xref: /openbmc/qemu/tcg/arm/tcg-target.c.inc (revision 238f43809a85a47cfbbc2e1d6aff4640fec30328)
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#include "../tcg-ldst.c.inc"
27#include "../tcg-pool.c.inc"
28
29int arm_arch = __ARM_ARCH;
30
31#ifndef use_idiv_instructions
32bool use_idiv_instructions;
33#endif
34#ifndef use_neon_instructions
35bool use_neon_instructions;
36#endif
37
38#ifdef CONFIG_DEBUG_TCG
39static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
40    "%r0",  "%r1",  "%r2",  "%r3",  "%r4",  "%r5",  "%r6",  "%r7",
41    "%r8",  "%r9",  "%r10", "%r11", "%r12", "%sp",  "%r14", "%pc",
42    "%q0",  "%q1",  "%q2",  "%q3",  "%q4",  "%q5",  "%q6",  "%q7",
43    "%q8",  "%q9",  "%q10", "%q11", "%q12", "%q13", "%q14", "%q15",
44};
45#endif
46
47static const int tcg_target_reg_alloc_order[] = {
48    TCG_REG_R4,
49    TCG_REG_R5,
50    TCG_REG_R6,
51    TCG_REG_R7,
52    TCG_REG_R8,
53    TCG_REG_R9,
54    TCG_REG_R10,
55    TCG_REG_R11,
56    TCG_REG_R13,
57    TCG_REG_R0,
58    TCG_REG_R1,
59    TCG_REG_R2,
60    TCG_REG_R3,
61    TCG_REG_R12,
62    TCG_REG_R14,
63
64    TCG_REG_Q0,
65    TCG_REG_Q1,
66    TCG_REG_Q2,
67    TCG_REG_Q3,
68    /* Q4 - Q7 are call-saved, and skipped. */
69    TCG_REG_Q8,
70    TCG_REG_Q9,
71    TCG_REG_Q10,
72    TCG_REG_Q11,
73    TCG_REG_Q12,
74    TCG_REG_Q13,
75    TCG_REG_Q14,
76    TCG_REG_Q15,
77};
78
79static const int tcg_target_call_iarg_regs[4] = {
80    TCG_REG_R0, TCG_REG_R1, TCG_REG_R2, TCG_REG_R3
81};
82
83static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
84{
85    tcg_debug_assert(kind == TCG_CALL_RET_NORMAL);
86    tcg_debug_assert(slot >= 0 && slot <= 3);
87    return TCG_REG_R0 + slot;
88}
89
90#define TCG_REG_TMP  TCG_REG_R12
91#define TCG_VEC_TMP  TCG_REG_Q15
92#ifndef CONFIG_SOFTMMU
93#define TCG_REG_GUEST_BASE  TCG_REG_R11
94#endif
95
96typedef enum {
97    COND_EQ = 0x0,
98    COND_NE = 0x1,
99    COND_CS = 0x2,	/* Unsigned greater or equal */
100    COND_CC = 0x3,	/* Unsigned less than */
101    COND_MI = 0x4,	/* Negative */
102    COND_PL = 0x5,	/* Zero or greater */
103    COND_VS = 0x6,	/* Overflow */
104    COND_VC = 0x7,	/* No overflow */
105    COND_HI = 0x8,	/* Unsigned greater than */
106    COND_LS = 0x9,	/* Unsigned less or equal */
107    COND_GE = 0xa,
108    COND_LT = 0xb,
109    COND_GT = 0xc,
110    COND_LE = 0xd,
111    COND_AL = 0xe,
112} ARMCond;
113
114#define TO_CPSR (1 << 20)
115
116#define SHIFT_IMM_LSL(im)	(((im) << 7) | 0x00)
117#define SHIFT_IMM_LSR(im)	(((im) << 7) | 0x20)
118#define SHIFT_IMM_ASR(im)	(((im) << 7) | 0x40)
119#define SHIFT_IMM_ROR(im)	(((im) << 7) | 0x60)
120#define SHIFT_REG_LSL(rs)	(((rs) << 8) | 0x10)
121#define SHIFT_REG_LSR(rs)	(((rs) << 8) | 0x30)
122#define SHIFT_REG_ASR(rs)	(((rs) << 8) | 0x50)
123#define SHIFT_REG_ROR(rs)	(((rs) << 8) | 0x70)
124
125typedef enum {
126    ARITH_AND = 0x0 << 21,
127    ARITH_EOR = 0x1 << 21,
128    ARITH_SUB = 0x2 << 21,
129    ARITH_RSB = 0x3 << 21,
130    ARITH_ADD = 0x4 << 21,
131    ARITH_ADC = 0x5 << 21,
132    ARITH_SBC = 0x6 << 21,
133    ARITH_RSC = 0x7 << 21,
134    ARITH_TST = 0x8 << 21 | TO_CPSR,
135    ARITH_CMP = 0xa << 21 | TO_CPSR,
136    ARITH_CMN = 0xb << 21 | TO_CPSR,
137    ARITH_ORR = 0xc << 21,
138    ARITH_MOV = 0xd << 21,
139    ARITH_BIC = 0xe << 21,
140    ARITH_MVN = 0xf << 21,
141
142    INSN_B         = 0x0a000000,
143
144    INSN_CLZ       = 0x016f0f10,
145    INSN_RBIT      = 0x06ff0f30,
146
147    INSN_LDMIA     = 0x08b00000,
148    INSN_STMDB     = 0x09200000,
149
150    INSN_LDR_IMM   = 0x04100000,
151    INSN_LDR_REG   = 0x06100000,
152    INSN_STR_IMM   = 0x04000000,
153    INSN_STR_REG   = 0x06000000,
154
155    INSN_LDRH_IMM  = 0x005000b0,
156    INSN_LDRH_REG  = 0x001000b0,
157    INSN_LDRSH_IMM = 0x005000f0,
158    INSN_LDRSH_REG = 0x001000f0,
159    INSN_STRH_IMM  = 0x004000b0,
160    INSN_STRH_REG  = 0x000000b0,
161
162    INSN_LDRB_IMM  = 0x04500000,
163    INSN_LDRB_REG  = 0x06500000,
164    INSN_LDRSB_IMM = 0x005000d0,
165    INSN_LDRSB_REG = 0x001000d0,
166    INSN_STRB_IMM  = 0x04400000,
167    INSN_STRB_REG  = 0x06400000,
168
169    INSN_LDRD_IMM  = 0x004000d0,
170    INSN_LDRD_REG  = 0x000000d0,
171    INSN_STRD_IMM  = 0x004000f0,
172    INSN_STRD_REG  = 0x000000f0,
173
174    INSN_DMB_ISH   = 0xf57ff05b,
175    INSN_DMB_MCR   = 0xee070fba,
176
177    /* Architected nop introduced in v6k.  */
178    /* ??? This is an MSR (imm) 0,0,0 insn.  Anyone know if this
179       also Just So Happened to do nothing on pre-v6k so that we
180       don't need to conditionalize it?  */
181    INSN_NOP_v6k   = 0xe320f000,
182    /* Otherwise the assembler uses mov r0,r0 */
183    INSN_NOP_v4    = (COND_AL << 28) | ARITH_MOV,
184
185    INSN_VADD      = 0xf2000800,
186    INSN_VAND      = 0xf2000110,
187    INSN_VBIC      = 0xf2100110,
188    INSN_VEOR      = 0xf3000110,
189    INSN_VORN      = 0xf2300110,
190    INSN_VORR      = 0xf2200110,
191    INSN_VSUB      = 0xf3000800,
192    INSN_VMUL      = 0xf2000910,
193    INSN_VQADD     = 0xf2000010,
194    INSN_VQADD_U   = 0xf3000010,
195    INSN_VQSUB     = 0xf2000210,
196    INSN_VQSUB_U   = 0xf3000210,
197    INSN_VMAX      = 0xf2000600,
198    INSN_VMAX_U    = 0xf3000600,
199    INSN_VMIN      = 0xf2000610,
200    INSN_VMIN_U    = 0xf3000610,
201
202    INSN_VABS      = 0xf3b10300,
203    INSN_VMVN      = 0xf3b00580,
204    INSN_VNEG      = 0xf3b10380,
205
206    INSN_VCEQ0     = 0xf3b10100,
207    INSN_VCGT0     = 0xf3b10000,
208    INSN_VCGE0     = 0xf3b10080,
209    INSN_VCLE0     = 0xf3b10180,
210    INSN_VCLT0     = 0xf3b10200,
211
212    INSN_VCEQ      = 0xf3000810,
213    INSN_VCGE      = 0xf2000310,
214    INSN_VCGT      = 0xf2000300,
215    INSN_VCGE_U    = 0xf3000310,
216    INSN_VCGT_U    = 0xf3000300,
217
218    INSN_VSHLI     = 0xf2800510,  /* VSHL (immediate) */
219    INSN_VSARI     = 0xf2800010,  /* VSHR.S */
220    INSN_VSHRI     = 0xf3800010,  /* VSHR.U */
221    INSN_VSLI      = 0xf3800510,
222    INSN_VSHL_S    = 0xf2000400,  /* VSHL.S (register) */
223    INSN_VSHL_U    = 0xf3000400,  /* VSHL.U (register) */
224
225    INSN_VBSL      = 0xf3100110,
226    INSN_VBIT      = 0xf3200110,
227    INSN_VBIF      = 0xf3300110,
228
229    INSN_VTST      = 0xf2000810,
230
231    INSN_VDUP_G    = 0xee800b10,  /* VDUP (ARM core register) */
232    INSN_VDUP_S    = 0xf3b00c00,  /* VDUP (scalar) */
233    INSN_VLDR_D    = 0xed100b00,  /* VLDR.64 */
234    INSN_VLD1      = 0xf4200000,  /* VLD1 (multiple single elements) */
235    INSN_VLD1R     = 0xf4a00c00,  /* VLD1 (single element to all lanes) */
236    INSN_VST1      = 0xf4000000,  /* VST1 (multiple single elements) */
237    INSN_VMOVI     = 0xf2800010,  /* VMOV (immediate) */
238} ARMInsn;
239
240#define INSN_NOP   (use_armv7_instructions ? INSN_NOP_v6k : INSN_NOP_v4)
241
242static const uint8_t tcg_cond_to_arm_cond[] = {
243    [TCG_COND_EQ] = COND_EQ,
244    [TCG_COND_NE] = COND_NE,
245    [TCG_COND_LT] = COND_LT,
246    [TCG_COND_GE] = COND_GE,
247    [TCG_COND_LE] = COND_LE,
248    [TCG_COND_GT] = COND_GT,
249    /* unsigned */
250    [TCG_COND_LTU] = COND_CC,
251    [TCG_COND_GEU] = COND_CS,
252    [TCG_COND_LEU] = COND_LS,
253    [TCG_COND_GTU] = COND_HI,
254};
255
256static int encode_imm(uint32_t imm);
257
258/* TCG private relocation type: add with pc+imm8 */
259#define R_ARM_PC8  11
260
261/* TCG private relocation type: vldr with imm8 << 2 */
262#define R_ARM_PC11 12
263
264static bool reloc_pc24(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
265{
266    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
267    ptrdiff_t offset = (tcg_ptr_byte_diff(target, src_rx) - 8) >> 2;
268
269    if (offset == sextract32(offset, 0, 24)) {
270        *src_rw = deposit32(*src_rw, 0, 24, offset);
271        return true;
272    }
273    return false;
274}
275
276static bool reloc_pc13(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
277{
278    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
279    ptrdiff_t offset = tcg_ptr_byte_diff(target, src_rx) - 8;
280
281    if (offset >= -0xfff && offset <= 0xfff) {
282        tcg_insn_unit insn = *src_rw;
283        bool u = (offset >= 0);
284        if (!u) {
285            offset = -offset;
286        }
287        insn = deposit32(insn, 23, 1, u);
288        insn = deposit32(insn, 0, 12, offset);
289        *src_rw = insn;
290        return true;
291    }
292    return false;
293}
294
295static bool reloc_pc11(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
296{
297    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
298    ptrdiff_t offset = (tcg_ptr_byte_diff(target, src_rx) - 8) / 4;
299
300    if (offset >= -0xff && offset <= 0xff) {
301        tcg_insn_unit insn = *src_rw;
302        bool u = (offset >= 0);
303        if (!u) {
304            offset = -offset;
305        }
306        insn = deposit32(insn, 23, 1, u);
307        insn = deposit32(insn, 0, 8, offset);
308        *src_rw = insn;
309        return true;
310    }
311    return false;
312}
313
314static bool reloc_pc8(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
315{
316    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
317    ptrdiff_t offset = tcg_ptr_byte_diff(target, src_rx) - 8;
318    int imm12 = encode_imm(offset);
319
320    if (imm12 >= 0) {
321        *src_rw = deposit32(*src_rw, 0, 12, imm12);
322        return true;
323    }
324    return false;
325}
326
327static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
328                        intptr_t value, intptr_t addend)
329{
330    tcg_debug_assert(addend == 0);
331    switch (type) {
332    case R_ARM_PC24:
333        return reloc_pc24(code_ptr, (const tcg_insn_unit *)value);
334    case R_ARM_PC13:
335        return reloc_pc13(code_ptr, (const tcg_insn_unit *)value);
336    case R_ARM_PC11:
337        return reloc_pc11(code_ptr, (const tcg_insn_unit *)value);
338    case R_ARM_PC8:
339        return reloc_pc8(code_ptr, (const tcg_insn_unit *)value);
340    default:
341        g_assert_not_reached();
342    }
343}
344
345#define TCG_CT_CONST_ARM  0x100
346#define TCG_CT_CONST_INV  0x200
347#define TCG_CT_CONST_NEG  0x400
348#define TCG_CT_CONST_ZERO 0x800
349#define TCG_CT_CONST_ORRI 0x1000
350#define TCG_CT_CONST_ANDI 0x2000
351
352#define ALL_GENERAL_REGS  0xffffu
353#define ALL_VECTOR_REGS   0xffff0000u
354
355/*
356 * r0-r3 will be overwritten when reading the tlb entry (softmmu only);
357 * r14 will be overwritten by the BLNE branching to the slow path.
358 */
359#ifdef CONFIG_SOFTMMU
360#define ALL_QLDST_REGS \
361    (ALL_GENERAL_REGS & ~((1 << TCG_REG_R0) | (1 << TCG_REG_R1) | \
362                          (1 << TCG_REG_R2) | (1 << TCG_REG_R3) | \
363                          (1 << TCG_REG_R14)))
364#else
365#define ALL_QLDST_REGS   (ALL_GENERAL_REGS & ~(1 << TCG_REG_R14))
366#endif
367
368/*
369 * ARM immediates for ALU instructions are made of an unsigned 8-bit
370 * right-rotated by an even amount between 0 and 30.
371 *
372 * Return < 0 if @imm cannot be encoded, else the entire imm12 field.
373 */
374static int encode_imm(uint32_t imm)
375{
376    uint32_t rot, imm8;
377
378    /* Simple case, no rotation required. */
379    if ((imm & ~0xff) == 0) {
380        return imm;
381    }
382
383    /* Next, try a simple even shift.  */
384    rot = ctz32(imm) & ~1;
385    imm8 = imm >> rot;
386    rot = 32 - rot;
387    if ((imm8 & ~0xff) == 0) {
388        goto found;
389    }
390
391    /*
392     * Finally, try harder with rotations.
393     * The ctz test above will have taken care of rotates >= 8.
394     */
395    for (rot = 2; rot < 8; rot += 2) {
396        imm8 = rol32(imm, rot);
397        if ((imm8 & ~0xff) == 0) {
398            goto found;
399        }
400    }
401    /* Fail: imm cannot be encoded. */
402    return -1;
403
404 found:
405    /* Note that rot is even, and we discard bit 0 by shifting by 7. */
406    return rot << 7 | imm8;
407}
408
409static int encode_imm_nofail(uint32_t imm)
410{
411    int ret = encode_imm(imm);
412    tcg_debug_assert(ret >= 0);
413    return ret;
414}
415
416static bool check_fit_imm(uint32_t imm)
417{
418    return encode_imm(imm) >= 0;
419}
420
421/* Return true if v16 is a valid 16-bit shifted immediate.  */
422static bool is_shimm16(uint16_t v16, int *cmode, int *imm8)
423{
424    if (v16 == (v16 & 0xff)) {
425        *cmode = 0x8;
426        *imm8 = v16 & 0xff;
427        return true;
428    } else if (v16 == (v16 & 0xff00)) {
429        *cmode = 0xa;
430        *imm8 = v16 >> 8;
431        return true;
432    }
433    return false;
434}
435
436/* Return true if v32 is a valid 32-bit shifted immediate.  */
437static bool is_shimm32(uint32_t v32, int *cmode, int *imm8)
438{
439    if (v32 == (v32 & 0xff)) {
440        *cmode = 0x0;
441        *imm8 = v32 & 0xff;
442        return true;
443    } else if (v32 == (v32 & 0xff00)) {
444        *cmode = 0x2;
445        *imm8 = (v32 >> 8) & 0xff;
446        return true;
447    } else if (v32 == (v32 & 0xff0000)) {
448        *cmode = 0x4;
449        *imm8 = (v32 >> 16) & 0xff;
450        return true;
451    } else if (v32 == (v32 & 0xff000000)) {
452        *cmode = 0x6;
453        *imm8 = v32 >> 24;
454        return true;
455    }
456    return false;
457}
458
459/* Return true if v32 is a valid 32-bit shifting ones immediate.  */
460static bool is_soimm32(uint32_t v32, int *cmode, int *imm8)
461{
462    if ((v32 & 0xffff00ff) == 0xff) {
463        *cmode = 0xc;
464        *imm8 = (v32 >> 8) & 0xff;
465        return true;
466    } else if ((v32 & 0xff00ffff) == 0xffff) {
467        *cmode = 0xd;
468        *imm8 = (v32 >> 16) & 0xff;
469        return true;
470    }
471    return false;
472}
473
474/*
475 * Return non-zero if v32 can be formed by MOVI+ORR.
476 * Place the parameters for MOVI in (cmode, imm8).
477 * Return the cmode for ORR; the imm8 can be had via extraction from v32.
478 */
479static int is_shimm32_pair(uint32_t v32, int *cmode, int *imm8)
480{
481    int i;
482
483    for (i = 6; i > 0; i -= 2) {
484        /* Mask out one byte we can add with ORR.  */
485        uint32_t tmp = v32 & ~(0xffu << (i * 4));
486        if (is_shimm32(tmp, cmode, imm8) ||
487            is_soimm32(tmp, cmode, imm8)) {
488            break;
489        }
490    }
491    return i;
492}
493
494/* Return true if V is a valid 16-bit or 32-bit shifted immediate.  */
495static bool is_shimm1632(uint32_t v32, int *cmode, int *imm8)
496{
497    if (v32 == deposit32(v32, 16, 16, v32)) {
498        return is_shimm16(v32, cmode, imm8);
499    } else {
500        return is_shimm32(v32, cmode, imm8);
501    }
502}
503
504/* Test if a constant matches the constraint.
505 * TODO: define constraints for:
506 *
507 * ldr/str offset:   between -0xfff and 0xfff
508 * ldrh/strh offset: between -0xff and 0xff
509 * mov operand2:     values represented with x << (2 * y), x < 0x100
510 * add, sub, eor...: ditto
511 */
512static bool tcg_target_const_match(int64_t val, TCGType type, int ct)
513{
514    if (ct & TCG_CT_CONST) {
515        return 1;
516    } else if ((ct & TCG_CT_CONST_ARM) && check_fit_imm(val)) {
517        return 1;
518    } else if ((ct & TCG_CT_CONST_INV) && check_fit_imm(~val)) {
519        return 1;
520    } else if ((ct & TCG_CT_CONST_NEG) && check_fit_imm(-val)) {
521        return 1;
522    } else if ((ct & TCG_CT_CONST_ZERO) && val == 0) {
523        return 1;
524    }
525
526    switch (ct & (TCG_CT_CONST_ORRI | TCG_CT_CONST_ANDI)) {
527    case 0:
528        break;
529    case TCG_CT_CONST_ANDI:
530        val = ~val;
531        /* fallthru */
532    case TCG_CT_CONST_ORRI:
533        if (val == deposit64(val, 32, 32, val)) {
534            int cmode, imm8;
535            return is_shimm1632(val, &cmode, &imm8);
536        }
537        break;
538    default:
539        /* Both bits should not be set for the same insn.  */
540        g_assert_not_reached();
541    }
542
543    return 0;
544}
545
546static void tcg_out_b_imm(TCGContext *s, ARMCond cond, int32_t offset)
547{
548    tcg_out32(s, (cond << 28) | INSN_B |
549                    (((offset - 8) >> 2) & 0x00ffffff));
550}
551
552static void tcg_out_bl_imm(TCGContext *s, ARMCond cond, int32_t offset)
553{
554    tcg_out32(s, (cond << 28) | 0x0b000000 |
555                    (((offset - 8) >> 2) & 0x00ffffff));
556}
557
558static void tcg_out_blx_reg(TCGContext *s, ARMCond cond, TCGReg rn)
559{
560    tcg_out32(s, (cond << 28) | 0x012fff30 | rn);
561}
562
563static void tcg_out_blx_imm(TCGContext *s, int32_t offset)
564{
565    tcg_out32(s, 0xfa000000 | ((offset & 2) << 23) |
566                (((offset - 8) >> 2) & 0x00ffffff));
567}
568
569static void tcg_out_dat_reg(TCGContext *s, ARMCond cond, ARMInsn opc,
570                            TCGReg rd, TCGReg rn, TCGReg rm, int shift)
571{
572    tcg_out32(s, (cond << 28) | (0 << 25) | opc |
573                    (rn << 16) | (rd << 12) | shift | rm);
574}
575
576static void tcg_out_mov_reg(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rm)
577{
578    /* Simple reg-reg move, optimising out the 'do nothing' case */
579    if (rd != rm) {
580        tcg_out_dat_reg(s, cond, ARITH_MOV, rd, 0, rm, SHIFT_IMM_LSL(0));
581    }
582}
583
584static void tcg_out_bx_reg(TCGContext *s, ARMCond cond, TCGReg rn)
585{
586    tcg_out32(s, (cond << 28) | 0x012fff10 | rn);
587}
588
589static void tcg_out_b_reg(TCGContext *s, ARMCond cond, TCGReg rn)
590{
591    /*
592     * Unless the C portion of QEMU is compiled as thumb, we don't need
593     * true BX semantics; merely a branch to an address held in a register.
594     */
595    tcg_out_bx_reg(s, cond, rn);
596}
597
598static void tcg_out_dat_imm(TCGContext *s, ARMCond cond, ARMInsn opc,
599                            TCGReg rd, TCGReg rn, int im)
600{
601    tcg_out32(s, (cond << 28) | (1 << 25) | opc |
602                    (rn << 16) | (rd << 12) | im);
603}
604
605static void tcg_out_ldstm(TCGContext *s, ARMCond cond, ARMInsn opc,
606                          TCGReg rn, uint16_t mask)
607{
608    tcg_out32(s, (cond << 28) | opc | (rn << 16) | mask);
609}
610
611/* Note that this routine is used for both LDR and LDRH formats, so we do
612   not wish to include an immediate shift at this point.  */
613static void tcg_out_memop_r(TCGContext *s, ARMCond cond, ARMInsn opc, TCGReg rt,
614                            TCGReg rn, TCGReg rm, bool u, bool p, bool w)
615{
616    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24)
617              | (w << 21) | (rn << 16) | (rt << 12) | rm);
618}
619
620static void tcg_out_memop_8(TCGContext *s, ARMCond cond, ARMInsn opc, TCGReg rt,
621                            TCGReg rn, int imm8, bool p, bool w)
622{
623    bool u = 1;
624    if (imm8 < 0) {
625        imm8 = -imm8;
626        u = 0;
627    }
628    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24) | (w << 21) |
629              (rn << 16) | (rt << 12) | ((imm8 & 0xf0) << 4) | (imm8 & 0xf));
630}
631
632static void tcg_out_memop_12(TCGContext *s, ARMCond cond, ARMInsn opc,
633                             TCGReg rt, TCGReg rn, int imm12, bool p, bool w)
634{
635    bool u = 1;
636    if (imm12 < 0) {
637        imm12 = -imm12;
638        u = 0;
639    }
640    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24) | (w << 21) |
641              (rn << 16) | (rt << 12) | imm12);
642}
643
644static void tcg_out_ld32_12(TCGContext *s, ARMCond cond, TCGReg rt,
645                            TCGReg rn, int imm12)
646{
647    tcg_out_memop_12(s, cond, INSN_LDR_IMM, rt, rn, imm12, 1, 0);
648}
649
650static void tcg_out_st32_12(TCGContext *s, ARMCond cond, TCGReg rt,
651                            TCGReg rn, int imm12)
652{
653    tcg_out_memop_12(s, cond, INSN_STR_IMM, rt, rn, imm12, 1, 0);
654}
655
656static void tcg_out_ld32_r(TCGContext *s, ARMCond cond, TCGReg rt,
657                           TCGReg rn, TCGReg rm)
658{
659    tcg_out_memop_r(s, cond, INSN_LDR_REG, rt, rn, rm, 1, 1, 0);
660}
661
662static void tcg_out_st32_r(TCGContext *s, ARMCond cond, TCGReg rt,
663                           TCGReg rn, TCGReg rm)
664{
665    tcg_out_memop_r(s, cond, INSN_STR_REG, rt, rn, rm, 1, 1, 0);
666}
667
668static void tcg_out_ldrd_8(TCGContext *s, ARMCond cond, TCGReg rt,
669                           TCGReg rn, int imm8)
670{
671    tcg_out_memop_8(s, cond, INSN_LDRD_IMM, rt, rn, imm8, 1, 0);
672}
673
674static void tcg_out_ldrd_r(TCGContext *s, ARMCond cond, TCGReg rt,
675                           TCGReg rn, TCGReg rm)
676{
677    tcg_out_memop_r(s, cond, INSN_LDRD_REG, rt, rn, rm, 1, 1, 0);
678}
679
680static void __attribute__((unused))
681tcg_out_ldrd_rwb(TCGContext *s, ARMCond cond, TCGReg rt, TCGReg rn, TCGReg rm)
682{
683    tcg_out_memop_r(s, cond, INSN_LDRD_REG, rt, rn, rm, 1, 1, 1);
684}
685
686static void __attribute__((unused))
687tcg_out_strd_8(TCGContext *s, ARMCond cond, TCGReg rt, TCGReg rn, int imm8)
688{
689    tcg_out_memop_8(s, cond, INSN_STRD_IMM, rt, rn, imm8, 1, 0);
690}
691
692static void tcg_out_strd_r(TCGContext *s, ARMCond cond, TCGReg rt,
693                           TCGReg rn, TCGReg rm)
694{
695    tcg_out_memop_r(s, cond, INSN_STRD_REG, rt, rn, rm, 1, 1, 0);
696}
697
698/* Register pre-increment with base writeback.  */
699static void tcg_out_ld32_rwb(TCGContext *s, ARMCond cond, TCGReg rt,
700                             TCGReg rn, TCGReg rm)
701{
702    tcg_out_memop_r(s, cond, INSN_LDR_REG, rt, rn, rm, 1, 1, 1);
703}
704
705static void tcg_out_st32_rwb(TCGContext *s, ARMCond cond, TCGReg rt,
706                             TCGReg rn, TCGReg rm)
707{
708    tcg_out_memop_r(s, cond, INSN_STR_REG, rt, rn, rm, 1, 1, 1);
709}
710
711static void tcg_out_ld16u_8(TCGContext *s, ARMCond cond, TCGReg rt,
712                            TCGReg rn, int imm8)
713{
714    tcg_out_memop_8(s, cond, INSN_LDRH_IMM, rt, rn, imm8, 1, 0);
715}
716
717static void tcg_out_st16_8(TCGContext *s, ARMCond cond, TCGReg rt,
718                           TCGReg rn, int imm8)
719{
720    tcg_out_memop_8(s, cond, INSN_STRH_IMM, rt, rn, imm8, 1, 0);
721}
722
723static void tcg_out_ld16u_r(TCGContext *s, ARMCond cond, TCGReg rt,
724                            TCGReg rn, TCGReg rm)
725{
726    tcg_out_memop_r(s, cond, INSN_LDRH_REG, rt, rn, rm, 1, 1, 0);
727}
728
729static void tcg_out_st16_r(TCGContext *s, ARMCond cond, TCGReg rt,
730                           TCGReg rn, TCGReg rm)
731{
732    tcg_out_memop_r(s, cond, INSN_STRH_REG, rt, rn, rm, 1, 1, 0);
733}
734
735static void tcg_out_ld16s_8(TCGContext *s, ARMCond cond, TCGReg rt,
736                            TCGReg rn, int imm8)
737{
738    tcg_out_memop_8(s, cond, INSN_LDRSH_IMM, rt, rn, imm8, 1, 0);
739}
740
741static void tcg_out_ld16s_r(TCGContext *s, ARMCond cond, TCGReg rt,
742                            TCGReg rn, TCGReg rm)
743{
744    tcg_out_memop_r(s, cond, INSN_LDRSH_REG, rt, rn, rm, 1, 1, 0);
745}
746
747static void tcg_out_ld8_12(TCGContext *s, ARMCond cond, TCGReg rt,
748                           TCGReg rn, int imm12)
749{
750    tcg_out_memop_12(s, cond, INSN_LDRB_IMM, rt, rn, imm12, 1, 0);
751}
752
753static void tcg_out_st8_12(TCGContext *s, ARMCond cond, TCGReg rt,
754                           TCGReg rn, int imm12)
755{
756    tcg_out_memop_12(s, cond, INSN_STRB_IMM, rt, rn, imm12, 1, 0);
757}
758
759static void tcg_out_ld8_r(TCGContext *s, ARMCond cond, TCGReg rt,
760                          TCGReg rn, TCGReg rm)
761{
762    tcg_out_memop_r(s, cond, INSN_LDRB_REG, rt, rn, rm, 1, 1, 0);
763}
764
765static void tcg_out_st8_r(TCGContext *s, ARMCond cond, TCGReg rt,
766                          TCGReg rn, TCGReg rm)
767{
768    tcg_out_memop_r(s, cond, INSN_STRB_REG, rt, rn, rm, 1, 1, 0);
769}
770
771static void tcg_out_ld8s_8(TCGContext *s, ARMCond cond, TCGReg rt,
772                           TCGReg rn, int imm8)
773{
774    tcg_out_memop_8(s, cond, INSN_LDRSB_IMM, rt, rn, imm8, 1, 0);
775}
776
777static void tcg_out_ld8s_r(TCGContext *s, ARMCond cond, TCGReg rt,
778                           TCGReg rn, TCGReg rm)
779{
780    tcg_out_memop_r(s, cond, INSN_LDRSB_REG, rt, rn, rm, 1, 1, 0);
781}
782
783static void tcg_out_movi_pool(TCGContext *s, ARMCond cond,
784                              TCGReg rd, uint32_t arg)
785{
786    new_pool_label(s, arg, R_ARM_PC13, s->code_ptr, 0);
787    tcg_out_ld32_12(s, cond, rd, TCG_REG_PC, 0);
788}
789
790static void tcg_out_movi32(TCGContext *s, ARMCond cond,
791                           TCGReg rd, uint32_t arg)
792{
793    int imm12, diff, opc, sh1, sh2;
794    uint32_t tt0, tt1, tt2;
795
796    /* Check a single MOV/MVN before anything else.  */
797    imm12 = encode_imm(arg);
798    if (imm12 >= 0) {
799        tcg_out_dat_imm(s, cond, ARITH_MOV, rd, 0, imm12);
800        return;
801    }
802    imm12 = encode_imm(~arg);
803    if (imm12 >= 0) {
804        tcg_out_dat_imm(s, cond, ARITH_MVN, rd, 0, imm12);
805        return;
806    }
807
808    /* Check for a pc-relative address.  This will usually be the TB,
809       or within the TB, which is immediately before the code block.  */
810    diff = tcg_pcrel_diff(s, (void *)arg) - 8;
811    if (diff >= 0) {
812        imm12 = encode_imm(diff);
813        if (imm12 >= 0) {
814            tcg_out_dat_imm(s, cond, ARITH_ADD, rd, TCG_REG_PC, imm12);
815            return;
816        }
817    } else {
818        imm12 = encode_imm(-diff);
819        if (imm12 >= 0) {
820            tcg_out_dat_imm(s, cond, ARITH_SUB, rd, TCG_REG_PC, imm12);
821            return;
822        }
823    }
824
825    /* Use movw + movt.  */
826    if (use_armv7_instructions) {
827        /* movw */
828        tcg_out32(s, (cond << 28) | 0x03000000 | (rd << 12)
829                  | ((arg << 4) & 0x000f0000) | (arg & 0xfff));
830        if (arg & 0xffff0000) {
831            /* movt */
832            tcg_out32(s, (cond << 28) | 0x03400000 | (rd << 12)
833                      | ((arg >> 12) & 0x000f0000) | ((arg >> 16) & 0xfff));
834        }
835        return;
836    }
837
838    /* Look for sequences of two insns.  If we have lots of 1's, we can
839       shorten the sequence by beginning with mvn and then clearing
840       higher bits with eor.  */
841    tt0 = arg;
842    opc = ARITH_MOV;
843    if (ctpop32(arg) > 16) {
844        tt0 = ~arg;
845        opc = ARITH_MVN;
846    }
847    sh1 = ctz32(tt0) & ~1;
848    tt1 = tt0 & ~(0xff << sh1);
849    sh2 = ctz32(tt1) & ~1;
850    tt2 = tt1 & ~(0xff << sh2);
851    if (tt2 == 0) {
852        int rot;
853
854        rot = ((32 - sh1) << 7) & 0xf00;
855        tcg_out_dat_imm(s, cond, opc, rd,  0, ((tt0 >> sh1) & 0xff) | rot);
856        rot = ((32 - sh2) << 7) & 0xf00;
857        tcg_out_dat_imm(s, cond, ARITH_EOR, rd, rd,
858                        ((tt0 >> sh2) & 0xff) | rot);
859        return;
860    }
861
862    /* Otherwise, drop it into the constant pool.  */
863    tcg_out_movi_pool(s, cond, rd, arg);
864}
865
866/*
867 * Emit either the reg,imm or reg,reg form of a data-processing insn.
868 * rhs must satisfy the "rI" constraint.
869 */
870static void tcg_out_dat_rI(TCGContext *s, ARMCond cond, ARMInsn opc,
871                           TCGReg dst, TCGReg lhs, TCGArg rhs, int rhs_is_const)
872{
873    if (rhs_is_const) {
874        tcg_out_dat_imm(s, cond, opc, dst, lhs, encode_imm_nofail(rhs));
875    } else {
876        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
877    }
878}
879
880/*
881 * Emit either the reg,imm or reg,reg form of a data-processing insn.
882 * rhs must satisfy the "rIK" constraint.
883 */
884static void tcg_out_dat_rIK(TCGContext *s, ARMCond cond, ARMInsn opc,
885                            ARMInsn opinv, TCGReg dst, TCGReg lhs, TCGArg rhs,
886                            bool rhs_is_const)
887{
888    if (rhs_is_const) {
889        int imm12 = encode_imm(rhs);
890        if (imm12 < 0) {
891            imm12 = encode_imm_nofail(~rhs);
892            opc = opinv;
893        }
894        tcg_out_dat_imm(s, cond, opc, dst, lhs, imm12);
895    } else {
896        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
897    }
898}
899
900static void tcg_out_dat_rIN(TCGContext *s, ARMCond cond, ARMInsn opc,
901                            ARMInsn opneg, TCGReg dst, TCGReg lhs, TCGArg rhs,
902                            bool rhs_is_const)
903{
904    /* Emit either the reg,imm or reg,reg form of a data-processing insn.
905     * rhs must satisfy the "rIN" constraint.
906     */
907    if (rhs_is_const) {
908        int imm12 = encode_imm(rhs);
909        if (imm12 < 0) {
910            imm12 = encode_imm_nofail(-rhs);
911            opc = opneg;
912        }
913        tcg_out_dat_imm(s, cond, opc, dst, lhs, imm12);
914    } else {
915        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
916    }
917}
918
919static void tcg_out_mul32(TCGContext *s, ARMCond cond, TCGReg rd,
920                          TCGReg rn, TCGReg rm)
921{
922    /* mul */
923    tcg_out32(s, (cond << 28) | 0x90 | (rd << 16) | (rm << 8) | rn);
924}
925
926static void tcg_out_umull32(TCGContext *s, ARMCond cond, TCGReg rd0,
927                            TCGReg rd1, TCGReg rn, TCGReg rm)
928{
929    /* umull */
930    tcg_out32(s, (cond << 28) | 0x00800090 |
931              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
932}
933
934static void tcg_out_smull32(TCGContext *s, ARMCond cond, TCGReg rd0,
935                            TCGReg rd1, TCGReg rn, TCGReg rm)
936{
937    /* smull */
938    tcg_out32(s, (cond << 28) | 0x00c00090 |
939              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
940}
941
942static void tcg_out_sdiv(TCGContext *s, ARMCond cond,
943                         TCGReg rd, TCGReg rn, TCGReg rm)
944{
945    tcg_out32(s, 0x0710f010 | (cond << 28) | (rd << 16) | rn | (rm << 8));
946}
947
948static void tcg_out_udiv(TCGContext *s, ARMCond cond,
949                         TCGReg rd, TCGReg rn, TCGReg rm)
950{
951    tcg_out32(s, 0x0730f010 | (cond << 28) | (rd << 16) | rn | (rm << 8));
952}
953
954static void tcg_out_ext8s(TCGContext *s, TCGType t, TCGReg rd, TCGReg rn)
955{
956    /* sxtb */
957    tcg_out32(s, 0x06af0070 | (COND_AL << 28) | (rd << 12) | rn);
958}
959
960static void tcg_out_ext8u(TCGContext *s, TCGReg rd, TCGReg rn)
961{
962    tcg_out_dat_imm(s, COND_AL, ARITH_AND, rd, rn, 0xff);
963}
964
965static void tcg_out_ext16s(TCGContext *s, TCGType t, TCGReg rd, TCGReg rn)
966{
967    /* sxth */
968    tcg_out32(s, 0x06bf0070 | (COND_AL << 28) | (rd << 12) | rn);
969}
970
971static void tcg_out_ext16u(TCGContext *s, TCGReg rd, TCGReg rn)
972{
973    /* uxth */
974    tcg_out32(s, 0x06ff0070 | (COND_AL << 28) | (rd << 12) | rn);
975}
976
977static void tcg_out_ext32s(TCGContext *s, TCGReg rd, TCGReg rn)
978{
979    g_assert_not_reached();
980}
981
982static void tcg_out_ext32u(TCGContext *s, TCGReg rd, TCGReg rn)
983{
984    g_assert_not_reached();
985}
986
987static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg rd, TCGReg rn)
988{
989    g_assert_not_reached();
990}
991
992static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg rd, TCGReg rn)
993{
994    g_assert_not_reached();
995}
996
997static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg rd, TCGReg rn)
998{
999    g_assert_not_reached();
1000}
1001
1002static void tcg_out_bswap16(TCGContext *s, ARMCond cond,
1003                            TCGReg rd, TCGReg rn, int flags)
1004{
1005    if (flags & TCG_BSWAP_OS) {
1006        /* revsh */
1007        tcg_out32(s, 0x06ff0fb0 | (cond << 28) | (rd << 12) | rn);
1008        return;
1009    }
1010
1011    /* rev16 */
1012    tcg_out32(s, 0x06bf0fb0 | (cond << 28) | (rd << 12) | rn);
1013    if ((flags & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1014        /* uxth */
1015        tcg_out32(s, 0x06ff0070 | (cond << 28) | (rd << 12) | rd);
1016    }
1017}
1018
1019static void tcg_out_bswap32(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rn)
1020{
1021    /* rev */
1022    tcg_out32(s, 0x06bf0f30 | (cond << 28) | (rd << 12) | rn);
1023}
1024
1025static void tcg_out_deposit(TCGContext *s, ARMCond cond, TCGReg rd,
1026                            TCGArg a1, int ofs, int len, bool const_a1)
1027{
1028    if (const_a1) {
1029        /* bfi becomes bfc with rn == 15.  */
1030        a1 = 15;
1031    }
1032    /* bfi/bfc */
1033    tcg_out32(s, 0x07c00010 | (cond << 28) | (rd << 12) | a1
1034              | (ofs << 7) | ((ofs + len - 1) << 16));
1035}
1036
1037static void tcg_out_extract(TCGContext *s, ARMCond cond, TCGReg rd,
1038                            TCGReg rn, int ofs, int len)
1039{
1040    /* ubfx */
1041    tcg_out32(s, 0x07e00050 | (cond << 28) | (rd << 12) | rn
1042              | (ofs << 7) | ((len - 1) << 16));
1043}
1044
1045static void tcg_out_sextract(TCGContext *s, ARMCond cond, TCGReg rd,
1046                             TCGReg rn, int ofs, int len)
1047{
1048    /* sbfx */
1049    tcg_out32(s, 0x07a00050 | (cond << 28) | (rd << 12) | rn
1050              | (ofs << 7) | ((len - 1) << 16));
1051}
1052
1053static void tcg_out_ld32u(TCGContext *s, ARMCond cond,
1054                          TCGReg rd, TCGReg rn, int32_t offset)
1055{
1056    if (offset > 0xfff || offset < -0xfff) {
1057        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1058        tcg_out_ld32_r(s, cond, rd, rn, TCG_REG_TMP);
1059    } else
1060        tcg_out_ld32_12(s, cond, rd, rn, offset);
1061}
1062
1063static void tcg_out_st32(TCGContext *s, ARMCond cond,
1064                         TCGReg rd, TCGReg rn, int32_t offset)
1065{
1066    if (offset > 0xfff || offset < -0xfff) {
1067        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1068        tcg_out_st32_r(s, cond, rd, rn, TCG_REG_TMP);
1069    } else
1070        tcg_out_st32_12(s, cond, rd, rn, offset);
1071}
1072
1073static void tcg_out_ld16u(TCGContext *s, ARMCond cond,
1074                          TCGReg rd, TCGReg rn, int32_t offset)
1075{
1076    if (offset > 0xff || offset < -0xff) {
1077        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1078        tcg_out_ld16u_r(s, cond, rd, rn, TCG_REG_TMP);
1079    } else
1080        tcg_out_ld16u_8(s, cond, rd, rn, offset);
1081}
1082
1083static void tcg_out_ld16s(TCGContext *s, ARMCond cond,
1084                          TCGReg rd, TCGReg rn, int32_t offset)
1085{
1086    if (offset > 0xff || offset < -0xff) {
1087        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1088        tcg_out_ld16s_r(s, cond, rd, rn, TCG_REG_TMP);
1089    } else
1090        tcg_out_ld16s_8(s, cond, rd, rn, offset);
1091}
1092
1093static void tcg_out_st16(TCGContext *s, ARMCond cond,
1094                         TCGReg rd, TCGReg rn, int32_t offset)
1095{
1096    if (offset > 0xff || offset < -0xff) {
1097        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1098        tcg_out_st16_r(s, cond, rd, rn, TCG_REG_TMP);
1099    } else
1100        tcg_out_st16_8(s, cond, rd, rn, offset);
1101}
1102
1103static void tcg_out_ld8u(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_ld8_r(s, cond, rd, rn, TCG_REG_TMP);
1109    } else
1110        tcg_out_ld8_12(s, cond, rd, rn, offset);
1111}
1112
1113static void tcg_out_ld8s(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_ld8s_r(s, cond, rd, rn, TCG_REG_TMP);
1119    } else
1120        tcg_out_ld8s_8(s, cond, rd, rn, offset);
1121}
1122
1123static void tcg_out_st8(TCGContext *s, ARMCond cond,
1124                        TCGReg rd, TCGReg rn, int32_t offset)
1125{
1126    if (offset > 0xfff || offset < -0xfff) {
1127        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1128        tcg_out_st8_r(s, cond, rd, rn, TCG_REG_TMP);
1129    } else
1130        tcg_out_st8_12(s, cond, rd, rn, offset);
1131}
1132
1133/*
1134 * The _goto case is normally between TBs within the same code buffer, and
1135 * with the code buffer limited to 16MB we wouldn't need the long case.
1136 * But we also use it for the tail-call to the qemu_ld/st helpers, which does.
1137 */
1138static void tcg_out_goto(TCGContext *s, ARMCond cond, const tcg_insn_unit *addr)
1139{
1140    intptr_t addri = (intptr_t)addr;
1141    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1142    bool arm_mode = !(addri & 1);
1143
1144    if (arm_mode && disp - 8 < 0x01fffffd && disp - 8 > -0x01fffffd) {
1145        tcg_out_b_imm(s, cond, disp);
1146        return;
1147    }
1148
1149    /* LDR is interworking from v5t. */
1150    tcg_out_movi_pool(s, cond, TCG_REG_PC, addri);
1151}
1152
1153/*
1154 * The call case is mostly used for helpers - so it's not unreasonable
1155 * for them to be beyond branch range.
1156 */
1157static void tcg_out_call_int(TCGContext *s, const tcg_insn_unit *addr)
1158{
1159    intptr_t addri = (intptr_t)addr;
1160    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1161    bool arm_mode = !(addri & 1);
1162
1163    if (disp - 8 < 0x02000000 && disp - 8 >= -0x02000000) {
1164        if (arm_mode) {
1165            tcg_out_bl_imm(s, COND_AL, disp);
1166        } else {
1167            tcg_out_blx_imm(s, disp);
1168        }
1169        return;
1170    }
1171
1172    tcg_out_movi32(s, COND_AL, TCG_REG_TMP, addri);
1173    tcg_out_blx_reg(s, COND_AL, TCG_REG_TMP);
1174}
1175
1176static void tcg_out_call(TCGContext *s, const tcg_insn_unit *addr,
1177                         const TCGHelperInfo *info)
1178{
1179    tcg_out_call_int(s, addr);
1180}
1181
1182static void tcg_out_goto_label(TCGContext *s, ARMCond cond, TCGLabel *l)
1183{
1184    if (l->has_value) {
1185        tcg_out_goto(s, cond, l->u.value_ptr);
1186    } else {
1187        tcg_out_reloc(s, s->code_ptr, R_ARM_PC24, l, 0);
1188        tcg_out_b_imm(s, cond, 0);
1189    }
1190}
1191
1192static void tcg_out_mb(TCGContext *s, TCGArg a0)
1193{
1194    if (use_armv7_instructions) {
1195        tcg_out32(s, INSN_DMB_ISH);
1196    } else {
1197        tcg_out32(s, INSN_DMB_MCR);
1198    }
1199}
1200
1201static TCGCond tcg_out_cmp2(TCGContext *s, const TCGArg *args,
1202                            const int *const_args)
1203{
1204    TCGReg al = args[0];
1205    TCGReg ah = args[1];
1206    TCGArg bl = args[2];
1207    TCGArg bh = args[3];
1208    TCGCond cond = args[4];
1209    int const_bl = const_args[2];
1210    int const_bh = const_args[3];
1211
1212    switch (cond) {
1213    case TCG_COND_EQ:
1214    case TCG_COND_NE:
1215    case TCG_COND_LTU:
1216    case TCG_COND_LEU:
1217    case TCG_COND_GTU:
1218    case TCG_COND_GEU:
1219        /* We perform a conditional comparision.  If the high half is
1220           equal, then overwrite the flags with the comparison of the
1221           low half.  The resulting flags cover the whole.  */
1222        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, ah, bh, const_bh);
1223        tcg_out_dat_rI(s, COND_EQ, ARITH_CMP, 0, al, bl, const_bl);
1224        return cond;
1225
1226    case TCG_COND_LT:
1227    case TCG_COND_GE:
1228        /* We perform a double-word subtraction and examine the result.
1229           We do not actually need the result of the subtract, so the
1230           low part "subtract" is a compare.  For the high half we have
1231           no choice but to compute into a temporary.  */
1232        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, al, bl, const_bl);
1233        tcg_out_dat_rI(s, COND_AL, ARITH_SBC | TO_CPSR,
1234                       TCG_REG_TMP, ah, bh, const_bh);
1235        return cond;
1236
1237    case TCG_COND_LE:
1238    case TCG_COND_GT:
1239        /* Similar, but with swapped arguments, via reversed subtract.  */
1240        tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR,
1241                       TCG_REG_TMP, al, bl, const_bl);
1242        tcg_out_dat_rI(s, COND_AL, ARITH_RSC | TO_CPSR,
1243                       TCG_REG_TMP, ah, bh, const_bh);
1244        return tcg_swap_cond(cond);
1245
1246    default:
1247        g_assert_not_reached();
1248    }
1249}
1250
1251/*
1252 * Note that TCGReg references Q-registers.
1253 * Q-regno = 2 * D-regno, so shift left by 1 whlie inserting.
1254 */
1255static uint32_t encode_vd(TCGReg rd)
1256{
1257    tcg_debug_assert(rd >= TCG_REG_Q0);
1258    return (extract32(rd, 3, 1) << 22) | (extract32(rd, 0, 3) << 13);
1259}
1260
1261static uint32_t encode_vn(TCGReg rn)
1262{
1263    tcg_debug_assert(rn >= TCG_REG_Q0);
1264    return (extract32(rn, 3, 1) << 7) | (extract32(rn, 0, 3) << 17);
1265}
1266
1267static uint32_t encode_vm(TCGReg rm)
1268{
1269    tcg_debug_assert(rm >= TCG_REG_Q0);
1270    return (extract32(rm, 3, 1) << 5) | (extract32(rm, 0, 3) << 1);
1271}
1272
1273static void tcg_out_vreg2(TCGContext *s, ARMInsn insn, int q, int vece,
1274                          TCGReg d, TCGReg m)
1275{
1276    tcg_out32(s, insn | (vece << 18) | (q << 6) |
1277              encode_vd(d) | encode_vm(m));
1278}
1279
1280static void tcg_out_vreg3(TCGContext *s, ARMInsn insn, int q, int vece,
1281                          TCGReg d, TCGReg n, TCGReg m)
1282{
1283    tcg_out32(s, insn | (vece << 20) | (q << 6) |
1284              encode_vd(d) | encode_vn(n) | encode_vm(m));
1285}
1286
1287static void tcg_out_vmovi(TCGContext *s, TCGReg rd,
1288                          int q, int op, int cmode, uint8_t imm8)
1289{
1290    tcg_out32(s, INSN_VMOVI | encode_vd(rd) | (q << 6) | (op << 5)
1291              | (cmode << 8) | extract32(imm8, 0, 4)
1292              | (extract32(imm8, 4, 3) << 16)
1293              | (extract32(imm8, 7, 1) << 24));
1294}
1295
1296static void tcg_out_vshifti(TCGContext *s, ARMInsn insn, int q,
1297                            TCGReg rd, TCGReg rm, int l_imm6)
1298{
1299    tcg_out32(s, insn | (q << 6) | encode_vd(rd) | encode_vm(rm) |
1300              (extract32(l_imm6, 6, 1) << 7) |
1301              (extract32(l_imm6, 0, 6) << 16));
1302}
1303
1304static void tcg_out_vldst(TCGContext *s, ARMInsn insn,
1305                          TCGReg rd, TCGReg rn, int offset)
1306{
1307    if (offset != 0) {
1308        if (check_fit_imm(offset) || check_fit_imm(-offset)) {
1309            tcg_out_dat_rIN(s, COND_AL, ARITH_ADD, ARITH_SUB,
1310                            TCG_REG_TMP, rn, offset, true);
1311        } else {
1312            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP, offset);
1313            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1314                            TCG_REG_TMP, TCG_REG_TMP, rn, 0);
1315        }
1316        rn = TCG_REG_TMP;
1317    }
1318    tcg_out32(s, insn | (rn << 16) | encode_vd(rd) | 0xf);
1319}
1320
1321typedef struct {
1322    ARMCond cond;
1323    TCGReg base;
1324    int index;
1325    bool index_scratch;
1326    TCGAtomAlign aa;
1327} HostAddress;
1328
1329bool tcg_target_has_memory_bswap(MemOp memop)
1330{
1331    return false;
1332}
1333
1334static TCGReg ldst_ra_gen(TCGContext *s, const TCGLabelQemuLdst *l, int arg)
1335{
1336    /* We arrive at the slow path via "BLNE", so R14 contains l->raddr. */
1337    return TCG_REG_R14;
1338}
1339
1340static const TCGLdstHelperParam ldst_helper_param = {
1341    .ra_gen = ldst_ra_gen,
1342    .ntmp = 1,
1343    .tmp = { TCG_REG_TMP },
1344};
1345
1346static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1347{
1348    MemOp opc = get_memop(lb->oi);
1349
1350    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1351        return false;
1352    }
1353
1354    tcg_out_ld_helper_args(s, lb, &ldst_helper_param);
1355    tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SIZE]);
1356    tcg_out_ld_helper_ret(s, lb, false, &ldst_helper_param);
1357
1358    tcg_out_goto(s, COND_AL, lb->raddr);
1359    return true;
1360}
1361
1362static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1363{
1364    MemOp opc = get_memop(lb->oi);
1365
1366    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1367        return false;
1368    }
1369
1370    tcg_out_st_helper_args(s, lb, &ldst_helper_param);
1371
1372    /* Tail-call to the helper, which will return to the fast path.  */
1373    tcg_out_goto(s, COND_AL, qemu_st_helpers[opc & MO_SIZE]);
1374    return true;
1375}
1376
1377static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
1378                                           TCGReg addrlo, TCGReg addrhi,
1379                                           MemOpIdx oi, bool is_ld)
1380{
1381    TCGLabelQemuLdst *ldst = NULL;
1382    MemOp opc = get_memop(oi);
1383    unsigned a_mask;
1384
1385#ifdef CONFIG_SOFTMMU
1386    *h = (HostAddress){
1387        .cond = COND_AL,
1388        .base = addrlo,
1389        .index = TCG_REG_R1,
1390        .index_scratch = true,
1391    };
1392#else
1393    *h = (HostAddress){
1394        .cond = COND_AL,
1395        .base = addrlo,
1396        .index = guest_base ? TCG_REG_GUEST_BASE : -1,
1397        .index_scratch = false,
1398    };
1399#endif
1400
1401    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, false);
1402    a_mask = (1 << h->aa.align) - 1;
1403
1404#ifdef CONFIG_SOFTMMU
1405    int mem_index = get_mmuidx(oi);
1406    int cmp_off = is_ld ? offsetof(CPUTLBEntry, addr_read)
1407                        : offsetof(CPUTLBEntry, addr_write);
1408    int fast_off = TLB_MASK_TABLE_OFS(mem_index);
1409    unsigned s_mask = (1 << (opc & MO_SIZE)) - 1;
1410    TCGReg t_addr;
1411
1412    ldst = new_ldst_label(s);
1413    ldst->is_ld = is_ld;
1414    ldst->oi = oi;
1415    ldst->addrlo_reg = addrlo;
1416    ldst->addrhi_reg = addrhi;
1417
1418    /* Load env_tlb(env)->f[mmu_idx].{mask,table} into {r0,r1}.  */
1419    QEMU_BUILD_BUG_ON(TLB_MASK_TABLE_OFS(0) > 0);
1420    QEMU_BUILD_BUG_ON(TLB_MASK_TABLE_OFS(0) < -256);
1421    QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, mask) != 0);
1422    QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, table) != 4);
1423    tcg_out_ldrd_8(s, COND_AL, TCG_REG_R0, TCG_AREG0, fast_off);
1424
1425    /* Extract the tlb index from the address into R0.  */
1426    tcg_out_dat_reg(s, COND_AL, ARITH_AND, TCG_REG_R0, TCG_REG_R0, addrlo,
1427                    SHIFT_IMM_LSR(s->page_bits - CPU_TLB_ENTRY_BITS));
1428
1429    /*
1430     * Add the tlb_table pointer, creating the CPUTLBEntry address in R1.
1431     * Load the tlb comparator into R2/R3 and the fast path addend into R1.
1432     */
1433    QEMU_BUILD_BUG_ON(HOST_BIG_ENDIAN);
1434    if (cmp_off == 0) {
1435        if (s->addr_type == TCG_TYPE_I32) {
1436            tcg_out_ld32_rwb(s, COND_AL, TCG_REG_R2, TCG_REG_R1, TCG_REG_R0);
1437        } else {
1438            tcg_out_ldrd_rwb(s, COND_AL, TCG_REG_R2, TCG_REG_R1, TCG_REG_R0);
1439        }
1440    } else {
1441        tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1442                        TCG_REG_R1, TCG_REG_R1, TCG_REG_R0, 0);
1443        if (s->addr_type == TCG_TYPE_I32) {
1444            tcg_out_ld32_12(s, COND_AL, TCG_REG_R2, TCG_REG_R1, cmp_off);
1445        } else {
1446            tcg_out_ldrd_8(s, COND_AL, TCG_REG_R2, TCG_REG_R1, cmp_off);
1447        }
1448    }
1449
1450    /* Load the tlb addend.  */
1451    tcg_out_ld32_12(s, COND_AL, TCG_REG_R1, TCG_REG_R1,
1452                    offsetof(CPUTLBEntry, addend));
1453
1454    /*
1455     * Check alignment, check comparators.
1456     * Do this in 2-4 insns.  Use MOVW for v7, if possible,
1457     * to reduce the number of sequential conditional instructions.
1458     * Almost all guests have at least 4k pages, which means that we need
1459     * to clear at least 9 bits even for an 8-byte memory, which means it
1460     * isn't worth checking for an immediate operand for BIC.
1461     *
1462     * For unaligned accesses, test the page of the last unit of alignment.
1463     * This leaves the least significant alignment bits unchanged, and of
1464     * course must be zero.
1465     */
1466    t_addr = addrlo;
1467    if (a_mask < s_mask) {
1468        t_addr = TCG_REG_R0;
1469        tcg_out_dat_imm(s, COND_AL, ARITH_ADD, t_addr,
1470                        addrlo, s_mask - a_mask);
1471    }
1472    if (use_armv7_instructions && s->page_bits <= 16) {
1473        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, ~(s->page_mask | a_mask));
1474        tcg_out_dat_reg(s, COND_AL, ARITH_BIC, TCG_REG_TMP,
1475                        t_addr, TCG_REG_TMP, 0);
1476        tcg_out_dat_reg(s, COND_AL, ARITH_CMP, 0, TCG_REG_R2, TCG_REG_TMP, 0);
1477    } else {
1478        if (a_mask) {
1479            tcg_debug_assert(a_mask <= 0xff);
1480            tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addrlo, a_mask);
1481        }
1482        tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0, t_addr,
1483                        SHIFT_IMM_LSR(s->page_bits));
1484        tcg_out_dat_reg(s, (a_mask ? COND_EQ : COND_AL), ARITH_CMP,
1485                        0, TCG_REG_R2, TCG_REG_TMP,
1486                        SHIFT_IMM_LSL(s->page_bits));
1487    }
1488
1489    if (s->addr_type != TCG_TYPE_I32) {
1490        tcg_out_dat_reg(s, COND_EQ, ARITH_CMP, 0, TCG_REG_R3, addrhi, 0);
1491    }
1492#else
1493    if (a_mask) {
1494        ldst = new_ldst_label(s);
1495        ldst->is_ld = is_ld;
1496        ldst->oi = oi;
1497        ldst->addrlo_reg = addrlo;
1498        ldst->addrhi_reg = addrhi;
1499
1500        /* We are expecting alignment to max out at 7 */
1501        tcg_debug_assert(a_mask <= 0xff);
1502        /* tst addr, #mask */
1503        tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addrlo, a_mask);
1504    }
1505#endif
1506
1507    return ldst;
1508}
1509
1510static void tcg_out_qemu_ld_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1511                                   TCGReg datahi, HostAddress h)
1512{
1513    TCGReg base;
1514
1515    /* Byte swapping is left to middle-end expansion. */
1516    tcg_debug_assert((opc & MO_BSWAP) == 0);
1517
1518    switch (opc & MO_SSIZE) {
1519    case MO_UB:
1520        if (h.index < 0) {
1521            tcg_out_ld8_12(s, h.cond, datalo, h.base, 0);
1522        } else {
1523            tcg_out_ld8_r(s, h.cond, datalo, h.base, h.index);
1524        }
1525        break;
1526    case MO_SB:
1527        if (h.index < 0) {
1528            tcg_out_ld8s_8(s, h.cond, datalo, h.base, 0);
1529        } else {
1530            tcg_out_ld8s_r(s, h.cond, datalo, h.base, h.index);
1531        }
1532        break;
1533    case MO_UW:
1534        if (h.index < 0) {
1535            tcg_out_ld16u_8(s, h.cond, datalo, h.base, 0);
1536        } else {
1537            tcg_out_ld16u_r(s, h.cond, datalo, h.base, h.index);
1538        }
1539        break;
1540    case MO_SW:
1541        if (h.index < 0) {
1542            tcg_out_ld16s_8(s, h.cond, datalo, h.base, 0);
1543        } else {
1544            tcg_out_ld16s_r(s, h.cond, datalo, h.base, h.index);
1545        }
1546        break;
1547    case MO_UL:
1548        if (h.index < 0) {
1549            tcg_out_ld32_12(s, h.cond, datalo, h.base, 0);
1550        } else {
1551            tcg_out_ld32_r(s, h.cond, datalo, h.base, h.index);
1552        }
1553        break;
1554    case MO_UQ:
1555        /* We used pair allocation for datalo, so already should be aligned. */
1556        tcg_debug_assert((datalo & 1) == 0);
1557        tcg_debug_assert(datahi == datalo + 1);
1558        /* LDRD requires alignment; double-check that. */
1559        if (get_alignment_bits(opc) >= MO_64) {
1560            if (h.index < 0) {
1561                tcg_out_ldrd_8(s, h.cond, datalo, h.base, 0);
1562                break;
1563            }
1564            /*
1565             * Rm (the second address op) must not overlap Rt or Rt + 1.
1566             * Since datalo is aligned, we can simplify the test via alignment.
1567             * Flip the two address arguments if that works.
1568             */
1569            if ((h.index & ~1) != datalo) {
1570                tcg_out_ldrd_r(s, h.cond, datalo, h.base, h.index);
1571                break;
1572            }
1573            if ((h.base & ~1) != datalo) {
1574                tcg_out_ldrd_r(s, h.cond, datalo, h.index, h.base);
1575                break;
1576            }
1577        }
1578        if (h.index < 0) {
1579            base = h.base;
1580            if (datalo == h.base) {
1581                tcg_out_mov_reg(s, h.cond, TCG_REG_TMP, base);
1582                base = TCG_REG_TMP;
1583            }
1584        } else if (h.index_scratch) {
1585            tcg_out_ld32_rwb(s, h.cond, datalo, h.index, h.base);
1586            tcg_out_ld32_12(s, h.cond, datahi, h.index, 4);
1587            break;
1588        } else {
1589            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1590                            h.base, h.index, SHIFT_IMM_LSL(0));
1591            base = TCG_REG_TMP;
1592        }
1593        tcg_out_ld32_12(s, h.cond, datalo, base, 0);
1594        tcg_out_ld32_12(s, h.cond, datahi, base, 4);
1595        break;
1596    default:
1597        g_assert_not_reached();
1598    }
1599}
1600
1601static void tcg_out_qemu_ld(TCGContext *s, TCGReg datalo, TCGReg datahi,
1602                            TCGReg addrlo, TCGReg addrhi,
1603                            MemOpIdx oi, TCGType data_type)
1604{
1605    MemOp opc = get_memop(oi);
1606    TCGLabelQemuLdst *ldst;
1607    HostAddress h;
1608
1609    ldst = prepare_host_addr(s, &h, addrlo, addrhi, oi, true);
1610    if (ldst) {
1611        ldst->type = data_type;
1612        ldst->datalo_reg = datalo;
1613        ldst->datahi_reg = datahi;
1614
1615        /*
1616         * This a conditional BL only to load a pointer within this
1617         * opcode into LR for the slow path.  We will not be using
1618         * the value for a tail call.
1619         */
1620        ldst->label_ptr[0] = s->code_ptr;
1621        tcg_out_bl_imm(s, COND_NE, 0);
1622
1623        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1624        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1625    } else {
1626        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1627    }
1628}
1629
1630static void tcg_out_qemu_st_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1631                                   TCGReg datahi, HostAddress h)
1632{
1633    /* Byte swapping is left to middle-end expansion. */
1634    tcg_debug_assert((opc & MO_BSWAP) == 0);
1635
1636    switch (opc & MO_SIZE) {
1637    case MO_8:
1638        if (h.index < 0) {
1639            tcg_out_st8_12(s, h.cond, datalo, h.base, 0);
1640        } else {
1641            tcg_out_st8_r(s, h.cond, datalo, h.base, h.index);
1642        }
1643        break;
1644    case MO_16:
1645        if (h.index < 0) {
1646            tcg_out_st16_8(s, h.cond, datalo, h.base, 0);
1647        } else {
1648            tcg_out_st16_r(s, h.cond, datalo, h.base, h.index);
1649        }
1650        break;
1651    case MO_32:
1652        if (h.index < 0) {
1653            tcg_out_st32_12(s, h.cond, datalo, h.base, 0);
1654        } else {
1655            tcg_out_st32_r(s, h.cond, datalo, h.base, h.index);
1656        }
1657        break;
1658    case MO_64:
1659        /* We used pair allocation for datalo, so already should be aligned. */
1660        tcg_debug_assert((datalo & 1) == 0);
1661        tcg_debug_assert(datahi == datalo + 1);
1662        /* STRD requires alignment; double-check that. */
1663        if (get_alignment_bits(opc) >= MO_64) {
1664            if (h.index < 0) {
1665                tcg_out_strd_8(s, h.cond, datalo, h.base, 0);
1666            } else {
1667                tcg_out_strd_r(s, h.cond, datalo, h.base, h.index);
1668            }
1669        } else if (h.index_scratch) {
1670            tcg_out_st32_rwb(s, h.cond, datalo, h.index, h.base);
1671            tcg_out_st32_12(s, h.cond, datahi, h.index, 4);
1672        } else {
1673            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1674                            h.base, h.index, SHIFT_IMM_LSL(0));
1675            tcg_out_st32_12(s, h.cond, datalo, TCG_REG_TMP, 0);
1676            tcg_out_st32_12(s, h.cond, datahi, TCG_REG_TMP, 4);
1677        }
1678        break;
1679    default:
1680        g_assert_not_reached();
1681    }
1682}
1683
1684static void tcg_out_qemu_st(TCGContext *s, TCGReg datalo, TCGReg datahi,
1685                            TCGReg addrlo, TCGReg addrhi,
1686                            MemOpIdx oi, TCGType data_type)
1687{
1688    MemOp opc = get_memop(oi);
1689    TCGLabelQemuLdst *ldst;
1690    HostAddress h;
1691
1692    ldst = prepare_host_addr(s, &h, addrlo, addrhi, oi, false);
1693    if (ldst) {
1694        ldst->type = data_type;
1695        ldst->datalo_reg = datalo;
1696        ldst->datahi_reg = datahi;
1697
1698        h.cond = COND_EQ;
1699        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1700
1701        /* The conditional call is last, as we're going to return here. */
1702        ldst->label_ptr[0] = s->code_ptr;
1703        tcg_out_bl_imm(s, COND_NE, 0);
1704        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1705    } else {
1706        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1707    }
1708}
1709
1710static void tcg_out_epilogue(TCGContext *s);
1711
1712static void tcg_out_exit_tb(TCGContext *s, uintptr_t arg)
1713{
1714    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, arg);
1715    tcg_out_epilogue(s);
1716}
1717
1718static void tcg_out_goto_tb(TCGContext *s, int which)
1719{
1720    uintptr_t i_addr;
1721    intptr_t i_disp;
1722
1723    /* Direct branch will be patched by tb_target_set_jmp_target. */
1724    set_jmp_insn_offset(s, which);
1725    tcg_out32(s, INSN_NOP);
1726
1727    /* When branch is out of range, fall through to indirect. */
1728    i_addr = get_jmp_target_addr(s, which);
1729    i_disp = tcg_pcrel_diff(s, (void *)i_addr) - 8;
1730    tcg_debug_assert(i_disp < 0);
1731    if (i_disp >= -0xfff) {
1732        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_PC, i_disp);
1733    } else {
1734        /*
1735         * The TB is close, but outside the 12 bits addressable by
1736         * the load.  We can extend this to 20 bits with a sub of a
1737         * shifted immediate from pc.
1738         */
1739        int h = -i_disp;
1740        int l = h & 0xfff;
1741
1742        h = encode_imm_nofail(h - l);
1743        tcg_out_dat_imm(s, COND_AL, ARITH_SUB, TCG_REG_R0, TCG_REG_PC, h);
1744        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_R0, l);
1745    }
1746    set_jmp_reset_offset(s, which);
1747}
1748
1749void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
1750                              uintptr_t jmp_rx, uintptr_t jmp_rw)
1751{
1752    uintptr_t addr = tb->jmp_target_addr[n];
1753    ptrdiff_t offset = addr - (jmp_rx + 8);
1754    tcg_insn_unit insn;
1755
1756    /* Either directly branch, or fall through to indirect branch. */
1757    if (offset == sextract64(offset, 0, 26)) {
1758        /* B <addr> */
1759        insn = deposit32((COND_AL << 28) | INSN_B, 0, 24, offset >> 2);
1760    } else {
1761        insn = INSN_NOP;
1762    }
1763
1764    qatomic_set((uint32_t *)jmp_rw, insn);
1765    flush_idcache_range(jmp_rx, jmp_rw, 4);
1766}
1767
1768static void tcg_out_op(TCGContext *s, TCGOpcode opc,
1769                       const TCGArg args[TCG_MAX_OP_ARGS],
1770                       const int const_args[TCG_MAX_OP_ARGS])
1771{
1772    TCGArg a0, a1, a2, a3, a4, a5;
1773    int c;
1774
1775    switch (opc) {
1776    case INDEX_op_goto_ptr:
1777        tcg_out_b_reg(s, COND_AL, args[0]);
1778        break;
1779    case INDEX_op_br:
1780        tcg_out_goto_label(s, COND_AL, arg_label(args[0]));
1781        break;
1782
1783    case INDEX_op_ld8u_i32:
1784        tcg_out_ld8u(s, COND_AL, args[0], args[1], args[2]);
1785        break;
1786    case INDEX_op_ld8s_i32:
1787        tcg_out_ld8s(s, COND_AL, args[0], args[1], args[2]);
1788        break;
1789    case INDEX_op_ld16u_i32:
1790        tcg_out_ld16u(s, COND_AL, args[0], args[1], args[2]);
1791        break;
1792    case INDEX_op_ld16s_i32:
1793        tcg_out_ld16s(s, COND_AL, args[0], args[1], args[2]);
1794        break;
1795    case INDEX_op_ld_i32:
1796        tcg_out_ld32u(s, COND_AL, args[0], args[1], args[2]);
1797        break;
1798    case INDEX_op_st8_i32:
1799        tcg_out_st8(s, COND_AL, args[0], args[1], args[2]);
1800        break;
1801    case INDEX_op_st16_i32:
1802        tcg_out_st16(s, COND_AL, args[0], args[1], args[2]);
1803        break;
1804    case INDEX_op_st_i32:
1805        tcg_out_st32(s, COND_AL, args[0], args[1], args[2]);
1806        break;
1807
1808    case INDEX_op_movcond_i32:
1809        /* Constraints mean that v2 is always in the same register as dest,
1810         * so we only need to do "if condition passed, move v1 to dest".
1811         */
1812        tcg_out_dat_rIN(s, COND_AL, ARITH_CMP, ARITH_CMN, 0,
1813                        args[1], args[2], const_args[2]);
1814        tcg_out_dat_rIK(s, tcg_cond_to_arm_cond[args[5]], ARITH_MOV,
1815                        ARITH_MVN, args[0], 0, args[3], const_args[3]);
1816        break;
1817    case INDEX_op_add_i32:
1818        tcg_out_dat_rIN(s, COND_AL, ARITH_ADD, ARITH_SUB,
1819                        args[0], args[1], args[2], const_args[2]);
1820        break;
1821    case INDEX_op_sub_i32:
1822        if (const_args[1]) {
1823            if (const_args[2]) {
1824                tcg_out_movi32(s, COND_AL, args[0], args[1] - args[2]);
1825            } else {
1826                tcg_out_dat_rI(s, COND_AL, ARITH_RSB,
1827                               args[0], args[2], args[1], 1);
1828            }
1829        } else {
1830            tcg_out_dat_rIN(s, COND_AL, ARITH_SUB, ARITH_ADD,
1831                            args[0], args[1], args[2], const_args[2]);
1832        }
1833        break;
1834    case INDEX_op_and_i32:
1835        tcg_out_dat_rIK(s, COND_AL, ARITH_AND, ARITH_BIC,
1836                        args[0], args[1], args[2], const_args[2]);
1837        break;
1838    case INDEX_op_andc_i32:
1839        tcg_out_dat_rIK(s, COND_AL, ARITH_BIC, ARITH_AND,
1840                        args[0], args[1], args[2], const_args[2]);
1841        break;
1842    case INDEX_op_or_i32:
1843        c = ARITH_ORR;
1844        goto gen_arith;
1845    case INDEX_op_xor_i32:
1846        c = ARITH_EOR;
1847        /* Fall through.  */
1848    gen_arith:
1849        tcg_out_dat_rI(s, COND_AL, c, args[0], args[1], args[2], const_args[2]);
1850        break;
1851    case INDEX_op_add2_i32:
1852        a0 = args[0], a1 = args[1], a2 = args[2];
1853        a3 = args[3], a4 = args[4], a5 = args[5];
1854        if (a0 == a3 || (a0 == a5 && !const_args[5])) {
1855            a0 = TCG_REG_TMP;
1856        }
1857        tcg_out_dat_rIN(s, COND_AL, ARITH_ADD | TO_CPSR, ARITH_SUB | TO_CPSR,
1858                        a0, a2, a4, const_args[4]);
1859        tcg_out_dat_rIK(s, COND_AL, ARITH_ADC, ARITH_SBC,
1860                        a1, a3, a5, const_args[5]);
1861        tcg_out_mov_reg(s, COND_AL, args[0], a0);
1862        break;
1863    case INDEX_op_sub2_i32:
1864        a0 = args[0], a1 = args[1], a2 = args[2];
1865        a3 = args[3], a4 = args[4], a5 = args[5];
1866        if ((a0 == a3 && !const_args[3]) || (a0 == a5 && !const_args[5])) {
1867            a0 = TCG_REG_TMP;
1868        }
1869        if (const_args[2]) {
1870            if (const_args[4]) {
1871                tcg_out_movi32(s, COND_AL, a0, a4);
1872                a4 = a0;
1873            }
1874            tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR, a0, a4, a2, 1);
1875        } else {
1876            tcg_out_dat_rIN(s, COND_AL, ARITH_SUB | TO_CPSR,
1877                            ARITH_ADD | TO_CPSR, a0, a2, a4, const_args[4]);
1878        }
1879        if (const_args[3]) {
1880            if (const_args[5]) {
1881                tcg_out_movi32(s, COND_AL, a1, a5);
1882                a5 = a1;
1883            }
1884            tcg_out_dat_rI(s, COND_AL, ARITH_RSC, a1, a5, a3, 1);
1885        } else {
1886            tcg_out_dat_rIK(s, COND_AL, ARITH_SBC, ARITH_ADC,
1887                            a1, a3, a5, const_args[5]);
1888        }
1889        tcg_out_mov_reg(s, COND_AL, args[0], a0);
1890        break;
1891    case INDEX_op_neg_i32:
1892        tcg_out_dat_imm(s, COND_AL, ARITH_RSB, args[0], args[1], 0);
1893        break;
1894    case INDEX_op_not_i32:
1895        tcg_out_dat_reg(s, COND_AL,
1896                        ARITH_MVN, args[0], 0, args[1], SHIFT_IMM_LSL(0));
1897        break;
1898    case INDEX_op_mul_i32:
1899        tcg_out_mul32(s, COND_AL, args[0], args[1], args[2]);
1900        break;
1901    case INDEX_op_mulu2_i32:
1902        tcg_out_umull32(s, COND_AL, args[0], args[1], args[2], args[3]);
1903        break;
1904    case INDEX_op_muls2_i32:
1905        tcg_out_smull32(s, COND_AL, args[0], args[1], args[2], args[3]);
1906        break;
1907    /* XXX: Perhaps args[2] & 0x1f is wrong */
1908    case INDEX_op_shl_i32:
1909        c = const_args[2] ?
1910                SHIFT_IMM_LSL(args[2] & 0x1f) : SHIFT_REG_LSL(args[2]);
1911        goto gen_shift32;
1912    case INDEX_op_shr_i32:
1913        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_LSR(args[2] & 0x1f) :
1914                SHIFT_IMM_LSL(0) : SHIFT_REG_LSR(args[2]);
1915        goto gen_shift32;
1916    case INDEX_op_sar_i32:
1917        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_ASR(args[2] & 0x1f) :
1918                SHIFT_IMM_LSL(0) : SHIFT_REG_ASR(args[2]);
1919        goto gen_shift32;
1920    case INDEX_op_rotr_i32:
1921        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_ROR(args[2] & 0x1f) :
1922                SHIFT_IMM_LSL(0) : SHIFT_REG_ROR(args[2]);
1923        /* Fall through.  */
1924    gen_shift32:
1925        tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1], c);
1926        break;
1927
1928    case INDEX_op_rotl_i32:
1929        if (const_args[2]) {
1930            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1],
1931                            ((0x20 - args[2]) & 0x1f) ?
1932                            SHIFT_IMM_ROR((0x20 - args[2]) & 0x1f) :
1933                            SHIFT_IMM_LSL(0));
1934        } else {
1935            tcg_out_dat_imm(s, COND_AL, ARITH_RSB, TCG_REG_TMP, args[2], 0x20);
1936            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1],
1937                            SHIFT_REG_ROR(TCG_REG_TMP));
1938        }
1939        break;
1940
1941    case INDEX_op_ctz_i32:
1942        tcg_out_dat_reg(s, COND_AL, INSN_RBIT, TCG_REG_TMP, 0, args[1], 0);
1943        a1 = TCG_REG_TMP;
1944        goto do_clz;
1945
1946    case INDEX_op_clz_i32:
1947        a1 = args[1];
1948    do_clz:
1949        a0 = args[0];
1950        a2 = args[2];
1951        c = const_args[2];
1952        if (c && a2 == 32) {
1953            tcg_out_dat_reg(s, COND_AL, INSN_CLZ, a0, 0, a1, 0);
1954            break;
1955        }
1956        tcg_out_dat_imm(s, COND_AL, ARITH_CMP, 0, a1, 0);
1957        tcg_out_dat_reg(s, COND_NE, INSN_CLZ, a0, 0, a1, 0);
1958        if (c || a0 != a2) {
1959            tcg_out_dat_rIK(s, COND_EQ, ARITH_MOV, ARITH_MVN, a0, 0, a2, c);
1960        }
1961        break;
1962
1963    case INDEX_op_brcond_i32:
1964        tcg_out_dat_rIN(s, COND_AL, ARITH_CMP, ARITH_CMN, 0,
1965                       args[0], args[1], const_args[1]);
1966        tcg_out_goto_label(s, tcg_cond_to_arm_cond[args[2]],
1967                           arg_label(args[3]));
1968        break;
1969    case INDEX_op_setcond_i32:
1970        tcg_out_dat_rIN(s, COND_AL, ARITH_CMP, ARITH_CMN, 0,
1971                        args[1], args[2], const_args[2]);
1972        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[args[3]],
1973                        ARITH_MOV, args[0], 0, 1);
1974        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(args[3])],
1975                        ARITH_MOV, args[0], 0, 0);
1976        break;
1977
1978    case INDEX_op_brcond2_i32:
1979        c = tcg_out_cmp2(s, args, const_args);
1980        tcg_out_goto_label(s, tcg_cond_to_arm_cond[c], arg_label(args[5]));
1981        break;
1982    case INDEX_op_setcond2_i32:
1983        c = tcg_out_cmp2(s, args + 1, const_args + 1);
1984        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c], ARITH_MOV, args[0], 0, 1);
1985        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
1986                        ARITH_MOV, args[0], 0, 0);
1987        break;
1988
1989    case INDEX_op_qemu_ld_a32_i32:
1990        tcg_out_qemu_ld(s, args[0], -1, args[1], -1, args[2], TCG_TYPE_I32);
1991        break;
1992    case INDEX_op_qemu_ld_a64_i32:
1993        tcg_out_qemu_ld(s, args[0], -1, args[1], args[2],
1994                        args[3], TCG_TYPE_I32);
1995        break;
1996    case INDEX_op_qemu_ld_a32_i64:
1997        tcg_out_qemu_ld(s, args[0], args[1], args[2], -1,
1998                        args[3], TCG_TYPE_I64);
1999        break;
2000    case INDEX_op_qemu_ld_a64_i64:
2001        tcg_out_qemu_ld(s, args[0], args[1], args[2], args[3],
2002                        args[4], TCG_TYPE_I64);
2003        break;
2004
2005    case INDEX_op_qemu_st_a32_i32:
2006        tcg_out_qemu_st(s, args[0], -1, args[1], -1, args[2], TCG_TYPE_I32);
2007        break;
2008    case INDEX_op_qemu_st_a64_i32:
2009        tcg_out_qemu_st(s, args[0], -1, args[1], args[2],
2010                        args[3], TCG_TYPE_I32);
2011        break;
2012    case INDEX_op_qemu_st_a32_i64:
2013        tcg_out_qemu_st(s, args[0], args[1], args[2], -1,
2014                        args[3], TCG_TYPE_I64);
2015        break;
2016    case INDEX_op_qemu_st_a64_i64:
2017        tcg_out_qemu_st(s, args[0], args[1], args[2], args[3],
2018                        args[4], TCG_TYPE_I64);
2019        break;
2020
2021    case INDEX_op_bswap16_i32:
2022        tcg_out_bswap16(s, COND_AL, args[0], args[1], args[2]);
2023        break;
2024    case INDEX_op_bswap32_i32:
2025        tcg_out_bswap32(s, COND_AL, args[0], args[1]);
2026        break;
2027
2028    case INDEX_op_deposit_i32:
2029        tcg_out_deposit(s, COND_AL, args[0], args[2],
2030                        args[3], args[4], const_args[2]);
2031        break;
2032    case INDEX_op_extract_i32:
2033        tcg_out_extract(s, COND_AL, args[0], args[1], args[2], args[3]);
2034        break;
2035    case INDEX_op_sextract_i32:
2036        tcg_out_sextract(s, COND_AL, args[0], args[1], args[2], args[3]);
2037        break;
2038    case INDEX_op_extract2_i32:
2039        /* ??? These optimization vs zero should be generic.  */
2040        /* ??? But we can't substitute 2 for 1 in the opcode stream yet.  */
2041        if (const_args[1]) {
2042            if (const_args[2]) {
2043                tcg_out_movi(s, TCG_TYPE_REG, args[0], 0);
2044            } else {
2045                tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2046                                args[2], SHIFT_IMM_LSL(32 - args[3]));
2047            }
2048        } else if (const_args[2]) {
2049            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2050                            args[1], SHIFT_IMM_LSR(args[3]));
2051        } else {
2052            /* We can do extract2 in 2 insns, vs the 3 required otherwise.  */
2053            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0,
2054                            args[2], SHIFT_IMM_LSL(32 - args[3]));
2055            tcg_out_dat_reg(s, COND_AL, ARITH_ORR, args[0], TCG_REG_TMP,
2056                            args[1], SHIFT_IMM_LSR(args[3]));
2057        }
2058        break;
2059
2060    case INDEX_op_div_i32:
2061        tcg_out_sdiv(s, COND_AL, args[0], args[1], args[2]);
2062        break;
2063    case INDEX_op_divu_i32:
2064        tcg_out_udiv(s, COND_AL, args[0], args[1], args[2]);
2065        break;
2066
2067    case INDEX_op_mb:
2068        tcg_out_mb(s, args[0]);
2069        break;
2070
2071    case INDEX_op_mov_i32:  /* Always emitted via tcg_out_mov.  */
2072    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
2073    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
2074    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
2075    case INDEX_op_ext8s_i32:  /* Always emitted via tcg_reg_alloc_op.  */
2076    case INDEX_op_ext8u_i32:
2077    case INDEX_op_ext16s_i32:
2078    case INDEX_op_ext16u_i32:
2079    default:
2080        g_assert_not_reached();
2081    }
2082}
2083
2084static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
2085{
2086    switch (op) {
2087    case INDEX_op_goto_ptr:
2088        return C_O0_I1(r);
2089
2090    case INDEX_op_ld8u_i32:
2091    case INDEX_op_ld8s_i32:
2092    case INDEX_op_ld16u_i32:
2093    case INDEX_op_ld16s_i32:
2094    case INDEX_op_ld_i32:
2095    case INDEX_op_neg_i32:
2096    case INDEX_op_not_i32:
2097    case INDEX_op_bswap16_i32:
2098    case INDEX_op_bswap32_i32:
2099    case INDEX_op_ext8s_i32:
2100    case INDEX_op_ext16s_i32:
2101    case INDEX_op_ext16u_i32:
2102    case INDEX_op_extract_i32:
2103    case INDEX_op_sextract_i32:
2104        return C_O1_I1(r, r);
2105
2106    case INDEX_op_st8_i32:
2107    case INDEX_op_st16_i32:
2108    case INDEX_op_st_i32:
2109        return C_O0_I2(r, r);
2110
2111    case INDEX_op_add_i32:
2112    case INDEX_op_sub_i32:
2113    case INDEX_op_setcond_i32:
2114        return C_O1_I2(r, r, rIN);
2115
2116    case INDEX_op_and_i32:
2117    case INDEX_op_andc_i32:
2118    case INDEX_op_clz_i32:
2119    case INDEX_op_ctz_i32:
2120        return C_O1_I2(r, r, rIK);
2121
2122    case INDEX_op_mul_i32:
2123    case INDEX_op_div_i32:
2124    case INDEX_op_divu_i32:
2125        return C_O1_I2(r, r, r);
2126
2127    case INDEX_op_mulu2_i32:
2128    case INDEX_op_muls2_i32:
2129        return C_O2_I2(r, r, r, r);
2130
2131    case INDEX_op_or_i32:
2132    case INDEX_op_xor_i32:
2133        return C_O1_I2(r, r, rI);
2134
2135    case INDEX_op_shl_i32:
2136    case INDEX_op_shr_i32:
2137    case INDEX_op_sar_i32:
2138    case INDEX_op_rotl_i32:
2139    case INDEX_op_rotr_i32:
2140        return C_O1_I2(r, r, ri);
2141
2142    case INDEX_op_brcond_i32:
2143        return C_O0_I2(r, rIN);
2144    case INDEX_op_deposit_i32:
2145        return C_O1_I2(r, 0, rZ);
2146    case INDEX_op_extract2_i32:
2147        return C_O1_I2(r, rZ, rZ);
2148    case INDEX_op_movcond_i32:
2149        return C_O1_I4(r, r, rIN, rIK, 0);
2150    case INDEX_op_add2_i32:
2151        return C_O2_I4(r, r, r, r, rIN, rIK);
2152    case INDEX_op_sub2_i32:
2153        return C_O2_I4(r, r, rI, rI, rIN, rIK);
2154    case INDEX_op_brcond2_i32:
2155        return C_O0_I4(r, r, rI, rI);
2156    case INDEX_op_setcond2_i32:
2157        return C_O1_I4(r, r, r, rI, rI);
2158
2159    case INDEX_op_qemu_ld_a32_i32:
2160        return C_O1_I1(r, q);
2161    case INDEX_op_qemu_ld_a64_i32:
2162        return C_O1_I2(r, q, q);
2163    case INDEX_op_qemu_ld_a32_i64:
2164        return C_O2_I1(e, p, q);
2165    case INDEX_op_qemu_ld_a64_i64:
2166        return C_O2_I2(e, p, q, q);
2167    case INDEX_op_qemu_st_a32_i32:
2168        return C_O0_I2(q, q);
2169    case INDEX_op_qemu_st_a64_i32:
2170        return C_O0_I3(q, q, q);
2171    case INDEX_op_qemu_st_a32_i64:
2172        return C_O0_I3(Q, p, q);
2173    case INDEX_op_qemu_st_a64_i64:
2174        return C_O0_I4(Q, p, q, q);
2175
2176    case INDEX_op_st_vec:
2177        return C_O0_I2(w, r);
2178    case INDEX_op_ld_vec:
2179    case INDEX_op_dupm_vec:
2180        return C_O1_I1(w, r);
2181    case INDEX_op_dup_vec:
2182        return C_O1_I1(w, wr);
2183    case INDEX_op_abs_vec:
2184    case INDEX_op_neg_vec:
2185    case INDEX_op_not_vec:
2186    case INDEX_op_shli_vec:
2187    case INDEX_op_shri_vec:
2188    case INDEX_op_sari_vec:
2189        return C_O1_I1(w, w);
2190    case INDEX_op_dup2_vec:
2191    case INDEX_op_add_vec:
2192    case INDEX_op_mul_vec:
2193    case INDEX_op_smax_vec:
2194    case INDEX_op_smin_vec:
2195    case INDEX_op_ssadd_vec:
2196    case INDEX_op_sssub_vec:
2197    case INDEX_op_sub_vec:
2198    case INDEX_op_umax_vec:
2199    case INDEX_op_umin_vec:
2200    case INDEX_op_usadd_vec:
2201    case INDEX_op_ussub_vec:
2202    case INDEX_op_xor_vec:
2203    case INDEX_op_arm_sshl_vec:
2204    case INDEX_op_arm_ushl_vec:
2205        return C_O1_I2(w, w, w);
2206    case INDEX_op_arm_sli_vec:
2207        return C_O1_I2(w, 0, w);
2208    case INDEX_op_or_vec:
2209    case INDEX_op_andc_vec:
2210        return C_O1_I2(w, w, wO);
2211    case INDEX_op_and_vec:
2212    case INDEX_op_orc_vec:
2213        return C_O1_I2(w, w, wV);
2214    case INDEX_op_cmp_vec:
2215        return C_O1_I2(w, w, wZ);
2216    case INDEX_op_bitsel_vec:
2217        return C_O1_I3(w, w, w, w);
2218    default:
2219        g_assert_not_reached();
2220    }
2221}
2222
2223static void tcg_target_init(TCGContext *s)
2224{
2225    /*
2226     * Only probe for the platform and capabilities if we haven't already
2227     * determined maximum values at compile time.
2228     */
2229#if !defined(use_idiv_instructions) || !defined(use_neon_instructions)
2230    {
2231        unsigned long hwcap = qemu_getauxval(AT_HWCAP);
2232#ifndef use_idiv_instructions
2233        use_idiv_instructions = (hwcap & HWCAP_ARM_IDIVA) != 0;
2234#endif
2235#ifndef use_neon_instructions
2236        use_neon_instructions = (hwcap & HWCAP_ARM_NEON) != 0;
2237#endif
2238    }
2239#endif
2240
2241    if (__ARM_ARCH < 7) {
2242        const char *pl = (const char *)qemu_getauxval(AT_PLATFORM);
2243        if (pl != NULL && pl[0] == 'v' && pl[1] >= '4' && pl[1] <= '9') {
2244            arm_arch = pl[1] - '0';
2245        }
2246
2247        if (arm_arch < 6) {
2248            error_report("TCG: ARMv%d is unsupported; exiting", arm_arch);
2249            exit(EXIT_FAILURE);
2250        }
2251    }
2252
2253    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
2254
2255    tcg_target_call_clobber_regs = 0;
2256    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R0);
2257    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R1);
2258    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R2);
2259    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R3);
2260    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R12);
2261    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R14);
2262
2263    if (use_neon_instructions) {
2264        tcg_target_available_regs[TCG_TYPE_V64]  = ALL_VECTOR_REGS;
2265        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
2266
2267        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q0);
2268        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q1);
2269        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q2);
2270        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q3);
2271        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q8);
2272        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q9);
2273        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q10);
2274        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q11);
2275        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q12);
2276        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q13);
2277        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q14);
2278        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q15);
2279    }
2280
2281    s->reserved_regs = 0;
2282    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
2283    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP);
2284    tcg_regset_set_reg(s->reserved_regs, TCG_REG_PC);
2285    tcg_regset_set_reg(s->reserved_regs, TCG_VEC_TMP);
2286}
2287
2288static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg arg,
2289                       TCGReg arg1, intptr_t arg2)
2290{
2291    switch (type) {
2292    case TCG_TYPE_I32:
2293        tcg_out_ld32u(s, COND_AL, arg, arg1, arg2);
2294        return;
2295    case TCG_TYPE_V64:
2296        /* regs 1; size 8; align 8 */
2297        tcg_out_vldst(s, INSN_VLD1 | 0x7d0, arg, arg1, arg2);
2298        return;
2299    case TCG_TYPE_V128:
2300        /*
2301         * We have only 8-byte alignment for the stack per the ABI.
2302         * Rather than dynamically re-align the stack, it's easier
2303         * to simply not request alignment beyond that.  So:
2304         * regs 2; size 8; align 8
2305         */
2306        tcg_out_vldst(s, INSN_VLD1 | 0xad0, arg, arg1, arg2);
2307        return;
2308    default:
2309        g_assert_not_reached();
2310    }
2311}
2312
2313static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
2314                       TCGReg arg1, intptr_t arg2)
2315{
2316    switch (type) {
2317    case TCG_TYPE_I32:
2318        tcg_out_st32(s, COND_AL, arg, arg1, arg2);
2319        return;
2320    case TCG_TYPE_V64:
2321        /* regs 1; size 8; align 8 */
2322        tcg_out_vldst(s, INSN_VST1 | 0x7d0, arg, arg1, arg2);
2323        return;
2324    case TCG_TYPE_V128:
2325        /* See tcg_out_ld re alignment: regs 2; size 8; align 8 */
2326        tcg_out_vldst(s, INSN_VST1 | 0xad0, arg, arg1, arg2);
2327        return;
2328    default:
2329        g_assert_not_reached();
2330    }
2331}
2332
2333static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
2334                        TCGReg base, intptr_t ofs)
2335{
2336    return false;
2337}
2338
2339static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
2340{
2341    if (ret == arg) {
2342        return true;
2343    }
2344    switch (type) {
2345    case TCG_TYPE_I32:
2346        if (ret < TCG_REG_Q0 && arg < TCG_REG_Q0) {
2347            tcg_out_mov_reg(s, COND_AL, ret, arg);
2348            return true;
2349        }
2350        return false;
2351
2352    case TCG_TYPE_V64:
2353    case TCG_TYPE_V128:
2354        /* "VMOV D,N" is an alias for "VORR D,N,N". */
2355        tcg_out_vreg3(s, INSN_VORR, type - TCG_TYPE_V64, 0, ret, arg, arg);
2356        return true;
2357
2358    default:
2359        g_assert_not_reached();
2360    }
2361}
2362
2363static void tcg_out_movi(TCGContext *s, TCGType type,
2364                         TCGReg ret, tcg_target_long arg)
2365{
2366    tcg_debug_assert(type == TCG_TYPE_I32);
2367    tcg_debug_assert(ret < TCG_REG_Q0);
2368    tcg_out_movi32(s, COND_AL, ret, arg);
2369}
2370
2371static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
2372{
2373    return false;
2374}
2375
2376static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
2377                             tcg_target_long imm)
2378{
2379    int enc, opc = ARITH_ADD;
2380
2381    /* All of the easiest immediates to encode are positive. */
2382    if (imm < 0) {
2383        imm = -imm;
2384        opc = ARITH_SUB;
2385    }
2386    enc = encode_imm(imm);
2387    if (enc >= 0) {
2388        tcg_out_dat_imm(s, COND_AL, opc, rd, rs, enc);
2389    } else {
2390        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, imm);
2391        tcg_out_dat_reg(s, COND_AL, opc, rd, rs,
2392                        TCG_REG_TMP, SHIFT_IMM_LSL(0));
2393    }
2394}
2395
2396/* Type is always V128, with I64 elements.  */
2397static void tcg_out_dup2_vec(TCGContext *s, TCGReg rd, TCGReg rl, TCGReg rh)
2398{
2399    /* Move high element into place first. */
2400    /* VMOV Dd+1, Ds */
2401    tcg_out_vreg3(s, INSN_VORR | (1 << 12), 0, 0, rd, rh, rh);
2402    /* Move low element into place; tcg_out_mov will check for nop. */
2403    tcg_out_mov(s, TCG_TYPE_V64, rd, rl);
2404}
2405
2406static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
2407                            TCGReg rd, TCGReg rs)
2408{
2409    int q = type - TCG_TYPE_V64;
2410
2411    if (vece == MO_64) {
2412        if (type == TCG_TYPE_V128) {
2413            tcg_out_dup2_vec(s, rd, rs, rs);
2414        } else {
2415            tcg_out_mov(s, TCG_TYPE_V64, rd, rs);
2416        }
2417    } else if (rs < TCG_REG_Q0) {
2418        int b = (vece == MO_8);
2419        int e = (vece == MO_16);
2420        tcg_out32(s, INSN_VDUP_G | (b << 22) | (q << 21) | (e << 5) |
2421                  encode_vn(rd) | (rs << 12));
2422    } else {
2423        int imm4 = 1 << vece;
2424        tcg_out32(s, INSN_VDUP_S | (imm4 << 16) | (q << 6) |
2425                  encode_vd(rd) | encode_vm(rs));
2426    }
2427    return true;
2428}
2429
2430static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
2431                             TCGReg rd, TCGReg base, intptr_t offset)
2432{
2433    if (vece == MO_64) {
2434        tcg_out_ld(s, TCG_TYPE_V64, rd, base, offset);
2435        if (type == TCG_TYPE_V128) {
2436            tcg_out_dup2_vec(s, rd, rd, rd);
2437        }
2438    } else {
2439        int q = type - TCG_TYPE_V64;
2440        tcg_out_vldst(s, INSN_VLD1R | (vece << 6) | (q << 5),
2441                      rd, base, offset);
2442    }
2443    return true;
2444}
2445
2446static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
2447                             TCGReg rd, int64_t v64)
2448{
2449    int q = type - TCG_TYPE_V64;
2450    int cmode, imm8, i;
2451
2452    /* Test all bytes equal first.  */
2453    if (vece == MO_8) {
2454        tcg_out_vmovi(s, rd, q, 0, 0xe, v64);
2455        return;
2456    }
2457
2458    /*
2459     * Test all bytes 0x00 or 0xff second.  This can match cases that
2460     * might otherwise take 2 or 3 insns for MO_16 or MO_32 below.
2461     */
2462    for (i = imm8 = 0; i < 8; i++) {
2463        uint8_t byte = v64 >> (i * 8);
2464        if (byte == 0xff) {
2465            imm8 |= 1 << i;
2466        } else if (byte != 0) {
2467            goto fail_bytes;
2468        }
2469    }
2470    tcg_out_vmovi(s, rd, q, 1, 0xe, imm8);
2471    return;
2472 fail_bytes:
2473
2474    /*
2475     * Tests for various replications.  For each element width, if we
2476     * cannot find an expansion there's no point checking a larger
2477     * width because we already know by replication it cannot match.
2478     */
2479    if (vece == MO_16) {
2480        uint16_t v16 = v64;
2481
2482        if (is_shimm16(v16, &cmode, &imm8)) {
2483            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2484            return;
2485        }
2486        if (is_shimm16(~v16, &cmode, &imm8)) {
2487            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2488            return;
2489        }
2490
2491        /*
2492         * Otherwise, all remaining constants can be loaded in two insns:
2493         * rd = v16 & 0xff, rd |= v16 & 0xff00.
2494         */
2495        tcg_out_vmovi(s, rd, q, 0, 0x8, v16 & 0xff);
2496        tcg_out_vmovi(s, rd, q, 0, 0xb, v16 >> 8);   /* VORRI */
2497        return;
2498    }
2499
2500    if (vece == MO_32) {
2501        uint32_t v32 = v64;
2502
2503        if (is_shimm32(v32, &cmode, &imm8) ||
2504            is_soimm32(v32, &cmode, &imm8)) {
2505            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2506            return;
2507        }
2508        if (is_shimm32(~v32, &cmode, &imm8) ||
2509            is_soimm32(~v32, &cmode, &imm8)) {
2510            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2511            return;
2512        }
2513
2514        /*
2515         * Restrict the set of constants to those we can load with
2516         * two instructions.  Others we load from the pool.
2517         */
2518        i = is_shimm32_pair(v32, &cmode, &imm8);
2519        if (i) {
2520            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2521            tcg_out_vmovi(s, rd, q, 0, i | 1, extract32(v32, i * 4, 8));
2522            return;
2523        }
2524        i = is_shimm32_pair(~v32, &cmode, &imm8);
2525        if (i) {
2526            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2527            tcg_out_vmovi(s, rd, q, 1, i | 1, extract32(~v32, i * 4, 8));
2528            return;
2529        }
2530    }
2531
2532    /*
2533     * As a last resort, load from the constant pool.
2534     */
2535    if (!q || vece == MO_64) {
2536        new_pool_l2(s, R_ARM_PC11, s->code_ptr, 0, v64, v64 >> 32);
2537        /* VLDR Dd, [pc + offset] */
2538        tcg_out32(s, INSN_VLDR_D | encode_vd(rd) | (0xf << 16));
2539        if (q) {
2540            tcg_out_dup2_vec(s, rd, rd, rd);
2541        }
2542    } else {
2543        new_pool_label(s, (uint32_t)v64, R_ARM_PC8, s->code_ptr, 0);
2544        /* add tmp, pc, offset */
2545        tcg_out_dat_imm(s, COND_AL, ARITH_ADD, TCG_REG_TMP, TCG_REG_PC, 0);
2546        tcg_out_dupm_vec(s, type, MO_32, rd, TCG_REG_TMP, 0);
2547    }
2548}
2549
2550static const ARMInsn vec_cmp_insn[16] = {
2551    [TCG_COND_EQ] = INSN_VCEQ,
2552    [TCG_COND_GT] = INSN_VCGT,
2553    [TCG_COND_GE] = INSN_VCGE,
2554    [TCG_COND_GTU] = INSN_VCGT_U,
2555    [TCG_COND_GEU] = INSN_VCGE_U,
2556};
2557
2558static const ARMInsn vec_cmp0_insn[16] = {
2559    [TCG_COND_EQ] = INSN_VCEQ0,
2560    [TCG_COND_GT] = INSN_VCGT0,
2561    [TCG_COND_GE] = INSN_VCGE0,
2562    [TCG_COND_LT] = INSN_VCLT0,
2563    [TCG_COND_LE] = INSN_VCLE0,
2564};
2565
2566static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
2567                           unsigned vecl, unsigned vece,
2568                           const TCGArg args[TCG_MAX_OP_ARGS],
2569                           const int const_args[TCG_MAX_OP_ARGS])
2570{
2571    TCGType type = vecl + TCG_TYPE_V64;
2572    unsigned q = vecl;
2573    TCGArg a0, a1, a2, a3;
2574    int cmode, imm8;
2575
2576    a0 = args[0];
2577    a1 = args[1];
2578    a2 = args[2];
2579
2580    switch (opc) {
2581    case INDEX_op_ld_vec:
2582        tcg_out_ld(s, type, a0, a1, a2);
2583        return;
2584    case INDEX_op_st_vec:
2585        tcg_out_st(s, type, a0, a1, a2);
2586        return;
2587    case INDEX_op_dupm_vec:
2588        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
2589        return;
2590    case INDEX_op_dup2_vec:
2591        tcg_out_dup2_vec(s, a0, a1, a2);
2592        return;
2593    case INDEX_op_abs_vec:
2594        tcg_out_vreg2(s, INSN_VABS, q, vece, a0, a1);
2595        return;
2596    case INDEX_op_neg_vec:
2597        tcg_out_vreg2(s, INSN_VNEG, q, vece, a0, a1);
2598        return;
2599    case INDEX_op_not_vec:
2600        tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a1);
2601        return;
2602    case INDEX_op_add_vec:
2603        tcg_out_vreg3(s, INSN_VADD, q, vece, a0, a1, a2);
2604        return;
2605    case INDEX_op_mul_vec:
2606        tcg_out_vreg3(s, INSN_VMUL, q, vece, a0, a1, a2);
2607        return;
2608    case INDEX_op_smax_vec:
2609        tcg_out_vreg3(s, INSN_VMAX, q, vece, a0, a1, a2);
2610        return;
2611    case INDEX_op_smin_vec:
2612        tcg_out_vreg3(s, INSN_VMIN, q, vece, a0, a1, a2);
2613        return;
2614    case INDEX_op_sub_vec:
2615        tcg_out_vreg3(s, INSN_VSUB, q, vece, a0, a1, a2);
2616        return;
2617    case INDEX_op_ssadd_vec:
2618        tcg_out_vreg3(s, INSN_VQADD, q, vece, a0, a1, a2);
2619        return;
2620    case INDEX_op_sssub_vec:
2621        tcg_out_vreg3(s, INSN_VQSUB, q, vece, a0, a1, a2);
2622        return;
2623    case INDEX_op_umax_vec:
2624        tcg_out_vreg3(s, INSN_VMAX_U, q, vece, a0, a1, a2);
2625        return;
2626    case INDEX_op_umin_vec:
2627        tcg_out_vreg3(s, INSN_VMIN_U, q, vece, a0, a1, a2);
2628        return;
2629    case INDEX_op_usadd_vec:
2630        tcg_out_vreg3(s, INSN_VQADD_U, q, vece, a0, a1, a2);
2631        return;
2632    case INDEX_op_ussub_vec:
2633        tcg_out_vreg3(s, INSN_VQSUB_U, q, vece, a0, a1, a2);
2634        return;
2635    case INDEX_op_xor_vec:
2636        tcg_out_vreg3(s, INSN_VEOR, q, 0, a0, a1, a2);
2637        return;
2638    case INDEX_op_arm_sshl_vec:
2639        /*
2640         * Note that Vm is the data and Vn is the shift count,
2641         * therefore the arguments appear reversed.
2642         */
2643        tcg_out_vreg3(s, INSN_VSHL_S, q, vece, a0, a2, a1);
2644        return;
2645    case INDEX_op_arm_ushl_vec:
2646        /* See above. */
2647        tcg_out_vreg3(s, INSN_VSHL_U, q, vece, a0, a2, a1);
2648        return;
2649    case INDEX_op_shli_vec:
2650        tcg_out_vshifti(s, INSN_VSHLI, q, a0, a1, a2 + (8 << vece));
2651        return;
2652    case INDEX_op_shri_vec:
2653        tcg_out_vshifti(s, INSN_VSHRI, q, a0, a1, (16 << vece) - a2);
2654        return;
2655    case INDEX_op_sari_vec:
2656        tcg_out_vshifti(s, INSN_VSARI, q, a0, a1, (16 << vece) - a2);
2657        return;
2658    case INDEX_op_arm_sli_vec:
2659        tcg_out_vshifti(s, INSN_VSLI, q, a0, a2, args[3] + (8 << vece));
2660        return;
2661
2662    case INDEX_op_andc_vec:
2663        if (!const_args[2]) {
2664            tcg_out_vreg3(s, INSN_VBIC, q, 0, a0, a1, a2);
2665            return;
2666        }
2667        a2 = ~a2;
2668        /* fall through */
2669    case INDEX_op_and_vec:
2670        if (const_args[2]) {
2671            is_shimm1632(~a2, &cmode, &imm8);
2672            if (a0 == a1) {
2673                tcg_out_vmovi(s, a0, q, 1, cmode | 1, imm8); /* VBICI */
2674                return;
2675            }
2676            tcg_out_vmovi(s, a0, q, 1, cmode, imm8); /* VMVNI */
2677            a2 = a0;
2678        }
2679        tcg_out_vreg3(s, INSN_VAND, q, 0, a0, a1, a2);
2680        return;
2681
2682    case INDEX_op_orc_vec:
2683        if (!const_args[2]) {
2684            tcg_out_vreg3(s, INSN_VORN, q, 0, a0, a1, a2);
2685            return;
2686        }
2687        a2 = ~a2;
2688        /* fall through */
2689    case INDEX_op_or_vec:
2690        if (const_args[2]) {
2691            is_shimm1632(a2, &cmode, &imm8);
2692            if (a0 == a1) {
2693                tcg_out_vmovi(s, a0, q, 0, cmode | 1, imm8); /* VORRI */
2694                return;
2695            }
2696            tcg_out_vmovi(s, a0, q, 0, cmode, imm8); /* VMOVI */
2697            a2 = a0;
2698        }
2699        tcg_out_vreg3(s, INSN_VORR, q, 0, a0, a1, a2);
2700        return;
2701
2702    case INDEX_op_cmp_vec:
2703        {
2704            TCGCond cond = args[3];
2705
2706            if (cond == TCG_COND_NE) {
2707                if (const_args[2]) {
2708                    tcg_out_vreg3(s, INSN_VTST, q, vece, a0, a1, a1);
2709                } else {
2710                    tcg_out_vreg3(s, INSN_VCEQ, q, vece, a0, a1, a2);
2711                    tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a0);
2712                }
2713            } else {
2714                ARMInsn insn;
2715
2716                if (const_args[2]) {
2717                    insn = vec_cmp0_insn[cond];
2718                    if (insn) {
2719                        tcg_out_vreg2(s, insn, q, vece, a0, a1);
2720                        return;
2721                    }
2722                    tcg_out_dupi_vec(s, type, MO_8, TCG_VEC_TMP, 0);
2723                    a2 = TCG_VEC_TMP;
2724                }
2725                insn = vec_cmp_insn[cond];
2726                if (insn == 0) {
2727                    TCGArg t;
2728                    t = a1, a1 = a2, a2 = t;
2729                    cond = tcg_swap_cond(cond);
2730                    insn = vec_cmp_insn[cond];
2731                    tcg_debug_assert(insn != 0);
2732                }
2733                tcg_out_vreg3(s, insn, q, vece, a0, a1, a2);
2734            }
2735        }
2736        return;
2737
2738    case INDEX_op_bitsel_vec:
2739        a3 = args[3];
2740        if (a0 == a3) {
2741            tcg_out_vreg3(s, INSN_VBIT, q, 0, a0, a2, a1);
2742        } else if (a0 == a2) {
2743            tcg_out_vreg3(s, INSN_VBIF, q, 0, a0, a3, a1);
2744        } else {
2745            tcg_out_mov(s, type, a0, a1);
2746            tcg_out_vreg3(s, INSN_VBSL, q, 0, a0, a2, a3);
2747        }
2748        return;
2749
2750    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
2751    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
2752    default:
2753        g_assert_not_reached();
2754    }
2755}
2756
2757int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
2758{
2759    switch (opc) {
2760    case INDEX_op_add_vec:
2761    case INDEX_op_sub_vec:
2762    case INDEX_op_and_vec:
2763    case INDEX_op_andc_vec:
2764    case INDEX_op_or_vec:
2765    case INDEX_op_orc_vec:
2766    case INDEX_op_xor_vec:
2767    case INDEX_op_not_vec:
2768    case INDEX_op_shli_vec:
2769    case INDEX_op_shri_vec:
2770    case INDEX_op_sari_vec:
2771    case INDEX_op_ssadd_vec:
2772    case INDEX_op_sssub_vec:
2773    case INDEX_op_usadd_vec:
2774    case INDEX_op_ussub_vec:
2775    case INDEX_op_bitsel_vec:
2776        return 1;
2777    case INDEX_op_abs_vec:
2778    case INDEX_op_cmp_vec:
2779    case INDEX_op_mul_vec:
2780    case INDEX_op_neg_vec:
2781    case INDEX_op_smax_vec:
2782    case INDEX_op_smin_vec:
2783    case INDEX_op_umax_vec:
2784    case INDEX_op_umin_vec:
2785        return vece < MO_64;
2786    case INDEX_op_shlv_vec:
2787    case INDEX_op_shrv_vec:
2788    case INDEX_op_sarv_vec:
2789    case INDEX_op_rotli_vec:
2790    case INDEX_op_rotlv_vec:
2791    case INDEX_op_rotrv_vec:
2792        return -1;
2793    default:
2794        return 0;
2795    }
2796}
2797
2798void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
2799                       TCGArg a0, ...)
2800{
2801    va_list va;
2802    TCGv_vec v0, v1, v2, t1, t2, c1;
2803    TCGArg a2;
2804
2805    va_start(va, a0);
2806    v0 = temp_tcgv_vec(arg_temp(a0));
2807    v1 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
2808    a2 = va_arg(va, TCGArg);
2809    va_end(va);
2810
2811    switch (opc) {
2812    case INDEX_op_shlv_vec:
2813        /*
2814         * Merely propagate shlv_vec to arm_ushl_vec.
2815         * In this way we don't set TCG_TARGET_HAS_shv_vec
2816         * because everything is done via expansion.
2817         */
2818        v2 = temp_tcgv_vec(arg_temp(a2));
2819        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
2820                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
2821        break;
2822
2823    case INDEX_op_shrv_vec:
2824    case INDEX_op_sarv_vec:
2825        /* Right shifts are negative left shifts for NEON.  */
2826        v2 = temp_tcgv_vec(arg_temp(a2));
2827        t1 = tcg_temp_new_vec(type);
2828        tcg_gen_neg_vec(vece, t1, v2);
2829        if (opc == INDEX_op_shrv_vec) {
2830            opc = INDEX_op_arm_ushl_vec;
2831        } else {
2832            opc = INDEX_op_arm_sshl_vec;
2833        }
2834        vec_gen_3(opc, type, vece, tcgv_vec_arg(v0),
2835                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
2836        tcg_temp_free_vec(t1);
2837        break;
2838
2839    case INDEX_op_rotli_vec:
2840        t1 = tcg_temp_new_vec(type);
2841        tcg_gen_shri_vec(vece, t1, v1, -a2 & ((8 << vece) - 1));
2842        vec_gen_4(INDEX_op_arm_sli_vec, type, vece,
2843                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(v1), a2);
2844        tcg_temp_free_vec(t1);
2845        break;
2846
2847    case INDEX_op_rotlv_vec:
2848        v2 = temp_tcgv_vec(arg_temp(a2));
2849        t1 = tcg_temp_new_vec(type);
2850        c1 = tcg_constant_vec(type, vece, 8 << vece);
2851        tcg_gen_sub_vec(vece, t1, v2, c1);
2852        /* Right shifts are negative left shifts for NEON.  */
2853        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
2854                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
2855        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
2856                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
2857        tcg_gen_or_vec(vece, v0, v0, t1);
2858        tcg_temp_free_vec(t1);
2859        break;
2860
2861    case INDEX_op_rotrv_vec:
2862        v2 = temp_tcgv_vec(arg_temp(a2));
2863        t1 = tcg_temp_new_vec(type);
2864        t2 = tcg_temp_new_vec(type);
2865        c1 = tcg_constant_vec(type, vece, 8 << vece);
2866        tcg_gen_neg_vec(vece, t1, v2);
2867        tcg_gen_sub_vec(vece, t2, c1, v2);
2868        /* Right shifts are negative left shifts for NEON.  */
2869        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
2870                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
2871        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t2),
2872                  tcgv_vec_arg(v1), tcgv_vec_arg(t2));
2873        tcg_gen_or_vec(vece, v0, t1, t2);
2874        tcg_temp_free_vec(t1);
2875        tcg_temp_free_vec(t2);
2876        break;
2877
2878    default:
2879        g_assert_not_reached();
2880    }
2881}
2882
2883static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
2884{
2885    int i;
2886    for (i = 0; i < count; ++i) {
2887        p[i] = INSN_NOP;
2888    }
2889}
2890
2891/* Compute frame size via macros, to share between tcg_target_qemu_prologue
2892   and tcg_register_jit.  */
2893
2894#define PUSH_SIZE  ((11 - 4 + 1 + 1) * sizeof(tcg_target_long))
2895
2896#define FRAME_SIZE \
2897    ((PUSH_SIZE \
2898      + TCG_STATIC_CALL_ARGS_SIZE \
2899      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
2900      + TCG_TARGET_STACK_ALIGN - 1) \
2901     & -TCG_TARGET_STACK_ALIGN)
2902
2903#define STACK_ADDEND  (FRAME_SIZE - PUSH_SIZE)
2904
2905static void tcg_target_qemu_prologue(TCGContext *s)
2906{
2907    /* Calling convention requires us to save r4-r11 and lr.  */
2908    /* stmdb sp!, { r4 - r11, lr } */
2909    tcg_out_ldstm(s, COND_AL, INSN_STMDB, TCG_REG_CALL_STACK,
2910                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
2911                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
2912                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_R14));
2913
2914    /* Reserve callee argument and tcg temp space.  */
2915    tcg_out_dat_rI(s, COND_AL, ARITH_SUB, TCG_REG_CALL_STACK,
2916                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
2917    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
2918                  CPU_TEMP_BUF_NLONGS * sizeof(long));
2919
2920    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
2921
2922#ifndef CONFIG_SOFTMMU
2923    if (guest_base) {
2924        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_GUEST_BASE, guest_base);
2925        tcg_regset_set_reg(s->reserved_regs, TCG_REG_GUEST_BASE);
2926    }
2927#endif
2928
2929    tcg_out_b_reg(s, COND_AL, tcg_target_call_iarg_regs[1]);
2930
2931    /*
2932     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
2933     * and fall through to the rest of the epilogue.
2934     */
2935    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
2936    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, 0);
2937    tcg_out_epilogue(s);
2938}
2939
2940static void tcg_out_epilogue(TCGContext *s)
2941{
2942    /* Release local stack frame.  */
2943    tcg_out_dat_rI(s, COND_AL, ARITH_ADD, TCG_REG_CALL_STACK,
2944                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
2945
2946    /* ldmia sp!, { r4 - r11, pc } */
2947    tcg_out_ldstm(s, COND_AL, INSN_LDMIA, TCG_REG_CALL_STACK,
2948                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
2949                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
2950                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_PC));
2951}
2952
2953typedef struct {
2954    DebugFrameHeader h;
2955    uint8_t fde_def_cfa[4];
2956    uint8_t fde_reg_ofs[18];
2957} DebugFrame;
2958
2959#define ELF_HOST_MACHINE EM_ARM
2960
2961/* We're expecting a 2 byte uleb128 encoded value.  */
2962QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
2963
2964static const DebugFrame debug_frame = {
2965    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
2966    .h.cie.id = -1,
2967    .h.cie.version = 1,
2968    .h.cie.code_align = 1,
2969    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
2970    .h.cie.return_column = 14,
2971
2972    /* Total FDE size does not include the "len" member.  */
2973    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
2974
2975    .fde_def_cfa = {
2976        12, 13,                         /* DW_CFA_def_cfa sp, ... */
2977        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
2978        (FRAME_SIZE >> 7)
2979    },
2980    .fde_reg_ofs = {
2981        /* The following must match the stmdb in the prologue.  */
2982        0x8e, 1,                        /* DW_CFA_offset, lr, -4 */
2983        0x8b, 2,                        /* DW_CFA_offset, r11, -8 */
2984        0x8a, 3,                        /* DW_CFA_offset, r10, -12 */
2985        0x89, 4,                        /* DW_CFA_offset, r9, -16 */
2986        0x88, 5,                        /* DW_CFA_offset, r8, -20 */
2987        0x87, 6,                        /* DW_CFA_offset, r7, -24 */
2988        0x86, 7,                        /* DW_CFA_offset, r6, -28 */
2989        0x85, 8,                        /* DW_CFA_offset, r5, -32 */
2990        0x84, 9,                        /* DW_CFA_offset, r4, -36 */
2991    }
2992};
2993
2994void tcg_register_jit(const void *buf, size_t buf_size)
2995{
2996    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
2997}
2998