xref: /openbmc/qemu/tcg/loongarch64/tcg-target.c.inc (revision 4d137ff819bae33d045f13bb9186e3a2c71cb7e4)
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 <asm/hwcap.h>
33
34/* used for function call generation */
35#define TCG_REG_CALL_STACK              TCG_REG_SP
36#define TCG_TARGET_STACK_ALIGN          16
37#define TCG_TARGET_CALL_STACK_OFFSET    0
38#define TCG_TARGET_CALL_ARG_I32         TCG_CALL_ARG_NORMAL
39#define TCG_TARGET_CALL_ARG_I64         TCG_CALL_ARG_NORMAL
40#define TCG_TARGET_CALL_ARG_I128        TCG_CALL_ARG_NORMAL
41#define TCG_TARGET_CALL_RET_I128        TCG_CALL_RET_NORMAL
42
43#ifdef CONFIG_DEBUG_TCG
44static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
45    "zero",
46    "ra",
47    "tp",
48    "sp",
49    "a0",
50    "a1",
51    "a2",
52    "a3",
53    "a4",
54    "a5",
55    "a6",
56    "a7",
57    "t0",
58    "t1",
59    "t2",
60    "t3",
61    "t4",
62    "t5",
63    "t6",
64    "t7",
65    "t8",
66    "r21", /* reserved in the LP64* ABI, hence no ABI name */
67    "s9",
68    "s0",
69    "s1",
70    "s2",
71    "s3",
72    "s4",
73    "s5",
74    "s6",
75    "s7",
76    "s8",
77    "vr0",
78    "vr1",
79    "vr2",
80    "vr3",
81    "vr4",
82    "vr5",
83    "vr6",
84    "vr7",
85    "vr8",
86    "vr9",
87    "vr10",
88    "vr11",
89    "vr12",
90    "vr13",
91    "vr14",
92    "vr15",
93    "vr16",
94    "vr17",
95    "vr18",
96    "vr19",
97    "vr20",
98    "vr21",
99    "vr22",
100    "vr23",
101    "vr24",
102    "vr25",
103    "vr26",
104    "vr27",
105    "vr28",
106    "vr29",
107    "vr30",
108    "vr31",
109};
110#endif
111
112static const int tcg_target_reg_alloc_order[] = {
113    /* Registers preserved across calls */
114    /* TCG_REG_S0 reserved for TCG_AREG0 */
115    TCG_REG_S1,
116    TCG_REG_S2,
117    TCG_REG_S3,
118    TCG_REG_S4,
119    TCG_REG_S5,
120    TCG_REG_S6,
121    TCG_REG_S7,
122    TCG_REG_S8,
123    TCG_REG_S9,
124
125    /* Registers (potentially) clobbered across calls */
126    TCG_REG_T0,
127    TCG_REG_T1,
128    TCG_REG_T2,
129    TCG_REG_T3,
130    TCG_REG_T4,
131    TCG_REG_T5,
132    TCG_REG_T6,
133    TCG_REG_T7,
134    TCG_REG_T8,
135
136    /* Argument registers, opposite order of allocation.  */
137    TCG_REG_A7,
138    TCG_REG_A6,
139    TCG_REG_A5,
140    TCG_REG_A4,
141    TCG_REG_A3,
142    TCG_REG_A2,
143    TCG_REG_A1,
144    TCG_REG_A0,
145
146    /* Vector registers */
147    TCG_REG_V0, TCG_REG_V1, TCG_REG_V2, TCG_REG_V3,
148    TCG_REG_V4, TCG_REG_V5, TCG_REG_V6, TCG_REG_V7,
149    TCG_REG_V8, TCG_REG_V9, TCG_REG_V10, TCG_REG_V11,
150    TCG_REG_V12, TCG_REG_V13, TCG_REG_V14, TCG_REG_V15,
151    TCG_REG_V16, TCG_REG_V17, TCG_REG_V18, TCG_REG_V19,
152    TCG_REG_V20, TCG_REG_V21, TCG_REG_V22, TCG_REG_V23,
153    /* V24 - V31 are caller-saved, and skipped.  */
154};
155
156static const int tcg_target_call_iarg_regs[] = {
157    TCG_REG_A0,
158    TCG_REG_A1,
159    TCG_REG_A2,
160    TCG_REG_A3,
161    TCG_REG_A4,
162    TCG_REG_A5,
163    TCG_REG_A6,
164    TCG_REG_A7,
165};
166
167static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
168{
169    tcg_debug_assert(kind == TCG_CALL_RET_NORMAL);
170    tcg_debug_assert(slot >= 0 && slot <= 1);
171    return TCG_REG_A0 + slot;
172}
173
174#define TCG_GUEST_BASE_REG TCG_REG_S1
175
176#define TCG_CT_CONST_S12   0x100
177#define TCG_CT_CONST_S32   0x200
178#define TCG_CT_CONST_U12   0x400
179#define TCG_CT_CONST_WSZ   0x800
180#define TCG_CT_CONST_VCMP  0x1000
181#define TCG_CT_CONST_VADD  0x2000
182
183#define ALL_GENERAL_REGS   MAKE_64BIT_MASK(0, 32)
184#define ALL_VECTOR_REGS    MAKE_64BIT_MASK(32, 32)
185
186static inline tcg_target_long sextreg(tcg_target_long val, int pos, int len)
187{
188    return sextract64(val, pos, len);
189}
190
191/* test if a constant matches the constraint */
192static bool tcg_target_const_match(int64_t val, int ct,
193                                   TCGType type, TCGCond cond, int vece)
194{
195    if (ct & TCG_CT_CONST) {
196        return true;
197    }
198    if ((ct & TCG_CT_CONST_S12) && val == sextreg(val, 0, 12)) {
199        return true;
200    }
201    if ((ct & TCG_CT_CONST_S32) && val == (int32_t)val) {
202        return true;
203    }
204    if ((ct & TCG_CT_CONST_U12) && 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    if (ct & (TCG_CT_CONST_VCMP | TCG_CT_CONST_VADD)) {
211        int64_t vec_val = sextract64(val, 0, 8 << vece);
212        if (ct & TCG_CT_CONST_VCMP) {
213            switch (cond) {
214            case TCG_COND_EQ:
215            case TCG_COND_LE:
216            case TCG_COND_LT:
217                return -0x10 <= vec_val && vec_val <= 0x0f;
218            case TCG_COND_LEU:
219            case TCG_COND_LTU:
220                return 0x00 <= vec_val && vec_val <= 0x1f;
221            default:
222                return false;
223            }
224        }
225        if ((ct & TCG_CT_CONST_VADD) && -0x1f <= vec_val && vec_val <= 0x1f) {
226            return true;
227        }
228    }
229    return false;
230}
231
232/*
233 * Relocations
234 */
235
236/*
237 * Relocation records defined in LoongArch ELF psABI v1.00 is way too
238 * complicated; a whopping stack machine is needed to stuff the fields, at
239 * the very least one SOP_PUSH and one SOP_POP (of the correct format) are
240 * needed.
241 *
242 * Hence, define our own simpler relocation types. Numbers are chosen as to
243 * not collide with potential future additions to the true ELF relocation
244 * type enum.
245 */
246
247/* Field Sk16, shifted right by 2; suitable for conditional jumps */
248#define R_LOONGARCH_BR_SK16     256
249/* Field Sd10k16, shifted right by 2; suitable for B and BL */
250#define R_LOONGARCH_BR_SD10K16  257
251
252static bool reloc_br_sk16(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
253{
254    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
255    intptr_t offset = (intptr_t)target - (intptr_t)src_rx;
256
257    tcg_debug_assert((offset & 3) == 0);
258    offset >>= 2;
259    if (offset == sextreg(offset, 0, 16)) {
260        *src_rw = deposit64(*src_rw, 10, 16, offset);
261        return true;
262    }
263
264    return false;
265}
266
267static bool reloc_br_sd10k16(tcg_insn_unit *src_rw,
268                             const tcg_insn_unit *target)
269{
270    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
271    intptr_t offset = (intptr_t)target - (intptr_t)src_rx;
272
273    tcg_debug_assert((offset & 3) == 0);
274    offset >>= 2;
275    if (offset == sextreg(offset, 0, 26)) {
276        *src_rw = deposit64(*src_rw, 0, 10, offset >> 16); /* slot d10 */
277        *src_rw = deposit64(*src_rw, 10, 16, offset); /* slot k16 */
278        return true;
279    }
280
281    return false;
282}
283
284static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
285                        intptr_t value, intptr_t addend)
286{
287    tcg_debug_assert(addend == 0);
288    switch (type) {
289    case R_LOONGARCH_BR_SK16:
290        return reloc_br_sk16(code_ptr, (tcg_insn_unit *)value);
291    case R_LOONGARCH_BR_SD10K16:
292        return reloc_br_sd10k16(code_ptr, (tcg_insn_unit *)value);
293    default:
294        g_assert_not_reached();
295    }
296}
297
298#include "tcg-insn-defs.c.inc"
299
300/*
301 * TCG intrinsics
302 */
303
304static void tcg_out_mb(TCGContext *s, TCGArg a0)
305{
306    /* Baseline LoongArch only has the full barrier, unfortunately.  */
307    tcg_out_opc_dbar(s, 0);
308}
309
310static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
311{
312    if (ret == arg) {
313        return true;
314    }
315    switch (type) {
316    case TCG_TYPE_I32:
317    case TCG_TYPE_I64:
318        if (ret < TCG_REG_V0) {
319            if (arg < TCG_REG_V0) {
320                /*
321                 * Conventional register-register move used in LoongArch is
322                 * `or dst, src, zero`.
323                 */
324                tcg_out_opc_or(s, ret, arg, TCG_REG_ZERO);
325            } else {
326                tcg_out_opc_movfr2gr_d(s, ret, arg);
327            }
328        } else {
329            if (arg < TCG_REG_V0) {
330                tcg_out_opc_movgr2fr_d(s, ret, arg);
331            } else {
332                tcg_out_opc_fmov_d(s, ret, arg);
333            }
334        }
335        break;
336    case TCG_TYPE_V64:
337    case TCG_TYPE_V128:
338        tcg_out_opc_vori_b(s, ret, arg, 0);
339        break;
340    case TCG_TYPE_V256:
341        tcg_out_opc_xvori_b(s, ret, arg, 0);
342        break;
343    default:
344        g_assert_not_reached();
345    }
346    return true;
347}
348
349/* Loads a 32-bit immediate into rd, sign-extended.  */
350static void tcg_out_movi_i32(TCGContext *s, TCGReg rd, int32_t val)
351{
352    tcg_target_long lo = sextreg(val, 0, 12);
353    tcg_target_long hi12 = sextreg(val, 12, 20);
354
355    /* Single-instruction cases.  */
356    if (hi12 == 0) {
357        /* val fits in uimm12: ori rd, zero, val */
358        tcg_out_opc_ori(s, rd, TCG_REG_ZERO, val);
359        return;
360    }
361    if (hi12 == sextreg(lo, 12, 20)) {
362        /* val fits in simm12: addi.w rd, zero, val */
363        tcg_out_opc_addi_w(s, rd, TCG_REG_ZERO, val);
364        return;
365    }
366
367    /* High bits must be set; load with lu12i.w + optional ori.  */
368    tcg_out_opc_lu12i_w(s, rd, hi12);
369    if (lo != 0) {
370        tcg_out_opc_ori(s, rd, rd, lo & 0xfff);
371    }
372}
373
374static void tcg_out_movi(TCGContext *s, TCGType type, TCGReg rd,
375                         tcg_target_long val)
376{
377    /*
378     * LoongArch conventionally loads 64-bit immediates in at most 4 steps,
379     * with dedicated instructions for filling the respective bitfields
380     * below:
381     *
382     *        6                   5                   4               3
383     *  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
384     * +-----------------------+---------------------------------------+...
385     * |          hi52         |                  hi32                 |
386     * +-----------------------+---------------------------------------+...
387     *       3                   2                   1
388     *     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
389     * ...+-------------------------------------+-------------------------+
390     *    |                 hi12                |            lo           |
391     * ...+-------------------------------------+-------------------------+
392     *
393     * Check if val belong to one of the several fast cases, before falling
394     * back to the slow path.
395     */
396
397    intptr_t src_rx, pc_offset;
398    tcg_target_long hi12, hi32, hi52;
399
400    /* Value fits in signed i32.  */
401    if (type == TCG_TYPE_I32 || val == (int32_t)val) {
402        tcg_out_movi_i32(s, rd, val);
403        return;
404    }
405
406    /* PC-relative cases.  */
407    src_rx = (intptr_t)tcg_splitwx_to_rx(s->code_ptr);
408    if ((val & 3) == 0) {
409        pc_offset = val - src_rx;
410        if (pc_offset == sextreg(pc_offset, 0, 22)) {
411            /* Single pcaddu2i.  */
412            tcg_out_opc_pcaddu2i(s, rd, pc_offset >> 2);
413            return;
414        }
415    }
416
417    pc_offset = (val >> 12) - (src_rx >> 12);
418    if (pc_offset == sextreg(pc_offset, 0, 20)) {
419        /* Load with pcalau12i + ori.  */
420        tcg_target_long val_lo = val & 0xfff;
421        tcg_out_opc_pcalau12i(s, rd, pc_offset);
422        if (val_lo != 0) {
423            tcg_out_opc_ori(s, rd, rd, val_lo);
424        }
425        return;
426    }
427
428    hi12 = sextreg(val, 12, 20);
429    hi32 = sextreg(val, 32, 20);
430    hi52 = sextreg(val, 52, 12);
431
432    /* Single cu52i.d case.  */
433    if ((hi52 != 0) && (ctz64(val) >= 52)) {
434        tcg_out_opc_cu52i_d(s, rd, TCG_REG_ZERO, hi52);
435        return;
436    }
437
438    /* Slow path.  Initialize the low 32 bits, then concat high bits.  */
439    tcg_out_movi_i32(s, rd, val);
440
441    /* Load hi32 and hi52 explicitly when they are unexpected values. */
442    if (hi32 != sextreg(hi12, 20, 20)) {
443        tcg_out_opc_cu32i_d(s, rd, hi32);
444    }
445
446    if (hi52 != sextreg(hi32, 20, 12)) {
447        tcg_out_opc_cu52i_d(s, rd, rd, hi52);
448    }
449}
450
451static void tcg_out_addi(TCGContext *s, TCGType type, TCGReg rd,
452                         TCGReg rs, tcg_target_long imm)
453{
454    tcg_target_long lo12 = sextreg(imm, 0, 12);
455    tcg_target_long hi16 = sextreg(imm - lo12, 16, 16);
456
457    /*
458     * Note that there's a hole in between hi16 and lo12:
459     *
460     *       3                   2                   1                   0
461     *     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
462     * ...+-------------------------------+-------+-----------------------+
463     *    |             hi16              |       |          lo12         |
464     * ...+-------------------------------+-------+-----------------------+
465     *
466     * For bits within that hole, it's more efficient to use LU12I and ADD.
467     */
468    if (imm == (hi16 << 16) + lo12) {
469        if (hi16) {
470            tcg_out_opc_addu16i_d(s, rd, rs, hi16);
471            rs = rd;
472        }
473        if (type == TCG_TYPE_I32) {
474            tcg_out_opc_addi_w(s, rd, rs, lo12);
475        } else if (lo12) {
476            tcg_out_opc_addi_d(s, rd, rs, lo12);
477        } else {
478            tcg_out_mov(s, type, rd, rs);
479        }
480    } else {
481        tcg_out_movi(s, type, TCG_REG_TMP0, imm);
482        if (type == TCG_TYPE_I32) {
483            tcg_out_opc_add_w(s, rd, rs, TCG_REG_TMP0);
484        } else {
485            tcg_out_opc_add_d(s, rd, rs, TCG_REG_TMP0);
486        }
487    }
488}
489
490static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
491{
492    return false;
493}
494
495static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
496                             tcg_target_long imm)
497{
498    /* This function is only used for passing structs by reference. */
499    g_assert_not_reached();
500}
501
502static void tcg_out_ext8u(TCGContext *s, TCGReg ret, TCGReg arg)
503{
504    tcg_out_opc_andi(s, ret, arg, 0xff);
505}
506
507static void tcg_out_ext16u(TCGContext *s, TCGReg ret, TCGReg arg)
508{
509    tcg_out_opc_bstrpick_w(s, ret, arg, 0, 15);
510}
511
512static void tcg_out_ext32u(TCGContext *s, TCGReg ret, TCGReg arg)
513{
514    tcg_out_opc_bstrpick_d(s, ret, arg, 0, 31);
515}
516
517static void tcg_out_ext8s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
518{
519    tcg_out_opc_sext_b(s, ret, arg);
520}
521
522static void tcg_out_ext16s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
523{
524    tcg_out_opc_sext_h(s, ret, arg);
525}
526
527static void tcg_out_ext32s(TCGContext *s, TCGReg ret, TCGReg arg)
528{
529    tcg_out_opc_addi_w(s, ret, arg, 0);
530}
531
532static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg)
533{
534    if (ret != arg) {
535        tcg_out_ext32s(s, ret, arg);
536    }
537}
538
539static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg)
540{
541    tcg_out_ext32u(s, ret, arg);
542}
543
544static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg ret, TCGReg arg)
545{
546    tcg_out_ext32s(s, ret, arg);
547}
548
549#define SETCOND_INV    TCG_TARGET_NB_REGS
550#define SETCOND_NEZ    (SETCOND_INV << 1)
551#define SETCOND_FLAGS  (SETCOND_INV | SETCOND_NEZ)
552
553static int tcg_out_setcond_int(TCGContext *s, TCGCond cond, TCGReg ret,
554                               TCGReg arg1, tcg_target_long arg2, bool c2)
555{
556    int flags = 0;
557
558    switch (cond) {
559    case TCG_COND_EQ:    /* -> NE  */
560    case TCG_COND_GE:    /* -> LT  */
561    case TCG_COND_GEU:   /* -> LTU */
562    case TCG_COND_GT:    /* -> LE  */
563    case TCG_COND_GTU:   /* -> LEU */
564        cond = tcg_invert_cond(cond);
565        flags ^= SETCOND_INV;
566        break;
567    default:
568        break;
569    }
570
571    switch (cond) {
572    case TCG_COND_LE:
573    case TCG_COND_LEU:
574        /*
575         * If we have a constant input, the most efficient way to implement
576         * LE is by adding 1 and using LT.  Watch out for wrap around for LEU.
577         * We don't need to care for this for LE because the constant input
578         * is still constrained to int32_t, and INT32_MAX+1 is representable
579         * in the 64-bit temporary register.
580         */
581        if (c2) {
582            if (cond == TCG_COND_LEU) {
583                /* unsigned <= -1 is true */
584                if (arg2 == -1) {
585                    tcg_out_movi(s, TCG_TYPE_REG, ret, !(flags & SETCOND_INV));
586                    return ret;
587                }
588                cond = TCG_COND_LTU;
589            } else {
590                cond = TCG_COND_LT;
591            }
592            arg2 += 1;
593        } else {
594            TCGReg tmp = arg2;
595            arg2 = arg1;
596            arg1 = tmp;
597            cond = tcg_swap_cond(cond);    /* LE -> GE */
598            cond = tcg_invert_cond(cond);  /* GE -> LT */
599            flags ^= SETCOND_INV;
600        }
601        break;
602    default:
603        break;
604    }
605
606    switch (cond) {
607    case TCG_COND_NE:
608        flags |= SETCOND_NEZ;
609        if (!c2) {
610            tcg_out_opc_xor(s, ret, arg1, arg2);
611        } else if (arg2 == 0) {
612            ret = arg1;
613        } else if (arg2 >= 0 && arg2 <= 0xfff) {
614            tcg_out_opc_xori(s, ret, arg1, arg2);
615        } else {
616            tcg_out_addi(s, TCG_TYPE_REG, ret, arg1, -arg2);
617        }
618        break;
619
620    case TCG_COND_LT:
621    case TCG_COND_LTU:
622        if (c2) {
623            if (arg2 >= -0x800 && arg2 <= 0x7ff) {
624                if (cond == TCG_COND_LT) {
625                    tcg_out_opc_slti(s, ret, arg1, arg2);
626                } else {
627                    tcg_out_opc_sltui(s, ret, arg1, arg2);
628                }
629                break;
630            }
631            tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_TMP0, arg2);
632            arg2 = TCG_REG_TMP0;
633        }
634        if (cond == TCG_COND_LT) {
635            tcg_out_opc_slt(s, ret, arg1, arg2);
636        } else {
637            tcg_out_opc_sltu(s, ret, arg1, arg2);
638        }
639        break;
640
641    default:
642        g_assert_not_reached();
643    }
644
645    return ret | flags;
646}
647
648static void tcg_out_setcond(TCGContext *s, TCGCond cond, TCGReg ret,
649                            TCGReg arg1, tcg_target_long arg2,
650                            bool c2, bool neg)
651{
652    int tmpflags = tcg_out_setcond_int(s, cond, ret, arg1, arg2, c2);
653    TCGReg tmp = tmpflags & ~SETCOND_FLAGS;
654
655    if (neg) {
656        /* If intermediate result is zero/non-zero: test != 0. */
657        if (tmpflags & SETCOND_NEZ) {
658            tcg_out_opc_sltu(s, ret, TCG_REG_ZERO, tmp);
659            tmp = ret;
660        }
661        /* Produce the 0/-1 result. */
662        if (tmpflags & SETCOND_INV) {
663            tcg_out_opc_addi_d(s, ret, tmp, -1);
664        } else {
665            tcg_out_opc_sub_d(s, ret, TCG_REG_ZERO, tmp);
666        }
667    } else {
668        switch (tmpflags & SETCOND_FLAGS) {
669        case 0:
670            tcg_debug_assert(tmp == ret);
671            break;
672        case SETCOND_INV:
673            /* Intermediate result is boolean: simply invert. */
674            tcg_out_opc_xori(s, ret, tmp, 1);
675            break;
676        case SETCOND_NEZ:
677            /* Intermediate result is zero/non-zero: test != 0. */
678            tcg_out_opc_sltu(s, ret, TCG_REG_ZERO, tmp);
679            break;
680        case SETCOND_NEZ | SETCOND_INV:
681            /* Intermediate result is zero/non-zero: test == 0. */
682            tcg_out_opc_sltui(s, ret, tmp, 1);
683            break;
684        default:
685            g_assert_not_reached();
686        }
687    }
688}
689
690static void tgen_setcond(TCGContext *s, TCGType type, TCGCond cond,
691                         TCGReg dest, TCGReg arg1, TCGReg arg2)
692{
693    tcg_out_setcond(s, cond, dest, arg1, arg2, false, false);
694}
695
696static void tgen_setcondi(TCGContext *s, TCGType type, TCGCond cond,
697                          TCGReg dest, TCGReg arg1, tcg_target_long arg2)
698{
699    tcg_out_setcond(s, cond, dest, arg1, arg2, true, false);
700}
701
702static const TCGOutOpSetcond outop_setcond = {
703    .base.static_constraint = C_O1_I2(r, r, rJ),
704    .out_rrr = tgen_setcond,
705    .out_rri = tgen_setcondi,
706};
707
708static void tgen_negsetcond(TCGContext *s, TCGType type, TCGCond cond,
709                            TCGReg dest, TCGReg arg1, TCGReg arg2)
710{
711    tcg_out_setcond(s, cond, dest, arg1, arg2, false, true);
712}
713
714static void tgen_negsetcondi(TCGContext *s, TCGType type, TCGCond cond,
715                             TCGReg dest, TCGReg arg1, tcg_target_long arg2)
716{
717    tcg_out_setcond(s, cond, dest, arg1, arg2, true, true);
718}
719
720static const TCGOutOpSetcond outop_negsetcond = {
721    .base.static_constraint = C_O1_I2(r, r, rJ),
722    .out_rrr = tgen_negsetcond,
723    .out_rri = tgen_negsetcondi,
724};
725
726static void tgen_movcond(TCGContext *s, TCGType type, TCGCond cond,
727                         TCGReg ret, TCGReg c1, TCGArg c2, bool const_c2,
728                         TCGArg v1, bool const_v1, TCGArg v2, bool const_v2)
729{
730    int tmpflags = tcg_out_setcond_int(s, cond, TCG_REG_TMP0, c1, c2, const_c2);
731    TCGReg t;
732
733    /* Standardize the test below to t != 0. */
734    if (tmpflags & SETCOND_INV) {
735        t = v1, v1 = v2, v2 = t;
736    }
737
738    t = tmpflags & ~SETCOND_FLAGS;
739    if (v1 == TCG_REG_ZERO) {
740        tcg_out_opc_masknez(s, ret, v2, t);
741    } else if (v2 == TCG_REG_ZERO) {
742        tcg_out_opc_maskeqz(s, ret, v1, t);
743    } else {
744        tcg_out_opc_masknez(s, TCG_REG_TMP2, v2, t); /* t ? 0 : v2 */
745        tcg_out_opc_maskeqz(s, TCG_REG_TMP1, v1, t); /* t ? v1 : 0 */
746        tcg_out_opc_or(s, ret, TCG_REG_TMP1, TCG_REG_TMP2);
747    }
748}
749
750static const TCGOutOpMovcond outop_movcond = {
751    .base.static_constraint = C_O1_I4(r, r, rJ, rz, rz),
752    .out = tgen_movcond,
753};
754
755/*
756 * Branch helpers
757 */
758
759static const struct {
760    LoongArchInsn op;
761    bool swap;
762} tcg_brcond_to_loongarch[] = {
763    [TCG_COND_EQ] =  { OPC_BEQ,  false },
764    [TCG_COND_NE] =  { OPC_BNE,  false },
765    [TCG_COND_LT] =  { OPC_BGT,  true  },
766    [TCG_COND_GE] =  { OPC_BLE,  true  },
767    [TCG_COND_LE] =  { OPC_BLE,  false },
768    [TCG_COND_GT] =  { OPC_BGT,  false },
769    [TCG_COND_LTU] = { OPC_BGTU, true  },
770    [TCG_COND_GEU] = { OPC_BLEU, true  },
771    [TCG_COND_LEU] = { OPC_BLEU, false },
772    [TCG_COND_GTU] = { OPC_BGTU, false }
773};
774
775static void tgen_brcond(TCGContext *s, TCGType type, TCGCond cond,
776                        TCGReg arg1, TCGReg arg2, TCGLabel *l)
777{
778    LoongArchInsn op = tcg_brcond_to_loongarch[cond].op;
779
780    tcg_debug_assert(op != 0);
781
782    if (tcg_brcond_to_loongarch[cond].swap) {
783        TCGReg t = arg1;
784        arg1 = arg2;
785        arg2 = t;
786    }
787
788    /* all conditional branch insns belong to DJSk16-format */
789    tcg_out_reloc(s, s->code_ptr, R_LOONGARCH_BR_SK16, l, 0);
790    tcg_out32(s, encode_djsk16_insn(op, arg1, arg2, 0));
791}
792
793static const TCGOutOpBrcond outop_brcond = {
794    .base.static_constraint = C_O0_I2(r, rz),
795    .out_rr = tgen_brcond,
796};
797
798static void tcg_out_call_int(TCGContext *s, const tcg_insn_unit *arg, bool tail)
799{
800    TCGReg link = tail ? TCG_REG_ZERO : TCG_REG_RA;
801    ptrdiff_t offset = tcg_pcrel_diff(s, arg);
802
803    tcg_debug_assert((offset & 3) == 0);
804    if (offset == sextreg(offset, 0, 28)) {
805        /* short jump: +/- 256MiB */
806        if (tail) {
807            tcg_out_opc_b(s, offset >> 2);
808        } else {
809            tcg_out_opc_bl(s, offset >> 2);
810        }
811    } else if (offset == sextreg(offset, 0, 38)) {
812        /* long jump: +/- 256GiB */
813        tcg_target_long lo = sextreg(offset, 0, 18);
814        tcg_target_long hi = offset - lo;
815        tcg_out_opc_pcaddu18i(s, TCG_REG_TMP0, hi >> 18);
816        tcg_out_opc_jirl(s, link, TCG_REG_TMP0, lo >> 2);
817    } else {
818        /* far jump: 64-bit */
819        tcg_target_long lo = sextreg((tcg_target_long)arg, 0, 18);
820        tcg_target_long hi = (tcg_target_long)arg - lo;
821        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, hi);
822        tcg_out_opc_jirl(s, link, TCG_REG_TMP0, lo >> 2);
823    }
824}
825
826static void tcg_out_call(TCGContext *s, const tcg_insn_unit *arg,
827                         const TCGHelperInfo *info)
828{
829    tcg_out_call_int(s, arg, false);
830}
831
832/*
833 * Load/store helpers
834 */
835
836static void tcg_out_ldst(TCGContext *s, LoongArchInsn opc, TCGReg data,
837                         TCGReg addr, intptr_t offset)
838{
839    intptr_t imm12 = sextreg(offset, 0, 12);
840
841    if (offset != imm12) {
842        intptr_t diff = tcg_pcrel_diff(s, (void *)offset);
843
844        if (addr == TCG_REG_ZERO && diff == (int32_t)diff) {
845            imm12 = sextreg(diff, 0, 12);
846            tcg_out_opc_pcaddu12i(s, TCG_REG_TMP2, (diff - imm12) >> 12);
847        } else {
848            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP2, offset - imm12);
849            if (addr != TCG_REG_ZERO) {
850                tcg_out_opc_add_d(s, TCG_REG_TMP2, TCG_REG_TMP2, addr);
851            }
852        }
853        addr = TCG_REG_TMP2;
854    }
855
856    switch (opc) {
857    case OPC_LD_B:
858    case OPC_LD_BU:
859    case OPC_LD_H:
860    case OPC_LD_HU:
861    case OPC_LD_W:
862    case OPC_LD_WU:
863    case OPC_LD_D:
864    case OPC_ST_B:
865    case OPC_ST_H:
866    case OPC_ST_W:
867    case OPC_ST_D:
868        tcg_out32(s, encode_djsk12_insn(opc, data, addr, imm12));
869        break;
870    case OPC_FLD_S:
871    case OPC_FLD_D:
872    case OPC_FST_S:
873    case OPC_FST_D:
874        tcg_out32(s, encode_fdjsk12_insn(opc, data, addr, imm12));
875        break;
876    default:
877        g_assert_not_reached();
878    }
879}
880
881static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg dest,
882                       TCGReg base, intptr_t offset)
883{
884    switch (type) {
885    case TCG_TYPE_I32:
886        if (dest < TCG_REG_V0) {
887            tcg_out_ldst(s, OPC_LD_W, dest, base, offset);
888        } else {
889            tcg_out_ldst(s, OPC_FLD_S, dest, base, offset);
890        }
891        break;
892    case TCG_TYPE_I64:
893    case TCG_TYPE_V64:
894        if (dest < TCG_REG_V0) {
895            tcg_out_ldst(s, OPC_LD_D, dest, base, offset);
896        } else {
897            tcg_out_ldst(s, OPC_FLD_D, dest, base, offset);
898        }
899        break;
900    case TCG_TYPE_V128:
901        if (-0x800 <= offset && offset <= 0x7ff) {
902            tcg_out_opc_vld(s, dest, base, offset);
903        } else {
904            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, offset);
905            tcg_out_opc_vldx(s, dest, base, TCG_REG_TMP0);
906        }
907        break;
908    case TCG_TYPE_V256:
909        if (-0x800 <= offset && offset <= 0x7ff) {
910            tcg_out_opc_xvld(s, dest, base, offset);
911        } else {
912            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, offset);
913            tcg_out_opc_xvldx(s, dest, base, TCG_REG_TMP0);
914        }
915        break;
916    default:
917        g_assert_not_reached();
918    }
919}
920
921static void tcg_out_st(TCGContext *s, TCGType type, TCGReg src,
922                       TCGReg base, intptr_t offset)
923{
924    switch (type) {
925    case TCG_TYPE_I32:
926        if (src < TCG_REG_V0) {
927            tcg_out_ldst(s, OPC_ST_W, src, base, offset);
928        } else {
929            tcg_out_ldst(s, OPC_FST_S, src, base, offset);
930        }
931        break;
932    case TCG_TYPE_I64:
933    case TCG_TYPE_V64:
934        if (src < TCG_REG_V0) {
935            tcg_out_ldst(s, OPC_ST_D, src, base, offset);
936        } else {
937            tcg_out_ldst(s, OPC_FST_D, src, base, offset);
938        }
939        break;
940    case TCG_TYPE_V128:
941        if (-0x800 <= offset && offset <= 0x7ff) {
942            tcg_out_opc_vst(s, src, base, offset);
943        } else {
944            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, offset);
945            tcg_out_opc_vstx(s, src, base, TCG_REG_TMP0);
946        }
947        break;
948    case TCG_TYPE_V256:
949        if (-0x800 <= offset && offset <= 0x7ff) {
950            tcg_out_opc_xvst(s, src, base, offset);
951        } else {
952            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, offset);
953            tcg_out_opc_xvstx(s, src, base, TCG_REG_TMP0);
954        }
955        break;
956    default:
957        g_assert_not_reached();
958    }
959}
960
961static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
962                        TCGReg base, intptr_t ofs)
963{
964    if (val == 0) {
965        tcg_out_st(s, type, TCG_REG_ZERO, base, ofs);
966        return true;
967    }
968    return false;
969}
970
971/*
972 * Load/store helpers for SoftMMU, and qemu_ld/st implementations
973 */
974
975static bool tcg_out_goto(TCGContext *s, const tcg_insn_unit *target)
976{
977    tcg_out_opc_b(s, 0);
978    return reloc_br_sd10k16(s->code_ptr - 1, target);
979}
980
981static const TCGLdstHelperParam ldst_helper_param = {
982    .ntmp = 1, .tmp = { TCG_REG_TMP0 }
983};
984
985static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
986{
987    MemOp opc = get_memop(l->oi);
988
989    /* resolve label address */
990    if (!reloc_br_sk16(l->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
991        return false;
992    }
993
994    tcg_out_ld_helper_args(s, l, &ldst_helper_param);
995    tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SIZE], false);
996    tcg_out_ld_helper_ret(s, l, false, &ldst_helper_param);
997    return tcg_out_goto(s, l->raddr);
998}
999
1000static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
1001{
1002    MemOp opc = get_memop(l->oi);
1003
1004    /* resolve label address */
1005    if (!reloc_br_sk16(l->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1006        return false;
1007    }
1008
1009    tcg_out_st_helper_args(s, l, &ldst_helper_param);
1010    tcg_out_call_int(s, qemu_st_helpers[opc & MO_SIZE], false);
1011    return tcg_out_goto(s, l->raddr);
1012}
1013
1014typedef struct {
1015    TCGReg base;
1016    TCGReg index;
1017    TCGAtomAlign aa;
1018} HostAddress;
1019
1020bool tcg_target_has_memory_bswap(MemOp memop)
1021{
1022    return false;
1023}
1024
1025/* We expect to use a 12-bit negative offset from ENV.  */
1026#define MIN_TLB_MASK_TABLE_OFS  -(1 << 11)
1027
1028/*
1029 * For system-mode, perform the TLB load and compare.
1030 * For user-mode, perform any required alignment tests.
1031 * In both cases, return a TCGLabelQemuLdst structure if the slow path
1032 * is required and fill in @h with the host address for the fast path.
1033 */
1034static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
1035                                           TCGReg addr_reg, MemOpIdx oi,
1036                                           bool is_ld)
1037{
1038    TCGType addr_type = s->addr_type;
1039    TCGLabelQemuLdst *ldst = NULL;
1040    MemOp opc = get_memop(oi);
1041    MemOp a_bits;
1042
1043    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, false);
1044    a_bits = h->aa.align;
1045
1046    if (tcg_use_softmmu) {
1047        unsigned s_bits = opc & MO_SIZE;
1048        int mem_index = get_mmuidx(oi);
1049        int fast_ofs = tlb_mask_table_ofs(s, mem_index);
1050        int mask_ofs = fast_ofs + offsetof(CPUTLBDescFast, mask);
1051        int table_ofs = fast_ofs + offsetof(CPUTLBDescFast, table);
1052
1053        ldst = new_ldst_label(s);
1054        ldst->is_ld = is_ld;
1055        ldst->oi = oi;
1056        ldst->addr_reg = addr_reg;
1057
1058        tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP0, TCG_AREG0, mask_ofs);
1059        tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP1, TCG_AREG0, table_ofs);
1060
1061        tcg_out_opc_srli_d(s, TCG_REG_TMP2, addr_reg,
1062                           s->page_bits - CPU_TLB_ENTRY_BITS);
1063        tcg_out_opc_and(s, TCG_REG_TMP2, TCG_REG_TMP2, TCG_REG_TMP0);
1064        tcg_out_opc_add_d(s, TCG_REG_TMP2, TCG_REG_TMP2, TCG_REG_TMP1);
1065
1066        /* Load the tlb comparator and the addend.  */
1067        QEMU_BUILD_BUG_ON(HOST_BIG_ENDIAN);
1068        tcg_out_ld(s, addr_type, TCG_REG_TMP0, TCG_REG_TMP2,
1069                   is_ld ? offsetof(CPUTLBEntry, addr_read)
1070                         : offsetof(CPUTLBEntry, addr_write));
1071        tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP2, TCG_REG_TMP2,
1072                   offsetof(CPUTLBEntry, addend));
1073
1074        /*
1075         * For aligned accesses, we check the first byte and include the
1076         * alignment bits within the address.  For unaligned access, we
1077         * check that we don't cross pages using the address of the last
1078         * byte of the access.
1079         */
1080        if (a_bits < s_bits) {
1081            unsigned a_mask = (1u << a_bits) - 1;
1082            unsigned s_mask = (1u << s_bits) - 1;
1083            tcg_out_addi(s, addr_type, TCG_REG_TMP1, addr_reg, s_mask - a_mask);
1084        } else {
1085            tcg_out_mov(s, addr_type, TCG_REG_TMP1, addr_reg);
1086        }
1087        tcg_out_opc_bstrins_d(s, TCG_REG_TMP1, TCG_REG_ZERO,
1088                              a_bits, s->page_bits - 1);
1089
1090        /* Compare masked address with the TLB entry.  */
1091        ldst->label_ptr[0] = s->code_ptr;
1092        tcg_out_opc_bne(s, TCG_REG_TMP0, TCG_REG_TMP1, 0);
1093
1094        h->index = TCG_REG_TMP2;
1095    } else {
1096        if (a_bits) {
1097            ldst = new_ldst_label(s);
1098
1099            ldst->is_ld = is_ld;
1100            ldst->oi = oi;
1101            ldst->addr_reg = addr_reg;
1102
1103            /*
1104             * Without micro-architecture details, we don't know which of
1105             * bstrpick or andi is faster, so use bstrpick as it's not
1106             * constrained by imm field width. Not to say alignments >= 2^12
1107             * are going to happen any time soon.
1108             */
1109            tcg_out_opc_bstrpick_d(s, TCG_REG_TMP1, addr_reg, 0, a_bits - 1);
1110
1111            ldst->label_ptr[0] = s->code_ptr;
1112            tcg_out_opc_bne(s, TCG_REG_TMP1, TCG_REG_ZERO, 0);
1113        }
1114
1115        h->index = guest_base ? TCG_GUEST_BASE_REG : TCG_REG_ZERO;
1116    }
1117
1118    if (addr_type == TCG_TYPE_I32) {
1119        h->base = TCG_REG_TMP0;
1120        tcg_out_ext32u(s, h->base, addr_reg);
1121    } else {
1122        h->base = addr_reg;
1123    }
1124
1125    return ldst;
1126}
1127
1128static void tcg_out_qemu_ld_indexed(TCGContext *s, MemOp opc, TCGType type,
1129                                    TCGReg rd, HostAddress h)
1130{
1131    /* Byte swapping is left to middle-end expansion.  */
1132    tcg_debug_assert((opc & MO_BSWAP) == 0);
1133
1134    switch (opc & MO_SSIZE) {
1135    case MO_UB:
1136        tcg_out_opc_ldx_bu(s, rd, h.base, h.index);
1137        break;
1138    case MO_SB:
1139        tcg_out_opc_ldx_b(s, rd, h.base, h.index);
1140        break;
1141    case MO_UW:
1142        tcg_out_opc_ldx_hu(s, rd, h.base, h.index);
1143        break;
1144    case MO_SW:
1145        tcg_out_opc_ldx_h(s, rd, h.base, h.index);
1146        break;
1147    case MO_UL:
1148        if (type == TCG_TYPE_I64) {
1149            tcg_out_opc_ldx_wu(s, rd, h.base, h.index);
1150            break;
1151        }
1152        /* fallthrough */
1153    case MO_SL:
1154        tcg_out_opc_ldx_w(s, rd, h.base, h.index);
1155        break;
1156    case MO_UQ:
1157        tcg_out_opc_ldx_d(s, rd, h.base, h.index);
1158        break;
1159    default:
1160        g_assert_not_reached();
1161    }
1162}
1163
1164static void tcg_out_qemu_ld(TCGContext *s, TCGReg data_reg, TCGReg addr_reg,
1165                            MemOpIdx oi, TCGType data_type)
1166{
1167    TCGLabelQemuLdst *ldst;
1168    HostAddress h;
1169
1170    ldst = prepare_host_addr(s, &h, addr_reg, oi, true);
1171    tcg_out_qemu_ld_indexed(s, get_memop(oi), data_type, data_reg, h);
1172
1173    if (ldst) {
1174        ldst->type = data_type;
1175        ldst->datalo_reg = data_reg;
1176        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1177    }
1178}
1179
1180static void tcg_out_qemu_st_indexed(TCGContext *s, MemOp opc,
1181                                    TCGReg rd, HostAddress h)
1182{
1183    /* Byte swapping is left to middle-end expansion.  */
1184    tcg_debug_assert((opc & MO_BSWAP) == 0);
1185
1186    switch (opc & MO_SIZE) {
1187    case MO_8:
1188        tcg_out_opc_stx_b(s, rd, h.base, h.index);
1189        break;
1190    case MO_16:
1191        tcg_out_opc_stx_h(s, rd, h.base, h.index);
1192        break;
1193    case MO_32:
1194        tcg_out_opc_stx_w(s, rd, h.base, h.index);
1195        break;
1196    case MO_64:
1197        tcg_out_opc_stx_d(s, rd, h.base, h.index);
1198        break;
1199    default:
1200        g_assert_not_reached();
1201    }
1202}
1203
1204static void tcg_out_qemu_st(TCGContext *s, TCGReg data_reg, TCGReg addr_reg,
1205                            MemOpIdx oi, TCGType data_type)
1206{
1207    TCGLabelQemuLdst *ldst;
1208    HostAddress h;
1209
1210    ldst = prepare_host_addr(s, &h, addr_reg, oi, false);
1211    tcg_out_qemu_st_indexed(s, get_memop(oi), data_reg, h);
1212
1213    if (ldst) {
1214        ldst->type = data_type;
1215        ldst->datalo_reg = data_reg;
1216        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1217    }
1218}
1219
1220static void tcg_out_qemu_ldst_i128(TCGContext *s, TCGReg data_lo, TCGReg data_hi,
1221                                   TCGReg addr_reg, MemOpIdx oi, bool is_ld)
1222{
1223    TCGLabelQemuLdst *ldst;
1224    HostAddress h;
1225
1226    ldst = prepare_host_addr(s, &h, addr_reg, oi, is_ld);
1227
1228    if (h.aa.atom == MO_128) {
1229        /*
1230         * Use VLDX/VSTX when 128-bit atomicity is required.
1231         * If address is aligned to 16-bytes, the 128-bit load/store is atomic.
1232         */
1233        if (is_ld) {
1234            tcg_out_opc_vldx(s, TCG_VEC_TMP0, h.base, h.index);
1235            tcg_out_opc_vpickve2gr_d(s, data_lo, TCG_VEC_TMP0, 0);
1236            tcg_out_opc_vpickve2gr_d(s, data_hi, TCG_VEC_TMP0, 1);
1237        } else {
1238            tcg_out_opc_vinsgr2vr_d(s, TCG_VEC_TMP0, data_lo, 0);
1239            tcg_out_opc_vinsgr2vr_d(s, TCG_VEC_TMP0, data_hi, 1);
1240            tcg_out_opc_vstx(s, TCG_VEC_TMP0, h.base, h.index);
1241        }
1242    } else {
1243        /* Otherwise use a pair of LD/ST. */
1244        TCGReg base = h.base;
1245        if (h.index != TCG_REG_ZERO) {
1246            base = TCG_REG_TMP0;
1247            tcg_out_opc_add_d(s, base, h.base, h.index);
1248        }
1249        if (is_ld) {
1250            tcg_debug_assert(base != data_lo);
1251            tcg_out_opc_ld_d(s, data_lo, base, 0);
1252            tcg_out_opc_ld_d(s, data_hi, base, 8);
1253        } else {
1254            tcg_out_opc_st_d(s, data_lo, base, 0);
1255            tcg_out_opc_st_d(s, data_hi, base, 8);
1256        }
1257    }
1258
1259    if (ldst) {
1260        ldst->type = TCG_TYPE_I128;
1261        ldst->datalo_reg = data_lo;
1262        ldst->datahi_reg = data_hi;
1263        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1264    }
1265}
1266
1267/*
1268 * Entry-points
1269 */
1270
1271static const tcg_insn_unit *tb_ret_addr;
1272
1273static void tcg_out_exit_tb(TCGContext *s, uintptr_t a0)
1274{
1275    /* Reuse the zeroing that exists for goto_ptr.  */
1276    if (a0 == 0) {
1277        tcg_out_call_int(s, tcg_code_gen_epilogue, true);
1278    } else {
1279        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_A0, a0);
1280        tcg_out_call_int(s, tb_ret_addr, true);
1281    }
1282}
1283
1284static void tcg_out_goto_tb(TCGContext *s, int which)
1285{
1286    /*
1287     * Direct branch, or load indirect address, to be patched
1288     * by tb_target_set_jmp_target.  Check indirect load offset
1289     * in range early, regardless of direct branch distance,
1290     * via assert within tcg_out_opc_pcaddu2i.
1291     */
1292    uintptr_t i_addr = get_jmp_target_addr(s, which);
1293    intptr_t i_disp = tcg_pcrel_diff(s, (void *)i_addr);
1294
1295    set_jmp_insn_offset(s, which);
1296    tcg_out_opc_pcaddu2i(s, TCG_REG_TMP0, i_disp >> 2);
1297
1298    /* Finish the load and indirect branch. */
1299    tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP0, TCG_REG_TMP0, 0);
1300    tcg_out_opc_jirl(s, TCG_REG_ZERO, TCG_REG_TMP0, 0);
1301    set_jmp_reset_offset(s, which);
1302}
1303
1304void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
1305                              uintptr_t jmp_rx, uintptr_t jmp_rw)
1306{
1307    uintptr_t d_addr = tb->jmp_target_addr[n];
1308    ptrdiff_t d_disp = (ptrdiff_t)(d_addr - jmp_rx) >> 2;
1309    tcg_insn_unit insn;
1310
1311    /* Either directly branch, or load slot address for indirect branch. */
1312    if (d_disp == sextreg(d_disp, 0, 26)) {
1313        insn = encode_sd10k16_insn(OPC_B, d_disp);
1314    } else {
1315        uintptr_t i_addr = (uintptr_t)&tb->jmp_target_addr[n];
1316        intptr_t i_disp = i_addr - jmp_rx;
1317        insn = encode_dsj20_insn(OPC_PCADDU2I, TCG_REG_TMP0, i_disp >> 2);
1318    }
1319
1320    qatomic_set((tcg_insn_unit *)jmp_rw, insn);
1321    flush_idcache_range(jmp_rx, jmp_rw, 4);
1322}
1323
1324
1325static void tgen_add(TCGContext *s, TCGType type,
1326                     TCGReg a0, TCGReg a1, TCGReg a2)
1327{
1328    if (type == TCG_TYPE_I32) {
1329        tcg_out_opc_add_w(s, a0, a1, a2);
1330    } else {
1331        tcg_out_opc_add_d(s, a0, a1, a2);
1332    }
1333}
1334
1335static const TCGOutOpBinary outop_add = {
1336    .base.static_constraint = C_O1_I2(r, r, rJ),
1337    .out_rrr = tgen_add,
1338    .out_rri = tcg_out_addi,
1339};
1340
1341static void tgen_and(TCGContext *s, TCGType type,
1342                     TCGReg a0, TCGReg a1, TCGReg a2)
1343{
1344    tcg_out_opc_and(s, a0, a1, a2);
1345}
1346
1347static void tgen_andi(TCGContext *s, TCGType type,
1348                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1349{
1350    tcg_out_opc_andi(s, a0, a1, a2);
1351}
1352
1353static const TCGOutOpBinary outop_and = {
1354    .base.static_constraint = C_O1_I2(r, r, rU),
1355    .out_rrr = tgen_and,
1356    .out_rri = tgen_andi,
1357};
1358
1359static void tgen_andc(TCGContext *s, TCGType type,
1360                      TCGReg a0, TCGReg a1, TCGReg a2)
1361{
1362    tcg_out_opc_andn(s, a0, a1, a2);
1363}
1364
1365static const TCGOutOpBinary outop_andc = {
1366    .base.static_constraint = C_O1_I2(r, r, r),
1367    .out_rrr = tgen_andc,
1368};
1369
1370static void tgen_clzi(TCGContext *s, TCGType type,
1371                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1372{
1373    /* a2 is constrained to exactly the type width. */
1374    if (type == TCG_TYPE_I32) {
1375        tcg_out_opc_clz_w(s, a0, a1);
1376    } else {
1377        tcg_out_opc_clz_d(s, a0, a1);
1378    }
1379}
1380
1381static void tgen_clz(TCGContext *s, TCGType type,
1382                     TCGReg a0, TCGReg a1, TCGReg a2)
1383{
1384    tgen_clzi(s, type, TCG_REG_TMP0, a1, /* ignored */ 0);
1385    /* a0 = a1 ? REG_TMP0 : a2 */
1386    tcg_out_opc_maskeqz(s, TCG_REG_TMP0, TCG_REG_TMP0, a1);
1387    tcg_out_opc_masknez(s, a0, a2, a1);
1388    tcg_out_opc_or(s, a0, a0, TCG_REG_TMP0);
1389}
1390
1391static const TCGOutOpBinary outop_clz = {
1392    .base.static_constraint = C_O1_I2(r, r, rW),
1393    .out_rrr = tgen_clz,
1394    .out_rri = tgen_clzi,
1395};
1396
1397static const TCGOutOpUnary outop_ctpop = {
1398    .base.static_constraint = C_NotImplemented,
1399};
1400
1401static void tgen_ctzi(TCGContext *s, TCGType type,
1402                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1403{
1404    /* a2 is constrained to exactly the type width. */
1405    if (type == TCG_TYPE_I32) {
1406        tcg_out_opc_ctz_w(s, a0, a1);
1407    } else {
1408        tcg_out_opc_ctz_d(s, a0, a1);
1409    }
1410}
1411
1412static void tgen_ctz(TCGContext *s, TCGType type,
1413                     TCGReg a0, TCGReg a1, TCGReg a2)
1414{
1415    tgen_ctzi(s, type, TCG_REG_TMP0, a1, /* ignored */ 0);
1416    /* a0 = a1 ? REG_TMP0 : a2 */
1417    tcg_out_opc_maskeqz(s, TCG_REG_TMP0, TCG_REG_TMP0, a1);
1418    tcg_out_opc_masknez(s, a0, a2, a1);
1419    tcg_out_opc_or(s, a0, a0, TCG_REG_TMP0);
1420}
1421
1422static const TCGOutOpBinary outop_ctz = {
1423    .base.static_constraint = C_O1_I2(r, r, rW),
1424    .out_rrr = tgen_ctz,
1425    .out_rri = tgen_ctzi,
1426};
1427
1428static void tgen_divs(TCGContext *s, TCGType type,
1429                      TCGReg a0, TCGReg a1, TCGReg a2)
1430{
1431    if (type == TCG_TYPE_I32) {
1432        tcg_out_opc_div_w(s, a0, a1, a2);
1433    } else {
1434        tcg_out_opc_div_d(s, a0, a1, a2);
1435    }
1436}
1437
1438static const TCGOutOpBinary outop_divs = {
1439    .base.static_constraint = C_O1_I2(r, r, r),
1440    .out_rrr = tgen_divs,
1441};
1442
1443static const TCGOutOpDivRem outop_divs2 = {
1444    .base.static_constraint = C_NotImplemented,
1445};
1446
1447static void tgen_divu(TCGContext *s, TCGType type,
1448                      TCGReg a0, TCGReg a1, TCGReg a2)
1449{
1450    if (type == TCG_TYPE_I32) {
1451        tcg_out_opc_div_wu(s, a0, a1, a2);
1452    } else {
1453        tcg_out_opc_div_du(s, a0, a1, a2);
1454    }
1455}
1456
1457static const TCGOutOpBinary outop_divu = {
1458    .base.static_constraint = C_O1_I2(r, r, r),
1459    .out_rrr = tgen_divu,
1460};
1461
1462static const TCGOutOpDivRem outop_divu2 = {
1463    .base.static_constraint = C_NotImplemented,
1464};
1465
1466static const TCGOutOpBinary outop_eqv = {
1467    .base.static_constraint = C_NotImplemented,
1468};
1469
1470static void tgen_extrh_i64_i32(TCGContext *s, TCGType t, TCGReg a0, TCGReg a1)
1471{
1472    tcg_out_opc_srai_d(s, a0, a1, 32);
1473}
1474
1475static const TCGOutOpUnary outop_extrh_i64_i32 = {
1476    .base.static_constraint = C_O1_I1(r, r),
1477    .out_rr = tgen_extrh_i64_i32,
1478};
1479
1480static void tgen_mul(TCGContext *s, TCGType type,
1481                     TCGReg a0, TCGReg a1, TCGReg a2)
1482{
1483    if (type == TCG_TYPE_I32) {
1484        tcg_out_opc_mul_w(s, a0, a1, a2);
1485    } else {
1486        tcg_out_opc_mul_d(s, a0, a1, a2);
1487    }
1488}
1489
1490static const TCGOutOpBinary outop_mul = {
1491    .base.static_constraint = C_O1_I2(r, r, r),
1492    .out_rrr = tgen_mul,
1493};
1494
1495static const TCGOutOpMul2 outop_muls2 = {
1496    .base.static_constraint = C_NotImplemented,
1497};
1498
1499static void tgen_mulsh(TCGContext *s, TCGType type,
1500                       TCGReg a0, TCGReg a1, TCGReg a2)
1501{
1502    if (type == TCG_TYPE_I32) {
1503        tcg_out_opc_mulh_w(s, a0, a1, a2);
1504    } else {
1505        tcg_out_opc_mulh_d(s, a0, a1, a2);
1506    }
1507}
1508
1509static const TCGOutOpBinary outop_mulsh = {
1510    .base.static_constraint = C_O1_I2(r, r, r),
1511    .out_rrr = tgen_mulsh,
1512};
1513
1514static const TCGOutOpMul2 outop_mulu2 = {
1515    .base.static_constraint = C_NotImplemented,
1516};
1517
1518static void tgen_muluh(TCGContext *s, TCGType type,
1519                       TCGReg a0, TCGReg a1, TCGReg a2)
1520{
1521    if (type == TCG_TYPE_I32) {
1522        tcg_out_opc_mulh_wu(s, a0, a1, a2);
1523    } else {
1524        tcg_out_opc_mulh_du(s, a0, a1, a2);
1525    }
1526}
1527
1528static const TCGOutOpBinary outop_muluh = {
1529    .base.static_constraint = C_O1_I2(r, r, r),
1530    .out_rrr = tgen_muluh,
1531};
1532
1533static const TCGOutOpBinary outop_nand = {
1534    .base.static_constraint = C_NotImplemented,
1535};
1536
1537static void tgen_nor(TCGContext *s, TCGType type,
1538                      TCGReg a0, TCGReg a1, TCGReg a2)
1539{
1540    tcg_out_opc_nor(s, a0, a1, a2);
1541}
1542
1543static const TCGOutOpBinary outop_nor = {
1544    .base.static_constraint = C_O1_I2(r, r, r),
1545    .out_rrr = tgen_nor,
1546};
1547
1548static void tgen_or(TCGContext *s, TCGType type,
1549                    TCGReg a0, TCGReg a1, TCGReg a2)
1550{
1551    tcg_out_opc_or(s, a0, a1, a2);
1552}
1553
1554static void tgen_ori(TCGContext *s, TCGType type,
1555                     TCGReg a0, TCGReg a1, tcg_target_long a2)
1556{
1557    tcg_out_opc_ori(s, a0, a1, a2);
1558}
1559
1560static const TCGOutOpBinary outop_or = {
1561    .base.static_constraint = C_O1_I2(r, r, rU),
1562    .out_rrr = tgen_or,
1563    .out_rri = tgen_ori,
1564};
1565
1566static void tgen_orc(TCGContext *s, TCGType type,
1567                     TCGReg a0, TCGReg a1, TCGReg a2)
1568{
1569    tcg_out_opc_orn(s, a0, a1, a2);
1570}
1571
1572static const TCGOutOpBinary outop_orc = {
1573    .base.static_constraint = C_O1_I2(r, r, r),
1574    .out_rrr = tgen_orc,
1575};
1576
1577static void tgen_rems(TCGContext *s, TCGType type,
1578                      TCGReg a0, TCGReg a1, TCGReg a2)
1579{
1580    if (type == TCG_TYPE_I32) {
1581        tcg_out_opc_mod_w(s, a0, a1, a2);
1582    } else {
1583        tcg_out_opc_mod_d(s, a0, a1, a2);
1584    }
1585}
1586
1587static const TCGOutOpBinary outop_rems = {
1588    .base.static_constraint = C_O1_I2(r, r, r),
1589    .out_rrr = tgen_rems,
1590};
1591
1592static void tgen_remu(TCGContext *s, TCGType type,
1593                      TCGReg a0, TCGReg a1, TCGReg a2)
1594{
1595    if (type == TCG_TYPE_I32) {
1596        tcg_out_opc_mod_wu(s, a0, a1, a2);
1597    } else {
1598        tcg_out_opc_mod_du(s, a0, a1, a2);
1599    }
1600}
1601
1602static const TCGOutOpBinary outop_remu = {
1603    .base.static_constraint = C_O1_I2(r, r, r),
1604    .out_rrr = tgen_remu,
1605};
1606
1607static const TCGOutOpBinary outop_rotl = {
1608    .base.static_constraint = C_NotImplemented,
1609};
1610
1611static void tgen_rotr(TCGContext *s, TCGType type,
1612                      TCGReg a0, TCGReg a1, TCGReg a2)
1613{
1614    if (type == TCG_TYPE_I32) {
1615        tcg_out_opc_rotr_w(s, a0, a1, a2);
1616    } else {
1617        tcg_out_opc_rotr_d(s, a0, a1, a2);
1618    }
1619}
1620
1621static void tgen_rotri(TCGContext *s, TCGType type,
1622                       TCGReg a0, TCGReg a1, tcg_target_long a2)
1623{
1624    if (type == TCG_TYPE_I32) {
1625        tcg_out_opc_rotri_w(s, a0, a1, a2 & 0x1f);
1626    } else {
1627        tcg_out_opc_rotri_d(s, a0, a1, a2 & 0x3f);
1628    }
1629}
1630
1631static const TCGOutOpBinary outop_rotr = {
1632    .base.static_constraint = C_O1_I2(r, r, ri),
1633    .out_rrr = tgen_rotr,
1634    .out_rri = tgen_rotri,
1635};
1636
1637static void tgen_sar(TCGContext *s, TCGType type,
1638                     TCGReg a0, TCGReg a1, TCGReg a2)
1639{
1640    if (type == TCG_TYPE_I32) {
1641        tcg_out_opc_sra_w(s, a0, a1, a2);
1642    } else {
1643        tcg_out_opc_sra_d(s, a0, a1, a2);
1644    }
1645}
1646
1647static void tgen_sari(TCGContext *s, TCGType type,
1648                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1649{
1650    if (type == TCG_TYPE_I32) {
1651        tcg_out_opc_srai_w(s, a0, a1, a2 & 0x1f);
1652    } else {
1653        tcg_out_opc_srai_d(s, a0, a1, a2 & 0x3f);
1654    }
1655}
1656
1657static const TCGOutOpBinary outop_sar = {
1658    .base.static_constraint = C_O1_I2(r, r, ri),
1659    .out_rrr = tgen_sar,
1660    .out_rri = tgen_sari,
1661};
1662
1663static void tgen_shl(TCGContext *s, TCGType type,
1664                     TCGReg a0, TCGReg a1, TCGReg a2)
1665{
1666    if (type == TCG_TYPE_I32) {
1667        tcg_out_opc_sll_w(s, a0, a1, a2);
1668    } else {
1669        tcg_out_opc_sll_d(s, a0, a1, a2);
1670    }
1671}
1672
1673static void tgen_shli(TCGContext *s, TCGType type,
1674                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1675{
1676    if (type == TCG_TYPE_I32) {
1677        tcg_out_opc_slli_w(s, a0, a1, a2 & 0x1f);
1678    } else {
1679        tcg_out_opc_slli_d(s, a0, a1, a2 & 0x3f);
1680    }
1681}
1682
1683static const TCGOutOpBinary outop_shl = {
1684    .base.static_constraint = C_O1_I2(r, r, ri),
1685    .out_rrr = tgen_shl,
1686    .out_rri = tgen_shli,
1687};
1688
1689static void tgen_shr(TCGContext *s, TCGType type,
1690                     TCGReg a0, TCGReg a1, TCGReg a2)
1691{
1692    if (type == TCG_TYPE_I32) {
1693        tcg_out_opc_srl_w(s, a0, a1, a2);
1694    } else {
1695        tcg_out_opc_srl_d(s, a0, a1, a2);
1696    }
1697}
1698
1699static void tgen_shri(TCGContext *s, TCGType type,
1700                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1701{
1702    if (type == TCG_TYPE_I32) {
1703        tcg_out_opc_srli_w(s, a0, a1, a2 & 0x1f);
1704    } else {
1705        tcg_out_opc_srli_d(s, a0, a1, a2 & 0x3f);
1706    }
1707}
1708
1709static const TCGOutOpBinary outop_shr = {
1710    .base.static_constraint = C_O1_I2(r, r, ri),
1711    .out_rrr = tgen_shr,
1712    .out_rri = tgen_shri,
1713};
1714
1715static void tgen_sub(TCGContext *s, TCGType type,
1716                     TCGReg a0, TCGReg a1, TCGReg a2)
1717{
1718    if (type == TCG_TYPE_I32) {
1719        tcg_out_opc_sub_w(s, a0, a1, a2);
1720    } else {
1721        tcg_out_opc_sub_d(s, a0, a1, a2);
1722    }
1723}
1724
1725static const TCGOutOpSubtract outop_sub = {
1726    .base.static_constraint = C_O1_I2(r, r, r),
1727    .out_rrr = tgen_sub,
1728};
1729
1730static void tgen_xor(TCGContext *s, TCGType type,
1731                     TCGReg a0, TCGReg a1, TCGReg a2)
1732{
1733    tcg_out_opc_xor(s, a0, a1, a2);
1734}
1735
1736static void tgen_xori(TCGContext *s, TCGType type,
1737                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1738{
1739    tcg_out_opc_xori(s, a0, a1, a2);
1740}
1741
1742static const TCGOutOpBinary outop_xor = {
1743    .base.static_constraint = C_O1_I2(r, r, rU),
1744    .out_rrr = tgen_xor,
1745    .out_rri = tgen_xori,
1746};
1747
1748static void tgen_bswap16(TCGContext *s, TCGType type,
1749                         TCGReg a0, TCGReg a1, unsigned flags)
1750{
1751    tcg_out_opc_revb_2h(s, a0, a1);
1752    if (flags & TCG_BSWAP_OS) {
1753        tcg_out_ext16s(s, TCG_TYPE_REG, a0, a0);
1754    } else if ((flags & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1755        tcg_out_ext16u(s, a0, a0);
1756    }
1757}
1758
1759static const TCGOutOpBswap outop_bswap16 = {
1760    .base.static_constraint = C_O1_I1(r, r),
1761    .out_rr = tgen_bswap16,
1762};
1763
1764static void tgen_bswap32(TCGContext *s, TCGType type,
1765                         TCGReg a0, TCGReg a1, unsigned flags)
1766{
1767    tcg_out_opc_revb_2w(s, a0, a1);
1768
1769    /* All 32-bit values are computed sign-extended in the register.  */
1770    if (type == TCG_TYPE_I32 || (flags & TCG_BSWAP_OS)) {
1771        tcg_out_ext32s(s, a0, a0);
1772    } else if ((flags & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1773        tcg_out_ext32u(s, a0, a0);
1774    }
1775}
1776
1777static const TCGOutOpBswap outop_bswap32 = {
1778    .base.static_constraint = C_O1_I1(r, r),
1779    .out_rr = tgen_bswap32,
1780};
1781
1782static void tgen_bswap64(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
1783{
1784    tcg_out_opc_revb_d(s, a0, a1);
1785}
1786
1787static const TCGOutOpUnary outop_bswap64 = {
1788    .base.static_constraint = C_O1_I1(r, r),
1789    .out_rr = tgen_bswap64,
1790};
1791
1792static void tgen_neg(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
1793{
1794    tgen_sub(s, type, a0, TCG_REG_ZERO, a1);
1795}
1796
1797static const TCGOutOpUnary outop_neg = {
1798    .base.static_constraint = C_O1_I1(r, r),
1799    .out_rr = tgen_neg,
1800};
1801
1802static void tgen_not(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
1803{
1804    tgen_nor(s, type, a0, a1, TCG_REG_ZERO);
1805}
1806
1807static const TCGOutOpUnary outop_not = {
1808    .base.static_constraint = C_O1_I1(r, r),
1809    .out_rr = tgen_not,
1810};
1811
1812static void tgen_deposit(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1,
1813                         TCGReg a2, unsigned ofs, unsigned len)
1814{
1815    if (type == TCG_TYPE_I32) {
1816        tcg_out_opc_bstrins_w(s, a0, a2, ofs, ofs + len - 1);
1817    } else {
1818        tcg_out_opc_bstrins_d(s, a0, a2, ofs, ofs + len - 1);
1819    }
1820}
1821
1822static const TCGOutOpDeposit outop_deposit = {
1823    .base.static_constraint = C_O1_I2(r, 0, rz),
1824    .out_rrr = tgen_deposit,
1825};
1826
1827static void tgen_extract(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1,
1828                         unsigned ofs, unsigned len)
1829{
1830    if (ofs == 0 && len <= 12) {
1831        tcg_out_opc_andi(s, a0, a1, (1 << len) - 1);
1832    } else if (type == TCG_TYPE_I32) {
1833        tcg_out_opc_bstrpick_w(s, a0, a1, ofs, ofs + len - 1);
1834    } else {
1835        tcg_out_opc_bstrpick_d(s, a0, a1, ofs, ofs + len - 1);
1836    }
1837}
1838
1839static const TCGOutOpExtract outop_extract = {
1840    .base.static_constraint = C_O1_I1(r, r),
1841    .out_rr = tgen_extract,
1842};
1843
1844static void tgen_sextract(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1,
1845                          unsigned ofs, unsigned len)
1846{
1847    if (ofs == 0) {
1848        switch (len) {
1849        case 8:
1850            tcg_out_ext8s(s, type, a0, a1);
1851            return;
1852        case 16:
1853            tcg_out_ext16s(s, type, a0, a1);
1854            return;
1855        case 32:
1856            tcg_out_ext32s(s, a0, a1);
1857            return;
1858        }
1859    } else if (ofs + len == 32) {
1860        tcg_out_opc_srai_w(s, a0, a1, ofs);
1861        return;
1862    }
1863    g_assert_not_reached();
1864}
1865
1866static const TCGOutOpExtract outop_sextract = {
1867    .base.static_constraint = C_O1_I1(r, r),
1868    .out_rr = tgen_sextract,
1869};
1870
1871static void tcg_out_op(TCGContext *s, TCGOpcode opc, TCGType type,
1872                       const TCGArg args[TCG_MAX_OP_ARGS],
1873                       const int const_args[TCG_MAX_OP_ARGS])
1874{
1875    TCGArg a0 = args[0];
1876    TCGArg a1 = args[1];
1877    TCGArg a2 = args[2];
1878    TCGArg a3 = args[3];
1879
1880    switch (opc) {
1881    case INDEX_op_mb:
1882        tcg_out_mb(s, a0);
1883        break;
1884
1885    case INDEX_op_goto_ptr:
1886        tcg_out_opc_jirl(s, TCG_REG_ZERO, a0, 0);
1887        break;
1888
1889    case INDEX_op_br:
1890        tcg_out_reloc(s, s->code_ptr, R_LOONGARCH_BR_SD10K16, arg_label(a0),
1891                      0);
1892        tcg_out_opc_b(s, 0);
1893        break;
1894
1895    case INDEX_op_ld8s_i32:
1896    case INDEX_op_ld8s_i64:
1897        tcg_out_ldst(s, OPC_LD_B, a0, a1, a2);
1898        break;
1899    case INDEX_op_ld8u_i32:
1900    case INDEX_op_ld8u_i64:
1901        tcg_out_ldst(s, OPC_LD_BU, a0, a1, a2);
1902        break;
1903    case INDEX_op_ld16s_i32:
1904    case INDEX_op_ld16s_i64:
1905        tcg_out_ldst(s, OPC_LD_H, a0, a1, a2);
1906        break;
1907    case INDEX_op_ld16u_i32:
1908    case INDEX_op_ld16u_i64:
1909        tcg_out_ldst(s, OPC_LD_HU, a0, a1, a2);
1910        break;
1911    case INDEX_op_ld_i32:
1912    case INDEX_op_ld32s_i64:
1913        tcg_out_ldst(s, OPC_LD_W, a0, a1, a2);
1914        break;
1915    case INDEX_op_ld32u_i64:
1916        tcg_out_ldst(s, OPC_LD_WU, a0, a1, a2);
1917        break;
1918    case INDEX_op_ld_i64:
1919        tcg_out_ldst(s, OPC_LD_D, a0, a1, a2);
1920        break;
1921
1922    case INDEX_op_st8_i32:
1923    case INDEX_op_st8_i64:
1924        tcg_out_ldst(s, OPC_ST_B, a0, a1, a2);
1925        break;
1926    case INDEX_op_st16_i32:
1927    case INDEX_op_st16_i64:
1928        tcg_out_ldst(s, OPC_ST_H, a0, a1, a2);
1929        break;
1930    case INDEX_op_st_i32:
1931    case INDEX_op_st32_i64:
1932        tcg_out_ldst(s, OPC_ST_W, a0, a1, a2);
1933        break;
1934    case INDEX_op_st_i64:
1935        tcg_out_ldst(s, OPC_ST_D, a0, a1, a2);
1936        break;
1937
1938    case INDEX_op_qemu_ld_i32:
1939        tcg_out_qemu_ld(s, a0, a1, a2, TCG_TYPE_I32);
1940        break;
1941    case INDEX_op_qemu_ld_i64:
1942        tcg_out_qemu_ld(s, a0, a1, a2, TCG_TYPE_I64);
1943        break;
1944    case INDEX_op_qemu_ld_i128:
1945        tcg_out_qemu_ldst_i128(s, a0, a1, a2, a3, true);
1946        break;
1947    case INDEX_op_qemu_st_i32:
1948        tcg_out_qemu_st(s, a0, a1, a2, TCG_TYPE_I32);
1949        break;
1950    case INDEX_op_qemu_st_i64:
1951        tcg_out_qemu_st(s, a0, a1, a2, TCG_TYPE_I64);
1952        break;
1953    case INDEX_op_qemu_st_i128:
1954        tcg_out_qemu_ldst_i128(s, a0, a1, a2, a3, false);
1955        break;
1956
1957    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
1958    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
1959    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
1960    default:
1961        g_assert_not_reached();
1962    }
1963}
1964
1965static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
1966                            TCGReg rd, TCGReg rs)
1967{
1968    static const LoongArchInsn repl_insn[2][4] = {
1969        { OPC_VREPLGR2VR_B, OPC_VREPLGR2VR_H,
1970          OPC_VREPLGR2VR_W, OPC_VREPLGR2VR_D },
1971        { OPC_XVREPLGR2VR_B, OPC_XVREPLGR2VR_H,
1972          OPC_XVREPLGR2VR_W, OPC_XVREPLGR2VR_D },
1973    };
1974    bool lasx = type == TCG_TYPE_V256;
1975
1976    tcg_debug_assert(vece <= MO_64);
1977    tcg_out32(s, encode_vdj_insn(repl_insn[lasx][vece], rd, rs));
1978    return true;
1979}
1980
1981static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
1982                             TCGReg r, TCGReg base, intptr_t offset)
1983{
1984    bool lasx = type == TCG_TYPE_V256;
1985
1986    /* Handle imm overflow and division (vldrepl.d imm is divided by 8). */
1987    if (offset < -0x800 || offset > 0x7ff ||
1988        (offset & ((1 << vece) - 1)) != 0) {
1989        tcg_out_addi(s, TCG_TYPE_I64, TCG_REG_TMP0, base, offset);
1990        base = TCG_REG_TMP0;
1991        offset = 0;
1992    }
1993    offset >>= vece;
1994
1995    switch (vece) {
1996    case MO_8:
1997        if (lasx) {
1998            tcg_out_opc_xvldrepl_b(s, r, base, offset);
1999        } else {
2000            tcg_out_opc_vldrepl_b(s, r, base, offset);
2001        }
2002        break;
2003    case MO_16:
2004        if (lasx) {
2005            tcg_out_opc_xvldrepl_h(s, r, base, offset);
2006        } else {
2007            tcg_out_opc_vldrepl_h(s, r, base, offset);
2008        }
2009        break;
2010    case MO_32:
2011        if (lasx) {
2012            tcg_out_opc_xvldrepl_w(s, r, base, offset);
2013        } else {
2014            tcg_out_opc_vldrepl_w(s, r, base, offset);
2015        }
2016        break;
2017    case MO_64:
2018        if (lasx) {
2019            tcg_out_opc_xvldrepl_d(s, r, base, offset);
2020        } else {
2021            tcg_out_opc_vldrepl_d(s, r, base, offset);
2022        }
2023        break;
2024    default:
2025        g_assert_not_reached();
2026    }
2027    return true;
2028}
2029
2030static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
2031                             TCGReg rd, int64_t v64)
2032{
2033    /* Try vldi if imm can fit */
2034    int64_t value = sextract64(v64, 0, 8 << vece);
2035    if (-0x200 <= value && value <= 0x1FF) {
2036        uint32_t imm = (vece << 10) | ((uint32_t)v64 & 0x3FF);
2037
2038        if (type == TCG_TYPE_V256) {
2039            tcg_out_opc_xvldi(s, rd, imm);
2040        } else {
2041            tcg_out_opc_vldi(s, rd, imm);
2042        }
2043        return;
2044    }
2045
2046    /* TODO: vldi patterns when imm 12 is set */
2047
2048    tcg_out_movi(s, TCG_TYPE_I64, TCG_REG_TMP0, value);
2049    tcg_out_dup_vec(s, type, vece, rd, TCG_REG_TMP0);
2050}
2051
2052static void tcg_out_addsub_vec(TCGContext *s, bool lasx, unsigned vece,
2053                               TCGArg a0, TCGArg a1, TCGArg a2,
2054                               bool a2_is_const, bool is_add)
2055{
2056    static const LoongArchInsn add_vec_insn[2][4] = {
2057        { OPC_VADD_B, OPC_VADD_H, OPC_VADD_W, OPC_VADD_D },
2058        { OPC_XVADD_B, OPC_XVADD_H, OPC_XVADD_W, OPC_XVADD_D },
2059    };
2060    static const LoongArchInsn add_vec_imm_insn[2][4] = {
2061        { OPC_VADDI_BU, OPC_VADDI_HU, OPC_VADDI_WU, OPC_VADDI_DU },
2062        { OPC_XVADDI_BU, OPC_XVADDI_HU, OPC_XVADDI_WU, OPC_XVADDI_DU },
2063    };
2064    static const LoongArchInsn sub_vec_insn[2][4] = {
2065        { OPC_VSUB_B, OPC_VSUB_H, OPC_VSUB_W, OPC_VSUB_D },
2066        { OPC_XVSUB_B, OPC_XVSUB_H, OPC_XVSUB_W, OPC_XVSUB_D },
2067    };
2068    static const LoongArchInsn sub_vec_imm_insn[2][4] = {
2069        { OPC_VSUBI_BU, OPC_VSUBI_HU, OPC_VSUBI_WU, OPC_VSUBI_DU },
2070        { OPC_XVSUBI_BU, OPC_XVSUBI_HU, OPC_XVSUBI_WU, OPC_XVSUBI_DU },
2071    };
2072    LoongArchInsn insn;
2073
2074    if (a2_is_const) {
2075        int64_t value = sextract64(a2, 0, 8 << vece);
2076
2077        if (!is_add) {
2078            value = -value;
2079        }
2080        if (value < 0) {
2081            insn = sub_vec_imm_insn[lasx][vece];
2082            value = -value;
2083        } else {
2084            insn = add_vec_imm_insn[lasx][vece];
2085        }
2086
2087        /* Constraint TCG_CT_CONST_VADD ensures validity. */
2088        tcg_debug_assert(0 <= value && value <= 0x1f);
2089
2090        tcg_out32(s, encode_vdvjuk5_insn(insn, a0, a1, value));
2091        return;
2092    }
2093
2094    if (is_add) {
2095        insn = add_vec_insn[lasx][vece];
2096    } else {
2097        insn = sub_vec_insn[lasx][vece];
2098    }
2099    tcg_out32(s, encode_vdvjvk_insn(insn, a0, a1, a2));
2100}
2101
2102static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
2103                           unsigned vecl, unsigned vece,
2104                           const TCGArg args[TCG_MAX_OP_ARGS],
2105                           const int const_args[TCG_MAX_OP_ARGS])
2106{
2107    TCGType type = vecl + TCG_TYPE_V64;
2108    bool lasx = type == TCG_TYPE_V256;
2109    TCGArg a0, a1, a2, a3;
2110    LoongArchInsn insn;
2111
2112    static const LoongArchInsn cmp_vec_insn[16][2][4] = {
2113        [TCG_COND_EQ] = {
2114            { OPC_VSEQ_B, OPC_VSEQ_H, OPC_VSEQ_W, OPC_VSEQ_D },
2115            { OPC_XVSEQ_B, OPC_XVSEQ_H, OPC_XVSEQ_W, OPC_XVSEQ_D },
2116        },
2117        [TCG_COND_LE] = {
2118            { OPC_VSLE_B, OPC_VSLE_H, OPC_VSLE_W, OPC_VSLE_D },
2119            { OPC_XVSLE_B, OPC_XVSLE_H, OPC_XVSLE_W, OPC_XVSLE_D },
2120        },
2121        [TCG_COND_LEU] = {
2122            { OPC_VSLE_BU, OPC_VSLE_HU, OPC_VSLE_WU, OPC_VSLE_DU },
2123            { OPC_XVSLE_BU, OPC_XVSLE_HU, OPC_XVSLE_WU, OPC_XVSLE_DU },
2124        },
2125        [TCG_COND_LT] = {
2126            { OPC_VSLT_B, OPC_VSLT_H, OPC_VSLT_W, OPC_VSLT_D },
2127            { OPC_XVSLT_B, OPC_XVSLT_H, OPC_XVSLT_W, OPC_XVSLT_D },
2128        },
2129        [TCG_COND_LTU] = {
2130            { OPC_VSLT_BU, OPC_VSLT_HU, OPC_VSLT_WU, OPC_VSLT_DU },
2131            { OPC_XVSLT_BU, OPC_XVSLT_HU, OPC_XVSLT_WU, OPC_XVSLT_DU },
2132        }
2133    };
2134    static const LoongArchInsn cmp_vec_imm_insn[16][2][4] = {
2135        [TCG_COND_EQ] = {
2136            { OPC_VSEQI_B, OPC_VSEQI_H, OPC_VSEQI_W, OPC_VSEQI_D },
2137            { OPC_XVSEQI_B, OPC_XVSEQI_H, OPC_XVSEQI_W, OPC_XVSEQI_D },
2138        },
2139        [TCG_COND_LE] = {
2140            { OPC_VSLEI_B, OPC_VSLEI_H, OPC_VSLEI_W, OPC_VSLEI_D },
2141            { OPC_XVSLEI_B, OPC_XVSLEI_H, OPC_XVSLEI_W, OPC_XVSLEI_D },
2142        },
2143        [TCG_COND_LEU] = {
2144            { OPC_VSLEI_BU, OPC_VSLEI_HU, OPC_VSLEI_WU, OPC_VSLEI_DU },
2145            { OPC_XVSLEI_BU, OPC_XVSLEI_HU, OPC_XVSLEI_WU, OPC_XVSLEI_DU },
2146        },
2147        [TCG_COND_LT] = {
2148            { OPC_VSLTI_B, OPC_VSLTI_H, OPC_VSLTI_W, OPC_VSLTI_D },
2149            { OPC_XVSLTI_B, OPC_XVSLTI_H, OPC_XVSLTI_W, OPC_XVSLTI_D },
2150        },
2151        [TCG_COND_LTU] = {
2152            { OPC_VSLTI_BU, OPC_VSLTI_HU, OPC_VSLTI_WU, OPC_VSLTI_DU },
2153            { OPC_XVSLTI_BU, OPC_XVSLTI_HU, OPC_XVSLTI_WU, OPC_XVSLTI_DU },
2154        }
2155    };
2156    static const LoongArchInsn neg_vec_insn[2][4] = {
2157        { OPC_VNEG_B, OPC_VNEG_H, OPC_VNEG_W, OPC_VNEG_D },
2158        { OPC_XVNEG_B, OPC_XVNEG_H, OPC_XVNEG_W, OPC_XVNEG_D },
2159    };
2160    static const LoongArchInsn mul_vec_insn[2][4] = {
2161        { OPC_VMUL_B, OPC_VMUL_H, OPC_VMUL_W, OPC_VMUL_D },
2162        { OPC_XVMUL_B, OPC_XVMUL_H, OPC_XVMUL_W, OPC_XVMUL_D },
2163    };
2164    static const LoongArchInsn smin_vec_insn[2][4] = {
2165        { OPC_VMIN_B, OPC_VMIN_H, OPC_VMIN_W, OPC_VMIN_D },
2166        { OPC_XVMIN_B, OPC_XVMIN_H, OPC_XVMIN_W, OPC_XVMIN_D },
2167    };
2168    static const LoongArchInsn umin_vec_insn[2][4] = {
2169        { OPC_VMIN_BU, OPC_VMIN_HU, OPC_VMIN_WU, OPC_VMIN_DU },
2170        { OPC_XVMIN_BU, OPC_XVMIN_HU, OPC_XVMIN_WU, OPC_XVMIN_DU },
2171    };
2172    static const LoongArchInsn smax_vec_insn[2][4] = {
2173        { OPC_VMAX_B, OPC_VMAX_H, OPC_VMAX_W, OPC_VMAX_D },
2174        { OPC_XVMAX_B, OPC_XVMAX_H, OPC_XVMAX_W, OPC_XVMAX_D },
2175    };
2176    static const LoongArchInsn umax_vec_insn[2][4] = {
2177        { OPC_VMAX_BU, OPC_VMAX_HU, OPC_VMAX_WU, OPC_VMAX_DU },
2178        { OPC_XVMAX_BU, OPC_XVMAX_HU, OPC_XVMAX_WU, OPC_XVMAX_DU },
2179    };
2180    static const LoongArchInsn ssadd_vec_insn[2][4] = {
2181        { OPC_VSADD_B, OPC_VSADD_H, OPC_VSADD_W, OPC_VSADD_D },
2182        { OPC_XVSADD_B, OPC_XVSADD_H, OPC_XVSADD_W, OPC_XVSADD_D },
2183    };
2184    static const LoongArchInsn usadd_vec_insn[2][4] = {
2185        { OPC_VSADD_BU, OPC_VSADD_HU, OPC_VSADD_WU, OPC_VSADD_DU },
2186        { OPC_XVSADD_BU, OPC_XVSADD_HU, OPC_XVSADD_WU, OPC_XVSADD_DU },
2187    };
2188    static const LoongArchInsn sssub_vec_insn[2][4] = {
2189        { OPC_VSSUB_B, OPC_VSSUB_H, OPC_VSSUB_W, OPC_VSSUB_D },
2190        { OPC_XVSSUB_B, OPC_XVSSUB_H, OPC_XVSSUB_W, OPC_XVSSUB_D },
2191    };
2192    static const LoongArchInsn ussub_vec_insn[2][4] = {
2193        { OPC_VSSUB_BU, OPC_VSSUB_HU, OPC_VSSUB_WU, OPC_VSSUB_DU },
2194        { OPC_XVSSUB_BU, OPC_XVSSUB_HU, OPC_XVSSUB_WU, OPC_XVSSUB_DU },
2195    };
2196    static const LoongArchInsn shlv_vec_insn[2][4] = {
2197        { OPC_VSLL_B, OPC_VSLL_H, OPC_VSLL_W, OPC_VSLL_D },
2198        { OPC_XVSLL_B, OPC_XVSLL_H, OPC_XVSLL_W, OPC_XVSLL_D },
2199    };
2200    static const LoongArchInsn shrv_vec_insn[2][4] = {
2201        { OPC_VSRL_B, OPC_VSRL_H, OPC_VSRL_W, OPC_VSRL_D },
2202        { OPC_XVSRL_B, OPC_XVSRL_H, OPC_XVSRL_W, OPC_XVSRL_D },
2203    };
2204    static const LoongArchInsn sarv_vec_insn[2][4] = {
2205        { OPC_VSRA_B, OPC_VSRA_H, OPC_VSRA_W, OPC_VSRA_D },
2206        { OPC_XVSRA_B, OPC_XVSRA_H, OPC_XVSRA_W, OPC_XVSRA_D },
2207    };
2208    static const LoongArchInsn shli_vec_insn[2][4] = {
2209        { OPC_VSLLI_B, OPC_VSLLI_H, OPC_VSLLI_W, OPC_VSLLI_D },
2210        { OPC_XVSLLI_B, OPC_XVSLLI_H, OPC_XVSLLI_W, OPC_XVSLLI_D },
2211    };
2212    static const LoongArchInsn shri_vec_insn[2][4] = {
2213        { OPC_VSRLI_B, OPC_VSRLI_H, OPC_VSRLI_W, OPC_VSRLI_D },
2214        { OPC_XVSRLI_B, OPC_XVSRLI_H, OPC_XVSRLI_W, OPC_XVSRLI_D },
2215    };
2216    static const LoongArchInsn sari_vec_insn[2][4] = {
2217        { OPC_VSRAI_B, OPC_VSRAI_H, OPC_VSRAI_W, OPC_VSRAI_D },
2218        { OPC_XVSRAI_B, OPC_XVSRAI_H, OPC_XVSRAI_W, OPC_XVSRAI_D },
2219    };
2220    static const LoongArchInsn rotrv_vec_insn[2][4] = {
2221        { OPC_VROTR_B, OPC_VROTR_H, OPC_VROTR_W, OPC_VROTR_D },
2222        { OPC_XVROTR_B, OPC_XVROTR_H, OPC_XVROTR_W, OPC_XVROTR_D },
2223    };
2224    static const LoongArchInsn rotri_vec_insn[2][4] = {
2225        { OPC_VROTRI_B, OPC_VROTRI_H, OPC_VROTRI_W, OPC_VROTRI_D },
2226        { OPC_XVROTRI_B, OPC_XVROTRI_H, OPC_XVROTRI_W, OPC_XVROTRI_D },
2227    };
2228
2229    a0 = args[0];
2230    a1 = args[1];
2231    a2 = args[2];
2232    a3 = args[3];
2233
2234    switch (opc) {
2235    case INDEX_op_st_vec:
2236        tcg_out_st(s, type, a0, a1, a2);
2237        break;
2238    case INDEX_op_ld_vec:
2239        tcg_out_ld(s, type, a0, a1, a2);
2240        break;
2241    case INDEX_op_and_vec:
2242        insn = lasx ? OPC_XVAND_V : OPC_VAND_V;
2243        goto vdvjvk;
2244    case INDEX_op_andc_vec:
2245        /*
2246         * vandn vd, vj, vk: vd = vk & ~vj
2247         * andc_vec vd, vj, vk: vd = vj & ~vk
2248         * vj and vk are swapped
2249         */
2250        a1 = a2;
2251        a2 = args[1];
2252        insn = lasx ? OPC_XVANDN_V : OPC_VANDN_V;
2253        goto vdvjvk;
2254    case INDEX_op_or_vec:
2255        insn = lasx ? OPC_XVOR_V : OPC_VOR_V;
2256        goto vdvjvk;
2257    case INDEX_op_orc_vec:
2258        insn = lasx ? OPC_XVORN_V : OPC_VORN_V;
2259        goto vdvjvk;
2260    case INDEX_op_xor_vec:
2261        insn = lasx ? OPC_XVXOR_V : OPC_VXOR_V;
2262        goto vdvjvk;
2263    case INDEX_op_not_vec:
2264        a2 = a1;
2265        /* fall through */
2266    case INDEX_op_nor_vec:
2267        insn = lasx ? OPC_XVNOR_V : OPC_VNOR_V;
2268        goto vdvjvk;
2269    case INDEX_op_cmp_vec:
2270        {
2271            TCGCond cond = args[3];
2272
2273            if (const_args[2]) {
2274                /*
2275                 * cmp_vec dest, src, value
2276                 * Try vseqi/vslei/vslti
2277                 */
2278                int64_t value = sextract64(a2, 0, 8 << vece);
2279                switch (cond) {
2280                case TCG_COND_EQ:
2281                case TCG_COND_LE:
2282                case TCG_COND_LT:
2283                    insn = cmp_vec_imm_insn[cond][lasx][vece];
2284                    tcg_out32(s, encode_vdvjsk5_insn(insn, a0, a1, value));
2285                    break;
2286                case TCG_COND_LEU:
2287                case TCG_COND_LTU:
2288                    insn = cmp_vec_imm_insn[cond][lasx][vece];
2289                    tcg_out32(s, encode_vdvjuk5_insn(insn, a0, a1, value));
2290                    break;
2291                default:
2292                    g_assert_not_reached();
2293                }
2294                break;
2295            }
2296
2297            insn = cmp_vec_insn[cond][lasx][vece];
2298            if (insn == 0) {
2299                TCGArg t;
2300                t = a1, a1 = a2, a2 = t;
2301                cond = tcg_swap_cond(cond);
2302                insn = cmp_vec_insn[cond][lasx][vece];
2303                tcg_debug_assert(insn != 0);
2304            }
2305        }
2306        goto vdvjvk;
2307    case INDEX_op_add_vec:
2308        tcg_out_addsub_vec(s, lasx, vece, a0, a1, a2, const_args[2], true);
2309        break;
2310    case INDEX_op_sub_vec:
2311        tcg_out_addsub_vec(s, lasx, vece, a0, a1, a2, const_args[2], false);
2312        break;
2313    case INDEX_op_neg_vec:
2314        tcg_out32(s, encode_vdvj_insn(neg_vec_insn[lasx][vece], a0, a1));
2315        break;
2316    case INDEX_op_mul_vec:
2317        insn = mul_vec_insn[lasx][vece];
2318        goto vdvjvk;
2319    case INDEX_op_smin_vec:
2320        insn = smin_vec_insn[lasx][vece];
2321        goto vdvjvk;
2322    case INDEX_op_smax_vec:
2323        insn = smax_vec_insn[lasx][vece];
2324        goto vdvjvk;
2325    case INDEX_op_umin_vec:
2326        insn = umin_vec_insn[lasx][vece];
2327        goto vdvjvk;
2328    case INDEX_op_umax_vec:
2329        insn = umax_vec_insn[lasx][vece];
2330        goto vdvjvk;
2331    case INDEX_op_ssadd_vec:
2332        insn = ssadd_vec_insn[lasx][vece];
2333        goto vdvjvk;
2334    case INDEX_op_usadd_vec:
2335        insn = usadd_vec_insn[lasx][vece];
2336        goto vdvjvk;
2337    case INDEX_op_sssub_vec:
2338        insn = sssub_vec_insn[lasx][vece];
2339        goto vdvjvk;
2340    case INDEX_op_ussub_vec:
2341        insn = ussub_vec_insn[lasx][vece];
2342        goto vdvjvk;
2343    case INDEX_op_shlv_vec:
2344        insn = shlv_vec_insn[lasx][vece];
2345        goto vdvjvk;
2346    case INDEX_op_shrv_vec:
2347        insn = shrv_vec_insn[lasx][vece];
2348        goto vdvjvk;
2349    case INDEX_op_sarv_vec:
2350        insn = sarv_vec_insn[lasx][vece];
2351        goto vdvjvk;
2352    case INDEX_op_rotlv_vec:
2353        /* rotlv_vec a1, a2 = rotrv_vec a1, -a2 */
2354        tcg_out32(s, encode_vdvj_insn(neg_vec_insn[lasx][vece],
2355                                      TCG_VEC_TMP0, a2));
2356        a2 = TCG_VEC_TMP0;
2357        /* fall through */
2358    case INDEX_op_rotrv_vec:
2359        insn = rotrv_vec_insn[lasx][vece];
2360        goto vdvjvk;
2361    case INDEX_op_shli_vec:
2362        insn = shli_vec_insn[lasx][vece];
2363        goto vdvjukN;
2364    case INDEX_op_shri_vec:
2365        insn = shri_vec_insn[lasx][vece];
2366        goto vdvjukN;
2367    case INDEX_op_sari_vec:
2368        insn = sari_vec_insn[lasx][vece];
2369        goto vdvjukN;
2370    case INDEX_op_rotli_vec:
2371        /* rotli_vec a1, a2 = rotri_vec a1, -a2 */
2372        a2 = extract32(-a2, 0, 3 + vece);
2373        insn = rotri_vec_insn[lasx][vece];
2374        goto vdvjukN;
2375    case INDEX_op_bitsel_vec:
2376        /* vbitsel vd, vj, vk, va = bitsel_vec vd, va, vk, vj */
2377        if (lasx) {
2378            tcg_out_opc_xvbitsel_v(s, a0, a3, a2, a1);
2379        } else {
2380            tcg_out_opc_vbitsel_v(s, a0, a3, a2, a1);
2381        }
2382        break;
2383    case INDEX_op_dupm_vec:
2384        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
2385        break;
2386    default:
2387        g_assert_not_reached();
2388    vdvjvk:
2389        tcg_out32(s, encode_vdvjvk_insn(insn, a0, a1, a2));
2390        break;
2391    vdvjukN:
2392        switch (vece) {
2393        case MO_8:
2394            tcg_out32(s, encode_vdvjuk3_insn(insn, a0, a1, a2));
2395            break;
2396        case MO_16:
2397            tcg_out32(s, encode_vdvjuk4_insn(insn, a0, a1, a2));
2398            break;
2399        case MO_32:
2400            tcg_out32(s, encode_vdvjuk5_insn(insn, a0, a1, a2));
2401            break;
2402        case MO_64:
2403            tcg_out32(s, encode_vdvjuk6_insn(insn, a0, a1, a2));
2404            break;
2405        default:
2406            g_assert_not_reached();
2407        }
2408        break;
2409    }
2410}
2411
2412int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
2413{
2414    switch (opc) {
2415    case INDEX_op_ld_vec:
2416    case INDEX_op_st_vec:
2417    case INDEX_op_dup_vec:
2418    case INDEX_op_dupm_vec:
2419    case INDEX_op_cmp_vec:
2420    case INDEX_op_add_vec:
2421    case INDEX_op_sub_vec:
2422    case INDEX_op_and_vec:
2423    case INDEX_op_andc_vec:
2424    case INDEX_op_or_vec:
2425    case INDEX_op_orc_vec:
2426    case INDEX_op_xor_vec:
2427    case INDEX_op_nor_vec:
2428    case INDEX_op_not_vec:
2429    case INDEX_op_neg_vec:
2430    case INDEX_op_mul_vec:
2431    case INDEX_op_smin_vec:
2432    case INDEX_op_smax_vec:
2433    case INDEX_op_umin_vec:
2434    case INDEX_op_umax_vec:
2435    case INDEX_op_ssadd_vec:
2436    case INDEX_op_usadd_vec:
2437    case INDEX_op_sssub_vec:
2438    case INDEX_op_ussub_vec:
2439    case INDEX_op_shlv_vec:
2440    case INDEX_op_shrv_vec:
2441    case INDEX_op_sarv_vec:
2442    case INDEX_op_bitsel_vec:
2443        return 1;
2444    default:
2445        return 0;
2446    }
2447}
2448
2449void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
2450                       TCGArg a0, ...)
2451{
2452    g_assert_not_reached();
2453}
2454
2455static TCGConstraintSetIndex
2456tcg_target_op_def(TCGOpcode op, TCGType type, unsigned flags)
2457{
2458    switch (op) {
2459    case INDEX_op_goto_ptr:
2460        return C_O0_I1(r);
2461
2462    case INDEX_op_st8_i32:
2463    case INDEX_op_st8_i64:
2464    case INDEX_op_st16_i32:
2465    case INDEX_op_st16_i64:
2466    case INDEX_op_st32_i64:
2467    case INDEX_op_st_i32:
2468    case INDEX_op_st_i64:
2469    case INDEX_op_qemu_st_i32:
2470    case INDEX_op_qemu_st_i64:
2471        return C_O0_I2(rz, r);
2472
2473    case INDEX_op_qemu_ld_i128:
2474        return C_N2_I1(r, r, r);
2475
2476    case INDEX_op_qemu_st_i128:
2477        return C_O0_I3(r, r, r);
2478
2479    case INDEX_op_ld8s_i32:
2480    case INDEX_op_ld8s_i64:
2481    case INDEX_op_ld8u_i32:
2482    case INDEX_op_ld8u_i64:
2483    case INDEX_op_ld16s_i32:
2484    case INDEX_op_ld16s_i64:
2485    case INDEX_op_ld16u_i32:
2486    case INDEX_op_ld16u_i64:
2487    case INDEX_op_ld32s_i64:
2488    case INDEX_op_ld32u_i64:
2489    case INDEX_op_ld_i32:
2490    case INDEX_op_ld_i64:
2491    case INDEX_op_qemu_ld_i32:
2492    case INDEX_op_qemu_ld_i64:
2493        return C_O1_I1(r, r);
2494
2495    case INDEX_op_ld_vec:
2496    case INDEX_op_dupm_vec:
2497    case INDEX_op_dup_vec:
2498        return C_O1_I1(w, r);
2499
2500    case INDEX_op_st_vec:
2501        return C_O0_I2(w, r);
2502
2503    case INDEX_op_cmp_vec:
2504        return C_O1_I2(w, w, wM);
2505
2506    case INDEX_op_add_vec:
2507    case INDEX_op_sub_vec:
2508        return C_O1_I2(w, w, wA);
2509
2510    case INDEX_op_and_vec:
2511    case INDEX_op_andc_vec:
2512    case INDEX_op_or_vec:
2513    case INDEX_op_orc_vec:
2514    case INDEX_op_xor_vec:
2515    case INDEX_op_nor_vec:
2516    case INDEX_op_mul_vec:
2517    case INDEX_op_smin_vec:
2518    case INDEX_op_smax_vec:
2519    case INDEX_op_umin_vec:
2520    case INDEX_op_umax_vec:
2521    case INDEX_op_ssadd_vec:
2522    case INDEX_op_usadd_vec:
2523    case INDEX_op_sssub_vec:
2524    case INDEX_op_ussub_vec:
2525    case INDEX_op_shlv_vec:
2526    case INDEX_op_shrv_vec:
2527    case INDEX_op_sarv_vec:
2528    case INDEX_op_rotrv_vec:
2529    case INDEX_op_rotlv_vec:
2530        return C_O1_I2(w, w, w);
2531
2532    case INDEX_op_not_vec:
2533    case INDEX_op_neg_vec:
2534    case INDEX_op_shli_vec:
2535    case INDEX_op_shri_vec:
2536    case INDEX_op_sari_vec:
2537    case INDEX_op_rotli_vec:
2538        return C_O1_I1(w, w);
2539
2540    case INDEX_op_bitsel_vec:
2541        return C_O1_I3(w, w, w, w);
2542
2543    default:
2544        return C_NotImplemented;
2545    }
2546}
2547
2548static const int tcg_target_callee_save_regs[] = {
2549    TCG_REG_S0,     /* used for the global env (TCG_AREG0) */
2550    TCG_REG_S1,
2551    TCG_REG_S2,
2552    TCG_REG_S3,
2553    TCG_REG_S4,
2554    TCG_REG_S5,
2555    TCG_REG_S6,
2556    TCG_REG_S7,
2557    TCG_REG_S8,
2558    TCG_REG_S9,
2559    TCG_REG_RA,     /* should be last for ABI compliance */
2560};
2561
2562/* Stack frame parameters.  */
2563#define REG_SIZE   (TCG_TARGET_REG_BITS / 8)
2564#define SAVE_SIZE  ((int)ARRAY_SIZE(tcg_target_callee_save_regs) * REG_SIZE)
2565#define TEMP_SIZE  (CPU_TEMP_BUF_NLONGS * (int)sizeof(long))
2566#define FRAME_SIZE ((TCG_STATIC_CALL_ARGS_SIZE + TEMP_SIZE + SAVE_SIZE \
2567                     + TCG_TARGET_STACK_ALIGN - 1) \
2568                    & -TCG_TARGET_STACK_ALIGN)
2569#define SAVE_OFS   (TCG_STATIC_CALL_ARGS_SIZE + TEMP_SIZE)
2570
2571/* We're expecting to be able to use an immediate for frame allocation.  */
2572QEMU_BUILD_BUG_ON(FRAME_SIZE > 0x7ff);
2573
2574/* Generate global QEMU prologue and epilogue code */
2575static void tcg_target_qemu_prologue(TCGContext *s)
2576{
2577    int i;
2578
2579    tcg_set_frame(s, TCG_REG_SP, TCG_STATIC_CALL_ARGS_SIZE, TEMP_SIZE);
2580
2581    /* TB prologue */
2582    tcg_out_opc_addi_d(s, TCG_REG_SP, TCG_REG_SP, -FRAME_SIZE);
2583    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
2584        tcg_out_st(s, TCG_TYPE_REG, tcg_target_callee_save_regs[i],
2585                   TCG_REG_SP, SAVE_OFS + i * REG_SIZE);
2586    }
2587
2588    if (!tcg_use_softmmu && guest_base) {
2589        tcg_out_movi(s, TCG_TYPE_PTR, TCG_GUEST_BASE_REG, guest_base);
2590        tcg_regset_set_reg(s->reserved_regs, TCG_GUEST_BASE_REG);
2591    }
2592
2593    /* Call generated code */
2594    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
2595    tcg_out_opc_jirl(s, TCG_REG_ZERO, tcg_target_call_iarg_regs[1], 0);
2596
2597    /* Return path for goto_ptr. Set return value to 0 */
2598    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
2599    tcg_out_mov(s, TCG_TYPE_REG, TCG_REG_A0, TCG_REG_ZERO);
2600
2601    /* TB epilogue */
2602    tb_ret_addr = tcg_splitwx_to_rx(s->code_ptr);
2603    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
2604        tcg_out_ld(s, TCG_TYPE_REG, tcg_target_callee_save_regs[i],
2605                   TCG_REG_SP, SAVE_OFS + i * REG_SIZE);
2606    }
2607
2608    tcg_out_opc_addi_d(s, TCG_REG_SP, TCG_REG_SP, FRAME_SIZE);
2609    tcg_out_opc_jirl(s, TCG_REG_ZERO, TCG_REG_RA, 0);
2610}
2611
2612static void tcg_out_tb_start(TCGContext *s)
2613{
2614    /* nothing to do */
2615}
2616
2617static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
2618{
2619    for (int i = 0; i < count; ++i) {
2620        /* Canonical nop is andi r0,r0,0 */
2621        p[i] = OPC_ANDI;
2622    }
2623}
2624
2625static void tcg_target_init(TCGContext *s)
2626{
2627    unsigned long hwcap = qemu_getauxval(AT_HWCAP);
2628
2629    /* Server and desktop class cpus have UAL; embedded cpus do not. */
2630    if (!(hwcap & HWCAP_LOONGARCH_UAL)) {
2631        error_report("TCG: unaligned access support required; exiting");
2632        exit(EXIT_FAILURE);
2633    }
2634
2635    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
2636    tcg_target_available_regs[TCG_TYPE_I64] = ALL_GENERAL_REGS;
2637
2638    tcg_target_call_clobber_regs = ALL_GENERAL_REGS | ALL_VECTOR_REGS;
2639    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S0);
2640    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S1);
2641    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S2);
2642    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S3);
2643    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S4);
2644    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S5);
2645    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S6);
2646    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S7);
2647    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S8);
2648    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S9);
2649
2650    if (cpuinfo & CPUINFO_LSX) {
2651        tcg_target_available_regs[TCG_TYPE_V64] = ALL_VECTOR_REGS;
2652        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
2653        if (cpuinfo & CPUINFO_LASX) {
2654            tcg_target_available_regs[TCG_TYPE_V256] = ALL_VECTOR_REGS;
2655        }
2656        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V24);
2657        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V25);
2658        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V26);
2659        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V27);
2660        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V28);
2661        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V29);
2662        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V30);
2663        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V31);
2664    }
2665
2666    s->reserved_regs = 0;
2667    tcg_regset_set_reg(s->reserved_regs, TCG_REG_ZERO);
2668    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP0);
2669    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP1);
2670    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP2);
2671    tcg_regset_set_reg(s->reserved_regs, TCG_REG_SP);
2672    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TP);
2673    tcg_regset_set_reg(s->reserved_regs, TCG_REG_RESERVED);
2674    tcg_regset_set_reg(s->reserved_regs, TCG_VEC_TMP0);
2675}
2676
2677typedef struct {
2678    DebugFrameHeader h;
2679    uint8_t fde_def_cfa[4];
2680    uint8_t fde_reg_ofs[ARRAY_SIZE(tcg_target_callee_save_regs) * 2];
2681} DebugFrame;
2682
2683#define ELF_HOST_MACHINE EM_LOONGARCH
2684
2685static const DebugFrame debug_frame = {
2686    .h.cie.len = sizeof(DebugFrameCIE) - 4, /* length after .len member */
2687    .h.cie.id = -1,
2688    .h.cie.version = 1,
2689    .h.cie.code_align = 1,
2690    .h.cie.data_align = -(TCG_TARGET_REG_BITS / 8) & 0x7f, /* sleb128 */
2691    .h.cie.return_column = TCG_REG_RA,
2692
2693    /* Total FDE size does not include the "len" member.  */
2694    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
2695
2696    .fde_def_cfa = {
2697        12, TCG_REG_SP,                 /* DW_CFA_def_cfa sp, ...  */
2698        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
2699        (FRAME_SIZE >> 7)
2700    },
2701    .fde_reg_ofs = {
2702        0x80 + 23, 11,                  /* DW_CFA_offset, s0, -88 */
2703        0x80 + 24, 10,                  /* DW_CFA_offset, s1, -80 */
2704        0x80 + 25, 9,                   /* DW_CFA_offset, s2, -72 */
2705        0x80 + 26, 8,                   /* DW_CFA_offset, s3, -64 */
2706        0x80 + 27, 7,                   /* DW_CFA_offset, s4, -56 */
2707        0x80 + 28, 6,                   /* DW_CFA_offset, s5, -48 */
2708        0x80 + 29, 5,                   /* DW_CFA_offset, s6, -40 */
2709        0x80 + 30, 4,                   /* DW_CFA_offset, s7, -32 */
2710        0x80 + 31, 3,                   /* DW_CFA_offset, s8, -24 */
2711        0x80 + 22, 2,                   /* DW_CFA_offset, s9, -16 */
2712        0x80 + 1 , 1,                   /* DW_CFA_offset, ra, -8 */
2713    }
2714};
2715
2716void tcg_register_jit(const void *buf, size_t buf_size)
2717{
2718    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
2719}
2720