xref: /openbmc/qemu/tcg/arm/tcg-target.c.inc (revision aa28c9ef8e109db40d4781d82452805486f2a2bf)
1/*
2 * Tiny Code Generator for QEMU
3 *
4 * Copyright (c) 2008 Andrzej Zaborowski
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to deal
8 * in the Software without restriction, including without limitation the rights
9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10 * copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22 * THE SOFTWARE.
23 */
24
25#include "elf.h"
26
27int arm_arch = __ARM_ARCH;
28
29#ifndef use_idiv_instructions
30bool use_idiv_instructions;
31#endif
32#ifndef use_neon_instructions
33bool use_neon_instructions;
34#endif
35
36/* Used for function call generation. */
37#define TCG_TARGET_STACK_ALIGN          8
38#define TCG_TARGET_CALL_STACK_OFFSET    0
39#define TCG_TARGET_CALL_ARG_I32         TCG_CALL_ARG_NORMAL
40#define TCG_TARGET_CALL_ARG_I64         TCG_CALL_ARG_EVEN
41#define TCG_TARGET_CALL_ARG_I128        TCG_CALL_ARG_EVEN
42#define TCG_TARGET_CALL_RET_I128        TCG_CALL_RET_BY_REF
43
44#ifdef CONFIG_DEBUG_TCG
45static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
46    "%r0",  "%r1",  "%r2",  "%r3",  "%r4",  "%r5",  "%r6",  "%r7",
47    "%r8",  "%r9",  "%r10", "%r11", "%r12", "%sp",  "%r14", "%pc",
48    "%q0",  "%q1",  "%q2",  "%q3",  "%q4",  "%q5",  "%q6",  "%q7",
49    "%q8",  "%q9",  "%q10", "%q11", "%q12", "%q13", "%q14", "%q15",
50};
51#endif
52
53static const int tcg_target_reg_alloc_order[] = {
54    TCG_REG_R4,
55    TCG_REG_R5,
56    TCG_REG_R6,
57    TCG_REG_R7,
58    TCG_REG_R8,
59    TCG_REG_R9,
60    TCG_REG_R10,
61    TCG_REG_R11,
62    TCG_REG_R13,
63    TCG_REG_R0,
64    TCG_REG_R1,
65    TCG_REG_R2,
66    TCG_REG_R3,
67    TCG_REG_R12,
68    TCG_REG_R14,
69
70    TCG_REG_Q0,
71    TCG_REG_Q1,
72    TCG_REG_Q2,
73    TCG_REG_Q3,
74    /* Q4 - Q7 are call-saved, and skipped. */
75    TCG_REG_Q8,
76    TCG_REG_Q9,
77    TCG_REG_Q10,
78    TCG_REG_Q11,
79    TCG_REG_Q12,
80    TCG_REG_Q13,
81    TCG_REG_Q14,
82    TCG_REG_Q15,
83};
84
85static const int tcg_target_call_iarg_regs[4] = {
86    TCG_REG_R0, TCG_REG_R1, TCG_REG_R2, TCG_REG_R3
87};
88
89static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
90{
91    tcg_debug_assert(kind == TCG_CALL_RET_NORMAL);
92    tcg_debug_assert(slot >= 0 && slot <= 3);
93    return TCG_REG_R0 + slot;
94}
95
96#define TCG_REG_TMP  TCG_REG_R12
97#define TCG_VEC_TMP  TCG_REG_Q15
98#define TCG_REG_GUEST_BASE  TCG_REG_R11
99
100typedef enum {
101    COND_EQ = 0x0,
102    COND_NE = 0x1,
103    COND_CS = 0x2,	/* Unsigned greater or equal */
104    COND_CC = 0x3,	/* Unsigned less than */
105    COND_MI = 0x4,	/* Negative */
106    COND_PL = 0x5,	/* Zero or greater */
107    COND_VS = 0x6,	/* Overflow */
108    COND_VC = 0x7,	/* No overflow */
109    COND_HI = 0x8,	/* Unsigned greater than */
110    COND_LS = 0x9,	/* Unsigned less or equal */
111    COND_GE = 0xa,
112    COND_LT = 0xb,
113    COND_GT = 0xc,
114    COND_LE = 0xd,
115    COND_AL = 0xe,
116} ARMCond;
117
118#define TO_CPSR (1 << 20)
119
120#define SHIFT_IMM_LSL(im)	(((im) << 7) | 0x00)
121#define SHIFT_IMM_LSR(im)	(((im) << 7) | 0x20)
122#define SHIFT_IMM_ASR(im)	(((im) << 7) | 0x40)
123#define SHIFT_IMM_ROR(im)	(((im) << 7) | 0x60)
124#define SHIFT_REG_LSL(rs)	(((rs) << 8) | 0x10)
125#define SHIFT_REG_LSR(rs)	(((rs) << 8) | 0x30)
126#define SHIFT_REG_ASR(rs)	(((rs) << 8) | 0x50)
127#define SHIFT_REG_ROR(rs)	(((rs) << 8) | 0x70)
128
129typedef enum {
130    ARITH_AND = 0x0 << 21,
131    ARITH_EOR = 0x1 << 21,
132    ARITH_SUB = 0x2 << 21,
133    ARITH_RSB = 0x3 << 21,
134    ARITH_ADD = 0x4 << 21,
135    ARITH_ADC = 0x5 << 21,
136    ARITH_SBC = 0x6 << 21,
137    ARITH_RSC = 0x7 << 21,
138    ARITH_TST = 0x8 << 21 | TO_CPSR,
139    ARITH_CMP = 0xa << 21 | TO_CPSR,
140    ARITH_CMN = 0xb << 21 | TO_CPSR,
141    ARITH_ORR = 0xc << 21,
142    ARITH_MOV = 0xd << 21,
143    ARITH_BIC = 0xe << 21,
144    ARITH_MVN = 0xf << 21,
145
146    INSN_B         = 0x0a000000,
147
148    INSN_CLZ       = 0x016f0f10,
149    INSN_RBIT      = 0x06ff0f30,
150
151    INSN_LDMIA     = 0x08b00000,
152    INSN_STMDB     = 0x09200000,
153
154    INSN_LDR_IMM   = 0x04100000,
155    INSN_LDR_REG   = 0x06100000,
156    INSN_STR_IMM   = 0x04000000,
157    INSN_STR_REG   = 0x06000000,
158
159    INSN_LDRH_IMM  = 0x005000b0,
160    INSN_LDRH_REG  = 0x001000b0,
161    INSN_LDRSH_IMM = 0x005000f0,
162    INSN_LDRSH_REG = 0x001000f0,
163    INSN_STRH_IMM  = 0x004000b0,
164    INSN_STRH_REG  = 0x000000b0,
165
166    INSN_LDRB_IMM  = 0x04500000,
167    INSN_LDRB_REG  = 0x06500000,
168    INSN_LDRSB_IMM = 0x005000d0,
169    INSN_LDRSB_REG = 0x001000d0,
170    INSN_STRB_IMM  = 0x04400000,
171    INSN_STRB_REG  = 0x06400000,
172
173    INSN_LDRD_IMM  = 0x004000d0,
174    INSN_LDRD_REG  = 0x000000d0,
175    INSN_STRD_IMM  = 0x004000f0,
176    INSN_STRD_REG  = 0x000000f0,
177
178    INSN_DMB_ISH   = 0xf57ff05b,
179    INSN_DMB_MCR   = 0xee070fba,
180
181    /* Architected nop introduced in v6k.  */
182    /* ??? This is an MSR (imm) 0,0,0 insn.  Anyone know if this
183       also Just So Happened to do nothing on pre-v6k so that we
184       don't need to conditionalize it?  */
185    INSN_NOP_v6k   = 0xe320f000,
186    /* Otherwise the assembler uses mov r0,r0 */
187    INSN_NOP_v4    = (COND_AL << 28) | ARITH_MOV,
188
189    INSN_VADD      = 0xf2000800,
190    INSN_VAND      = 0xf2000110,
191    INSN_VBIC      = 0xf2100110,
192    INSN_VEOR      = 0xf3000110,
193    INSN_VORN      = 0xf2300110,
194    INSN_VORR      = 0xf2200110,
195    INSN_VSUB      = 0xf3000800,
196    INSN_VMUL      = 0xf2000910,
197    INSN_VQADD     = 0xf2000010,
198    INSN_VQADD_U   = 0xf3000010,
199    INSN_VQSUB     = 0xf2000210,
200    INSN_VQSUB_U   = 0xf3000210,
201    INSN_VMAX      = 0xf2000600,
202    INSN_VMAX_U    = 0xf3000600,
203    INSN_VMIN      = 0xf2000610,
204    INSN_VMIN_U    = 0xf3000610,
205
206    INSN_VABS      = 0xf3b10300,
207    INSN_VMVN      = 0xf3b00580,
208    INSN_VNEG      = 0xf3b10380,
209
210    INSN_VCEQ0     = 0xf3b10100,
211    INSN_VCGT0     = 0xf3b10000,
212    INSN_VCGE0     = 0xf3b10080,
213    INSN_VCLE0     = 0xf3b10180,
214    INSN_VCLT0     = 0xf3b10200,
215
216    INSN_VCEQ      = 0xf3000810,
217    INSN_VCGE      = 0xf2000310,
218    INSN_VCGT      = 0xf2000300,
219    INSN_VCGE_U    = 0xf3000310,
220    INSN_VCGT_U    = 0xf3000300,
221
222    INSN_VSHLI     = 0xf2800510,  /* VSHL (immediate) */
223    INSN_VSARI     = 0xf2800010,  /* VSHR.S */
224    INSN_VSHRI     = 0xf3800010,  /* VSHR.U */
225    INSN_VSLI      = 0xf3800510,
226    INSN_VSHL_S    = 0xf2000400,  /* VSHL.S (register) */
227    INSN_VSHL_U    = 0xf3000400,  /* VSHL.U (register) */
228
229    INSN_VBSL      = 0xf3100110,
230    INSN_VBIT      = 0xf3200110,
231    INSN_VBIF      = 0xf3300110,
232
233    INSN_VTST      = 0xf2000810,
234
235    INSN_VDUP_G    = 0xee800b10,  /* VDUP (ARM core register) */
236    INSN_VDUP_S    = 0xf3b00c00,  /* VDUP (scalar) */
237    INSN_VLDR_D    = 0xed100b00,  /* VLDR.64 */
238    INSN_VLD1      = 0xf4200000,  /* VLD1 (multiple single elements) */
239    INSN_VLD1R     = 0xf4a00c00,  /* VLD1 (single element to all lanes) */
240    INSN_VST1      = 0xf4000000,  /* VST1 (multiple single elements) */
241    INSN_VMOVI     = 0xf2800010,  /* VMOV (immediate) */
242} ARMInsn;
243
244#define INSN_NOP   (use_armv7_instructions ? INSN_NOP_v6k : INSN_NOP_v4)
245
246static const uint8_t tcg_cond_to_arm_cond[] = {
247    [TCG_COND_EQ] = COND_EQ,
248    [TCG_COND_NE] = COND_NE,
249    [TCG_COND_LT] = COND_LT,
250    [TCG_COND_GE] = COND_GE,
251    [TCG_COND_LE] = COND_LE,
252    [TCG_COND_GT] = COND_GT,
253    /* unsigned */
254    [TCG_COND_LTU] = COND_CC,
255    [TCG_COND_GEU] = COND_CS,
256    [TCG_COND_LEU] = COND_LS,
257    [TCG_COND_GTU] = COND_HI,
258};
259
260static int encode_imm(uint32_t imm);
261
262/* TCG private relocation type: add with pc+imm8 */
263#define R_ARM_PC8  11
264
265/* TCG private relocation type: vldr with imm8 << 2 */
266#define R_ARM_PC11 12
267
268static bool reloc_pc24(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
269{
270    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
271    ptrdiff_t offset = (tcg_ptr_byte_diff(target, src_rx) - 8) >> 2;
272
273    if (offset == sextract32(offset, 0, 24)) {
274        *src_rw = deposit32(*src_rw, 0, 24, offset);
275        return true;
276    }
277    return false;
278}
279
280static bool reloc_pc13(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
281{
282    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
283    ptrdiff_t offset = tcg_ptr_byte_diff(target, src_rx) - 8;
284
285    if (offset >= -0xfff && offset <= 0xfff) {
286        tcg_insn_unit insn = *src_rw;
287        bool u = (offset >= 0);
288        if (!u) {
289            offset = -offset;
290        }
291        insn = deposit32(insn, 23, 1, u);
292        insn = deposit32(insn, 0, 12, offset);
293        *src_rw = insn;
294        return true;
295    }
296    return false;
297}
298
299static bool reloc_pc11(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
300{
301    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
302    ptrdiff_t offset = (tcg_ptr_byte_diff(target, src_rx) - 8) / 4;
303
304    if (offset >= -0xff && offset <= 0xff) {
305        tcg_insn_unit insn = *src_rw;
306        bool u = (offset >= 0);
307        if (!u) {
308            offset = -offset;
309        }
310        insn = deposit32(insn, 23, 1, u);
311        insn = deposit32(insn, 0, 8, offset);
312        *src_rw = insn;
313        return true;
314    }
315    return false;
316}
317
318static bool reloc_pc8(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
319{
320    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
321    ptrdiff_t offset = tcg_ptr_byte_diff(target, src_rx) - 8;
322    int imm12 = encode_imm(offset);
323
324    if (imm12 >= 0) {
325        *src_rw = deposit32(*src_rw, 0, 12, imm12);
326        return true;
327    }
328    return false;
329}
330
331static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
332                        intptr_t value, intptr_t addend)
333{
334    tcg_debug_assert(addend == 0);
335    switch (type) {
336    case R_ARM_PC24:
337        return reloc_pc24(code_ptr, (const tcg_insn_unit *)value);
338    case R_ARM_PC13:
339        return reloc_pc13(code_ptr, (const tcg_insn_unit *)value);
340    case R_ARM_PC11:
341        return reloc_pc11(code_ptr, (const tcg_insn_unit *)value);
342    case R_ARM_PC8:
343        return reloc_pc8(code_ptr, (const tcg_insn_unit *)value);
344    default:
345        g_assert_not_reached();
346    }
347}
348
349#define TCG_CT_CONST_ARM  0x100
350#define TCG_CT_CONST_INV  0x200
351#define TCG_CT_CONST_NEG  0x400
352#define TCG_CT_CONST_ZERO 0x800
353#define TCG_CT_CONST_ORRI 0x1000
354#define TCG_CT_CONST_ANDI 0x2000
355
356#define ALL_GENERAL_REGS  0xffffu
357#define ALL_VECTOR_REGS   0xffff0000u
358
359/*
360 * r0-r3 will be overwritten when reading the tlb entry (system-mode only);
361 * r14 will be overwritten by the BLNE branching to the slow path.
362 */
363#define ALL_QLDST_REGS \
364    (ALL_GENERAL_REGS & ~((tcg_use_softmmu ? 0xf : 0) | (1 << TCG_REG_R14)))
365
366/*
367 * ARM immediates for ALU instructions are made of an unsigned 8-bit
368 * right-rotated by an even amount between 0 and 30.
369 *
370 * Return < 0 if @imm cannot be encoded, else the entire imm12 field.
371 */
372static int encode_imm(uint32_t imm)
373{
374    uint32_t rot, imm8;
375
376    /* Simple case, no rotation required. */
377    if ((imm & ~0xff) == 0) {
378        return imm;
379    }
380
381    /* Next, try a simple even shift.  */
382    rot = ctz32(imm) & ~1;
383    imm8 = imm >> rot;
384    rot = 32 - rot;
385    if ((imm8 & ~0xff) == 0) {
386        goto found;
387    }
388
389    /*
390     * Finally, try harder with rotations.
391     * The ctz test above will have taken care of rotates >= 8.
392     */
393    for (rot = 2; rot < 8; rot += 2) {
394        imm8 = rol32(imm, rot);
395        if ((imm8 & ~0xff) == 0) {
396            goto found;
397        }
398    }
399    /* Fail: imm cannot be encoded. */
400    return -1;
401
402 found:
403    /* Note that rot is even, and we discard bit 0 by shifting by 7. */
404    return rot << 7 | imm8;
405}
406
407static int encode_imm_nofail(uint32_t imm)
408{
409    int ret = encode_imm(imm);
410    tcg_debug_assert(ret >= 0);
411    return ret;
412}
413
414static bool check_fit_imm(uint32_t imm)
415{
416    return encode_imm(imm) >= 0;
417}
418
419/* Return true if v16 is a valid 16-bit shifted immediate.  */
420static bool is_shimm16(uint16_t v16, int *cmode, int *imm8)
421{
422    if (v16 == (v16 & 0xff)) {
423        *cmode = 0x8;
424        *imm8 = v16 & 0xff;
425        return true;
426    } else if (v16 == (v16 & 0xff00)) {
427        *cmode = 0xa;
428        *imm8 = v16 >> 8;
429        return true;
430    }
431    return false;
432}
433
434/* Return true if v32 is a valid 32-bit shifted immediate.  */
435static bool is_shimm32(uint32_t v32, int *cmode, int *imm8)
436{
437    if (v32 == (v32 & 0xff)) {
438        *cmode = 0x0;
439        *imm8 = v32 & 0xff;
440        return true;
441    } else if (v32 == (v32 & 0xff00)) {
442        *cmode = 0x2;
443        *imm8 = (v32 >> 8) & 0xff;
444        return true;
445    } else if (v32 == (v32 & 0xff0000)) {
446        *cmode = 0x4;
447        *imm8 = (v32 >> 16) & 0xff;
448        return true;
449    } else if (v32 == (v32 & 0xff000000)) {
450        *cmode = 0x6;
451        *imm8 = v32 >> 24;
452        return true;
453    }
454    return false;
455}
456
457/* Return true if v32 is a valid 32-bit shifting ones immediate.  */
458static bool is_soimm32(uint32_t v32, int *cmode, int *imm8)
459{
460    if ((v32 & 0xffff00ff) == 0xff) {
461        *cmode = 0xc;
462        *imm8 = (v32 >> 8) & 0xff;
463        return true;
464    } else if ((v32 & 0xff00ffff) == 0xffff) {
465        *cmode = 0xd;
466        *imm8 = (v32 >> 16) & 0xff;
467        return true;
468    }
469    return false;
470}
471
472/*
473 * Return non-zero if v32 can be formed by MOVI+ORR.
474 * Place the parameters for MOVI in (cmode, imm8).
475 * Return the cmode for ORR; the imm8 can be had via extraction from v32.
476 */
477static int is_shimm32_pair(uint32_t v32, int *cmode, int *imm8)
478{
479    int i;
480
481    for (i = 6; i > 0; i -= 2) {
482        /* Mask out one byte we can add with ORR.  */
483        uint32_t tmp = v32 & ~(0xffu << (i * 4));
484        if (is_shimm32(tmp, cmode, imm8) ||
485            is_soimm32(tmp, cmode, imm8)) {
486            break;
487        }
488    }
489    return i;
490}
491
492/* Return true if V is a valid 16-bit or 32-bit shifted immediate.  */
493static bool is_shimm1632(uint32_t v32, int *cmode, int *imm8)
494{
495    if (v32 == deposit32(v32, 16, 16, v32)) {
496        return is_shimm16(v32, cmode, imm8);
497    } else {
498        return is_shimm32(v32, cmode, imm8);
499    }
500}
501
502/* Test if a constant matches the constraint.
503 * TODO: define constraints for:
504 *
505 * ldr/str offset:   between -0xfff and 0xfff
506 * ldrh/strh offset: between -0xff and 0xff
507 * mov operand2:     values represented with x << (2 * y), x < 0x100
508 * add, sub, eor...: ditto
509 */
510static bool tcg_target_const_match(int64_t val, int ct,
511                                   TCGType type, TCGCond cond, int vece)
512{
513    if (ct & TCG_CT_CONST) {
514        return 1;
515    } else if ((ct & TCG_CT_CONST_ARM) && check_fit_imm(val)) {
516        return 1;
517    } else if ((ct & TCG_CT_CONST_INV) && check_fit_imm(~val)) {
518        return 1;
519    } else if ((ct & TCG_CT_CONST_NEG) && check_fit_imm(-val)) {
520        return 1;
521    } else if ((ct & TCG_CT_CONST_ZERO) && val == 0) {
522        return 1;
523    }
524
525    switch (ct & (TCG_CT_CONST_ORRI | TCG_CT_CONST_ANDI)) {
526    case 0:
527        break;
528    case TCG_CT_CONST_ANDI:
529        val = ~val;
530        /* fallthru */
531    case TCG_CT_CONST_ORRI:
532        if (val == deposit64(val, 32, 32, val)) {
533            int cmode, imm8;
534            return is_shimm1632(val, &cmode, &imm8);
535        }
536        break;
537    default:
538        /* Both bits should not be set for the same insn.  */
539        g_assert_not_reached();
540    }
541
542    return 0;
543}
544
545static void tcg_out_b_imm(TCGContext *s, ARMCond cond, int32_t offset)
546{
547    tcg_out32(s, (cond << 28) | INSN_B |
548                    (((offset - 8) >> 2) & 0x00ffffff));
549}
550
551static void tcg_out_bl_imm(TCGContext *s, ARMCond cond, int32_t offset)
552{
553    tcg_out32(s, (cond << 28) | 0x0b000000 |
554                    (((offset - 8) >> 2) & 0x00ffffff));
555}
556
557static void tcg_out_blx_reg(TCGContext *s, ARMCond cond, TCGReg rn)
558{
559    tcg_out32(s, (cond << 28) | 0x012fff30 | rn);
560}
561
562static void tcg_out_blx_imm(TCGContext *s, int32_t offset)
563{
564    tcg_out32(s, 0xfa000000 | ((offset & 2) << 23) |
565                (((offset - 8) >> 2) & 0x00ffffff));
566}
567
568static void tcg_out_dat_reg(TCGContext *s, ARMCond cond, ARMInsn opc,
569                            TCGReg rd, TCGReg rn, TCGReg rm, int shift)
570{
571    tcg_out32(s, (cond << 28) | (0 << 25) | opc |
572                    (rn << 16) | (rd << 12) | shift | rm);
573}
574
575static void tcg_out_mov_reg(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rm)
576{
577    /* Simple reg-reg move, optimising out the 'do nothing' case */
578    if (rd != rm) {
579        tcg_out_dat_reg(s, cond, ARITH_MOV, rd, 0, rm, SHIFT_IMM_LSL(0));
580    }
581}
582
583static void tcg_out_bx_reg(TCGContext *s, ARMCond cond, TCGReg rn)
584{
585    tcg_out32(s, (cond << 28) | 0x012fff10 | rn);
586}
587
588static void tcg_out_b_reg(TCGContext *s, ARMCond cond, TCGReg rn)
589{
590    /*
591     * Unless the C portion of QEMU is compiled as thumb, we don't need
592     * true BX semantics; merely a branch to an address held in a register.
593     */
594    tcg_out_bx_reg(s, cond, rn);
595}
596
597static void tcg_out_dat_imm(TCGContext *s, ARMCond cond, ARMInsn opc,
598                            TCGReg rd, TCGReg rn, int im)
599{
600    tcg_out32(s, (cond << 28) | (1 << 25) | opc |
601                    (rn << 16) | (rd << 12) | im);
602}
603
604static void tcg_out_ldstm(TCGContext *s, ARMCond cond, ARMInsn opc,
605                          TCGReg rn, uint16_t mask)
606{
607    tcg_out32(s, (cond << 28) | opc | (rn << 16) | mask);
608}
609
610/* Note that this routine is used for both LDR and LDRH formats, so we do
611   not wish to include an immediate shift at this point.  */
612static void tcg_out_memop_r(TCGContext *s, ARMCond cond, ARMInsn opc, TCGReg rt,
613                            TCGReg rn, TCGReg rm, bool u, bool p, bool w)
614{
615    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24)
616              | (w << 21) | (rn << 16) | (rt << 12) | rm);
617}
618
619static void tcg_out_memop_8(TCGContext *s, ARMCond cond, ARMInsn opc, TCGReg rt,
620                            TCGReg rn, int imm8, bool p, bool w)
621{
622    bool u = 1;
623    if (imm8 < 0) {
624        imm8 = -imm8;
625        u = 0;
626    }
627    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24) | (w << 21) |
628              (rn << 16) | (rt << 12) | ((imm8 & 0xf0) << 4) | (imm8 & 0xf));
629}
630
631static void tcg_out_memop_12(TCGContext *s, ARMCond cond, ARMInsn opc,
632                             TCGReg rt, TCGReg rn, int imm12, bool p, bool w)
633{
634    bool u = 1;
635    if (imm12 < 0) {
636        imm12 = -imm12;
637        u = 0;
638    }
639    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24) | (w << 21) |
640              (rn << 16) | (rt << 12) | imm12);
641}
642
643static void tcg_out_ld32_12(TCGContext *s, ARMCond cond, TCGReg rt,
644                            TCGReg rn, int imm12)
645{
646    tcg_out_memop_12(s, cond, INSN_LDR_IMM, rt, rn, imm12, 1, 0);
647}
648
649static void tcg_out_st32_12(TCGContext *s, ARMCond cond, TCGReg rt,
650                            TCGReg rn, int imm12)
651{
652    tcg_out_memop_12(s, cond, INSN_STR_IMM, rt, rn, imm12, 1, 0);
653}
654
655static void tcg_out_ld32_r(TCGContext *s, ARMCond cond, TCGReg rt,
656                           TCGReg rn, TCGReg rm)
657{
658    tcg_out_memop_r(s, cond, INSN_LDR_REG, rt, rn, rm, 1, 1, 0);
659}
660
661static void tcg_out_st32_r(TCGContext *s, ARMCond cond, TCGReg rt,
662                           TCGReg rn, TCGReg rm)
663{
664    tcg_out_memop_r(s, cond, INSN_STR_REG, rt, rn, rm, 1, 1, 0);
665}
666
667static void tcg_out_ldrd_8(TCGContext *s, ARMCond cond, TCGReg rt,
668                           TCGReg rn, int imm8)
669{
670    tcg_out_memop_8(s, cond, INSN_LDRD_IMM, rt, rn, imm8, 1, 0);
671}
672
673static void tcg_out_ldrd_r(TCGContext *s, ARMCond cond, TCGReg rt,
674                           TCGReg rn, TCGReg rm)
675{
676    tcg_out_memop_r(s, cond, INSN_LDRD_REG, rt, rn, rm, 1, 1, 0);
677}
678
679static void tcg_out_strd_8(TCGContext *s, ARMCond cond, TCGReg rt,
680                           TCGReg rn, int imm8)
681{
682    tcg_out_memop_8(s, cond, INSN_STRD_IMM, rt, rn, imm8, 1, 0);
683}
684
685static void tcg_out_strd_r(TCGContext *s, ARMCond cond, TCGReg rt,
686                           TCGReg rn, TCGReg rm)
687{
688    tcg_out_memop_r(s, cond, INSN_STRD_REG, rt, rn, rm, 1, 1, 0);
689}
690
691/* Register pre-increment with base writeback.  */
692static void tcg_out_ld32_rwb(TCGContext *s, ARMCond cond, TCGReg rt,
693                             TCGReg rn, TCGReg rm)
694{
695    tcg_out_memop_r(s, cond, INSN_LDR_REG, rt, rn, rm, 1, 1, 1);
696}
697
698static void tcg_out_st32_rwb(TCGContext *s, ARMCond cond, TCGReg rt,
699                             TCGReg rn, TCGReg rm)
700{
701    tcg_out_memop_r(s, cond, INSN_STR_REG, rt, rn, rm, 1, 1, 1);
702}
703
704static void tcg_out_ld16u_8(TCGContext *s, ARMCond cond, TCGReg rt,
705                            TCGReg rn, int imm8)
706{
707    tcg_out_memop_8(s, cond, INSN_LDRH_IMM, rt, rn, imm8, 1, 0);
708}
709
710static void tcg_out_st16_8(TCGContext *s, ARMCond cond, TCGReg rt,
711                           TCGReg rn, int imm8)
712{
713    tcg_out_memop_8(s, cond, INSN_STRH_IMM, rt, rn, imm8, 1, 0);
714}
715
716static void tcg_out_ld16u_r(TCGContext *s, ARMCond cond, TCGReg rt,
717                            TCGReg rn, TCGReg rm)
718{
719    tcg_out_memop_r(s, cond, INSN_LDRH_REG, rt, rn, rm, 1, 1, 0);
720}
721
722static void tcg_out_st16_r(TCGContext *s, ARMCond cond, TCGReg rt,
723                           TCGReg rn, TCGReg rm)
724{
725    tcg_out_memop_r(s, cond, INSN_STRH_REG, rt, rn, rm, 1, 1, 0);
726}
727
728static void tcg_out_ld16s_8(TCGContext *s, ARMCond cond, TCGReg rt,
729                            TCGReg rn, int imm8)
730{
731    tcg_out_memop_8(s, cond, INSN_LDRSH_IMM, rt, rn, imm8, 1, 0);
732}
733
734static void tcg_out_ld16s_r(TCGContext *s, ARMCond cond, TCGReg rt,
735                            TCGReg rn, TCGReg rm)
736{
737    tcg_out_memop_r(s, cond, INSN_LDRSH_REG, rt, rn, rm, 1, 1, 0);
738}
739
740static void tcg_out_ld8_12(TCGContext *s, ARMCond cond, TCGReg rt,
741                           TCGReg rn, int imm12)
742{
743    tcg_out_memop_12(s, cond, INSN_LDRB_IMM, rt, rn, imm12, 1, 0);
744}
745
746static void tcg_out_st8_12(TCGContext *s, ARMCond cond, TCGReg rt,
747                           TCGReg rn, int imm12)
748{
749    tcg_out_memop_12(s, cond, INSN_STRB_IMM, rt, rn, imm12, 1, 0);
750}
751
752static void tcg_out_ld8_r(TCGContext *s, ARMCond cond, TCGReg rt,
753                          TCGReg rn, TCGReg rm)
754{
755    tcg_out_memop_r(s, cond, INSN_LDRB_REG, rt, rn, rm, 1, 1, 0);
756}
757
758static void tcg_out_st8_r(TCGContext *s, ARMCond cond, TCGReg rt,
759                          TCGReg rn, TCGReg rm)
760{
761    tcg_out_memop_r(s, cond, INSN_STRB_REG, rt, rn, rm, 1, 1, 0);
762}
763
764static void tcg_out_ld8s_8(TCGContext *s, ARMCond cond, TCGReg rt,
765                           TCGReg rn, int imm8)
766{
767    tcg_out_memop_8(s, cond, INSN_LDRSB_IMM, rt, rn, imm8, 1, 0);
768}
769
770static void tcg_out_ld8s_r(TCGContext *s, ARMCond cond, TCGReg rt,
771                           TCGReg rn, TCGReg rm)
772{
773    tcg_out_memop_r(s, cond, INSN_LDRSB_REG, rt, rn, rm, 1, 1, 0);
774}
775
776static void tcg_out_movi_pool(TCGContext *s, ARMCond cond,
777                              TCGReg rd, uint32_t arg)
778{
779    new_pool_label(s, arg, R_ARM_PC13, s->code_ptr, 0);
780    tcg_out_ld32_12(s, cond, rd, TCG_REG_PC, 0);
781}
782
783static void tcg_out_movi32(TCGContext *s, ARMCond cond,
784                           TCGReg rd, uint32_t arg)
785{
786    int imm12, diff, opc, sh1, sh2;
787    uint32_t tt0, tt1, tt2;
788
789    /* Check a single MOV/MVN before anything else.  */
790    imm12 = encode_imm(arg);
791    if (imm12 >= 0) {
792        tcg_out_dat_imm(s, cond, ARITH_MOV, rd, 0, imm12);
793        return;
794    }
795    imm12 = encode_imm(~arg);
796    if (imm12 >= 0) {
797        tcg_out_dat_imm(s, cond, ARITH_MVN, rd, 0, imm12);
798        return;
799    }
800
801    /* Check for a pc-relative address.  This will usually be the TB,
802       or within the TB, which is immediately before the code block.  */
803    diff = tcg_pcrel_diff(s, (void *)arg) - 8;
804    if (diff >= 0) {
805        imm12 = encode_imm(diff);
806        if (imm12 >= 0) {
807            tcg_out_dat_imm(s, cond, ARITH_ADD, rd, TCG_REG_PC, imm12);
808            return;
809        }
810    } else {
811        imm12 = encode_imm(-diff);
812        if (imm12 >= 0) {
813            tcg_out_dat_imm(s, cond, ARITH_SUB, rd, TCG_REG_PC, imm12);
814            return;
815        }
816    }
817
818    /* Use movw + movt.  */
819    if (use_armv7_instructions) {
820        /* movw */
821        tcg_out32(s, (cond << 28) | 0x03000000 | (rd << 12)
822                  | ((arg << 4) & 0x000f0000) | (arg & 0xfff));
823        if (arg & 0xffff0000) {
824            /* movt */
825            tcg_out32(s, (cond << 28) | 0x03400000 | (rd << 12)
826                      | ((arg >> 12) & 0x000f0000) | ((arg >> 16) & 0xfff));
827        }
828        return;
829    }
830
831    /* Look for sequences of two insns.  If we have lots of 1's, we can
832       shorten the sequence by beginning with mvn and then clearing
833       higher bits with eor.  */
834    tt0 = arg;
835    opc = ARITH_MOV;
836    if (ctpop32(arg) > 16) {
837        tt0 = ~arg;
838        opc = ARITH_MVN;
839    }
840    sh1 = ctz32(tt0) & ~1;
841    tt1 = tt0 & ~(0xff << sh1);
842    sh2 = ctz32(tt1) & ~1;
843    tt2 = tt1 & ~(0xff << sh2);
844    if (tt2 == 0) {
845        int rot;
846
847        rot = ((32 - sh1) << 7) & 0xf00;
848        tcg_out_dat_imm(s, cond, opc, rd,  0, ((tt0 >> sh1) & 0xff) | rot);
849        rot = ((32 - sh2) << 7) & 0xf00;
850        tcg_out_dat_imm(s, cond, ARITH_EOR, rd, rd,
851                        ((tt0 >> sh2) & 0xff) | rot);
852        return;
853    }
854
855    /* Otherwise, drop it into the constant pool.  */
856    tcg_out_movi_pool(s, cond, rd, arg);
857}
858
859/*
860 * Emit either the reg,imm or reg,reg form of a data-processing insn.
861 * rhs must satisfy the "rI" constraint.
862 */
863static void tcg_out_dat_rI(TCGContext *s, ARMCond cond, ARMInsn opc,
864                           TCGReg dst, TCGReg lhs, TCGArg rhs, int rhs_is_const)
865{
866    if (rhs_is_const) {
867        tcg_out_dat_imm(s, cond, opc, dst, lhs, encode_imm_nofail(rhs));
868    } else {
869        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
870    }
871}
872
873/*
874 * Emit either the reg,imm or reg,reg form of a data-processing insn.
875 * rhs must satisfy the "rIK" constraint.
876 */
877static void tcg_out_dat_IK(TCGContext *s, ARMCond cond, ARMInsn opc,
878                            ARMInsn opinv, TCGReg dst, TCGReg lhs, TCGArg rhs)
879{
880    int imm12 = encode_imm(rhs);
881    if (imm12 < 0) {
882        imm12 = encode_imm_nofail(~rhs);
883        opc = opinv;
884    }
885    tcg_out_dat_imm(s, cond, opc, dst, lhs, imm12);
886}
887
888static void tcg_out_dat_rIK(TCGContext *s, ARMCond cond, ARMInsn opc,
889                            ARMInsn opinv, TCGReg dst, TCGReg lhs, TCGArg rhs,
890                            bool rhs_is_const)
891{
892    if (rhs_is_const) {
893        tcg_out_dat_IK(s, cond, opc, opinv, dst, lhs, rhs);
894    } else {
895        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
896    }
897}
898
899static void tcg_out_dat_IN(TCGContext *s, ARMCond cond, ARMInsn opc,
900                           ARMInsn opneg, TCGReg dst, TCGReg lhs, TCGArg rhs)
901{
902    int imm12 = encode_imm(rhs);
903    if (imm12 < 0) {
904        imm12 = encode_imm_nofail(-rhs);
905        opc = opneg;
906    }
907    tcg_out_dat_imm(s, cond, opc, dst, lhs, imm12);
908}
909
910static void tcg_out_dat_rIN(TCGContext *s, ARMCond cond, ARMInsn opc,
911                            ARMInsn opneg, TCGReg dst, TCGReg lhs, TCGArg rhs,
912                            bool rhs_is_const)
913{
914    /* Emit either the reg,imm or reg,reg form of a data-processing insn.
915     * rhs must satisfy the "rIN" constraint.
916     */
917    if (rhs_is_const) {
918        tcg_out_dat_IN(s, cond, opc, opneg, dst, lhs, rhs);
919    } else {
920        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
921    }
922}
923
924static void tcg_out_umull32(TCGContext *s, ARMCond cond, TCGReg rd0,
925                            TCGReg rd1, TCGReg rn, TCGReg rm)
926{
927    /* umull */
928    tcg_out32(s, (cond << 28) | 0x00800090 |
929              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
930}
931
932static void tcg_out_smull32(TCGContext *s, ARMCond cond, TCGReg rd0,
933                            TCGReg rd1, TCGReg rn, TCGReg rm)
934{
935    /* smull */
936    tcg_out32(s, (cond << 28) | 0x00c00090 |
937              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
938}
939
940static void tcg_out_sdiv(TCGContext *s, ARMCond cond,
941                         TCGReg rd, TCGReg rn, TCGReg rm)
942{
943    tcg_out32(s, 0x0710f010 | (cond << 28) | (rd << 16) | rn | (rm << 8));
944}
945
946static void tcg_out_udiv(TCGContext *s, ARMCond cond,
947                         TCGReg rd, TCGReg rn, TCGReg rm)
948{
949    tcg_out32(s, 0x0730f010 | (cond << 28) | (rd << 16) | rn | (rm << 8));
950}
951
952static void tcg_out_ext8s(TCGContext *s, TCGType t, TCGReg rd, TCGReg rn)
953{
954    /* sxtb */
955    tcg_out32(s, 0x06af0070 | (COND_AL << 28) | (rd << 12) | rn);
956}
957
958static void tcg_out_ext8u(TCGContext *s, TCGReg rd, TCGReg rn)
959{
960    tcg_out_dat_imm(s, COND_AL, ARITH_AND, rd, rn, 0xff);
961}
962
963static void tcg_out_ext16s(TCGContext *s, TCGType t, TCGReg rd, TCGReg rn)
964{
965    /* sxth */
966    tcg_out32(s, 0x06bf0070 | (COND_AL << 28) | (rd << 12) | rn);
967}
968
969static void tcg_out_ext16u(TCGContext *s, TCGReg rd, TCGReg rn)
970{
971    /* uxth */
972    tcg_out32(s, 0x06ff0070 | (COND_AL << 28) | (rd << 12) | rn);
973}
974
975static void tcg_out_ext32s(TCGContext *s, TCGReg rd, TCGReg rn)
976{
977    g_assert_not_reached();
978}
979
980static void tcg_out_ext32u(TCGContext *s, TCGReg rd, TCGReg rn)
981{
982    g_assert_not_reached();
983}
984
985static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg rd, TCGReg rn)
986{
987    g_assert_not_reached();
988}
989
990static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg rd, TCGReg rn)
991{
992    g_assert_not_reached();
993}
994
995static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg rd, TCGReg rn)
996{
997    g_assert_not_reached();
998}
999
1000static void tcg_out_bswap16(TCGContext *s, ARMCond cond,
1001                            TCGReg rd, TCGReg rn, int flags)
1002{
1003    if (flags & TCG_BSWAP_OS) {
1004        /* revsh */
1005        tcg_out32(s, 0x06ff0fb0 | (cond << 28) | (rd << 12) | rn);
1006        return;
1007    }
1008
1009    /* rev16 */
1010    tcg_out32(s, 0x06bf0fb0 | (cond << 28) | (rd << 12) | rn);
1011    if ((flags & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1012        /* uxth */
1013        tcg_out32(s, 0x06ff0070 | (cond << 28) | (rd << 12) | rd);
1014    }
1015}
1016
1017static void tcg_out_bswap32(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rn)
1018{
1019    /* rev */
1020    tcg_out32(s, 0x06bf0f30 | (cond << 28) | (rd << 12) | rn);
1021}
1022
1023static void tcg_out_deposit(TCGContext *s, ARMCond cond, TCGReg rd,
1024                            TCGArg a1, int ofs, int len, bool const_a1)
1025{
1026    if (const_a1) {
1027        /* bfi becomes bfc with rn == 15.  */
1028        a1 = 15;
1029    }
1030    /* bfi/bfc */
1031    tcg_out32(s, 0x07c00010 | (cond << 28) | (rd << 12) | a1
1032              | (ofs << 7) | ((ofs + len - 1) << 16));
1033}
1034
1035static void tcg_out_extract(TCGContext *s, ARMCond cond, TCGReg rd,
1036                            TCGReg rn, int ofs, int len)
1037{
1038    /* According to gcc, AND can be faster. */
1039    if (ofs == 0 && len <= 8) {
1040        tcg_out_dat_imm(s, cond, ARITH_AND, rd, rn,
1041                        encode_imm_nofail((1 << len) - 1));
1042        return;
1043    }
1044
1045    if (use_armv7_instructions) {
1046        /* ubfx */
1047        tcg_out32(s, 0x07e00050 | (cond << 28) | (rd << 12) | rn
1048                  | (ofs << 7) | ((len - 1) << 16));
1049        return;
1050    }
1051
1052    assert(ofs % 8 == 0);
1053    switch (len) {
1054    case 8:
1055        /* uxtb */
1056        tcg_out32(s, 0x06ef0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1057        break;
1058    case 16:
1059        /* uxth */
1060        tcg_out32(s, 0x06ff0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1061        break;
1062    default:
1063        g_assert_not_reached();
1064    }
1065}
1066
1067static void tcg_out_sextract(TCGContext *s, ARMCond cond, TCGReg rd,
1068                             TCGReg rn, int ofs, int len)
1069{
1070    if (use_armv7_instructions) {
1071        /* sbfx */
1072        tcg_out32(s, 0x07a00050 | (cond << 28) | (rd << 12) | rn
1073                  | (ofs << 7) | ((len - 1) << 16));
1074        return;
1075    }
1076
1077    assert(ofs % 8 == 0);
1078    switch (len) {
1079    case 8:
1080        /* sxtb */
1081        tcg_out32(s, 0x06af0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1082        break;
1083    case 16:
1084        /* sxth */
1085        tcg_out32(s, 0x06bf0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1086        break;
1087    default:
1088        g_assert_not_reached();
1089    }
1090}
1091
1092
1093static void tcg_out_ld32u(TCGContext *s, ARMCond cond,
1094                          TCGReg rd, TCGReg rn, int32_t offset)
1095{
1096    if (offset > 0xfff || offset < -0xfff) {
1097        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1098        tcg_out_ld32_r(s, cond, rd, rn, TCG_REG_TMP);
1099    } else
1100        tcg_out_ld32_12(s, cond, rd, rn, offset);
1101}
1102
1103static void tcg_out_st32(TCGContext *s, ARMCond cond,
1104                         TCGReg rd, TCGReg rn, int32_t offset)
1105{
1106    if (offset > 0xfff || offset < -0xfff) {
1107        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1108        tcg_out_st32_r(s, cond, rd, rn, TCG_REG_TMP);
1109    } else
1110        tcg_out_st32_12(s, cond, rd, rn, offset);
1111}
1112
1113static void tcg_out_ld16u(TCGContext *s, ARMCond cond,
1114                          TCGReg rd, TCGReg rn, int32_t offset)
1115{
1116    if (offset > 0xff || offset < -0xff) {
1117        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1118        tcg_out_ld16u_r(s, cond, rd, rn, TCG_REG_TMP);
1119    } else
1120        tcg_out_ld16u_8(s, cond, rd, rn, offset);
1121}
1122
1123static void tcg_out_ld16s(TCGContext *s, ARMCond cond,
1124                          TCGReg rd, TCGReg rn, int32_t offset)
1125{
1126    if (offset > 0xff || offset < -0xff) {
1127        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1128        tcg_out_ld16s_r(s, cond, rd, rn, TCG_REG_TMP);
1129    } else
1130        tcg_out_ld16s_8(s, cond, rd, rn, offset);
1131}
1132
1133static void tcg_out_st16(TCGContext *s, ARMCond cond,
1134                         TCGReg rd, TCGReg rn, int32_t offset)
1135{
1136    if (offset > 0xff || offset < -0xff) {
1137        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1138        tcg_out_st16_r(s, cond, rd, rn, TCG_REG_TMP);
1139    } else
1140        tcg_out_st16_8(s, cond, rd, rn, offset);
1141}
1142
1143static void tcg_out_ld8u(TCGContext *s, ARMCond cond,
1144                         TCGReg rd, TCGReg rn, int32_t offset)
1145{
1146    if (offset > 0xfff || offset < -0xfff) {
1147        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1148        tcg_out_ld8_r(s, cond, rd, rn, TCG_REG_TMP);
1149    } else
1150        tcg_out_ld8_12(s, cond, rd, rn, offset);
1151}
1152
1153static void tcg_out_ld8s(TCGContext *s, ARMCond cond,
1154                         TCGReg rd, TCGReg rn, int32_t offset)
1155{
1156    if (offset > 0xff || offset < -0xff) {
1157        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1158        tcg_out_ld8s_r(s, cond, rd, rn, TCG_REG_TMP);
1159    } else
1160        tcg_out_ld8s_8(s, cond, rd, rn, offset);
1161}
1162
1163static void tcg_out_st8(TCGContext *s, ARMCond cond,
1164                        TCGReg rd, TCGReg rn, int32_t offset)
1165{
1166    if (offset > 0xfff || offset < -0xfff) {
1167        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1168        tcg_out_st8_r(s, cond, rd, rn, TCG_REG_TMP);
1169    } else
1170        tcg_out_st8_12(s, cond, rd, rn, offset);
1171}
1172
1173/*
1174 * The _goto case is normally between TBs within the same code buffer, and
1175 * with the code buffer limited to 16MB we wouldn't need the long case.
1176 * But we also use it for the tail-call to the qemu_ld/st helpers, which does.
1177 */
1178static void tcg_out_goto(TCGContext *s, ARMCond cond, const tcg_insn_unit *addr)
1179{
1180    intptr_t addri = (intptr_t)addr;
1181    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1182    bool arm_mode = !(addri & 1);
1183
1184    if (arm_mode && disp - 8 < 0x01fffffd && disp - 8 > -0x01fffffd) {
1185        tcg_out_b_imm(s, cond, disp);
1186        return;
1187    }
1188
1189    /* LDR is interworking from v5t. */
1190    tcg_out_movi_pool(s, cond, TCG_REG_PC, addri);
1191}
1192
1193/*
1194 * The call case is mostly used for helpers - so it's not unreasonable
1195 * for them to be beyond branch range.
1196 */
1197static void tcg_out_call_int(TCGContext *s, const tcg_insn_unit *addr)
1198{
1199    intptr_t addri = (intptr_t)addr;
1200    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1201    bool arm_mode = !(addri & 1);
1202
1203    if (disp - 8 < 0x02000000 && disp - 8 >= -0x02000000) {
1204        if (arm_mode) {
1205            tcg_out_bl_imm(s, COND_AL, disp);
1206        } else {
1207            tcg_out_blx_imm(s, disp);
1208        }
1209        return;
1210    }
1211
1212    tcg_out_movi32(s, COND_AL, TCG_REG_TMP, addri);
1213    tcg_out_blx_reg(s, COND_AL, TCG_REG_TMP);
1214}
1215
1216static void tcg_out_call(TCGContext *s, const tcg_insn_unit *addr,
1217                         const TCGHelperInfo *info)
1218{
1219    tcg_out_call_int(s, addr);
1220}
1221
1222static void tcg_out_goto_label(TCGContext *s, ARMCond cond, TCGLabel *l)
1223{
1224    if (l->has_value) {
1225        tcg_out_goto(s, cond, l->u.value_ptr);
1226    } else {
1227        tcg_out_reloc(s, s->code_ptr, R_ARM_PC24, l, 0);
1228        tcg_out_b_imm(s, cond, 0);
1229    }
1230}
1231
1232static void tcg_out_mb(TCGContext *s, TCGArg a0)
1233{
1234    if (use_armv7_instructions) {
1235        tcg_out32(s, INSN_DMB_ISH);
1236    } else {
1237        tcg_out32(s, INSN_DMB_MCR);
1238    }
1239}
1240
1241static TCGCond tcg_out_cmp(TCGContext *s, TCGCond cond, TCGReg a,
1242                           TCGArg b, int b_const)
1243{
1244    if (!is_tst_cond(cond)) {
1245        tcg_out_dat_rIN(s, COND_AL, ARITH_CMP, ARITH_CMN, 0, a, b, b_const);
1246        return cond;
1247    }
1248
1249    cond = tcg_tst_eqne_cond(cond);
1250    if (b_const) {
1251        int imm12 = encode_imm(b);
1252
1253        /*
1254         * The compare constraints allow rIN, but TST does not support N.
1255         * Be prepared to load the constant into a scratch register.
1256         */
1257        if (imm12 >= 0) {
1258            tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, a, imm12);
1259            return cond;
1260        }
1261        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, b);
1262        b = TCG_REG_TMP;
1263    }
1264    tcg_out_dat_reg(s, COND_AL, ARITH_TST, 0, a, b, SHIFT_IMM_LSL(0));
1265    return cond;
1266}
1267
1268static TCGCond tcg_out_cmp2(TCGContext *s, const TCGArg *args,
1269                            const int *const_args)
1270{
1271    TCGReg al = args[0];
1272    TCGReg ah = args[1];
1273    TCGArg bl = args[2];
1274    TCGArg bh = args[3];
1275    TCGCond cond = args[4];
1276    int const_bl = const_args[2];
1277    int const_bh = const_args[3];
1278
1279    switch (cond) {
1280    case TCG_COND_EQ:
1281    case TCG_COND_NE:
1282    case TCG_COND_LTU:
1283    case TCG_COND_LEU:
1284    case TCG_COND_GTU:
1285    case TCG_COND_GEU:
1286        /*
1287         * We perform a conditional comparison.  If the high half is
1288         * equal, then overwrite the flags with the comparison of the
1289         * low half.  The resulting flags cover the whole.
1290         */
1291        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, ah, bh, const_bh);
1292        tcg_out_dat_rI(s, COND_EQ, ARITH_CMP, 0, al, bl, const_bl);
1293        return cond;
1294
1295    case TCG_COND_TSTEQ:
1296    case TCG_COND_TSTNE:
1297        /* Similar, but with TST instead of CMP. */
1298        tcg_out_dat_rI(s, COND_AL, ARITH_TST, 0, ah, bh, const_bh);
1299        tcg_out_dat_rI(s, COND_EQ, ARITH_TST, 0, al, bl, const_bl);
1300        return tcg_tst_eqne_cond(cond);
1301
1302    case TCG_COND_LT:
1303    case TCG_COND_GE:
1304        /* We perform a double-word subtraction and examine the result.
1305           We do not actually need the result of the subtract, so the
1306           low part "subtract" is a compare.  For the high half we have
1307           no choice but to compute into a temporary.  */
1308        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, al, bl, const_bl);
1309        tcg_out_dat_rI(s, COND_AL, ARITH_SBC | TO_CPSR,
1310                       TCG_REG_TMP, ah, bh, const_bh);
1311        return cond;
1312
1313    case TCG_COND_LE:
1314    case TCG_COND_GT:
1315        /* Similar, but with swapped arguments, via reversed subtract.  */
1316        tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR,
1317                       TCG_REG_TMP, al, bl, const_bl);
1318        tcg_out_dat_rI(s, COND_AL, ARITH_RSC | TO_CPSR,
1319                       TCG_REG_TMP, ah, bh, const_bh);
1320        return tcg_swap_cond(cond);
1321
1322    default:
1323        g_assert_not_reached();
1324    }
1325}
1326
1327/*
1328 * Note that TCGReg references Q-registers.
1329 * Q-regno = 2 * D-regno, so shift left by 1 while inserting.
1330 */
1331static uint32_t encode_vd(TCGReg rd)
1332{
1333    tcg_debug_assert(rd >= TCG_REG_Q0);
1334    return (extract32(rd, 3, 1) << 22) | (extract32(rd, 0, 3) << 13);
1335}
1336
1337static uint32_t encode_vn(TCGReg rn)
1338{
1339    tcg_debug_assert(rn >= TCG_REG_Q0);
1340    return (extract32(rn, 3, 1) << 7) | (extract32(rn, 0, 3) << 17);
1341}
1342
1343static uint32_t encode_vm(TCGReg rm)
1344{
1345    tcg_debug_assert(rm >= TCG_REG_Q0);
1346    return (extract32(rm, 3, 1) << 5) | (extract32(rm, 0, 3) << 1);
1347}
1348
1349static void tcg_out_vreg2(TCGContext *s, ARMInsn insn, int q, int vece,
1350                          TCGReg d, TCGReg m)
1351{
1352    tcg_out32(s, insn | (vece << 18) | (q << 6) |
1353              encode_vd(d) | encode_vm(m));
1354}
1355
1356static void tcg_out_vreg3(TCGContext *s, ARMInsn insn, int q, int vece,
1357                          TCGReg d, TCGReg n, TCGReg m)
1358{
1359    tcg_out32(s, insn | (vece << 20) | (q << 6) |
1360              encode_vd(d) | encode_vn(n) | encode_vm(m));
1361}
1362
1363static void tcg_out_vmovi(TCGContext *s, TCGReg rd,
1364                          int q, int op, int cmode, uint8_t imm8)
1365{
1366    tcg_out32(s, INSN_VMOVI | encode_vd(rd) | (q << 6) | (op << 5)
1367              | (cmode << 8) | extract32(imm8, 0, 4)
1368              | (extract32(imm8, 4, 3) << 16)
1369              | (extract32(imm8, 7, 1) << 24));
1370}
1371
1372static void tcg_out_vshifti(TCGContext *s, ARMInsn insn, int q,
1373                            TCGReg rd, TCGReg rm, int l_imm6)
1374{
1375    tcg_out32(s, insn | (q << 6) | encode_vd(rd) | encode_vm(rm) |
1376              (extract32(l_imm6, 6, 1) << 7) |
1377              (extract32(l_imm6, 0, 6) << 16));
1378}
1379
1380static void tcg_out_vldst(TCGContext *s, ARMInsn insn,
1381                          TCGReg rd, TCGReg rn, int offset)
1382{
1383    if (offset != 0) {
1384        if (check_fit_imm(offset) || check_fit_imm(-offset)) {
1385            tcg_out_dat_rIN(s, COND_AL, ARITH_ADD, ARITH_SUB,
1386                            TCG_REG_TMP, rn, offset, true);
1387        } else {
1388            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP, offset);
1389            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1390                            TCG_REG_TMP, TCG_REG_TMP, rn, 0);
1391        }
1392        rn = TCG_REG_TMP;
1393    }
1394    tcg_out32(s, insn | (rn << 16) | encode_vd(rd) | 0xf);
1395}
1396
1397typedef struct {
1398    ARMCond cond;
1399    TCGReg base;
1400    int index;
1401    bool index_scratch;
1402    TCGAtomAlign aa;
1403} HostAddress;
1404
1405bool tcg_target_has_memory_bswap(MemOp memop)
1406{
1407    return false;
1408}
1409
1410static TCGReg ldst_ra_gen(TCGContext *s, const TCGLabelQemuLdst *l, int arg)
1411{
1412    /* We arrive at the slow path via "BLNE", so R14 contains l->raddr. */
1413    return TCG_REG_R14;
1414}
1415
1416static const TCGLdstHelperParam ldst_helper_param = {
1417    .ra_gen = ldst_ra_gen,
1418    .ntmp = 1,
1419    .tmp = { TCG_REG_TMP },
1420};
1421
1422static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1423{
1424    MemOp opc = get_memop(lb->oi);
1425
1426    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1427        return false;
1428    }
1429
1430    tcg_out_ld_helper_args(s, lb, &ldst_helper_param);
1431    tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SIZE]);
1432    tcg_out_ld_helper_ret(s, lb, false, &ldst_helper_param);
1433
1434    tcg_out_goto(s, COND_AL, lb->raddr);
1435    return true;
1436}
1437
1438static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1439{
1440    MemOp opc = get_memop(lb->oi);
1441
1442    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1443        return false;
1444    }
1445
1446    tcg_out_st_helper_args(s, lb, &ldst_helper_param);
1447
1448    /* Tail-call to the helper, which will return to the fast path.  */
1449    tcg_out_goto(s, COND_AL, qemu_st_helpers[opc & MO_SIZE]);
1450    return true;
1451}
1452
1453/* We expect to use an 9-bit sign-magnitude negative offset from ENV.  */
1454#define MIN_TLB_MASK_TABLE_OFS  -256
1455
1456static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
1457                                           TCGReg addr, MemOpIdx oi, bool is_ld)
1458{
1459    TCGLabelQemuLdst *ldst = NULL;
1460    MemOp opc = get_memop(oi);
1461    unsigned a_mask;
1462
1463    if (tcg_use_softmmu) {
1464        *h = (HostAddress){
1465            .cond = COND_AL,
1466            .base = addr,
1467            .index = TCG_REG_R1,
1468            .index_scratch = true,
1469        };
1470    } else {
1471        *h = (HostAddress){
1472            .cond = COND_AL,
1473            .base = addr,
1474            .index = guest_base ? TCG_REG_GUEST_BASE : -1,
1475            .index_scratch = false,
1476        };
1477    }
1478
1479    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, false);
1480    a_mask = (1 << h->aa.align) - 1;
1481
1482    if (tcg_use_softmmu) {
1483        int mem_index = get_mmuidx(oi);
1484        int cmp_off = is_ld ? offsetof(CPUTLBEntry, addr_read)
1485                            : offsetof(CPUTLBEntry, addr_write);
1486        int fast_off = tlb_mask_table_ofs(s, mem_index);
1487        unsigned s_mask = (1 << (opc & MO_SIZE)) - 1;
1488        TCGReg t_addr;
1489
1490        ldst = new_ldst_label(s);
1491        ldst->is_ld = is_ld;
1492        ldst->oi = oi;
1493        ldst->addr_reg = addr;
1494
1495        /* Load cpu->neg.tlb.f[mmu_idx].{mask,table} into {r0,r1}.  */
1496        QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, mask) != 0);
1497        QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, table) != 4);
1498        tcg_out_ldrd_8(s, COND_AL, TCG_REG_R0, TCG_AREG0, fast_off);
1499
1500        /* Extract the tlb index from the address into R0.  */
1501        tcg_out_dat_reg(s, COND_AL, ARITH_AND, TCG_REG_R0, TCG_REG_R0, addr,
1502                        SHIFT_IMM_LSR(s->page_bits - CPU_TLB_ENTRY_BITS));
1503
1504        /*
1505         * Add the tlb_table pointer, creating the CPUTLBEntry address in R1.
1506         * Load the tlb comparator into R2 and the fast path addend into R1.
1507         */
1508        if (cmp_off == 0) {
1509            tcg_out_ld32_rwb(s, COND_AL, TCG_REG_R2, TCG_REG_R1, TCG_REG_R0);
1510        } else {
1511            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1512                            TCG_REG_R1, TCG_REG_R1, TCG_REG_R0, 0);
1513            tcg_out_ld32_12(s, COND_AL, TCG_REG_R2, TCG_REG_R1, cmp_off);
1514        }
1515
1516        /* Load the tlb addend.  */
1517        tcg_out_ld32_12(s, COND_AL, TCG_REG_R1, TCG_REG_R1,
1518                        offsetof(CPUTLBEntry, addend));
1519
1520        /*
1521         * Check alignment, check comparators.
1522         * Do this in 2-4 insns.  Use MOVW for v7, if possible,
1523         * to reduce the number of sequential conditional instructions.
1524         * Almost all guests have at least 4k pages, which means that we need
1525         * to clear at least 9 bits even for an 8-byte memory, which means it
1526         * isn't worth checking for an immediate operand for BIC.
1527         *
1528         * For unaligned accesses, test the page of the last unit of alignment.
1529         * This leaves the least significant alignment bits unchanged, and of
1530         * course must be zero.
1531         */
1532        t_addr = addr;
1533        if (a_mask < s_mask) {
1534            t_addr = TCG_REG_R0;
1535            tcg_out_dat_imm(s, COND_AL, ARITH_ADD, t_addr,
1536                            addr, s_mask - a_mask);
1537        }
1538        if (use_armv7_instructions && s->page_bits <= 16) {
1539            tcg_out_movi32(s, COND_AL, TCG_REG_TMP, ~(s->page_mask | a_mask));
1540            tcg_out_dat_reg(s, COND_AL, ARITH_BIC, TCG_REG_TMP,
1541                            t_addr, TCG_REG_TMP, 0);
1542            tcg_out_dat_reg(s, COND_AL, ARITH_CMP, 0,
1543                            TCG_REG_R2, TCG_REG_TMP, 0);
1544        } else {
1545            if (a_mask) {
1546                tcg_debug_assert(a_mask <= 0xff);
1547                tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addr, a_mask);
1548            }
1549            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0, t_addr,
1550                            SHIFT_IMM_LSR(s->page_bits));
1551            tcg_out_dat_reg(s, (a_mask ? COND_EQ : COND_AL), ARITH_CMP,
1552                            0, TCG_REG_R2, TCG_REG_TMP,
1553                            SHIFT_IMM_LSL(s->page_bits));
1554        }
1555    } else if (a_mask) {
1556        ldst = new_ldst_label(s);
1557        ldst->is_ld = is_ld;
1558        ldst->oi = oi;
1559        ldst->addr_reg = addr;
1560
1561        /* We are expecting alignment to max out at 7 */
1562        tcg_debug_assert(a_mask <= 0xff);
1563        /* tst addr, #mask */
1564        tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addr, a_mask);
1565    }
1566
1567    return ldst;
1568}
1569
1570static void tcg_out_qemu_ld_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1571                                   TCGReg datahi, HostAddress h)
1572{
1573    TCGReg base;
1574
1575    /* Byte swapping is left to middle-end expansion. */
1576    tcg_debug_assert((opc & MO_BSWAP) == 0);
1577
1578    switch (opc & MO_SSIZE) {
1579    case MO_UB:
1580        if (h.index < 0) {
1581            tcg_out_ld8_12(s, h.cond, datalo, h.base, 0);
1582        } else {
1583            tcg_out_ld8_r(s, h.cond, datalo, h.base, h.index);
1584        }
1585        break;
1586    case MO_SB:
1587        if (h.index < 0) {
1588            tcg_out_ld8s_8(s, h.cond, datalo, h.base, 0);
1589        } else {
1590            tcg_out_ld8s_r(s, h.cond, datalo, h.base, h.index);
1591        }
1592        break;
1593    case MO_UW:
1594        if (h.index < 0) {
1595            tcg_out_ld16u_8(s, h.cond, datalo, h.base, 0);
1596        } else {
1597            tcg_out_ld16u_r(s, h.cond, datalo, h.base, h.index);
1598        }
1599        break;
1600    case MO_SW:
1601        if (h.index < 0) {
1602            tcg_out_ld16s_8(s, h.cond, datalo, h.base, 0);
1603        } else {
1604            tcg_out_ld16s_r(s, h.cond, datalo, h.base, h.index);
1605        }
1606        break;
1607    case MO_UL:
1608        if (h.index < 0) {
1609            tcg_out_ld32_12(s, h.cond, datalo, h.base, 0);
1610        } else {
1611            tcg_out_ld32_r(s, h.cond, datalo, h.base, h.index);
1612        }
1613        break;
1614    case MO_UQ:
1615        /* We used pair allocation for datalo, so already should be aligned. */
1616        tcg_debug_assert((datalo & 1) == 0);
1617        tcg_debug_assert(datahi == datalo + 1);
1618        /* LDRD requires alignment; double-check that. */
1619        if (memop_alignment_bits(opc) >= MO_64) {
1620            if (h.index < 0) {
1621                tcg_out_ldrd_8(s, h.cond, datalo, h.base, 0);
1622                break;
1623            }
1624            /*
1625             * Rm (the second address op) must not overlap Rt or Rt + 1.
1626             * Since datalo is aligned, we can simplify the test via alignment.
1627             * Flip the two address arguments if that works.
1628             */
1629            if ((h.index & ~1) != datalo) {
1630                tcg_out_ldrd_r(s, h.cond, datalo, h.base, h.index);
1631                break;
1632            }
1633            if ((h.base & ~1) != datalo) {
1634                tcg_out_ldrd_r(s, h.cond, datalo, h.index, h.base);
1635                break;
1636            }
1637        }
1638        if (h.index < 0) {
1639            base = h.base;
1640            if (datalo == h.base) {
1641                tcg_out_mov_reg(s, h.cond, TCG_REG_TMP, base);
1642                base = TCG_REG_TMP;
1643            }
1644        } else if (h.index_scratch) {
1645            tcg_out_ld32_rwb(s, h.cond, datalo, h.index, h.base);
1646            tcg_out_ld32_12(s, h.cond, datahi, h.index, 4);
1647            break;
1648        } else {
1649            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1650                            h.base, h.index, SHIFT_IMM_LSL(0));
1651            base = TCG_REG_TMP;
1652        }
1653        tcg_out_ld32_12(s, h.cond, datalo, base, 0);
1654        tcg_out_ld32_12(s, h.cond, datahi, base, 4);
1655        break;
1656    default:
1657        g_assert_not_reached();
1658    }
1659}
1660
1661static void tcg_out_qemu_ld(TCGContext *s, TCGReg datalo, TCGReg datahi,
1662                            TCGReg addr, MemOpIdx oi, TCGType data_type)
1663{
1664    MemOp opc = get_memop(oi);
1665    TCGLabelQemuLdst *ldst;
1666    HostAddress h;
1667
1668    ldst = prepare_host_addr(s, &h, addr, oi, true);
1669    if (ldst) {
1670        ldst->type = data_type;
1671        ldst->datalo_reg = datalo;
1672        ldst->datahi_reg = datahi;
1673
1674        /*
1675         * This a conditional BL only to load a pointer within this
1676         * opcode into LR for the slow path.  We will not be using
1677         * the value for a tail call.
1678         */
1679        ldst->label_ptr[0] = s->code_ptr;
1680        tcg_out_bl_imm(s, COND_NE, 0);
1681
1682        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1683        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1684    } else {
1685        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1686    }
1687}
1688
1689static void tcg_out_qemu_st_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1690                                   TCGReg datahi, HostAddress h)
1691{
1692    /* Byte swapping is left to middle-end expansion. */
1693    tcg_debug_assert((opc & MO_BSWAP) == 0);
1694
1695    switch (opc & MO_SIZE) {
1696    case MO_8:
1697        if (h.index < 0) {
1698            tcg_out_st8_12(s, h.cond, datalo, h.base, 0);
1699        } else {
1700            tcg_out_st8_r(s, h.cond, datalo, h.base, h.index);
1701        }
1702        break;
1703    case MO_16:
1704        if (h.index < 0) {
1705            tcg_out_st16_8(s, h.cond, datalo, h.base, 0);
1706        } else {
1707            tcg_out_st16_r(s, h.cond, datalo, h.base, h.index);
1708        }
1709        break;
1710    case MO_32:
1711        if (h.index < 0) {
1712            tcg_out_st32_12(s, h.cond, datalo, h.base, 0);
1713        } else {
1714            tcg_out_st32_r(s, h.cond, datalo, h.base, h.index);
1715        }
1716        break;
1717    case MO_64:
1718        /* We used pair allocation for datalo, so already should be aligned. */
1719        tcg_debug_assert((datalo & 1) == 0);
1720        tcg_debug_assert(datahi == datalo + 1);
1721        /* STRD requires alignment; double-check that. */
1722        if (memop_alignment_bits(opc) >= MO_64) {
1723            if (h.index < 0) {
1724                tcg_out_strd_8(s, h.cond, datalo, h.base, 0);
1725            } else {
1726                tcg_out_strd_r(s, h.cond, datalo, h.base, h.index);
1727            }
1728        } else if (h.index < 0) {
1729            tcg_out_st32_12(s, h.cond, datalo, h.base, 0);
1730            tcg_out_st32_12(s, h.cond, datahi, h.base, 4);
1731        } else if (h.index_scratch) {
1732            tcg_out_st32_rwb(s, h.cond, datalo, h.index, h.base);
1733            tcg_out_st32_12(s, h.cond, datahi, h.index, 4);
1734        } else {
1735            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1736                            h.base, h.index, SHIFT_IMM_LSL(0));
1737            tcg_out_st32_12(s, h.cond, datalo, TCG_REG_TMP, 0);
1738            tcg_out_st32_12(s, h.cond, datahi, TCG_REG_TMP, 4);
1739        }
1740        break;
1741    default:
1742        g_assert_not_reached();
1743    }
1744}
1745
1746static void tcg_out_qemu_st(TCGContext *s, TCGReg datalo, TCGReg datahi,
1747                            TCGReg addr, MemOpIdx oi, TCGType data_type)
1748{
1749    MemOp opc = get_memop(oi);
1750    TCGLabelQemuLdst *ldst;
1751    HostAddress h;
1752
1753    ldst = prepare_host_addr(s, &h, addr, oi, false);
1754    if (ldst) {
1755        ldst->type = data_type;
1756        ldst->datalo_reg = datalo;
1757        ldst->datahi_reg = datahi;
1758
1759        h.cond = COND_EQ;
1760        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1761
1762        /* The conditional call is last, as we're going to return here. */
1763        ldst->label_ptr[0] = s->code_ptr;
1764        tcg_out_bl_imm(s, COND_NE, 0);
1765        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1766    } else {
1767        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1768    }
1769}
1770
1771static void tcg_out_epilogue(TCGContext *s);
1772
1773static void tcg_out_exit_tb(TCGContext *s, uintptr_t arg)
1774{
1775    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, arg);
1776    tcg_out_epilogue(s);
1777}
1778
1779static void tcg_out_goto_tb(TCGContext *s, int which)
1780{
1781    uintptr_t i_addr;
1782    intptr_t i_disp;
1783
1784    /* Direct branch will be patched by tb_target_set_jmp_target. */
1785    set_jmp_insn_offset(s, which);
1786    tcg_out32(s, INSN_NOP);
1787
1788    /* When branch is out of range, fall through to indirect. */
1789    i_addr = get_jmp_target_addr(s, which);
1790    i_disp = tcg_pcrel_diff(s, (void *)i_addr) - 8;
1791    tcg_debug_assert(i_disp < 0);
1792    if (i_disp >= -0xfff) {
1793        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_PC, i_disp);
1794    } else {
1795        /*
1796         * The TB is close, but outside the 12 bits addressable by
1797         * the load.  We can extend this to 20 bits with a sub of a
1798         * shifted immediate from pc.
1799         */
1800        int h = -i_disp;
1801        int l = -(h & 0xfff);
1802
1803        h = encode_imm_nofail(h + l);
1804        tcg_out_dat_imm(s, COND_AL, ARITH_SUB, TCG_REG_R0, TCG_REG_PC, h);
1805        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_R0, l);
1806    }
1807    set_jmp_reset_offset(s, which);
1808}
1809
1810void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
1811                              uintptr_t jmp_rx, uintptr_t jmp_rw)
1812{
1813    uintptr_t addr = tb->jmp_target_addr[n];
1814    ptrdiff_t offset = addr - (jmp_rx + 8);
1815    tcg_insn_unit insn;
1816
1817    /* Either directly branch, or fall through to indirect branch. */
1818    if (offset == sextract64(offset, 0, 26)) {
1819        /* B <addr> */
1820        insn = deposit32((COND_AL << 28) | INSN_B, 0, 24, offset >> 2);
1821    } else {
1822        insn = INSN_NOP;
1823    }
1824
1825    qatomic_set((uint32_t *)jmp_rw, insn);
1826    flush_idcache_range(jmp_rx, jmp_rw, 4);
1827}
1828
1829
1830static void tgen_add(TCGContext *s, TCGType type,
1831                     TCGReg a0, TCGReg a1, TCGReg a2)
1832{
1833    tcg_out_dat_reg(s, COND_AL, ARITH_ADD, a0, a1, a2, SHIFT_IMM_LSL(0));
1834}
1835
1836static void tgen_addi(TCGContext *s, TCGType type,
1837                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1838{
1839    tcg_out_dat_IN(s, COND_AL, ARITH_ADD, ARITH_SUB, a0, a1, a2);
1840}
1841
1842static const TCGOutOpBinary outop_add = {
1843    .base.static_constraint = C_O1_I2(r, r, rIN),
1844    .out_rrr = tgen_add,
1845    .out_rri = tgen_addi,
1846};
1847
1848static void tgen_and(TCGContext *s, TCGType type,
1849                     TCGReg a0, TCGReg a1, TCGReg a2)
1850{
1851    tcg_out_dat_reg(s, COND_AL, ARITH_AND, a0, a1, a2, SHIFT_IMM_LSL(0));
1852}
1853
1854static void tgen_andi(TCGContext *s, TCGType type,
1855                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1856{
1857    tcg_out_dat_IK(s, COND_AL, ARITH_AND, ARITH_BIC, a0, a1, a2);
1858}
1859
1860static const TCGOutOpBinary outop_and = {
1861    .base.static_constraint = C_O1_I2(r, r, rIK),
1862    .out_rrr = tgen_and,
1863    .out_rri = tgen_andi,
1864};
1865
1866static void tgen_andc(TCGContext *s, TCGType type,
1867                      TCGReg a0, TCGReg a1, TCGReg a2)
1868{
1869    tcg_out_dat_reg(s, COND_AL, ARITH_BIC, a0, a1, a2, SHIFT_IMM_LSL(0));
1870}
1871
1872static const TCGOutOpBinary outop_andc = {
1873    .base.static_constraint = C_O1_I2(r, r, r),
1874    .out_rrr = tgen_andc,
1875};
1876
1877static const TCGOutOpBinary outop_eqv = {
1878    .base.static_constraint = C_NotImplemented,
1879};
1880
1881static void tgen_mul(TCGContext *s, TCGType type,
1882                     TCGReg a0, TCGReg a1, TCGReg a2)
1883{
1884    /* mul */
1885    tcg_out32(s, (COND_AL << 28) | 0x90 | (a0 << 16) | (a1 << 8) | a2);
1886}
1887
1888static const TCGOutOpBinary outop_mul = {
1889    .base.static_constraint = C_O1_I2(r, r, r),
1890    .out_rrr = tgen_mul,
1891};
1892
1893static const TCGOutOpBinary outop_muluh = {
1894    .base.static_constraint = C_NotImplemented,
1895};
1896
1897static const TCGOutOpBinary outop_nand = {
1898    .base.static_constraint = C_NotImplemented,
1899};
1900
1901static const TCGOutOpBinary outop_nor = {
1902    .base.static_constraint = C_NotImplemented,
1903};
1904
1905static void tgen_or(TCGContext *s, TCGType type,
1906                     TCGReg a0, TCGReg a1, TCGReg a2)
1907{
1908    tcg_out_dat_reg(s, COND_AL, ARITH_ORR, a0, a1, a2, SHIFT_IMM_LSL(0));
1909}
1910
1911static void tgen_ori(TCGContext *s, TCGType type,
1912                     TCGReg a0, TCGReg a1, tcg_target_long a2)
1913{
1914    tcg_out_dat_imm(s, COND_AL, ARITH_ORR, a0, a1, encode_imm_nofail(a2));
1915}
1916
1917static const TCGOutOpBinary outop_or = {
1918    .base.static_constraint = C_O1_I2(r, r, rI),
1919    .out_rrr = tgen_or,
1920    .out_rri = tgen_ori,
1921};
1922
1923static const TCGOutOpBinary outop_orc = {
1924    .base.static_constraint = C_NotImplemented,
1925};
1926
1927static void tgen_sub(TCGContext *s, TCGType type,
1928                     TCGReg a0, TCGReg a1, TCGReg a2)
1929{
1930    tcg_out_dat_reg(s, COND_AL, ARITH_SUB, a0, a1, a2, SHIFT_IMM_LSL(0));
1931}
1932
1933static void tgen_subfi(TCGContext *s, TCGType type,
1934                       TCGReg a0, tcg_target_long a1, TCGReg a2)
1935{
1936    tcg_out_dat_imm(s, COND_AL, ARITH_RSB, a0, a2, encode_imm_nofail(a1));
1937}
1938
1939static const TCGOutOpSubtract outop_sub = {
1940    .base.static_constraint = C_O1_I2(r, rI, r),
1941    .out_rrr = tgen_sub,
1942    .out_rir = tgen_subfi,
1943};
1944
1945static void tgen_xor(TCGContext *s, TCGType type,
1946                     TCGReg a0, TCGReg a1, TCGReg a2)
1947{
1948    tcg_out_dat_reg(s, COND_AL, ARITH_EOR, a0, a1, a2, SHIFT_IMM_LSL(0));
1949}
1950
1951static void tgen_xori(TCGContext *s, TCGType type,
1952                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1953{
1954    tcg_out_dat_imm(s, COND_AL, ARITH_EOR, a0, a1, encode_imm_nofail(a2));
1955}
1956
1957static const TCGOutOpBinary outop_xor = {
1958    .base.static_constraint = C_O1_I2(r, r, rI),
1959    .out_rrr = tgen_xor,
1960    .out_rri = tgen_xori,
1961};
1962
1963static void tgen_neg(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
1964{
1965    tgen_subfi(s, type, a0, 0, a1);
1966}
1967
1968static const TCGOutOpUnary outop_neg = {
1969    .base.static_constraint = C_O1_I1(r, r),
1970    .out_rr = tgen_neg,
1971};
1972
1973static void tgen_not(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
1974{
1975    tcg_out_dat_reg(s, COND_AL, ARITH_MVN, a0, 0, a1, SHIFT_IMM_LSL(0));
1976}
1977
1978static const TCGOutOpUnary outop_not = {
1979    .base.static_constraint = C_O1_I1(r, r),
1980    .out_rr = tgen_not,
1981};
1982
1983
1984static void tcg_out_op(TCGContext *s, TCGOpcode opc, TCGType type,
1985                       const TCGArg args[TCG_MAX_OP_ARGS],
1986                       const int const_args[TCG_MAX_OP_ARGS])
1987{
1988    TCGArg a0, a1, a2, a3, a4, a5;
1989    int c;
1990
1991    switch (opc) {
1992    case INDEX_op_goto_ptr:
1993        tcg_out_b_reg(s, COND_AL, args[0]);
1994        break;
1995    case INDEX_op_br:
1996        tcg_out_goto_label(s, COND_AL, arg_label(args[0]));
1997        break;
1998
1999    case INDEX_op_ld8u_i32:
2000        tcg_out_ld8u(s, COND_AL, args[0], args[1], args[2]);
2001        break;
2002    case INDEX_op_ld8s_i32:
2003        tcg_out_ld8s(s, COND_AL, args[0], args[1], args[2]);
2004        break;
2005    case INDEX_op_ld16u_i32:
2006        tcg_out_ld16u(s, COND_AL, args[0], args[1], args[2]);
2007        break;
2008    case INDEX_op_ld16s_i32:
2009        tcg_out_ld16s(s, COND_AL, args[0], args[1], args[2]);
2010        break;
2011    case INDEX_op_ld_i32:
2012        tcg_out_ld32u(s, COND_AL, args[0], args[1], args[2]);
2013        break;
2014    case INDEX_op_st8_i32:
2015        tcg_out_st8(s, COND_AL, args[0], args[1], args[2]);
2016        break;
2017    case INDEX_op_st16_i32:
2018        tcg_out_st16(s, COND_AL, args[0], args[1], args[2]);
2019        break;
2020    case INDEX_op_st_i32:
2021        tcg_out_st32(s, COND_AL, args[0], args[1], args[2]);
2022        break;
2023
2024    case INDEX_op_movcond_i32:
2025        /* Constraints mean that v2 is always in the same register as dest,
2026         * so we only need to do "if condition passed, move v1 to dest".
2027         */
2028        c = tcg_out_cmp(s, args[5], args[1], args[2], const_args[2]);
2029        tcg_out_dat_rIK(s, tcg_cond_to_arm_cond[c], ARITH_MOV,
2030                        ARITH_MVN, args[0], 0, args[3], const_args[3]);
2031        break;
2032    case INDEX_op_add2_i32:
2033        a0 = args[0], a1 = args[1], a2 = args[2];
2034        a3 = args[3], a4 = args[4], a5 = args[5];
2035        if (a0 == a3 || (a0 == a5 && !const_args[5])) {
2036            a0 = TCG_REG_TMP;
2037        }
2038        tcg_out_dat_rIN(s, COND_AL, ARITH_ADD | TO_CPSR, ARITH_SUB | TO_CPSR,
2039                        a0, a2, a4, const_args[4]);
2040        tcg_out_dat_rIK(s, COND_AL, ARITH_ADC, ARITH_SBC,
2041                        a1, a3, a5, const_args[5]);
2042        tcg_out_mov_reg(s, COND_AL, args[0], a0);
2043        break;
2044    case INDEX_op_sub2_i32:
2045        a0 = args[0], a1 = args[1], a2 = args[2];
2046        a3 = args[3], a4 = args[4], a5 = args[5];
2047        if ((a0 == a3 && !const_args[3]) || (a0 == a5 && !const_args[5])) {
2048            a0 = TCG_REG_TMP;
2049        }
2050        if (const_args[2]) {
2051            if (const_args[4]) {
2052                tcg_out_movi32(s, COND_AL, a0, a4);
2053                a4 = a0;
2054            }
2055            tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR, a0, a4, a2, 1);
2056        } else {
2057            tcg_out_dat_rIN(s, COND_AL, ARITH_SUB | TO_CPSR,
2058                            ARITH_ADD | TO_CPSR, a0, a2, a4, const_args[4]);
2059        }
2060        if (const_args[3]) {
2061            if (const_args[5]) {
2062                tcg_out_movi32(s, COND_AL, a1, a5);
2063                a5 = a1;
2064            }
2065            tcg_out_dat_rI(s, COND_AL, ARITH_RSC, a1, a5, a3, 1);
2066        } else {
2067            tcg_out_dat_rIK(s, COND_AL, ARITH_SBC, ARITH_ADC,
2068                            a1, a3, a5, const_args[5]);
2069        }
2070        tcg_out_mov_reg(s, COND_AL, args[0], a0);
2071        break;
2072    case INDEX_op_mulu2_i32:
2073        tcg_out_umull32(s, COND_AL, args[0], args[1], args[2], args[3]);
2074        break;
2075    case INDEX_op_muls2_i32:
2076        tcg_out_smull32(s, COND_AL, args[0], args[1], args[2], args[3]);
2077        break;
2078    /* XXX: Perhaps args[2] & 0x1f is wrong */
2079    case INDEX_op_shl_i32:
2080        c = const_args[2] ?
2081                SHIFT_IMM_LSL(args[2] & 0x1f) : SHIFT_REG_LSL(args[2]);
2082        goto gen_shift32;
2083    case INDEX_op_shr_i32:
2084        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_LSR(args[2] & 0x1f) :
2085                SHIFT_IMM_LSL(0) : SHIFT_REG_LSR(args[2]);
2086        goto gen_shift32;
2087    case INDEX_op_sar_i32:
2088        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_ASR(args[2] & 0x1f) :
2089                SHIFT_IMM_LSL(0) : SHIFT_REG_ASR(args[2]);
2090        goto gen_shift32;
2091    case INDEX_op_rotr_i32:
2092        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_ROR(args[2] & 0x1f) :
2093                SHIFT_IMM_LSL(0) : SHIFT_REG_ROR(args[2]);
2094        /* Fall through.  */
2095    gen_shift32:
2096        tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1], c);
2097        break;
2098
2099    case INDEX_op_rotl_i32:
2100        if (const_args[2]) {
2101            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1],
2102                            ((0x20 - args[2]) & 0x1f) ?
2103                            SHIFT_IMM_ROR((0x20 - args[2]) & 0x1f) :
2104                            SHIFT_IMM_LSL(0));
2105        } else {
2106            tcg_out_dat_imm(s, COND_AL, ARITH_RSB, TCG_REG_TMP, args[2], 0x20);
2107            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1],
2108                            SHIFT_REG_ROR(TCG_REG_TMP));
2109        }
2110        break;
2111
2112    case INDEX_op_ctz_i32:
2113        tcg_out_dat_reg(s, COND_AL, INSN_RBIT, TCG_REG_TMP, 0, args[1], 0);
2114        a1 = TCG_REG_TMP;
2115        goto do_clz;
2116
2117    case INDEX_op_clz_i32:
2118        a1 = args[1];
2119    do_clz:
2120        a0 = args[0];
2121        a2 = args[2];
2122        c = const_args[2];
2123        if (c && a2 == 32) {
2124            tcg_out_dat_reg(s, COND_AL, INSN_CLZ, a0, 0, a1, 0);
2125            break;
2126        }
2127        tcg_out_dat_imm(s, COND_AL, ARITH_CMP, 0, a1, 0);
2128        tcg_out_dat_reg(s, COND_NE, INSN_CLZ, a0, 0, a1, 0);
2129        if (c || a0 != a2) {
2130            tcg_out_dat_rIK(s, COND_EQ, ARITH_MOV, ARITH_MVN, a0, 0, a2, c);
2131        }
2132        break;
2133
2134    case INDEX_op_brcond_i32:
2135        c = tcg_out_cmp(s, args[2], args[0], args[1], const_args[1]);
2136        tcg_out_goto_label(s, tcg_cond_to_arm_cond[c], arg_label(args[3]));
2137        break;
2138    case INDEX_op_setcond_i32:
2139        c = tcg_out_cmp(s, args[3], args[1], args[2], const_args[2]);
2140        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c],
2141                        ARITH_MOV, args[0], 0, 1);
2142        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
2143                        ARITH_MOV, args[0], 0, 0);
2144        break;
2145    case INDEX_op_negsetcond_i32:
2146        c = tcg_out_cmp(s, args[3], args[1], args[2], const_args[2]);
2147        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c],
2148                        ARITH_MVN, args[0], 0, 0);
2149        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
2150                        ARITH_MOV, args[0], 0, 0);
2151        break;
2152
2153    case INDEX_op_brcond2_i32:
2154        c = tcg_out_cmp2(s, args, const_args);
2155        tcg_out_goto_label(s, tcg_cond_to_arm_cond[c], arg_label(args[5]));
2156        break;
2157    case INDEX_op_setcond2_i32:
2158        c = tcg_out_cmp2(s, args + 1, const_args + 1);
2159        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c], ARITH_MOV, args[0], 0, 1);
2160        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
2161                        ARITH_MOV, args[0], 0, 0);
2162        break;
2163
2164    case INDEX_op_qemu_ld_i32:
2165        tcg_out_qemu_ld(s, args[0], -1, args[1], args[2], TCG_TYPE_I32);
2166        break;
2167    case INDEX_op_qemu_ld_i64:
2168        tcg_out_qemu_ld(s, args[0], args[1], args[2], args[3], TCG_TYPE_I64);
2169        break;
2170
2171    case INDEX_op_qemu_st_i32:
2172        tcg_out_qemu_st(s, args[0], -1, args[1], args[2], TCG_TYPE_I32);
2173        break;
2174    case INDEX_op_qemu_st_i64:
2175        tcg_out_qemu_st(s, args[0], args[1], args[2], args[3], TCG_TYPE_I64);
2176        break;
2177
2178    case INDEX_op_bswap16_i32:
2179        tcg_out_bswap16(s, COND_AL, args[0], args[1], args[2]);
2180        break;
2181    case INDEX_op_bswap32_i32:
2182        tcg_out_bswap32(s, COND_AL, args[0], args[1]);
2183        break;
2184
2185    case INDEX_op_deposit_i32:
2186        tcg_out_deposit(s, COND_AL, args[0], args[2],
2187                        args[3], args[4], const_args[2]);
2188        break;
2189    case INDEX_op_extract_i32:
2190        tcg_out_extract(s, COND_AL, args[0], args[1], args[2], args[3]);
2191        break;
2192    case INDEX_op_sextract_i32:
2193        tcg_out_sextract(s, COND_AL, args[0], args[1], args[2], args[3]);
2194        break;
2195    case INDEX_op_extract2_i32:
2196        /* ??? These optimization vs zero should be generic.  */
2197        /* ??? But we can't substitute 2 for 1 in the opcode stream yet.  */
2198        if (const_args[1]) {
2199            if (const_args[2]) {
2200                tcg_out_movi(s, TCG_TYPE_REG, args[0], 0);
2201            } else {
2202                tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2203                                args[2], SHIFT_IMM_LSL(32 - args[3]));
2204            }
2205        } else if (const_args[2]) {
2206            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2207                            args[1], SHIFT_IMM_LSR(args[3]));
2208        } else {
2209            /* We can do extract2 in 2 insns, vs the 3 required otherwise.  */
2210            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0,
2211                            args[2], SHIFT_IMM_LSL(32 - args[3]));
2212            tcg_out_dat_reg(s, COND_AL, ARITH_ORR, args[0], TCG_REG_TMP,
2213                            args[1], SHIFT_IMM_LSR(args[3]));
2214        }
2215        break;
2216
2217    case INDEX_op_div_i32:
2218        tcg_out_sdiv(s, COND_AL, args[0], args[1], args[2]);
2219        break;
2220    case INDEX_op_divu_i32:
2221        tcg_out_udiv(s, COND_AL, args[0], args[1], args[2]);
2222        break;
2223
2224    case INDEX_op_mb:
2225        tcg_out_mb(s, args[0]);
2226        break;
2227
2228    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
2229    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
2230    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
2231    default:
2232        g_assert_not_reached();
2233    }
2234}
2235
2236static TCGConstraintSetIndex
2237tcg_target_op_def(TCGOpcode op, TCGType type, unsigned flags)
2238{
2239    switch (op) {
2240    case INDEX_op_goto_ptr:
2241        return C_O0_I1(r);
2242
2243    case INDEX_op_ld8u_i32:
2244    case INDEX_op_ld8s_i32:
2245    case INDEX_op_ld16u_i32:
2246    case INDEX_op_ld16s_i32:
2247    case INDEX_op_ld_i32:
2248    case INDEX_op_bswap16_i32:
2249    case INDEX_op_bswap32_i32:
2250    case INDEX_op_extract_i32:
2251    case INDEX_op_sextract_i32:
2252        return C_O1_I1(r, r);
2253
2254    case INDEX_op_st8_i32:
2255    case INDEX_op_st16_i32:
2256    case INDEX_op_st_i32:
2257        return C_O0_I2(r, r);
2258
2259    case INDEX_op_setcond_i32:
2260    case INDEX_op_negsetcond_i32:
2261        return C_O1_I2(r, r, rIN);
2262
2263    case INDEX_op_clz_i32:
2264    case INDEX_op_ctz_i32:
2265        return C_O1_I2(r, r, rIK);
2266
2267    case INDEX_op_div_i32:
2268    case INDEX_op_divu_i32:
2269        return C_O1_I2(r, r, r);
2270
2271    case INDEX_op_mulu2_i32:
2272    case INDEX_op_muls2_i32:
2273        return C_O2_I2(r, r, r, r);
2274
2275    case INDEX_op_shl_i32:
2276    case INDEX_op_shr_i32:
2277    case INDEX_op_sar_i32:
2278    case INDEX_op_rotl_i32:
2279    case INDEX_op_rotr_i32:
2280        return C_O1_I2(r, r, ri);
2281
2282    case INDEX_op_brcond_i32:
2283        return C_O0_I2(r, rIN);
2284    case INDEX_op_deposit_i32:
2285        return C_O1_I2(r, 0, rZ);
2286    case INDEX_op_extract2_i32:
2287        return C_O1_I2(r, rZ, rZ);
2288    case INDEX_op_movcond_i32:
2289        return C_O1_I4(r, r, rIN, rIK, 0);
2290    case INDEX_op_add2_i32:
2291        return C_O2_I4(r, r, r, r, rIN, rIK);
2292    case INDEX_op_sub2_i32:
2293        return C_O2_I4(r, r, rI, rI, rIN, rIK);
2294    case INDEX_op_brcond2_i32:
2295        return C_O0_I4(r, r, rI, rI);
2296    case INDEX_op_setcond2_i32:
2297        return C_O1_I4(r, r, r, rI, rI);
2298
2299    case INDEX_op_qemu_ld_i32:
2300        return C_O1_I1(r, q);
2301    case INDEX_op_qemu_ld_i64:
2302        return C_O2_I1(e, p, q);
2303    case INDEX_op_qemu_st_i32:
2304        return C_O0_I2(q, q);
2305    case INDEX_op_qemu_st_i64:
2306        return C_O0_I3(Q, p, q);
2307
2308    case INDEX_op_st_vec:
2309        return C_O0_I2(w, r);
2310    case INDEX_op_ld_vec:
2311    case INDEX_op_dupm_vec:
2312        return C_O1_I1(w, r);
2313    case INDEX_op_dup_vec:
2314        return C_O1_I1(w, wr);
2315    case INDEX_op_abs_vec:
2316    case INDEX_op_neg_vec:
2317    case INDEX_op_not_vec:
2318    case INDEX_op_shli_vec:
2319    case INDEX_op_shri_vec:
2320    case INDEX_op_sari_vec:
2321        return C_O1_I1(w, w);
2322    case INDEX_op_dup2_vec:
2323    case INDEX_op_add_vec:
2324    case INDEX_op_mul_vec:
2325    case INDEX_op_smax_vec:
2326    case INDEX_op_smin_vec:
2327    case INDEX_op_ssadd_vec:
2328    case INDEX_op_sssub_vec:
2329    case INDEX_op_sub_vec:
2330    case INDEX_op_umax_vec:
2331    case INDEX_op_umin_vec:
2332    case INDEX_op_usadd_vec:
2333    case INDEX_op_ussub_vec:
2334    case INDEX_op_xor_vec:
2335    case INDEX_op_arm_sshl_vec:
2336    case INDEX_op_arm_ushl_vec:
2337        return C_O1_I2(w, w, w);
2338    case INDEX_op_arm_sli_vec:
2339        return C_O1_I2(w, 0, w);
2340    case INDEX_op_or_vec:
2341    case INDEX_op_andc_vec:
2342        return C_O1_I2(w, w, wO);
2343    case INDEX_op_and_vec:
2344    case INDEX_op_orc_vec:
2345        return C_O1_I2(w, w, wV);
2346    case INDEX_op_cmp_vec:
2347        return C_O1_I2(w, w, wZ);
2348    case INDEX_op_bitsel_vec:
2349        return C_O1_I3(w, w, w, w);
2350    default:
2351        return C_NotImplemented;
2352    }
2353}
2354
2355static void tcg_target_init(TCGContext *s)
2356{
2357    /*
2358     * Only probe for the platform and capabilities if we haven't already
2359     * determined maximum values at compile time.
2360     */
2361#if !defined(use_idiv_instructions) || !defined(use_neon_instructions)
2362    {
2363        unsigned long hwcap = qemu_getauxval(AT_HWCAP);
2364#ifndef use_idiv_instructions
2365        use_idiv_instructions = (hwcap & HWCAP_ARM_IDIVA) != 0;
2366#endif
2367#ifndef use_neon_instructions
2368        use_neon_instructions = (hwcap & HWCAP_ARM_NEON) != 0;
2369#endif
2370    }
2371#endif
2372
2373    if (__ARM_ARCH < 7) {
2374        const char *pl = (const char *)qemu_getauxval(AT_PLATFORM);
2375        if (pl != NULL && pl[0] == 'v' && pl[1] >= '4' && pl[1] <= '9') {
2376            arm_arch = pl[1] - '0';
2377        }
2378
2379        if (arm_arch < 6) {
2380            error_report("TCG: ARMv%d is unsupported; exiting", arm_arch);
2381            exit(EXIT_FAILURE);
2382        }
2383    }
2384
2385    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
2386
2387    tcg_target_call_clobber_regs = 0;
2388    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R0);
2389    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R1);
2390    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R2);
2391    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R3);
2392    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R12);
2393    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R14);
2394
2395    if (use_neon_instructions) {
2396        tcg_target_available_regs[TCG_TYPE_V64]  = ALL_VECTOR_REGS;
2397        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
2398
2399        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q0);
2400        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q1);
2401        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q2);
2402        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q3);
2403        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q8);
2404        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q9);
2405        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q10);
2406        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q11);
2407        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q12);
2408        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q13);
2409        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q14);
2410        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q15);
2411    }
2412
2413    s->reserved_regs = 0;
2414    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
2415    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP);
2416    tcg_regset_set_reg(s->reserved_regs, TCG_REG_PC);
2417    tcg_regset_set_reg(s->reserved_regs, TCG_VEC_TMP);
2418}
2419
2420static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg arg,
2421                       TCGReg arg1, intptr_t arg2)
2422{
2423    switch (type) {
2424    case TCG_TYPE_I32:
2425        tcg_out_ld32u(s, COND_AL, arg, arg1, arg2);
2426        return;
2427    case TCG_TYPE_V64:
2428        /* regs 1; size 8; align 8 */
2429        tcg_out_vldst(s, INSN_VLD1 | 0x7d0, arg, arg1, arg2);
2430        return;
2431    case TCG_TYPE_V128:
2432        /*
2433         * We have only 8-byte alignment for the stack per the ABI.
2434         * Rather than dynamically re-align the stack, it's easier
2435         * to simply not request alignment beyond that.  So:
2436         * regs 2; size 8; align 8
2437         */
2438        tcg_out_vldst(s, INSN_VLD1 | 0xad0, arg, arg1, arg2);
2439        return;
2440    default:
2441        g_assert_not_reached();
2442    }
2443}
2444
2445static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
2446                       TCGReg arg1, intptr_t arg2)
2447{
2448    switch (type) {
2449    case TCG_TYPE_I32:
2450        tcg_out_st32(s, COND_AL, arg, arg1, arg2);
2451        return;
2452    case TCG_TYPE_V64:
2453        /* regs 1; size 8; align 8 */
2454        tcg_out_vldst(s, INSN_VST1 | 0x7d0, arg, arg1, arg2);
2455        return;
2456    case TCG_TYPE_V128:
2457        /* See tcg_out_ld re alignment: regs 2; size 8; align 8 */
2458        tcg_out_vldst(s, INSN_VST1 | 0xad0, arg, arg1, arg2);
2459        return;
2460    default:
2461        g_assert_not_reached();
2462    }
2463}
2464
2465static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
2466                        TCGReg base, intptr_t ofs)
2467{
2468    return false;
2469}
2470
2471static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
2472{
2473    if (ret == arg) {
2474        return true;
2475    }
2476    switch (type) {
2477    case TCG_TYPE_I32:
2478        if (ret < TCG_REG_Q0 && arg < TCG_REG_Q0) {
2479            tcg_out_mov_reg(s, COND_AL, ret, arg);
2480            return true;
2481        }
2482        return false;
2483
2484    case TCG_TYPE_V64:
2485    case TCG_TYPE_V128:
2486        /* "VMOV D,N" is an alias for "VORR D,N,N". */
2487        tcg_out_vreg3(s, INSN_VORR, type - TCG_TYPE_V64, 0, ret, arg, arg);
2488        return true;
2489
2490    default:
2491        g_assert_not_reached();
2492    }
2493}
2494
2495static void tcg_out_movi(TCGContext *s, TCGType type,
2496                         TCGReg ret, tcg_target_long arg)
2497{
2498    tcg_debug_assert(type == TCG_TYPE_I32);
2499    tcg_debug_assert(ret < TCG_REG_Q0);
2500    tcg_out_movi32(s, COND_AL, ret, arg);
2501}
2502
2503static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
2504{
2505    return false;
2506}
2507
2508static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
2509                             tcg_target_long imm)
2510{
2511    int enc, opc = ARITH_ADD;
2512
2513    /* All of the easiest immediates to encode are positive. */
2514    if (imm < 0) {
2515        imm = -imm;
2516        opc = ARITH_SUB;
2517    }
2518    enc = encode_imm(imm);
2519    if (enc >= 0) {
2520        tcg_out_dat_imm(s, COND_AL, opc, rd, rs, enc);
2521    } else {
2522        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, imm);
2523        tcg_out_dat_reg(s, COND_AL, opc, rd, rs,
2524                        TCG_REG_TMP, SHIFT_IMM_LSL(0));
2525    }
2526}
2527
2528/* Type is always V128, with I64 elements.  */
2529static void tcg_out_dup2_vec(TCGContext *s, TCGReg rd, TCGReg rl, TCGReg rh)
2530{
2531    /* Move high element into place first. */
2532    /* VMOV Dd+1, Ds */
2533    tcg_out_vreg3(s, INSN_VORR | (1 << 12), 0, 0, rd, rh, rh);
2534    /* Move low element into place; tcg_out_mov will check for nop. */
2535    tcg_out_mov(s, TCG_TYPE_V64, rd, rl);
2536}
2537
2538static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
2539                            TCGReg rd, TCGReg rs)
2540{
2541    int q = type - TCG_TYPE_V64;
2542
2543    if (vece == MO_64) {
2544        if (type == TCG_TYPE_V128) {
2545            tcg_out_dup2_vec(s, rd, rs, rs);
2546        } else {
2547            tcg_out_mov(s, TCG_TYPE_V64, rd, rs);
2548        }
2549    } else if (rs < TCG_REG_Q0) {
2550        int b = (vece == MO_8);
2551        int e = (vece == MO_16);
2552        tcg_out32(s, INSN_VDUP_G | (b << 22) | (q << 21) | (e << 5) |
2553                  encode_vn(rd) | (rs << 12));
2554    } else {
2555        int imm4 = 1 << vece;
2556        tcg_out32(s, INSN_VDUP_S | (imm4 << 16) | (q << 6) |
2557                  encode_vd(rd) | encode_vm(rs));
2558    }
2559    return true;
2560}
2561
2562static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
2563                             TCGReg rd, TCGReg base, intptr_t offset)
2564{
2565    if (vece == MO_64) {
2566        tcg_out_ld(s, TCG_TYPE_V64, rd, base, offset);
2567        if (type == TCG_TYPE_V128) {
2568            tcg_out_dup2_vec(s, rd, rd, rd);
2569        }
2570    } else {
2571        int q = type - TCG_TYPE_V64;
2572        tcg_out_vldst(s, INSN_VLD1R | (vece << 6) | (q << 5),
2573                      rd, base, offset);
2574    }
2575    return true;
2576}
2577
2578static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
2579                             TCGReg rd, int64_t v64)
2580{
2581    int q = type - TCG_TYPE_V64;
2582    int cmode, imm8, i;
2583
2584    /* Test all bytes equal first.  */
2585    if (vece == MO_8) {
2586        tcg_out_vmovi(s, rd, q, 0, 0xe, v64);
2587        return;
2588    }
2589
2590    /*
2591     * Test all bytes 0x00 or 0xff second.  This can match cases that
2592     * might otherwise take 2 or 3 insns for MO_16 or MO_32 below.
2593     */
2594    for (i = imm8 = 0; i < 8; i++) {
2595        uint8_t byte = v64 >> (i * 8);
2596        if (byte == 0xff) {
2597            imm8 |= 1 << i;
2598        } else if (byte != 0) {
2599            goto fail_bytes;
2600        }
2601    }
2602    tcg_out_vmovi(s, rd, q, 1, 0xe, imm8);
2603    return;
2604 fail_bytes:
2605
2606    /*
2607     * Tests for various replications.  For each element width, if we
2608     * cannot find an expansion there's no point checking a larger
2609     * width because we already know by replication it cannot match.
2610     */
2611    if (vece == MO_16) {
2612        uint16_t v16 = v64;
2613
2614        if (is_shimm16(v16, &cmode, &imm8)) {
2615            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2616            return;
2617        }
2618        if (is_shimm16(~v16, &cmode, &imm8)) {
2619            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2620            return;
2621        }
2622
2623        /*
2624         * Otherwise, all remaining constants can be loaded in two insns:
2625         * rd = v16 & 0xff, rd |= v16 & 0xff00.
2626         */
2627        tcg_out_vmovi(s, rd, q, 0, 0x8, v16 & 0xff);
2628        tcg_out_vmovi(s, rd, q, 0, 0xb, v16 >> 8);   /* VORRI */
2629        return;
2630    }
2631
2632    if (vece == MO_32) {
2633        uint32_t v32 = v64;
2634
2635        if (is_shimm32(v32, &cmode, &imm8) ||
2636            is_soimm32(v32, &cmode, &imm8)) {
2637            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2638            return;
2639        }
2640        if (is_shimm32(~v32, &cmode, &imm8) ||
2641            is_soimm32(~v32, &cmode, &imm8)) {
2642            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2643            return;
2644        }
2645
2646        /*
2647         * Restrict the set of constants to those we can load with
2648         * two instructions.  Others we load from the pool.
2649         */
2650        i = is_shimm32_pair(v32, &cmode, &imm8);
2651        if (i) {
2652            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2653            tcg_out_vmovi(s, rd, q, 0, i | 1, extract32(v32, i * 4, 8));
2654            return;
2655        }
2656        i = is_shimm32_pair(~v32, &cmode, &imm8);
2657        if (i) {
2658            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2659            tcg_out_vmovi(s, rd, q, 1, i | 1, extract32(~v32, i * 4, 8));
2660            return;
2661        }
2662    }
2663
2664    /*
2665     * As a last resort, load from the constant pool.
2666     */
2667    if (!q || vece == MO_64) {
2668        new_pool_l2(s, R_ARM_PC11, s->code_ptr, 0, v64, v64 >> 32);
2669        /* VLDR Dd, [pc + offset] */
2670        tcg_out32(s, INSN_VLDR_D | encode_vd(rd) | (0xf << 16));
2671        if (q) {
2672            tcg_out_dup2_vec(s, rd, rd, rd);
2673        }
2674    } else {
2675        new_pool_label(s, (uint32_t)v64, R_ARM_PC8, s->code_ptr, 0);
2676        /* add tmp, pc, offset */
2677        tcg_out_dat_imm(s, COND_AL, ARITH_ADD, TCG_REG_TMP, TCG_REG_PC, 0);
2678        tcg_out_dupm_vec(s, type, MO_32, rd, TCG_REG_TMP, 0);
2679    }
2680}
2681
2682static const ARMInsn vec_cmp_insn[16] = {
2683    [TCG_COND_EQ] = INSN_VCEQ,
2684    [TCG_COND_GT] = INSN_VCGT,
2685    [TCG_COND_GE] = INSN_VCGE,
2686    [TCG_COND_GTU] = INSN_VCGT_U,
2687    [TCG_COND_GEU] = INSN_VCGE_U,
2688};
2689
2690static const ARMInsn vec_cmp0_insn[16] = {
2691    [TCG_COND_EQ] = INSN_VCEQ0,
2692    [TCG_COND_GT] = INSN_VCGT0,
2693    [TCG_COND_GE] = INSN_VCGE0,
2694    [TCG_COND_LT] = INSN_VCLT0,
2695    [TCG_COND_LE] = INSN_VCLE0,
2696};
2697
2698static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
2699                           unsigned vecl, unsigned vece,
2700                           const TCGArg args[TCG_MAX_OP_ARGS],
2701                           const int const_args[TCG_MAX_OP_ARGS])
2702{
2703    TCGType type = vecl + TCG_TYPE_V64;
2704    unsigned q = vecl;
2705    TCGArg a0, a1, a2, a3;
2706    int cmode, imm8;
2707
2708    a0 = args[0];
2709    a1 = args[1];
2710    a2 = args[2];
2711
2712    switch (opc) {
2713    case INDEX_op_ld_vec:
2714        tcg_out_ld(s, type, a0, a1, a2);
2715        return;
2716    case INDEX_op_st_vec:
2717        tcg_out_st(s, type, a0, a1, a2);
2718        return;
2719    case INDEX_op_dupm_vec:
2720        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
2721        return;
2722    case INDEX_op_dup2_vec:
2723        tcg_out_dup2_vec(s, a0, a1, a2);
2724        return;
2725    case INDEX_op_abs_vec:
2726        tcg_out_vreg2(s, INSN_VABS, q, vece, a0, a1);
2727        return;
2728    case INDEX_op_neg_vec:
2729        tcg_out_vreg2(s, INSN_VNEG, q, vece, a0, a1);
2730        return;
2731    case INDEX_op_not_vec:
2732        tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a1);
2733        return;
2734    case INDEX_op_add_vec:
2735        tcg_out_vreg3(s, INSN_VADD, q, vece, a0, a1, a2);
2736        return;
2737    case INDEX_op_mul_vec:
2738        tcg_out_vreg3(s, INSN_VMUL, q, vece, a0, a1, a2);
2739        return;
2740    case INDEX_op_smax_vec:
2741        tcg_out_vreg3(s, INSN_VMAX, q, vece, a0, a1, a2);
2742        return;
2743    case INDEX_op_smin_vec:
2744        tcg_out_vreg3(s, INSN_VMIN, q, vece, a0, a1, a2);
2745        return;
2746    case INDEX_op_sub_vec:
2747        tcg_out_vreg3(s, INSN_VSUB, q, vece, a0, a1, a2);
2748        return;
2749    case INDEX_op_ssadd_vec:
2750        tcg_out_vreg3(s, INSN_VQADD, q, vece, a0, a1, a2);
2751        return;
2752    case INDEX_op_sssub_vec:
2753        tcg_out_vreg3(s, INSN_VQSUB, q, vece, a0, a1, a2);
2754        return;
2755    case INDEX_op_umax_vec:
2756        tcg_out_vreg3(s, INSN_VMAX_U, q, vece, a0, a1, a2);
2757        return;
2758    case INDEX_op_umin_vec:
2759        tcg_out_vreg3(s, INSN_VMIN_U, q, vece, a0, a1, a2);
2760        return;
2761    case INDEX_op_usadd_vec:
2762        tcg_out_vreg3(s, INSN_VQADD_U, q, vece, a0, a1, a2);
2763        return;
2764    case INDEX_op_ussub_vec:
2765        tcg_out_vreg3(s, INSN_VQSUB_U, q, vece, a0, a1, a2);
2766        return;
2767    case INDEX_op_xor_vec:
2768        tcg_out_vreg3(s, INSN_VEOR, q, 0, a0, a1, a2);
2769        return;
2770    case INDEX_op_arm_sshl_vec:
2771        /*
2772         * Note that Vm is the data and Vn is the shift count,
2773         * therefore the arguments appear reversed.
2774         */
2775        tcg_out_vreg3(s, INSN_VSHL_S, q, vece, a0, a2, a1);
2776        return;
2777    case INDEX_op_arm_ushl_vec:
2778        /* See above. */
2779        tcg_out_vreg3(s, INSN_VSHL_U, q, vece, a0, a2, a1);
2780        return;
2781    case INDEX_op_shli_vec:
2782        tcg_out_vshifti(s, INSN_VSHLI, q, a0, a1, a2 + (8 << vece));
2783        return;
2784    case INDEX_op_shri_vec:
2785        tcg_out_vshifti(s, INSN_VSHRI, q, a0, a1, (16 << vece) - a2);
2786        return;
2787    case INDEX_op_sari_vec:
2788        tcg_out_vshifti(s, INSN_VSARI, q, a0, a1, (16 << vece) - a2);
2789        return;
2790    case INDEX_op_arm_sli_vec:
2791        tcg_out_vshifti(s, INSN_VSLI, q, a0, a2, args[3] + (8 << vece));
2792        return;
2793
2794    case INDEX_op_andc_vec:
2795        if (!const_args[2]) {
2796            tcg_out_vreg3(s, INSN_VBIC, q, 0, a0, a1, a2);
2797            return;
2798        }
2799        a2 = ~a2;
2800        /* fall through */
2801    case INDEX_op_and_vec:
2802        if (const_args[2]) {
2803            is_shimm1632(~a2, &cmode, &imm8);
2804            if (a0 == a1) {
2805                tcg_out_vmovi(s, a0, q, 1, cmode | 1, imm8); /* VBICI */
2806                return;
2807            }
2808            tcg_out_vmovi(s, a0, q, 1, cmode, imm8); /* VMVNI */
2809            a2 = a0;
2810        }
2811        tcg_out_vreg3(s, INSN_VAND, q, 0, a0, a1, a2);
2812        return;
2813
2814    case INDEX_op_orc_vec:
2815        if (!const_args[2]) {
2816            tcg_out_vreg3(s, INSN_VORN, q, 0, a0, a1, a2);
2817            return;
2818        }
2819        a2 = ~a2;
2820        /* fall through */
2821    case INDEX_op_or_vec:
2822        if (const_args[2]) {
2823            is_shimm1632(a2, &cmode, &imm8);
2824            if (a0 == a1) {
2825                tcg_out_vmovi(s, a0, q, 0, cmode | 1, imm8); /* VORRI */
2826                return;
2827            }
2828            tcg_out_vmovi(s, a0, q, 0, cmode, imm8); /* VMOVI */
2829            a2 = a0;
2830        }
2831        tcg_out_vreg3(s, INSN_VORR, q, 0, a0, a1, a2);
2832        return;
2833
2834    case INDEX_op_cmp_vec:
2835        {
2836            TCGCond cond = args[3];
2837            ARMInsn insn;
2838
2839            switch (cond) {
2840            case TCG_COND_NE:
2841                if (const_args[2]) {
2842                    tcg_out_vreg3(s, INSN_VTST, q, vece, a0, a1, a1);
2843                } else {
2844                    tcg_out_vreg3(s, INSN_VCEQ, q, vece, a0, a1, a2);
2845                    tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a0);
2846                }
2847                break;
2848
2849            case TCG_COND_TSTNE:
2850            case TCG_COND_TSTEQ:
2851                if (const_args[2]) {
2852                    /* (x & 0) == 0 */
2853                    tcg_out_dupi_vec(s, type, MO_8, a0,
2854                                     -(cond == TCG_COND_TSTEQ));
2855                    break;
2856                }
2857                tcg_out_vreg3(s, INSN_VTST, q, vece, a0, a1, a2);
2858                if (cond == TCG_COND_TSTEQ) {
2859                    tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a0);
2860                }
2861                break;
2862
2863            default:
2864                if (const_args[2]) {
2865                    insn = vec_cmp0_insn[cond];
2866                    if (insn) {
2867                        tcg_out_vreg2(s, insn, q, vece, a0, a1);
2868                        return;
2869                    }
2870                    tcg_out_dupi_vec(s, type, MO_8, TCG_VEC_TMP, 0);
2871                    a2 = TCG_VEC_TMP;
2872                }
2873                insn = vec_cmp_insn[cond];
2874                if (insn == 0) {
2875                    TCGArg t;
2876                    t = a1, a1 = a2, a2 = t;
2877                    cond = tcg_swap_cond(cond);
2878                    insn = vec_cmp_insn[cond];
2879                    tcg_debug_assert(insn != 0);
2880                }
2881                tcg_out_vreg3(s, insn, q, vece, a0, a1, a2);
2882                break;
2883            }
2884        }
2885        return;
2886
2887    case INDEX_op_bitsel_vec:
2888        a3 = args[3];
2889        if (a0 == a3) {
2890            tcg_out_vreg3(s, INSN_VBIT, q, 0, a0, a2, a1);
2891        } else if (a0 == a2) {
2892            tcg_out_vreg3(s, INSN_VBIF, q, 0, a0, a3, a1);
2893        } else {
2894            tcg_out_mov(s, type, a0, a1);
2895            tcg_out_vreg3(s, INSN_VBSL, q, 0, a0, a2, a3);
2896        }
2897        return;
2898
2899    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
2900    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
2901    default:
2902        g_assert_not_reached();
2903    }
2904}
2905
2906int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
2907{
2908    switch (opc) {
2909    case INDEX_op_add_vec:
2910    case INDEX_op_sub_vec:
2911    case INDEX_op_and_vec:
2912    case INDEX_op_andc_vec:
2913    case INDEX_op_or_vec:
2914    case INDEX_op_orc_vec:
2915    case INDEX_op_xor_vec:
2916    case INDEX_op_not_vec:
2917    case INDEX_op_shli_vec:
2918    case INDEX_op_shri_vec:
2919    case INDEX_op_sari_vec:
2920    case INDEX_op_ssadd_vec:
2921    case INDEX_op_sssub_vec:
2922    case INDEX_op_usadd_vec:
2923    case INDEX_op_ussub_vec:
2924    case INDEX_op_bitsel_vec:
2925        return 1;
2926    case INDEX_op_abs_vec:
2927    case INDEX_op_cmp_vec:
2928    case INDEX_op_mul_vec:
2929    case INDEX_op_neg_vec:
2930    case INDEX_op_smax_vec:
2931    case INDEX_op_smin_vec:
2932    case INDEX_op_umax_vec:
2933    case INDEX_op_umin_vec:
2934        return vece < MO_64;
2935    case INDEX_op_shlv_vec:
2936    case INDEX_op_shrv_vec:
2937    case INDEX_op_sarv_vec:
2938    case INDEX_op_rotli_vec:
2939    case INDEX_op_rotlv_vec:
2940    case INDEX_op_rotrv_vec:
2941        return -1;
2942    default:
2943        return 0;
2944    }
2945}
2946
2947void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
2948                       TCGArg a0, ...)
2949{
2950    va_list va;
2951    TCGv_vec v0, v1, v2, t1, t2, c1;
2952    TCGArg a2;
2953
2954    va_start(va, a0);
2955    v0 = temp_tcgv_vec(arg_temp(a0));
2956    v1 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
2957    a2 = va_arg(va, TCGArg);
2958    va_end(va);
2959
2960    switch (opc) {
2961    case INDEX_op_shlv_vec:
2962        /*
2963         * Merely propagate shlv_vec to arm_ushl_vec.
2964         * In this way we don't set TCG_TARGET_HAS_shv_vec
2965         * because everything is done via expansion.
2966         */
2967        v2 = temp_tcgv_vec(arg_temp(a2));
2968        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
2969                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
2970        break;
2971
2972    case INDEX_op_shrv_vec:
2973    case INDEX_op_sarv_vec:
2974        /* Right shifts are negative left shifts for NEON.  */
2975        v2 = temp_tcgv_vec(arg_temp(a2));
2976        t1 = tcg_temp_new_vec(type);
2977        tcg_gen_neg_vec(vece, t1, v2);
2978        if (opc == INDEX_op_shrv_vec) {
2979            opc = INDEX_op_arm_ushl_vec;
2980        } else {
2981            opc = INDEX_op_arm_sshl_vec;
2982        }
2983        vec_gen_3(opc, type, vece, tcgv_vec_arg(v0),
2984                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
2985        tcg_temp_free_vec(t1);
2986        break;
2987
2988    case INDEX_op_rotli_vec:
2989        t1 = tcg_temp_new_vec(type);
2990        tcg_gen_shri_vec(vece, t1, v1, -a2 & ((8 << vece) - 1));
2991        vec_gen_4(INDEX_op_arm_sli_vec, type, vece,
2992                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(v1), a2);
2993        tcg_temp_free_vec(t1);
2994        break;
2995
2996    case INDEX_op_rotlv_vec:
2997        v2 = temp_tcgv_vec(arg_temp(a2));
2998        t1 = tcg_temp_new_vec(type);
2999        c1 = tcg_constant_vec(type, vece, 8 << vece);
3000        tcg_gen_sub_vec(vece, t1, v2, c1);
3001        /* Right shifts are negative left shifts for NEON.  */
3002        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
3003                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
3004        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
3005                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
3006        tcg_gen_or_vec(vece, v0, v0, t1);
3007        tcg_temp_free_vec(t1);
3008        break;
3009
3010    case INDEX_op_rotrv_vec:
3011        v2 = temp_tcgv_vec(arg_temp(a2));
3012        t1 = tcg_temp_new_vec(type);
3013        t2 = tcg_temp_new_vec(type);
3014        c1 = tcg_constant_vec(type, vece, 8 << vece);
3015        tcg_gen_neg_vec(vece, t1, v2);
3016        tcg_gen_sub_vec(vece, t2, c1, v2);
3017        /* Right shifts are negative left shifts for NEON.  */
3018        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
3019                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
3020        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t2),
3021                  tcgv_vec_arg(v1), tcgv_vec_arg(t2));
3022        tcg_gen_or_vec(vece, v0, t1, t2);
3023        tcg_temp_free_vec(t1);
3024        tcg_temp_free_vec(t2);
3025        break;
3026
3027    default:
3028        g_assert_not_reached();
3029    }
3030}
3031
3032static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
3033{
3034    int i;
3035    for (i = 0; i < count; ++i) {
3036        p[i] = INSN_NOP;
3037    }
3038}
3039
3040/* Compute frame size via macros, to share between tcg_target_qemu_prologue
3041   and tcg_register_jit.  */
3042
3043#define PUSH_SIZE  ((11 - 4 + 1 + 1) * sizeof(tcg_target_long))
3044
3045#define FRAME_SIZE \
3046    ((PUSH_SIZE \
3047      + TCG_STATIC_CALL_ARGS_SIZE \
3048      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
3049      + TCG_TARGET_STACK_ALIGN - 1) \
3050     & -TCG_TARGET_STACK_ALIGN)
3051
3052#define STACK_ADDEND  (FRAME_SIZE - PUSH_SIZE)
3053
3054static void tcg_target_qemu_prologue(TCGContext *s)
3055{
3056    /* Calling convention requires us to save r4-r11 and lr.  */
3057    /* stmdb sp!, { r4 - r11, lr } */
3058    tcg_out_ldstm(s, COND_AL, INSN_STMDB, TCG_REG_CALL_STACK,
3059                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
3060                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
3061                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_R14));
3062
3063    /* Reserve callee argument and tcg temp space.  */
3064    tcg_out_dat_rI(s, COND_AL, ARITH_SUB, TCG_REG_CALL_STACK,
3065                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
3066    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
3067                  CPU_TEMP_BUF_NLONGS * sizeof(long));
3068
3069    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
3070
3071    if (!tcg_use_softmmu && guest_base) {
3072        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_GUEST_BASE, guest_base);
3073        tcg_regset_set_reg(s->reserved_regs, TCG_REG_GUEST_BASE);
3074    }
3075
3076    tcg_out_b_reg(s, COND_AL, tcg_target_call_iarg_regs[1]);
3077
3078    /*
3079     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
3080     * and fall through to the rest of the epilogue.
3081     */
3082    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
3083    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, 0);
3084    tcg_out_epilogue(s);
3085}
3086
3087static void tcg_out_epilogue(TCGContext *s)
3088{
3089    /* Release local stack frame.  */
3090    tcg_out_dat_rI(s, COND_AL, ARITH_ADD, TCG_REG_CALL_STACK,
3091                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
3092
3093    /* ldmia sp!, { r4 - r11, pc } */
3094    tcg_out_ldstm(s, COND_AL, INSN_LDMIA, TCG_REG_CALL_STACK,
3095                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
3096                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
3097                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_PC));
3098}
3099
3100static void tcg_out_tb_start(TCGContext *s)
3101{
3102    /* nothing to do */
3103}
3104
3105typedef struct {
3106    DebugFrameHeader h;
3107    uint8_t fde_def_cfa[4];
3108    uint8_t fde_reg_ofs[18];
3109} DebugFrame;
3110
3111#define ELF_HOST_MACHINE EM_ARM
3112
3113/* We're expecting a 2 byte uleb128 encoded value.  */
3114QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
3115
3116static const DebugFrame debug_frame = {
3117    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
3118    .h.cie.id = -1,
3119    .h.cie.version = 1,
3120    .h.cie.code_align = 1,
3121    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
3122    .h.cie.return_column = 14,
3123
3124    /* Total FDE size does not include the "len" member.  */
3125    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
3126
3127    .fde_def_cfa = {
3128        12, 13,                         /* DW_CFA_def_cfa sp, ... */
3129        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
3130        (FRAME_SIZE >> 7)
3131    },
3132    .fde_reg_ofs = {
3133        /* The following must match the stmdb in the prologue.  */
3134        0x8e, 1,                        /* DW_CFA_offset, lr, -4 */
3135        0x8b, 2,                        /* DW_CFA_offset, r11, -8 */
3136        0x8a, 3,                        /* DW_CFA_offset, r10, -12 */
3137        0x89, 4,                        /* DW_CFA_offset, r9, -16 */
3138        0x88, 5,                        /* DW_CFA_offset, r8, -20 */
3139        0x87, 6,                        /* DW_CFA_offset, r7, -24 */
3140        0x86, 7,                        /* DW_CFA_offset, r6, -28 */
3141        0x85, 8,                        /* DW_CFA_offset, r5, -32 */
3142        0x84, 9,                        /* DW_CFA_offset, r4, -36 */
3143    }
3144};
3145
3146void tcg_register_jit(const void *buf, size_t buf_size)
3147{
3148    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
3149}
3150