xref: /openbmc/qemu/tcg/aarch64/tcg-target.c.inc (revision 46f96bff163512f9f8f9959de4a18c0799001422)
1/*
2 * Initial TCG Implementation for aarch64
3 *
4 * Copyright (c) 2013 Huawei Technologies Duesseldorf GmbH
5 * Written by Claudio Fontana
6 *
7 * This work is licensed under the terms of the GNU GPL, version 2 or
8 * (at your option) any later version.
9 *
10 * See the COPYING file in the top-level directory for details.
11 */
12
13#include "qemu/bitops.h"
14
15/* Used for function call generation. */
16#define TCG_REG_CALL_STACK              TCG_REG_SP
17#define TCG_TARGET_STACK_ALIGN          16
18#define TCG_TARGET_CALL_STACK_OFFSET    0
19#define TCG_TARGET_CALL_ARG_I32         TCG_CALL_ARG_NORMAL
20#define TCG_TARGET_CALL_ARG_I64         TCG_CALL_ARG_NORMAL
21#ifdef CONFIG_DARWIN
22# define TCG_TARGET_CALL_ARG_I128       TCG_CALL_ARG_NORMAL
23#else
24# define TCG_TARGET_CALL_ARG_I128       TCG_CALL_ARG_EVEN
25#endif
26#define TCG_TARGET_CALL_RET_I128        TCG_CALL_RET_NORMAL
27
28/* We're going to re-use TCGType in setting of the SF bit, which controls
29   the size of the operation performed.  If we know the values match, it
30   makes things much cleaner.  */
31QEMU_BUILD_BUG_ON(TCG_TYPE_I32 != 0 || TCG_TYPE_I64 != 1);
32
33#ifdef CONFIG_DEBUG_TCG
34static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
35    "x0", "x1", "x2", "x3", "x4", "x5", "x6", "x7",
36    "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15",
37    "x16", "x17", "x18", "x19", "x20", "x21", "x22", "x23",
38    "x24", "x25", "x26", "x27", "x28", "fp", "x30", "sp",
39
40    "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
41    "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15",
42    "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23",
43    "v24", "v25", "v26", "v27", "v28", "fp", "v30", "v31",
44};
45#endif /* CONFIG_DEBUG_TCG */
46
47static const int tcg_target_reg_alloc_order[] = {
48    TCG_REG_X20, TCG_REG_X21, TCG_REG_X22, TCG_REG_X23,
49    TCG_REG_X24, TCG_REG_X25, TCG_REG_X26, TCG_REG_X27,
50    TCG_REG_X28, /* we will reserve this for guest_base if configured */
51
52    TCG_REG_X8, TCG_REG_X9, TCG_REG_X10, TCG_REG_X11,
53    TCG_REG_X12, TCG_REG_X13, TCG_REG_X14, TCG_REG_X15,
54
55    TCG_REG_X0, TCG_REG_X1, TCG_REG_X2, TCG_REG_X3,
56    TCG_REG_X4, TCG_REG_X5, TCG_REG_X6, TCG_REG_X7,
57
58    /* X16 reserved as temporary */
59    /* X17 reserved as temporary */
60    /* X18 reserved by system */
61    /* X19 reserved for AREG0 */
62    /* X29 reserved as fp */
63    /* X30 reserved as temporary */
64
65    TCG_REG_V0, TCG_REG_V1, TCG_REG_V2, TCG_REG_V3,
66    TCG_REG_V4, TCG_REG_V5, TCG_REG_V6, TCG_REG_V7,
67    /* V8 - V15 are call-saved, and skipped.  */
68    TCG_REG_V16, TCG_REG_V17, TCG_REG_V18, TCG_REG_V19,
69    TCG_REG_V20, TCG_REG_V21, TCG_REG_V22, TCG_REG_V23,
70    TCG_REG_V24, TCG_REG_V25, TCG_REG_V26, TCG_REG_V27,
71    TCG_REG_V28, TCG_REG_V29, TCG_REG_V30, TCG_REG_V31,
72};
73
74static const int tcg_target_call_iarg_regs[8] = {
75    TCG_REG_X0, TCG_REG_X1, TCG_REG_X2, TCG_REG_X3,
76    TCG_REG_X4, TCG_REG_X5, TCG_REG_X6, TCG_REG_X7
77};
78
79static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
80{
81    tcg_debug_assert(kind == TCG_CALL_RET_NORMAL);
82    tcg_debug_assert(slot >= 0 && slot <= 1);
83    return TCG_REG_X0 + slot;
84}
85
86#define TCG_REG_TMP0 TCG_REG_X16
87#define TCG_REG_TMP1 TCG_REG_X17
88#define TCG_REG_TMP2 TCG_REG_X30
89#define TCG_VEC_TMP0 TCG_REG_V31
90
91#define TCG_REG_GUEST_BASE TCG_REG_X28
92
93static bool reloc_pc26(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
94{
95    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
96    ptrdiff_t offset = target - src_rx;
97
98    if (offset == sextract64(offset, 0, 26)) {
99        /* read instruction, mask away previous PC_REL26 parameter contents,
100           set the proper offset, then write back the instruction. */
101        *src_rw = deposit32(*src_rw, 0, 26, offset);
102        return true;
103    }
104    return false;
105}
106
107static bool reloc_pc19(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
108{
109    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
110    ptrdiff_t offset = target - src_rx;
111
112    if (offset == sextract64(offset, 0, 19)) {
113        *src_rw = deposit32(*src_rw, 5, 19, offset);
114        return true;
115    }
116    return false;
117}
118
119static bool reloc_pc14(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
120{
121    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
122    ptrdiff_t offset = target - src_rx;
123
124    if (offset == sextract64(offset, 0, 14)) {
125        *src_rw = deposit32(*src_rw, 5, 14, offset);
126        return true;
127    }
128    return false;
129}
130
131static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
132                        intptr_t value, intptr_t addend)
133{
134    tcg_debug_assert(addend == 0);
135    switch (type) {
136    case R_AARCH64_JUMP26:
137    case R_AARCH64_CALL26:
138        return reloc_pc26(code_ptr, (const tcg_insn_unit *)value);
139    case R_AARCH64_CONDBR19:
140        return reloc_pc19(code_ptr, (const tcg_insn_unit *)value);
141    case R_AARCH64_TSTBR14:
142        return reloc_pc14(code_ptr, (const tcg_insn_unit *)value);
143    default:
144        g_assert_not_reached();
145    }
146}
147
148#define TCG_CT_CONST_AIMM 0x100
149#define TCG_CT_CONST_LIMM 0x200
150#define TCG_CT_CONST_ZERO 0x400
151#define TCG_CT_CONST_MONE 0x800
152#define TCG_CT_CONST_ORRI 0x1000
153#define TCG_CT_CONST_ANDI 0x2000
154#define TCG_CT_CONST_CMP  0x4000
155
156#define ALL_GENERAL_REGS  0xffffffffu
157#define ALL_VECTOR_REGS   0xffffffff00000000ull
158
159/* Match a constant valid for addition (12-bit, optionally shifted).  */
160static inline bool is_aimm(uint64_t val)
161{
162    return (val & ~0xfff) == 0 || (val & ~0xfff000) == 0;
163}
164
165/* Match a constant valid for logical operations.  */
166static inline bool is_limm(uint64_t val)
167{
168    /* Taking a simplified view of the logical immediates for now, ignoring
169       the replication that can happen across the field.  Match bit patterns
170       of the forms
171           0....01....1
172           0..01..10..0
173       and their inverses.  */
174
175    /* Make things easier below, by testing the form with msb clear. */
176    if ((int64_t)val < 0) {
177        val = ~val;
178    }
179    if (val == 0) {
180        return false;
181    }
182    val += val & -val;
183    return (val & (val - 1)) == 0;
184}
185
186/* Return true if v16 is a valid 16-bit shifted immediate.  */
187static bool is_shimm16(uint16_t v16, int *cmode, int *imm8)
188{
189    if (v16 == (v16 & 0xff)) {
190        *cmode = 0x8;
191        *imm8 = v16 & 0xff;
192        return true;
193    } else if (v16 == (v16 & 0xff00)) {
194        *cmode = 0xa;
195        *imm8 = v16 >> 8;
196        return true;
197    }
198    return false;
199}
200
201/* Return true if v32 is a valid 32-bit shifted immediate.  */
202static bool is_shimm32(uint32_t v32, int *cmode, int *imm8)
203{
204    if (v32 == (v32 & 0xff)) {
205        *cmode = 0x0;
206        *imm8 = v32 & 0xff;
207        return true;
208    } else if (v32 == (v32 & 0xff00)) {
209        *cmode = 0x2;
210        *imm8 = (v32 >> 8) & 0xff;
211        return true;
212    } else if (v32 == (v32 & 0xff0000)) {
213        *cmode = 0x4;
214        *imm8 = (v32 >> 16) & 0xff;
215        return true;
216    } else if (v32 == (v32 & 0xff000000)) {
217        *cmode = 0x6;
218        *imm8 = v32 >> 24;
219        return true;
220    }
221    return false;
222}
223
224/* Return true if v32 is a valid 32-bit shifting ones immediate.  */
225static bool is_soimm32(uint32_t v32, int *cmode, int *imm8)
226{
227    if ((v32 & 0xffff00ff) == 0xff) {
228        *cmode = 0xc;
229        *imm8 = (v32 >> 8) & 0xff;
230        return true;
231    } else if ((v32 & 0xff00ffff) == 0xffff) {
232        *cmode = 0xd;
233        *imm8 = (v32 >> 16) & 0xff;
234        return true;
235    }
236    return false;
237}
238
239/* Return true if v32 is a valid float32 immediate.  */
240static bool is_fimm32(uint32_t v32, int *cmode, int *imm8)
241{
242    if (extract32(v32, 0, 19) == 0
243        && (extract32(v32, 25, 6) == 0x20
244            || extract32(v32, 25, 6) == 0x1f)) {
245        *cmode = 0xf;
246        *imm8 = (extract32(v32, 31, 1) << 7)
247              | (extract32(v32, 25, 1) << 6)
248              | extract32(v32, 19, 6);
249        return true;
250    }
251    return false;
252}
253
254/* Return true if v64 is a valid float64 immediate.  */
255static bool is_fimm64(uint64_t v64, int *cmode, int *imm8)
256{
257    if (extract64(v64, 0, 48) == 0
258        && (extract64(v64, 54, 9) == 0x100
259            || extract64(v64, 54, 9) == 0x0ff)) {
260        *cmode = 0xf;
261        *imm8 = (extract64(v64, 63, 1) << 7)
262              | (extract64(v64, 54, 1) << 6)
263              | extract64(v64, 48, 6);
264        return true;
265    }
266    return false;
267}
268
269/*
270 * Return non-zero if v32 can be formed by MOVI+ORR.
271 * Place the parameters for MOVI in (cmode, imm8).
272 * Return the cmode for ORR; the imm8 can be had via extraction from v32.
273 */
274static int is_shimm32_pair(uint32_t v32, int *cmode, int *imm8)
275{
276    int i;
277
278    for (i = 6; i > 0; i -= 2) {
279        /* Mask out one byte we can add with ORR.  */
280        uint32_t tmp = v32 & ~(0xffu << (i * 4));
281        if (is_shimm32(tmp, cmode, imm8) ||
282            is_soimm32(tmp, cmode, imm8)) {
283            break;
284        }
285    }
286    return i;
287}
288
289/* Return true if V is a valid 16-bit or 32-bit shifted immediate.  */
290static bool is_shimm1632(uint32_t v32, int *cmode, int *imm8)
291{
292    if (v32 == deposit32(v32, 16, 16, v32)) {
293        return is_shimm16(v32, cmode, imm8);
294    } else {
295        return is_shimm32(v32, cmode, imm8);
296    }
297}
298
299static bool tcg_target_const_match(int64_t val, int ct,
300                                   TCGType type, TCGCond cond, int vece)
301{
302    if (ct & TCG_CT_CONST) {
303        return 1;
304    }
305    if (type == TCG_TYPE_I32) {
306        val = (int32_t)val;
307    }
308
309    if (ct & TCG_CT_CONST_CMP) {
310        if (is_tst_cond(cond)) {
311            ct |= TCG_CT_CONST_LIMM;
312        } else {
313            ct |= TCG_CT_CONST_AIMM;
314        }
315    }
316
317    if ((ct & TCG_CT_CONST_AIMM) && (is_aimm(val) || is_aimm(-val))) {
318        return 1;
319    }
320    if ((ct & TCG_CT_CONST_LIMM) && is_limm(val)) {
321        return 1;
322    }
323    if ((ct & TCG_CT_CONST_ZERO) && val == 0) {
324        return 1;
325    }
326    if ((ct & TCG_CT_CONST_MONE) && val == -1) {
327        return 1;
328    }
329
330    switch (ct & (TCG_CT_CONST_ORRI | TCG_CT_CONST_ANDI)) {
331    case 0:
332        break;
333    case TCG_CT_CONST_ANDI:
334        val = ~val;
335        /* fallthru */
336    case TCG_CT_CONST_ORRI:
337        if (val == deposit64(val, 32, 32, val)) {
338            int cmode, imm8;
339            return is_shimm1632(val, &cmode, &imm8);
340        }
341        break;
342    default:
343        /* Both bits should not be set for the same insn.  */
344        g_assert_not_reached();
345    }
346
347    return 0;
348}
349
350enum aarch64_cond_code {
351    COND_EQ = 0x0,
352    COND_NE = 0x1,
353    COND_CS = 0x2,     /* Unsigned greater or equal */
354    COND_HS = COND_CS, /* ALIAS greater or equal */
355    COND_CC = 0x3,     /* Unsigned less than */
356    COND_LO = COND_CC, /* ALIAS Lower */
357    COND_MI = 0x4,     /* Negative */
358    COND_PL = 0x5,     /* Zero or greater */
359    COND_VS = 0x6,     /* Overflow */
360    COND_VC = 0x7,     /* No overflow */
361    COND_HI = 0x8,     /* Unsigned greater than */
362    COND_LS = 0x9,     /* Unsigned less or equal */
363    COND_GE = 0xa,
364    COND_LT = 0xb,
365    COND_GT = 0xc,
366    COND_LE = 0xd,
367    COND_AL = 0xe,
368    COND_NV = 0xf, /* behaves like COND_AL here */
369};
370
371static const enum aarch64_cond_code tcg_cond_to_aarch64[] = {
372    [TCG_COND_EQ] = COND_EQ,
373    [TCG_COND_NE] = COND_NE,
374    [TCG_COND_LT] = COND_LT,
375    [TCG_COND_GE] = COND_GE,
376    [TCG_COND_LE] = COND_LE,
377    [TCG_COND_GT] = COND_GT,
378    /* unsigned */
379    [TCG_COND_LTU] = COND_LO,
380    [TCG_COND_GTU] = COND_HI,
381    [TCG_COND_GEU] = COND_HS,
382    [TCG_COND_LEU] = COND_LS,
383    /* bit test */
384    [TCG_COND_TSTEQ] = COND_EQ,
385    [TCG_COND_TSTNE] = COND_NE,
386};
387
388typedef enum {
389    LDST_ST = 0,    /* store */
390    LDST_LD = 1,    /* load */
391    LDST_LD_S_X = 2,  /* load and sign-extend into Xt */
392    LDST_LD_S_W = 3,  /* load and sign-extend into Wt */
393} AArch64LdstType;
394
395/* We encode the format of the insn into the beginning of the name, so that
396   we can have the preprocessor help "typecheck" the insn vs the output
397   function.  Arm didn't provide us with nice names for the formats, so we
398   use the section number of the architecture reference manual in which the
399   instruction group is described.  */
400typedef enum {
401    /* Compare and branch (immediate).  */
402    I3201_CBZ       = 0x34000000,
403    I3201_CBNZ      = 0x35000000,
404
405    /* Conditional branch (immediate).  */
406    I3202_B_C       = 0x54000000,
407
408    /* Test and branch (immediate).  */
409    I3205_TBZ       = 0x36000000,
410    I3205_TBNZ      = 0x37000000,
411
412    /* Unconditional branch (immediate).  */
413    I3206_B         = 0x14000000,
414    I3206_BL        = 0x94000000,
415
416    /* Unconditional branch (register).  */
417    I3207_BR        = 0xd61f0000,
418    I3207_BLR       = 0xd63f0000,
419    I3207_RET       = 0xd65f0000,
420
421    /* AdvSIMD load/store single structure.  */
422    I3303_LD1R      = 0x0d40c000,
423
424    /* Load literal for loading the address at pc-relative offset */
425    I3305_LDR       = 0x58000000,
426    I3305_LDR_v64   = 0x5c000000,
427    I3305_LDR_v128  = 0x9c000000,
428
429    /* Load/store exclusive. */
430    I3306_LDXP      = 0xc8600000,
431    I3306_STXP      = 0xc8200000,
432
433    /* Load/store register.  Described here as 3.3.12, but the helper
434       that emits them can transform to 3.3.10 or 3.3.13.  */
435    I3312_STRB      = 0x38000000 | LDST_ST << 22 | MO_8 << 30,
436    I3312_STRH      = 0x38000000 | LDST_ST << 22 | MO_16 << 30,
437    I3312_STRW      = 0x38000000 | LDST_ST << 22 | MO_32 << 30,
438    I3312_STRX      = 0x38000000 | LDST_ST << 22 | MO_64 << 30,
439
440    I3312_LDRB      = 0x38000000 | LDST_LD << 22 | MO_8 << 30,
441    I3312_LDRH      = 0x38000000 | LDST_LD << 22 | MO_16 << 30,
442    I3312_LDRW      = 0x38000000 | LDST_LD << 22 | MO_32 << 30,
443    I3312_LDRX      = 0x38000000 | LDST_LD << 22 | MO_64 << 30,
444
445    I3312_LDRSBW    = 0x38000000 | LDST_LD_S_W << 22 | MO_8 << 30,
446    I3312_LDRSHW    = 0x38000000 | LDST_LD_S_W << 22 | MO_16 << 30,
447
448    I3312_LDRSBX    = 0x38000000 | LDST_LD_S_X << 22 | MO_8 << 30,
449    I3312_LDRSHX    = 0x38000000 | LDST_LD_S_X << 22 | MO_16 << 30,
450    I3312_LDRSWX    = 0x38000000 | LDST_LD_S_X << 22 | MO_32 << 30,
451
452    I3312_LDRVS     = 0x3c000000 | LDST_LD << 22 | MO_32 << 30,
453    I3312_STRVS     = 0x3c000000 | LDST_ST << 22 | MO_32 << 30,
454
455    I3312_LDRVD     = 0x3c000000 | LDST_LD << 22 | MO_64 << 30,
456    I3312_STRVD     = 0x3c000000 | LDST_ST << 22 | MO_64 << 30,
457
458    I3312_LDRVQ     = 0x3c000000 | 3 << 22 | 0 << 30,
459    I3312_STRVQ     = 0x3c000000 | 2 << 22 | 0 << 30,
460
461    I3312_TO_I3310  = 0x00200800,
462    I3312_TO_I3313  = 0x01000000,
463
464    /* Load/store register pair instructions.  */
465    I3314_LDP       = 0x28400000,
466    I3314_STP       = 0x28000000,
467
468    /* Add/subtract immediate instructions.  */
469    I3401_ADDI      = 0x11000000,
470    I3401_ADDSI     = 0x31000000,
471    I3401_SUBI      = 0x51000000,
472    I3401_SUBSI     = 0x71000000,
473
474    /* Bitfield instructions.  */
475    I3402_BFM       = 0x33000000,
476    I3402_SBFM      = 0x13000000,
477    I3402_UBFM      = 0x53000000,
478
479    /* Extract instruction.  */
480    I3403_EXTR      = 0x13800000,
481
482    /* Logical immediate instructions.  */
483    I3404_ANDI      = 0x12000000,
484    I3404_ORRI      = 0x32000000,
485    I3404_EORI      = 0x52000000,
486    I3404_ANDSI     = 0x72000000,
487
488    /* Move wide immediate instructions.  */
489    I3405_MOVN      = 0x12800000,
490    I3405_MOVZ      = 0x52800000,
491    I3405_MOVK      = 0x72800000,
492
493    /* PC relative addressing instructions.  */
494    I3406_ADR       = 0x10000000,
495    I3406_ADRP      = 0x90000000,
496
497    /* Add/subtract extended register instructions. */
498    I3501_ADD       = 0x0b200000,
499
500    /* Add/subtract shifted register instructions (without a shift).  */
501    I3502_ADD       = 0x0b000000,
502    I3502_ADDS      = 0x2b000000,
503    I3502_SUB       = 0x4b000000,
504    I3502_SUBS      = 0x6b000000,
505
506    /* Add/subtract shifted register instructions (with a shift).  */
507    I3502S_ADD_LSL  = I3502_ADD,
508
509    /* Add/subtract with carry instructions.  */
510    I3503_ADC       = 0x1a000000,
511    I3503_SBC       = 0x5a000000,
512
513    /* Conditional select instructions.  */
514    I3506_CSEL      = 0x1a800000,
515    I3506_CSINC     = 0x1a800400,
516    I3506_CSINV     = 0x5a800000,
517    I3506_CSNEG     = 0x5a800400,
518
519    /* Data-processing (1 source) instructions.  */
520    I3507_CLZ       = 0x5ac01000,
521    I3507_RBIT      = 0x5ac00000,
522    I3507_REV       = 0x5ac00000, /* + size << 10 */
523
524    /* Data-processing (2 source) instructions.  */
525    I3508_LSLV      = 0x1ac02000,
526    I3508_LSRV      = 0x1ac02400,
527    I3508_ASRV      = 0x1ac02800,
528    I3508_RORV      = 0x1ac02c00,
529    I3508_SMULH     = 0x9b407c00,
530    I3508_UMULH     = 0x9bc07c00,
531    I3508_UDIV      = 0x1ac00800,
532    I3508_SDIV      = 0x1ac00c00,
533
534    /* Data-processing (3 source) instructions.  */
535    I3509_MADD      = 0x1b000000,
536    I3509_MSUB      = 0x1b008000,
537
538    /* Logical shifted register instructions (without a shift).  */
539    I3510_AND       = 0x0a000000,
540    I3510_BIC       = 0x0a200000,
541    I3510_ORR       = 0x2a000000,
542    I3510_ORN       = 0x2a200000,
543    I3510_EOR       = 0x4a000000,
544    I3510_EON       = 0x4a200000,
545    I3510_ANDS      = 0x6a000000,
546
547    /* Logical shifted register instructions (with a shift).  */
548    I3502S_AND_LSR  = I3510_AND | (1 << 22),
549
550    /* AdvSIMD copy */
551    I3605_DUP      = 0x0e000400,
552    I3605_INS      = 0x4e001c00,
553    I3605_UMOV     = 0x0e003c00,
554
555    /* AdvSIMD modified immediate */
556    I3606_MOVI      = 0x0f000400,
557    I3606_MVNI      = 0x2f000400,
558    I3606_BIC       = 0x2f001400,
559    I3606_ORR       = 0x0f001400,
560
561    /* AdvSIMD scalar shift by immediate */
562    I3609_SSHR      = 0x5f000400,
563    I3609_SSRA      = 0x5f001400,
564    I3609_SHL       = 0x5f005400,
565    I3609_USHR      = 0x7f000400,
566    I3609_USRA      = 0x7f001400,
567    I3609_SLI       = 0x7f005400,
568
569    /* AdvSIMD scalar three same */
570    I3611_SQADD     = 0x5e200c00,
571    I3611_SQSUB     = 0x5e202c00,
572    I3611_CMGT      = 0x5e203400,
573    I3611_CMGE      = 0x5e203c00,
574    I3611_SSHL      = 0x5e204400,
575    I3611_ADD       = 0x5e208400,
576    I3611_CMTST     = 0x5e208c00,
577    I3611_UQADD     = 0x7e200c00,
578    I3611_UQSUB     = 0x7e202c00,
579    I3611_CMHI      = 0x7e203400,
580    I3611_CMHS      = 0x7e203c00,
581    I3611_USHL      = 0x7e204400,
582    I3611_SUB       = 0x7e208400,
583    I3611_CMEQ      = 0x7e208c00,
584
585    /* AdvSIMD scalar two-reg misc */
586    I3612_CMGT0     = 0x5e208800,
587    I3612_CMEQ0     = 0x5e209800,
588    I3612_CMLT0     = 0x5e20a800,
589    I3612_ABS       = 0x5e20b800,
590    I3612_CMGE0     = 0x7e208800,
591    I3612_CMLE0     = 0x7e209800,
592    I3612_NEG       = 0x7e20b800,
593
594    /* AdvSIMD shift by immediate */
595    I3614_SSHR      = 0x0f000400,
596    I3614_SSRA      = 0x0f001400,
597    I3614_SHL       = 0x0f005400,
598    I3614_SLI       = 0x2f005400,
599    I3614_USHR      = 0x2f000400,
600    I3614_USRA      = 0x2f001400,
601
602    /* AdvSIMD three same.  */
603    I3616_ADD       = 0x0e208400,
604    I3616_AND       = 0x0e201c00,
605    I3616_BIC       = 0x0e601c00,
606    I3616_BIF       = 0x2ee01c00,
607    I3616_BIT       = 0x2ea01c00,
608    I3616_BSL       = 0x2e601c00,
609    I3616_EOR       = 0x2e201c00,
610    I3616_MUL       = 0x0e209c00,
611    I3616_ORR       = 0x0ea01c00,
612    I3616_ORN       = 0x0ee01c00,
613    I3616_SUB       = 0x2e208400,
614    I3616_CMGT      = 0x0e203400,
615    I3616_CMGE      = 0x0e203c00,
616    I3616_CMTST     = 0x0e208c00,
617    I3616_CMHI      = 0x2e203400,
618    I3616_CMHS      = 0x2e203c00,
619    I3616_CMEQ      = 0x2e208c00,
620    I3616_SMAX      = 0x0e206400,
621    I3616_SMIN      = 0x0e206c00,
622    I3616_SSHL      = 0x0e204400,
623    I3616_SQADD     = 0x0e200c00,
624    I3616_SQSUB     = 0x0e202c00,
625    I3616_UMAX      = 0x2e206400,
626    I3616_UMIN      = 0x2e206c00,
627    I3616_UQADD     = 0x2e200c00,
628    I3616_UQSUB     = 0x2e202c00,
629    I3616_USHL      = 0x2e204400,
630
631    /* AdvSIMD two-reg misc.  */
632    I3617_CMGT0     = 0x0e208800,
633    I3617_CMEQ0     = 0x0e209800,
634    I3617_CMLT0     = 0x0e20a800,
635    I3617_CMGE0     = 0x2e208800,
636    I3617_CMLE0     = 0x2e209800,
637    I3617_NOT       = 0x2e205800,
638    I3617_ABS       = 0x0e20b800,
639    I3617_NEG       = 0x2e20b800,
640
641    /* System instructions.  */
642    NOP             = 0xd503201f,
643    DMB_ISH         = 0xd50338bf,
644    DMB_LD          = 0x00000100,
645    DMB_ST          = 0x00000200,
646
647    BTI_C           = 0xd503245f,
648    BTI_J           = 0xd503249f,
649    BTI_JC          = 0xd50324df,
650} AArch64Insn;
651
652static inline uint32_t tcg_in32(TCGContext *s)
653{
654    uint32_t v = *(uint32_t *)s->code_ptr;
655    return v;
656}
657
658/* Emit an opcode with "type-checking" of the format.  */
659#define tcg_out_insn(S, FMT, OP, ...) \
660    glue(tcg_out_insn_,FMT)(S, glue(glue(glue(I,FMT),_),OP), ## __VA_ARGS__)
661
662static void tcg_out_insn_3303(TCGContext *s, AArch64Insn insn, bool q,
663                              TCGReg rt, TCGReg rn, unsigned size)
664{
665    tcg_out32(s, insn | (rt & 0x1f) | (rn << 5) | (size << 10) | (q << 30));
666}
667
668static void tcg_out_insn_3305(TCGContext *s, AArch64Insn insn,
669                              int imm19, TCGReg rt)
670{
671    tcg_out32(s, insn | (imm19 & 0x7ffff) << 5 | rt);
672}
673
674static void tcg_out_insn_3306(TCGContext *s, AArch64Insn insn, TCGReg rs,
675                              TCGReg rt, TCGReg rt2, TCGReg rn)
676{
677    tcg_out32(s, insn | rs << 16 | rt2 << 10 | rn << 5 | rt);
678}
679
680static void tcg_out_insn_3201(TCGContext *s, AArch64Insn insn, TCGType ext,
681                              TCGReg rt, int imm19)
682{
683    tcg_out32(s, insn | ext << 31 | (imm19 & 0x7ffff) << 5 | rt);
684}
685
686static void tcg_out_insn_3202(TCGContext *s, AArch64Insn insn,
687                              TCGCond c, int imm19)
688{
689    tcg_out32(s, insn | tcg_cond_to_aarch64[c] | (imm19 & 0x7ffff) << 5);
690}
691
692static void tcg_out_insn_3205(TCGContext *s, AArch64Insn insn,
693                              TCGReg rt, int imm6, int imm14)
694{
695    insn |= (imm6 & 0x20) << (31 - 5);
696    insn |= (imm6 & 0x1f) << 19;
697    tcg_out32(s, insn | (imm14 & 0x3fff) << 5 | rt);
698}
699
700static void tcg_out_insn_3206(TCGContext *s, AArch64Insn insn, int imm26)
701{
702    tcg_out32(s, insn | (imm26 & 0x03ffffff));
703}
704
705static void tcg_out_insn_3207(TCGContext *s, AArch64Insn insn, TCGReg rn)
706{
707    tcg_out32(s, insn | rn << 5);
708}
709
710static void tcg_out_insn_3314(TCGContext *s, AArch64Insn insn,
711                              TCGReg r1, TCGReg r2, TCGReg rn,
712                              tcg_target_long ofs, bool pre, bool w)
713{
714    insn |= 1u << 31; /* ext */
715    insn |= pre << 24;
716    insn |= w << 23;
717
718    tcg_debug_assert(ofs >= -0x200 && ofs < 0x200 && (ofs & 7) == 0);
719    insn |= (ofs & (0x7f << 3)) << (15 - 3);
720
721    tcg_out32(s, insn | r2 << 10 | rn << 5 | r1);
722}
723
724static void tcg_out_insn_3401(TCGContext *s, AArch64Insn insn, TCGType ext,
725                              TCGReg rd, TCGReg rn, uint64_t aimm)
726{
727    if (aimm > 0xfff) {
728        tcg_debug_assert((aimm & 0xfff) == 0);
729        aimm >>= 12;
730        tcg_debug_assert(aimm <= 0xfff);
731        aimm |= 1 << 12;  /* apply LSL 12 */
732    }
733    tcg_out32(s, insn | ext << 31 | aimm << 10 | rn << 5 | rd);
734}
735
736/* This function can be used for both 3.4.2 (Bitfield) and 3.4.4
737   (Logical immediate).  Both insn groups have N, IMMR and IMMS fields
738   that feed the DecodeBitMasks pseudo function.  */
739static void tcg_out_insn_3402(TCGContext *s, AArch64Insn insn, TCGType ext,
740                              TCGReg rd, TCGReg rn, int n, int immr, int imms)
741{
742    tcg_out32(s, insn | ext << 31 | n << 22 | immr << 16 | imms << 10
743              | rn << 5 | rd);
744}
745
746#define tcg_out_insn_3404  tcg_out_insn_3402
747
748static void tcg_out_insn_3403(TCGContext *s, AArch64Insn insn, TCGType ext,
749                              TCGReg rd, TCGReg rn, TCGReg rm, int imms)
750{
751    tcg_out32(s, insn | ext << 31 | ext << 22 | rm << 16 | imms << 10
752              | rn << 5 | rd);
753}
754
755/* This function is used for the Move (wide immediate) instruction group.
756   Note that SHIFT is a full shift count, not the 2 bit HW field. */
757static void tcg_out_insn_3405(TCGContext *s, AArch64Insn insn, TCGType ext,
758                              TCGReg rd, uint16_t half, unsigned shift)
759{
760    tcg_debug_assert((shift & ~0x30) == 0);
761    tcg_out32(s, insn | ext << 31 | shift << (21 - 4) | half << 5 | rd);
762}
763
764static void tcg_out_insn_3406(TCGContext *s, AArch64Insn insn,
765                              TCGReg rd, int64_t disp)
766{
767    tcg_out32(s, insn | (disp & 3) << 29 | (disp & 0x1ffffc) << (5 - 2) | rd);
768}
769
770static inline void tcg_out_insn_3501(TCGContext *s, AArch64Insn insn,
771                                     TCGType sf, TCGReg rd, TCGReg rn,
772                                     TCGReg rm, int opt, int imm3)
773{
774    tcg_out32(s, insn | sf << 31 | rm << 16 | opt << 13 |
775              imm3 << 10 | rn << 5 | rd);
776}
777
778/* This function is for both 3.5.2 (Add/Subtract shifted register), for
779   the rare occasion when we actually want to supply a shift amount.  */
780static inline void tcg_out_insn_3502S(TCGContext *s, AArch64Insn insn,
781                                      TCGType ext, TCGReg rd, TCGReg rn,
782                                      TCGReg rm, int imm6)
783{
784    tcg_out32(s, insn | ext << 31 | rm << 16 | imm6 << 10 | rn << 5 | rd);
785}
786
787/* This function is for 3.5.2 (Add/subtract shifted register),
788   and 3.5.10 (Logical shifted register), for the vast majorty of cases
789   when we don't want to apply a shift.  Thus it can also be used for
790   3.5.3 (Add/subtract with carry) and 3.5.8 (Data processing 2 source).  */
791static void tcg_out_insn_3502(TCGContext *s, AArch64Insn insn, TCGType ext,
792                              TCGReg rd, TCGReg rn, TCGReg rm)
793{
794    tcg_out32(s, insn | ext << 31 | rm << 16 | rn << 5 | rd);
795}
796
797#define tcg_out_insn_3503  tcg_out_insn_3502
798#define tcg_out_insn_3508  tcg_out_insn_3502
799#define tcg_out_insn_3510  tcg_out_insn_3502
800
801static void tcg_out_insn_3506(TCGContext *s, AArch64Insn insn, TCGType ext,
802                              TCGReg rd, TCGReg rn, TCGReg rm, TCGCond c)
803{
804    tcg_out32(s, insn | ext << 31 | rm << 16 | rn << 5 | rd
805              | tcg_cond_to_aarch64[c] << 12);
806}
807
808static void tcg_out_insn_3507(TCGContext *s, AArch64Insn insn, TCGType ext,
809                              TCGReg rd, TCGReg rn)
810{
811    tcg_out32(s, insn | ext << 31 | rn << 5 | rd);
812}
813
814static void tcg_out_insn_3509(TCGContext *s, AArch64Insn insn, TCGType ext,
815                              TCGReg rd, TCGReg rn, TCGReg rm, TCGReg ra)
816{
817    tcg_out32(s, insn | ext << 31 | rm << 16 | ra << 10 | rn << 5 | rd);
818}
819
820static void tcg_out_insn_3605(TCGContext *s, AArch64Insn insn, bool q,
821                              TCGReg rd, TCGReg rn, int dst_idx, int src_idx)
822{
823    /* Note that bit 11 set means general register input.  Therefore
824       we can handle both register sets with one function.  */
825    tcg_out32(s, insn | q << 30 | (dst_idx << 16) | (src_idx << 11)
826              | (rd & 0x1f) | (~rn & 0x20) << 6 | (rn & 0x1f) << 5);
827}
828
829static void tcg_out_insn_3606(TCGContext *s, AArch64Insn insn, bool q,
830                              TCGReg rd, bool op, int cmode, uint8_t imm8)
831{
832    tcg_out32(s, insn | q << 30 | op << 29 | cmode << 12 | (rd & 0x1f)
833              | (imm8 & 0xe0) << (16 - 5) | (imm8 & 0x1f) << 5);
834}
835
836static void tcg_out_insn_3609(TCGContext *s, AArch64Insn insn,
837                              TCGReg rd, TCGReg rn, unsigned immhb)
838{
839    tcg_out32(s, insn | immhb << 16 | (rn & 0x1f) << 5 | (rd & 0x1f));
840}
841
842static void tcg_out_insn_3611(TCGContext *s, AArch64Insn insn,
843                              unsigned size, TCGReg rd, TCGReg rn, TCGReg rm)
844{
845    tcg_out32(s, insn | (size << 22) | (rm & 0x1f) << 16
846              | (rn & 0x1f) << 5 | (rd & 0x1f));
847}
848
849static void tcg_out_insn_3612(TCGContext *s, AArch64Insn insn,
850                              unsigned size, TCGReg rd, TCGReg rn)
851{
852    tcg_out32(s, insn | (size << 22) | (rn & 0x1f) << 5 | (rd & 0x1f));
853}
854
855static void tcg_out_insn_3614(TCGContext *s, AArch64Insn insn, bool q,
856                              TCGReg rd, TCGReg rn, unsigned immhb)
857{
858    tcg_out32(s, insn | q << 30 | immhb << 16
859              | (rn & 0x1f) << 5 | (rd & 0x1f));
860}
861
862static void tcg_out_insn_3616(TCGContext *s, AArch64Insn insn, bool q,
863                              unsigned size, TCGReg rd, TCGReg rn, TCGReg rm)
864{
865    tcg_out32(s, insn | q << 30 | (size << 22) | (rm & 0x1f) << 16
866              | (rn & 0x1f) << 5 | (rd & 0x1f));
867}
868
869static void tcg_out_insn_3617(TCGContext *s, AArch64Insn insn, bool q,
870                              unsigned size, TCGReg rd, TCGReg rn)
871{
872    tcg_out32(s, insn | q << 30 | (size << 22)
873              | (rn & 0x1f) << 5 | (rd & 0x1f));
874}
875
876static void tcg_out_insn_3310(TCGContext *s, AArch64Insn insn,
877                              TCGReg rd, TCGReg base, TCGType ext,
878                              TCGReg regoff)
879{
880    /* Note the AArch64Insn constants above are for C3.3.12.  Adjust.  */
881    tcg_out32(s, insn | I3312_TO_I3310 | regoff << 16 |
882              0x4000 | ext << 13 | base << 5 | (rd & 0x1f));
883}
884
885static void tcg_out_insn_3312(TCGContext *s, AArch64Insn insn,
886                              TCGReg rd, TCGReg rn, intptr_t offset)
887{
888    tcg_out32(s, insn | (offset & 0x1ff) << 12 | rn << 5 | (rd & 0x1f));
889}
890
891static void tcg_out_insn_3313(TCGContext *s, AArch64Insn insn,
892                              TCGReg rd, TCGReg rn, uintptr_t scaled_uimm)
893{
894    /* Note the AArch64Insn constants above are for C3.3.12.  Adjust.  */
895    tcg_out32(s, insn | I3312_TO_I3313 | scaled_uimm << 10
896              | rn << 5 | (rd & 0x1f));
897}
898
899static void tcg_out_bti(TCGContext *s, AArch64Insn insn)
900{
901    /*
902     * While BTI insns are nops on hosts without FEAT_BTI,
903     * there is no point in emitting them in that case either.
904     */
905    if (cpuinfo & CPUINFO_BTI) {
906        tcg_out32(s, insn);
907    }
908}
909
910/* Register to register move using ORR (shifted register with no shift). */
911static void tcg_out_movr(TCGContext *s, TCGType ext, TCGReg rd, TCGReg rm)
912{
913    tcg_out_insn(s, 3510, ORR, ext, rd, TCG_REG_XZR, rm);
914}
915
916/* Register to register move using ADDI (move to/from SP).  */
917static void tcg_out_movr_sp(TCGContext *s, TCGType ext, TCGReg rd, TCGReg rn)
918{
919    tcg_out_insn(s, 3401, ADDI, ext, rd, rn, 0);
920}
921
922/* This function is used for the Logical (immediate) instruction group.
923   The value of LIMM must satisfy IS_LIMM.  See the comment above about
924   only supporting simplified logical immediates.  */
925static void tcg_out_logicali(TCGContext *s, AArch64Insn insn, TCGType ext,
926                             TCGReg rd, TCGReg rn, uint64_t limm)
927{
928    unsigned h, l, r, c;
929
930    tcg_debug_assert(is_limm(limm));
931
932    h = clz64(limm);
933    l = ctz64(limm);
934    if (l == 0) {
935        r = 0;                  /* form 0....01....1 */
936        c = ctz64(~limm) - 1;
937        if (h == 0) {
938            r = clz64(~limm);   /* form 1..10..01..1 */
939            c += r;
940        }
941    } else {
942        r = 64 - l;             /* form 1....10....0 or 0..01..10..0 */
943        c = r - h - 1;
944    }
945    if (ext == TCG_TYPE_I32) {
946        r &= 31;
947        c &= 31;
948    }
949
950    tcg_out_insn_3404(s, insn, ext, rd, rn, ext, r, c);
951}
952
953static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
954                             TCGReg rd, int64_t v64)
955{
956    bool q = type == TCG_TYPE_V128;
957    int cmode, imm8, i;
958
959    /* Test all bytes equal first.  */
960    if (vece == MO_8) {
961        imm8 = (uint8_t)v64;
962        tcg_out_insn(s, 3606, MOVI, q, rd, 0, 0xe, imm8);
963        return;
964    }
965
966    /*
967     * Test all bytes 0x00 or 0xff second.  This can match cases that
968     * might otherwise take 2 or 3 insns for MO_16 or MO_32 below.
969     */
970    for (i = imm8 = 0; i < 8; i++) {
971        uint8_t byte = v64 >> (i * 8);
972        if (byte == 0xff) {
973            imm8 |= 1 << i;
974        } else if (byte != 0) {
975            goto fail_bytes;
976        }
977    }
978    tcg_out_insn(s, 3606, MOVI, q, rd, 1, 0xe, imm8);
979    return;
980 fail_bytes:
981
982    /*
983     * Tests for various replications.  For each element width, if we
984     * cannot find an expansion there's no point checking a larger
985     * width because we already know by replication it cannot match.
986     */
987    if (vece == MO_16) {
988        uint16_t v16 = v64;
989
990        if (is_shimm16(v16, &cmode, &imm8)) {
991            tcg_out_insn(s, 3606, MOVI, q, rd, 0, cmode, imm8);
992            return;
993        }
994        if (is_shimm16(~v16, &cmode, &imm8)) {
995            tcg_out_insn(s, 3606, MVNI, q, rd, 0, cmode, imm8);
996            return;
997        }
998
999        /*
1000         * Otherwise, all remaining constants can be loaded in two insns:
1001         * rd = v16 & 0xff, rd |= v16 & 0xff00.
1002         */
1003        tcg_out_insn(s, 3606, MOVI, q, rd, 0, 0x8, v16 & 0xff);
1004        tcg_out_insn(s, 3606, ORR, q, rd, 0, 0xa, v16 >> 8);
1005        return;
1006    } else if (vece == MO_32) {
1007        uint32_t v32 = v64;
1008        uint32_t n32 = ~v32;
1009
1010        if (is_shimm32(v32, &cmode, &imm8) ||
1011            is_soimm32(v32, &cmode, &imm8) ||
1012            is_fimm32(v32, &cmode, &imm8)) {
1013            tcg_out_insn(s, 3606, MOVI, q, rd, 0, cmode, imm8);
1014            return;
1015        }
1016        if (is_shimm32(n32, &cmode, &imm8) ||
1017            is_soimm32(n32, &cmode, &imm8)) {
1018            tcg_out_insn(s, 3606, MVNI, q, rd, 0, cmode, imm8);
1019            return;
1020        }
1021
1022        /*
1023         * Restrict the set of constants to those we can load with
1024         * two instructions.  Others we load from the pool.
1025         */
1026        i = is_shimm32_pair(v32, &cmode, &imm8);
1027        if (i) {
1028            tcg_out_insn(s, 3606, MOVI, q, rd, 0, cmode, imm8);
1029            tcg_out_insn(s, 3606, ORR, q, rd, 0, i, extract32(v32, i * 4, 8));
1030            return;
1031        }
1032        i = is_shimm32_pair(n32, &cmode, &imm8);
1033        if (i) {
1034            tcg_out_insn(s, 3606, MVNI, q, rd, 0, cmode, imm8);
1035            tcg_out_insn(s, 3606, BIC, q, rd, 0, i, extract32(n32, i * 4, 8));
1036            return;
1037        }
1038    } else if (is_fimm64(v64, &cmode, &imm8)) {
1039        tcg_out_insn(s, 3606, MOVI, q, rd, 1, cmode, imm8);
1040        return;
1041    }
1042
1043    /*
1044     * As a last resort, load from the constant pool.  Sadly there
1045     * is no LD1R (literal), so store the full 16-byte vector.
1046     */
1047    if (type == TCG_TYPE_V128) {
1048        new_pool_l2(s, R_AARCH64_CONDBR19, s->code_ptr, 0, v64, v64);
1049        tcg_out_insn(s, 3305, LDR_v128, 0, rd);
1050    } else {
1051        new_pool_label(s, v64, R_AARCH64_CONDBR19, s->code_ptr, 0);
1052        tcg_out_insn(s, 3305, LDR_v64, 0, rd);
1053    }
1054}
1055
1056static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
1057                            TCGReg rd, TCGReg rs)
1058{
1059    int is_q = type - TCG_TYPE_V64;
1060    tcg_out_insn(s, 3605, DUP, is_q, rd, rs, 1 << vece, 0);
1061    return true;
1062}
1063
1064static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
1065                             TCGReg r, TCGReg base, intptr_t offset)
1066{
1067    TCGReg temp = TCG_REG_TMP0;
1068
1069    if (offset < -0xffffff || offset > 0xffffff) {
1070        tcg_out_movi(s, TCG_TYPE_PTR, temp, offset);
1071        tcg_out_insn(s, 3502, ADD, 1, temp, temp, base);
1072        base = temp;
1073    } else {
1074        AArch64Insn add_insn = I3401_ADDI;
1075
1076        if (offset < 0) {
1077            add_insn = I3401_SUBI;
1078            offset = -offset;
1079        }
1080        if (offset & 0xfff000) {
1081            tcg_out_insn_3401(s, add_insn, 1, temp, base, offset & 0xfff000);
1082            base = temp;
1083        }
1084        if (offset & 0xfff) {
1085            tcg_out_insn_3401(s, add_insn, 1, temp, base, offset & 0xfff);
1086            base = temp;
1087        }
1088    }
1089    tcg_out_insn(s, 3303, LD1R, type == TCG_TYPE_V128, r, base, vece);
1090    return true;
1091}
1092
1093static void tcg_out_movi(TCGContext *s, TCGType type, TCGReg rd,
1094                         tcg_target_long value)
1095{
1096    tcg_target_long svalue = value;
1097    tcg_target_long ivalue = ~value;
1098    tcg_target_long t0, t1, t2;
1099    int s0, s1;
1100    AArch64Insn opc;
1101
1102    switch (type) {
1103    case TCG_TYPE_I32:
1104    case TCG_TYPE_I64:
1105        tcg_debug_assert(rd < 32);
1106        break;
1107    default:
1108        g_assert_not_reached();
1109    }
1110
1111    /* For 32-bit values, discard potential garbage in value.  For 64-bit
1112       values within [2**31, 2**32-1], we can create smaller sequences by
1113       interpreting this as a negative 32-bit number, while ensuring that
1114       the high 32 bits are cleared by setting SF=0.  */
1115    if (type == TCG_TYPE_I32 || (value & ~0xffffffffull) == 0) {
1116        svalue = (int32_t)value;
1117        value = (uint32_t)value;
1118        ivalue = (uint32_t)ivalue;
1119        type = TCG_TYPE_I32;
1120    }
1121
1122    /* Speed things up by handling the common case of small positive
1123       and negative values specially.  */
1124    if ((value & ~0xffffull) == 0) {
1125        tcg_out_insn(s, 3405, MOVZ, type, rd, value, 0);
1126        return;
1127    } else if ((ivalue & ~0xffffull) == 0) {
1128        tcg_out_insn(s, 3405, MOVN, type, rd, ivalue, 0);
1129        return;
1130    }
1131
1132    /* Check for bitfield immediates.  For the benefit of 32-bit quantities,
1133       use the sign-extended value.  That lets us match rotated values such
1134       as 0xff0000ff with the same 64-bit logic matching 0xffffffffff0000ff. */
1135    if (is_limm(svalue)) {
1136        tcg_out_logicali(s, I3404_ORRI, type, rd, TCG_REG_XZR, svalue);
1137        return;
1138    }
1139
1140    /* Look for host pointer values within 4G of the PC.  This happens
1141       often when loading pointers to QEMU's own data structures.  */
1142    if (type == TCG_TYPE_I64) {
1143        intptr_t src_rx = (intptr_t)tcg_splitwx_to_rx(s->code_ptr);
1144        tcg_target_long disp = value - src_rx;
1145        if (disp == sextract64(disp, 0, 21)) {
1146            tcg_out_insn(s, 3406, ADR, rd, disp);
1147            return;
1148        }
1149        disp = (value >> 12) - (src_rx >> 12);
1150        if (disp == sextract64(disp, 0, 21)) {
1151            tcg_out_insn(s, 3406, ADRP, rd, disp);
1152            if (value & 0xfff) {
1153                tcg_out_insn(s, 3401, ADDI, type, rd, rd, value & 0xfff);
1154            }
1155            return;
1156        }
1157    }
1158
1159    /* Would it take fewer insns to begin with MOVN?  */
1160    if (ctpop64(value) >= 32) {
1161        t0 = ivalue;
1162        opc = I3405_MOVN;
1163    } else {
1164        t0 = value;
1165        opc = I3405_MOVZ;
1166    }
1167    s0 = ctz64(t0) & (63 & -16);
1168    t1 = t0 & ~(0xffffull << s0);
1169    s1 = ctz64(t1) & (63 & -16);
1170    t2 = t1 & ~(0xffffull << s1);
1171    if (t2 == 0) {
1172        tcg_out_insn_3405(s, opc, type, rd, t0 >> s0, s0);
1173        if (t1 != 0) {
1174            tcg_out_insn(s, 3405, MOVK, type, rd, value >> s1, s1);
1175        }
1176        return;
1177    }
1178
1179    /* For more than 2 insns, dump it into the constant pool.  */
1180    new_pool_label(s, value, R_AARCH64_CONDBR19, s->code_ptr, 0);
1181    tcg_out_insn(s, 3305, LDR, 0, rd);
1182}
1183
1184static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
1185{
1186    return false;
1187}
1188
1189static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
1190                             tcg_target_long imm)
1191{
1192    /* This function is only used for passing structs by reference. */
1193    g_assert_not_reached();
1194}
1195
1196/* Define something more legible for general use.  */
1197#define tcg_out_ldst_r  tcg_out_insn_3310
1198
1199static void tcg_out_ldst(TCGContext *s, AArch64Insn insn, TCGReg rd,
1200                         TCGReg rn, intptr_t offset, int lgsize)
1201{
1202    /* If the offset is naturally aligned and in range, then we can
1203       use the scaled uimm12 encoding */
1204    if (offset >= 0 && !(offset & ((1 << lgsize) - 1))) {
1205        uintptr_t scaled_uimm = offset >> lgsize;
1206        if (scaled_uimm <= 0xfff) {
1207            tcg_out_insn_3313(s, insn, rd, rn, scaled_uimm);
1208            return;
1209        }
1210    }
1211
1212    /* Small signed offsets can use the unscaled encoding.  */
1213    if (offset >= -256 && offset < 256) {
1214        tcg_out_insn_3312(s, insn, rd, rn, offset);
1215        return;
1216    }
1217
1218    /* Worst-case scenario, move offset to temp register, use reg offset.  */
1219    tcg_out_movi(s, TCG_TYPE_I64, TCG_REG_TMP0, offset);
1220    tcg_out_ldst_r(s, insn, rd, rn, TCG_TYPE_I64, TCG_REG_TMP0);
1221}
1222
1223static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
1224{
1225    if (ret == arg) {
1226        return true;
1227    }
1228    switch (type) {
1229    case TCG_TYPE_I32:
1230    case TCG_TYPE_I64:
1231        if (ret < 32 && arg < 32) {
1232            tcg_out_movr(s, type, ret, arg);
1233            break;
1234        } else if (ret < 32) {
1235            tcg_out_insn(s, 3605, UMOV, type, ret, arg, 0, 0);
1236            break;
1237        } else if (arg < 32) {
1238            tcg_out_insn(s, 3605, INS, 0, ret, arg, 4 << type, 0);
1239            break;
1240        }
1241        /* FALLTHRU */
1242
1243    case TCG_TYPE_V64:
1244        tcg_debug_assert(ret >= 32 && arg >= 32);
1245        tcg_out_insn(s, 3616, ORR, 0, 0, ret, arg, arg);
1246        break;
1247    case TCG_TYPE_V128:
1248        tcg_debug_assert(ret >= 32 && arg >= 32);
1249        tcg_out_insn(s, 3616, ORR, 1, 0, ret, arg, arg);
1250        break;
1251
1252    default:
1253        g_assert_not_reached();
1254    }
1255    return true;
1256}
1257
1258static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg ret,
1259                       TCGReg base, intptr_t ofs)
1260{
1261    AArch64Insn insn;
1262    int lgsz;
1263
1264    switch (type) {
1265    case TCG_TYPE_I32:
1266        insn = (ret < 32 ? I3312_LDRW : I3312_LDRVS);
1267        lgsz = 2;
1268        break;
1269    case TCG_TYPE_I64:
1270        insn = (ret < 32 ? I3312_LDRX : I3312_LDRVD);
1271        lgsz = 3;
1272        break;
1273    case TCG_TYPE_V64:
1274        insn = I3312_LDRVD;
1275        lgsz = 3;
1276        break;
1277    case TCG_TYPE_V128:
1278        insn = I3312_LDRVQ;
1279        lgsz = 4;
1280        break;
1281    default:
1282        g_assert_not_reached();
1283    }
1284    tcg_out_ldst(s, insn, ret, base, ofs, lgsz);
1285}
1286
1287static void tcg_out_st(TCGContext *s, TCGType type, TCGReg src,
1288                       TCGReg base, intptr_t ofs)
1289{
1290    AArch64Insn insn;
1291    int lgsz;
1292
1293    switch (type) {
1294    case TCG_TYPE_I32:
1295        insn = (src < 32 ? I3312_STRW : I3312_STRVS);
1296        lgsz = 2;
1297        break;
1298    case TCG_TYPE_I64:
1299        insn = (src < 32 ? I3312_STRX : I3312_STRVD);
1300        lgsz = 3;
1301        break;
1302    case TCG_TYPE_V64:
1303        insn = I3312_STRVD;
1304        lgsz = 3;
1305        break;
1306    case TCG_TYPE_V128:
1307        insn = I3312_STRVQ;
1308        lgsz = 4;
1309        break;
1310    default:
1311        g_assert_not_reached();
1312    }
1313    tcg_out_ldst(s, insn, src, base, ofs, lgsz);
1314}
1315
1316static inline bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
1317                               TCGReg base, intptr_t ofs)
1318{
1319    if (type <= TCG_TYPE_I64 && val == 0) {
1320        tcg_out_st(s, type, TCG_REG_XZR, base, ofs);
1321        return true;
1322    }
1323    return false;
1324}
1325
1326static inline void tcg_out_bfm(TCGContext *s, TCGType ext, TCGReg rd,
1327                               TCGReg rn, unsigned int a, unsigned int b)
1328{
1329    tcg_out_insn(s, 3402, BFM, ext, rd, rn, ext, a, b);
1330}
1331
1332static inline void tcg_out_ubfm(TCGContext *s, TCGType ext, TCGReg rd,
1333                                TCGReg rn, unsigned int a, unsigned int b)
1334{
1335    tcg_out_insn(s, 3402, UBFM, ext, rd, rn, ext, a, b);
1336}
1337
1338static inline void tcg_out_sbfm(TCGContext *s, TCGType ext, TCGReg rd,
1339                                TCGReg rn, unsigned int a, unsigned int b)
1340{
1341    tcg_out_insn(s, 3402, SBFM, ext, rd, rn, ext, a, b);
1342}
1343
1344static inline void tcg_out_extr(TCGContext *s, TCGType ext, TCGReg rd,
1345                                TCGReg rn, TCGReg rm, unsigned int a)
1346{
1347    tcg_out_insn(s, 3403, EXTR, ext, rd, rn, rm, a);
1348}
1349
1350static inline void tcg_out_shl(TCGContext *s, TCGType ext,
1351                               TCGReg rd, TCGReg rn, unsigned int m)
1352{
1353    int bits = ext ? 64 : 32;
1354    int max = bits - 1;
1355    tcg_out_ubfm(s, ext, rd, rn, (bits - m) & max, (max - m) & max);
1356}
1357
1358static inline void tcg_out_shr(TCGContext *s, TCGType ext,
1359                               TCGReg rd, TCGReg rn, unsigned int m)
1360{
1361    int max = ext ? 63 : 31;
1362    tcg_out_ubfm(s, ext, rd, rn, m & max, max);
1363}
1364
1365static inline void tcg_out_sar(TCGContext *s, TCGType ext,
1366                               TCGReg rd, TCGReg rn, unsigned int m)
1367{
1368    int max = ext ? 63 : 31;
1369    tcg_out_sbfm(s, ext, rd, rn, m & max, max);
1370}
1371
1372static inline void tcg_out_rotr(TCGContext *s, TCGType ext,
1373                                TCGReg rd, TCGReg rn, unsigned int m)
1374{
1375    int max = ext ? 63 : 31;
1376    tcg_out_extr(s, ext, rd, rn, rn, m & max);
1377}
1378
1379static inline void tcg_out_rotl(TCGContext *s, TCGType ext,
1380                                TCGReg rd, TCGReg rn, unsigned int m)
1381{
1382    int max = ext ? 63 : 31;
1383    tcg_out_extr(s, ext, rd, rn, rn, -m & max);
1384}
1385
1386static inline void tcg_out_dep(TCGContext *s, TCGType ext, TCGReg rd,
1387                               TCGReg rn, unsigned lsb, unsigned width)
1388{
1389    unsigned size = ext ? 64 : 32;
1390    unsigned a = (size - lsb) & (size - 1);
1391    unsigned b = width - 1;
1392    tcg_out_bfm(s, ext, rd, rn, a, b);
1393}
1394
1395static void tcg_out_cmp(TCGContext *s, TCGType ext, TCGCond cond, TCGReg a,
1396                        tcg_target_long b, bool const_b)
1397{
1398    if (is_tst_cond(cond)) {
1399        if (!const_b) {
1400            tcg_out_insn(s, 3510, ANDS, ext, TCG_REG_XZR, a, b);
1401        } else {
1402            tcg_out_logicali(s, I3404_ANDSI, ext, TCG_REG_XZR, a, b);
1403        }
1404    } else {
1405        if (!const_b) {
1406            tcg_out_insn(s, 3502, SUBS, ext, TCG_REG_XZR, a, b);
1407        } else if (b >= 0) {
1408            tcg_debug_assert(is_aimm(b));
1409            tcg_out_insn(s, 3401, SUBSI, ext, TCG_REG_XZR, a, b);
1410        } else {
1411            tcg_debug_assert(is_aimm(-b));
1412            tcg_out_insn(s, 3401, ADDSI, ext, TCG_REG_XZR, a, -b);
1413        }
1414    }
1415}
1416
1417static void tcg_out_goto(TCGContext *s, const tcg_insn_unit *target)
1418{
1419    ptrdiff_t offset = tcg_pcrel_diff(s, target) >> 2;
1420    tcg_debug_assert(offset == sextract64(offset, 0, 26));
1421    tcg_out_insn(s, 3206, B, offset);
1422}
1423
1424static void tcg_out_call_int(TCGContext *s, const tcg_insn_unit *target)
1425{
1426    ptrdiff_t offset = tcg_pcrel_diff(s, target) >> 2;
1427    if (offset == sextract64(offset, 0, 26)) {
1428        tcg_out_insn(s, 3206, BL, offset);
1429    } else {
1430        tcg_out_movi(s, TCG_TYPE_I64, TCG_REG_TMP0, (intptr_t)target);
1431        tcg_out_insn(s, 3207, BLR, TCG_REG_TMP0);
1432    }
1433}
1434
1435static void tcg_out_call(TCGContext *s, const tcg_insn_unit *target,
1436                         const TCGHelperInfo *info)
1437{
1438    tcg_out_call_int(s, target);
1439}
1440
1441static inline void tcg_out_goto_label(TCGContext *s, TCGLabel *l)
1442{
1443    if (!l->has_value) {
1444        tcg_out_reloc(s, s->code_ptr, R_AARCH64_JUMP26, l, 0);
1445        tcg_out_insn(s, 3206, B, 0);
1446    } else {
1447        tcg_out_goto(s, l->u.value_ptr);
1448    }
1449}
1450
1451static void tcg_out_brcond(TCGContext *s, TCGType ext, TCGCond c, TCGArg a,
1452                           TCGArg b, bool b_const, TCGLabel *l)
1453{
1454    int tbit = -1;
1455    bool need_cmp = true;
1456
1457    switch (c) {
1458    case TCG_COND_EQ:
1459    case TCG_COND_NE:
1460        /* cmp xN,0; b.ne L -> cbnz xN,L */
1461        if (b_const && b == 0) {
1462            need_cmp = false;
1463        }
1464        break;
1465    case TCG_COND_LT:
1466    case TCG_COND_GE:
1467        /* cmp xN,0; b.mi L -> tbnz xN,63,L */
1468        if (b_const && b == 0) {
1469            c = (c == TCG_COND_LT ? TCG_COND_TSTNE : TCG_COND_TSTEQ);
1470            tbit = ext ? 63 : 31;
1471            need_cmp = false;
1472        }
1473        break;
1474    case TCG_COND_TSTEQ:
1475    case TCG_COND_TSTNE:
1476        /* tst xN,0xffffffff; b.ne L -> cbnz wN,L */
1477        if (b_const && b == UINT32_MAX) {
1478            c = tcg_tst_eqne_cond(c);
1479            ext = TCG_TYPE_I32;
1480            need_cmp = false;
1481            break;
1482        }
1483        /* tst xN,1<<B; b.ne L -> tbnz xN,B,L */
1484        if (b_const && is_power_of_2(b)) {
1485            tbit = ctz64(b);
1486            need_cmp = false;
1487        }
1488        break;
1489    default:
1490        break;
1491    }
1492
1493    if (need_cmp) {
1494        tcg_out_cmp(s, ext, c, a, b, b_const);
1495        tcg_out_reloc(s, s->code_ptr, R_AARCH64_CONDBR19, l, 0);
1496        tcg_out_insn(s, 3202, B_C, c, 0);
1497        return;
1498    }
1499
1500    if (tbit >= 0) {
1501        tcg_out_reloc(s, s->code_ptr, R_AARCH64_TSTBR14, l, 0);
1502        switch (c) {
1503        case TCG_COND_TSTEQ:
1504            tcg_out_insn(s, 3205, TBZ, a, tbit, 0);
1505            break;
1506        case TCG_COND_TSTNE:
1507            tcg_out_insn(s, 3205, TBNZ, a, tbit, 0);
1508            break;
1509        default:
1510            g_assert_not_reached();
1511        }
1512    } else {
1513        tcg_out_reloc(s, s->code_ptr, R_AARCH64_CONDBR19, l, 0);
1514        switch (c) {
1515        case TCG_COND_EQ:
1516            tcg_out_insn(s, 3201, CBZ, ext, a, 0);
1517            break;
1518        case TCG_COND_NE:
1519            tcg_out_insn(s, 3201, CBNZ, ext, a, 0);
1520            break;
1521        default:
1522            g_assert_not_reached();
1523        }
1524    }
1525}
1526
1527static inline void tcg_out_rev(TCGContext *s, int ext, MemOp s_bits,
1528                               TCGReg rd, TCGReg rn)
1529{
1530    /* REV, REV16, REV32 */
1531    tcg_out_insn_3507(s, I3507_REV | (s_bits << 10), ext, rd, rn);
1532}
1533
1534static inline void tcg_out_sxt(TCGContext *s, TCGType ext, MemOp s_bits,
1535                               TCGReg rd, TCGReg rn)
1536{
1537    /* Using ALIASes SXTB, SXTH, SXTW, of SBFM Xd, Xn, #0, #7|15|31 */
1538    int bits = (8 << s_bits) - 1;
1539    tcg_out_sbfm(s, ext, rd, rn, 0, bits);
1540}
1541
1542static void tcg_out_ext8s(TCGContext *s, TCGType type, TCGReg rd, TCGReg rn)
1543{
1544    tcg_out_sxt(s, type, MO_8, rd, rn);
1545}
1546
1547static void tcg_out_ext16s(TCGContext *s, TCGType type, TCGReg rd, TCGReg rn)
1548{
1549    tcg_out_sxt(s, type, MO_16, rd, rn);
1550}
1551
1552static void tcg_out_ext32s(TCGContext *s, TCGReg rd, TCGReg rn)
1553{
1554    tcg_out_sxt(s, TCG_TYPE_I64, MO_32, rd, rn);
1555}
1556
1557static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg rd, TCGReg rn)
1558{
1559    tcg_out_ext32s(s, rd, rn);
1560}
1561
1562static inline void tcg_out_uxt(TCGContext *s, MemOp s_bits,
1563                               TCGReg rd, TCGReg rn)
1564{
1565    /* Using ALIASes UXTB, UXTH of UBFM Wd, Wn, #0, #7|15 */
1566    int bits = (8 << s_bits) - 1;
1567    tcg_out_ubfm(s, 0, rd, rn, 0, bits);
1568}
1569
1570static void tcg_out_ext8u(TCGContext *s, TCGReg rd, TCGReg rn)
1571{
1572    tcg_out_uxt(s, MO_8, rd, rn);
1573}
1574
1575static void tcg_out_ext16u(TCGContext *s, TCGReg rd, TCGReg rn)
1576{
1577    tcg_out_uxt(s, MO_16, rd, rn);
1578}
1579
1580static void tcg_out_ext32u(TCGContext *s, TCGReg rd, TCGReg rn)
1581{
1582    tcg_out_movr(s, TCG_TYPE_I32, rd, rn);
1583}
1584
1585static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg rd, TCGReg rn)
1586{
1587    tcg_out_ext32u(s, rd, rn);
1588}
1589
1590static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg rd, TCGReg rn)
1591{
1592    tcg_out_mov(s, TCG_TYPE_I32, rd, rn);
1593}
1594
1595static void tcg_out_addsub2(TCGContext *s, TCGType ext, TCGReg rl,
1596                            TCGReg rh, TCGReg al, TCGReg ah,
1597                            tcg_target_long bl, tcg_target_long bh,
1598                            bool const_bl, bool const_bh, bool sub)
1599{
1600    TCGReg orig_rl = rl;
1601    AArch64Insn insn;
1602
1603    if (rl == ah || (!const_bh && rl == bh)) {
1604        rl = TCG_REG_TMP0;
1605    }
1606
1607    if (const_bl) {
1608        if (bl < 0) {
1609            bl = -bl;
1610            insn = sub ? I3401_ADDSI : I3401_SUBSI;
1611        } else {
1612            insn = sub ? I3401_SUBSI : I3401_ADDSI;
1613        }
1614
1615        if (unlikely(al == TCG_REG_XZR)) {
1616            /* ??? We want to allow al to be zero for the benefit of
1617               negation via subtraction.  However, that leaves open the
1618               possibility of adding 0+const in the low part, and the
1619               immediate add instructions encode XSP not XZR.  Don't try
1620               anything more elaborate here than loading another zero.  */
1621            al = TCG_REG_TMP0;
1622            tcg_out_movi(s, ext, al, 0);
1623        }
1624        tcg_out_insn_3401(s, insn, ext, rl, al, bl);
1625    } else {
1626        tcg_out_insn_3502(s, sub ? I3502_SUBS : I3502_ADDS, ext, rl, al, bl);
1627    }
1628
1629    insn = I3503_ADC;
1630    if (const_bh) {
1631        /* Note that the only two constants we support are 0 and -1, and
1632           that SBC = rn + ~rm + c, so adc -1 is sbc 0, and vice-versa.  */
1633        if ((bh != 0) ^ sub) {
1634            insn = I3503_SBC;
1635        }
1636        bh = TCG_REG_XZR;
1637    } else if (sub) {
1638        insn = I3503_SBC;
1639    }
1640    tcg_out_insn_3503(s, insn, ext, rh, ah, bh);
1641
1642    tcg_out_mov(s, ext, orig_rl, rl);
1643}
1644
1645static inline void tcg_out_mb(TCGContext *s, TCGArg a0)
1646{
1647    static const uint32_t sync[] = {
1648        [0 ... TCG_MO_ALL]            = DMB_ISH | DMB_LD | DMB_ST,
1649        [TCG_MO_ST_ST]                = DMB_ISH | DMB_ST,
1650        [TCG_MO_LD_LD]                = DMB_ISH | DMB_LD,
1651        [TCG_MO_LD_ST]                = DMB_ISH | DMB_LD,
1652        [TCG_MO_LD_ST | TCG_MO_LD_LD] = DMB_ISH | DMB_LD,
1653    };
1654    tcg_out32(s, sync[a0 & TCG_MO_ALL]);
1655}
1656
1657static void tcg_out_cltz(TCGContext *s, TCGType ext, TCGReg d,
1658                         TCGReg a0, TCGArg b, bool const_b, bool is_ctz)
1659{
1660    TCGReg a1 = a0;
1661    if (is_ctz) {
1662        a1 = TCG_REG_TMP0;
1663        tcg_out_insn(s, 3507, RBIT, ext, a1, a0);
1664    }
1665    if (const_b && b == (ext ? 64 : 32)) {
1666        tcg_out_insn(s, 3507, CLZ, ext, d, a1);
1667    } else {
1668        AArch64Insn sel = I3506_CSEL;
1669
1670        tcg_out_cmp(s, ext, TCG_COND_NE, a0, 0, 1);
1671        tcg_out_insn(s, 3507, CLZ, ext, TCG_REG_TMP0, a1);
1672
1673        if (const_b) {
1674            if (b == -1) {
1675                b = TCG_REG_XZR;
1676                sel = I3506_CSINV;
1677            } else if (b == 0) {
1678                b = TCG_REG_XZR;
1679            } else {
1680                tcg_out_movi(s, ext, d, b);
1681                b = d;
1682            }
1683        }
1684        tcg_out_insn_3506(s, sel, ext, d, TCG_REG_TMP0, b, TCG_COND_NE);
1685    }
1686}
1687
1688typedef struct {
1689    TCGReg base;
1690    TCGReg index;
1691    TCGType index_ext;
1692    TCGAtomAlign aa;
1693} HostAddress;
1694
1695bool tcg_target_has_memory_bswap(MemOp memop)
1696{
1697    return false;
1698}
1699
1700static const TCGLdstHelperParam ldst_helper_param = {
1701    .ntmp = 1, .tmp = { TCG_REG_TMP0 }
1702};
1703
1704static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1705{
1706    MemOp opc = get_memop(lb->oi);
1707
1708    if (!reloc_pc19(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1709        return false;
1710    }
1711
1712    tcg_out_ld_helper_args(s, lb, &ldst_helper_param);
1713    tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SIZE]);
1714    tcg_out_ld_helper_ret(s, lb, false, &ldst_helper_param);
1715    tcg_out_goto(s, lb->raddr);
1716    return true;
1717}
1718
1719static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1720{
1721    MemOp opc = get_memop(lb->oi);
1722
1723    if (!reloc_pc19(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1724        return false;
1725    }
1726
1727    tcg_out_st_helper_args(s, lb, &ldst_helper_param);
1728    tcg_out_call_int(s, qemu_st_helpers[opc & MO_SIZE]);
1729    tcg_out_goto(s, lb->raddr);
1730    return true;
1731}
1732
1733/* We expect to use a 7-bit scaled negative offset from ENV.  */
1734#define MIN_TLB_MASK_TABLE_OFS  -512
1735
1736/*
1737 * For system-mode, perform the TLB load and compare.
1738 * For user-mode, perform any required alignment tests.
1739 * In both cases, return a TCGLabelQemuLdst structure if the slow path
1740 * is required and fill in @h with the host address for the fast path.
1741 */
1742static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
1743                                           TCGReg addr_reg, MemOpIdx oi,
1744                                           bool is_ld)
1745{
1746    TCGType addr_type = s->addr_type;
1747    TCGLabelQemuLdst *ldst = NULL;
1748    MemOp opc = get_memop(oi);
1749    MemOp s_bits = opc & MO_SIZE;
1750    unsigned a_mask;
1751
1752    h->aa = atom_and_align_for_opc(s, opc,
1753                                   have_lse2 ? MO_ATOM_WITHIN16
1754                                             : MO_ATOM_IFALIGN,
1755                                   s_bits == MO_128);
1756    a_mask = (1 << h->aa.align) - 1;
1757
1758    if (tcg_use_softmmu) {
1759        unsigned s_mask = (1u << s_bits) - 1;
1760        unsigned mem_index = get_mmuidx(oi);
1761        TCGReg addr_adj;
1762        TCGType mask_type;
1763        uint64_t compare_mask;
1764
1765        ldst = new_ldst_label(s);
1766        ldst->is_ld = is_ld;
1767        ldst->oi = oi;
1768        ldst->addr_reg = addr_reg;
1769
1770        mask_type = (s->page_bits + s->tlb_dyn_max_bits > 32
1771                     ? TCG_TYPE_I64 : TCG_TYPE_I32);
1772
1773        /* Load cpu->neg.tlb.f[mmu_idx].{mask,table} into {tmp0,tmp1}. */
1774        QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, mask) != 0);
1775        QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, table) != 8);
1776        tcg_out_insn(s, 3314, LDP, TCG_REG_TMP0, TCG_REG_TMP1, TCG_AREG0,
1777                     tlb_mask_table_ofs(s, mem_index), 1, 0);
1778
1779        /* Extract the TLB index from the address into X0.  */
1780        tcg_out_insn(s, 3502S, AND_LSR, mask_type == TCG_TYPE_I64,
1781                     TCG_REG_TMP0, TCG_REG_TMP0, addr_reg,
1782                     s->page_bits - CPU_TLB_ENTRY_BITS);
1783
1784        /* Add the tlb_table pointer, forming the CPUTLBEntry address. */
1785        tcg_out_insn(s, 3502, ADD, 1, TCG_REG_TMP1, TCG_REG_TMP1, TCG_REG_TMP0);
1786
1787        /* Load the tlb comparator into TMP0, and the fast path addend. */
1788        QEMU_BUILD_BUG_ON(HOST_BIG_ENDIAN);
1789        tcg_out_ld(s, addr_type, TCG_REG_TMP0, TCG_REG_TMP1,
1790                   is_ld ? offsetof(CPUTLBEntry, addr_read)
1791                         : offsetof(CPUTLBEntry, addr_write));
1792        tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_TMP1, TCG_REG_TMP1,
1793                   offsetof(CPUTLBEntry, addend));
1794
1795        /*
1796         * For aligned accesses, we check the first byte and include
1797         * the alignment bits within the address.  For unaligned access,
1798         * we check that we don't cross pages using the address of the
1799         * last byte of the access.
1800         */
1801        if (a_mask >= s_mask) {
1802            addr_adj = addr_reg;
1803        } else {
1804            addr_adj = TCG_REG_TMP2;
1805            tcg_out_insn(s, 3401, ADDI, addr_type,
1806                         addr_adj, addr_reg, s_mask - a_mask);
1807        }
1808        compare_mask = (uint64_t)s->page_mask | a_mask;
1809
1810        /* Store the page mask part of the address into TMP2.  */
1811        tcg_out_logicali(s, I3404_ANDI, addr_type, TCG_REG_TMP2,
1812                         addr_adj, compare_mask);
1813
1814        /* Perform the address comparison. */
1815        tcg_out_cmp(s, addr_type, TCG_COND_NE, TCG_REG_TMP0, TCG_REG_TMP2, 0);
1816
1817        /* If not equal, we jump to the slow path. */
1818        ldst->label_ptr[0] = s->code_ptr;
1819        tcg_out_insn(s, 3202, B_C, TCG_COND_NE, 0);
1820
1821        h->base = TCG_REG_TMP1;
1822        h->index = addr_reg;
1823        h->index_ext = addr_type;
1824    } else {
1825        if (a_mask) {
1826            ldst = new_ldst_label(s);
1827
1828            ldst->is_ld = is_ld;
1829            ldst->oi = oi;
1830            ldst->addr_reg = addr_reg;
1831
1832            /* tst addr, #mask */
1833            tcg_out_logicali(s, I3404_ANDSI, 0, TCG_REG_XZR, addr_reg, a_mask);
1834
1835            /* b.ne slow_path */
1836            ldst->label_ptr[0] = s->code_ptr;
1837            tcg_out_insn(s, 3202, B_C, TCG_COND_NE, 0);
1838        }
1839
1840        if (guest_base || addr_type == TCG_TYPE_I32) {
1841            h->base = TCG_REG_GUEST_BASE;
1842            h->index = addr_reg;
1843            h->index_ext = addr_type;
1844        } else {
1845            h->base = addr_reg;
1846            h->index = TCG_REG_XZR;
1847            h->index_ext = TCG_TYPE_I64;
1848        }
1849    }
1850
1851    return ldst;
1852}
1853
1854static void tcg_out_qemu_ld_direct(TCGContext *s, MemOp memop, TCGType ext,
1855                                   TCGReg data_r, HostAddress h)
1856{
1857    switch (memop & MO_SSIZE) {
1858    case MO_UB:
1859        tcg_out_ldst_r(s, I3312_LDRB, data_r, h.base, h.index_ext, h.index);
1860        break;
1861    case MO_SB:
1862        tcg_out_ldst_r(s, ext ? I3312_LDRSBX : I3312_LDRSBW,
1863                       data_r, h.base, h.index_ext, h.index);
1864        break;
1865    case MO_UW:
1866        tcg_out_ldst_r(s, I3312_LDRH, data_r, h.base, h.index_ext, h.index);
1867        break;
1868    case MO_SW:
1869        tcg_out_ldst_r(s, (ext ? I3312_LDRSHX : I3312_LDRSHW),
1870                       data_r, h.base, h.index_ext, h.index);
1871        break;
1872    case MO_UL:
1873        tcg_out_ldst_r(s, I3312_LDRW, data_r, h.base, h.index_ext, h.index);
1874        break;
1875    case MO_SL:
1876        tcg_out_ldst_r(s, I3312_LDRSWX, data_r, h.base, h.index_ext, h.index);
1877        break;
1878    case MO_UQ:
1879        tcg_out_ldst_r(s, I3312_LDRX, data_r, h.base, h.index_ext, h.index);
1880        break;
1881    default:
1882        g_assert_not_reached();
1883    }
1884}
1885
1886static void tcg_out_qemu_st_direct(TCGContext *s, MemOp memop,
1887                                   TCGReg data_r, HostAddress h)
1888{
1889    switch (memop & MO_SIZE) {
1890    case MO_8:
1891        tcg_out_ldst_r(s, I3312_STRB, data_r, h.base, h.index_ext, h.index);
1892        break;
1893    case MO_16:
1894        tcg_out_ldst_r(s, I3312_STRH, data_r, h.base, h.index_ext, h.index);
1895        break;
1896    case MO_32:
1897        tcg_out_ldst_r(s, I3312_STRW, data_r, h.base, h.index_ext, h.index);
1898        break;
1899    case MO_64:
1900        tcg_out_ldst_r(s, I3312_STRX, data_r, h.base, h.index_ext, h.index);
1901        break;
1902    default:
1903        g_assert_not_reached();
1904    }
1905}
1906
1907static void tcg_out_qemu_ld(TCGContext *s, TCGReg data_reg, TCGReg addr_reg,
1908                            MemOpIdx oi, TCGType data_type)
1909{
1910    TCGLabelQemuLdst *ldst;
1911    HostAddress h;
1912
1913    ldst = prepare_host_addr(s, &h, addr_reg, oi, true);
1914    tcg_out_qemu_ld_direct(s, get_memop(oi), data_type, data_reg, h);
1915
1916    if (ldst) {
1917        ldst->type = data_type;
1918        ldst->datalo_reg = data_reg;
1919        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1920    }
1921}
1922
1923static void tcg_out_qemu_st(TCGContext *s, TCGReg data_reg, TCGReg addr_reg,
1924                            MemOpIdx oi, TCGType data_type)
1925{
1926    TCGLabelQemuLdst *ldst;
1927    HostAddress h;
1928
1929    ldst = prepare_host_addr(s, &h, addr_reg, oi, false);
1930    tcg_out_qemu_st_direct(s, get_memop(oi), data_reg, h);
1931
1932    if (ldst) {
1933        ldst->type = data_type;
1934        ldst->datalo_reg = data_reg;
1935        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1936    }
1937}
1938
1939static void tcg_out_qemu_ldst_i128(TCGContext *s, TCGReg datalo, TCGReg datahi,
1940                                   TCGReg addr_reg, MemOpIdx oi, bool is_ld)
1941{
1942    TCGLabelQemuLdst *ldst;
1943    HostAddress h;
1944    TCGReg base;
1945    bool use_pair;
1946
1947    ldst = prepare_host_addr(s, &h, addr_reg, oi, is_ld);
1948
1949    /* Compose the final address, as LDP/STP have no indexing. */
1950    if (h.index == TCG_REG_XZR) {
1951        base = h.base;
1952    } else {
1953        base = TCG_REG_TMP2;
1954        if (h.index_ext == TCG_TYPE_I32) {
1955            /* add base, base, index, uxtw */
1956            tcg_out_insn(s, 3501, ADD, TCG_TYPE_I64, base,
1957                         h.base, h.index, MO_32, 0);
1958        } else {
1959            /* add base, base, index */
1960            tcg_out_insn(s, 3502, ADD, 1, base, h.base, h.index);
1961        }
1962    }
1963
1964    use_pair = h.aa.atom < MO_128 || have_lse2;
1965
1966    if (!use_pair) {
1967        tcg_insn_unit *branch = NULL;
1968        TCGReg ll, lh, sl, sh;
1969
1970        /*
1971         * If we have already checked for 16-byte alignment, that's all
1972         * we need. Otherwise we have determined that misaligned atomicity
1973         * may be handled with two 8-byte loads.
1974         */
1975        if (h.aa.align < MO_128) {
1976            /*
1977             * TODO: align should be MO_64, so we only need test bit 3,
1978             * which means we could use TBNZ instead of ANDS+B_C.
1979             */
1980            tcg_out_logicali(s, I3404_ANDSI, 0, TCG_REG_XZR, addr_reg, 15);
1981            branch = s->code_ptr;
1982            tcg_out_insn(s, 3202, B_C, TCG_COND_NE, 0);
1983            use_pair = true;
1984        }
1985
1986        if (is_ld) {
1987            /*
1988             * 16-byte atomicity without LSE2 requires LDXP+STXP loop:
1989             *    ldxp lo, hi, [base]
1990             *    stxp t0, lo, hi, [base]
1991             *    cbnz t0, .-8
1992             * Require no overlap between data{lo,hi} and base.
1993             */
1994            if (datalo == base || datahi == base) {
1995                tcg_out_mov(s, TCG_TYPE_REG, TCG_REG_TMP2, base);
1996                base = TCG_REG_TMP2;
1997            }
1998            ll = sl = datalo;
1999            lh = sh = datahi;
2000        } else {
2001            /*
2002             * 16-byte atomicity without LSE2 requires LDXP+STXP loop:
2003             * 1: ldxp t0, t1, [base]
2004             *    stxp t0, lo, hi, [base]
2005             *    cbnz t0, 1b
2006             */
2007            tcg_debug_assert(base != TCG_REG_TMP0 && base != TCG_REG_TMP1);
2008            ll = TCG_REG_TMP0;
2009            lh = TCG_REG_TMP1;
2010            sl = datalo;
2011            sh = datahi;
2012        }
2013
2014        tcg_out_insn(s, 3306, LDXP, TCG_REG_XZR, ll, lh, base);
2015        tcg_out_insn(s, 3306, STXP, TCG_REG_TMP0, sl, sh, base);
2016        tcg_out_insn(s, 3201, CBNZ, 0, TCG_REG_TMP0, -2);
2017
2018        if (use_pair) {
2019            /* "b .+8", branching across the one insn of use_pair. */
2020            tcg_out_insn(s, 3206, B, 2);
2021            reloc_pc19(branch, tcg_splitwx_to_rx(s->code_ptr));
2022        }
2023    }
2024
2025    if (use_pair) {
2026        if (is_ld) {
2027            tcg_out_insn(s, 3314, LDP, datalo, datahi, base, 0, 1, 0);
2028        } else {
2029            tcg_out_insn(s, 3314, STP, datalo, datahi, base, 0, 1, 0);
2030        }
2031    }
2032
2033    if (ldst) {
2034        ldst->type = TCG_TYPE_I128;
2035        ldst->datalo_reg = datalo;
2036        ldst->datahi_reg = datahi;
2037        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2038    }
2039}
2040
2041static const tcg_insn_unit *tb_ret_addr;
2042
2043static void tcg_out_exit_tb(TCGContext *s, uintptr_t a0)
2044{
2045    const tcg_insn_unit *target;
2046    ptrdiff_t offset;
2047
2048    /* Reuse the zeroing that exists for goto_ptr.  */
2049    if (a0 == 0) {
2050        target = tcg_code_gen_epilogue;
2051    } else {
2052        tcg_out_movi(s, TCG_TYPE_I64, TCG_REG_X0, a0);
2053        target = tb_ret_addr;
2054    }
2055
2056    offset = tcg_pcrel_diff(s, target) >> 2;
2057    if (offset == sextract64(offset, 0, 26)) {
2058        tcg_out_insn(s, 3206, B, offset);
2059    } else {
2060        /*
2061         * Only x16/x17 generate BTI type Jump (2),
2062         * other registers generate BTI type Jump|Call (3).
2063         */
2064        QEMU_BUILD_BUG_ON(TCG_REG_TMP0 != TCG_REG_X16);
2065        tcg_out_movi(s, TCG_TYPE_I64, TCG_REG_TMP0, (intptr_t)target);
2066        tcg_out_insn(s, 3207, BR, TCG_REG_TMP0);
2067    }
2068}
2069
2070static void tcg_out_goto_tb(TCGContext *s, int which)
2071{
2072    /*
2073     * Direct branch, or indirect address load, will be patched
2074     * by tb_target_set_jmp_target.  Assert indirect load offset
2075     * in range early, regardless of direct branch distance.
2076     */
2077    intptr_t i_off = tcg_pcrel_diff(s, (void *)get_jmp_target_addr(s, which));
2078    tcg_debug_assert(i_off == sextract64(i_off, 0, 21));
2079
2080    set_jmp_insn_offset(s, which);
2081    tcg_out32(s, I3206_B);
2082    tcg_out_insn(s, 3207, BR, TCG_REG_TMP0);
2083    set_jmp_reset_offset(s, which);
2084    tcg_out_bti(s, BTI_J);
2085}
2086
2087void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
2088                              uintptr_t jmp_rx, uintptr_t jmp_rw)
2089{
2090    uintptr_t d_addr = tb->jmp_target_addr[n];
2091    ptrdiff_t d_offset = d_addr - jmp_rx;
2092    tcg_insn_unit insn;
2093
2094    /* Either directly branch, or indirect branch load. */
2095    if (d_offset == sextract64(d_offset, 0, 28)) {
2096        insn = deposit32(I3206_B, 0, 26, d_offset >> 2);
2097    } else {
2098        uintptr_t i_addr = (uintptr_t)&tb->jmp_target_addr[n];
2099        ptrdiff_t i_offset = i_addr - jmp_rx;
2100
2101        /* Note that we asserted this in range in tcg_out_goto_tb. */
2102        insn = deposit32(I3305_LDR | TCG_REG_TMP0, 5, 19, i_offset >> 2);
2103    }
2104    qatomic_set((uint32_t *)jmp_rw, insn);
2105    flush_idcache_range(jmp_rx, jmp_rw, 4);
2106}
2107
2108
2109static void tgen_add(TCGContext *s, TCGType type,
2110                     TCGReg a0, TCGReg a1, TCGReg a2)
2111{
2112    tcg_out_insn(s, 3502, ADD, type, a0, a1, a2);
2113}
2114
2115static void tgen_addi(TCGContext *s, TCGType type,
2116                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2117{
2118    if (a2 >= 0) {
2119        tcg_out_insn(s, 3401, ADDI, type, a0, a1, a2);
2120    } else {
2121        tcg_out_insn(s, 3401, SUBI, type, a0, a1, -a2);
2122    }
2123}
2124
2125static const TCGOutOpBinary outop_add = {
2126    .base.static_constraint = C_O1_I2(r, r, rA),
2127    .out_rrr = tgen_add,
2128    .out_rri = tgen_addi,
2129};
2130
2131static void tgen_and(TCGContext *s, TCGType type,
2132                     TCGReg a0, TCGReg a1, TCGReg a2)
2133{
2134    tcg_out_insn(s, 3510, AND, type, a0, a1, a2);
2135}
2136
2137static void tgen_andi(TCGContext *s, TCGType type,
2138                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2139{
2140    tcg_out_logicali(s, I3404_ANDI, type, a0, a1, a2);
2141}
2142
2143static const TCGOutOpBinary outop_and = {
2144    .base.static_constraint = C_O1_I2(r, r, rL),
2145    .out_rrr = tgen_and,
2146    .out_rri = tgen_andi,
2147};
2148
2149static void tgen_andc(TCGContext *s, TCGType type,
2150                      TCGReg a0, TCGReg a1, TCGReg a2)
2151{
2152    tcg_out_insn(s, 3510, BIC, type, a0, a1, a2);
2153}
2154
2155static const TCGOutOpBinary outop_andc = {
2156    .base.static_constraint = C_O1_I2(r, r, r),
2157    .out_rrr = tgen_andc,
2158};
2159
2160
2161static void tcg_out_op(TCGContext *s, TCGOpcode opc, TCGType ext,
2162                       const TCGArg args[TCG_MAX_OP_ARGS],
2163                       const int const_args[TCG_MAX_OP_ARGS])
2164{
2165    /* Hoist the loads of the most common arguments.  */
2166    TCGArg a0 = args[0];
2167    TCGArg a1 = args[1];
2168    TCGArg a2 = args[2];
2169    int c2 = const_args[2];
2170
2171    switch (opc) {
2172    case INDEX_op_goto_ptr:
2173        tcg_out_insn(s, 3207, BR, a0);
2174        break;
2175
2176    case INDEX_op_br:
2177        tcg_out_goto_label(s, arg_label(a0));
2178        break;
2179
2180    case INDEX_op_ld8u_i32:
2181    case INDEX_op_ld8u_i64:
2182        tcg_out_ldst(s, I3312_LDRB, a0, a1, a2, 0);
2183        break;
2184    case INDEX_op_ld8s_i32:
2185        tcg_out_ldst(s, I3312_LDRSBW, a0, a1, a2, 0);
2186        break;
2187    case INDEX_op_ld8s_i64:
2188        tcg_out_ldst(s, I3312_LDRSBX, a0, a1, a2, 0);
2189        break;
2190    case INDEX_op_ld16u_i32:
2191    case INDEX_op_ld16u_i64:
2192        tcg_out_ldst(s, I3312_LDRH, a0, a1, a2, 1);
2193        break;
2194    case INDEX_op_ld16s_i32:
2195        tcg_out_ldst(s, I3312_LDRSHW, a0, a1, a2, 1);
2196        break;
2197    case INDEX_op_ld16s_i64:
2198        tcg_out_ldst(s, I3312_LDRSHX, a0, a1, a2, 1);
2199        break;
2200    case INDEX_op_ld_i32:
2201    case INDEX_op_ld32u_i64:
2202        tcg_out_ldst(s, I3312_LDRW, a0, a1, a2, 2);
2203        break;
2204    case INDEX_op_ld32s_i64:
2205        tcg_out_ldst(s, I3312_LDRSWX, a0, a1, a2, 2);
2206        break;
2207    case INDEX_op_ld_i64:
2208        tcg_out_ldst(s, I3312_LDRX, a0, a1, a2, 3);
2209        break;
2210
2211    case INDEX_op_st8_i32:
2212    case INDEX_op_st8_i64:
2213        tcg_out_ldst(s, I3312_STRB, a0, a1, a2, 0);
2214        break;
2215    case INDEX_op_st16_i32:
2216    case INDEX_op_st16_i64:
2217        tcg_out_ldst(s, I3312_STRH, a0, a1, a2, 1);
2218        break;
2219    case INDEX_op_st_i32:
2220    case INDEX_op_st32_i64:
2221        tcg_out_ldst(s, I3312_STRW, a0, a1, a2, 2);
2222        break;
2223    case INDEX_op_st_i64:
2224        tcg_out_ldst(s, I3312_STRX, a0, a1, a2, 3);
2225        break;
2226
2227    case INDEX_op_sub_i32:
2228    case INDEX_op_sub_i64:
2229        if (c2) {
2230            tgen_addi(s, ext, a0, a1, -a2);
2231        } else {
2232            tcg_out_insn(s, 3502, SUB, ext, a0, a1, a2);
2233        }
2234        break;
2235
2236    case INDEX_op_neg_i64:
2237    case INDEX_op_neg_i32:
2238        tcg_out_insn(s, 3502, SUB, ext, a0, TCG_REG_XZR, a1);
2239        break;
2240
2241    case INDEX_op_or_i32:
2242        a2 = (int32_t)a2;
2243        /* FALLTHRU */
2244    case INDEX_op_or_i64:
2245        if (c2) {
2246            tcg_out_logicali(s, I3404_ORRI, ext, a0, a1, a2);
2247        } else {
2248            tcg_out_insn(s, 3510, ORR, ext, a0, a1, a2);
2249        }
2250        break;
2251
2252    case INDEX_op_orc_i32:
2253        a2 = (int32_t)a2;
2254        /* FALLTHRU */
2255    case INDEX_op_orc_i64:
2256        if (c2) {
2257            tcg_out_logicali(s, I3404_ORRI, ext, a0, a1, ~a2);
2258        } else {
2259            tcg_out_insn(s, 3510, ORN, ext, a0, a1, a2);
2260        }
2261        break;
2262
2263    case INDEX_op_xor_i32:
2264        a2 = (int32_t)a2;
2265        /* FALLTHRU */
2266    case INDEX_op_xor_i64:
2267        if (c2) {
2268            tcg_out_logicali(s, I3404_EORI, ext, a0, a1, a2);
2269        } else {
2270            tcg_out_insn(s, 3510, EOR, ext, a0, a1, a2);
2271        }
2272        break;
2273
2274    case INDEX_op_eqv_i32:
2275        a2 = (int32_t)a2;
2276        /* FALLTHRU */
2277    case INDEX_op_eqv_i64:
2278        if (c2) {
2279            tcg_out_logicali(s, I3404_EORI, ext, a0, a1, ~a2);
2280        } else {
2281            tcg_out_insn(s, 3510, EON, ext, a0, a1, a2);
2282        }
2283        break;
2284
2285    case INDEX_op_not_i64:
2286    case INDEX_op_not_i32:
2287        tcg_out_insn(s, 3510, ORN, ext, a0, TCG_REG_XZR, a1);
2288        break;
2289
2290    case INDEX_op_mul_i64:
2291    case INDEX_op_mul_i32:
2292        tcg_out_insn(s, 3509, MADD, ext, a0, a1, a2, TCG_REG_XZR);
2293        break;
2294
2295    case INDEX_op_div_i64:
2296    case INDEX_op_div_i32:
2297        tcg_out_insn(s, 3508, SDIV, ext, a0, a1, a2);
2298        break;
2299    case INDEX_op_divu_i64:
2300    case INDEX_op_divu_i32:
2301        tcg_out_insn(s, 3508, UDIV, ext, a0, a1, a2);
2302        break;
2303
2304    case INDEX_op_rem_i64:
2305    case INDEX_op_rem_i32:
2306        tcg_out_insn(s, 3508, SDIV, ext, TCG_REG_TMP0, a1, a2);
2307        tcg_out_insn(s, 3509, MSUB, ext, a0, TCG_REG_TMP0, a2, a1);
2308        break;
2309    case INDEX_op_remu_i64:
2310    case INDEX_op_remu_i32:
2311        tcg_out_insn(s, 3508, UDIV, ext, TCG_REG_TMP0, a1, a2);
2312        tcg_out_insn(s, 3509, MSUB, ext, a0, TCG_REG_TMP0, a2, a1);
2313        break;
2314
2315    case INDEX_op_shl_i64:
2316    case INDEX_op_shl_i32:
2317        if (c2) {
2318            tcg_out_shl(s, ext, a0, a1, a2);
2319        } else {
2320            tcg_out_insn(s, 3508, LSLV, ext, a0, a1, a2);
2321        }
2322        break;
2323
2324    case INDEX_op_shr_i64:
2325    case INDEX_op_shr_i32:
2326        if (c2) {
2327            tcg_out_shr(s, ext, a0, a1, a2);
2328        } else {
2329            tcg_out_insn(s, 3508, LSRV, ext, a0, a1, a2);
2330        }
2331        break;
2332
2333    case INDEX_op_sar_i64:
2334    case INDEX_op_sar_i32:
2335        if (c2) {
2336            tcg_out_sar(s, ext, a0, a1, a2);
2337        } else {
2338            tcg_out_insn(s, 3508, ASRV, ext, a0, a1, a2);
2339        }
2340        break;
2341
2342    case INDEX_op_rotr_i64:
2343    case INDEX_op_rotr_i32:
2344        if (c2) {
2345            tcg_out_rotr(s, ext, a0, a1, a2);
2346        } else {
2347            tcg_out_insn(s, 3508, RORV, ext, a0, a1, a2);
2348        }
2349        break;
2350
2351    case INDEX_op_rotl_i64:
2352    case INDEX_op_rotl_i32:
2353        if (c2) {
2354            tcg_out_rotl(s, ext, a0, a1, a2);
2355        } else {
2356            tcg_out_insn(s, 3502, SUB, 0, TCG_REG_TMP0, TCG_REG_XZR, a2);
2357            tcg_out_insn(s, 3508, RORV, ext, a0, a1, TCG_REG_TMP0);
2358        }
2359        break;
2360
2361    case INDEX_op_clz_i64:
2362    case INDEX_op_clz_i32:
2363        tcg_out_cltz(s, ext, a0, a1, a2, c2, false);
2364        break;
2365    case INDEX_op_ctz_i64:
2366    case INDEX_op_ctz_i32:
2367        tcg_out_cltz(s, ext, a0, a1, a2, c2, true);
2368        break;
2369
2370    case INDEX_op_brcond_i32:
2371        a1 = (int32_t)a1;
2372        /* FALLTHRU */
2373    case INDEX_op_brcond_i64:
2374        tcg_out_brcond(s, ext, a2, a0, a1, const_args[1], arg_label(args[3]));
2375        break;
2376
2377    case INDEX_op_setcond_i32:
2378        a2 = (int32_t)a2;
2379        /* FALLTHRU */
2380    case INDEX_op_setcond_i64:
2381        tcg_out_cmp(s, ext, args[3], a1, a2, c2);
2382        /* Use CSET alias of CSINC Wd, WZR, WZR, invert(cond).  */
2383        tcg_out_insn(s, 3506, CSINC, TCG_TYPE_I32, a0, TCG_REG_XZR,
2384                     TCG_REG_XZR, tcg_invert_cond(args[3]));
2385        break;
2386
2387    case INDEX_op_negsetcond_i32:
2388        a2 = (int32_t)a2;
2389        /* FALLTHRU */
2390    case INDEX_op_negsetcond_i64:
2391        tcg_out_cmp(s, ext, args[3], a1, a2, c2);
2392        /* Use CSETM alias of CSINV Wd, WZR, WZR, invert(cond).  */
2393        tcg_out_insn(s, 3506, CSINV, ext, a0, TCG_REG_XZR,
2394                     TCG_REG_XZR, tcg_invert_cond(args[3]));
2395        break;
2396
2397    case INDEX_op_movcond_i32:
2398        a2 = (int32_t)a2;
2399        /* FALLTHRU */
2400    case INDEX_op_movcond_i64:
2401        tcg_out_cmp(s, ext, args[5], a1, a2, c2);
2402        tcg_out_insn(s, 3506, CSEL, ext, a0, args[3], args[4], args[5]);
2403        break;
2404
2405    case INDEX_op_qemu_ld_i32:
2406    case INDEX_op_qemu_ld_i64:
2407        tcg_out_qemu_ld(s, a0, a1, a2, ext);
2408        break;
2409    case INDEX_op_qemu_st_i32:
2410    case INDEX_op_qemu_st_i64:
2411        tcg_out_qemu_st(s, a0, a1, a2, ext);
2412        break;
2413    case INDEX_op_qemu_ld_i128:
2414        tcg_out_qemu_ldst_i128(s, a0, a1, a2, args[3], true);
2415        break;
2416    case INDEX_op_qemu_st_i128:
2417        tcg_out_qemu_ldst_i128(s, a0, a1, a2, args[3], false);
2418        break;
2419
2420    case INDEX_op_bswap64_i64:
2421        tcg_out_rev(s, TCG_TYPE_I64, MO_64, a0, a1);
2422        break;
2423    case INDEX_op_bswap32_i64:
2424        tcg_out_rev(s, TCG_TYPE_I32, MO_32, a0, a1);
2425        if (a2 & TCG_BSWAP_OS) {
2426            tcg_out_ext32s(s, a0, a0);
2427        }
2428        break;
2429    case INDEX_op_bswap32_i32:
2430        tcg_out_rev(s, TCG_TYPE_I32, MO_32, a0, a1);
2431        break;
2432    case INDEX_op_bswap16_i64:
2433    case INDEX_op_bswap16_i32:
2434        tcg_out_rev(s, TCG_TYPE_I32, MO_16, a0, a1);
2435        if (a2 & TCG_BSWAP_OS) {
2436            /* Output must be sign-extended. */
2437            tcg_out_ext16s(s, ext, a0, a0);
2438        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
2439            /* Output must be zero-extended, but input isn't. */
2440            tcg_out_ext16u(s, a0, a0);
2441        }
2442        break;
2443
2444    case INDEX_op_deposit_i64:
2445    case INDEX_op_deposit_i32:
2446        tcg_out_dep(s, ext, a0, a2, args[3], args[4]);
2447        break;
2448
2449    case INDEX_op_extract_i64:
2450    case INDEX_op_extract_i32:
2451        if (a2 == 0) {
2452            uint64_t mask = MAKE_64BIT_MASK(0, args[3]);
2453            tcg_out_logicali(s, I3404_ANDI, ext, a0, a1, mask);
2454        } else {
2455            tcg_out_ubfm(s, ext, a0, a1, a2, a2 + args[3] - 1);
2456        }
2457        break;
2458
2459    case INDEX_op_sextract_i64:
2460    case INDEX_op_sextract_i32:
2461        tcg_out_sbfm(s, ext, a0, a1, a2, a2 + args[3] - 1);
2462        break;
2463
2464    case INDEX_op_extract2_i64:
2465    case INDEX_op_extract2_i32:
2466        tcg_out_extr(s, ext, a0, a2, a1, args[3]);
2467        break;
2468
2469    case INDEX_op_add2_i32:
2470        tcg_out_addsub2(s, TCG_TYPE_I32, a0, a1, a2, args[3],
2471                        (int32_t)args[4], args[5], const_args[4],
2472                        const_args[5], false);
2473        break;
2474    case INDEX_op_add2_i64:
2475        tcg_out_addsub2(s, TCG_TYPE_I64, a0, a1, a2, args[3], args[4],
2476                        args[5], const_args[4], const_args[5], false);
2477        break;
2478    case INDEX_op_sub2_i32:
2479        tcg_out_addsub2(s, TCG_TYPE_I32, a0, a1, a2, args[3],
2480                        (int32_t)args[4], args[5], const_args[4],
2481                        const_args[5], true);
2482        break;
2483    case INDEX_op_sub2_i64:
2484        tcg_out_addsub2(s, TCG_TYPE_I64, a0, a1, a2, args[3], args[4],
2485                        args[5], const_args[4], const_args[5], true);
2486        break;
2487
2488    case INDEX_op_muluh_i64:
2489        tcg_out_insn(s, 3508, UMULH, TCG_TYPE_I64, a0, a1, a2);
2490        break;
2491    case INDEX_op_mulsh_i64:
2492        tcg_out_insn(s, 3508, SMULH, TCG_TYPE_I64, a0, a1, a2);
2493        break;
2494
2495    case INDEX_op_mb:
2496        tcg_out_mb(s, a0);
2497        break;
2498
2499    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
2500    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
2501    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
2502    case INDEX_op_ext_i32_i64:  /* Always emitted via tcg_reg_alloc_op.  */
2503    case INDEX_op_extu_i32_i64:
2504    case INDEX_op_extrl_i64_i32:
2505    default:
2506        g_assert_not_reached();
2507    }
2508}
2509
2510static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
2511                           unsigned vecl, unsigned vece,
2512                           const TCGArg args[TCG_MAX_OP_ARGS],
2513                           const int const_args[TCG_MAX_OP_ARGS])
2514{
2515    static const AArch64Insn cmp_vec_insn[16] = {
2516        [TCG_COND_EQ] = I3616_CMEQ,
2517        [TCG_COND_GT] = I3616_CMGT,
2518        [TCG_COND_GE] = I3616_CMGE,
2519        [TCG_COND_GTU] = I3616_CMHI,
2520        [TCG_COND_GEU] = I3616_CMHS,
2521    };
2522    static const AArch64Insn cmp_scalar_insn[16] = {
2523        [TCG_COND_EQ] = I3611_CMEQ,
2524        [TCG_COND_GT] = I3611_CMGT,
2525        [TCG_COND_GE] = I3611_CMGE,
2526        [TCG_COND_GTU] = I3611_CMHI,
2527        [TCG_COND_GEU] = I3611_CMHS,
2528    };
2529    static const AArch64Insn cmp0_vec_insn[16] = {
2530        [TCG_COND_EQ] = I3617_CMEQ0,
2531        [TCG_COND_GT] = I3617_CMGT0,
2532        [TCG_COND_GE] = I3617_CMGE0,
2533        [TCG_COND_LT] = I3617_CMLT0,
2534        [TCG_COND_LE] = I3617_CMLE0,
2535    };
2536    static const AArch64Insn cmp0_scalar_insn[16] = {
2537        [TCG_COND_EQ] = I3612_CMEQ0,
2538        [TCG_COND_GT] = I3612_CMGT0,
2539        [TCG_COND_GE] = I3612_CMGE0,
2540        [TCG_COND_LT] = I3612_CMLT0,
2541        [TCG_COND_LE] = I3612_CMLE0,
2542    };
2543
2544    TCGType type = vecl + TCG_TYPE_V64;
2545    unsigned is_q = vecl;
2546    bool is_scalar = !is_q && vece == MO_64;
2547    TCGArg a0, a1, a2, a3;
2548    int cmode, imm8;
2549
2550    a0 = args[0];
2551    a1 = args[1];
2552    a2 = args[2];
2553
2554    switch (opc) {
2555    case INDEX_op_ld_vec:
2556        tcg_out_ld(s, type, a0, a1, a2);
2557        break;
2558    case INDEX_op_st_vec:
2559        tcg_out_st(s, type, a0, a1, a2);
2560        break;
2561    case INDEX_op_dupm_vec:
2562        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
2563        break;
2564    case INDEX_op_add_vec:
2565        if (is_scalar) {
2566            tcg_out_insn(s, 3611, ADD, vece, a0, a1, a2);
2567        } else {
2568            tcg_out_insn(s, 3616, ADD, is_q, vece, a0, a1, a2);
2569        }
2570        break;
2571    case INDEX_op_sub_vec:
2572        if (is_scalar) {
2573            tcg_out_insn(s, 3611, SUB, vece, a0, a1, a2);
2574        } else {
2575            tcg_out_insn(s, 3616, SUB, is_q, vece, a0, a1, a2);
2576        }
2577        break;
2578    case INDEX_op_mul_vec:
2579        tcg_out_insn(s, 3616, MUL, is_q, vece, a0, a1, a2);
2580        break;
2581    case INDEX_op_neg_vec:
2582        if (is_scalar) {
2583            tcg_out_insn(s, 3612, NEG, vece, a0, a1);
2584        } else {
2585            tcg_out_insn(s, 3617, NEG, is_q, vece, a0, a1);
2586        }
2587        break;
2588    case INDEX_op_abs_vec:
2589        if (is_scalar) {
2590            tcg_out_insn(s, 3612, ABS, vece, a0, a1);
2591        } else {
2592            tcg_out_insn(s, 3617, ABS, is_q, vece, a0, a1);
2593        }
2594        break;
2595    case INDEX_op_and_vec:
2596        if (const_args[2]) {
2597            is_shimm1632(~a2, &cmode, &imm8);
2598            if (a0 == a1) {
2599                tcg_out_insn(s, 3606, BIC, is_q, a0, 0, cmode, imm8);
2600                return;
2601            }
2602            tcg_out_insn(s, 3606, MVNI, is_q, a0, 0, cmode, imm8);
2603            a2 = a0;
2604        }
2605        tcg_out_insn(s, 3616, AND, is_q, 0, a0, a1, a2);
2606        break;
2607    case INDEX_op_or_vec:
2608        if (const_args[2]) {
2609            is_shimm1632(a2, &cmode, &imm8);
2610            if (a0 == a1) {
2611                tcg_out_insn(s, 3606, ORR, is_q, a0, 0, cmode, imm8);
2612                return;
2613            }
2614            tcg_out_insn(s, 3606, MOVI, is_q, a0, 0, cmode, imm8);
2615            a2 = a0;
2616        }
2617        tcg_out_insn(s, 3616, ORR, is_q, 0, a0, a1, a2);
2618        break;
2619    case INDEX_op_andc_vec:
2620        if (const_args[2]) {
2621            is_shimm1632(a2, &cmode, &imm8);
2622            if (a0 == a1) {
2623                tcg_out_insn(s, 3606, BIC, is_q, a0, 0, cmode, imm8);
2624                return;
2625            }
2626            tcg_out_insn(s, 3606, MOVI, is_q, a0, 0, cmode, imm8);
2627            a2 = a0;
2628        }
2629        tcg_out_insn(s, 3616, BIC, is_q, 0, a0, a1, a2);
2630        break;
2631    case INDEX_op_orc_vec:
2632        if (const_args[2]) {
2633            is_shimm1632(~a2, &cmode, &imm8);
2634            if (a0 == a1) {
2635                tcg_out_insn(s, 3606, ORR, is_q, a0, 0, cmode, imm8);
2636                return;
2637            }
2638            tcg_out_insn(s, 3606, MVNI, is_q, a0, 0, cmode, imm8);
2639            a2 = a0;
2640        }
2641        tcg_out_insn(s, 3616, ORN, is_q, 0, a0, a1, a2);
2642        break;
2643    case INDEX_op_xor_vec:
2644        tcg_out_insn(s, 3616, EOR, is_q, 0, a0, a1, a2);
2645        break;
2646    case INDEX_op_ssadd_vec:
2647        if (is_scalar) {
2648            tcg_out_insn(s, 3611, SQADD, vece, a0, a1, a2);
2649        } else {
2650            tcg_out_insn(s, 3616, SQADD, is_q, vece, a0, a1, a2);
2651        }
2652        break;
2653    case INDEX_op_sssub_vec:
2654        if (is_scalar) {
2655            tcg_out_insn(s, 3611, SQSUB, vece, a0, a1, a2);
2656        } else {
2657            tcg_out_insn(s, 3616, SQSUB, is_q, vece, a0, a1, a2);
2658        }
2659        break;
2660    case INDEX_op_usadd_vec:
2661        if (is_scalar) {
2662            tcg_out_insn(s, 3611, UQADD, vece, a0, a1, a2);
2663        } else {
2664            tcg_out_insn(s, 3616, UQADD, is_q, vece, a0, a1, a2);
2665        }
2666        break;
2667    case INDEX_op_ussub_vec:
2668        if (is_scalar) {
2669            tcg_out_insn(s, 3611, UQSUB, vece, a0, a1, a2);
2670        } else {
2671            tcg_out_insn(s, 3616, UQSUB, is_q, vece, a0, a1, a2);
2672        }
2673        break;
2674    case INDEX_op_smax_vec:
2675        tcg_out_insn(s, 3616, SMAX, is_q, vece, a0, a1, a2);
2676        break;
2677    case INDEX_op_smin_vec:
2678        tcg_out_insn(s, 3616, SMIN, is_q, vece, a0, a1, a2);
2679        break;
2680    case INDEX_op_umax_vec:
2681        tcg_out_insn(s, 3616, UMAX, is_q, vece, a0, a1, a2);
2682        break;
2683    case INDEX_op_umin_vec:
2684        tcg_out_insn(s, 3616, UMIN, is_q, vece, a0, a1, a2);
2685        break;
2686    case INDEX_op_not_vec:
2687        tcg_out_insn(s, 3617, NOT, is_q, 0, a0, a1);
2688        break;
2689    case INDEX_op_shli_vec:
2690        if (is_scalar) {
2691            tcg_out_insn(s, 3609, SHL, a0, a1, a2 + (8 << vece));
2692        } else {
2693            tcg_out_insn(s, 3614, SHL, is_q, a0, a1, a2 + (8 << vece));
2694        }
2695        break;
2696    case INDEX_op_shri_vec:
2697        if (is_scalar) {
2698            tcg_out_insn(s, 3609, USHR, a0, a1, (16 << vece) - a2);
2699        } else {
2700            tcg_out_insn(s, 3614, USHR, is_q, a0, a1, (16 << vece) - a2);
2701        }
2702        break;
2703    case INDEX_op_sari_vec:
2704        if (is_scalar) {
2705            tcg_out_insn(s, 3609, SSHR, a0, a1, (16 << vece) - a2);
2706        } else {
2707            tcg_out_insn(s, 3614, SSHR, is_q, a0, a1, (16 << vece) - a2);
2708        }
2709        break;
2710    case INDEX_op_aa64_sli_vec:
2711        if (is_scalar) {
2712            tcg_out_insn(s, 3609, SLI, a0, a2, args[3] + (8 << vece));
2713        } else {
2714            tcg_out_insn(s, 3614, SLI, is_q, a0, a2, args[3] + (8 << vece));
2715        }
2716        break;
2717    case INDEX_op_shlv_vec:
2718        if (is_scalar) {
2719            tcg_out_insn(s, 3611, USHL, vece, a0, a1, a2);
2720        } else {
2721            tcg_out_insn(s, 3616, USHL, is_q, vece, a0, a1, a2);
2722        }
2723        break;
2724    case INDEX_op_aa64_sshl_vec:
2725        if (is_scalar) {
2726            tcg_out_insn(s, 3611, SSHL, vece, a0, a1, a2);
2727        } else {
2728            tcg_out_insn(s, 3616, SSHL, is_q, vece, a0, a1, a2);
2729        }
2730        break;
2731    case INDEX_op_cmp_vec:
2732        {
2733            TCGCond cond = args[3];
2734            AArch64Insn insn;
2735
2736            switch (cond) {
2737            case TCG_COND_NE:
2738                if (const_args[2]) {
2739                    if (is_scalar) {
2740                        tcg_out_insn(s, 3611, CMTST, vece, a0, a1, a1);
2741                    } else {
2742                        tcg_out_insn(s, 3616, CMTST, is_q, vece, a0, a1, a1);
2743                    }
2744                } else {
2745                    if (is_scalar) {
2746                        tcg_out_insn(s, 3611, CMEQ, vece, a0, a1, a2);
2747                    } else {
2748                        tcg_out_insn(s, 3616, CMEQ, is_q, vece, a0, a1, a2);
2749                    }
2750                    tcg_out_insn(s, 3617, NOT, is_q, 0, a0, a0);
2751                }
2752                break;
2753
2754            case TCG_COND_TSTNE:
2755            case TCG_COND_TSTEQ:
2756                if (const_args[2]) {
2757                    /* (x & 0) == 0 */
2758                    tcg_out_dupi_vec(s, type, MO_8, a0,
2759                                     -(cond == TCG_COND_TSTEQ));
2760                    break;
2761                }
2762                if (is_scalar) {
2763                    tcg_out_insn(s, 3611, CMTST, vece, a0, a1, a2);
2764                } else {
2765                    tcg_out_insn(s, 3616, CMTST, is_q, vece, a0, a1, a2);
2766                }
2767                if (cond == TCG_COND_TSTEQ) {
2768                    tcg_out_insn(s, 3617, NOT, is_q, 0, a0, a0);
2769                }
2770                break;
2771
2772            default:
2773                if (const_args[2]) {
2774                    if (is_scalar) {
2775                        insn = cmp0_scalar_insn[cond];
2776                        if (insn) {
2777                            tcg_out_insn_3612(s, insn, vece, a0, a1);
2778                            break;
2779                        }
2780                    } else {
2781                        insn = cmp0_vec_insn[cond];
2782                        if (insn) {
2783                            tcg_out_insn_3617(s, insn, is_q, vece, a0, a1);
2784                            break;
2785                        }
2786                    }
2787                    tcg_out_dupi_vec(s, type, MO_8, TCG_VEC_TMP0, 0);
2788                    a2 = TCG_VEC_TMP0;
2789                }
2790                if (is_scalar) {
2791                    insn = cmp_scalar_insn[cond];
2792                    if (insn == 0) {
2793                        TCGArg t;
2794                        t = a1, a1 = a2, a2 = t;
2795                        cond = tcg_swap_cond(cond);
2796                        insn = cmp_scalar_insn[cond];
2797                        tcg_debug_assert(insn != 0);
2798                    }
2799                    tcg_out_insn_3611(s, insn, vece, a0, a1, a2);
2800                } else {
2801                    insn = cmp_vec_insn[cond];
2802                    if (insn == 0) {
2803                        TCGArg t;
2804                        t = a1, a1 = a2, a2 = t;
2805                        cond = tcg_swap_cond(cond);
2806                        insn = cmp_vec_insn[cond];
2807                        tcg_debug_assert(insn != 0);
2808                    }
2809                    tcg_out_insn_3616(s, insn, is_q, vece, a0, a1, a2);
2810                }
2811                break;
2812            }
2813        }
2814        break;
2815
2816    case INDEX_op_bitsel_vec:
2817        a3 = args[3];
2818        if (a0 == a3) {
2819            tcg_out_insn(s, 3616, BIT, is_q, 0, a0, a2, a1);
2820        } else if (a0 == a2) {
2821            tcg_out_insn(s, 3616, BIF, is_q, 0, a0, a3, a1);
2822        } else {
2823            if (a0 != a1) {
2824                tcg_out_mov(s, type, a0, a1);
2825            }
2826            tcg_out_insn(s, 3616, BSL, is_q, 0, a0, a2, a3);
2827        }
2828        break;
2829
2830    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
2831    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
2832    default:
2833        g_assert_not_reached();
2834    }
2835}
2836
2837int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
2838{
2839    switch (opc) {
2840    case INDEX_op_add_vec:
2841    case INDEX_op_sub_vec:
2842    case INDEX_op_and_vec:
2843    case INDEX_op_or_vec:
2844    case INDEX_op_xor_vec:
2845    case INDEX_op_andc_vec:
2846    case INDEX_op_orc_vec:
2847    case INDEX_op_neg_vec:
2848    case INDEX_op_abs_vec:
2849    case INDEX_op_not_vec:
2850    case INDEX_op_cmp_vec:
2851    case INDEX_op_shli_vec:
2852    case INDEX_op_shri_vec:
2853    case INDEX_op_sari_vec:
2854    case INDEX_op_ssadd_vec:
2855    case INDEX_op_sssub_vec:
2856    case INDEX_op_usadd_vec:
2857    case INDEX_op_ussub_vec:
2858    case INDEX_op_shlv_vec:
2859    case INDEX_op_bitsel_vec:
2860        return 1;
2861    case INDEX_op_rotli_vec:
2862    case INDEX_op_shrv_vec:
2863    case INDEX_op_sarv_vec:
2864    case INDEX_op_rotlv_vec:
2865    case INDEX_op_rotrv_vec:
2866        return -1;
2867    case INDEX_op_mul_vec:
2868    case INDEX_op_smax_vec:
2869    case INDEX_op_smin_vec:
2870    case INDEX_op_umax_vec:
2871    case INDEX_op_umin_vec:
2872        return vece < MO_64;
2873
2874    default:
2875        return 0;
2876    }
2877}
2878
2879void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
2880                       TCGArg a0, ...)
2881{
2882    va_list va;
2883    TCGv_vec v0, v1, v2, t1, t2, c1;
2884    TCGArg a2;
2885
2886    va_start(va, a0);
2887    v0 = temp_tcgv_vec(arg_temp(a0));
2888    v1 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
2889    a2 = va_arg(va, TCGArg);
2890    va_end(va);
2891
2892    switch (opc) {
2893    case INDEX_op_rotli_vec:
2894        t1 = tcg_temp_new_vec(type);
2895        tcg_gen_shri_vec(vece, t1, v1, -a2 & ((8 << vece) - 1));
2896        vec_gen_4(INDEX_op_aa64_sli_vec, type, vece,
2897                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(v1), a2);
2898        tcg_temp_free_vec(t1);
2899        break;
2900
2901    case INDEX_op_shrv_vec:
2902    case INDEX_op_sarv_vec:
2903        /* Right shifts are negative left shifts for AArch64.  */
2904        v2 = temp_tcgv_vec(arg_temp(a2));
2905        t1 = tcg_temp_new_vec(type);
2906        tcg_gen_neg_vec(vece, t1, v2);
2907        opc = (opc == INDEX_op_shrv_vec
2908               ? INDEX_op_shlv_vec : INDEX_op_aa64_sshl_vec);
2909        vec_gen_3(opc, type, vece, tcgv_vec_arg(v0),
2910                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
2911        tcg_temp_free_vec(t1);
2912        break;
2913
2914    case INDEX_op_rotlv_vec:
2915        v2 = temp_tcgv_vec(arg_temp(a2));
2916        t1 = tcg_temp_new_vec(type);
2917        c1 = tcg_constant_vec(type, vece, 8 << vece);
2918        tcg_gen_sub_vec(vece, t1, v2, c1);
2919        /* Right shifts are negative left shifts for AArch64.  */
2920        vec_gen_3(INDEX_op_shlv_vec, type, vece, tcgv_vec_arg(t1),
2921                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
2922        vec_gen_3(INDEX_op_shlv_vec, type, vece, tcgv_vec_arg(v0),
2923                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
2924        tcg_gen_or_vec(vece, v0, v0, t1);
2925        tcg_temp_free_vec(t1);
2926        break;
2927
2928    case INDEX_op_rotrv_vec:
2929        v2 = temp_tcgv_vec(arg_temp(a2));
2930        t1 = tcg_temp_new_vec(type);
2931        t2 = tcg_temp_new_vec(type);
2932        c1 = tcg_constant_vec(type, vece, 8 << vece);
2933        tcg_gen_neg_vec(vece, t1, v2);
2934        tcg_gen_sub_vec(vece, t2, c1, v2);
2935        /* Right shifts are negative left shifts for AArch64.  */
2936        vec_gen_3(INDEX_op_shlv_vec, type, vece, tcgv_vec_arg(t1),
2937                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
2938        vec_gen_3(INDEX_op_shlv_vec, type, vece, tcgv_vec_arg(t2),
2939                  tcgv_vec_arg(v1), tcgv_vec_arg(t2));
2940        tcg_gen_or_vec(vece, v0, t1, t2);
2941        tcg_temp_free_vec(t1);
2942        tcg_temp_free_vec(t2);
2943        break;
2944
2945    default:
2946        g_assert_not_reached();
2947    }
2948}
2949
2950static TCGConstraintSetIndex
2951tcg_target_op_def(TCGOpcode op, TCGType type, unsigned flags)
2952{
2953    switch (op) {
2954    case INDEX_op_goto_ptr:
2955        return C_O0_I1(r);
2956
2957    case INDEX_op_ld8u_i32:
2958    case INDEX_op_ld8s_i32:
2959    case INDEX_op_ld16u_i32:
2960    case INDEX_op_ld16s_i32:
2961    case INDEX_op_ld_i32:
2962    case INDEX_op_ld8u_i64:
2963    case INDEX_op_ld8s_i64:
2964    case INDEX_op_ld16u_i64:
2965    case INDEX_op_ld16s_i64:
2966    case INDEX_op_ld32u_i64:
2967    case INDEX_op_ld32s_i64:
2968    case INDEX_op_ld_i64:
2969    case INDEX_op_neg_i32:
2970    case INDEX_op_neg_i64:
2971    case INDEX_op_not_i32:
2972    case INDEX_op_not_i64:
2973    case INDEX_op_bswap16_i32:
2974    case INDEX_op_bswap32_i32:
2975    case INDEX_op_bswap16_i64:
2976    case INDEX_op_bswap32_i64:
2977    case INDEX_op_bswap64_i64:
2978    case INDEX_op_ext_i32_i64:
2979    case INDEX_op_extu_i32_i64:
2980    case INDEX_op_extract_i32:
2981    case INDEX_op_extract_i64:
2982    case INDEX_op_sextract_i32:
2983    case INDEX_op_sextract_i64:
2984        return C_O1_I1(r, r);
2985
2986    case INDEX_op_st8_i32:
2987    case INDEX_op_st16_i32:
2988    case INDEX_op_st_i32:
2989    case INDEX_op_st8_i64:
2990    case INDEX_op_st16_i64:
2991    case INDEX_op_st32_i64:
2992    case INDEX_op_st_i64:
2993        return C_O0_I2(rz, r);
2994
2995    case INDEX_op_sub_i32:
2996    case INDEX_op_sub_i64:
2997        return C_O1_I2(r, r, rA);
2998
2999    case INDEX_op_setcond_i32:
3000    case INDEX_op_setcond_i64:
3001    case INDEX_op_negsetcond_i32:
3002    case INDEX_op_negsetcond_i64:
3003        return C_O1_I2(r, r, rC);
3004
3005    case INDEX_op_mul_i32:
3006    case INDEX_op_mul_i64:
3007    case INDEX_op_div_i32:
3008    case INDEX_op_div_i64:
3009    case INDEX_op_divu_i32:
3010    case INDEX_op_divu_i64:
3011    case INDEX_op_rem_i32:
3012    case INDEX_op_rem_i64:
3013    case INDEX_op_remu_i32:
3014    case INDEX_op_remu_i64:
3015    case INDEX_op_muluh_i64:
3016    case INDEX_op_mulsh_i64:
3017        return C_O1_I2(r, r, r);
3018
3019    case INDEX_op_or_i32:
3020    case INDEX_op_or_i64:
3021    case INDEX_op_xor_i32:
3022    case INDEX_op_xor_i64:
3023    case INDEX_op_orc_i32:
3024    case INDEX_op_orc_i64:
3025    case INDEX_op_eqv_i32:
3026    case INDEX_op_eqv_i64:
3027        return C_O1_I2(r, r, rL);
3028
3029    case INDEX_op_shl_i32:
3030    case INDEX_op_shr_i32:
3031    case INDEX_op_sar_i32:
3032    case INDEX_op_rotl_i32:
3033    case INDEX_op_rotr_i32:
3034    case INDEX_op_shl_i64:
3035    case INDEX_op_shr_i64:
3036    case INDEX_op_sar_i64:
3037    case INDEX_op_rotl_i64:
3038    case INDEX_op_rotr_i64:
3039        return C_O1_I2(r, r, ri);
3040
3041    case INDEX_op_clz_i32:
3042    case INDEX_op_ctz_i32:
3043    case INDEX_op_clz_i64:
3044    case INDEX_op_ctz_i64:
3045        return C_O1_I2(r, r, rAL);
3046
3047    case INDEX_op_brcond_i32:
3048    case INDEX_op_brcond_i64:
3049        return C_O0_I2(r, rC);
3050
3051    case INDEX_op_movcond_i32:
3052    case INDEX_op_movcond_i64:
3053        return C_O1_I4(r, r, rC, rz, rz);
3054
3055    case INDEX_op_qemu_ld_i32:
3056    case INDEX_op_qemu_ld_i64:
3057        return C_O1_I1(r, r);
3058    case INDEX_op_qemu_ld_i128:
3059        return C_O2_I1(r, r, r);
3060    case INDEX_op_qemu_st_i32:
3061    case INDEX_op_qemu_st_i64:
3062        return C_O0_I2(rz, r);
3063    case INDEX_op_qemu_st_i128:
3064        return C_O0_I3(rz, rz, r);
3065
3066    case INDEX_op_deposit_i32:
3067    case INDEX_op_deposit_i64:
3068        return C_O1_I2(r, 0, rz);
3069
3070    case INDEX_op_extract2_i32:
3071    case INDEX_op_extract2_i64:
3072        return C_O1_I2(r, rz, rz);
3073
3074    case INDEX_op_add2_i32:
3075    case INDEX_op_add2_i64:
3076    case INDEX_op_sub2_i32:
3077    case INDEX_op_sub2_i64:
3078        return C_O2_I4(r, r, rz, rz, rA, rMZ);
3079
3080    case INDEX_op_add_vec:
3081    case INDEX_op_sub_vec:
3082    case INDEX_op_mul_vec:
3083    case INDEX_op_xor_vec:
3084    case INDEX_op_ssadd_vec:
3085    case INDEX_op_sssub_vec:
3086    case INDEX_op_usadd_vec:
3087    case INDEX_op_ussub_vec:
3088    case INDEX_op_smax_vec:
3089    case INDEX_op_smin_vec:
3090    case INDEX_op_umax_vec:
3091    case INDEX_op_umin_vec:
3092    case INDEX_op_shlv_vec:
3093    case INDEX_op_shrv_vec:
3094    case INDEX_op_sarv_vec:
3095    case INDEX_op_aa64_sshl_vec:
3096        return C_O1_I2(w, w, w);
3097    case INDEX_op_not_vec:
3098    case INDEX_op_neg_vec:
3099    case INDEX_op_abs_vec:
3100    case INDEX_op_shli_vec:
3101    case INDEX_op_shri_vec:
3102    case INDEX_op_sari_vec:
3103        return C_O1_I1(w, w);
3104    case INDEX_op_ld_vec:
3105    case INDEX_op_dupm_vec:
3106        return C_O1_I1(w, r);
3107    case INDEX_op_st_vec:
3108        return C_O0_I2(w, r);
3109    case INDEX_op_dup_vec:
3110        return C_O1_I1(w, wr);
3111    case INDEX_op_or_vec:
3112    case INDEX_op_andc_vec:
3113        return C_O1_I2(w, w, wO);
3114    case INDEX_op_and_vec:
3115    case INDEX_op_orc_vec:
3116        return C_O1_I2(w, w, wN);
3117    case INDEX_op_cmp_vec:
3118        return C_O1_I2(w, w, wZ);
3119    case INDEX_op_bitsel_vec:
3120        return C_O1_I3(w, w, w, w);
3121    case INDEX_op_aa64_sli_vec:
3122        return C_O1_I2(w, 0, w);
3123
3124    default:
3125        return C_NotImplemented;
3126    }
3127}
3128
3129static void tcg_target_init(TCGContext *s)
3130{
3131    tcg_target_available_regs[TCG_TYPE_I32] = 0xffffffffu;
3132    tcg_target_available_regs[TCG_TYPE_I64] = 0xffffffffu;
3133    tcg_target_available_regs[TCG_TYPE_V64] = 0xffffffff00000000ull;
3134    tcg_target_available_regs[TCG_TYPE_V128] = 0xffffffff00000000ull;
3135
3136    tcg_target_call_clobber_regs = -1ull;
3137    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_X19);
3138    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_X20);
3139    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_X21);
3140    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_X22);
3141    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_X23);
3142    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_X24);
3143    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_X25);
3144    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_X26);
3145    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_X27);
3146    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_X28);
3147    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_X29);
3148    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V8);
3149    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V9);
3150    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V10);
3151    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V11);
3152    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V12);
3153    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V13);
3154    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V14);
3155    tcg_regset_reset_reg(tcg_target_call_clobber_regs, TCG_REG_V15);
3156
3157    s->reserved_regs = 0;
3158    tcg_regset_set_reg(s->reserved_regs, TCG_REG_SP);
3159    tcg_regset_set_reg(s->reserved_regs, TCG_REG_FP);
3160    tcg_regset_set_reg(s->reserved_regs, TCG_REG_X18); /* platform register */
3161    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP0);
3162    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP1);
3163    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP2);
3164    tcg_regset_set_reg(s->reserved_regs, TCG_VEC_TMP0);
3165}
3166
3167/* Saving pairs: (X19, X20) .. (X27, X28), (X29(fp), X30(lr)).  */
3168#define PUSH_SIZE  ((30 - 19 + 1) * 8)
3169
3170#define FRAME_SIZE \
3171    ((PUSH_SIZE \
3172      + TCG_STATIC_CALL_ARGS_SIZE \
3173      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
3174      + TCG_TARGET_STACK_ALIGN - 1) \
3175     & ~(TCG_TARGET_STACK_ALIGN - 1))
3176
3177/* We're expecting a 2 byte uleb128 encoded value.  */
3178QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
3179
3180/* We're expecting to use a single ADDI insn.  */
3181QEMU_BUILD_BUG_ON(FRAME_SIZE - PUSH_SIZE > 0xfff);
3182
3183static void tcg_target_qemu_prologue(TCGContext *s)
3184{
3185    TCGReg r;
3186
3187    tcg_out_bti(s, BTI_C);
3188
3189    /* Push (FP, LR) and allocate space for all saved registers.  */
3190    tcg_out_insn(s, 3314, STP, TCG_REG_FP, TCG_REG_LR,
3191                 TCG_REG_SP, -PUSH_SIZE, 1, 1);
3192
3193    /* Set up frame pointer for canonical unwinding.  */
3194    tcg_out_movr_sp(s, TCG_TYPE_I64, TCG_REG_FP, TCG_REG_SP);
3195
3196    /* Store callee-preserved regs x19..x28.  */
3197    for (r = TCG_REG_X19; r <= TCG_REG_X27; r += 2) {
3198        int ofs = (r - TCG_REG_X19 + 2) * 8;
3199        tcg_out_insn(s, 3314, STP, r, r + 1, TCG_REG_SP, ofs, 1, 0);
3200    }
3201
3202    /* Make stack space for TCG locals.  */
3203    tcg_out_insn(s, 3401, SUBI, TCG_TYPE_I64, TCG_REG_SP, TCG_REG_SP,
3204                 FRAME_SIZE - PUSH_SIZE);
3205
3206    /* Inform TCG about how to find TCG locals with register, offset, size.  */
3207    tcg_set_frame(s, TCG_REG_SP, TCG_STATIC_CALL_ARGS_SIZE,
3208                  CPU_TEMP_BUF_NLONGS * sizeof(long));
3209
3210    if (!tcg_use_softmmu) {
3211        /*
3212         * Note that XZR cannot be encoded in the address base register slot,
3213         * as that actually encodes SP.  Depending on the guest, we may need
3214         * to zero-extend the guest address via the address index register slot,
3215         * therefore we need to load even a zero guest base into a register.
3216         */
3217        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_GUEST_BASE, guest_base);
3218        tcg_regset_set_reg(s->reserved_regs, TCG_REG_GUEST_BASE);
3219    }
3220
3221    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
3222    tcg_out_insn(s, 3207, BR, tcg_target_call_iarg_regs[1]);
3223
3224    /*
3225     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
3226     * and fall through to the rest of the epilogue.
3227     */
3228    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
3229    tcg_out_bti(s, BTI_J);
3230    tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_X0, 0);
3231
3232    /* TB epilogue */
3233    tb_ret_addr = tcg_splitwx_to_rx(s->code_ptr);
3234    tcg_out_bti(s, BTI_J);
3235
3236    /* Remove TCG locals stack space.  */
3237    tcg_out_insn(s, 3401, ADDI, TCG_TYPE_I64, TCG_REG_SP, TCG_REG_SP,
3238                 FRAME_SIZE - PUSH_SIZE);
3239
3240    /* Restore registers x19..x28.  */
3241    for (r = TCG_REG_X19; r <= TCG_REG_X27; r += 2) {
3242        int ofs = (r - TCG_REG_X19 + 2) * 8;
3243        tcg_out_insn(s, 3314, LDP, r, r + 1, TCG_REG_SP, ofs, 1, 0);
3244    }
3245
3246    /* Pop (FP, LR), restore SP to previous frame.  */
3247    tcg_out_insn(s, 3314, LDP, TCG_REG_FP, TCG_REG_LR,
3248                 TCG_REG_SP, PUSH_SIZE, 0, 1);
3249    tcg_out_insn(s, 3207, RET, TCG_REG_LR);
3250}
3251
3252static void tcg_out_tb_start(TCGContext *s)
3253{
3254    tcg_out_bti(s, BTI_J);
3255}
3256
3257static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
3258{
3259    int i;
3260    for (i = 0; i < count; ++i) {
3261        p[i] = NOP;
3262    }
3263}
3264
3265typedef struct {
3266    DebugFrameHeader h;
3267    uint8_t fde_def_cfa[4];
3268    uint8_t fde_reg_ofs[24];
3269} DebugFrame;
3270
3271#define ELF_HOST_MACHINE EM_AARCH64
3272
3273static const DebugFrame debug_frame = {
3274    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
3275    .h.cie.id = -1,
3276    .h.cie.version = 1,
3277    .h.cie.code_align = 1,
3278    .h.cie.data_align = 0x78,             /* sleb128 -8 */
3279    .h.cie.return_column = TCG_REG_LR,
3280
3281    /* Total FDE size does not include the "len" member.  */
3282    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
3283
3284    .fde_def_cfa = {
3285        12, TCG_REG_SP,                 /* DW_CFA_def_cfa sp, ... */
3286        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
3287        (FRAME_SIZE >> 7)
3288    },
3289    .fde_reg_ofs = {
3290        0x80 + 28, 1,                   /* DW_CFA_offset, x28,  -8 */
3291        0x80 + 27, 2,                   /* DW_CFA_offset, x27, -16 */
3292        0x80 + 26, 3,                   /* DW_CFA_offset, x26, -24 */
3293        0x80 + 25, 4,                   /* DW_CFA_offset, x25, -32 */
3294        0x80 + 24, 5,                   /* DW_CFA_offset, x24, -40 */
3295        0x80 + 23, 6,                   /* DW_CFA_offset, x23, -48 */
3296        0x80 + 22, 7,                   /* DW_CFA_offset, x22, -56 */
3297        0x80 + 21, 8,                   /* DW_CFA_offset, x21, -64 */
3298        0x80 + 20, 9,                   /* DW_CFA_offset, x20, -72 */
3299        0x80 + 19, 10,                  /* DW_CFA_offset, x1p, -80 */
3300        0x80 + 30, 11,                  /* DW_CFA_offset,  lr, -88 */
3301        0x80 + 29, 12,                  /* DW_CFA_offset,  fp, -96 */
3302    }
3303};
3304
3305void tcg_register_jit(const void *buf, size_t buf_size)
3306{
3307    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
3308}
3309