xref: /openbmc/qemu/tcg/loongarch64/tcg-target.c.inc (revision 6aba25ebb9eb6e1e86398294694aa0ab1f12076f)
1/*
2 * Tiny Code Generator for QEMU
3 *
4 * Copyright (c) 2021 WANG Xuerui <git@xen0n.name>
5 *
6 * Based on tcg/riscv/tcg-target.c.inc
7 *
8 * Copyright (c) 2018 SiFive, Inc
9 * Copyright (c) 2008-2009 Arnaud Patard <arnaud.patard@rtp-net.org>
10 * Copyright (c) 2009 Aurelien Jarno <aurelien@aurel32.net>
11 * Copyright (c) 2008 Fabrice Bellard
12 *
13 * Permission is hereby granted, free of charge, to any person obtaining a copy
14 * of this software and associated documentation files (the "Software"), to deal
15 * in the Software without restriction, including without limitation the rights
16 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
17 * copies of the Software, and to permit persons to whom the Software is
18 * furnished to do so, subject to the following conditions:
19 *
20 * The above copyright notice and this permission notice shall be included in
21 * all copies or substantial portions of the Software.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
24 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
25 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
26 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
27 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
28 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
29 * THE SOFTWARE.
30 */
31
32#include <asm/hwcap.h>
33
34/* used for function call generation */
35#define TCG_REG_CALL_STACK              TCG_REG_SP
36#define TCG_TARGET_STACK_ALIGN          16
37#define TCG_TARGET_CALL_STACK_OFFSET    0
38#define TCG_TARGET_CALL_ARG_I32         TCG_CALL_ARG_NORMAL
39#define TCG_TARGET_CALL_ARG_I64         TCG_CALL_ARG_NORMAL
40#define TCG_TARGET_CALL_ARG_I128        TCG_CALL_ARG_NORMAL
41#define TCG_TARGET_CALL_RET_I128        TCG_CALL_RET_NORMAL
42
43#ifdef CONFIG_DEBUG_TCG
44static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
45    "zero",
46    "ra",
47    "tp",
48    "sp",
49    "a0",
50    "a1",
51    "a2",
52    "a3",
53    "a4",
54    "a5",
55    "a6",
56    "a7",
57    "t0",
58    "t1",
59    "t2",
60    "t3",
61    "t4",
62    "t5",
63    "t6",
64    "t7",
65    "t8",
66    "r21", /* reserved in the LP64* ABI, hence no ABI name */
67    "s9",
68    "s0",
69    "s1",
70    "s2",
71    "s3",
72    "s4",
73    "s5",
74    "s6",
75    "s7",
76    "s8",
77    "vr0",
78    "vr1",
79    "vr2",
80    "vr3",
81    "vr4",
82    "vr5",
83    "vr6",
84    "vr7",
85    "vr8",
86    "vr9",
87    "vr10",
88    "vr11",
89    "vr12",
90    "vr13",
91    "vr14",
92    "vr15",
93    "vr16",
94    "vr17",
95    "vr18",
96    "vr19",
97    "vr20",
98    "vr21",
99    "vr22",
100    "vr23",
101    "vr24",
102    "vr25",
103    "vr26",
104    "vr27",
105    "vr28",
106    "vr29",
107    "vr30",
108    "vr31",
109};
110#endif
111
112static const int tcg_target_reg_alloc_order[] = {
113    /* Registers preserved across calls */
114    /* TCG_REG_S0 reserved for TCG_AREG0 */
115    TCG_REG_S1,
116    TCG_REG_S2,
117    TCG_REG_S3,
118    TCG_REG_S4,
119    TCG_REG_S5,
120    TCG_REG_S6,
121    TCG_REG_S7,
122    TCG_REG_S8,
123    TCG_REG_S9,
124
125    /* Registers (potentially) clobbered across calls */
126    TCG_REG_T0,
127    TCG_REG_T1,
128    TCG_REG_T2,
129    TCG_REG_T3,
130    TCG_REG_T4,
131    TCG_REG_T5,
132    TCG_REG_T6,
133    TCG_REG_T7,
134    TCG_REG_T8,
135
136    /* Argument registers, opposite order of allocation.  */
137    TCG_REG_A7,
138    TCG_REG_A6,
139    TCG_REG_A5,
140    TCG_REG_A4,
141    TCG_REG_A3,
142    TCG_REG_A2,
143    TCG_REG_A1,
144    TCG_REG_A0,
145
146    /* Vector registers */
147    TCG_REG_V0, TCG_REG_V1, TCG_REG_V2, TCG_REG_V3,
148    TCG_REG_V4, TCG_REG_V5, TCG_REG_V6, TCG_REG_V7,
149    TCG_REG_V8, TCG_REG_V9, TCG_REG_V10, TCG_REG_V11,
150    TCG_REG_V12, TCG_REG_V13, TCG_REG_V14, TCG_REG_V15,
151    TCG_REG_V16, TCG_REG_V17, TCG_REG_V18, TCG_REG_V19,
152    TCG_REG_V20, TCG_REG_V21, TCG_REG_V22, TCG_REG_V23,
153    /* V24 - V31 are caller-saved, and skipped.  */
154};
155
156static const int tcg_target_call_iarg_regs[] = {
157    TCG_REG_A0,
158    TCG_REG_A1,
159    TCG_REG_A2,
160    TCG_REG_A3,
161    TCG_REG_A4,
162    TCG_REG_A5,
163    TCG_REG_A6,
164    TCG_REG_A7,
165};
166
167static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
168{
169    tcg_debug_assert(kind == TCG_CALL_RET_NORMAL);
170    tcg_debug_assert(slot >= 0 && slot <= 1);
171    return TCG_REG_A0 + slot;
172}
173
174#define TCG_GUEST_BASE_REG TCG_REG_S1
175
176#define TCG_CT_CONST_S12   0x100
177#define TCG_CT_CONST_S32   0x200
178#define TCG_CT_CONST_U12   0x400
179#define TCG_CT_CONST_WSZ   0x800
180#define TCG_CT_CONST_VCMP  0x1000
181#define TCG_CT_CONST_VADD  0x2000
182
183#define ALL_GENERAL_REGS   MAKE_64BIT_MASK(0, 32)
184#define ALL_VECTOR_REGS    MAKE_64BIT_MASK(32, 32)
185
186static inline tcg_target_long sextreg(tcg_target_long val, int pos, int len)
187{
188    return sextract64(val, pos, len);
189}
190
191/* test if a constant matches the constraint */
192static bool tcg_target_const_match(int64_t val, int ct,
193                                   TCGType type, TCGCond cond, int vece)
194{
195    if (ct & TCG_CT_CONST) {
196        return true;
197    }
198    if ((ct & TCG_CT_CONST_S12) && val == sextreg(val, 0, 12)) {
199        return true;
200    }
201    if ((ct & TCG_CT_CONST_S32) && val == (int32_t)val) {
202        return true;
203    }
204    if ((ct & TCG_CT_CONST_U12) && val >= 0 && val <= 0xfff) {
205        return true;
206    }
207    if ((ct & TCG_CT_CONST_WSZ) && val == (type == TCG_TYPE_I32 ? 32 : 64)) {
208        return true;
209    }
210    if (ct & (TCG_CT_CONST_VCMP | TCG_CT_CONST_VADD)) {
211        int64_t vec_val = sextract64(val, 0, 8 << vece);
212        if (ct & TCG_CT_CONST_VCMP) {
213            switch (cond) {
214            case TCG_COND_EQ:
215            case TCG_COND_LE:
216            case TCG_COND_LT:
217                return -0x10 <= vec_val && vec_val <= 0x0f;
218            case TCG_COND_LEU:
219            case TCG_COND_LTU:
220                return 0x00 <= vec_val && vec_val <= 0x1f;
221            default:
222                return false;
223            }
224        }
225        if ((ct & TCG_CT_CONST_VADD) && -0x1f <= vec_val && vec_val <= 0x1f) {
226            return true;
227        }
228    }
229    return false;
230}
231
232/*
233 * Relocations
234 */
235
236/*
237 * Relocation records defined in LoongArch ELF psABI v1.00 is way too
238 * complicated; a whopping stack machine is needed to stuff the fields, at
239 * the very least one SOP_PUSH and one SOP_POP (of the correct format) are
240 * needed.
241 *
242 * Hence, define our own simpler relocation types. Numbers are chosen as to
243 * not collide with potential future additions to the true ELF relocation
244 * type enum.
245 */
246
247/* Field Sk16, shifted right by 2; suitable for conditional jumps */
248#define R_LOONGARCH_BR_SK16     256
249/* Field Sd10k16, shifted right by 2; suitable for B and BL */
250#define R_LOONGARCH_BR_SD10K16  257
251
252static bool reloc_br_sk16(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
253{
254    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
255    intptr_t offset = (intptr_t)target - (intptr_t)src_rx;
256
257    tcg_debug_assert((offset & 3) == 0);
258    offset >>= 2;
259    if (offset == sextreg(offset, 0, 16)) {
260        *src_rw = deposit64(*src_rw, 10, 16, offset);
261        return true;
262    }
263
264    return false;
265}
266
267static bool reloc_br_sd10k16(tcg_insn_unit *src_rw,
268                             const tcg_insn_unit *target)
269{
270    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
271    intptr_t offset = (intptr_t)target - (intptr_t)src_rx;
272
273    tcg_debug_assert((offset & 3) == 0);
274    offset >>= 2;
275    if (offset == sextreg(offset, 0, 26)) {
276        *src_rw = deposit64(*src_rw, 0, 10, offset >> 16); /* slot d10 */
277        *src_rw = deposit64(*src_rw, 10, 16, offset); /* slot k16 */
278        return true;
279    }
280
281    return false;
282}
283
284static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
285                        intptr_t value, intptr_t addend)
286{
287    tcg_debug_assert(addend == 0);
288    switch (type) {
289    case R_LOONGARCH_BR_SK16:
290        return reloc_br_sk16(code_ptr, (tcg_insn_unit *)value);
291    case R_LOONGARCH_BR_SD10K16:
292        return reloc_br_sd10k16(code_ptr, (tcg_insn_unit *)value);
293    default:
294        g_assert_not_reached();
295    }
296}
297
298#include "tcg-insn-defs.c.inc"
299
300/*
301 * TCG intrinsics
302 */
303
304static void tcg_out_mb(TCGContext *s, TCGArg a0)
305{
306    /* Baseline LoongArch only has the full barrier, unfortunately.  */
307    tcg_out_opc_dbar(s, 0);
308}
309
310static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
311{
312    if (ret == arg) {
313        return true;
314    }
315    switch (type) {
316    case TCG_TYPE_I32:
317    case TCG_TYPE_I64:
318        if (ret < TCG_REG_V0) {
319            if (arg < TCG_REG_V0) {
320                /*
321                 * Conventional register-register move used in LoongArch is
322                 * `or dst, src, zero`.
323                 */
324                tcg_out_opc_or(s, ret, arg, TCG_REG_ZERO);
325            } else {
326                tcg_out_opc_movfr2gr_d(s, ret, arg);
327            }
328        } else {
329            if (arg < TCG_REG_V0) {
330                tcg_out_opc_movgr2fr_d(s, ret, arg);
331            } else {
332                tcg_out_opc_fmov_d(s, ret, arg);
333            }
334        }
335        break;
336    case TCG_TYPE_V64:
337    case TCG_TYPE_V128:
338        tcg_out_opc_vori_b(s, ret, arg, 0);
339        break;
340    case TCG_TYPE_V256:
341        tcg_out_opc_xvori_b(s, ret, arg, 0);
342        break;
343    default:
344        g_assert_not_reached();
345    }
346    return true;
347}
348
349/* Loads a 32-bit immediate into rd, sign-extended.  */
350static void tcg_out_movi_i32(TCGContext *s, TCGReg rd, int32_t val)
351{
352    tcg_target_long lo = sextreg(val, 0, 12);
353    tcg_target_long hi12 = sextreg(val, 12, 20);
354
355    /* Single-instruction cases.  */
356    if (hi12 == 0) {
357        /* val fits in uimm12: ori rd, zero, val */
358        tcg_out_opc_ori(s, rd, TCG_REG_ZERO, val);
359        return;
360    }
361    if (hi12 == sextreg(lo, 12, 20)) {
362        /* val fits in simm12: addi.w rd, zero, val */
363        tcg_out_opc_addi_w(s, rd, TCG_REG_ZERO, val);
364        return;
365    }
366
367    /* High bits must be set; load with lu12i.w + optional ori.  */
368    tcg_out_opc_lu12i_w(s, rd, hi12);
369    if (lo != 0) {
370        tcg_out_opc_ori(s, rd, rd, lo & 0xfff);
371    }
372}
373
374static void tcg_out_movi(TCGContext *s, TCGType type, TCGReg rd,
375                         tcg_target_long val)
376{
377    /*
378     * LoongArch conventionally loads 64-bit immediates in at most 4 steps,
379     * with dedicated instructions for filling the respective bitfields
380     * below:
381     *
382     *        6                   5                   4               3
383     *  3 2 1 0 9 8 7 6 5 4 3 2 1 0 9 8 7 6 5 4 3 2 1 0 9 8 7 6 5 4 3 2
384     * +-----------------------+---------------------------------------+...
385     * |          hi52         |                  hi32                 |
386     * +-----------------------+---------------------------------------+...
387     *       3                   2                   1
388     *     1 0 9 8 7 6 5 4 3 2 1 0 9 8 7 6 5 4 3 2 1 0 9 8 7 6 5 4 3 2 1 0
389     * ...+-------------------------------------+-------------------------+
390     *    |                 hi12                |            lo           |
391     * ...+-------------------------------------+-------------------------+
392     *
393     * Check if val belong to one of the several fast cases, before falling
394     * back to the slow path.
395     */
396
397    intptr_t src_rx, pc_offset;
398    tcg_target_long hi12, hi32, hi52;
399
400    /* Value fits in signed i32.  */
401    if (type == TCG_TYPE_I32 || val == (int32_t)val) {
402        tcg_out_movi_i32(s, rd, val);
403        return;
404    }
405
406    /* PC-relative cases.  */
407    src_rx = (intptr_t)tcg_splitwx_to_rx(s->code_ptr);
408    if ((val & 3) == 0) {
409        pc_offset = val - src_rx;
410        if (pc_offset == sextreg(pc_offset, 0, 22)) {
411            /* Single pcaddu2i.  */
412            tcg_out_opc_pcaddu2i(s, rd, pc_offset >> 2);
413            return;
414        }
415    }
416
417    pc_offset = (val >> 12) - (src_rx >> 12);
418    if (pc_offset == sextreg(pc_offset, 0, 20)) {
419        /* Load with pcalau12i + ori.  */
420        tcg_target_long val_lo = val & 0xfff;
421        tcg_out_opc_pcalau12i(s, rd, pc_offset);
422        if (val_lo != 0) {
423            tcg_out_opc_ori(s, rd, rd, val_lo);
424        }
425        return;
426    }
427
428    hi12 = sextreg(val, 12, 20);
429    hi32 = sextreg(val, 32, 20);
430    hi52 = sextreg(val, 52, 12);
431
432    /* Single cu52i.d case.  */
433    if ((hi52 != 0) && (ctz64(val) >= 52)) {
434        tcg_out_opc_cu52i_d(s, rd, TCG_REG_ZERO, hi52);
435        return;
436    }
437
438    /* Slow path.  Initialize the low 32 bits, then concat high bits.  */
439    tcg_out_movi_i32(s, rd, val);
440
441    /* Load hi32 and hi52 explicitly when they are unexpected values. */
442    if (hi32 != sextreg(hi12, 20, 20)) {
443        tcg_out_opc_cu32i_d(s, rd, hi32);
444    }
445
446    if (hi52 != sextreg(hi32, 20, 12)) {
447        tcg_out_opc_cu52i_d(s, rd, rd, hi52);
448    }
449}
450
451static void tcg_out_addi(TCGContext *s, TCGType type, TCGReg rd,
452                         TCGReg rs, tcg_target_long imm)
453{
454    tcg_target_long lo12 = sextreg(imm, 0, 12);
455    tcg_target_long hi16 = sextreg(imm - lo12, 16, 16);
456
457    /*
458     * Note that there's a hole in between hi16 and lo12:
459     *
460     *       3                   2                   1                   0
461     *     1 0 9 8 7 6 5 4 3 2 1 0 9 8 7 6 5 4 3 2 1 0 9 8 7 6 5 4 3 2 1 0
462     * ...+-------------------------------+-------+-----------------------+
463     *    |             hi16              |       |          lo12         |
464     * ...+-------------------------------+-------+-----------------------+
465     *
466     * For bits within that hole, it's more efficient to use LU12I and ADD.
467     */
468    if (imm == (hi16 << 16) + lo12) {
469        if (hi16) {
470            tcg_out_opc_addu16i_d(s, rd, rs, hi16);
471            rs = rd;
472        }
473        if (type == TCG_TYPE_I32) {
474            tcg_out_opc_addi_w(s, rd, rs, lo12);
475        } else if (lo12) {
476            tcg_out_opc_addi_d(s, rd, rs, lo12);
477        } else {
478            tcg_out_mov(s, type, rd, rs);
479        }
480    } else {
481        tcg_out_movi(s, type, TCG_REG_TMP0, imm);
482        if (type == TCG_TYPE_I32) {
483            tcg_out_opc_add_w(s, rd, rs, TCG_REG_TMP0);
484        } else {
485            tcg_out_opc_add_d(s, rd, rs, TCG_REG_TMP0);
486        }
487    }
488}
489
490static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
491{
492    return false;
493}
494
495static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
496                             tcg_target_long imm)
497{
498    /* This function is only used for passing structs by reference. */
499    g_assert_not_reached();
500}
501
502static void tcg_out_ext8u(TCGContext *s, TCGReg ret, TCGReg arg)
503{
504    tcg_out_opc_andi(s, ret, arg, 0xff);
505}
506
507static void tcg_out_ext16u(TCGContext *s, TCGReg ret, TCGReg arg)
508{
509    tcg_out_opc_bstrpick_w(s, ret, arg, 0, 15);
510}
511
512static void tcg_out_ext32u(TCGContext *s, TCGReg ret, TCGReg arg)
513{
514    tcg_out_opc_bstrpick_d(s, ret, arg, 0, 31);
515}
516
517static void tcg_out_ext8s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
518{
519    tcg_out_opc_sext_b(s, ret, arg);
520}
521
522static void tcg_out_ext16s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
523{
524    tcg_out_opc_sext_h(s, ret, arg);
525}
526
527static void tcg_out_ext32s(TCGContext *s, TCGReg ret, TCGReg arg)
528{
529    tcg_out_opc_addi_w(s, ret, arg, 0);
530}
531
532static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg)
533{
534    if (ret != arg) {
535        tcg_out_ext32s(s, ret, arg);
536    }
537}
538
539static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg)
540{
541    tcg_out_ext32u(s, ret, arg);
542}
543
544static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg ret, TCGReg arg)
545{
546    tcg_out_ext32s(s, ret, arg);
547}
548
549static void tcg_out_clzctz(TCGContext *s, LoongArchInsn opc,
550                           TCGReg a0, TCGReg a1, TCGReg a2,
551                           bool c2, bool is_32bit)
552{
553    if (c2) {
554        /*
555         * Fast path: semantics already satisfied due to constraint and
556         * insn behavior, single instruction is enough.
557         */
558        tcg_debug_assert(a2 == (is_32bit ? 32 : 64));
559        /* all clz/ctz insns belong to DJ-format */
560        tcg_out32(s, encode_dj_insn(opc, a0, a1));
561        return;
562    }
563
564    tcg_out32(s, encode_dj_insn(opc, TCG_REG_TMP0, a1));
565    /* a0 = a1 ? REG_TMP0 : a2 */
566    tcg_out_opc_maskeqz(s, TCG_REG_TMP0, TCG_REG_TMP0, a1);
567    tcg_out_opc_masknez(s, a0, a2, a1);
568    tcg_out_opc_or(s, a0, TCG_REG_TMP0, a0);
569}
570
571#define SETCOND_INV    TCG_TARGET_NB_REGS
572#define SETCOND_NEZ    (SETCOND_INV << 1)
573#define SETCOND_FLAGS  (SETCOND_INV | SETCOND_NEZ)
574
575static int tcg_out_setcond_int(TCGContext *s, TCGCond cond, TCGReg ret,
576                               TCGReg arg1, tcg_target_long arg2, bool c2)
577{
578    int flags = 0;
579
580    switch (cond) {
581    case TCG_COND_EQ:    /* -> NE  */
582    case TCG_COND_GE:    /* -> LT  */
583    case TCG_COND_GEU:   /* -> LTU */
584    case TCG_COND_GT:    /* -> LE  */
585    case TCG_COND_GTU:   /* -> LEU */
586        cond = tcg_invert_cond(cond);
587        flags ^= SETCOND_INV;
588        break;
589    default:
590        break;
591    }
592
593    switch (cond) {
594    case TCG_COND_LE:
595    case TCG_COND_LEU:
596        /*
597         * If we have a constant input, the most efficient way to implement
598         * LE is by adding 1 and using LT.  Watch out for wrap around for LEU.
599         * We don't need to care for this for LE because the constant input
600         * is still constrained to int32_t, and INT32_MAX+1 is representable
601         * in the 64-bit temporary register.
602         */
603        if (c2) {
604            if (cond == TCG_COND_LEU) {
605                /* unsigned <= -1 is true */
606                if (arg2 == -1) {
607                    tcg_out_movi(s, TCG_TYPE_REG, ret, !(flags & SETCOND_INV));
608                    return ret;
609                }
610                cond = TCG_COND_LTU;
611            } else {
612                cond = TCG_COND_LT;
613            }
614            arg2 += 1;
615        } else {
616            TCGReg tmp = arg2;
617            arg2 = arg1;
618            arg1 = tmp;
619            cond = tcg_swap_cond(cond);    /* LE -> GE */
620            cond = tcg_invert_cond(cond);  /* GE -> LT */
621            flags ^= SETCOND_INV;
622        }
623        break;
624    default:
625        break;
626    }
627
628    switch (cond) {
629    case TCG_COND_NE:
630        flags |= SETCOND_NEZ;
631        if (!c2) {
632            tcg_out_opc_xor(s, ret, arg1, arg2);
633        } else if (arg2 == 0) {
634            ret = arg1;
635        } else if (arg2 >= 0 && arg2 <= 0xfff) {
636            tcg_out_opc_xori(s, ret, arg1, arg2);
637        } else {
638            tcg_out_addi(s, TCG_TYPE_REG, ret, arg1, -arg2);
639        }
640        break;
641
642    case TCG_COND_LT:
643    case TCG_COND_LTU:
644        if (c2) {
645            if (arg2 >= -0x800 && arg2 <= 0x7ff) {
646                if (cond == TCG_COND_LT) {
647                    tcg_out_opc_slti(s, ret, arg1, arg2);
648                } else {
649                    tcg_out_opc_sltui(s, ret, arg1, arg2);
650                }
651                break;
652            }
653            tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_TMP0, arg2);
654            arg2 = TCG_REG_TMP0;
655        }
656        if (cond == TCG_COND_LT) {
657            tcg_out_opc_slt(s, ret, arg1, arg2);
658        } else {
659            tcg_out_opc_sltu(s, ret, arg1, arg2);
660        }
661        break;
662
663    default:
664        g_assert_not_reached();
665    }
666
667    return ret | flags;
668}
669
670static void tcg_out_setcond(TCGContext *s, TCGCond cond, TCGReg ret,
671                            TCGReg arg1, tcg_target_long arg2, bool c2)
672{
673    int tmpflags = tcg_out_setcond_int(s, cond, ret, arg1, arg2, c2);
674
675    if (tmpflags != ret) {
676        TCGReg tmp = tmpflags & ~SETCOND_FLAGS;
677
678        switch (tmpflags & SETCOND_FLAGS) {
679        case SETCOND_INV:
680            /* Intermediate result is boolean: simply invert. */
681            tcg_out_opc_xori(s, ret, tmp, 1);
682            break;
683        case SETCOND_NEZ:
684            /* Intermediate result is zero/non-zero: test != 0. */
685            tcg_out_opc_sltu(s, ret, TCG_REG_ZERO, tmp);
686            break;
687        case SETCOND_NEZ | SETCOND_INV:
688            /* Intermediate result is zero/non-zero: test == 0. */
689            tcg_out_opc_sltui(s, ret, tmp, 1);
690            break;
691        default:
692            g_assert_not_reached();
693        }
694    }
695}
696
697static void tcg_out_movcond(TCGContext *s, TCGCond cond, TCGReg ret,
698                            TCGReg c1, tcg_target_long c2, bool const2,
699                            TCGReg v1, TCGReg v2)
700{
701    int tmpflags = tcg_out_setcond_int(s, cond, TCG_REG_TMP0, c1, c2, const2);
702    TCGReg t;
703
704    /* Standardize the test below to t != 0. */
705    if (tmpflags & SETCOND_INV) {
706        t = v1, v1 = v2, v2 = t;
707    }
708
709    t = tmpflags & ~SETCOND_FLAGS;
710    if (v1 == TCG_REG_ZERO) {
711        tcg_out_opc_masknez(s, ret, v2, t);
712    } else if (v2 == TCG_REG_ZERO) {
713        tcg_out_opc_maskeqz(s, ret, v1, t);
714    } else {
715        tcg_out_opc_masknez(s, TCG_REG_TMP2, v2, t); /* t ? 0 : v2 */
716        tcg_out_opc_maskeqz(s, TCG_REG_TMP1, v1, t); /* t ? v1 : 0 */
717        tcg_out_opc_or(s, ret, TCG_REG_TMP1, TCG_REG_TMP2);
718    }
719}
720
721/*
722 * Branch helpers
723 */
724
725static const struct {
726    LoongArchInsn op;
727    bool swap;
728} tcg_brcond_to_loongarch[] = {
729    [TCG_COND_EQ] =  { OPC_BEQ,  false },
730    [TCG_COND_NE] =  { OPC_BNE,  false },
731    [TCG_COND_LT] =  { OPC_BGT,  true  },
732    [TCG_COND_GE] =  { OPC_BLE,  true  },
733    [TCG_COND_LE] =  { OPC_BLE,  false },
734    [TCG_COND_GT] =  { OPC_BGT,  false },
735    [TCG_COND_LTU] = { OPC_BGTU, true  },
736    [TCG_COND_GEU] = { OPC_BLEU, true  },
737    [TCG_COND_LEU] = { OPC_BLEU, false },
738    [TCG_COND_GTU] = { OPC_BGTU, false }
739};
740
741static void tcg_out_brcond(TCGContext *s, TCGCond cond, TCGReg arg1,
742                           TCGReg arg2, TCGLabel *l)
743{
744    LoongArchInsn op = tcg_brcond_to_loongarch[cond].op;
745
746    tcg_debug_assert(op != 0);
747
748    if (tcg_brcond_to_loongarch[cond].swap) {
749        TCGReg t = arg1;
750        arg1 = arg2;
751        arg2 = t;
752    }
753
754    /* all conditional branch insns belong to DJSk16-format */
755    tcg_out_reloc(s, s->code_ptr, R_LOONGARCH_BR_SK16, l, 0);
756    tcg_out32(s, encode_djsk16_insn(op, arg1, arg2, 0));
757}
758
759static void tcg_out_call_int(TCGContext *s, const tcg_insn_unit *arg, bool tail)
760{
761    TCGReg link = tail ? TCG_REG_ZERO : TCG_REG_RA;
762    ptrdiff_t offset = tcg_pcrel_diff(s, arg);
763
764    tcg_debug_assert((offset & 3) == 0);
765    if (offset == sextreg(offset, 0, 28)) {
766        /* short jump: +/- 256MiB */
767        if (tail) {
768            tcg_out_opc_b(s, offset >> 2);
769        } else {
770            tcg_out_opc_bl(s, offset >> 2);
771        }
772    } else if (offset == sextreg(offset, 0, 38)) {
773        /* long jump: +/- 256GiB */
774        tcg_target_long lo = sextreg(offset, 0, 18);
775        tcg_target_long hi = offset - lo;
776        tcg_out_opc_pcaddu18i(s, TCG_REG_TMP0, hi >> 18);
777        tcg_out_opc_jirl(s, link, TCG_REG_TMP0, lo >> 2);
778    } else {
779        /* far jump: 64-bit */
780        tcg_target_long lo = sextreg((tcg_target_long)arg, 0, 18);
781        tcg_target_long hi = (tcg_target_long)arg - lo;
782        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, hi);
783        tcg_out_opc_jirl(s, link, TCG_REG_TMP0, lo >> 2);
784    }
785}
786
787static void tcg_out_call(TCGContext *s, const tcg_insn_unit *arg,
788                         const TCGHelperInfo *info)
789{
790    tcg_out_call_int(s, arg, false);
791}
792
793/*
794 * Load/store helpers
795 */
796
797static void tcg_out_ldst(TCGContext *s, LoongArchInsn opc, TCGReg data,
798                         TCGReg addr, intptr_t offset)
799{
800    intptr_t imm12 = sextreg(offset, 0, 12);
801
802    if (offset != imm12) {
803        intptr_t diff = tcg_pcrel_diff(s, (void *)offset);
804
805        if (addr == TCG_REG_ZERO && diff == (int32_t)diff) {
806            imm12 = sextreg(diff, 0, 12);
807            tcg_out_opc_pcaddu12i(s, TCG_REG_TMP2, (diff - imm12) >> 12);
808        } else {
809            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP2, offset - imm12);
810            if (addr != TCG_REG_ZERO) {
811                tcg_out_opc_add_d(s, TCG_REG_TMP2, TCG_REG_TMP2, addr);
812            }
813        }
814        addr = TCG_REG_TMP2;
815    }
816
817    switch (opc) {
818    case OPC_LD_B:
819    case OPC_LD_BU:
820    case OPC_LD_H:
821    case OPC_LD_HU:
822    case OPC_LD_W:
823    case OPC_LD_WU:
824    case OPC_LD_D:
825    case OPC_ST_B:
826    case OPC_ST_H:
827    case OPC_ST_W:
828    case OPC_ST_D:
829        tcg_out32(s, encode_djsk12_insn(opc, data, addr, imm12));
830        break;
831    case OPC_FLD_S:
832    case OPC_FLD_D:
833    case OPC_FST_S:
834    case OPC_FST_D:
835        tcg_out32(s, encode_fdjsk12_insn(opc, data, addr, imm12));
836        break;
837    default:
838        g_assert_not_reached();
839    }
840}
841
842static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg dest,
843                       TCGReg base, intptr_t offset)
844{
845    switch (type) {
846    case TCG_TYPE_I32:
847        if (dest < TCG_REG_V0) {
848            tcg_out_ldst(s, OPC_LD_W, dest, base, offset);
849        } else {
850            tcg_out_ldst(s, OPC_FLD_S, dest, base, offset);
851        }
852        break;
853    case TCG_TYPE_I64:
854    case TCG_TYPE_V64:
855        if (dest < TCG_REG_V0) {
856            tcg_out_ldst(s, OPC_LD_D, dest, base, offset);
857        } else {
858            tcg_out_ldst(s, OPC_FLD_D, dest, base, offset);
859        }
860        break;
861    case TCG_TYPE_V128:
862        if (-0x800 <= offset && offset <= 0x7ff) {
863            tcg_out_opc_vld(s, dest, base, offset);
864        } else {
865            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, offset);
866            tcg_out_opc_vldx(s, dest, base, TCG_REG_TMP0);
867        }
868        break;
869    case TCG_TYPE_V256:
870        if (-0x800 <= offset && offset <= 0x7ff) {
871            tcg_out_opc_xvld(s, dest, base, offset);
872        } else {
873            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, offset);
874            tcg_out_opc_xvldx(s, dest, base, TCG_REG_TMP0);
875        }
876        break;
877    default:
878        g_assert_not_reached();
879    }
880}
881
882static void tcg_out_st(TCGContext *s, TCGType type, TCGReg src,
883                       TCGReg base, intptr_t offset)
884{
885    switch (type) {
886    case TCG_TYPE_I32:
887        if (src < TCG_REG_V0) {
888            tcg_out_ldst(s, OPC_ST_W, src, base, offset);
889        } else {
890            tcg_out_ldst(s, OPC_FST_S, src, base, offset);
891        }
892        break;
893    case TCG_TYPE_I64:
894    case TCG_TYPE_V64:
895        if (src < TCG_REG_V0) {
896            tcg_out_ldst(s, OPC_ST_D, src, base, offset);
897        } else {
898            tcg_out_ldst(s, OPC_FST_D, src, base, offset);
899        }
900        break;
901    case TCG_TYPE_V128:
902        if (-0x800 <= offset && offset <= 0x7ff) {
903            tcg_out_opc_vst(s, src, base, offset);
904        } else {
905            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, offset);
906            tcg_out_opc_vstx(s, src, base, TCG_REG_TMP0);
907        }
908        break;
909    case TCG_TYPE_V256:
910        if (-0x800 <= offset && offset <= 0x7ff) {
911            tcg_out_opc_xvst(s, src, base, offset);
912        } else {
913            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, offset);
914            tcg_out_opc_xvstx(s, src, base, TCG_REG_TMP0);
915        }
916        break;
917    default:
918        g_assert_not_reached();
919    }
920}
921
922static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
923                        TCGReg base, intptr_t ofs)
924{
925    if (val == 0) {
926        tcg_out_st(s, type, TCG_REG_ZERO, base, ofs);
927        return true;
928    }
929    return false;
930}
931
932/*
933 * Load/store helpers for SoftMMU, and qemu_ld/st implementations
934 */
935
936static bool tcg_out_goto(TCGContext *s, const tcg_insn_unit *target)
937{
938    tcg_out_opc_b(s, 0);
939    return reloc_br_sd10k16(s->code_ptr - 1, target);
940}
941
942static const TCGLdstHelperParam ldst_helper_param = {
943    .ntmp = 1, .tmp = { TCG_REG_TMP0 }
944};
945
946static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
947{
948    MemOp opc = get_memop(l->oi);
949
950    /* resolve label address */
951    if (!reloc_br_sk16(l->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
952        return false;
953    }
954
955    tcg_out_ld_helper_args(s, l, &ldst_helper_param);
956    tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SIZE], false);
957    tcg_out_ld_helper_ret(s, l, false, &ldst_helper_param);
958    return tcg_out_goto(s, l->raddr);
959}
960
961static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
962{
963    MemOp opc = get_memop(l->oi);
964
965    /* resolve label address */
966    if (!reloc_br_sk16(l->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
967        return false;
968    }
969
970    tcg_out_st_helper_args(s, l, &ldst_helper_param);
971    tcg_out_call_int(s, qemu_st_helpers[opc & MO_SIZE], false);
972    return tcg_out_goto(s, l->raddr);
973}
974
975typedef struct {
976    TCGReg base;
977    TCGReg index;
978    TCGAtomAlign aa;
979} HostAddress;
980
981bool tcg_target_has_memory_bswap(MemOp memop)
982{
983    return false;
984}
985
986/* We expect to use a 12-bit negative offset from ENV.  */
987#define MIN_TLB_MASK_TABLE_OFS  -(1 << 11)
988
989/*
990 * For system-mode, perform the TLB load and compare.
991 * For user-mode, perform any required alignment tests.
992 * In both cases, return a TCGLabelQemuLdst structure if the slow path
993 * is required and fill in @h with the host address for the fast path.
994 */
995static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
996                                           TCGReg addr_reg, MemOpIdx oi,
997                                           bool is_ld)
998{
999    TCGType addr_type = s->addr_type;
1000    TCGLabelQemuLdst *ldst = NULL;
1001    MemOp opc = get_memop(oi);
1002    MemOp a_bits;
1003
1004    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, false);
1005    a_bits = h->aa.align;
1006
1007    if (tcg_use_softmmu) {
1008        unsigned s_bits = opc & MO_SIZE;
1009        int mem_index = get_mmuidx(oi);
1010        int fast_ofs = tlb_mask_table_ofs(s, mem_index);
1011        int mask_ofs = fast_ofs + offsetof(CPUTLBDescFast, mask);
1012        int table_ofs = fast_ofs + offsetof(CPUTLBDescFast, table);
1013
1014        ldst = new_ldst_label(s);
1015        ldst->is_ld = is_ld;
1016        ldst->oi = oi;
1017        ldst->addr_reg = addr_reg;
1018
1019        tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP0, TCG_AREG0, mask_ofs);
1020        tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP1, TCG_AREG0, table_ofs);
1021
1022        tcg_out_opc_srli_d(s, TCG_REG_TMP2, addr_reg,
1023                           s->page_bits - CPU_TLB_ENTRY_BITS);
1024        tcg_out_opc_and(s, TCG_REG_TMP2, TCG_REG_TMP2, TCG_REG_TMP0);
1025        tcg_out_opc_add_d(s, TCG_REG_TMP2, TCG_REG_TMP2, TCG_REG_TMP1);
1026
1027        /* Load the tlb comparator and the addend.  */
1028        QEMU_BUILD_BUG_ON(HOST_BIG_ENDIAN);
1029        tcg_out_ld(s, addr_type, TCG_REG_TMP0, TCG_REG_TMP2,
1030                   is_ld ? offsetof(CPUTLBEntry, addr_read)
1031                         : offsetof(CPUTLBEntry, addr_write));
1032        tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP2, TCG_REG_TMP2,
1033                   offsetof(CPUTLBEntry, addend));
1034
1035        /*
1036         * For aligned accesses, we check the first byte and include the
1037         * alignment bits within the address.  For unaligned access, we
1038         * check that we don't cross pages using the address of the last
1039         * byte of the access.
1040         */
1041        if (a_bits < s_bits) {
1042            unsigned a_mask = (1u << a_bits) - 1;
1043            unsigned s_mask = (1u << s_bits) - 1;
1044            tcg_out_addi(s, addr_type, TCG_REG_TMP1, addr_reg, s_mask - a_mask);
1045        } else {
1046            tcg_out_mov(s, addr_type, TCG_REG_TMP1, addr_reg);
1047        }
1048        tcg_out_opc_bstrins_d(s, TCG_REG_TMP1, TCG_REG_ZERO,
1049                              a_bits, s->page_bits - 1);
1050
1051        /* Compare masked address with the TLB entry.  */
1052        ldst->label_ptr[0] = s->code_ptr;
1053        tcg_out_opc_bne(s, TCG_REG_TMP0, TCG_REG_TMP1, 0);
1054
1055        h->index = TCG_REG_TMP2;
1056    } else {
1057        if (a_bits) {
1058            ldst = new_ldst_label(s);
1059
1060            ldst->is_ld = is_ld;
1061            ldst->oi = oi;
1062            ldst->addr_reg = addr_reg;
1063
1064            /*
1065             * Without micro-architecture details, we don't know which of
1066             * bstrpick or andi is faster, so use bstrpick as it's not
1067             * constrained by imm field width. Not to say alignments >= 2^12
1068             * are going to happen any time soon.
1069             */
1070            tcg_out_opc_bstrpick_d(s, TCG_REG_TMP1, addr_reg, 0, a_bits - 1);
1071
1072            ldst->label_ptr[0] = s->code_ptr;
1073            tcg_out_opc_bne(s, TCG_REG_TMP1, TCG_REG_ZERO, 0);
1074        }
1075
1076        h->index = guest_base ? TCG_GUEST_BASE_REG : TCG_REG_ZERO;
1077    }
1078
1079    if (addr_type == TCG_TYPE_I32) {
1080        h->base = TCG_REG_TMP0;
1081        tcg_out_ext32u(s, h->base, addr_reg);
1082    } else {
1083        h->base = addr_reg;
1084    }
1085
1086    return ldst;
1087}
1088
1089static void tcg_out_qemu_ld_indexed(TCGContext *s, MemOp opc, TCGType type,
1090                                    TCGReg rd, HostAddress h)
1091{
1092    /* Byte swapping is left to middle-end expansion.  */
1093    tcg_debug_assert((opc & MO_BSWAP) == 0);
1094
1095    switch (opc & MO_SSIZE) {
1096    case MO_UB:
1097        tcg_out_opc_ldx_bu(s, rd, h.base, h.index);
1098        break;
1099    case MO_SB:
1100        tcg_out_opc_ldx_b(s, rd, h.base, h.index);
1101        break;
1102    case MO_UW:
1103        tcg_out_opc_ldx_hu(s, rd, h.base, h.index);
1104        break;
1105    case MO_SW:
1106        tcg_out_opc_ldx_h(s, rd, h.base, h.index);
1107        break;
1108    case MO_UL:
1109        if (type == TCG_TYPE_I64) {
1110            tcg_out_opc_ldx_wu(s, rd, h.base, h.index);
1111            break;
1112        }
1113        /* fallthrough */
1114    case MO_SL:
1115        tcg_out_opc_ldx_w(s, rd, h.base, h.index);
1116        break;
1117    case MO_UQ:
1118        tcg_out_opc_ldx_d(s, rd, h.base, h.index);
1119        break;
1120    default:
1121        g_assert_not_reached();
1122    }
1123}
1124
1125static void tcg_out_qemu_ld(TCGContext *s, TCGReg data_reg, TCGReg addr_reg,
1126                            MemOpIdx oi, TCGType data_type)
1127{
1128    TCGLabelQemuLdst *ldst;
1129    HostAddress h;
1130
1131    ldst = prepare_host_addr(s, &h, addr_reg, oi, true);
1132    tcg_out_qemu_ld_indexed(s, get_memop(oi), data_type, data_reg, h);
1133
1134    if (ldst) {
1135        ldst->type = data_type;
1136        ldst->datalo_reg = data_reg;
1137        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1138    }
1139}
1140
1141static void tcg_out_qemu_st_indexed(TCGContext *s, MemOp opc,
1142                                    TCGReg rd, HostAddress h)
1143{
1144    /* Byte swapping is left to middle-end expansion.  */
1145    tcg_debug_assert((opc & MO_BSWAP) == 0);
1146
1147    switch (opc & MO_SIZE) {
1148    case MO_8:
1149        tcg_out_opc_stx_b(s, rd, h.base, h.index);
1150        break;
1151    case MO_16:
1152        tcg_out_opc_stx_h(s, rd, h.base, h.index);
1153        break;
1154    case MO_32:
1155        tcg_out_opc_stx_w(s, rd, h.base, h.index);
1156        break;
1157    case MO_64:
1158        tcg_out_opc_stx_d(s, rd, h.base, h.index);
1159        break;
1160    default:
1161        g_assert_not_reached();
1162    }
1163}
1164
1165static void tcg_out_qemu_st(TCGContext *s, TCGReg data_reg, TCGReg addr_reg,
1166                            MemOpIdx oi, TCGType data_type)
1167{
1168    TCGLabelQemuLdst *ldst;
1169    HostAddress h;
1170
1171    ldst = prepare_host_addr(s, &h, addr_reg, oi, false);
1172    tcg_out_qemu_st_indexed(s, get_memop(oi), data_reg, h);
1173
1174    if (ldst) {
1175        ldst->type = data_type;
1176        ldst->datalo_reg = data_reg;
1177        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1178    }
1179}
1180
1181static void tcg_out_qemu_ldst_i128(TCGContext *s, TCGReg data_lo, TCGReg data_hi,
1182                                   TCGReg addr_reg, MemOpIdx oi, bool is_ld)
1183{
1184    TCGLabelQemuLdst *ldst;
1185    HostAddress h;
1186
1187    ldst = prepare_host_addr(s, &h, addr_reg, oi, is_ld);
1188
1189    if (h.aa.atom == MO_128) {
1190        /*
1191         * Use VLDX/VSTX when 128-bit atomicity is required.
1192         * If address is aligned to 16-bytes, the 128-bit load/store is atomic.
1193         */
1194        if (is_ld) {
1195            tcg_out_opc_vldx(s, TCG_VEC_TMP0, h.base, h.index);
1196            tcg_out_opc_vpickve2gr_d(s, data_lo, TCG_VEC_TMP0, 0);
1197            tcg_out_opc_vpickve2gr_d(s, data_hi, TCG_VEC_TMP0, 1);
1198        } else {
1199            tcg_out_opc_vinsgr2vr_d(s, TCG_VEC_TMP0, data_lo, 0);
1200            tcg_out_opc_vinsgr2vr_d(s, TCG_VEC_TMP0, data_hi, 1);
1201            tcg_out_opc_vstx(s, TCG_VEC_TMP0, h.base, h.index);
1202        }
1203    } else {
1204        /* Otherwise use a pair of LD/ST. */
1205        TCGReg base = h.base;
1206        if (h.index != TCG_REG_ZERO) {
1207            base = TCG_REG_TMP0;
1208            tcg_out_opc_add_d(s, base, h.base, h.index);
1209        }
1210        if (is_ld) {
1211            tcg_debug_assert(base != data_lo);
1212            tcg_out_opc_ld_d(s, data_lo, base, 0);
1213            tcg_out_opc_ld_d(s, data_hi, base, 8);
1214        } else {
1215            tcg_out_opc_st_d(s, data_lo, base, 0);
1216            tcg_out_opc_st_d(s, data_hi, base, 8);
1217        }
1218    }
1219
1220    if (ldst) {
1221        ldst->type = TCG_TYPE_I128;
1222        ldst->datalo_reg = data_lo;
1223        ldst->datahi_reg = data_hi;
1224        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1225    }
1226}
1227
1228/*
1229 * Entry-points
1230 */
1231
1232static const tcg_insn_unit *tb_ret_addr;
1233
1234static void tcg_out_exit_tb(TCGContext *s, uintptr_t a0)
1235{
1236    /* Reuse the zeroing that exists for goto_ptr.  */
1237    if (a0 == 0) {
1238        tcg_out_call_int(s, tcg_code_gen_epilogue, true);
1239    } else {
1240        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_A0, a0);
1241        tcg_out_call_int(s, tb_ret_addr, true);
1242    }
1243}
1244
1245static void tcg_out_goto_tb(TCGContext *s, int which)
1246{
1247    /*
1248     * Direct branch, or load indirect address, to be patched
1249     * by tb_target_set_jmp_target.  Check indirect load offset
1250     * in range early, regardless of direct branch distance,
1251     * via assert within tcg_out_opc_pcaddu2i.
1252     */
1253    uintptr_t i_addr = get_jmp_target_addr(s, which);
1254    intptr_t i_disp = tcg_pcrel_diff(s, (void *)i_addr);
1255
1256    set_jmp_insn_offset(s, which);
1257    tcg_out_opc_pcaddu2i(s, TCG_REG_TMP0, i_disp >> 2);
1258
1259    /* Finish the load and indirect branch. */
1260    tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP0, TCG_REG_TMP0, 0);
1261    tcg_out_opc_jirl(s, TCG_REG_ZERO, TCG_REG_TMP0, 0);
1262    set_jmp_reset_offset(s, which);
1263}
1264
1265void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
1266                              uintptr_t jmp_rx, uintptr_t jmp_rw)
1267{
1268    uintptr_t d_addr = tb->jmp_target_addr[n];
1269    ptrdiff_t d_disp = (ptrdiff_t)(d_addr - jmp_rx) >> 2;
1270    tcg_insn_unit insn;
1271
1272    /* Either directly branch, or load slot address for indirect branch. */
1273    if (d_disp == sextreg(d_disp, 0, 26)) {
1274        insn = encode_sd10k16_insn(OPC_B, d_disp);
1275    } else {
1276        uintptr_t i_addr = (uintptr_t)&tb->jmp_target_addr[n];
1277        intptr_t i_disp = i_addr - jmp_rx;
1278        insn = encode_dsj20_insn(OPC_PCADDU2I, TCG_REG_TMP0, i_disp >> 2);
1279    }
1280
1281    qatomic_set((tcg_insn_unit *)jmp_rw, insn);
1282    flush_idcache_range(jmp_rx, jmp_rw, 4);
1283}
1284
1285
1286static void tgen_add(TCGContext *s, TCGType type,
1287                     TCGReg a0, TCGReg a1, TCGReg a2)
1288{
1289    if (type == TCG_TYPE_I32) {
1290        tcg_out_opc_add_w(s, a0, a1, a2);
1291    } else {
1292        tcg_out_opc_add_d(s, a0, a1, a2);
1293    }
1294}
1295
1296static const TCGOutOpBinary outop_add = {
1297    .base.static_constraint = C_O1_I2(r, r, rJ),
1298    .out_rrr = tgen_add,
1299    .out_rri = tcg_out_addi,
1300};
1301
1302static void tgen_and(TCGContext *s, TCGType type,
1303                     TCGReg a0, TCGReg a1, TCGReg a2)
1304{
1305    tcg_out_opc_and(s, a0, a1, a2);
1306}
1307
1308static void tgen_andi(TCGContext *s, TCGType type,
1309                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1310{
1311    tcg_out_opc_andi(s, a0, a1, a2);
1312}
1313
1314static const TCGOutOpBinary outop_and = {
1315    .base.static_constraint = C_O1_I2(r, r, rU),
1316    .out_rrr = tgen_and,
1317    .out_rri = tgen_andi,
1318};
1319
1320static void tgen_andc(TCGContext *s, TCGType type,
1321                      TCGReg a0, TCGReg a1, TCGReg a2)
1322{
1323    tcg_out_opc_andn(s, a0, a1, a2);
1324}
1325
1326static const TCGOutOpBinary outop_andc = {
1327    .base.static_constraint = C_O1_I2(r, r, r),
1328    .out_rrr = tgen_andc,
1329};
1330
1331static void tgen_or(TCGContext *s, TCGType type,
1332                    TCGReg a0, TCGReg a1, TCGReg a2)
1333{
1334    tcg_out_opc_or(s, a0, a1, a2);
1335}
1336
1337static void tgen_ori(TCGContext *s, TCGType type,
1338                     TCGReg a0, TCGReg a1, tcg_target_long a2)
1339{
1340    tcg_out_opc_ori(s, a0, a1, a2);
1341}
1342
1343static const TCGOutOpBinary outop_or = {
1344    .base.static_constraint = C_O1_I2(r, r, rU),
1345    .out_rrr = tgen_or,
1346    .out_rri = tgen_ori,
1347};
1348
1349static void tgen_orc(TCGContext *s, TCGType type,
1350                     TCGReg a0, TCGReg a1, TCGReg a2)
1351{
1352    tcg_out_opc_orn(s, a0, a1, a2);
1353}
1354
1355static const TCGOutOpBinary outop_orc = {
1356    .base.static_constraint = C_O1_I2(r, r, r),
1357    .out_rrr = tgen_orc,
1358};
1359
1360
1361static void tcg_out_op(TCGContext *s, TCGOpcode opc, TCGType type,
1362                       const TCGArg args[TCG_MAX_OP_ARGS],
1363                       const int const_args[TCG_MAX_OP_ARGS])
1364{
1365    TCGArg a0 = args[0];
1366    TCGArg a1 = args[1];
1367    TCGArg a2 = args[2];
1368    TCGArg a3 = args[3];
1369    int c2 = const_args[2];
1370
1371    switch (opc) {
1372    case INDEX_op_mb:
1373        tcg_out_mb(s, a0);
1374        break;
1375
1376    case INDEX_op_goto_ptr:
1377        tcg_out_opc_jirl(s, TCG_REG_ZERO, a0, 0);
1378        break;
1379
1380    case INDEX_op_br:
1381        tcg_out_reloc(s, s->code_ptr, R_LOONGARCH_BR_SD10K16, arg_label(a0),
1382                      0);
1383        tcg_out_opc_b(s, 0);
1384        break;
1385
1386    case INDEX_op_brcond_i32:
1387    case INDEX_op_brcond_i64:
1388        tcg_out_brcond(s, a2, a0, a1, arg_label(args[3]));
1389        break;
1390
1391    case INDEX_op_extrh_i64_i32:
1392        tcg_out_opc_srai_d(s, a0, a1, 32);
1393        break;
1394
1395    case INDEX_op_not_i32:
1396    case INDEX_op_not_i64:
1397        tcg_out_opc_nor(s, a0, a1, TCG_REG_ZERO);
1398        break;
1399
1400    case INDEX_op_nor_i32:
1401    case INDEX_op_nor_i64:
1402        if (c2) {
1403            tcg_out_opc_ori(s, a0, a1, a2);
1404            tcg_out_opc_nor(s, a0, a0, TCG_REG_ZERO);
1405        } else {
1406            tcg_out_opc_nor(s, a0, a1, a2);
1407        }
1408        break;
1409
1410    case INDEX_op_xor_i32:
1411    case INDEX_op_xor_i64:
1412        if (c2) {
1413            tcg_out_opc_xori(s, a0, a1, a2);
1414        } else {
1415            tcg_out_opc_xor(s, a0, a1, a2);
1416        }
1417        break;
1418
1419    case INDEX_op_extract_i32:
1420        if (a2 == 0 && args[3] <= 12) {
1421            tcg_out_opc_andi(s, a0, a1, (1 << args[3]) - 1);
1422        } else {
1423            tcg_out_opc_bstrpick_w(s, a0, a1, a2, a2 + args[3] - 1);
1424        }
1425        break;
1426    case INDEX_op_extract_i64:
1427        if (a2 == 0 && args[3] <= 12) {
1428            tcg_out_opc_andi(s, a0, a1, (1 << args[3]) - 1);
1429        } else {
1430            tcg_out_opc_bstrpick_d(s, a0, a1, a2, a2 + args[3] - 1);
1431        }
1432        break;
1433
1434    case INDEX_op_sextract_i64:
1435        if (a2 + args[3] == 32) {
1436            if (a2 == 0) {
1437                tcg_out_ext32s(s, a0, a1);
1438            } else {
1439                tcg_out_opc_srai_w(s, a0, a1, a2);
1440            }
1441            break;
1442        }
1443        /* FALLTHRU */
1444    case INDEX_op_sextract_i32:
1445        if (a2 == 0 && args[3] == 8) {
1446            tcg_out_ext8s(s, TCG_TYPE_REG, a0, a1);
1447        } else if (a2 == 0 && args[3] == 16) {
1448            tcg_out_ext16s(s, TCG_TYPE_REG, a0, a1);
1449        } else {
1450            g_assert_not_reached();
1451        }
1452        break;
1453
1454    case INDEX_op_deposit_i32:
1455        tcg_out_opc_bstrins_w(s, a0, a2, args[3], args[3] + args[4] - 1);
1456        break;
1457    case INDEX_op_deposit_i64:
1458        tcg_out_opc_bstrins_d(s, a0, a2, args[3], args[3] + args[4] - 1);
1459        break;
1460
1461    case INDEX_op_bswap16_i32:
1462    case INDEX_op_bswap16_i64:
1463        tcg_out_opc_revb_2h(s, a0, a1);
1464        if (a2 & TCG_BSWAP_OS) {
1465            tcg_out_ext16s(s, TCG_TYPE_REG, a0, a0);
1466        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1467            tcg_out_ext16u(s, a0, a0);
1468        }
1469        break;
1470
1471    case INDEX_op_bswap32_i32:
1472        /* All 32-bit values are computed sign-extended in the register.  */
1473        a2 = TCG_BSWAP_OS;
1474        /* fallthrough */
1475    case INDEX_op_bswap32_i64:
1476        tcg_out_opc_revb_2w(s, a0, a1);
1477        if (a2 & TCG_BSWAP_OS) {
1478            tcg_out_ext32s(s, a0, a0);
1479        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1480            tcg_out_ext32u(s, a0, a0);
1481        }
1482        break;
1483
1484    case INDEX_op_bswap64_i64:
1485        tcg_out_opc_revb_d(s, a0, a1);
1486        break;
1487
1488    case INDEX_op_clz_i32:
1489        tcg_out_clzctz(s, OPC_CLZ_W, a0, a1, a2, c2, true);
1490        break;
1491    case INDEX_op_clz_i64:
1492        tcg_out_clzctz(s, OPC_CLZ_D, a0, a1, a2, c2, false);
1493        break;
1494
1495    case INDEX_op_ctz_i32:
1496        tcg_out_clzctz(s, OPC_CTZ_W, a0, a1, a2, c2, true);
1497        break;
1498    case INDEX_op_ctz_i64:
1499        tcg_out_clzctz(s, OPC_CTZ_D, a0, a1, a2, c2, false);
1500        break;
1501
1502    case INDEX_op_shl_i32:
1503        if (c2) {
1504            tcg_out_opc_slli_w(s, a0, a1, a2 & 0x1f);
1505        } else {
1506            tcg_out_opc_sll_w(s, a0, a1, a2);
1507        }
1508        break;
1509    case INDEX_op_shl_i64:
1510        if (c2) {
1511            tcg_out_opc_slli_d(s, a0, a1, a2 & 0x3f);
1512        } else {
1513            tcg_out_opc_sll_d(s, a0, a1, a2);
1514        }
1515        break;
1516
1517    case INDEX_op_shr_i32:
1518        if (c2) {
1519            tcg_out_opc_srli_w(s, a0, a1, a2 & 0x1f);
1520        } else {
1521            tcg_out_opc_srl_w(s, a0, a1, a2);
1522        }
1523        break;
1524    case INDEX_op_shr_i64:
1525        if (c2) {
1526            tcg_out_opc_srli_d(s, a0, a1, a2 & 0x3f);
1527        } else {
1528            tcg_out_opc_srl_d(s, a0, a1, a2);
1529        }
1530        break;
1531
1532    case INDEX_op_sar_i32:
1533        if (c2) {
1534            tcg_out_opc_srai_w(s, a0, a1, a2 & 0x1f);
1535        } else {
1536            tcg_out_opc_sra_w(s, a0, a1, a2);
1537        }
1538        break;
1539    case INDEX_op_sar_i64:
1540        if (c2) {
1541            tcg_out_opc_srai_d(s, a0, a1, a2 & 0x3f);
1542        } else {
1543            tcg_out_opc_sra_d(s, a0, a1, a2);
1544        }
1545        break;
1546
1547    case INDEX_op_rotl_i32:
1548        /* transform into equivalent rotr/rotri */
1549        if (c2) {
1550            tcg_out_opc_rotri_w(s, a0, a1, (32 - a2) & 0x1f);
1551        } else {
1552            tcg_out_opc_sub_w(s, TCG_REG_TMP0, TCG_REG_ZERO, a2);
1553            tcg_out_opc_rotr_w(s, a0, a1, TCG_REG_TMP0);
1554        }
1555        break;
1556    case INDEX_op_rotl_i64:
1557        /* transform into equivalent rotr/rotri */
1558        if (c2) {
1559            tcg_out_opc_rotri_d(s, a0, a1, (64 - a2) & 0x3f);
1560        } else {
1561            tcg_out_opc_sub_w(s, TCG_REG_TMP0, TCG_REG_ZERO, a2);
1562            tcg_out_opc_rotr_d(s, a0, a1, TCG_REG_TMP0);
1563        }
1564        break;
1565
1566    case INDEX_op_rotr_i32:
1567        if (c2) {
1568            tcg_out_opc_rotri_w(s, a0, a1, a2 & 0x1f);
1569        } else {
1570            tcg_out_opc_rotr_w(s, a0, a1, a2);
1571        }
1572        break;
1573    case INDEX_op_rotr_i64:
1574        if (c2) {
1575            tcg_out_opc_rotri_d(s, a0, a1, a2 & 0x3f);
1576        } else {
1577            tcg_out_opc_rotr_d(s, a0, a1, a2);
1578        }
1579        break;
1580
1581    case INDEX_op_sub_i32:
1582        if (c2) {
1583            tcg_out_addi(s, TCG_TYPE_I32, a0, a1, -a2);
1584        } else {
1585            tcg_out_opc_sub_w(s, a0, a1, a2);
1586        }
1587        break;
1588    case INDEX_op_sub_i64:
1589        if (c2) {
1590            tcg_out_addi(s, TCG_TYPE_I64, a0, a1, -a2);
1591        } else {
1592            tcg_out_opc_sub_d(s, a0, a1, a2);
1593        }
1594        break;
1595
1596    case INDEX_op_neg_i32:
1597        tcg_out_opc_sub_w(s, a0, TCG_REG_ZERO, a1);
1598        break;
1599    case INDEX_op_neg_i64:
1600        tcg_out_opc_sub_d(s, a0, TCG_REG_ZERO, a1);
1601        break;
1602
1603    case INDEX_op_mul_i32:
1604        tcg_out_opc_mul_w(s, a0, a1, a2);
1605        break;
1606    case INDEX_op_mul_i64:
1607        tcg_out_opc_mul_d(s, a0, a1, a2);
1608        break;
1609
1610    case INDEX_op_mulsh_i32:
1611        tcg_out_opc_mulh_w(s, a0, a1, a2);
1612        break;
1613    case INDEX_op_mulsh_i64:
1614        tcg_out_opc_mulh_d(s, a0, a1, a2);
1615        break;
1616
1617    case INDEX_op_muluh_i32:
1618        tcg_out_opc_mulh_wu(s, a0, a1, a2);
1619        break;
1620    case INDEX_op_muluh_i64:
1621        tcg_out_opc_mulh_du(s, a0, a1, a2);
1622        break;
1623
1624    case INDEX_op_div_i32:
1625        tcg_out_opc_div_w(s, a0, a1, a2);
1626        break;
1627    case INDEX_op_div_i64:
1628        tcg_out_opc_div_d(s, a0, a1, a2);
1629        break;
1630
1631    case INDEX_op_divu_i32:
1632        tcg_out_opc_div_wu(s, a0, a1, a2);
1633        break;
1634    case INDEX_op_divu_i64:
1635        tcg_out_opc_div_du(s, a0, a1, a2);
1636        break;
1637
1638    case INDEX_op_rem_i32:
1639        tcg_out_opc_mod_w(s, a0, a1, a2);
1640        break;
1641    case INDEX_op_rem_i64:
1642        tcg_out_opc_mod_d(s, a0, a1, a2);
1643        break;
1644
1645    case INDEX_op_remu_i32:
1646        tcg_out_opc_mod_wu(s, a0, a1, a2);
1647        break;
1648    case INDEX_op_remu_i64:
1649        tcg_out_opc_mod_du(s, a0, a1, a2);
1650        break;
1651
1652    case INDEX_op_setcond_i32:
1653    case INDEX_op_setcond_i64:
1654        tcg_out_setcond(s, args[3], a0, a1, a2, c2);
1655        break;
1656
1657    case INDEX_op_movcond_i32:
1658    case INDEX_op_movcond_i64:
1659        tcg_out_movcond(s, args[5], a0, a1, a2, c2, args[3], args[4]);
1660        break;
1661
1662    case INDEX_op_ld8s_i32:
1663    case INDEX_op_ld8s_i64:
1664        tcg_out_ldst(s, OPC_LD_B, a0, a1, a2);
1665        break;
1666    case INDEX_op_ld8u_i32:
1667    case INDEX_op_ld8u_i64:
1668        tcg_out_ldst(s, OPC_LD_BU, a0, a1, a2);
1669        break;
1670    case INDEX_op_ld16s_i32:
1671    case INDEX_op_ld16s_i64:
1672        tcg_out_ldst(s, OPC_LD_H, a0, a1, a2);
1673        break;
1674    case INDEX_op_ld16u_i32:
1675    case INDEX_op_ld16u_i64:
1676        tcg_out_ldst(s, OPC_LD_HU, a0, a1, a2);
1677        break;
1678    case INDEX_op_ld_i32:
1679    case INDEX_op_ld32s_i64:
1680        tcg_out_ldst(s, OPC_LD_W, a0, a1, a2);
1681        break;
1682    case INDEX_op_ld32u_i64:
1683        tcg_out_ldst(s, OPC_LD_WU, a0, a1, a2);
1684        break;
1685    case INDEX_op_ld_i64:
1686        tcg_out_ldst(s, OPC_LD_D, a0, a1, a2);
1687        break;
1688
1689    case INDEX_op_st8_i32:
1690    case INDEX_op_st8_i64:
1691        tcg_out_ldst(s, OPC_ST_B, a0, a1, a2);
1692        break;
1693    case INDEX_op_st16_i32:
1694    case INDEX_op_st16_i64:
1695        tcg_out_ldst(s, OPC_ST_H, a0, a1, a2);
1696        break;
1697    case INDEX_op_st_i32:
1698    case INDEX_op_st32_i64:
1699        tcg_out_ldst(s, OPC_ST_W, a0, a1, a2);
1700        break;
1701    case INDEX_op_st_i64:
1702        tcg_out_ldst(s, OPC_ST_D, a0, a1, a2);
1703        break;
1704
1705    case INDEX_op_qemu_ld_i32:
1706        tcg_out_qemu_ld(s, a0, a1, a2, TCG_TYPE_I32);
1707        break;
1708    case INDEX_op_qemu_ld_i64:
1709        tcg_out_qemu_ld(s, a0, a1, a2, TCG_TYPE_I64);
1710        break;
1711    case INDEX_op_qemu_ld_i128:
1712        tcg_out_qemu_ldst_i128(s, a0, a1, a2, a3, true);
1713        break;
1714    case INDEX_op_qemu_st_i32:
1715        tcg_out_qemu_st(s, a0, a1, a2, TCG_TYPE_I32);
1716        break;
1717    case INDEX_op_qemu_st_i64:
1718        tcg_out_qemu_st(s, a0, a1, a2, TCG_TYPE_I64);
1719        break;
1720    case INDEX_op_qemu_st_i128:
1721        tcg_out_qemu_ldst_i128(s, a0, a1, a2, a3, false);
1722        break;
1723
1724    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
1725    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
1726    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
1727    case INDEX_op_ext_i32_i64:  /* Always emitted via tcg_reg_alloc_op.  */
1728    case INDEX_op_extu_i32_i64:
1729    case INDEX_op_extrl_i64_i32:
1730    default:
1731        g_assert_not_reached();
1732    }
1733}
1734
1735static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
1736                            TCGReg rd, TCGReg rs)
1737{
1738    static const LoongArchInsn repl_insn[2][4] = {
1739        { OPC_VREPLGR2VR_B, OPC_VREPLGR2VR_H,
1740          OPC_VREPLGR2VR_W, OPC_VREPLGR2VR_D },
1741        { OPC_XVREPLGR2VR_B, OPC_XVREPLGR2VR_H,
1742          OPC_XVREPLGR2VR_W, OPC_XVREPLGR2VR_D },
1743    };
1744    bool lasx = type == TCG_TYPE_V256;
1745
1746    tcg_debug_assert(vece <= MO_64);
1747    tcg_out32(s, encode_vdj_insn(repl_insn[lasx][vece], rd, rs));
1748    return true;
1749}
1750
1751static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
1752                             TCGReg r, TCGReg base, intptr_t offset)
1753{
1754    bool lasx = type == TCG_TYPE_V256;
1755
1756    /* Handle imm overflow and division (vldrepl.d imm is divided by 8). */
1757    if (offset < -0x800 || offset > 0x7ff ||
1758        (offset & ((1 << vece) - 1)) != 0) {
1759        tcg_out_addi(s, TCG_TYPE_I64, TCG_REG_TMP0, base, offset);
1760        base = TCG_REG_TMP0;
1761        offset = 0;
1762    }
1763    offset >>= vece;
1764
1765    switch (vece) {
1766    case MO_8:
1767        if (lasx) {
1768            tcg_out_opc_xvldrepl_b(s, r, base, offset);
1769        } else {
1770            tcg_out_opc_vldrepl_b(s, r, base, offset);
1771        }
1772        break;
1773    case MO_16:
1774        if (lasx) {
1775            tcg_out_opc_xvldrepl_h(s, r, base, offset);
1776        } else {
1777            tcg_out_opc_vldrepl_h(s, r, base, offset);
1778        }
1779        break;
1780    case MO_32:
1781        if (lasx) {
1782            tcg_out_opc_xvldrepl_w(s, r, base, offset);
1783        } else {
1784            tcg_out_opc_vldrepl_w(s, r, base, offset);
1785        }
1786        break;
1787    case MO_64:
1788        if (lasx) {
1789            tcg_out_opc_xvldrepl_d(s, r, base, offset);
1790        } else {
1791            tcg_out_opc_vldrepl_d(s, r, base, offset);
1792        }
1793        break;
1794    default:
1795        g_assert_not_reached();
1796    }
1797    return true;
1798}
1799
1800static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
1801                             TCGReg rd, int64_t v64)
1802{
1803    /* Try vldi if imm can fit */
1804    int64_t value = sextract64(v64, 0, 8 << vece);
1805    if (-0x200 <= value && value <= 0x1FF) {
1806        uint32_t imm = (vece << 10) | ((uint32_t)v64 & 0x3FF);
1807
1808        if (type == TCG_TYPE_V256) {
1809            tcg_out_opc_xvldi(s, rd, imm);
1810        } else {
1811            tcg_out_opc_vldi(s, rd, imm);
1812        }
1813        return;
1814    }
1815
1816    /* TODO: vldi patterns when imm 12 is set */
1817
1818    tcg_out_movi(s, TCG_TYPE_I64, TCG_REG_TMP0, value);
1819    tcg_out_dup_vec(s, type, vece, rd, TCG_REG_TMP0);
1820}
1821
1822static void tcg_out_addsub_vec(TCGContext *s, bool lasx, unsigned vece,
1823                               TCGArg a0, TCGArg a1, TCGArg a2,
1824                               bool a2_is_const, bool is_add)
1825{
1826    static const LoongArchInsn add_vec_insn[2][4] = {
1827        { OPC_VADD_B, OPC_VADD_H, OPC_VADD_W, OPC_VADD_D },
1828        { OPC_XVADD_B, OPC_XVADD_H, OPC_XVADD_W, OPC_XVADD_D },
1829    };
1830    static const LoongArchInsn add_vec_imm_insn[2][4] = {
1831        { OPC_VADDI_BU, OPC_VADDI_HU, OPC_VADDI_WU, OPC_VADDI_DU },
1832        { OPC_XVADDI_BU, OPC_XVADDI_HU, OPC_XVADDI_WU, OPC_XVADDI_DU },
1833    };
1834    static const LoongArchInsn sub_vec_insn[2][4] = {
1835        { OPC_VSUB_B, OPC_VSUB_H, OPC_VSUB_W, OPC_VSUB_D },
1836        { OPC_XVSUB_B, OPC_XVSUB_H, OPC_XVSUB_W, OPC_XVSUB_D },
1837    };
1838    static const LoongArchInsn sub_vec_imm_insn[2][4] = {
1839        { OPC_VSUBI_BU, OPC_VSUBI_HU, OPC_VSUBI_WU, OPC_VSUBI_DU },
1840        { OPC_XVSUBI_BU, OPC_XVSUBI_HU, OPC_XVSUBI_WU, OPC_XVSUBI_DU },
1841    };
1842    LoongArchInsn insn;
1843
1844    if (a2_is_const) {
1845        int64_t value = sextract64(a2, 0, 8 << vece);
1846
1847        if (!is_add) {
1848            value = -value;
1849        }
1850        if (value < 0) {
1851            insn = sub_vec_imm_insn[lasx][vece];
1852            value = -value;
1853        } else {
1854            insn = add_vec_imm_insn[lasx][vece];
1855        }
1856
1857        /* Constraint TCG_CT_CONST_VADD ensures validity. */
1858        tcg_debug_assert(0 <= value && value <= 0x1f);
1859
1860        tcg_out32(s, encode_vdvjuk5_insn(insn, a0, a1, value));
1861        return;
1862    }
1863
1864    if (is_add) {
1865        insn = add_vec_insn[lasx][vece];
1866    } else {
1867        insn = sub_vec_insn[lasx][vece];
1868    }
1869    tcg_out32(s, encode_vdvjvk_insn(insn, a0, a1, a2));
1870}
1871
1872static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
1873                           unsigned vecl, unsigned vece,
1874                           const TCGArg args[TCG_MAX_OP_ARGS],
1875                           const int const_args[TCG_MAX_OP_ARGS])
1876{
1877    TCGType type = vecl + TCG_TYPE_V64;
1878    bool lasx = type == TCG_TYPE_V256;
1879    TCGArg a0, a1, a2, a3;
1880    LoongArchInsn insn;
1881
1882    static const LoongArchInsn cmp_vec_insn[16][2][4] = {
1883        [TCG_COND_EQ] = {
1884            { OPC_VSEQ_B, OPC_VSEQ_H, OPC_VSEQ_W, OPC_VSEQ_D },
1885            { OPC_XVSEQ_B, OPC_XVSEQ_H, OPC_XVSEQ_W, OPC_XVSEQ_D },
1886        },
1887        [TCG_COND_LE] = {
1888            { OPC_VSLE_B, OPC_VSLE_H, OPC_VSLE_W, OPC_VSLE_D },
1889            { OPC_XVSLE_B, OPC_XVSLE_H, OPC_XVSLE_W, OPC_XVSLE_D },
1890        },
1891        [TCG_COND_LEU] = {
1892            { OPC_VSLE_BU, OPC_VSLE_HU, OPC_VSLE_WU, OPC_VSLE_DU },
1893            { OPC_XVSLE_BU, OPC_XVSLE_HU, OPC_XVSLE_WU, OPC_XVSLE_DU },
1894        },
1895        [TCG_COND_LT] = {
1896            { OPC_VSLT_B, OPC_VSLT_H, OPC_VSLT_W, OPC_VSLT_D },
1897            { OPC_XVSLT_B, OPC_XVSLT_H, OPC_XVSLT_W, OPC_XVSLT_D },
1898        },
1899        [TCG_COND_LTU] = {
1900            { OPC_VSLT_BU, OPC_VSLT_HU, OPC_VSLT_WU, OPC_VSLT_DU },
1901            { OPC_XVSLT_BU, OPC_XVSLT_HU, OPC_XVSLT_WU, OPC_XVSLT_DU },
1902        }
1903    };
1904    static const LoongArchInsn cmp_vec_imm_insn[16][2][4] = {
1905        [TCG_COND_EQ] = {
1906            { OPC_VSEQI_B, OPC_VSEQI_H, OPC_VSEQI_W, OPC_VSEQI_D },
1907            { OPC_XVSEQI_B, OPC_XVSEQI_H, OPC_XVSEQI_W, OPC_XVSEQI_D },
1908        },
1909        [TCG_COND_LE] = {
1910            { OPC_VSLEI_B, OPC_VSLEI_H, OPC_VSLEI_W, OPC_VSLEI_D },
1911            { OPC_XVSLEI_B, OPC_XVSLEI_H, OPC_XVSLEI_W, OPC_XVSLEI_D },
1912        },
1913        [TCG_COND_LEU] = {
1914            { OPC_VSLEI_BU, OPC_VSLEI_HU, OPC_VSLEI_WU, OPC_VSLEI_DU },
1915            { OPC_XVSLEI_BU, OPC_XVSLEI_HU, OPC_XVSLEI_WU, OPC_XVSLEI_DU },
1916        },
1917        [TCG_COND_LT] = {
1918            { OPC_VSLTI_B, OPC_VSLTI_H, OPC_VSLTI_W, OPC_VSLTI_D },
1919            { OPC_XVSLTI_B, OPC_XVSLTI_H, OPC_XVSLTI_W, OPC_XVSLTI_D },
1920        },
1921        [TCG_COND_LTU] = {
1922            { OPC_VSLTI_BU, OPC_VSLTI_HU, OPC_VSLTI_WU, OPC_VSLTI_DU },
1923            { OPC_XVSLTI_BU, OPC_XVSLTI_HU, OPC_XVSLTI_WU, OPC_XVSLTI_DU },
1924        }
1925    };
1926    static const LoongArchInsn neg_vec_insn[2][4] = {
1927        { OPC_VNEG_B, OPC_VNEG_H, OPC_VNEG_W, OPC_VNEG_D },
1928        { OPC_XVNEG_B, OPC_XVNEG_H, OPC_XVNEG_W, OPC_XVNEG_D },
1929    };
1930    static const LoongArchInsn mul_vec_insn[2][4] = {
1931        { OPC_VMUL_B, OPC_VMUL_H, OPC_VMUL_W, OPC_VMUL_D },
1932        { OPC_XVMUL_B, OPC_XVMUL_H, OPC_XVMUL_W, OPC_XVMUL_D },
1933    };
1934    static const LoongArchInsn smin_vec_insn[2][4] = {
1935        { OPC_VMIN_B, OPC_VMIN_H, OPC_VMIN_W, OPC_VMIN_D },
1936        { OPC_XVMIN_B, OPC_XVMIN_H, OPC_XVMIN_W, OPC_XVMIN_D },
1937    };
1938    static const LoongArchInsn umin_vec_insn[2][4] = {
1939        { OPC_VMIN_BU, OPC_VMIN_HU, OPC_VMIN_WU, OPC_VMIN_DU },
1940        { OPC_XVMIN_BU, OPC_XVMIN_HU, OPC_XVMIN_WU, OPC_XVMIN_DU },
1941    };
1942    static const LoongArchInsn smax_vec_insn[2][4] = {
1943        { OPC_VMAX_B, OPC_VMAX_H, OPC_VMAX_W, OPC_VMAX_D },
1944        { OPC_XVMAX_B, OPC_XVMAX_H, OPC_XVMAX_W, OPC_XVMAX_D },
1945    };
1946    static const LoongArchInsn umax_vec_insn[2][4] = {
1947        { OPC_VMAX_BU, OPC_VMAX_HU, OPC_VMAX_WU, OPC_VMAX_DU },
1948        { OPC_XVMAX_BU, OPC_XVMAX_HU, OPC_XVMAX_WU, OPC_XVMAX_DU },
1949    };
1950    static const LoongArchInsn ssadd_vec_insn[2][4] = {
1951        { OPC_VSADD_B, OPC_VSADD_H, OPC_VSADD_W, OPC_VSADD_D },
1952        { OPC_XVSADD_B, OPC_XVSADD_H, OPC_XVSADD_W, OPC_XVSADD_D },
1953    };
1954    static const LoongArchInsn usadd_vec_insn[2][4] = {
1955        { OPC_VSADD_BU, OPC_VSADD_HU, OPC_VSADD_WU, OPC_VSADD_DU },
1956        { OPC_XVSADD_BU, OPC_XVSADD_HU, OPC_XVSADD_WU, OPC_XVSADD_DU },
1957    };
1958    static const LoongArchInsn sssub_vec_insn[2][4] = {
1959        { OPC_VSSUB_B, OPC_VSSUB_H, OPC_VSSUB_W, OPC_VSSUB_D },
1960        { OPC_XVSSUB_B, OPC_XVSSUB_H, OPC_XVSSUB_W, OPC_XVSSUB_D },
1961    };
1962    static const LoongArchInsn ussub_vec_insn[2][4] = {
1963        { OPC_VSSUB_BU, OPC_VSSUB_HU, OPC_VSSUB_WU, OPC_VSSUB_DU },
1964        { OPC_XVSSUB_BU, OPC_XVSSUB_HU, OPC_XVSSUB_WU, OPC_XVSSUB_DU },
1965    };
1966    static const LoongArchInsn shlv_vec_insn[2][4] = {
1967        { OPC_VSLL_B, OPC_VSLL_H, OPC_VSLL_W, OPC_VSLL_D },
1968        { OPC_XVSLL_B, OPC_XVSLL_H, OPC_XVSLL_W, OPC_XVSLL_D },
1969    };
1970    static const LoongArchInsn shrv_vec_insn[2][4] = {
1971        { OPC_VSRL_B, OPC_VSRL_H, OPC_VSRL_W, OPC_VSRL_D },
1972        { OPC_XVSRL_B, OPC_XVSRL_H, OPC_XVSRL_W, OPC_XVSRL_D },
1973    };
1974    static const LoongArchInsn sarv_vec_insn[2][4] = {
1975        { OPC_VSRA_B, OPC_VSRA_H, OPC_VSRA_W, OPC_VSRA_D },
1976        { OPC_XVSRA_B, OPC_XVSRA_H, OPC_XVSRA_W, OPC_XVSRA_D },
1977    };
1978    static const LoongArchInsn shli_vec_insn[2][4] = {
1979        { OPC_VSLLI_B, OPC_VSLLI_H, OPC_VSLLI_W, OPC_VSLLI_D },
1980        { OPC_XVSLLI_B, OPC_XVSLLI_H, OPC_XVSLLI_W, OPC_XVSLLI_D },
1981    };
1982    static const LoongArchInsn shri_vec_insn[2][4] = {
1983        { OPC_VSRLI_B, OPC_VSRLI_H, OPC_VSRLI_W, OPC_VSRLI_D },
1984        { OPC_XVSRLI_B, OPC_XVSRLI_H, OPC_XVSRLI_W, OPC_XVSRLI_D },
1985    };
1986    static const LoongArchInsn sari_vec_insn[2][4] = {
1987        { OPC_VSRAI_B, OPC_VSRAI_H, OPC_VSRAI_W, OPC_VSRAI_D },
1988        { OPC_XVSRAI_B, OPC_XVSRAI_H, OPC_XVSRAI_W, OPC_XVSRAI_D },
1989    };
1990    static const LoongArchInsn rotrv_vec_insn[2][4] = {
1991        { OPC_VROTR_B, OPC_VROTR_H, OPC_VROTR_W, OPC_VROTR_D },
1992        { OPC_XVROTR_B, OPC_XVROTR_H, OPC_XVROTR_W, OPC_XVROTR_D },
1993    };
1994    static const LoongArchInsn rotri_vec_insn[2][4] = {
1995        { OPC_VROTRI_B, OPC_VROTRI_H, OPC_VROTRI_W, OPC_VROTRI_D },
1996        { OPC_XVROTRI_B, OPC_XVROTRI_H, OPC_XVROTRI_W, OPC_XVROTRI_D },
1997    };
1998
1999    a0 = args[0];
2000    a1 = args[1];
2001    a2 = args[2];
2002    a3 = args[3];
2003
2004    switch (opc) {
2005    case INDEX_op_st_vec:
2006        tcg_out_st(s, type, a0, a1, a2);
2007        break;
2008    case INDEX_op_ld_vec:
2009        tcg_out_ld(s, type, a0, a1, a2);
2010        break;
2011    case INDEX_op_and_vec:
2012        insn = lasx ? OPC_XVAND_V : OPC_VAND_V;
2013        goto vdvjvk;
2014    case INDEX_op_andc_vec:
2015        /*
2016         * vandn vd, vj, vk: vd = vk & ~vj
2017         * andc_vec vd, vj, vk: vd = vj & ~vk
2018         * vj and vk are swapped
2019         */
2020        a1 = a2;
2021        a2 = args[1];
2022        insn = lasx ? OPC_XVANDN_V : OPC_VANDN_V;
2023        goto vdvjvk;
2024    case INDEX_op_or_vec:
2025        insn = lasx ? OPC_XVOR_V : OPC_VOR_V;
2026        goto vdvjvk;
2027    case INDEX_op_orc_vec:
2028        insn = lasx ? OPC_XVORN_V : OPC_VORN_V;
2029        goto vdvjvk;
2030    case INDEX_op_xor_vec:
2031        insn = lasx ? OPC_XVXOR_V : OPC_VXOR_V;
2032        goto vdvjvk;
2033    case INDEX_op_not_vec:
2034        a2 = a1;
2035        /* fall through */
2036    case INDEX_op_nor_vec:
2037        insn = lasx ? OPC_XVNOR_V : OPC_VNOR_V;
2038        goto vdvjvk;
2039    case INDEX_op_cmp_vec:
2040        {
2041            TCGCond cond = args[3];
2042
2043            if (const_args[2]) {
2044                /*
2045                 * cmp_vec dest, src, value
2046                 * Try vseqi/vslei/vslti
2047                 */
2048                int64_t value = sextract64(a2, 0, 8 << vece);
2049                switch (cond) {
2050                case TCG_COND_EQ:
2051                case TCG_COND_LE:
2052                case TCG_COND_LT:
2053                    insn = cmp_vec_imm_insn[cond][lasx][vece];
2054                    tcg_out32(s, encode_vdvjsk5_insn(insn, a0, a1, value));
2055                    break;
2056                case TCG_COND_LEU:
2057                case TCG_COND_LTU:
2058                    insn = cmp_vec_imm_insn[cond][lasx][vece];
2059                    tcg_out32(s, encode_vdvjuk5_insn(insn, a0, a1, value));
2060                    break;
2061                default:
2062                    g_assert_not_reached();
2063                }
2064                break;
2065            }
2066
2067            insn = cmp_vec_insn[cond][lasx][vece];
2068            if (insn == 0) {
2069                TCGArg t;
2070                t = a1, a1 = a2, a2 = t;
2071                cond = tcg_swap_cond(cond);
2072                insn = cmp_vec_insn[cond][lasx][vece];
2073                tcg_debug_assert(insn != 0);
2074            }
2075        }
2076        goto vdvjvk;
2077    case INDEX_op_add_vec:
2078        tcg_out_addsub_vec(s, lasx, vece, a0, a1, a2, const_args[2], true);
2079        break;
2080    case INDEX_op_sub_vec:
2081        tcg_out_addsub_vec(s, lasx, vece, a0, a1, a2, const_args[2], false);
2082        break;
2083    case INDEX_op_neg_vec:
2084        tcg_out32(s, encode_vdvj_insn(neg_vec_insn[lasx][vece], a0, a1));
2085        break;
2086    case INDEX_op_mul_vec:
2087        insn = mul_vec_insn[lasx][vece];
2088        goto vdvjvk;
2089    case INDEX_op_smin_vec:
2090        insn = smin_vec_insn[lasx][vece];
2091        goto vdvjvk;
2092    case INDEX_op_smax_vec:
2093        insn = smax_vec_insn[lasx][vece];
2094        goto vdvjvk;
2095    case INDEX_op_umin_vec:
2096        insn = umin_vec_insn[lasx][vece];
2097        goto vdvjvk;
2098    case INDEX_op_umax_vec:
2099        insn = umax_vec_insn[lasx][vece];
2100        goto vdvjvk;
2101    case INDEX_op_ssadd_vec:
2102        insn = ssadd_vec_insn[lasx][vece];
2103        goto vdvjvk;
2104    case INDEX_op_usadd_vec:
2105        insn = usadd_vec_insn[lasx][vece];
2106        goto vdvjvk;
2107    case INDEX_op_sssub_vec:
2108        insn = sssub_vec_insn[lasx][vece];
2109        goto vdvjvk;
2110    case INDEX_op_ussub_vec:
2111        insn = ussub_vec_insn[lasx][vece];
2112        goto vdvjvk;
2113    case INDEX_op_shlv_vec:
2114        insn = shlv_vec_insn[lasx][vece];
2115        goto vdvjvk;
2116    case INDEX_op_shrv_vec:
2117        insn = shrv_vec_insn[lasx][vece];
2118        goto vdvjvk;
2119    case INDEX_op_sarv_vec:
2120        insn = sarv_vec_insn[lasx][vece];
2121        goto vdvjvk;
2122    case INDEX_op_rotlv_vec:
2123        /* rotlv_vec a1, a2 = rotrv_vec a1, -a2 */
2124        tcg_out32(s, encode_vdvj_insn(neg_vec_insn[lasx][vece],
2125                                      TCG_VEC_TMP0, a2));
2126        a2 = TCG_VEC_TMP0;
2127        /* fall through */
2128    case INDEX_op_rotrv_vec:
2129        insn = rotrv_vec_insn[lasx][vece];
2130        goto vdvjvk;
2131    case INDEX_op_shli_vec:
2132        insn = shli_vec_insn[lasx][vece];
2133        goto vdvjukN;
2134    case INDEX_op_shri_vec:
2135        insn = shri_vec_insn[lasx][vece];
2136        goto vdvjukN;
2137    case INDEX_op_sari_vec:
2138        insn = sari_vec_insn[lasx][vece];
2139        goto vdvjukN;
2140    case INDEX_op_rotli_vec:
2141        /* rotli_vec a1, a2 = rotri_vec a1, -a2 */
2142        a2 = extract32(-a2, 0, 3 + vece);
2143        insn = rotri_vec_insn[lasx][vece];
2144        goto vdvjukN;
2145    case INDEX_op_bitsel_vec:
2146        /* vbitsel vd, vj, vk, va = bitsel_vec vd, va, vk, vj */
2147        if (lasx) {
2148            tcg_out_opc_xvbitsel_v(s, a0, a3, a2, a1);
2149        } else {
2150            tcg_out_opc_vbitsel_v(s, a0, a3, a2, a1);
2151        }
2152        break;
2153    case INDEX_op_dupm_vec:
2154        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
2155        break;
2156    default:
2157        g_assert_not_reached();
2158    vdvjvk:
2159        tcg_out32(s, encode_vdvjvk_insn(insn, a0, a1, a2));
2160        break;
2161    vdvjukN:
2162        switch (vece) {
2163        case MO_8:
2164            tcg_out32(s, encode_vdvjuk3_insn(insn, a0, a1, a2));
2165            break;
2166        case MO_16:
2167            tcg_out32(s, encode_vdvjuk4_insn(insn, a0, a1, a2));
2168            break;
2169        case MO_32:
2170            tcg_out32(s, encode_vdvjuk5_insn(insn, a0, a1, a2));
2171            break;
2172        case MO_64:
2173            tcg_out32(s, encode_vdvjuk6_insn(insn, a0, a1, a2));
2174            break;
2175        default:
2176            g_assert_not_reached();
2177        }
2178        break;
2179    }
2180}
2181
2182int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
2183{
2184    switch (opc) {
2185    case INDEX_op_ld_vec:
2186    case INDEX_op_st_vec:
2187    case INDEX_op_dup_vec:
2188    case INDEX_op_dupm_vec:
2189    case INDEX_op_cmp_vec:
2190    case INDEX_op_add_vec:
2191    case INDEX_op_sub_vec:
2192    case INDEX_op_and_vec:
2193    case INDEX_op_andc_vec:
2194    case INDEX_op_or_vec:
2195    case INDEX_op_orc_vec:
2196    case INDEX_op_xor_vec:
2197    case INDEX_op_nor_vec:
2198    case INDEX_op_not_vec:
2199    case INDEX_op_neg_vec:
2200    case INDEX_op_mul_vec:
2201    case INDEX_op_smin_vec:
2202    case INDEX_op_smax_vec:
2203    case INDEX_op_umin_vec:
2204    case INDEX_op_umax_vec:
2205    case INDEX_op_ssadd_vec:
2206    case INDEX_op_usadd_vec:
2207    case INDEX_op_sssub_vec:
2208    case INDEX_op_ussub_vec:
2209    case INDEX_op_shlv_vec:
2210    case INDEX_op_shrv_vec:
2211    case INDEX_op_sarv_vec:
2212    case INDEX_op_bitsel_vec:
2213        return 1;
2214    default:
2215        return 0;
2216    }
2217}
2218
2219void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
2220                       TCGArg a0, ...)
2221{
2222    g_assert_not_reached();
2223}
2224
2225static TCGConstraintSetIndex
2226tcg_target_op_def(TCGOpcode op, TCGType type, unsigned flags)
2227{
2228    switch (op) {
2229    case INDEX_op_goto_ptr:
2230        return C_O0_I1(r);
2231
2232    case INDEX_op_st8_i32:
2233    case INDEX_op_st8_i64:
2234    case INDEX_op_st16_i32:
2235    case INDEX_op_st16_i64:
2236    case INDEX_op_st32_i64:
2237    case INDEX_op_st_i32:
2238    case INDEX_op_st_i64:
2239    case INDEX_op_qemu_st_i32:
2240    case INDEX_op_qemu_st_i64:
2241        return C_O0_I2(rz, r);
2242
2243    case INDEX_op_qemu_ld_i128:
2244        return C_N2_I1(r, r, r);
2245
2246    case INDEX_op_qemu_st_i128:
2247        return C_O0_I3(r, r, r);
2248
2249    case INDEX_op_brcond_i32:
2250    case INDEX_op_brcond_i64:
2251        return C_O0_I2(rz, rz);
2252
2253    case INDEX_op_extu_i32_i64:
2254    case INDEX_op_extrl_i64_i32:
2255    case INDEX_op_extrh_i64_i32:
2256    case INDEX_op_ext_i32_i64:
2257    case INDEX_op_neg_i32:
2258    case INDEX_op_neg_i64:
2259    case INDEX_op_not_i32:
2260    case INDEX_op_not_i64:
2261    case INDEX_op_extract_i32:
2262    case INDEX_op_extract_i64:
2263    case INDEX_op_sextract_i32:
2264    case INDEX_op_sextract_i64:
2265    case INDEX_op_bswap16_i32:
2266    case INDEX_op_bswap16_i64:
2267    case INDEX_op_bswap32_i32:
2268    case INDEX_op_bswap32_i64:
2269    case INDEX_op_bswap64_i64:
2270    case INDEX_op_ld8s_i32:
2271    case INDEX_op_ld8s_i64:
2272    case INDEX_op_ld8u_i32:
2273    case INDEX_op_ld8u_i64:
2274    case INDEX_op_ld16s_i32:
2275    case INDEX_op_ld16s_i64:
2276    case INDEX_op_ld16u_i32:
2277    case INDEX_op_ld16u_i64:
2278    case INDEX_op_ld32s_i64:
2279    case INDEX_op_ld32u_i64:
2280    case INDEX_op_ld_i32:
2281    case INDEX_op_ld_i64:
2282    case INDEX_op_qemu_ld_i32:
2283    case INDEX_op_qemu_ld_i64:
2284        return C_O1_I1(r, r);
2285
2286    case INDEX_op_shl_i32:
2287    case INDEX_op_shl_i64:
2288    case INDEX_op_shr_i32:
2289    case INDEX_op_shr_i64:
2290    case INDEX_op_sar_i32:
2291    case INDEX_op_sar_i64:
2292    case INDEX_op_rotl_i32:
2293    case INDEX_op_rotl_i64:
2294    case INDEX_op_rotr_i32:
2295    case INDEX_op_rotr_i64:
2296        return C_O1_I2(r, r, ri);
2297
2298    case INDEX_op_nor_i32:
2299    case INDEX_op_nor_i64:
2300    case INDEX_op_xor_i32:
2301    case INDEX_op_xor_i64:
2302        /* LoongArch reg-imm bitops have their imms ZERO-extended */
2303        return C_O1_I2(r, r, rU);
2304
2305    case INDEX_op_clz_i32:
2306    case INDEX_op_clz_i64:
2307    case INDEX_op_ctz_i32:
2308    case INDEX_op_ctz_i64:
2309        return C_O1_I2(r, r, rW);
2310
2311    case INDEX_op_deposit_i32:
2312    case INDEX_op_deposit_i64:
2313        /* Must deposit into the same register as input */
2314        return C_O1_I2(r, 0, rz);
2315
2316    case INDEX_op_sub_i32:
2317    case INDEX_op_setcond_i32:
2318        return C_O1_I2(r, rz, ri);
2319    case INDEX_op_sub_i64:
2320    case INDEX_op_setcond_i64:
2321        return C_O1_I2(r, rz, rJ);
2322
2323    case INDEX_op_mul_i32:
2324    case INDEX_op_mul_i64:
2325    case INDEX_op_mulsh_i32:
2326    case INDEX_op_mulsh_i64:
2327    case INDEX_op_muluh_i32:
2328    case INDEX_op_muluh_i64:
2329    case INDEX_op_div_i32:
2330    case INDEX_op_div_i64:
2331    case INDEX_op_divu_i32:
2332    case INDEX_op_divu_i64:
2333    case INDEX_op_rem_i32:
2334    case INDEX_op_rem_i64:
2335    case INDEX_op_remu_i32:
2336    case INDEX_op_remu_i64:
2337        return C_O1_I2(r, rz, rz);
2338
2339    case INDEX_op_movcond_i32:
2340    case INDEX_op_movcond_i64:
2341        return C_O1_I4(r, rz, rJ, rz, rz);
2342
2343    case INDEX_op_ld_vec:
2344    case INDEX_op_dupm_vec:
2345    case INDEX_op_dup_vec:
2346        return C_O1_I1(w, r);
2347
2348    case INDEX_op_st_vec:
2349        return C_O0_I2(w, r);
2350
2351    case INDEX_op_cmp_vec:
2352        return C_O1_I2(w, w, wM);
2353
2354    case INDEX_op_add_vec:
2355    case INDEX_op_sub_vec:
2356        return C_O1_I2(w, w, wA);
2357
2358    case INDEX_op_and_vec:
2359    case INDEX_op_andc_vec:
2360    case INDEX_op_or_vec:
2361    case INDEX_op_orc_vec:
2362    case INDEX_op_xor_vec:
2363    case INDEX_op_nor_vec:
2364    case INDEX_op_mul_vec:
2365    case INDEX_op_smin_vec:
2366    case INDEX_op_smax_vec:
2367    case INDEX_op_umin_vec:
2368    case INDEX_op_umax_vec:
2369    case INDEX_op_ssadd_vec:
2370    case INDEX_op_usadd_vec:
2371    case INDEX_op_sssub_vec:
2372    case INDEX_op_ussub_vec:
2373    case INDEX_op_shlv_vec:
2374    case INDEX_op_shrv_vec:
2375    case INDEX_op_sarv_vec:
2376    case INDEX_op_rotrv_vec:
2377    case INDEX_op_rotlv_vec:
2378        return C_O1_I2(w, w, w);
2379
2380    case INDEX_op_not_vec:
2381    case INDEX_op_neg_vec:
2382    case INDEX_op_shli_vec:
2383    case INDEX_op_shri_vec:
2384    case INDEX_op_sari_vec:
2385    case INDEX_op_rotli_vec:
2386        return C_O1_I1(w, w);
2387
2388    case INDEX_op_bitsel_vec:
2389        return C_O1_I3(w, w, w, w);
2390
2391    default:
2392        return C_NotImplemented;
2393    }
2394}
2395
2396static const int tcg_target_callee_save_regs[] = {
2397    TCG_REG_S0,     /* used for the global env (TCG_AREG0) */
2398    TCG_REG_S1,
2399    TCG_REG_S2,
2400    TCG_REG_S3,
2401    TCG_REG_S4,
2402    TCG_REG_S5,
2403    TCG_REG_S6,
2404    TCG_REG_S7,
2405    TCG_REG_S8,
2406    TCG_REG_S9,
2407    TCG_REG_RA,     /* should be last for ABI compliance */
2408};
2409
2410/* Stack frame parameters.  */
2411#define REG_SIZE   (TCG_TARGET_REG_BITS / 8)
2412#define SAVE_SIZE  ((int)ARRAY_SIZE(tcg_target_callee_save_regs) * REG_SIZE)
2413#define TEMP_SIZE  (CPU_TEMP_BUF_NLONGS * (int)sizeof(long))
2414#define FRAME_SIZE ((TCG_STATIC_CALL_ARGS_SIZE + TEMP_SIZE + SAVE_SIZE \
2415                     + TCG_TARGET_STACK_ALIGN - 1) \
2416                    & -TCG_TARGET_STACK_ALIGN)
2417#define SAVE_OFS   (TCG_STATIC_CALL_ARGS_SIZE + TEMP_SIZE)
2418
2419/* We're expecting to be able to use an immediate for frame allocation.  */
2420QEMU_BUILD_BUG_ON(FRAME_SIZE > 0x7ff);
2421
2422/* Generate global QEMU prologue and epilogue code */
2423static void tcg_target_qemu_prologue(TCGContext *s)
2424{
2425    int i;
2426
2427    tcg_set_frame(s, TCG_REG_SP, TCG_STATIC_CALL_ARGS_SIZE, TEMP_SIZE);
2428
2429    /* TB prologue */
2430    tcg_out_opc_addi_d(s, TCG_REG_SP, TCG_REG_SP, -FRAME_SIZE);
2431    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
2432        tcg_out_st(s, TCG_TYPE_REG, tcg_target_callee_save_regs[i],
2433                   TCG_REG_SP, SAVE_OFS + i * REG_SIZE);
2434    }
2435
2436    if (!tcg_use_softmmu && guest_base) {
2437        tcg_out_movi(s, TCG_TYPE_PTR, TCG_GUEST_BASE_REG, guest_base);
2438        tcg_regset_set_reg(s->reserved_regs, TCG_GUEST_BASE_REG);
2439    }
2440
2441    /* Call generated code */
2442    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
2443    tcg_out_opc_jirl(s, TCG_REG_ZERO, tcg_target_call_iarg_regs[1], 0);
2444
2445    /* Return path for goto_ptr. Set return value to 0 */
2446    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
2447    tcg_out_mov(s, TCG_TYPE_REG, TCG_REG_A0, TCG_REG_ZERO);
2448
2449    /* TB epilogue */
2450    tb_ret_addr = tcg_splitwx_to_rx(s->code_ptr);
2451    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
2452        tcg_out_ld(s, TCG_TYPE_REG, tcg_target_callee_save_regs[i],
2453                   TCG_REG_SP, SAVE_OFS + i * REG_SIZE);
2454    }
2455
2456    tcg_out_opc_addi_d(s, TCG_REG_SP, TCG_REG_SP, FRAME_SIZE);
2457    tcg_out_opc_jirl(s, TCG_REG_ZERO, TCG_REG_RA, 0);
2458}
2459
2460static void tcg_out_tb_start(TCGContext *s)
2461{
2462    /* nothing to do */
2463}
2464
2465static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
2466{
2467    for (int i = 0; i < count; ++i) {
2468        /* Canonical nop is andi r0,r0,0 */
2469        p[i] = OPC_ANDI;
2470    }
2471}
2472
2473static void tcg_target_init(TCGContext *s)
2474{
2475    unsigned long hwcap = qemu_getauxval(AT_HWCAP);
2476
2477    /* Server and desktop class cpus have UAL; embedded cpus do not. */
2478    if (!(hwcap & HWCAP_LOONGARCH_UAL)) {
2479        error_report("TCG: unaligned access support required; exiting");
2480        exit(EXIT_FAILURE);
2481    }
2482
2483    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
2484    tcg_target_available_regs[TCG_TYPE_I64] = ALL_GENERAL_REGS;
2485
2486    tcg_target_call_clobber_regs = ALL_GENERAL_REGS | ALL_VECTOR_REGS;
2487    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S0);
2488    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S1);
2489    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S2);
2490    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S3);
2491    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S4);
2492    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S5);
2493    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S6);
2494    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S7);
2495    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S8);
2496    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S9);
2497
2498    if (cpuinfo & CPUINFO_LSX) {
2499        tcg_target_available_regs[TCG_TYPE_V64] = ALL_VECTOR_REGS;
2500        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
2501        if (cpuinfo & CPUINFO_LASX) {
2502            tcg_target_available_regs[TCG_TYPE_V256] = ALL_VECTOR_REGS;
2503        }
2504        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V24);
2505        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V25);
2506        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V26);
2507        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V27);
2508        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V28);
2509        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V29);
2510        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V30);
2511        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V31);
2512    }
2513
2514    s->reserved_regs = 0;
2515    tcg_regset_set_reg(s->reserved_regs, TCG_REG_ZERO);
2516    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP0);
2517    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP1);
2518    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP2);
2519    tcg_regset_set_reg(s->reserved_regs, TCG_REG_SP);
2520    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TP);
2521    tcg_regset_set_reg(s->reserved_regs, TCG_REG_RESERVED);
2522    tcg_regset_set_reg(s->reserved_regs, TCG_VEC_TMP0);
2523}
2524
2525typedef struct {
2526    DebugFrameHeader h;
2527    uint8_t fde_def_cfa[4];
2528    uint8_t fde_reg_ofs[ARRAY_SIZE(tcg_target_callee_save_regs) * 2];
2529} DebugFrame;
2530
2531#define ELF_HOST_MACHINE EM_LOONGARCH
2532
2533static const DebugFrame debug_frame = {
2534    .h.cie.len = sizeof(DebugFrameCIE) - 4, /* length after .len member */
2535    .h.cie.id = -1,
2536    .h.cie.version = 1,
2537    .h.cie.code_align = 1,
2538    .h.cie.data_align = -(TCG_TARGET_REG_BITS / 8) & 0x7f, /* sleb128 */
2539    .h.cie.return_column = TCG_REG_RA,
2540
2541    /* Total FDE size does not include the "len" member.  */
2542    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
2543
2544    .fde_def_cfa = {
2545        12, TCG_REG_SP,                 /* DW_CFA_def_cfa sp, ...  */
2546        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
2547        (FRAME_SIZE >> 7)
2548    },
2549    .fde_reg_ofs = {
2550        0x80 + 23, 11,                  /* DW_CFA_offset, s0, -88 */
2551        0x80 + 24, 10,                  /* DW_CFA_offset, s1, -80 */
2552        0x80 + 25, 9,                   /* DW_CFA_offset, s2, -72 */
2553        0x80 + 26, 8,                   /* DW_CFA_offset, s3, -64 */
2554        0x80 + 27, 7,                   /* DW_CFA_offset, s4, -56 */
2555        0x80 + 28, 6,                   /* DW_CFA_offset, s5, -48 */
2556        0x80 + 29, 5,                   /* DW_CFA_offset, s6, -40 */
2557        0x80 + 30, 4,                   /* DW_CFA_offset, s7, -32 */
2558        0x80 + 31, 3,                   /* DW_CFA_offset, s8, -24 */
2559        0x80 + 22, 2,                   /* DW_CFA_offset, s9, -16 */
2560        0x80 + 1 , 1,                   /* DW_CFA_offset, ra, -8 */
2561    }
2562};
2563
2564void tcg_register_jit(const void *buf, size_t buf_size)
2565{
2566    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
2567}
2568