1/*
2 * Tiny Code Generator for QEMU
3 *
4 * Copyright (c) 2021 WANG Xuerui <git@xen0n.name>
5 *
6 * Based on tcg/riscv/tcg-target.c.inc
7 *
8 * Copyright (c) 2018 SiFive, Inc
9 * Copyright (c) 2008-2009 Arnaud Patard <arnaud.patard@rtp-net.org>
10 * Copyright (c) 2009 Aurelien Jarno <aurelien@aurel32.net>
11 * Copyright (c) 2008 Fabrice Bellard
12 *
13 * Permission is hereby granted, free of charge, to any person obtaining a copy
14 * of this software and associated documentation files (the "Software"), to deal
15 * in the Software without restriction, including without limitation the rights
16 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
17 * copies of the Software, and to permit persons to whom the Software is
18 * furnished to do so, subject to the following conditions:
19 *
20 * The above copyright notice and this permission notice shall be included in
21 * all copies or substantial portions of the Software.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
24 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
25 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
26 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
27 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
28 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
29 * THE SOFTWARE.
30 */
31
32#include "../tcg-ldst.c.inc"
33#include <asm/hwcap.h>
34
35bool use_lsx_instructions;
36
37#ifdef CONFIG_DEBUG_TCG
38static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
39    "zero",
40    "ra",
41    "tp",
42    "sp",
43    "a0",
44    "a1",
45    "a2",
46    "a3",
47    "a4",
48    "a5",
49    "a6",
50    "a7",
51    "t0",
52    "t1",
53    "t2",
54    "t3",
55    "t4",
56    "t5",
57    "t6",
58    "t7",
59    "t8",
60    "r21", /* reserved in the LP64* ABI, hence no ABI name */
61    "s9",
62    "s0",
63    "s1",
64    "s2",
65    "s3",
66    "s4",
67    "s5",
68    "s6",
69    "s7",
70    "s8",
71    "vr0",
72    "vr1",
73    "vr2",
74    "vr3",
75    "vr4",
76    "vr5",
77    "vr6",
78    "vr7",
79    "vr8",
80    "vr9",
81    "vr10",
82    "vr11",
83    "vr12",
84    "vr13",
85    "vr14",
86    "vr15",
87    "vr16",
88    "vr17",
89    "vr18",
90    "vr19",
91    "vr20",
92    "vr21",
93    "vr22",
94    "vr23",
95    "vr24",
96    "vr25",
97    "vr26",
98    "vr27",
99    "vr28",
100    "vr29",
101    "vr30",
102    "vr31",
103};
104#endif
105
106static const int tcg_target_reg_alloc_order[] = {
107    /* Registers preserved across calls */
108    /* TCG_REG_S0 reserved for TCG_AREG0 */
109    TCG_REG_S1,
110    TCG_REG_S2,
111    TCG_REG_S3,
112    TCG_REG_S4,
113    TCG_REG_S5,
114    TCG_REG_S6,
115    TCG_REG_S7,
116    TCG_REG_S8,
117    TCG_REG_S9,
118
119    /* Registers (potentially) clobbered across calls */
120    TCG_REG_T0,
121    TCG_REG_T1,
122    TCG_REG_T2,
123    TCG_REG_T3,
124    TCG_REG_T4,
125    TCG_REG_T5,
126    TCG_REG_T6,
127    TCG_REG_T7,
128    TCG_REG_T8,
129
130    /* Argument registers, opposite order of allocation.  */
131    TCG_REG_A7,
132    TCG_REG_A6,
133    TCG_REG_A5,
134    TCG_REG_A4,
135    TCG_REG_A3,
136    TCG_REG_A2,
137    TCG_REG_A1,
138    TCG_REG_A0,
139
140    /* Vector registers */
141    TCG_REG_V0, TCG_REG_V1, TCG_REG_V2, TCG_REG_V3,
142    TCG_REG_V4, TCG_REG_V5, TCG_REG_V6, TCG_REG_V7,
143    TCG_REG_V8, TCG_REG_V9, TCG_REG_V10, TCG_REG_V11,
144    TCG_REG_V12, TCG_REG_V13, TCG_REG_V14, TCG_REG_V15,
145    TCG_REG_V16, TCG_REG_V17, TCG_REG_V18, TCG_REG_V19,
146    TCG_REG_V20, TCG_REG_V21, TCG_REG_V22, TCG_REG_V23,
147    /* V24 - V31 are caller-saved, and skipped.  */
148};
149
150static const int tcg_target_call_iarg_regs[] = {
151    TCG_REG_A0,
152    TCG_REG_A1,
153    TCG_REG_A2,
154    TCG_REG_A3,
155    TCG_REG_A4,
156    TCG_REG_A5,
157    TCG_REG_A6,
158    TCG_REG_A7,
159};
160
161static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
162{
163    tcg_debug_assert(kind == TCG_CALL_RET_NORMAL);
164    tcg_debug_assert(slot >= 0 && slot <= 1);
165    return TCG_REG_A0 + slot;
166}
167
168#ifndef CONFIG_SOFTMMU
169#define USE_GUEST_BASE     (guest_base != 0)
170#define TCG_GUEST_BASE_REG TCG_REG_S1
171#endif
172
173#define TCG_CT_CONST_ZERO  0x100
174#define TCG_CT_CONST_S12   0x200
175#define TCG_CT_CONST_S32   0x400
176#define TCG_CT_CONST_U12   0x800
177#define TCG_CT_CONST_C12   0x1000
178#define TCG_CT_CONST_WSZ   0x2000
179#define TCG_CT_CONST_VCMP  0x4000
180#define TCG_CT_CONST_VADD  0x8000
181
182#define ALL_GENERAL_REGS   MAKE_64BIT_MASK(0, 32)
183#define ALL_VECTOR_REGS    MAKE_64BIT_MASK(32, 32)
184
185static inline tcg_target_long sextreg(tcg_target_long val, int pos, int len)
186{
187    return sextract64(val, pos, len);
188}
189
190/* test if a constant matches the constraint */
191static bool tcg_target_const_match(int64_t val, TCGType type, int ct, int vece)
192{
193    if (ct & TCG_CT_CONST) {
194        return true;
195    }
196    if ((ct & TCG_CT_CONST_ZERO) && val == 0) {
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    int64_t vec_val = sextract64(val, 0, 8 << vece);
215    if ((ct & TCG_CT_CONST_VCMP) && -0x10 <= vec_val && vec_val <= 0x1f) {
216        return true;
217    }
218    if ((ct & TCG_CT_CONST_VADD) && -0x1f <= vec_val && vec_val <= 0x1f) {
219        return true;
220    }
221    return false;
222}
223
224/*
225 * Relocations
226 */
227
228/*
229 * Relocation records defined in LoongArch ELF psABI v1.00 is way too
230 * complicated; a whopping stack machine is needed to stuff the fields, at
231 * the very least one SOP_PUSH and one SOP_POP (of the correct format) are
232 * needed.
233 *
234 * Hence, define our own simpler relocation types. Numbers are chosen as to
235 * not collide with potential future additions to the true ELF relocation
236 * type enum.
237 */
238
239/* Field Sk16, shifted right by 2; suitable for conditional jumps */
240#define R_LOONGARCH_BR_SK16     256
241/* Field Sd10k16, shifted right by 2; suitable for B and BL */
242#define R_LOONGARCH_BR_SD10K16  257
243
244static bool reloc_br_sk16(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
245{
246    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
247    intptr_t offset = (intptr_t)target - (intptr_t)src_rx;
248
249    tcg_debug_assert((offset & 3) == 0);
250    offset >>= 2;
251    if (offset == sextreg(offset, 0, 16)) {
252        *src_rw = deposit64(*src_rw, 10, 16, offset);
253        return true;
254    }
255
256    return false;
257}
258
259static bool reloc_br_sd10k16(tcg_insn_unit *src_rw,
260                             const tcg_insn_unit *target)
261{
262    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
263    intptr_t offset = (intptr_t)target - (intptr_t)src_rx;
264
265    tcg_debug_assert((offset & 3) == 0);
266    offset >>= 2;
267    if (offset == sextreg(offset, 0, 26)) {
268        *src_rw = deposit64(*src_rw, 0, 10, offset >> 16); /* slot d10 */
269        *src_rw = deposit64(*src_rw, 10, 16, offset); /* slot k16 */
270        return true;
271    }
272
273    return false;
274}
275
276static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
277                        intptr_t value, intptr_t addend)
278{
279    tcg_debug_assert(addend == 0);
280    switch (type) {
281    case R_LOONGARCH_BR_SK16:
282        return reloc_br_sk16(code_ptr, (tcg_insn_unit *)value);
283    case R_LOONGARCH_BR_SD10K16:
284        return reloc_br_sd10k16(code_ptr, (tcg_insn_unit *)value);
285    default:
286        g_assert_not_reached();
287    }
288}
289
290#include "tcg-insn-defs.c.inc"
291
292/*
293 * TCG intrinsics
294 */
295
296static void tcg_out_mb(TCGContext *s, TCGArg a0)
297{
298    /* Baseline LoongArch only has the full barrier, unfortunately.  */
299    tcg_out_opc_dbar(s, 0);
300}
301
302static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
303{
304    if (ret == arg) {
305        return true;
306    }
307    switch (type) {
308    case TCG_TYPE_I32:
309    case TCG_TYPE_I64:
310        /*
311         * Conventional register-register move used in LoongArch is
312         * `or dst, src, zero`.
313         */
314        tcg_out_opc_or(s, ret, arg, TCG_REG_ZERO);
315        break;
316    default:
317        g_assert_not_reached();
318    }
319    return true;
320}
321
322/* Loads a 32-bit immediate into rd, sign-extended.  */
323static void tcg_out_movi_i32(TCGContext *s, TCGReg rd, int32_t val)
324{
325    tcg_target_long lo = sextreg(val, 0, 12);
326    tcg_target_long hi12 = sextreg(val, 12, 20);
327
328    /* Single-instruction cases.  */
329    if (hi12 == 0) {
330        /* val fits in uimm12: ori rd, zero, val */
331        tcg_out_opc_ori(s, rd, TCG_REG_ZERO, val);
332        return;
333    }
334    if (hi12 == sextreg(lo, 12, 20)) {
335        /* val fits in simm12: addi.w rd, zero, val */
336        tcg_out_opc_addi_w(s, rd, TCG_REG_ZERO, val);
337        return;
338    }
339
340    /* High bits must be set; load with lu12i.w + optional ori.  */
341    tcg_out_opc_lu12i_w(s, rd, hi12);
342    if (lo != 0) {
343        tcg_out_opc_ori(s, rd, rd, lo & 0xfff);
344    }
345}
346
347static void tcg_out_movi(TCGContext *s, TCGType type, TCGReg rd,
348                         tcg_target_long val)
349{
350    /*
351     * LoongArch conventionally loads 64-bit immediates in at most 4 steps,
352     * with dedicated instructions for filling the respective bitfields
353     * below:
354     *
355     *        6                   5                   4               3
356     *  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
357     * +-----------------------+---------------------------------------+...
358     * |          hi52         |                  hi32                 |
359     * +-----------------------+---------------------------------------+...
360     *       3                   2                   1
361     *     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
362     * ...+-------------------------------------+-------------------------+
363     *    |                 hi12                |            lo           |
364     * ...+-------------------------------------+-------------------------+
365     *
366     * Check if val belong to one of the several fast cases, before falling
367     * back to the slow path.
368     */
369
370    intptr_t pc_offset;
371    tcg_target_long val_lo, val_hi, pc_hi, offset_hi;
372    tcg_target_long hi12, hi32, hi52;
373
374    /* Value fits in signed i32.  */
375    if (type == TCG_TYPE_I32 || val == (int32_t)val) {
376        tcg_out_movi_i32(s, rd, val);
377        return;
378    }
379
380    /* PC-relative cases.  */
381    pc_offset = tcg_pcrel_diff(s, (void *)val);
382    if (pc_offset == sextreg(pc_offset, 0, 22) && (pc_offset & 3) == 0) {
383        /* Single pcaddu2i.  */
384        tcg_out_opc_pcaddu2i(s, rd, pc_offset >> 2);
385        return;
386    }
387
388    if (pc_offset == (int32_t)pc_offset) {
389        /* Offset within 32 bits; load with pcalau12i + ori.  */
390        val_lo = sextreg(val, 0, 12);
391        val_hi = val >> 12;
392        pc_hi = (val - pc_offset) >> 12;
393        offset_hi = val_hi - pc_hi;
394
395        tcg_debug_assert(offset_hi == sextreg(offset_hi, 0, 20));
396        tcg_out_opc_pcalau12i(s, rd, offset_hi);
397        if (val_lo != 0) {
398            tcg_out_opc_ori(s, rd, rd, val_lo & 0xfff);
399        }
400        return;
401    }
402
403    hi12 = sextreg(val, 12, 20);
404    hi32 = sextreg(val, 32, 20);
405    hi52 = sextreg(val, 52, 12);
406
407    /* Single cu52i.d case.  */
408    if ((hi52 != 0) && (ctz64(val) >= 52)) {
409        tcg_out_opc_cu52i_d(s, rd, TCG_REG_ZERO, hi52);
410        return;
411    }
412
413    /* Slow path.  Initialize the low 32 bits, then concat high bits.  */
414    tcg_out_movi_i32(s, rd, val);
415
416    /* Load hi32 and hi52 explicitly when they are unexpected values. */
417    if (hi32 != sextreg(hi12, 20, 20)) {
418        tcg_out_opc_cu32i_d(s, rd, hi32);
419    }
420
421    if (hi52 != sextreg(hi32, 20, 12)) {
422        tcg_out_opc_cu52i_d(s, rd, rd, hi52);
423    }
424}
425
426static void tcg_out_addi(TCGContext *s, TCGType type, TCGReg rd,
427                         TCGReg rs, tcg_target_long imm)
428{
429    tcg_target_long lo12 = sextreg(imm, 0, 12);
430    tcg_target_long hi16 = sextreg(imm - lo12, 16, 16);
431
432    /*
433     * Note that there's a hole in between hi16 and lo12:
434     *
435     *       3                   2                   1                   0
436     *     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
437     * ...+-------------------------------+-------+-----------------------+
438     *    |             hi16              |       |          lo12         |
439     * ...+-------------------------------+-------+-----------------------+
440     *
441     * For bits within that hole, it's more efficient to use LU12I and ADD.
442     */
443    if (imm == (hi16 << 16) + lo12) {
444        if (hi16) {
445            tcg_out_opc_addu16i_d(s, rd, rs, hi16);
446            rs = rd;
447        }
448        if (type == TCG_TYPE_I32) {
449            tcg_out_opc_addi_w(s, rd, rs, lo12);
450        } else if (lo12) {
451            tcg_out_opc_addi_d(s, rd, rs, lo12);
452        } else {
453            tcg_out_mov(s, type, rd, rs);
454        }
455    } else {
456        tcg_out_movi(s, type, TCG_REG_TMP0, imm);
457        if (type == TCG_TYPE_I32) {
458            tcg_out_opc_add_w(s, rd, rs, TCG_REG_TMP0);
459        } else {
460            tcg_out_opc_add_d(s, rd, rs, TCG_REG_TMP0);
461        }
462    }
463}
464
465static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
466{
467    return false;
468}
469
470static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
471                             tcg_target_long imm)
472{
473    /* This function is only used for passing structs by reference. */
474    g_assert_not_reached();
475}
476
477static void tcg_out_ext8u(TCGContext *s, TCGReg ret, TCGReg arg)
478{
479    tcg_out_opc_andi(s, ret, arg, 0xff);
480}
481
482static void tcg_out_ext16u(TCGContext *s, TCGReg ret, TCGReg arg)
483{
484    tcg_out_opc_bstrpick_w(s, ret, arg, 0, 15);
485}
486
487static void tcg_out_ext32u(TCGContext *s, TCGReg ret, TCGReg arg)
488{
489    tcg_out_opc_bstrpick_d(s, ret, arg, 0, 31);
490}
491
492static void tcg_out_ext8s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
493{
494    tcg_out_opc_sext_b(s, ret, arg);
495}
496
497static void tcg_out_ext16s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
498{
499    tcg_out_opc_sext_h(s, ret, arg);
500}
501
502static void tcg_out_ext32s(TCGContext *s, TCGReg ret, TCGReg arg)
503{
504    tcg_out_opc_addi_w(s, ret, arg, 0);
505}
506
507static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg)
508{
509    if (ret != arg) {
510        tcg_out_ext32s(s, ret, arg);
511    }
512}
513
514static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg)
515{
516    tcg_out_ext32u(s, ret, arg);
517}
518
519static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg ret, TCGReg arg)
520{
521    tcg_out_ext32s(s, ret, arg);
522}
523
524static void tcg_out_clzctz(TCGContext *s, LoongArchInsn opc,
525                           TCGReg a0, TCGReg a1, TCGReg a2,
526                           bool c2, bool is_32bit)
527{
528    if (c2) {
529        /*
530         * Fast path: semantics already satisfied due to constraint and
531         * insn behavior, single instruction is enough.
532         */
533        tcg_debug_assert(a2 == (is_32bit ? 32 : 64));
534        /* all clz/ctz insns belong to DJ-format */
535        tcg_out32(s, encode_dj_insn(opc, a0, a1));
536        return;
537    }
538
539    tcg_out32(s, encode_dj_insn(opc, TCG_REG_TMP0, a1));
540    /* a0 = a1 ? REG_TMP0 : a2 */
541    tcg_out_opc_maskeqz(s, TCG_REG_TMP0, TCG_REG_TMP0, a1);
542    tcg_out_opc_masknez(s, a0, a2, a1);
543    tcg_out_opc_or(s, a0, TCG_REG_TMP0, a0);
544}
545
546#define SETCOND_INV    TCG_TARGET_NB_REGS
547#define SETCOND_NEZ    (SETCOND_INV << 1)
548#define SETCOND_FLAGS  (SETCOND_INV | SETCOND_NEZ)
549
550static int tcg_out_setcond_int(TCGContext *s, TCGCond cond, TCGReg ret,
551                               TCGReg arg1, tcg_target_long arg2, bool c2)
552{
553    int flags = 0;
554
555    switch (cond) {
556    case TCG_COND_EQ:    /* -> NE  */
557    case TCG_COND_GE:    /* -> LT  */
558    case TCG_COND_GEU:   /* -> LTU */
559    case TCG_COND_GT:    /* -> LE  */
560    case TCG_COND_GTU:   /* -> LEU */
561        cond = tcg_invert_cond(cond);
562        flags ^= SETCOND_INV;
563        break;
564    default:
565        break;
566    }
567
568    switch (cond) {
569    case TCG_COND_LE:
570    case TCG_COND_LEU:
571        /*
572         * If we have a constant input, the most efficient way to implement
573         * LE is by adding 1 and using LT.  Watch out for wrap around for LEU.
574         * We don't need to care for this for LE because the constant input
575         * is still constrained to int32_t, and INT32_MAX+1 is representable
576         * in the 64-bit temporary register.
577         */
578        if (c2) {
579            if (cond == TCG_COND_LEU) {
580                /* unsigned <= -1 is true */
581                if (arg2 == -1) {
582                    tcg_out_movi(s, TCG_TYPE_REG, ret, !(flags & SETCOND_INV));
583                    return ret;
584                }
585                cond = TCG_COND_LTU;
586            } else {
587                cond = TCG_COND_LT;
588            }
589            arg2 += 1;
590        } else {
591            TCGReg tmp = arg2;
592            arg2 = arg1;
593            arg1 = tmp;
594            cond = tcg_swap_cond(cond);    /* LE -> GE */
595            cond = tcg_invert_cond(cond);  /* GE -> LT */
596            flags ^= SETCOND_INV;
597        }
598        break;
599    default:
600        break;
601    }
602
603    switch (cond) {
604    case TCG_COND_NE:
605        flags |= SETCOND_NEZ;
606        if (!c2) {
607            tcg_out_opc_xor(s, ret, arg1, arg2);
608        } else if (arg2 == 0) {
609            ret = arg1;
610        } else if (arg2 >= 0 && arg2 <= 0xfff) {
611            tcg_out_opc_xori(s, ret, arg1, arg2);
612        } else {
613            tcg_out_addi(s, TCG_TYPE_REG, ret, arg1, -arg2);
614        }
615        break;
616
617    case TCG_COND_LT:
618    case TCG_COND_LTU:
619        if (c2) {
620            if (arg2 >= -0x800 && arg2 <= 0x7ff) {
621                if (cond == TCG_COND_LT) {
622                    tcg_out_opc_slti(s, ret, arg1, arg2);
623                } else {
624                    tcg_out_opc_sltui(s, ret, arg1, arg2);
625                }
626                break;
627            }
628            tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_TMP0, arg2);
629            arg2 = TCG_REG_TMP0;
630        }
631        if (cond == TCG_COND_LT) {
632            tcg_out_opc_slt(s, ret, arg1, arg2);
633        } else {
634            tcg_out_opc_sltu(s, ret, arg1, arg2);
635        }
636        break;
637
638    default:
639        g_assert_not_reached();
640        break;
641    }
642
643    return ret | flags;
644}
645
646static void tcg_out_setcond(TCGContext *s, TCGCond cond, TCGReg ret,
647                            TCGReg arg1, tcg_target_long arg2, bool c2)
648{
649    int tmpflags = tcg_out_setcond_int(s, cond, ret, arg1, arg2, c2);
650
651    if (tmpflags != ret) {
652        TCGReg tmp = tmpflags & ~SETCOND_FLAGS;
653
654        switch (tmpflags & SETCOND_FLAGS) {
655        case SETCOND_INV:
656            /* Intermediate result is boolean: simply invert. */
657            tcg_out_opc_xori(s, ret, tmp, 1);
658            break;
659        case SETCOND_NEZ:
660            /* Intermediate result is zero/non-zero: test != 0. */
661            tcg_out_opc_sltu(s, ret, TCG_REG_ZERO, tmp);
662            break;
663        case SETCOND_NEZ | SETCOND_INV:
664            /* Intermediate result is zero/non-zero: test == 0. */
665            tcg_out_opc_sltui(s, ret, tmp, 1);
666            break;
667        default:
668            g_assert_not_reached();
669        }
670    }
671}
672
673static void tcg_out_movcond(TCGContext *s, TCGCond cond, TCGReg ret,
674                            TCGReg c1, tcg_target_long c2, bool const2,
675                            TCGReg v1, TCGReg v2)
676{
677    int tmpflags = tcg_out_setcond_int(s, cond, TCG_REG_TMP0, c1, c2, const2);
678    TCGReg t;
679
680    /* Standardize the test below to t != 0. */
681    if (tmpflags & SETCOND_INV) {
682        t = v1, v1 = v2, v2 = t;
683    }
684
685    t = tmpflags & ~SETCOND_FLAGS;
686    if (v1 == TCG_REG_ZERO) {
687        tcg_out_opc_masknez(s, ret, v2, t);
688    } else if (v2 == TCG_REG_ZERO) {
689        tcg_out_opc_maskeqz(s, ret, v1, t);
690    } else {
691        tcg_out_opc_masknez(s, TCG_REG_TMP2, v2, t); /* t ? 0 : v2 */
692        tcg_out_opc_maskeqz(s, TCG_REG_TMP1, v1, t); /* t ? v1 : 0 */
693        tcg_out_opc_or(s, ret, TCG_REG_TMP1, TCG_REG_TMP2);
694    }
695}
696
697/*
698 * Branch helpers
699 */
700
701static const struct {
702    LoongArchInsn op;
703    bool swap;
704} tcg_brcond_to_loongarch[] = {
705    [TCG_COND_EQ] =  { OPC_BEQ,  false },
706    [TCG_COND_NE] =  { OPC_BNE,  false },
707    [TCG_COND_LT] =  { OPC_BGT,  true  },
708    [TCG_COND_GE] =  { OPC_BLE,  true  },
709    [TCG_COND_LE] =  { OPC_BLE,  false },
710    [TCG_COND_GT] =  { OPC_BGT,  false },
711    [TCG_COND_LTU] = { OPC_BGTU, true  },
712    [TCG_COND_GEU] = { OPC_BLEU, true  },
713    [TCG_COND_LEU] = { OPC_BLEU, false },
714    [TCG_COND_GTU] = { OPC_BGTU, false }
715};
716
717static void tcg_out_brcond(TCGContext *s, TCGCond cond, TCGReg arg1,
718                           TCGReg arg2, TCGLabel *l)
719{
720    LoongArchInsn op = tcg_brcond_to_loongarch[cond].op;
721
722    tcg_debug_assert(op != 0);
723
724    if (tcg_brcond_to_loongarch[cond].swap) {
725        TCGReg t = arg1;
726        arg1 = arg2;
727        arg2 = t;
728    }
729
730    /* all conditional branch insns belong to DJSk16-format */
731    tcg_out_reloc(s, s->code_ptr, R_LOONGARCH_BR_SK16, l, 0);
732    tcg_out32(s, encode_djsk16_insn(op, arg1, arg2, 0));
733}
734
735static void tcg_out_call_int(TCGContext *s, const tcg_insn_unit *arg, bool tail)
736{
737    TCGReg link = tail ? TCG_REG_ZERO : TCG_REG_RA;
738    ptrdiff_t offset = tcg_pcrel_diff(s, arg);
739
740    tcg_debug_assert((offset & 3) == 0);
741    if (offset == sextreg(offset, 0, 28)) {
742        /* short jump: +/- 256MiB */
743        if (tail) {
744            tcg_out_opc_b(s, offset >> 2);
745        } else {
746            tcg_out_opc_bl(s, offset >> 2);
747        }
748    } else if (offset == sextreg(offset, 0, 38)) {
749        /* long jump: +/- 256GiB */
750        tcg_target_long lo = sextreg(offset, 0, 18);
751        tcg_target_long hi = offset - lo;
752        tcg_out_opc_pcaddu18i(s, TCG_REG_TMP0, hi >> 18);
753        tcg_out_opc_jirl(s, link, TCG_REG_TMP0, lo >> 2);
754    } else {
755        /* far jump: 64-bit */
756        tcg_target_long lo = sextreg((tcg_target_long)arg, 0, 18);
757        tcg_target_long hi = (tcg_target_long)arg - lo;
758        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, hi);
759        tcg_out_opc_jirl(s, link, TCG_REG_TMP0, lo >> 2);
760    }
761}
762
763static void tcg_out_call(TCGContext *s, const tcg_insn_unit *arg,
764                         const TCGHelperInfo *info)
765{
766    tcg_out_call_int(s, arg, false);
767}
768
769/*
770 * Load/store helpers
771 */
772
773static void tcg_out_ldst(TCGContext *s, LoongArchInsn opc, TCGReg data,
774                         TCGReg addr, intptr_t offset)
775{
776    intptr_t imm12 = sextreg(offset, 0, 12);
777
778    if (offset != imm12) {
779        intptr_t diff = tcg_pcrel_diff(s, (void *)offset);
780
781        if (addr == TCG_REG_ZERO && diff == (int32_t)diff) {
782            imm12 = sextreg(diff, 0, 12);
783            tcg_out_opc_pcaddu12i(s, TCG_REG_TMP2, (diff - imm12) >> 12);
784        } else {
785            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP2, offset - imm12);
786            if (addr != TCG_REG_ZERO) {
787                tcg_out_opc_add_d(s, TCG_REG_TMP2, TCG_REG_TMP2, addr);
788            }
789        }
790        addr = TCG_REG_TMP2;
791    }
792
793    switch (opc) {
794    case OPC_LD_B:
795    case OPC_LD_BU:
796    case OPC_LD_H:
797    case OPC_LD_HU:
798    case OPC_LD_W:
799    case OPC_LD_WU:
800    case OPC_LD_D:
801    case OPC_ST_B:
802    case OPC_ST_H:
803    case OPC_ST_W:
804    case OPC_ST_D:
805        tcg_out32(s, encode_djsk12_insn(opc, data, addr, imm12));
806        break;
807    default:
808        g_assert_not_reached();
809    }
810}
811
812static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg arg,
813                       TCGReg arg1, intptr_t arg2)
814{
815    bool is_32bit = type == TCG_TYPE_I32;
816    tcg_out_ldst(s, is_32bit ? OPC_LD_W : OPC_LD_D, arg, arg1, arg2);
817}
818
819static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
820                       TCGReg arg1, intptr_t arg2)
821{
822    bool is_32bit = type == TCG_TYPE_I32;
823    tcg_out_ldst(s, is_32bit ? OPC_ST_W : OPC_ST_D, arg, arg1, arg2);
824}
825
826static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
827                        TCGReg base, intptr_t ofs)
828{
829    if (val == 0) {
830        tcg_out_st(s, type, TCG_REG_ZERO, base, ofs);
831        return true;
832    }
833    return false;
834}
835
836/*
837 * Load/store helpers for SoftMMU, and qemu_ld/st implementations
838 */
839
840static bool tcg_out_goto(TCGContext *s, const tcg_insn_unit *target)
841{
842    tcg_out_opc_b(s, 0);
843    return reloc_br_sd10k16(s->code_ptr - 1, target);
844}
845
846static const TCGLdstHelperParam ldst_helper_param = {
847    .ntmp = 1, .tmp = { TCG_REG_TMP0 }
848};
849
850static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
851{
852    MemOp opc = get_memop(l->oi);
853
854    /* resolve label address */
855    if (!reloc_br_sk16(l->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
856        return false;
857    }
858
859    tcg_out_ld_helper_args(s, l, &ldst_helper_param);
860    tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SIZE], false);
861    tcg_out_ld_helper_ret(s, l, false, &ldst_helper_param);
862    return tcg_out_goto(s, l->raddr);
863}
864
865static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
866{
867    MemOp opc = get_memop(l->oi);
868
869    /* resolve label address */
870    if (!reloc_br_sk16(l->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
871        return false;
872    }
873
874    tcg_out_st_helper_args(s, l, &ldst_helper_param);
875    tcg_out_call_int(s, qemu_st_helpers[opc & MO_SIZE], false);
876    return tcg_out_goto(s, l->raddr);
877}
878
879typedef struct {
880    TCGReg base;
881    TCGReg index;
882    TCGAtomAlign aa;
883} HostAddress;
884
885bool tcg_target_has_memory_bswap(MemOp memop)
886{
887    return false;
888}
889
890/* We expect to use a 12-bit negative offset from ENV.  */
891#define MIN_TLB_MASK_TABLE_OFS  -(1 << 11)
892
893/*
894 * For softmmu, perform the TLB load and compare.
895 * For useronly, perform any required alignment tests.
896 * In both cases, return a TCGLabelQemuLdst structure if the slow path
897 * is required and fill in @h with the host address for the fast path.
898 */
899static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
900                                           TCGReg addr_reg, MemOpIdx oi,
901                                           bool is_ld)
902{
903    TCGType addr_type = s->addr_type;
904    TCGLabelQemuLdst *ldst = NULL;
905    MemOp opc = get_memop(oi);
906    MemOp a_bits;
907
908    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, false);
909    a_bits = h->aa.align;
910
911#ifdef CONFIG_SOFTMMU
912    unsigned s_bits = opc & MO_SIZE;
913    int mem_index = get_mmuidx(oi);
914    int fast_ofs = tlb_mask_table_ofs(s, mem_index);
915    int mask_ofs = fast_ofs + offsetof(CPUTLBDescFast, mask);
916    int table_ofs = fast_ofs + offsetof(CPUTLBDescFast, table);
917
918    ldst = new_ldst_label(s);
919    ldst->is_ld = is_ld;
920    ldst->oi = oi;
921    ldst->addrlo_reg = addr_reg;
922
923    tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP0, TCG_AREG0, mask_ofs);
924    tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP1, TCG_AREG0, table_ofs);
925
926    tcg_out_opc_srli_d(s, TCG_REG_TMP2, addr_reg,
927                    s->page_bits - CPU_TLB_ENTRY_BITS);
928    tcg_out_opc_and(s, TCG_REG_TMP2, TCG_REG_TMP2, TCG_REG_TMP0);
929    tcg_out_opc_add_d(s, TCG_REG_TMP2, TCG_REG_TMP2, TCG_REG_TMP1);
930
931    /* Load the tlb comparator and the addend.  */
932    QEMU_BUILD_BUG_ON(HOST_BIG_ENDIAN);
933    tcg_out_ld(s, addr_type, TCG_REG_TMP0, TCG_REG_TMP2,
934               is_ld ? offsetof(CPUTLBEntry, addr_read)
935                     : offsetof(CPUTLBEntry, addr_write));
936    tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP2, TCG_REG_TMP2,
937               offsetof(CPUTLBEntry, addend));
938
939    /*
940     * For aligned accesses, we check the first byte and include the alignment
941     * bits within the address.  For unaligned access, we check that we don't
942     * cross pages using the address of the last byte of the access.
943     */
944    if (a_bits < s_bits) {
945        unsigned a_mask = (1u << a_bits) - 1;
946        unsigned s_mask = (1u << s_bits) - 1;
947        tcg_out_addi(s, addr_type, TCG_REG_TMP1, addr_reg, s_mask - a_mask);
948    } else {
949        tcg_out_mov(s, addr_type, TCG_REG_TMP1, addr_reg);
950    }
951    tcg_out_opc_bstrins_d(s, TCG_REG_TMP1, TCG_REG_ZERO,
952                          a_bits, s->page_bits - 1);
953
954    /* Compare masked address with the TLB entry.  */
955    ldst->label_ptr[0] = s->code_ptr;
956    tcg_out_opc_bne(s, TCG_REG_TMP0, TCG_REG_TMP1, 0);
957
958    h->index = TCG_REG_TMP2;
959#else
960    if (a_bits) {
961        ldst = new_ldst_label(s);
962
963        ldst->is_ld = is_ld;
964        ldst->oi = oi;
965        ldst->addrlo_reg = addr_reg;
966
967        /*
968         * Without micro-architecture details, we don't know which of
969         * bstrpick or andi is faster, so use bstrpick as it's not
970         * constrained by imm field width. Not to say alignments >= 2^12
971         * are going to happen any time soon.
972         */
973        tcg_out_opc_bstrpick_d(s, TCG_REG_TMP1, addr_reg, 0, a_bits - 1);
974
975        ldst->label_ptr[0] = s->code_ptr;
976        tcg_out_opc_bne(s, TCG_REG_TMP1, TCG_REG_ZERO, 0);
977    }
978
979    h->index = USE_GUEST_BASE ? TCG_GUEST_BASE_REG : TCG_REG_ZERO;
980#endif
981
982    if (addr_type == TCG_TYPE_I32) {
983        h->base = TCG_REG_TMP0;
984        tcg_out_ext32u(s, h->base, addr_reg);
985    } else {
986        h->base = addr_reg;
987    }
988
989    return ldst;
990}
991
992static void tcg_out_qemu_ld_indexed(TCGContext *s, MemOp opc, TCGType type,
993                                    TCGReg rd, HostAddress h)
994{
995    /* Byte swapping is left to middle-end expansion.  */
996    tcg_debug_assert((opc & MO_BSWAP) == 0);
997
998    switch (opc & MO_SSIZE) {
999    case MO_UB:
1000        tcg_out_opc_ldx_bu(s, rd, h.base, h.index);
1001        break;
1002    case MO_SB:
1003        tcg_out_opc_ldx_b(s, rd, h.base, h.index);
1004        break;
1005    case MO_UW:
1006        tcg_out_opc_ldx_hu(s, rd, h.base, h.index);
1007        break;
1008    case MO_SW:
1009        tcg_out_opc_ldx_h(s, rd, h.base, h.index);
1010        break;
1011    case MO_UL:
1012        if (type == TCG_TYPE_I64) {
1013            tcg_out_opc_ldx_wu(s, rd, h.base, h.index);
1014            break;
1015        }
1016        /* fallthrough */
1017    case MO_SL:
1018        tcg_out_opc_ldx_w(s, rd, h.base, h.index);
1019        break;
1020    case MO_UQ:
1021        tcg_out_opc_ldx_d(s, rd, h.base, h.index);
1022        break;
1023    default:
1024        g_assert_not_reached();
1025    }
1026}
1027
1028static void tcg_out_qemu_ld(TCGContext *s, TCGReg data_reg, TCGReg addr_reg,
1029                            MemOpIdx oi, TCGType data_type)
1030{
1031    TCGLabelQemuLdst *ldst;
1032    HostAddress h;
1033
1034    ldst = prepare_host_addr(s, &h, addr_reg, oi, true);
1035    tcg_out_qemu_ld_indexed(s, get_memop(oi), data_type, data_reg, h);
1036
1037    if (ldst) {
1038        ldst->type = data_type;
1039        ldst->datalo_reg = data_reg;
1040        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1041    }
1042}
1043
1044static void tcg_out_qemu_st_indexed(TCGContext *s, MemOp opc,
1045                                    TCGReg rd, HostAddress h)
1046{
1047    /* Byte swapping is left to middle-end expansion.  */
1048    tcg_debug_assert((opc & MO_BSWAP) == 0);
1049
1050    switch (opc & MO_SIZE) {
1051    case MO_8:
1052        tcg_out_opc_stx_b(s, rd, h.base, h.index);
1053        break;
1054    case MO_16:
1055        tcg_out_opc_stx_h(s, rd, h.base, h.index);
1056        break;
1057    case MO_32:
1058        tcg_out_opc_stx_w(s, rd, h.base, h.index);
1059        break;
1060    case MO_64:
1061        tcg_out_opc_stx_d(s, rd, h.base, h.index);
1062        break;
1063    default:
1064        g_assert_not_reached();
1065    }
1066}
1067
1068static void tcg_out_qemu_st(TCGContext *s, TCGReg data_reg, TCGReg addr_reg,
1069                            MemOpIdx oi, TCGType data_type)
1070{
1071    TCGLabelQemuLdst *ldst;
1072    HostAddress h;
1073
1074    ldst = prepare_host_addr(s, &h, addr_reg, oi, false);
1075    tcg_out_qemu_st_indexed(s, get_memop(oi), data_reg, h);
1076
1077    if (ldst) {
1078        ldst->type = data_type;
1079        ldst->datalo_reg = data_reg;
1080        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1081    }
1082}
1083
1084static void tcg_out_qemu_ldst_i128(TCGContext *s, TCGReg data_lo, TCGReg data_hi,
1085                                   TCGReg addr_reg, MemOpIdx oi, bool is_ld)
1086{
1087    TCGLabelQemuLdst *ldst;
1088    HostAddress h;
1089
1090    ldst = prepare_host_addr(s, &h, addr_reg, oi, is_ld);
1091
1092    if (h.aa.atom == MO_128) {
1093        /*
1094         * Use VLDX/VSTX when 128-bit atomicity is required.
1095         * If address is aligned to 16-bytes, the 128-bit load/store is atomic.
1096         */
1097        if (is_ld) {
1098            tcg_out_opc_vldx(s, TCG_VEC_TMP0, h.base, h.index);
1099            tcg_out_opc_vpickve2gr_d(s, data_lo, TCG_VEC_TMP0, 0);
1100            tcg_out_opc_vpickve2gr_d(s, data_hi, TCG_VEC_TMP0, 1);
1101        } else {
1102            tcg_out_opc_vinsgr2vr_d(s, TCG_VEC_TMP0, data_lo, 0);
1103            tcg_out_opc_vinsgr2vr_d(s, TCG_VEC_TMP0, data_hi, 1);
1104            tcg_out_opc_vstx(s, TCG_VEC_TMP0, h.base, h.index);
1105        }
1106    } else {
1107        /* Otherwise use a pair of LD/ST. */
1108        tcg_out_opc_add_d(s, TCG_REG_TMP0, h.base, h.index);
1109        if (is_ld) {
1110            tcg_out_opc_ld_d(s, data_lo, TCG_REG_TMP0, 0);
1111            tcg_out_opc_ld_d(s, data_hi, TCG_REG_TMP0, 8);
1112        } else {
1113            tcg_out_opc_st_d(s, data_lo, TCG_REG_TMP0, 0);
1114            tcg_out_opc_st_d(s, data_hi, TCG_REG_TMP0, 8);
1115        }
1116    }
1117
1118    if (ldst) {
1119        ldst->type = TCG_TYPE_I128;
1120        ldst->datalo_reg = data_lo;
1121        ldst->datahi_reg = data_hi;
1122        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1123    }
1124}
1125
1126/*
1127 * Entry-points
1128 */
1129
1130static const tcg_insn_unit *tb_ret_addr;
1131
1132static void tcg_out_exit_tb(TCGContext *s, uintptr_t a0)
1133{
1134    /* Reuse the zeroing that exists for goto_ptr.  */
1135    if (a0 == 0) {
1136        tcg_out_call_int(s, tcg_code_gen_epilogue, true);
1137    } else {
1138        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_A0, a0);
1139        tcg_out_call_int(s, tb_ret_addr, true);
1140    }
1141}
1142
1143static void tcg_out_goto_tb(TCGContext *s, int which)
1144{
1145    /*
1146     * Direct branch, or load indirect address, to be patched
1147     * by tb_target_set_jmp_target.  Check indirect load offset
1148     * in range early, regardless of direct branch distance,
1149     * via assert within tcg_out_opc_pcaddu2i.
1150     */
1151    uintptr_t i_addr = get_jmp_target_addr(s, which);
1152    intptr_t i_disp = tcg_pcrel_diff(s, (void *)i_addr);
1153
1154    set_jmp_insn_offset(s, which);
1155    tcg_out_opc_pcaddu2i(s, TCG_REG_TMP0, i_disp >> 2);
1156
1157    /* Finish the load and indirect branch. */
1158    tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP0, TCG_REG_TMP0, 0);
1159    tcg_out_opc_jirl(s, TCG_REG_ZERO, TCG_REG_TMP0, 0);
1160    set_jmp_reset_offset(s, which);
1161}
1162
1163void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
1164                              uintptr_t jmp_rx, uintptr_t jmp_rw)
1165{
1166    uintptr_t d_addr = tb->jmp_target_addr[n];
1167    ptrdiff_t d_disp = (ptrdiff_t)(d_addr - jmp_rx) >> 2;
1168    tcg_insn_unit insn;
1169
1170    /* Either directly branch, or load slot address for indirect branch. */
1171    if (d_disp == sextreg(d_disp, 0, 26)) {
1172        insn = encode_sd10k16_insn(OPC_B, d_disp);
1173    } else {
1174        uintptr_t i_addr = (uintptr_t)&tb->jmp_target_addr[n];
1175        intptr_t i_disp = i_addr - jmp_rx;
1176        insn = encode_dsj20_insn(OPC_PCADDU2I, TCG_REG_TMP0, i_disp >> 2);
1177    }
1178
1179    qatomic_set((tcg_insn_unit *)jmp_rw, insn);
1180    flush_idcache_range(jmp_rx, jmp_rw, 4);
1181}
1182
1183static void tcg_out_op(TCGContext *s, TCGOpcode opc,
1184                       const TCGArg args[TCG_MAX_OP_ARGS],
1185                       const int const_args[TCG_MAX_OP_ARGS])
1186{
1187    TCGArg a0 = args[0];
1188    TCGArg a1 = args[1];
1189    TCGArg a2 = args[2];
1190    TCGArg a3 = args[3];
1191    int c2 = const_args[2];
1192
1193    switch (opc) {
1194    case INDEX_op_mb:
1195        tcg_out_mb(s, a0);
1196        break;
1197
1198    case INDEX_op_goto_ptr:
1199        tcg_out_opc_jirl(s, TCG_REG_ZERO, a0, 0);
1200        break;
1201
1202    case INDEX_op_br:
1203        tcg_out_reloc(s, s->code_ptr, R_LOONGARCH_BR_SD10K16, arg_label(a0),
1204                      0);
1205        tcg_out_opc_b(s, 0);
1206        break;
1207
1208    case INDEX_op_brcond_i32:
1209    case INDEX_op_brcond_i64:
1210        tcg_out_brcond(s, a2, a0, a1, arg_label(args[3]));
1211        break;
1212
1213    case INDEX_op_extrh_i64_i32:
1214        tcg_out_opc_srai_d(s, a0, a1, 32);
1215        break;
1216
1217    case INDEX_op_not_i32:
1218    case INDEX_op_not_i64:
1219        tcg_out_opc_nor(s, a0, a1, TCG_REG_ZERO);
1220        break;
1221
1222    case INDEX_op_nor_i32:
1223    case INDEX_op_nor_i64:
1224        if (c2) {
1225            tcg_out_opc_ori(s, a0, a1, a2);
1226            tcg_out_opc_nor(s, a0, a0, TCG_REG_ZERO);
1227        } else {
1228            tcg_out_opc_nor(s, a0, a1, a2);
1229        }
1230        break;
1231
1232    case INDEX_op_andc_i32:
1233    case INDEX_op_andc_i64:
1234        if (c2) {
1235            /* guaranteed to fit due to constraint */
1236            tcg_out_opc_andi(s, a0, a1, ~a2);
1237        } else {
1238            tcg_out_opc_andn(s, a0, a1, a2);
1239        }
1240        break;
1241
1242    case INDEX_op_orc_i32:
1243    case INDEX_op_orc_i64:
1244        if (c2) {
1245            /* guaranteed to fit due to constraint */
1246            tcg_out_opc_ori(s, a0, a1, ~a2);
1247        } else {
1248            tcg_out_opc_orn(s, a0, a1, a2);
1249        }
1250        break;
1251
1252    case INDEX_op_and_i32:
1253    case INDEX_op_and_i64:
1254        if (c2) {
1255            tcg_out_opc_andi(s, a0, a1, a2);
1256        } else {
1257            tcg_out_opc_and(s, a0, a1, a2);
1258        }
1259        break;
1260
1261    case INDEX_op_or_i32:
1262    case INDEX_op_or_i64:
1263        if (c2) {
1264            tcg_out_opc_ori(s, a0, a1, a2);
1265        } else {
1266            tcg_out_opc_or(s, a0, a1, a2);
1267        }
1268        break;
1269
1270    case INDEX_op_xor_i32:
1271    case INDEX_op_xor_i64:
1272        if (c2) {
1273            tcg_out_opc_xori(s, a0, a1, a2);
1274        } else {
1275            tcg_out_opc_xor(s, a0, a1, a2);
1276        }
1277        break;
1278
1279    case INDEX_op_extract_i32:
1280        tcg_out_opc_bstrpick_w(s, a0, a1, a2, a2 + args[3] - 1);
1281        break;
1282    case INDEX_op_extract_i64:
1283        tcg_out_opc_bstrpick_d(s, a0, a1, a2, a2 + args[3] - 1);
1284        break;
1285
1286    case INDEX_op_deposit_i32:
1287        tcg_out_opc_bstrins_w(s, a0, a2, args[3], args[3] + args[4] - 1);
1288        break;
1289    case INDEX_op_deposit_i64:
1290        tcg_out_opc_bstrins_d(s, a0, a2, args[3], args[3] + args[4] - 1);
1291        break;
1292
1293    case INDEX_op_bswap16_i32:
1294    case INDEX_op_bswap16_i64:
1295        tcg_out_opc_revb_2h(s, a0, a1);
1296        if (a2 & TCG_BSWAP_OS) {
1297            tcg_out_ext16s(s, TCG_TYPE_REG, a0, a0);
1298        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1299            tcg_out_ext16u(s, a0, a0);
1300        }
1301        break;
1302
1303    case INDEX_op_bswap32_i32:
1304        /* All 32-bit values are computed sign-extended in the register.  */
1305        a2 = TCG_BSWAP_OS;
1306        /* fallthrough */
1307    case INDEX_op_bswap32_i64:
1308        tcg_out_opc_revb_2w(s, a0, a1);
1309        if (a2 & TCG_BSWAP_OS) {
1310            tcg_out_ext32s(s, a0, a0);
1311        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1312            tcg_out_ext32u(s, a0, a0);
1313        }
1314        break;
1315
1316    case INDEX_op_bswap64_i64:
1317        tcg_out_opc_revb_d(s, a0, a1);
1318        break;
1319
1320    case INDEX_op_clz_i32:
1321        tcg_out_clzctz(s, OPC_CLZ_W, a0, a1, a2, c2, true);
1322        break;
1323    case INDEX_op_clz_i64:
1324        tcg_out_clzctz(s, OPC_CLZ_D, a0, a1, a2, c2, false);
1325        break;
1326
1327    case INDEX_op_ctz_i32:
1328        tcg_out_clzctz(s, OPC_CTZ_W, a0, a1, a2, c2, true);
1329        break;
1330    case INDEX_op_ctz_i64:
1331        tcg_out_clzctz(s, OPC_CTZ_D, a0, a1, a2, c2, false);
1332        break;
1333
1334    case INDEX_op_shl_i32:
1335        if (c2) {
1336            tcg_out_opc_slli_w(s, a0, a1, a2 & 0x1f);
1337        } else {
1338            tcg_out_opc_sll_w(s, a0, a1, a2);
1339        }
1340        break;
1341    case INDEX_op_shl_i64:
1342        if (c2) {
1343            tcg_out_opc_slli_d(s, a0, a1, a2 & 0x3f);
1344        } else {
1345            tcg_out_opc_sll_d(s, a0, a1, a2);
1346        }
1347        break;
1348
1349    case INDEX_op_shr_i32:
1350        if (c2) {
1351            tcg_out_opc_srli_w(s, a0, a1, a2 & 0x1f);
1352        } else {
1353            tcg_out_opc_srl_w(s, a0, a1, a2);
1354        }
1355        break;
1356    case INDEX_op_shr_i64:
1357        if (c2) {
1358            tcg_out_opc_srli_d(s, a0, a1, a2 & 0x3f);
1359        } else {
1360            tcg_out_opc_srl_d(s, a0, a1, a2);
1361        }
1362        break;
1363
1364    case INDEX_op_sar_i32:
1365        if (c2) {
1366            tcg_out_opc_srai_w(s, a0, a1, a2 & 0x1f);
1367        } else {
1368            tcg_out_opc_sra_w(s, a0, a1, a2);
1369        }
1370        break;
1371    case INDEX_op_sar_i64:
1372        if (c2) {
1373            tcg_out_opc_srai_d(s, a0, a1, a2 & 0x3f);
1374        } else {
1375            tcg_out_opc_sra_d(s, a0, a1, a2);
1376        }
1377        break;
1378
1379    case INDEX_op_rotl_i32:
1380        /* transform into equivalent rotr/rotri */
1381        if (c2) {
1382            tcg_out_opc_rotri_w(s, a0, a1, (32 - a2) & 0x1f);
1383        } else {
1384            tcg_out_opc_sub_w(s, TCG_REG_TMP0, TCG_REG_ZERO, a2);
1385            tcg_out_opc_rotr_w(s, a0, a1, TCG_REG_TMP0);
1386        }
1387        break;
1388    case INDEX_op_rotl_i64:
1389        /* transform into equivalent rotr/rotri */
1390        if (c2) {
1391            tcg_out_opc_rotri_d(s, a0, a1, (64 - a2) & 0x3f);
1392        } else {
1393            tcg_out_opc_sub_w(s, TCG_REG_TMP0, TCG_REG_ZERO, a2);
1394            tcg_out_opc_rotr_d(s, a0, a1, TCG_REG_TMP0);
1395        }
1396        break;
1397
1398    case INDEX_op_rotr_i32:
1399        if (c2) {
1400            tcg_out_opc_rotri_w(s, a0, a1, a2 & 0x1f);
1401        } else {
1402            tcg_out_opc_rotr_w(s, a0, a1, a2);
1403        }
1404        break;
1405    case INDEX_op_rotr_i64:
1406        if (c2) {
1407            tcg_out_opc_rotri_d(s, a0, a1, a2 & 0x3f);
1408        } else {
1409            tcg_out_opc_rotr_d(s, a0, a1, a2);
1410        }
1411        break;
1412
1413    case INDEX_op_add_i32:
1414        if (c2) {
1415            tcg_out_addi(s, TCG_TYPE_I32, a0, a1, a2);
1416        } else {
1417            tcg_out_opc_add_w(s, a0, a1, a2);
1418        }
1419        break;
1420    case INDEX_op_add_i64:
1421        if (c2) {
1422            tcg_out_addi(s, TCG_TYPE_I64, a0, a1, a2);
1423        } else {
1424            tcg_out_opc_add_d(s, a0, a1, a2);
1425        }
1426        break;
1427
1428    case INDEX_op_sub_i32:
1429        if (c2) {
1430            tcg_out_addi(s, TCG_TYPE_I32, a0, a1, -a2);
1431        } else {
1432            tcg_out_opc_sub_w(s, a0, a1, a2);
1433        }
1434        break;
1435    case INDEX_op_sub_i64:
1436        if (c2) {
1437            tcg_out_addi(s, TCG_TYPE_I64, a0, a1, -a2);
1438        } else {
1439            tcg_out_opc_sub_d(s, a0, a1, a2);
1440        }
1441        break;
1442
1443    case INDEX_op_mul_i32:
1444        tcg_out_opc_mul_w(s, a0, a1, a2);
1445        break;
1446    case INDEX_op_mul_i64:
1447        tcg_out_opc_mul_d(s, a0, a1, a2);
1448        break;
1449
1450    case INDEX_op_mulsh_i32:
1451        tcg_out_opc_mulh_w(s, a0, a1, a2);
1452        break;
1453    case INDEX_op_mulsh_i64:
1454        tcg_out_opc_mulh_d(s, a0, a1, a2);
1455        break;
1456
1457    case INDEX_op_muluh_i32:
1458        tcg_out_opc_mulh_wu(s, a0, a1, a2);
1459        break;
1460    case INDEX_op_muluh_i64:
1461        tcg_out_opc_mulh_du(s, a0, a1, a2);
1462        break;
1463
1464    case INDEX_op_div_i32:
1465        tcg_out_opc_div_w(s, a0, a1, a2);
1466        break;
1467    case INDEX_op_div_i64:
1468        tcg_out_opc_div_d(s, a0, a1, a2);
1469        break;
1470
1471    case INDEX_op_divu_i32:
1472        tcg_out_opc_div_wu(s, a0, a1, a2);
1473        break;
1474    case INDEX_op_divu_i64:
1475        tcg_out_opc_div_du(s, a0, a1, a2);
1476        break;
1477
1478    case INDEX_op_rem_i32:
1479        tcg_out_opc_mod_w(s, a0, a1, a2);
1480        break;
1481    case INDEX_op_rem_i64:
1482        tcg_out_opc_mod_d(s, a0, a1, a2);
1483        break;
1484
1485    case INDEX_op_remu_i32:
1486        tcg_out_opc_mod_wu(s, a0, a1, a2);
1487        break;
1488    case INDEX_op_remu_i64:
1489        tcg_out_opc_mod_du(s, a0, a1, a2);
1490        break;
1491
1492    case INDEX_op_setcond_i32:
1493    case INDEX_op_setcond_i64:
1494        tcg_out_setcond(s, args[3], a0, a1, a2, c2);
1495        break;
1496
1497    case INDEX_op_movcond_i32:
1498    case INDEX_op_movcond_i64:
1499        tcg_out_movcond(s, args[5], a0, a1, a2, c2, args[3], args[4]);
1500        break;
1501
1502    case INDEX_op_ld8s_i32:
1503    case INDEX_op_ld8s_i64:
1504        tcg_out_ldst(s, OPC_LD_B, a0, a1, a2);
1505        break;
1506    case INDEX_op_ld8u_i32:
1507    case INDEX_op_ld8u_i64:
1508        tcg_out_ldst(s, OPC_LD_BU, a0, a1, a2);
1509        break;
1510    case INDEX_op_ld16s_i32:
1511    case INDEX_op_ld16s_i64:
1512        tcg_out_ldst(s, OPC_LD_H, a0, a1, a2);
1513        break;
1514    case INDEX_op_ld16u_i32:
1515    case INDEX_op_ld16u_i64:
1516        tcg_out_ldst(s, OPC_LD_HU, a0, a1, a2);
1517        break;
1518    case INDEX_op_ld_i32:
1519    case INDEX_op_ld32s_i64:
1520        tcg_out_ldst(s, OPC_LD_W, a0, a1, a2);
1521        break;
1522    case INDEX_op_ld32u_i64:
1523        tcg_out_ldst(s, OPC_LD_WU, a0, a1, a2);
1524        break;
1525    case INDEX_op_ld_i64:
1526        tcg_out_ldst(s, OPC_LD_D, a0, a1, a2);
1527        break;
1528
1529    case INDEX_op_st8_i32:
1530    case INDEX_op_st8_i64:
1531        tcg_out_ldst(s, OPC_ST_B, a0, a1, a2);
1532        break;
1533    case INDEX_op_st16_i32:
1534    case INDEX_op_st16_i64:
1535        tcg_out_ldst(s, OPC_ST_H, a0, a1, a2);
1536        break;
1537    case INDEX_op_st_i32:
1538    case INDEX_op_st32_i64:
1539        tcg_out_ldst(s, OPC_ST_W, a0, a1, a2);
1540        break;
1541    case INDEX_op_st_i64:
1542        tcg_out_ldst(s, OPC_ST_D, a0, a1, a2);
1543        break;
1544
1545    case INDEX_op_qemu_ld_a32_i32:
1546    case INDEX_op_qemu_ld_a64_i32:
1547        tcg_out_qemu_ld(s, a0, a1, a2, TCG_TYPE_I32);
1548        break;
1549    case INDEX_op_qemu_ld_a32_i64:
1550    case INDEX_op_qemu_ld_a64_i64:
1551        tcg_out_qemu_ld(s, a0, a1, a2, TCG_TYPE_I64);
1552        break;
1553    case INDEX_op_qemu_ld_a32_i128:
1554    case INDEX_op_qemu_ld_a64_i128:
1555        tcg_out_qemu_ldst_i128(s, a0, a1, a2, a3, true);
1556        break;
1557    case INDEX_op_qemu_st_a32_i32:
1558    case INDEX_op_qemu_st_a64_i32:
1559        tcg_out_qemu_st(s, a0, a1, a2, TCG_TYPE_I32);
1560        break;
1561    case INDEX_op_qemu_st_a32_i64:
1562    case INDEX_op_qemu_st_a64_i64:
1563        tcg_out_qemu_st(s, a0, a1, a2, TCG_TYPE_I64);
1564        break;
1565    case INDEX_op_qemu_st_a32_i128:
1566    case INDEX_op_qemu_st_a64_i128:
1567        tcg_out_qemu_ldst_i128(s, a0, a1, a2, a3, false);
1568        break;
1569
1570    case INDEX_op_mov_i32:  /* Always emitted via tcg_out_mov.  */
1571    case INDEX_op_mov_i64:
1572    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
1573    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
1574    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
1575    case INDEX_op_ext8s_i32:  /* Always emitted via tcg_reg_alloc_op.  */
1576    case INDEX_op_ext8s_i64:
1577    case INDEX_op_ext8u_i32:
1578    case INDEX_op_ext8u_i64:
1579    case INDEX_op_ext16s_i32:
1580    case INDEX_op_ext16s_i64:
1581    case INDEX_op_ext16u_i32:
1582    case INDEX_op_ext16u_i64:
1583    case INDEX_op_ext32s_i64:
1584    case INDEX_op_ext32u_i64:
1585    case INDEX_op_ext_i32_i64:
1586    case INDEX_op_extu_i32_i64:
1587    case INDEX_op_extrl_i64_i32:
1588    default:
1589        g_assert_not_reached();
1590    }
1591}
1592
1593static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
1594                            TCGReg rd, TCGReg rs)
1595{
1596    switch (vece) {
1597    case MO_8:
1598        tcg_out_opc_vreplgr2vr_b(s, rd, rs);
1599        break;
1600    case MO_16:
1601        tcg_out_opc_vreplgr2vr_h(s, rd, rs);
1602        break;
1603    case MO_32:
1604        tcg_out_opc_vreplgr2vr_w(s, rd, rs);
1605        break;
1606    case MO_64:
1607        tcg_out_opc_vreplgr2vr_d(s, rd, rs);
1608        break;
1609    default:
1610        g_assert_not_reached();
1611    }
1612    return true;
1613}
1614
1615static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
1616                             TCGReg r, TCGReg base, intptr_t offset)
1617{
1618    /* Handle imm overflow and division (vldrepl.d imm is divided by 8) */
1619    if (offset < -0x800 || offset > 0x7ff || \
1620        (offset & ((1 << vece) - 1)) != 0) {
1621        tcg_out_addi(s, TCG_TYPE_I64, TCG_REG_TMP0, base, offset);
1622        base = TCG_REG_TMP0;
1623        offset = 0;
1624    }
1625    offset >>= vece;
1626
1627    switch (vece) {
1628    case MO_8:
1629        tcg_out_opc_vldrepl_b(s, r, base, offset);
1630        break;
1631    case MO_16:
1632        tcg_out_opc_vldrepl_h(s, r, base, offset);
1633        break;
1634    case MO_32:
1635        tcg_out_opc_vldrepl_w(s, r, base, offset);
1636        break;
1637    case MO_64:
1638        tcg_out_opc_vldrepl_d(s, r, base, offset);
1639        break;
1640    default:
1641        g_assert_not_reached();
1642    }
1643    return true;
1644}
1645
1646static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
1647                             TCGReg rd, int64_t v64)
1648{
1649    /* Try vldi if imm can fit */
1650    int64_t value = sextract64(v64, 0, 8 << vece);
1651    if (-0x200 <= value && value <= 0x1FF) {
1652        uint32_t imm = (vece << 10) | ((uint32_t)v64 & 0x3FF);
1653        tcg_out_opc_vldi(s, rd, imm);
1654        return;
1655    }
1656
1657    /* TODO: vldi patterns when imm 12 is set */
1658
1659    /* Fallback to vreplgr2vr */
1660    tcg_out_movi(s, TCG_TYPE_I64, TCG_REG_TMP0, value);
1661    switch (vece) {
1662    case MO_8:
1663        tcg_out_opc_vreplgr2vr_b(s, rd, TCG_REG_TMP0);
1664        break;
1665    case MO_16:
1666        tcg_out_opc_vreplgr2vr_h(s, rd, TCG_REG_TMP0);
1667        break;
1668    case MO_32:
1669        tcg_out_opc_vreplgr2vr_w(s, rd, TCG_REG_TMP0);
1670        break;
1671    case MO_64:
1672        tcg_out_opc_vreplgr2vr_d(s, rd, TCG_REG_TMP0);
1673        break;
1674    default:
1675        g_assert_not_reached();
1676    }
1677}
1678
1679static void tcg_out_addsub_vec(TCGContext *s, unsigned vece, const TCGArg a0,
1680                               const TCGArg a1, const TCGArg a2,
1681                               bool a2_is_const, bool is_add)
1682{
1683    static const LoongArchInsn add_vec_insn[4] = {
1684        OPC_VADD_B, OPC_VADD_H, OPC_VADD_W, OPC_VADD_D
1685    };
1686    static const LoongArchInsn add_vec_imm_insn[4] = {
1687        OPC_VADDI_BU, OPC_VADDI_HU, OPC_VADDI_WU, OPC_VADDI_DU
1688    };
1689    static const LoongArchInsn sub_vec_insn[4] = {
1690        OPC_VSUB_B, OPC_VSUB_H, OPC_VSUB_W, OPC_VSUB_D
1691    };
1692    static const LoongArchInsn sub_vec_imm_insn[4] = {
1693        OPC_VSUBI_BU, OPC_VSUBI_HU, OPC_VSUBI_WU, OPC_VSUBI_DU
1694    };
1695
1696    if (a2_is_const) {
1697        int64_t value = sextract64(a2, 0, 8 << vece);
1698        if (!is_add) {
1699            value = -value;
1700        }
1701
1702        /* Try vaddi/vsubi */
1703        if (0 <= value && value <= 0x1f) {
1704            tcg_out32(s, encode_vdvjuk5_insn(add_vec_imm_insn[vece], a0, \
1705                                             a1, value));
1706            return;
1707        } else if (-0x1f <= value && value < 0) {
1708            tcg_out32(s, encode_vdvjuk5_insn(sub_vec_imm_insn[vece], a0, \
1709                                             a1, -value));
1710            return;
1711        }
1712
1713        /* constraint TCG_CT_CONST_VADD ensures unreachable */
1714        g_assert_not_reached();
1715    }
1716
1717    if (is_add) {
1718        tcg_out32(s, encode_vdvjvk_insn(add_vec_insn[vece], a0, a1, a2));
1719    } else {
1720        tcg_out32(s, encode_vdvjvk_insn(sub_vec_insn[vece], a0, a1, a2));
1721    }
1722}
1723
1724static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
1725                           unsigned vecl, unsigned vece,
1726                           const TCGArg args[TCG_MAX_OP_ARGS],
1727                           const int const_args[TCG_MAX_OP_ARGS])
1728{
1729    TCGType type = vecl + TCG_TYPE_V64;
1730    TCGArg a0, a1, a2, a3;
1731    TCGReg temp = TCG_REG_TMP0;
1732    TCGReg temp_vec = TCG_VEC_TMP0;
1733
1734    static const LoongArchInsn cmp_vec_insn[16][4] = {
1735        [TCG_COND_EQ] = {OPC_VSEQ_B, OPC_VSEQ_H, OPC_VSEQ_W, OPC_VSEQ_D},
1736        [TCG_COND_LE] = {OPC_VSLE_B, OPC_VSLE_H, OPC_VSLE_W, OPC_VSLE_D},
1737        [TCG_COND_LEU] = {OPC_VSLE_BU, OPC_VSLE_HU, OPC_VSLE_WU, OPC_VSLE_DU},
1738        [TCG_COND_LT] = {OPC_VSLT_B, OPC_VSLT_H, OPC_VSLT_W, OPC_VSLT_D},
1739        [TCG_COND_LTU] = {OPC_VSLT_BU, OPC_VSLT_HU, OPC_VSLT_WU, OPC_VSLT_DU},
1740    };
1741    static const LoongArchInsn cmp_vec_imm_insn[16][4] = {
1742        [TCG_COND_EQ] = {OPC_VSEQI_B, OPC_VSEQI_H, OPC_VSEQI_W, OPC_VSEQI_D},
1743        [TCG_COND_LE] = {OPC_VSLEI_B, OPC_VSLEI_H, OPC_VSLEI_W, OPC_VSLEI_D},
1744        [TCG_COND_LEU] = {OPC_VSLEI_BU, OPC_VSLEI_HU, OPC_VSLEI_WU, OPC_VSLEI_DU},
1745        [TCG_COND_LT] = {OPC_VSLTI_B, OPC_VSLTI_H, OPC_VSLTI_W, OPC_VSLTI_D},
1746        [TCG_COND_LTU] = {OPC_VSLTI_BU, OPC_VSLTI_HU, OPC_VSLTI_WU, OPC_VSLTI_DU},
1747    };
1748    LoongArchInsn insn;
1749    static const LoongArchInsn neg_vec_insn[4] = {
1750        OPC_VNEG_B, OPC_VNEG_H, OPC_VNEG_W, OPC_VNEG_D
1751    };
1752    static const LoongArchInsn mul_vec_insn[4] = {
1753        OPC_VMUL_B, OPC_VMUL_H, OPC_VMUL_W, OPC_VMUL_D
1754    };
1755    static const LoongArchInsn smin_vec_insn[4] = {
1756        OPC_VMIN_B, OPC_VMIN_H, OPC_VMIN_W, OPC_VMIN_D
1757    };
1758    static const LoongArchInsn umin_vec_insn[4] = {
1759        OPC_VMIN_BU, OPC_VMIN_HU, OPC_VMIN_WU, OPC_VMIN_DU
1760    };
1761    static const LoongArchInsn smax_vec_insn[4] = {
1762        OPC_VMAX_B, OPC_VMAX_H, OPC_VMAX_W, OPC_VMAX_D
1763    };
1764    static const LoongArchInsn umax_vec_insn[4] = {
1765        OPC_VMAX_BU, OPC_VMAX_HU, OPC_VMAX_WU, OPC_VMAX_DU
1766    };
1767    static const LoongArchInsn ssadd_vec_insn[4] = {
1768        OPC_VSADD_B, OPC_VSADD_H, OPC_VSADD_W, OPC_VSADD_D
1769    };
1770    static const LoongArchInsn usadd_vec_insn[4] = {
1771        OPC_VSADD_BU, OPC_VSADD_HU, OPC_VSADD_WU, OPC_VSADD_DU
1772    };
1773    static const LoongArchInsn sssub_vec_insn[4] = {
1774        OPC_VSSUB_B, OPC_VSSUB_H, OPC_VSSUB_W, OPC_VSSUB_D
1775    };
1776    static const LoongArchInsn ussub_vec_insn[4] = {
1777        OPC_VSSUB_BU, OPC_VSSUB_HU, OPC_VSSUB_WU, OPC_VSSUB_DU
1778    };
1779    static const LoongArchInsn shlv_vec_insn[4] = {
1780        OPC_VSLL_B, OPC_VSLL_H, OPC_VSLL_W, OPC_VSLL_D
1781    };
1782    static const LoongArchInsn shrv_vec_insn[4] = {
1783        OPC_VSRL_B, OPC_VSRL_H, OPC_VSRL_W, OPC_VSRL_D
1784    };
1785    static const LoongArchInsn sarv_vec_insn[4] = {
1786        OPC_VSRA_B, OPC_VSRA_H, OPC_VSRA_W, OPC_VSRA_D
1787    };
1788    static const LoongArchInsn shli_vec_insn[4] = {
1789        OPC_VSLLI_B, OPC_VSLLI_H, OPC_VSLLI_W, OPC_VSLLI_D
1790    };
1791    static const LoongArchInsn shri_vec_insn[4] = {
1792        OPC_VSRLI_B, OPC_VSRLI_H, OPC_VSRLI_W, OPC_VSRLI_D
1793    };
1794    static const LoongArchInsn sari_vec_insn[4] = {
1795        OPC_VSRAI_B, OPC_VSRAI_H, OPC_VSRAI_W, OPC_VSRAI_D
1796    };
1797    static const LoongArchInsn rotrv_vec_insn[4] = {
1798        OPC_VROTR_B, OPC_VROTR_H, OPC_VROTR_W, OPC_VROTR_D
1799    };
1800
1801    a0 = args[0];
1802    a1 = args[1];
1803    a2 = args[2];
1804    a3 = args[3];
1805
1806    /* Currently only supports V128 */
1807    tcg_debug_assert(type == TCG_TYPE_V128);
1808
1809    switch (opc) {
1810    case INDEX_op_st_vec:
1811        /* Try to fit vst imm */
1812        if (-0x800 <= a2 && a2 <= 0x7ff) {
1813            tcg_out_opc_vst(s, a0, a1, a2);
1814        } else {
1815            tcg_out_movi(s, TCG_TYPE_I64, temp, a2);
1816            tcg_out_opc_vstx(s, a0, a1, temp);
1817        }
1818        break;
1819    case INDEX_op_ld_vec:
1820        /* Try to fit vld imm */
1821        if (-0x800 <= a2 && a2 <= 0x7ff) {
1822            tcg_out_opc_vld(s, a0, a1, a2);
1823        } else {
1824            tcg_out_movi(s, TCG_TYPE_I64, temp, a2);
1825            tcg_out_opc_vldx(s, a0, a1, temp);
1826        }
1827        break;
1828    case INDEX_op_and_vec:
1829        tcg_out_opc_vand_v(s, a0, a1, a2);
1830        break;
1831    case INDEX_op_andc_vec:
1832        /*
1833         * vandn vd, vj, vk: vd = vk & ~vj
1834         * andc_vec vd, vj, vk: vd = vj & ~vk
1835         * vk and vk are swapped
1836         */
1837        tcg_out_opc_vandn_v(s, a0, a2, a1);
1838        break;
1839    case INDEX_op_or_vec:
1840        tcg_out_opc_vor_v(s, a0, a1, a2);
1841        break;
1842    case INDEX_op_orc_vec:
1843        tcg_out_opc_vorn_v(s, a0, a1, a2);
1844        break;
1845    case INDEX_op_xor_vec:
1846        tcg_out_opc_vxor_v(s, a0, a1, a2);
1847        break;
1848    case INDEX_op_nor_vec:
1849        tcg_out_opc_vnor_v(s, a0, a1, a2);
1850        break;
1851    case INDEX_op_not_vec:
1852        tcg_out_opc_vnor_v(s, a0, a1, a1);
1853        break;
1854    case INDEX_op_cmp_vec:
1855        TCGCond cond = args[3];
1856        if (const_args[2]) {
1857            /*
1858             * cmp_vec dest, src, value
1859             * Try vseqi/vslei/vslti
1860             */
1861            int64_t value = sextract64(a2, 0, 8 << vece);
1862            if ((cond == TCG_COND_EQ || cond == TCG_COND_LE || \
1863                 cond == TCG_COND_LT) && (-0x10 <= value && value <= 0x0f)) {
1864                tcg_out32(s, encode_vdvjsk5_insn(cmp_vec_imm_insn[cond][vece], \
1865                                                 a0, a1, value));
1866                break;
1867            } else if ((cond == TCG_COND_LEU || cond == TCG_COND_LTU) &&
1868                (0x00 <= value && value <= 0x1f)) {
1869                tcg_out32(s, encode_vdvjuk5_insn(cmp_vec_imm_insn[cond][vece], \
1870                                                 a0, a1, value));
1871                break;
1872            }
1873
1874            /*
1875             * Fallback to:
1876             * dupi_vec temp, a2
1877             * cmp_vec a0, a1, temp, cond
1878             */
1879            tcg_out_dupi_vec(s, type, vece, temp_vec, a2);
1880            a2 = temp_vec;
1881        }
1882
1883        insn = cmp_vec_insn[cond][vece];
1884        if (insn == 0) {
1885            TCGArg t;
1886            t = a1, a1 = a2, a2 = t;
1887            cond = tcg_swap_cond(cond);
1888            insn = cmp_vec_insn[cond][vece];
1889            tcg_debug_assert(insn != 0);
1890        }
1891        tcg_out32(s, encode_vdvjvk_insn(insn, a0, a1, a2));
1892        break;
1893    case INDEX_op_add_vec:
1894        tcg_out_addsub_vec(s, vece, a0, a1, a2, const_args[2], true);
1895        break;
1896    case INDEX_op_sub_vec:
1897        tcg_out_addsub_vec(s, vece, a0, a1, a2, const_args[2], false);
1898        break;
1899    case INDEX_op_neg_vec:
1900        tcg_out32(s, encode_vdvj_insn(neg_vec_insn[vece], a0, a1));
1901        break;
1902    case INDEX_op_mul_vec:
1903        tcg_out32(s, encode_vdvjvk_insn(mul_vec_insn[vece], a0, a1, a2));
1904        break;
1905    case INDEX_op_smin_vec:
1906        tcg_out32(s, encode_vdvjvk_insn(smin_vec_insn[vece], a0, a1, a2));
1907        break;
1908    case INDEX_op_smax_vec:
1909        tcg_out32(s, encode_vdvjvk_insn(smax_vec_insn[vece], a0, a1, a2));
1910        break;
1911    case INDEX_op_umin_vec:
1912        tcg_out32(s, encode_vdvjvk_insn(umin_vec_insn[vece], a0, a1, a2));
1913        break;
1914    case INDEX_op_umax_vec:
1915        tcg_out32(s, encode_vdvjvk_insn(umax_vec_insn[vece], a0, a1, a2));
1916        break;
1917    case INDEX_op_ssadd_vec:
1918        tcg_out32(s, encode_vdvjvk_insn(ssadd_vec_insn[vece], a0, a1, a2));
1919        break;
1920    case INDEX_op_usadd_vec:
1921        tcg_out32(s, encode_vdvjvk_insn(usadd_vec_insn[vece], a0, a1, a2));
1922        break;
1923    case INDEX_op_sssub_vec:
1924        tcg_out32(s, encode_vdvjvk_insn(sssub_vec_insn[vece], a0, a1, a2));
1925        break;
1926    case INDEX_op_ussub_vec:
1927        tcg_out32(s, encode_vdvjvk_insn(ussub_vec_insn[vece], a0, a1, a2));
1928        break;
1929    case INDEX_op_shlv_vec:
1930        tcg_out32(s, encode_vdvjvk_insn(shlv_vec_insn[vece], a0, a1, a2));
1931        break;
1932    case INDEX_op_shrv_vec:
1933        tcg_out32(s, encode_vdvjvk_insn(shrv_vec_insn[vece], a0, a1, a2));
1934        break;
1935    case INDEX_op_sarv_vec:
1936        tcg_out32(s, encode_vdvjvk_insn(sarv_vec_insn[vece], a0, a1, a2));
1937        break;
1938    case INDEX_op_shli_vec:
1939        tcg_out32(s, encode_vdvjuk3_insn(shli_vec_insn[vece], a0, a1, a2));
1940        break;
1941    case INDEX_op_shri_vec:
1942        tcg_out32(s, encode_vdvjuk3_insn(shri_vec_insn[vece], a0, a1, a2));
1943        break;
1944    case INDEX_op_sari_vec:
1945        tcg_out32(s, encode_vdvjuk3_insn(sari_vec_insn[vece], a0, a1, a2));
1946        break;
1947    case INDEX_op_rotrv_vec:
1948        tcg_out32(s, encode_vdvjvk_insn(rotrv_vec_insn[vece], a0, a1, a2));
1949        break;
1950    case INDEX_op_rotlv_vec:
1951        /* rotlv_vec a1, a2 = rotrv_vec a1, -a2 */
1952        tcg_out32(s, encode_vdvj_insn(neg_vec_insn[vece], temp_vec, a2));
1953        tcg_out32(s, encode_vdvjvk_insn(rotrv_vec_insn[vece], a0, a1,
1954                                        temp_vec));
1955        break;
1956    case INDEX_op_rotli_vec:
1957        /* rotli_vec a1, a2 = rotri_vec a1, -a2 */
1958        a2 = extract32(-a2, 0, 3 + vece);
1959        switch (vece) {
1960        case MO_8:
1961            tcg_out_opc_vrotri_b(s, a0, a1, a2);
1962            break;
1963        case MO_16:
1964            tcg_out_opc_vrotri_h(s, a0, a1, a2);
1965            break;
1966        case MO_32:
1967            tcg_out_opc_vrotri_w(s, a0, a1, a2);
1968            break;
1969        case MO_64:
1970            tcg_out_opc_vrotri_d(s, a0, a1, a2);
1971            break;
1972        default:
1973            g_assert_not_reached();
1974        }
1975        break;
1976    case INDEX_op_bitsel_vec:
1977        /* vbitsel vd, vj, vk, va = bitsel_vec vd, va, vk, vj */
1978        tcg_out_opc_vbitsel_v(s, a0, a3, a2, a1);
1979        break;
1980    case INDEX_op_dupm_vec:
1981        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
1982        break;
1983    default:
1984        g_assert_not_reached();
1985    }
1986}
1987
1988int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
1989{
1990    switch (opc) {
1991    case INDEX_op_ld_vec:
1992    case INDEX_op_st_vec:
1993    case INDEX_op_dup_vec:
1994    case INDEX_op_dupm_vec:
1995    case INDEX_op_cmp_vec:
1996    case INDEX_op_add_vec:
1997    case INDEX_op_sub_vec:
1998    case INDEX_op_and_vec:
1999    case INDEX_op_andc_vec:
2000    case INDEX_op_or_vec:
2001    case INDEX_op_orc_vec:
2002    case INDEX_op_xor_vec:
2003    case INDEX_op_nor_vec:
2004    case INDEX_op_not_vec:
2005    case INDEX_op_neg_vec:
2006    case INDEX_op_mul_vec:
2007    case INDEX_op_smin_vec:
2008    case INDEX_op_smax_vec:
2009    case INDEX_op_umin_vec:
2010    case INDEX_op_umax_vec:
2011    case INDEX_op_ssadd_vec:
2012    case INDEX_op_usadd_vec:
2013    case INDEX_op_sssub_vec:
2014    case INDEX_op_ussub_vec:
2015    case INDEX_op_shlv_vec:
2016    case INDEX_op_shrv_vec:
2017    case INDEX_op_sarv_vec:
2018    case INDEX_op_bitsel_vec:
2019        return 1;
2020    default:
2021        return 0;
2022    }
2023}
2024
2025void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
2026                       TCGArg a0, ...)
2027{
2028    g_assert_not_reached();
2029}
2030
2031static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
2032{
2033    switch (op) {
2034    case INDEX_op_goto_ptr:
2035        return C_O0_I1(r);
2036
2037    case INDEX_op_st8_i32:
2038    case INDEX_op_st8_i64:
2039    case INDEX_op_st16_i32:
2040    case INDEX_op_st16_i64:
2041    case INDEX_op_st32_i64:
2042    case INDEX_op_st_i32:
2043    case INDEX_op_st_i64:
2044    case INDEX_op_qemu_st_a32_i32:
2045    case INDEX_op_qemu_st_a64_i32:
2046    case INDEX_op_qemu_st_a32_i64:
2047    case INDEX_op_qemu_st_a64_i64:
2048        return C_O0_I2(rZ, r);
2049
2050    case INDEX_op_qemu_ld_a32_i128:
2051    case INDEX_op_qemu_ld_a64_i128:
2052        return C_O2_I1(r, r, r);
2053
2054    case INDEX_op_qemu_st_a32_i128:
2055    case INDEX_op_qemu_st_a64_i128:
2056        return C_O0_I3(r, r, r);
2057
2058    case INDEX_op_brcond_i32:
2059    case INDEX_op_brcond_i64:
2060        return C_O0_I2(rZ, rZ);
2061
2062    case INDEX_op_ext8s_i32:
2063    case INDEX_op_ext8s_i64:
2064    case INDEX_op_ext8u_i32:
2065    case INDEX_op_ext8u_i64:
2066    case INDEX_op_ext16s_i32:
2067    case INDEX_op_ext16s_i64:
2068    case INDEX_op_ext16u_i32:
2069    case INDEX_op_ext16u_i64:
2070    case INDEX_op_ext32s_i64:
2071    case INDEX_op_ext32u_i64:
2072    case INDEX_op_extu_i32_i64:
2073    case INDEX_op_extrl_i64_i32:
2074    case INDEX_op_extrh_i64_i32:
2075    case INDEX_op_ext_i32_i64:
2076    case INDEX_op_not_i32:
2077    case INDEX_op_not_i64:
2078    case INDEX_op_extract_i32:
2079    case INDEX_op_extract_i64:
2080    case INDEX_op_bswap16_i32:
2081    case INDEX_op_bswap16_i64:
2082    case INDEX_op_bswap32_i32:
2083    case INDEX_op_bswap32_i64:
2084    case INDEX_op_bswap64_i64:
2085    case INDEX_op_ld8s_i32:
2086    case INDEX_op_ld8s_i64:
2087    case INDEX_op_ld8u_i32:
2088    case INDEX_op_ld8u_i64:
2089    case INDEX_op_ld16s_i32:
2090    case INDEX_op_ld16s_i64:
2091    case INDEX_op_ld16u_i32:
2092    case INDEX_op_ld16u_i64:
2093    case INDEX_op_ld32s_i64:
2094    case INDEX_op_ld32u_i64:
2095    case INDEX_op_ld_i32:
2096    case INDEX_op_ld_i64:
2097    case INDEX_op_qemu_ld_a32_i32:
2098    case INDEX_op_qemu_ld_a64_i32:
2099    case INDEX_op_qemu_ld_a32_i64:
2100    case INDEX_op_qemu_ld_a64_i64:
2101        return C_O1_I1(r, r);
2102
2103    case INDEX_op_andc_i32:
2104    case INDEX_op_andc_i64:
2105    case INDEX_op_orc_i32:
2106    case INDEX_op_orc_i64:
2107        /*
2108         * LoongArch insns for these ops don't have reg-imm forms, but we
2109         * can express using andi/ori if ~constant satisfies
2110         * TCG_CT_CONST_U12.
2111         */
2112        return C_O1_I2(r, r, rC);
2113
2114    case INDEX_op_shl_i32:
2115    case INDEX_op_shl_i64:
2116    case INDEX_op_shr_i32:
2117    case INDEX_op_shr_i64:
2118    case INDEX_op_sar_i32:
2119    case INDEX_op_sar_i64:
2120    case INDEX_op_rotl_i32:
2121    case INDEX_op_rotl_i64:
2122    case INDEX_op_rotr_i32:
2123    case INDEX_op_rotr_i64:
2124        return C_O1_I2(r, r, ri);
2125
2126    case INDEX_op_add_i32:
2127        return C_O1_I2(r, r, ri);
2128    case INDEX_op_add_i64:
2129        return C_O1_I2(r, r, rJ);
2130
2131    case INDEX_op_and_i32:
2132    case INDEX_op_and_i64:
2133    case INDEX_op_nor_i32:
2134    case INDEX_op_nor_i64:
2135    case INDEX_op_or_i32:
2136    case INDEX_op_or_i64:
2137    case INDEX_op_xor_i32:
2138    case INDEX_op_xor_i64:
2139        /* LoongArch reg-imm bitops have their imms ZERO-extended */
2140        return C_O1_I2(r, r, rU);
2141
2142    case INDEX_op_clz_i32:
2143    case INDEX_op_clz_i64:
2144    case INDEX_op_ctz_i32:
2145    case INDEX_op_ctz_i64:
2146        return C_O1_I2(r, r, rW);
2147
2148    case INDEX_op_deposit_i32:
2149    case INDEX_op_deposit_i64:
2150        /* Must deposit into the same register as input */
2151        return C_O1_I2(r, 0, rZ);
2152
2153    case INDEX_op_sub_i32:
2154    case INDEX_op_setcond_i32:
2155        return C_O1_I2(r, rZ, ri);
2156    case INDEX_op_sub_i64:
2157    case INDEX_op_setcond_i64:
2158        return C_O1_I2(r, rZ, rJ);
2159
2160    case INDEX_op_mul_i32:
2161    case INDEX_op_mul_i64:
2162    case INDEX_op_mulsh_i32:
2163    case INDEX_op_mulsh_i64:
2164    case INDEX_op_muluh_i32:
2165    case INDEX_op_muluh_i64:
2166    case INDEX_op_div_i32:
2167    case INDEX_op_div_i64:
2168    case INDEX_op_divu_i32:
2169    case INDEX_op_divu_i64:
2170    case INDEX_op_rem_i32:
2171    case INDEX_op_rem_i64:
2172    case INDEX_op_remu_i32:
2173    case INDEX_op_remu_i64:
2174        return C_O1_I2(r, rZ, rZ);
2175
2176    case INDEX_op_movcond_i32:
2177    case INDEX_op_movcond_i64:
2178        return C_O1_I4(r, rZ, rJ, rZ, rZ);
2179
2180    case INDEX_op_ld_vec:
2181    case INDEX_op_dupm_vec:
2182    case INDEX_op_dup_vec:
2183        return C_O1_I1(w, r);
2184
2185    case INDEX_op_st_vec:
2186        return C_O0_I2(w, r);
2187
2188    case INDEX_op_cmp_vec:
2189        return C_O1_I2(w, w, wM);
2190
2191    case INDEX_op_add_vec:
2192    case INDEX_op_sub_vec:
2193        return C_O1_I2(w, w, wA);
2194
2195    case INDEX_op_and_vec:
2196    case INDEX_op_andc_vec:
2197    case INDEX_op_or_vec:
2198    case INDEX_op_orc_vec:
2199    case INDEX_op_xor_vec:
2200    case INDEX_op_nor_vec:
2201    case INDEX_op_mul_vec:
2202    case INDEX_op_smin_vec:
2203    case INDEX_op_smax_vec:
2204    case INDEX_op_umin_vec:
2205    case INDEX_op_umax_vec:
2206    case INDEX_op_ssadd_vec:
2207    case INDEX_op_usadd_vec:
2208    case INDEX_op_sssub_vec:
2209    case INDEX_op_ussub_vec:
2210    case INDEX_op_shlv_vec:
2211    case INDEX_op_shrv_vec:
2212    case INDEX_op_sarv_vec:
2213    case INDEX_op_rotrv_vec:
2214    case INDEX_op_rotlv_vec:
2215        return C_O1_I2(w, w, w);
2216
2217    case INDEX_op_not_vec:
2218    case INDEX_op_neg_vec:
2219    case INDEX_op_shli_vec:
2220    case INDEX_op_shri_vec:
2221    case INDEX_op_sari_vec:
2222    case INDEX_op_rotli_vec:
2223        return C_O1_I1(w, w);
2224
2225    case INDEX_op_bitsel_vec:
2226        return C_O1_I3(w, w, w, w);
2227
2228    default:
2229        g_assert_not_reached();
2230    }
2231}
2232
2233static const int tcg_target_callee_save_regs[] = {
2234    TCG_REG_S0,     /* used for the global env (TCG_AREG0) */
2235    TCG_REG_S1,
2236    TCG_REG_S2,
2237    TCG_REG_S3,
2238    TCG_REG_S4,
2239    TCG_REG_S5,
2240    TCG_REG_S6,
2241    TCG_REG_S7,
2242    TCG_REG_S8,
2243    TCG_REG_S9,
2244    TCG_REG_RA,     /* should be last for ABI compliance */
2245};
2246
2247/* Stack frame parameters.  */
2248#define REG_SIZE   (TCG_TARGET_REG_BITS / 8)
2249#define SAVE_SIZE  ((int)ARRAY_SIZE(tcg_target_callee_save_regs) * REG_SIZE)
2250#define TEMP_SIZE  (CPU_TEMP_BUF_NLONGS * (int)sizeof(long))
2251#define FRAME_SIZE ((TCG_STATIC_CALL_ARGS_SIZE + TEMP_SIZE + SAVE_SIZE \
2252                     + TCG_TARGET_STACK_ALIGN - 1) \
2253                    & -TCG_TARGET_STACK_ALIGN)
2254#define SAVE_OFS   (TCG_STATIC_CALL_ARGS_SIZE + TEMP_SIZE)
2255
2256/* We're expecting to be able to use an immediate for frame allocation.  */
2257QEMU_BUILD_BUG_ON(FRAME_SIZE > 0x7ff);
2258
2259/* Generate global QEMU prologue and epilogue code */
2260static void tcg_target_qemu_prologue(TCGContext *s)
2261{
2262    int i;
2263
2264    tcg_set_frame(s, TCG_REG_SP, TCG_STATIC_CALL_ARGS_SIZE, TEMP_SIZE);
2265
2266    /* TB prologue */
2267    tcg_out_opc_addi_d(s, TCG_REG_SP, TCG_REG_SP, -FRAME_SIZE);
2268    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
2269        tcg_out_st(s, TCG_TYPE_REG, tcg_target_callee_save_regs[i],
2270                   TCG_REG_SP, SAVE_OFS + i * REG_SIZE);
2271    }
2272
2273#if !defined(CONFIG_SOFTMMU)
2274    if (USE_GUEST_BASE) {
2275        tcg_out_movi(s, TCG_TYPE_PTR, TCG_GUEST_BASE_REG, guest_base);
2276        tcg_regset_set_reg(s->reserved_regs, TCG_GUEST_BASE_REG);
2277    }
2278#endif
2279
2280    /* Call generated code */
2281    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
2282    tcg_out_opc_jirl(s, TCG_REG_ZERO, tcg_target_call_iarg_regs[1], 0);
2283
2284    /* Return path for goto_ptr. Set return value to 0 */
2285    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
2286    tcg_out_mov(s, TCG_TYPE_REG, TCG_REG_A0, TCG_REG_ZERO);
2287
2288    /* TB epilogue */
2289    tb_ret_addr = tcg_splitwx_to_rx(s->code_ptr);
2290    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
2291        tcg_out_ld(s, TCG_TYPE_REG, tcg_target_callee_save_regs[i],
2292                   TCG_REG_SP, SAVE_OFS + i * REG_SIZE);
2293    }
2294
2295    tcg_out_opc_addi_d(s, TCG_REG_SP, TCG_REG_SP, FRAME_SIZE);
2296    tcg_out_opc_jirl(s, TCG_REG_ZERO, TCG_REG_RA, 0);
2297}
2298
2299static void tcg_target_init(TCGContext *s)
2300{
2301    unsigned long hwcap = qemu_getauxval(AT_HWCAP);
2302
2303    /* Server and desktop class cpus have UAL; embedded cpus do not. */
2304    if (!(hwcap & HWCAP_LOONGARCH_UAL)) {
2305        error_report("TCG: unaligned access support required; exiting");
2306        exit(EXIT_FAILURE);
2307    }
2308
2309    if (hwcap & HWCAP_LOONGARCH_LSX) {
2310        use_lsx_instructions = 1;
2311    }
2312
2313    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
2314    tcg_target_available_regs[TCG_TYPE_I64] = ALL_GENERAL_REGS;
2315
2316    tcg_target_call_clobber_regs = ALL_GENERAL_REGS;
2317    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S0);
2318    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S1);
2319    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S2);
2320    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S3);
2321    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S4);
2322    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S5);
2323    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S6);
2324    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S7);
2325    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S8);
2326    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S9);
2327
2328    if (use_lsx_instructions) {
2329        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
2330        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V24);
2331        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V25);
2332        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V26);
2333        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V27);
2334        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V28);
2335        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V29);
2336        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V30);
2337        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V31);
2338    }
2339
2340    s->reserved_regs = 0;
2341    tcg_regset_set_reg(s->reserved_regs, TCG_REG_ZERO);
2342    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP0);
2343    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP1);
2344    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP2);
2345    tcg_regset_set_reg(s->reserved_regs, TCG_REG_SP);
2346    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TP);
2347    tcg_regset_set_reg(s->reserved_regs, TCG_REG_RESERVED);
2348    tcg_regset_set_reg(s->reserved_regs, TCG_VEC_TMP0);
2349}
2350
2351typedef struct {
2352    DebugFrameHeader h;
2353    uint8_t fde_def_cfa[4];
2354    uint8_t fde_reg_ofs[ARRAY_SIZE(tcg_target_callee_save_regs) * 2];
2355} DebugFrame;
2356
2357#define ELF_HOST_MACHINE EM_LOONGARCH
2358
2359static const DebugFrame debug_frame = {
2360    .h.cie.len = sizeof(DebugFrameCIE) - 4, /* length after .len member */
2361    .h.cie.id = -1,
2362    .h.cie.version = 1,
2363    .h.cie.code_align = 1,
2364    .h.cie.data_align = -(TCG_TARGET_REG_BITS / 8) & 0x7f, /* sleb128 */
2365    .h.cie.return_column = TCG_REG_RA,
2366
2367    /* Total FDE size does not include the "len" member.  */
2368    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
2369
2370    .fde_def_cfa = {
2371        12, TCG_REG_SP,                 /* DW_CFA_def_cfa sp, ...  */
2372        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
2373        (FRAME_SIZE >> 7)
2374    },
2375    .fde_reg_ofs = {
2376        0x80 + 23, 11,                  /* DW_CFA_offset, s0, -88 */
2377        0x80 + 24, 10,                  /* DW_CFA_offset, s1, -80 */
2378        0x80 + 25, 9,                   /* DW_CFA_offset, s2, -72 */
2379        0x80 + 26, 8,                   /* DW_CFA_offset, s3, -64 */
2380        0x80 + 27, 7,                   /* DW_CFA_offset, s4, -56 */
2381        0x80 + 28, 6,                   /* DW_CFA_offset, s5, -48 */
2382        0x80 + 29, 5,                   /* DW_CFA_offset, s6, -40 */
2383        0x80 + 30, 4,                   /* DW_CFA_offset, s7, -32 */
2384        0x80 + 31, 3,                   /* DW_CFA_offset, s8, -24 */
2385        0x80 + 22, 2,                   /* DW_CFA_offset, s9, -16 */
2386        0x80 + 1 , 1,                   /* DW_CFA_offset, ra, -8 */
2387    }
2388};
2389
2390void tcg_register_jit(const void *buf, size_t buf_size)
2391{
2392    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
2393}
2394