xref: /openbmc/qemu/tcg/arm/tcg-target.c.inc (revision d31f1185fb029b44c439a6961a6cb087df6567d9)
1/*
2 * Tiny Code Generator for QEMU
3 *
4 * Copyright (c) 2008 Andrzej Zaborowski
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to deal
8 * in the Software without restriction, including without limitation the rights
9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10 * copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22 * THE SOFTWARE.
23 */
24
25#include "elf.h"
26
27int arm_arch = __ARM_ARCH;
28
29#ifndef use_idiv_instructions
30bool use_idiv_instructions;
31#endif
32#ifndef use_neon_instructions
33bool use_neon_instructions;
34#endif
35
36/* Used for function call generation. */
37#define TCG_TARGET_STACK_ALIGN          8
38#define TCG_TARGET_CALL_STACK_OFFSET    0
39#define TCG_TARGET_CALL_ARG_I32         TCG_CALL_ARG_NORMAL
40#define TCG_TARGET_CALL_ARG_I64         TCG_CALL_ARG_EVEN
41#define TCG_TARGET_CALL_ARG_I128        TCG_CALL_ARG_EVEN
42#define TCG_TARGET_CALL_RET_I128        TCG_CALL_RET_BY_REF
43
44#ifdef CONFIG_DEBUG_TCG
45static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
46    "%r0",  "%r1",  "%r2",  "%r3",  "%r4",  "%r5",  "%r6",  "%r7",
47    "%r8",  "%r9",  "%r10", "%r11", "%r12", "%sp",  "%r14", "%pc",
48    "%q0",  "%q1",  "%q2",  "%q3",  "%q4",  "%q5",  "%q6",  "%q7",
49    "%q8",  "%q9",  "%q10", "%q11", "%q12", "%q13", "%q14", "%q15",
50};
51#endif
52
53static const int tcg_target_reg_alloc_order[] = {
54    TCG_REG_R4,
55    TCG_REG_R5,
56    TCG_REG_R6,
57    TCG_REG_R7,
58    TCG_REG_R8,
59    TCG_REG_R9,
60    TCG_REG_R10,
61    TCG_REG_R11,
62    TCG_REG_R13,
63    TCG_REG_R0,
64    TCG_REG_R1,
65    TCG_REG_R2,
66    TCG_REG_R3,
67    TCG_REG_R12,
68    TCG_REG_R14,
69
70    TCG_REG_Q0,
71    TCG_REG_Q1,
72    TCG_REG_Q2,
73    TCG_REG_Q3,
74    /* Q4 - Q7 are call-saved, and skipped. */
75    TCG_REG_Q8,
76    TCG_REG_Q9,
77    TCG_REG_Q10,
78    TCG_REG_Q11,
79    TCG_REG_Q12,
80    TCG_REG_Q13,
81    TCG_REG_Q14,
82    TCG_REG_Q15,
83};
84
85static const int tcg_target_call_iarg_regs[4] = {
86    TCG_REG_R0, TCG_REG_R1, TCG_REG_R2, TCG_REG_R3
87};
88
89static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
90{
91    tcg_debug_assert(kind == TCG_CALL_RET_NORMAL);
92    tcg_debug_assert(slot >= 0 && slot <= 3);
93    return TCG_REG_R0 + slot;
94}
95
96#define TCG_REG_TMP  TCG_REG_R12
97#define TCG_VEC_TMP  TCG_REG_Q15
98#define TCG_REG_GUEST_BASE  TCG_REG_R11
99
100typedef enum {
101    COND_EQ = 0x0,
102    COND_NE = 0x1,
103    COND_CS = 0x2,	/* Unsigned greater or equal */
104    COND_CC = 0x3,	/* Unsigned less than */
105    COND_MI = 0x4,	/* Negative */
106    COND_PL = 0x5,	/* Zero or greater */
107    COND_VS = 0x6,	/* Overflow */
108    COND_VC = 0x7,	/* No overflow */
109    COND_HI = 0x8,	/* Unsigned greater than */
110    COND_LS = 0x9,	/* Unsigned less or equal */
111    COND_GE = 0xa,
112    COND_LT = 0xb,
113    COND_GT = 0xc,
114    COND_LE = 0xd,
115    COND_AL = 0xe,
116} ARMCond;
117
118#define TO_CPSR (1 << 20)
119
120#define SHIFT_IMM_LSL(im)	(((im) << 7) | 0x00)
121#define SHIFT_IMM_LSR(im)	(((im) << 7) | 0x20)
122#define SHIFT_IMM_ASR(im)	(((im) << 7) | 0x40)
123#define SHIFT_IMM_ROR(im)	(((im) << 7) | 0x60)
124#define SHIFT_REG_LSL(rs)	(((rs) << 8) | 0x10)
125#define SHIFT_REG_LSR(rs)	(((rs) << 8) | 0x30)
126#define SHIFT_REG_ASR(rs)	(((rs) << 8) | 0x50)
127#define SHIFT_REG_ROR(rs)	(((rs) << 8) | 0x70)
128
129typedef enum {
130    ARITH_AND = 0x0 << 21,
131    ARITH_EOR = 0x1 << 21,
132    ARITH_SUB = 0x2 << 21,
133    ARITH_RSB = 0x3 << 21,
134    ARITH_ADD = 0x4 << 21,
135    ARITH_ADC = 0x5 << 21,
136    ARITH_SBC = 0x6 << 21,
137    ARITH_RSC = 0x7 << 21,
138    ARITH_TST = 0x8 << 21 | TO_CPSR,
139    ARITH_CMP = 0xa << 21 | TO_CPSR,
140    ARITH_CMN = 0xb << 21 | TO_CPSR,
141    ARITH_ORR = 0xc << 21,
142    ARITH_MOV = 0xd << 21,
143    ARITH_BIC = 0xe << 21,
144    ARITH_MVN = 0xf << 21,
145
146    INSN_B         = 0x0a000000,
147
148    INSN_CLZ       = 0x016f0f10,
149    INSN_RBIT      = 0x06ff0f30,
150
151    INSN_LDMIA     = 0x08b00000,
152    INSN_STMDB     = 0x09200000,
153
154    INSN_LDR_IMM   = 0x04100000,
155    INSN_LDR_REG   = 0x06100000,
156    INSN_STR_IMM   = 0x04000000,
157    INSN_STR_REG   = 0x06000000,
158
159    INSN_LDRH_IMM  = 0x005000b0,
160    INSN_LDRH_REG  = 0x001000b0,
161    INSN_LDRSH_IMM = 0x005000f0,
162    INSN_LDRSH_REG = 0x001000f0,
163    INSN_STRH_IMM  = 0x004000b0,
164    INSN_STRH_REG  = 0x000000b0,
165
166    INSN_LDRB_IMM  = 0x04500000,
167    INSN_LDRB_REG  = 0x06500000,
168    INSN_LDRSB_IMM = 0x005000d0,
169    INSN_LDRSB_REG = 0x001000d0,
170    INSN_STRB_IMM  = 0x04400000,
171    INSN_STRB_REG  = 0x06400000,
172
173    INSN_LDRD_IMM  = 0x004000d0,
174    INSN_LDRD_REG  = 0x000000d0,
175    INSN_STRD_IMM  = 0x004000f0,
176    INSN_STRD_REG  = 0x000000f0,
177
178    INSN_DMB_ISH   = 0xf57ff05b,
179    INSN_DMB_MCR   = 0xee070fba,
180
181    /* Architected nop introduced in v6k.  */
182    /* ??? This is an MSR (imm) 0,0,0 insn.  Anyone know if this
183       also Just So Happened to do nothing on pre-v6k so that we
184       don't need to conditionalize it?  */
185    INSN_NOP_v6k   = 0xe320f000,
186    /* Otherwise the assembler uses mov r0,r0 */
187    INSN_NOP_v4    = (COND_AL << 28) | ARITH_MOV,
188
189    INSN_VADD      = 0xf2000800,
190    INSN_VAND      = 0xf2000110,
191    INSN_VBIC      = 0xf2100110,
192    INSN_VEOR      = 0xf3000110,
193    INSN_VORN      = 0xf2300110,
194    INSN_VORR      = 0xf2200110,
195    INSN_VSUB      = 0xf3000800,
196    INSN_VMUL      = 0xf2000910,
197    INSN_VQADD     = 0xf2000010,
198    INSN_VQADD_U   = 0xf3000010,
199    INSN_VQSUB     = 0xf2000210,
200    INSN_VQSUB_U   = 0xf3000210,
201    INSN_VMAX      = 0xf2000600,
202    INSN_VMAX_U    = 0xf3000600,
203    INSN_VMIN      = 0xf2000610,
204    INSN_VMIN_U    = 0xf3000610,
205
206    INSN_VABS      = 0xf3b10300,
207    INSN_VMVN      = 0xf3b00580,
208    INSN_VNEG      = 0xf3b10380,
209
210    INSN_VCEQ0     = 0xf3b10100,
211    INSN_VCGT0     = 0xf3b10000,
212    INSN_VCGE0     = 0xf3b10080,
213    INSN_VCLE0     = 0xf3b10180,
214    INSN_VCLT0     = 0xf3b10200,
215
216    INSN_VCEQ      = 0xf3000810,
217    INSN_VCGE      = 0xf2000310,
218    INSN_VCGT      = 0xf2000300,
219    INSN_VCGE_U    = 0xf3000310,
220    INSN_VCGT_U    = 0xf3000300,
221
222    INSN_VSHLI     = 0xf2800510,  /* VSHL (immediate) */
223    INSN_VSARI     = 0xf2800010,  /* VSHR.S */
224    INSN_VSHRI     = 0xf3800010,  /* VSHR.U */
225    INSN_VSLI      = 0xf3800510,
226    INSN_VSHL_S    = 0xf2000400,  /* VSHL.S (register) */
227    INSN_VSHL_U    = 0xf3000400,  /* VSHL.U (register) */
228
229    INSN_VBSL      = 0xf3100110,
230    INSN_VBIT      = 0xf3200110,
231    INSN_VBIF      = 0xf3300110,
232
233    INSN_VTST      = 0xf2000810,
234
235    INSN_VDUP_G    = 0xee800b10,  /* VDUP (ARM core register) */
236    INSN_VDUP_S    = 0xf3b00c00,  /* VDUP (scalar) */
237    INSN_VLDR_D    = 0xed100b00,  /* VLDR.64 */
238    INSN_VLD1      = 0xf4200000,  /* VLD1 (multiple single elements) */
239    INSN_VLD1R     = 0xf4a00c00,  /* VLD1 (single element to all lanes) */
240    INSN_VST1      = 0xf4000000,  /* VST1 (multiple single elements) */
241    INSN_VMOVI     = 0xf2800010,  /* VMOV (immediate) */
242} ARMInsn;
243
244#define INSN_NOP   (use_armv7_instructions ? INSN_NOP_v6k : INSN_NOP_v4)
245
246static const uint8_t tcg_cond_to_arm_cond[] = {
247    [TCG_COND_EQ] = COND_EQ,
248    [TCG_COND_NE] = COND_NE,
249    [TCG_COND_LT] = COND_LT,
250    [TCG_COND_GE] = COND_GE,
251    [TCG_COND_LE] = COND_LE,
252    [TCG_COND_GT] = COND_GT,
253    /* unsigned */
254    [TCG_COND_LTU] = COND_CC,
255    [TCG_COND_GEU] = COND_CS,
256    [TCG_COND_LEU] = COND_LS,
257    [TCG_COND_GTU] = COND_HI,
258};
259
260static int encode_imm(uint32_t imm);
261
262/* TCG private relocation type: add with pc+imm8 */
263#define R_ARM_PC8  11
264
265/* TCG private relocation type: vldr with imm8 << 2 */
266#define R_ARM_PC11 12
267
268static bool reloc_pc24(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
269{
270    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
271    ptrdiff_t offset = (tcg_ptr_byte_diff(target, src_rx) - 8) >> 2;
272
273    if (offset == sextract32(offset, 0, 24)) {
274        *src_rw = deposit32(*src_rw, 0, 24, offset);
275        return true;
276    }
277    return false;
278}
279
280static bool reloc_pc13(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
281{
282    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
283    ptrdiff_t offset = tcg_ptr_byte_diff(target, src_rx) - 8;
284
285    if (offset >= -0xfff && offset <= 0xfff) {
286        tcg_insn_unit insn = *src_rw;
287        bool u = (offset >= 0);
288        if (!u) {
289            offset = -offset;
290        }
291        insn = deposit32(insn, 23, 1, u);
292        insn = deposit32(insn, 0, 12, offset);
293        *src_rw = insn;
294        return true;
295    }
296    return false;
297}
298
299static bool reloc_pc11(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
300{
301    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
302    ptrdiff_t offset = (tcg_ptr_byte_diff(target, src_rx) - 8) / 4;
303
304    if (offset >= -0xff && offset <= 0xff) {
305        tcg_insn_unit insn = *src_rw;
306        bool u = (offset >= 0);
307        if (!u) {
308            offset = -offset;
309        }
310        insn = deposit32(insn, 23, 1, u);
311        insn = deposit32(insn, 0, 8, offset);
312        *src_rw = insn;
313        return true;
314    }
315    return false;
316}
317
318static bool reloc_pc8(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
319{
320    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
321    ptrdiff_t offset = tcg_ptr_byte_diff(target, src_rx) - 8;
322    int imm12 = encode_imm(offset);
323
324    if (imm12 >= 0) {
325        *src_rw = deposit32(*src_rw, 0, 12, imm12);
326        return true;
327    }
328    return false;
329}
330
331static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
332                        intptr_t value, intptr_t addend)
333{
334    tcg_debug_assert(addend == 0);
335    switch (type) {
336    case R_ARM_PC24:
337        return reloc_pc24(code_ptr, (const tcg_insn_unit *)value);
338    case R_ARM_PC13:
339        return reloc_pc13(code_ptr, (const tcg_insn_unit *)value);
340    case R_ARM_PC11:
341        return reloc_pc11(code_ptr, (const tcg_insn_unit *)value);
342    case R_ARM_PC8:
343        return reloc_pc8(code_ptr, (const tcg_insn_unit *)value);
344    default:
345        g_assert_not_reached();
346    }
347}
348
349#define TCG_CT_CONST_ARM  0x100
350#define TCG_CT_CONST_INV  0x200
351#define TCG_CT_CONST_NEG  0x400
352#define TCG_CT_CONST_ZERO 0x800
353#define TCG_CT_CONST_ORRI 0x1000
354#define TCG_CT_CONST_ANDI 0x2000
355
356#define ALL_GENERAL_REGS  0xffffu
357#define ALL_VECTOR_REGS   0xffff0000u
358
359/*
360 * r0-r3 will be overwritten when reading the tlb entry (system-mode only);
361 * r14 will be overwritten by the BLNE branching to the slow path.
362 */
363#define ALL_QLDST_REGS \
364    (ALL_GENERAL_REGS & ~((tcg_use_softmmu ? 0xf : 0) | (1 << TCG_REG_R14)))
365
366/*
367 * ARM immediates for ALU instructions are made of an unsigned 8-bit
368 * right-rotated by an even amount between 0 and 30.
369 *
370 * Return < 0 if @imm cannot be encoded, else the entire imm12 field.
371 */
372static int encode_imm(uint32_t imm)
373{
374    uint32_t rot, imm8;
375
376    /* Simple case, no rotation required. */
377    if ((imm & ~0xff) == 0) {
378        return imm;
379    }
380
381    /* Next, try a simple even shift.  */
382    rot = ctz32(imm) & ~1;
383    imm8 = imm >> rot;
384    rot = 32 - rot;
385    if ((imm8 & ~0xff) == 0) {
386        goto found;
387    }
388
389    /*
390     * Finally, try harder with rotations.
391     * The ctz test above will have taken care of rotates >= 8.
392     */
393    for (rot = 2; rot < 8; rot += 2) {
394        imm8 = rol32(imm, rot);
395        if ((imm8 & ~0xff) == 0) {
396            goto found;
397        }
398    }
399    /* Fail: imm cannot be encoded. */
400    return -1;
401
402 found:
403    /* Note that rot is even, and we discard bit 0 by shifting by 7. */
404    return rot << 7 | imm8;
405}
406
407static int encode_imm_nofail(uint32_t imm)
408{
409    int ret = encode_imm(imm);
410    tcg_debug_assert(ret >= 0);
411    return ret;
412}
413
414static bool check_fit_imm(uint32_t imm)
415{
416    return encode_imm(imm) >= 0;
417}
418
419/* Return true if v16 is a valid 16-bit shifted immediate.  */
420static bool is_shimm16(uint16_t v16, int *cmode, int *imm8)
421{
422    if (v16 == (v16 & 0xff)) {
423        *cmode = 0x8;
424        *imm8 = v16 & 0xff;
425        return true;
426    } else if (v16 == (v16 & 0xff00)) {
427        *cmode = 0xa;
428        *imm8 = v16 >> 8;
429        return true;
430    }
431    return false;
432}
433
434/* Return true if v32 is a valid 32-bit shifted immediate.  */
435static bool is_shimm32(uint32_t v32, int *cmode, int *imm8)
436{
437    if (v32 == (v32 & 0xff)) {
438        *cmode = 0x0;
439        *imm8 = v32 & 0xff;
440        return true;
441    } else if (v32 == (v32 & 0xff00)) {
442        *cmode = 0x2;
443        *imm8 = (v32 >> 8) & 0xff;
444        return true;
445    } else if (v32 == (v32 & 0xff0000)) {
446        *cmode = 0x4;
447        *imm8 = (v32 >> 16) & 0xff;
448        return true;
449    } else if (v32 == (v32 & 0xff000000)) {
450        *cmode = 0x6;
451        *imm8 = v32 >> 24;
452        return true;
453    }
454    return false;
455}
456
457/* Return true if v32 is a valid 32-bit shifting ones immediate.  */
458static bool is_soimm32(uint32_t v32, int *cmode, int *imm8)
459{
460    if ((v32 & 0xffff00ff) == 0xff) {
461        *cmode = 0xc;
462        *imm8 = (v32 >> 8) & 0xff;
463        return true;
464    } else if ((v32 & 0xff00ffff) == 0xffff) {
465        *cmode = 0xd;
466        *imm8 = (v32 >> 16) & 0xff;
467        return true;
468    }
469    return false;
470}
471
472/*
473 * Return non-zero if v32 can be formed by MOVI+ORR.
474 * Place the parameters for MOVI in (cmode, imm8).
475 * Return the cmode for ORR; the imm8 can be had via extraction from v32.
476 */
477static int is_shimm32_pair(uint32_t v32, int *cmode, int *imm8)
478{
479    int i;
480
481    for (i = 6; i > 0; i -= 2) {
482        /* Mask out one byte we can add with ORR.  */
483        uint32_t tmp = v32 & ~(0xffu << (i * 4));
484        if (is_shimm32(tmp, cmode, imm8) ||
485            is_soimm32(tmp, cmode, imm8)) {
486            break;
487        }
488    }
489    return i;
490}
491
492/* Return true if V is a valid 16-bit or 32-bit shifted immediate.  */
493static bool is_shimm1632(uint32_t v32, int *cmode, int *imm8)
494{
495    if (v32 == deposit32(v32, 16, 16, v32)) {
496        return is_shimm16(v32, cmode, imm8);
497    } else {
498        return is_shimm32(v32, cmode, imm8);
499    }
500}
501
502/* Test if a constant matches the constraint.
503 * TODO: define constraints for:
504 *
505 * ldr/str offset:   between -0xfff and 0xfff
506 * ldrh/strh offset: between -0xff and 0xff
507 * mov operand2:     values represented with x << (2 * y), x < 0x100
508 * add, sub, eor...: ditto
509 */
510static bool tcg_target_const_match(int64_t val, int ct,
511                                   TCGType type, TCGCond cond, int vece)
512{
513    if (ct & TCG_CT_CONST) {
514        return 1;
515    } else if ((ct & TCG_CT_CONST_ARM) && check_fit_imm(val)) {
516        return 1;
517    } else if ((ct & TCG_CT_CONST_INV) && check_fit_imm(~val)) {
518        return 1;
519    } else if ((ct & TCG_CT_CONST_NEG) && check_fit_imm(-val)) {
520        return 1;
521    } else if ((ct & TCG_CT_CONST_ZERO) && val == 0) {
522        return 1;
523    }
524
525    switch (ct & (TCG_CT_CONST_ORRI | TCG_CT_CONST_ANDI)) {
526    case 0:
527        break;
528    case TCG_CT_CONST_ANDI:
529        val = ~val;
530        /* fallthru */
531    case TCG_CT_CONST_ORRI:
532        if (val == deposit64(val, 32, 32, val)) {
533            int cmode, imm8;
534            return is_shimm1632(val, &cmode, &imm8);
535        }
536        break;
537    default:
538        /* Both bits should not be set for the same insn.  */
539        g_assert_not_reached();
540    }
541
542    return 0;
543}
544
545static void tcg_out_b_imm(TCGContext *s, ARMCond cond, int32_t offset)
546{
547    tcg_out32(s, (cond << 28) | INSN_B |
548                    (((offset - 8) >> 2) & 0x00ffffff));
549}
550
551static void tcg_out_bl_imm(TCGContext *s, ARMCond cond, int32_t offset)
552{
553    tcg_out32(s, (cond << 28) | 0x0b000000 |
554                    (((offset - 8) >> 2) & 0x00ffffff));
555}
556
557static void tcg_out_blx_reg(TCGContext *s, ARMCond cond, TCGReg rn)
558{
559    tcg_out32(s, (cond << 28) | 0x012fff30 | rn);
560}
561
562static void tcg_out_blx_imm(TCGContext *s, int32_t offset)
563{
564    tcg_out32(s, 0xfa000000 | ((offset & 2) << 23) |
565                (((offset - 8) >> 2) & 0x00ffffff));
566}
567
568static void tcg_out_dat_reg(TCGContext *s, ARMCond cond, ARMInsn opc,
569                            TCGReg rd, TCGReg rn, TCGReg rm, int shift)
570{
571    tcg_out32(s, (cond << 28) | (0 << 25) | opc |
572                    (rn << 16) | (rd << 12) | shift | rm);
573}
574
575static void tcg_out_mov_reg(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rm)
576{
577    /* Simple reg-reg move, optimising out the 'do nothing' case */
578    if (rd != rm) {
579        tcg_out_dat_reg(s, cond, ARITH_MOV, rd, 0, rm, SHIFT_IMM_LSL(0));
580    }
581}
582
583static void tcg_out_bx_reg(TCGContext *s, ARMCond cond, TCGReg rn)
584{
585    tcg_out32(s, (cond << 28) | 0x012fff10 | rn);
586}
587
588static void tcg_out_b_reg(TCGContext *s, ARMCond cond, TCGReg rn)
589{
590    /*
591     * Unless the C portion of QEMU is compiled as thumb, we don't need
592     * true BX semantics; merely a branch to an address held in a register.
593     */
594    tcg_out_bx_reg(s, cond, rn);
595}
596
597static void tcg_out_dat_imm(TCGContext *s, ARMCond cond, ARMInsn opc,
598                            TCGReg rd, TCGReg rn, int im)
599{
600    tcg_out32(s, (cond << 28) | (1 << 25) | opc |
601                    (rn << 16) | (rd << 12) | im);
602}
603
604static void tcg_out_ldstm(TCGContext *s, ARMCond cond, ARMInsn opc,
605                          TCGReg rn, uint16_t mask)
606{
607    tcg_out32(s, (cond << 28) | opc | (rn << 16) | mask);
608}
609
610/* Note that this routine is used for both LDR and LDRH formats, so we do
611   not wish to include an immediate shift at this point.  */
612static void tcg_out_memop_r(TCGContext *s, ARMCond cond, ARMInsn opc, TCGReg rt,
613                            TCGReg rn, TCGReg rm, bool u, bool p, bool w)
614{
615    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24)
616              | (w << 21) | (rn << 16) | (rt << 12) | rm);
617}
618
619static void tcg_out_memop_8(TCGContext *s, ARMCond cond, ARMInsn opc, TCGReg rt,
620                            TCGReg rn, int imm8, bool p, bool w)
621{
622    bool u = 1;
623    if (imm8 < 0) {
624        imm8 = -imm8;
625        u = 0;
626    }
627    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24) | (w << 21) |
628              (rn << 16) | (rt << 12) | ((imm8 & 0xf0) << 4) | (imm8 & 0xf));
629}
630
631static void tcg_out_memop_12(TCGContext *s, ARMCond cond, ARMInsn opc,
632                             TCGReg rt, TCGReg rn, int imm12, bool p, bool w)
633{
634    bool u = 1;
635    if (imm12 < 0) {
636        imm12 = -imm12;
637        u = 0;
638    }
639    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24) | (w << 21) |
640              (rn << 16) | (rt << 12) | imm12);
641}
642
643static void tcg_out_ld32_12(TCGContext *s, ARMCond cond, TCGReg rt,
644                            TCGReg rn, int imm12)
645{
646    tcg_out_memop_12(s, cond, INSN_LDR_IMM, rt, rn, imm12, 1, 0);
647}
648
649static void tcg_out_st32_12(TCGContext *s, ARMCond cond, TCGReg rt,
650                            TCGReg rn, int imm12)
651{
652    tcg_out_memop_12(s, cond, INSN_STR_IMM, rt, rn, imm12, 1, 0);
653}
654
655static void tcg_out_ld32_r(TCGContext *s, ARMCond cond, TCGReg rt,
656                           TCGReg rn, TCGReg rm)
657{
658    tcg_out_memop_r(s, cond, INSN_LDR_REG, rt, rn, rm, 1, 1, 0);
659}
660
661static void tcg_out_st32_r(TCGContext *s, ARMCond cond, TCGReg rt,
662                           TCGReg rn, TCGReg rm)
663{
664    tcg_out_memop_r(s, cond, INSN_STR_REG, rt, rn, rm, 1, 1, 0);
665}
666
667static void tcg_out_ldrd_8(TCGContext *s, ARMCond cond, TCGReg rt,
668                           TCGReg rn, int imm8)
669{
670    tcg_out_memop_8(s, cond, INSN_LDRD_IMM, rt, rn, imm8, 1, 0);
671}
672
673static void tcg_out_ldrd_r(TCGContext *s, ARMCond cond, TCGReg rt,
674                           TCGReg rn, TCGReg rm)
675{
676    tcg_out_memop_r(s, cond, INSN_LDRD_REG, rt, rn, rm, 1, 1, 0);
677}
678
679static void __attribute__((unused))
680tcg_out_ldrd_rwb(TCGContext *s, ARMCond cond, TCGReg rt, TCGReg rn, TCGReg rm)
681{
682    tcg_out_memop_r(s, cond, INSN_LDRD_REG, rt, rn, rm, 1, 1, 1);
683}
684
685static void __attribute__((unused))
686tcg_out_strd_8(TCGContext *s, ARMCond cond, TCGReg rt, TCGReg rn, int imm8)
687{
688    tcg_out_memop_8(s, cond, INSN_STRD_IMM, rt, rn, imm8, 1, 0);
689}
690
691static void tcg_out_strd_r(TCGContext *s, ARMCond cond, TCGReg rt,
692                           TCGReg rn, TCGReg rm)
693{
694    tcg_out_memop_r(s, cond, INSN_STRD_REG, rt, rn, rm, 1, 1, 0);
695}
696
697/* Register pre-increment with base writeback.  */
698static void tcg_out_ld32_rwb(TCGContext *s, ARMCond cond, TCGReg rt,
699                             TCGReg rn, TCGReg rm)
700{
701    tcg_out_memop_r(s, cond, INSN_LDR_REG, rt, rn, rm, 1, 1, 1);
702}
703
704static void tcg_out_st32_rwb(TCGContext *s, ARMCond cond, TCGReg rt,
705                             TCGReg rn, TCGReg rm)
706{
707    tcg_out_memop_r(s, cond, INSN_STR_REG, rt, rn, rm, 1, 1, 1);
708}
709
710static void tcg_out_ld16u_8(TCGContext *s, ARMCond cond, TCGReg rt,
711                            TCGReg rn, int imm8)
712{
713    tcg_out_memop_8(s, cond, INSN_LDRH_IMM, rt, rn, imm8, 1, 0);
714}
715
716static void tcg_out_st16_8(TCGContext *s, ARMCond cond, TCGReg rt,
717                           TCGReg rn, int imm8)
718{
719    tcg_out_memop_8(s, cond, INSN_STRH_IMM, rt, rn, imm8, 1, 0);
720}
721
722static void tcg_out_ld16u_r(TCGContext *s, ARMCond cond, TCGReg rt,
723                            TCGReg rn, TCGReg rm)
724{
725    tcg_out_memop_r(s, cond, INSN_LDRH_REG, rt, rn, rm, 1, 1, 0);
726}
727
728static void tcg_out_st16_r(TCGContext *s, ARMCond cond, TCGReg rt,
729                           TCGReg rn, TCGReg rm)
730{
731    tcg_out_memop_r(s, cond, INSN_STRH_REG, rt, rn, rm, 1, 1, 0);
732}
733
734static void tcg_out_ld16s_8(TCGContext *s, ARMCond cond, TCGReg rt,
735                            TCGReg rn, int imm8)
736{
737    tcg_out_memop_8(s, cond, INSN_LDRSH_IMM, rt, rn, imm8, 1, 0);
738}
739
740static void tcg_out_ld16s_r(TCGContext *s, ARMCond cond, TCGReg rt,
741                            TCGReg rn, TCGReg rm)
742{
743    tcg_out_memop_r(s, cond, INSN_LDRSH_REG, rt, rn, rm, 1, 1, 0);
744}
745
746static void tcg_out_ld8_12(TCGContext *s, ARMCond cond, TCGReg rt,
747                           TCGReg rn, int imm12)
748{
749    tcg_out_memop_12(s, cond, INSN_LDRB_IMM, rt, rn, imm12, 1, 0);
750}
751
752static void tcg_out_st8_12(TCGContext *s, ARMCond cond, TCGReg rt,
753                           TCGReg rn, int imm12)
754{
755    tcg_out_memop_12(s, cond, INSN_STRB_IMM, rt, rn, imm12, 1, 0);
756}
757
758static void tcg_out_ld8_r(TCGContext *s, ARMCond cond, TCGReg rt,
759                          TCGReg rn, TCGReg rm)
760{
761    tcg_out_memop_r(s, cond, INSN_LDRB_REG, rt, rn, rm, 1, 1, 0);
762}
763
764static void tcg_out_st8_r(TCGContext *s, ARMCond cond, TCGReg rt,
765                          TCGReg rn, TCGReg rm)
766{
767    tcg_out_memop_r(s, cond, INSN_STRB_REG, rt, rn, rm, 1, 1, 0);
768}
769
770static void tcg_out_ld8s_8(TCGContext *s, ARMCond cond, TCGReg rt,
771                           TCGReg rn, int imm8)
772{
773    tcg_out_memop_8(s, cond, INSN_LDRSB_IMM, rt, rn, imm8, 1, 0);
774}
775
776static void tcg_out_ld8s_r(TCGContext *s, ARMCond cond, TCGReg rt,
777                           TCGReg rn, TCGReg rm)
778{
779    tcg_out_memop_r(s, cond, INSN_LDRSB_REG, rt, rn, rm, 1, 1, 0);
780}
781
782static void tcg_out_movi_pool(TCGContext *s, ARMCond cond,
783                              TCGReg rd, uint32_t arg)
784{
785    new_pool_label(s, arg, R_ARM_PC13, s->code_ptr, 0);
786    tcg_out_ld32_12(s, cond, rd, TCG_REG_PC, 0);
787}
788
789static void tcg_out_movi32(TCGContext *s, ARMCond cond,
790                           TCGReg rd, uint32_t arg)
791{
792    int imm12, diff, opc, sh1, sh2;
793    uint32_t tt0, tt1, tt2;
794
795    /* Check a single MOV/MVN before anything else.  */
796    imm12 = encode_imm(arg);
797    if (imm12 >= 0) {
798        tcg_out_dat_imm(s, cond, ARITH_MOV, rd, 0, imm12);
799        return;
800    }
801    imm12 = encode_imm(~arg);
802    if (imm12 >= 0) {
803        tcg_out_dat_imm(s, cond, ARITH_MVN, rd, 0, imm12);
804        return;
805    }
806
807    /* Check for a pc-relative address.  This will usually be the TB,
808       or within the TB, which is immediately before the code block.  */
809    diff = tcg_pcrel_diff(s, (void *)arg) - 8;
810    if (diff >= 0) {
811        imm12 = encode_imm(diff);
812        if (imm12 >= 0) {
813            tcg_out_dat_imm(s, cond, ARITH_ADD, rd, TCG_REG_PC, imm12);
814            return;
815        }
816    } else {
817        imm12 = encode_imm(-diff);
818        if (imm12 >= 0) {
819            tcg_out_dat_imm(s, cond, ARITH_SUB, rd, TCG_REG_PC, imm12);
820            return;
821        }
822    }
823
824    /* Use movw + movt.  */
825    if (use_armv7_instructions) {
826        /* movw */
827        tcg_out32(s, (cond << 28) | 0x03000000 | (rd << 12)
828                  | ((arg << 4) & 0x000f0000) | (arg & 0xfff));
829        if (arg & 0xffff0000) {
830            /* movt */
831            tcg_out32(s, (cond << 28) | 0x03400000 | (rd << 12)
832                      | ((arg >> 12) & 0x000f0000) | ((arg >> 16) & 0xfff));
833        }
834        return;
835    }
836
837    /* Look for sequences of two insns.  If we have lots of 1's, we can
838       shorten the sequence by beginning with mvn and then clearing
839       higher bits with eor.  */
840    tt0 = arg;
841    opc = ARITH_MOV;
842    if (ctpop32(arg) > 16) {
843        tt0 = ~arg;
844        opc = ARITH_MVN;
845    }
846    sh1 = ctz32(tt0) & ~1;
847    tt1 = tt0 & ~(0xff << sh1);
848    sh2 = ctz32(tt1) & ~1;
849    tt2 = tt1 & ~(0xff << sh2);
850    if (tt2 == 0) {
851        int rot;
852
853        rot = ((32 - sh1) << 7) & 0xf00;
854        tcg_out_dat_imm(s, cond, opc, rd,  0, ((tt0 >> sh1) & 0xff) | rot);
855        rot = ((32 - sh2) << 7) & 0xf00;
856        tcg_out_dat_imm(s, cond, ARITH_EOR, rd, rd,
857                        ((tt0 >> sh2) & 0xff) | rot);
858        return;
859    }
860
861    /* Otherwise, drop it into the constant pool.  */
862    tcg_out_movi_pool(s, cond, rd, arg);
863}
864
865/*
866 * Emit either the reg,imm or reg,reg form of a data-processing insn.
867 * rhs must satisfy the "rI" constraint.
868 */
869static void tcg_out_dat_rI(TCGContext *s, ARMCond cond, ARMInsn opc,
870                           TCGReg dst, TCGReg lhs, TCGArg rhs, int rhs_is_const)
871{
872    if (rhs_is_const) {
873        tcg_out_dat_imm(s, cond, opc, dst, lhs, encode_imm_nofail(rhs));
874    } else {
875        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
876    }
877}
878
879/*
880 * Emit either the reg,imm or reg,reg form of a data-processing insn.
881 * rhs must satisfy the "rIK" constraint.
882 */
883static void tcg_out_dat_rIK(TCGContext *s, ARMCond cond, ARMInsn opc,
884                            ARMInsn opinv, TCGReg dst, TCGReg lhs, TCGArg rhs,
885                            bool rhs_is_const)
886{
887    if (rhs_is_const) {
888        int imm12 = encode_imm(rhs);
889        if (imm12 < 0) {
890            imm12 = encode_imm_nofail(~rhs);
891            opc = opinv;
892        }
893        tcg_out_dat_imm(s, cond, opc, dst, lhs, imm12);
894    } else {
895        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
896    }
897}
898
899static void tcg_out_dat_rIN(TCGContext *s, ARMCond cond, ARMInsn opc,
900                            ARMInsn opneg, TCGReg dst, TCGReg lhs, TCGArg rhs,
901                            bool rhs_is_const)
902{
903    /* Emit either the reg,imm or reg,reg form of a data-processing insn.
904     * rhs must satisfy the "rIN" constraint.
905     */
906    if (rhs_is_const) {
907        int imm12 = encode_imm(rhs);
908        if (imm12 < 0) {
909            imm12 = encode_imm_nofail(-rhs);
910            opc = opneg;
911        }
912        tcg_out_dat_imm(s, cond, opc, dst, lhs, imm12);
913    } else {
914        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
915    }
916}
917
918static void tcg_out_mul32(TCGContext *s, ARMCond cond, TCGReg rd,
919                          TCGReg rn, TCGReg rm)
920{
921    /* mul */
922    tcg_out32(s, (cond << 28) | 0x90 | (rd << 16) | (rm << 8) | rn);
923}
924
925static void tcg_out_umull32(TCGContext *s, ARMCond cond, TCGReg rd0,
926                            TCGReg rd1, TCGReg rn, TCGReg rm)
927{
928    /* umull */
929    tcg_out32(s, (cond << 28) | 0x00800090 |
930              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
931}
932
933static void tcg_out_smull32(TCGContext *s, ARMCond cond, TCGReg rd0,
934                            TCGReg rd1, TCGReg rn, TCGReg rm)
935{
936    /* smull */
937    tcg_out32(s, (cond << 28) | 0x00c00090 |
938              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
939}
940
941static void tcg_out_sdiv(TCGContext *s, ARMCond cond,
942                         TCGReg rd, TCGReg rn, TCGReg rm)
943{
944    tcg_out32(s, 0x0710f010 | (cond << 28) | (rd << 16) | rn | (rm << 8));
945}
946
947static void tcg_out_udiv(TCGContext *s, ARMCond cond,
948                         TCGReg rd, TCGReg rn, TCGReg rm)
949{
950    tcg_out32(s, 0x0730f010 | (cond << 28) | (rd << 16) | rn | (rm << 8));
951}
952
953static void tcg_out_ext8s(TCGContext *s, TCGType t, TCGReg rd, TCGReg rn)
954{
955    /* sxtb */
956    tcg_out32(s, 0x06af0070 | (COND_AL << 28) | (rd << 12) | rn);
957}
958
959static void tcg_out_ext8u(TCGContext *s, TCGReg rd, TCGReg rn)
960{
961    tcg_out_dat_imm(s, COND_AL, ARITH_AND, rd, rn, 0xff);
962}
963
964static void tcg_out_ext16s(TCGContext *s, TCGType t, TCGReg rd, TCGReg rn)
965{
966    /* sxth */
967    tcg_out32(s, 0x06bf0070 | (COND_AL << 28) | (rd << 12) | rn);
968}
969
970static void tcg_out_ext16u(TCGContext *s, TCGReg rd, TCGReg rn)
971{
972    /* uxth */
973    tcg_out32(s, 0x06ff0070 | (COND_AL << 28) | (rd << 12) | rn);
974}
975
976static void tcg_out_ext32s(TCGContext *s, TCGReg rd, TCGReg rn)
977{
978    g_assert_not_reached();
979}
980
981static void tcg_out_ext32u(TCGContext *s, TCGReg rd, TCGReg rn)
982{
983    g_assert_not_reached();
984}
985
986static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg rd, TCGReg rn)
987{
988    g_assert_not_reached();
989}
990
991static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg rd, TCGReg rn)
992{
993    g_assert_not_reached();
994}
995
996static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg rd, TCGReg rn)
997{
998    g_assert_not_reached();
999}
1000
1001static void tcg_out_bswap16(TCGContext *s, ARMCond cond,
1002                            TCGReg rd, TCGReg rn, int flags)
1003{
1004    if (flags & TCG_BSWAP_OS) {
1005        /* revsh */
1006        tcg_out32(s, 0x06ff0fb0 | (cond << 28) | (rd << 12) | rn);
1007        return;
1008    }
1009
1010    /* rev16 */
1011    tcg_out32(s, 0x06bf0fb0 | (cond << 28) | (rd << 12) | rn);
1012    if ((flags & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1013        /* uxth */
1014        tcg_out32(s, 0x06ff0070 | (cond << 28) | (rd << 12) | rd);
1015    }
1016}
1017
1018static void tcg_out_bswap32(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rn)
1019{
1020    /* rev */
1021    tcg_out32(s, 0x06bf0f30 | (cond << 28) | (rd << 12) | rn);
1022}
1023
1024static void tcg_out_deposit(TCGContext *s, ARMCond cond, TCGReg rd,
1025                            TCGArg a1, int ofs, int len, bool const_a1)
1026{
1027    if (const_a1) {
1028        /* bfi becomes bfc with rn == 15.  */
1029        a1 = 15;
1030    }
1031    /* bfi/bfc */
1032    tcg_out32(s, 0x07c00010 | (cond << 28) | (rd << 12) | a1
1033              | (ofs << 7) | ((ofs + len - 1) << 16));
1034}
1035
1036static void tcg_out_extract(TCGContext *s, ARMCond cond, TCGReg rd,
1037                            TCGReg rn, int ofs, int len)
1038{
1039    /* According to gcc, AND can be faster. */
1040    if (ofs == 0 && len <= 8) {
1041        tcg_out_dat_imm(s, cond, ARITH_AND, rd, rn,
1042                        encode_imm_nofail((1 << len) - 1));
1043        return;
1044    }
1045
1046    if (use_armv7_instructions) {
1047        /* ubfx */
1048        tcg_out32(s, 0x07e00050 | (cond << 28) | (rd << 12) | rn
1049                  | (ofs << 7) | ((len - 1) << 16));
1050        return;
1051    }
1052
1053    assert(ofs % 8 == 0);
1054    switch (len) {
1055    case 8:
1056        /* uxtb */
1057        tcg_out32(s, 0x06ef0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1058        break;
1059    case 16:
1060        /* uxth */
1061        tcg_out32(s, 0x06ff0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1062        break;
1063    default:
1064        g_assert_not_reached();
1065    }
1066}
1067
1068static void tcg_out_sextract(TCGContext *s, ARMCond cond, TCGReg rd,
1069                             TCGReg rn, int ofs, int len)
1070{
1071    if (use_armv7_instructions) {
1072        /* sbfx */
1073        tcg_out32(s, 0x07a00050 | (cond << 28) | (rd << 12) | rn
1074                  | (ofs << 7) | ((len - 1) << 16));
1075        return;
1076    }
1077
1078    assert(ofs % 8 == 0);
1079    switch (len) {
1080    case 8:
1081        /* sxtb */
1082        tcg_out32(s, 0x06af0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1083        break;
1084    case 16:
1085        /* sxth */
1086        tcg_out32(s, 0x06bf0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1087        break;
1088    default:
1089        g_assert_not_reached();
1090    }
1091}
1092
1093
1094static void tcg_out_ld32u(TCGContext *s, ARMCond cond,
1095                          TCGReg rd, TCGReg rn, int32_t offset)
1096{
1097    if (offset > 0xfff || offset < -0xfff) {
1098        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1099        tcg_out_ld32_r(s, cond, rd, rn, TCG_REG_TMP);
1100    } else
1101        tcg_out_ld32_12(s, cond, rd, rn, offset);
1102}
1103
1104static void tcg_out_st32(TCGContext *s, ARMCond cond,
1105                         TCGReg rd, TCGReg rn, int32_t offset)
1106{
1107    if (offset > 0xfff || offset < -0xfff) {
1108        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1109        tcg_out_st32_r(s, cond, rd, rn, TCG_REG_TMP);
1110    } else
1111        tcg_out_st32_12(s, cond, rd, rn, offset);
1112}
1113
1114static void tcg_out_ld16u(TCGContext *s, ARMCond cond,
1115                          TCGReg rd, TCGReg rn, int32_t offset)
1116{
1117    if (offset > 0xff || offset < -0xff) {
1118        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1119        tcg_out_ld16u_r(s, cond, rd, rn, TCG_REG_TMP);
1120    } else
1121        tcg_out_ld16u_8(s, cond, rd, rn, offset);
1122}
1123
1124static void tcg_out_ld16s(TCGContext *s, ARMCond cond,
1125                          TCGReg rd, TCGReg rn, int32_t offset)
1126{
1127    if (offset > 0xff || offset < -0xff) {
1128        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1129        tcg_out_ld16s_r(s, cond, rd, rn, TCG_REG_TMP);
1130    } else
1131        tcg_out_ld16s_8(s, cond, rd, rn, offset);
1132}
1133
1134static void tcg_out_st16(TCGContext *s, ARMCond cond,
1135                         TCGReg rd, TCGReg rn, int32_t offset)
1136{
1137    if (offset > 0xff || offset < -0xff) {
1138        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1139        tcg_out_st16_r(s, cond, rd, rn, TCG_REG_TMP);
1140    } else
1141        tcg_out_st16_8(s, cond, rd, rn, offset);
1142}
1143
1144static void tcg_out_ld8u(TCGContext *s, ARMCond cond,
1145                         TCGReg rd, TCGReg rn, int32_t offset)
1146{
1147    if (offset > 0xfff || offset < -0xfff) {
1148        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1149        tcg_out_ld8_r(s, cond, rd, rn, TCG_REG_TMP);
1150    } else
1151        tcg_out_ld8_12(s, cond, rd, rn, offset);
1152}
1153
1154static void tcg_out_ld8s(TCGContext *s, ARMCond cond,
1155                         TCGReg rd, TCGReg rn, int32_t offset)
1156{
1157    if (offset > 0xff || offset < -0xff) {
1158        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1159        tcg_out_ld8s_r(s, cond, rd, rn, TCG_REG_TMP);
1160    } else
1161        tcg_out_ld8s_8(s, cond, rd, rn, offset);
1162}
1163
1164static void tcg_out_st8(TCGContext *s, ARMCond cond,
1165                        TCGReg rd, TCGReg rn, int32_t offset)
1166{
1167    if (offset > 0xfff || offset < -0xfff) {
1168        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1169        tcg_out_st8_r(s, cond, rd, rn, TCG_REG_TMP);
1170    } else
1171        tcg_out_st8_12(s, cond, rd, rn, offset);
1172}
1173
1174/*
1175 * The _goto case is normally between TBs within the same code buffer, and
1176 * with the code buffer limited to 16MB we wouldn't need the long case.
1177 * But we also use it for the tail-call to the qemu_ld/st helpers, which does.
1178 */
1179static void tcg_out_goto(TCGContext *s, ARMCond cond, const tcg_insn_unit *addr)
1180{
1181    intptr_t addri = (intptr_t)addr;
1182    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1183    bool arm_mode = !(addri & 1);
1184
1185    if (arm_mode && disp - 8 < 0x01fffffd && disp - 8 > -0x01fffffd) {
1186        tcg_out_b_imm(s, cond, disp);
1187        return;
1188    }
1189
1190    /* LDR is interworking from v5t. */
1191    tcg_out_movi_pool(s, cond, TCG_REG_PC, addri);
1192}
1193
1194/*
1195 * The call case is mostly used for helpers - so it's not unreasonable
1196 * for them to be beyond branch range.
1197 */
1198static void tcg_out_call_int(TCGContext *s, const tcg_insn_unit *addr)
1199{
1200    intptr_t addri = (intptr_t)addr;
1201    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1202    bool arm_mode = !(addri & 1);
1203
1204    if (disp - 8 < 0x02000000 && disp - 8 >= -0x02000000) {
1205        if (arm_mode) {
1206            tcg_out_bl_imm(s, COND_AL, disp);
1207        } else {
1208            tcg_out_blx_imm(s, disp);
1209        }
1210        return;
1211    }
1212
1213    tcg_out_movi32(s, COND_AL, TCG_REG_TMP, addri);
1214    tcg_out_blx_reg(s, COND_AL, TCG_REG_TMP);
1215}
1216
1217static void tcg_out_call(TCGContext *s, const tcg_insn_unit *addr,
1218                         const TCGHelperInfo *info)
1219{
1220    tcg_out_call_int(s, addr);
1221}
1222
1223static void tcg_out_goto_label(TCGContext *s, ARMCond cond, TCGLabel *l)
1224{
1225    if (l->has_value) {
1226        tcg_out_goto(s, cond, l->u.value_ptr);
1227    } else {
1228        tcg_out_reloc(s, s->code_ptr, R_ARM_PC24, l, 0);
1229        tcg_out_b_imm(s, cond, 0);
1230    }
1231}
1232
1233static void tcg_out_mb(TCGContext *s, TCGArg a0)
1234{
1235    if (use_armv7_instructions) {
1236        tcg_out32(s, INSN_DMB_ISH);
1237    } else {
1238        tcg_out32(s, INSN_DMB_MCR);
1239    }
1240}
1241
1242static TCGCond tcg_out_cmp(TCGContext *s, TCGCond cond, TCGReg a,
1243                           TCGArg b, int b_const)
1244{
1245    if (!is_tst_cond(cond)) {
1246        tcg_out_dat_rIN(s, COND_AL, ARITH_CMP, ARITH_CMN, 0, a, b, b_const);
1247        return cond;
1248    }
1249
1250    cond = tcg_tst_eqne_cond(cond);
1251    if (b_const) {
1252        int imm12 = encode_imm(b);
1253
1254        /*
1255         * The compare constraints allow rIN, but TST does not support N.
1256         * Be prepared to load the constant into a scratch register.
1257         */
1258        if (imm12 >= 0) {
1259            tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, a, imm12);
1260            return cond;
1261        }
1262        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, b);
1263        b = TCG_REG_TMP;
1264    }
1265    tcg_out_dat_reg(s, COND_AL, ARITH_TST, 0, a, b, SHIFT_IMM_LSL(0));
1266    return cond;
1267}
1268
1269static TCGCond tcg_out_cmp2(TCGContext *s, const TCGArg *args,
1270                            const int *const_args)
1271{
1272    TCGReg al = args[0];
1273    TCGReg ah = args[1];
1274    TCGArg bl = args[2];
1275    TCGArg bh = args[3];
1276    TCGCond cond = args[4];
1277    int const_bl = const_args[2];
1278    int const_bh = const_args[3];
1279
1280    switch (cond) {
1281    case TCG_COND_EQ:
1282    case TCG_COND_NE:
1283    case TCG_COND_LTU:
1284    case TCG_COND_LEU:
1285    case TCG_COND_GTU:
1286    case TCG_COND_GEU:
1287        /*
1288         * We perform a conditional comparison.  If the high half is
1289         * equal, then overwrite the flags with the comparison of the
1290         * low half.  The resulting flags cover the whole.
1291         */
1292        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, ah, bh, const_bh);
1293        tcg_out_dat_rI(s, COND_EQ, ARITH_CMP, 0, al, bl, const_bl);
1294        return cond;
1295
1296    case TCG_COND_TSTEQ:
1297    case TCG_COND_TSTNE:
1298        /* Similar, but with TST instead of CMP. */
1299        tcg_out_dat_rI(s, COND_AL, ARITH_TST, 0, ah, bh, const_bh);
1300        tcg_out_dat_rI(s, COND_EQ, ARITH_TST, 0, al, bl, const_bl);
1301        return tcg_tst_eqne_cond(cond);
1302
1303    case TCG_COND_LT:
1304    case TCG_COND_GE:
1305        /* We perform a double-word subtraction and examine the result.
1306           We do not actually need the result of the subtract, so the
1307           low part "subtract" is a compare.  For the high half we have
1308           no choice but to compute into a temporary.  */
1309        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, al, bl, const_bl);
1310        tcg_out_dat_rI(s, COND_AL, ARITH_SBC | TO_CPSR,
1311                       TCG_REG_TMP, ah, bh, const_bh);
1312        return cond;
1313
1314    case TCG_COND_LE:
1315    case TCG_COND_GT:
1316        /* Similar, but with swapped arguments, via reversed subtract.  */
1317        tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR,
1318                       TCG_REG_TMP, al, bl, const_bl);
1319        tcg_out_dat_rI(s, COND_AL, ARITH_RSC | TO_CPSR,
1320                       TCG_REG_TMP, ah, bh, const_bh);
1321        return tcg_swap_cond(cond);
1322
1323    default:
1324        g_assert_not_reached();
1325    }
1326}
1327
1328/*
1329 * Note that TCGReg references Q-registers.
1330 * Q-regno = 2 * D-regno, so shift left by 1 while inserting.
1331 */
1332static uint32_t encode_vd(TCGReg rd)
1333{
1334    tcg_debug_assert(rd >= TCG_REG_Q0);
1335    return (extract32(rd, 3, 1) << 22) | (extract32(rd, 0, 3) << 13);
1336}
1337
1338static uint32_t encode_vn(TCGReg rn)
1339{
1340    tcg_debug_assert(rn >= TCG_REG_Q0);
1341    return (extract32(rn, 3, 1) << 7) | (extract32(rn, 0, 3) << 17);
1342}
1343
1344static uint32_t encode_vm(TCGReg rm)
1345{
1346    tcg_debug_assert(rm >= TCG_REG_Q0);
1347    return (extract32(rm, 3, 1) << 5) | (extract32(rm, 0, 3) << 1);
1348}
1349
1350static void tcg_out_vreg2(TCGContext *s, ARMInsn insn, int q, int vece,
1351                          TCGReg d, TCGReg m)
1352{
1353    tcg_out32(s, insn | (vece << 18) | (q << 6) |
1354              encode_vd(d) | encode_vm(m));
1355}
1356
1357static void tcg_out_vreg3(TCGContext *s, ARMInsn insn, int q, int vece,
1358                          TCGReg d, TCGReg n, TCGReg m)
1359{
1360    tcg_out32(s, insn | (vece << 20) | (q << 6) |
1361              encode_vd(d) | encode_vn(n) | encode_vm(m));
1362}
1363
1364static void tcg_out_vmovi(TCGContext *s, TCGReg rd,
1365                          int q, int op, int cmode, uint8_t imm8)
1366{
1367    tcg_out32(s, INSN_VMOVI | encode_vd(rd) | (q << 6) | (op << 5)
1368              | (cmode << 8) | extract32(imm8, 0, 4)
1369              | (extract32(imm8, 4, 3) << 16)
1370              | (extract32(imm8, 7, 1) << 24));
1371}
1372
1373static void tcg_out_vshifti(TCGContext *s, ARMInsn insn, int q,
1374                            TCGReg rd, TCGReg rm, int l_imm6)
1375{
1376    tcg_out32(s, insn | (q << 6) | encode_vd(rd) | encode_vm(rm) |
1377              (extract32(l_imm6, 6, 1) << 7) |
1378              (extract32(l_imm6, 0, 6) << 16));
1379}
1380
1381static void tcg_out_vldst(TCGContext *s, ARMInsn insn,
1382                          TCGReg rd, TCGReg rn, int offset)
1383{
1384    if (offset != 0) {
1385        if (check_fit_imm(offset) || check_fit_imm(-offset)) {
1386            tcg_out_dat_rIN(s, COND_AL, ARITH_ADD, ARITH_SUB,
1387                            TCG_REG_TMP, rn, offset, true);
1388        } else {
1389            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP, offset);
1390            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1391                            TCG_REG_TMP, TCG_REG_TMP, rn, 0);
1392        }
1393        rn = TCG_REG_TMP;
1394    }
1395    tcg_out32(s, insn | (rn << 16) | encode_vd(rd) | 0xf);
1396}
1397
1398typedef struct {
1399    ARMCond cond;
1400    TCGReg base;
1401    int index;
1402    bool index_scratch;
1403    TCGAtomAlign aa;
1404} HostAddress;
1405
1406bool tcg_target_has_memory_bswap(MemOp memop)
1407{
1408    return false;
1409}
1410
1411static TCGReg ldst_ra_gen(TCGContext *s, const TCGLabelQemuLdst *l, int arg)
1412{
1413    /* We arrive at the slow path via "BLNE", so R14 contains l->raddr. */
1414    return TCG_REG_R14;
1415}
1416
1417static const TCGLdstHelperParam ldst_helper_param = {
1418    .ra_gen = ldst_ra_gen,
1419    .ntmp = 1,
1420    .tmp = { TCG_REG_TMP },
1421};
1422
1423static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1424{
1425    MemOp opc = get_memop(lb->oi);
1426
1427    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1428        return false;
1429    }
1430
1431    tcg_out_ld_helper_args(s, lb, &ldst_helper_param);
1432    tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SIZE]);
1433    tcg_out_ld_helper_ret(s, lb, false, &ldst_helper_param);
1434
1435    tcg_out_goto(s, COND_AL, lb->raddr);
1436    return true;
1437}
1438
1439static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1440{
1441    MemOp opc = get_memop(lb->oi);
1442
1443    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1444        return false;
1445    }
1446
1447    tcg_out_st_helper_args(s, lb, &ldst_helper_param);
1448
1449    /* Tail-call to the helper, which will return to the fast path.  */
1450    tcg_out_goto(s, COND_AL, qemu_st_helpers[opc & MO_SIZE]);
1451    return true;
1452}
1453
1454/* We expect to use an 9-bit sign-magnitude negative offset from ENV.  */
1455#define MIN_TLB_MASK_TABLE_OFS  -256
1456
1457static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
1458                                           TCGReg addrlo, TCGReg addrhi,
1459                                           MemOpIdx oi, bool is_ld)
1460{
1461    TCGLabelQemuLdst *ldst = NULL;
1462    MemOp opc = get_memop(oi);
1463    unsigned a_mask;
1464
1465    if (tcg_use_softmmu) {
1466        *h = (HostAddress){
1467            .cond = COND_AL,
1468            .base = addrlo,
1469            .index = TCG_REG_R1,
1470            .index_scratch = true,
1471        };
1472    } else {
1473        *h = (HostAddress){
1474            .cond = COND_AL,
1475            .base = addrlo,
1476            .index = guest_base ? TCG_REG_GUEST_BASE : -1,
1477            .index_scratch = false,
1478        };
1479    }
1480
1481    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, false);
1482    a_mask = (1 << h->aa.align) - 1;
1483
1484    if (tcg_use_softmmu) {
1485        int mem_index = get_mmuidx(oi);
1486        int cmp_off = is_ld ? offsetof(CPUTLBEntry, addr_read)
1487                            : offsetof(CPUTLBEntry, addr_write);
1488        int fast_off = tlb_mask_table_ofs(s, mem_index);
1489        unsigned s_mask = (1 << (opc & MO_SIZE)) - 1;
1490        TCGReg t_addr;
1491
1492        ldst = new_ldst_label(s);
1493        ldst->is_ld = is_ld;
1494        ldst->oi = oi;
1495        ldst->addrlo_reg = addrlo;
1496        ldst->addrhi_reg = addrhi;
1497
1498        /* Load cpu->neg.tlb.f[mmu_idx].{mask,table} into {r0,r1}.  */
1499        QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, mask) != 0);
1500        QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, table) != 4);
1501        tcg_out_ldrd_8(s, COND_AL, TCG_REG_R0, TCG_AREG0, fast_off);
1502
1503        /* Extract the tlb index from the address into R0.  */
1504        tcg_out_dat_reg(s, COND_AL, ARITH_AND, TCG_REG_R0, TCG_REG_R0, addrlo,
1505                        SHIFT_IMM_LSR(s->page_bits - CPU_TLB_ENTRY_BITS));
1506
1507        /*
1508         * Add the tlb_table pointer, creating the CPUTLBEntry address in R1.
1509         * Load the tlb comparator into R2/R3 and the fast path addend into R1.
1510         */
1511        QEMU_BUILD_BUG_ON(HOST_BIG_ENDIAN);
1512        if (cmp_off == 0) {
1513            if (s->addr_type == TCG_TYPE_I32) {
1514                tcg_out_ld32_rwb(s, COND_AL, TCG_REG_R2,
1515                                 TCG_REG_R1, TCG_REG_R0);
1516            } else {
1517                tcg_out_ldrd_rwb(s, COND_AL, TCG_REG_R2,
1518                                 TCG_REG_R1, TCG_REG_R0);
1519            }
1520        } else {
1521            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1522                            TCG_REG_R1, TCG_REG_R1, TCG_REG_R0, 0);
1523            if (s->addr_type == TCG_TYPE_I32) {
1524                tcg_out_ld32_12(s, COND_AL, TCG_REG_R2, TCG_REG_R1, cmp_off);
1525            } else {
1526                tcg_out_ldrd_8(s, COND_AL, TCG_REG_R2, TCG_REG_R1, cmp_off);
1527            }
1528        }
1529
1530        /* Load the tlb addend.  */
1531        tcg_out_ld32_12(s, COND_AL, TCG_REG_R1, TCG_REG_R1,
1532                        offsetof(CPUTLBEntry, addend));
1533
1534        /*
1535         * Check alignment, check comparators.
1536         * Do this in 2-4 insns.  Use MOVW for v7, if possible,
1537         * to reduce the number of sequential conditional instructions.
1538         * Almost all guests have at least 4k pages, which means that we need
1539         * to clear at least 9 bits even for an 8-byte memory, which means it
1540         * isn't worth checking for an immediate operand for BIC.
1541         *
1542         * For unaligned accesses, test the page of the last unit of alignment.
1543         * This leaves the least significant alignment bits unchanged, and of
1544         * course must be zero.
1545         */
1546        t_addr = addrlo;
1547        if (a_mask < s_mask) {
1548            t_addr = TCG_REG_R0;
1549            tcg_out_dat_imm(s, COND_AL, ARITH_ADD, t_addr,
1550                            addrlo, s_mask - a_mask);
1551        }
1552        if (use_armv7_instructions && s->page_bits <= 16) {
1553            tcg_out_movi32(s, COND_AL, TCG_REG_TMP, ~(s->page_mask | a_mask));
1554            tcg_out_dat_reg(s, COND_AL, ARITH_BIC, TCG_REG_TMP,
1555                            t_addr, TCG_REG_TMP, 0);
1556            tcg_out_dat_reg(s, COND_AL, ARITH_CMP, 0,
1557                            TCG_REG_R2, TCG_REG_TMP, 0);
1558        } else {
1559            if (a_mask) {
1560                tcg_debug_assert(a_mask <= 0xff);
1561                tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addrlo, a_mask);
1562            }
1563            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0, t_addr,
1564                            SHIFT_IMM_LSR(s->page_bits));
1565            tcg_out_dat_reg(s, (a_mask ? COND_EQ : COND_AL), ARITH_CMP,
1566                            0, TCG_REG_R2, TCG_REG_TMP,
1567                            SHIFT_IMM_LSL(s->page_bits));
1568        }
1569
1570        if (s->addr_type != TCG_TYPE_I32) {
1571            tcg_out_dat_reg(s, COND_EQ, ARITH_CMP, 0, TCG_REG_R3, addrhi, 0);
1572        }
1573    } else if (a_mask) {
1574        ldst = new_ldst_label(s);
1575        ldst->is_ld = is_ld;
1576        ldst->oi = oi;
1577        ldst->addrlo_reg = addrlo;
1578        ldst->addrhi_reg = addrhi;
1579
1580        /* We are expecting alignment to max out at 7 */
1581        tcg_debug_assert(a_mask <= 0xff);
1582        /* tst addr, #mask */
1583        tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addrlo, a_mask);
1584    }
1585
1586    return ldst;
1587}
1588
1589static void tcg_out_qemu_ld_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1590                                   TCGReg datahi, HostAddress h)
1591{
1592    TCGReg base;
1593
1594    /* Byte swapping is left to middle-end expansion. */
1595    tcg_debug_assert((opc & MO_BSWAP) == 0);
1596
1597    switch (opc & MO_SSIZE) {
1598    case MO_UB:
1599        if (h.index < 0) {
1600            tcg_out_ld8_12(s, h.cond, datalo, h.base, 0);
1601        } else {
1602            tcg_out_ld8_r(s, h.cond, datalo, h.base, h.index);
1603        }
1604        break;
1605    case MO_SB:
1606        if (h.index < 0) {
1607            tcg_out_ld8s_8(s, h.cond, datalo, h.base, 0);
1608        } else {
1609            tcg_out_ld8s_r(s, h.cond, datalo, h.base, h.index);
1610        }
1611        break;
1612    case MO_UW:
1613        if (h.index < 0) {
1614            tcg_out_ld16u_8(s, h.cond, datalo, h.base, 0);
1615        } else {
1616            tcg_out_ld16u_r(s, h.cond, datalo, h.base, h.index);
1617        }
1618        break;
1619    case MO_SW:
1620        if (h.index < 0) {
1621            tcg_out_ld16s_8(s, h.cond, datalo, h.base, 0);
1622        } else {
1623            tcg_out_ld16s_r(s, h.cond, datalo, h.base, h.index);
1624        }
1625        break;
1626    case MO_UL:
1627        if (h.index < 0) {
1628            tcg_out_ld32_12(s, h.cond, datalo, h.base, 0);
1629        } else {
1630            tcg_out_ld32_r(s, h.cond, datalo, h.base, h.index);
1631        }
1632        break;
1633    case MO_UQ:
1634        /* We used pair allocation for datalo, so already should be aligned. */
1635        tcg_debug_assert((datalo & 1) == 0);
1636        tcg_debug_assert(datahi == datalo + 1);
1637        /* LDRD requires alignment; double-check that. */
1638        if (memop_alignment_bits(opc) >= MO_64) {
1639            if (h.index < 0) {
1640                tcg_out_ldrd_8(s, h.cond, datalo, h.base, 0);
1641                break;
1642            }
1643            /*
1644             * Rm (the second address op) must not overlap Rt or Rt + 1.
1645             * Since datalo is aligned, we can simplify the test via alignment.
1646             * Flip the two address arguments if that works.
1647             */
1648            if ((h.index & ~1) != datalo) {
1649                tcg_out_ldrd_r(s, h.cond, datalo, h.base, h.index);
1650                break;
1651            }
1652            if ((h.base & ~1) != datalo) {
1653                tcg_out_ldrd_r(s, h.cond, datalo, h.index, h.base);
1654                break;
1655            }
1656        }
1657        if (h.index < 0) {
1658            base = h.base;
1659            if (datalo == h.base) {
1660                tcg_out_mov_reg(s, h.cond, TCG_REG_TMP, base);
1661                base = TCG_REG_TMP;
1662            }
1663        } else if (h.index_scratch) {
1664            tcg_out_ld32_rwb(s, h.cond, datalo, h.index, h.base);
1665            tcg_out_ld32_12(s, h.cond, datahi, h.index, 4);
1666            break;
1667        } else {
1668            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1669                            h.base, h.index, SHIFT_IMM_LSL(0));
1670            base = TCG_REG_TMP;
1671        }
1672        tcg_out_ld32_12(s, h.cond, datalo, base, 0);
1673        tcg_out_ld32_12(s, h.cond, datahi, base, 4);
1674        break;
1675    default:
1676        g_assert_not_reached();
1677    }
1678}
1679
1680static void tcg_out_qemu_ld(TCGContext *s, TCGReg datalo, TCGReg datahi,
1681                            TCGReg addrlo, TCGReg addrhi,
1682                            MemOpIdx oi, TCGType data_type)
1683{
1684    MemOp opc = get_memop(oi);
1685    TCGLabelQemuLdst *ldst;
1686    HostAddress h;
1687
1688    ldst = prepare_host_addr(s, &h, addrlo, addrhi, oi, true);
1689    if (ldst) {
1690        ldst->type = data_type;
1691        ldst->datalo_reg = datalo;
1692        ldst->datahi_reg = datahi;
1693
1694        /*
1695         * This a conditional BL only to load a pointer within this
1696         * opcode into LR for the slow path.  We will not be using
1697         * the value for a tail call.
1698         */
1699        ldst->label_ptr[0] = s->code_ptr;
1700        tcg_out_bl_imm(s, COND_NE, 0);
1701
1702        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1703        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1704    } else {
1705        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1706    }
1707}
1708
1709static void tcg_out_qemu_st_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1710                                   TCGReg datahi, HostAddress h)
1711{
1712    /* Byte swapping is left to middle-end expansion. */
1713    tcg_debug_assert((opc & MO_BSWAP) == 0);
1714
1715    switch (opc & MO_SIZE) {
1716    case MO_8:
1717        if (h.index < 0) {
1718            tcg_out_st8_12(s, h.cond, datalo, h.base, 0);
1719        } else {
1720            tcg_out_st8_r(s, h.cond, datalo, h.base, h.index);
1721        }
1722        break;
1723    case MO_16:
1724        if (h.index < 0) {
1725            tcg_out_st16_8(s, h.cond, datalo, h.base, 0);
1726        } else {
1727            tcg_out_st16_r(s, h.cond, datalo, h.base, h.index);
1728        }
1729        break;
1730    case MO_32:
1731        if (h.index < 0) {
1732            tcg_out_st32_12(s, h.cond, datalo, h.base, 0);
1733        } else {
1734            tcg_out_st32_r(s, h.cond, datalo, h.base, h.index);
1735        }
1736        break;
1737    case MO_64:
1738        /* We used pair allocation for datalo, so already should be aligned. */
1739        tcg_debug_assert((datalo & 1) == 0);
1740        tcg_debug_assert(datahi == datalo + 1);
1741        /* STRD requires alignment; double-check that. */
1742        if (memop_alignment_bits(opc) >= MO_64) {
1743            if (h.index < 0) {
1744                tcg_out_strd_8(s, h.cond, datalo, h.base, 0);
1745            } else {
1746                tcg_out_strd_r(s, h.cond, datalo, h.base, h.index);
1747            }
1748        } else if (h.index < 0) {
1749            tcg_out_st32_12(s, h.cond, datalo, h.base, 0);
1750            tcg_out_st32_12(s, h.cond, datahi, h.base, 4);
1751        } else if (h.index_scratch) {
1752            tcg_out_st32_rwb(s, h.cond, datalo, h.index, h.base);
1753            tcg_out_st32_12(s, h.cond, datahi, h.index, 4);
1754        } else {
1755            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1756                            h.base, h.index, SHIFT_IMM_LSL(0));
1757            tcg_out_st32_12(s, h.cond, datalo, TCG_REG_TMP, 0);
1758            tcg_out_st32_12(s, h.cond, datahi, TCG_REG_TMP, 4);
1759        }
1760        break;
1761    default:
1762        g_assert_not_reached();
1763    }
1764}
1765
1766static void tcg_out_qemu_st(TCGContext *s, TCGReg datalo, TCGReg datahi,
1767                            TCGReg addrlo, TCGReg addrhi,
1768                            MemOpIdx oi, TCGType data_type)
1769{
1770    MemOp opc = get_memop(oi);
1771    TCGLabelQemuLdst *ldst;
1772    HostAddress h;
1773
1774    ldst = prepare_host_addr(s, &h, addrlo, addrhi, oi, false);
1775    if (ldst) {
1776        ldst->type = data_type;
1777        ldst->datalo_reg = datalo;
1778        ldst->datahi_reg = datahi;
1779
1780        h.cond = COND_EQ;
1781        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1782
1783        /* The conditional call is last, as we're going to return here. */
1784        ldst->label_ptr[0] = s->code_ptr;
1785        tcg_out_bl_imm(s, COND_NE, 0);
1786        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1787    } else {
1788        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1789    }
1790}
1791
1792static void tcg_out_epilogue(TCGContext *s);
1793
1794static void tcg_out_exit_tb(TCGContext *s, uintptr_t arg)
1795{
1796    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, arg);
1797    tcg_out_epilogue(s);
1798}
1799
1800static void tcg_out_goto_tb(TCGContext *s, int which)
1801{
1802    uintptr_t i_addr;
1803    intptr_t i_disp;
1804
1805    /* Direct branch will be patched by tb_target_set_jmp_target. */
1806    set_jmp_insn_offset(s, which);
1807    tcg_out32(s, INSN_NOP);
1808
1809    /* When branch is out of range, fall through to indirect. */
1810    i_addr = get_jmp_target_addr(s, which);
1811    i_disp = tcg_pcrel_diff(s, (void *)i_addr) - 8;
1812    tcg_debug_assert(i_disp < 0);
1813    if (i_disp >= -0xfff) {
1814        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_PC, i_disp);
1815    } else {
1816        /*
1817         * The TB is close, but outside the 12 bits addressable by
1818         * the load.  We can extend this to 20 bits with a sub of a
1819         * shifted immediate from pc.
1820         */
1821        int h = -i_disp;
1822        int l = -(h & 0xfff);
1823
1824        h = encode_imm_nofail(h + l);
1825        tcg_out_dat_imm(s, COND_AL, ARITH_SUB, TCG_REG_R0, TCG_REG_PC, h);
1826        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_R0, l);
1827    }
1828    set_jmp_reset_offset(s, which);
1829}
1830
1831void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
1832                              uintptr_t jmp_rx, uintptr_t jmp_rw)
1833{
1834    uintptr_t addr = tb->jmp_target_addr[n];
1835    ptrdiff_t offset = addr - (jmp_rx + 8);
1836    tcg_insn_unit insn;
1837
1838    /* Either directly branch, or fall through to indirect branch. */
1839    if (offset == sextract64(offset, 0, 26)) {
1840        /* B <addr> */
1841        insn = deposit32((COND_AL << 28) | INSN_B, 0, 24, offset >> 2);
1842    } else {
1843        insn = INSN_NOP;
1844    }
1845
1846    qatomic_set((uint32_t *)jmp_rw, insn);
1847    flush_idcache_range(jmp_rx, jmp_rw, 4);
1848}
1849
1850static void tcg_out_op(TCGContext *s, TCGOpcode opc, TCGType type,
1851                       const TCGArg args[TCG_MAX_OP_ARGS],
1852                       const int const_args[TCG_MAX_OP_ARGS])
1853{
1854    TCGArg a0, a1, a2, a3, a4, a5;
1855    int c;
1856
1857    switch (opc) {
1858    case INDEX_op_goto_ptr:
1859        tcg_out_b_reg(s, COND_AL, args[0]);
1860        break;
1861    case INDEX_op_br:
1862        tcg_out_goto_label(s, COND_AL, arg_label(args[0]));
1863        break;
1864
1865    case INDEX_op_ld8u_i32:
1866        tcg_out_ld8u(s, COND_AL, args[0], args[1], args[2]);
1867        break;
1868    case INDEX_op_ld8s_i32:
1869        tcg_out_ld8s(s, COND_AL, args[0], args[1], args[2]);
1870        break;
1871    case INDEX_op_ld16u_i32:
1872        tcg_out_ld16u(s, COND_AL, args[0], args[1], args[2]);
1873        break;
1874    case INDEX_op_ld16s_i32:
1875        tcg_out_ld16s(s, COND_AL, args[0], args[1], args[2]);
1876        break;
1877    case INDEX_op_ld_i32:
1878        tcg_out_ld32u(s, COND_AL, args[0], args[1], args[2]);
1879        break;
1880    case INDEX_op_st8_i32:
1881        tcg_out_st8(s, COND_AL, args[0], args[1], args[2]);
1882        break;
1883    case INDEX_op_st16_i32:
1884        tcg_out_st16(s, COND_AL, args[0], args[1], args[2]);
1885        break;
1886    case INDEX_op_st_i32:
1887        tcg_out_st32(s, COND_AL, args[0], args[1], args[2]);
1888        break;
1889
1890    case INDEX_op_movcond_i32:
1891        /* Constraints mean that v2 is always in the same register as dest,
1892         * so we only need to do "if condition passed, move v1 to dest".
1893         */
1894        c = tcg_out_cmp(s, args[5], args[1], args[2], const_args[2]);
1895        tcg_out_dat_rIK(s, tcg_cond_to_arm_cond[c], ARITH_MOV,
1896                        ARITH_MVN, args[0], 0, args[3], const_args[3]);
1897        break;
1898    case INDEX_op_add_i32:
1899        tcg_out_dat_rIN(s, COND_AL, ARITH_ADD, ARITH_SUB,
1900                        args[0], args[1], args[2], const_args[2]);
1901        break;
1902    case INDEX_op_sub_i32:
1903        if (const_args[1]) {
1904            if (const_args[2]) {
1905                tcg_out_movi32(s, COND_AL, args[0], args[1] - args[2]);
1906            } else {
1907                tcg_out_dat_rI(s, COND_AL, ARITH_RSB,
1908                               args[0], args[2], args[1], 1);
1909            }
1910        } else {
1911            tcg_out_dat_rIN(s, COND_AL, ARITH_SUB, ARITH_ADD,
1912                            args[0], args[1], args[2], const_args[2]);
1913        }
1914        break;
1915    case INDEX_op_and_i32:
1916        tcg_out_dat_rIK(s, COND_AL, ARITH_AND, ARITH_BIC,
1917                        args[0], args[1], args[2], const_args[2]);
1918        break;
1919    case INDEX_op_andc_i32:
1920        tcg_out_dat_rIK(s, COND_AL, ARITH_BIC, ARITH_AND,
1921                        args[0], args[1], args[2], const_args[2]);
1922        break;
1923    case INDEX_op_or_i32:
1924        c = ARITH_ORR;
1925        goto gen_arith;
1926    case INDEX_op_xor_i32:
1927        c = ARITH_EOR;
1928        /* Fall through.  */
1929    gen_arith:
1930        tcg_out_dat_rI(s, COND_AL, c, args[0], args[1], args[2], const_args[2]);
1931        break;
1932    case INDEX_op_add2_i32:
1933        a0 = args[0], a1 = args[1], a2 = args[2];
1934        a3 = args[3], a4 = args[4], a5 = args[5];
1935        if (a0 == a3 || (a0 == a5 && !const_args[5])) {
1936            a0 = TCG_REG_TMP;
1937        }
1938        tcg_out_dat_rIN(s, COND_AL, ARITH_ADD | TO_CPSR, ARITH_SUB | TO_CPSR,
1939                        a0, a2, a4, const_args[4]);
1940        tcg_out_dat_rIK(s, COND_AL, ARITH_ADC, ARITH_SBC,
1941                        a1, a3, a5, const_args[5]);
1942        tcg_out_mov_reg(s, COND_AL, args[0], a0);
1943        break;
1944    case INDEX_op_sub2_i32:
1945        a0 = args[0], a1 = args[1], a2 = args[2];
1946        a3 = args[3], a4 = args[4], a5 = args[5];
1947        if ((a0 == a3 && !const_args[3]) || (a0 == a5 && !const_args[5])) {
1948            a0 = TCG_REG_TMP;
1949        }
1950        if (const_args[2]) {
1951            if (const_args[4]) {
1952                tcg_out_movi32(s, COND_AL, a0, a4);
1953                a4 = a0;
1954            }
1955            tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR, a0, a4, a2, 1);
1956        } else {
1957            tcg_out_dat_rIN(s, COND_AL, ARITH_SUB | TO_CPSR,
1958                            ARITH_ADD | TO_CPSR, a0, a2, a4, const_args[4]);
1959        }
1960        if (const_args[3]) {
1961            if (const_args[5]) {
1962                tcg_out_movi32(s, COND_AL, a1, a5);
1963                a5 = a1;
1964            }
1965            tcg_out_dat_rI(s, COND_AL, ARITH_RSC, a1, a5, a3, 1);
1966        } else {
1967            tcg_out_dat_rIK(s, COND_AL, ARITH_SBC, ARITH_ADC,
1968                            a1, a3, a5, const_args[5]);
1969        }
1970        tcg_out_mov_reg(s, COND_AL, args[0], a0);
1971        break;
1972    case INDEX_op_neg_i32:
1973        tcg_out_dat_imm(s, COND_AL, ARITH_RSB, args[0], args[1], 0);
1974        break;
1975    case INDEX_op_not_i32:
1976        tcg_out_dat_reg(s, COND_AL,
1977                        ARITH_MVN, args[0], 0, args[1], SHIFT_IMM_LSL(0));
1978        break;
1979    case INDEX_op_mul_i32:
1980        tcg_out_mul32(s, COND_AL, args[0], args[1], args[2]);
1981        break;
1982    case INDEX_op_mulu2_i32:
1983        tcg_out_umull32(s, COND_AL, args[0], args[1], args[2], args[3]);
1984        break;
1985    case INDEX_op_muls2_i32:
1986        tcg_out_smull32(s, COND_AL, args[0], args[1], args[2], args[3]);
1987        break;
1988    /* XXX: Perhaps args[2] & 0x1f is wrong */
1989    case INDEX_op_shl_i32:
1990        c = const_args[2] ?
1991                SHIFT_IMM_LSL(args[2] & 0x1f) : SHIFT_REG_LSL(args[2]);
1992        goto gen_shift32;
1993    case INDEX_op_shr_i32:
1994        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_LSR(args[2] & 0x1f) :
1995                SHIFT_IMM_LSL(0) : SHIFT_REG_LSR(args[2]);
1996        goto gen_shift32;
1997    case INDEX_op_sar_i32:
1998        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_ASR(args[2] & 0x1f) :
1999                SHIFT_IMM_LSL(0) : SHIFT_REG_ASR(args[2]);
2000        goto gen_shift32;
2001    case INDEX_op_rotr_i32:
2002        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_ROR(args[2] & 0x1f) :
2003                SHIFT_IMM_LSL(0) : SHIFT_REG_ROR(args[2]);
2004        /* Fall through.  */
2005    gen_shift32:
2006        tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1], c);
2007        break;
2008
2009    case INDEX_op_rotl_i32:
2010        if (const_args[2]) {
2011            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1],
2012                            ((0x20 - args[2]) & 0x1f) ?
2013                            SHIFT_IMM_ROR((0x20 - args[2]) & 0x1f) :
2014                            SHIFT_IMM_LSL(0));
2015        } else {
2016            tcg_out_dat_imm(s, COND_AL, ARITH_RSB, TCG_REG_TMP, args[2], 0x20);
2017            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1],
2018                            SHIFT_REG_ROR(TCG_REG_TMP));
2019        }
2020        break;
2021
2022    case INDEX_op_ctz_i32:
2023        tcg_out_dat_reg(s, COND_AL, INSN_RBIT, TCG_REG_TMP, 0, args[1], 0);
2024        a1 = TCG_REG_TMP;
2025        goto do_clz;
2026
2027    case INDEX_op_clz_i32:
2028        a1 = args[1];
2029    do_clz:
2030        a0 = args[0];
2031        a2 = args[2];
2032        c = const_args[2];
2033        if (c && a2 == 32) {
2034            tcg_out_dat_reg(s, COND_AL, INSN_CLZ, a0, 0, a1, 0);
2035            break;
2036        }
2037        tcg_out_dat_imm(s, COND_AL, ARITH_CMP, 0, a1, 0);
2038        tcg_out_dat_reg(s, COND_NE, INSN_CLZ, a0, 0, a1, 0);
2039        if (c || a0 != a2) {
2040            tcg_out_dat_rIK(s, COND_EQ, ARITH_MOV, ARITH_MVN, a0, 0, a2, c);
2041        }
2042        break;
2043
2044    case INDEX_op_brcond_i32:
2045        c = tcg_out_cmp(s, args[2], args[0], args[1], const_args[1]);
2046        tcg_out_goto_label(s, tcg_cond_to_arm_cond[c], arg_label(args[3]));
2047        break;
2048    case INDEX_op_setcond_i32:
2049        c = tcg_out_cmp(s, args[3], args[1], args[2], const_args[2]);
2050        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c],
2051                        ARITH_MOV, args[0], 0, 1);
2052        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
2053                        ARITH_MOV, args[0], 0, 0);
2054        break;
2055    case INDEX_op_negsetcond_i32:
2056        c = tcg_out_cmp(s, args[3], args[1], args[2], const_args[2]);
2057        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c],
2058                        ARITH_MVN, args[0], 0, 0);
2059        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
2060                        ARITH_MOV, args[0], 0, 0);
2061        break;
2062
2063    case INDEX_op_brcond2_i32:
2064        c = tcg_out_cmp2(s, args, const_args);
2065        tcg_out_goto_label(s, tcg_cond_to_arm_cond[c], arg_label(args[5]));
2066        break;
2067    case INDEX_op_setcond2_i32:
2068        c = tcg_out_cmp2(s, args + 1, const_args + 1);
2069        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c], ARITH_MOV, args[0], 0, 1);
2070        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
2071                        ARITH_MOV, args[0], 0, 0);
2072        break;
2073
2074    case INDEX_op_qemu_ld_a32_i32:
2075        tcg_out_qemu_ld(s, args[0], -1, args[1], -1, args[2], TCG_TYPE_I32);
2076        break;
2077    case INDEX_op_qemu_ld_a64_i32:
2078        tcg_out_qemu_ld(s, args[0], -1, args[1], args[2],
2079                        args[3], TCG_TYPE_I32);
2080        break;
2081    case INDEX_op_qemu_ld_a32_i64:
2082        tcg_out_qemu_ld(s, args[0], args[1], args[2], -1,
2083                        args[3], TCG_TYPE_I64);
2084        break;
2085    case INDEX_op_qemu_ld_a64_i64:
2086        tcg_out_qemu_ld(s, args[0], args[1], args[2], args[3],
2087                        args[4], TCG_TYPE_I64);
2088        break;
2089
2090    case INDEX_op_qemu_st_a32_i32:
2091        tcg_out_qemu_st(s, args[0], -1, args[1], -1, args[2], TCG_TYPE_I32);
2092        break;
2093    case INDEX_op_qemu_st_a64_i32:
2094        tcg_out_qemu_st(s, args[0], -1, args[1], args[2],
2095                        args[3], TCG_TYPE_I32);
2096        break;
2097    case INDEX_op_qemu_st_a32_i64:
2098        tcg_out_qemu_st(s, args[0], args[1], args[2], -1,
2099                        args[3], TCG_TYPE_I64);
2100        break;
2101    case INDEX_op_qemu_st_a64_i64:
2102        tcg_out_qemu_st(s, args[0], args[1], args[2], args[3],
2103                        args[4], TCG_TYPE_I64);
2104        break;
2105
2106    case INDEX_op_bswap16_i32:
2107        tcg_out_bswap16(s, COND_AL, args[0], args[1], args[2]);
2108        break;
2109    case INDEX_op_bswap32_i32:
2110        tcg_out_bswap32(s, COND_AL, args[0], args[1]);
2111        break;
2112
2113    case INDEX_op_deposit_i32:
2114        tcg_out_deposit(s, COND_AL, args[0], args[2],
2115                        args[3], args[4], const_args[2]);
2116        break;
2117    case INDEX_op_extract_i32:
2118        tcg_out_extract(s, COND_AL, args[0], args[1], args[2], args[3]);
2119        break;
2120    case INDEX_op_sextract_i32:
2121        tcg_out_sextract(s, COND_AL, args[0], args[1], args[2], args[3]);
2122        break;
2123    case INDEX_op_extract2_i32:
2124        /* ??? These optimization vs zero should be generic.  */
2125        /* ??? But we can't substitute 2 for 1 in the opcode stream yet.  */
2126        if (const_args[1]) {
2127            if (const_args[2]) {
2128                tcg_out_movi(s, TCG_TYPE_REG, args[0], 0);
2129            } else {
2130                tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2131                                args[2], SHIFT_IMM_LSL(32 - args[3]));
2132            }
2133        } else if (const_args[2]) {
2134            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2135                            args[1], SHIFT_IMM_LSR(args[3]));
2136        } else {
2137            /* We can do extract2 in 2 insns, vs the 3 required otherwise.  */
2138            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0,
2139                            args[2], SHIFT_IMM_LSL(32 - args[3]));
2140            tcg_out_dat_reg(s, COND_AL, ARITH_ORR, args[0], TCG_REG_TMP,
2141                            args[1], SHIFT_IMM_LSR(args[3]));
2142        }
2143        break;
2144
2145    case INDEX_op_div_i32:
2146        tcg_out_sdiv(s, COND_AL, args[0], args[1], args[2]);
2147        break;
2148    case INDEX_op_divu_i32:
2149        tcg_out_udiv(s, COND_AL, args[0], args[1], args[2]);
2150        break;
2151
2152    case INDEX_op_mb:
2153        tcg_out_mb(s, args[0]);
2154        break;
2155
2156    case INDEX_op_mov_i32:  /* Always emitted via tcg_out_mov.  */
2157    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
2158    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
2159    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
2160    case INDEX_op_ext8s_i32:  /* Always emitted via tcg_reg_alloc_op.  */
2161    case INDEX_op_ext8u_i32:
2162    case INDEX_op_ext16s_i32:
2163    case INDEX_op_ext16u_i32:
2164    default:
2165        g_assert_not_reached();
2166    }
2167}
2168
2169static TCGConstraintSetIndex
2170tcg_target_op_def(TCGOpcode op, TCGType type, unsigned flags)
2171{
2172    switch (op) {
2173    case INDEX_op_goto_ptr:
2174        return C_O0_I1(r);
2175
2176    case INDEX_op_ld8u_i32:
2177    case INDEX_op_ld8s_i32:
2178    case INDEX_op_ld16u_i32:
2179    case INDEX_op_ld16s_i32:
2180    case INDEX_op_ld_i32:
2181    case INDEX_op_neg_i32:
2182    case INDEX_op_not_i32:
2183    case INDEX_op_bswap16_i32:
2184    case INDEX_op_bswap32_i32:
2185    case INDEX_op_ext8s_i32:
2186    case INDEX_op_ext16s_i32:
2187    case INDEX_op_ext16u_i32:
2188    case INDEX_op_extract_i32:
2189    case INDEX_op_sextract_i32:
2190        return C_O1_I1(r, r);
2191
2192    case INDEX_op_st8_i32:
2193    case INDEX_op_st16_i32:
2194    case INDEX_op_st_i32:
2195        return C_O0_I2(r, r);
2196
2197    case INDEX_op_add_i32:
2198    case INDEX_op_sub_i32:
2199    case INDEX_op_setcond_i32:
2200    case INDEX_op_negsetcond_i32:
2201        return C_O1_I2(r, r, rIN);
2202
2203    case INDEX_op_and_i32:
2204    case INDEX_op_andc_i32:
2205    case INDEX_op_clz_i32:
2206    case INDEX_op_ctz_i32:
2207        return C_O1_I2(r, r, rIK);
2208
2209    case INDEX_op_mul_i32:
2210    case INDEX_op_div_i32:
2211    case INDEX_op_divu_i32:
2212        return C_O1_I2(r, r, r);
2213
2214    case INDEX_op_mulu2_i32:
2215    case INDEX_op_muls2_i32:
2216        return C_O2_I2(r, r, r, r);
2217
2218    case INDEX_op_or_i32:
2219    case INDEX_op_xor_i32:
2220        return C_O1_I2(r, r, rI);
2221
2222    case INDEX_op_shl_i32:
2223    case INDEX_op_shr_i32:
2224    case INDEX_op_sar_i32:
2225    case INDEX_op_rotl_i32:
2226    case INDEX_op_rotr_i32:
2227        return C_O1_I2(r, r, ri);
2228
2229    case INDEX_op_brcond_i32:
2230        return C_O0_I2(r, rIN);
2231    case INDEX_op_deposit_i32:
2232        return C_O1_I2(r, 0, rZ);
2233    case INDEX_op_extract2_i32:
2234        return C_O1_I2(r, rZ, rZ);
2235    case INDEX_op_movcond_i32:
2236        return C_O1_I4(r, r, rIN, rIK, 0);
2237    case INDEX_op_add2_i32:
2238        return C_O2_I4(r, r, r, r, rIN, rIK);
2239    case INDEX_op_sub2_i32:
2240        return C_O2_I4(r, r, rI, rI, rIN, rIK);
2241    case INDEX_op_brcond2_i32:
2242        return C_O0_I4(r, r, rI, rI);
2243    case INDEX_op_setcond2_i32:
2244        return C_O1_I4(r, r, r, rI, rI);
2245
2246    case INDEX_op_qemu_ld_a32_i32:
2247        return C_O1_I1(r, q);
2248    case INDEX_op_qemu_ld_a64_i32:
2249        return C_O1_I2(r, q, q);
2250    case INDEX_op_qemu_ld_a32_i64:
2251        return C_O2_I1(e, p, q);
2252    case INDEX_op_qemu_ld_a64_i64:
2253        return C_O2_I2(e, p, q, q);
2254    case INDEX_op_qemu_st_a32_i32:
2255        return C_O0_I2(q, q);
2256    case INDEX_op_qemu_st_a64_i32:
2257        return C_O0_I3(q, q, q);
2258    case INDEX_op_qemu_st_a32_i64:
2259        return C_O0_I3(Q, p, q);
2260    case INDEX_op_qemu_st_a64_i64:
2261        return C_O0_I4(Q, p, q, q);
2262
2263    case INDEX_op_st_vec:
2264        return C_O0_I2(w, r);
2265    case INDEX_op_ld_vec:
2266    case INDEX_op_dupm_vec:
2267        return C_O1_I1(w, r);
2268    case INDEX_op_dup_vec:
2269        return C_O1_I1(w, wr);
2270    case INDEX_op_abs_vec:
2271    case INDEX_op_neg_vec:
2272    case INDEX_op_not_vec:
2273    case INDEX_op_shli_vec:
2274    case INDEX_op_shri_vec:
2275    case INDEX_op_sari_vec:
2276        return C_O1_I1(w, w);
2277    case INDEX_op_dup2_vec:
2278    case INDEX_op_add_vec:
2279    case INDEX_op_mul_vec:
2280    case INDEX_op_smax_vec:
2281    case INDEX_op_smin_vec:
2282    case INDEX_op_ssadd_vec:
2283    case INDEX_op_sssub_vec:
2284    case INDEX_op_sub_vec:
2285    case INDEX_op_umax_vec:
2286    case INDEX_op_umin_vec:
2287    case INDEX_op_usadd_vec:
2288    case INDEX_op_ussub_vec:
2289    case INDEX_op_xor_vec:
2290    case INDEX_op_arm_sshl_vec:
2291    case INDEX_op_arm_ushl_vec:
2292        return C_O1_I2(w, w, w);
2293    case INDEX_op_arm_sli_vec:
2294        return C_O1_I2(w, 0, w);
2295    case INDEX_op_or_vec:
2296    case INDEX_op_andc_vec:
2297        return C_O1_I2(w, w, wO);
2298    case INDEX_op_and_vec:
2299    case INDEX_op_orc_vec:
2300        return C_O1_I2(w, w, wV);
2301    case INDEX_op_cmp_vec:
2302        return C_O1_I2(w, w, wZ);
2303    case INDEX_op_bitsel_vec:
2304        return C_O1_I3(w, w, w, w);
2305    default:
2306        return C_NotImplemented;
2307    }
2308}
2309
2310static void tcg_target_init(TCGContext *s)
2311{
2312    /*
2313     * Only probe for the platform and capabilities if we haven't already
2314     * determined maximum values at compile time.
2315     */
2316#if !defined(use_idiv_instructions) || !defined(use_neon_instructions)
2317    {
2318        unsigned long hwcap = qemu_getauxval(AT_HWCAP);
2319#ifndef use_idiv_instructions
2320        use_idiv_instructions = (hwcap & HWCAP_ARM_IDIVA) != 0;
2321#endif
2322#ifndef use_neon_instructions
2323        use_neon_instructions = (hwcap & HWCAP_ARM_NEON) != 0;
2324#endif
2325    }
2326#endif
2327
2328    if (__ARM_ARCH < 7) {
2329        const char *pl = (const char *)qemu_getauxval(AT_PLATFORM);
2330        if (pl != NULL && pl[0] == 'v' && pl[1] >= '4' && pl[1] <= '9') {
2331            arm_arch = pl[1] - '0';
2332        }
2333
2334        if (arm_arch < 6) {
2335            error_report("TCG: ARMv%d is unsupported; exiting", arm_arch);
2336            exit(EXIT_FAILURE);
2337        }
2338    }
2339
2340    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
2341
2342    tcg_target_call_clobber_regs = 0;
2343    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R0);
2344    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R1);
2345    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R2);
2346    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R3);
2347    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R12);
2348    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R14);
2349
2350    if (use_neon_instructions) {
2351        tcg_target_available_regs[TCG_TYPE_V64]  = ALL_VECTOR_REGS;
2352        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
2353
2354        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q0);
2355        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q1);
2356        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q2);
2357        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q3);
2358        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q8);
2359        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q9);
2360        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q10);
2361        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q11);
2362        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q12);
2363        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q13);
2364        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q14);
2365        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q15);
2366    }
2367
2368    s->reserved_regs = 0;
2369    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
2370    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP);
2371    tcg_regset_set_reg(s->reserved_regs, TCG_REG_PC);
2372    tcg_regset_set_reg(s->reserved_regs, TCG_VEC_TMP);
2373}
2374
2375static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg arg,
2376                       TCGReg arg1, intptr_t arg2)
2377{
2378    switch (type) {
2379    case TCG_TYPE_I32:
2380        tcg_out_ld32u(s, COND_AL, arg, arg1, arg2);
2381        return;
2382    case TCG_TYPE_V64:
2383        /* regs 1; size 8; align 8 */
2384        tcg_out_vldst(s, INSN_VLD1 | 0x7d0, arg, arg1, arg2);
2385        return;
2386    case TCG_TYPE_V128:
2387        /*
2388         * We have only 8-byte alignment for the stack per the ABI.
2389         * Rather than dynamically re-align the stack, it's easier
2390         * to simply not request alignment beyond that.  So:
2391         * regs 2; size 8; align 8
2392         */
2393        tcg_out_vldst(s, INSN_VLD1 | 0xad0, arg, arg1, arg2);
2394        return;
2395    default:
2396        g_assert_not_reached();
2397    }
2398}
2399
2400static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
2401                       TCGReg arg1, intptr_t arg2)
2402{
2403    switch (type) {
2404    case TCG_TYPE_I32:
2405        tcg_out_st32(s, COND_AL, arg, arg1, arg2);
2406        return;
2407    case TCG_TYPE_V64:
2408        /* regs 1; size 8; align 8 */
2409        tcg_out_vldst(s, INSN_VST1 | 0x7d0, arg, arg1, arg2);
2410        return;
2411    case TCG_TYPE_V128:
2412        /* See tcg_out_ld re alignment: regs 2; size 8; align 8 */
2413        tcg_out_vldst(s, INSN_VST1 | 0xad0, arg, arg1, arg2);
2414        return;
2415    default:
2416        g_assert_not_reached();
2417    }
2418}
2419
2420static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
2421                        TCGReg base, intptr_t ofs)
2422{
2423    return false;
2424}
2425
2426static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
2427{
2428    if (ret == arg) {
2429        return true;
2430    }
2431    switch (type) {
2432    case TCG_TYPE_I32:
2433        if (ret < TCG_REG_Q0 && arg < TCG_REG_Q0) {
2434            tcg_out_mov_reg(s, COND_AL, ret, arg);
2435            return true;
2436        }
2437        return false;
2438
2439    case TCG_TYPE_V64:
2440    case TCG_TYPE_V128:
2441        /* "VMOV D,N" is an alias for "VORR D,N,N". */
2442        tcg_out_vreg3(s, INSN_VORR, type - TCG_TYPE_V64, 0, ret, arg, arg);
2443        return true;
2444
2445    default:
2446        g_assert_not_reached();
2447    }
2448}
2449
2450static void tcg_out_movi(TCGContext *s, TCGType type,
2451                         TCGReg ret, tcg_target_long arg)
2452{
2453    tcg_debug_assert(type == TCG_TYPE_I32);
2454    tcg_debug_assert(ret < TCG_REG_Q0);
2455    tcg_out_movi32(s, COND_AL, ret, arg);
2456}
2457
2458static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
2459{
2460    return false;
2461}
2462
2463static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
2464                             tcg_target_long imm)
2465{
2466    int enc, opc = ARITH_ADD;
2467
2468    /* All of the easiest immediates to encode are positive. */
2469    if (imm < 0) {
2470        imm = -imm;
2471        opc = ARITH_SUB;
2472    }
2473    enc = encode_imm(imm);
2474    if (enc >= 0) {
2475        tcg_out_dat_imm(s, COND_AL, opc, rd, rs, enc);
2476    } else {
2477        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, imm);
2478        tcg_out_dat_reg(s, COND_AL, opc, rd, rs,
2479                        TCG_REG_TMP, SHIFT_IMM_LSL(0));
2480    }
2481}
2482
2483/* Type is always V128, with I64 elements.  */
2484static void tcg_out_dup2_vec(TCGContext *s, TCGReg rd, TCGReg rl, TCGReg rh)
2485{
2486    /* Move high element into place first. */
2487    /* VMOV Dd+1, Ds */
2488    tcg_out_vreg3(s, INSN_VORR | (1 << 12), 0, 0, rd, rh, rh);
2489    /* Move low element into place; tcg_out_mov will check for nop. */
2490    tcg_out_mov(s, TCG_TYPE_V64, rd, rl);
2491}
2492
2493static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
2494                            TCGReg rd, TCGReg rs)
2495{
2496    int q = type - TCG_TYPE_V64;
2497
2498    if (vece == MO_64) {
2499        if (type == TCG_TYPE_V128) {
2500            tcg_out_dup2_vec(s, rd, rs, rs);
2501        } else {
2502            tcg_out_mov(s, TCG_TYPE_V64, rd, rs);
2503        }
2504    } else if (rs < TCG_REG_Q0) {
2505        int b = (vece == MO_8);
2506        int e = (vece == MO_16);
2507        tcg_out32(s, INSN_VDUP_G | (b << 22) | (q << 21) | (e << 5) |
2508                  encode_vn(rd) | (rs << 12));
2509    } else {
2510        int imm4 = 1 << vece;
2511        tcg_out32(s, INSN_VDUP_S | (imm4 << 16) | (q << 6) |
2512                  encode_vd(rd) | encode_vm(rs));
2513    }
2514    return true;
2515}
2516
2517static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
2518                             TCGReg rd, TCGReg base, intptr_t offset)
2519{
2520    if (vece == MO_64) {
2521        tcg_out_ld(s, TCG_TYPE_V64, rd, base, offset);
2522        if (type == TCG_TYPE_V128) {
2523            tcg_out_dup2_vec(s, rd, rd, rd);
2524        }
2525    } else {
2526        int q = type - TCG_TYPE_V64;
2527        tcg_out_vldst(s, INSN_VLD1R | (vece << 6) | (q << 5),
2528                      rd, base, offset);
2529    }
2530    return true;
2531}
2532
2533static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
2534                             TCGReg rd, int64_t v64)
2535{
2536    int q = type - TCG_TYPE_V64;
2537    int cmode, imm8, i;
2538
2539    /* Test all bytes equal first.  */
2540    if (vece == MO_8) {
2541        tcg_out_vmovi(s, rd, q, 0, 0xe, v64);
2542        return;
2543    }
2544
2545    /*
2546     * Test all bytes 0x00 or 0xff second.  This can match cases that
2547     * might otherwise take 2 or 3 insns for MO_16 or MO_32 below.
2548     */
2549    for (i = imm8 = 0; i < 8; i++) {
2550        uint8_t byte = v64 >> (i * 8);
2551        if (byte == 0xff) {
2552            imm8 |= 1 << i;
2553        } else if (byte != 0) {
2554            goto fail_bytes;
2555        }
2556    }
2557    tcg_out_vmovi(s, rd, q, 1, 0xe, imm8);
2558    return;
2559 fail_bytes:
2560
2561    /*
2562     * Tests for various replications.  For each element width, if we
2563     * cannot find an expansion there's no point checking a larger
2564     * width because we already know by replication it cannot match.
2565     */
2566    if (vece == MO_16) {
2567        uint16_t v16 = v64;
2568
2569        if (is_shimm16(v16, &cmode, &imm8)) {
2570            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2571            return;
2572        }
2573        if (is_shimm16(~v16, &cmode, &imm8)) {
2574            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2575            return;
2576        }
2577
2578        /*
2579         * Otherwise, all remaining constants can be loaded in two insns:
2580         * rd = v16 & 0xff, rd |= v16 & 0xff00.
2581         */
2582        tcg_out_vmovi(s, rd, q, 0, 0x8, v16 & 0xff);
2583        tcg_out_vmovi(s, rd, q, 0, 0xb, v16 >> 8);   /* VORRI */
2584        return;
2585    }
2586
2587    if (vece == MO_32) {
2588        uint32_t v32 = v64;
2589
2590        if (is_shimm32(v32, &cmode, &imm8) ||
2591            is_soimm32(v32, &cmode, &imm8)) {
2592            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2593            return;
2594        }
2595        if (is_shimm32(~v32, &cmode, &imm8) ||
2596            is_soimm32(~v32, &cmode, &imm8)) {
2597            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2598            return;
2599        }
2600
2601        /*
2602         * Restrict the set of constants to those we can load with
2603         * two instructions.  Others we load from the pool.
2604         */
2605        i = is_shimm32_pair(v32, &cmode, &imm8);
2606        if (i) {
2607            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2608            tcg_out_vmovi(s, rd, q, 0, i | 1, extract32(v32, i * 4, 8));
2609            return;
2610        }
2611        i = is_shimm32_pair(~v32, &cmode, &imm8);
2612        if (i) {
2613            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2614            tcg_out_vmovi(s, rd, q, 1, i | 1, extract32(~v32, i * 4, 8));
2615            return;
2616        }
2617    }
2618
2619    /*
2620     * As a last resort, load from the constant pool.
2621     */
2622    if (!q || vece == MO_64) {
2623        new_pool_l2(s, R_ARM_PC11, s->code_ptr, 0, v64, v64 >> 32);
2624        /* VLDR Dd, [pc + offset] */
2625        tcg_out32(s, INSN_VLDR_D | encode_vd(rd) | (0xf << 16));
2626        if (q) {
2627            tcg_out_dup2_vec(s, rd, rd, rd);
2628        }
2629    } else {
2630        new_pool_label(s, (uint32_t)v64, R_ARM_PC8, s->code_ptr, 0);
2631        /* add tmp, pc, offset */
2632        tcg_out_dat_imm(s, COND_AL, ARITH_ADD, TCG_REG_TMP, TCG_REG_PC, 0);
2633        tcg_out_dupm_vec(s, type, MO_32, rd, TCG_REG_TMP, 0);
2634    }
2635}
2636
2637static const ARMInsn vec_cmp_insn[16] = {
2638    [TCG_COND_EQ] = INSN_VCEQ,
2639    [TCG_COND_GT] = INSN_VCGT,
2640    [TCG_COND_GE] = INSN_VCGE,
2641    [TCG_COND_GTU] = INSN_VCGT_U,
2642    [TCG_COND_GEU] = INSN_VCGE_U,
2643};
2644
2645static const ARMInsn vec_cmp0_insn[16] = {
2646    [TCG_COND_EQ] = INSN_VCEQ0,
2647    [TCG_COND_GT] = INSN_VCGT0,
2648    [TCG_COND_GE] = INSN_VCGE0,
2649    [TCG_COND_LT] = INSN_VCLT0,
2650    [TCG_COND_LE] = INSN_VCLE0,
2651};
2652
2653static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
2654                           unsigned vecl, unsigned vece,
2655                           const TCGArg args[TCG_MAX_OP_ARGS],
2656                           const int const_args[TCG_MAX_OP_ARGS])
2657{
2658    TCGType type = vecl + TCG_TYPE_V64;
2659    unsigned q = vecl;
2660    TCGArg a0, a1, a2, a3;
2661    int cmode, imm8;
2662
2663    a0 = args[0];
2664    a1 = args[1];
2665    a2 = args[2];
2666
2667    switch (opc) {
2668    case INDEX_op_ld_vec:
2669        tcg_out_ld(s, type, a0, a1, a2);
2670        return;
2671    case INDEX_op_st_vec:
2672        tcg_out_st(s, type, a0, a1, a2);
2673        return;
2674    case INDEX_op_dupm_vec:
2675        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
2676        return;
2677    case INDEX_op_dup2_vec:
2678        tcg_out_dup2_vec(s, a0, a1, a2);
2679        return;
2680    case INDEX_op_abs_vec:
2681        tcg_out_vreg2(s, INSN_VABS, q, vece, a0, a1);
2682        return;
2683    case INDEX_op_neg_vec:
2684        tcg_out_vreg2(s, INSN_VNEG, q, vece, a0, a1);
2685        return;
2686    case INDEX_op_not_vec:
2687        tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a1);
2688        return;
2689    case INDEX_op_add_vec:
2690        tcg_out_vreg3(s, INSN_VADD, q, vece, a0, a1, a2);
2691        return;
2692    case INDEX_op_mul_vec:
2693        tcg_out_vreg3(s, INSN_VMUL, q, vece, a0, a1, a2);
2694        return;
2695    case INDEX_op_smax_vec:
2696        tcg_out_vreg3(s, INSN_VMAX, q, vece, a0, a1, a2);
2697        return;
2698    case INDEX_op_smin_vec:
2699        tcg_out_vreg3(s, INSN_VMIN, q, vece, a0, a1, a2);
2700        return;
2701    case INDEX_op_sub_vec:
2702        tcg_out_vreg3(s, INSN_VSUB, q, vece, a0, a1, a2);
2703        return;
2704    case INDEX_op_ssadd_vec:
2705        tcg_out_vreg3(s, INSN_VQADD, q, vece, a0, a1, a2);
2706        return;
2707    case INDEX_op_sssub_vec:
2708        tcg_out_vreg3(s, INSN_VQSUB, q, vece, a0, a1, a2);
2709        return;
2710    case INDEX_op_umax_vec:
2711        tcg_out_vreg3(s, INSN_VMAX_U, q, vece, a0, a1, a2);
2712        return;
2713    case INDEX_op_umin_vec:
2714        tcg_out_vreg3(s, INSN_VMIN_U, q, vece, a0, a1, a2);
2715        return;
2716    case INDEX_op_usadd_vec:
2717        tcg_out_vreg3(s, INSN_VQADD_U, q, vece, a0, a1, a2);
2718        return;
2719    case INDEX_op_ussub_vec:
2720        tcg_out_vreg3(s, INSN_VQSUB_U, q, vece, a0, a1, a2);
2721        return;
2722    case INDEX_op_xor_vec:
2723        tcg_out_vreg3(s, INSN_VEOR, q, 0, a0, a1, a2);
2724        return;
2725    case INDEX_op_arm_sshl_vec:
2726        /*
2727         * Note that Vm is the data and Vn is the shift count,
2728         * therefore the arguments appear reversed.
2729         */
2730        tcg_out_vreg3(s, INSN_VSHL_S, q, vece, a0, a2, a1);
2731        return;
2732    case INDEX_op_arm_ushl_vec:
2733        /* See above. */
2734        tcg_out_vreg3(s, INSN_VSHL_U, q, vece, a0, a2, a1);
2735        return;
2736    case INDEX_op_shli_vec:
2737        tcg_out_vshifti(s, INSN_VSHLI, q, a0, a1, a2 + (8 << vece));
2738        return;
2739    case INDEX_op_shri_vec:
2740        tcg_out_vshifti(s, INSN_VSHRI, q, a0, a1, (16 << vece) - a2);
2741        return;
2742    case INDEX_op_sari_vec:
2743        tcg_out_vshifti(s, INSN_VSARI, q, a0, a1, (16 << vece) - a2);
2744        return;
2745    case INDEX_op_arm_sli_vec:
2746        tcg_out_vshifti(s, INSN_VSLI, q, a0, a2, args[3] + (8 << vece));
2747        return;
2748
2749    case INDEX_op_andc_vec:
2750        if (!const_args[2]) {
2751            tcg_out_vreg3(s, INSN_VBIC, q, 0, a0, a1, a2);
2752            return;
2753        }
2754        a2 = ~a2;
2755        /* fall through */
2756    case INDEX_op_and_vec:
2757        if (const_args[2]) {
2758            is_shimm1632(~a2, &cmode, &imm8);
2759            if (a0 == a1) {
2760                tcg_out_vmovi(s, a0, q, 1, cmode | 1, imm8); /* VBICI */
2761                return;
2762            }
2763            tcg_out_vmovi(s, a0, q, 1, cmode, imm8); /* VMVNI */
2764            a2 = a0;
2765        }
2766        tcg_out_vreg3(s, INSN_VAND, q, 0, a0, a1, a2);
2767        return;
2768
2769    case INDEX_op_orc_vec:
2770        if (!const_args[2]) {
2771            tcg_out_vreg3(s, INSN_VORN, q, 0, a0, a1, a2);
2772            return;
2773        }
2774        a2 = ~a2;
2775        /* fall through */
2776    case INDEX_op_or_vec:
2777        if (const_args[2]) {
2778            is_shimm1632(a2, &cmode, &imm8);
2779            if (a0 == a1) {
2780                tcg_out_vmovi(s, a0, q, 0, cmode | 1, imm8); /* VORRI */
2781                return;
2782            }
2783            tcg_out_vmovi(s, a0, q, 0, cmode, imm8); /* VMOVI */
2784            a2 = a0;
2785        }
2786        tcg_out_vreg3(s, INSN_VORR, q, 0, a0, a1, a2);
2787        return;
2788
2789    case INDEX_op_cmp_vec:
2790        {
2791            TCGCond cond = args[3];
2792            ARMInsn insn;
2793
2794            switch (cond) {
2795            case TCG_COND_NE:
2796                if (const_args[2]) {
2797                    tcg_out_vreg3(s, INSN_VTST, q, vece, a0, a1, a1);
2798                } else {
2799                    tcg_out_vreg3(s, INSN_VCEQ, q, vece, a0, a1, a2);
2800                    tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a0);
2801                }
2802                break;
2803
2804            case TCG_COND_TSTNE:
2805            case TCG_COND_TSTEQ:
2806                if (const_args[2]) {
2807                    /* (x & 0) == 0 */
2808                    tcg_out_dupi_vec(s, type, MO_8, a0,
2809                                     -(cond == TCG_COND_TSTEQ));
2810                    break;
2811                }
2812                tcg_out_vreg3(s, INSN_VTST, q, vece, a0, a1, a2);
2813                if (cond == TCG_COND_TSTEQ) {
2814                    tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a0);
2815                }
2816                break;
2817
2818            default:
2819                if (const_args[2]) {
2820                    insn = vec_cmp0_insn[cond];
2821                    if (insn) {
2822                        tcg_out_vreg2(s, insn, q, vece, a0, a1);
2823                        return;
2824                    }
2825                    tcg_out_dupi_vec(s, type, MO_8, TCG_VEC_TMP, 0);
2826                    a2 = TCG_VEC_TMP;
2827                }
2828                insn = vec_cmp_insn[cond];
2829                if (insn == 0) {
2830                    TCGArg t;
2831                    t = a1, a1 = a2, a2 = t;
2832                    cond = tcg_swap_cond(cond);
2833                    insn = vec_cmp_insn[cond];
2834                    tcg_debug_assert(insn != 0);
2835                }
2836                tcg_out_vreg3(s, insn, q, vece, a0, a1, a2);
2837                break;
2838            }
2839        }
2840        return;
2841
2842    case INDEX_op_bitsel_vec:
2843        a3 = args[3];
2844        if (a0 == a3) {
2845            tcg_out_vreg3(s, INSN_VBIT, q, 0, a0, a2, a1);
2846        } else if (a0 == a2) {
2847            tcg_out_vreg3(s, INSN_VBIF, q, 0, a0, a3, a1);
2848        } else {
2849            tcg_out_mov(s, type, a0, a1);
2850            tcg_out_vreg3(s, INSN_VBSL, q, 0, a0, a2, a3);
2851        }
2852        return;
2853
2854    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
2855    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
2856    default:
2857        g_assert_not_reached();
2858    }
2859}
2860
2861int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
2862{
2863    switch (opc) {
2864    case INDEX_op_add_vec:
2865    case INDEX_op_sub_vec:
2866    case INDEX_op_and_vec:
2867    case INDEX_op_andc_vec:
2868    case INDEX_op_or_vec:
2869    case INDEX_op_orc_vec:
2870    case INDEX_op_xor_vec:
2871    case INDEX_op_not_vec:
2872    case INDEX_op_shli_vec:
2873    case INDEX_op_shri_vec:
2874    case INDEX_op_sari_vec:
2875    case INDEX_op_ssadd_vec:
2876    case INDEX_op_sssub_vec:
2877    case INDEX_op_usadd_vec:
2878    case INDEX_op_ussub_vec:
2879    case INDEX_op_bitsel_vec:
2880        return 1;
2881    case INDEX_op_abs_vec:
2882    case INDEX_op_cmp_vec:
2883    case INDEX_op_mul_vec:
2884    case INDEX_op_neg_vec:
2885    case INDEX_op_smax_vec:
2886    case INDEX_op_smin_vec:
2887    case INDEX_op_umax_vec:
2888    case INDEX_op_umin_vec:
2889        return vece < MO_64;
2890    case INDEX_op_shlv_vec:
2891    case INDEX_op_shrv_vec:
2892    case INDEX_op_sarv_vec:
2893    case INDEX_op_rotli_vec:
2894    case INDEX_op_rotlv_vec:
2895    case INDEX_op_rotrv_vec:
2896        return -1;
2897    default:
2898        return 0;
2899    }
2900}
2901
2902void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
2903                       TCGArg a0, ...)
2904{
2905    va_list va;
2906    TCGv_vec v0, v1, v2, t1, t2, c1;
2907    TCGArg a2;
2908
2909    va_start(va, a0);
2910    v0 = temp_tcgv_vec(arg_temp(a0));
2911    v1 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
2912    a2 = va_arg(va, TCGArg);
2913    va_end(va);
2914
2915    switch (opc) {
2916    case INDEX_op_shlv_vec:
2917        /*
2918         * Merely propagate shlv_vec to arm_ushl_vec.
2919         * In this way we don't set TCG_TARGET_HAS_shv_vec
2920         * because everything is done via expansion.
2921         */
2922        v2 = temp_tcgv_vec(arg_temp(a2));
2923        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
2924                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
2925        break;
2926
2927    case INDEX_op_shrv_vec:
2928    case INDEX_op_sarv_vec:
2929        /* Right shifts are negative left shifts for NEON.  */
2930        v2 = temp_tcgv_vec(arg_temp(a2));
2931        t1 = tcg_temp_new_vec(type);
2932        tcg_gen_neg_vec(vece, t1, v2);
2933        if (opc == INDEX_op_shrv_vec) {
2934            opc = INDEX_op_arm_ushl_vec;
2935        } else {
2936            opc = INDEX_op_arm_sshl_vec;
2937        }
2938        vec_gen_3(opc, type, vece, tcgv_vec_arg(v0),
2939                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
2940        tcg_temp_free_vec(t1);
2941        break;
2942
2943    case INDEX_op_rotli_vec:
2944        t1 = tcg_temp_new_vec(type);
2945        tcg_gen_shri_vec(vece, t1, v1, -a2 & ((8 << vece) - 1));
2946        vec_gen_4(INDEX_op_arm_sli_vec, type, vece,
2947                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(v1), a2);
2948        tcg_temp_free_vec(t1);
2949        break;
2950
2951    case INDEX_op_rotlv_vec:
2952        v2 = temp_tcgv_vec(arg_temp(a2));
2953        t1 = tcg_temp_new_vec(type);
2954        c1 = tcg_constant_vec(type, vece, 8 << vece);
2955        tcg_gen_sub_vec(vece, t1, v2, c1);
2956        /* Right shifts are negative left shifts for NEON.  */
2957        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
2958                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
2959        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
2960                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
2961        tcg_gen_or_vec(vece, v0, v0, t1);
2962        tcg_temp_free_vec(t1);
2963        break;
2964
2965    case INDEX_op_rotrv_vec:
2966        v2 = temp_tcgv_vec(arg_temp(a2));
2967        t1 = tcg_temp_new_vec(type);
2968        t2 = tcg_temp_new_vec(type);
2969        c1 = tcg_constant_vec(type, vece, 8 << vece);
2970        tcg_gen_neg_vec(vece, t1, v2);
2971        tcg_gen_sub_vec(vece, t2, c1, v2);
2972        /* Right shifts are negative left shifts for NEON.  */
2973        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
2974                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
2975        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t2),
2976                  tcgv_vec_arg(v1), tcgv_vec_arg(t2));
2977        tcg_gen_or_vec(vece, v0, t1, t2);
2978        tcg_temp_free_vec(t1);
2979        tcg_temp_free_vec(t2);
2980        break;
2981
2982    default:
2983        g_assert_not_reached();
2984    }
2985}
2986
2987static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
2988{
2989    int i;
2990    for (i = 0; i < count; ++i) {
2991        p[i] = INSN_NOP;
2992    }
2993}
2994
2995/* Compute frame size via macros, to share between tcg_target_qemu_prologue
2996   and tcg_register_jit.  */
2997
2998#define PUSH_SIZE  ((11 - 4 + 1 + 1) * sizeof(tcg_target_long))
2999
3000#define FRAME_SIZE \
3001    ((PUSH_SIZE \
3002      + TCG_STATIC_CALL_ARGS_SIZE \
3003      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
3004      + TCG_TARGET_STACK_ALIGN - 1) \
3005     & -TCG_TARGET_STACK_ALIGN)
3006
3007#define STACK_ADDEND  (FRAME_SIZE - PUSH_SIZE)
3008
3009static void tcg_target_qemu_prologue(TCGContext *s)
3010{
3011    /* Calling convention requires us to save r4-r11 and lr.  */
3012    /* stmdb sp!, { r4 - r11, lr } */
3013    tcg_out_ldstm(s, COND_AL, INSN_STMDB, TCG_REG_CALL_STACK,
3014                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
3015                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
3016                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_R14));
3017
3018    /* Reserve callee argument and tcg temp space.  */
3019    tcg_out_dat_rI(s, COND_AL, ARITH_SUB, TCG_REG_CALL_STACK,
3020                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
3021    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
3022                  CPU_TEMP_BUF_NLONGS * sizeof(long));
3023
3024    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
3025
3026    if (!tcg_use_softmmu && guest_base) {
3027        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_GUEST_BASE, guest_base);
3028        tcg_regset_set_reg(s->reserved_regs, TCG_REG_GUEST_BASE);
3029    }
3030
3031    tcg_out_b_reg(s, COND_AL, tcg_target_call_iarg_regs[1]);
3032
3033    /*
3034     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
3035     * and fall through to the rest of the epilogue.
3036     */
3037    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
3038    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, 0);
3039    tcg_out_epilogue(s);
3040}
3041
3042static void tcg_out_epilogue(TCGContext *s)
3043{
3044    /* Release local stack frame.  */
3045    tcg_out_dat_rI(s, COND_AL, ARITH_ADD, TCG_REG_CALL_STACK,
3046                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
3047
3048    /* ldmia sp!, { r4 - r11, pc } */
3049    tcg_out_ldstm(s, COND_AL, INSN_LDMIA, TCG_REG_CALL_STACK,
3050                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
3051                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
3052                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_PC));
3053}
3054
3055static void tcg_out_tb_start(TCGContext *s)
3056{
3057    /* nothing to do */
3058}
3059
3060typedef struct {
3061    DebugFrameHeader h;
3062    uint8_t fde_def_cfa[4];
3063    uint8_t fde_reg_ofs[18];
3064} DebugFrame;
3065
3066#define ELF_HOST_MACHINE EM_ARM
3067
3068/* We're expecting a 2 byte uleb128 encoded value.  */
3069QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
3070
3071static const DebugFrame debug_frame = {
3072    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
3073    .h.cie.id = -1,
3074    .h.cie.version = 1,
3075    .h.cie.code_align = 1,
3076    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
3077    .h.cie.return_column = 14,
3078
3079    /* Total FDE size does not include the "len" member.  */
3080    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
3081
3082    .fde_def_cfa = {
3083        12, 13,                         /* DW_CFA_def_cfa sp, ... */
3084        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
3085        (FRAME_SIZE >> 7)
3086    },
3087    .fde_reg_ofs = {
3088        /* The following must match the stmdb in the prologue.  */
3089        0x8e, 1,                        /* DW_CFA_offset, lr, -4 */
3090        0x8b, 2,                        /* DW_CFA_offset, r11, -8 */
3091        0x8a, 3,                        /* DW_CFA_offset, r10, -12 */
3092        0x89, 4,                        /* DW_CFA_offset, r9, -16 */
3093        0x88, 5,                        /* DW_CFA_offset, r8, -20 */
3094        0x87, 6,                        /* DW_CFA_offset, r7, -24 */
3095        0x86, 7,                        /* DW_CFA_offset, r6, -28 */
3096        0x85, 8,                        /* DW_CFA_offset, r5, -32 */
3097        0x84, 9,                        /* DW_CFA_offset, r4, -36 */
3098    }
3099};
3100
3101void tcg_register_jit(const void *buf, size_t buf_size)
3102{
3103    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
3104}
3105