xref: /openbmc/qemu/tcg/loongarch64/tcg-target.c.inc (revision 12fde9bcdb52118495d10c32ed375679f23e323c)
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
35#ifdef CONFIG_DEBUG_TCG
36static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
37    "zero",
38    "ra",
39    "tp",
40    "sp",
41    "a0",
42    "a1",
43    "a2",
44    "a3",
45    "a4",
46    "a5",
47    "a6",
48    "a7",
49    "t0",
50    "t1",
51    "t2",
52    "t3",
53    "t4",
54    "t5",
55    "t6",
56    "t7",
57    "t8",
58    "r21", /* reserved in the LP64* ABI, hence no ABI name */
59    "s9",
60    "s0",
61    "s1",
62    "s2",
63    "s3",
64    "s4",
65    "s5",
66    "s6",
67    "s7",
68    "s8"
69};
70#endif
71
72static const int tcg_target_reg_alloc_order[] = {
73    /* Registers preserved across calls */
74    /* TCG_REG_S0 reserved for TCG_AREG0 */
75    TCG_REG_S1,
76    TCG_REG_S2,
77    TCG_REG_S3,
78    TCG_REG_S4,
79    TCG_REG_S5,
80    TCG_REG_S6,
81    TCG_REG_S7,
82    TCG_REG_S8,
83    TCG_REG_S9,
84
85    /* Registers (potentially) clobbered across calls */
86    TCG_REG_T0,
87    TCG_REG_T1,
88    TCG_REG_T2,
89    TCG_REG_T3,
90    TCG_REG_T4,
91    TCG_REG_T5,
92    TCG_REG_T6,
93    TCG_REG_T7,
94    TCG_REG_T8,
95
96    /* Argument registers, opposite order of allocation.  */
97    TCG_REG_A7,
98    TCG_REG_A6,
99    TCG_REG_A5,
100    TCG_REG_A4,
101    TCG_REG_A3,
102    TCG_REG_A2,
103    TCG_REG_A1,
104    TCG_REG_A0,
105};
106
107static const int tcg_target_call_iarg_regs[] = {
108    TCG_REG_A0,
109    TCG_REG_A1,
110    TCG_REG_A2,
111    TCG_REG_A3,
112    TCG_REG_A4,
113    TCG_REG_A5,
114    TCG_REG_A6,
115    TCG_REG_A7,
116};
117
118static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
119{
120    tcg_debug_assert(kind == TCG_CALL_RET_NORMAL);
121    tcg_debug_assert(slot >= 0 && slot <= 1);
122    return TCG_REG_A0 + slot;
123}
124
125#ifndef CONFIG_SOFTMMU
126#define USE_GUEST_BASE     (guest_base != 0)
127#define TCG_GUEST_BASE_REG TCG_REG_S1
128#endif
129
130#define TCG_CT_CONST_ZERO  0x100
131#define TCG_CT_CONST_S12   0x200
132#define TCG_CT_CONST_S32   0x400
133#define TCG_CT_CONST_U12   0x800
134#define TCG_CT_CONST_C12   0x1000
135#define TCG_CT_CONST_WSZ   0x2000
136
137#define ALL_GENERAL_REGS   MAKE_64BIT_MASK(0, 32)
138
139static inline tcg_target_long sextreg(tcg_target_long val, int pos, int len)
140{
141    return sextract64(val, pos, len);
142}
143
144/* test if a constant matches the constraint */
145static bool tcg_target_const_match(int64_t val, TCGType type, int ct)
146{
147    if (ct & TCG_CT_CONST) {
148        return true;
149    }
150    if ((ct & TCG_CT_CONST_ZERO) && val == 0) {
151        return true;
152    }
153    if ((ct & TCG_CT_CONST_S12) && val == sextreg(val, 0, 12)) {
154        return true;
155    }
156    if ((ct & TCG_CT_CONST_S32) && val == (int32_t)val) {
157        return true;
158    }
159    if ((ct & TCG_CT_CONST_U12) && val >= 0 && val <= 0xfff) {
160        return true;
161    }
162    if ((ct & TCG_CT_CONST_C12) && ~val >= 0 && ~val <= 0xfff) {
163        return true;
164    }
165    if ((ct & TCG_CT_CONST_WSZ) && val == (type == TCG_TYPE_I32 ? 32 : 64)) {
166        return true;
167    }
168    return false;
169}
170
171/*
172 * Relocations
173 */
174
175/*
176 * Relocation records defined in LoongArch ELF psABI v1.00 is way too
177 * complicated; a whopping stack machine is needed to stuff the fields, at
178 * the very least one SOP_PUSH and one SOP_POP (of the correct format) are
179 * needed.
180 *
181 * Hence, define our own simpler relocation types. Numbers are chosen as to
182 * not collide with potential future additions to the true ELF relocation
183 * type enum.
184 */
185
186/* Field Sk16, shifted right by 2; suitable for conditional jumps */
187#define R_LOONGARCH_BR_SK16     256
188/* Field Sd10k16, shifted right by 2; suitable for B and BL */
189#define R_LOONGARCH_BR_SD10K16  257
190
191static bool reloc_br_sk16(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
192{
193    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
194    intptr_t offset = (intptr_t)target - (intptr_t)src_rx;
195
196    tcg_debug_assert((offset & 3) == 0);
197    offset >>= 2;
198    if (offset == sextreg(offset, 0, 16)) {
199        *src_rw = deposit64(*src_rw, 10, 16, offset);
200        return true;
201    }
202
203    return false;
204}
205
206static bool reloc_br_sd10k16(tcg_insn_unit *src_rw,
207                             const tcg_insn_unit *target)
208{
209    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
210    intptr_t offset = (intptr_t)target - (intptr_t)src_rx;
211
212    tcg_debug_assert((offset & 3) == 0);
213    offset >>= 2;
214    if (offset == sextreg(offset, 0, 26)) {
215        *src_rw = deposit64(*src_rw, 0, 10, offset >> 16); /* slot d10 */
216        *src_rw = deposit64(*src_rw, 10, 16, offset); /* slot k16 */
217        return true;
218    }
219
220    return false;
221}
222
223static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
224                        intptr_t value, intptr_t addend)
225{
226    tcg_debug_assert(addend == 0);
227    switch (type) {
228    case R_LOONGARCH_BR_SK16:
229        return reloc_br_sk16(code_ptr, (tcg_insn_unit *)value);
230    case R_LOONGARCH_BR_SD10K16:
231        return reloc_br_sd10k16(code_ptr, (tcg_insn_unit *)value);
232    default:
233        g_assert_not_reached();
234    }
235}
236
237#include "tcg-insn-defs.c.inc"
238
239/*
240 * TCG intrinsics
241 */
242
243static void tcg_out_mb(TCGContext *s, TCGArg a0)
244{
245    /* Baseline LoongArch only has the full barrier, unfortunately.  */
246    tcg_out_opc_dbar(s, 0);
247}
248
249static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
250{
251    if (ret == arg) {
252        return true;
253    }
254    switch (type) {
255    case TCG_TYPE_I32:
256    case TCG_TYPE_I64:
257        /*
258         * Conventional register-register move used in LoongArch is
259         * `or dst, src, zero`.
260         */
261        tcg_out_opc_or(s, ret, arg, TCG_REG_ZERO);
262        break;
263    default:
264        g_assert_not_reached();
265    }
266    return true;
267}
268
269/* Loads a 32-bit immediate into rd, sign-extended.  */
270static void tcg_out_movi_i32(TCGContext *s, TCGReg rd, int32_t val)
271{
272    tcg_target_long lo = sextreg(val, 0, 12);
273    tcg_target_long hi12 = sextreg(val, 12, 20);
274
275    /* Single-instruction cases.  */
276    if (hi12 == 0) {
277        /* val fits in uimm12: ori rd, zero, val */
278        tcg_out_opc_ori(s, rd, TCG_REG_ZERO, val);
279        return;
280    }
281    if (hi12 == sextreg(lo, 12, 20)) {
282        /* val fits in simm12: addi.w rd, zero, val */
283        tcg_out_opc_addi_w(s, rd, TCG_REG_ZERO, val);
284        return;
285    }
286
287    /* High bits must be set; load with lu12i.w + optional ori.  */
288    tcg_out_opc_lu12i_w(s, rd, hi12);
289    if (lo != 0) {
290        tcg_out_opc_ori(s, rd, rd, lo & 0xfff);
291    }
292}
293
294static void tcg_out_movi(TCGContext *s, TCGType type, TCGReg rd,
295                         tcg_target_long val)
296{
297    /*
298     * LoongArch conventionally loads 64-bit immediates in at most 4 steps,
299     * with dedicated instructions for filling the respective bitfields
300     * below:
301     *
302     *        6                   5                   4               3
303     *  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
304     * +-----------------------+---------------------------------------+...
305     * |          hi52         |                  hi32                 |
306     * +-----------------------+---------------------------------------+...
307     *       3                   2                   1
308     *     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
309     * ...+-------------------------------------+-------------------------+
310     *    |                 hi12                |            lo           |
311     * ...+-------------------------------------+-------------------------+
312     *
313     * Check if val belong to one of the several fast cases, before falling
314     * back to the slow path.
315     */
316
317    intptr_t pc_offset;
318    tcg_target_long val_lo, val_hi, pc_hi, offset_hi;
319    tcg_target_long hi12, hi32, hi52;
320
321    /* Value fits in signed i32.  */
322    if (type == TCG_TYPE_I32 || val == (int32_t)val) {
323        tcg_out_movi_i32(s, rd, val);
324        return;
325    }
326
327    /* PC-relative cases.  */
328    pc_offset = tcg_pcrel_diff(s, (void *)val);
329    if (pc_offset == sextreg(pc_offset, 0, 22) && (pc_offset & 3) == 0) {
330        /* Single pcaddu2i.  */
331        tcg_out_opc_pcaddu2i(s, rd, pc_offset >> 2);
332        return;
333    }
334
335    if (pc_offset == (int32_t)pc_offset) {
336        /* Offset within 32 bits; load with pcalau12i + ori.  */
337        val_lo = sextreg(val, 0, 12);
338        val_hi = val >> 12;
339        pc_hi = (val - pc_offset) >> 12;
340        offset_hi = val_hi - pc_hi;
341
342        tcg_debug_assert(offset_hi == sextreg(offset_hi, 0, 20));
343        tcg_out_opc_pcalau12i(s, rd, offset_hi);
344        if (val_lo != 0) {
345            tcg_out_opc_ori(s, rd, rd, val_lo & 0xfff);
346        }
347        return;
348    }
349
350    hi12 = sextreg(val, 12, 20);
351    hi32 = sextreg(val, 32, 20);
352    hi52 = sextreg(val, 52, 12);
353
354    /* Single cu52i.d case.  */
355    if ((hi52 != 0) && (ctz64(val) >= 52)) {
356        tcg_out_opc_cu52i_d(s, rd, TCG_REG_ZERO, hi52);
357        return;
358    }
359
360    /* Slow path.  Initialize the low 32 bits, then concat high bits.  */
361    tcg_out_movi_i32(s, rd, val);
362
363    /* Load hi32 and hi52 explicitly when they are unexpected values. */
364    if (hi32 != sextreg(hi12, 20, 20)) {
365        tcg_out_opc_cu32i_d(s, rd, hi32);
366    }
367
368    if (hi52 != sextreg(hi32, 20, 12)) {
369        tcg_out_opc_cu52i_d(s, rd, rd, hi52);
370    }
371}
372
373static void tcg_out_addi(TCGContext *s, TCGType type, TCGReg rd,
374                         TCGReg rs, tcg_target_long imm)
375{
376    tcg_target_long lo12 = sextreg(imm, 0, 12);
377    tcg_target_long hi16 = sextreg(imm - lo12, 16, 16);
378
379    /*
380     * Note that there's a hole in between hi16 and lo12:
381     *
382     *       3                   2                   1                   0
383     *     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
384     * ...+-------------------------------+-------+-----------------------+
385     *    |             hi16              |       |          lo12         |
386     * ...+-------------------------------+-------+-----------------------+
387     *
388     * For bits within that hole, it's more efficient to use LU12I and ADD.
389     */
390    if (imm == (hi16 << 16) + lo12) {
391        if (hi16) {
392            tcg_out_opc_addu16i_d(s, rd, rs, hi16);
393            rs = rd;
394        }
395        if (type == TCG_TYPE_I32) {
396            tcg_out_opc_addi_w(s, rd, rs, lo12);
397        } else if (lo12) {
398            tcg_out_opc_addi_d(s, rd, rs, lo12);
399        } else {
400            tcg_out_mov(s, type, rd, rs);
401        }
402    } else {
403        tcg_out_movi(s, type, TCG_REG_TMP0, imm);
404        if (type == TCG_TYPE_I32) {
405            tcg_out_opc_add_w(s, rd, rs, TCG_REG_TMP0);
406        } else {
407            tcg_out_opc_add_d(s, rd, rs, TCG_REG_TMP0);
408        }
409    }
410}
411
412static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
413{
414    return false;
415}
416
417static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
418                             tcg_target_long imm)
419{
420    /* This function is only used for passing structs by reference. */
421    g_assert_not_reached();
422}
423
424static void tcg_out_ext8u(TCGContext *s, TCGReg ret, TCGReg arg)
425{
426    tcg_out_opc_andi(s, ret, arg, 0xff);
427}
428
429static void tcg_out_ext16u(TCGContext *s, TCGReg ret, TCGReg arg)
430{
431    tcg_out_opc_bstrpick_w(s, ret, arg, 0, 15);
432}
433
434static void tcg_out_ext32u(TCGContext *s, TCGReg ret, TCGReg arg)
435{
436    tcg_out_opc_bstrpick_d(s, ret, arg, 0, 31);
437}
438
439static void tcg_out_ext8s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
440{
441    tcg_out_opc_sext_b(s, ret, arg);
442}
443
444static void tcg_out_ext16s(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
445{
446    tcg_out_opc_sext_h(s, ret, arg);
447}
448
449static void tcg_out_ext32s(TCGContext *s, TCGReg ret, TCGReg arg)
450{
451    tcg_out_opc_addi_w(s, ret, arg, 0);
452}
453
454static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg)
455{
456    if (ret != arg) {
457        tcg_out_ext32s(s, ret, arg);
458    }
459}
460
461static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg ret, TCGReg arg)
462{
463    tcg_out_ext32u(s, ret, arg);
464}
465
466static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg ret, TCGReg arg)
467{
468    tcg_out_ext32s(s, ret, arg);
469}
470
471static void tcg_out_clzctz(TCGContext *s, LoongArchInsn opc,
472                           TCGReg a0, TCGReg a1, TCGReg a2,
473                           bool c2, bool is_32bit)
474{
475    if (c2) {
476        /*
477         * Fast path: semantics already satisfied due to constraint and
478         * insn behavior, single instruction is enough.
479         */
480        tcg_debug_assert(a2 == (is_32bit ? 32 : 64));
481        /* all clz/ctz insns belong to DJ-format */
482        tcg_out32(s, encode_dj_insn(opc, a0, a1));
483        return;
484    }
485
486    tcg_out32(s, encode_dj_insn(opc, TCG_REG_TMP0, a1));
487    /* a0 = a1 ? REG_TMP0 : a2 */
488    tcg_out_opc_maskeqz(s, TCG_REG_TMP0, TCG_REG_TMP0, a1);
489    tcg_out_opc_masknez(s, a0, a2, a1);
490    tcg_out_opc_or(s, a0, TCG_REG_TMP0, a0);
491}
492
493#define SETCOND_INV    TCG_TARGET_NB_REGS
494#define SETCOND_NEZ    (SETCOND_INV << 1)
495#define SETCOND_FLAGS  (SETCOND_INV | SETCOND_NEZ)
496
497static int tcg_out_setcond_int(TCGContext *s, TCGCond cond, TCGReg ret,
498                               TCGReg arg1, tcg_target_long arg2, bool c2)
499{
500    int flags = 0;
501
502    switch (cond) {
503    case TCG_COND_EQ:    /* -> NE  */
504    case TCG_COND_GE:    /* -> LT  */
505    case TCG_COND_GEU:   /* -> LTU */
506    case TCG_COND_GT:    /* -> LE  */
507    case TCG_COND_GTU:   /* -> LEU */
508        cond = tcg_invert_cond(cond);
509        flags ^= SETCOND_INV;
510        break;
511    default:
512        break;
513    }
514
515    switch (cond) {
516    case TCG_COND_LE:
517    case TCG_COND_LEU:
518        /*
519         * If we have a constant input, the most efficient way to implement
520         * LE is by adding 1 and using LT.  Watch out for wrap around for LEU.
521         * We don't need to care for this for LE because the constant input
522         * is still constrained to int32_t, and INT32_MAX+1 is representable
523         * in the 64-bit temporary register.
524         */
525        if (c2) {
526            if (cond == TCG_COND_LEU) {
527                /* unsigned <= -1 is true */
528                if (arg2 == -1) {
529                    tcg_out_movi(s, TCG_TYPE_REG, ret, !(flags & SETCOND_INV));
530                    return ret;
531                }
532                cond = TCG_COND_LTU;
533            } else {
534                cond = TCG_COND_LT;
535            }
536            arg2 += 1;
537        } else {
538            TCGReg tmp = arg2;
539            arg2 = arg1;
540            arg1 = tmp;
541            cond = tcg_swap_cond(cond);    /* LE -> GE */
542            cond = tcg_invert_cond(cond);  /* GE -> LT */
543            flags ^= SETCOND_INV;
544        }
545        break;
546    default:
547        break;
548    }
549
550    switch (cond) {
551    case TCG_COND_NE:
552        flags |= SETCOND_NEZ;
553        if (!c2) {
554            tcg_out_opc_xor(s, ret, arg1, arg2);
555        } else if (arg2 == 0) {
556            ret = arg1;
557        } else if (arg2 >= 0 && arg2 <= 0xfff) {
558            tcg_out_opc_xori(s, ret, arg1, arg2);
559        } else {
560            tcg_out_addi(s, TCG_TYPE_REG, ret, arg1, -arg2);
561        }
562        break;
563
564    case TCG_COND_LT:
565    case TCG_COND_LTU:
566        if (c2) {
567            if (arg2 >= -0x800 && arg2 <= 0x7ff) {
568                if (cond == TCG_COND_LT) {
569                    tcg_out_opc_slti(s, ret, arg1, arg2);
570                } else {
571                    tcg_out_opc_sltui(s, ret, arg1, arg2);
572                }
573                break;
574            }
575            tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_TMP0, arg2);
576            arg2 = TCG_REG_TMP0;
577        }
578        if (cond == TCG_COND_LT) {
579            tcg_out_opc_slt(s, ret, arg1, arg2);
580        } else {
581            tcg_out_opc_sltu(s, ret, arg1, arg2);
582        }
583        break;
584
585    default:
586        g_assert_not_reached();
587        break;
588    }
589
590    return ret | flags;
591}
592
593static void tcg_out_setcond(TCGContext *s, TCGCond cond, TCGReg ret,
594                            TCGReg arg1, tcg_target_long arg2, bool c2)
595{
596    int tmpflags = tcg_out_setcond_int(s, cond, ret, arg1, arg2, c2);
597
598    if (tmpflags != ret) {
599        TCGReg tmp = tmpflags & ~SETCOND_FLAGS;
600
601        switch (tmpflags & SETCOND_FLAGS) {
602        case SETCOND_INV:
603            /* Intermediate result is boolean: simply invert. */
604            tcg_out_opc_xori(s, ret, tmp, 1);
605            break;
606        case SETCOND_NEZ:
607            /* Intermediate result is zero/non-zero: test != 0. */
608            tcg_out_opc_sltu(s, ret, TCG_REG_ZERO, tmp);
609            break;
610        case SETCOND_NEZ | SETCOND_INV:
611            /* Intermediate result is zero/non-zero: test == 0. */
612            tcg_out_opc_sltui(s, ret, tmp, 1);
613            break;
614        default:
615            g_assert_not_reached();
616        }
617    }
618}
619
620static void tcg_out_movcond(TCGContext *s, TCGCond cond, TCGReg ret,
621                            TCGReg c1, tcg_target_long c2, bool const2,
622                            TCGReg v1, TCGReg v2)
623{
624    int tmpflags = tcg_out_setcond_int(s, cond, TCG_REG_TMP0, c1, c2, const2);
625    TCGReg t;
626
627    /* Standardize the test below to t != 0. */
628    if (tmpflags & SETCOND_INV) {
629        t = v1, v1 = v2, v2 = t;
630    }
631
632    t = tmpflags & ~SETCOND_FLAGS;
633    if (v1 == TCG_REG_ZERO) {
634        tcg_out_opc_masknez(s, ret, v2, t);
635    } else if (v2 == TCG_REG_ZERO) {
636        tcg_out_opc_maskeqz(s, ret, v1, t);
637    } else {
638        tcg_out_opc_masknez(s, TCG_REG_TMP2, v2, t); /* t ? 0 : v2 */
639        tcg_out_opc_maskeqz(s, TCG_REG_TMP1, v1, t); /* t ? v1 : 0 */
640        tcg_out_opc_or(s, ret, TCG_REG_TMP1, TCG_REG_TMP2);
641    }
642}
643
644/*
645 * Branch helpers
646 */
647
648static const struct {
649    LoongArchInsn op;
650    bool swap;
651} tcg_brcond_to_loongarch[] = {
652    [TCG_COND_EQ] =  { OPC_BEQ,  false },
653    [TCG_COND_NE] =  { OPC_BNE,  false },
654    [TCG_COND_LT] =  { OPC_BGT,  true  },
655    [TCG_COND_GE] =  { OPC_BLE,  true  },
656    [TCG_COND_LE] =  { OPC_BLE,  false },
657    [TCG_COND_GT] =  { OPC_BGT,  false },
658    [TCG_COND_LTU] = { OPC_BGTU, true  },
659    [TCG_COND_GEU] = { OPC_BLEU, true  },
660    [TCG_COND_LEU] = { OPC_BLEU, false },
661    [TCG_COND_GTU] = { OPC_BGTU, false }
662};
663
664static void tcg_out_brcond(TCGContext *s, TCGCond cond, TCGReg arg1,
665                           TCGReg arg2, TCGLabel *l)
666{
667    LoongArchInsn op = tcg_brcond_to_loongarch[cond].op;
668
669    tcg_debug_assert(op != 0);
670
671    if (tcg_brcond_to_loongarch[cond].swap) {
672        TCGReg t = arg1;
673        arg1 = arg2;
674        arg2 = t;
675    }
676
677    /* all conditional branch insns belong to DJSk16-format */
678    tcg_out_reloc(s, s->code_ptr, R_LOONGARCH_BR_SK16, l, 0);
679    tcg_out32(s, encode_djsk16_insn(op, arg1, arg2, 0));
680}
681
682static void tcg_out_call_int(TCGContext *s, const tcg_insn_unit *arg, bool tail)
683{
684    TCGReg link = tail ? TCG_REG_ZERO : TCG_REG_RA;
685    ptrdiff_t offset = tcg_pcrel_diff(s, arg);
686
687    tcg_debug_assert((offset & 3) == 0);
688    if (offset == sextreg(offset, 0, 28)) {
689        /* short jump: +/- 256MiB */
690        if (tail) {
691            tcg_out_opc_b(s, offset >> 2);
692        } else {
693            tcg_out_opc_bl(s, offset >> 2);
694        }
695    } else if (offset == sextreg(offset, 0, 38)) {
696        /* long jump: +/- 256GiB */
697        tcg_target_long lo = sextreg(offset, 0, 18);
698        tcg_target_long hi = offset - lo;
699        tcg_out_opc_pcaddu18i(s, TCG_REG_TMP0, hi >> 18);
700        tcg_out_opc_jirl(s, link, TCG_REG_TMP0, lo >> 2);
701    } else {
702        /* far jump: 64-bit */
703        tcg_target_long lo = sextreg((tcg_target_long)arg, 0, 18);
704        tcg_target_long hi = (tcg_target_long)arg - lo;
705        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP0, hi);
706        tcg_out_opc_jirl(s, link, TCG_REG_TMP0, lo >> 2);
707    }
708}
709
710static void tcg_out_call(TCGContext *s, const tcg_insn_unit *arg,
711                         const TCGHelperInfo *info)
712{
713    tcg_out_call_int(s, arg, false);
714}
715
716/*
717 * Load/store helpers
718 */
719
720static void tcg_out_ldst(TCGContext *s, LoongArchInsn opc, TCGReg data,
721                         TCGReg addr, intptr_t offset)
722{
723    intptr_t imm12 = sextreg(offset, 0, 12);
724
725    if (offset != imm12) {
726        intptr_t diff = tcg_pcrel_diff(s, (void *)offset);
727
728        if (addr == TCG_REG_ZERO && diff == (int32_t)diff) {
729            imm12 = sextreg(diff, 0, 12);
730            tcg_out_opc_pcaddu12i(s, TCG_REG_TMP2, (diff - imm12) >> 12);
731        } else {
732            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP2, offset - imm12);
733            if (addr != TCG_REG_ZERO) {
734                tcg_out_opc_add_d(s, TCG_REG_TMP2, TCG_REG_TMP2, addr);
735            }
736        }
737        addr = TCG_REG_TMP2;
738    }
739
740    switch (opc) {
741    case OPC_LD_B:
742    case OPC_LD_BU:
743    case OPC_LD_H:
744    case OPC_LD_HU:
745    case OPC_LD_W:
746    case OPC_LD_WU:
747    case OPC_LD_D:
748    case OPC_ST_B:
749    case OPC_ST_H:
750    case OPC_ST_W:
751    case OPC_ST_D:
752        tcg_out32(s, encode_djsk12_insn(opc, data, addr, imm12));
753        break;
754    default:
755        g_assert_not_reached();
756    }
757}
758
759static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg arg,
760                       TCGReg arg1, intptr_t arg2)
761{
762    bool is_32bit = type == TCG_TYPE_I32;
763    tcg_out_ldst(s, is_32bit ? OPC_LD_W : OPC_LD_D, arg, arg1, arg2);
764}
765
766static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
767                       TCGReg arg1, intptr_t arg2)
768{
769    bool is_32bit = type == TCG_TYPE_I32;
770    tcg_out_ldst(s, is_32bit ? OPC_ST_W : OPC_ST_D, arg, arg1, arg2);
771}
772
773static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
774                        TCGReg base, intptr_t ofs)
775{
776    if (val == 0) {
777        tcg_out_st(s, type, TCG_REG_ZERO, base, ofs);
778        return true;
779    }
780    return false;
781}
782
783/*
784 * Load/store helpers for SoftMMU, and qemu_ld/st implementations
785 */
786
787static bool tcg_out_goto(TCGContext *s, const tcg_insn_unit *target)
788{
789    tcg_out_opc_b(s, 0);
790    return reloc_br_sd10k16(s->code_ptr - 1, target);
791}
792
793static const TCGLdstHelperParam ldst_helper_param = {
794    .ntmp = 1, .tmp = { TCG_REG_TMP0 }
795};
796
797static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
798{
799    MemOp opc = get_memop(l->oi);
800
801    /* resolve label address */
802    if (!reloc_br_sk16(l->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
803        return false;
804    }
805
806    tcg_out_ld_helper_args(s, l, &ldst_helper_param);
807    tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SIZE], false);
808    tcg_out_ld_helper_ret(s, l, false, &ldst_helper_param);
809    return tcg_out_goto(s, l->raddr);
810}
811
812static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
813{
814    MemOp opc = get_memop(l->oi);
815
816    /* resolve label address */
817    if (!reloc_br_sk16(l->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
818        return false;
819    }
820
821    tcg_out_st_helper_args(s, l, &ldst_helper_param);
822    tcg_out_call_int(s, qemu_st_helpers[opc & MO_SIZE], false);
823    return tcg_out_goto(s, l->raddr);
824}
825
826typedef struct {
827    TCGReg base;
828    TCGReg index;
829} HostAddress;
830
831bool tcg_target_has_memory_bswap(MemOp memop)
832{
833    return false;
834}
835
836/*
837 * For softmmu, perform the TLB load and compare.
838 * For useronly, perform any required alignment tests.
839 * In both cases, return a TCGLabelQemuLdst structure if the slow path
840 * is required and fill in @h with the host address for the fast path.
841 */
842static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
843                                           TCGReg addr_reg, MemOpIdx oi,
844                                           bool is_ld)
845{
846    TCGLabelQemuLdst *ldst = NULL;
847    MemOp opc = get_memop(oi);
848    unsigned a_bits = get_alignment_bits(opc);
849
850#ifdef CONFIG_SOFTMMU
851    unsigned s_bits = opc & MO_SIZE;
852    int mem_index = get_mmuidx(oi);
853    int fast_ofs = TLB_MASK_TABLE_OFS(mem_index);
854    int mask_ofs = fast_ofs + offsetof(CPUTLBDescFast, mask);
855    int table_ofs = fast_ofs + offsetof(CPUTLBDescFast, table);
856
857    ldst = new_ldst_label(s);
858    ldst->is_ld = is_ld;
859    ldst->oi = oi;
860    ldst->addrlo_reg = addr_reg;
861
862    QEMU_BUILD_BUG_ON(TLB_MASK_TABLE_OFS(0) > 0);
863    QEMU_BUILD_BUG_ON(TLB_MASK_TABLE_OFS(0) < -(1 << 11));
864    tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP0, TCG_AREG0, mask_ofs);
865    tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP1, TCG_AREG0, table_ofs);
866
867    tcg_out_opc_srli_d(s, TCG_REG_TMP2, addr_reg,
868                    TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS);
869    tcg_out_opc_and(s, TCG_REG_TMP2, TCG_REG_TMP2, TCG_REG_TMP0);
870    tcg_out_opc_add_d(s, TCG_REG_TMP2, TCG_REG_TMP2, TCG_REG_TMP1);
871
872    /* Load the tlb comparator and the addend.  */
873    tcg_out_ld(s, TCG_TYPE_TL, TCG_REG_TMP0, TCG_REG_TMP2,
874               is_ld ? offsetof(CPUTLBEntry, addr_read)
875                     : offsetof(CPUTLBEntry, addr_write));
876    tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP2, TCG_REG_TMP2,
877               offsetof(CPUTLBEntry, addend));
878
879    /*
880     * For aligned accesses, we check the first byte and include the alignment
881     * bits within the address.  For unaligned access, we check that we don't
882     * cross pages using the address of the last byte of the access.
883     */
884    if (a_bits < s_bits) {
885        unsigned a_mask = (1u << a_bits) - 1;
886        unsigned s_mask = (1u << s_bits) - 1;
887        tcg_out_addi(s, TCG_TYPE_TL, TCG_REG_TMP1, addr_reg, s_mask - a_mask);
888    } else {
889        tcg_out_mov(s, TCG_TYPE_TL, TCG_REG_TMP1, addr_reg);
890    }
891    tcg_out_opc_bstrins_d(s, TCG_REG_TMP1, TCG_REG_ZERO,
892                          a_bits, TARGET_PAGE_BITS - 1);
893
894    /* Compare masked address with the TLB entry.  */
895    ldst->label_ptr[0] = s->code_ptr;
896    tcg_out_opc_bne(s, TCG_REG_TMP0, TCG_REG_TMP1, 0);
897
898    h->index = TCG_REG_TMP2;
899#else
900    if (a_bits) {
901        ldst = new_ldst_label(s);
902
903        ldst->is_ld = is_ld;
904        ldst->oi = oi;
905        ldst->addrlo_reg = addr_reg;
906
907        /*
908         * Without micro-architecture details, we don't know which of
909         * bstrpick or andi is faster, so use bstrpick as it's not
910         * constrained by imm field width. Not to say alignments >= 2^12
911         * are going to happen any time soon.
912         */
913        tcg_out_opc_bstrpick_d(s, TCG_REG_TMP1, addr_reg, 0, a_bits - 1);
914
915        ldst->label_ptr[0] = s->code_ptr;
916        tcg_out_opc_bne(s, TCG_REG_TMP1, TCG_REG_ZERO, 0);
917    }
918
919    h->index = USE_GUEST_BASE ? TCG_GUEST_BASE_REG : TCG_REG_ZERO;
920#endif
921
922    if (TARGET_LONG_BITS == 32) {
923        h->base = TCG_REG_TMP0;
924        tcg_out_ext32u(s, h->base, addr_reg);
925    } else {
926        h->base = addr_reg;
927    }
928
929    return ldst;
930}
931
932static void tcg_out_qemu_ld_indexed(TCGContext *s, MemOp opc, TCGType type,
933                                    TCGReg rd, HostAddress h)
934{
935    /* Byte swapping is left to middle-end expansion.  */
936    tcg_debug_assert((opc & MO_BSWAP) == 0);
937
938    switch (opc & MO_SSIZE) {
939    case MO_UB:
940        tcg_out_opc_ldx_bu(s, rd, h.base, h.index);
941        break;
942    case MO_SB:
943        tcg_out_opc_ldx_b(s, rd, h.base, h.index);
944        break;
945    case MO_UW:
946        tcg_out_opc_ldx_hu(s, rd, h.base, h.index);
947        break;
948    case MO_SW:
949        tcg_out_opc_ldx_h(s, rd, h.base, h.index);
950        break;
951    case MO_UL:
952        if (type == TCG_TYPE_I64) {
953            tcg_out_opc_ldx_wu(s, rd, h.base, h.index);
954            break;
955        }
956        /* fallthrough */
957    case MO_SL:
958        tcg_out_opc_ldx_w(s, rd, h.base, h.index);
959        break;
960    case MO_UQ:
961        tcg_out_opc_ldx_d(s, rd, h.base, h.index);
962        break;
963    default:
964        g_assert_not_reached();
965    }
966}
967
968static void tcg_out_qemu_ld(TCGContext *s, TCGReg data_reg, TCGReg addr_reg,
969                            MemOpIdx oi, TCGType data_type)
970{
971    TCGLabelQemuLdst *ldst;
972    HostAddress h;
973
974    ldst = prepare_host_addr(s, &h, addr_reg, oi, true);
975    tcg_out_qemu_ld_indexed(s, get_memop(oi), data_type, data_reg, h);
976
977    if (ldst) {
978        ldst->type = data_type;
979        ldst->datalo_reg = data_reg;
980        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
981    }
982}
983
984static void tcg_out_qemu_st_indexed(TCGContext *s, MemOp opc,
985                                    TCGReg rd, HostAddress h)
986{
987    /* Byte swapping is left to middle-end expansion.  */
988    tcg_debug_assert((opc & MO_BSWAP) == 0);
989
990    switch (opc & MO_SIZE) {
991    case MO_8:
992        tcg_out_opc_stx_b(s, rd, h.base, h.index);
993        break;
994    case MO_16:
995        tcg_out_opc_stx_h(s, rd, h.base, h.index);
996        break;
997    case MO_32:
998        tcg_out_opc_stx_w(s, rd, h.base, h.index);
999        break;
1000    case MO_64:
1001        tcg_out_opc_stx_d(s, rd, h.base, h.index);
1002        break;
1003    default:
1004        g_assert_not_reached();
1005    }
1006}
1007
1008static void tcg_out_qemu_st(TCGContext *s, TCGReg data_reg, TCGReg addr_reg,
1009                            MemOpIdx oi, TCGType data_type)
1010{
1011    TCGLabelQemuLdst *ldst;
1012    HostAddress h;
1013
1014    ldst = prepare_host_addr(s, &h, addr_reg, oi, false);
1015    tcg_out_qemu_st_indexed(s, get_memop(oi), data_reg, h);
1016
1017    if (ldst) {
1018        ldst->type = data_type;
1019        ldst->datalo_reg = data_reg;
1020        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1021    }
1022}
1023
1024/*
1025 * Entry-points
1026 */
1027
1028static const tcg_insn_unit *tb_ret_addr;
1029
1030static void tcg_out_exit_tb(TCGContext *s, uintptr_t a0)
1031{
1032    /* Reuse the zeroing that exists for goto_ptr.  */
1033    if (a0 == 0) {
1034        tcg_out_call_int(s, tcg_code_gen_epilogue, true);
1035    } else {
1036        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_A0, a0);
1037        tcg_out_call_int(s, tb_ret_addr, true);
1038    }
1039}
1040
1041static void tcg_out_goto_tb(TCGContext *s, int which)
1042{
1043    /*
1044     * Direct branch, or load indirect address, to be patched
1045     * by tb_target_set_jmp_target.  Check indirect load offset
1046     * in range early, regardless of direct branch distance,
1047     * via assert within tcg_out_opc_pcaddu2i.
1048     */
1049    uintptr_t i_addr = get_jmp_target_addr(s, which);
1050    intptr_t i_disp = tcg_pcrel_diff(s, (void *)i_addr);
1051
1052    set_jmp_insn_offset(s, which);
1053    tcg_out_opc_pcaddu2i(s, TCG_REG_TMP0, i_disp >> 2);
1054
1055    /* Finish the load and indirect branch. */
1056    tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP0, TCG_REG_TMP0, 0);
1057    tcg_out_opc_jirl(s, TCG_REG_ZERO, TCG_REG_TMP0, 0);
1058    set_jmp_reset_offset(s, which);
1059}
1060
1061void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
1062                              uintptr_t jmp_rx, uintptr_t jmp_rw)
1063{
1064    uintptr_t d_addr = tb->jmp_target_addr[n];
1065    ptrdiff_t d_disp = (ptrdiff_t)(d_addr - jmp_rx) >> 2;
1066    tcg_insn_unit insn;
1067
1068    /* Either directly branch, or load slot address for indirect branch. */
1069    if (d_disp == sextreg(d_disp, 0, 26)) {
1070        insn = encode_sd10k16_insn(OPC_B, d_disp);
1071    } else {
1072        uintptr_t i_addr = (uintptr_t)&tb->jmp_target_addr[n];
1073        intptr_t i_disp = i_addr - jmp_rx;
1074        insn = encode_dsj20_insn(OPC_PCADDU2I, TCG_REG_TMP0, i_disp >> 2);
1075    }
1076
1077    qatomic_set((tcg_insn_unit *)jmp_rw, insn);
1078    flush_idcache_range(jmp_rx, jmp_rw, 4);
1079}
1080
1081static void tcg_out_op(TCGContext *s, TCGOpcode opc,
1082                       const TCGArg args[TCG_MAX_OP_ARGS],
1083                       const int const_args[TCG_MAX_OP_ARGS])
1084{
1085    TCGArg a0 = args[0];
1086    TCGArg a1 = args[1];
1087    TCGArg a2 = args[2];
1088    int c2 = const_args[2];
1089
1090    switch (opc) {
1091    case INDEX_op_mb:
1092        tcg_out_mb(s, a0);
1093        break;
1094
1095    case INDEX_op_goto_ptr:
1096        tcg_out_opc_jirl(s, TCG_REG_ZERO, a0, 0);
1097        break;
1098
1099    case INDEX_op_br:
1100        tcg_out_reloc(s, s->code_ptr, R_LOONGARCH_BR_SD10K16, arg_label(a0),
1101                      0);
1102        tcg_out_opc_b(s, 0);
1103        break;
1104
1105    case INDEX_op_brcond_i32:
1106    case INDEX_op_brcond_i64:
1107        tcg_out_brcond(s, a2, a0, a1, arg_label(args[3]));
1108        break;
1109
1110    case INDEX_op_extrh_i64_i32:
1111        tcg_out_opc_srai_d(s, a0, a1, 32);
1112        break;
1113
1114    case INDEX_op_not_i32:
1115    case INDEX_op_not_i64:
1116        tcg_out_opc_nor(s, a0, a1, TCG_REG_ZERO);
1117        break;
1118
1119    case INDEX_op_nor_i32:
1120    case INDEX_op_nor_i64:
1121        if (c2) {
1122            tcg_out_opc_ori(s, a0, a1, a2);
1123            tcg_out_opc_nor(s, a0, a0, TCG_REG_ZERO);
1124        } else {
1125            tcg_out_opc_nor(s, a0, a1, a2);
1126        }
1127        break;
1128
1129    case INDEX_op_andc_i32:
1130    case INDEX_op_andc_i64:
1131        if (c2) {
1132            /* guaranteed to fit due to constraint */
1133            tcg_out_opc_andi(s, a0, a1, ~a2);
1134        } else {
1135            tcg_out_opc_andn(s, a0, a1, a2);
1136        }
1137        break;
1138
1139    case INDEX_op_orc_i32:
1140    case INDEX_op_orc_i64:
1141        if (c2) {
1142            /* guaranteed to fit due to constraint */
1143            tcg_out_opc_ori(s, a0, a1, ~a2);
1144        } else {
1145            tcg_out_opc_orn(s, a0, a1, a2);
1146        }
1147        break;
1148
1149    case INDEX_op_and_i32:
1150    case INDEX_op_and_i64:
1151        if (c2) {
1152            tcg_out_opc_andi(s, a0, a1, a2);
1153        } else {
1154            tcg_out_opc_and(s, a0, a1, a2);
1155        }
1156        break;
1157
1158    case INDEX_op_or_i32:
1159    case INDEX_op_or_i64:
1160        if (c2) {
1161            tcg_out_opc_ori(s, a0, a1, a2);
1162        } else {
1163            tcg_out_opc_or(s, a0, a1, a2);
1164        }
1165        break;
1166
1167    case INDEX_op_xor_i32:
1168    case INDEX_op_xor_i64:
1169        if (c2) {
1170            tcg_out_opc_xori(s, a0, a1, a2);
1171        } else {
1172            tcg_out_opc_xor(s, a0, a1, a2);
1173        }
1174        break;
1175
1176    case INDEX_op_extract_i32:
1177        tcg_out_opc_bstrpick_w(s, a0, a1, a2, a2 + args[3] - 1);
1178        break;
1179    case INDEX_op_extract_i64:
1180        tcg_out_opc_bstrpick_d(s, a0, a1, a2, a2 + args[3] - 1);
1181        break;
1182
1183    case INDEX_op_deposit_i32:
1184        tcg_out_opc_bstrins_w(s, a0, a2, args[3], args[3] + args[4] - 1);
1185        break;
1186    case INDEX_op_deposit_i64:
1187        tcg_out_opc_bstrins_d(s, a0, a2, args[3], args[3] + args[4] - 1);
1188        break;
1189
1190    case INDEX_op_bswap16_i32:
1191    case INDEX_op_bswap16_i64:
1192        tcg_out_opc_revb_2h(s, a0, a1);
1193        if (a2 & TCG_BSWAP_OS) {
1194            tcg_out_ext16s(s, TCG_TYPE_REG, a0, a0);
1195        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1196            tcg_out_ext16u(s, a0, a0);
1197        }
1198        break;
1199
1200    case INDEX_op_bswap32_i32:
1201        /* All 32-bit values are computed sign-extended in the register.  */
1202        a2 = TCG_BSWAP_OS;
1203        /* fallthrough */
1204    case INDEX_op_bswap32_i64:
1205        tcg_out_opc_revb_2w(s, a0, a1);
1206        if (a2 & TCG_BSWAP_OS) {
1207            tcg_out_ext32s(s, a0, a0);
1208        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1209            tcg_out_ext32u(s, a0, a0);
1210        }
1211        break;
1212
1213    case INDEX_op_bswap64_i64:
1214        tcg_out_opc_revb_d(s, a0, a1);
1215        break;
1216
1217    case INDEX_op_clz_i32:
1218        tcg_out_clzctz(s, OPC_CLZ_W, a0, a1, a2, c2, true);
1219        break;
1220    case INDEX_op_clz_i64:
1221        tcg_out_clzctz(s, OPC_CLZ_D, a0, a1, a2, c2, false);
1222        break;
1223
1224    case INDEX_op_ctz_i32:
1225        tcg_out_clzctz(s, OPC_CTZ_W, a0, a1, a2, c2, true);
1226        break;
1227    case INDEX_op_ctz_i64:
1228        tcg_out_clzctz(s, OPC_CTZ_D, a0, a1, a2, c2, false);
1229        break;
1230
1231    case INDEX_op_shl_i32:
1232        if (c2) {
1233            tcg_out_opc_slli_w(s, a0, a1, a2 & 0x1f);
1234        } else {
1235            tcg_out_opc_sll_w(s, a0, a1, a2);
1236        }
1237        break;
1238    case INDEX_op_shl_i64:
1239        if (c2) {
1240            tcg_out_opc_slli_d(s, a0, a1, a2 & 0x3f);
1241        } else {
1242            tcg_out_opc_sll_d(s, a0, a1, a2);
1243        }
1244        break;
1245
1246    case INDEX_op_shr_i32:
1247        if (c2) {
1248            tcg_out_opc_srli_w(s, a0, a1, a2 & 0x1f);
1249        } else {
1250            tcg_out_opc_srl_w(s, a0, a1, a2);
1251        }
1252        break;
1253    case INDEX_op_shr_i64:
1254        if (c2) {
1255            tcg_out_opc_srli_d(s, a0, a1, a2 & 0x3f);
1256        } else {
1257            tcg_out_opc_srl_d(s, a0, a1, a2);
1258        }
1259        break;
1260
1261    case INDEX_op_sar_i32:
1262        if (c2) {
1263            tcg_out_opc_srai_w(s, a0, a1, a2 & 0x1f);
1264        } else {
1265            tcg_out_opc_sra_w(s, a0, a1, a2);
1266        }
1267        break;
1268    case INDEX_op_sar_i64:
1269        if (c2) {
1270            tcg_out_opc_srai_d(s, a0, a1, a2 & 0x3f);
1271        } else {
1272            tcg_out_opc_sra_d(s, a0, a1, a2);
1273        }
1274        break;
1275
1276    case INDEX_op_rotl_i32:
1277        /* transform into equivalent rotr/rotri */
1278        if (c2) {
1279            tcg_out_opc_rotri_w(s, a0, a1, (32 - a2) & 0x1f);
1280        } else {
1281            tcg_out_opc_sub_w(s, TCG_REG_TMP0, TCG_REG_ZERO, a2);
1282            tcg_out_opc_rotr_w(s, a0, a1, TCG_REG_TMP0);
1283        }
1284        break;
1285    case INDEX_op_rotl_i64:
1286        /* transform into equivalent rotr/rotri */
1287        if (c2) {
1288            tcg_out_opc_rotri_d(s, a0, a1, (64 - a2) & 0x3f);
1289        } else {
1290            tcg_out_opc_sub_w(s, TCG_REG_TMP0, TCG_REG_ZERO, a2);
1291            tcg_out_opc_rotr_d(s, a0, a1, TCG_REG_TMP0);
1292        }
1293        break;
1294
1295    case INDEX_op_rotr_i32:
1296        if (c2) {
1297            tcg_out_opc_rotri_w(s, a0, a1, a2 & 0x1f);
1298        } else {
1299            tcg_out_opc_rotr_w(s, a0, a1, a2);
1300        }
1301        break;
1302    case INDEX_op_rotr_i64:
1303        if (c2) {
1304            tcg_out_opc_rotri_d(s, a0, a1, a2 & 0x3f);
1305        } else {
1306            tcg_out_opc_rotr_d(s, a0, a1, a2);
1307        }
1308        break;
1309
1310    case INDEX_op_add_i32:
1311        if (c2) {
1312            tcg_out_addi(s, TCG_TYPE_I32, a0, a1, a2);
1313        } else {
1314            tcg_out_opc_add_w(s, a0, a1, a2);
1315        }
1316        break;
1317    case INDEX_op_add_i64:
1318        if (c2) {
1319            tcg_out_addi(s, TCG_TYPE_I64, a0, a1, a2);
1320        } else {
1321            tcg_out_opc_add_d(s, a0, a1, a2);
1322        }
1323        break;
1324
1325    case INDEX_op_sub_i32:
1326        if (c2) {
1327            tcg_out_addi(s, TCG_TYPE_I32, a0, a1, -a2);
1328        } else {
1329            tcg_out_opc_sub_w(s, a0, a1, a2);
1330        }
1331        break;
1332    case INDEX_op_sub_i64:
1333        if (c2) {
1334            tcg_out_addi(s, TCG_TYPE_I64, a0, a1, -a2);
1335        } else {
1336            tcg_out_opc_sub_d(s, a0, a1, a2);
1337        }
1338        break;
1339
1340    case INDEX_op_mul_i32:
1341        tcg_out_opc_mul_w(s, a0, a1, a2);
1342        break;
1343    case INDEX_op_mul_i64:
1344        tcg_out_opc_mul_d(s, a0, a1, a2);
1345        break;
1346
1347    case INDEX_op_mulsh_i32:
1348        tcg_out_opc_mulh_w(s, a0, a1, a2);
1349        break;
1350    case INDEX_op_mulsh_i64:
1351        tcg_out_opc_mulh_d(s, a0, a1, a2);
1352        break;
1353
1354    case INDEX_op_muluh_i32:
1355        tcg_out_opc_mulh_wu(s, a0, a1, a2);
1356        break;
1357    case INDEX_op_muluh_i64:
1358        tcg_out_opc_mulh_du(s, a0, a1, a2);
1359        break;
1360
1361    case INDEX_op_div_i32:
1362        tcg_out_opc_div_w(s, a0, a1, a2);
1363        break;
1364    case INDEX_op_div_i64:
1365        tcg_out_opc_div_d(s, a0, a1, a2);
1366        break;
1367
1368    case INDEX_op_divu_i32:
1369        tcg_out_opc_div_wu(s, a0, a1, a2);
1370        break;
1371    case INDEX_op_divu_i64:
1372        tcg_out_opc_div_du(s, a0, a1, a2);
1373        break;
1374
1375    case INDEX_op_rem_i32:
1376        tcg_out_opc_mod_w(s, a0, a1, a2);
1377        break;
1378    case INDEX_op_rem_i64:
1379        tcg_out_opc_mod_d(s, a0, a1, a2);
1380        break;
1381
1382    case INDEX_op_remu_i32:
1383        tcg_out_opc_mod_wu(s, a0, a1, a2);
1384        break;
1385    case INDEX_op_remu_i64:
1386        tcg_out_opc_mod_du(s, a0, a1, a2);
1387        break;
1388
1389    case INDEX_op_setcond_i32:
1390    case INDEX_op_setcond_i64:
1391        tcg_out_setcond(s, args[3], a0, a1, a2, c2);
1392        break;
1393
1394    case INDEX_op_movcond_i32:
1395    case INDEX_op_movcond_i64:
1396        tcg_out_movcond(s, args[5], a0, a1, a2, c2, args[3], args[4]);
1397        break;
1398
1399    case INDEX_op_ld8s_i32:
1400    case INDEX_op_ld8s_i64:
1401        tcg_out_ldst(s, OPC_LD_B, a0, a1, a2);
1402        break;
1403    case INDEX_op_ld8u_i32:
1404    case INDEX_op_ld8u_i64:
1405        tcg_out_ldst(s, OPC_LD_BU, a0, a1, a2);
1406        break;
1407    case INDEX_op_ld16s_i32:
1408    case INDEX_op_ld16s_i64:
1409        tcg_out_ldst(s, OPC_LD_H, a0, a1, a2);
1410        break;
1411    case INDEX_op_ld16u_i32:
1412    case INDEX_op_ld16u_i64:
1413        tcg_out_ldst(s, OPC_LD_HU, a0, a1, a2);
1414        break;
1415    case INDEX_op_ld_i32:
1416    case INDEX_op_ld32s_i64:
1417        tcg_out_ldst(s, OPC_LD_W, a0, a1, a2);
1418        break;
1419    case INDEX_op_ld32u_i64:
1420        tcg_out_ldst(s, OPC_LD_WU, a0, a1, a2);
1421        break;
1422    case INDEX_op_ld_i64:
1423        tcg_out_ldst(s, OPC_LD_D, a0, a1, a2);
1424        break;
1425
1426    case INDEX_op_st8_i32:
1427    case INDEX_op_st8_i64:
1428        tcg_out_ldst(s, OPC_ST_B, a0, a1, a2);
1429        break;
1430    case INDEX_op_st16_i32:
1431    case INDEX_op_st16_i64:
1432        tcg_out_ldst(s, OPC_ST_H, a0, a1, a2);
1433        break;
1434    case INDEX_op_st_i32:
1435    case INDEX_op_st32_i64:
1436        tcg_out_ldst(s, OPC_ST_W, a0, a1, a2);
1437        break;
1438    case INDEX_op_st_i64:
1439        tcg_out_ldst(s, OPC_ST_D, a0, a1, a2);
1440        break;
1441
1442    case INDEX_op_qemu_ld_i32:
1443        tcg_out_qemu_ld(s, a0, a1, a2, TCG_TYPE_I32);
1444        break;
1445    case INDEX_op_qemu_ld_i64:
1446        tcg_out_qemu_ld(s, a0, a1, a2, TCG_TYPE_I64);
1447        break;
1448    case INDEX_op_qemu_st_i32:
1449        tcg_out_qemu_st(s, a0, a1, a2, TCG_TYPE_I32);
1450        break;
1451    case INDEX_op_qemu_st_i64:
1452        tcg_out_qemu_st(s, a0, a1, a2, TCG_TYPE_I64);
1453        break;
1454
1455    case INDEX_op_mov_i32:  /* Always emitted via tcg_out_mov.  */
1456    case INDEX_op_mov_i64:
1457    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
1458    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
1459    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
1460    case INDEX_op_ext8s_i32:  /* Always emitted via tcg_reg_alloc_op.  */
1461    case INDEX_op_ext8s_i64:
1462    case INDEX_op_ext8u_i32:
1463    case INDEX_op_ext8u_i64:
1464    case INDEX_op_ext16s_i32:
1465    case INDEX_op_ext16s_i64:
1466    case INDEX_op_ext16u_i32:
1467    case INDEX_op_ext16u_i64:
1468    case INDEX_op_ext32s_i64:
1469    case INDEX_op_ext32u_i64:
1470    case INDEX_op_ext_i32_i64:
1471    case INDEX_op_extu_i32_i64:
1472    case INDEX_op_extrl_i64_i32:
1473    default:
1474        g_assert_not_reached();
1475    }
1476}
1477
1478static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
1479{
1480    switch (op) {
1481    case INDEX_op_goto_ptr:
1482        return C_O0_I1(r);
1483
1484    case INDEX_op_st8_i32:
1485    case INDEX_op_st8_i64:
1486    case INDEX_op_st16_i32:
1487    case INDEX_op_st16_i64:
1488    case INDEX_op_st32_i64:
1489    case INDEX_op_st_i32:
1490    case INDEX_op_st_i64:
1491    case INDEX_op_qemu_st_i32:
1492    case INDEX_op_qemu_st_i64:
1493        return C_O0_I2(rZ, r);
1494
1495    case INDEX_op_brcond_i32:
1496    case INDEX_op_brcond_i64:
1497        return C_O0_I2(rZ, rZ);
1498
1499    case INDEX_op_ext8s_i32:
1500    case INDEX_op_ext8s_i64:
1501    case INDEX_op_ext8u_i32:
1502    case INDEX_op_ext8u_i64:
1503    case INDEX_op_ext16s_i32:
1504    case INDEX_op_ext16s_i64:
1505    case INDEX_op_ext16u_i32:
1506    case INDEX_op_ext16u_i64:
1507    case INDEX_op_ext32s_i64:
1508    case INDEX_op_ext32u_i64:
1509    case INDEX_op_extu_i32_i64:
1510    case INDEX_op_extrl_i64_i32:
1511    case INDEX_op_extrh_i64_i32:
1512    case INDEX_op_ext_i32_i64:
1513    case INDEX_op_not_i32:
1514    case INDEX_op_not_i64:
1515    case INDEX_op_extract_i32:
1516    case INDEX_op_extract_i64:
1517    case INDEX_op_bswap16_i32:
1518    case INDEX_op_bswap16_i64:
1519    case INDEX_op_bswap32_i32:
1520    case INDEX_op_bswap32_i64:
1521    case INDEX_op_bswap64_i64:
1522    case INDEX_op_ld8s_i32:
1523    case INDEX_op_ld8s_i64:
1524    case INDEX_op_ld8u_i32:
1525    case INDEX_op_ld8u_i64:
1526    case INDEX_op_ld16s_i32:
1527    case INDEX_op_ld16s_i64:
1528    case INDEX_op_ld16u_i32:
1529    case INDEX_op_ld16u_i64:
1530    case INDEX_op_ld32s_i64:
1531    case INDEX_op_ld32u_i64:
1532    case INDEX_op_ld_i32:
1533    case INDEX_op_ld_i64:
1534    case INDEX_op_qemu_ld_i32:
1535    case INDEX_op_qemu_ld_i64:
1536        return C_O1_I1(r, r);
1537
1538    case INDEX_op_andc_i32:
1539    case INDEX_op_andc_i64:
1540    case INDEX_op_orc_i32:
1541    case INDEX_op_orc_i64:
1542        /*
1543         * LoongArch insns for these ops don't have reg-imm forms, but we
1544         * can express using andi/ori if ~constant satisfies
1545         * TCG_CT_CONST_U12.
1546         */
1547        return C_O1_I2(r, r, rC);
1548
1549    case INDEX_op_shl_i32:
1550    case INDEX_op_shl_i64:
1551    case INDEX_op_shr_i32:
1552    case INDEX_op_shr_i64:
1553    case INDEX_op_sar_i32:
1554    case INDEX_op_sar_i64:
1555    case INDEX_op_rotl_i32:
1556    case INDEX_op_rotl_i64:
1557    case INDEX_op_rotr_i32:
1558    case INDEX_op_rotr_i64:
1559        return C_O1_I2(r, r, ri);
1560
1561    case INDEX_op_add_i32:
1562        return C_O1_I2(r, r, ri);
1563    case INDEX_op_add_i64:
1564        return C_O1_I2(r, r, rJ);
1565
1566    case INDEX_op_and_i32:
1567    case INDEX_op_and_i64:
1568    case INDEX_op_nor_i32:
1569    case INDEX_op_nor_i64:
1570    case INDEX_op_or_i32:
1571    case INDEX_op_or_i64:
1572    case INDEX_op_xor_i32:
1573    case INDEX_op_xor_i64:
1574        /* LoongArch reg-imm bitops have their imms ZERO-extended */
1575        return C_O1_I2(r, r, rU);
1576
1577    case INDEX_op_clz_i32:
1578    case INDEX_op_clz_i64:
1579    case INDEX_op_ctz_i32:
1580    case INDEX_op_ctz_i64:
1581        return C_O1_I2(r, r, rW);
1582
1583    case INDEX_op_deposit_i32:
1584    case INDEX_op_deposit_i64:
1585        /* Must deposit into the same register as input */
1586        return C_O1_I2(r, 0, rZ);
1587
1588    case INDEX_op_sub_i32:
1589    case INDEX_op_setcond_i32:
1590        return C_O1_I2(r, rZ, ri);
1591    case INDEX_op_sub_i64:
1592    case INDEX_op_setcond_i64:
1593        return C_O1_I2(r, rZ, rJ);
1594
1595    case INDEX_op_mul_i32:
1596    case INDEX_op_mul_i64:
1597    case INDEX_op_mulsh_i32:
1598    case INDEX_op_mulsh_i64:
1599    case INDEX_op_muluh_i32:
1600    case INDEX_op_muluh_i64:
1601    case INDEX_op_div_i32:
1602    case INDEX_op_div_i64:
1603    case INDEX_op_divu_i32:
1604    case INDEX_op_divu_i64:
1605    case INDEX_op_rem_i32:
1606    case INDEX_op_rem_i64:
1607    case INDEX_op_remu_i32:
1608    case INDEX_op_remu_i64:
1609        return C_O1_I2(r, rZ, rZ);
1610
1611    case INDEX_op_movcond_i32:
1612    case INDEX_op_movcond_i64:
1613        return C_O1_I4(r, rZ, rJ, rZ, rZ);
1614
1615    default:
1616        g_assert_not_reached();
1617    }
1618}
1619
1620static const int tcg_target_callee_save_regs[] = {
1621    TCG_REG_S0,     /* used for the global env (TCG_AREG0) */
1622    TCG_REG_S1,
1623    TCG_REG_S2,
1624    TCG_REG_S3,
1625    TCG_REG_S4,
1626    TCG_REG_S5,
1627    TCG_REG_S6,
1628    TCG_REG_S7,
1629    TCG_REG_S8,
1630    TCG_REG_S9,
1631    TCG_REG_RA,     /* should be last for ABI compliance */
1632};
1633
1634/* Stack frame parameters.  */
1635#define REG_SIZE   (TCG_TARGET_REG_BITS / 8)
1636#define SAVE_SIZE  ((int)ARRAY_SIZE(tcg_target_callee_save_regs) * REG_SIZE)
1637#define TEMP_SIZE  (CPU_TEMP_BUF_NLONGS * (int)sizeof(long))
1638#define FRAME_SIZE ((TCG_STATIC_CALL_ARGS_SIZE + TEMP_SIZE + SAVE_SIZE \
1639                     + TCG_TARGET_STACK_ALIGN - 1) \
1640                    & -TCG_TARGET_STACK_ALIGN)
1641#define SAVE_OFS   (TCG_STATIC_CALL_ARGS_SIZE + TEMP_SIZE)
1642
1643/* We're expecting to be able to use an immediate for frame allocation.  */
1644QEMU_BUILD_BUG_ON(FRAME_SIZE > 0x7ff);
1645
1646/* Generate global QEMU prologue and epilogue code */
1647static void tcg_target_qemu_prologue(TCGContext *s)
1648{
1649    int i;
1650
1651    tcg_set_frame(s, TCG_REG_SP, TCG_STATIC_CALL_ARGS_SIZE, TEMP_SIZE);
1652
1653    /* TB prologue */
1654    tcg_out_opc_addi_d(s, TCG_REG_SP, TCG_REG_SP, -FRAME_SIZE);
1655    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
1656        tcg_out_st(s, TCG_TYPE_REG, tcg_target_callee_save_regs[i],
1657                   TCG_REG_SP, SAVE_OFS + i * REG_SIZE);
1658    }
1659
1660#if !defined(CONFIG_SOFTMMU)
1661    if (USE_GUEST_BASE) {
1662        tcg_out_movi(s, TCG_TYPE_PTR, TCG_GUEST_BASE_REG, guest_base);
1663        tcg_regset_set_reg(s->reserved_regs, TCG_GUEST_BASE_REG);
1664    }
1665#endif
1666
1667    /* Call generated code */
1668    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
1669    tcg_out_opc_jirl(s, TCG_REG_ZERO, tcg_target_call_iarg_regs[1], 0);
1670
1671    /* Return path for goto_ptr. Set return value to 0 */
1672    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
1673    tcg_out_mov(s, TCG_TYPE_REG, TCG_REG_A0, TCG_REG_ZERO);
1674
1675    /* TB epilogue */
1676    tb_ret_addr = tcg_splitwx_to_rx(s->code_ptr);
1677    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
1678        tcg_out_ld(s, TCG_TYPE_REG, tcg_target_callee_save_regs[i],
1679                   TCG_REG_SP, SAVE_OFS + i * REG_SIZE);
1680    }
1681
1682    tcg_out_opc_addi_d(s, TCG_REG_SP, TCG_REG_SP, FRAME_SIZE);
1683    tcg_out_opc_jirl(s, TCG_REG_ZERO, TCG_REG_RA, 0);
1684}
1685
1686static void tcg_target_init(TCGContext *s)
1687{
1688    unsigned long hwcap = qemu_getauxval(AT_HWCAP);
1689
1690    /* Server and desktop class cpus have UAL; embedded cpus do not. */
1691    if (!(hwcap & HWCAP_LOONGARCH_UAL)) {
1692        error_report("TCG: unaligned access support required; exiting");
1693        exit(EXIT_FAILURE);
1694    }
1695
1696    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
1697    tcg_target_available_regs[TCG_TYPE_I64] = ALL_GENERAL_REGS;
1698
1699    tcg_target_call_clobber_regs = ALL_GENERAL_REGS;
1700    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S0);
1701    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S1);
1702    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S2);
1703    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S3);
1704    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S4);
1705    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S5);
1706    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S6);
1707    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S7);
1708    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S8);
1709    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_S9);
1710
1711    s->reserved_regs = 0;
1712    tcg_regset_set_reg(s->reserved_regs, TCG_REG_ZERO);
1713    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP0);
1714    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP1);
1715    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP2);
1716    tcg_regset_set_reg(s->reserved_regs, TCG_REG_SP);
1717    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TP);
1718    tcg_regset_set_reg(s->reserved_regs, TCG_REG_RESERVED);
1719}
1720
1721typedef struct {
1722    DebugFrameHeader h;
1723    uint8_t fde_def_cfa[4];
1724    uint8_t fde_reg_ofs[ARRAY_SIZE(tcg_target_callee_save_regs) * 2];
1725} DebugFrame;
1726
1727#define ELF_HOST_MACHINE EM_LOONGARCH
1728
1729static const DebugFrame debug_frame = {
1730    .h.cie.len = sizeof(DebugFrameCIE) - 4, /* length after .len member */
1731    .h.cie.id = -1,
1732    .h.cie.version = 1,
1733    .h.cie.code_align = 1,
1734    .h.cie.data_align = -(TCG_TARGET_REG_BITS / 8) & 0x7f, /* sleb128 */
1735    .h.cie.return_column = TCG_REG_RA,
1736
1737    /* Total FDE size does not include the "len" member.  */
1738    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
1739
1740    .fde_def_cfa = {
1741        12, TCG_REG_SP,                 /* DW_CFA_def_cfa sp, ...  */
1742        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
1743        (FRAME_SIZE >> 7)
1744    },
1745    .fde_reg_ofs = {
1746        0x80 + 23, 11,                  /* DW_CFA_offset, s0, -88 */
1747        0x80 + 24, 10,                  /* DW_CFA_offset, s1, -80 */
1748        0x80 + 25, 9,                   /* DW_CFA_offset, s2, -72 */
1749        0x80 + 26, 8,                   /* DW_CFA_offset, s3, -64 */
1750        0x80 + 27, 7,                   /* DW_CFA_offset, s4, -56 */
1751        0x80 + 28, 6,                   /* DW_CFA_offset, s5, -48 */
1752        0x80 + 29, 5,                   /* DW_CFA_offset, s6, -40 */
1753        0x80 + 30, 4,                   /* DW_CFA_offset, s7, -32 */
1754        0x80 + 31, 3,                   /* DW_CFA_offset, s8, -24 */
1755        0x80 + 22, 2,                   /* DW_CFA_offset, s9, -16 */
1756        0x80 + 1 , 1,                   /* DW_CFA_offset, ra, -8 */
1757    }
1758};
1759
1760void tcg_register_jit(const void *buf, size_t buf_size)
1761{
1762    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
1763}
1764