TCG_AREG0, mask_ofs);
992        tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP1, TCG_AREG0, table_ofs);
993
994        tcg_out_opc_srli_d(s, TCG_REG_TMP2, addr_reg,
995                           s->page_bits - CPU_TLB_ENTRY_BITS);
996        tcg_out_opc_and(s, TCG_REG_TMP2, TCG_REG_TMP2, TCG_REG_TMP0);
997        tcg_out_opc_add_d(s, TCG_REG_TMP2, TCG_REG_TMP2, TCG_REG_TMP1);
998
999        /* Load the tlb comparator and the addend.  */
1000        QEMU_BUILD_BUG_ON(HOST_BIG_ENDIAN);
1001        tcg_out_ld(s, addr_type, TCG_REG_TMP0, TCG_REG_TMP2,
1002                   is_ld ? offsetof(CPUTLBEntry, addr_read)
1003                         : offsetof(CPUTLBEntry, addr_write));
1004        tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP2, TCG_REG_TMP2,
1005                   offsetof(CPUTLBEntry, addend));
1006
1007        /*
1008         * For aligned accesses, we check the first byte and include the
1009         * alignment bits within the address.  For unaligned access, we
1010         * check that we don't cross pages using the address of the last
1011         * byte of the access.
1012         */
1013        if (a_bits < s_bits) {
1014            unsigned a_mask = (1u << a_bits) - 1;
1015            unsigned s_mask = (1u << s_bits) - 1;
1016            tcg_out_addi(s, addr_type, TCG_REG_TMP1, addr_reg, s_mask - a_mask);
1017        } else {
1018            tcg_out_mov(s, addr_type, TCG_REG_TMP1, addr_reg);
1019        }
1020        tcg_out_opc_bstrins_d(s, TCG_REG_TMP1, TCG_REG_ZERO,
1021                              a_bits, s->page_bits - 1);
1022
1023        /* Compare masked address with the TLB entry.  */
1024        ldst->label_ptr[0] = s->code_ptr;
1025        tcg_out_opc_bne(s, TCG_REG_TMP0, TCG_REG_TMP1, 0);
1026
1027        h->index = TCG_REG_TMP2;
1028    } else {
1029        if (a_bits) {
1030            ldst = new_ldst_label(s);
1031
1032            ldst->is_ld = is_ld;
1033            ldst->oi = oi;
1034            ldst->addrlo_reg = addr_reg;
1035
1036            /*
1037             * Without micro-architecture details, we don't know which of
1038             * bstrpick or andi is faster, so use bstrpick as it's not
1039             * constrained by imm field width. Not to say alignments >= 2^12
1040             * are going to happen any time soon.
1041             */
1042            tcg_out_opc_bstrpick_d(s, TCG_REG_TMP1, addr_reg, 0, a_bits - 1);
1043
1044            ldst->label_ptr[0] = s->code_ptr;
1045            tcg_out_opc_bne(s, TCG_REG_TMP1, TCG_REG_ZERO, 0);
1046        }
1047
1048        h->index = guest_base ? TCG_GUEST_BASE_REG : TCG_REG_ZERO;
1049    }
1050
1051    if (addr_type == TCG_TYPE_I32) {
1052        h->base = TCG_REG_TMP0;
1053        tcg_out_ext32u(s, h->base, addr_reg);
1054    } else {
1055        h->base = addr_reg;
1056    }
1057
1058    return ldst;
1059}
1060
1061static void tcg_out_qemu_ld_indexed(TCGContext *s, MemOp opc, TCGType type,
1062                                    TCGReg rd, HostAddress h)
1063{
1064    /* Byte swapping is left to middle-end expansion.  */
1065    tcg_debug_assert((opc & MO_BSWAP) == 0);
1066
1067    switch (opc & MO_SSIZE) {
1068    case MO_UB:
1069        tcg_out_opc_ldx_bu(s, rd, h.base, h.index);
1070        break;
1071    case MO_SB:
1072        tcg_out_opc_ldx_b(s, rd, h.base, h.index);
1073        break;
1074    case MO_UW:
1075        tcg_out_opc_ldx_hu(s, rd, h.base, h.index);
1076        break;
1077    case MO_SW:
1078        tcg_out_opc_ldx_h(s, rd, h.base, h.index);
1079        break;
1080    case MO_UL:
1081        if (type == TCG_TYPE_I64) {
1082            tcg_out_opc_ldx_wu(s, rd, h.base, h.index);
1083            break;
1084        }
1085        /* fallthrough */
1086    case MO_SL:
1087        tcg_out_opc_ldx_w(s, rd, h.base, h.index);
1088        break;
1089    case MO_UQ:
1090        tcg_out_opc_ldx_d(s, rd, h.base, h.index);
1091        break;
1092    default:
1093        g_assert_not_reached();
1094    }
1095}
1096
1097static void tcg_out_qemu_ld(TCGContext *s, TCGReg data_reg, TCGReg addr_reg,
1098                            MemOpIdx oi, TCGType data_type)
1099{
1100    TCGLabelQemuLdst *ldst;
1101    HostAddress h;
1102
1103    ldst = prepare_host_addr(s, &h, addr_reg, oi, true);
1104    tcg_out_qemu_ld_indexed(s, get_memop(oi), data_type, data_reg, h);
1105
1106    if (ldst) {
1107        ldst->type = data_type;
1108        ldst->datalo_reg = data_reg;
1109        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1110    }
1111}
1112
1113static void tcg_out_qemu_st_indexed(TCGContext *s, MemOp opc,
1114                                    TCGReg rd, HostAddress h)
1115{
1116    /* Byte swapping is left to middle-end expansion.  */
1117    tcg_debug_assert((opc & MO_BSWAP) == 0);
1118
1119    switch (opc & MO_SIZE) {
1120    case MO_8:
1121        tcg_out_opc_stx_b(s, rd, h.base, h.index);
1122        break;
1123    case MO_16:
1124        tcg_out_opc_stx_h(s, rd, h.base, h.index);
1125        break;
1126    case MO_32:
1127        tcg_out_opc_stx_w(s, rd, h.base, h.index);
1128        break;
1129    case MO_64:
1130        tcg_out_opc_stx_d(s, rd, h.base, h.index);
1131        break;
1132    default:
1133        g_assert_not_reached();
1134    }
1135}
1136
1137static void tcg_out_qemu_st(TCGContext *s, TCGReg data_reg, TCGReg addr_reg,
1138                            MemOpIdx oi, TCGType data_type)
1139{
1140    TCGLabelQemuLdst *ldst;
1141    HostAddress h;
1142
1143    ldst = prepare_host_addr(s, &h, addr_reg, oi, false);
1144    tcg_out_qemu_st_indexed(s, get_memop(oi), data_reg, h);
1145
1146    if (ldst) {
1147        ldst->type = data_type;
1148        ldst->datalo_reg = data_reg;
1149        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1150    }
1151}
1152
1153static void tcg_out_qemu_ldst_i128(TCGContext *s, TCGReg data_lo, TCGReg data_hi,
1154                                   TCGReg addr_reg, MemOpIdx oi, bool is_ld)
1155{
1156    TCGLabelQemuLdst *ldst;
1157    HostAddress h;
1158
1159    ldst = prepare_host_addr(s, &h, addr_reg, oi, is_ld);
1160
1161    if (h.aa.atom == MO_128) {
1162        /*
1163         * Use VLDX/VSTX when 128-bit atomicity is required.
1164         * If address is aligned to 16-bytes, the 128-bit load/store is atomic.
1165         */
1166        if (is_ld) {
1167            tcg_out_opc_vldx(s, TCG_VEC_TMP0, h.base, h.index);
1168            tcg_out_opc_vpickve2gr_d(s, data_lo, TCG_VEC_TMP0, 0);
1169            tcg_out_opc_vpickve2gr_d(s, data_hi, TCG_VEC_TMP0, 1);
1170        } else {
1171            tcg_out_opc_vinsgr2vr_d(s, TCG_VEC_TMP0, data_lo, 0);
1172            tcg_out_opc_vinsgr2vr_d(s, TCG_VEC_TMP0, data_hi, 1);
1173            tcg_out_opc_vstx(s, TCG_VEC_TMP0, h.base, h.index);
1174        }
1175    } else {
1176        /* Otherwise use a pair of LD/ST. */
1177        TCGReg base = h.base;
1178        if (h.index != TCG_REG_ZERO) {
1179            base = TCG_REG_TMP0;
1180            tcg_out_opc_add_d(s, base, h.base, h.index);
1181        }
1182        if (is_ld) {
1183            tcg_debug_assert(base != data_lo);
1184            tcg_out_opc_ld_d(s, data_lo, base, 0);
1185            tcg_out_opc_ld_d(s, data_hi, base, 8);
1186        } else {
1187            tcg_out_opc_st_d(s, data_lo, base, 0);
1188            tcg_out_opc_st_d(s, data_hi, base, 8);
1189        }
1190    }
1191
1192    if (ldst) {
1193        ldst->type = TCG_TYPE_I128;
1194        ldst->datalo_reg = data_lo;
1195        ldst->datahi_reg = data_hi;
1196        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1197    }
1198}
1199
1200/*
1201 * Entry-points
1202 */
1203
1204static const tcg_insn_unit *tb_ret_addr;
1205
1206static void tcg_out_exit_tb(TCGContext *s, uintptr_t a0)
1207{
1208    /* Reuse the zeroing that exists for goto_ptr.  */
1209    if (a0 == 0) {
1210        tcg_out_call_int(s, tcg_code_gen_epilogue, true);
1211    } else {
1212        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_A0, a0);
1213        tcg_out_call_int(s, tb_ret_addr, true);
1214    }
1215}
1216
1217static void tcg_out_goto_tb(TCGContext *s, int which)
1218{
1219    /*
1220     * Direct branch, or load indirect address, to be patched
1221     * by tb_target_set_jmp_target.  Check indirect load offset
1222     * in range early, regardless of direct branch distance,
1223     * via assert within tcg_out_opc_pcaddu2i.
1224     */
1225    uintptr_t i_addr = get_jmp_target_addr(s, which);
1226    intptr_t i_disp = tcg_pcrel_diff(s, (void *)i_addr);
1227
1228    set_jmp_insn_offset(s, which);
1229    tcg_out_opc_pcaddu2i(s, TCG_REG_TMP0, i_disp >> 2);
1230
1231    /* Finish the load and indirect branch. */
1232    tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP0, TCG_REG_TMP0, 0);
1233    tcg_out_opc_jirl(s, TCG_REG_ZERO, TCG_REG_TMP0, 0);
1234    set_jmp_reset_offset(s, which);
1235}
1236
1237void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
1238                              uintptr_t jmp_rx, uintptr_t jmp_rw)
1239{
1240    uintptr_t d_addr = tb->jmp_target_addr[n];
1241    ptrdiff_t d_disp = (ptrdiff_t)(d_addr - jmp_rx) >> 2;
1242    tcg_insn_unit insn;
1243
1244    /* Either directly branch, or load slot address for indirect branch. */
1245    if (d_disp == sextreg(d_disp, 0, 26)) {
1246        insn = encode_sd10k16_insn(OPC_B, d_disp);
1247    } else {
1248        uintptr_t i_addr = (uintptr_t)&tb->jmp_target_addr[n];
1249        intptr_t i_disp = i_addr - jmp_rx;
1250        insn = encode_dsj20_insn(OPC_PCADDU2I, TCG_REG_TMP0, i_disp >> 2);
1251    }
1252
1253    qatomic_set((tcg_insn_unit *)jmp_rw, insn);
1254    flush_idcache_range(jmp_rx, jmp_rw, 4);
1255}
1256
1257static void tcg_out_op(TCGContext *s, TCGOpcode opc,
1258                       const TCGArg args[TCG_MAX_OP_ARGS],
1259                       const int const_args[TCG_MAX_OP_ARGS])
1260{
1261    TCGArg a0 = args[0];
1262    TCGArg a1 = args[1];
1263    TCGArg a2 = args[2];
1264    TCGArg a3 = args[3];
1265    int c2 = const_args[2];
1266
1267    switch (opc) {
1268    case INDEX_op_mb:
1269        tcg_out_mb(s, a0);
1270        break;
1271
1272    case INDEX_op_goto_ptr:
1273        tcg_out_opc_jirl(s, TCG_REG_ZERO, a0, 0);
1274        break;
1275
1276    case INDEX_op_br:
1277        tcg_out_reloc(s, s->code_ptr, R_LOONGARCH_BR_SD10K16, arg_label(a0),
1278                      0);
1279        tcg_out_opc_b(s, 0);
1280        break;
1281
1282    case INDEX_op_brcond_i32:
1283    case INDEX_op_brcond_i64:
1284        tcg_out_brcond(s, a2, a0, a1, arg_label(args[3]));
1285        break;
1286
1287    case INDEX_op_extrh_i64_i32:
1288        tcg_out_opc_srai_d(s, a0, a1, 32);
1289        break;
1290
1291    case INDEX_op_not_i32:
1292    case INDEX_op_not_i64:
1293        tcg_out_opc_nor(s, a0, a1, TCG_REG_ZERO);
1294        break;
1295
1296    case INDEX_op_nor_i32:
1297    case INDEX_op_nor_i64:
1298        if (c2) {
1299            tcg_out_opc_ori(s, a0, a1, a2);
1300            tcg_out_opc_nor(s, a0, a0, TCG_REG_ZERO);
1301        } else {
1302            tcg_out_opc_nor(s, a0, a1, a2);
1303        }
1304        break;
1305
1306    case INDEX_op_andc_i32:
1307    case INDEX_op_andc_i64:
1308        if (c2) {
1309            /* guaranteed to fit due to constraint */
1310            tcg_out_opc_andi(s, a0, a1, ~a2);
1311        } else {
1312            tcg_out_opc_andn(s, a0, a1, a2);
1313        }
1314        break;
1315
1316    case INDEX_op_orc_i32:
1317    case INDEX_op_orc_i64:
1318        if (c2) {
1319            /* guaranteed to fit due to constraint */
1320            tcg_out_opc_ori(s, a0, a1, ~a2);
1321        } else {
1322            tcg_out_opc_orn(s, a0, a1, a2);
1323        }
1324        break;
1325
1326    case INDEX_op_and_i32:
1327    case INDEX_op_and_i64:
1328        if (c2) {
1329            tcg_out_opc_andi(s, a0, a1, a2);
1330        } else {
1331            tcg_out_opc_and(s, a0, a1, a2);
1332        }
1333        break;
1334
1335    case INDEX_op_or_i32:
1336    case INDEX_op_or_i64:
1337        if (c2) {
1338            tcg_out_opc_ori(s, a0, a1, a2);
1339        } else {
1340            tcg_out_opc_or(s, a0, a1, a2);
1341        }
1342        break;
1343
1344    case INDEX_op_xor_i32:
1345    case INDEX_op_xor_i64:
1346        if (c2) {
1347            tcg_out_opc_xori(s, a0, a1, a2);
1348        } else {
1349            tcg_out_opc_xor(s, a0, a1, a2);
1350        }
1351        break;
1352
1353    case INDEX_op_extract_i32:
1354        tcg_out_opc_bstrpick_w(s, a0, a1, a2, a2 + args[3] - 1);
1355        break;
1356    case INDEX_op_extract_i64:
1357        tcg_out_opc_bstrpick_d(s, a0, a1, a2, a2 + args[3] - 1);
1358        break;
1359
1360    case INDEX_op_deposit_i32:
1361        tcg_out_opc_bstrins_w(s, a0, a2, args[3], args[3] + args[4] - 1);
1362        break;
1363    case INDEX_op_deposit_i64:
1364        tcg_out_opc_bstrins_d(s, a0, a2, args[3], args[3] + args[4] - 1);
1365        break;
1366
1367    case INDEX_op_bswap16_i32:
1368    case INDEX_op_bswap16_i64:
1369        tcg_out_opc_revb_2h(s, a0, a1);
1370        if (a2 & TCG_BSWAP_OS) {
1371            tcg_out_ext16s(s, TCG_TYPE_REG, a0, a0);
1372        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1373            tcg_out_ext16u(s, a0, a0);
1374        }
1375        break;
1376
1377    case INDEX_op_bswap32_i32:
1378        /* All 32-bit values are computed sign-extended in the register.  */
1379        a2 = TCG_BSWAP_OS;
1380        /* fallthrough */
1381    case INDEX_op_bswap32_i64:
1382        tcg_out_opc_revb_2w(s, a0, a1);
1383        if (a2 & TCG_BSWAP_OS) {
1384            tcg_out_ext32s(s, a0, a0);
1385        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1386            tcg_out_ext32u(s, a0, a0);
1387        }
1388        break;
1389
1390    case INDEX_op_bswap64_i64:
1391        tcg_out_opc_revb_d(s, a0, a1);
1392        break;
1393
1394    case INDEX_op_clz_i32:
1395        tcg_out_clzctz(s, OPC_CLZ_W, a0, a1, a2, c2, true);
1396        break;
1397    case INDEX_op_clz_i64:
1398        tcg_out_clzctz(s, OPC_CLZ_D, a0, a1, a2, c2, false);
1399        break;
1400
1401    case INDEX_op_ctz_i32:
1402        tcg_out_clzctz(s, OPC_CTZ_W, a0, a1, a2, c2, true);
1403        break;
1404    case INDEX_op_ctz_i64:
1405        tcg_out_clzctz(s, OPC_CTZ_D, a0, a1, a2, c2, false);
1406        break;
1407
1408    case INDEX_op_shl_i32:
1409        if (c2) {
1410            tcg_out_opc_slli_w(s, a0, a1, a2 & 0x1f);
1411        } else {
1412            tcg_out_opc_sll_w(s, a0, a1, a2);
1413        }
1414        break;
1415    case INDEX_op_shl_i64:
1416        if (c2) {
1417            tcg_out_opc_slli_d(s, a0, a1, a2 & 0x3f);
1418        } else {
1419            tcg_out_opc_sll_d(s, a0, a1, a2);
1420        }
1421        break;
1422
1423    case INDEX_op_shr_i32:
1424        if (c2) {
1425            tcg_out_opc_srli_w(s, a0, a1, a2 & 0x1f);
1426        } else {
1427            tcg_out_opc_srl_w(s, a0, a1, a2);
1428        }
1429        break;
1430    case INDEX_op_shr_i64:
1431        if (c2) {
1432            tcg_out_opc_srli_d(s, a0, a1, a2 & 0x3f);
1433        } else {
1434            tcg_out_opc_srl_d(s, a0, a1, a2);
1435        }
1436        break;
1437
1438    case INDEX_op_sar_i32:
1439        if (c2) {
1440            tcg_out_opc_srai_w(s, a0, a1, a2 & 0x1f);
1441        } else {
1442            tcg_out_opc_sra_w(s, a0, a1, a2);
1443        }
1444        break;
1445    case INDEX_op_sar_i64:
1446        if (c2) {
1447            tcg_out_opc_srai_d(s, a0, a1, a2 & 0x3f);
1448        } else {
1449            tcg_out_opc_sra_d(s, a0, a1, a2);
1450        }
1451        break;
1452
1453    case INDEX_op_rotl_i32:
1454        /* transform into equivalent rotr/rotri */
1455        if (c2) {
1456            tcg_out_opc_rotri_w(s, a0, a1, (32 - a2) & 0x1f);
1457        } else {
1458            tcg_out_opc_sub_w(s, TCG_REG_TMP0, TCG_REG_ZERO, a2);
1459            tcg_out_opc_rotr_w(s, a0, a1, TCG_REG_TMP0);
1460        }
1461        break;
1462    case INDEX_op_rotl_i64:
1463        /* transform into equivalent rotr/rotri */
1464        if (c2) {
1465            tcg_out_opc_rotri_d(s, a0, a1, (64 - a2) & 0x3f);
1466        } else {
1467            tcg_out_opc_sub_w(s, TCG_REG_TMP0, TCG_REG_ZERO, a2);
1468            tcg_out_opc_rotr_d(s, a0, a1, TCG_REG_TMP0);
1469        }
1470        break;
1471
1472    case INDEX_op_rotr_i32:
1473        if (c2) {
1474            tcg_out_opc_rotri_w(s, a0, a1, a2 & 0x1f);
1475        } else {
1476            tcg_out_opc_rotr_w(s, a0, a1, a2);
1477        }
1478        break;
1479    case INDEX_op_rotr_i64:
1480        if (c2) {
1481            tcg_out_opc_rotri_d(s, a0, a1, a2 & 0x3f);
1482        } else {
1483            tcg_out_opc_rotr_d(s, a0, a1, a2);
1484        }
1485        break;
1486
1487    case INDEX_op_add_i32:
1488        if (c2) {
1489            tcg_out_addi(s, TCG_TYPE_I32, a0, a1, a2);
1490        } else {
1491            tcg_out_opc_add_w(s, a0, a1, a2);
1492        }
1493        break;
1494    case INDEX_op_add_i64:
1495        if (c2) {
1496            tcg_out_addi(s, TCG_TYPE_I64, a0, a1, a2);
1497        } else {
1498            tcg_out_opc_add_d(s, a0, a1, a2);
1499        }
1500        break;
1501
1502    case INDEX_op_sub_i32:
1503        if (c2) {
1504            tcg_out_addi(s, TCG_TYPE_I32, a0, a1, -a2);
1505        } else {
1506            tcg_out_opc_sub_w(s, a0, a1, a2);
1507        }
1508        break;
1509    case INDEX_op_sub_i64:
1510        if (c2) {
1511            tcg_out_addi(s, TCG_TYPE_I64, a0, a1, -a2);
1512        } else {
1513            tcg_out_opc_sub_d(s, a0, a1, a2);
1514        }
1515        break;
1516
1517    case INDEX_op_neg_i32:
1518        tcg_out_opc_sub_w(s, a0, TCG_REG_ZERO, a1);
1519        break;
1520    case INDEX_op_neg_i64:
1521        tcg_out_opc_sub_d(s, a0, TCG_REG_ZERO, a1);
1522        break;
1523
1524    case INDEX_op_mul_i32:
1525        tcg_out_opc_mul_w(s, a0, a1, a2);
1526        break;
1527    case INDEX_op_mul_i64:
1528        tcg_out_opc_mul_d(s, a0, a1, a2);
1529        break;
1530
1531    case INDEX_op_mulsh_i32:
1532        tcg_out_opc_mulh_w(s, a0, a1, a2);
1533        break;
1534    case INDEX_op_mulsh_i64:
1535        tcg_out_opc_mulh_d(s, a0, a1, a2);
1536        break;
1537
1538    case INDEX_op_muluh_i32:
1539        tcg_out_opc_mulh_wu(s, a0, a1, a2);
1540        break;
1541    case INDEX_op_muluh_i64:
1542        tcg_out_opc_mulh_du(s, a0, a1, a2);
1543        break;
1544
1545    case INDEX_op_div_i32:
1546        tcg_out_opc_div_w(s, a0, a1, a2);
1547        break;
1548    case INDEX_op_div_i64:
1549        tcg_out_opc_div_d(s, a0, a1, a2);
1550        break;
1551
1552    case INDEX_op_divu_i32:
1553        tcg_out_opc_div_wu(s, a0, a1, a2);
1554        break;
1555    case INDEX_op_divu_i64:
1556        tcg_out_opc_div_du(s, a0, a1, a2);
1557        break;
1558
1559    case INDEX_op_rem_i32:
1560        tcg_out_opc_mod_w(s, a0, a1, a2);
1561        break;
1562    case INDEX_op_rem_i64:
1563        tcg_out_opc_mod_d(s, a0, a1, a2);
1564        break;
1565
1566    case INDEX_op_remu_i32:
1567        tcg_out_opc_mod_wu(s, a0, a1, a2);
1568        break;
1569    case INDEX_op_remu_i64:
1570        tcg_out_opc_mod_du(s, a0, a1, a2);
1571        break;
1572
1573    case INDEX_op_setcond_i32:
1574    case INDEX_op_setcond_i64:
1575        tcg_out_setcond(s, args[3], a0, a1, a2, c2);
1576        break;
1577
1578    case INDEX_op_movcond_i32:
1579    case INDEX_op_movcond_i64:
1580        tcg_out_movcond(s, args[5], a0, a1, a2, c2, args[3], args[4]);
1581        break;
1582
1583    case INDEX_op_ld8s_i32:
1584    case INDEX_op_ld8s_i64:
1585        tcg_out_ldst(s, OPC_LD_B, a0, a1, a2);
1586        break;
1587    case INDEX_op_ld8u_i32:
1588    case INDEX_op_ld8u_i64:
1589        tcg_out_ldst(s, OPC_LD_BU, a0, a1, a2);
1590        break;
1591    case INDEX_op_ld16s_i32:
1592    case INDEX_op_ld16s_i64:
1593        tcg_out_ldst(s, OPC_LD_H, a0, a1, a2);
1594        break;
1595    case INDEX_op_ld16u_i32:
1596    case INDEX_op_ld16u_i64:
1597        tcg_out_ldst(s, OPC_LD_HU, a0, a1, a2);
1598        break;
1599    case INDEX_op_ld_i32:
1600    case INDEX_op_ld32s_i64:
1601        tcg_out_ldst(s, OPC_LD_W, a0, a1, a2);
1602        break;
1603    case INDEX_op_ld32u_i64:
1604        tcg_out_ldst(s, OPC_LD_WU, a0, a1, a2);
1605        break;
1606    case INDEX_op_ld_i64:
1607        tcg_out_ldst(s, OPC_LD_D, a0, a1, a2);
1608        break;
1609
1610    case INDEX_op_st8_i32:
1611    case INDEX_op_st8_i64:
1612        tcg_out_ldst(s, OPC_ST_B, a0, a1, a2);
1613        break;
1614    case INDEX_op_st16_i32:
1615    case INDEX_op_st16_i64:
1616        tcg_out_ldst(s, OPC_ST_H, a0, a1, a2);
1617        break;
1618    case INDEX_op_st_i32:
1619    case INDEX_op_st32_i64:
1620        tcg_out_ldst(s, OPC_ST_W, a0, a1, a2);
1621        break;
1622    case INDEX_op_st_i64:
1623        tcg_out_ldst(s, OPC_ST_D, a0, a1, a2);
1624        break;
1625
1626    case INDEX_op_qemu_ld_a32_i32:
1627    case INDEX_op_qemu_ld_a64_i32:
1628        tcg_out_qemu_ld(s, a0, a1, a2, TCG_TYPE_I32);
1629        break;
1630    case INDEX_op_qemu_ld_a32_i64:
1631    case INDEX_op_qemu_ld_a64_i64:
1632        tcg_out_qemu_ld(s, a0, a1, a2, TCG_TYPE_I64);
1633        break;
1634    case INDEX_op_qemu_ld_a32_i128:
1635    case INDEX_op_qemu_ld_a64_i128:
1636        tcg_out_qemu_ldst_i128(s, a0, a1, a2, a3, true);
1637        break;
1638    case INDEX_op_qemu_st_a32_i32:
1639    case INDEX_op_qemu_st_a64_i32:
1640        tcg_out_qemu_st(s, a0, a1, a2, TCG_TYPE_I32);
1641        break;
1642    case INDEX_op_qemu_st_a32_i64:
1643    case INDEX_op_qemu_st_a64_i64:
1644        tcg_out_qemu_st(s, a0, a1, a2, TCG_TYPE_I64);
1645        break;
1646    case INDEX_op_qemu_st_a32_i128:
1647    case INDEX_op_qemu_st_a64_i128:
1648        tcg_out_qemu_ldst_i128(s, a0, a1, a2, a3, false);
1649        break;
1650
1651    case INDEX_op_mov_i32:  /* Always emitted via tcg_out_mov.  */
1652    case INDEX_op_mov_i64:
1653    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
1654    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
1655    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
1656    case INDEX_op_ext8s_i32:  /* Always emitted via tcg_reg_alloc_op.  */
1657    case INDEX_op_ext8s_i64:
1658    case INDEX_op_ext8u_i32:
1659    case INDEX_op_ext8u_i64:
1660    case INDEX_op_ext16s_i32:
1661    case INDEX_op_ext16s_i64:
1662    case INDEX_op_ext16u_i32:
1663    case INDEX_op_ext16u_i64:
1664    case INDEX_op_ext32s_i64:
1665    case INDEX_op_ext32u_i64:
1666    case INDEX_op_ext_i32_i64:
1667    case INDEX_op_extu_i32_i64:
1668    case INDEX_op_extrl_i64_i32:
1669    default:
1670        g_assert_not_reached();
1671    }
1672}
1673
1674static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
1675                            TCGReg rd, TCGReg rs)
1676{
1677    static const LoongArchInsn repl_insn[2][4] = {
1678        { OPC_VREPLGR2VR_B, OPC_VREPLGR2VR_H,
1679          OPC_VREPLGR2VR_W, OPC_VREPLGR2VR_D },
1680        { OPC_XVREPLGR2VR_B, OPC_XVREPLGR2VR_H,
1681          OPC_XVREPLGR2VR_W, OPC_XVREPLGR2VR_D },
1682    };
1683    bool lasx = type == TCG_TYPE_V256;
1684
1685    tcg_debug_assert(vece <= MO_64);
1686    tcg_out32(s, encode_vdj_insn(repl_insn[lasx][vece], rd, rs));
1687    return true;
1688}
1689
1690static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
1691                             TCGReg r, TCGReg base, intptr_t offset)
1692{
1693    bool lasx = type == TCG_TYPE_V256;
1694
1695    /* Handle imm overflow and division (vldrepl.d imm is divided by 8). */
1696    if (offset < -0x800 || offset > 0x7ff ||
1697        (offset & ((1 << vece) - 1)) != 0) {
1698        tcg_out_addi(s, TCG_TYPE_I64, TCG_REG_TMP0, base, offset);
1699        base = TCG_REG_TMP0;
1700        offset = 0;
1701    }
1702    offset >>= vece;
1703
1704    switch (vece) {
1705    case MO_8:
1706        if (lasx) {
1707            tcg_out_opc_xvldrepl_b(s, r, base, offset);
1708        } else {
1709            tcg_out_opc_vldrepl_b(s, r, base, offset);
1710        }
1711        break;
1712    case MO_16:
1713        if (lasx) {
1714            tcg_out_opc_xvldrepl_h(s, r, base, offset);
1715        } else {
1716            tcg_out_opc_vldrepl_h(s, r, base, offset);
1717        }
1718        break;
1719    case MO_32:
1720        if (lasx) {
1721            tcg_out_opc_xvldrepl_w(s, r, base, offset);
1722        } else {
1723            tcg_out_opc_vldrepl_w(s, r, base, offset);
1724        }
1725        break;
1726    case MO_64:
1727        if (lasx) {
1728            tcg_out_opc_xvldrepl_d(s, r, base, offset);
1729        } else {
1730            tcg_out_opc_vldrepl_d(s, r, base, offset);
1731        }
1732        break;
1733    default:
1734        g_assert_not_reached();
1735    }
1736    return true;
1737}
1738
1739static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
1740                             TCGReg rd, int64_t v64)
1741{
1742    /* Try vldi if imm can fit */
1743    int64_t value = sextract64(v64, 0, 8 << vece);
1744    if (-0x200 <= value && value <= 0x1FF) {
1745        uint32_t imm = (vece << 10) | ((uint32_t)v64 & 0x3FF);
1746
1747        if (type == TCG_TYPE_V256) {
1748            tcg_out_opc_xvldi(s, rd, imm);
1749        } else {
1750            tcg_out_opc_vldi(s, rd, imm);
1751        }
1752        return;
1753    }
1754
1755    /* TODO: vldi patterns when imm 12 is set */
1756
1757    tcg_out_movi(s, TCG_TYPE_I64, TCG_REG_TMP0, value);
1758    tcg_out_dup_vec(s, type, vece, rd, TCG_REG_TMP0);
1759}
1760
1761static void tcg_out_addsub_vec(TCGContext *s, unsigned vece, const TCGArg a0,
1762                               const TCGArg a1, const TCGArg a2,
1763                               bool a2_is_const, bool is_add)
1764{
1765    static const LoongArchInsn add_vec_insn[4] = {
1766        OPC_VADD_B, OPC_VADD_H, OPC_VADD_W, OPC_VADD_D
1767    };
1768    static const LoongArchInsn add_vec_imm_insn[4] = {
1769        OPC_VADDI_BU, OPC_VADDI_HU, OPC_VADDI_WU, OPC_VADDI_DU
1770    };
1771    static const LoongArchInsn sub_vec_insn[4] = {
1772        OPC_VSUB_B, OPC_VSUB_H, OPC_VSUB_W, OPC_VSUB_D
1773    };
1774    static const LoongArchInsn sub_vec_imm_insn[4] = {
1775        OPC_VSUBI_BU, OPC_VSUBI_HU, OPC_VSUBI_WU, OPC_VSUBI_DU
1776    };
1777
1778    if (a2_is_const) {
1779        int64_t value = sextract64(a2, 0, 8 << vece);
1780        if (!is_add) {
1781            value = -value;
1782        }
1783
1784        /* Try vaddi/vsubi */
1785        if (0 <= value && value <= 0x1f) {
1786            tcg_out32(s, encode_vdvjuk5_insn(add_vec_imm_insn[vece], a0, \
1787                                             a1, value));
1788            return;
1789        } else if (-0x1f <= value && value < 0) {
1790            tcg_out32(s, encode_vdvjuk5_insn(sub_vec_imm_insn[vece], a0, \
1791                                             a1, -value));
1792            return;
1793        }
1794
1795        /* constraint TCG_CT_CONST_VADD ensures unreachable */
1796        g_assert_not_reached();
1797    }
1798
1799    if (is_add) {
1800        tcg_out32(s, encode_vdvjvk_insn(add_vec_insn[vece], a0, a1, a2));
1801    } else {
1802        tcg_out32(s, encode_vdvjvk_insn(sub_vec_insn[vece], a0, a1, a2));
1803    }
1804}
1805
1806static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
1807                           unsigned vecl, unsigned vece,
1808                           const TCGArg args[TCG_MAX_OP_ARGS],
1809                           const int const_args[TCG_MAX_OP_ARGS])
1810{
1811    TCGType type = vecl + TCG_TYPE_V64;
1812    TCGArg a0, a1, a2, a3;
1813    TCGReg temp_vec = TCG_VEC_TMP0;
1814
1815    static const LoongArchInsn cmp_vec_insn[16][4] = {
1816        [TCG_COND_EQ] = {OPC_VSEQ_B, OPC_VSEQ_H, OPC_VSEQ_W, OPC_VSEQ_D},
1817        [TCG_COND_LE] = {OPC_VSLE_B, OPC_VSLE_H, OPC_VSLE_W, OPC_VSLE_D},
1818        [TCG_COND_LEU] = {OPC_VSLE_BU, OPC_VSLE_HU, OPC_VSLE_WU, OPC_VSLE_DU},
1819        [TCG_COND_LT] = {OPC_VSLT_B, OPC_VSLT_H, OPC_VSLT_W, OPC_VSLT_D},
1820        [TCG_COND_LTU] = {OPC_VSLT_BU, OPC_VSLT_HU, OPC_VSLT_WU, OPC_VSLT_DU},
1821    };
1822    static const LoongArchInsn cmp_vec_imm_insn[16][4] = {
1823        [TCG_COND_EQ] = {OPC_VSEQI_B, OPC_VSEQI_H, OPC_VSEQI_W, OPC_VSEQI_D},
1824        [TCG_COND_LE] = {OPC_VSLEI_B, OPC_VSLEI_H, OPC_VSLEI_W, OPC_VSLEI_D},
1825        [TCG_COND_LEU] = {OPC_VSLEI_BU, OPC_VSLEI_HU, OPC_VSLEI_WU, OPC_VSLEI_DU},
1826        [TCG_COND_LT] = {OPC_VSLTI_B, OPC_VSLTI_H, OPC_VSLTI_W, OPC_VSLTI_D},
1827        [TCG_COND_LTU] = {OPC_VSLTI_BU, OPC_VSLTI_HU, OPC_VSLTI_WU, OPC_VSLTI_DU},
1828    };
1829    LoongArchInsn insn;
1830    static const LoongArchInsn neg_vec_insn[4] = {
1831        OPC_VNEG_B, OPC_VNEG_H, OPC_VNEG_W, OPC_VNEG_D
1832    };
1833    static const LoongArchInsn mul_vec_insn[4] = {
1834        OPC_VMUL_B, OPC_VMUL_H, OPC_VMUL_W, OPC_VMUL_D
1835    };
1836    static const LoongArchInsn smin_vec_insn[4] = {
1837        OPC_VMIN_B, OPC_VMIN_H, OPC_VMIN_W, OPC_VMIN_D
1838    };
1839    static const LoongArchInsn umin_vec_insn[4] = {
1840        OPC_VMIN_BU, OPC_VMIN_HU, OPC_VMIN_WU, OPC_VMIN_DU
1841    };
1842    static const LoongArchInsn smax_vec_insn[4] = {
1843        OPC_VMAX_B, OPC_VMAX_H, OPC_VMAX_W, OPC_VMAX_D
1844    };
1845    static const LoongArchInsn umax_vec_insn[4] = {
1846        OPC_VMAX_BU, OPC_VMAX_HU, OPC_VMAX_WU, OPC_VMAX_DU
1847    };
1848    static const LoongArchInsn ssadd_vec_insn[4] = {
1849        OPC_VSADD_B, OPC_VSADD_H, OPC_VSADD_W, OPC_VSADD_D
1850    };
1851    static const LoongArchInsn usadd_vec_insn[4] = {
1852        OPC_VSADD_BU, OPC_VSADD_HU, OPC_VSADD_WU, OPC_VSADD_DU
1853    };
1854    static const LoongArchInsn sssub_vec_insn[4] = {
1855        OPC_VSSUB_B, OPC_VSSUB_H, OPC_VSSUB_W, OPC_VSSUB_D
1856    };
1857    static const LoongArchInsn ussub_vec_insn[4] = {
1858        OPC_VSSUB_BU, OPC_VSSUB_HU, OPC_VSSUB_WU, OPC_VSSUB_DU
1859    };
1860    static const LoongArchInsn shlv_vec_insn[4] = {
1861        OPC_VSLL_B, OPC_VSLL_H, OPC_VSLL_W, OPC_VSLL_D
1862    };
1863    static const LoongArchInsn shrv_vec_insn[4] = {
1864        OPC_VSRL_B, OPC_VSRL_H, OPC_VSRL_W, OPC_VSRL_D
1865    };
1866    static const LoongArchInsn sarv_vec_insn[4] = {
1867        OPC_VSRA_B, OPC_VSRA_H, OPC_VSRA_W, OPC_VSRA_D
1868    };
1869    static const LoongArchInsn shli_vec_insn[4] = {
1870        OPC_VSLLI_B, OPC_VSLLI_H, OPC_VSLLI_W, OPC_VSLLI_D
1871    };
1872    static const LoongArchInsn shri_vec_insn[4] = {
1873        OPC_VSRLI_B, OPC_VSRLI_H, OPC_VSRLI_W, OPC_VSRLI_D
1874    };
1875    static const LoongArchInsn sari_vec_insn[4] = {
1876        OPC_VSRAI_B, OPC_VSRAI_H, OPC_VSRAI_W, OPC_VSRAI_D
1877    };
1878    static const LoongArchInsn rotrv_vec_insn[4] = {
1879        OPC_VROTR_B, OPC_VROTR_H, OPC_VROTR_W, OPC_VROTR_D
1880    };
1881
1882    a0 = args[0];
1883    a1 = args[1];
1884    a2 = args[2];
1885    a3 = args[3];
1886
1887    /* Currently only supports V64 & V128 */
1888    tcg_debug_assert(type == TCG_TYPE_V64 || type == TCG_TYPE_V128);
1889
1890    switch (opc) {
1891    case INDEX_op_st_vec:
1892        tcg_out_st(s, type, a0, a1, a2);
1893        break;
1894    case INDEX_op_ld_vec:
1895        tcg_out_ld(s, type, a0, a1, a2);
1896        break;
1897    case INDEX_op_and_vec:
1898        tcg_out_opc_vand_v(s, a0, a1, a2);
1899        break;
1900    case INDEX_op_andc_vec:
1901        /*
1902         * vandn vd, vj, vk: vd = vk & ~vj
1903         * andc_vec vd, vj, vk: vd = vj & ~vk
1904         * vk and vk are swapped
1905         */
1906        tcg_out_opc_vandn_v(s, a0, a2, a1);
1907        break;
1908    case INDEX_op_or_vec:
1909        tcg_out_opc_vor_v(s, a0, a1, a2);
1910        break;
1911    case INDEX_op_orc_vec:
1912        tcg_out_opc_vorn_v(s, a0, a1, a2);
1913        break;
1914    case INDEX_op_xor_vec:
1915        tcg_out_opc_vxor_v(s, a0, a1, a2);
1916        break;
1917    case INDEX_op_nor_vec:
1918        tcg_out_opc_vnor_v(s, a0, a1, a2);
1919        break;
1920    case INDEX_op_not_vec:
1921        tcg_out_opc_vnor_v(s, a0, a1, a1);
1922        break;
1923    case INDEX_op_cmp_vec:
1924        {
1925            TCGCond cond = args[3];
1926            if (const_args[2]) {
1927                /*
1928                 * cmp_vec dest, src, value
1929                 * Try vseqi/vslei/vslti
1930                 */
1931                int64_t value = sextract64(a2, 0, 8 << vece);
1932                if ((cond == TCG_COND_EQ || cond == TCG_COND_LE || \
1933                     cond == TCG_COND_LT) && (-0x10 <= value && value <= 0x0f)) {
1934                    tcg_out32(s, encode_vdvjsk5_insn(cmp_vec_imm_insn[cond][vece], \
1935                                                     a0, a1, value));
1936                    break;
1937                } else if ((cond == TCG_COND_LEU || cond == TCG_COND_LTU) &&
1938                    (0x00 <= value && value <= 0x1f)) {
1939                    tcg_out32(s, encode_vdvjuk5_insn(cmp_vec_imm_insn[cond][vece], \
1940                                                     a0, a1, value));
1941                    break;
1942                }
1943
1944                /*
1945                 * Fallback to:
1946                 * dupi_vec temp, a2
1947                 * cmp_vec a0, a1, temp, cond
1948                 */
1949                tcg_out_dupi_vec(s, type, vece, temp_vec, a2);
1950                a2 = temp_vec;
1951            }
1952
1953            insn = cmp_vec_insn[cond][vece];
1954            if (insn == 0) {
1955                TCGArg t;
1956                t = a1, a1 = a2, a2 = t;
1957                cond = tcg_swap_cond(cond);
1958                insn = cmp_vec_insn[cond][vece];
1959                tcg_debug_assert(insn != 0);
1960            }
1961            tcg_out32(s, encode_vdvjvk_insn(insn, a0, a1, a2));
1962        }
1963        break;
1964    case INDEX_op_add_vec:
1965        tcg_out_addsub_vec(s, vece, a0, a1, a2, const_args[2], true);
1966        break;
1967    case INDEX_op_sub_vec:
1968        tcg_out_addsub_vec(s, vece, a0, a1, a2, const_args[2], false);
1969        break;
1970    case INDEX_op_neg_vec:
1971        tcg_out32(s, encode_vdvj_insn(neg_vec_insn[vece], a0, a1));
1972        break;
1973    case INDEX_op_mul_vec:
1974        tcg_out32(s, encode_vdvjvk_insn(mul_vec_insn[vece], a0, a1, a2));
1975        break;
1976    case INDEX_op_smin_vec:
1977        tcg_out32(s, encode_vdvjvk_insn(smin_vec_insn[vece], a0, a1, a2));
1978        break;
1979    case INDEX_op_smax_vec:
1980        tcg_out32(s, encode_vdvjvk_insn(smax_vec_insn[vece], a0, a1, a2));
1981        break;
1982    case INDEX_op_umin_vec:
1983        tcg_out32(s, encode_vdvjvk_insn(umin_vec_insn[vece], a0, a1, a2));
1984        break;
1985    case INDEX_op_umax_vec:
1986        tcg_out32(s, encode_vdvjvk_insn(umax_vec_insn[vece], a0, a1, a2));
1987        break;
1988    case INDEX_op_ssadd_vec:
1989        tcg_out32(s, encode_vdvjvk_insn(ssadd_vec_insn[vece], a0, a1, a2));
1990        break;
1991    case INDEX_op_usadd_vec:
1992        tcg_out32(s, encode_vdvjvk_insn(usadd_vec_insn[vece], a0, a1, a2));
1993        break;
1994    case INDEX_op_sssub_vec:
1995        tcg_out32(s, encode_vdvjvk_insn(sssub_vec_insn[vece], a0, a1, a2));
1996        break;
1997    case INDEX_op_ussub_vec:
1998        tcg_out32(s, encode_vdvjvk_insn(ussub_vec_insn[vece], a0, a1, a2));
1999        break;
2000    case INDEX_op_shlv_vec:
2001        tcg_out32(s, encode_vdvjvk_insn(shlv_vec_insn[vece], a0, a1, a2));
2002        break;
2003    case INDEX_op_shrv_vec:
2004        tcg_out32(s, encode_vdvjvk_insn(shrv_vec_insn[vece], a0, a1, a2));
2005        break;
2006    case INDEX_op_sarv_vec:
2007        tcg_out32(s, encode_vdvjvk_insn(sarv_vec_insn[vece], a0, a1, a2));
2008        break;
2009    case INDEX_op_shli_vec:
2010        tcg_out32(s, encode_vdvjuk3_insn(shli_vec_insn[vece], a0, a1, a2));
2011        break;
2012    case INDEX_op_shri_vec:
2013        tcg_out32(s, encode_vdvjuk3_insn(shri_vec_insn[vece], a0, a1, a2));
2014        break;
2015    case INDEX_op_sari_vec:
2016        tcg_out32(s, encode_vdvjuk3_insn(sari_vec_insn[vece], a0, a1, a2));
2017        break;
2018    case INDEX_op_rotrv_vec:
2019        tcg_out32(s, encode_vdvjvk_insn(rotrv_vec_insn[vece], a0, a1, a2));
2020        break;
2021    case INDEX_op_rotlv_vec:
2022        /* rotlv_vec a1, a2 = rotrv_vec a1, -a2 */
2023        tcg_out32(s, encode_vdvj_insn(neg_vec_insn[vece], temp_vec, a2));
2024        tcg_out32(s, encode_vdvjvk_insn(rotrv_vec_insn[vece], a0, a1,
2025                                        temp_vec));
2026        break;
2027    case INDEX_op_rotli_vec:
2028        /* rotli_vec a1, a2 = rotri_vec a1, -a2 */
2029        a2 = extract32(-a2, 0, 3 + vece);
2030        switch (vece) {
2031        case MO_8:
2032            tcg_out_opc_vrotri_b(s, a0, a1, a2);
2033            break;
2034        case MO_16:
2035            tcg_out_opc_vrotri_h(s, a0, a1, a2);
2036            break;
2037        case MO_32:
2038            tcg_out_opc_vrotri_w(s, a0, a1, a2);
2039            break;
2040        case MO_64:
2041            tcg_out_opc_vrotri_d(s, a0, a1, a2);
2042            break;
2043        default:
2044            g_assert_not_reached();
2045        }
2046        break;
2047    case INDEX_op_bitsel_vec:
2048        /* vbitsel vd, vj, vk, va = bitsel_vec vd, va, vk, vj */
2049        tcg_out_opc_vbitsel_v(s, a0, a3, a2, a1);
2050        break;
2051    case INDEX_op_dupm_vec:
2052        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
2053        break;
2054    default:
2055        g_assert_not_reached();
2056    }
2057}
2058
2059int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
2060{
2061    switch (opc) {
2062    case INDEX_op_ld_vec:
2063    case INDEX_op_st_vec:
2064    case INDEX_op_dup_vec:
2065    case INDEX_op_dupm_vec:
2066    case INDEX_op_cmp_vec:
2067    case INDEX_op_add_vec:
2068    case INDEX_op_sub_vec:
2069    case INDEX_op_and_vec:
2070    case INDEX_op_andc_vec:
2071    case INDEX_op_or_vec:
2072    case INDEX_op_orc_vec:
2073    case INDEX_op_xor_vec:
2074    case INDEX_op_nor_vec:
2075    case INDEX_op_not_vec:
2076    case INDEX_op_neg_vec:
2077    case INDEX_op_mul_vec:
2078    case INDEX_op_smin_vec:
2079    case INDEX_op_smax_vec:
2080    case INDEX_op_umin_vec:
2081    case INDEX_op_umax_vec:
2082    case INDEX_op_ssadd_vec:
2083    case INDEX_op_usadd_vec:
2084    case INDEX_op_sssub_vec:
2085    case INDEX_op_ussub_vec:
2086    case INDEX_op_shlv_vec:
2087    case INDEX_op_shrv_vec:
2088    case INDEX_op_sarv_vec:
2089    case INDEX_op_bitsel_vec:
2090        return 1;
2091    default:
2092        return 0;
2093    }
2094}
2095
2096void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
2097                       TCGArg a0, ...)
2098{
2099    g_assert_not_reached();
2100}
2101
2102static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
2103{
2104    switch (op) {
2105    case INDEX_op_goto_ptr:
2106        return C_O0_I1(r);
2107
2108    case INDEX_op_st8_i32:
2109    case INDEX_op_st8_i64:
2110    case INDEX_op_st16_i32:
2111    case INDEX_op_st16_i64:
2112    case INDEX_op_st32_i64:
2113    case INDEX_op_st_i32:
2114    case INDEX_op_st_i64:
2115    case INDEX_op_qemu_st_a32_i32:
2116    case INDEX_op_qemu_st_a64_i32:
2117    case INDEX_op_qemu_st_a32_i64:
2118    case INDEX_op_qemu_st_a64_i64:
2119        return C_O0_I2(rZ, r);
2120
2121    case INDEX_op_qemu_ld_a32_i128:
2122    case INDEX_op_qemu_ld_a64_i128:
2123        return C_N2_I1(r, r, r);
2124
2125    case INDEX_op_qemu_st_a32_i128:
2126    case INDEX_op_qemu_st_a64_i128:
2127        return C_O0_I3(r, r, r);
2128
2129    case INDEX_op_brcond_i32:
2130    case INDEX_op_brcond_i64:
2131        return C_O0_I2(rZ, rZ);
2132
2133    case INDEX_op_ext8s_i32:
2134    case INDEX_op_ext8s_i64:
2135    case INDEX_op_ext8u_i32:
2136    case INDEX_op_ext8u_i64:
2137    case INDEX_op_ext16s_i32:
2138    case INDEX_op_ext16s_i64:
2139    case INDEX_op_ext16u_i32:
2140    case INDEX_op_ext16u_i64:
2141    case INDEX_op_ext32s_i64:
2142    case INDEX_op_ext32u_i64:
2143    case INDEX_op_extu_i32_i64:
2144    case INDEX_op_extrl_i64_i32:
2145    case INDEX_op_extrh_i64_i32:
2146    case INDEX_op_ext_i32_i64:
2147    case INDEX_op_neg_i32:
2148    case INDEX_op_neg_i64:
2149    case INDEX_op_not_i32:
2150    case INDEX_op_not_i64:
2151    case INDEX_op_extract_i32:
2152    case INDEX_op_extract_i64:
2153    case INDEX_op_bswap16_i32:
2154    case INDEX_op_bswap16_i64:
2155    case INDEX_op_bswap32_i32:
2156    case INDEX_op_bswap32_i64:
2157    case INDEX_op_bswap64_i64:
2158    case INDEX_op_ld8s_i32:
2159    case INDEX_op_ld8s_i64:
2160    case INDEX_op_ld8u_i32:
2161    case INDEX_op_ld8u_i64:
2162    case INDEX_op_ld16s_i32:
2163    case INDEX_op_ld16s_i64:
2164    case INDEX_op_ld16u_i32:
2165    case INDEX_op_ld16u_i64:
2166    case INDEX_op_ld32s_i64:
2167    case INDEX_op_ld32u_i64:
2168    case INDEX_op_ld_i32:
2169    case INDEX_op_ld_i64:
2170    case INDEX_op_qemu_ld_a32_i32:
2171    case INDEX_op_qemu_ld_a64_i32:
2172    case INDEX_op_qemu_ld_a32_i64:
2173    case INDEX_op_qemu_ld_a64_i64:
2174        return C_O1_I1(r, r);
2175
2176    case INDEX_op_andc_i32:
2177    case INDEX_op_andc_i64:
2178    case INDEX_op_orc_i32:
2179    case INDEX_op_orc_i64:
2180        /*
2181         * LoongArch insns for these ops don't have reg-imm forms, but we
2182         * can express using andi/ori if ~constant satisfies
2183         * TCG_CT_CONST_U12.
2184         */
2185        return C_O1_I2(r, r, rC);
2186
2187    case INDEX_op_shl_i32:
2188    case INDEX_op_shl_i64:
2189    case INDEX_op_shr_i32:
2190    case INDEX_op_shr_i64:
2191    case INDEX_op_sar_i32:
2192    case INDEX_op_sar_i64:
2193    case INDEX_op_rotl_i32:
2194    case INDEX_op_rotl_i64:
2195    case INDEX_op_rotr_i32:
2196    case INDEX_op_rotr_i64:
2197        return C_O1_I2(r, r, ri);
2198
2199    case INDEX_op_add_i32:
2200        return C_O1_I2(r, r, ri);
2201    case INDEX_op_add_i64:
2202        return C_O1_I2(r, r, rJ);
2203
2204    case INDEX_op_and_i32:
2205    case INDEX_op_and_i64:
2206    case INDEX_op_nor_i32:
2207    case INDEX_op_nor_i64:
2208    case INDEX_op_or_i32:
2209    case INDEX_op_or_i64:
2210    case INDEX_op_xor_i32:
2211    case INDEX_op_xor_i64:
2212        /* LoongArch reg-imm bitops have their imms ZERO-extended */
2213        return C_O1_I2(r, r, rU);
2214
2215    case INDEX_op_clz_i32:
2216    case INDEX_op_clz_i64:
2217    case INDEX_op_ctz_i32:
2218    case INDEX_op_ctz_i64:
2219        return C_O1_I2(r, r, rW);
2220
2221    case INDEX_op_deposit_i32:
2222    case INDEX_op_deposit_i64:
2223        /* Must deposit into the same register as input */
2224        return C_O1_I2(r, 0, rZ);
2225
2226    case INDEX_op_sub_i32:
2227    case INDEX_op_setcond_i32:
2228        return C_O1_I2(r, rZ, ri);
2229    case INDEX_op_sub_i64:
2230    case INDEX_op_setcond_i64:
2231        return C_O1_I2(r, rZ, rJ);
2232
2233    case INDEX_op_mul_i32:
2234    case INDEX_op_mul_i64:
2235    case INDEX_op_mulsh_i32:
2236    case INDEX_op_mulsh_i64:
2237    case INDEX_op_muluh_i32:
2238    case INDEX_op_muluh_i64:
2239    case INDEX_op_div_i32:
2240    case INDEX_op_div_i64:
2241    case INDEX_op_divu_i32:
2242    case INDEX_op_divu_i64:
2243    case INDEX_op_rem_i32:
2244    case INDEX_op_rem_i64:
2245    case INDEX_op_remu_i32:
2246    case INDEX_op_remu_i64:
2247        return C_O1_I2(r, rZ, rZ);
2248
2249    case INDEX_op_movcond_i32:
2250    case INDEX_op_movcond_i64:
2251        return C_O1_I4(r, rZ, rJ, rZ, rZ);
2252
2253    case INDEX_op_ld_vec:
2254    case INDEX_op_dupm_vec:
2255    case INDEX_op_dup_vec:
2256        return C_O1_I1(w, r);
2257
2258    case INDEX_op_st_vec:
2259        return C_O0_I2(w, r);
2260
2261    case INDEX_op_cmp_vec:
2262        return C_O1_I2(w, w, wM);
2263
2264    case INDEX_op_add_vec:
2265    case INDEX_op_sub_vec:
2266        return C_O1_I2(w, w, wA);
2267
2268    case INDEX_op_and_vec:
2269    case INDEX_op_andc_vec:
2270    case INDEX_op_or_vec:
2271    case INDEX_op_orc_vec:
2272    case INDEX_op_xor_vec:
2273    case INDEX_op_nor_vec:
2274    case INDEX_op_mul_vec:
2275    case INDEX_op_smin_vec:
2276    case INDEX_op_smax_vec:
2277    case INDEX_op_umin_vec:
2278    case INDEX_op_umax_vec:
2279    case INDEX_op_ssadd_vec:
2280    case INDEX_op_usadd_vec:
2281    case INDEX_op_sssub_vec:
2282    case INDEX_op_ussub_vec:
2283    case INDEX_op_shlv_vec:
2284    case INDEX_op_shrv_vec:
2285    case INDEX_op_sarv_vec:
2286    case INDEX_op_rotrv_vec:
2287    case INDEX_op_rotlv_vec:
2288        return C_O1_I2(w, w, w);
2289
2290    case INDEX_op_not_vec:
2291    case INDEX_op_neg_vec:
2292    case INDEX_op_shli_vec:
2293    case INDEX_op_shri_vec:
2294    case INDEX_op_sari_vec:
2295    case INDEX_op_rotli_vec:
2296        return C_O1_I1(w, w);
2297
2298    case INDEX_op_bitsel_vec:
2299        return C_O1_I3(w, w, w, w);
2300
2301    default:
2302        g_assert_not_reached();
2303    }
2304}
2305
2306static const int tcg_target_callee_save_regs[] = {
2307    TCG_REG_S0,     /* used for the global env (TCG_AREG0) */
2308    TCG_REG_S1,
2309    TCG_REG_S2,
2310    TCG_REG_S3,
2311    TCG_REG_S4,
2312    TCG_REG_S5,
2313    TCG_REG_S6,
2314    TCG_REG_S7,
2315    TCG_REG_S8,
2316    TCG_REG_S9,
2317    TCG_REG_RA,     /* should be last for ABI compliance */
2318};
2319
2320/* Stack frame parameters.  */
2321#define REG_SIZE   (TCG_TARGET_REG_BITS / 8)
2322#define SAVE_SIZE  ((int)ARRAY_SIZE(tcg_target_callee_save_regs) * REG_SIZE)
2323#define TEMP_SIZE  (CPU_TEMP_BUF_NLONGS * (int)sizeof(long))
2324#define FRAME_SIZE ((TCG_STATIC_CALL_ARGS_SIZE + TEMP_SIZE + SAVE_SIZE \
2325                     + TCG_TARGET_STACK_ALIGN - 1) \
2326                    & -TCG_TARGET_STACK_ALIGN)
2327#define SAVE_OFS   (TCG_STATIC_CALL_ARGS_SIZE + TEMP_SIZE)
2328
2329/* We're expecting to be able to use an immediate for frame allocation.  */
2330QEMU_BUILD_BUG_ON(FRAME_SIZE > 0x7ff);
2331
2332/* Generate global QEMU prologue and epilogue code */
2333static void tcg_target_qemu_prologue(TCGContext *s)
2334{
2335    int i;
2336
2337    tcg_set_frame(s, TCG_REG_SP, TCG_STATIC_CALL_ARGS_SIZE, TEMP_SIZE);
2338
2339    /* TB prologue */
2340    tcg_out_opc_addi_d(s, TCG_REG_SP, TCG_REG_SP, -FRAME_SIZE);
2341    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
2342        tcg_out_st(s, TCG_TYPE_REG, tcg_target_callee_save_regs[i],
2343                   TCG_REG_SP, SAVE_OFS + i * REG_SIZE);
2344    }
2345
2346    if (!tcg_use_softmmu && guest_base) {
2347        tcg_out_movi(s, TCG_TYPE_PTR, TCG_GUEST_BASE_REG, guest_base);
2348        tcg_regset_set_reg(s->reserved_regs, TCG_GUEST_BASE_REG);
2349    }
2350
2351    /* Call generated code */
2352    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
2353    tcg_out_opc_jirl(s, TCG_REG_ZERO, tcg_target_call_iarg_regs[1], 0);
2354
2355    /* Return path for goto_ptr. Set return value to 0 */
2356    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
2357    tcg_out_mov(s, TCG_TYPE_REG, TCG_REG_A0, TCG_REG_ZERO);
2358
2359    /* TB epilogue */
2360    tb_ret_addr = tcg_splitwx_to_rx(s->code_ptr);
2361    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
2362        tcg_out_ld(s, TCG_TYPE_REG, tcg_target_callee_save_regs[i],
2363                   TCG_REG_SP, SAVE_OFS + i * REG_SIZE);
2364    }
2365
2366    tcg_out_opc_addi_d(s, TCG_REG_SP, TCG_REG_SP, FRAME_SIZE);
2367    tcg_out_opc_jirl(s, TCG_REG_ZERO, TCG_REG_RA, 0);
2368}
2369
2370static void tcg_out_tb_start(TCGContext *s)
2371{
2372    /* nothing to do */
2373}
2374
2375static void tcg_target_init(TCGContext *s)
2376{
2377    unsigned long hwcap = qemu_getauxval(AT_HWCAP);
2378
2379    /* Server and desktop class cpus have UAL; embedded cpus do not. */
2380    if (!(hwcap & HWCAP_LOONGARCH_UAL)) {
2381        error_report("TCG: unaligned access support required; exiting");
2382        exit(EXIT_FAILURE);
2383    }
2384
2385    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
2386    tcg_target_available_regs[TCG_TYPE_I64] = ALL_GENERAL_REGS;
2387
2388    tcg_target_call_clobber_regs = ALL_GENERAL_REGS | ALL_VECTOR_REGS;
2389    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S0);
2390    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S1);
2391    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S2);
2392    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S3);
2393    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S4);
2394    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S5);
2395    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S6);
2396    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S7);
2397    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S8);
2398    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S9);
2399
2400    if (cpuinfo & CPUINFO_LSX) {
2401        tcg_target_available_regs[TCG_TYPE_V64] = ALL_VECTOR_REGS;
2402        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
2403        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V24);
2404        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V25);
2405        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V26);
2406        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V27);
2407        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V28);
2408        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V29);
2409        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V30);
2410        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V31);
2411    }
2412
2413    s->reserved_regs = 0;
2414    tcg_regset_set_reg(s->reserved_regs, TCG_REG_ZERO);
2415    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP0);
2416    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP1);
2417    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP2);
2418    tcg_regset_set_reg(s->reserved_regs, TCG_REG_SP);
2419    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TP);
2420    tcg_regset_set_reg(s->reserved_regs, TCG_REG_RESERVED);
2421    tcg_regset_set_reg(s->reserved_regs, TCG_VEC_TMP0);
2422}
2423
2424typedef struct {
2425    DebugFrameHeader h;
2426    uint8_t fde_def_cfa[4];
2427    uint8_t fde_reg_ofs[ARRAY_SIZE(tcg_target_callee_save_regs) * 2];
2428} DebugFrame;
2429
2430#define ELF_HOST_MACHINE EM_LOONGARCH
2431
2432static const DebugFrame debug_frame = {
2433    .h.cie.len = sizeof(DebugFrameCIE) - 4, /* length after .len member */
2434    .h.cie.id = -1,
2435    .h.cie.version = 1,
2436    .h.cie.code_align = 1,
2437    .h.cie.data_align = -(TCG_TARGET_REG_BITS / 8) & 0x7f, /* sleb128 */
2438    .h.cie.return_column = TCG_REG_RA,
2439
2440    /* Total FDE size does not include the "len" member.  */
2441    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
2442
2443    .fde_def_cfa = {
2444        12, TCG_REG_SP,                 /* DW_CFA_def_cfa sp, ...  */
2445        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
2446        (FRAME_SIZE >> 7)
2447    },
2448    .fde_reg_ofs = {
2449        0x80 + 23, 11,                  /* DW_CFA_offset, s0, -88 */
2450        0x80 + 24, 10,                  /* DW_CFA_offset, s1, -80 */
2451        0x80 + 25, 9,                   /* DW_CFA_offset, s2, -72 */
2452        0x80 + 26, 8,                   /* DW_CFA_offset, s3, -64 */
2453        0x80 + 27, 7,                   /* DW_CFA_offset, s4, -56 */
2454        0x80 + 28, 6,                   /* DW_CFA_offset, s5, -48 */
2455        0x80 + 29, 5,                   /* DW_CFA_offset, s6, -40 */
2456        0x80 + 30, 4,                   /* DW_CFA_offset, s7, -32 */
2457        0x80 + 31, 3,                   /* DW_CFA_offset, s8, -24 */
2458        0x80 + 22, 2,                   /* DW_CFA_offset, s9, -16 */
2459        0x80 + 1 , 1,                   /* DW_CFA_offset, ra, -8 */
2460    }
2461};
2462
2463void tcg_register_jit(const void *buf, size_t buf_size)
2464{
2465    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
2466}
2467