xref: /openbmc/qemu/tcg/loongarch64/tcg-target.c.inc (revision 48e8de684aff7ad112aafcf74f776d2a66ef192e)
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_C12   0x800
180#define TCG_CT_CONST_WSZ   0x1000
181#define TCG_CT_CONST_VCMP  0x2000
182#define TCG_CT_CONST_VADD  0x4000
183
184#define ALL_GENERAL_REGS   MAKE_64BIT_MASK(0, 32)
185#define ALL_VECTOR_REGS    MAKE_64BIT_MASK(32, 32)
186
187static inline tcg_target_long sextreg(tcg_target_long val, int pos, int len)
188{
189    return sextract64(val, pos, len);
190}
191
192/* test if a constant matches the constraint */
193static bool tcg_target_const_match(int64_t val, int ct,
194                                   TCGType type, TCGCond cond, int vece)
195{
196    if (ct & TCG_CT_CONST) {
197        return true;
198    }
199    if ((ct & TCG_CT_CONST_S12) && val == sextreg(val, 0, 12)) {
200        return true;
201    }
202    if ((ct & TCG_CT_CONST_S32) && val == (int32_t)val) {
203        return true;
204    }
205    if ((ct & TCG_CT_CONST_U12) && val >= 0 && val <= 0xfff) {
206        return true;
207    }
208    if ((ct & TCG_CT_CONST_C12) && ~val >= 0 && ~val <= 0xfff) {
209        return true;
210    }
211    if ((ct & TCG_CT_CONST_WSZ) && val == (type == TCG_TYPE_I32 ? 32 : 64)) {
212        return true;
213    }
214    if (ct & (TCG_CT_CONST_VCMP | TCG_CT_CONST_VADD)) {
215        int64_t vec_val = sextract64(val, 0, 8 << vece);
216        if (ct & TCG_CT_CONST_VCMP) {
217            switch (cond) {
218            case TCG_COND_EQ:
219            case TCG_COND_LE:
220            case TCG_COND_LT:
221                return -0x10 <= vec_val && vec_val <= 0x0f;
222            case TCG_COND_LEU:
223            case TCG_COND_LTU:
224                return 0x00 <= vec_val && vec_val <= 0x1f;
225            default:
226                return false;
227            }
228        }
229        if ((ct & TCG_CT_CONST_VADD) && -0x1f <= vec_val && vec_val <= 0x1f) {
230            return true;
231        }
232    }
233    return false;
234}
235
236/*
237 * Relocations
238 */
239
240/*
241 * Relocation records defined in LoongArch ELF psABI v1.00 is way too
242 * complicated; a whopping stack machine is needed to stuff the fields, at
243 * the very least one SOP_PUSH and one SOP_POP (of the correct format) are
244 * needed.
245 *
246 * Hence, define our own simpler relocation types. Numbers are chosen as to
247 * not collide with potential future additions to the true ELF relocation
248 * type enum.
249 */
250
251/* Field Sk16, shifted right by 2; suitable for conditional jumps */
252#define R_LOONGARCH_BR_SK16     256
253/* Field Sd10k16, shifted right by 2; suitable for B and BL */
254#define R_LOONGARCH_BR_SD10K16  257
255
256static bool reloc_br_sk16(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
257{
258    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
259    intptr_t offset = (intptr_t)target - (intptr_t)src_rx;
260
261    tcg_debug_assert((offset & 3) == 0);
262    offset >>= 2;
263    if (offset == sextreg(offset, 0, 16)) {
264        *src_rw = deposit64(*src_rw, 10, 16, offset);
265        return true;
266    }
267
268    return false;
269}
270
271static bool reloc_br_sd10k16(tcg_insn_unit *src_rw,
272                             const tcg_insn_unit *target)
273{
274    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
275    intptr_t offset = (intptr_t)target - (intptr_t)src_rx;
276
277    tcg_debug_assert((offset & 3) == 0);
278    offset >>= 2;
279    if (offset == sextreg(offset, 0, 26)) {
280        *src_rw = deposit64(*src_rw, 0, 10, offset >> 16); /* slot d10 */
281        *src_rw = deposit64(*src_rw, 10, 16, offset); /* slot k16 */
282        return true;
283    }
284
285    return false;
286}
287
288static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
289                        intptr_t value, intptr_t addend)
290{
291    tcg_debug_assert(addend == 0);
292    switch (type) {
293    case R_LOONGARCH_BR_SK16:
294        return reloc_br_sk16(code_ptr, (tcg_insn_unit *)value);
295    case R_LOONGARCH_BR_SD10K16:
296        return reloc_br_sd10k16(code_ptr, (tcg_insn_unit *)value);
297    default:
298        g_assert_not_reached();
299    }
300}
301
302#include "tcg-insn-defs.c.inc"
303
304/*
305 * TCG intrinsics
306 */
307
308static void tcg_out_mb(TCGContext *s, TCGArg a0)
309{
310    /* Baseline LoongArch only has the full barrier, unfortunately.  */
311    tcg_out_opc_dbar(s, 0);
312}
313
314static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
315{
316    if (ret == arg) {
317        return true;
318    }
319    switch (type) {
320    case TCG_TYPE_I32:
321    case TCG_TYPE_I64:
322        if (ret < TCG_REG_V0) {
323            if (arg < TCG_REG_V0) {
324                /*
325                 * Conventional register-register move used in LoongArch is
326                 * `or dst, src, zero`.
327                 */
328                tcg_out_opc_or(s, ret, arg, TCG_REG_ZERO);
329            } else {
330                tcg_out_opc_movfr2gr_d(s, ret, arg);
331            }
332        } else {
333            if (arg < TCG_REG_V0) {
334                tcg_out_opc_movgr2fr_d(s, ret, arg);
335            } else {
336                tcg_out_opc_fmov_d(s, ret, arg);
337            }
338        }
339        break;
340    case TCG_TYPE_V64:
341    case TCG_TYPE_V128:
342        tcg_out_opc_vori_b(s, ret, arg, 0);
343        break;
344    case TCG_TYPE_V256:
345        tcg_out_opc_xvori_b(s, ret, arg, 0);
346        break;
347    default:
348        g_assert_not_reached();
349    }
350    return true;
351}
352
353/* Loads a 32-bit immediate into rd, sign-extended.  */
354static void tcg_out_movi_i32(TCGContext *s, TCGReg rd, int32_t val)
355{
356    tcg_target_long lo = sextreg(val, 0, 12);
357    tcg_target_long hi12 = sextreg(val, 12, 20);
358
359    /* Single-instruction cases.  */
360    if (hi12 == 0) {
361        /* val fits in uimm12: ori rd, zero, val */
362        tcg_out_opc_ori(s, rd, TCG_REG_ZERO, val);
363        return;
364    }
365    if (hi12 == sextreg(lo, 12, 20)) {
366        /* val fits in simm12: addi.w rd, zero, val */
367        tcg_out_opc_addi_w(s, rd, TCG_REG_ZERO, val);
368        return;
369    }
370
371    /* High bits must be set; load with lu12i.w + optional ori.  */
372    tcg_out_opc_lu12i_w(s, rd, hi12);
373    if (lo != 0) {
374        tcg_out_opc_ori(s, rd, rd, lo & 0xfff);
375    }
376}
377
378static void tcg_out_movi(TCGContext *s, TCGType type, TCGReg rd,
379                         tcg_target_long val)
380{
381    /*
382     * LoongArch conventionally loads 64-bit immediates in at most 4 steps,
383     * with dedicated instructions for filling the respective bitfields
384     * below:
385     *
386     *        6                   5                   4               3
387     *  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
388     * +-----------------------+---------------------------------------+...
389     * |          hi52         |                  hi32                 |
390     * +-----------------------+---------------------------------------+...
391     *       3                   2                   1
392     *     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
393     * ...+-------------------------------------+-------------------------+
394     *    |                 hi12                |            lo           |
395     * ...+-------------------------------------+-------------------------+
396     *
397     * Check if val belong to one of the several fast cases, before falling
398     * back to the slow path.
399     */
400
401    intptr_t src_rx, pc_offset;
402    tcg_target_long hi12, hi32, hi52;
403
404    /* Value fits in signed i32.  */
405    if (type == TCG_TYPE_I32 || val == (int32_t)val) {
406        tcg_out_movi_i32(s, rd, val);
407        return;
408    }
409
410    /* PC-relative cases.  */
411    src_rx = (intptr_t)tcg_splitwx_to_rx(s->code_ptr);
412    if ((val & 3) == 0) {
413        pc_offset = val - src_rx;
414        if (pc_offset == sextreg(pc_offset, 0, 22)) {
415            /* Single pcaddu2i.  */
416            tcg_out_opc_pcaddu2i(s, rd, pc_offset >> 2);
417            return;
418        }
419    }
420
421    pc_offset = (val >> 12) - (src_rx >> 12);
422    if (pc_offset == sextreg(pc_offset, 0, 20)) {
423        /* Load with pcalau12i + ori.  */
424        tcg_target_long val_lo = val & 0xfff;
425        tcg_out_opc_pcalau12i(s, rd, pc_offset);
426        if (val_lo != 0) {
427            tcg_out_opc_ori(s, rd, rd, val_lo);
428        }
429        return;
430    }
431
432    hi12 = sextreg(val, 12, 20);
433    hi32 = sextreg(val, 32, 20);
434    hi52 = sextreg(val, 52, 12);
435
436    /* Single cu52i.d case.  */
437    if ((hi52 != 0) && (ctz64(val) >= 52)) {
438        tcg_out_opc_cu52i_d(s, rd, TCG_REG_ZERO, hi52);
439        return;
440    }
441
442    /* Slow path.  Initialize the low 32 bits, then concat high bits.  */
443    tcg_out_movi_i32(s, rd, val);
444
445    /* Load hi32 and hi52 explicitly when they are unexpected values. */
446    if (hi32 != sextreg(hi12, 20, 20)) {
447        tcg_out_opc_cu32i_d(s, rd, hi32);
448    }
449
450    if (hi52 != sextreg(hi32, 20, 12)) {
451        tcg_out_opc_cu52i_d(s, rd, rd, hi52);
452    }
453}
454
455static void tcg_out_addi(TCGContext *s, TCGType type, TCGReg rd,
456                         TCGReg rs, tcg_target_long imm)
457{
458    tcg_target_long lo12 = sextreg(imm, 0, 12);
459    tcg_target_long hi16 = sextreg(imm - lo12, 16, 16);
460
461    /*
462     * Note that there's a hole in between hi16 and lo12:
463     *
464     *       3                   2                   1                   0
465     *     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
466     * ...+-------------------------------+-------+-----------------------+
467     *    |             hi16              |       |          lo12         |
468     * ...+-------------------------------+-------+-----------------------+
469     *
470     * For bits within that hole, it's more efficient to use LU12I and ADD.
471     */
472    if (imm == (hi16 << 16) + lo12) {
473        if (hi16) {
474            tcg_out_opc_addu16i_d(s, rd, rs, hi16);
475            rs = rd;
476        }
477        if (type == TCG_TYPE_I32) {
478            tcg_out_opc_addi_w(s, rd, rs, lo12);
479        } else if (lo12) {
480            tcg_out_opc_addi_d(s, rd, rs, lo12);
481        } else {
482            tcg_out_mov(s, type, rd, rs);
483        }
484    } else {
485        tcg_out_movi(s, type, TCG_REG_TMP0, imm);
486        if (type == TCG_TYPE_I32) {
487            tcg_out_opc_add_w(s, rd, rs, TCG_REG_TMP0);
488        } else {
489            tcg_out_opc_add_d(s, rd, rs, TCG_REG_TMP0);
490        }
491    }
492}
493
494static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
495{
496    return false;
497}
498
499static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
500                             tcg_target_long imm)
501{
502    /* This function is only used for passing structs by reference. */
503    g_assert_not_reached();
504}
505
506static void tcg_out_ext8u(TCGContext *s, TCGReg ret, TCGReg arg)
507{
508    tcg_out_opc_andi(s, ret, arg, 0xff);
509}
510
511static void tcg_out_ext16u(TCGContext *s, TCGReg ret, TCGReg arg)
512{
513    tcg_out_opc_bstrpick_w(s, ret, arg, 0, 15);
514}
515
516static void tcg_out_ext32u(TCGContext *s, TCGReg ret, TCGReg arg)
517{
518    tcg_out_opc_bstrpick_d(s, ret, arg, 0, 31);
519}
520
521static void tcg_out_ext8s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
522{
523    tcg_out_opc_sext_b(s, ret, arg);
524}
525
526static void tcg_out_ext16s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
527{
528    tcg_out_opc_sext_h(s, ret, arg);
529}
530
531static void tcg_out_ext32s(TCGContext *s, TCGReg ret, TCGReg arg)
532{
533    tcg_out_opc_addi_w(s, ret, arg, 0);
534}
535
536static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg)
537{
538    if (ret != arg) {
539        tcg_out_ext32s(s, ret, arg);
540    }
541}
542
543static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg)
544{
545    tcg_out_ext32u(s, ret, arg);
546}
547
548static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg ret, TCGReg arg)
549{
550    tcg_out_ext32s(s, ret, arg);
551}
552
553static void tcg_out_clzctz(TCGContext *s, LoongArchInsn opc,
554                           TCGReg a0, TCGReg a1, TCGReg a2,
555                           bool c2, bool is_32bit)
556{
557    if (c2) {
558        /*
559         * Fast path: semantics already satisfied due to constraint and
560         * insn behavior, single instruction is enough.
561         */
562        tcg_debug_assert(a2 == (is_32bit ? 32 : 64));
563        /* all clz/ctz insns belong to DJ-format */
564        tcg_out32(s, encode_dj_insn(opc, a0, a1));
565        return;
566    }
567
568    tcg_out32(s, encode_dj_insn(opc, TCG_REG_TMP0, a1));
569    /* a0 = a1 ? REG_TMP0 : a2 */
570    tcg_out_opc_maskeqz(s, TCG_REG_TMP0, TCG_REG_TMP0, a1);
571    tcg_out_opc_masknez(s, a0, a2, a1);
572    tcg_out_opc_or(s, a0, TCG_REG_TMP0, a0);
573}
574
575#define SETCOND_INV    TCG_TARGET_NB_REGS
576#define SETCOND_NEZ    (SETCOND_INV << 1)
577#define SETCOND_FLAGS  (SETCOND_INV | SETCOND_NEZ)
578
579static int tcg_out_setcond_int(TCGContext *s, TCGCond cond, TCGReg ret,
580                               TCGReg arg1, tcg_target_long arg2, bool c2)
581{
582    int flags = 0;
583
584    switch (cond) {
585    case TCG_COND_EQ:    /* -> NE  */
586    case TCG_COND_GE:    /* -> LT  */
587    case TCG_COND_GEU:   /* -> LTU */
588    case TCG_COND_GT:    /* -> LE  */
589    case TCG_COND_GTU:   /* -> LEU */
590        cond = tcg_invert_cond(cond);
591        flags ^= SETCOND_INV;
592        break;
593    default:
594        break;
595    }
596
597    switch (cond) {
598    case TCG_COND_LE:
599    case TCG_COND_LEU:
600        /*
601         * If we have a constant input, the most efficient way to implement
602         * LE is by adding 1 and using LT.  Watch out for wrap around for LEU.
603         * We don't need to care for this for LE because the constant input
604         * is still constrained to int32_t, and INT32_MAX+1 is representable
605         * in the 64-bit temporary register.
606         */
607        if (c2) {
608            if (cond == TCG_COND_LEU) {
609                /* unsigned <= -1 is true */
610                if (arg2 == -1) {
611                    tcg_out_movi(s, TCG_TYPE_REG, ret, !(flags & SETCOND_INV));
612                    return ret;
613                }
614                cond = TCG_COND_LTU;
615            } else {
616                cond = TCG_COND_LT;
617            }
618            arg2 += 1;
619        } else {
620            TCGReg tmp = arg2;
621            arg2 = arg1;
622            arg1 = tmp;
623            cond = tcg_swap_cond(cond);    /* LE -> GE */
624            cond = tcg_invert_cond(cond);  /* GE -> LT */
625            flags ^= SETCOND_INV;
626        }
627        break;
628    default:
629        break;
630    }
631
632    switch (cond) {
633    case TCG_COND_NE:
634        flags |= SETCOND_NEZ;
635        if (!c2) {
636            tcg_out_opc_xor(s, ret, arg1, arg2);
637        } else if (arg2 == 0) {
638            ret = arg1;
639        } else if (arg2 >= 0 && arg2 <= 0xfff) {
640            tcg_out_opc_xori(s, ret, arg1, arg2);
641        } else {
642            tcg_out_addi(s, TCG_TYPE_REG, ret, arg1, -arg2);
643        }
644        break;
645
646    case TCG_COND_LT:
647    case TCG_COND_LTU:
648        if (c2) {
649            if (arg2 >= -0x800 && arg2 <= 0x7ff) {
650                if (cond == TCG_COND_LT) {
651                    tcg_out_opc_slti(s, ret, arg1, arg2);
652                } else {
653                    tcg_out_opc_sltui(s, ret, arg1, arg2);
654                }
655                break;
656            }
657            tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_TMP0, arg2);
658            arg2 = TCG_REG_TMP0;
659        }
660        if (cond == TCG_COND_LT) {
661            tcg_out_opc_slt(s, ret, arg1, arg2);
662        } else {
663            tcg_out_opc_sltu(s, ret, arg1, arg2);
664        }
665        break;
666
667    default:
668        g_assert_not_reached();
669    }
670
671    return ret | flags;
672}
673
674static void tcg_out_setcond(TCGContext *s, TCGCond cond, TCGReg ret,
675                            TCGReg arg1, tcg_target_long arg2, bool c2)
676{
677    int tmpflags = tcg_out_setcond_int(s, cond, ret, arg1, arg2, c2);
678
679    if (tmpflags != ret) {
680        TCGReg tmp = tmpflags & ~SETCOND_FLAGS;
681
682        switch (tmpflags & SETCOND_FLAGS) {
683        case SETCOND_INV:
684            /* Intermediate result is boolean: simply invert. */
685            tcg_out_opc_xori(s, ret, tmp, 1);
686            break;
687        case SETCOND_NEZ:
688            /* Intermediate result is zero/non-zero: test != 0. */
689            tcg_out_opc_sltu(s, ret, TCG_REG_ZERO, tmp);
690            break;
691        case SETCOND_NEZ | SETCOND_INV:
692            /* Intermediate result is zero/non-zero: test == 0. */
693            tcg_out_opc_sltui(s, ret, tmp, 1);
694            break;
695        default:
696            g_assert_not_reached();
697        }
698    }
699}
700
701static void tcg_out_movcond(TCGContext *s, TCGCond cond, TCGReg ret,
702                            TCGReg c1, tcg_target_long c2, bool const2,
703                            TCGReg v1, TCGReg v2)
704{
705    int tmpflags = tcg_out_setcond_int(s, cond, TCG_REG_TMP0, c1, c2, const2);
706    TCGReg t;
707
708    /* Standardize the test below to t != 0. */
709    if (tmpflags & SETCOND_INV) {
710        t = v1, v1 = v2, v2 = t;
711    }
712
713    t = tmpflags & ~SETCOND_FLAGS;
714    if (v1 == TCG_REG_ZERO) {
715        tcg_out_opc_masknez(s, ret, v2, t);
716    } else if (v2 == TCG_REG_ZERO) {
717        tcg_out_opc_maskeqz(s, ret, v1, t);
718    } else {
719        tcg_out_opc_masknez(s, TCG_REG_TMP2, v2, t); /* t ? 0 : v2 */
720        tcg_out_opc_maskeqz(s, TCG_REG_TMP1, v1, t); /* t ? v1 : 0 */
721        tcg_out_opc_or(s, ret, TCG_REG_TMP1, TCG_REG_TMP2);
722    }
723}
724
725/*
726 * Branch helpers
727 */
728
729static const struct {
730    LoongArchInsn op;
731    bool swap;
732} tcg_brcond_to_loongarch[] = {
733    [TCG_COND_EQ] =  { OPC_BEQ,  false },
734    [TCG_COND_NE] =  { OPC_BNE,  false },
735    [TCG_COND_LT] =  { OPC_BGT,  true  },
736    [TCG_COND_GE] =  { OPC_BLE,  true  },
737    [TCG_COND_LE] =  { OPC_BLE,  false },
738    [TCG_COND_GT] =  { OPC_BGT,  false },
739    [TCG_COND_LTU] = { OPC_BGTU, true  },
740    [TCG_COND_GEU] = { OPC_BLEU, true  },
741    [TCG_COND_LEU] = { OPC_BLEU, false },
742    [TCG_COND_GTU] = { OPC_BGTU, false }
743};
744
745static void tcg_out_brcond(TCGContext *s, TCGCond cond, TCGReg arg1,
746                           TCGReg arg2, TCGLabel *l)
747{
748    LoongArchInsn op = tcg_brcond_to_loongarch[cond].op;
749
750    tcg_debug_assert(op != 0);
751
752    if (tcg_brcond_to_loongarch[cond].swap) {
753        TCGReg t = arg1;
754        arg1 = arg2;
755        arg2 = t;
756    }
757
758    /* all conditional branch insns belong to DJSk16-format */
759    tcg_out_reloc(s, s->code_ptr, R_LOONGARCH_BR_SK16, l, 0);
760    tcg_out32(s, encode_djsk16_insn(op, arg1, arg2, 0));
761}
762
763static void tcg_out_call_int(TCGContext *s, const tcg_insn_unit *arg, bool tail)
764{
765    TCGReg link = tail ? TCG_REG_ZERO : TCG_REG_RA;
766    ptrdiff_t offset = tcg_pcrel_diff(s, arg);
767
768    tcg_debug_assert((offset & 3) == 0);
769    if (offset == sextreg(offset, 0, 28)) {
770        /* short jump: +/- 256MiB */
771        if (tail) {
772            tcg_out_opc_b(s, offset >> 2);
773        } else {
774            tcg_out_opc_bl(s, offset >> 2);
775        }
776    } else if (offset == sextreg(offset, 0, 38)) {
777        /* long jump: +/- 256GiB */
778        tcg_target_long lo = sextreg(offset, 0, 18);
779        tcg_target_long hi = offset - lo;
780        tcg_out_opc_pcaddu18i(s, TCG_REG_TMP0, hi >> 18);
781        tcg_out_opc_jirl(s, link, TCG_REG_TMP0, lo >> 2);
782    } else {
783        /* far jump: 64-bit */
784        tcg_target_long lo = sextreg((tcg_target_long)arg, 0, 18);
785        tcg_target_long hi = (tcg_target_long)arg - lo;
786        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, hi);
787        tcg_out_opc_jirl(s, link, TCG_REG_TMP0, lo >> 2);
788    }
789}
790
791static void tcg_out_call(TCGContext *s, const tcg_insn_unit *arg,
792                         const TCGHelperInfo *info)
793{
794    tcg_out_call_int(s, arg, false);
795}
796
797/*
798 * Load/store helpers
799 */
800
801static void tcg_out_ldst(TCGContext *s, LoongArchInsn opc, TCGReg data,
802                         TCGReg addr, intptr_t offset)
803{
804    intptr_t imm12 = sextreg(offset, 0, 12);
805
806    if (offset != imm12) {
807        intptr_t diff = tcg_pcrel_diff(s, (void *)offset);
808
809        if (addr == TCG_REG_ZERO && diff == (int32_t)diff) {
810            imm12 = sextreg(diff, 0, 12);
811            tcg_out_opc_pcaddu12i(s, TCG_REG_TMP2, (diff - imm12) >> 12);
812        } else {
813            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP2, offset - imm12);
814            if (addr != TCG_REG_ZERO) {
815                tcg_out_opc_add_d(s, TCG_REG_TMP2, TCG_REG_TMP2, addr);
816            }
817        }
818        addr = TCG_REG_TMP2;
819    }
820
821    switch (opc) {
822    case OPC_LD_B:
823    case OPC_LD_BU:
824    case OPC_LD_H:
825    case OPC_LD_HU:
826    case OPC_LD_W:
827    case OPC_LD_WU:
828    case OPC_LD_D:
829    case OPC_ST_B:
830    case OPC_ST_H:
831    case OPC_ST_W:
832    case OPC_ST_D:
833        tcg_out32(s, encode_djsk12_insn(opc, data, addr, imm12));
834        break;
835    case OPC_FLD_S:
836    case OPC_FLD_D:
837    case OPC_FST_S:
838    case OPC_FST_D:
839        tcg_out32(s, encode_fdjsk12_insn(opc, data, addr, imm12));
840        break;
841    default:
842        g_assert_not_reached();
843    }
844}
845
846static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg dest,
847                       TCGReg base, intptr_t offset)
848{
849    switch (type) {
850    case TCG_TYPE_I32:
851        if (dest < TCG_REG_V0) {
852            tcg_out_ldst(s, OPC_LD_W, dest, base, offset);
853        } else {
854            tcg_out_ldst(s, OPC_FLD_S, dest, base, offset);
855        }
856        break;
857    case TCG_TYPE_I64:
858    case TCG_TYPE_V64:
859        if (dest < TCG_REG_V0) {
860            tcg_out_ldst(s, OPC_LD_D, dest, base, offset);
861        } else {
862            tcg_out_ldst(s, OPC_FLD_D, dest, base, offset);
863        }
864        break;
865    case TCG_TYPE_V128:
866        if (-0x800 <= offset && offset <= 0x7ff) {
867            tcg_out_opc_vld(s, dest, base, offset);
868        } else {
869            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, offset);
870            tcg_out_opc_vldx(s, dest, base, TCG_REG_TMP0);
871        }
872        break;
873    case TCG_TYPE_V256:
874        if (-0x800 <= offset && offset <= 0x7ff) {
875            tcg_out_opc_xvld(s, dest, base, offset);
876        } else {
877            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, offset);
878            tcg_out_opc_xvldx(s, dest, base, TCG_REG_TMP0);
879        }
880        break;
881    default:
882        g_assert_not_reached();
883    }
884}
885
886static void tcg_out_st(TCGContext *s, TCGType type, TCGReg src,
887                       TCGReg base, intptr_t offset)
888{
889    switch (type) {
890    case TCG_TYPE_I32:
891        if (src < TCG_REG_V0) {
892            tcg_out_ldst(s, OPC_ST_W, src, base, offset);
893        } else {
894            tcg_out_ldst(s, OPC_FST_S, src, base, offset);
895        }
896        break;
897    case TCG_TYPE_I64:
898    case TCG_TYPE_V64:
899        if (src < TCG_REG_V0) {
900            tcg_out_ldst(s, OPC_ST_D, src, base, offset);
901        } else {
902            tcg_out_ldst(s, OPC_FST_D, src, base, offset);
903        }
904        break;
905    case TCG_TYPE_V128:
906        if (-0x800 <= offset && offset <= 0x7ff) {
907            tcg_out_opc_vst(s, src, base, offset);
908        } else {
909            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, offset);
910            tcg_out_opc_vstx(s, src, base, TCG_REG_TMP0);
911        }
912        break;
913    case TCG_TYPE_V256:
914        if (-0x800 <= offset && offset <= 0x7ff) {
915            tcg_out_opc_xvst(s, src, base, offset);
916        } else {
917            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, offset);
918            tcg_out_opc_xvstx(s, src, base, TCG_REG_TMP0);
919        }
920        break;
921    default:
922        g_assert_not_reached();
923    }
924}
925
926static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
927                        TCGReg base, intptr_t ofs)
928{
929    if (val == 0) {
930        tcg_out_st(s, type, TCG_REG_ZERO, base, ofs);
931        return true;
932    }
933    return false;
934}
935
936/*
937 * Load/store helpers for SoftMMU, and qemu_ld/st implementations
938 */
939
940static bool tcg_out_goto(TCGContext *s, const tcg_insn_unit *target)
941{
942    tcg_out_opc_b(s, 0);
943    return reloc_br_sd10k16(s->code_ptr - 1, target);
944}
945
946static const TCGLdstHelperParam ldst_helper_param = {
947    .ntmp = 1, .tmp = { TCG_REG_TMP0 }
948};
949
950static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
951{
952    MemOp opc = get_memop(l->oi);
953
954    /* resolve label address */
955    if (!reloc_br_sk16(l->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
956        return false;
957    }
958
959    tcg_out_ld_helper_args(s, l, &ldst_helper_param);
960    tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SIZE], false);
961    tcg_out_ld_helper_ret(s, l, false, &ldst_helper_param);
962    return tcg_out_goto(s, l->raddr);
963}
964
965static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
966{
967    MemOp opc = get_memop(l->oi);
968
969    /* resolve label address */
970    if (!reloc_br_sk16(l->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
971        return false;
972    }
973
974    tcg_out_st_helper_args(s, l, &ldst_helper_param);
975    tcg_out_call_int(s, qemu_st_helpers[opc & MO_SIZE], false);
976    return tcg_out_goto(s, l->raddr);
977}
978
979typedef struct {
980    TCGReg base;
981    TCGReg index;
982    TCGAtomAlign aa;
983} HostAddress;
984
985bool tcg_target_has_memory_bswap(MemOp memop)
986{
987    return false;
988}
989
990/* We expect to use a 12-bit negative offset from ENV.  */
991#define MIN_TLB_MASK_TABLE_OFS  -(1 << 11)
992
993/*
994 * For system-mode, perform the TLB load and compare.
995 * For user-mode, perform any required alignment tests.
996 * In both cases, return a TCGLabelQemuLdst structure if the slow path
997 * is required and fill in @h with the host address for the fast path.
998 */
999static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
1000                                           TCGReg addr_reg, MemOpIdx oi,
1001                                           bool is_ld)
1002{
1003    TCGType addr_type = s->addr_type;
1004    TCGLabelQemuLdst *ldst = NULL;
1005    MemOp opc = get_memop(oi);
1006    MemOp a_bits;
1007
1008    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, false);
1009    a_bits = h->aa.align;
1010
1011    if (tcg_use_softmmu) {
1012        unsigned s_bits = opc & MO_SIZE;
1013        int mem_index = get_mmuidx(oi);
1014        int fast_ofs = tlb_mask_table_ofs(s, mem_index);
1015        int mask_ofs = fast_ofs + offsetof(CPUTLBDescFast, mask);
1016        int table_ofs = fast_ofs + offsetof(CPUTLBDescFast, table);
1017
1018        ldst = new_ldst_label(s);
1019        ldst->is_ld = is_ld;
1020        ldst->oi = oi;
1021        ldst->addr_reg = addr_reg;
1022
1023        tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP0, TCG_AREG0, mask_ofs);
1024        tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP1, TCG_AREG0, table_ofs);
1025
1026        tcg_out_opc_srli_d(s, TCG_REG_TMP2, addr_reg,
1027                           s->page_bits - CPU_TLB_ENTRY_BITS);
1028        tcg_out_opc_and(s, TCG_REG_TMP2, TCG_REG_TMP2, TCG_REG_TMP0);
1029        tcg_out_opc_add_d(s, TCG_REG_TMP2, TCG_REG_TMP2, TCG_REG_TMP1);
1030
1031        /* Load the tlb comparator and the addend.  */
1032        QEMU_BUILD_BUG_ON(HOST_BIG_ENDIAN);
1033        tcg_out_ld(s, addr_type, TCG_REG_TMP0, TCG_REG_TMP2,
1034                   is_ld ? offsetof(CPUTLBEntry, addr_read)
1035                         : offsetof(CPUTLBEntry, addr_write));
1036        tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP2, TCG_REG_TMP2,
1037                   offsetof(CPUTLBEntry, addend));
1038
1039        /*
1040         * For aligned accesses, we check the first byte and include the
1041         * alignment bits within the address.  For unaligned access, we
1042         * check that we don't cross pages using the address of the last
1043         * byte of the access.
1044         */
1045        if (a_bits < s_bits) {
1046            unsigned a_mask = (1u << a_bits) - 1;
1047            unsigned s_mask = (1u << s_bits) - 1;
1048            tcg_out_addi(s, addr_type, TCG_REG_TMP1, addr_reg, s_mask - a_mask);
1049        } else {
1050            tcg_out_mov(s, addr_type, TCG_REG_TMP1, addr_reg);
1051        }
1052        tcg_out_opc_bstrins_d(s, TCG_REG_TMP1, TCG_REG_ZERO,
1053                              a_bits, s->page_bits - 1);
1054
1055        /* Compare masked address with the TLB entry.  */
1056        ldst->label_ptr[0] = s->code_ptr;
1057        tcg_out_opc_bne(s, TCG_REG_TMP0, TCG_REG_TMP1, 0);
1058
1059        h->index = TCG_REG_TMP2;
1060    } else {
1061        if (a_bits) {
1062            ldst = new_ldst_label(s);
1063
1064            ldst->is_ld = is_ld;
1065            ldst->oi = oi;
1066            ldst->addr_reg = addr_reg;
1067
1068            /*
1069             * Without micro-architecture details, we don't know which of
1070             * bstrpick or andi is faster, so use bstrpick as it's not
1071             * constrained by imm field width. Not to say alignments >= 2^12
1072             * are going to happen any time soon.
1073             */
1074            tcg_out_opc_bstrpick_d(s, TCG_REG_TMP1, addr_reg, 0, a_bits - 1);
1075
1076            ldst->label_ptr[0] = s->code_ptr;
1077            tcg_out_opc_bne(s, TCG_REG_TMP1, TCG_REG_ZERO, 0);
1078        }
1079
1080        h->index = guest_base ? TCG_GUEST_BASE_REG : TCG_REG_ZERO;
1081    }
1082
1083    if (addr_type == TCG_TYPE_I32) {
1084        h->base = TCG_REG_TMP0;
1085        tcg_out_ext32u(s, h->base, addr_reg);
1086    } else {
1087        h->base = addr_reg;
1088    }
1089
1090    return ldst;
1091}
1092
1093static void tcg_out_qemu_ld_indexed(TCGContext *s, MemOp opc, TCGType type,
1094                                    TCGReg rd, HostAddress h)
1095{
1096    /* Byte swapping is left to middle-end expansion.  */
1097    tcg_debug_assert((opc & MO_BSWAP) == 0);
1098
1099    switch (opc & MO_SSIZE) {
1100    case MO_UB:
1101        tcg_out_opc_ldx_bu(s, rd, h.base, h.index);
1102        break;
1103    case MO_SB:
1104        tcg_out_opc_ldx_b(s, rd, h.base, h.index);
1105        break;
1106    case MO_UW:
1107        tcg_out_opc_ldx_hu(s, rd, h.base, h.index);
1108        break;
1109    case MO_SW:
1110        tcg_out_opc_ldx_h(s, rd, h.base, h.index);
1111        break;
1112    case MO_UL:
1113        if (type == TCG_TYPE_I64) {
1114            tcg_out_opc_ldx_wu(s, rd, h.base, h.index);
1115            break;
1116        }
1117        /* fallthrough */
1118    case MO_SL:
1119        tcg_out_opc_ldx_w(s, rd, h.base, h.index);
1120        break;
1121    case MO_UQ:
1122        tcg_out_opc_ldx_d(s, rd, h.base, h.index);
1123        break;
1124    default:
1125        g_assert_not_reached();
1126    }
1127}
1128
1129static void tcg_out_qemu_ld(TCGContext *s, TCGReg data_reg, TCGReg addr_reg,
1130                            MemOpIdx oi, TCGType data_type)
1131{
1132    TCGLabelQemuLdst *ldst;
1133    HostAddress h;
1134
1135    ldst = prepare_host_addr(s, &h, addr_reg, oi, true);
1136    tcg_out_qemu_ld_indexed(s, get_memop(oi), data_type, data_reg, h);
1137
1138    if (ldst) {
1139        ldst->type = data_type;
1140        ldst->datalo_reg = data_reg;
1141        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1142    }
1143}
1144
1145static void tcg_out_qemu_st_indexed(TCGContext *s, MemOp opc,
1146                                    TCGReg rd, HostAddress h)
1147{
1148    /* Byte swapping is left to middle-end expansion.  */
1149    tcg_debug_assert((opc & MO_BSWAP) == 0);
1150
1151    switch (opc & MO_SIZE) {
1152    case MO_8:
1153        tcg_out_opc_stx_b(s, rd, h.base, h.index);
1154        break;
1155    case MO_16:
1156        tcg_out_opc_stx_h(s, rd, h.base, h.index);
1157        break;
1158    case MO_32:
1159        tcg_out_opc_stx_w(s, rd, h.base, h.index);
1160        break;
1161    case MO_64:
1162        tcg_out_opc_stx_d(s, rd, h.base, h.index);
1163        break;
1164    default:
1165        g_assert_not_reached();
1166    }
1167}
1168
1169static void tcg_out_qemu_st(TCGContext *s, TCGReg data_reg, TCGReg addr_reg,
1170                            MemOpIdx oi, TCGType data_type)
1171{
1172    TCGLabelQemuLdst *ldst;
1173    HostAddress h;
1174
1175    ldst = prepare_host_addr(s, &h, addr_reg, oi, false);
1176    tcg_out_qemu_st_indexed(s, get_memop(oi), data_reg, h);
1177
1178    if (ldst) {
1179        ldst->type = data_type;
1180        ldst->datalo_reg = data_reg;
1181        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1182    }
1183}
1184
1185static void tcg_out_qemu_ldst_i128(TCGContext *s, TCGReg data_lo, TCGReg data_hi,
1186                                   TCGReg addr_reg, MemOpIdx oi, bool is_ld)
1187{
1188    TCGLabelQemuLdst *ldst;
1189    HostAddress h;
1190
1191    ldst = prepare_host_addr(s, &h, addr_reg, oi, is_ld);
1192
1193    if (h.aa.atom == MO_128) {
1194        /*
1195         * Use VLDX/VSTX when 128-bit atomicity is required.
1196         * If address is aligned to 16-bytes, the 128-bit load/store is atomic.
1197         */
1198        if (is_ld) {
1199            tcg_out_opc_vldx(s, TCG_VEC_TMP0, h.base, h.index);
1200            tcg_out_opc_vpickve2gr_d(s, data_lo, TCG_VEC_TMP0, 0);
1201            tcg_out_opc_vpickve2gr_d(s, data_hi, TCG_VEC_TMP0, 1);
1202        } else {
1203            tcg_out_opc_vinsgr2vr_d(s, TCG_VEC_TMP0, data_lo, 0);
1204            tcg_out_opc_vinsgr2vr_d(s, TCG_VEC_TMP0, data_hi, 1);
1205            tcg_out_opc_vstx(s, TCG_VEC_TMP0, h.base, h.index);
1206        }
1207    } else {
1208        /* Otherwise use a pair of LD/ST. */
1209        TCGReg base = h.base;
1210        if (h.index != TCG_REG_ZERO) {
1211            base = TCG_REG_TMP0;
1212            tcg_out_opc_add_d(s, base, h.base, h.index);
1213        }
1214        if (is_ld) {
1215            tcg_debug_assert(base != data_lo);
1216            tcg_out_opc_ld_d(s, data_lo, base, 0);
1217            tcg_out_opc_ld_d(s, data_hi, base, 8);
1218        } else {
1219            tcg_out_opc_st_d(s, data_lo, base, 0);
1220            tcg_out_opc_st_d(s, data_hi, base, 8);
1221        }
1222    }
1223
1224    if (ldst) {
1225        ldst->type = TCG_TYPE_I128;
1226        ldst->datalo_reg = data_lo;
1227        ldst->datahi_reg = data_hi;
1228        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1229    }
1230}
1231
1232/*
1233 * Entry-points
1234 */
1235
1236static const tcg_insn_unit *tb_ret_addr;
1237
1238static void tcg_out_exit_tb(TCGContext *s, uintptr_t a0)
1239{
1240    /* Reuse the zeroing that exists for goto_ptr.  */
1241    if (a0 == 0) {
1242        tcg_out_call_int(s, tcg_code_gen_epilogue, true);
1243    } else {
1244        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_A0, a0);
1245        tcg_out_call_int(s, tb_ret_addr, true);
1246    }
1247}
1248
1249static void tcg_out_goto_tb(TCGContext *s, int which)
1250{
1251    /*
1252     * Direct branch, or load indirect address, to be patched
1253     * by tb_target_set_jmp_target.  Check indirect load offset
1254     * in range early, regardless of direct branch distance,
1255     * via assert within tcg_out_opc_pcaddu2i.
1256     */
1257    uintptr_t i_addr = get_jmp_target_addr(s, which);
1258    intptr_t i_disp = tcg_pcrel_diff(s, (void *)i_addr);
1259
1260    set_jmp_insn_offset(s, which);
1261    tcg_out_opc_pcaddu2i(s, TCG_REG_TMP0, i_disp >> 2);
1262
1263    /* Finish the load and indirect branch. */
1264    tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP0, TCG_REG_TMP0, 0);
1265    tcg_out_opc_jirl(s, TCG_REG_ZERO, TCG_REG_TMP0, 0);
1266    set_jmp_reset_offset(s, which);
1267}
1268
1269void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
1270                              uintptr_t jmp_rx, uintptr_t jmp_rw)
1271{
1272    uintptr_t d_addr = tb->jmp_target_addr[n];
1273    ptrdiff_t d_disp = (ptrdiff_t)(d_addr - jmp_rx) >> 2;
1274    tcg_insn_unit insn;
1275
1276    /* Either directly branch, or load slot address for indirect branch. */
1277    if (d_disp == sextreg(d_disp, 0, 26)) {
1278        insn = encode_sd10k16_insn(OPC_B, d_disp);
1279    } else {
1280        uintptr_t i_addr = (uintptr_t)&tb->jmp_target_addr[n];
1281        intptr_t i_disp = i_addr - jmp_rx;
1282        insn = encode_dsj20_insn(OPC_PCADDU2I, TCG_REG_TMP0, i_disp >> 2);
1283    }
1284
1285    qatomic_set((tcg_insn_unit *)jmp_rw, insn);
1286    flush_idcache_range(jmp_rx, jmp_rw, 4);
1287}
1288
1289static void tcg_out_op(TCGContext *s, TCGOpcode opc, TCGType type,
1290                       const TCGArg args[TCG_MAX_OP_ARGS],
1291                       const int const_args[TCG_MAX_OP_ARGS])
1292{
1293    TCGArg a0 = args[0];
1294    TCGArg a1 = args[1];
1295    TCGArg a2 = args[2];
1296    TCGArg a3 = args[3];
1297    int c2 = const_args[2];
1298
1299    switch (opc) {
1300    case INDEX_op_mb:
1301        tcg_out_mb(s, a0);
1302        break;
1303
1304    case INDEX_op_goto_ptr:
1305        tcg_out_opc_jirl(s, TCG_REG_ZERO, a0, 0);
1306        break;
1307
1308    case INDEX_op_br:
1309        tcg_out_reloc(s, s->code_ptr, R_LOONGARCH_BR_SD10K16, arg_label(a0),
1310                      0);
1311        tcg_out_opc_b(s, 0);
1312        break;
1313
1314    case INDEX_op_brcond_i32:
1315    case INDEX_op_brcond_i64:
1316        tcg_out_brcond(s, a2, a0, a1, arg_label(args[3]));
1317        break;
1318
1319    case INDEX_op_extrh_i64_i32:
1320        tcg_out_opc_srai_d(s, a0, a1, 32);
1321        break;
1322
1323    case INDEX_op_not_i32:
1324    case INDEX_op_not_i64:
1325        tcg_out_opc_nor(s, a0, a1, TCG_REG_ZERO);
1326        break;
1327
1328    case INDEX_op_nor_i32:
1329    case INDEX_op_nor_i64:
1330        if (c2) {
1331            tcg_out_opc_ori(s, a0, a1, a2);
1332            tcg_out_opc_nor(s, a0, a0, TCG_REG_ZERO);
1333        } else {
1334            tcg_out_opc_nor(s, a0, a1, a2);
1335        }
1336        break;
1337
1338    case INDEX_op_andc_i32:
1339    case INDEX_op_andc_i64:
1340        if (c2) {
1341            /* guaranteed to fit due to constraint */
1342            tcg_out_opc_andi(s, a0, a1, ~a2);
1343        } else {
1344            tcg_out_opc_andn(s, a0, a1, a2);
1345        }
1346        break;
1347
1348    case INDEX_op_orc_i32:
1349    case INDEX_op_orc_i64:
1350        if (c2) {
1351            /* guaranteed to fit due to constraint */
1352            tcg_out_opc_ori(s, a0, a1, ~a2);
1353        } else {
1354            tcg_out_opc_orn(s, a0, a1, a2);
1355        }
1356        break;
1357
1358    case INDEX_op_and_i32:
1359    case INDEX_op_and_i64:
1360        if (c2) {
1361            tcg_out_opc_andi(s, a0, a1, a2);
1362        } else {
1363            tcg_out_opc_and(s, a0, a1, a2);
1364        }
1365        break;
1366
1367    case INDEX_op_or_i32:
1368    case INDEX_op_or_i64:
1369        if (c2) {
1370            tcg_out_opc_ori(s, a0, a1, a2);
1371        } else {
1372            tcg_out_opc_or(s, a0, a1, a2);
1373        }
1374        break;
1375
1376    case INDEX_op_xor_i32:
1377    case INDEX_op_xor_i64:
1378        if (c2) {
1379            tcg_out_opc_xori(s, a0, a1, a2);
1380        } else {
1381            tcg_out_opc_xor(s, a0, a1, a2);
1382        }
1383        break;
1384
1385    case INDEX_op_extract_i32:
1386        if (a2 == 0 && args[3] <= 12) {
1387            tcg_out_opc_andi(s, a0, a1, (1 << args[3]) - 1);
1388        } else {
1389            tcg_out_opc_bstrpick_w(s, a0, a1, a2, a2 + args[3] - 1);
1390        }
1391        break;
1392    case INDEX_op_extract_i64:
1393        if (a2 == 0 && args[3] <= 12) {
1394            tcg_out_opc_andi(s, a0, a1, (1 << args[3]) - 1);
1395        } else {
1396            tcg_out_opc_bstrpick_d(s, a0, a1, a2, a2 + args[3] - 1);
1397        }
1398        break;
1399
1400    case INDEX_op_sextract_i64:
1401        if (a2 + args[3] == 32) {
1402            if (a2 == 0) {
1403                tcg_out_ext32s(s, a0, a1);
1404            } else {
1405                tcg_out_opc_srai_w(s, a0, a1, a2);
1406            }
1407            break;
1408        }
1409        /* FALLTHRU */
1410    case INDEX_op_sextract_i32:
1411        if (a2 == 0 && args[3] == 8) {
1412            tcg_out_ext8s(s, TCG_TYPE_REG, a0, a1);
1413        } else if (a2 == 0 && args[3] == 16) {
1414            tcg_out_ext16s(s, TCG_TYPE_REG, a0, a1);
1415        } else {
1416            g_assert_not_reached();
1417        }
1418        break;
1419
1420    case INDEX_op_deposit_i32:
1421        tcg_out_opc_bstrins_w(s, a0, a2, args[3], args[3] + args[4] - 1);
1422        break;
1423    case INDEX_op_deposit_i64:
1424        tcg_out_opc_bstrins_d(s, a0, a2, args[3], args[3] + args[4] - 1);
1425        break;
1426
1427    case INDEX_op_bswap16_i32:
1428    case INDEX_op_bswap16_i64:
1429        tcg_out_opc_revb_2h(s, a0, a1);
1430        if (a2 & TCG_BSWAP_OS) {
1431            tcg_out_ext16s(s, TCG_TYPE_REG, a0, a0);
1432        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1433            tcg_out_ext16u(s, a0, a0);
1434        }
1435        break;
1436
1437    case INDEX_op_bswap32_i32:
1438        /* All 32-bit values are computed sign-extended in the register.  */
1439        a2 = TCG_BSWAP_OS;
1440        /* fallthrough */
1441    case INDEX_op_bswap32_i64:
1442        tcg_out_opc_revb_2w(s, a0, a1);
1443        if (a2 & TCG_BSWAP_OS) {
1444            tcg_out_ext32s(s, a0, a0);
1445        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1446            tcg_out_ext32u(s, a0, a0);
1447        }
1448        break;
1449
1450    case INDEX_op_bswap64_i64:
1451        tcg_out_opc_revb_d(s, a0, a1);
1452        break;
1453
1454    case INDEX_op_clz_i32:
1455        tcg_out_clzctz(s, OPC_CLZ_W, a0, a1, a2, c2, true);
1456        break;
1457    case INDEX_op_clz_i64:
1458        tcg_out_clzctz(s, OPC_CLZ_D, a0, a1, a2, c2, false);
1459        break;
1460
1461    case INDEX_op_ctz_i32:
1462        tcg_out_clzctz(s, OPC_CTZ_W, a0, a1, a2, c2, true);
1463        break;
1464    case INDEX_op_ctz_i64:
1465        tcg_out_clzctz(s, OPC_CTZ_D, a0, a1, a2, c2, false);
1466        break;
1467
1468    case INDEX_op_shl_i32:
1469        if (c2) {
1470            tcg_out_opc_slli_w(s, a0, a1, a2 & 0x1f);
1471        } else {
1472            tcg_out_opc_sll_w(s, a0, a1, a2);
1473        }
1474        break;
1475    case INDEX_op_shl_i64:
1476        if (c2) {
1477            tcg_out_opc_slli_d(s, a0, a1, a2 & 0x3f);
1478        } else {
1479            tcg_out_opc_sll_d(s, a0, a1, a2);
1480        }
1481        break;
1482
1483    case INDEX_op_shr_i32:
1484        if (c2) {
1485            tcg_out_opc_srli_w(s, a0, a1, a2 & 0x1f);
1486        } else {
1487            tcg_out_opc_srl_w(s, a0, a1, a2);
1488        }
1489        break;
1490    case INDEX_op_shr_i64:
1491        if (c2) {
1492            tcg_out_opc_srli_d(s, a0, a1, a2 & 0x3f);
1493        } else {
1494            tcg_out_opc_srl_d(s, a0, a1, a2);
1495        }
1496        break;
1497
1498    case INDEX_op_sar_i32:
1499        if (c2) {
1500            tcg_out_opc_srai_w(s, a0, a1, a2 & 0x1f);
1501        } else {
1502            tcg_out_opc_sra_w(s, a0, a1, a2);
1503        }
1504        break;
1505    case INDEX_op_sar_i64:
1506        if (c2) {
1507            tcg_out_opc_srai_d(s, a0, a1, a2 & 0x3f);
1508        } else {
1509            tcg_out_opc_sra_d(s, a0, a1, a2);
1510        }
1511        break;
1512
1513    case INDEX_op_rotl_i32:
1514        /* transform into equivalent rotr/rotri */
1515        if (c2) {
1516            tcg_out_opc_rotri_w(s, a0, a1, (32 - a2) & 0x1f);
1517        } else {
1518            tcg_out_opc_sub_w(s, TCG_REG_TMP0, TCG_REG_ZERO, a2);
1519            tcg_out_opc_rotr_w(s, a0, a1, TCG_REG_TMP0);
1520        }
1521        break;
1522    case INDEX_op_rotl_i64:
1523        /* transform into equivalent rotr/rotri */
1524        if (c2) {
1525            tcg_out_opc_rotri_d(s, a0, a1, (64 - a2) & 0x3f);
1526        } else {
1527            tcg_out_opc_sub_w(s, TCG_REG_TMP0, TCG_REG_ZERO, a2);
1528            tcg_out_opc_rotr_d(s, a0, a1, TCG_REG_TMP0);
1529        }
1530        break;
1531
1532    case INDEX_op_rotr_i32:
1533        if (c2) {
1534            tcg_out_opc_rotri_w(s, a0, a1, a2 & 0x1f);
1535        } else {
1536            tcg_out_opc_rotr_w(s, a0, a1, a2);
1537        }
1538        break;
1539    case INDEX_op_rotr_i64:
1540        if (c2) {
1541            tcg_out_opc_rotri_d(s, a0, a1, a2 & 0x3f);
1542        } else {
1543            tcg_out_opc_rotr_d(s, a0, a1, a2);
1544        }
1545        break;
1546
1547    case INDEX_op_add_i32:
1548        if (c2) {
1549            tcg_out_addi(s, TCG_TYPE_I32, a0, a1, a2);
1550        } else {
1551            tcg_out_opc_add_w(s, a0, a1, a2);
1552        }
1553        break;
1554    case INDEX_op_add_i64:
1555        if (c2) {
1556            tcg_out_addi(s, TCG_TYPE_I64, a0, a1, a2);
1557        } else {
1558            tcg_out_opc_add_d(s, a0, a1, a2);
1559        }
1560        break;
1561
1562    case INDEX_op_sub_i32:
1563        if (c2) {
1564            tcg_out_addi(s, TCG_TYPE_I32, a0, a1, -a2);
1565        } else {
1566            tcg_out_opc_sub_w(s, a0, a1, a2);
1567        }
1568        break;
1569    case INDEX_op_sub_i64:
1570        if (c2) {
1571            tcg_out_addi(s, TCG_TYPE_I64, a0, a1, -a2);
1572        } else {
1573            tcg_out_opc_sub_d(s, a0, a1, a2);
1574        }
1575        break;
1576
1577    case INDEX_op_neg_i32:
1578        tcg_out_opc_sub_w(s, a0, TCG_REG_ZERO, a1);
1579        break;
1580    case INDEX_op_neg_i64:
1581        tcg_out_opc_sub_d(s, a0, TCG_REG_ZERO, a1);
1582        break;
1583
1584    case INDEX_op_mul_i32:
1585        tcg_out_opc_mul_w(s, a0, a1, a2);
1586        break;
1587    case INDEX_op_mul_i64:
1588        tcg_out_opc_mul_d(s, a0, a1, a2);
1589        break;
1590
1591    case INDEX_op_mulsh_i32:
1592        tcg_out_opc_mulh_w(s, a0, a1, a2);
1593        break;
1594    case INDEX_op_mulsh_i64:
1595        tcg_out_opc_mulh_d(s, a0, a1, a2);
1596        break;
1597
1598    case INDEX_op_muluh_i32:
1599        tcg_out_opc_mulh_wu(s, a0, a1, a2);
1600        break;
1601    case INDEX_op_muluh_i64:
1602        tcg_out_opc_mulh_du(s, a0, a1, a2);
1603        break;
1604
1605    case INDEX_op_div_i32:
1606        tcg_out_opc_div_w(s, a0, a1, a2);
1607        break;
1608    case INDEX_op_div_i64:
1609        tcg_out_opc_div_d(s, a0, a1, a2);
1610        break;
1611
1612    case INDEX_op_divu_i32:
1613        tcg_out_opc_div_wu(s, a0, a1, a2);
1614        break;
1615    case INDEX_op_divu_i64:
1616        tcg_out_opc_div_du(s, a0, a1, a2);
1617        break;
1618
1619    case INDEX_op_rem_i32:
1620        tcg_out_opc_mod_w(s, a0, a1, a2);
1621        break;
1622    case INDEX_op_rem_i64:
1623        tcg_out_opc_mod_d(s, a0, a1, a2);
1624        break;
1625
1626    case INDEX_op_remu_i32:
1627        tcg_out_opc_mod_wu(s, a0, a1, a2);
1628        break;
1629    case INDEX_op_remu_i64:
1630        tcg_out_opc_mod_du(s, a0, a1, a2);
1631        break;
1632
1633    case INDEX_op_setcond_i32:
1634    case INDEX_op_setcond_i64:
1635        tcg_out_setcond(s, args[3], a0, a1, a2, c2);
1636        break;
1637
1638    case INDEX_op_movcond_i32:
1639    case INDEX_op_movcond_i64:
1640        tcg_out_movcond(s, args[5], a0, a1, a2, c2, args[3], args[4]);
1641        break;
1642
1643    case INDEX_op_ld8s_i32:
1644    case INDEX_op_ld8s_i64:
1645        tcg_out_ldst(s, OPC_LD_B, a0, a1, a2);
1646        break;
1647    case INDEX_op_ld8u_i32:
1648    case INDEX_op_ld8u_i64:
1649        tcg_out_ldst(s, OPC_LD_BU, a0, a1, a2);
1650        break;
1651    case INDEX_op_ld16s_i32:
1652    case INDEX_op_ld16s_i64:
1653        tcg_out_ldst(s, OPC_LD_H, a0, a1, a2);
1654        break;
1655    case INDEX_op_ld16u_i32:
1656    case INDEX_op_ld16u_i64:
1657        tcg_out_ldst(s, OPC_LD_HU, a0, a1, a2);
1658        break;
1659    case INDEX_op_ld_i32:
1660    case INDEX_op_ld32s_i64:
1661        tcg_out_ldst(s, OPC_LD_W, a0, a1, a2);
1662        break;
1663    case INDEX_op_ld32u_i64:
1664        tcg_out_ldst(s, OPC_LD_WU, a0, a1, a2);
1665        break;
1666    case INDEX_op_ld_i64:
1667        tcg_out_ldst(s, OPC_LD_D, a0, a1, a2);
1668        break;
1669
1670    case INDEX_op_st8_i32:
1671    case INDEX_op_st8_i64:
1672        tcg_out_ldst(s, OPC_ST_B, a0, a1, a2);
1673        break;
1674    case INDEX_op_st16_i32:
1675    case INDEX_op_st16_i64:
1676        tcg_out_ldst(s, OPC_ST_H, a0, a1, a2);
1677        break;
1678    case INDEX_op_st_i32:
1679    case INDEX_op_st32_i64:
1680        tcg_out_ldst(s, OPC_ST_W, a0, a1, a2);
1681        break;
1682    case INDEX_op_st_i64:
1683        tcg_out_ldst(s, OPC_ST_D, a0, a1, a2);
1684        break;
1685
1686    case INDEX_op_qemu_ld_i32:
1687        tcg_out_qemu_ld(s, a0, a1, a2, TCG_TYPE_I32);
1688        break;
1689    case INDEX_op_qemu_ld_i64:
1690        tcg_out_qemu_ld(s, a0, a1, a2, TCG_TYPE_I64);
1691        break;
1692    case INDEX_op_qemu_ld_i128:
1693        tcg_out_qemu_ldst_i128(s, a0, a1, a2, a3, true);
1694        break;
1695    case INDEX_op_qemu_st_i32:
1696        tcg_out_qemu_st(s, a0, a1, a2, TCG_TYPE_I32);
1697        break;
1698    case INDEX_op_qemu_st_i64:
1699        tcg_out_qemu_st(s, a0, a1, a2, TCG_TYPE_I64);
1700        break;
1701    case INDEX_op_qemu_st_i128:
1702        tcg_out_qemu_ldst_i128(s, a0, a1, a2, a3, false);
1703        break;
1704
1705    case INDEX_op_mov_i32:  /* Always emitted via tcg_out_mov.  */
1706    case INDEX_op_mov_i64:
1707    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
1708    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
1709    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
1710    case INDEX_op_ext_i32_i64:  /* Always emitted via tcg_reg_alloc_op.  */
1711    case INDEX_op_extu_i32_i64:
1712    case INDEX_op_extrl_i64_i32:
1713    default:
1714        g_assert_not_reached();
1715    }
1716}
1717
1718static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
1719                            TCGReg rd, TCGReg rs)
1720{
1721    static const LoongArchInsn repl_insn[2][4] = {
1722        { OPC_VREPLGR2VR_B, OPC_VREPLGR2VR_H,
1723          OPC_VREPLGR2VR_W, OPC_VREPLGR2VR_D },
1724        { OPC_XVREPLGR2VR_B, OPC_XVREPLGR2VR_H,
1725          OPC_XVREPLGR2VR_W, OPC_XVREPLGR2VR_D },
1726    };
1727    bool lasx = type == TCG_TYPE_V256;
1728
1729    tcg_debug_assert(vece <= MO_64);
1730    tcg_out32(s, encode_vdj_insn(repl_insn[lasx][vece], rd, rs));
1731    return true;
1732}
1733
1734static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
1735                             TCGReg r, TCGReg base, intptr_t offset)
1736{
1737    bool lasx = type == TCG_TYPE_V256;
1738
1739    /* Handle imm overflow and division (vldrepl.d imm is divided by 8). */
1740    if (offset < -0x800 || offset > 0x7ff ||
1741        (offset & ((1 << vece) - 1)) != 0) {
1742        tcg_out_addi(s, TCG_TYPE_I64, TCG_REG_TMP0, base, offset);
1743        base = TCG_REG_TMP0;
1744        offset = 0;
1745    }
1746    offset >>= vece;
1747
1748    switch (vece) {
1749    case MO_8:
1750        if (lasx) {
1751            tcg_out_opc_xvldrepl_b(s, r, base, offset);
1752        } else {
1753            tcg_out_opc_vldrepl_b(s, r, base, offset);
1754        }
1755        break;
1756    case MO_16:
1757        if (lasx) {
1758            tcg_out_opc_xvldrepl_h(s, r, base, offset);
1759        } else {
1760            tcg_out_opc_vldrepl_h(s, r, base, offset);
1761        }
1762        break;
1763    case MO_32:
1764        if (lasx) {
1765            tcg_out_opc_xvldrepl_w(s, r, base, offset);
1766        } else {
1767            tcg_out_opc_vldrepl_w(s, r, base, offset);
1768        }
1769        break;
1770    case MO_64:
1771        if (lasx) {
1772            tcg_out_opc_xvldrepl_d(s, r, base, offset);
1773        } else {
1774            tcg_out_opc_vldrepl_d(s, r, base, offset);
1775        }
1776        break;
1777    default:
1778        g_assert_not_reached();
1779    }
1780    return true;
1781}
1782
1783static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
1784                             TCGReg rd, int64_t v64)
1785{
1786    /* Try vldi if imm can fit */
1787    int64_t value = sextract64(v64, 0, 8 << vece);
1788    if (-0x200 <= value && value <= 0x1FF) {
1789        uint32_t imm = (vece << 10) | ((uint32_t)v64 & 0x3FF);
1790
1791        if (type == TCG_TYPE_V256) {
1792            tcg_out_opc_xvldi(s, rd, imm);
1793        } else {
1794            tcg_out_opc_vldi(s, rd, imm);
1795        }
1796        return;
1797    }
1798
1799    /* TODO: vldi patterns when imm 12 is set */
1800
1801    tcg_out_movi(s, TCG_TYPE_I64, TCG_REG_TMP0, value);
1802    tcg_out_dup_vec(s, type, vece, rd, TCG_REG_TMP0);
1803}
1804
1805static void tcg_out_addsub_vec(TCGContext *s, bool lasx, unsigned vece,
1806                               TCGArg a0, TCGArg a1, TCGArg a2,
1807                               bool a2_is_const, bool is_add)
1808{
1809    static const LoongArchInsn add_vec_insn[2][4] = {
1810        { OPC_VADD_B, OPC_VADD_H, OPC_VADD_W, OPC_VADD_D },
1811        { OPC_XVADD_B, OPC_XVADD_H, OPC_XVADD_W, OPC_XVADD_D },
1812    };
1813    static const LoongArchInsn add_vec_imm_insn[2][4] = {
1814        { OPC_VADDI_BU, OPC_VADDI_HU, OPC_VADDI_WU, OPC_VADDI_DU },
1815        { OPC_XVADDI_BU, OPC_XVADDI_HU, OPC_XVADDI_WU, OPC_XVADDI_DU },
1816    };
1817    static const LoongArchInsn sub_vec_insn[2][4] = {
1818        { OPC_VSUB_B, OPC_VSUB_H, OPC_VSUB_W, OPC_VSUB_D },
1819        { OPC_XVSUB_B, OPC_XVSUB_H, OPC_XVSUB_W, OPC_XVSUB_D },
1820    };
1821    static const LoongArchInsn sub_vec_imm_insn[2][4] = {
1822        { OPC_VSUBI_BU, OPC_VSUBI_HU, OPC_VSUBI_WU, OPC_VSUBI_DU },
1823        { OPC_XVSUBI_BU, OPC_XVSUBI_HU, OPC_XVSUBI_WU, OPC_XVSUBI_DU },
1824    };
1825    LoongArchInsn insn;
1826
1827    if (a2_is_const) {
1828        int64_t value = sextract64(a2, 0, 8 << vece);
1829
1830        if (!is_add) {
1831            value = -value;
1832        }
1833        if (value < 0) {
1834            insn = sub_vec_imm_insn[lasx][vece];
1835            value = -value;
1836        } else {
1837            insn = add_vec_imm_insn[lasx][vece];
1838        }
1839
1840        /* Constraint TCG_CT_CONST_VADD ensures validity. */
1841        tcg_debug_assert(0 <= value && value <= 0x1f);
1842
1843        tcg_out32(s, encode_vdvjuk5_insn(insn, a0, a1, value));
1844        return;
1845    }
1846
1847    if (is_add) {
1848        insn = add_vec_insn[lasx][vece];
1849    } else {
1850        insn = sub_vec_insn[lasx][vece];
1851    }
1852    tcg_out32(s, encode_vdvjvk_insn(insn, a0, a1, a2));
1853}
1854
1855static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
1856                           unsigned vecl, unsigned vece,
1857                           const TCGArg args[TCG_MAX_OP_ARGS],
1858                           const int const_args[TCG_MAX_OP_ARGS])
1859{
1860    TCGType type = vecl + TCG_TYPE_V64;
1861    bool lasx = type == TCG_TYPE_V256;
1862    TCGArg a0, a1, a2, a3;
1863    LoongArchInsn insn;
1864
1865    static const LoongArchInsn cmp_vec_insn[16][2][4] = {
1866        [TCG_COND_EQ] = {
1867            { OPC_VSEQ_B, OPC_VSEQ_H, OPC_VSEQ_W, OPC_VSEQ_D },
1868            { OPC_XVSEQ_B, OPC_XVSEQ_H, OPC_XVSEQ_W, OPC_XVSEQ_D },
1869        },
1870        [TCG_COND_LE] = {
1871            { OPC_VSLE_B, OPC_VSLE_H, OPC_VSLE_W, OPC_VSLE_D },
1872            { OPC_XVSLE_B, OPC_XVSLE_H, OPC_XVSLE_W, OPC_XVSLE_D },
1873        },
1874        [TCG_COND_LEU] = {
1875            { OPC_VSLE_BU, OPC_VSLE_HU, OPC_VSLE_WU, OPC_VSLE_DU },
1876            { OPC_XVSLE_BU, OPC_XVSLE_HU, OPC_XVSLE_WU, OPC_XVSLE_DU },
1877        },
1878        [TCG_COND_LT] = {
1879            { OPC_VSLT_B, OPC_VSLT_H, OPC_VSLT_W, OPC_VSLT_D },
1880            { OPC_XVSLT_B, OPC_XVSLT_H, OPC_XVSLT_W, OPC_XVSLT_D },
1881        },
1882        [TCG_COND_LTU] = {
1883            { OPC_VSLT_BU, OPC_VSLT_HU, OPC_VSLT_WU, OPC_VSLT_DU },
1884            { OPC_XVSLT_BU, OPC_XVSLT_HU, OPC_XVSLT_WU, OPC_XVSLT_DU },
1885        }
1886    };
1887    static const LoongArchInsn cmp_vec_imm_insn[16][2][4] = {
1888        [TCG_COND_EQ] = {
1889            { OPC_VSEQI_B, OPC_VSEQI_H, OPC_VSEQI_W, OPC_VSEQI_D },
1890            { OPC_XVSEQI_B, OPC_XVSEQI_H, OPC_XVSEQI_W, OPC_XVSEQI_D },
1891        },
1892        [TCG_COND_LE] = {
1893            { OPC_VSLEI_B, OPC_VSLEI_H, OPC_VSLEI_W, OPC_VSLEI_D },
1894            { OPC_XVSLEI_B, OPC_XVSLEI_H, OPC_XVSLEI_W, OPC_XVSLEI_D },
1895        },
1896        [TCG_COND_LEU] = {
1897            { OPC_VSLEI_BU, OPC_VSLEI_HU, OPC_VSLEI_WU, OPC_VSLEI_DU },
1898            { OPC_XVSLEI_BU, OPC_XVSLEI_HU, OPC_XVSLEI_WU, OPC_XVSLEI_DU },
1899        },
1900        [TCG_COND_LT] = {
1901            { OPC_VSLTI_B, OPC_VSLTI_H, OPC_VSLTI_W, OPC_VSLTI_D },
1902            { OPC_XVSLTI_B, OPC_XVSLTI_H, OPC_XVSLTI_W, OPC_XVSLTI_D },
1903        },
1904        [TCG_COND_LTU] = {
1905            { OPC_VSLTI_BU, OPC_VSLTI_HU, OPC_VSLTI_WU, OPC_VSLTI_DU },
1906            { OPC_XVSLTI_BU, OPC_XVSLTI_HU, OPC_XVSLTI_WU, OPC_XVSLTI_DU },
1907        }
1908    };
1909    static const LoongArchInsn neg_vec_insn[2][4] = {
1910        { OPC_VNEG_B, OPC_VNEG_H, OPC_VNEG_W, OPC_VNEG_D },
1911        { OPC_XVNEG_B, OPC_XVNEG_H, OPC_XVNEG_W, OPC_XVNEG_D },
1912    };
1913    static const LoongArchInsn mul_vec_insn[2][4] = {
1914        { OPC_VMUL_B, OPC_VMUL_H, OPC_VMUL_W, OPC_VMUL_D },
1915        { OPC_XVMUL_B, OPC_XVMUL_H, OPC_XVMUL_W, OPC_XVMUL_D },
1916    };
1917    static const LoongArchInsn smin_vec_insn[2][4] = {
1918        { OPC_VMIN_B, OPC_VMIN_H, OPC_VMIN_W, OPC_VMIN_D },
1919        { OPC_XVMIN_B, OPC_XVMIN_H, OPC_XVMIN_W, OPC_XVMIN_D },
1920    };
1921    static const LoongArchInsn umin_vec_insn[2][4] = {
1922        { OPC_VMIN_BU, OPC_VMIN_HU, OPC_VMIN_WU, OPC_VMIN_DU },
1923        { OPC_XVMIN_BU, OPC_XVMIN_HU, OPC_XVMIN_WU, OPC_XVMIN_DU },
1924    };
1925    static const LoongArchInsn smax_vec_insn[2][4] = {
1926        { OPC_VMAX_B, OPC_VMAX_H, OPC_VMAX_W, OPC_VMAX_D },
1927        { OPC_XVMAX_B, OPC_XVMAX_H, OPC_XVMAX_W, OPC_XVMAX_D },
1928    };
1929    static const LoongArchInsn umax_vec_insn[2][4] = {
1930        { OPC_VMAX_BU, OPC_VMAX_HU, OPC_VMAX_WU, OPC_VMAX_DU },
1931        { OPC_XVMAX_BU, OPC_XVMAX_HU, OPC_XVMAX_WU, OPC_XVMAX_DU },
1932    };
1933    static const LoongArchInsn ssadd_vec_insn[2][4] = {
1934        { OPC_VSADD_B, OPC_VSADD_H, OPC_VSADD_W, OPC_VSADD_D },
1935        { OPC_XVSADD_B, OPC_XVSADD_H, OPC_XVSADD_W, OPC_XVSADD_D },
1936    };
1937    static const LoongArchInsn usadd_vec_insn[2][4] = {
1938        { OPC_VSADD_BU, OPC_VSADD_HU, OPC_VSADD_WU, OPC_VSADD_DU },
1939        { OPC_XVSADD_BU, OPC_XVSADD_HU, OPC_XVSADD_WU, OPC_XVSADD_DU },
1940    };
1941    static const LoongArchInsn sssub_vec_insn[2][4] = {
1942        { OPC_VSSUB_B, OPC_VSSUB_H, OPC_VSSUB_W, OPC_VSSUB_D },
1943        { OPC_XVSSUB_B, OPC_XVSSUB_H, OPC_XVSSUB_W, OPC_XVSSUB_D },
1944    };
1945    static const LoongArchInsn ussub_vec_insn[2][4] = {
1946        { OPC_VSSUB_BU, OPC_VSSUB_HU, OPC_VSSUB_WU, OPC_VSSUB_DU },
1947        { OPC_XVSSUB_BU, OPC_XVSSUB_HU, OPC_XVSSUB_WU, OPC_XVSSUB_DU },
1948    };
1949    static const LoongArchInsn shlv_vec_insn[2][4] = {
1950        { OPC_VSLL_B, OPC_VSLL_H, OPC_VSLL_W, OPC_VSLL_D },
1951        { OPC_XVSLL_B, OPC_XVSLL_H, OPC_XVSLL_W, OPC_XVSLL_D },
1952    };
1953    static const LoongArchInsn shrv_vec_insn[2][4] = {
1954        { OPC_VSRL_B, OPC_VSRL_H, OPC_VSRL_W, OPC_VSRL_D },
1955        { OPC_XVSRL_B, OPC_XVSRL_H, OPC_XVSRL_W, OPC_XVSRL_D },
1956    };
1957    static const LoongArchInsn sarv_vec_insn[2][4] = {
1958        { OPC_VSRA_B, OPC_VSRA_H, OPC_VSRA_W, OPC_VSRA_D },
1959        { OPC_XVSRA_B, OPC_XVSRA_H, OPC_XVSRA_W, OPC_XVSRA_D },
1960    };
1961    static const LoongArchInsn shli_vec_insn[2][4] = {
1962        { OPC_VSLLI_B, OPC_VSLLI_H, OPC_VSLLI_W, OPC_VSLLI_D },
1963        { OPC_XVSLLI_B, OPC_XVSLLI_H, OPC_XVSLLI_W, OPC_XVSLLI_D },
1964    };
1965    static const LoongArchInsn shri_vec_insn[2][4] = {
1966        { OPC_VSRLI_B, OPC_VSRLI_H, OPC_VSRLI_W, OPC_VSRLI_D },
1967        { OPC_XVSRLI_B, OPC_XVSRLI_H, OPC_XVSRLI_W, OPC_XVSRLI_D },
1968    };
1969    static const LoongArchInsn sari_vec_insn[2][4] = {
1970        { OPC_VSRAI_B, OPC_VSRAI_H, OPC_VSRAI_W, OPC_VSRAI_D },
1971        { OPC_XVSRAI_B, OPC_XVSRAI_H, OPC_XVSRAI_W, OPC_XVSRAI_D },
1972    };
1973    static const LoongArchInsn rotrv_vec_insn[2][4] = {
1974        { OPC_VROTR_B, OPC_VROTR_H, OPC_VROTR_W, OPC_VROTR_D },
1975        { OPC_XVROTR_B, OPC_XVROTR_H, OPC_XVROTR_W, OPC_XVROTR_D },
1976    };
1977    static const LoongArchInsn rotri_vec_insn[2][4] = {
1978        { OPC_VROTRI_B, OPC_VROTRI_H, OPC_VROTRI_W, OPC_VROTRI_D },
1979        { OPC_XVROTRI_B, OPC_XVROTRI_H, OPC_XVROTRI_W, OPC_XVROTRI_D },
1980    };
1981
1982    a0 = args[0];
1983    a1 = args[1];
1984    a2 = args[2];
1985    a3 = args[3];
1986
1987    switch (opc) {
1988    case INDEX_op_st_vec:
1989        tcg_out_st(s, type, a0, a1, a2);
1990        break;
1991    case INDEX_op_ld_vec:
1992        tcg_out_ld(s, type, a0, a1, a2);
1993        break;
1994    case INDEX_op_and_vec:
1995        insn = lasx ? OPC_XVAND_V : OPC_VAND_V;
1996        goto vdvjvk;
1997    case INDEX_op_andc_vec:
1998        /*
1999         * vandn vd, vj, vk: vd = vk & ~vj
2000         * andc_vec vd, vj, vk: vd = vj & ~vk
2001         * vj and vk are swapped
2002         */
2003        a1 = a2;
2004        a2 = args[1];
2005        insn = lasx ? OPC_XVANDN_V : OPC_VANDN_V;
2006        goto vdvjvk;
2007    case INDEX_op_or_vec:
2008        insn = lasx ? OPC_XVOR_V : OPC_VOR_V;
2009        goto vdvjvk;
2010    case INDEX_op_orc_vec:
2011        insn = lasx ? OPC_XVORN_V : OPC_VORN_V;
2012        goto vdvjvk;
2013    case INDEX_op_xor_vec:
2014        insn = lasx ? OPC_XVXOR_V : OPC_VXOR_V;
2015        goto vdvjvk;
2016    case INDEX_op_not_vec:
2017        a2 = a1;
2018        /* fall through */
2019    case INDEX_op_nor_vec:
2020        insn = lasx ? OPC_XVNOR_V : OPC_VNOR_V;
2021        goto vdvjvk;
2022    case INDEX_op_cmp_vec:
2023        {
2024            TCGCond cond = args[3];
2025
2026            if (const_args[2]) {
2027                /*
2028                 * cmp_vec dest, src, value
2029                 * Try vseqi/vslei/vslti
2030                 */
2031                int64_t value = sextract64(a2, 0, 8 << vece);
2032                switch (cond) {
2033                case TCG_COND_EQ:
2034                case TCG_COND_LE:
2035                case TCG_COND_LT:
2036                    insn = cmp_vec_imm_insn[cond][lasx][vece];
2037                    tcg_out32(s, encode_vdvjsk5_insn(insn, a0, a1, value));
2038                    break;
2039                case TCG_COND_LEU:
2040                case TCG_COND_LTU:
2041                    insn = cmp_vec_imm_insn[cond][lasx][vece];
2042                    tcg_out32(s, encode_vdvjuk5_insn(insn, a0, a1, value));
2043                    break;
2044                default:
2045                    g_assert_not_reached();
2046                }
2047                break;
2048            }
2049
2050            insn = cmp_vec_insn[cond][lasx][vece];
2051            if (insn == 0) {
2052                TCGArg t;
2053                t = a1, a1 = a2, a2 = t;
2054                cond = tcg_swap_cond(cond);
2055                insn = cmp_vec_insn[cond][lasx][vece];
2056                tcg_debug_assert(insn != 0);
2057            }
2058        }
2059        goto vdvjvk;
2060    case INDEX_op_add_vec:
2061        tcg_out_addsub_vec(s, lasx, vece, a0, a1, a2, const_args[2], true);
2062        break;
2063    case INDEX_op_sub_vec:
2064        tcg_out_addsub_vec(s, lasx, vece, a0, a1, a2, const_args[2], false);
2065        break;
2066    case INDEX_op_neg_vec:
2067        tcg_out32(s, encode_vdvj_insn(neg_vec_insn[lasx][vece], a0, a1));
2068        break;
2069    case INDEX_op_mul_vec:
2070        insn = mul_vec_insn[lasx][vece];
2071        goto vdvjvk;
2072    case INDEX_op_smin_vec:
2073        insn = smin_vec_insn[lasx][vece];
2074        goto vdvjvk;
2075    case INDEX_op_smax_vec:
2076        insn = smax_vec_insn[lasx][vece];
2077        goto vdvjvk;
2078    case INDEX_op_umin_vec:
2079        insn = umin_vec_insn[lasx][vece];
2080        goto vdvjvk;
2081    case INDEX_op_umax_vec:
2082        insn = umax_vec_insn[lasx][vece];
2083        goto vdvjvk;
2084    case INDEX_op_ssadd_vec:
2085        insn = ssadd_vec_insn[lasx][vece];
2086        goto vdvjvk;
2087    case INDEX_op_usadd_vec:
2088        insn = usadd_vec_insn[lasx][vece];
2089        goto vdvjvk;
2090    case INDEX_op_sssub_vec:
2091        insn = sssub_vec_insn[lasx][vece];
2092        goto vdvjvk;
2093    case INDEX_op_ussub_vec:
2094        insn = ussub_vec_insn[lasx][vece];
2095        goto vdvjvk;
2096    case INDEX_op_shlv_vec:
2097        insn = shlv_vec_insn[lasx][vece];
2098        goto vdvjvk;
2099    case INDEX_op_shrv_vec:
2100        insn = shrv_vec_insn[lasx][vece];
2101        goto vdvjvk;
2102    case INDEX_op_sarv_vec:
2103        insn = sarv_vec_insn[lasx][vece];
2104        goto vdvjvk;
2105    case INDEX_op_rotlv_vec:
2106        /* rotlv_vec a1, a2 = rotrv_vec a1, -a2 */
2107        tcg_out32(s, encode_vdvj_insn(neg_vec_insn[lasx][vece],
2108                                      TCG_VEC_TMP0, a2));
2109        a2 = TCG_VEC_TMP0;
2110        /* fall through */
2111    case INDEX_op_rotrv_vec:
2112        insn = rotrv_vec_insn[lasx][vece];
2113        goto vdvjvk;
2114    case INDEX_op_shli_vec:
2115        insn = shli_vec_insn[lasx][vece];
2116        goto vdvjukN;
2117    case INDEX_op_shri_vec:
2118        insn = shri_vec_insn[lasx][vece];
2119        goto vdvjukN;
2120    case INDEX_op_sari_vec:
2121        insn = sari_vec_insn[lasx][vece];
2122        goto vdvjukN;
2123    case INDEX_op_rotli_vec:
2124        /* rotli_vec a1, a2 = rotri_vec a1, -a2 */
2125        a2 = extract32(-a2, 0, 3 + vece);
2126        insn = rotri_vec_insn[lasx][vece];
2127        goto vdvjukN;
2128    case INDEX_op_bitsel_vec:
2129        /* vbitsel vd, vj, vk, va = bitsel_vec vd, va, vk, vj */
2130        if (lasx) {
2131            tcg_out_opc_xvbitsel_v(s, a0, a3, a2, a1);
2132        } else {
2133            tcg_out_opc_vbitsel_v(s, a0, a3, a2, a1);
2134        }
2135        break;
2136    case INDEX_op_dupm_vec:
2137        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
2138        break;
2139    default:
2140        g_assert_not_reached();
2141    vdvjvk:
2142        tcg_out32(s, encode_vdvjvk_insn(insn, a0, a1, a2));
2143        break;
2144    vdvjukN:
2145        switch (vece) {
2146        case MO_8:
2147            tcg_out32(s, encode_vdvjuk3_insn(insn, a0, a1, a2));
2148            break;
2149        case MO_16:
2150            tcg_out32(s, encode_vdvjuk4_insn(insn, a0, a1, a2));
2151            break;
2152        case MO_32:
2153            tcg_out32(s, encode_vdvjuk5_insn(insn, a0, a1, a2));
2154            break;
2155        case MO_64:
2156            tcg_out32(s, encode_vdvjuk6_insn(insn, a0, a1, a2));
2157            break;
2158        default:
2159            g_assert_not_reached();
2160        }
2161        break;
2162    }
2163}
2164
2165int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
2166{
2167    switch (opc) {
2168    case INDEX_op_ld_vec:
2169    case INDEX_op_st_vec:
2170    case INDEX_op_dup_vec:
2171    case INDEX_op_dupm_vec:
2172    case INDEX_op_cmp_vec:
2173    case INDEX_op_add_vec:
2174    case INDEX_op_sub_vec:
2175    case INDEX_op_and_vec:
2176    case INDEX_op_andc_vec:
2177    case INDEX_op_or_vec:
2178    case INDEX_op_orc_vec:
2179    case INDEX_op_xor_vec:
2180    case INDEX_op_nor_vec:
2181    case INDEX_op_not_vec:
2182    case INDEX_op_neg_vec:
2183    case INDEX_op_mul_vec:
2184    case INDEX_op_smin_vec:
2185    case INDEX_op_smax_vec:
2186    case INDEX_op_umin_vec:
2187    case INDEX_op_umax_vec:
2188    case INDEX_op_ssadd_vec:
2189    case INDEX_op_usadd_vec:
2190    case INDEX_op_sssub_vec:
2191    case INDEX_op_ussub_vec:
2192    case INDEX_op_shlv_vec:
2193    case INDEX_op_shrv_vec:
2194    case INDEX_op_sarv_vec:
2195    case INDEX_op_bitsel_vec:
2196        return 1;
2197    default:
2198        return 0;
2199    }
2200}
2201
2202void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
2203                       TCGArg a0, ...)
2204{
2205    g_assert_not_reached();
2206}
2207
2208static TCGConstraintSetIndex
2209tcg_target_op_def(TCGOpcode op, TCGType type, unsigned flags)
2210{
2211    switch (op) {
2212    case INDEX_op_goto_ptr:
2213        return C_O0_I1(r);
2214
2215    case INDEX_op_st8_i32:
2216    case INDEX_op_st8_i64:
2217    case INDEX_op_st16_i32:
2218    case INDEX_op_st16_i64:
2219    case INDEX_op_st32_i64:
2220    case INDEX_op_st_i32:
2221    case INDEX_op_st_i64:
2222    case INDEX_op_qemu_st_i32:
2223    case INDEX_op_qemu_st_i64:
2224        return C_O0_I2(rz, r);
2225
2226    case INDEX_op_qemu_ld_i128:
2227        return C_N2_I1(r, r, r);
2228
2229    case INDEX_op_qemu_st_i128:
2230        return C_O0_I3(r, r, r);
2231
2232    case INDEX_op_brcond_i32:
2233    case INDEX_op_brcond_i64:
2234        return C_O0_I2(rz, rz);
2235
2236    case INDEX_op_extu_i32_i64:
2237    case INDEX_op_extrl_i64_i32:
2238    case INDEX_op_extrh_i64_i32:
2239    case INDEX_op_ext_i32_i64:
2240    case INDEX_op_neg_i32:
2241    case INDEX_op_neg_i64:
2242    case INDEX_op_not_i32:
2243    case INDEX_op_not_i64:
2244    case INDEX_op_extract_i32:
2245    case INDEX_op_extract_i64:
2246    case INDEX_op_sextract_i32:
2247    case INDEX_op_sextract_i64:
2248    case INDEX_op_bswap16_i32:
2249    case INDEX_op_bswap16_i64:
2250    case INDEX_op_bswap32_i32:
2251    case INDEX_op_bswap32_i64:
2252    case INDEX_op_bswap64_i64:
2253    case INDEX_op_ld8s_i32:
2254    case INDEX_op_ld8s_i64:
2255    case INDEX_op_ld8u_i32:
2256    case INDEX_op_ld8u_i64:
2257    case INDEX_op_ld16s_i32:
2258    case INDEX_op_ld16s_i64:
2259    case INDEX_op_ld16u_i32:
2260    case INDEX_op_ld16u_i64:
2261    case INDEX_op_ld32s_i64:
2262    case INDEX_op_ld32u_i64:
2263    case INDEX_op_ld_i32:
2264    case INDEX_op_ld_i64:
2265    case INDEX_op_qemu_ld_i32:
2266    case INDEX_op_qemu_ld_i64:
2267        return C_O1_I1(r, r);
2268
2269    case INDEX_op_andc_i32:
2270    case INDEX_op_andc_i64:
2271    case INDEX_op_orc_i32:
2272    case INDEX_op_orc_i64:
2273        /*
2274         * LoongArch insns for these ops don't have reg-imm forms, but we
2275         * can express using andi/ori if ~constant satisfies
2276         * TCG_CT_CONST_U12.
2277         */
2278        return C_O1_I2(r, r, rC);
2279
2280    case INDEX_op_shl_i32:
2281    case INDEX_op_shl_i64:
2282    case INDEX_op_shr_i32:
2283    case INDEX_op_shr_i64:
2284    case INDEX_op_sar_i32:
2285    case INDEX_op_sar_i64:
2286    case INDEX_op_rotl_i32:
2287    case INDEX_op_rotl_i64:
2288    case INDEX_op_rotr_i32:
2289    case INDEX_op_rotr_i64:
2290        return C_O1_I2(r, r, ri);
2291
2292    case INDEX_op_add_i32:
2293        return C_O1_I2(r, r, ri);
2294    case INDEX_op_add_i64:
2295        return C_O1_I2(r, r, rJ);
2296
2297    case INDEX_op_and_i32:
2298    case INDEX_op_and_i64:
2299    case INDEX_op_nor_i32:
2300    case INDEX_op_nor_i64:
2301    case INDEX_op_or_i32:
2302    case INDEX_op_or_i64:
2303    case INDEX_op_xor_i32:
2304    case INDEX_op_xor_i64:
2305        /* LoongArch reg-imm bitops have their imms ZERO-extended */
2306        return C_O1_I2(r, r, rU);
2307
2308    case INDEX_op_clz_i32:
2309    case INDEX_op_clz_i64:
2310    case INDEX_op_ctz_i32:
2311    case INDEX_op_ctz_i64:
2312        return C_O1_I2(r, r, rW);
2313
2314    case INDEX_op_deposit_i32:
2315    case INDEX_op_deposit_i64:
2316        /* Must deposit into the same register as input */
2317        return C_O1_I2(r, 0, rz);
2318
2319    case INDEX_op_sub_i32:
2320    case INDEX_op_setcond_i32:
2321        return C_O1_I2(r, rz, ri);
2322    case INDEX_op_sub_i64:
2323    case INDEX_op_setcond_i64:
2324        return C_O1_I2(r, rz, rJ);
2325
2326    case INDEX_op_mul_i32:
2327    case INDEX_op_mul_i64:
2328    case INDEX_op_mulsh_i32:
2329    case INDEX_op_mulsh_i64:
2330    case INDEX_op_muluh_i32:
2331    case INDEX_op_muluh_i64:
2332    case INDEX_op_div_i32:
2333    case INDEX_op_div_i64:
2334    case INDEX_op_divu_i32:
2335    case INDEX_op_divu_i64:
2336    case INDEX_op_rem_i32:
2337    case INDEX_op_rem_i64:
2338    case INDEX_op_remu_i32:
2339    case INDEX_op_remu_i64:
2340        return C_O1_I2(r, rz, rz);
2341
2342    case INDEX_op_movcond_i32:
2343    case INDEX_op_movcond_i64:
2344        return C_O1_I4(r, rz, rJ, rz, rz);
2345
2346    case INDEX_op_ld_vec:
2347    case INDEX_op_dupm_vec:
2348    case INDEX_op_dup_vec:
2349        return C_O1_I1(w, r);
2350
2351    case INDEX_op_st_vec:
2352        return C_O0_I2(w, r);
2353
2354    case INDEX_op_cmp_vec:
2355        return C_O1_I2(w, w, wM);
2356
2357    case INDEX_op_add_vec:
2358    case INDEX_op_sub_vec:
2359        return C_O1_I2(w, w, wA);
2360
2361    case INDEX_op_and_vec:
2362    case INDEX_op_andc_vec:
2363    case INDEX_op_or_vec:
2364    case INDEX_op_orc_vec:
2365    case INDEX_op_xor_vec:
2366    case INDEX_op_nor_vec:
2367    case INDEX_op_mul_vec:
2368    case INDEX_op_smin_vec:
2369    case INDEX_op_smax_vec:
2370    case INDEX_op_umin_vec:
2371    case INDEX_op_umax_vec:
2372    case INDEX_op_ssadd_vec:
2373    case INDEX_op_usadd_vec:
2374    case INDEX_op_sssub_vec:
2375    case INDEX_op_ussub_vec:
2376    case INDEX_op_shlv_vec:
2377    case INDEX_op_shrv_vec:
2378    case INDEX_op_sarv_vec:
2379    case INDEX_op_rotrv_vec:
2380    case INDEX_op_rotlv_vec:
2381        return C_O1_I2(w, w, w);
2382
2383    case INDEX_op_not_vec:
2384    case INDEX_op_neg_vec:
2385    case INDEX_op_shli_vec:
2386    case INDEX_op_shri_vec:
2387    case INDEX_op_sari_vec:
2388    case INDEX_op_rotli_vec:
2389        return C_O1_I1(w, w);
2390
2391    case INDEX_op_bitsel_vec:
2392        return C_O1_I3(w, w, w, w);
2393
2394    default:
2395        return C_NotImplemented;
2396    }
2397}
2398
2399static const int tcg_target_callee_save_regs[] = {
2400    TCG_REG_S0,     /* used for the global env (TCG_AREG0) */
2401    TCG_REG_S1,
2402    TCG_REG_S2,
2403    TCG_REG_S3,
2404    TCG_REG_S4,
2405    TCG_REG_S5,
2406    TCG_REG_S6,
2407    TCG_REG_S7,
2408    TCG_REG_S8,
2409    TCG_REG_S9,
2410    TCG_REG_RA,     /* should be last for ABI compliance */
2411};
2412
2413/* Stack frame parameters.  */
2414#define REG_SIZE   (TCG_TARGET_REG_BITS / 8)
2415#define SAVE_SIZE  ((int)ARRAY_SIZE(tcg_target_callee_save_regs) * REG_SIZE)
2416#define TEMP_SIZE  (CPU_TEMP_BUF_NLONGS * (int)sizeof(long))
2417#define FRAME_SIZE ((TCG_STATIC_CALL_ARGS_SIZE + TEMP_SIZE + SAVE_SIZE \
2418                     + TCG_TARGET_STACK_ALIGN - 1) \
2419                    & -TCG_TARGET_STACK_ALIGN)
2420#define SAVE_OFS   (TCG_STATIC_CALL_ARGS_SIZE + TEMP_SIZE)
2421
2422/* We're expecting to be able to use an immediate for frame allocation.  */
2423QEMU_BUILD_BUG_ON(FRAME_SIZE > 0x7ff);
2424
2425/* Generate global QEMU prologue and epilogue code */
2426static void tcg_target_qemu_prologue(TCGContext *s)
2427{
2428    int i;
2429
2430    tcg_set_frame(s, TCG_REG_SP, TCG_STATIC_CALL_ARGS_SIZE, TEMP_SIZE);
2431
2432    /* TB prologue */
2433    tcg_out_opc_addi_d(s, TCG_REG_SP, TCG_REG_SP, -FRAME_SIZE);
2434    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
2435        tcg_out_st(s, TCG_TYPE_REG, tcg_target_callee_save_regs[i],
2436                   TCG_REG_SP, SAVE_OFS + i * REG_SIZE);
2437    }
2438
2439    if (!tcg_use_softmmu && guest_base) {
2440        tcg_out_movi(s, TCG_TYPE_PTR, TCG_GUEST_BASE_REG, guest_base);
2441        tcg_regset_set_reg(s->reserved_regs, TCG_GUEST_BASE_REG);
2442    }
2443
2444    /* Call generated code */
2445    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
2446    tcg_out_opc_jirl(s, TCG_REG_ZERO, tcg_target_call_iarg_regs[1], 0);
2447
2448    /* Return path for goto_ptr. Set return value to 0 */
2449    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
2450    tcg_out_mov(s, TCG_TYPE_REG, TCG_REG_A0, TCG_REG_ZERO);
2451
2452    /* TB epilogue */
2453    tb_ret_addr = tcg_splitwx_to_rx(s->code_ptr);
2454    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
2455        tcg_out_ld(s, TCG_TYPE_REG, tcg_target_callee_save_regs[i],
2456                   TCG_REG_SP, SAVE_OFS + i * REG_SIZE);
2457    }
2458
2459    tcg_out_opc_addi_d(s, TCG_REG_SP, TCG_REG_SP, FRAME_SIZE);
2460    tcg_out_opc_jirl(s, TCG_REG_ZERO, TCG_REG_RA, 0);
2461}
2462
2463static void tcg_out_tb_start(TCGContext *s)
2464{
2465    /* nothing to do */
2466}
2467
2468static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
2469{
2470    for (int i = 0; i < count; ++i) {
2471        /* Canonical nop is andi r0,r0,0 */
2472        p[i] = OPC_ANDI;
2473    }
2474}
2475
2476static void tcg_target_init(TCGContext *s)
2477{
2478    unsigned long hwcap = qemu_getauxval(AT_HWCAP);
2479
2480    /* Server and desktop class cpus have UAL; embedded cpus do not. */
2481    if (!(hwcap & HWCAP_LOONGARCH_UAL)) {
2482        error_report("TCG: unaligned access support required; exiting");
2483        exit(EXIT_FAILURE);
2484    }
2485
2486    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
2487    tcg_target_available_regs[TCG_TYPE_I64] = ALL_GENERAL_REGS;
2488
2489    tcg_target_call_clobber_regs = ALL_GENERAL_REGS | ALL_VECTOR_REGS;
2490    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S0);
2491    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S1);
2492    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S2);
2493    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S3);
2494    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S4);
2495    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S5);
2496    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S6);
2497    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S7);
2498    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S8);
2499    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S9);
2500
2501    if (cpuinfo & CPUINFO_LSX) {
2502        tcg_target_available_regs[TCG_TYPE_V64] = ALL_VECTOR_REGS;
2503        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
2504        if (cpuinfo & CPUINFO_LASX) {
2505            tcg_target_available_regs[TCG_TYPE_V256] = ALL_VECTOR_REGS;
2506        }
2507        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V24);
2508        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V25);
2509        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V26);
2510        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V27);
2511        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V28);
2512        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V29);
2513        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V30);
2514        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V31);
2515    }
2516
2517    s->reserved_regs = 0;
2518    tcg_regset_set_reg(s->reserved_regs, TCG_REG_ZERO);
2519    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP0);
2520    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP1);
2521    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP2);
2522    tcg_regset_set_reg(s->reserved_regs, TCG_REG_SP);
2523    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TP);
2524    tcg_regset_set_reg(s->reserved_regs, TCG_REG_RESERVED);
2525    tcg_regset_set_reg(s->reserved_regs, TCG_VEC_TMP0);
2526}
2527
2528typedef struct {
2529    DebugFrameHeader h;
2530    uint8_t fde_def_cfa[4];
2531    uint8_t fde_reg_ofs[ARRAY_SIZE(tcg_target_callee_save_regs) * 2];
2532} DebugFrame;
2533
2534#define ELF_HOST_MACHINE EM_LOONGARCH
2535
2536static const DebugFrame debug_frame = {
2537    .h.cie.len = sizeof(DebugFrameCIE) - 4, /* length after .len member */
2538    .h.cie.id = -1,
2539    .h.cie.version = 1,
2540    .h.cie.code_align = 1,
2541    .h.cie.data_align = -(TCG_TARGET_REG_BITS / 8) & 0x7f, /* sleb128 */
2542    .h.cie.return_column = TCG_REG_RA,
2543
2544    /* Total FDE size does not include the "len" member.  */
2545    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
2546
2547    .fde_def_cfa = {
2548        12, TCG_REG_SP,                 /* DW_CFA_def_cfa sp, ...  */
2549        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
2550        (FRAME_SIZE >> 7)
2551    },
2552    .fde_reg_ofs = {
2553        0x80 + 23, 11,                  /* DW_CFA_offset, s0, -88 */
2554        0x80 + 24, 10,                  /* DW_CFA_offset, s1, -80 */
2555        0x80 + 25, 9,                   /* DW_CFA_offset, s2, -72 */
2556        0x80 + 26, 8,                   /* DW_CFA_offset, s3, -64 */
2557        0x80 + 27, 7,                   /* DW_CFA_offset, s4, -56 */
2558        0x80 + 28, 6,                   /* DW_CFA_offset, s5, -48 */
2559        0x80 + 29, 5,                   /* DW_CFA_offset, s6, -40 */
2560        0x80 + 30, 4,                   /* DW_CFA_offset, s7, -32 */
2561        0x80 + 31, 3,                   /* DW_CFA_offset, s8, -24 */
2562        0x80 + 22, 2,                   /* DW_CFA_offset, s9, -16 */
2563        0x80 + 1 , 1,                   /* DW_CFA_offset, ra, -8 */
2564    }
2565};
2566
2567void tcg_register_jit(const void *buf, size_t buf_size)
2568{
2569    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
2570}
2571