xref: /openbmc/qemu/tcg/arm/tcg-target.c.inc (revision 46f96bff163512f9f8f9959de4a18c0799001422)
1/*
2 * Tiny Code Generator for QEMU
3 *
4 * Copyright (c) 2008 Andrzej Zaborowski
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to deal
8 * in the Software without restriction, including without limitation the rights
9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10 * copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22 * THE SOFTWARE.
23 */
24
25#include "elf.h"
26
27int arm_arch = __ARM_ARCH;
28
29#ifndef use_idiv_instructions
30bool use_idiv_instructions;
31#endif
32#ifndef use_neon_instructions
33bool use_neon_instructions;
34#endif
35
36/* Used for function call generation. */
37#define TCG_TARGET_STACK_ALIGN          8
38#define TCG_TARGET_CALL_STACK_OFFSET    0
39#define TCG_TARGET_CALL_ARG_I32         TCG_CALL_ARG_NORMAL
40#define TCG_TARGET_CALL_ARG_I64         TCG_CALL_ARG_EVEN
41#define TCG_TARGET_CALL_ARG_I128        TCG_CALL_ARG_EVEN
42#define TCG_TARGET_CALL_RET_I128        TCG_CALL_RET_BY_REF
43
44#ifdef CONFIG_DEBUG_TCG
45static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
46    "%r0",  "%r1",  "%r2",  "%r3",  "%r4",  "%r5",  "%r6",  "%r7",
47    "%r8",  "%r9",  "%r10", "%r11", "%r12", "%sp",  "%r14", "%pc",
48    "%q0",  "%q1",  "%q2",  "%q3",  "%q4",  "%q5",  "%q6",  "%q7",
49    "%q8",  "%q9",  "%q10", "%q11", "%q12", "%q13", "%q14", "%q15",
50};
51#endif
52
53static const int tcg_target_reg_alloc_order[] = {
54    TCG_REG_R4,
55    TCG_REG_R5,
56    TCG_REG_R6,
57    TCG_REG_R7,
58    TCG_REG_R8,
59    TCG_REG_R9,
60    TCG_REG_R10,
61    TCG_REG_R11,
62    TCG_REG_R13,
63    TCG_REG_R0,
64    TCG_REG_R1,
65    TCG_REG_R2,
66    TCG_REG_R3,
67    TCG_REG_R12,
68    TCG_REG_R14,
69
70    TCG_REG_Q0,
71    TCG_REG_Q1,
72    TCG_REG_Q2,
73    TCG_REG_Q3,
74    /* Q4 - Q7 are call-saved, and skipped. */
75    TCG_REG_Q8,
76    TCG_REG_Q9,
77    TCG_REG_Q10,
78    TCG_REG_Q11,
79    TCG_REG_Q12,
80    TCG_REG_Q13,
81    TCG_REG_Q14,
82    TCG_REG_Q15,
83};
84
85static const int tcg_target_call_iarg_regs[4] = {
86    TCG_REG_R0, TCG_REG_R1, TCG_REG_R2, TCG_REG_R3
87};
88
89static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
90{
91    tcg_debug_assert(kind == TCG_CALL_RET_NORMAL);
92    tcg_debug_assert(slot >= 0 && slot <= 3);
93    return TCG_REG_R0 + slot;
94}
95
96#define TCG_REG_TMP  TCG_REG_R12
97#define TCG_VEC_TMP  TCG_REG_Q15
98#define TCG_REG_GUEST_BASE  TCG_REG_R11
99
100typedef enum {
101    COND_EQ = 0x0,
102    COND_NE = 0x1,
103    COND_CS = 0x2,	/* Unsigned greater or equal */
104    COND_CC = 0x3,	/* Unsigned less than */
105    COND_MI = 0x4,	/* Negative */
106    COND_PL = 0x5,	/* Zero or greater */
107    COND_VS = 0x6,	/* Overflow */
108    COND_VC = 0x7,	/* No overflow */
109    COND_HI = 0x8,	/* Unsigned greater than */
110    COND_LS = 0x9,	/* Unsigned less or equal */
111    COND_GE = 0xa,
112    COND_LT = 0xb,
113    COND_GT = 0xc,
114    COND_LE = 0xd,
115    COND_AL = 0xe,
116} ARMCond;
117
118#define TO_CPSR (1 << 20)
119
120#define SHIFT_IMM_LSL(im)	(((im) << 7) | 0x00)
121#define SHIFT_IMM_LSR(im)	(((im) << 7) | 0x20)
122#define SHIFT_IMM_ASR(im)	(((im) << 7) | 0x40)
123#define SHIFT_IMM_ROR(im)	(((im) << 7) | 0x60)
124#define SHIFT_REG_LSL(rs)	(((rs) << 8) | 0x10)
125#define SHIFT_REG_LSR(rs)	(((rs) << 8) | 0x30)
126#define SHIFT_REG_ASR(rs)	(((rs) << 8) | 0x50)
127#define SHIFT_REG_ROR(rs)	(((rs) << 8) | 0x70)
128
129typedef enum {
130    ARITH_AND = 0x0 << 21,
131    ARITH_EOR = 0x1 << 21,
132    ARITH_SUB = 0x2 << 21,
133    ARITH_RSB = 0x3 << 21,
134    ARITH_ADD = 0x4 << 21,
135    ARITH_ADC = 0x5 << 21,
136    ARITH_SBC = 0x6 << 21,
137    ARITH_RSC = 0x7 << 21,
138    ARITH_TST = 0x8 << 21 | TO_CPSR,
139    ARITH_CMP = 0xa << 21 | TO_CPSR,
140    ARITH_CMN = 0xb << 21 | TO_CPSR,
141    ARITH_ORR = 0xc << 21,
142    ARITH_MOV = 0xd << 21,
143    ARITH_BIC = 0xe << 21,
144    ARITH_MVN = 0xf << 21,
145
146    INSN_B         = 0x0a000000,
147
148    INSN_CLZ       = 0x016f0f10,
149    INSN_RBIT      = 0x06ff0f30,
150
151    INSN_LDMIA     = 0x08b00000,
152    INSN_STMDB     = 0x09200000,
153
154    INSN_LDR_IMM   = 0x04100000,
155    INSN_LDR_REG   = 0x06100000,
156    INSN_STR_IMM   = 0x04000000,
157    INSN_STR_REG   = 0x06000000,
158
159    INSN_LDRH_IMM  = 0x005000b0,
160    INSN_LDRH_REG  = 0x001000b0,
161    INSN_LDRSH_IMM = 0x005000f0,
162    INSN_LDRSH_REG = 0x001000f0,
163    INSN_STRH_IMM  = 0x004000b0,
164    INSN_STRH_REG  = 0x000000b0,
165
166    INSN_LDRB_IMM  = 0x04500000,
167    INSN_LDRB_REG  = 0x06500000,
168    INSN_LDRSB_IMM = 0x005000d0,
169    INSN_LDRSB_REG = 0x001000d0,
170    INSN_STRB_IMM  = 0x04400000,
171    INSN_STRB_REG  = 0x06400000,
172
173    INSN_LDRD_IMM  = 0x004000d0,
174    INSN_LDRD_REG  = 0x000000d0,
175    INSN_STRD_IMM  = 0x004000f0,
176    INSN_STRD_REG  = 0x000000f0,
177
178    INSN_DMB_ISH   = 0xf57ff05b,
179    INSN_DMB_MCR   = 0xee070fba,
180
181    /* Architected nop introduced in v6k.  */
182    /* ??? This is an MSR (imm) 0,0,0 insn.  Anyone know if this
183       also Just So Happened to do nothing on pre-v6k so that we
184       don't need to conditionalize it?  */
185    INSN_NOP_v6k   = 0xe320f000,
186    /* Otherwise the assembler uses mov r0,r0 */
187    INSN_NOP_v4    = (COND_AL << 28) | ARITH_MOV,
188
189    INSN_VADD      = 0xf2000800,
190    INSN_VAND      = 0xf2000110,
191    INSN_VBIC      = 0xf2100110,
192    INSN_VEOR      = 0xf3000110,
193    INSN_VORN      = 0xf2300110,
194    INSN_VORR      = 0xf2200110,
195    INSN_VSUB      = 0xf3000800,
196    INSN_VMUL      = 0xf2000910,
197    INSN_VQADD     = 0xf2000010,
198    INSN_VQADD_U   = 0xf3000010,
199    INSN_VQSUB     = 0xf2000210,
200    INSN_VQSUB_U   = 0xf3000210,
201    INSN_VMAX      = 0xf2000600,
202    INSN_VMAX_U    = 0xf3000600,
203    INSN_VMIN      = 0xf2000610,
204    INSN_VMIN_U    = 0xf3000610,
205
206    INSN_VABS      = 0xf3b10300,
207    INSN_VMVN      = 0xf3b00580,
208    INSN_VNEG      = 0xf3b10380,
209
210    INSN_VCEQ0     = 0xf3b10100,
211    INSN_VCGT0     = 0xf3b10000,
212    INSN_VCGE0     = 0xf3b10080,
213    INSN_VCLE0     = 0xf3b10180,
214    INSN_VCLT0     = 0xf3b10200,
215
216    INSN_VCEQ      = 0xf3000810,
217    INSN_VCGE      = 0xf2000310,
218    INSN_VCGT      = 0xf2000300,
219    INSN_VCGE_U    = 0xf3000310,
220    INSN_VCGT_U    = 0xf3000300,
221
222    INSN_VSHLI     = 0xf2800510,  /* VSHL (immediate) */
223    INSN_VSARI     = 0xf2800010,  /* VSHR.S */
224    INSN_VSHRI     = 0xf3800010,  /* VSHR.U */
225    INSN_VSLI      = 0xf3800510,
226    INSN_VSHL_S    = 0xf2000400,  /* VSHL.S (register) */
227    INSN_VSHL_U    = 0xf3000400,  /* VSHL.U (register) */
228
229    INSN_VBSL      = 0xf3100110,
230    INSN_VBIT      = 0xf3200110,
231    INSN_VBIF      = 0xf3300110,
232
233    INSN_VTST      = 0xf2000810,
234
235    INSN_VDUP_G    = 0xee800b10,  /* VDUP (ARM core register) */
236    INSN_VDUP_S    = 0xf3b00c00,  /* VDUP (scalar) */
237    INSN_VLDR_D    = 0xed100b00,  /* VLDR.64 */
238    INSN_VLD1      = 0xf4200000,  /* VLD1 (multiple single elements) */
239    INSN_VLD1R     = 0xf4a00c00,  /* VLD1 (single element to all lanes) */
240    INSN_VST1      = 0xf4000000,  /* VST1 (multiple single elements) */
241    INSN_VMOVI     = 0xf2800010,  /* VMOV (immediate) */
242} ARMInsn;
243
244#define INSN_NOP   (use_armv7_instructions ? INSN_NOP_v6k : INSN_NOP_v4)
245
246static const uint8_t tcg_cond_to_arm_cond[] = {
247    [TCG_COND_EQ] = COND_EQ,
248    [TCG_COND_NE] = COND_NE,
249    [TCG_COND_LT] = COND_LT,
250    [TCG_COND_GE] = COND_GE,
251    [TCG_COND_LE] = COND_LE,
252    [TCG_COND_GT] = COND_GT,
253    /* unsigned */
254    [TCG_COND_LTU] = COND_CC,
255    [TCG_COND_GEU] = COND_CS,
256    [TCG_COND_LEU] = COND_LS,
257    [TCG_COND_GTU] = COND_HI,
258};
259
260static int encode_imm(uint32_t imm);
261
262/* TCG private relocation type: add with pc+imm8 */
263#define R_ARM_PC8  11
264
265/* TCG private relocation type: vldr with imm8 << 2 */
266#define R_ARM_PC11 12
267
268static bool reloc_pc24(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
269{
270    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
271    ptrdiff_t offset = (tcg_ptr_byte_diff(target, src_rx) - 8) >> 2;
272
273    if (offset == sextract32(offset, 0, 24)) {
274        *src_rw = deposit32(*src_rw, 0, 24, offset);
275        return true;
276    }
277    return false;
278}
279
280static bool reloc_pc13(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
281{
282    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
283    ptrdiff_t offset = tcg_ptr_byte_diff(target, src_rx) - 8;
284
285    if (offset >= -0xfff && offset <= 0xfff) {
286        tcg_insn_unit insn = *src_rw;
287        bool u = (offset >= 0);
288        if (!u) {
289            offset = -offset;
290        }
291        insn = deposit32(insn, 23, 1, u);
292        insn = deposit32(insn, 0, 12, offset);
293        *src_rw = insn;
294        return true;
295    }
296    return false;
297}
298
299static bool reloc_pc11(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
300{
301    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
302    ptrdiff_t offset = (tcg_ptr_byte_diff(target, src_rx) - 8) / 4;
303
304    if (offset >= -0xff && offset <= 0xff) {
305        tcg_insn_unit insn = *src_rw;
306        bool u = (offset >= 0);
307        if (!u) {
308            offset = -offset;
309        }
310        insn = deposit32(insn, 23, 1, u);
311        insn = deposit32(insn, 0, 8, offset);
312        *src_rw = insn;
313        return true;
314    }
315    return false;
316}
317
318static bool reloc_pc8(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
319{
320    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
321    ptrdiff_t offset = tcg_ptr_byte_diff(target, src_rx) - 8;
322    int imm12 = encode_imm(offset);
323
324    if (imm12 >= 0) {
325        *src_rw = deposit32(*src_rw, 0, 12, imm12);
326        return true;
327    }
328    return false;
329}
330
331static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
332                        intptr_t value, intptr_t addend)
333{
334    tcg_debug_assert(addend == 0);
335    switch (type) {
336    case R_ARM_PC24:
337        return reloc_pc24(code_ptr, (const tcg_insn_unit *)value);
338    case R_ARM_PC13:
339        return reloc_pc13(code_ptr, (const tcg_insn_unit *)value);
340    case R_ARM_PC11:
341        return reloc_pc11(code_ptr, (const tcg_insn_unit *)value);
342    case R_ARM_PC8:
343        return reloc_pc8(code_ptr, (const tcg_insn_unit *)value);
344    default:
345        g_assert_not_reached();
346    }
347}
348
349#define TCG_CT_CONST_ARM  0x100
350#define TCG_CT_CONST_INV  0x200
351#define TCG_CT_CONST_NEG  0x400
352#define TCG_CT_CONST_ZERO 0x800
353#define TCG_CT_CONST_ORRI 0x1000
354#define TCG_CT_CONST_ANDI 0x2000
355
356#define ALL_GENERAL_REGS  0xffffu
357#define ALL_VECTOR_REGS   0xffff0000u
358
359/*
360 * r0-r3 will be overwritten when reading the tlb entry (system-mode only);
361 * r14 will be overwritten by the BLNE branching to the slow path.
362 */
363#define ALL_QLDST_REGS \
364    (ALL_GENERAL_REGS & ~((tcg_use_softmmu ? 0xf : 0) | (1 << TCG_REG_R14)))
365
366/*
367 * ARM immediates for ALU instructions are made of an unsigned 8-bit
368 * right-rotated by an even amount between 0 and 30.
369 *
370 * Return < 0 if @imm cannot be encoded, else the entire imm12 field.
371 */
372static int encode_imm(uint32_t imm)
373{
374    uint32_t rot, imm8;
375
376    /* Simple case, no rotation required. */
377    if ((imm & ~0xff) == 0) {
378        return imm;
379    }
380
381    /* Next, try a simple even shift.  */
382    rot = ctz32(imm) & ~1;
383    imm8 = imm >> rot;
384    rot = 32 - rot;
385    if ((imm8 & ~0xff) == 0) {
386        goto found;
387    }
388
389    /*
390     * Finally, try harder with rotations.
391     * The ctz test above will have taken care of rotates >= 8.
392     */
393    for (rot = 2; rot < 8; rot += 2) {
394        imm8 = rol32(imm, rot);
395        if ((imm8 & ~0xff) == 0) {
396            goto found;
397        }
398    }
399    /* Fail: imm cannot be encoded. */
400    return -1;
401
402 found:
403    /* Note that rot is even, and we discard bit 0 by shifting by 7. */
404    return rot << 7 | imm8;
405}
406
407static int encode_imm_nofail(uint32_t imm)
408{
409    int ret = encode_imm(imm);
410    tcg_debug_assert(ret >= 0);
411    return ret;
412}
413
414static bool check_fit_imm(uint32_t imm)
415{
416    return encode_imm(imm) >= 0;
417}
418
419/* Return true if v16 is a valid 16-bit shifted immediate.  */
420static bool is_shimm16(uint16_t v16, int *cmode, int *imm8)
421{
422    if (v16 == (v16 & 0xff)) {
423        *cmode = 0x8;
424        *imm8 = v16 & 0xff;
425        return true;
426    } else if (v16 == (v16 & 0xff00)) {
427        *cmode = 0xa;
428        *imm8 = v16 >> 8;
429        return true;
430    }
431    return false;
432}
433
434/* Return true if v32 is a valid 32-bit shifted immediate.  */
435static bool is_shimm32(uint32_t v32, int *cmode, int *imm8)
436{
437    if (v32 == (v32 & 0xff)) {
438        *cmode = 0x0;
439        *imm8 = v32 & 0xff;
440        return true;
441    } else if (v32 == (v32 & 0xff00)) {
442        *cmode = 0x2;
443        *imm8 = (v32 >> 8) & 0xff;
444        return true;
445    } else if (v32 == (v32 & 0xff0000)) {
446        *cmode = 0x4;
447        *imm8 = (v32 >> 16) & 0xff;
448        return true;
449    } else if (v32 == (v32 & 0xff000000)) {
450        *cmode = 0x6;
451        *imm8 = v32 >> 24;
452        return true;
453    }
454    return false;
455}
456
457/* Return true if v32 is a valid 32-bit shifting ones immediate.  */
458static bool is_soimm32(uint32_t v32, int *cmode, int *imm8)
459{
460    if ((v32 & 0xffff00ff) == 0xff) {
461        *cmode = 0xc;
462        *imm8 = (v32 >> 8) & 0xff;
463        return true;
464    } else if ((v32 & 0xff00ffff) == 0xffff) {
465        *cmode = 0xd;
466        *imm8 = (v32 >> 16) & 0xff;
467        return true;
468    }
469    return false;
470}
471
472/*
473 * Return non-zero if v32 can be formed by MOVI+ORR.
474 * Place the parameters for MOVI in (cmode, imm8).
475 * Return the cmode for ORR; the imm8 can be had via extraction from v32.
476 */
477static int is_shimm32_pair(uint32_t v32, int *cmode, int *imm8)
478{
479    int i;
480
481    for (i = 6; i > 0; i -= 2) {
482        /* Mask out one byte we can add with ORR.  */
483        uint32_t tmp = v32 & ~(0xffu << (i * 4));
484        if (is_shimm32(tmp, cmode, imm8) ||
485            is_soimm32(tmp, cmode, imm8)) {
486            break;
487        }
488    }
489    return i;
490}
491
492/* Return true if V is a valid 16-bit or 32-bit shifted immediate.  */
493static bool is_shimm1632(uint32_t v32, int *cmode, int *imm8)
494{
495    if (v32 == deposit32(v32, 16, 16, v32)) {
496        return is_shimm16(v32, cmode, imm8);
497    } else {
498        return is_shimm32(v32, cmode, imm8);
499    }
500}
501
502/* Test if a constant matches the constraint.
503 * TODO: define constraints for:
504 *
505 * ldr/str offset:   between -0xfff and 0xfff
506 * ldrh/strh offset: between -0xff and 0xff
507 * mov operand2:     values represented with x << (2 * y), x < 0x100
508 * add, sub, eor...: ditto
509 */
510static bool tcg_target_const_match(int64_t val, int ct,
511                                   TCGType type, TCGCond cond, int vece)
512{
513    if (ct & TCG_CT_CONST) {
514        return 1;
515    } else if ((ct & TCG_CT_CONST_ARM) && check_fit_imm(val)) {
516        return 1;
517    } else if ((ct & TCG_CT_CONST_INV) && check_fit_imm(~val)) {
518        return 1;
519    } else if ((ct & TCG_CT_CONST_NEG) && check_fit_imm(-val)) {
520        return 1;
521    } else if ((ct & TCG_CT_CONST_ZERO) && val == 0) {
522        return 1;
523    }
524
525    switch (ct & (TCG_CT_CONST_ORRI | TCG_CT_CONST_ANDI)) {
526    case 0:
527        break;
528    case TCG_CT_CONST_ANDI:
529        val = ~val;
530        /* fallthru */
531    case TCG_CT_CONST_ORRI:
532        if (val == deposit64(val, 32, 32, val)) {
533            int cmode, imm8;
534            return is_shimm1632(val, &cmode, &imm8);
535        }
536        break;
537    default:
538        /* Both bits should not be set for the same insn.  */
539        g_assert_not_reached();
540    }
541
542    return 0;
543}
544
545static void tcg_out_b_imm(TCGContext *s, ARMCond cond, int32_t offset)
546{
547    tcg_out32(s, (cond << 28) | INSN_B |
548                    (((offset - 8) >> 2) & 0x00ffffff));
549}
550
551static void tcg_out_bl_imm(TCGContext *s, ARMCond cond, int32_t offset)
552{
553    tcg_out32(s, (cond << 28) | 0x0b000000 |
554                    (((offset - 8) >> 2) & 0x00ffffff));
555}
556
557static void tcg_out_blx_reg(TCGContext *s, ARMCond cond, TCGReg rn)
558{
559    tcg_out32(s, (cond << 28) | 0x012fff30 | rn);
560}
561
562static void tcg_out_blx_imm(TCGContext *s, int32_t offset)
563{
564    tcg_out32(s, 0xfa000000 | ((offset & 2) << 23) |
565                (((offset - 8) >> 2) & 0x00ffffff));
566}
567
568static void tcg_out_dat_reg(TCGContext *s, ARMCond cond, ARMInsn opc,
569                            TCGReg rd, TCGReg rn, TCGReg rm, int shift)
570{
571    tcg_out32(s, (cond << 28) | (0 << 25) | opc |
572                    (rn << 16) | (rd << 12) | shift | rm);
573}
574
575static void tcg_out_mov_reg(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rm)
576{
577    /* Simple reg-reg move, optimising out the 'do nothing' case */
578    if (rd != rm) {
579        tcg_out_dat_reg(s, cond, ARITH_MOV, rd, 0, rm, SHIFT_IMM_LSL(0));
580    }
581}
582
583static void tcg_out_bx_reg(TCGContext *s, ARMCond cond, TCGReg rn)
584{
585    tcg_out32(s, (cond << 28) | 0x012fff10 | rn);
586}
587
588static void tcg_out_b_reg(TCGContext *s, ARMCond cond, TCGReg rn)
589{
590    /*
591     * Unless the C portion of QEMU is compiled as thumb, we don't need
592     * true BX semantics; merely a branch to an address held in a register.
593     */
594    tcg_out_bx_reg(s, cond, rn);
595}
596
597static void tcg_out_dat_imm(TCGContext *s, ARMCond cond, ARMInsn opc,
598                            TCGReg rd, TCGReg rn, int im)
599{
600    tcg_out32(s, (cond << 28) | (1 << 25) | opc |
601                    (rn << 16) | (rd << 12) | im);
602}
603
604static void tcg_out_ldstm(TCGContext *s, ARMCond cond, ARMInsn opc,
605                          TCGReg rn, uint16_t mask)
606{
607    tcg_out32(s, (cond << 28) | opc | (rn << 16) | mask);
608}
609
610/* Note that this routine is used for both LDR and LDRH formats, so we do
611   not wish to include an immediate shift at this point.  */
612static void tcg_out_memop_r(TCGContext *s, ARMCond cond, ARMInsn opc, TCGReg rt,
613                            TCGReg rn, TCGReg rm, bool u, bool p, bool w)
614{
615    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24)
616              | (w << 21) | (rn << 16) | (rt << 12) | rm);
617}
618
619static void tcg_out_memop_8(TCGContext *s, ARMCond cond, ARMInsn opc, TCGReg rt,
620                            TCGReg rn, int imm8, bool p, bool w)
621{
622    bool u = 1;
623    if (imm8 < 0) {
624        imm8 = -imm8;
625        u = 0;
626    }
627    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24) | (w << 21) |
628              (rn << 16) | (rt << 12) | ((imm8 & 0xf0) << 4) | (imm8 & 0xf));
629}
630
631static void tcg_out_memop_12(TCGContext *s, ARMCond cond, ARMInsn opc,
632                             TCGReg rt, TCGReg rn, int imm12, bool p, bool w)
633{
634    bool u = 1;
635    if (imm12 < 0) {
636        imm12 = -imm12;
637        u = 0;
638    }
639    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24) | (w << 21) |
640              (rn << 16) | (rt << 12) | imm12);
641}
642
643static void tcg_out_ld32_12(TCGContext *s, ARMCond cond, TCGReg rt,
644                            TCGReg rn, int imm12)
645{
646    tcg_out_memop_12(s, cond, INSN_LDR_IMM, rt, rn, imm12, 1, 0);
647}
648
649static void tcg_out_st32_12(TCGContext *s, ARMCond cond, TCGReg rt,
650                            TCGReg rn, int imm12)
651{
652    tcg_out_memop_12(s, cond, INSN_STR_IMM, rt, rn, imm12, 1, 0);
653}
654
655static void tcg_out_ld32_r(TCGContext *s, ARMCond cond, TCGReg rt,
656                           TCGReg rn, TCGReg rm)
657{
658    tcg_out_memop_r(s, cond, INSN_LDR_REG, rt, rn, rm, 1, 1, 0);
659}
660
661static void tcg_out_st32_r(TCGContext *s, ARMCond cond, TCGReg rt,
662                           TCGReg rn, TCGReg rm)
663{
664    tcg_out_memop_r(s, cond, INSN_STR_REG, rt, rn, rm, 1, 1, 0);
665}
666
667static void tcg_out_ldrd_8(TCGContext *s, ARMCond cond, TCGReg rt,
668                           TCGReg rn, int imm8)
669{
670    tcg_out_memop_8(s, cond, INSN_LDRD_IMM, rt, rn, imm8, 1, 0);
671}
672
673static void tcg_out_ldrd_r(TCGContext *s, ARMCond cond, TCGReg rt,
674                           TCGReg rn, TCGReg rm)
675{
676    tcg_out_memop_r(s, cond, INSN_LDRD_REG, rt, rn, rm, 1, 1, 0);
677}
678
679static void tcg_out_strd_8(TCGContext *s, ARMCond cond, TCGReg rt,
680                           TCGReg rn, int imm8)
681{
682    tcg_out_memop_8(s, cond, INSN_STRD_IMM, rt, rn, imm8, 1, 0);
683}
684
685static void tcg_out_strd_r(TCGContext *s, ARMCond cond, TCGReg rt,
686                           TCGReg rn, TCGReg rm)
687{
688    tcg_out_memop_r(s, cond, INSN_STRD_REG, rt, rn, rm, 1, 1, 0);
689}
690
691/* Register pre-increment with base writeback.  */
692static void tcg_out_ld32_rwb(TCGContext *s, ARMCond cond, TCGReg rt,
693                             TCGReg rn, TCGReg rm)
694{
695    tcg_out_memop_r(s, cond, INSN_LDR_REG, rt, rn, rm, 1, 1, 1);
696}
697
698static void tcg_out_st32_rwb(TCGContext *s, ARMCond cond, TCGReg rt,
699                             TCGReg rn, TCGReg rm)
700{
701    tcg_out_memop_r(s, cond, INSN_STR_REG, rt, rn, rm, 1, 1, 1);
702}
703
704static void tcg_out_ld16u_8(TCGContext *s, ARMCond cond, TCGReg rt,
705                            TCGReg rn, int imm8)
706{
707    tcg_out_memop_8(s, cond, INSN_LDRH_IMM, rt, rn, imm8, 1, 0);
708}
709
710static void tcg_out_st16_8(TCGContext *s, ARMCond cond, TCGReg rt,
711                           TCGReg rn, int imm8)
712{
713    tcg_out_memop_8(s, cond, INSN_STRH_IMM, rt, rn, imm8, 1, 0);
714}
715
716static void tcg_out_ld16u_r(TCGContext *s, ARMCond cond, TCGReg rt,
717                            TCGReg rn, TCGReg rm)
718{
719    tcg_out_memop_r(s, cond, INSN_LDRH_REG, rt, rn, rm, 1, 1, 0);
720}
721
722static void tcg_out_st16_r(TCGContext *s, ARMCond cond, TCGReg rt,
723                           TCGReg rn, TCGReg rm)
724{
725    tcg_out_memop_r(s, cond, INSN_STRH_REG, rt, rn, rm, 1, 1, 0);
726}
727
728static void tcg_out_ld16s_8(TCGContext *s, ARMCond cond, TCGReg rt,
729                            TCGReg rn, int imm8)
730{
731    tcg_out_memop_8(s, cond, INSN_LDRSH_IMM, rt, rn, imm8, 1, 0);
732}
733
734static void tcg_out_ld16s_r(TCGContext *s, ARMCond cond, TCGReg rt,
735                            TCGReg rn, TCGReg rm)
736{
737    tcg_out_memop_r(s, cond, INSN_LDRSH_REG, rt, rn, rm, 1, 1, 0);
738}
739
740static void tcg_out_ld8_12(TCGContext *s, ARMCond cond, TCGReg rt,
741                           TCGReg rn, int imm12)
742{
743    tcg_out_memop_12(s, cond, INSN_LDRB_IMM, rt, rn, imm12, 1, 0);
744}
745
746static void tcg_out_st8_12(TCGContext *s, ARMCond cond, TCGReg rt,
747                           TCGReg rn, int imm12)
748{
749    tcg_out_memop_12(s, cond, INSN_STRB_IMM, rt, rn, imm12, 1, 0);
750}
751
752static void tcg_out_ld8_r(TCGContext *s, ARMCond cond, TCGReg rt,
753                          TCGReg rn, TCGReg rm)
754{
755    tcg_out_memop_r(s, cond, INSN_LDRB_REG, rt, rn, rm, 1, 1, 0);
756}
757
758static void tcg_out_st8_r(TCGContext *s, ARMCond cond, TCGReg rt,
759                          TCGReg rn, TCGReg rm)
760{
761    tcg_out_memop_r(s, cond, INSN_STRB_REG, rt, rn, rm, 1, 1, 0);
762}
763
764static void tcg_out_ld8s_8(TCGContext *s, ARMCond cond, TCGReg rt,
765                           TCGReg rn, int imm8)
766{
767    tcg_out_memop_8(s, cond, INSN_LDRSB_IMM, rt, rn, imm8, 1, 0);
768}
769
770static void tcg_out_ld8s_r(TCGContext *s, ARMCond cond, TCGReg rt,
771                           TCGReg rn, TCGReg rm)
772{
773    tcg_out_memop_r(s, cond, INSN_LDRSB_REG, rt, rn, rm, 1, 1, 0);
774}
775
776static void tcg_out_movi_pool(TCGContext *s, ARMCond cond,
777                              TCGReg rd, uint32_t arg)
778{
779    new_pool_label(s, arg, R_ARM_PC13, s->code_ptr, 0);
780    tcg_out_ld32_12(s, cond, rd, TCG_REG_PC, 0);
781}
782
783static void tcg_out_movi32(TCGContext *s, ARMCond cond,
784                           TCGReg rd, uint32_t arg)
785{
786    int imm12, diff, opc, sh1, sh2;
787    uint32_t tt0, tt1, tt2;
788
789    /* Check a single MOV/MVN before anything else.  */
790    imm12 = encode_imm(arg);
791    if (imm12 >= 0) {
792        tcg_out_dat_imm(s, cond, ARITH_MOV, rd, 0, imm12);
793        return;
794    }
795    imm12 = encode_imm(~arg);
796    if (imm12 >= 0) {
797        tcg_out_dat_imm(s, cond, ARITH_MVN, rd, 0, imm12);
798        return;
799    }
800
801    /* Check for a pc-relative address.  This will usually be the TB,
802       or within the TB, which is immediately before the code block.  */
803    diff = tcg_pcrel_diff(s, (void *)arg) - 8;
804    if (diff >= 0) {
805        imm12 = encode_imm(diff);
806        if (imm12 >= 0) {
807            tcg_out_dat_imm(s, cond, ARITH_ADD, rd, TCG_REG_PC, imm12);
808            return;
809        }
810    } else {
811        imm12 = encode_imm(-diff);
812        if (imm12 >= 0) {
813            tcg_out_dat_imm(s, cond, ARITH_SUB, rd, TCG_REG_PC, imm12);
814            return;
815        }
816    }
817
818    /* Use movw + movt.  */
819    if (use_armv7_instructions) {
820        /* movw */
821        tcg_out32(s, (cond << 28) | 0x03000000 | (rd << 12)
822                  | ((arg << 4) & 0x000f0000) | (arg & 0xfff));
823        if (arg & 0xffff0000) {
824            /* movt */
825            tcg_out32(s, (cond << 28) | 0x03400000 | (rd << 12)
826                      | ((arg >> 12) & 0x000f0000) | ((arg >> 16) & 0xfff));
827        }
828        return;
829    }
830
831    /* Look for sequences of two insns.  If we have lots of 1's, we can
832       shorten the sequence by beginning with mvn and then clearing
833       higher bits with eor.  */
834    tt0 = arg;
835    opc = ARITH_MOV;
836    if (ctpop32(arg) > 16) {
837        tt0 = ~arg;
838        opc = ARITH_MVN;
839    }
840    sh1 = ctz32(tt0) & ~1;
841    tt1 = tt0 & ~(0xff << sh1);
842    sh2 = ctz32(tt1) & ~1;
843    tt2 = tt1 & ~(0xff << sh2);
844    if (tt2 == 0) {
845        int rot;
846
847        rot = ((32 - sh1) << 7) & 0xf00;
848        tcg_out_dat_imm(s, cond, opc, rd,  0, ((tt0 >> sh1) & 0xff) | rot);
849        rot = ((32 - sh2) << 7) & 0xf00;
850        tcg_out_dat_imm(s, cond, ARITH_EOR, rd, rd,
851                        ((tt0 >> sh2) & 0xff) | rot);
852        return;
853    }
854
855    /* Otherwise, drop it into the constant pool.  */
856    tcg_out_movi_pool(s, cond, rd, arg);
857}
858
859/*
860 * Emit either the reg,imm or reg,reg form of a data-processing insn.
861 * rhs must satisfy the "rI" constraint.
862 */
863static void tcg_out_dat_rI(TCGContext *s, ARMCond cond, ARMInsn opc,
864                           TCGReg dst, TCGReg lhs, TCGArg rhs, int rhs_is_const)
865{
866    if (rhs_is_const) {
867        tcg_out_dat_imm(s, cond, opc, dst, lhs, encode_imm_nofail(rhs));
868    } else {
869        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
870    }
871}
872
873/*
874 * Emit either the reg,imm or reg,reg form of a data-processing insn.
875 * rhs must satisfy the "rIK" constraint.
876 */
877static void tcg_out_dat_IK(TCGContext *s, ARMCond cond, ARMInsn opc,
878                            ARMInsn opinv, TCGReg dst, TCGReg lhs, TCGArg rhs)
879{
880    int imm12 = encode_imm(rhs);
881    if (imm12 < 0) {
882        imm12 = encode_imm_nofail(~rhs);
883        opc = opinv;
884    }
885    tcg_out_dat_imm(s, cond, opc, dst, lhs, imm12);
886}
887
888static void tcg_out_dat_rIK(TCGContext *s, ARMCond cond, ARMInsn opc,
889                            ARMInsn opinv, TCGReg dst, TCGReg lhs, TCGArg rhs,
890                            bool rhs_is_const)
891{
892    if (rhs_is_const) {
893        tcg_out_dat_IK(s, cond, opc, opinv, dst, lhs, rhs);
894    } else {
895        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
896    }
897}
898
899static void tcg_out_dat_IN(TCGContext *s, ARMCond cond, ARMInsn opc,
900                           ARMInsn opneg, TCGReg dst, TCGReg lhs, TCGArg rhs)
901{
902    int imm12 = encode_imm(rhs);
903    if (imm12 < 0) {
904        imm12 = encode_imm_nofail(-rhs);
905        opc = opneg;
906    }
907    tcg_out_dat_imm(s, cond, opc, dst, lhs, imm12);
908}
909
910static void tcg_out_dat_rIN(TCGContext *s, ARMCond cond, ARMInsn opc,
911                            ARMInsn opneg, TCGReg dst, TCGReg lhs, TCGArg rhs,
912                            bool rhs_is_const)
913{
914    /* Emit either the reg,imm or reg,reg form of a data-processing insn.
915     * rhs must satisfy the "rIN" constraint.
916     */
917    if (rhs_is_const) {
918        tcg_out_dat_IN(s, cond, opc, opneg, dst, lhs, rhs);
919    } else {
920        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
921    }
922}
923
924static void tcg_out_mul32(TCGContext *s, ARMCond cond, TCGReg rd,
925                          TCGReg rn, TCGReg rm)
926{
927    /* mul */
928    tcg_out32(s, (cond << 28) | 0x90 | (rd << 16) | (rm << 8) | rn);
929}
930
931static void tcg_out_umull32(TCGContext *s, ARMCond cond, TCGReg rd0,
932                            TCGReg rd1, TCGReg rn, TCGReg rm)
933{
934    /* umull */
935    tcg_out32(s, (cond << 28) | 0x00800090 |
936              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
937}
938
939static void tcg_out_smull32(TCGContext *s, ARMCond cond, TCGReg rd0,
940                            TCGReg rd1, TCGReg rn, TCGReg rm)
941{
942    /* smull */
943    tcg_out32(s, (cond << 28) | 0x00c00090 |
944              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
945}
946
947static void tcg_out_sdiv(TCGContext *s, ARMCond cond,
948                         TCGReg rd, TCGReg rn, TCGReg rm)
949{
950    tcg_out32(s, 0x0710f010 | (cond << 28) | (rd << 16) | rn | (rm << 8));
951}
952
953static void tcg_out_udiv(TCGContext *s, ARMCond cond,
954                         TCGReg rd, TCGReg rn, TCGReg rm)
955{
956    tcg_out32(s, 0x0730f010 | (cond << 28) | (rd << 16) | rn | (rm << 8));
957}
958
959static void tcg_out_ext8s(TCGContext *s, TCGType t, TCGReg rd, TCGReg rn)
960{
961    /* sxtb */
962    tcg_out32(s, 0x06af0070 | (COND_AL << 28) | (rd << 12) | rn);
963}
964
965static void tcg_out_ext8u(TCGContext *s, TCGReg rd, TCGReg rn)
966{
967    tcg_out_dat_imm(s, COND_AL, ARITH_AND, rd, rn, 0xff);
968}
969
970static void tcg_out_ext16s(TCGContext *s, TCGType t, TCGReg rd, TCGReg rn)
971{
972    /* sxth */
973    tcg_out32(s, 0x06bf0070 | (COND_AL << 28) | (rd << 12) | rn);
974}
975
976static void tcg_out_ext16u(TCGContext *s, TCGReg rd, TCGReg rn)
977{
978    /* uxth */
979    tcg_out32(s, 0x06ff0070 | (COND_AL << 28) | (rd << 12) | rn);
980}
981
982static void tcg_out_ext32s(TCGContext *s, TCGReg rd, TCGReg rn)
983{
984    g_assert_not_reached();
985}
986
987static void tcg_out_ext32u(TCGContext *s, TCGReg rd, TCGReg rn)
988{
989    g_assert_not_reached();
990}
991
992static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg rd, TCGReg rn)
993{
994    g_assert_not_reached();
995}
996
997static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg rd, TCGReg rn)
998{
999    g_assert_not_reached();
1000}
1001
1002static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg rd, TCGReg rn)
1003{
1004    g_assert_not_reached();
1005}
1006
1007static void tcg_out_bswap16(TCGContext *s, ARMCond cond,
1008                            TCGReg rd, TCGReg rn, int flags)
1009{
1010    if (flags & TCG_BSWAP_OS) {
1011        /* revsh */
1012        tcg_out32(s, 0x06ff0fb0 | (cond << 28) | (rd << 12) | rn);
1013        return;
1014    }
1015
1016    /* rev16 */
1017    tcg_out32(s, 0x06bf0fb0 | (cond << 28) | (rd << 12) | rn);
1018    if ((flags & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1019        /* uxth */
1020        tcg_out32(s, 0x06ff0070 | (cond << 28) | (rd << 12) | rd);
1021    }
1022}
1023
1024static void tcg_out_bswap32(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rn)
1025{
1026    /* rev */
1027    tcg_out32(s, 0x06bf0f30 | (cond << 28) | (rd << 12) | rn);
1028}
1029
1030static void tcg_out_deposit(TCGContext *s, ARMCond cond, TCGReg rd,
1031                            TCGArg a1, int ofs, int len, bool const_a1)
1032{
1033    if (const_a1) {
1034        /* bfi becomes bfc with rn == 15.  */
1035        a1 = 15;
1036    }
1037    /* bfi/bfc */
1038    tcg_out32(s, 0x07c00010 | (cond << 28) | (rd << 12) | a1
1039              | (ofs << 7) | ((ofs + len - 1) << 16));
1040}
1041
1042static void tcg_out_extract(TCGContext *s, ARMCond cond, TCGReg rd,
1043                            TCGReg rn, int ofs, int len)
1044{
1045    /* According to gcc, AND can be faster. */
1046    if (ofs == 0 && len <= 8) {
1047        tcg_out_dat_imm(s, cond, ARITH_AND, rd, rn,
1048                        encode_imm_nofail((1 << len) - 1));
1049        return;
1050    }
1051
1052    if (use_armv7_instructions) {
1053        /* ubfx */
1054        tcg_out32(s, 0x07e00050 | (cond << 28) | (rd << 12) | rn
1055                  | (ofs << 7) | ((len - 1) << 16));
1056        return;
1057    }
1058
1059    assert(ofs % 8 == 0);
1060    switch (len) {
1061    case 8:
1062        /* uxtb */
1063        tcg_out32(s, 0x06ef0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1064        break;
1065    case 16:
1066        /* uxth */
1067        tcg_out32(s, 0x06ff0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1068        break;
1069    default:
1070        g_assert_not_reached();
1071    }
1072}
1073
1074static void tcg_out_sextract(TCGContext *s, ARMCond cond, TCGReg rd,
1075                             TCGReg rn, int ofs, int len)
1076{
1077    if (use_armv7_instructions) {
1078        /* sbfx */
1079        tcg_out32(s, 0x07a00050 | (cond << 28) | (rd << 12) | rn
1080                  | (ofs << 7) | ((len - 1) << 16));
1081        return;
1082    }
1083
1084    assert(ofs % 8 == 0);
1085    switch (len) {
1086    case 8:
1087        /* sxtb */
1088        tcg_out32(s, 0x06af0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1089        break;
1090    case 16:
1091        /* sxth */
1092        tcg_out32(s, 0x06bf0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1093        break;
1094    default:
1095        g_assert_not_reached();
1096    }
1097}
1098
1099
1100static void tcg_out_ld32u(TCGContext *s, ARMCond cond,
1101                          TCGReg rd, TCGReg rn, int32_t offset)
1102{
1103    if (offset > 0xfff || offset < -0xfff) {
1104        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1105        tcg_out_ld32_r(s, cond, rd, rn, TCG_REG_TMP);
1106    } else
1107        tcg_out_ld32_12(s, cond, rd, rn, offset);
1108}
1109
1110static void tcg_out_st32(TCGContext *s, ARMCond cond,
1111                         TCGReg rd, TCGReg rn, int32_t offset)
1112{
1113    if (offset > 0xfff || offset < -0xfff) {
1114        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1115        tcg_out_st32_r(s, cond, rd, rn, TCG_REG_TMP);
1116    } else
1117        tcg_out_st32_12(s, cond, rd, rn, offset);
1118}
1119
1120static void tcg_out_ld16u(TCGContext *s, ARMCond cond,
1121                          TCGReg rd, TCGReg rn, int32_t offset)
1122{
1123    if (offset > 0xff || offset < -0xff) {
1124        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1125        tcg_out_ld16u_r(s, cond, rd, rn, TCG_REG_TMP);
1126    } else
1127        tcg_out_ld16u_8(s, cond, rd, rn, offset);
1128}
1129
1130static void tcg_out_ld16s(TCGContext *s, ARMCond cond,
1131                          TCGReg rd, TCGReg rn, int32_t offset)
1132{
1133    if (offset > 0xff || offset < -0xff) {
1134        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1135        tcg_out_ld16s_r(s, cond, rd, rn, TCG_REG_TMP);
1136    } else
1137        tcg_out_ld16s_8(s, cond, rd, rn, offset);
1138}
1139
1140static void tcg_out_st16(TCGContext *s, ARMCond cond,
1141                         TCGReg rd, TCGReg rn, int32_t offset)
1142{
1143    if (offset > 0xff || offset < -0xff) {
1144        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1145        tcg_out_st16_r(s, cond, rd, rn, TCG_REG_TMP);
1146    } else
1147        tcg_out_st16_8(s, cond, rd, rn, offset);
1148}
1149
1150static void tcg_out_ld8u(TCGContext *s, ARMCond cond,
1151                         TCGReg rd, TCGReg rn, int32_t offset)
1152{
1153    if (offset > 0xfff || offset < -0xfff) {
1154        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1155        tcg_out_ld8_r(s, cond, rd, rn, TCG_REG_TMP);
1156    } else
1157        tcg_out_ld8_12(s, cond, rd, rn, offset);
1158}
1159
1160static void tcg_out_ld8s(TCGContext *s, ARMCond cond,
1161                         TCGReg rd, TCGReg rn, int32_t offset)
1162{
1163    if (offset > 0xff || offset < -0xff) {
1164        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1165        tcg_out_ld8s_r(s, cond, rd, rn, TCG_REG_TMP);
1166    } else
1167        tcg_out_ld8s_8(s, cond, rd, rn, offset);
1168}
1169
1170static void tcg_out_st8(TCGContext *s, ARMCond cond,
1171                        TCGReg rd, TCGReg rn, int32_t offset)
1172{
1173    if (offset > 0xfff || offset < -0xfff) {
1174        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1175        tcg_out_st8_r(s, cond, rd, rn, TCG_REG_TMP);
1176    } else
1177        tcg_out_st8_12(s, cond, rd, rn, offset);
1178}
1179
1180/*
1181 * The _goto case is normally between TBs within the same code buffer, and
1182 * with the code buffer limited to 16MB we wouldn't need the long case.
1183 * But we also use it for the tail-call to the qemu_ld/st helpers, which does.
1184 */
1185static void tcg_out_goto(TCGContext *s, ARMCond cond, const tcg_insn_unit *addr)
1186{
1187    intptr_t addri = (intptr_t)addr;
1188    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1189    bool arm_mode = !(addri & 1);
1190
1191    if (arm_mode && disp - 8 < 0x01fffffd && disp - 8 > -0x01fffffd) {
1192        tcg_out_b_imm(s, cond, disp);
1193        return;
1194    }
1195
1196    /* LDR is interworking from v5t. */
1197    tcg_out_movi_pool(s, cond, TCG_REG_PC, addri);
1198}
1199
1200/*
1201 * The call case is mostly used for helpers - so it's not unreasonable
1202 * for them to be beyond branch range.
1203 */
1204static void tcg_out_call_int(TCGContext *s, const tcg_insn_unit *addr)
1205{
1206    intptr_t addri = (intptr_t)addr;
1207    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1208    bool arm_mode = !(addri & 1);
1209
1210    if (disp - 8 < 0x02000000 && disp - 8 >= -0x02000000) {
1211        if (arm_mode) {
1212            tcg_out_bl_imm(s, COND_AL, disp);
1213        } else {
1214            tcg_out_blx_imm(s, disp);
1215        }
1216        return;
1217    }
1218
1219    tcg_out_movi32(s, COND_AL, TCG_REG_TMP, addri);
1220    tcg_out_blx_reg(s, COND_AL, TCG_REG_TMP);
1221}
1222
1223static void tcg_out_call(TCGContext *s, const tcg_insn_unit *addr,
1224                         const TCGHelperInfo *info)
1225{
1226    tcg_out_call_int(s, addr);
1227}
1228
1229static void tcg_out_goto_label(TCGContext *s, ARMCond cond, TCGLabel *l)
1230{
1231    if (l->has_value) {
1232        tcg_out_goto(s, cond, l->u.value_ptr);
1233    } else {
1234        tcg_out_reloc(s, s->code_ptr, R_ARM_PC24, l, 0);
1235        tcg_out_b_imm(s, cond, 0);
1236    }
1237}
1238
1239static void tcg_out_mb(TCGContext *s, TCGArg a0)
1240{
1241    if (use_armv7_instructions) {
1242        tcg_out32(s, INSN_DMB_ISH);
1243    } else {
1244        tcg_out32(s, INSN_DMB_MCR);
1245    }
1246}
1247
1248static TCGCond tcg_out_cmp(TCGContext *s, TCGCond cond, TCGReg a,
1249                           TCGArg b, int b_const)
1250{
1251    if (!is_tst_cond(cond)) {
1252        tcg_out_dat_rIN(s, COND_AL, ARITH_CMP, ARITH_CMN, 0, a, b, b_const);
1253        return cond;
1254    }
1255
1256    cond = tcg_tst_eqne_cond(cond);
1257    if (b_const) {
1258        int imm12 = encode_imm(b);
1259
1260        /*
1261         * The compare constraints allow rIN, but TST does not support N.
1262         * Be prepared to load the constant into a scratch register.
1263         */
1264        if (imm12 >= 0) {
1265            tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, a, imm12);
1266            return cond;
1267        }
1268        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, b);
1269        b = TCG_REG_TMP;
1270    }
1271    tcg_out_dat_reg(s, COND_AL, ARITH_TST, 0, a, b, SHIFT_IMM_LSL(0));
1272    return cond;
1273}
1274
1275static TCGCond tcg_out_cmp2(TCGContext *s, const TCGArg *args,
1276                            const int *const_args)
1277{
1278    TCGReg al = args[0];
1279    TCGReg ah = args[1];
1280    TCGArg bl = args[2];
1281    TCGArg bh = args[3];
1282    TCGCond cond = args[4];
1283    int const_bl = const_args[2];
1284    int const_bh = const_args[3];
1285
1286    switch (cond) {
1287    case TCG_COND_EQ:
1288    case TCG_COND_NE:
1289    case TCG_COND_LTU:
1290    case TCG_COND_LEU:
1291    case TCG_COND_GTU:
1292    case TCG_COND_GEU:
1293        /*
1294         * We perform a conditional comparison.  If the high half is
1295         * equal, then overwrite the flags with the comparison of the
1296         * low half.  The resulting flags cover the whole.
1297         */
1298        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, ah, bh, const_bh);
1299        tcg_out_dat_rI(s, COND_EQ, ARITH_CMP, 0, al, bl, const_bl);
1300        return cond;
1301
1302    case TCG_COND_TSTEQ:
1303    case TCG_COND_TSTNE:
1304        /* Similar, but with TST instead of CMP. */
1305        tcg_out_dat_rI(s, COND_AL, ARITH_TST, 0, ah, bh, const_bh);
1306        tcg_out_dat_rI(s, COND_EQ, ARITH_TST, 0, al, bl, const_bl);
1307        return tcg_tst_eqne_cond(cond);
1308
1309    case TCG_COND_LT:
1310    case TCG_COND_GE:
1311        /* We perform a double-word subtraction and examine the result.
1312           We do not actually need the result of the subtract, so the
1313           low part "subtract" is a compare.  For the high half we have
1314           no choice but to compute into a temporary.  */
1315        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, al, bl, const_bl);
1316        tcg_out_dat_rI(s, COND_AL, ARITH_SBC | TO_CPSR,
1317                       TCG_REG_TMP, ah, bh, const_bh);
1318        return cond;
1319
1320    case TCG_COND_LE:
1321    case TCG_COND_GT:
1322        /* Similar, but with swapped arguments, via reversed subtract.  */
1323        tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR,
1324                       TCG_REG_TMP, al, bl, const_bl);
1325        tcg_out_dat_rI(s, COND_AL, ARITH_RSC | TO_CPSR,
1326                       TCG_REG_TMP, ah, bh, const_bh);
1327        return tcg_swap_cond(cond);
1328
1329    default:
1330        g_assert_not_reached();
1331    }
1332}
1333
1334/*
1335 * Note that TCGReg references Q-registers.
1336 * Q-regno = 2 * D-regno, so shift left by 1 while inserting.
1337 */
1338static uint32_t encode_vd(TCGReg rd)
1339{
1340    tcg_debug_assert(rd >= TCG_REG_Q0);
1341    return (extract32(rd, 3, 1) << 22) | (extract32(rd, 0, 3) << 13);
1342}
1343
1344static uint32_t encode_vn(TCGReg rn)
1345{
1346    tcg_debug_assert(rn >= TCG_REG_Q0);
1347    return (extract32(rn, 3, 1) << 7) | (extract32(rn, 0, 3) << 17);
1348}
1349
1350static uint32_t encode_vm(TCGReg rm)
1351{
1352    tcg_debug_assert(rm >= TCG_REG_Q0);
1353    return (extract32(rm, 3, 1) << 5) | (extract32(rm, 0, 3) << 1);
1354}
1355
1356static void tcg_out_vreg2(TCGContext *s, ARMInsn insn, int q, int vece,
1357                          TCGReg d, TCGReg m)
1358{
1359    tcg_out32(s, insn | (vece << 18) | (q << 6) |
1360              encode_vd(d) | encode_vm(m));
1361}
1362
1363static void tcg_out_vreg3(TCGContext *s, ARMInsn insn, int q, int vece,
1364                          TCGReg d, TCGReg n, TCGReg m)
1365{
1366    tcg_out32(s, insn | (vece << 20) | (q << 6) |
1367              encode_vd(d) | encode_vn(n) | encode_vm(m));
1368}
1369
1370static void tcg_out_vmovi(TCGContext *s, TCGReg rd,
1371                          int q, int op, int cmode, uint8_t imm8)
1372{
1373    tcg_out32(s, INSN_VMOVI | encode_vd(rd) | (q << 6) | (op << 5)
1374              | (cmode << 8) | extract32(imm8, 0, 4)
1375              | (extract32(imm8, 4, 3) << 16)
1376              | (extract32(imm8, 7, 1) << 24));
1377}
1378
1379static void tcg_out_vshifti(TCGContext *s, ARMInsn insn, int q,
1380                            TCGReg rd, TCGReg rm, int l_imm6)
1381{
1382    tcg_out32(s, insn | (q << 6) | encode_vd(rd) | encode_vm(rm) |
1383              (extract32(l_imm6, 6, 1) << 7) |
1384              (extract32(l_imm6, 0, 6) << 16));
1385}
1386
1387static void tcg_out_vldst(TCGContext *s, ARMInsn insn,
1388                          TCGReg rd, TCGReg rn, int offset)
1389{
1390    if (offset != 0) {
1391        if (check_fit_imm(offset) || check_fit_imm(-offset)) {
1392            tcg_out_dat_rIN(s, COND_AL, ARITH_ADD, ARITH_SUB,
1393                            TCG_REG_TMP, rn, offset, true);
1394        } else {
1395            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP, offset);
1396            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1397                            TCG_REG_TMP, TCG_REG_TMP, rn, 0);
1398        }
1399        rn = TCG_REG_TMP;
1400    }
1401    tcg_out32(s, insn | (rn << 16) | encode_vd(rd) | 0xf);
1402}
1403
1404typedef struct {
1405    ARMCond cond;
1406    TCGReg base;
1407    int index;
1408    bool index_scratch;
1409    TCGAtomAlign aa;
1410} HostAddress;
1411
1412bool tcg_target_has_memory_bswap(MemOp memop)
1413{
1414    return false;
1415}
1416
1417static TCGReg ldst_ra_gen(TCGContext *s, const TCGLabelQemuLdst *l, int arg)
1418{
1419    /* We arrive at the slow path via "BLNE", so R14 contains l->raddr. */
1420    return TCG_REG_R14;
1421}
1422
1423static const TCGLdstHelperParam ldst_helper_param = {
1424    .ra_gen = ldst_ra_gen,
1425    .ntmp = 1,
1426    .tmp = { TCG_REG_TMP },
1427};
1428
1429static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1430{
1431    MemOp opc = get_memop(lb->oi);
1432
1433    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1434        return false;
1435    }
1436
1437    tcg_out_ld_helper_args(s, lb, &ldst_helper_param);
1438    tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SIZE]);
1439    tcg_out_ld_helper_ret(s, lb, false, &ldst_helper_param);
1440
1441    tcg_out_goto(s, COND_AL, lb->raddr);
1442    return true;
1443}
1444
1445static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1446{
1447    MemOp opc = get_memop(lb->oi);
1448
1449    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1450        return false;
1451    }
1452
1453    tcg_out_st_helper_args(s, lb, &ldst_helper_param);
1454
1455    /* Tail-call to the helper, which will return to the fast path.  */
1456    tcg_out_goto(s, COND_AL, qemu_st_helpers[opc & MO_SIZE]);
1457    return true;
1458}
1459
1460/* We expect to use an 9-bit sign-magnitude negative offset from ENV.  */
1461#define MIN_TLB_MASK_TABLE_OFS  -256
1462
1463static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
1464                                           TCGReg addr, MemOpIdx oi, bool is_ld)
1465{
1466    TCGLabelQemuLdst *ldst = NULL;
1467    MemOp opc = get_memop(oi);
1468    unsigned a_mask;
1469
1470    if (tcg_use_softmmu) {
1471        *h = (HostAddress){
1472            .cond = COND_AL,
1473            .base = addr,
1474            .index = TCG_REG_R1,
1475            .index_scratch = true,
1476        };
1477    } else {
1478        *h = (HostAddress){
1479            .cond = COND_AL,
1480            .base = addr,
1481            .index = guest_base ? TCG_REG_GUEST_BASE : -1,
1482            .index_scratch = false,
1483        };
1484    }
1485
1486    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, false);
1487    a_mask = (1 << h->aa.align) - 1;
1488
1489    if (tcg_use_softmmu) {
1490        int mem_index = get_mmuidx(oi);
1491        int cmp_off = is_ld ? offsetof(CPUTLBEntry, addr_read)
1492                            : offsetof(CPUTLBEntry, addr_write);
1493        int fast_off = tlb_mask_table_ofs(s, mem_index);
1494        unsigned s_mask = (1 << (opc & MO_SIZE)) - 1;
1495        TCGReg t_addr;
1496
1497        ldst = new_ldst_label(s);
1498        ldst->is_ld = is_ld;
1499        ldst->oi = oi;
1500        ldst->addr_reg = addr;
1501
1502        /* Load cpu->neg.tlb.f[mmu_idx].{mask,table} into {r0,r1}.  */
1503        QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, mask) != 0);
1504        QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, table) != 4);
1505        tcg_out_ldrd_8(s, COND_AL, TCG_REG_R0, TCG_AREG0, fast_off);
1506
1507        /* Extract the tlb index from the address into R0.  */
1508        tcg_out_dat_reg(s, COND_AL, ARITH_AND, TCG_REG_R0, TCG_REG_R0, addr,
1509                        SHIFT_IMM_LSR(s->page_bits - CPU_TLB_ENTRY_BITS));
1510
1511        /*
1512         * Add the tlb_table pointer, creating the CPUTLBEntry address in R1.
1513         * Load the tlb comparator into R2 and the fast path addend into R1.
1514         */
1515        if (cmp_off == 0) {
1516            tcg_out_ld32_rwb(s, COND_AL, TCG_REG_R2, TCG_REG_R1, TCG_REG_R0);
1517        } else {
1518            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1519                            TCG_REG_R1, TCG_REG_R1, TCG_REG_R0, 0);
1520            tcg_out_ld32_12(s, COND_AL, TCG_REG_R2, TCG_REG_R1, cmp_off);
1521        }
1522
1523        /* Load the tlb addend.  */
1524        tcg_out_ld32_12(s, COND_AL, TCG_REG_R1, TCG_REG_R1,
1525                        offsetof(CPUTLBEntry, addend));
1526
1527        /*
1528         * Check alignment, check comparators.
1529         * Do this in 2-4 insns.  Use MOVW for v7, if possible,
1530         * to reduce the number of sequential conditional instructions.
1531         * Almost all guests have at least 4k pages, which means that we need
1532         * to clear at least 9 bits even for an 8-byte memory, which means it
1533         * isn't worth checking for an immediate operand for BIC.
1534         *
1535         * For unaligned accesses, test the page of the last unit of alignment.
1536         * This leaves the least significant alignment bits unchanged, and of
1537         * course must be zero.
1538         */
1539        t_addr = addr;
1540        if (a_mask < s_mask) {
1541            t_addr = TCG_REG_R0;
1542            tcg_out_dat_imm(s, COND_AL, ARITH_ADD, t_addr,
1543                            addr, s_mask - a_mask);
1544        }
1545        if (use_armv7_instructions && s->page_bits <= 16) {
1546            tcg_out_movi32(s, COND_AL, TCG_REG_TMP, ~(s->page_mask | a_mask));
1547            tcg_out_dat_reg(s, COND_AL, ARITH_BIC, TCG_REG_TMP,
1548                            t_addr, TCG_REG_TMP, 0);
1549            tcg_out_dat_reg(s, COND_AL, ARITH_CMP, 0,
1550                            TCG_REG_R2, TCG_REG_TMP, 0);
1551        } else {
1552            if (a_mask) {
1553                tcg_debug_assert(a_mask <= 0xff);
1554                tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addr, a_mask);
1555            }
1556            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0, t_addr,
1557                            SHIFT_IMM_LSR(s->page_bits));
1558            tcg_out_dat_reg(s, (a_mask ? COND_EQ : COND_AL), ARITH_CMP,
1559                            0, TCG_REG_R2, TCG_REG_TMP,
1560                            SHIFT_IMM_LSL(s->page_bits));
1561        }
1562    } else if (a_mask) {
1563        ldst = new_ldst_label(s);
1564        ldst->is_ld = is_ld;
1565        ldst->oi = oi;
1566        ldst->addr_reg = addr;
1567
1568        /* We are expecting alignment to max out at 7 */
1569        tcg_debug_assert(a_mask <= 0xff);
1570        /* tst addr, #mask */
1571        tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addr, a_mask);
1572    }
1573
1574    return ldst;
1575}
1576
1577static void tcg_out_qemu_ld_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1578                                   TCGReg datahi, HostAddress h)
1579{
1580    TCGReg base;
1581
1582    /* Byte swapping is left to middle-end expansion. */
1583    tcg_debug_assert((opc & MO_BSWAP) == 0);
1584
1585    switch (opc & MO_SSIZE) {
1586    case MO_UB:
1587        if (h.index < 0) {
1588            tcg_out_ld8_12(s, h.cond, datalo, h.base, 0);
1589        } else {
1590            tcg_out_ld8_r(s, h.cond, datalo, h.base, h.index);
1591        }
1592        break;
1593    case MO_SB:
1594        if (h.index < 0) {
1595            tcg_out_ld8s_8(s, h.cond, datalo, h.base, 0);
1596        } else {
1597            tcg_out_ld8s_r(s, h.cond, datalo, h.base, h.index);
1598        }
1599        break;
1600    case MO_UW:
1601        if (h.index < 0) {
1602            tcg_out_ld16u_8(s, h.cond, datalo, h.base, 0);
1603        } else {
1604            tcg_out_ld16u_r(s, h.cond, datalo, h.base, h.index);
1605        }
1606        break;
1607    case MO_SW:
1608        if (h.index < 0) {
1609            tcg_out_ld16s_8(s, h.cond, datalo, h.base, 0);
1610        } else {
1611            tcg_out_ld16s_r(s, h.cond, datalo, h.base, h.index);
1612        }
1613        break;
1614    case MO_UL:
1615        if (h.index < 0) {
1616            tcg_out_ld32_12(s, h.cond, datalo, h.base, 0);
1617        } else {
1618            tcg_out_ld32_r(s, h.cond, datalo, h.base, h.index);
1619        }
1620        break;
1621    case MO_UQ:
1622        /* We used pair allocation for datalo, so already should be aligned. */
1623        tcg_debug_assert((datalo & 1) == 0);
1624        tcg_debug_assert(datahi == datalo + 1);
1625        /* LDRD requires alignment; double-check that. */
1626        if (memop_alignment_bits(opc) >= MO_64) {
1627            if (h.index < 0) {
1628                tcg_out_ldrd_8(s, h.cond, datalo, h.base, 0);
1629                break;
1630            }
1631            /*
1632             * Rm (the second address op) must not overlap Rt or Rt + 1.
1633             * Since datalo is aligned, we can simplify the test via alignment.
1634             * Flip the two address arguments if that works.
1635             */
1636            if ((h.index & ~1) != datalo) {
1637                tcg_out_ldrd_r(s, h.cond, datalo, h.base, h.index);
1638                break;
1639            }
1640            if ((h.base & ~1) != datalo) {
1641                tcg_out_ldrd_r(s, h.cond, datalo, h.index, h.base);
1642                break;
1643            }
1644        }
1645        if (h.index < 0) {
1646            base = h.base;
1647            if (datalo == h.base) {
1648                tcg_out_mov_reg(s, h.cond, TCG_REG_TMP, base);
1649                base = TCG_REG_TMP;
1650            }
1651        } else if (h.index_scratch) {
1652            tcg_out_ld32_rwb(s, h.cond, datalo, h.index, h.base);
1653            tcg_out_ld32_12(s, h.cond, datahi, h.index, 4);
1654            break;
1655        } else {
1656            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1657                            h.base, h.index, SHIFT_IMM_LSL(0));
1658            base = TCG_REG_TMP;
1659        }
1660        tcg_out_ld32_12(s, h.cond, datalo, base, 0);
1661        tcg_out_ld32_12(s, h.cond, datahi, base, 4);
1662        break;
1663    default:
1664        g_assert_not_reached();
1665    }
1666}
1667
1668static void tcg_out_qemu_ld(TCGContext *s, TCGReg datalo, TCGReg datahi,
1669                            TCGReg addr, MemOpIdx oi, TCGType data_type)
1670{
1671    MemOp opc = get_memop(oi);
1672    TCGLabelQemuLdst *ldst;
1673    HostAddress h;
1674
1675    ldst = prepare_host_addr(s, &h, addr, oi, true);
1676    if (ldst) {
1677        ldst->type = data_type;
1678        ldst->datalo_reg = datalo;
1679        ldst->datahi_reg = datahi;
1680
1681        /*
1682         * This a conditional BL only to load a pointer within this
1683         * opcode into LR for the slow path.  We will not be using
1684         * the value for a tail call.
1685         */
1686        ldst->label_ptr[0] = s->code_ptr;
1687        tcg_out_bl_imm(s, COND_NE, 0);
1688
1689        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1690        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1691    } else {
1692        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1693    }
1694}
1695
1696static void tcg_out_qemu_st_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1697                                   TCGReg datahi, HostAddress h)
1698{
1699    /* Byte swapping is left to middle-end expansion. */
1700    tcg_debug_assert((opc & MO_BSWAP) == 0);
1701
1702    switch (opc & MO_SIZE) {
1703    case MO_8:
1704        if (h.index < 0) {
1705            tcg_out_st8_12(s, h.cond, datalo, h.base, 0);
1706        } else {
1707            tcg_out_st8_r(s, h.cond, datalo, h.base, h.index);
1708        }
1709        break;
1710    case MO_16:
1711        if (h.index < 0) {
1712            tcg_out_st16_8(s, h.cond, datalo, h.base, 0);
1713        } else {
1714            tcg_out_st16_r(s, h.cond, datalo, h.base, h.index);
1715        }
1716        break;
1717    case MO_32:
1718        if (h.index < 0) {
1719            tcg_out_st32_12(s, h.cond, datalo, h.base, 0);
1720        } else {
1721            tcg_out_st32_r(s, h.cond, datalo, h.base, h.index);
1722        }
1723        break;
1724    case MO_64:
1725        /* We used pair allocation for datalo, so already should be aligned. */
1726        tcg_debug_assert((datalo & 1) == 0);
1727        tcg_debug_assert(datahi == datalo + 1);
1728        /* STRD requires alignment; double-check that. */
1729        if (memop_alignment_bits(opc) >= MO_64) {
1730            if (h.index < 0) {
1731                tcg_out_strd_8(s, h.cond, datalo, h.base, 0);
1732            } else {
1733                tcg_out_strd_r(s, h.cond, datalo, h.base, h.index);
1734            }
1735        } else if (h.index < 0) {
1736            tcg_out_st32_12(s, h.cond, datalo, h.base, 0);
1737            tcg_out_st32_12(s, h.cond, datahi, h.base, 4);
1738        } else if (h.index_scratch) {
1739            tcg_out_st32_rwb(s, h.cond, datalo, h.index, h.base);
1740            tcg_out_st32_12(s, h.cond, datahi, h.index, 4);
1741        } else {
1742            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1743                            h.base, h.index, SHIFT_IMM_LSL(0));
1744            tcg_out_st32_12(s, h.cond, datalo, TCG_REG_TMP, 0);
1745            tcg_out_st32_12(s, h.cond, datahi, TCG_REG_TMP, 4);
1746        }
1747        break;
1748    default:
1749        g_assert_not_reached();
1750    }
1751}
1752
1753static void tcg_out_qemu_st(TCGContext *s, TCGReg datalo, TCGReg datahi,
1754                            TCGReg addr, MemOpIdx oi, TCGType data_type)
1755{
1756    MemOp opc = get_memop(oi);
1757    TCGLabelQemuLdst *ldst;
1758    HostAddress h;
1759
1760    ldst = prepare_host_addr(s, &h, addr, oi, false);
1761    if (ldst) {
1762        ldst->type = data_type;
1763        ldst->datalo_reg = datalo;
1764        ldst->datahi_reg = datahi;
1765
1766        h.cond = COND_EQ;
1767        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1768
1769        /* The conditional call is last, as we're going to return here. */
1770        ldst->label_ptr[0] = s->code_ptr;
1771        tcg_out_bl_imm(s, COND_NE, 0);
1772        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1773    } else {
1774        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1775    }
1776}
1777
1778static void tcg_out_epilogue(TCGContext *s);
1779
1780static void tcg_out_exit_tb(TCGContext *s, uintptr_t arg)
1781{
1782    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, arg);
1783    tcg_out_epilogue(s);
1784}
1785
1786static void tcg_out_goto_tb(TCGContext *s, int which)
1787{
1788    uintptr_t i_addr;
1789    intptr_t i_disp;
1790
1791    /* Direct branch will be patched by tb_target_set_jmp_target. */
1792    set_jmp_insn_offset(s, which);
1793    tcg_out32(s, INSN_NOP);
1794
1795    /* When branch is out of range, fall through to indirect. */
1796    i_addr = get_jmp_target_addr(s, which);
1797    i_disp = tcg_pcrel_diff(s, (void *)i_addr) - 8;
1798    tcg_debug_assert(i_disp < 0);
1799    if (i_disp >= -0xfff) {
1800        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_PC, i_disp);
1801    } else {
1802        /*
1803         * The TB is close, but outside the 12 bits addressable by
1804         * the load.  We can extend this to 20 bits with a sub of a
1805         * shifted immediate from pc.
1806         */
1807        int h = -i_disp;
1808        int l = -(h & 0xfff);
1809
1810        h = encode_imm_nofail(h + l);
1811        tcg_out_dat_imm(s, COND_AL, ARITH_SUB, TCG_REG_R0, TCG_REG_PC, h);
1812        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_R0, l);
1813    }
1814    set_jmp_reset_offset(s, which);
1815}
1816
1817void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
1818                              uintptr_t jmp_rx, uintptr_t jmp_rw)
1819{
1820    uintptr_t addr = tb->jmp_target_addr[n];
1821    ptrdiff_t offset = addr - (jmp_rx + 8);
1822    tcg_insn_unit insn;
1823
1824    /* Either directly branch, or fall through to indirect branch. */
1825    if (offset == sextract64(offset, 0, 26)) {
1826        /* B <addr> */
1827        insn = deposit32((COND_AL << 28) | INSN_B, 0, 24, offset >> 2);
1828    } else {
1829        insn = INSN_NOP;
1830    }
1831
1832    qatomic_set((uint32_t *)jmp_rw, insn);
1833    flush_idcache_range(jmp_rx, jmp_rw, 4);
1834}
1835
1836
1837static void tgen_add(TCGContext *s, TCGType type,
1838                     TCGReg a0, TCGReg a1, TCGReg a2)
1839{
1840    tcg_out_dat_reg(s, COND_AL, ARITH_ADD, a0, a1, a2, SHIFT_IMM_LSL(0));
1841}
1842
1843static void tgen_addi(TCGContext *s, TCGType type,
1844                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1845{
1846    tcg_out_dat_IN(s, COND_AL, ARITH_ADD, ARITH_SUB, a0, a1, a2);
1847}
1848
1849static const TCGOutOpBinary outop_add = {
1850    .base.static_constraint = C_O1_I2(r, r, rIN),
1851    .out_rrr = tgen_add,
1852    .out_rri = tgen_addi,
1853};
1854
1855static void tgen_and(TCGContext *s, TCGType type,
1856                     TCGReg a0, TCGReg a1, TCGReg a2)
1857{
1858    tcg_out_dat_reg(s, COND_AL, ARITH_AND, a0, a1, a2, SHIFT_IMM_LSL(0));
1859}
1860
1861static void tgen_andi(TCGContext *s, TCGType type,
1862                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1863{
1864    tcg_out_dat_IK(s, COND_AL, ARITH_AND, ARITH_BIC, a0, a1, a2);
1865}
1866
1867static const TCGOutOpBinary outop_and = {
1868    .base.static_constraint = C_O1_I2(r, r, rIK),
1869    .out_rrr = tgen_and,
1870    .out_rri = tgen_andi,
1871};
1872
1873static void tgen_andc(TCGContext *s, TCGType type,
1874                      TCGReg a0, TCGReg a1, TCGReg a2)
1875{
1876    tcg_out_dat_reg(s, COND_AL, ARITH_BIC, a0, a1, a2, SHIFT_IMM_LSL(0));
1877}
1878
1879static const TCGOutOpBinary outop_andc = {
1880    .base.static_constraint = C_O1_I2(r, r, r),
1881    .out_rrr = tgen_andc,
1882};
1883
1884
1885static void tcg_out_op(TCGContext *s, TCGOpcode opc, TCGType type,
1886                       const TCGArg args[TCG_MAX_OP_ARGS],
1887                       const int const_args[TCG_MAX_OP_ARGS])
1888{
1889    TCGArg a0, a1, a2, a3, a4, a5;
1890    int c;
1891
1892    switch (opc) {
1893    case INDEX_op_goto_ptr:
1894        tcg_out_b_reg(s, COND_AL, args[0]);
1895        break;
1896    case INDEX_op_br:
1897        tcg_out_goto_label(s, COND_AL, arg_label(args[0]));
1898        break;
1899
1900    case INDEX_op_ld8u_i32:
1901        tcg_out_ld8u(s, COND_AL, args[0], args[1], args[2]);
1902        break;
1903    case INDEX_op_ld8s_i32:
1904        tcg_out_ld8s(s, COND_AL, args[0], args[1], args[2]);
1905        break;
1906    case INDEX_op_ld16u_i32:
1907        tcg_out_ld16u(s, COND_AL, args[0], args[1], args[2]);
1908        break;
1909    case INDEX_op_ld16s_i32:
1910        tcg_out_ld16s(s, COND_AL, args[0], args[1], args[2]);
1911        break;
1912    case INDEX_op_ld_i32:
1913        tcg_out_ld32u(s, COND_AL, args[0], args[1], args[2]);
1914        break;
1915    case INDEX_op_st8_i32:
1916        tcg_out_st8(s, COND_AL, args[0], args[1], args[2]);
1917        break;
1918    case INDEX_op_st16_i32:
1919        tcg_out_st16(s, COND_AL, args[0], args[1], args[2]);
1920        break;
1921    case INDEX_op_st_i32:
1922        tcg_out_st32(s, COND_AL, args[0], args[1], args[2]);
1923        break;
1924
1925    case INDEX_op_movcond_i32:
1926        /* Constraints mean that v2 is always in the same register as dest,
1927         * so we only need to do "if condition passed, move v1 to dest".
1928         */
1929        c = tcg_out_cmp(s, args[5], args[1], args[2], const_args[2]);
1930        tcg_out_dat_rIK(s, tcg_cond_to_arm_cond[c], ARITH_MOV,
1931                        ARITH_MVN, args[0], 0, args[3], const_args[3]);
1932        break;
1933    case INDEX_op_sub_i32:
1934        if (const_args[1]) {
1935            if (const_args[2]) {
1936                tcg_out_movi32(s, COND_AL, args[0], args[1] - args[2]);
1937            } else {
1938                tcg_out_dat_rI(s, COND_AL, ARITH_RSB,
1939                               args[0], args[2], args[1], 1);
1940            }
1941        } else {
1942            tcg_out_dat_rIN(s, COND_AL, ARITH_SUB, ARITH_ADD,
1943                            args[0], args[1], args[2], const_args[2]);
1944        }
1945        break;
1946    case INDEX_op_or_i32:
1947        c = ARITH_ORR;
1948        goto gen_arith;
1949    case INDEX_op_xor_i32:
1950        c = ARITH_EOR;
1951        /* Fall through.  */
1952    gen_arith:
1953        tcg_out_dat_rI(s, COND_AL, c, args[0], args[1], args[2], const_args[2]);
1954        break;
1955    case INDEX_op_add2_i32:
1956        a0 = args[0], a1 = args[1], a2 = args[2];
1957        a3 = args[3], a4 = args[4], a5 = args[5];
1958        if (a0 == a3 || (a0 == a5 && !const_args[5])) {
1959            a0 = TCG_REG_TMP;
1960        }
1961        tcg_out_dat_rIN(s, COND_AL, ARITH_ADD | TO_CPSR, ARITH_SUB | TO_CPSR,
1962                        a0, a2, a4, const_args[4]);
1963        tcg_out_dat_rIK(s, COND_AL, ARITH_ADC, ARITH_SBC,
1964                        a1, a3, a5, const_args[5]);
1965        tcg_out_mov_reg(s, COND_AL, args[0], a0);
1966        break;
1967    case INDEX_op_sub2_i32:
1968        a0 = args[0], a1 = args[1], a2 = args[2];
1969        a3 = args[3], a4 = args[4], a5 = args[5];
1970        if ((a0 == a3 && !const_args[3]) || (a0 == a5 && !const_args[5])) {
1971            a0 = TCG_REG_TMP;
1972        }
1973        if (const_args[2]) {
1974            if (const_args[4]) {
1975                tcg_out_movi32(s, COND_AL, a0, a4);
1976                a4 = a0;
1977            }
1978            tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR, a0, a4, a2, 1);
1979        } else {
1980            tcg_out_dat_rIN(s, COND_AL, ARITH_SUB | TO_CPSR,
1981                            ARITH_ADD | TO_CPSR, a0, a2, a4, const_args[4]);
1982        }
1983        if (const_args[3]) {
1984            if (const_args[5]) {
1985                tcg_out_movi32(s, COND_AL, a1, a5);
1986                a5 = a1;
1987            }
1988            tcg_out_dat_rI(s, COND_AL, ARITH_RSC, a1, a5, a3, 1);
1989        } else {
1990            tcg_out_dat_rIK(s, COND_AL, ARITH_SBC, ARITH_ADC,
1991                            a1, a3, a5, const_args[5]);
1992        }
1993        tcg_out_mov_reg(s, COND_AL, args[0], a0);
1994        break;
1995    case INDEX_op_neg_i32:
1996        tcg_out_dat_imm(s, COND_AL, ARITH_RSB, args[0], args[1], 0);
1997        break;
1998    case INDEX_op_not_i32:
1999        tcg_out_dat_reg(s, COND_AL,
2000                        ARITH_MVN, args[0], 0, args[1], SHIFT_IMM_LSL(0));
2001        break;
2002    case INDEX_op_mul_i32:
2003        tcg_out_mul32(s, COND_AL, args[0], args[1], args[2]);
2004        break;
2005    case INDEX_op_mulu2_i32:
2006        tcg_out_umull32(s, COND_AL, args[0], args[1], args[2], args[3]);
2007        break;
2008    case INDEX_op_muls2_i32:
2009        tcg_out_smull32(s, COND_AL, args[0], args[1], args[2], args[3]);
2010        break;
2011    /* XXX: Perhaps args[2] & 0x1f is wrong */
2012    case INDEX_op_shl_i32:
2013        c = const_args[2] ?
2014                SHIFT_IMM_LSL(args[2] & 0x1f) : SHIFT_REG_LSL(args[2]);
2015        goto gen_shift32;
2016    case INDEX_op_shr_i32:
2017        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_LSR(args[2] & 0x1f) :
2018                SHIFT_IMM_LSL(0) : SHIFT_REG_LSR(args[2]);
2019        goto gen_shift32;
2020    case INDEX_op_sar_i32:
2021        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_ASR(args[2] & 0x1f) :
2022                SHIFT_IMM_LSL(0) : SHIFT_REG_ASR(args[2]);
2023        goto gen_shift32;
2024    case INDEX_op_rotr_i32:
2025        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_ROR(args[2] & 0x1f) :
2026                SHIFT_IMM_LSL(0) : SHIFT_REG_ROR(args[2]);
2027        /* Fall through.  */
2028    gen_shift32:
2029        tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1], c);
2030        break;
2031
2032    case INDEX_op_rotl_i32:
2033        if (const_args[2]) {
2034            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1],
2035                            ((0x20 - args[2]) & 0x1f) ?
2036                            SHIFT_IMM_ROR((0x20 - args[2]) & 0x1f) :
2037                            SHIFT_IMM_LSL(0));
2038        } else {
2039            tcg_out_dat_imm(s, COND_AL, ARITH_RSB, TCG_REG_TMP, args[2], 0x20);
2040            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1],
2041                            SHIFT_REG_ROR(TCG_REG_TMP));
2042        }
2043        break;
2044
2045    case INDEX_op_ctz_i32:
2046        tcg_out_dat_reg(s, COND_AL, INSN_RBIT, TCG_REG_TMP, 0, args[1], 0);
2047        a1 = TCG_REG_TMP;
2048        goto do_clz;
2049
2050    case INDEX_op_clz_i32:
2051        a1 = args[1];
2052    do_clz:
2053        a0 = args[0];
2054        a2 = args[2];
2055        c = const_args[2];
2056        if (c && a2 == 32) {
2057            tcg_out_dat_reg(s, COND_AL, INSN_CLZ, a0, 0, a1, 0);
2058            break;
2059        }
2060        tcg_out_dat_imm(s, COND_AL, ARITH_CMP, 0, a1, 0);
2061        tcg_out_dat_reg(s, COND_NE, INSN_CLZ, a0, 0, a1, 0);
2062        if (c || a0 != a2) {
2063            tcg_out_dat_rIK(s, COND_EQ, ARITH_MOV, ARITH_MVN, a0, 0, a2, c);
2064        }
2065        break;
2066
2067    case INDEX_op_brcond_i32:
2068        c = tcg_out_cmp(s, args[2], args[0], args[1], const_args[1]);
2069        tcg_out_goto_label(s, tcg_cond_to_arm_cond[c], arg_label(args[3]));
2070        break;
2071    case INDEX_op_setcond_i32:
2072        c = tcg_out_cmp(s, args[3], args[1], args[2], const_args[2]);
2073        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c],
2074                        ARITH_MOV, args[0], 0, 1);
2075        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
2076                        ARITH_MOV, args[0], 0, 0);
2077        break;
2078    case INDEX_op_negsetcond_i32:
2079        c = tcg_out_cmp(s, args[3], args[1], args[2], const_args[2]);
2080        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c],
2081                        ARITH_MVN, args[0], 0, 0);
2082        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
2083                        ARITH_MOV, args[0], 0, 0);
2084        break;
2085
2086    case INDEX_op_brcond2_i32:
2087        c = tcg_out_cmp2(s, args, const_args);
2088        tcg_out_goto_label(s, tcg_cond_to_arm_cond[c], arg_label(args[5]));
2089        break;
2090    case INDEX_op_setcond2_i32:
2091        c = tcg_out_cmp2(s, args + 1, const_args + 1);
2092        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c], ARITH_MOV, args[0], 0, 1);
2093        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
2094                        ARITH_MOV, args[0], 0, 0);
2095        break;
2096
2097    case INDEX_op_qemu_ld_i32:
2098        tcg_out_qemu_ld(s, args[0], -1, args[1], args[2], TCG_TYPE_I32);
2099        break;
2100    case INDEX_op_qemu_ld_i64:
2101        tcg_out_qemu_ld(s, args[0], args[1], args[2], args[3], TCG_TYPE_I64);
2102        break;
2103
2104    case INDEX_op_qemu_st_i32:
2105        tcg_out_qemu_st(s, args[0], -1, args[1], args[2], TCG_TYPE_I32);
2106        break;
2107    case INDEX_op_qemu_st_i64:
2108        tcg_out_qemu_st(s, args[0], args[1], args[2], args[3], TCG_TYPE_I64);
2109        break;
2110
2111    case INDEX_op_bswap16_i32:
2112        tcg_out_bswap16(s, COND_AL, args[0], args[1], args[2]);
2113        break;
2114    case INDEX_op_bswap32_i32:
2115        tcg_out_bswap32(s, COND_AL, args[0], args[1]);
2116        break;
2117
2118    case INDEX_op_deposit_i32:
2119        tcg_out_deposit(s, COND_AL, args[0], args[2],
2120                        args[3], args[4], const_args[2]);
2121        break;
2122    case INDEX_op_extract_i32:
2123        tcg_out_extract(s, COND_AL, args[0], args[1], args[2], args[3]);
2124        break;
2125    case INDEX_op_sextract_i32:
2126        tcg_out_sextract(s, COND_AL, args[0], args[1], args[2], args[3]);
2127        break;
2128    case INDEX_op_extract2_i32:
2129        /* ??? These optimization vs zero should be generic.  */
2130        /* ??? But we can't substitute 2 for 1 in the opcode stream yet.  */
2131        if (const_args[1]) {
2132            if (const_args[2]) {
2133                tcg_out_movi(s, TCG_TYPE_REG, args[0], 0);
2134            } else {
2135                tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2136                                args[2], SHIFT_IMM_LSL(32 - args[3]));
2137            }
2138        } else if (const_args[2]) {
2139            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2140                            args[1], SHIFT_IMM_LSR(args[3]));
2141        } else {
2142            /* We can do extract2 in 2 insns, vs the 3 required otherwise.  */
2143            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0,
2144                            args[2], SHIFT_IMM_LSL(32 - args[3]));
2145            tcg_out_dat_reg(s, COND_AL, ARITH_ORR, args[0], TCG_REG_TMP,
2146                            args[1], SHIFT_IMM_LSR(args[3]));
2147        }
2148        break;
2149
2150    case INDEX_op_div_i32:
2151        tcg_out_sdiv(s, COND_AL, args[0], args[1], args[2]);
2152        break;
2153    case INDEX_op_divu_i32:
2154        tcg_out_udiv(s, COND_AL, args[0], args[1], args[2]);
2155        break;
2156
2157    case INDEX_op_mb:
2158        tcg_out_mb(s, args[0]);
2159        break;
2160
2161    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
2162    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
2163    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
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_extract_i32:
2186    case INDEX_op_sextract_i32:
2187        return C_O1_I1(r, r);
2188
2189    case INDEX_op_st8_i32:
2190    case INDEX_op_st16_i32:
2191    case INDEX_op_st_i32:
2192        return C_O0_I2(r, r);
2193
2194    case INDEX_op_sub_i32:
2195    case INDEX_op_setcond_i32:
2196    case INDEX_op_negsetcond_i32:
2197        return C_O1_I2(r, r, rIN);
2198
2199    case INDEX_op_clz_i32:
2200    case INDEX_op_ctz_i32:
2201        return C_O1_I2(r, r, rIK);
2202
2203    case INDEX_op_mul_i32:
2204    case INDEX_op_div_i32:
2205    case INDEX_op_divu_i32:
2206        return C_O1_I2(r, r, r);
2207
2208    case INDEX_op_mulu2_i32:
2209    case INDEX_op_muls2_i32:
2210        return C_O2_I2(r, r, r, r);
2211
2212    case INDEX_op_or_i32:
2213    case INDEX_op_xor_i32:
2214        return C_O1_I2(r, r, rI);
2215
2216    case INDEX_op_shl_i32:
2217    case INDEX_op_shr_i32:
2218    case INDEX_op_sar_i32:
2219    case INDEX_op_rotl_i32:
2220    case INDEX_op_rotr_i32:
2221        return C_O1_I2(r, r, ri);
2222
2223    case INDEX_op_brcond_i32:
2224        return C_O0_I2(r, rIN);
2225    case INDEX_op_deposit_i32:
2226        return C_O1_I2(r, 0, rZ);
2227    case INDEX_op_extract2_i32:
2228        return C_O1_I2(r, rZ, rZ);
2229    case INDEX_op_movcond_i32:
2230        return C_O1_I4(r, r, rIN, rIK, 0);
2231    case INDEX_op_add2_i32:
2232        return C_O2_I4(r, r, r, r, rIN, rIK);
2233    case INDEX_op_sub2_i32:
2234        return C_O2_I4(r, r, rI, rI, rIN, rIK);
2235    case INDEX_op_brcond2_i32:
2236        return C_O0_I4(r, r, rI, rI);
2237    case INDEX_op_setcond2_i32:
2238        return C_O1_I4(r, r, r, rI, rI);
2239
2240    case INDEX_op_qemu_ld_i32:
2241        return C_O1_I1(r, q);
2242    case INDEX_op_qemu_ld_i64:
2243        return C_O2_I1(e, p, q);
2244    case INDEX_op_qemu_st_i32:
2245        return C_O0_I2(q, q);
2246    case INDEX_op_qemu_st_i64:
2247        return C_O0_I3(Q, p, q);
2248
2249    case INDEX_op_st_vec:
2250        return C_O0_I2(w, r);
2251    case INDEX_op_ld_vec:
2252    case INDEX_op_dupm_vec:
2253        return C_O1_I1(w, r);
2254    case INDEX_op_dup_vec:
2255        return C_O1_I1(w, wr);
2256    case INDEX_op_abs_vec:
2257    case INDEX_op_neg_vec:
2258    case INDEX_op_not_vec:
2259    case INDEX_op_shli_vec:
2260    case INDEX_op_shri_vec:
2261    case INDEX_op_sari_vec:
2262        return C_O1_I1(w, w);
2263    case INDEX_op_dup2_vec:
2264    case INDEX_op_add_vec:
2265    case INDEX_op_mul_vec:
2266    case INDEX_op_smax_vec:
2267    case INDEX_op_smin_vec:
2268    case INDEX_op_ssadd_vec:
2269    case INDEX_op_sssub_vec:
2270    case INDEX_op_sub_vec:
2271    case INDEX_op_umax_vec:
2272    case INDEX_op_umin_vec:
2273    case INDEX_op_usadd_vec:
2274    case INDEX_op_ussub_vec:
2275    case INDEX_op_xor_vec:
2276    case INDEX_op_arm_sshl_vec:
2277    case INDEX_op_arm_ushl_vec:
2278        return C_O1_I2(w, w, w);
2279    case INDEX_op_arm_sli_vec:
2280        return C_O1_I2(w, 0, w);
2281    case INDEX_op_or_vec:
2282    case INDEX_op_andc_vec:
2283        return C_O1_I2(w, w, wO);
2284    case INDEX_op_and_vec:
2285    case INDEX_op_orc_vec:
2286        return C_O1_I2(w, w, wV);
2287    case INDEX_op_cmp_vec:
2288        return C_O1_I2(w, w, wZ);
2289    case INDEX_op_bitsel_vec:
2290        return C_O1_I3(w, w, w, w);
2291    default:
2292        return C_NotImplemented;
2293    }
2294}
2295
2296static void tcg_target_init(TCGContext *s)
2297{
2298    /*
2299     * Only probe for the platform and capabilities if we haven't already
2300     * determined maximum values at compile time.
2301     */
2302#if !defined(use_idiv_instructions) || !defined(use_neon_instructions)
2303    {
2304        unsigned long hwcap = qemu_getauxval(AT_HWCAP);
2305#ifndef use_idiv_instructions
2306        use_idiv_instructions = (hwcap & HWCAP_ARM_IDIVA) != 0;
2307#endif
2308#ifndef use_neon_instructions
2309        use_neon_instructions = (hwcap & HWCAP_ARM_NEON) != 0;
2310#endif
2311    }
2312#endif
2313
2314    if (__ARM_ARCH < 7) {
2315        const char *pl = (const char *)qemu_getauxval(AT_PLATFORM);
2316        if (pl != NULL && pl[0] == 'v' && pl[1] >= '4' && pl[1] <= '9') {
2317            arm_arch = pl[1] - '0';
2318        }
2319
2320        if (arm_arch < 6) {
2321            error_report("TCG: ARMv%d is unsupported; exiting", arm_arch);
2322            exit(EXIT_FAILURE);
2323        }
2324    }
2325
2326    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
2327
2328    tcg_target_call_clobber_regs = 0;
2329    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R0);
2330    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R1);
2331    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R2);
2332    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R3);
2333    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R12);
2334    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R14);
2335
2336    if (use_neon_instructions) {
2337        tcg_target_available_regs[TCG_TYPE_V64]  = ALL_VECTOR_REGS;
2338        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
2339
2340        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q0);
2341        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q1);
2342        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q2);
2343        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q3);
2344        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q8);
2345        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q9);
2346        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q10);
2347        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q11);
2348        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q12);
2349        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q13);
2350        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q14);
2351        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q15);
2352    }
2353
2354    s->reserved_regs = 0;
2355    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
2356    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP);
2357    tcg_regset_set_reg(s->reserved_regs, TCG_REG_PC);
2358    tcg_regset_set_reg(s->reserved_regs, TCG_VEC_TMP);
2359}
2360
2361static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg arg,
2362                       TCGReg arg1, intptr_t arg2)
2363{
2364    switch (type) {
2365    case TCG_TYPE_I32:
2366        tcg_out_ld32u(s, COND_AL, arg, arg1, arg2);
2367        return;
2368    case TCG_TYPE_V64:
2369        /* regs 1; size 8; align 8 */
2370        tcg_out_vldst(s, INSN_VLD1 | 0x7d0, arg, arg1, arg2);
2371        return;
2372    case TCG_TYPE_V128:
2373        /*
2374         * We have only 8-byte alignment for the stack per the ABI.
2375         * Rather than dynamically re-align the stack, it's easier
2376         * to simply not request alignment beyond that.  So:
2377         * regs 2; size 8; align 8
2378         */
2379        tcg_out_vldst(s, INSN_VLD1 | 0xad0, arg, arg1, arg2);
2380        return;
2381    default:
2382        g_assert_not_reached();
2383    }
2384}
2385
2386static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
2387                       TCGReg arg1, intptr_t arg2)
2388{
2389    switch (type) {
2390    case TCG_TYPE_I32:
2391        tcg_out_st32(s, COND_AL, arg, arg1, arg2);
2392        return;
2393    case TCG_TYPE_V64:
2394        /* regs 1; size 8; align 8 */
2395        tcg_out_vldst(s, INSN_VST1 | 0x7d0, arg, arg1, arg2);
2396        return;
2397    case TCG_TYPE_V128:
2398        /* See tcg_out_ld re alignment: regs 2; size 8; align 8 */
2399        tcg_out_vldst(s, INSN_VST1 | 0xad0, arg, arg1, arg2);
2400        return;
2401    default:
2402        g_assert_not_reached();
2403    }
2404}
2405
2406static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
2407                        TCGReg base, intptr_t ofs)
2408{
2409    return false;
2410}
2411
2412static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
2413{
2414    if (ret == arg) {
2415        return true;
2416    }
2417    switch (type) {
2418    case TCG_TYPE_I32:
2419        if (ret < TCG_REG_Q0 && arg < TCG_REG_Q0) {
2420            tcg_out_mov_reg(s, COND_AL, ret, arg);
2421            return true;
2422        }
2423        return false;
2424
2425    case TCG_TYPE_V64:
2426    case TCG_TYPE_V128:
2427        /* "VMOV D,N" is an alias for "VORR D,N,N". */
2428        tcg_out_vreg3(s, INSN_VORR, type - TCG_TYPE_V64, 0, ret, arg, arg);
2429        return true;
2430
2431    default:
2432        g_assert_not_reached();
2433    }
2434}
2435
2436static void tcg_out_movi(TCGContext *s, TCGType type,
2437                         TCGReg ret, tcg_target_long arg)
2438{
2439    tcg_debug_assert(type == TCG_TYPE_I32);
2440    tcg_debug_assert(ret < TCG_REG_Q0);
2441    tcg_out_movi32(s, COND_AL, ret, arg);
2442}
2443
2444static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
2445{
2446    return false;
2447}
2448
2449static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
2450                             tcg_target_long imm)
2451{
2452    int enc, opc = ARITH_ADD;
2453
2454    /* All of the easiest immediates to encode are positive. */
2455    if (imm < 0) {
2456        imm = -imm;
2457        opc = ARITH_SUB;
2458    }
2459    enc = encode_imm(imm);
2460    if (enc >= 0) {
2461        tcg_out_dat_imm(s, COND_AL, opc, rd, rs, enc);
2462    } else {
2463        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, imm);
2464        tcg_out_dat_reg(s, COND_AL, opc, rd, rs,
2465                        TCG_REG_TMP, SHIFT_IMM_LSL(0));
2466    }
2467}
2468
2469/* Type is always V128, with I64 elements.  */
2470static void tcg_out_dup2_vec(TCGContext *s, TCGReg rd, TCGReg rl, TCGReg rh)
2471{
2472    /* Move high element into place first. */
2473    /* VMOV Dd+1, Ds */
2474    tcg_out_vreg3(s, INSN_VORR | (1 << 12), 0, 0, rd, rh, rh);
2475    /* Move low element into place; tcg_out_mov will check for nop. */
2476    tcg_out_mov(s, TCG_TYPE_V64, rd, rl);
2477}
2478
2479static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
2480                            TCGReg rd, TCGReg rs)
2481{
2482    int q = type - TCG_TYPE_V64;
2483
2484    if (vece == MO_64) {
2485        if (type == TCG_TYPE_V128) {
2486            tcg_out_dup2_vec(s, rd, rs, rs);
2487        } else {
2488            tcg_out_mov(s, TCG_TYPE_V64, rd, rs);
2489        }
2490    } else if (rs < TCG_REG_Q0) {
2491        int b = (vece == MO_8);
2492        int e = (vece == MO_16);
2493        tcg_out32(s, INSN_VDUP_G | (b << 22) | (q << 21) | (e << 5) |
2494                  encode_vn(rd) | (rs << 12));
2495    } else {
2496        int imm4 = 1 << vece;
2497        tcg_out32(s, INSN_VDUP_S | (imm4 << 16) | (q << 6) |
2498                  encode_vd(rd) | encode_vm(rs));
2499    }
2500    return true;
2501}
2502
2503static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
2504                             TCGReg rd, TCGReg base, intptr_t offset)
2505{
2506    if (vece == MO_64) {
2507        tcg_out_ld(s, TCG_TYPE_V64, rd, base, offset);
2508        if (type == TCG_TYPE_V128) {
2509            tcg_out_dup2_vec(s, rd, rd, rd);
2510        }
2511    } else {
2512        int q = type - TCG_TYPE_V64;
2513        tcg_out_vldst(s, INSN_VLD1R | (vece << 6) | (q << 5),
2514                      rd, base, offset);
2515    }
2516    return true;
2517}
2518
2519static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
2520                             TCGReg rd, int64_t v64)
2521{
2522    int q = type - TCG_TYPE_V64;
2523    int cmode, imm8, i;
2524
2525    /* Test all bytes equal first.  */
2526    if (vece == MO_8) {
2527        tcg_out_vmovi(s, rd, q, 0, 0xe, v64);
2528        return;
2529    }
2530
2531    /*
2532     * Test all bytes 0x00 or 0xff second.  This can match cases that
2533     * might otherwise take 2 or 3 insns for MO_16 or MO_32 below.
2534     */
2535    for (i = imm8 = 0; i < 8; i++) {
2536        uint8_t byte = v64 >> (i * 8);
2537        if (byte == 0xff) {
2538            imm8 |= 1 << i;
2539        } else if (byte != 0) {
2540            goto fail_bytes;
2541        }
2542    }
2543    tcg_out_vmovi(s, rd, q, 1, 0xe, imm8);
2544    return;
2545 fail_bytes:
2546
2547    /*
2548     * Tests for various replications.  For each element width, if we
2549     * cannot find an expansion there's no point checking a larger
2550     * width because we already know by replication it cannot match.
2551     */
2552    if (vece == MO_16) {
2553        uint16_t v16 = v64;
2554
2555        if (is_shimm16(v16, &cmode, &imm8)) {
2556            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2557            return;
2558        }
2559        if (is_shimm16(~v16, &cmode, &imm8)) {
2560            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2561            return;
2562        }
2563
2564        /*
2565         * Otherwise, all remaining constants can be loaded in two insns:
2566         * rd = v16 & 0xff, rd |= v16 & 0xff00.
2567         */
2568        tcg_out_vmovi(s, rd, q, 0, 0x8, v16 & 0xff);
2569        tcg_out_vmovi(s, rd, q, 0, 0xb, v16 >> 8);   /* VORRI */
2570        return;
2571    }
2572
2573    if (vece == MO_32) {
2574        uint32_t v32 = v64;
2575
2576        if (is_shimm32(v32, &cmode, &imm8) ||
2577            is_soimm32(v32, &cmode, &imm8)) {
2578            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2579            return;
2580        }
2581        if (is_shimm32(~v32, &cmode, &imm8) ||
2582            is_soimm32(~v32, &cmode, &imm8)) {
2583            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2584            return;
2585        }
2586
2587        /*
2588         * Restrict the set of constants to those we can load with
2589         * two instructions.  Others we load from the pool.
2590         */
2591        i = is_shimm32_pair(v32, &cmode, &imm8);
2592        if (i) {
2593            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2594            tcg_out_vmovi(s, rd, q, 0, i | 1, extract32(v32, i * 4, 8));
2595            return;
2596        }
2597        i = is_shimm32_pair(~v32, &cmode, &imm8);
2598        if (i) {
2599            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2600            tcg_out_vmovi(s, rd, q, 1, i | 1, extract32(~v32, i * 4, 8));
2601            return;
2602        }
2603    }
2604
2605    /*
2606     * As a last resort, load from the constant pool.
2607     */
2608    if (!q || vece == MO_64) {
2609        new_pool_l2(s, R_ARM_PC11, s->code_ptr, 0, v64, v64 >> 32);
2610        /* VLDR Dd, [pc + offset] */
2611        tcg_out32(s, INSN_VLDR_D | encode_vd(rd) | (0xf << 16));
2612        if (q) {
2613            tcg_out_dup2_vec(s, rd, rd, rd);
2614        }
2615    } else {
2616        new_pool_label(s, (uint32_t)v64, R_ARM_PC8, s->code_ptr, 0);
2617        /* add tmp, pc, offset */
2618        tcg_out_dat_imm(s, COND_AL, ARITH_ADD, TCG_REG_TMP, TCG_REG_PC, 0);
2619        tcg_out_dupm_vec(s, type, MO_32, rd, TCG_REG_TMP, 0);
2620    }
2621}
2622
2623static const ARMInsn vec_cmp_insn[16] = {
2624    [TCG_COND_EQ] = INSN_VCEQ,
2625    [TCG_COND_GT] = INSN_VCGT,
2626    [TCG_COND_GE] = INSN_VCGE,
2627    [TCG_COND_GTU] = INSN_VCGT_U,
2628    [TCG_COND_GEU] = INSN_VCGE_U,
2629};
2630
2631static const ARMInsn vec_cmp0_insn[16] = {
2632    [TCG_COND_EQ] = INSN_VCEQ0,
2633    [TCG_COND_GT] = INSN_VCGT0,
2634    [TCG_COND_GE] = INSN_VCGE0,
2635    [TCG_COND_LT] = INSN_VCLT0,
2636    [TCG_COND_LE] = INSN_VCLE0,
2637};
2638
2639static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
2640                           unsigned vecl, unsigned vece,
2641                           const TCGArg args[TCG_MAX_OP_ARGS],
2642                           const int const_args[TCG_MAX_OP_ARGS])
2643{
2644    TCGType type = vecl + TCG_TYPE_V64;
2645    unsigned q = vecl;
2646    TCGArg a0, a1, a2, a3;
2647    int cmode, imm8;
2648
2649    a0 = args[0];
2650    a1 = args[1];
2651    a2 = args[2];
2652
2653    switch (opc) {
2654    case INDEX_op_ld_vec:
2655        tcg_out_ld(s, type, a0, a1, a2);
2656        return;
2657    case INDEX_op_st_vec:
2658        tcg_out_st(s, type, a0, a1, a2);
2659        return;
2660    case INDEX_op_dupm_vec:
2661        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
2662        return;
2663    case INDEX_op_dup2_vec:
2664        tcg_out_dup2_vec(s, a0, a1, a2);
2665        return;
2666    case INDEX_op_abs_vec:
2667        tcg_out_vreg2(s, INSN_VABS, q, vece, a0, a1);
2668        return;
2669    case INDEX_op_neg_vec:
2670        tcg_out_vreg2(s, INSN_VNEG, q, vece, a0, a1);
2671        return;
2672    case INDEX_op_not_vec:
2673        tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a1);
2674        return;
2675    case INDEX_op_add_vec:
2676        tcg_out_vreg3(s, INSN_VADD, q, vece, a0, a1, a2);
2677        return;
2678    case INDEX_op_mul_vec:
2679        tcg_out_vreg3(s, INSN_VMUL, q, vece, a0, a1, a2);
2680        return;
2681    case INDEX_op_smax_vec:
2682        tcg_out_vreg3(s, INSN_VMAX, q, vece, a0, a1, a2);
2683        return;
2684    case INDEX_op_smin_vec:
2685        tcg_out_vreg3(s, INSN_VMIN, q, vece, a0, a1, a2);
2686        return;
2687    case INDEX_op_sub_vec:
2688        tcg_out_vreg3(s, INSN_VSUB, q, vece, a0, a1, a2);
2689        return;
2690    case INDEX_op_ssadd_vec:
2691        tcg_out_vreg3(s, INSN_VQADD, q, vece, a0, a1, a2);
2692        return;
2693    case INDEX_op_sssub_vec:
2694        tcg_out_vreg3(s, INSN_VQSUB, q, vece, a0, a1, a2);
2695        return;
2696    case INDEX_op_umax_vec:
2697        tcg_out_vreg3(s, INSN_VMAX_U, q, vece, a0, a1, a2);
2698        return;
2699    case INDEX_op_umin_vec:
2700        tcg_out_vreg3(s, INSN_VMIN_U, q, vece, a0, a1, a2);
2701        return;
2702    case INDEX_op_usadd_vec:
2703        tcg_out_vreg3(s, INSN_VQADD_U, q, vece, a0, a1, a2);
2704        return;
2705    case INDEX_op_ussub_vec:
2706        tcg_out_vreg3(s, INSN_VQSUB_U, q, vece, a0, a1, a2);
2707        return;
2708    case INDEX_op_xor_vec:
2709        tcg_out_vreg3(s, INSN_VEOR, q, 0, a0, a1, a2);
2710        return;
2711    case INDEX_op_arm_sshl_vec:
2712        /*
2713         * Note that Vm is the data and Vn is the shift count,
2714         * therefore the arguments appear reversed.
2715         */
2716        tcg_out_vreg3(s, INSN_VSHL_S, q, vece, a0, a2, a1);
2717        return;
2718    case INDEX_op_arm_ushl_vec:
2719        /* See above. */
2720        tcg_out_vreg3(s, INSN_VSHL_U, q, vece, a0, a2, a1);
2721        return;
2722    case INDEX_op_shli_vec:
2723        tcg_out_vshifti(s, INSN_VSHLI, q, a0, a1, a2 + (8 << vece));
2724        return;
2725    case INDEX_op_shri_vec:
2726        tcg_out_vshifti(s, INSN_VSHRI, q, a0, a1, (16 << vece) - a2);
2727        return;
2728    case INDEX_op_sari_vec:
2729        tcg_out_vshifti(s, INSN_VSARI, q, a0, a1, (16 << vece) - a2);
2730        return;
2731    case INDEX_op_arm_sli_vec:
2732        tcg_out_vshifti(s, INSN_VSLI, q, a0, a2, args[3] + (8 << vece));
2733        return;
2734
2735    case INDEX_op_andc_vec:
2736        if (!const_args[2]) {
2737            tcg_out_vreg3(s, INSN_VBIC, q, 0, a0, a1, a2);
2738            return;
2739        }
2740        a2 = ~a2;
2741        /* fall through */
2742    case INDEX_op_and_vec:
2743        if (const_args[2]) {
2744            is_shimm1632(~a2, &cmode, &imm8);
2745            if (a0 == a1) {
2746                tcg_out_vmovi(s, a0, q, 1, cmode | 1, imm8); /* VBICI */
2747                return;
2748            }
2749            tcg_out_vmovi(s, a0, q, 1, cmode, imm8); /* VMVNI */
2750            a2 = a0;
2751        }
2752        tcg_out_vreg3(s, INSN_VAND, q, 0, a0, a1, a2);
2753        return;
2754
2755    case INDEX_op_orc_vec:
2756        if (!const_args[2]) {
2757            tcg_out_vreg3(s, INSN_VORN, q, 0, a0, a1, a2);
2758            return;
2759        }
2760        a2 = ~a2;
2761        /* fall through */
2762    case INDEX_op_or_vec:
2763        if (const_args[2]) {
2764            is_shimm1632(a2, &cmode, &imm8);
2765            if (a0 == a1) {
2766                tcg_out_vmovi(s, a0, q, 0, cmode | 1, imm8); /* VORRI */
2767                return;
2768            }
2769            tcg_out_vmovi(s, a0, q, 0, cmode, imm8); /* VMOVI */
2770            a2 = a0;
2771        }
2772        tcg_out_vreg3(s, INSN_VORR, q, 0, a0, a1, a2);
2773        return;
2774
2775    case INDEX_op_cmp_vec:
2776        {
2777            TCGCond cond = args[3];
2778            ARMInsn insn;
2779
2780            switch (cond) {
2781            case TCG_COND_NE:
2782                if (const_args[2]) {
2783                    tcg_out_vreg3(s, INSN_VTST, q, vece, a0, a1, a1);
2784                } else {
2785                    tcg_out_vreg3(s, INSN_VCEQ, q, vece, a0, a1, a2);
2786                    tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a0);
2787                }
2788                break;
2789
2790            case TCG_COND_TSTNE:
2791            case TCG_COND_TSTEQ:
2792                if (const_args[2]) {
2793                    /* (x & 0) == 0 */
2794                    tcg_out_dupi_vec(s, type, MO_8, a0,
2795                                     -(cond == TCG_COND_TSTEQ));
2796                    break;
2797                }
2798                tcg_out_vreg3(s, INSN_VTST, q, vece, a0, a1, a2);
2799                if (cond == TCG_COND_TSTEQ) {
2800                    tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a0);
2801                }
2802                break;
2803
2804            default:
2805                if (const_args[2]) {
2806                    insn = vec_cmp0_insn[cond];
2807                    if (insn) {
2808                        tcg_out_vreg2(s, insn, q, vece, a0, a1);
2809                        return;
2810                    }
2811                    tcg_out_dupi_vec(s, type, MO_8, TCG_VEC_TMP, 0);
2812                    a2 = TCG_VEC_TMP;
2813                }
2814                insn = vec_cmp_insn[cond];
2815                if (insn == 0) {
2816                    TCGArg t;
2817                    t = a1, a1 = a2, a2 = t;
2818                    cond = tcg_swap_cond(cond);
2819                    insn = vec_cmp_insn[cond];
2820                    tcg_debug_assert(insn != 0);
2821                }
2822                tcg_out_vreg3(s, insn, q, vece, a0, a1, a2);
2823                break;
2824            }
2825        }
2826        return;
2827
2828    case INDEX_op_bitsel_vec:
2829        a3 = args[3];
2830        if (a0 == a3) {
2831            tcg_out_vreg3(s, INSN_VBIT, q, 0, a0, a2, a1);
2832        } else if (a0 == a2) {
2833            tcg_out_vreg3(s, INSN_VBIF, q, 0, a0, a3, a1);
2834        } else {
2835            tcg_out_mov(s, type, a0, a1);
2836            tcg_out_vreg3(s, INSN_VBSL, q, 0, a0, a2, a3);
2837        }
2838        return;
2839
2840    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
2841    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
2842    default:
2843        g_assert_not_reached();
2844    }
2845}
2846
2847int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
2848{
2849    switch (opc) {
2850    case INDEX_op_add_vec:
2851    case INDEX_op_sub_vec:
2852    case INDEX_op_and_vec:
2853    case INDEX_op_andc_vec:
2854    case INDEX_op_or_vec:
2855    case INDEX_op_orc_vec:
2856    case INDEX_op_xor_vec:
2857    case INDEX_op_not_vec:
2858    case INDEX_op_shli_vec:
2859    case INDEX_op_shri_vec:
2860    case INDEX_op_sari_vec:
2861    case INDEX_op_ssadd_vec:
2862    case INDEX_op_sssub_vec:
2863    case INDEX_op_usadd_vec:
2864    case INDEX_op_ussub_vec:
2865    case INDEX_op_bitsel_vec:
2866        return 1;
2867    case INDEX_op_abs_vec:
2868    case INDEX_op_cmp_vec:
2869    case INDEX_op_mul_vec:
2870    case INDEX_op_neg_vec:
2871    case INDEX_op_smax_vec:
2872    case INDEX_op_smin_vec:
2873    case INDEX_op_umax_vec:
2874    case INDEX_op_umin_vec:
2875        return vece < MO_64;
2876    case INDEX_op_shlv_vec:
2877    case INDEX_op_shrv_vec:
2878    case INDEX_op_sarv_vec:
2879    case INDEX_op_rotli_vec:
2880    case INDEX_op_rotlv_vec:
2881    case INDEX_op_rotrv_vec:
2882        return -1;
2883    default:
2884        return 0;
2885    }
2886}
2887
2888void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
2889                       TCGArg a0, ...)
2890{
2891    va_list va;
2892    TCGv_vec v0, v1, v2, t1, t2, c1;
2893    TCGArg a2;
2894
2895    va_start(va, a0);
2896    v0 = temp_tcgv_vec(arg_temp(a0));
2897    v1 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
2898    a2 = va_arg(va, TCGArg);
2899    va_end(va);
2900
2901    switch (opc) {
2902    case INDEX_op_shlv_vec:
2903        /*
2904         * Merely propagate shlv_vec to arm_ushl_vec.
2905         * In this way we don't set TCG_TARGET_HAS_shv_vec
2906         * because everything is done via expansion.
2907         */
2908        v2 = temp_tcgv_vec(arg_temp(a2));
2909        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
2910                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
2911        break;
2912
2913    case INDEX_op_shrv_vec:
2914    case INDEX_op_sarv_vec:
2915        /* Right shifts are negative left shifts for NEON.  */
2916        v2 = temp_tcgv_vec(arg_temp(a2));
2917        t1 = tcg_temp_new_vec(type);
2918        tcg_gen_neg_vec(vece, t1, v2);
2919        if (opc == INDEX_op_shrv_vec) {
2920            opc = INDEX_op_arm_ushl_vec;
2921        } else {
2922            opc = INDEX_op_arm_sshl_vec;
2923        }
2924        vec_gen_3(opc, type, vece, tcgv_vec_arg(v0),
2925                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
2926        tcg_temp_free_vec(t1);
2927        break;
2928
2929    case INDEX_op_rotli_vec:
2930        t1 = tcg_temp_new_vec(type);
2931        tcg_gen_shri_vec(vece, t1, v1, -a2 & ((8 << vece) - 1));
2932        vec_gen_4(INDEX_op_arm_sli_vec, type, vece,
2933                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(v1), a2);
2934        tcg_temp_free_vec(t1);
2935        break;
2936
2937    case INDEX_op_rotlv_vec:
2938        v2 = temp_tcgv_vec(arg_temp(a2));
2939        t1 = tcg_temp_new_vec(type);
2940        c1 = tcg_constant_vec(type, vece, 8 << vece);
2941        tcg_gen_sub_vec(vece, t1, v2, c1);
2942        /* Right shifts are negative left shifts for NEON.  */
2943        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
2944                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
2945        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
2946                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
2947        tcg_gen_or_vec(vece, v0, v0, t1);
2948        tcg_temp_free_vec(t1);
2949        break;
2950
2951    case INDEX_op_rotrv_vec:
2952        v2 = temp_tcgv_vec(arg_temp(a2));
2953        t1 = tcg_temp_new_vec(type);
2954        t2 = tcg_temp_new_vec(type);
2955        c1 = tcg_constant_vec(type, vece, 8 << vece);
2956        tcg_gen_neg_vec(vece, t1, v2);
2957        tcg_gen_sub_vec(vece, t2, c1, v2);
2958        /* Right shifts are negative left shifts for NEON.  */
2959        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
2960                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
2961        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t2),
2962                  tcgv_vec_arg(v1), tcgv_vec_arg(t2));
2963        tcg_gen_or_vec(vece, v0, t1, t2);
2964        tcg_temp_free_vec(t1);
2965        tcg_temp_free_vec(t2);
2966        break;
2967
2968    default:
2969        g_assert_not_reached();
2970    }
2971}
2972
2973static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
2974{
2975    int i;
2976    for (i = 0; i < count; ++i) {
2977        p[i] = INSN_NOP;
2978    }
2979}
2980
2981/* Compute frame size via macros, to share between tcg_target_qemu_prologue
2982   and tcg_register_jit.  */
2983
2984#define PUSH_SIZE  ((11 - 4 + 1 + 1) * sizeof(tcg_target_long))
2985
2986#define FRAME_SIZE \
2987    ((PUSH_SIZE \
2988      + TCG_STATIC_CALL_ARGS_SIZE \
2989      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
2990      + TCG_TARGET_STACK_ALIGN - 1) \
2991     & -TCG_TARGET_STACK_ALIGN)
2992
2993#define STACK_ADDEND  (FRAME_SIZE - PUSH_SIZE)
2994
2995static void tcg_target_qemu_prologue(TCGContext *s)
2996{
2997    /* Calling convention requires us to save r4-r11 and lr.  */
2998    /* stmdb sp!, { r4 - r11, lr } */
2999    tcg_out_ldstm(s, COND_AL, INSN_STMDB, TCG_REG_CALL_STACK,
3000                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
3001                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
3002                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_R14));
3003
3004    /* Reserve callee argument and tcg temp space.  */
3005    tcg_out_dat_rI(s, COND_AL, ARITH_SUB, TCG_REG_CALL_STACK,
3006                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
3007    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
3008                  CPU_TEMP_BUF_NLONGS * sizeof(long));
3009
3010    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
3011
3012    if (!tcg_use_softmmu && guest_base) {
3013        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_GUEST_BASE, guest_base);
3014        tcg_regset_set_reg(s->reserved_regs, TCG_REG_GUEST_BASE);
3015    }
3016
3017    tcg_out_b_reg(s, COND_AL, tcg_target_call_iarg_regs[1]);
3018
3019    /*
3020     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
3021     * and fall through to the rest of the epilogue.
3022     */
3023    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
3024    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, 0);
3025    tcg_out_epilogue(s);
3026}
3027
3028static void tcg_out_epilogue(TCGContext *s)
3029{
3030    /* Release local stack frame.  */
3031    tcg_out_dat_rI(s, COND_AL, ARITH_ADD, TCG_REG_CALL_STACK,
3032                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
3033
3034    /* ldmia sp!, { r4 - r11, pc } */
3035    tcg_out_ldstm(s, COND_AL, INSN_LDMIA, TCG_REG_CALL_STACK,
3036                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
3037                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
3038                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_PC));
3039}
3040
3041static void tcg_out_tb_start(TCGContext *s)
3042{
3043    /* nothing to do */
3044}
3045
3046typedef struct {
3047    DebugFrameHeader h;
3048    uint8_t fde_def_cfa[4];
3049    uint8_t fde_reg_ofs[18];
3050} DebugFrame;
3051
3052#define ELF_HOST_MACHINE EM_ARM
3053
3054/* We're expecting a 2 byte uleb128 encoded value.  */
3055QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
3056
3057static const DebugFrame debug_frame = {
3058    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
3059    .h.cie.id = -1,
3060    .h.cie.version = 1,
3061    .h.cie.code_align = 1,
3062    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
3063    .h.cie.return_column = 14,
3064
3065    /* Total FDE size does not include the "len" member.  */
3066    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
3067
3068    .fde_def_cfa = {
3069        12, 13,                         /* DW_CFA_def_cfa sp, ... */
3070        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
3071        (FRAME_SIZE >> 7)
3072    },
3073    .fde_reg_ofs = {
3074        /* The following must match the stmdb in the prologue.  */
3075        0x8e, 1,                        /* DW_CFA_offset, lr, -4 */
3076        0x8b, 2,                        /* DW_CFA_offset, r11, -8 */
3077        0x8a, 3,                        /* DW_CFA_offset, r10, -12 */
3078        0x89, 4,                        /* DW_CFA_offset, r9, -16 */
3079        0x88, 5,                        /* DW_CFA_offset, r8, -20 */
3080        0x87, 6,                        /* DW_CFA_offset, r7, -24 */
3081        0x86, 7,                        /* DW_CFA_offset, r6, -28 */
3082        0x85, 8,                        /* DW_CFA_offset, r5, -32 */
3083        0x84, 9,                        /* DW_CFA_offset, r4, -36 */
3084    }
3085};
3086
3087void tcg_register_jit(const void *buf, size_t buf_size)
3088{
3089    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
3090}
3091