xref: /openbmc/qemu/tcg/loongarch64/tcg-target.c.inc (revision ebe92db2ccf791208c58d15652f9460a67cc0fdd)
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
180#define ALL_GENERAL_REGS   MAKE_64BIT_MASK(0, 32)
181#define ALL_VECTOR_REGS    MAKE_64BIT_MASK(32, 32)
182
183static inline tcg_target_long sextreg(tcg_target_long val, int pos, int len)
184{
185    return sextract64(val, pos, len);
186}
187
188/* test if a constant matches the constraint */
189static bool tcg_target_const_match(int64_t val, TCGType type, int ct, int vece)
190{
191    if (ct & TCG_CT_CONST) {
192        return true;
193    }
194    if ((ct & TCG_CT_CONST_ZERO) && val == 0) {
195        return true;
196    }
197    if ((ct & TCG_CT_CONST_S12) && val == sextreg(val, 0, 12)) {
198        return true;
199    }
200    if ((ct & TCG_CT_CONST_S32) && val == (int32_t)val) {
201        return true;
202    }
203    if ((ct & TCG_CT_CONST_U12) && val >= 0 && val <= 0xfff) {
204        return true;
205    }
206    if ((ct & TCG_CT_CONST_C12) && ~val >= 0 && ~val <= 0xfff) {
207        return true;
208    }
209    if ((ct & TCG_CT_CONST_WSZ) && val == (type == TCG_TYPE_I32 ? 32 : 64)) {
210        return true;
211    }
212    return false;
213}
214
215/*
216 * Relocations
217 */
218
219/*
220 * Relocation records defined in LoongArch ELF psABI v1.00 is way too
221 * complicated; a whopping stack machine is needed to stuff the fields, at
222 * the very least one SOP_PUSH and one SOP_POP (of the correct format) are
223 * needed.
224 *
225 * Hence, define our own simpler relocation types. Numbers are chosen as to
226 * not collide with potential future additions to the true ELF relocation
227 * type enum.
228 */
229
230/* Field Sk16, shifted right by 2; suitable for conditional jumps */
231#define R_LOONGARCH_BR_SK16     256
232/* Field Sd10k16, shifted right by 2; suitable for B and BL */
233#define R_LOONGARCH_BR_SD10K16  257
234
235static bool reloc_br_sk16(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
236{
237    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
238    intptr_t offset = (intptr_t)target - (intptr_t)src_rx;
239
240    tcg_debug_assert((offset & 3) == 0);
241    offset >>= 2;
242    if (offset == sextreg(offset, 0, 16)) {
243        *src_rw = deposit64(*src_rw, 10, 16, offset);
244        return true;
245    }
246
247    return false;
248}
249
250static bool reloc_br_sd10k16(tcg_insn_unit *src_rw,
251                             const tcg_insn_unit *target)
252{
253    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
254    intptr_t offset = (intptr_t)target - (intptr_t)src_rx;
255
256    tcg_debug_assert((offset & 3) == 0);
257    offset >>= 2;
258    if (offset == sextreg(offset, 0, 26)) {
259        *src_rw = deposit64(*src_rw, 0, 10, offset >> 16); /* slot d10 */
260        *src_rw = deposit64(*src_rw, 10, 16, offset); /* slot k16 */
261        return true;
262    }
263
264    return false;
265}
266
267static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
268                        intptr_t value, intptr_t addend)
269{
270    tcg_debug_assert(addend == 0);
271    switch (type) {
272    case R_LOONGARCH_BR_SK16:
273        return reloc_br_sk16(code_ptr, (tcg_insn_unit *)value);
274    case R_LOONGARCH_BR_SD10K16:
275        return reloc_br_sd10k16(code_ptr, (tcg_insn_unit *)value);
276    default:
277        g_assert_not_reached();
278    }
279}
280
281#include "tcg-insn-defs.c.inc"
282
283/*
284 * TCG intrinsics
285 */
286
287static void tcg_out_mb(TCGContext *s, TCGArg a0)
288{
289    /* Baseline LoongArch only has the full barrier, unfortunately.  */
290    tcg_out_opc_dbar(s, 0);
291}
292
293static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
294{
295    if (ret == arg) {
296        return true;
297    }
298    switch (type) {
299    case TCG_TYPE_I32:
300    case TCG_TYPE_I64:
301        /*
302         * Conventional register-register move used in LoongArch is
303         * `or dst, src, zero`.
304         */
305        tcg_out_opc_or(s, ret, arg, TCG_REG_ZERO);
306        break;
307    default:
308        g_assert_not_reached();
309    }
310    return true;
311}
312
313/* Loads a 32-bit immediate into rd, sign-extended.  */
314static void tcg_out_movi_i32(TCGContext *s, TCGReg rd, int32_t val)
315{
316    tcg_target_long lo = sextreg(val, 0, 12);
317    tcg_target_long hi12 = sextreg(val, 12, 20);
318
319    /* Single-instruction cases.  */
320    if (hi12 == 0) {
321        /* val fits in uimm12: ori rd, zero, val */
322        tcg_out_opc_ori(s, rd, TCG_REG_ZERO, val);
323        return;
324    }
325    if (hi12 == sextreg(lo, 12, 20)) {
326        /* val fits in simm12: addi.w rd, zero, val */
327        tcg_out_opc_addi_w(s, rd, TCG_REG_ZERO, val);
328        return;
329    }
330
331    /* High bits must be set; load with lu12i.w + optional ori.  */
332    tcg_out_opc_lu12i_w(s, rd, hi12);
333    if (lo != 0) {
334        tcg_out_opc_ori(s, rd, rd, lo & 0xfff);
335    }
336}
337
338static void tcg_out_movi(TCGContext *s, TCGType type, TCGReg rd,
339                         tcg_target_long val)
340{
341    /*
342     * LoongArch conventionally loads 64-bit immediates in at most 4 steps,
343     * with dedicated instructions for filling the respective bitfields
344     * below:
345     *
346     *        6                   5                   4               3
347     *  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
348     * +-----------------------+---------------------------------------+...
349     * |          hi52         |                  hi32                 |
350     * +-----------------------+---------------------------------------+...
351     *       3                   2                   1
352     *     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
353     * ...+-------------------------------------+-------------------------+
354     *    |                 hi12                |            lo           |
355     * ...+-------------------------------------+-------------------------+
356     *
357     * Check if val belong to one of the several fast cases, before falling
358     * back to the slow path.
359     */
360
361    intptr_t pc_offset;
362    tcg_target_long val_lo, val_hi, pc_hi, offset_hi;
363    tcg_target_long hi12, hi32, hi52;
364
365    /* Value fits in signed i32.  */
366    if (type == TCG_TYPE_I32 || val == (int32_t)val) {
367        tcg_out_movi_i32(s, rd, val);
368        return;
369    }
370
371    /* PC-relative cases.  */
372    pc_offset = tcg_pcrel_diff(s, (void *)val);
373    if (pc_offset == sextreg(pc_offset, 0, 22) && (pc_offset & 3) == 0) {
374        /* Single pcaddu2i.  */
375        tcg_out_opc_pcaddu2i(s, rd, pc_offset >> 2);
376        return;
377    }
378
379    if (pc_offset == (int32_t)pc_offset) {
380        /* Offset within 32 bits; load with pcalau12i + ori.  */
381        val_lo = sextreg(val, 0, 12);
382        val_hi = val >> 12;
383        pc_hi = (val - pc_offset) >> 12;
384        offset_hi = val_hi - pc_hi;
385
386        tcg_debug_assert(offset_hi == sextreg(offset_hi, 0, 20));
387        tcg_out_opc_pcalau12i(s, rd, offset_hi);
388        if (val_lo != 0) {
389            tcg_out_opc_ori(s, rd, rd, val_lo & 0xfff);
390        }
391        return;
392    }
393
394    hi12 = sextreg(val, 12, 20);
395    hi32 = sextreg(val, 32, 20);
396    hi52 = sextreg(val, 52, 12);
397
398    /* Single cu52i.d case.  */
399    if ((hi52 != 0) && (ctz64(val) >= 52)) {
400        tcg_out_opc_cu52i_d(s, rd, TCG_REG_ZERO, hi52);
401        return;
402    }
403
404    /* Slow path.  Initialize the low 32 bits, then concat high bits.  */
405    tcg_out_movi_i32(s, rd, val);
406
407    /* Load hi32 and hi52 explicitly when they are unexpected values. */
408    if (hi32 != sextreg(hi12, 20, 20)) {
409        tcg_out_opc_cu32i_d(s, rd, hi32);
410    }
411
412    if (hi52 != sextreg(hi32, 20, 12)) {
413        tcg_out_opc_cu52i_d(s, rd, rd, hi52);
414    }
415}
416
417static void tcg_out_addi(TCGContext *s, TCGType type, TCGReg rd,
418                         TCGReg rs, tcg_target_long imm)
419{
420    tcg_target_long lo12 = sextreg(imm, 0, 12);
421    tcg_target_long hi16 = sextreg(imm - lo12, 16, 16);
422
423    /*
424     * Note that there's a hole in between hi16 and lo12:
425     *
426     *       3                   2                   1                   0
427     *     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
428     * ...+-------------------------------+-------+-----------------------+
429     *    |             hi16              |       |          lo12         |
430     * ...+-------------------------------+-------+-----------------------+
431     *
432     * For bits within that hole, it's more efficient to use LU12I and ADD.
433     */
434    if (imm == (hi16 << 16) + lo12) {
435        if (hi16) {
436            tcg_out_opc_addu16i_d(s, rd, rs, hi16);
437            rs = rd;
438        }
439        if (type == TCG_TYPE_I32) {
440            tcg_out_opc_addi_w(s, rd, rs, lo12);
441        } else if (lo12) {
442            tcg_out_opc_addi_d(s, rd, rs, lo12);
443        } else {
444            tcg_out_mov(s, type, rd, rs);
445        }
446    } else {
447        tcg_out_movi(s, type, TCG_REG_TMP0, imm);
448        if (type == TCG_TYPE_I32) {
449            tcg_out_opc_add_w(s, rd, rs, TCG_REG_TMP0);
450        } else {
451            tcg_out_opc_add_d(s, rd, rs, TCG_REG_TMP0);
452        }
453    }
454}
455
456static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
457{
458    return false;
459}
460
461static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
462                             tcg_target_long imm)
463{
464    /* This function is only used for passing structs by reference. */
465    g_assert_not_reached();
466}
467
468static void tcg_out_ext8u(TCGContext *s, TCGReg ret, TCGReg arg)
469{
470    tcg_out_opc_andi(s, ret, arg, 0xff);
471}
472
473static void tcg_out_ext16u(TCGContext *s, TCGReg ret, TCGReg arg)
474{
475    tcg_out_opc_bstrpick_w(s, ret, arg, 0, 15);
476}
477
478static void tcg_out_ext32u(TCGContext *s, TCGReg ret, TCGReg arg)
479{
480    tcg_out_opc_bstrpick_d(s, ret, arg, 0, 31);
481}
482
483static void tcg_out_ext8s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
484{
485    tcg_out_opc_sext_b(s, ret, arg);
486}
487
488static void tcg_out_ext16s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
489{
490    tcg_out_opc_sext_h(s, ret, arg);
491}
492
493static void tcg_out_ext32s(TCGContext *s, TCGReg ret, TCGReg arg)
494{
495    tcg_out_opc_addi_w(s, ret, arg, 0);
496}
497
498static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg)
499{
500    if (ret != arg) {
501        tcg_out_ext32s(s, ret, arg);
502    }
503}
504
505static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg)
506{
507    tcg_out_ext32u(s, ret, arg);
508}
509
510static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg ret, TCGReg arg)
511{
512    tcg_out_ext32s(s, ret, arg);
513}
514
515static void tcg_out_clzctz(TCGContext *s, LoongArchInsn opc,
516                           TCGReg a0, TCGReg a1, TCGReg a2,
517                           bool c2, bool is_32bit)
518{
519    if (c2) {
520        /*
521         * Fast path: semantics already satisfied due to constraint and
522         * insn behavior, single instruction is enough.
523         */
524        tcg_debug_assert(a2 == (is_32bit ? 32 : 64));
525        /* all clz/ctz insns belong to DJ-format */
526        tcg_out32(s, encode_dj_insn(opc, a0, a1));
527        return;
528    }
529
530    tcg_out32(s, encode_dj_insn(opc, TCG_REG_TMP0, a1));
531    /* a0 = a1 ? REG_TMP0 : a2 */
532    tcg_out_opc_maskeqz(s, TCG_REG_TMP0, TCG_REG_TMP0, a1);
533    tcg_out_opc_masknez(s, a0, a2, a1);
534    tcg_out_opc_or(s, a0, TCG_REG_TMP0, a0);
535}
536
537#define SETCOND_INV    TCG_TARGET_NB_REGS
538#define SETCOND_NEZ    (SETCOND_INV << 1)
539#define SETCOND_FLAGS  (SETCOND_INV | SETCOND_NEZ)
540
541static int tcg_out_setcond_int(TCGContext *s, TCGCond cond, TCGReg ret,
542                               TCGReg arg1, tcg_target_long arg2, bool c2)
543{
544    int flags = 0;
545
546    switch (cond) {
547    case TCG_COND_EQ:    /* -> NE  */
548    case TCG_COND_GE:    /* -> LT  */
549    case TCG_COND_GEU:   /* -> LTU */
550    case TCG_COND_GT:    /* -> LE  */
551    case TCG_COND_GTU:   /* -> LEU */
552        cond = tcg_invert_cond(cond);
553        flags ^= SETCOND_INV;
554        break;
555    default:
556        break;
557    }
558
559    switch (cond) {
560    case TCG_COND_LE:
561    case TCG_COND_LEU:
562        /*
563         * If we have a constant input, the most efficient way to implement
564         * LE is by adding 1 and using LT.  Watch out for wrap around for LEU.
565         * We don't need to care for this for LE because the constant input
566         * is still constrained to int32_t, and INT32_MAX+1 is representable
567         * in the 64-bit temporary register.
568         */
569        if (c2) {
570            if (cond == TCG_COND_LEU) {
571                /* unsigned <= -1 is true */
572                if (arg2 == -1) {
573                    tcg_out_movi(s, TCG_TYPE_REG, ret, !(flags & SETCOND_INV));
574                    return ret;
575                }
576                cond = TCG_COND_LTU;
577            } else {
578                cond = TCG_COND_LT;
579            }
580            arg2 += 1;
581        } else {
582            TCGReg tmp = arg2;
583            arg2 = arg1;
584            arg1 = tmp;
585            cond = tcg_swap_cond(cond);    /* LE -> GE */
586            cond = tcg_invert_cond(cond);  /* GE -> LT */
587            flags ^= SETCOND_INV;
588        }
589        break;
590    default:
591        break;
592    }
593
594    switch (cond) {
595    case TCG_COND_NE:
596        flags |= SETCOND_NEZ;
597        if (!c2) {
598            tcg_out_opc_xor(s, ret, arg1, arg2);
599        } else if (arg2 == 0) {
600            ret = arg1;
601        } else if (arg2 >= 0 && arg2 <= 0xfff) {
602            tcg_out_opc_xori(s, ret, arg1, arg2);
603        } else {
604            tcg_out_addi(s, TCG_TYPE_REG, ret, arg1, -arg2);
605        }
606        break;
607
608    case TCG_COND_LT:
609    case TCG_COND_LTU:
610        if (c2) {
611            if (arg2 >= -0x800 && arg2 <= 0x7ff) {
612                if (cond == TCG_COND_LT) {
613                    tcg_out_opc_slti(s, ret, arg1, arg2);
614                } else {
615                    tcg_out_opc_sltui(s, ret, arg1, arg2);
616                }
617                break;
618            }
619            tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_TMP0, arg2);
620            arg2 = TCG_REG_TMP0;
621        }
622        if (cond == TCG_COND_LT) {
623            tcg_out_opc_slt(s, ret, arg1, arg2);
624        } else {
625            tcg_out_opc_sltu(s, ret, arg1, arg2);
626        }
627        break;
628
629    default:
630        g_assert_not_reached();
631        break;
632    }
633
634    return ret | flags;
635}
636
637static void tcg_out_setcond(TCGContext *s, TCGCond cond, TCGReg ret,
638                            TCGReg arg1, tcg_target_long arg2, bool c2)
639{
640    int tmpflags = tcg_out_setcond_int(s, cond, ret, arg1, arg2, c2);
641
642    if (tmpflags != ret) {
643        TCGReg tmp = tmpflags & ~SETCOND_FLAGS;
644
645        switch (tmpflags & SETCOND_FLAGS) {
646        case SETCOND_INV:
647            /* Intermediate result is boolean: simply invert. */
648            tcg_out_opc_xori(s, ret, tmp, 1);
649            break;
650        case SETCOND_NEZ:
651            /* Intermediate result is zero/non-zero: test != 0. */
652            tcg_out_opc_sltu(s, ret, TCG_REG_ZERO, tmp);
653            break;
654        case SETCOND_NEZ | SETCOND_INV:
655            /* Intermediate result is zero/non-zero: test == 0. */
656            tcg_out_opc_sltui(s, ret, tmp, 1);
657            break;
658        default:
659            g_assert_not_reached();
660        }
661    }
662}
663
664static void tcg_out_movcond(TCGContext *s, TCGCond cond, TCGReg ret,
665                            TCGReg c1, tcg_target_long c2, bool const2,
666                            TCGReg v1, TCGReg v2)
667{
668    int tmpflags = tcg_out_setcond_int(s, cond, TCG_REG_TMP0, c1, c2, const2);
669    TCGReg t;
670
671    /* Standardize the test below to t != 0. */
672    if (tmpflags & SETCOND_INV) {
673        t = v1, v1 = v2, v2 = t;
674    }
675
676    t = tmpflags & ~SETCOND_FLAGS;
677    if (v1 == TCG_REG_ZERO) {
678        tcg_out_opc_masknez(s, ret, v2, t);
679    } else if (v2 == TCG_REG_ZERO) {
680        tcg_out_opc_maskeqz(s, ret, v1, t);
681    } else {
682        tcg_out_opc_masknez(s, TCG_REG_TMP2, v2, t); /* t ? 0 : v2 */
683        tcg_out_opc_maskeqz(s, TCG_REG_TMP1, v1, t); /* t ? v1 : 0 */
684        tcg_out_opc_or(s, ret, TCG_REG_TMP1, TCG_REG_TMP2);
685    }
686}
687
688/*
689 * Branch helpers
690 */
691
692static const struct {
693    LoongArchInsn op;
694    bool swap;
695} tcg_brcond_to_loongarch[] = {
696    [TCG_COND_EQ] =  { OPC_BEQ,  false },
697    [TCG_COND_NE] =  { OPC_BNE,  false },
698    [TCG_COND_LT] =  { OPC_BGT,  true  },
699    [TCG_COND_GE] =  { OPC_BLE,  true  },
700    [TCG_COND_LE] =  { OPC_BLE,  false },
701    [TCG_COND_GT] =  { OPC_BGT,  false },
702    [TCG_COND_LTU] = { OPC_BGTU, true  },
703    [TCG_COND_GEU] = { OPC_BLEU, true  },
704    [TCG_COND_LEU] = { OPC_BLEU, false },
705    [TCG_COND_GTU] = { OPC_BGTU, false }
706};
707
708static void tcg_out_brcond(TCGContext *s, TCGCond cond, TCGReg arg1,
709                           TCGReg arg2, TCGLabel *l)
710{
711    LoongArchInsn op = tcg_brcond_to_loongarch[cond].op;
712
713    tcg_debug_assert(op != 0);
714
715    if (tcg_brcond_to_loongarch[cond].swap) {
716        TCGReg t = arg1;
717        arg1 = arg2;
718        arg2 = t;
719    }
720
721    /* all conditional branch insns belong to DJSk16-format */
722    tcg_out_reloc(s, s->code_ptr, R_LOONGARCH_BR_SK16, l, 0);
723    tcg_out32(s, encode_djsk16_insn(op, arg1, arg2, 0));
724}
725
726static void tcg_out_call_int(TCGContext *s, const tcg_insn_unit *arg, bool tail)
727{
728    TCGReg link = tail ? TCG_REG_ZERO : TCG_REG_RA;
729    ptrdiff_t offset = tcg_pcrel_diff(s, arg);
730
731    tcg_debug_assert((offset & 3) == 0);
732    if (offset == sextreg(offset, 0, 28)) {
733        /* short jump: +/- 256MiB */
734        if (tail) {
735            tcg_out_opc_b(s, offset >> 2);
736        } else {
737            tcg_out_opc_bl(s, offset >> 2);
738        }
739    } else if (offset == sextreg(offset, 0, 38)) {
740        /* long jump: +/- 256GiB */
741        tcg_target_long lo = sextreg(offset, 0, 18);
742        tcg_target_long hi = offset - lo;
743        tcg_out_opc_pcaddu18i(s, TCG_REG_TMP0, hi >> 18);
744        tcg_out_opc_jirl(s, link, TCG_REG_TMP0, lo >> 2);
745    } else {
746        /* far jump: 64-bit */
747        tcg_target_long lo = sextreg((tcg_target_long)arg, 0, 18);
748        tcg_target_long hi = (tcg_target_long)arg - lo;
749        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, hi);
750        tcg_out_opc_jirl(s, link, TCG_REG_TMP0, lo >> 2);
751    }
752}
753
754static void tcg_out_call(TCGContext *s, const tcg_insn_unit *arg,
755                         const TCGHelperInfo *info)
756{
757    tcg_out_call_int(s, arg, false);
758}
759
760/*
761 * Load/store helpers
762 */
763
764static void tcg_out_ldst(TCGContext *s, LoongArchInsn opc, TCGReg data,
765                         TCGReg addr, intptr_t offset)
766{
767    intptr_t imm12 = sextreg(offset, 0, 12);
768
769    if (offset != imm12) {
770        intptr_t diff = tcg_pcrel_diff(s, (void *)offset);
771
772        if (addr == TCG_REG_ZERO && diff == (int32_t)diff) {
773            imm12 = sextreg(diff, 0, 12);
774            tcg_out_opc_pcaddu12i(s, TCG_REG_TMP2, (diff - imm12) >> 12);
775        } else {
776            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP2, offset - imm12);
777            if (addr != TCG_REG_ZERO) {
778                tcg_out_opc_add_d(s, TCG_REG_TMP2, TCG_REG_TMP2, addr);
779            }
780        }
781        addr = TCG_REG_TMP2;
782    }
783
784    switch (opc) {
785    case OPC_LD_B:
786    case OPC_LD_BU:
787    case OPC_LD_H:
788    case OPC_LD_HU:
789    case OPC_LD_W:
790    case OPC_LD_WU:
791    case OPC_LD_D:
792    case OPC_ST_B:
793    case OPC_ST_H:
794    case OPC_ST_W:
795    case OPC_ST_D:
796        tcg_out32(s, encode_djsk12_insn(opc, data, addr, imm12));
797        break;
798    default:
799        g_assert_not_reached();
800    }
801}
802
803static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg arg,
804                       TCGReg arg1, intptr_t arg2)
805{
806    bool is_32bit = type == TCG_TYPE_I32;
807    tcg_out_ldst(s, is_32bit ? OPC_LD_W : OPC_LD_D, arg, arg1, arg2);
808}
809
810static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
811                       TCGReg arg1, intptr_t arg2)
812{
813    bool is_32bit = type == TCG_TYPE_I32;
814    tcg_out_ldst(s, is_32bit ? OPC_ST_W : OPC_ST_D, arg, arg1, arg2);
815}
816
817static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
818                        TCGReg base, intptr_t ofs)
819{
820    if (val == 0) {
821        tcg_out_st(s, type, TCG_REG_ZERO, base, ofs);
822        return true;
823    }
824    return false;
825}
826
827/*
828 * Load/store helpers for SoftMMU, and qemu_ld/st implementations
829 */
830
831static bool tcg_out_goto(TCGContext *s, const tcg_insn_unit *target)
832{
833    tcg_out_opc_b(s, 0);
834    return reloc_br_sd10k16(s->code_ptr - 1, target);
835}
836
837static const TCGLdstHelperParam ldst_helper_param = {
838    .ntmp = 1, .tmp = { TCG_REG_TMP0 }
839};
840
841static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
842{
843    MemOp opc = get_memop(l->oi);
844
845    /* resolve label address */
846    if (!reloc_br_sk16(l->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
847        return false;
848    }
849
850    tcg_out_ld_helper_args(s, l, &ldst_helper_param);
851    tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SIZE], false);
852    tcg_out_ld_helper_ret(s, l, false, &ldst_helper_param);
853    return tcg_out_goto(s, l->raddr);
854}
855
856static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
857{
858    MemOp opc = get_memop(l->oi);
859
860    /* resolve label address */
861    if (!reloc_br_sk16(l->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
862        return false;
863    }
864
865    tcg_out_st_helper_args(s, l, &ldst_helper_param);
866    tcg_out_call_int(s, qemu_st_helpers[opc & MO_SIZE], false);
867    return tcg_out_goto(s, l->raddr);
868}
869
870typedef struct {
871    TCGReg base;
872    TCGReg index;
873    TCGAtomAlign aa;
874} HostAddress;
875
876bool tcg_target_has_memory_bswap(MemOp memop)
877{
878    return false;
879}
880
881/* We expect to use a 12-bit negative offset from ENV.  */
882#define MIN_TLB_MASK_TABLE_OFS  -(1 << 11)
883
884/*
885 * For softmmu, perform the TLB load and compare.
886 * For useronly, perform any required alignment tests.
887 * In both cases, return a TCGLabelQemuLdst structure if the slow path
888 * is required and fill in @h with the host address for the fast path.
889 */
890static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
891                                           TCGReg addr_reg, MemOpIdx oi,
892                                           bool is_ld)
893{
894    TCGType addr_type = s->addr_type;
895    TCGLabelQemuLdst *ldst = NULL;
896    MemOp opc = get_memop(oi);
897    MemOp a_bits;
898
899    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, false);
900    a_bits = h->aa.align;
901
902#ifdef CONFIG_SOFTMMU
903    unsigned s_bits = opc & MO_SIZE;
904    int mem_index = get_mmuidx(oi);
905    int fast_ofs = tlb_mask_table_ofs(s, mem_index);
906    int mask_ofs = fast_ofs + offsetof(CPUTLBDescFast, mask);
907    int table_ofs = fast_ofs + offsetof(CPUTLBDescFast, table);
908
909    ldst = new_ldst_label(s);
910    ldst->is_ld = is_ld;
911    ldst->oi = oi;
912    ldst->addrlo_reg = addr_reg;
913
914    tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP0, TCG_AREG0, mask_ofs);
915    tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP1, TCG_AREG0, table_ofs);
916
917    tcg_out_opc_srli_d(s, TCG_REG_TMP2, addr_reg,
918                    s->page_bits - CPU_TLB_ENTRY_BITS);
919    tcg_out_opc_and(s, TCG_REG_TMP2, TCG_REG_TMP2, TCG_REG_TMP0);
920    tcg_out_opc_add_d(s, TCG_REG_TMP2, TCG_REG_TMP2, TCG_REG_TMP1);
921
922    /* Load the tlb comparator and the addend.  */
923    QEMU_BUILD_BUG_ON(HOST_BIG_ENDIAN);
924    tcg_out_ld(s, addr_type, TCG_REG_TMP0, TCG_REG_TMP2,
925               is_ld ? offsetof(CPUTLBEntry, addr_read)
926                     : offsetof(CPUTLBEntry, addr_write));
927    tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP2, TCG_REG_TMP2,
928               offsetof(CPUTLBEntry, addend));
929
930    /*
931     * For aligned accesses, we check the first byte and include the alignment
932     * bits within the address.  For unaligned access, we check that we don't
933     * cross pages using the address of the last byte of the access.
934     */
935    if (a_bits < s_bits) {
936        unsigned a_mask = (1u << a_bits) - 1;
937        unsigned s_mask = (1u << s_bits) - 1;
938        tcg_out_addi(s, addr_type, TCG_REG_TMP1, addr_reg, s_mask - a_mask);
939    } else {
940        tcg_out_mov(s, addr_type, TCG_REG_TMP1, addr_reg);
941    }
942    tcg_out_opc_bstrins_d(s, TCG_REG_TMP1, TCG_REG_ZERO,
943                          a_bits, s->page_bits - 1);
944
945    /* Compare masked address with the TLB entry.  */
946    ldst->label_ptr[0] = s->code_ptr;
947    tcg_out_opc_bne(s, TCG_REG_TMP0, TCG_REG_TMP1, 0);
948
949    h->index = TCG_REG_TMP2;
950#else
951    if (a_bits) {
952        ldst = new_ldst_label(s);
953
954        ldst->is_ld = is_ld;
955        ldst->oi = oi;
956        ldst->addrlo_reg = addr_reg;
957
958        /*
959         * Without micro-architecture details, we don't know which of
960         * bstrpick or andi is faster, so use bstrpick as it's not
961         * constrained by imm field width. Not to say alignments >= 2^12
962         * are going to happen any time soon.
963         */
964        tcg_out_opc_bstrpick_d(s, TCG_REG_TMP1, addr_reg, 0, a_bits - 1);
965
966        ldst->label_ptr[0] = s->code_ptr;
967        tcg_out_opc_bne(s, TCG_REG_TMP1, TCG_REG_ZERO, 0);
968    }
969
970    h->index = USE_GUEST_BASE ? TCG_GUEST_BASE_REG : TCG_REG_ZERO;
971#endif
972
973    if (addr_type == TCG_TYPE_I32) {
974        h->base = TCG_REG_TMP0;
975        tcg_out_ext32u(s, h->base, addr_reg);
976    } else {
977        h->base = addr_reg;
978    }
979
980    return ldst;
981}
982
983static void tcg_out_qemu_ld_indexed(TCGContext *s, MemOp opc, TCGType type,
984                                    TCGReg rd, HostAddress h)
985{
986    /* Byte swapping is left to middle-end expansion.  */
987    tcg_debug_assert((opc & MO_BSWAP) == 0);
988
989    switch (opc & MO_SSIZE) {
990    case MO_UB:
991        tcg_out_opc_ldx_bu(s, rd, h.base, h.index);
992        break;
993    case MO_SB:
994        tcg_out_opc_ldx_b(s, rd, h.base, h.index);
995        break;
996    case MO_UW:
997        tcg_out_opc_ldx_hu(s, rd, h.base, h.index);
998        break;
999    case MO_SW:
1000        tcg_out_opc_ldx_h(s, rd, h.base, h.index);
1001        break;
1002    case MO_UL:
1003        if (type == TCG_TYPE_I64) {
1004            tcg_out_opc_ldx_wu(s, rd, h.base, h.index);
1005            break;
1006        }
1007        /* fallthrough */
1008    case MO_SL:
1009        tcg_out_opc_ldx_w(s, rd, h.base, h.index);
1010        break;
1011    case MO_UQ:
1012        tcg_out_opc_ldx_d(s, rd, h.base, h.index);
1013        break;
1014    default:
1015        g_assert_not_reached();
1016    }
1017}
1018
1019static void tcg_out_qemu_ld(TCGContext *s, TCGReg data_reg, TCGReg addr_reg,
1020                            MemOpIdx oi, TCGType data_type)
1021{
1022    TCGLabelQemuLdst *ldst;
1023    HostAddress h;
1024
1025    ldst = prepare_host_addr(s, &h, addr_reg, oi, true);
1026    tcg_out_qemu_ld_indexed(s, get_memop(oi), data_type, data_reg, h);
1027
1028    if (ldst) {
1029        ldst->type = data_type;
1030        ldst->datalo_reg = data_reg;
1031        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1032    }
1033}
1034
1035static void tcg_out_qemu_st_indexed(TCGContext *s, MemOp opc,
1036                                    TCGReg rd, HostAddress h)
1037{
1038    /* Byte swapping is left to middle-end expansion.  */
1039    tcg_debug_assert((opc & MO_BSWAP) == 0);
1040
1041    switch (opc & MO_SIZE) {
1042    case MO_8:
1043        tcg_out_opc_stx_b(s, rd, h.base, h.index);
1044        break;
1045    case MO_16:
1046        tcg_out_opc_stx_h(s, rd, h.base, h.index);
1047        break;
1048    case MO_32:
1049        tcg_out_opc_stx_w(s, rd, h.base, h.index);
1050        break;
1051    case MO_64:
1052        tcg_out_opc_stx_d(s, rd, h.base, h.index);
1053        break;
1054    default:
1055        g_assert_not_reached();
1056    }
1057}
1058
1059static void tcg_out_qemu_st(TCGContext *s, TCGReg data_reg, TCGReg addr_reg,
1060                            MemOpIdx oi, TCGType data_type)
1061{
1062    TCGLabelQemuLdst *ldst;
1063    HostAddress h;
1064
1065    ldst = prepare_host_addr(s, &h, addr_reg, oi, false);
1066    tcg_out_qemu_st_indexed(s, get_memop(oi), data_reg, h);
1067
1068    if (ldst) {
1069        ldst->type = data_type;
1070        ldst->datalo_reg = data_reg;
1071        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1072    }
1073}
1074
1075/*
1076 * Entry-points
1077 */
1078
1079static const tcg_insn_unit *tb_ret_addr;
1080
1081static void tcg_out_exit_tb(TCGContext *s, uintptr_t a0)
1082{
1083    /* Reuse the zeroing that exists for goto_ptr.  */
1084    if (a0 == 0) {
1085        tcg_out_call_int(s, tcg_code_gen_epilogue, true);
1086    } else {
1087        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_A0, a0);
1088        tcg_out_call_int(s, tb_ret_addr, true);
1089    }
1090}
1091
1092static void tcg_out_goto_tb(TCGContext *s, int which)
1093{
1094    /*
1095     * Direct branch, or load indirect address, to be patched
1096     * by tb_target_set_jmp_target.  Check indirect load offset
1097     * in range early, regardless of direct branch distance,
1098     * via assert within tcg_out_opc_pcaddu2i.
1099     */
1100    uintptr_t i_addr = get_jmp_target_addr(s, which);
1101    intptr_t i_disp = tcg_pcrel_diff(s, (void *)i_addr);
1102
1103    set_jmp_insn_offset(s, which);
1104    tcg_out_opc_pcaddu2i(s, TCG_REG_TMP0, i_disp >> 2);
1105
1106    /* Finish the load and indirect branch. */
1107    tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP0, TCG_REG_TMP0, 0);
1108    tcg_out_opc_jirl(s, TCG_REG_ZERO, TCG_REG_TMP0, 0);
1109    set_jmp_reset_offset(s, which);
1110}
1111
1112void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
1113                              uintptr_t jmp_rx, uintptr_t jmp_rw)
1114{
1115    uintptr_t d_addr = tb->jmp_target_addr[n];
1116    ptrdiff_t d_disp = (ptrdiff_t)(d_addr - jmp_rx) >> 2;
1117    tcg_insn_unit insn;
1118
1119    /* Either directly branch, or load slot address for indirect branch. */
1120    if (d_disp == sextreg(d_disp, 0, 26)) {
1121        insn = encode_sd10k16_insn(OPC_B, d_disp);
1122    } else {
1123        uintptr_t i_addr = (uintptr_t)&tb->jmp_target_addr[n];
1124        intptr_t i_disp = i_addr - jmp_rx;
1125        insn = encode_dsj20_insn(OPC_PCADDU2I, TCG_REG_TMP0, i_disp >> 2);
1126    }
1127
1128    qatomic_set((tcg_insn_unit *)jmp_rw, insn);
1129    flush_idcache_range(jmp_rx, jmp_rw, 4);
1130}
1131
1132static void tcg_out_op(TCGContext *s, TCGOpcode opc,
1133                       const TCGArg args[TCG_MAX_OP_ARGS],
1134                       const int const_args[TCG_MAX_OP_ARGS])
1135{
1136    TCGArg a0 = args[0];
1137    TCGArg a1 = args[1];
1138    TCGArg a2 = args[2];
1139    int c2 = const_args[2];
1140
1141    switch (opc) {
1142    case INDEX_op_mb:
1143        tcg_out_mb(s, a0);
1144        break;
1145
1146    case INDEX_op_goto_ptr:
1147        tcg_out_opc_jirl(s, TCG_REG_ZERO, a0, 0);
1148        break;
1149
1150    case INDEX_op_br:
1151        tcg_out_reloc(s, s->code_ptr, R_LOONGARCH_BR_SD10K16, arg_label(a0),
1152                      0);
1153        tcg_out_opc_b(s, 0);
1154        break;
1155
1156    case INDEX_op_brcond_i32:
1157    case INDEX_op_brcond_i64:
1158        tcg_out_brcond(s, a2, a0, a1, arg_label(args[3]));
1159        break;
1160
1161    case INDEX_op_extrh_i64_i32:
1162        tcg_out_opc_srai_d(s, a0, a1, 32);
1163        break;
1164
1165    case INDEX_op_not_i32:
1166    case INDEX_op_not_i64:
1167        tcg_out_opc_nor(s, a0, a1, TCG_REG_ZERO);
1168        break;
1169
1170    case INDEX_op_nor_i32:
1171    case INDEX_op_nor_i64:
1172        if (c2) {
1173            tcg_out_opc_ori(s, a0, a1, a2);
1174            tcg_out_opc_nor(s, a0, a0, TCG_REG_ZERO);
1175        } else {
1176            tcg_out_opc_nor(s, a0, a1, a2);
1177        }
1178        break;
1179
1180    case INDEX_op_andc_i32:
1181    case INDEX_op_andc_i64:
1182        if (c2) {
1183            /* guaranteed to fit due to constraint */
1184            tcg_out_opc_andi(s, a0, a1, ~a2);
1185        } else {
1186            tcg_out_opc_andn(s, a0, a1, a2);
1187        }
1188        break;
1189
1190    case INDEX_op_orc_i32:
1191    case INDEX_op_orc_i64:
1192        if (c2) {
1193            /* guaranteed to fit due to constraint */
1194            tcg_out_opc_ori(s, a0, a1, ~a2);
1195        } else {
1196            tcg_out_opc_orn(s, a0, a1, a2);
1197        }
1198        break;
1199
1200    case INDEX_op_and_i32:
1201    case INDEX_op_and_i64:
1202        if (c2) {
1203            tcg_out_opc_andi(s, a0, a1, a2);
1204        } else {
1205            tcg_out_opc_and(s, a0, a1, a2);
1206        }
1207        break;
1208
1209    case INDEX_op_or_i32:
1210    case INDEX_op_or_i64:
1211        if (c2) {
1212            tcg_out_opc_ori(s, a0, a1, a2);
1213        } else {
1214            tcg_out_opc_or(s, a0, a1, a2);
1215        }
1216        break;
1217
1218    case INDEX_op_xor_i32:
1219    case INDEX_op_xor_i64:
1220        if (c2) {
1221            tcg_out_opc_xori(s, a0, a1, a2);
1222        } else {
1223            tcg_out_opc_xor(s, a0, a1, a2);
1224        }
1225        break;
1226
1227    case INDEX_op_extract_i32:
1228        tcg_out_opc_bstrpick_w(s, a0, a1, a2, a2 + args[3] - 1);
1229        break;
1230    case INDEX_op_extract_i64:
1231        tcg_out_opc_bstrpick_d(s, a0, a1, a2, a2 + args[3] - 1);
1232        break;
1233
1234    case INDEX_op_deposit_i32:
1235        tcg_out_opc_bstrins_w(s, a0, a2, args[3], args[3] + args[4] - 1);
1236        break;
1237    case INDEX_op_deposit_i64:
1238        tcg_out_opc_bstrins_d(s, a0, a2, args[3], args[3] + args[4] - 1);
1239        break;
1240
1241    case INDEX_op_bswap16_i32:
1242    case INDEX_op_bswap16_i64:
1243        tcg_out_opc_revb_2h(s, a0, a1);
1244        if (a2 & TCG_BSWAP_OS) {
1245            tcg_out_ext16s(s, TCG_TYPE_REG, a0, a0);
1246        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1247            tcg_out_ext16u(s, a0, a0);
1248        }
1249        break;
1250
1251    case INDEX_op_bswap32_i32:
1252        /* All 32-bit values are computed sign-extended in the register.  */
1253        a2 = TCG_BSWAP_OS;
1254        /* fallthrough */
1255    case INDEX_op_bswap32_i64:
1256        tcg_out_opc_revb_2w(s, a0, a1);
1257        if (a2 & TCG_BSWAP_OS) {
1258            tcg_out_ext32s(s, a0, a0);
1259        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1260            tcg_out_ext32u(s, a0, a0);
1261        }
1262        break;
1263
1264    case INDEX_op_bswap64_i64:
1265        tcg_out_opc_revb_d(s, a0, a1);
1266        break;
1267
1268    case INDEX_op_clz_i32:
1269        tcg_out_clzctz(s, OPC_CLZ_W, a0, a1, a2, c2, true);
1270        break;
1271    case INDEX_op_clz_i64:
1272        tcg_out_clzctz(s, OPC_CLZ_D, a0, a1, a2, c2, false);
1273        break;
1274
1275    case INDEX_op_ctz_i32:
1276        tcg_out_clzctz(s, OPC_CTZ_W, a0, a1, a2, c2, true);
1277        break;
1278    case INDEX_op_ctz_i64:
1279        tcg_out_clzctz(s, OPC_CTZ_D, a0, a1, a2, c2, false);
1280        break;
1281
1282    case INDEX_op_shl_i32:
1283        if (c2) {
1284            tcg_out_opc_slli_w(s, a0, a1, a2 & 0x1f);
1285        } else {
1286            tcg_out_opc_sll_w(s, a0, a1, a2);
1287        }
1288        break;
1289    case INDEX_op_shl_i64:
1290        if (c2) {
1291            tcg_out_opc_slli_d(s, a0, a1, a2 & 0x3f);
1292        } else {
1293            tcg_out_opc_sll_d(s, a0, a1, a2);
1294        }
1295        break;
1296
1297    case INDEX_op_shr_i32:
1298        if (c2) {
1299            tcg_out_opc_srli_w(s, a0, a1, a2 & 0x1f);
1300        } else {
1301            tcg_out_opc_srl_w(s, a0, a1, a2);
1302        }
1303        break;
1304    case INDEX_op_shr_i64:
1305        if (c2) {
1306            tcg_out_opc_srli_d(s, a0, a1, a2 & 0x3f);
1307        } else {
1308            tcg_out_opc_srl_d(s, a0, a1, a2);
1309        }
1310        break;
1311
1312    case INDEX_op_sar_i32:
1313        if (c2) {
1314            tcg_out_opc_srai_w(s, a0, a1, a2 & 0x1f);
1315        } else {
1316            tcg_out_opc_sra_w(s, a0, a1, a2);
1317        }
1318        break;
1319    case INDEX_op_sar_i64:
1320        if (c2) {
1321            tcg_out_opc_srai_d(s, a0, a1, a2 & 0x3f);
1322        } else {
1323            tcg_out_opc_sra_d(s, a0, a1, a2);
1324        }
1325        break;
1326
1327    case INDEX_op_rotl_i32:
1328        /* transform into equivalent rotr/rotri */
1329        if (c2) {
1330            tcg_out_opc_rotri_w(s, a0, a1, (32 - a2) & 0x1f);
1331        } else {
1332            tcg_out_opc_sub_w(s, TCG_REG_TMP0, TCG_REG_ZERO, a2);
1333            tcg_out_opc_rotr_w(s, a0, a1, TCG_REG_TMP0);
1334        }
1335        break;
1336    case INDEX_op_rotl_i64:
1337        /* transform into equivalent rotr/rotri */
1338        if (c2) {
1339            tcg_out_opc_rotri_d(s, a0, a1, (64 - a2) & 0x3f);
1340        } else {
1341            tcg_out_opc_sub_w(s, TCG_REG_TMP0, TCG_REG_ZERO, a2);
1342            tcg_out_opc_rotr_d(s, a0, a1, TCG_REG_TMP0);
1343        }
1344        break;
1345
1346    case INDEX_op_rotr_i32:
1347        if (c2) {
1348            tcg_out_opc_rotri_w(s, a0, a1, a2 & 0x1f);
1349        } else {
1350            tcg_out_opc_rotr_w(s, a0, a1, a2);
1351        }
1352        break;
1353    case INDEX_op_rotr_i64:
1354        if (c2) {
1355            tcg_out_opc_rotri_d(s, a0, a1, a2 & 0x3f);
1356        } else {
1357            tcg_out_opc_rotr_d(s, a0, a1, a2);
1358        }
1359        break;
1360
1361    case INDEX_op_add_i32:
1362        if (c2) {
1363            tcg_out_addi(s, TCG_TYPE_I32, a0, a1, a2);
1364        } else {
1365            tcg_out_opc_add_w(s, a0, a1, a2);
1366        }
1367        break;
1368    case INDEX_op_add_i64:
1369        if (c2) {
1370            tcg_out_addi(s, TCG_TYPE_I64, a0, a1, a2);
1371        } else {
1372            tcg_out_opc_add_d(s, a0, a1, a2);
1373        }
1374        break;
1375
1376    case INDEX_op_sub_i32:
1377        if (c2) {
1378            tcg_out_addi(s, TCG_TYPE_I32, a0, a1, -a2);
1379        } else {
1380            tcg_out_opc_sub_w(s, a0, a1, a2);
1381        }
1382        break;
1383    case INDEX_op_sub_i64:
1384        if (c2) {
1385            tcg_out_addi(s, TCG_TYPE_I64, a0, a1, -a2);
1386        } else {
1387            tcg_out_opc_sub_d(s, a0, a1, a2);
1388        }
1389        break;
1390
1391    case INDEX_op_mul_i32:
1392        tcg_out_opc_mul_w(s, a0, a1, a2);
1393        break;
1394    case INDEX_op_mul_i64:
1395        tcg_out_opc_mul_d(s, a0, a1, a2);
1396        break;
1397
1398    case INDEX_op_mulsh_i32:
1399        tcg_out_opc_mulh_w(s, a0, a1, a2);
1400        break;
1401    case INDEX_op_mulsh_i64:
1402        tcg_out_opc_mulh_d(s, a0, a1, a2);
1403        break;
1404
1405    case INDEX_op_muluh_i32:
1406        tcg_out_opc_mulh_wu(s, a0, a1, a2);
1407        break;
1408    case INDEX_op_muluh_i64:
1409        tcg_out_opc_mulh_du(s, a0, a1, a2);
1410        break;
1411
1412    case INDEX_op_div_i32:
1413        tcg_out_opc_div_w(s, a0, a1, a2);
1414        break;
1415    case INDEX_op_div_i64:
1416        tcg_out_opc_div_d(s, a0, a1, a2);
1417        break;
1418
1419    case INDEX_op_divu_i32:
1420        tcg_out_opc_div_wu(s, a0, a1, a2);
1421        break;
1422    case INDEX_op_divu_i64:
1423        tcg_out_opc_div_du(s, a0, a1, a2);
1424        break;
1425
1426    case INDEX_op_rem_i32:
1427        tcg_out_opc_mod_w(s, a0, a1, a2);
1428        break;
1429    case INDEX_op_rem_i64:
1430        tcg_out_opc_mod_d(s, a0, a1, a2);
1431        break;
1432
1433    case INDEX_op_remu_i32:
1434        tcg_out_opc_mod_wu(s, a0, a1, a2);
1435        break;
1436    case INDEX_op_remu_i64:
1437        tcg_out_opc_mod_du(s, a0, a1, a2);
1438        break;
1439
1440    case INDEX_op_setcond_i32:
1441    case INDEX_op_setcond_i64:
1442        tcg_out_setcond(s, args[3], a0, a1, a2, c2);
1443        break;
1444
1445    case INDEX_op_movcond_i32:
1446    case INDEX_op_movcond_i64:
1447        tcg_out_movcond(s, args[5], a0, a1, a2, c2, args[3], args[4]);
1448        break;
1449
1450    case INDEX_op_ld8s_i32:
1451    case INDEX_op_ld8s_i64:
1452        tcg_out_ldst(s, OPC_LD_B, a0, a1, a2);
1453        break;
1454    case INDEX_op_ld8u_i32:
1455    case INDEX_op_ld8u_i64:
1456        tcg_out_ldst(s, OPC_LD_BU, a0, a1, a2);
1457        break;
1458    case INDEX_op_ld16s_i32:
1459    case INDEX_op_ld16s_i64:
1460        tcg_out_ldst(s, OPC_LD_H, a0, a1, a2);
1461        break;
1462    case INDEX_op_ld16u_i32:
1463    case INDEX_op_ld16u_i64:
1464        tcg_out_ldst(s, OPC_LD_HU, a0, a1, a2);
1465        break;
1466    case INDEX_op_ld_i32:
1467    case INDEX_op_ld32s_i64:
1468        tcg_out_ldst(s, OPC_LD_W, a0, a1, a2);
1469        break;
1470    case INDEX_op_ld32u_i64:
1471        tcg_out_ldst(s, OPC_LD_WU, a0, a1, a2);
1472        break;
1473    case INDEX_op_ld_i64:
1474        tcg_out_ldst(s, OPC_LD_D, a0, a1, a2);
1475        break;
1476
1477    case INDEX_op_st8_i32:
1478    case INDEX_op_st8_i64:
1479        tcg_out_ldst(s, OPC_ST_B, a0, a1, a2);
1480        break;
1481    case INDEX_op_st16_i32:
1482    case INDEX_op_st16_i64:
1483        tcg_out_ldst(s, OPC_ST_H, a0, a1, a2);
1484        break;
1485    case INDEX_op_st_i32:
1486    case INDEX_op_st32_i64:
1487        tcg_out_ldst(s, OPC_ST_W, a0, a1, a2);
1488        break;
1489    case INDEX_op_st_i64:
1490        tcg_out_ldst(s, OPC_ST_D, a0, a1, a2);
1491        break;
1492
1493    case INDEX_op_qemu_ld_a32_i32:
1494    case INDEX_op_qemu_ld_a64_i32:
1495        tcg_out_qemu_ld(s, a0, a1, a2, TCG_TYPE_I32);
1496        break;
1497    case INDEX_op_qemu_ld_a32_i64:
1498    case INDEX_op_qemu_ld_a64_i64:
1499        tcg_out_qemu_ld(s, a0, a1, a2, TCG_TYPE_I64);
1500        break;
1501    case INDEX_op_qemu_st_a32_i32:
1502    case INDEX_op_qemu_st_a64_i32:
1503        tcg_out_qemu_st(s, a0, a1, a2, TCG_TYPE_I32);
1504        break;
1505    case INDEX_op_qemu_st_a32_i64:
1506    case INDEX_op_qemu_st_a64_i64:
1507        tcg_out_qemu_st(s, a0, a1, a2, TCG_TYPE_I64);
1508        break;
1509
1510    case INDEX_op_mov_i32:  /* Always emitted via tcg_out_mov.  */
1511    case INDEX_op_mov_i64:
1512    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
1513    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
1514    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
1515    case INDEX_op_ext8s_i32:  /* Always emitted via tcg_reg_alloc_op.  */
1516    case INDEX_op_ext8s_i64:
1517    case INDEX_op_ext8u_i32:
1518    case INDEX_op_ext8u_i64:
1519    case INDEX_op_ext16s_i32:
1520    case INDEX_op_ext16s_i64:
1521    case INDEX_op_ext16u_i32:
1522    case INDEX_op_ext16u_i64:
1523    case INDEX_op_ext32s_i64:
1524    case INDEX_op_ext32u_i64:
1525    case INDEX_op_ext_i32_i64:
1526    case INDEX_op_extu_i32_i64:
1527    case INDEX_op_extrl_i64_i32:
1528    default:
1529        g_assert_not_reached();
1530    }
1531}
1532
1533static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
1534                            TCGReg rd, TCGReg rs)
1535{
1536    switch (vece) {
1537    case MO_8:
1538        tcg_out_opc_vreplgr2vr_b(s, rd, rs);
1539        break;
1540    case MO_16:
1541        tcg_out_opc_vreplgr2vr_h(s, rd, rs);
1542        break;
1543    case MO_32:
1544        tcg_out_opc_vreplgr2vr_w(s, rd, rs);
1545        break;
1546    case MO_64:
1547        tcg_out_opc_vreplgr2vr_d(s, rd, rs);
1548        break;
1549    default:
1550        g_assert_not_reached();
1551    }
1552    return true;
1553}
1554
1555static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
1556                             TCGReg r, TCGReg base, intptr_t offset)
1557{
1558    /* Handle imm overflow and division (vldrepl.d imm is divided by 8) */
1559    if (offset < -0x800 || offset > 0x7ff || \
1560        (offset & ((1 << vece) - 1)) != 0) {
1561        tcg_out_addi(s, TCG_TYPE_I64, TCG_REG_TMP0, base, offset);
1562        base = TCG_REG_TMP0;
1563        offset = 0;
1564    }
1565    offset >>= vece;
1566
1567    switch (vece) {
1568    case MO_8:
1569        tcg_out_opc_vldrepl_b(s, r, base, offset);
1570        break;
1571    case MO_16:
1572        tcg_out_opc_vldrepl_h(s, r, base, offset);
1573        break;
1574    case MO_32:
1575        tcg_out_opc_vldrepl_w(s, r, base, offset);
1576        break;
1577    case MO_64:
1578        tcg_out_opc_vldrepl_d(s, r, base, offset);
1579        break;
1580    default:
1581        g_assert_not_reached();
1582    }
1583    return true;
1584}
1585
1586static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
1587                             TCGReg rd, int64_t v64)
1588{
1589    /* Try vldi if imm can fit */
1590    int64_t value = sextract64(v64, 0, 8 << vece);
1591    if (-0x200 <= value && value <= 0x1FF) {
1592        uint32_t imm = (vece << 10) | ((uint32_t)v64 & 0x3FF);
1593        tcg_out_opc_vldi(s, rd, imm);
1594        return;
1595    }
1596
1597    /* TODO: vldi patterns when imm 12 is set */
1598
1599    /* Fallback to vreplgr2vr */
1600    tcg_out_movi(s, TCG_TYPE_I64, TCG_REG_TMP0, value);
1601    switch (vece) {
1602    case MO_8:
1603        tcg_out_opc_vreplgr2vr_b(s, rd, TCG_REG_TMP0);
1604        break;
1605    case MO_16:
1606        tcg_out_opc_vreplgr2vr_h(s, rd, TCG_REG_TMP0);
1607        break;
1608    case MO_32:
1609        tcg_out_opc_vreplgr2vr_w(s, rd, TCG_REG_TMP0);
1610        break;
1611    case MO_64:
1612        tcg_out_opc_vreplgr2vr_d(s, rd, TCG_REG_TMP0);
1613        break;
1614    default:
1615        g_assert_not_reached();
1616    }
1617}
1618
1619static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
1620                           unsigned vecl, unsigned vece,
1621                           const TCGArg args[TCG_MAX_OP_ARGS],
1622                           const int const_args[TCG_MAX_OP_ARGS])
1623{
1624    TCGType type = vecl + TCG_TYPE_V64;
1625    TCGArg a0, a1, a2;
1626    TCGReg temp = TCG_REG_TMP0;
1627
1628    a0 = args[0];
1629    a1 = args[1];
1630    a2 = args[2];
1631
1632    /* Currently only supports V128 */
1633    tcg_debug_assert(type == TCG_TYPE_V128);
1634
1635    switch (opc) {
1636    case INDEX_op_st_vec:
1637        /* Try to fit vst imm */
1638        if (-0x800 <= a2 && a2 <= 0x7ff) {
1639            tcg_out_opc_vst(s, a0, a1, a2);
1640        } else {
1641            tcg_out_movi(s, TCG_TYPE_I64, temp, a2);
1642            tcg_out_opc_vstx(s, a0, a1, temp);
1643        }
1644        break;
1645    case INDEX_op_ld_vec:
1646        /* Try to fit vld imm */
1647        if (-0x800 <= a2 && a2 <= 0x7ff) {
1648            tcg_out_opc_vld(s, a0, a1, a2);
1649        } else {
1650            tcg_out_movi(s, TCG_TYPE_I64, temp, a2);
1651            tcg_out_opc_vldx(s, a0, a1, temp);
1652        }
1653        break;
1654    case INDEX_op_dupm_vec:
1655        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
1656        break;
1657    default:
1658        g_assert_not_reached();
1659    }
1660}
1661
1662int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
1663{
1664    switch (opc) {
1665    case INDEX_op_ld_vec:
1666    case INDEX_op_st_vec:
1667    case INDEX_op_dup_vec:
1668    case INDEX_op_dupm_vec:
1669        return 1;
1670    default:
1671        return 0;
1672    }
1673}
1674
1675void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
1676                       TCGArg a0, ...)
1677{
1678    g_assert_not_reached();
1679}
1680
1681static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
1682{
1683    switch (op) {
1684    case INDEX_op_goto_ptr:
1685        return C_O0_I1(r);
1686
1687    case INDEX_op_st8_i32:
1688    case INDEX_op_st8_i64:
1689    case INDEX_op_st16_i32:
1690    case INDEX_op_st16_i64:
1691    case INDEX_op_st32_i64:
1692    case INDEX_op_st_i32:
1693    case INDEX_op_st_i64:
1694    case INDEX_op_qemu_st_a32_i32:
1695    case INDEX_op_qemu_st_a64_i32:
1696    case INDEX_op_qemu_st_a32_i64:
1697    case INDEX_op_qemu_st_a64_i64:
1698        return C_O0_I2(rZ, r);
1699
1700    case INDEX_op_brcond_i32:
1701    case INDEX_op_brcond_i64:
1702        return C_O0_I2(rZ, rZ);
1703
1704    case INDEX_op_ext8s_i32:
1705    case INDEX_op_ext8s_i64:
1706    case INDEX_op_ext8u_i32:
1707    case INDEX_op_ext8u_i64:
1708    case INDEX_op_ext16s_i32:
1709    case INDEX_op_ext16s_i64:
1710    case INDEX_op_ext16u_i32:
1711    case INDEX_op_ext16u_i64:
1712    case INDEX_op_ext32s_i64:
1713    case INDEX_op_ext32u_i64:
1714    case INDEX_op_extu_i32_i64:
1715    case INDEX_op_extrl_i64_i32:
1716    case INDEX_op_extrh_i64_i32:
1717    case INDEX_op_ext_i32_i64:
1718    case INDEX_op_not_i32:
1719    case INDEX_op_not_i64:
1720    case INDEX_op_extract_i32:
1721    case INDEX_op_extract_i64:
1722    case INDEX_op_bswap16_i32:
1723    case INDEX_op_bswap16_i64:
1724    case INDEX_op_bswap32_i32:
1725    case INDEX_op_bswap32_i64:
1726    case INDEX_op_bswap64_i64:
1727    case INDEX_op_ld8s_i32:
1728    case INDEX_op_ld8s_i64:
1729    case INDEX_op_ld8u_i32:
1730    case INDEX_op_ld8u_i64:
1731    case INDEX_op_ld16s_i32:
1732    case INDEX_op_ld16s_i64:
1733    case INDEX_op_ld16u_i32:
1734    case INDEX_op_ld16u_i64:
1735    case INDEX_op_ld32s_i64:
1736    case INDEX_op_ld32u_i64:
1737    case INDEX_op_ld_i32:
1738    case INDEX_op_ld_i64:
1739    case INDEX_op_qemu_ld_a32_i32:
1740    case INDEX_op_qemu_ld_a64_i32:
1741    case INDEX_op_qemu_ld_a32_i64:
1742    case INDEX_op_qemu_ld_a64_i64:
1743        return C_O1_I1(r, r);
1744
1745    case INDEX_op_andc_i32:
1746    case INDEX_op_andc_i64:
1747    case INDEX_op_orc_i32:
1748    case INDEX_op_orc_i64:
1749        /*
1750         * LoongArch insns for these ops don't have reg-imm forms, but we
1751         * can express using andi/ori if ~constant satisfies
1752         * TCG_CT_CONST_U12.
1753         */
1754        return C_O1_I2(r, r, rC);
1755
1756    case INDEX_op_shl_i32:
1757    case INDEX_op_shl_i64:
1758    case INDEX_op_shr_i32:
1759    case INDEX_op_shr_i64:
1760    case INDEX_op_sar_i32:
1761    case INDEX_op_sar_i64:
1762    case INDEX_op_rotl_i32:
1763    case INDEX_op_rotl_i64:
1764    case INDEX_op_rotr_i32:
1765    case INDEX_op_rotr_i64:
1766        return C_O1_I2(r, r, ri);
1767
1768    case INDEX_op_add_i32:
1769        return C_O1_I2(r, r, ri);
1770    case INDEX_op_add_i64:
1771        return C_O1_I2(r, r, rJ);
1772
1773    case INDEX_op_and_i32:
1774    case INDEX_op_and_i64:
1775    case INDEX_op_nor_i32:
1776    case INDEX_op_nor_i64:
1777    case INDEX_op_or_i32:
1778    case INDEX_op_or_i64:
1779    case INDEX_op_xor_i32:
1780    case INDEX_op_xor_i64:
1781        /* LoongArch reg-imm bitops have their imms ZERO-extended */
1782        return C_O1_I2(r, r, rU);
1783
1784    case INDEX_op_clz_i32:
1785    case INDEX_op_clz_i64:
1786    case INDEX_op_ctz_i32:
1787    case INDEX_op_ctz_i64:
1788        return C_O1_I2(r, r, rW);
1789
1790    case INDEX_op_deposit_i32:
1791    case INDEX_op_deposit_i64:
1792        /* Must deposit into the same register as input */
1793        return C_O1_I2(r, 0, rZ);
1794
1795    case INDEX_op_sub_i32:
1796    case INDEX_op_setcond_i32:
1797        return C_O1_I2(r, rZ, ri);
1798    case INDEX_op_sub_i64:
1799    case INDEX_op_setcond_i64:
1800        return C_O1_I2(r, rZ, rJ);
1801
1802    case INDEX_op_mul_i32:
1803    case INDEX_op_mul_i64:
1804    case INDEX_op_mulsh_i32:
1805    case INDEX_op_mulsh_i64:
1806    case INDEX_op_muluh_i32:
1807    case INDEX_op_muluh_i64:
1808    case INDEX_op_div_i32:
1809    case INDEX_op_div_i64:
1810    case INDEX_op_divu_i32:
1811    case INDEX_op_divu_i64:
1812    case INDEX_op_rem_i32:
1813    case INDEX_op_rem_i64:
1814    case INDEX_op_remu_i32:
1815    case INDEX_op_remu_i64:
1816        return C_O1_I2(r, rZ, rZ);
1817
1818    case INDEX_op_movcond_i32:
1819    case INDEX_op_movcond_i64:
1820        return C_O1_I4(r, rZ, rJ, rZ, rZ);
1821
1822    case INDEX_op_ld_vec:
1823    case INDEX_op_dupm_vec:
1824    case INDEX_op_dup_vec:
1825        return C_O1_I1(w, r);
1826
1827    case INDEX_op_st_vec:
1828        return C_O0_I2(w, r);
1829
1830    default:
1831        g_assert_not_reached();
1832    }
1833}
1834
1835static const int tcg_target_callee_save_regs[] = {
1836    TCG_REG_S0,     /* used for the global env (TCG_AREG0) */
1837    TCG_REG_S1,
1838    TCG_REG_S2,
1839    TCG_REG_S3,
1840    TCG_REG_S4,
1841    TCG_REG_S5,
1842    TCG_REG_S6,
1843    TCG_REG_S7,
1844    TCG_REG_S8,
1845    TCG_REG_S9,
1846    TCG_REG_RA,     /* should be last for ABI compliance */
1847};
1848
1849/* Stack frame parameters.  */
1850#define REG_SIZE   (TCG_TARGET_REG_BITS / 8)
1851#define SAVE_SIZE  ((int)ARRAY_SIZE(tcg_target_callee_save_regs) * REG_SIZE)
1852#define TEMP_SIZE  (CPU_TEMP_BUF_NLONGS * (int)sizeof(long))
1853#define FRAME_SIZE ((TCG_STATIC_CALL_ARGS_SIZE + TEMP_SIZE + SAVE_SIZE \
1854                     + TCG_TARGET_STACK_ALIGN - 1) \
1855                    & -TCG_TARGET_STACK_ALIGN)
1856#define SAVE_OFS   (TCG_STATIC_CALL_ARGS_SIZE + TEMP_SIZE)
1857
1858/* We're expecting to be able to use an immediate for frame allocation.  */
1859QEMU_BUILD_BUG_ON(FRAME_SIZE > 0x7ff);
1860
1861/* Generate global QEMU prologue and epilogue code */
1862static void tcg_target_qemu_prologue(TCGContext *s)
1863{
1864    int i;
1865
1866    tcg_set_frame(s, TCG_REG_SP, TCG_STATIC_CALL_ARGS_SIZE, TEMP_SIZE);
1867
1868    /* TB prologue */
1869    tcg_out_opc_addi_d(s, TCG_REG_SP, TCG_REG_SP, -FRAME_SIZE);
1870    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
1871        tcg_out_st(s, TCG_TYPE_REG, tcg_target_callee_save_regs[i],
1872                   TCG_REG_SP, SAVE_OFS + i * REG_SIZE);
1873    }
1874
1875#if !defined(CONFIG_SOFTMMU)
1876    if (USE_GUEST_BASE) {
1877        tcg_out_movi(s, TCG_TYPE_PTR, TCG_GUEST_BASE_REG, guest_base);
1878        tcg_regset_set_reg(s->reserved_regs, TCG_GUEST_BASE_REG);
1879    }
1880#endif
1881
1882    /* Call generated code */
1883    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
1884    tcg_out_opc_jirl(s, TCG_REG_ZERO, tcg_target_call_iarg_regs[1], 0);
1885
1886    /* Return path for goto_ptr. Set return value to 0 */
1887    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
1888    tcg_out_mov(s, TCG_TYPE_REG, TCG_REG_A0, TCG_REG_ZERO);
1889
1890    /* TB epilogue */
1891    tb_ret_addr = tcg_splitwx_to_rx(s->code_ptr);
1892    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
1893        tcg_out_ld(s, TCG_TYPE_REG, tcg_target_callee_save_regs[i],
1894                   TCG_REG_SP, SAVE_OFS + i * REG_SIZE);
1895    }
1896
1897    tcg_out_opc_addi_d(s, TCG_REG_SP, TCG_REG_SP, FRAME_SIZE);
1898    tcg_out_opc_jirl(s, TCG_REG_ZERO, TCG_REG_RA, 0);
1899}
1900
1901static void tcg_target_init(TCGContext *s)
1902{
1903    unsigned long hwcap = qemu_getauxval(AT_HWCAP);
1904
1905    /* Server and desktop class cpus have UAL; embedded cpus do not. */
1906    if (!(hwcap & HWCAP_LOONGARCH_UAL)) {
1907        error_report("TCG: unaligned access support required; exiting");
1908        exit(EXIT_FAILURE);
1909    }
1910
1911    if (hwcap & HWCAP_LOONGARCH_LSX) {
1912        use_lsx_instructions = 1;
1913    }
1914
1915    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
1916    tcg_target_available_regs[TCG_TYPE_I64] = ALL_GENERAL_REGS;
1917
1918    tcg_target_call_clobber_regs = ALL_GENERAL_REGS;
1919    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S0);
1920    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S1);
1921    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S2);
1922    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S3);
1923    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S4);
1924    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S5);
1925    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S6);
1926    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S7);
1927    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S8);
1928    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S9);
1929
1930    if (use_lsx_instructions) {
1931        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
1932        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V24);
1933        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V25);
1934        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V26);
1935        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V27);
1936        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V28);
1937        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V29);
1938        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V30);
1939        tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V31);
1940    }
1941
1942    s->reserved_regs = 0;
1943    tcg_regset_set_reg(s->reserved_regs, TCG_REG_ZERO);
1944    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP0);
1945    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP1);
1946    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP2);
1947    tcg_regset_set_reg(s->reserved_regs, TCG_REG_SP);
1948    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TP);
1949    tcg_regset_set_reg(s->reserved_regs, TCG_REG_RESERVED);
1950    tcg_regset_set_reg(s->reserved_regs, TCG_VEC_TMP0);
1951}
1952
1953typedef struct {
1954    DebugFrameHeader h;
1955    uint8_t fde_def_cfa[4];
1956    uint8_t fde_reg_ofs[ARRAY_SIZE(tcg_target_callee_save_regs) * 2];
1957} DebugFrame;
1958
1959#define ELF_HOST_MACHINE EM_LOONGARCH
1960
1961static const DebugFrame debug_frame = {
1962    .h.cie.len = sizeof(DebugFrameCIE) - 4, /* length after .len member */
1963    .h.cie.id = -1,
1964    .h.cie.version = 1,
1965    .h.cie.code_align = 1,
1966    .h.cie.data_align = -(TCG_TARGET_REG_BITS / 8) & 0x7f, /* sleb128 */
1967    .h.cie.return_column = TCG_REG_RA,
1968
1969    /* Total FDE size does not include the "len" member.  */
1970    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
1971
1972    .fde_def_cfa = {
1973        12, TCG_REG_SP,                 /* DW_CFA_def_cfa sp, ...  */
1974        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
1975        (FRAME_SIZE >> 7)
1976    },
1977    .fde_reg_ofs = {
1978        0x80 + 23, 11,                  /* DW_CFA_offset, s0, -88 */
1979        0x80 + 24, 10,                  /* DW_CFA_offset, s1, -80 */
1980        0x80 + 25, 9,                   /* DW_CFA_offset, s2, -72 */
1981        0x80 + 26, 8,                   /* DW_CFA_offset, s3, -64 */
1982        0x80 + 27, 7,                   /* DW_CFA_offset, s4, -56 */
1983        0x80 + 28, 6,                   /* DW_CFA_offset, s5, -48 */
1984        0x80 + 29, 5,                   /* DW_CFA_offset, s6, -40 */
1985        0x80 + 30, 4,                   /* DW_CFA_offset, s7, -32 */
1986        0x80 + 31, 3,                   /* DW_CFA_offset, s8, -24 */
1987        0x80 + 22, 2,                   /* DW_CFA_offset, s9, -16 */
1988        0x80 + 1 , 1,                   /* DW_CFA_offset, ra, -8 */
1989    }
1990};
1991
1992void tcg_register_jit(const void *buf, size_t buf_size)
1993{
1994    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
1995}
1996