xref: /openbmc/qemu/tcg/loongarch64/tcg-target.c.inc (revision 60f34f55f1a708c071774bd7f837163d6b686867)
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
549static void tcg_out_clzctz(TCGContext *s, LoongArchInsn opc,
550                           TCGReg a0, TCGReg a1, TCGReg a2,
551                           bool c2, bool is_32bit)
552{
553    if (c2) {
554        /*
555         * Fast path: semantics already satisfied due to constraint and
556         * insn behavior, single instruction is enough.
557         */
558        tcg_debug_assert(a2 == (is_32bit ? 32 : 64));
559        /* all clz/ctz insns belong to DJ-format */
560        tcg_out32(s, encode_dj_insn(opc, a0, a1));
561        return;
562    }
563
564    tcg_out32(s, encode_dj_insn(opc, TCG_REG_TMP0, a1));
565    /* a0 = a1 ? REG_TMP0 : a2 */
566    tcg_out_opc_maskeqz(s, TCG_REG_TMP0, TCG_REG_TMP0, a1);
567    tcg_out_opc_masknez(s, a0, a2, a1);
568    tcg_out_opc_or(s, a0, TCG_REG_TMP0, a0);
569}
570
571#define SETCOND_INV    TCG_TARGET_NB_REGS
572#define SETCOND_NEZ    (SETCOND_INV << 1)
573#define SETCOND_FLAGS  (SETCOND_INV | SETCOND_NEZ)
574
575static int tcg_out_setcond_int(TCGContext *s, TCGCond cond, TCGReg ret,
576                               TCGReg arg1, tcg_target_long arg2, bool c2)
577{
578    int flags = 0;
579
580    switch (cond) {
581    case TCG_COND_EQ:    /* -> NE  */
582    case TCG_COND_GE:    /* -> LT  */
583    case TCG_COND_GEU:   /* -> LTU */
584    case TCG_COND_GT:    /* -> LE  */
585    case TCG_COND_GTU:   /* -> LEU */
586        cond = tcg_invert_cond(cond);
587        flags ^= SETCOND_INV;
588        break;
589    default:
590        break;
591    }
592
593    switch (cond) {
594    case TCG_COND_LE:
595    case TCG_COND_LEU:
596        /*
597         * If we have a constant input, the most efficient way to implement
598         * LE is by adding 1 and using LT.  Watch out for wrap around for LEU.
599         * We don't need to care for this for LE because the constant input
600         * is still constrained to int32_t, and INT32_MAX+1 is representable
601         * in the 64-bit temporary register.
602         */
603        if (c2) {
604            if (cond == TCG_COND_LEU) {
605                /* unsigned <= -1 is true */
606                if (arg2 == -1) {
607                    tcg_out_movi(s, TCG_TYPE_REG, ret, !(flags & SETCOND_INV));
608                    return ret;
609                }
610                cond = TCG_COND_LTU;
611            } else {
612                cond = TCG_COND_LT;
613            }
614            arg2 += 1;
615        } else {
616            TCGReg tmp = arg2;
617            arg2 = arg1;
618            arg1 = tmp;
619            cond = tcg_swap_cond(cond);    /* LE -> GE */
620            cond = tcg_invert_cond(cond);  /* GE -> LT */
621            flags ^= SETCOND_INV;
622        }
623        break;
624    default:
625        break;
626    }
627
628    switch (cond) {
629    case TCG_COND_NE:
630        flags |= SETCOND_NEZ;
631        if (!c2) {
632            tcg_out_opc_xor(s, ret, arg1, arg2);
633        } else if (arg2 == 0) {
634            ret = arg1;
635        } else if (arg2 >= 0 && arg2 <= 0xfff) {
636            tcg_out_opc_xori(s, ret, arg1, arg2);
637        } else {
638            tcg_out_addi(s, TCG_TYPE_REG, ret, arg1, -arg2);
639        }
640        break;
641
642    case TCG_COND_LT:
643    case TCG_COND_LTU:
644        if (c2) {
645            if (arg2 >= -0x800 && arg2 <= 0x7ff) {
646                if (cond == TCG_COND_LT) {
647                    tcg_out_opc_slti(s, ret, arg1, arg2);
648                } else {
649                    tcg_out_opc_sltui(s, ret, arg1, arg2);
650                }
651                break;
652            }
653            tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_TMP0, arg2);
654            arg2 = TCG_REG_TMP0;
655        }
656        if (cond == TCG_COND_LT) {
657            tcg_out_opc_slt(s, ret, arg1, arg2);
658        } else {
659            tcg_out_opc_sltu(s, ret, arg1, arg2);
660        }
661        break;
662
663    default:
664        g_assert_not_reached();
665    }
666
667    return ret | flags;
668}
669
670static void tcg_out_setcond(TCGContext *s, TCGCond cond, TCGReg ret,
671                            TCGReg arg1, tcg_target_long arg2, bool c2)
672{
673    int tmpflags = tcg_out_setcond_int(s, cond, ret, arg1, arg2, c2);
674
675    if (tmpflags != ret) {
676        TCGReg tmp = tmpflags & ~SETCOND_FLAGS;
677
678        switch (tmpflags & SETCOND_FLAGS) {
679        case SETCOND_INV:
680            /* Intermediate result is boolean: simply invert. */
681            tcg_out_opc_xori(s, ret, tmp, 1);
682            break;
683        case SETCOND_NEZ:
684            /* Intermediate result is zero/non-zero: test != 0. */
685            tcg_out_opc_sltu(s, ret, TCG_REG_ZERO, tmp);
686            break;
687        case SETCOND_NEZ | SETCOND_INV:
688            /* Intermediate result is zero/non-zero: test == 0. */
689            tcg_out_opc_sltui(s, ret, tmp, 1);
690            break;
691        default:
692            g_assert_not_reached();
693        }
694    }
695}
696
697static void tcg_out_movcond(TCGContext *s, TCGCond cond, TCGReg ret,
698                            TCGReg c1, tcg_target_long c2, bool const2,
699                            TCGReg v1, TCGReg v2)
700{
701    int tmpflags = tcg_out_setcond_int(s, cond, TCG_REG_TMP0, c1, c2, const2);
702    TCGReg t;
703
704    /* Standardize the test below to t != 0. */
705    if (tmpflags & SETCOND_INV) {
706        t = v1, v1 = v2, v2 = t;
707    }
708
709    t = tmpflags & ~SETCOND_FLAGS;
710    if (v1 == TCG_REG_ZERO) {
711        tcg_out_opc_masknez(s, ret, v2, t);
712    } else if (v2 == TCG_REG_ZERO) {
713        tcg_out_opc_maskeqz(s, ret, v1, t);
714    } else {
715        tcg_out_opc_masknez(s, TCG_REG_TMP2, v2, t); /* t ? 0 : v2 */
716        tcg_out_opc_maskeqz(s, TCG_REG_TMP1, v1, t); /* t ? v1 : 0 */
717        tcg_out_opc_or(s, ret, TCG_REG_TMP1, TCG_REG_TMP2);
718    }
719}
720
721/*
722 * Branch helpers
723 */
724
725static const struct {
726    LoongArchInsn op;
727    bool swap;
728} tcg_brcond_to_loongarch[] = {
729    [TCG_COND_EQ] =  { OPC_BEQ,  false },
730    [TCG_COND_NE] =  { OPC_BNE,  false },
731    [TCG_COND_LT] =  { OPC_BGT,  true  },
732    [TCG_COND_GE] =  { OPC_BLE,  true  },
733    [TCG_COND_LE] =  { OPC_BLE,  false },
734    [TCG_COND_GT] =  { OPC_BGT,  false },
735    [TCG_COND_LTU] = { OPC_BGTU, true  },
736    [TCG_COND_GEU] = { OPC_BLEU, true  },
737    [TCG_COND_LEU] = { OPC_BLEU, false },
738    [TCG_COND_GTU] = { OPC_BGTU, false }
739};
740
741static void tcg_out_brcond(TCGContext *s, TCGCond cond, TCGReg arg1,
742                           TCGReg arg2, TCGLabel *l)
743{
744    LoongArchInsn op = tcg_brcond_to_loongarch[cond].op;
745
746    tcg_debug_assert(op != 0);
747
748    if (tcg_brcond_to_loongarch[cond].swap) {
749        TCGReg t = arg1;
750        arg1 = arg2;
751        arg2 = t;
752    }
753
754    /* all conditional branch insns belong to DJSk16-format */
755    tcg_out_reloc(s, s->code_ptr, R_LOONGARCH_BR_SK16, l, 0);
756    tcg_out32(s, encode_djsk16_insn(op, arg1, arg2, 0));
757}
758
759static void tcg_out_call_int(TCGContext *s, const tcg_insn_unit *arg, bool tail)
760{
761    TCGReg link = tail ? TCG_REG_ZERO : TCG_REG_RA;
762    ptrdiff_t offset = tcg_pcrel_diff(s, arg);
763
764    tcg_debug_assert((offset & 3) == 0);
765    if (offset == sextreg(offset, 0, 28)) {
766        /* short jump: +/- 256MiB */
767        if (tail) {
768            tcg_out_opc_b(s, offset >> 2);
769        } else {
770            tcg_out_opc_bl(s, offset >> 2);
771        }
772    } else if (offset == sextreg(offset, 0, 38)) {
773        /* long jump: +/- 256GiB */
774        tcg_target_long lo = sextreg(offset, 0, 18);
775        tcg_target_long hi = offset - lo;
776        tcg_out_opc_pcaddu18i(s, TCG_REG_TMP0, hi >> 18);
777        tcg_out_opc_jirl(s, link, TCG_REG_TMP0, lo >> 2);
778    } else {
779        /* far jump: 64-bit */
780        tcg_target_long lo = sextreg((tcg_target_long)arg, 0, 18);
781        tcg_target_long hi = (tcg_target_long)arg - lo;
782        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, hi);
783        tcg_out_opc_jirl(s, link, TCG_REG_TMP0, lo >> 2);
784    }
785}
786
787static void tcg_out_call(TCGContext *s, const tcg_insn_unit *arg,
788                         const TCGHelperInfo *info)
789{
790    tcg_out_call_int(s, arg, false);
791}
792
793/*
794 * Load/store helpers
795 */
796
797static void tcg_out_ldst(TCGContext *s, LoongArchInsn opc, TCGReg data,
798                         TCGReg addr, intptr_t offset)
799{
800    intptr_t imm12 = sextreg(offset, 0, 12);
801
802    if (offset != imm12) {
803        intptr_t diff = tcg_pcrel_diff(s, (void *)offset);
804
805        if (addr == TCG_REG_ZERO && diff == (int32_t)diff) {
806            imm12 = sextreg(diff, 0, 12);
807            tcg_out_opc_pcaddu12i(s, TCG_REG_TMP2, (diff - imm12) >> 12);
808        } else {
809            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP2, offset - imm12);
810            if (addr != TCG_REG_ZERO) {
811                tcg_out_opc_add_d(s, TCG_REG_TMP2, TCG_REG_TMP2, addr);
812            }
813        }
814        addr = TCG_REG_TMP2;
815    }
816
817    switch (opc) {
818    case OPC_LD_B:
819    case OPC_LD_BU:
820    case OPC_LD_H:
821    case OPC_LD_HU:
822    case OPC_LD_W:
823    case OPC_LD_WU:
824    case OPC_LD_D:
825    case OPC_ST_B:
826    case OPC_ST_H:
827    case OPC_ST_W:
828    case OPC_ST_D:
829        tcg_out32(s, encode_djsk12_insn(opc, data, addr, imm12));
830        break;
831    case OPC_FLD_S:
832    case OPC_FLD_D:
833    case OPC_FST_S:
834    case OPC_FST_D:
835        tcg_out32(s, encode_fdjsk12_insn(opc, data, addr, imm12));
836        break;
837    default:
838        g_assert_not_reached();
839    }
840}
841
842static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg dest,
843                       TCGReg base, intptr_t offset)
844{
845    switch (type) {
846    case TCG_TYPE_I32:
847        if (dest < TCG_REG_V0) {
848            tcg_out_ldst(s, OPC_LD_W, dest, base, offset);
849        } else {
850            tcg_out_ldst(s, OPC_FLD_S, dest, base, offset);
851        }
852        break;
853    case TCG_TYPE_I64:
854    case TCG_TYPE_V64:
855        if (dest < TCG_REG_V0) {
856            tcg_out_ldst(s, OPC_LD_D, dest, base, offset);
857        } else {
858            tcg_out_ldst(s, OPC_FLD_D, dest, base, offset);
859        }
860        break;
861    case TCG_TYPE_V128:
862        if (-0x800 <= offset && offset <= 0x7ff) {
863            tcg_out_opc_vld(s, dest, base, offset);
864        } else {
865            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, offset);
866            tcg_out_opc_vldx(s, dest, base, TCG_REG_TMP0);
867        }
868        break;
869    case TCG_TYPE_V256:
870        if (-0x800 <= offset && offset <= 0x7ff) {
871            tcg_out_opc_xvld(s, dest, base, offset);
872        } else {
873            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, offset);
874            tcg_out_opc_xvldx(s, dest, base, TCG_REG_TMP0);
875        }
876        break;
877    default:
878        g_assert_not_reached();
879    }
880}
881
882static void tcg_out_st(TCGContext *s, TCGType type, TCGReg src,
883                       TCGReg base, intptr_t offset)
884{
885    switch (type) {
886    case TCG_TYPE_I32:
887        if (src < TCG_REG_V0) {
888            tcg_out_ldst(s, OPC_ST_W, src, base, offset);
889        } else {
890            tcg_out_ldst(s, OPC_FST_S, src, base, offset);
891        }
892        break;
893    case TCG_TYPE_I64:
894    case TCG_TYPE_V64:
895        if (src < TCG_REG_V0) {
896            tcg_out_ldst(s, OPC_ST_D, src, base, offset);
897        } else {
898            tcg_out_ldst(s, OPC_FST_D, src, base, offset);
899        }
900        break;
901    case TCG_TYPE_V128:
902        if (-0x800 <= offset && offset <= 0x7ff) {
903            tcg_out_opc_vst(s, src, base, offset);
904        } else {
905            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, offset);
906            tcg_out_opc_vstx(s, src, base, TCG_REG_TMP0);
907        }
908        break;
909    case TCG_TYPE_V256:
910        if (-0x800 <= offset && offset <= 0x7ff) {
911            tcg_out_opc_xvst(s, src, base, offset);
912        } else {
913            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, offset);
914            tcg_out_opc_xvstx(s, src, base, TCG_REG_TMP0);
915        }
916        break;
917    default:
918        g_assert_not_reached();
919    }
920}
921
922static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
923                        TCGReg base, intptr_t ofs)
924{
925    if (val == 0) {
926        tcg_out_st(s, type, TCG_REG_ZERO, base, ofs);
927        return true;
928    }
929    return false;
930}
931
932/*
933 * Load/store helpers for SoftMMU, and qemu_ld/st implementations
934 */
935
936static bool tcg_out_goto(TCGContext *s, const tcg_insn_unit *target)
937{
938    tcg_out_opc_b(s, 0);
939    return reloc_br_sd10k16(s->code_ptr - 1, target);
940}
941
942static const TCGLdstHelperParam ldst_helper_param = {
943    .ntmp = 1, .tmp = { TCG_REG_TMP0 }
944};
945
946static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
947{
948    MemOp opc = get_memop(l->oi);
949
950    /* resolve label address */
951    if (!reloc_br_sk16(l->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
952        return false;
953    }
954
955    tcg_out_ld_helper_args(s, l, &ldst_helper_param);
956    tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SIZE], false);
957    tcg_out_ld_helper_ret(s, l, false, &ldst_helper_param);
958    return tcg_out_goto(s, l->raddr);
959}
960
961static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
962{
963    MemOp opc = get_memop(l->oi);
964
965    /* resolve label address */
966    if (!reloc_br_sk16(l->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
967        return false;
968    }
969
970    tcg_out_st_helper_args(s, l, &ldst_helper_param);
971    tcg_out_call_int(s, qemu_st_helpers[opc & MO_SIZE], false);
972    return tcg_out_goto(s, l->raddr);
973}
974
975typedef struct {
976    TCGReg base;
977    TCGReg index;
978    TCGAtomAlign aa;
979} HostAddress;
980
981bool tcg_target_has_memory_bswap(MemOp memop)
982{
983    return false;
984}
985
986/* We expect to use a 12-bit negative offset from ENV.  */
987#define MIN_TLB_MASK_TABLE_OFS  -(1 << 11)
988
989/*
990 * For system-mode, perform the TLB load and compare.
991 * For user-mode, perform any required alignment tests.
992 * In both cases, return a TCGLabelQemuLdst structure if the slow path
993 * is required and fill in @h with the host address for the fast path.
994 */
995static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
996                                           TCGReg addr_reg, MemOpIdx oi,
997                                           bool is_ld)
998{
999    TCGType addr_type = s->addr_type;
1000    TCGLabelQemuLdst *ldst = NULL;
1001    MemOp opc = get_memop(oi);
1002    MemOp a_bits;
1003
1004    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, false);
1005    a_bits = h->aa.align;
1006
1007    if (tcg_use_softmmu) {
1008        unsigned s_bits = opc & MO_SIZE;
1009        int mem_index = get_mmuidx(oi);
1010        int fast_ofs = tlb_mask_table_ofs(s, mem_index);
1011        int mask_ofs = fast_ofs + offsetof(CPUTLBDescFast, mask);
1012        int table_ofs = fast_ofs + offsetof(CPUTLBDescFast, table);
1013
1014        ldst = new_ldst_label(s);
1015        ldst->is_ld = is_ld;
1016        ldst->oi = oi;
1017        ldst->addr_reg = addr_reg;
1018
1019        tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP0, TCG_AREG0, mask_ofs);
1020        tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP1, TCG_AREG0, table_ofs);
1021
1022        tcg_out_opc_srli_d(s, TCG_REG_TMP2, addr_reg,
1023                           s->page_bits - CPU_TLB_ENTRY_BITS);
1024        tcg_out_opc_and(s, TCG_REG_TMP2, TCG_REG_TMP2, TCG_REG_TMP0);
1025        tcg_out_opc_add_d(s, TCG_REG_TMP2, TCG_REG_TMP2, TCG_REG_TMP1);
1026
1027        /* Load the tlb comparator and the addend.  */
1028        QEMU_BUILD_BUG_ON(HOST_BIG_ENDIAN);
1029        tcg_out_ld(s, addr_type, TCG_REG_TMP0, TCG_REG_TMP2,
1030                   is_ld ? offsetof(CPUTLBEntry, addr_read)
1031                         : offsetof(CPUTLBEntry, addr_write));
1032        tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP2, TCG_REG_TMP2,
1033                   offsetof(CPUTLBEntry, addend));
1034
1035        /*
1036         * For aligned accesses, we check the first byte and include the
1037         * alignment bits within the address.  For unaligned access, we
1038         * check that we don't cross pages using the address of the last
1039         * byte of the access.
1040         */
1041        if (a_bits < s_bits) {
1042            unsigned a_mask = (1u << a_bits) - 1;
1043            unsigned s_mask = (1u << s_bits) - 1;
1044            tcg_out_addi(s, addr_type, TCG_REG_TMP1, addr_reg, s_mask - a_mask);
1045        } else {
1046            tcg_out_mov(s, addr_type, TCG_REG_TMP1, addr_reg);
1047        }
1048        tcg_out_opc_bstrins_d(s, TCG_REG_TMP1, TCG_REG_ZERO,
1049                              a_bits, s->page_bits - 1);
1050
1051        /* Compare masked address with the TLB entry.  */
1052        ldst->label_ptr[0] = s->code_ptr;
1053        tcg_out_opc_bne(s, TCG_REG_TMP0, TCG_REG_TMP1, 0);
1054
1055        h->index = TCG_REG_TMP2;
1056    } else {
1057        if (a_bits) {
1058            ldst = new_ldst_label(s);
1059
1060            ldst->is_ld = is_ld;
1061            ldst->oi = oi;
1062            ldst->addr_reg = addr_reg;
1063
1064            /*
1065             * Without micro-architecture details, we don't know which of
1066             * bstrpick or andi is faster, so use bstrpick as it's not
1067             * constrained by imm field width. Not to say alignments >= 2^12
1068             * are going to happen any time soon.
1069             */
1070            tcg_out_opc_bstrpick_d(s, TCG_REG_TMP1, addr_reg, 0, a_bits - 1);
1071
1072            ldst->label_ptr[0] = s->code_ptr;
1073            tcg_out_opc_bne(s, TCG_REG_TMP1, TCG_REG_ZERO, 0);
1074        }
1075
1076        h->index = guest_base ? TCG_GUEST_BASE_REG : TCG_REG_ZERO;
1077    }
1078
1079    if (addr_type == TCG_TYPE_I32) {
1080        h->base = TCG_REG_TMP0;
1081        tcg_out_ext32u(s, h->base, addr_reg);
1082    } else {
1083        h->base = addr_reg;
1084    }
1085
1086    return ldst;
1087}
1088
1089static void tcg_out_qemu_ld_indexed(TCGContext *s, MemOp opc, TCGType type,
1090                                    TCGReg rd, HostAddress h)
1091{
1092    /* Byte swapping is left to middle-end expansion.  */
1093    tcg_debug_assert((opc & MO_BSWAP) == 0);
1094
1095    switch (opc & MO_SSIZE) {
1096    case MO_UB:
1097        tcg_out_opc_ldx_bu(s, rd, h.base, h.index);
1098        break;
1099    case MO_SB:
1100        tcg_out_opc_ldx_b(s, rd, h.base, h.index);
1101        break;
1102    case MO_UW:
1103        tcg_out_opc_ldx_hu(s, rd, h.base, h.index);
1104        break;
1105    case MO_SW:
1106        tcg_out_opc_ldx_h(s, rd, h.base, h.index);
1107        break;
1108    case MO_UL:
1109        if (type == TCG_TYPE_I64) {
1110            tcg_out_opc_ldx_wu(s, rd, h.base, h.index);
1111            break;
1112        }
1113        /* fallthrough */
1114    case MO_SL:
1115        tcg_out_opc_ldx_w(s, rd, h.base, h.index);
1116        break;
1117    case MO_UQ:
1118        tcg_out_opc_ldx_d(s, rd, h.base, h.index);
1119        break;
1120    default:
1121        g_assert_not_reached();
1122    }
1123}
1124
1125static void tcg_out_qemu_ld(TCGContext *s, TCGReg data_reg, TCGReg addr_reg,
1126                            MemOpIdx oi, TCGType data_type)
1127{
1128    TCGLabelQemuLdst *ldst;
1129    HostAddress h;
1130
1131    ldst = prepare_host_addr(s, &h, addr_reg, oi, true);
1132    tcg_out_qemu_ld_indexed(s, get_memop(oi), data_type, data_reg, h);
1133
1134    if (ldst) {
1135        ldst->type = data_type;
1136        ldst->datalo_reg = data_reg;
1137        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1138    }
1139}
1140
1141static void tcg_out_qemu_st_indexed(TCGContext *s, MemOp opc,
1142                                    TCGReg rd, HostAddress h)
1143{
1144    /* Byte swapping is left to middle-end expansion.  */
1145    tcg_debug_assert((opc & MO_BSWAP) == 0);
1146
1147    switch (opc & MO_SIZE) {
1148    case MO_8:
1149        tcg_out_opc_stx_b(s, rd, h.base, h.index);
1150        break;
1151    case MO_16:
1152        tcg_out_opc_stx_h(s, rd, h.base, h.index);
1153        break;
1154    case MO_32:
1155        tcg_out_opc_stx_w(s, rd, h.base, h.index);
1156        break;
1157    case MO_64:
1158        tcg_out_opc_stx_d(s, rd, h.base, h.index);
1159        break;
1160    default:
1161        g_assert_not_reached();
1162    }
1163}
1164
1165static void tcg_out_qemu_st(TCGContext *s, TCGReg data_reg, TCGReg addr_reg,
1166                            MemOpIdx oi, TCGType data_type)
1167{
1168    TCGLabelQemuLdst *ldst;
1169    HostAddress h;
1170
1171    ldst = prepare_host_addr(s, &h, addr_reg, oi, false);
1172    tcg_out_qemu_st_indexed(s, get_memop(oi), data_reg, h);
1173
1174    if (ldst) {
1175        ldst->type = data_type;
1176        ldst->datalo_reg = data_reg;
1177        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1178    }
1179}
1180
1181static void tcg_out_qemu_ldst_i128(TCGContext *s, TCGReg data_lo, TCGReg data_hi,
1182                                   TCGReg addr_reg, MemOpIdx oi, bool is_ld)
1183{
1184    TCGLabelQemuLdst *ldst;
1185    HostAddress h;
1186
1187    ldst = prepare_host_addr(s, &h, addr_reg, oi, is_ld);
1188
1189    if (h.aa.atom == MO_128) {
1190        /*
1191         * Use VLDX/VSTX when 128-bit atomicity is required.
1192         * If address is aligned to 16-bytes, the 128-bit load/store is atomic.
1193         */
1194        if (is_ld) {
1195            tcg_out_opc_vldx(s, TCG_VEC_TMP0, h.base, h.index);
1196            tcg_out_opc_vpickve2gr_d(s, data_lo, TCG_VEC_TMP0, 0);
1197            tcg_out_opc_vpickve2gr_d(s, data_hi, TCG_VEC_TMP0, 1);
1198        } else {
1199            tcg_out_opc_vinsgr2vr_d(s, TCG_VEC_TMP0, data_lo, 0);
1200            tcg_out_opc_vinsgr2vr_d(s, TCG_VEC_TMP0, data_hi, 1);
1201            tcg_out_opc_vstx(s, TCG_VEC_TMP0, h.base, h.index);
1202        }
1203    } else {
1204        /* Otherwise use a pair of LD/ST. */
1205        TCGReg base = h.base;
1206        if (h.index != TCG_REG_ZERO) {
1207            base = TCG_REG_TMP0;
1208            tcg_out_opc_add_d(s, base, h.base, h.index);
1209        }
1210        if (is_ld) {
1211            tcg_debug_assert(base != data_lo);
1212            tcg_out_opc_ld_d(s, data_lo, base, 0);
1213            tcg_out_opc_ld_d(s, data_hi, base, 8);
1214        } else {
1215            tcg_out_opc_st_d(s, data_lo, base, 0);
1216            tcg_out_opc_st_d(s, data_hi, base, 8);
1217        }
1218    }
1219
1220    if (ldst) {
1221        ldst->type = TCG_TYPE_I128;
1222        ldst->datalo_reg = data_lo;
1223        ldst->datahi_reg = data_hi;
1224        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1225    }
1226}
1227
1228/*
1229 * Entry-points
1230 */
1231
1232static const tcg_insn_unit *tb_ret_addr;
1233
1234static void tcg_out_exit_tb(TCGContext *s, uintptr_t a0)
1235{
1236    /* Reuse the zeroing that exists for goto_ptr.  */
1237    if (a0 == 0) {
1238        tcg_out_call_int(s, tcg_code_gen_epilogue, true);
1239    } else {
1240        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_A0, a0);
1241        tcg_out_call_int(s, tb_ret_addr, true);
1242    }
1243}
1244
1245static void tcg_out_goto_tb(TCGContext *s, int which)
1246{
1247    /*
1248     * Direct branch, or load indirect address, to be patched
1249     * by tb_target_set_jmp_target.  Check indirect load offset
1250     * in range early, regardless of direct branch distance,
1251     * via assert within tcg_out_opc_pcaddu2i.
1252     */
1253    uintptr_t i_addr = get_jmp_target_addr(s, which);
1254    intptr_t i_disp = tcg_pcrel_diff(s, (void *)i_addr);
1255
1256    set_jmp_insn_offset(s, which);
1257    tcg_out_opc_pcaddu2i(s, TCG_REG_TMP0, i_disp >> 2);
1258
1259    /* Finish the load and indirect branch. */
1260    tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP0, TCG_REG_TMP0, 0);
1261    tcg_out_opc_jirl(s, TCG_REG_ZERO, TCG_REG_TMP0, 0);
1262    set_jmp_reset_offset(s, which);
1263}
1264
1265void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
1266                              uintptr_t jmp_rx, uintptr_t jmp_rw)
1267{
1268    uintptr_t d_addr = tb->jmp_target_addr[n];
1269    ptrdiff_t d_disp = (ptrdiff_t)(d_addr - jmp_rx) >> 2;
1270    tcg_insn_unit insn;
1271
1272    /* Either directly branch, or load slot address for indirect branch. */
1273    if (d_disp == sextreg(d_disp, 0, 26)) {
1274        insn = encode_sd10k16_insn(OPC_B, d_disp);
1275    } else {
1276        uintptr_t i_addr = (uintptr_t)&tb->jmp_target_addr[n];
1277        intptr_t i_disp = i_addr - jmp_rx;
1278        insn = encode_dsj20_insn(OPC_PCADDU2I, TCG_REG_TMP0, i_disp >> 2);
1279    }
1280
1281    qatomic_set((tcg_insn_unit *)jmp_rw, insn);
1282    flush_idcache_range(jmp_rx, jmp_rw, 4);
1283}
1284
1285
1286static void tgen_add(TCGContext *s, TCGType type,
1287                     TCGReg a0, TCGReg a1, TCGReg a2)
1288{
1289    if (type == TCG_TYPE_I32) {
1290        tcg_out_opc_add_w(s, a0, a1, a2);
1291    } else {
1292        tcg_out_opc_add_d(s, a0, a1, a2);
1293    }
1294}
1295
1296static const TCGOutOpBinary outop_add = {
1297    .base.static_constraint = C_O1_I2(r, r, rJ),
1298    .out_rrr = tgen_add,
1299    .out_rri = tcg_out_addi,
1300};
1301
1302static void tgen_and(TCGContext *s, TCGType type,
1303                     TCGReg a0, TCGReg a1, TCGReg a2)
1304{
1305    tcg_out_opc_and(s, a0, a1, a2);
1306}
1307
1308static void tgen_andi(TCGContext *s, TCGType type,
1309                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1310{
1311    tcg_out_opc_andi(s, a0, a1, a2);
1312}
1313
1314static const TCGOutOpBinary outop_and = {
1315    .base.static_constraint = C_O1_I2(r, r, rU),
1316    .out_rrr = tgen_and,
1317    .out_rri = tgen_andi,
1318};
1319
1320static void tgen_andc(TCGContext *s, TCGType type,
1321                      TCGReg a0, TCGReg a1, TCGReg a2)
1322{
1323    tcg_out_opc_andn(s, a0, a1, a2);
1324}
1325
1326static const TCGOutOpBinary outop_andc = {
1327    .base.static_constraint = C_O1_I2(r, r, r),
1328    .out_rrr = tgen_andc,
1329};
1330
1331static const TCGOutOpBinary outop_eqv = {
1332    .base.static_constraint = C_NotImplemented,
1333};
1334
1335static const TCGOutOpBinary outop_nand = {
1336    .base.static_constraint = C_NotImplemented,
1337};
1338
1339static void tgen_nor(TCGContext *s, TCGType type,
1340                      TCGReg a0, TCGReg a1, TCGReg a2)
1341{
1342    tcg_out_opc_nor(s, a0, a1, a2);
1343}
1344
1345static const TCGOutOpBinary outop_nor = {
1346    .base.static_constraint = C_O1_I2(r, r, r),
1347    .out_rrr = tgen_nor,
1348};
1349
1350static void tgen_or(TCGContext *s, TCGType type,
1351                    TCGReg a0, TCGReg a1, TCGReg a2)
1352{
1353    tcg_out_opc_or(s, a0, a1, a2);
1354}
1355
1356static void tgen_ori(TCGContext *s, TCGType type,
1357                     TCGReg a0, TCGReg a1, tcg_target_long a2)
1358{
1359    tcg_out_opc_ori(s, a0, a1, a2);
1360}
1361
1362static const TCGOutOpBinary outop_or = {
1363    .base.static_constraint = C_O1_I2(r, r, rU),
1364    .out_rrr = tgen_or,
1365    .out_rri = tgen_ori,
1366};
1367
1368static void tgen_orc(TCGContext *s, TCGType type,
1369                     TCGReg a0, TCGReg a1, TCGReg a2)
1370{
1371    tcg_out_opc_orn(s, a0, a1, a2);
1372}
1373
1374static const TCGOutOpBinary outop_orc = {
1375    .base.static_constraint = C_O1_I2(r, r, r),
1376    .out_rrr = tgen_orc,
1377};
1378
1379static void tgen_sub(TCGContext *s, TCGType type,
1380                     TCGReg a0, TCGReg a1, TCGReg a2)
1381{
1382    if (type == TCG_TYPE_I32) {
1383        tcg_out_opc_sub_w(s, a0, a1, a2);
1384    } else {
1385        tcg_out_opc_sub_d(s, a0, a1, a2);
1386    }
1387}
1388
1389static const TCGOutOpSubtract outop_sub = {
1390    .base.static_constraint = C_O1_I2(r, r, r),
1391    .out_rrr = tgen_sub,
1392};
1393
1394static void tgen_xor(TCGContext *s, TCGType type,
1395                     TCGReg a0, TCGReg a1, TCGReg a2)
1396{
1397    tcg_out_opc_xor(s, a0, a1, a2);
1398}
1399
1400static void tgen_xori(TCGContext *s, TCGType type,
1401                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1402{
1403    tcg_out_opc_xori(s, a0, a1, a2);
1404}
1405
1406static const TCGOutOpBinary outop_xor = {
1407    .base.static_constraint = C_O1_I2(r, r, rU),
1408    .out_rrr = tgen_xor,
1409    .out_rri = tgen_xori,
1410};
1411
1412
1413static void tcg_out_op(TCGContext *s, TCGOpcode opc, TCGType type,
1414                       const TCGArg args[TCG_MAX_OP_ARGS],
1415                       const int const_args[TCG_MAX_OP_ARGS])
1416{
1417    TCGArg a0 = args[0];
1418    TCGArg a1 = args[1];
1419    TCGArg a2 = args[2];
1420    TCGArg a3 = args[3];
1421    int c2 = const_args[2];
1422
1423    switch (opc) {
1424    case INDEX_op_mb:
1425        tcg_out_mb(s, a0);
1426        break;
1427
1428    case INDEX_op_goto_ptr:
1429        tcg_out_opc_jirl(s, TCG_REG_ZERO, a0, 0);
1430        break;
1431
1432    case INDEX_op_br:
1433        tcg_out_reloc(s, s->code_ptr, R_LOONGARCH_BR_SD10K16, arg_label(a0),
1434                      0);
1435        tcg_out_opc_b(s, 0);
1436        break;
1437
1438    case INDEX_op_brcond_i32:
1439    case INDEX_op_brcond_i64:
1440        tcg_out_brcond(s, a2, a0, a1, arg_label(args[3]));
1441        break;
1442
1443    case INDEX_op_extrh_i64_i32:
1444        tcg_out_opc_srai_d(s, a0, a1, 32);
1445        break;
1446
1447    case INDEX_op_not_i32:
1448    case INDEX_op_not_i64:
1449        tcg_out_opc_nor(s, a0, a1, TCG_REG_ZERO);
1450        break;
1451
1452    case INDEX_op_extract_i32:
1453        if (a2 == 0 && args[3] <= 12) {
1454            tcg_out_opc_andi(s, a0, a1, (1 << args[3]) - 1);
1455        } else {
1456            tcg_out_opc_bstrpick_w(s, a0, a1, a2, a2 + args[3] - 1);
1457        }
1458        break;
1459    case INDEX_op_extract_i64:
1460        if (a2 == 0 && args[3] <= 12) {
1461            tcg_out_opc_andi(s, a0, a1, (1 << args[3]) - 1);
1462        } else {
1463            tcg_out_opc_bstrpick_d(s, a0, a1, a2, a2 + args[3] - 1);
1464        }
1465        break;
1466
1467    case INDEX_op_sextract_i64:
1468        if (a2 + args[3] == 32) {
1469            if (a2 == 0) {
1470                tcg_out_ext32s(s, a0, a1);
1471            } else {
1472                tcg_out_opc_srai_w(s, a0, a1, a2);
1473            }
1474            break;
1475        }
1476        /* FALLTHRU */
1477    case INDEX_op_sextract_i32:
1478        if (a2 == 0 && args[3] == 8) {
1479            tcg_out_ext8s(s, TCG_TYPE_REG, a0, a1);
1480        } else if (a2 == 0 && args[3] == 16) {
1481            tcg_out_ext16s(s, TCG_TYPE_REG, a0, a1);
1482        } else {
1483            g_assert_not_reached();
1484        }
1485        break;
1486
1487    case INDEX_op_deposit_i32:
1488        tcg_out_opc_bstrins_w(s, a0, a2, args[3], args[3] + args[4] - 1);
1489        break;
1490    case INDEX_op_deposit_i64:
1491        tcg_out_opc_bstrins_d(s, a0, a2, args[3], args[3] + args[4] - 1);
1492        break;
1493
1494    case INDEX_op_bswap16_i32:
1495    case INDEX_op_bswap16_i64:
1496        tcg_out_opc_revb_2h(s, a0, a1);
1497        if (a2 & TCG_BSWAP_OS) {
1498            tcg_out_ext16s(s, TCG_TYPE_REG, a0, a0);
1499        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1500            tcg_out_ext16u(s, a0, a0);
1501        }
1502        break;
1503
1504    case INDEX_op_bswap32_i32:
1505        /* All 32-bit values are computed sign-extended in the register.  */
1506        a2 = TCG_BSWAP_OS;
1507        /* fallthrough */
1508    case INDEX_op_bswap32_i64:
1509        tcg_out_opc_revb_2w(s, a0, a1);
1510        if (a2 & TCG_BSWAP_OS) {
1511            tcg_out_ext32s(s, a0, a0);
1512        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1513            tcg_out_ext32u(s, a0, a0);
1514        }
1515        break;
1516
1517    case INDEX_op_bswap64_i64:
1518        tcg_out_opc_revb_d(s, a0, a1);
1519        break;
1520
1521    case INDEX_op_clz_i32:
1522        tcg_out_clzctz(s, OPC_CLZ_W, a0, a1, a2, c2, true);
1523        break;
1524    case INDEX_op_clz_i64:
1525        tcg_out_clzctz(s, OPC_CLZ_D, a0, a1, a2, c2, false);
1526        break;
1527
1528    case INDEX_op_ctz_i32:
1529        tcg_out_clzctz(s, OPC_CTZ_W, a0, a1, a2, c2, true);
1530        break;
1531    case INDEX_op_ctz_i64:
1532        tcg_out_clzctz(s, OPC_CTZ_D, a0, a1, a2, c2, false);
1533        break;
1534
1535    case INDEX_op_shl_i32:
1536        if (c2) {
1537            tcg_out_opc_slli_w(s, a0, a1, a2 & 0x1f);
1538        } else {
1539            tcg_out_opc_sll_w(s, a0, a1, a2);
1540        }
1541        break;
1542    case INDEX_op_shl_i64:
1543        if (c2) {
1544            tcg_out_opc_slli_d(s, a0, a1, a2 & 0x3f);
1545        } else {
1546            tcg_out_opc_sll_d(s, a0, a1, a2);
1547        }
1548        break;
1549
1550    case INDEX_op_shr_i32:
1551        if (c2) {
1552            tcg_out_opc_srli_w(s, a0, a1, a2 & 0x1f);
1553        } else {
1554            tcg_out_opc_srl_w(s, a0, a1, a2);
1555        }
1556        break;
1557    case INDEX_op_shr_i64:
1558        if (c2) {
1559            tcg_out_opc_srli_d(s, a0, a1, a2 & 0x3f);
1560        } else {
1561            tcg_out_opc_srl_d(s, a0, a1, a2);
1562        }
1563        break;
1564
1565    case INDEX_op_sar_i32:
1566        if (c2) {
1567            tcg_out_opc_srai_w(s, a0, a1, a2 & 0x1f);
1568        } else {
1569            tcg_out_opc_sra_w(s, a0, a1, a2);
1570        }
1571        break;
1572    case INDEX_op_sar_i64:
1573        if (c2) {
1574            tcg_out_opc_srai_d(s, a0, a1, a2 & 0x3f);
1575        } else {
1576            tcg_out_opc_sra_d(s, a0, a1, a2);
1577        }
1578        break;
1579
1580    case INDEX_op_rotl_i32:
1581        /* transform into equivalent rotr/rotri */
1582        if (c2) {
1583            tcg_out_opc_rotri_w(s, a0, a1, (32 - a2) & 0x1f);
1584        } else {
1585            tcg_out_opc_sub_w(s, TCG_REG_TMP0, TCG_REG_ZERO, a2);
1586            tcg_out_opc_rotr_w(s, a0, a1, TCG_REG_TMP0);
1587        }
1588        break;
1589    case INDEX_op_rotl_i64:
1590        /* transform into equivalent rotr/rotri */
1591        if (c2) {
1592            tcg_out_opc_rotri_d(s, a0, a1, (64 - a2) & 0x3f);
1593        } else {
1594            tcg_out_opc_sub_w(s, TCG_REG_TMP0, TCG_REG_ZERO, a2);
1595            tcg_out_opc_rotr_d(s, a0, a1, TCG_REG_TMP0);
1596        }
1597        break;
1598
1599    case INDEX_op_rotr_i32:
1600        if (c2) {
1601            tcg_out_opc_rotri_w(s, a0, a1, a2 & 0x1f);
1602        } else {
1603            tcg_out_opc_rotr_w(s, a0, a1, a2);
1604        }
1605        break;
1606    case INDEX_op_rotr_i64:
1607        if (c2) {
1608            tcg_out_opc_rotri_d(s, a0, a1, a2 & 0x3f);
1609        } else {
1610            tcg_out_opc_rotr_d(s, a0, a1, a2);
1611        }
1612        break;
1613
1614    case INDEX_op_neg_i32:
1615        tcg_out_opc_sub_w(s, a0, TCG_REG_ZERO, a1);
1616        break;
1617    case INDEX_op_neg_i64:
1618        tcg_out_opc_sub_d(s, a0, TCG_REG_ZERO, a1);
1619        break;
1620
1621    case INDEX_op_mul_i32:
1622        tcg_out_opc_mul_w(s, a0, a1, a2);
1623        break;
1624    case INDEX_op_mul_i64:
1625        tcg_out_opc_mul_d(s, a0, a1, a2);
1626        break;
1627
1628    case INDEX_op_mulsh_i32:
1629        tcg_out_opc_mulh_w(s, a0, a1, a2);
1630        break;
1631    case INDEX_op_mulsh_i64:
1632        tcg_out_opc_mulh_d(s, a0, a1, a2);
1633        break;
1634
1635    case INDEX_op_muluh_i32:
1636        tcg_out_opc_mulh_wu(s, a0, a1, a2);
1637        break;
1638    case INDEX_op_muluh_i64:
1639        tcg_out_opc_mulh_du(s, a0, a1, a2);
1640        break;
1641
1642    case INDEX_op_div_i32:
1643        tcg_out_opc_div_w(s, a0, a1, a2);
1644        break;
1645    case INDEX_op_div_i64:
1646        tcg_out_opc_div_d(s, a0, a1, a2);
1647        break;
1648
1649    case INDEX_op_divu_i32:
1650        tcg_out_opc_div_wu(s, a0, a1, a2);
1651        break;
1652    case INDEX_op_divu_i64:
1653        tcg_out_opc_div_du(s, a0, a1, a2);
1654        break;
1655
1656    case INDEX_op_rem_i32:
1657        tcg_out_opc_mod_w(s, a0, a1, a2);
1658        break;
1659    case INDEX_op_rem_i64:
1660        tcg_out_opc_mod_d(s, a0, a1, a2);
1661        break;
1662
1663    case INDEX_op_remu_i32:
1664        tcg_out_opc_mod_wu(s, a0, a1, a2);
1665        break;
1666    case INDEX_op_remu_i64:
1667        tcg_out_opc_mod_du(s, a0, a1, a2);
1668        break;
1669
1670    case INDEX_op_setcond_i32:
1671    case INDEX_op_setcond_i64:
1672        tcg_out_setcond(s, args[3], a0, a1, a2, c2);
1673        break;
1674
1675    case INDEX_op_movcond_i32:
1676    case INDEX_op_movcond_i64:
1677        tcg_out_movcond(s, args[5], a0, a1, a2, c2, args[3], args[4]);
1678        break;
1679
1680    case INDEX_op_ld8s_i32:
1681    case INDEX_op_ld8s_i64:
1682        tcg_out_ldst(s, OPC_LD_B, a0, a1, a2);
1683        break;
1684    case INDEX_op_ld8u_i32:
1685    case INDEX_op_ld8u_i64:
1686        tcg_out_ldst(s, OPC_LD_BU, a0, a1, a2);
1687        break;
1688    case INDEX_op_ld16s_i32:
1689    case INDEX_op_ld16s_i64:
1690        tcg_out_ldst(s, OPC_LD_H, a0, a1, a2);
1691        break;
1692    case INDEX_op_ld16u_i32:
1693    case INDEX_op_ld16u_i64:
1694        tcg_out_ldst(s, OPC_LD_HU, a0, a1, a2);
1695        break;
1696    case INDEX_op_ld_i32:
1697    case INDEX_op_ld32s_i64:
1698        tcg_out_ldst(s, OPC_LD_W, a0, a1, a2);
1699        break;
1700    case INDEX_op_ld32u_i64:
1701        tcg_out_ldst(s, OPC_LD_WU, a0, a1, a2);
1702        break;
1703    case INDEX_op_ld_i64:
1704        tcg_out_ldst(s, OPC_LD_D, a0, a1, a2);
1705        break;
1706
1707    case INDEX_op_st8_i32:
1708    case INDEX_op_st8_i64:
1709        tcg_out_ldst(s, OPC_ST_B, a0, a1, a2);
1710        break;
1711    case INDEX_op_st16_i32:
1712    case INDEX_op_st16_i64:
1713        tcg_out_ldst(s, OPC_ST_H, a0, a1, a2);
1714        break;
1715    case INDEX_op_st_i32:
1716    case INDEX_op_st32_i64:
1717        tcg_out_ldst(s, OPC_ST_W, a0, a1, a2);
1718        break;
1719    case INDEX_op_st_i64:
1720        tcg_out_ldst(s, OPC_ST_D, a0, a1, a2);
1721        break;
1722
1723    case INDEX_op_qemu_ld_i32:
1724        tcg_out_qemu_ld(s, a0, a1, a2, TCG_TYPE_I32);
1725        break;
1726    case INDEX_op_qemu_ld_i64:
1727        tcg_out_qemu_ld(s, a0, a1, a2, TCG_TYPE_I64);
1728        break;
1729    case INDEX_op_qemu_ld_i128:
1730        tcg_out_qemu_ldst_i128(s, a0, a1, a2, a3, true);
1731        break;
1732    case INDEX_op_qemu_st_i32:
1733        tcg_out_qemu_st(s, a0, a1, a2, TCG_TYPE_I32);
1734        break;
1735    case INDEX_op_qemu_st_i64:
1736        tcg_out_qemu_st(s, a0, a1, a2, TCG_TYPE_I64);
1737        break;
1738    case INDEX_op_qemu_st_i128:
1739        tcg_out_qemu_ldst_i128(s, a0, a1, a2, a3, false);
1740        break;
1741
1742    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
1743    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
1744    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
1745    case INDEX_op_ext_i32_i64:  /* Always emitted via tcg_reg_alloc_op.  */
1746    case INDEX_op_extu_i32_i64:
1747    case INDEX_op_extrl_i64_i32:
1748    default:
1749        g_assert_not_reached();
1750    }
1751}
1752
1753static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
1754                            TCGReg rd, TCGReg rs)
1755{
1756    static const LoongArchInsn repl_insn[2][4] = {
1757        { OPC_VREPLGR2VR_B, OPC_VREPLGR2VR_H,
1758          OPC_VREPLGR2VR_W, OPC_VREPLGR2VR_D },
1759        { OPC_XVREPLGR2VR_B, OPC_XVREPLGR2VR_H,
1760          OPC_XVREPLGR2VR_W, OPC_XVREPLGR2VR_D },
1761    };
1762    bool lasx = type == TCG_TYPE_V256;
1763
1764    tcg_debug_assert(vece <= MO_64);
1765    tcg_out32(s, encode_vdj_insn(repl_insn[lasx][vece], rd, rs));
1766    return true;
1767}
1768
1769static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
1770                             TCGReg r, TCGReg base, intptr_t offset)
1771{
1772    bool lasx = type == TCG_TYPE_V256;
1773
1774    /* Handle imm overflow and division (vldrepl.d imm is divided by 8). */
1775    if (offset < -0x800 || offset > 0x7ff ||
1776        (offset & ((1 << vece) - 1)) != 0) {
1777        tcg_out_addi(s, TCG_TYPE_I64, TCG_REG_TMP0, base, offset);
1778        base = TCG_REG_TMP0;
1779        offset = 0;
1780    }
1781    offset >>= vece;
1782
1783    switch (vece) {
1784    case MO_8:
1785        if (lasx) {
1786            tcg_out_opc_xvldrepl_b(s, r, base, offset);
1787        } else {
1788            tcg_out_opc_vldrepl_b(s, r, base, offset);
1789        }
1790        break;
1791    case MO_16:
1792        if (lasx) {
1793            tcg_out_opc_xvldrepl_h(s, r, base, offset);
1794        } else {
1795            tcg_out_opc_vldrepl_h(s, r, base, offset);
1796        }
1797        break;
1798    case MO_32:
1799        if (lasx) {
1800            tcg_out_opc_xvldrepl_w(s, r, base, offset);
1801        } else {
1802            tcg_out_opc_vldrepl_w(s, r, base, offset);
1803        }
1804        break;
1805    case MO_64:
1806        if (lasx) {
1807            tcg_out_opc_xvldrepl_d(s, r, base, offset);
1808        } else {
1809            tcg_out_opc_vldrepl_d(s, r, base, offset);
1810        }
1811        break;
1812    default:
1813        g_assert_not_reached();
1814    }
1815    return true;
1816}
1817
1818static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
1819                             TCGReg rd, int64_t v64)
1820{
1821    /* Try vldi if imm can fit */
1822    int64_t value = sextract64(v64, 0, 8 << vece);
1823    if (-0x200 <= value && value <= 0x1FF) {
1824        uint32_t imm = (vece << 10) | ((uint32_t)v64 & 0x3FF);
1825
1826        if (type == TCG_TYPE_V256) {
1827            tcg_out_opc_xvldi(s, rd, imm);
1828        } else {
1829            tcg_out_opc_vldi(s, rd, imm);
1830        }
1831        return;
1832    }
1833
1834    /* TODO: vldi patterns when imm 12 is set */
1835
1836    tcg_out_movi(s, TCG_TYPE_I64, TCG_REG_TMP0, value);
1837    tcg_out_dup_vec(s, type, vece, rd, TCG_REG_TMP0);
1838}
1839
1840static void tcg_out_addsub_vec(TCGContext *s, bool lasx, unsigned vece,
1841                               TCGArg a0, TCGArg a1, TCGArg a2,
1842                               bool a2_is_const, bool is_add)
1843{
1844    static const LoongArchInsn add_vec_insn[2][4] = {
1845        { OPC_VADD_B, OPC_VADD_H, OPC_VADD_W, OPC_VADD_D },
1846        { OPC_XVADD_B, OPC_XVADD_H, OPC_XVADD_W, OPC_XVADD_D },
1847    };
1848    static const LoongArchInsn add_vec_imm_insn[2][4] = {
1849        { OPC_VADDI_BU, OPC_VADDI_HU, OPC_VADDI_WU, OPC_VADDI_DU },
1850        { OPC_XVADDI_BU, OPC_XVADDI_HU, OPC_XVADDI_WU, OPC_XVADDI_DU },
1851    };
1852    static const LoongArchInsn sub_vec_insn[2][4] = {
1853        { OPC_VSUB_B, OPC_VSUB_H, OPC_VSUB_W, OPC_VSUB_D },
1854        { OPC_XVSUB_B, OPC_XVSUB_H, OPC_XVSUB_W, OPC_XVSUB_D },
1855    };
1856    static const LoongArchInsn sub_vec_imm_insn[2][4] = {
1857        { OPC_VSUBI_BU, OPC_VSUBI_HU, OPC_VSUBI_WU, OPC_VSUBI_DU },
1858        { OPC_XVSUBI_BU, OPC_XVSUBI_HU, OPC_XVSUBI_WU, OPC_XVSUBI_DU },
1859    };
1860    LoongArchInsn insn;
1861
1862    if (a2_is_const) {
1863        int64_t value = sextract64(a2, 0, 8 << vece);
1864
1865        if (!is_add) {
1866            value = -value;
1867        }
1868        if (value < 0) {
1869            insn = sub_vec_imm_insn[lasx][vece];
1870            value = -value;
1871        } else {
1872            insn = add_vec_imm_insn[lasx][vece];
1873        }
1874
1875        /* Constraint TCG_CT_CONST_VADD ensures validity. */
1876        tcg_debug_assert(0 <= value && value <= 0x1f);
1877
1878        tcg_out32(s, encode_vdvjuk5_insn(insn, a0, a1, value));
1879        return;
1880    }
1881
1882    if (is_add) {
1883        insn = add_vec_insn[lasx][vece];
1884    } else {
1885        insn = sub_vec_insn[lasx][vece];
1886    }
1887    tcg_out32(s, encode_vdvjvk_insn(insn, a0, a1, a2));
1888}
1889
1890static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
1891                           unsigned vecl, unsigned vece,
1892                           const TCGArg args[TCG_MAX_OP_ARGS],
1893                           const int const_args[TCG_MAX_OP_ARGS])
1894{
1895    TCGType type = vecl + TCG_TYPE_V64;
1896    bool lasx = type == TCG_TYPE_V256;
1897    TCGArg a0, a1, a2, a3;
1898    LoongArchInsn insn;
1899
1900    static const LoongArchInsn cmp_vec_insn[16][2][4] = {
1901        [TCG_COND_EQ] = {
1902            { OPC_VSEQ_B, OPC_VSEQ_H, OPC_VSEQ_W, OPC_VSEQ_D },
1903            { OPC_XVSEQ_B, OPC_XVSEQ_H, OPC_XVSEQ_W, OPC_XVSEQ_D },
1904        },
1905        [TCG_COND_LE] = {
1906            { OPC_VSLE_B, OPC_VSLE_H, OPC_VSLE_W, OPC_VSLE_D },
1907            { OPC_XVSLE_B, OPC_XVSLE_H, OPC_XVSLE_W, OPC_XVSLE_D },
1908        },
1909        [TCG_COND_LEU] = {
1910            { OPC_VSLE_BU, OPC_VSLE_HU, OPC_VSLE_WU, OPC_VSLE_DU },
1911            { OPC_XVSLE_BU, OPC_XVSLE_HU, OPC_XVSLE_WU, OPC_XVSLE_DU },
1912        },
1913        [TCG_COND_LT] = {
1914            { OPC_VSLT_B, OPC_VSLT_H, OPC_VSLT_W, OPC_VSLT_D },
1915            { OPC_XVSLT_B, OPC_XVSLT_H, OPC_XVSLT_W, OPC_XVSLT_D },
1916        },
1917        [TCG_COND_LTU] = {
1918            { OPC_VSLT_BU, OPC_VSLT_HU, OPC_VSLT_WU, OPC_VSLT_DU },
1919            { OPC_XVSLT_BU, OPC_XVSLT_HU, OPC_XVSLT_WU, OPC_XVSLT_DU },
1920        }
1921    };
1922    static const LoongArchInsn cmp_vec_imm_insn[16][2][4] = {
1923        [TCG_COND_EQ] = {
1924            { OPC_VSEQI_B, OPC_VSEQI_H, OPC_VSEQI_W, OPC_VSEQI_D },
1925            { OPC_XVSEQI_B, OPC_XVSEQI_H, OPC_XVSEQI_W, OPC_XVSEQI_D },
1926        },
1927        [TCG_COND_LE] = {
1928            { OPC_VSLEI_B, OPC_VSLEI_H, OPC_VSLEI_W, OPC_VSLEI_D },
1929            { OPC_XVSLEI_B, OPC_XVSLEI_H, OPC_XVSLEI_W, OPC_XVSLEI_D },
1930        },
1931        [TCG_COND_LEU] = {
1932            { OPC_VSLEI_BU, OPC_VSLEI_HU, OPC_VSLEI_WU, OPC_VSLEI_DU },
1933            { OPC_XVSLEI_BU, OPC_XVSLEI_HU, OPC_XVSLEI_WU, OPC_XVSLEI_DU },
1934        },
1935        [TCG_COND_LT] = {
1936            { OPC_VSLTI_B, OPC_VSLTI_H, OPC_VSLTI_W, OPC_VSLTI_D },
1937            { OPC_XVSLTI_B, OPC_XVSLTI_H, OPC_XVSLTI_W, OPC_XVSLTI_D },
1938        },
1939        [TCG_COND_LTU] = {
1940            { OPC_VSLTI_BU, OPC_VSLTI_HU, OPC_VSLTI_WU, OPC_VSLTI_DU },
1941            { OPC_XVSLTI_BU, OPC_XVSLTI_HU, OPC_XVSLTI_WU, OPC_XVSLTI_DU },
1942        }
1943    };
1944    static const LoongArchInsn neg_vec_insn[2][4] = {
1945        { OPC_VNEG_B, OPC_VNEG_H, OPC_VNEG_W, OPC_VNEG_D },
1946        { OPC_XVNEG_B, OPC_XVNEG_H, OPC_XVNEG_W, OPC_XVNEG_D },
1947    };
1948    static const LoongArchInsn mul_vec_insn[2][4] = {
1949        { OPC_VMUL_B, OPC_VMUL_H, OPC_VMUL_W, OPC_VMUL_D },
1950        { OPC_XVMUL_B, OPC_XVMUL_H, OPC_XVMUL_W, OPC_XVMUL_D },
1951    };
1952    static const LoongArchInsn smin_vec_insn[2][4] = {
1953        { OPC_VMIN_B, OPC_VMIN_H, OPC_VMIN_W, OPC_VMIN_D },
1954        { OPC_XVMIN_B, OPC_XVMIN_H, OPC_XVMIN_W, OPC_XVMIN_D },
1955    };
1956    static const LoongArchInsn umin_vec_insn[2][4] = {
1957        { OPC_VMIN_BU, OPC_VMIN_HU, OPC_VMIN_WU, OPC_VMIN_DU },
1958        { OPC_XVMIN_BU, OPC_XVMIN_HU, OPC_XVMIN_WU, OPC_XVMIN_DU },
1959    };
1960    static const LoongArchInsn smax_vec_insn[2][4] = {
1961        { OPC_VMAX_B, OPC_VMAX_H, OPC_VMAX_W, OPC_VMAX_D },
1962        { OPC_XVMAX_B, OPC_XVMAX_H, OPC_XVMAX_W, OPC_XVMAX_D },
1963    };
1964    static const LoongArchInsn umax_vec_insn[2][4] = {
1965        { OPC_VMAX_BU, OPC_VMAX_HU, OPC_VMAX_WU, OPC_VMAX_DU },
1966        { OPC_XVMAX_BU, OPC_XVMAX_HU, OPC_XVMAX_WU, OPC_XVMAX_DU },
1967    };
1968    static const LoongArchInsn ssadd_vec_insn[2][4] = {
1969        { OPC_VSADD_B, OPC_VSADD_H, OPC_VSADD_W, OPC_VSADD_D },
1970        { OPC_XVSADD_B, OPC_XVSADD_H, OPC_XVSADD_W, OPC_XVSADD_D },
1971    };
1972    static const LoongArchInsn usadd_vec_insn[2][4] = {
1973        { OPC_VSADD_BU, OPC_VSADD_HU, OPC_VSADD_WU, OPC_VSADD_DU },
1974        { OPC_XVSADD_BU, OPC_XVSADD_HU, OPC_XVSADD_WU, OPC_XVSADD_DU },
1975    };
1976    static const LoongArchInsn sssub_vec_insn[2][4] = {
1977        { OPC_VSSUB_B, OPC_VSSUB_H, OPC_VSSUB_W, OPC_VSSUB_D },
1978        { OPC_XVSSUB_B, OPC_XVSSUB_H, OPC_XVSSUB_W, OPC_XVSSUB_D },
1979    };
1980    static const LoongArchInsn ussub_vec_insn[2][4] = {
1981        { OPC_VSSUB_BU, OPC_VSSUB_HU, OPC_VSSUB_WU, OPC_VSSUB_DU },
1982        { OPC_XVSSUB_BU, OPC_XVSSUB_HU, OPC_XVSSUB_WU, OPC_XVSSUB_DU },
1983    };
1984    static const LoongArchInsn shlv_vec_insn[2][4] = {
1985        { OPC_VSLL_B, OPC_VSLL_H, OPC_VSLL_W, OPC_VSLL_D },
1986        { OPC_XVSLL_B, OPC_XVSLL_H, OPC_XVSLL_W, OPC_XVSLL_D },
1987    };
1988    static const LoongArchInsn shrv_vec_insn[2][4] = {
1989        { OPC_VSRL_B, OPC_VSRL_H, OPC_VSRL_W, OPC_VSRL_D },
1990        { OPC_XVSRL_B, OPC_XVSRL_H, OPC_XVSRL_W, OPC_XVSRL_D },
1991    };
1992    static const LoongArchInsn sarv_vec_insn[2][4] = {
1993        { OPC_VSRA_B, OPC_VSRA_H, OPC_VSRA_W, OPC_VSRA_D },
1994        { OPC_XVSRA_B, OPC_XVSRA_H, OPC_XVSRA_W, OPC_XVSRA_D },
1995    };
1996    static const LoongArchInsn shli_vec_insn[2][4] = {
1997        { OPC_VSLLI_B, OPC_VSLLI_H, OPC_VSLLI_W, OPC_VSLLI_D },
1998        { OPC_XVSLLI_B, OPC_XVSLLI_H, OPC_XVSLLI_W, OPC_XVSLLI_D },
1999    };
2000    static const LoongArchInsn shri_vec_insn[2][4] = {
2001        { OPC_VSRLI_B, OPC_VSRLI_H, OPC_VSRLI_W, OPC_VSRLI_D },
2002        { OPC_XVSRLI_B, OPC_XVSRLI_H, OPC_XVSRLI_W, OPC_XVSRLI_D },
2003    };
2004    static const LoongArchInsn sari_vec_insn[2][4] = {
2005        { OPC_VSRAI_B, OPC_VSRAI_H, OPC_VSRAI_W, OPC_VSRAI_D },
2006        { OPC_XVSRAI_B, OPC_XVSRAI_H, OPC_XVSRAI_W, OPC_XVSRAI_D },
2007    };
2008    static const LoongArchInsn rotrv_vec_insn[2][4] = {
2009        { OPC_VROTR_B, OPC_VROTR_H, OPC_VROTR_W, OPC_VROTR_D },
2010        { OPC_XVROTR_B, OPC_XVROTR_H, OPC_XVROTR_W, OPC_XVROTR_D },
2011    };
2012    static const LoongArchInsn rotri_vec_insn[2][4] = {
2013        { OPC_VROTRI_B, OPC_VROTRI_H, OPC_VROTRI_W, OPC_VROTRI_D },
2014        { OPC_XVROTRI_B, OPC_XVROTRI_H, OPC_XVROTRI_W, OPC_XVROTRI_D },
2015    };
2016
2017    a0 = args[0];
2018    a1 = args[1];
2019    a2 = args[2];
2020    a3 = args[3];
2021
2022    switch (opc) {
2023    case INDEX_op_st_vec:
2024        tcg_out_st(s, type, a0, a1, a2);
2025        break;
2026    case INDEX_op_ld_vec:
2027        tcg_out_ld(s, type, a0, a1, a2);
2028        break;
2029    case INDEX_op_and_vec:
2030        insn = lasx ? OPC_XVAND_V : OPC_VAND_V;
2031        goto vdvjvk;
2032    case INDEX_op_andc_vec:
2033        /*
2034         * vandn vd, vj, vk: vd = vk & ~vj
2035         * andc_vec vd, vj, vk: vd = vj & ~vk
2036         * vj and vk are swapped
2037         */
2038        a1 = a2;
2039        a2 = args[1];
2040        insn = lasx ? OPC_XVANDN_V : OPC_VANDN_V;
2041        goto vdvjvk;
2042    case INDEX_op_or_vec:
2043        insn = lasx ? OPC_XVOR_V : OPC_VOR_V;
2044        goto vdvjvk;
2045    case INDEX_op_orc_vec:
2046        insn = lasx ? OPC_XVORN_V : OPC_VORN_V;
2047        goto vdvjvk;
2048    case INDEX_op_xor_vec:
2049        insn = lasx ? OPC_XVXOR_V : OPC_VXOR_V;
2050        goto vdvjvk;
2051    case INDEX_op_not_vec:
2052        a2 = a1;
2053        /* fall through */
2054    case INDEX_op_nor_vec:
2055        insn = lasx ? OPC_XVNOR_V : OPC_VNOR_V;
2056        goto vdvjvk;
2057    case INDEX_op_cmp_vec:
2058        {
2059            TCGCond cond = args[3];
2060
2061            if (const_args[2]) {
2062                /*
2063                 * cmp_vec dest, src, value
2064                 * Try vseqi/vslei/vslti
2065                 */
2066                int64_t value = sextract64(a2, 0, 8 << vece);
2067                switch (cond) {
2068                case TCG_COND_EQ:
2069                case TCG_COND_LE:
2070                case TCG_COND_LT:
2071                    insn = cmp_vec_imm_insn[cond][lasx][vece];
2072                    tcg_out32(s, encode_vdvjsk5_insn(insn, a0, a1, value));
2073                    break;
2074                case TCG_COND_LEU:
2075                case TCG_COND_LTU:
2076                    insn = cmp_vec_imm_insn[cond][lasx][vece];
2077                    tcg_out32(s, encode_vdvjuk5_insn(insn, a0, a1, value));
2078                    break;
2079                default:
2080                    g_assert_not_reached();
2081                }
2082                break;
2083            }
2084
2085            insn = cmp_vec_insn[cond][lasx][vece];
2086            if (insn == 0) {
2087                TCGArg t;
2088                t = a1, a1 = a2, a2 = t;
2089                cond = tcg_swap_cond(cond);
2090                insn = cmp_vec_insn[cond][lasx][vece];
2091                tcg_debug_assert(insn != 0);
2092            }
2093        }
2094        goto vdvjvk;
2095    case INDEX_op_add_vec:
2096        tcg_out_addsub_vec(s, lasx, vece, a0, a1, a2, const_args[2], true);
2097        break;
2098    case INDEX_op_sub_vec:
2099        tcg_out_addsub_vec(s, lasx, vece, a0, a1, a2, const_args[2], false);
2100        break;
2101    case INDEX_op_neg_vec:
2102        tcg_out32(s, encode_vdvj_insn(neg_vec_insn[lasx][vece], a0, a1));
2103        break;
2104    case INDEX_op_mul_vec:
2105        insn = mul_vec_insn[lasx][vece];
2106        goto vdvjvk;
2107    case INDEX_op_smin_vec:
2108        insn = smin_vec_insn[lasx][vece];
2109        goto vdvjvk;
2110    case INDEX_op_smax_vec:
2111        insn = smax_vec_insn[lasx][vece];
2112        goto vdvjvk;
2113    case INDEX_op_umin_vec:
2114        insn = umin_vec_insn[lasx][vece];
2115        goto vdvjvk;
2116    case INDEX_op_umax_vec:
2117        insn = umax_vec_insn[lasx][vece];
2118        goto vdvjvk;
2119    case INDEX_op_ssadd_vec:
2120        insn = ssadd_vec_insn[lasx][vece];
2121        goto vdvjvk;
2122    case INDEX_op_usadd_vec:
2123        insn = usadd_vec_insn[lasx][vece];
2124        goto vdvjvk;
2125    case INDEX_op_sssub_vec:
2126        insn = sssub_vec_insn[lasx][vece];
2127        goto vdvjvk;
2128    case INDEX_op_ussub_vec:
2129        insn = ussub_vec_insn[lasx][vece];
2130        goto vdvjvk;
2131    case INDEX_op_shlv_vec:
2132        insn = shlv_vec_insn[lasx][vece];
2133        goto vdvjvk;
2134    case INDEX_op_shrv_vec:
2135        insn = shrv_vec_insn[lasx][vece];
2136        goto vdvjvk;
2137    case INDEX_op_sarv_vec:
2138        insn = sarv_vec_insn[lasx][vece];
2139        goto vdvjvk;
2140    case INDEX_op_rotlv_vec:
2141        /* rotlv_vec a1, a2 = rotrv_vec a1, -a2 */
2142        tcg_out32(s, encode_vdvj_insn(neg_vec_insn[lasx][vece],
2143                                      TCG_VEC_TMP0, a2));
2144        a2 = TCG_VEC_TMP0;
2145        /* fall through */
2146    case INDEX_op_rotrv_vec:
2147        insn = rotrv_vec_insn[lasx][vece];
2148        goto vdvjvk;
2149    case INDEX_op_shli_vec:
2150        insn = shli_vec_insn[lasx][vece];
2151        goto vdvjukN;
2152    case INDEX_op_shri_vec:
2153        insn = shri_vec_insn[lasx][vece];
2154        goto vdvjukN;
2155    case INDEX_op_sari_vec:
2156        insn = sari_vec_insn[lasx][vece];
2157        goto vdvjukN;
2158    case INDEX_op_rotli_vec:
2159        /* rotli_vec a1, a2 = rotri_vec a1, -a2 */
2160        a2 = extract32(-a2, 0, 3 + vece);
2161        insn = rotri_vec_insn[lasx][vece];
2162        goto vdvjukN;
2163    case INDEX_op_bitsel_vec:
2164        /* vbitsel vd, vj, vk, va = bitsel_vec vd, va, vk, vj */
2165        if (lasx) {
2166            tcg_out_opc_xvbitsel_v(s, a0, a3, a2, a1);
2167        } else {
2168            tcg_out_opc_vbitsel_v(s, a0, a3, a2, a1);
2169        }
2170        break;
2171    case INDEX_op_dupm_vec:
2172        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
2173        break;
2174    default:
2175        g_assert_not_reached();
2176    vdvjvk:
2177        tcg_out32(s, encode_vdvjvk_insn(insn, a0, a1, a2));
2178        break;
2179    vdvjukN:
2180        switch (vece) {
2181        case MO_8:
2182            tcg_out32(s, encode_vdvjuk3_insn(insn, a0, a1, a2));
2183            break;
2184        case MO_16:
2185            tcg_out32(s, encode_vdvjuk4_insn(insn, a0, a1, a2));
2186            break;
2187        case MO_32:
2188            tcg_out32(s, encode_vdvjuk5_insn(insn, a0, a1, a2));
2189            break;
2190        case MO_64:
2191            tcg_out32(s, encode_vdvjuk6_insn(insn, a0, a1, a2));
2192            break;
2193        default:
2194            g_assert_not_reached();
2195        }
2196        break;
2197    }
2198}
2199
2200int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
2201{
2202    switch (opc) {
2203    case INDEX_op_ld_vec:
2204    case INDEX_op_st_vec:
2205    case INDEX_op_dup_vec:
2206    case INDEX_op_dupm_vec:
2207    case INDEX_op_cmp_vec:
2208    case INDEX_op_add_vec:
2209    case INDEX_op_sub_vec:
2210    case INDEX_op_and_vec:
2211    case INDEX_op_andc_vec:
2212    case INDEX_op_or_vec:
2213    case INDEX_op_orc_vec:
2214    case INDEX_op_xor_vec:
2215    case INDEX_op_nor_vec:
2216    case INDEX_op_not_vec:
2217    case INDEX_op_neg_vec:
2218    case INDEX_op_mul_vec:
2219    case INDEX_op_smin_vec:
2220    case INDEX_op_smax_vec:
2221    case INDEX_op_umin_vec:
2222    case INDEX_op_umax_vec:
2223    case INDEX_op_ssadd_vec:
2224    case INDEX_op_usadd_vec:
2225    case INDEX_op_sssub_vec:
2226    case INDEX_op_ussub_vec:
2227    case INDEX_op_shlv_vec:
2228    case INDEX_op_shrv_vec:
2229    case INDEX_op_sarv_vec:
2230    case INDEX_op_bitsel_vec:
2231        return 1;
2232    default:
2233        return 0;
2234    }
2235}
2236
2237void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
2238                       TCGArg a0, ...)
2239{
2240    g_assert_not_reached();
2241}
2242
2243static TCGConstraintSetIndex
2244tcg_target_op_def(TCGOpcode op, TCGType type, unsigned flags)
2245{
2246    switch (op) {
2247    case INDEX_op_goto_ptr:
2248        return C_O0_I1(r);
2249
2250    case INDEX_op_st8_i32:
2251    case INDEX_op_st8_i64:
2252    case INDEX_op_st16_i32:
2253    case INDEX_op_st16_i64:
2254    case INDEX_op_st32_i64:
2255    case INDEX_op_st_i32:
2256    case INDEX_op_st_i64:
2257    case INDEX_op_qemu_st_i32:
2258    case INDEX_op_qemu_st_i64:
2259        return C_O0_I2(rz, r);
2260
2261    case INDEX_op_qemu_ld_i128:
2262        return C_N2_I1(r, r, r);
2263
2264    case INDEX_op_qemu_st_i128:
2265        return C_O0_I3(r, r, r);
2266
2267    case INDEX_op_brcond_i32:
2268    case INDEX_op_brcond_i64:
2269        return C_O0_I2(rz, rz);
2270
2271    case INDEX_op_extu_i32_i64:
2272    case INDEX_op_extrl_i64_i32:
2273    case INDEX_op_extrh_i64_i32:
2274    case INDEX_op_ext_i32_i64:
2275    case INDEX_op_neg_i32:
2276    case INDEX_op_neg_i64:
2277    case INDEX_op_not_i32:
2278    case INDEX_op_not_i64:
2279    case INDEX_op_extract_i32:
2280    case INDEX_op_extract_i64:
2281    case INDEX_op_sextract_i32:
2282    case INDEX_op_sextract_i64:
2283    case INDEX_op_bswap16_i32:
2284    case INDEX_op_bswap16_i64:
2285    case INDEX_op_bswap32_i32:
2286    case INDEX_op_bswap32_i64:
2287    case INDEX_op_bswap64_i64:
2288    case INDEX_op_ld8s_i32:
2289    case INDEX_op_ld8s_i64:
2290    case INDEX_op_ld8u_i32:
2291    case INDEX_op_ld8u_i64:
2292    case INDEX_op_ld16s_i32:
2293    case INDEX_op_ld16s_i64:
2294    case INDEX_op_ld16u_i32:
2295    case INDEX_op_ld16u_i64:
2296    case INDEX_op_ld32s_i64:
2297    case INDEX_op_ld32u_i64:
2298    case INDEX_op_ld_i32:
2299    case INDEX_op_ld_i64:
2300    case INDEX_op_qemu_ld_i32:
2301    case INDEX_op_qemu_ld_i64:
2302        return C_O1_I1(r, r);
2303
2304    case INDEX_op_shl_i32:
2305    case INDEX_op_shl_i64:
2306    case INDEX_op_shr_i32:
2307    case INDEX_op_shr_i64:
2308    case INDEX_op_sar_i32:
2309    case INDEX_op_sar_i64:
2310    case INDEX_op_rotl_i32:
2311    case INDEX_op_rotl_i64:
2312    case INDEX_op_rotr_i32:
2313    case INDEX_op_rotr_i64:
2314        return C_O1_I2(r, r, ri);
2315
2316    case INDEX_op_clz_i32:
2317    case INDEX_op_clz_i64:
2318    case INDEX_op_ctz_i32:
2319    case INDEX_op_ctz_i64:
2320        return C_O1_I2(r, r, rW);
2321
2322    case INDEX_op_deposit_i32:
2323    case INDEX_op_deposit_i64:
2324        /* Must deposit into the same register as input */
2325        return C_O1_I2(r, 0, rz);
2326
2327    case INDEX_op_setcond_i32:
2328        return C_O1_I2(r, rz, ri);
2329    case INDEX_op_setcond_i64:
2330        return C_O1_I2(r, rz, rJ);
2331
2332    case INDEX_op_mul_i32:
2333    case INDEX_op_mul_i64:
2334    case INDEX_op_mulsh_i32:
2335    case INDEX_op_mulsh_i64:
2336    case INDEX_op_muluh_i32:
2337    case INDEX_op_muluh_i64:
2338    case INDEX_op_div_i32:
2339    case INDEX_op_div_i64:
2340    case INDEX_op_divu_i32:
2341    case INDEX_op_divu_i64:
2342    case INDEX_op_rem_i32:
2343    case INDEX_op_rem_i64:
2344    case INDEX_op_remu_i32:
2345    case INDEX_op_remu_i64:
2346        return C_O1_I2(r, rz, rz);
2347
2348    case INDEX_op_movcond_i32:
2349    case INDEX_op_movcond_i64:
2350        return C_O1_I4(r, rz, rJ, rz, rz);
2351
2352    case INDEX_op_ld_vec:
2353    case INDEX_op_dupm_vec:
2354    case INDEX_op_dup_vec:
2355        return C_O1_I1(w, r);
2356
2357    case INDEX_op_st_vec:
2358        return C_O0_I2(w, r);
2359
2360    case INDEX_op_cmp_vec:
2361        return C_O1_I2(w, w, wM);
2362
2363    case INDEX_op_add_vec:
2364    case INDEX_op_sub_vec:
2365        return C_O1_I2(w, w, wA);
2366
2367    case INDEX_op_and_vec:
2368    case INDEX_op_andc_vec:
2369    case INDEX_op_or_vec:
2370    case INDEX_op_orc_vec:
2371    case INDEX_op_xor_vec:
2372    case INDEX_op_nor_vec:
2373    case INDEX_op_mul_vec:
2374    case INDEX_op_smin_vec:
2375    case INDEX_op_smax_vec:
2376    case INDEX_op_umin_vec:
2377    case INDEX_op_umax_vec:
2378    case INDEX_op_ssadd_vec:
2379    case INDEX_op_usadd_vec:
2380    case INDEX_op_sssub_vec:
2381    case INDEX_op_ussub_vec:
2382    case INDEX_op_shlv_vec:
2383    case INDEX_op_shrv_vec:
2384    case INDEX_op_sarv_vec:
2385    case INDEX_op_rotrv_vec:
2386    case INDEX_op_rotlv_vec:
2387        return C_O1_I2(w, w, w);
2388
2389    case INDEX_op_not_vec:
2390    case INDEX_op_neg_vec:
2391    case INDEX_op_shli_vec:
2392    case INDEX_op_shri_vec:
2393    case INDEX_op_sari_vec:
2394    case INDEX_op_rotli_vec:
2395        return C_O1_I1(w, w);
2396
2397    case INDEX_op_bitsel_vec:
2398        return C_O1_I3(w, w, w, w);
2399
2400    default:
2401        return C_NotImplemented;
2402    }
2403}
2404
2405static const int tcg_target_callee_save_regs[] = {
2406    TCG_REG_S0,     /* used for the global env (TCG_AREG0) */
2407    TCG_REG_S1,
2408    TCG_REG_S2,
2409    TCG_REG_S3,
2410    TCG_REG_S4,
2411    TCG_REG_S5,
2412    TCG_REG_S6,
2413    TCG_REG_S7,
2414    TCG_REG_S8,
2415    TCG_REG_S9,
2416    TCG_REG_RA,     /* should be last for ABI compliance */
2417};
2418
2419/* Stack frame parameters.  */
2420#define REG_SIZE   (TCG_TARGET_REG_BITS / 8)
2421#define SAVE_SIZE  ((int)ARRAY_SIZE(tcg_target_callee_save_regs) * REG_SIZE)
2422#define TEMP_SIZE  (CPU_TEMP_BUF_NLONGS * (int)sizeof(long))
2423#define FRAME_SIZE ((TCG_STATIC_CALL_ARGS_SIZE + TEMP_SIZE + SAVE_SIZE \
2424                     + TCG_TARGET_STACK_ALIGN - 1) \
2425                    & -TCG_TARGET_STACK_ALIGN)
2426#define SAVE_OFS   (TCG_STATIC_CALL_ARGS_SIZE + TEMP_SIZE)
2427
2428/* We're expecting to be able to use an immediate for frame allocation.  */
2429QEMU_BUILD_BUG_ON(FRAME_SIZE > 0x7ff);
2430
2431/* Generate global QEMU prologue and epilogue code */
2432static void tcg_target_qemu_prologue(TCGContext *s)
2433{
2434    int i;
2435
2436    tcg_set_frame(s, TCG_REG_SP, TCG_STATIC_CALL_ARGS_SIZE, TEMP_SIZE);
2437
2438    /* TB prologue */
2439    tcg_out_opc_addi_d(s, TCG_REG_SP, TCG_REG_SP, -FRAME_SIZE);
2440    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
2441        tcg_out_st(s, TCG_TYPE_REG, tcg_target_callee_save_regs[i],
2442                   TCG_REG_SP, SAVE_OFS + i * REG_SIZE);
2443    }
2444
2445    if (!tcg_use_softmmu && guest_base) {
2446        tcg_out_movi(s, TCG_TYPE_PTR, TCG_GUEST_BASE_REG, guest_base);
2447        tcg_regset_set_reg(s->reserved_regs, TCG_GUEST_BASE_REG);
2448    }
2449
2450    /* Call generated code */
2451    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
2452    tcg_out_opc_jirl(s, TCG_REG_ZERO, tcg_target_call_iarg_regs[1], 0);
2453
2454    /* Return path for goto_ptr. Set return value to 0 */
2455    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
2456    tcg_out_mov(s, TCG_TYPE_REG, TCG_REG_A0, TCG_REG_ZERO);
2457
2458    /* TB epilogue */
2459    tb_ret_addr = tcg_splitwx_to_rx(s->code_ptr);
2460    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
2461        tcg_out_ld(s, TCG_TYPE_REG, tcg_target_callee_save_regs[i],
2462                   TCG_REG_SP, SAVE_OFS + i * REG_SIZE);
2463    }
2464
2465    tcg_out_opc_addi_d(s, TCG_REG_SP, TCG_REG_SP, FRAME_SIZE);
2466    tcg_out_opc_jirl(s, TCG_REG_ZERO, TCG_REG_RA, 0);
2467}
2468
2469static void tcg_out_tb_start(TCGContext *s)
2470{
2471    /* nothing to do */
2472}
2473
2474static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
2475{
2476    for (int i = 0; i < count; ++i) {
2477        /* Canonical nop is andi r0,r0,0 */
2478        p[i] = OPC_ANDI;
2479    }
2480}
2481
2482static void tcg_target_init(TCGContext *s)
2483{
2484    unsigned long hwcap = qemu_getauxval(AT_HWCAP);
2485
2486    /* Server and desktop class cpus have UAL; embedded cpus do not. */
2487    if (!(hwcap & HWCAP_LOONGARCH_UAL)) {
2488        error_report("TCG: unaligned access support required; exiting");
2489        exit(EXIT_FAILURE);
2490    }
2491
2492    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
2493    tcg_target_available_regs[TCG_TYPE_I64] = ALL_GENERAL_REGS;
2494
2495    tcg_target_call_clobber_regs = ALL_GENERAL_REGS | ALL_VECTOR_REGS;
2496    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S0);
2497    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S1);
2498    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S2);
2499    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S3);
2500    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S4);
2501    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S5);
2502    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S6);
2503    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S7);
2504    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S8);
2505    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S9);
2506
2507    if (cpuinfo & CPUINFO_LSX) {
2508        tcg_target_available_regs[TCG_TYPE_V64] = ALL_VECTOR_REGS;
2509        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
2510        if (cpuinfo & CPUINFO_LASX) {
2511            tcg_target_available_regs[TCG_TYPE_V256] = ALL_VECTOR_REGS;
2512        }
2513        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V24);
2514        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V25);
2515        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V26);
2516        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V27);
2517        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V28);
2518        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V29);
2519        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V30);
2520        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V31);
2521    }
2522
2523    s->reserved_regs = 0;
2524    tcg_regset_set_reg(s->reserved_regs, TCG_REG_ZERO);
2525    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP0);
2526    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP1);
2527    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP2);
2528    tcg_regset_set_reg(s->reserved_regs, TCG_REG_SP);
2529    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TP);
2530    tcg_regset_set_reg(s->reserved_regs, TCG_REG_RESERVED);
2531    tcg_regset_set_reg(s->reserved_regs, TCG_VEC_TMP0);
2532}
2533
2534typedef struct {
2535    DebugFrameHeader h;
2536    uint8_t fde_def_cfa[4];
2537    uint8_t fde_reg_ofs[ARRAY_SIZE(tcg_target_callee_save_regs) * 2];
2538} DebugFrame;
2539
2540#define ELF_HOST_MACHINE EM_LOONGARCH
2541
2542static const DebugFrame debug_frame = {
2543    .h.cie.len = sizeof(DebugFrameCIE) - 4, /* length after .len member */
2544    .h.cie.id = -1,
2545    .h.cie.version = 1,
2546    .h.cie.code_align = 1,
2547    .h.cie.data_align = -(TCG_TARGET_REG_BITS / 8) & 0x7f, /* sleb128 */
2548    .h.cie.return_column = TCG_REG_RA,
2549
2550    /* Total FDE size does not include the "len" member.  */
2551    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
2552
2553    .fde_def_cfa = {
2554        12, TCG_REG_SP,                 /* DW_CFA_def_cfa sp, ...  */
2555        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
2556        (FRAME_SIZE >> 7)
2557    },
2558    .fde_reg_ofs = {
2559        0x80 + 23, 11,                  /* DW_CFA_offset, s0, -88 */
2560        0x80 + 24, 10,                  /* DW_CFA_offset, s1, -80 */
2561        0x80 + 25, 9,                   /* DW_CFA_offset, s2, -72 */
2562        0x80 + 26, 8,                   /* DW_CFA_offset, s3, -64 */
2563        0x80 + 27, 7,                   /* DW_CFA_offset, s4, -56 */
2564        0x80 + 28, 6,                   /* DW_CFA_offset, s5, -48 */
2565        0x80 + 29, 5,                   /* DW_CFA_offset, s6, -40 */
2566        0x80 + 30, 4,                   /* DW_CFA_offset, s7, -32 */
2567        0x80 + 31, 3,                   /* DW_CFA_offset, s8, -24 */
2568        0x80 + 22, 2,                   /* DW_CFA_offset, s9, -16 */
2569        0x80 + 1 , 1,                   /* DW_CFA_offset, ra, -8 */
2570    }
2571};
2572
2573void tcg_register_jit(const void *buf, size_t buf_size)
2574{
2575    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
2576}
2577