xref: /openbmc/qemu/tcg/arm/tcg-target.c.inc (revision 79602f632a20228fc963161cd53f7d5f6a3bd953)
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_rIK(TCGContext *s, ARMCond cond, ARMInsn opc,
878                            ARMInsn opinv, TCGReg dst, TCGReg lhs, TCGArg rhs,
879                            bool rhs_is_const)
880{
881    if (rhs_is_const) {
882        int imm12 = encode_imm(rhs);
883        if (imm12 < 0) {
884            imm12 = encode_imm_nofail(~rhs);
885            opc = opinv;
886        }
887        tcg_out_dat_imm(s, cond, opc, dst, lhs, imm12);
888    } else {
889        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
890    }
891}
892
893static void tcg_out_dat_IN(TCGContext *s, ARMCond cond, ARMInsn opc,
894                           ARMInsn opneg, TCGReg dst, TCGReg lhs, TCGArg rhs)
895{
896    int imm12 = encode_imm(rhs);
897    if (imm12 < 0) {
898        imm12 = encode_imm_nofail(-rhs);
899        opc = opneg;
900    }
901    tcg_out_dat_imm(s, cond, opc, dst, lhs, imm12);
902}
903
904static void tcg_out_dat_rIN(TCGContext *s, ARMCond cond, ARMInsn opc,
905                            ARMInsn opneg, TCGReg dst, TCGReg lhs, TCGArg rhs,
906                            bool rhs_is_const)
907{
908    /* Emit either the reg,imm or reg,reg form of a data-processing insn.
909     * rhs must satisfy the "rIN" constraint.
910     */
911    if (rhs_is_const) {
912        tcg_out_dat_IN(s, cond, opc, opneg, dst, lhs, rhs);
913    } else {
914        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
915    }
916}
917
918static void tcg_out_mul32(TCGContext *s, ARMCond cond, TCGReg rd,
919                          TCGReg rn, TCGReg rm)
920{
921    /* mul */
922    tcg_out32(s, (cond << 28) | 0x90 | (rd << 16) | (rm << 8) | rn);
923}
924
925static void tcg_out_umull32(TCGContext *s, ARMCond cond, TCGReg rd0,
926                            TCGReg rd1, TCGReg rn, TCGReg rm)
927{
928    /* umull */
929    tcg_out32(s, (cond << 28) | 0x00800090 |
930              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
931}
932
933static void tcg_out_smull32(TCGContext *s, ARMCond cond, TCGReg rd0,
934                            TCGReg rd1, TCGReg rn, TCGReg rm)
935{
936    /* smull */
937    tcg_out32(s, (cond << 28) | 0x00c00090 |
938              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
939}
940
941static void tcg_out_sdiv(TCGContext *s, ARMCond cond,
942                         TCGReg rd, TCGReg rn, TCGReg rm)
943{
944    tcg_out32(s, 0x0710f010 | (cond << 28) | (rd << 16) | rn | (rm << 8));
945}
946
947static void tcg_out_udiv(TCGContext *s, ARMCond cond,
948                         TCGReg rd, TCGReg rn, TCGReg rm)
949{
950    tcg_out32(s, 0x0730f010 | (cond << 28) | (rd << 16) | rn | (rm << 8));
951}
952
953static void tcg_out_ext8s(TCGContext *s, TCGType t, TCGReg rd, TCGReg rn)
954{
955    /* sxtb */
956    tcg_out32(s, 0x06af0070 | (COND_AL << 28) | (rd << 12) | rn);
957}
958
959static void tcg_out_ext8u(TCGContext *s, TCGReg rd, TCGReg rn)
960{
961    tcg_out_dat_imm(s, COND_AL, ARITH_AND, rd, rn, 0xff);
962}
963
964static void tcg_out_ext16s(TCGContext *s, TCGType t, TCGReg rd, TCGReg rn)
965{
966    /* sxth */
967    tcg_out32(s, 0x06bf0070 | (COND_AL << 28) | (rd << 12) | rn);
968}
969
970static void tcg_out_ext16u(TCGContext *s, TCGReg rd, TCGReg rn)
971{
972    /* uxth */
973    tcg_out32(s, 0x06ff0070 | (COND_AL << 28) | (rd << 12) | rn);
974}
975
976static void tcg_out_ext32s(TCGContext *s, TCGReg rd, TCGReg rn)
977{
978    g_assert_not_reached();
979}
980
981static void tcg_out_ext32u(TCGContext *s, TCGReg rd, TCGReg rn)
982{
983    g_assert_not_reached();
984}
985
986static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg rd, TCGReg rn)
987{
988    g_assert_not_reached();
989}
990
991static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg rd, TCGReg rn)
992{
993    g_assert_not_reached();
994}
995
996static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg rd, TCGReg rn)
997{
998    g_assert_not_reached();
999}
1000
1001static void tcg_out_bswap16(TCGContext *s, ARMCond cond,
1002                            TCGReg rd, TCGReg rn, int flags)
1003{
1004    if (flags & TCG_BSWAP_OS) {
1005        /* revsh */
1006        tcg_out32(s, 0x06ff0fb0 | (cond << 28) | (rd << 12) | rn);
1007        return;
1008    }
1009
1010    /* rev16 */
1011    tcg_out32(s, 0x06bf0fb0 | (cond << 28) | (rd << 12) | rn);
1012    if ((flags & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1013        /* uxth */
1014        tcg_out32(s, 0x06ff0070 | (cond << 28) | (rd << 12) | rd);
1015    }
1016}
1017
1018static void tcg_out_bswap32(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rn)
1019{
1020    /* rev */
1021    tcg_out32(s, 0x06bf0f30 | (cond << 28) | (rd << 12) | rn);
1022}
1023
1024static void tcg_out_deposit(TCGContext *s, ARMCond cond, TCGReg rd,
1025                            TCGArg a1, int ofs, int len, bool const_a1)
1026{
1027    if (const_a1) {
1028        /* bfi becomes bfc with rn == 15.  */
1029        a1 = 15;
1030    }
1031    /* bfi/bfc */
1032    tcg_out32(s, 0x07c00010 | (cond << 28) | (rd << 12) | a1
1033              | (ofs << 7) | ((ofs + len - 1) << 16));
1034}
1035
1036static void tcg_out_extract(TCGContext *s, ARMCond cond, TCGReg rd,
1037                            TCGReg rn, int ofs, int len)
1038{
1039    /* According to gcc, AND can be faster. */
1040    if (ofs == 0 && len <= 8) {
1041        tcg_out_dat_imm(s, cond, ARITH_AND, rd, rn,
1042                        encode_imm_nofail((1 << len) - 1));
1043        return;
1044    }
1045
1046    if (use_armv7_instructions) {
1047        /* ubfx */
1048        tcg_out32(s, 0x07e00050 | (cond << 28) | (rd << 12) | rn
1049                  | (ofs << 7) | ((len - 1) << 16));
1050        return;
1051    }
1052
1053    assert(ofs % 8 == 0);
1054    switch (len) {
1055    case 8:
1056        /* uxtb */
1057        tcg_out32(s, 0x06ef0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1058        break;
1059    case 16:
1060        /* uxth */
1061        tcg_out32(s, 0x06ff0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1062        break;
1063    default:
1064        g_assert_not_reached();
1065    }
1066}
1067
1068static void tcg_out_sextract(TCGContext *s, ARMCond cond, TCGReg rd,
1069                             TCGReg rn, int ofs, int len)
1070{
1071    if (use_armv7_instructions) {
1072        /* sbfx */
1073        tcg_out32(s, 0x07a00050 | (cond << 28) | (rd << 12) | rn
1074                  | (ofs << 7) | ((len - 1) << 16));
1075        return;
1076    }
1077
1078    assert(ofs % 8 == 0);
1079    switch (len) {
1080    case 8:
1081        /* sxtb */
1082        tcg_out32(s, 0x06af0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1083        break;
1084    case 16:
1085        /* sxth */
1086        tcg_out32(s, 0x06bf0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1087        break;
1088    default:
1089        g_assert_not_reached();
1090    }
1091}
1092
1093
1094static void tcg_out_ld32u(TCGContext *s, ARMCond cond,
1095                          TCGReg rd, TCGReg rn, int32_t offset)
1096{
1097    if (offset > 0xfff || offset < -0xfff) {
1098        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1099        tcg_out_ld32_r(s, cond, rd, rn, TCG_REG_TMP);
1100    } else
1101        tcg_out_ld32_12(s, cond, rd, rn, offset);
1102}
1103
1104static void tcg_out_st32(TCGContext *s, ARMCond cond,
1105                         TCGReg rd, TCGReg rn, int32_t offset)
1106{
1107    if (offset > 0xfff || offset < -0xfff) {
1108        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1109        tcg_out_st32_r(s, cond, rd, rn, TCG_REG_TMP);
1110    } else
1111        tcg_out_st32_12(s, cond, rd, rn, offset);
1112}
1113
1114static void tcg_out_ld16u(TCGContext *s, ARMCond cond,
1115                          TCGReg rd, TCGReg rn, int32_t offset)
1116{
1117    if (offset > 0xff || offset < -0xff) {
1118        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1119        tcg_out_ld16u_r(s, cond, rd, rn, TCG_REG_TMP);
1120    } else
1121        tcg_out_ld16u_8(s, cond, rd, rn, offset);
1122}
1123
1124static void tcg_out_ld16s(TCGContext *s, ARMCond cond,
1125                          TCGReg rd, TCGReg rn, int32_t offset)
1126{
1127    if (offset > 0xff || offset < -0xff) {
1128        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1129        tcg_out_ld16s_r(s, cond, rd, rn, TCG_REG_TMP);
1130    } else
1131        tcg_out_ld16s_8(s, cond, rd, rn, offset);
1132}
1133
1134static void tcg_out_st16(TCGContext *s, ARMCond cond,
1135                         TCGReg rd, TCGReg rn, int32_t offset)
1136{
1137    if (offset > 0xff || offset < -0xff) {
1138        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1139        tcg_out_st16_r(s, cond, rd, rn, TCG_REG_TMP);
1140    } else
1141        tcg_out_st16_8(s, cond, rd, rn, offset);
1142}
1143
1144static void tcg_out_ld8u(TCGContext *s, ARMCond cond,
1145                         TCGReg rd, TCGReg rn, int32_t offset)
1146{
1147    if (offset > 0xfff || offset < -0xfff) {
1148        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1149        tcg_out_ld8_r(s, cond, rd, rn, TCG_REG_TMP);
1150    } else
1151        tcg_out_ld8_12(s, cond, rd, rn, offset);
1152}
1153
1154static void tcg_out_ld8s(TCGContext *s, ARMCond cond,
1155                         TCGReg rd, TCGReg rn, int32_t offset)
1156{
1157    if (offset > 0xff || offset < -0xff) {
1158        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1159        tcg_out_ld8s_r(s, cond, rd, rn, TCG_REG_TMP);
1160    } else
1161        tcg_out_ld8s_8(s, cond, rd, rn, offset);
1162}
1163
1164static void tcg_out_st8(TCGContext *s, ARMCond cond,
1165                        TCGReg rd, TCGReg rn, int32_t offset)
1166{
1167    if (offset > 0xfff || offset < -0xfff) {
1168        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1169        tcg_out_st8_r(s, cond, rd, rn, TCG_REG_TMP);
1170    } else
1171        tcg_out_st8_12(s, cond, rd, rn, offset);
1172}
1173
1174/*
1175 * The _goto case is normally between TBs within the same code buffer, and
1176 * with the code buffer limited to 16MB we wouldn't need the long case.
1177 * But we also use it for the tail-call to the qemu_ld/st helpers, which does.
1178 */
1179static void tcg_out_goto(TCGContext *s, ARMCond cond, const tcg_insn_unit *addr)
1180{
1181    intptr_t addri = (intptr_t)addr;
1182    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1183    bool arm_mode = !(addri & 1);
1184
1185    if (arm_mode && disp - 8 < 0x01fffffd && disp - 8 > -0x01fffffd) {
1186        tcg_out_b_imm(s, cond, disp);
1187        return;
1188    }
1189
1190    /* LDR is interworking from v5t. */
1191    tcg_out_movi_pool(s, cond, TCG_REG_PC, addri);
1192}
1193
1194/*
1195 * The call case is mostly used for helpers - so it's not unreasonable
1196 * for them to be beyond branch range.
1197 */
1198static void tcg_out_call_int(TCGContext *s, const tcg_insn_unit *addr)
1199{
1200    intptr_t addri = (intptr_t)addr;
1201    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1202    bool arm_mode = !(addri & 1);
1203
1204    if (disp - 8 < 0x02000000 && disp - 8 >= -0x02000000) {
1205        if (arm_mode) {
1206            tcg_out_bl_imm(s, COND_AL, disp);
1207        } else {
1208            tcg_out_blx_imm(s, disp);
1209        }
1210        return;
1211    }
1212
1213    tcg_out_movi32(s, COND_AL, TCG_REG_TMP, addri);
1214    tcg_out_blx_reg(s, COND_AL, TCG_REG_TMP);
1215}
1216
1217static void tcg_out_call(TCGContext *s, const tcg_insn_unit *addr,
1218                         const TCGHelperInfo *info)
1219{
1220    tcg_out_call_int(s, addr);
1221}
1222
1223static void tcg_out_goto_label(TCGContext *s, ARMCond cond, TCGLabel *l)
1224{
1225    if (l->has_value) {
1226        tcg_out_goto(s, cond, l->u.value_ptr);
1227    } else {
1228        tcg_out_reloc(s, s->code_ptr, R_ARM_PC24, l, 0);
1229        tcg_out_b_imm(s, cond, 0);
1230    }
1231}
1232
1233static void tcg_out_mb(TCGContext *s, TCGArg a0)
1234{
1235    if (use_armv7_instructions) {
1236        tcg_out32(s, INSN_DMB_ISH);
1237    } else {
1238        tcg_out32(s, INSN_DMB_MCR);
1239    }
1240}
1241
1242static TCGCond tcg_out_cmp(TCGContext *s, TCGCond cond, TCGReg a,
1243                           TCGArg b, int b_const)
1244{
1245    if (!is_tst_cond(cond)) {
1246        tcg_out_dat_rIN(s, COND_AL, ARITH_CMP, ARITH_CMN, 0, a, b, b_const);
1247        return cond;
1248    }
1249
1250    cond = tcg_tst_eqne_cond(cond);
1251    if (b_const) {
1252        int imm12 = encode_imm(b);
1253
1254        /*
1255         * The compare constraints allow rIN, but TST does not support N.
1256         * Be prepared to load the constant into a scratch register.
1257         */
1258        if (imm12 >= 0) {
1259            tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, a, imm12);
1260            return cond;
1261        }
1262        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, b);
1263        b = TCG_REG_TMP;
1264    }
1265    tcg_out_dat_reg(s, COND_AL, ARITH_TST, 0, a, b, SHIFT_IMM_LSL(0));
1266    return cond;
1267}
1268
1269static TCGCond tcg_out_cmp2(TCGContext *s, const TCGArg *args,
1270                            const int *const_args)
1271{
1272    TCGReg al = args[0];
1273    TCGReg ah = args[1];
1274    TCGArg bl = args[2];
1275    TCGArg bh = args[3];
1276    TCGCond cond = args[4];
1277    int const_bl = const_args[2];
1278    int const_bh = const_args[3];
1279
1280    switch (cond) {
1281    case TCG_COND_EQ:
1282    case TCG_COND_NE:
1283    case TCG_COND_LTU:
1284    case TCG_COND_LEU:
1285    case TCG_COND_GTU:
1286    case TCG_COND_GEU:
1287        /*
1288         * We perform a conditional comparison.  If the high half is
1289         * equal, then overwrite the flags with the comparison of the
1290         * low half.  The resulting flags cover the whole.
1291         */
1292        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, ah, bh, const_bh);
1293        tcg_out_dat_rI(s, COND_EQ, ARITH_CMP, 0, al, bl, const_bl);
1294        return cond;
1295
1296    case TCG_COND_TSTEQ:
1297    case TCG_COND_TSTNE:
1298        /* Similar, but with TST instead of CMP. */
1299        tcg_out_dat_rI(s, COND_AL, ARITH_TST, 0, ah, bh, const_bh);
1300        tcg_out_dat_rI(s, COND_EQ, ARITH_TST, 0, al, bl, const_bl);
1301        return tcg_tst_eqne_cond(cond);
1302
1303    case TCG_COND_LT:
1304    case TCG_COND_GE:
1305        /* We perform a double-word subtraction and examine the result.
1306           We do not actually need the result of the subtract, so the
1307           low part "subtract" is a compare.  For the high half we have
1308           no choice but to compute into a temporary.  */
1309        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, al, bl, const_bl);
1310        tcg_out_dat_rI(s, COND_AL, ARITH_SBC | TO_CPSR,
1311                       TCG_REG_TMP, ah, bh, const_bh);
1312        return cond;
1313
1314    case TCG_COND_LE:
1315    case TCG_COND_GT:
1316        /* Similar, but with swapped arguments, via reversed subtract.  */
1317        tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR,
1318                       TCG_REG_TMP, al, bl, const_bl);
1319        tcg_out_dat_rI(s, COND_AL, ARITH_RSC | TO_CPSR,
1320                       TCG_REG_TMP, ah, bh, const_bh);
1321        return tcg_swap_cond(cond);
1322
1323    default:
1324        g_assert_not_reached();
1325    }
1326}
1327
1328/*
1329 * Note that TCGReg references Q-registers.
1330 * Q-regno = 2 * D-regno, so shift left by 1 while inserting.
1331 */
1332static uint32_t encode_vd(TCGReg rd)
1333{
1334    tcg_debug_assert(rd >= TCG_REG_Q0);
1335    return (extract32(rd, 3, 1) << 22) | (extract32(rd, 0, 3) << 13);
1336}
1337
1338static uint32_t encode_vn(TCGReg rn)
1339{
1340    tcg_debug_assert(rn >= TCG_REG_Q0);
1341    return (extract32(rn, 3, 1) << 7) | (extract32(rn, 0, 3) << 17);
1342}
1343
1344static uint32_t encode_vm(TCGReg rm)
1345{
1346    tcg_debug_assert(rm >= TCG_REG_Q0);
1347    return (extract32(rm, 3, 1) << 5) | (extract32(rm, 0, 3) << 1);
1348}
1349
1350static void tcg_out_vreg2(TCGContext *s, ARMInsn insn, int q, int vece,
1351                          TCGReg d, TCGReg m)
1352{
1353    tcg_out32(s, insn | (vece << 18) | (q << 6) |
1354              encode_vd(d) | encode_vm(m));
1355}
1356
1357static void tcg_out_vreg3(TCGContext *s, ARMInsn insn, int q, int vece,
1358                          TCGReg d, TCGReg n, TCGReg m)
1359{
1360    tcg_out32(s, insn | (vece << 20) | (q << 6) |
1361              encode_vd(d) | encode_vn(n) | encode_vm(m));
1362}
1363
1364static void tcg_out_vmovi(TCGContext *s, TCGReg rd,
1365                          int q, int op, int cmode, uint8_t imm8)
1366{
1367    tcg_out32(s, INSN_VMOVI | encode_vd(rd) | (q << 6) | (op << 5)
1368              | (cmode << 8) | extract32(imm8, 0, 4)
1369              | (extract32(imm8, 4, 3) << 16)
1370              | (extract32(imm8, 7, 1) << 24));
1371}
1372
1373static void tcg_out_vshifti(TCGContext *s, ARMInsn insn, int q,
1374                            TCGReg rd, TCGReg rm, int l_imm6)
1375{
1376    tcg_out32(s, insn | (q << 6) | encode_vd(rd) | encode_vm(rm) |
1377              (extract32(l_imm6, 6, 1) << 7) |
1378              (extract32(l_imm6, 0, 6) << 16));
1379}
1380
1381static void tcg_out_vldst(TCGContext *s, ARMInsn insn,
1382                          TCGReg rd, TCGReg rn, int offset)
1383{
1384    if (offset != 0) {
1385        if (check_fit_imm(offset) || check_fit_imm(-offset)) {
1386            tcg_out_dat_rIN(s, COND_AL, ARITH_ADD, ARITH_SUB,
1387                            TCG_REG_TMP, rn, offset, true);
1388        } else {
1389            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP, offset);
1390            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1391                            TCG_REG_TMP, TCG_REG_TMP, rn, 0);
1392        }
1393        rn = TCG_REG_TMP;
1394    }
1395    tcg_out32(s, insn | (rn << 16) | encode_vd(rd) | 0xf);
1396}
1397
1398typedef struct {
1399    ARMCond cond;
1400    TCGReg base;
1401    int index;
1402    bool index_scratch;
1403    TCGAtomAlign aa;
1404} HostAddress;
1405
1406bool tcg_target_has_memory_bswap(MemOp memop)
1407{
1408    return false;
1409}
1410
1411static TCGReg ldst_ra_gen(TCGContext *s, const TCGLabelQemuLdst *l, int arg)
1412{
1413    /* We arrive at the slow path via "BLNE", so R14 contains l->raddr. */
1414    return TCG_REG_R14;
1415}
1416
1417static const TCGLdstHelperParam ldst_helper_param = {
1418    .ra_gen = ldst_ra_gen,
1419    .ntmp = 1,
1420    .tmp = { TCG_REG_TMP },
1421};
1422
1423static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1424{
1425    MemOp opc = get_memop(lb->oi);
1426
1427    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1428        return false;
1429    }
1430
1431    tcg_out_ld_helper_args(s, lb, &ldst_helper_param);
1432    tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SIZE]);
1433    tcg_out_ld_helper_ret(s, lb, false, &ldst_helper_param);
1434
1435    tcg_out_goto(s, COND_AL, lb->raddr);
1436    return true;
1437}
1438
1439static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1440{
1441    MemOp opc = get_memop(lb->oi);
1442
1443    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1444        return false;
1445    }
1446
1447    tcg_out_st_helper_args(s, lb, &ldst_helper_param);
1448
1449    /* Tail-call to the helper, which will return to the fast path.  */
1450    tcg_out_goto(s, COND_AL, qemu_st_helpers[opc & MO_SIZE]);
1451    return true;
1452}
1453
1454/* We expect to use an 9-bit sign-magnitude negative offset from ENV.  */
1455#define MIN_TLB_MASK_TABLE_OFS  -256
1456
1457static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
1458                                           TCGReg addr, MemOpIdx oi, bool is_ld)
1459{
1460    TCGLabelQemuLdst *ldst = NULL;
1461    MemOp opc = get_memop(oi);
1462    unsigned a_mask;
1463
1464    if (tcg_use_softmmu) {
1465        *h = (HostAddress){
1466            .cond = COND_AL,
1467            .base = addr,
1468            .index = TCG_REG_R1,
1469            .index_scratch = true,
1470        };
1471    } else {
1472        *h = (HostAddress){
1473            .cond = COND_AL,
1474            .base = addr,
1475            .index = guest_base ? TCG_REG_GUEST_BASE : -1,
1476            .index_scratch = false,
1477        };
1478    }
1479
1480    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, false);
1481    a_mask = (1 << h->aa.align) - 1;
1482
1483    if (tcg_use_softmmu) {
1484        int mem_index = get_mmuidx(oi);
1485        int cmp_off = is_ld ? offsetof(CPUTLBEntry, addr_read)
1486                            : offsetof(CPUTLBEntry, addr_write);
1487        int fast_off = tlb_mask_table_ofs(s, mem_index);
1488        unsigned s_mask = (1 << (opc & MO_SIZE)) - 1;
1489        TCGReg t_addr;
1490
1491        ldst = new_ldst_label(s);
1492        ldst->is_ld = is_ld;
1493        ldst->oi = oi;
1494        ldst->addr_reg = addr;
1495
1496        /* Load cpu->neg.tlb.f[mmu_idx].{mask,table} into {r0,r1}.  */
1497        QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, mask) != 0);
1498        QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, table) != 4);
1499        tcg_out_ldrd_8(s, COND_AL, TCG_REG_R0, TCG_AREG0, fast_off);
1500
1501        /* Extract the tlb index from the address into R0.  */
1502        tcg_out_dat_reg(s, COND_AL, ARITH_AND, TCG_REG_R0, TCG_REG_R0, addr,
1503                        SHIFT_IMM_LSR(s->page_bits - CPU_TLB_ENTRY_BITS));
1504
1505        /*
1506         * Add the tlb_table pointer, creating the CPUTLBEntry address in R1.
1507         * Load the tlb comparator into R2 and the fast path addend into R1.
1508         */
1509        if (cmp_off == 0) {
1510            tcg_out_ld32_rwb(s, COND_AL, TCG_REG_R2, TCG_REG_R1, TCG_REG_R0);
1511        } else {
1512            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1513                            TCG_REG_R1, TCG_REG_R1, TCG_REG_R0, 0);
1514            tcg_out_ld32_12(s, COND_AL, TCG_REG_R2, TCG_REG_R1, cmp_off);
1515        }
1516
1517        /* Load the tlb addend.  */
1518        tcg_out_ld32_12(s, COND_AL, TCG_REG_R1, TCG_REG_R1,
1519                        offsetof(CPUTLBEntry, addend));
1520
1521        /*
1522         * Check alignment, check comparators.
1523         * Do this in 2-4 insns.  Use MOVW for v7, if possible,
1524         * to reduce the number of sequential conditional instructions.
1525         * Almost all guests have at least 4k pages, which means that we need
1526         * to clear at least 9 bits even for an 8-byte memory, which means it
1527         * isn't worth checking for an immediate operand for BIC.
1528         *
1529         * For unaligned accesses, test the page of the last unit of alignment.
1530         * This leaves the least significant alignment bits unchanged, and of
1531         * course must be zero.
1532         */
1533        t_addr = addr;
1534        if (a_mask < s_mask) {
1535            t_addr = TCG_REG_R0;
1536            tcg_out_dat_imm(s, COND_AL, ARITH_ADD, t_addr,
1537                            addr, s_mask - a_mask);
1538        }
1539        if (use_armv7_instructions && s->page_bits <= 16) {
1540            tcg_out_movi32(s, COND_AL, TCG_REG_TMP, ~(s->page_mask | a_mask));
1541            tcg_out_dat_reg(s, COND_AL, ARITH_BIC, TCG_REG_TMP,
1542                            t_addr, TCG_REG_TMP, 0);
1543            tcg_out_dat_reg(s, COND_AL, ARITH_CMP, 0,
1544                            TCG_REG_R2, TCG_REG_TMP, 0);
1545        } else {
1546            if (a_mask) {
1547                tcg_debug_assert(a_mask <= 0xff);
1548                tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addr, a_mask);
1549            }
1550            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0, t_addr,
1551                            SHIFT_IMM_LSR(s->page_bits));
1552            tcg_out_dat_reg(s, (a_mask ? COND_EQ : COND_AL), ARITH_CMP,
1553                            0, TCG_REG_R2, TCG_REG_TMP,
1554                            SHIFT_IMM_LSL(s->page_bits));
1555        }
1556    } else if (a_mask) {
1557        ldst = new_ldst_label(s);
1558        ldst->is_ld = is_ld;
1559        ldst->oi = oi;
1560        ldst->addr_reg = addr;
1561
1562        /* We are expecting alignment to max out at 7 */
1563        tcg_debug_assert(a_mask <= 0xff);
1564        /* tst addr, #mask */
1565        tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addr, a_mask);
1566    }
1567
1568    return ldst;
1569}
1570
1571static void tcg_out_qemu_ld_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1572                                   TCGReg datahi, HostAddress h)
1573{
1574    TCGReg base;
1575
1576    /* Byte swapping is left to middle-end expansion. */
1577    tcg_debug_assert((opc & MO_BSWAP) == 0);
1578
1579    switch (opc & MO_SSIZE) {
1580    case MO_UB:
1581        if (h.index < 0) {
1582            tcg_out_ld8_12(s, h.cond, datalo, h.base, 0);
1583        } else {
1584            tcg_out_ld8_r(s, h.cond, datalo, h.base, h.index);
1585        }
1586        break;
1587    case MO_SB:
1588        if (h.index < 0) {
1589            tcg_out_ld8s_8(s, h.cond, datalo, h.base, 0);
1590        } else {
1591            tcg_out_ld8s_r(s, h.cond, datalo, h.base, h.index);
1592        }
1593        break;
1594    case MO_UW:
1595        if (h.index < 0) {
1596            tcg_out_ld16u_8(s, h.cond, datalo, h.base, 0);
1597        } else {
1598            tcg_out_ld16u_r(s, h.cond, datalo, h.base, h.index);
1599        }
1600        break;
1601    case MO_SW:
1602        if (h.index < 0) {
1603            tcg_out_ld16s_8(s, h.cond, datalo, h.base, 0);
1604        } else {
1605            tcg_out_ld16s_r(s, h.cond, datalo, h.base, h.index);
1606        }
1607        break;
1608    case MO_UL:
1609        if (h.index < 0) {
1610            tcg_out_ld32_12(s, h.cond, datalo, h.base, 0);
1611        } else {
1612            tcg_out_ld32_r(s, h.cond, datalo, h.base, h.index);
1613        }
1614        break;
1615    case MO_UQ:
1616        /* We used pair allocation for datalo, so already should be aligned. */
1617        tcg_debug_assert((datalo & 1) == 0);
1618        tcg_debug_assert(datahi == datalo + 1);
1619        /* LDRD requires alignment; double-check that. */
1620        if (memop_alignment_bits(opc) >= MO_64) {
1621            if (h.index < 0) {
1622                tcg_out_ldrd_8(s, h.cond, datalo, h.base, 0);
1623                break;
1624            }
1625            /*
1626             * Rm (the second address op) must not overlap Rt or Rt + 1.
1627             * Since datalo is aligned, we can simplify the test via alignment.
1628             * Flip the two address arguments if that works.
1629             */
1630            if ((h.index & ~1) != datalo) {
1631                tcg_out_ldrd_r(s, h.cond, datalo, h.base, h.index);
1632                break;
1633            }
1634            if ((h.base & ~1) != datalo) {
1635                tcg_out_ldrd_r(s, h.cond, datalo, h.index, h.base);
1636                break;
1637            }
1638        }
1639        if (h.index < 0) {
1640            base = h.base;
1641            if (datalo == h.base) {
1642                tcg_out_mov_reg(s, h.cond, TCG_REG_TMP, base);
1643                base = TCG_REG_TMP;
1644            }
1645        } else if (h.index_scratch) {
1646            tcg_out_ld32_rwb(s, h.cond, datalo, h.index, h.base);
1647            tcg_out_ld32_12(s, h.cond, datahi, h.index, 4);
1648            break;
1649        } else {
1650            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1651                            h.base, h.index, SHIFT_IMM_LSL(0));
1652            base = TCG_REG_TMP;
1653        }
1654        tcg_out_ld32_12(s, h.cond, datalo, base, 0);
1655        tcg_out_ld32_12(s, h.cond, datahi, base, 4);
1656        break;
1657    default:
1658        g_assert_not_reached();
1659    }
1660}
1661
1662static void tcg_out_qemu_ld(TCGContext *s, TCGReg datalo, TCGReg datahi,
1663                            TCGReg addr, MemOpIdx oi, TCGType data_type)
1664{
1665    MemOp opc = get_memop(oi);
1666    TCGLabelQemuLdst *ldst;
1667    HostAddress h;
1668
1669    ldst = prepare_host_addr(s, &h, addr, oi, true);
1670    if (ldst) {
1671        ldst->type = data_type;
1672        ldst->datalo_reg = datalo;
1673        ldst->datahi_reg = datahi;
1674
1675        /*
1676         * This a conditional BL only to load a pointer within this
1677         * opcode into LR for the slow path.  We will not be using
1678         * the value for a tail call.
1679         */
1680        ldst->label_ptr[0] = s->code_ptr;
1681        tcg_out_bl_imm(s, COND_NE, 0);
1682
1683        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1684        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1685    } else {
1686        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1687    }
1688}
1689
1690static void tcg_out_qemu_st_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1691                                   TCGReg datahi, HostAddress h)
1692{
1693    /* Byte swapping is left to middle-end expansion. */
1694    tcg_debug_assert((opc & MO_BSWAP) == 0);
1695
1696    switch (opc & MO_SIZE) {
1697    case MO_8:
1698        if (h.index < 0) {
1699            tcg_out_st8_12(s, h.cond, datalo, h.base, 0);
1700        } else {
1701            tcg_out_st8_r(s, h.cond, datalo, h.base, h.index);
1702        }
1703        break;
1704    case MO_16:
1705        if (h.index < 0) {
1706            tcg_out_st16_8(s, h.cond, datalo, h.base, 0);
1707        } else {
1708            tcg_out_st16_r(s, h.cond, datalo, h.base, h.index);
1709        }
1710        break;
1711    case MO_32:
1712        if (h.index < 0) {
1713            tcg_out_st32_12(s, h.cond, datalo, h.base, 0);
1714        } else {
1715            tcg_out_st32_r(s, h.cond, datalo, h.base, h.index);
1716        }
1717        break;
1718    case MO_64:
1719        /* We used pair allocation for datalo, so already should be aligned. */
1720        tcg_debug_assert((datalo & 1) == 0);
1721        tcg_debug_assert(datahi == datalo + 1);
1722        /* STRD requires alignment; double-check that. */
1723        if (memop_alignment_bits(opc) >= MO_64) {
1724            if (h.index < 0) {
1725                tcg_out_strd_8(s, h.cond, datalo, h.base, 0);
1726            } else {
1727                tcg_out_strd_r(s, h.cond, datalo, h.base, h.index);
1728            }
1729        } else if (h.index < 0) {
1730            tcg_out_st32_12(s, h.cond, datalo, h.base, 0);
1731            tcg_out_st32_12(s, h.cond, datahi, h.base, 4);
1732        } else if (h.index_scratch) {
1733            tcg_out_st32_rwb(s, h.cond, datalo, h.index, h.base);
1734            tcg_out_st32_12(s, h.cond, datahi, h.index, 4);
1735        } else {
1736            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1737                            h.base, h.index, SHIFT_IMM_LSL(0));
1738            tcg_out_st32_12(s, h.cond, datalo, TCG_REG_TMP, 0);
1739            tcg_out_st32_12(s, h.cond, datahi, TCG_REG_TMP, 4);
1740        }
1741        break;
1742    default:
1743        g_assert_not_reached();
1744    }
1745}
1746
1747static void tcg_out_qemu_st(TCGContext *s, TCGReg datalo, TCGReg datahi,
1748                            TCGReg addr, MemOpIdx oi, TCGType data_type)
1749{
1750    MemOp opc = get_memop(oi);
1751    TCGLabelQemuLdst *ldst;
1752    HostAddress h;
1753
1754    ldst = prepare_host_addr(s, &h, addr, oi, false);
1755    if (ldst) {
1756        ldst->type = data_type;
1757        ldst->datalo_reg = datalo;
1758        ldst->datahi_reg = datahi;
1759
1760        h.cond = COND_EQ;
1761        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1762
1763        /* The conditional call is last, as we're going to return here. */
1764        ldst->label_ptr[0] = s->code_ptr;
1765        tcg_out_bl_imm(s, COND_NE, 0);
1766        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1767    } else {
1768        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1769    }
1770}
1771
1772static void tcg_out_epilogue(TCGContext *s);
1773
1774static void tcg_out_exit_tb(TCGContext *s, uintptr_t arg)
1775{
1776    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, arg);
1777    tcg_out_epilogue(s);
1778}
1779
1780static void tcg_out_goto_tb(TCGContext *s, int which)
1781{
1782    uintptr_t i_addr;
1783    intptr_t i_disp;
1784
1785    /* Direct branch will be patched by tb_target_set_jmp_target. */
1786    set_jmp_insn_offset(s, which);
1787    tcg_out32(s, INSN_NOP);
1788
1789    /* When branch is out of range, fall through to indirect. */
1790    i_addr = get_jmp_target_addr(s, which);
1791    i_disp = tcg_pcrel_diff(s, (void *)i_addr) - 8;
1792    tcg_debug_assert(i_disp < 0);
1793    if (i_disp >= -0xfff) {
1794        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_PC, i_disp);
1795    } else {
1796        /*
1797         * The TB is close, but outside the 12 bits addressable by
1798         * the load.  We can extend this to 20 bits with a sub of a
1799         * shifted immediate from pc.
1800         */
1801        int h = -i_disp;
1802        int l = -(h & 0xfff);
1803
1804        h = encode_imm_nofail(h + l);
1805        tcg_out_dat_imm(s, COND_AL, ARITH_SUB, TCG_REG_R0, TCG_REG_PC, h);
1806        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_R0, l);
1807    }
1808    set_jmp_reset_offset(s, which);
1809}
1810
1811void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
1812                              uintptr_t jmp_rx, uintptr_t jmp_rw)
1813{
1814    uintptr_t addr = tb->jmp_target_addr[n];
1815    ptrdiff_t offset = addr - (jmp_rx + 8);
1816    tcg_insn_unit insn;
1817
1818    /* Either directly branch, or fall through to indirect branch. */
1819    if (offset == sextract64(offset, 0, 26)) {
1820        /* B <addr> */
1821        insn = deposit32((COND_AL << 28) | INSN_B, 0, 24, offset >> 2);
1822    } else {
1823        insn = INSN_NOP;
1824    }
1825
1826    qatomic_set((uint32_t *)jmp_rw, insn);
1827    flush_idcache_range(jmp_rx, jmp_rw, 4);
1828}
1829
1830
1831static void tgen_add(TCGContext *s, TCGType type,
1832                     TCGReg a0, TCGReg a1, TCGReg a2)
1833{
1834    tcg_out_dat_reg(s, COND_AL, ARITH_ADD, a0, a1, a2, SHIFT_IMM_LSL(0));
1835}
1836
1837static void tgen_addi(TCGContext *s, TCGType type,
1838                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1839{
1840    tcg_out_dat_IN(s, COND_AL, ARITH_ADD, ARITH_SUB, a0, a1, a2);
1841}
1842
1843static const TCGOutOpBinary outop_add = {
1844    .base.static_constraint = C_O1_I2(r, r, rIN),
1845    .out_rrr = tgen_add,
1846    .out_rri = tgen_addi,
1847};
1848
1849
1850static void tcg_out_op(TCGContext *s, TCGOpcode opc, TCGType type,
1851                       const TCGArg args[TCG_MAX_OP_ARGS],
1852                       const int const_args[TCG_MAX_OP_ARGS])
1853{
1854    TCGArg a0, a1, a2, a3, a4, a5;
1855    int c;
1856
1857    switch (opc) {
1858    case INDEX_op_goto_ptr:
1859        tcg_out_b_reg(s, COND_AL, args[0]);
1860        break;
1861    case INDEX_op_br:
1862        tcg_out_goto_label(s, COND_AL, arg_label(args[0]));
1863        break;
1864
1865    case INDEX_op_ld8u_i32:
1866        tcg_out_ld8u(s, COND_AL, args[0], args[1], args[2]);
1867        break;
1868    case INDEX_op_ld8s_i32:
1869        tcg_out_ld8s(s, COND_AL, args[0], args[1], args[2]);
1870        break;
1871    case INDEX_op_ld16u_i32:
1872        tcg_out_ld16u(s, COND_AL, args[0], args[1], args[2]);
1873        break;
1874    case INDEX_op_ld16s_i32:
1875        tcg_out_ld16s(s, COND_AL, args[0], args[1], args[2]);
1876        break;
1877    case INDEX_op_ld_i32:
1878        tcg_out_ld32u(s, COND_AL, args[0], args[1], args[2]);
1879        break;
1880    case INDEX_op_st8_i32:
1881        tcg_out_st8(s, COND_AL, args[0], args[1], args[2]);
1882        break;
1883    case INDEX_op_st16_i32:
1884        tcg_out_st16(s, COND_AL, args[0], args[1], args[2]);
1885        break;
1886    case INDEX_op_st_i32:
1887        tcg_out_st32(s, COND_AL, args[0], args[1], args[2]);
1888        break;
1889
1890    case INDEX_op_movcond_i32:
1891        /* Constraints mean that v2 is always in the same register as dest,
1892         * so we only need to do "if condition passed, move v1 to dest".
1893         */
1894        c = tcg_out_cmp(s, args[5], args[1], args[2], const_args[2]);
1895        tcg_out_dat_rIK(s, tcg_cond_to_arm_cond[c], ARITH_MOV,
1896                        ARITH_MVN, args[0], 0, args[3], const_args[3]);
1897        break;
1898    case INDEX_op_sub_i32:
1899        if (const_args[1]) {
1900            if (const_args[2]) {
1901                tcg_out_movi32(s, COND_AL, args[0], args[1] - args[2]);
1902            } else {
1903                tcg_out_dat_rI(s, COND_AL, ARITH_RSB,
1904                               args[0], args[2], args[1], 1);
1905            }
1906        } else {
1907            tcg_out_dat_rIN(s, COND_AL, ARITH_SUB, ARITH_ADD,
1908                            args[0], args[1], args[2], const_args[2]);
1909        }
1910        break;
1911    case INDEX_op_and_i32:
1912        tcg_out_dat_rIK(s, COND_AL, ARITH_AND, ARITH_BIC,
1913                        args[0], args[1], args[2], const_args[2]);
1914        break;
1915    case INDEX_op_andc_i32:
1916        tcg_out_dat_rIK(s, COND_AL, ARITH_BIC, ARITH_AND,
1917                        args[0], args[1], args[2], const_args[2]);
1918        break;
1919    case INDEX_op_or_i32:
1920        c = ARITH_ORR;
1921        goto gen_arith;
1922    case INDEX_op_xor_i32:
1923        c = ARITH_EOR;
1924        /* Fall through.  */
1925    gen_arith:
1926        tcg_out_dat_rI(s, COND_AL, c, args[0], args[1], args[2], const_args[2]);
1927        break;
1928    case INDEX_op_add2_i32:
1929        a0 = args[0], a1 = args[1], a2 = args[2];
1930        a3 = args[3], a4 = args[4], a5 = args[5];
1931        if (a0 == a3 || (a0 == a5 && !const_args[5])) {
1932            a0 = TCG_REG_TMP;
1933        }
1934        tcg_out_dat_rIN(s, COND_AL, ARITH_ADD | TO_CPSR, ARITH_SUB | TO_CPSR,
1935                        a0, a2, a4, const_args[4]);
1936        tcg_out_dat_rIK(s, COND_AL, ARITH_ADC, ARITH_SBC,
1937                        a1, a3, a5, const_args[5]);
1938        tcg_out_mov_reg(s, COND_AL, args[0], a0);
1939        break;
1940    case INDEX_op_sub2_i32:
1941        a0 = args[0], a1 = args[1], a2 = args[2];
1942        a3 = args[3], a4 = args[4], a5 = args[5];
1943        if ((a0 == a3 && !const_args[3]) || (a0 == a5 && !const_args[5])) {
1944            a0 = TCG_REG_TMP;
1945        }
1946        if (const_args[2]) {
1947            if (const_args[4]) {
1948                tcg_out_movi32(s, COND_AL, a0, a4);
1949                a4 = a0;
1950            }
1951            tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR, a0, a4, a2, 1);
1952        } else {
1953            tcg_out_dat_rIN(s, COND_AL, ARITH_SUB | TO_CPSR,
1954                            ARITH_ADD | TO_CPSR, a0, a2, a4, const_args[4]);
1955        }
1956        if (const_args[3]) {
1957            if (const_args[5]) {
1958                tcg_out_movi32(s, COND_AL, a1, a5);
1959                a5 = a1;
1960            }
1961            tcg_out_dat_rI(s, COND_AL, ARITH_RSC, a1, a5, a3, 1);
1962        } else {
1963            tcg_out_dat_rIK(s, COND_AL, ARITH_SBC, ARITH_ADC,
1964                            a1, a3, a5, const_args[5]);
1965        }
1966        tcg_out_mov_reg(s, COND_AL, args[0], a0);
1967        break;
1968    case INDEX_op_neg_i32:
1969        tcg_out_dat_imm(s, COND_AL, ARITH_RSB, args[0], args[1], 0);
1970        break;
1971    case INDEX_op_not_i32:
1972        tcg_out_dat_reg(s, COND_AL,
1973                        ARITH_MVN, args[0], 0, args[1], SHIFT_IMM_LSL(0));
1974        break;
1975    case INDEX_op_mul_i32:
1976        tcg_out_mul32(s, COND_AL, args[0], args[1], args[2]);
1977        break;
1978    case INDEX_op_mulu2_i32:
1979        tcg_out_umull32(s, COND_AL, args[0], args[1], args[2], args[3]);
1980        break;
1981    case INDEX_op_muls2_i32:
1982        tcg_out_smull32(s, COND_AL, args[0], args[1], args[2], args[3]);
1983        break;
1984    /* XXX: Perhaps args[2] & 0x1f is wrong */
1985    case INDEX_op_shl_i32:
1986        c = const_args[2] ?
1987                SHIFT_IMM_LSL(args[2] & 0x1f) : SHIFT_REG_LSL(args[2]);
1988        goto gen_shift32;
1989    case INDEX_op_shr_i32:
1990        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_LSR(args[2] & 0x1f) :
1991                SHIFT_IMM_LSL(0) : SHIFT_REG_LSR(args[2]);
1992        goto gen_shift32;
1993    case INDEX_op_sar_i32:
1994        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_ASR(args[2] & 0x1f) :
1995                SHIFT_IMM_LSL(0) : SHIFT_REG_ASR(args[2]);
1996        goto gen_shift32;
1997    case INDEX_op_rotr_i32:
1998        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_ROR(args[2] & 0x1f) :
1999                SHIFT_IMM_LSL(0) : SHIFT_REG_ROR(args[2]);
2000        /* Fall through.  */
2001    gen_shift32:
2002        tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1], c);
2003        break;
2004
2005    case INDEX_op_rotl_i32:
2006        if (const_args[2]) {
2007            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1],
2008                            ((0x20 - args[2]) & 0x1f) ?
2009                            SHIFT_IMM_ROR((0x20 - args[2]) & 0x1f) :
2010                            SHIFT_IMM_LSL(0));
2011        } else {
2012            tcg_out_dat_imm(s, COND_AL, ARITH_RSB, TCG_REG_TMP, args[2], 0x20);
2013            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1],
2014                            SHIFT_REG_ROR(TCG_REG_TMP));
2015        }
2016        break;
2017
2018    case INDEX_op_ctz_i32:
2019        tcg_out_dat_reg(s, COND_AL, INSN_RBIT, TCG_REG_TMP, 0, args[1], 0);
2020        a1 = TCG_REG_TMP;
2021        goto do_clz;
2022
2023    case INDEX_op_clz_i32:
2024        a1 = args[1];
2025    do_clz:
2026        a0 = args[0];
2027        a2 = args[2];
2028        c = const_args[2];
2029        if (c && a2 == 32) {
2030            tcg_out_dat_reg(s, COND_AL, INSN_CLZ, a0, 0, a1, 0);
2031            break;
2032        }
2033        tcg_out_dat_imm(s, COND_AL, ARITH_CMP, 0, a1, 0);
2034        tcg_out_dat_reg(s, COND_NE, INSN_CLZ, a0, 0, a1, 0);
2035        if (c || a0 != a2) {
2036            tcg_out_dat_rIK(s, COND_EQ, ARITH_MOV, ARITH_MVN, a0, 0, a2, c);
2037        }
2038        break;
2039
2040    case INDEX_op_brcond_i32:
2041        c = tcg_out_cmp(s, args[2], args[0], args[1], const_args[1]);
2042        tcg_out_goto_label(s, tcg_cond_to_arm_cond[c], arg_label(args[3]));
2043        break;
2044    case INDEX_op_setcond_i32:
2045        c = tcg_out_cmp(s, args[3], args[1], args[2], const_args[2]);
2046        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c],
2047                        ARITH_MOV, args[0], 0, 1);
2048        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
2049                        ARITH_MOV, args[0], 0, 0);
2050        break;
2051    case INDEX_op_negsetcond_i32:
2052        c = tcg_out_cmp(s, args[3], args[1], args[2], const_args[2]);
2053        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c],
2054                        ARITH_MVN, args[0], 0, 0);
2055        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
2056                        ARITH_MOV, args[0], 0, 0);
2057        break;
2058
2059    case INDEX_op_brcond2_i32:
2060        c = tcg_out_cmp2(s, args, const_args);
2061        tcg_out_goto_label(s, tcg_cond_to_arm_cond[c], arg_label(args[5]));
2062        break;
2063    case INDEX_op_setcond2_i32:
2064        c = tcg_out_cmp2(s, args + 1, const_args + 1);
2065        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c], ARITH_MOV, args[0], 0, 1);
2066        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
2067                        ARITH_MOV, args[0], 0, 0);
2068        break;
2069
2070    case INDEX_op_qemu_ld_i32:
2071        tcg_out_qemu_ld(s, args[0], -1, args[1], args[2], TCG_TYPE_I32);
2072        break;
2073    case INDEX_op_qemu_ld_i64:
2074        tcg_out_qemu_ld(s, args[0], args[1], args[2], args[3], TCG_TYPE_I64);
2075        break;
2076
2077    case INDEX_op_qemu_st_i32:
2078        tcg_out_qemu_st(s, args[0], -1, args[1], args[2], TCG_TYPE_I32);
2079        break;
2080    case INDEX_op_qemu_st_i64:
2081        tcg_out_qemu_st(s, args[0], args[1], args[2], args[3], TCG_TYPE_I64);
2082        break;
2083
2084    case INDEX_op_bswap16_i32:
2085        tcg_out_bswap16(s, COND_AL, args[0], args[1], args[2]);
2086        break;
2087    case INDEX_op_bswap32_i32:
2088        tcg_out_bswap32(s, COND_AL, args[0], args[1]);
2089        break;
2090
2091    case INDEX_op_deposit_i32:
2092        tcg_out_deposit(s, COND_AL, args[0], args[2],
2093                        args[3], args[4], const_args[2]);
2094        break;
2095    case INDEX_op_extract_i32:
2096        tcg_out_extract(s, COND_AL, args[0], args[1], args[2], args[3]);
2097        break;
2098    case INDEX_op_sextract_i32:
2099        tcg_out_sextract(s, COND_AL, args[0], args[1], args[2], args[3]);
2100        break;
2101    case INDEX_op_extract2_i32:
2102        /* ??? These optimization vs zero should be generic.  */
2103        /* ??? But we can't substitute 2 for 1 in the opcode stream yet.  */
2104        if (const_args[1]) {
2105            if (const_args[2]) {
2106                tcg_out_movi(s, TCG_TYPE_REG, args[0], 0);
2107            } else {
2108                tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2109                                args[2], SHIFT_IMM_LSL(32 - args[3]));
2110            }
2111        } else if (const_args[2]) {
2112            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2113                            args[1], SHIFT_IMM_LSR(args[3]));
2114        } else {
2115            /* We can do extract2 in 2 insns, vs the 3 required otherwise.  */
2116            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0,
2117                            args[2], SHIFT_IMM_LSL(32 - args[3]));
2118            tcg_out_dat_reg(s, COND_AL, ARITH_ORR, args[0], TCG_REG_TMP,
2119                            args[1], SHIFT_IMM_LSR(args[3]));
2120        }
2121        break;
2122
2123    case INDEX_op_div_i32:
2124        tcg_out_sdiv(s, COND_AL, args[0], args[1], args[2]);
2125        break;
2126    case INDEX_op_divu_i32:
2127        tcg_out_udiv(s, COND_AL, args[0], args[1], args[2]);
2128        break;
2129
2130    case INDEX_op_mb:
2131        tcg_out_mb(s, args[0]);
2132        break;
2133
2134    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
2135    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
2136    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
2137    default:
2138        g_assert_not_reached();
2139    }
2140}
2141
2142static TCGConstraintSetIndex
2143tcg_target_op_def(TCGOpcode op, TCGType type, unsigned flags)
2144{
2145    switch (op) {
2146    case INDEX_op_goto_ptr:
2147        return C_O0_I1(r);
2148
2149    case INDEX_op_ld8u_i32:
2150    case INDEX_op_ld8s_i32:
2151    case INDEX_op_ld16u_i32:
2152    case INDEX_op_ld16s_i32:
2153    case INDEX_op_ld_i32:
2154    case INDEX_op_neg_i32:
2155    case INDEX_op_not_i32:
2156    case INDEX_op_bswap16_i32:
2157    case INDEX_op_bswap32_i32:
2158    case INDEX_op_extract_i32:
2159    case INDEX_op_sextract_i32:
2160        return C_O1_I1(r, r);
2161
2162    case INDEX_op_st8_i32:
2163    case INDEX_op_st16_i32:
2164    case INDEX_op_st_i32:
2165        return C_O0_I2(r, r);
2166
2167    case INDEX_op_sub_i32:
2168    case INDEX_op_setcond_i32:
2169    case INDEX_op_negsetcond_i32:
2170        return C_O1_I2(r, r, rIN);
2171
2172    case INDEX_op_and_i32:
2173    case INDEX_op_andc_i32:
2174    case INDEX_op_clz_i32:
2175    case INDEX_op_ctz_i32:
2176        return C_O1_I2(r, r, rIK);
2177
2178    case INDEX_op_mul_i32:
2179    case INDEX_op_div_i32:
2180    case INDEX_op_divu_i32:
2181        return C_O1_I2(r, r, r);
2182
2183    case INDEX_op_mulu2_i32:
2184    case INDEX_op_muls2_i32:
2185        return C_O2_I2(r, r, r, r);
2186
2187    case INDEX_op_or_i32:
2188    case INDEX_op_xor_i32:
2189        return C_O1_I2(r, r, rI);
2190
2191    case INDEX_op_shl_i32:
2192    case INDEX_op_shr_i32:
2193    case INDEX_op_sar_i32:
2194    case INDEX_op_rotl_i32:
2195    case INDEX_op_rotr_i32:
2196        return C_O1_I2(r, r, ri);
2197
2198    case INDEX_op_brcond_i32:
2199        return C_O0_I2(r, rIN);
2200    case INDEX_op_deposit_i32:
2201        return C_O1_I2(r, 0, rZ);
2202    case INDEX_op_extract2_i32:
2203        return C_O1_I2(r, rZ, rZ);
2204    case INDEX_op_movcond_i32:
2205        return C_O1_I4(r, r, rIN, rIK, 0);
2206    case INDEX_op_add2_i32:
2207        return C_O2_I4(r, r, r, r, rIN, rIK);
2208    case INDEX_op_sub2_i32:
2209        return C_O2_I4(r, r, rI, rI, rIN, rIK);
2210    case INDEX_op_brcond2_i32:
2211        return C_O0_I4(r, r, rI, rI);
2212    case INDEX_op_setcond2_i32:
2213        return C_O1_I4(r, r, r, rI, rI);
2214
2215    case INDEX_op_qemu_ld_i32:
2216        return C_O1_I1(r, q);
2217    case INDEX_op_qemu_ld_i64:
2218        return C_O2_I1(e, p, q);
2219    case INDEX_op_qemu_st_i32:
2220        return C_O0_I2(q, q);
2221    case INDEX_op_qemu_st_i64:
2222        return C_O0_I3(Q, p, q);
2223
2224    case INDEX_op_st_vec:
2225        return C_O0_I2(w, r);
2226    case INDEX_op_ld_vec:
2227    case INDEX_op_dupm_vec:
2228        return C_O1_I1(w, r);
2229    case INDEX_op_dup_vec:
2230        return C_O1_I1(w, wr);
2231    case INDEX_op_abs_vec:
2232    case INDEX_op_neg_vec:
2233    case INDEX_op_not_vec:
2234    case INDEX_op_shli_vec:
2235    case INDEX_op_shri_vec:
2236    case INDEX_op_sari_vec:
2237        return C_O1_I1(w, w);
2238    case INDEX_op_dup2_vec:
2239    case INDEX_op_add_vec:
2240    case INDEX_op_mul_vec:
2241    case INDEX_op_smax_vec:
2242    case INDEX_op_smin_vec:
2243    case INDEX_op_ssadd_vec:
2244    case INDEX_op_sssub_vec:
2245    case INDEX_op_sub_vec:
2246    case INDEX_op_umax_vec:
2247    case INDEX_op_umin_vec:
2248    case INDEX_op_usadd_vec:
2249    case INDEX_op_ussub_vec:
2250    case INDEX_op_xor_vec:
2251    case INDEX_op_arm_sshl_vec:
2252    case INDEX_op_arm_ushl_vec:
2253        return C_O1_I2(w, w, w);
2254    case INDEX_op_arm_sli_vec:
2255        return C_O1_I2(w, 0, w);
2256    case INDEX_op_or_vec:
2257    case INDEX_op_andc_vec:
2258        return C_O1_I2(w, w, wO);
2259    case INDEX_op_and_vec:
2260    case INDEX_op_orc_vec:
2261        return C_O1_I2(w, w, wV);
2262    case INDEX_op_cmp_vec:
2263        return C_O1_I2(w, w, wZ);
2264    case INDEX_op_bitsel_vec:
2265        return C_O1_I3(w, w, w, w);
2266    default:
2267        return C_NotImplemented;
2268    }
2269}
2270
2271static void tcg_target_init(TCGContext *s)
2272{
2273    /*
2274     * Only probe for the platform and capabilities if we haven't already
2275     * determined maximum values at compile time.
2276     */
2277#if !defined(use_idiv_instructions) || !defined(use_neon_instructions)
2278    {
2279        unsigned long hwcap = qemu_getauxval(AT_HWCAP);
2280#ifndef use_idiv_instructions
2281        use_idiv_instructions = (hwcap & HWCAP_ARM_IDIVA) != 0;
2282#endif
2283#ifndef use_neon_instructions
2284        use_neon_instructions = (hwcap & HWCAP_ARM_NEON) != 0;
2285#endif
2286    }
2287#endif
2288
2289    if (__ARM_ARCH < 7) {
2290        const char *pl = (const char *)qemu_getauxval(AT_PLATFORM);
2291        if (pl != NULL && pl[0] == 'v' && pl[1] >= '4' && pl[1] <= '9') {
2292            arm_arch = pl[1] - '0';
2293        }
2294
2295        if (arm_arch < 6) {
2296            error_report("TCG: ARMv%d is unsupported; exiting", arm_arch);
2297            exit(EXIT_FAILURE);
2298        }
2299    }
2300
2301    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
2302
2303    tcg_target_call_clobber_regs = 0;
2304    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R0);
2305    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R1);
2306    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R2);
2307    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R3);
2308    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R12);
2309    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R14);
2310
2311    if (use_neon_instructions) {
2312        tcg_target_available_regs[TCG_TYPE_V64]  = ALL_VECTOR_REGS;
2313        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
2314
2315        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q0);
2316        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q1);
2317        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q2);
2318        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q3);
2319        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q8);
2320        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q9);
2321        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q10);
2322        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q11);
2323        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q12);
2324        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q13);
2325        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q14);
2326        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q15);
2327    }
2328
2329    s->reserved_regs = 0;
2330    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
2331    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP);
2332    tcg_regset_set_reg(s->reserved_regs, TCG_REG_PC);
2333    tcg_regset_set_reg(s->reserved_regs, TCG_VEC_TMP);
2334}
2335
2336static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg arg,
2337                       TCGReg arg1, intptr_t arg2)
2338{
2339    switch (type) {
2340    case TCG_TYPE_I32:
2341        tcg_out_ld32u(s, COND_AL, arg, arg1, arg2);
2342        return;
2343    case TCG_TYPE_V64:
2344        /* regs 1; size 8; align 8 */
2345        tcg_out_vldst(s, INSN_VLD1 | 0x7d0, arg, arg1, arg2);
2346        return;
2347    case TCG_TYPE_V128:
2348        /*
2349         * We have only 8-byte alignment for the stack per the ABI.
2350         * Rather than dynamically re-align the stack, it's easier
2351         * to simply not request alignment beyond that.  So:
2352         * regs 2; size 8; align 8
2353         */
2354        tcg_out_vldst(s, INSN_VLD1 | 0xad0, arg, arg1, arg2);
2355        return;
2356    default:
2357        g_assert_not_reached();
2358    }
2359}
2360
2361static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
2362                       TCGReg arg1, intptr_t arg2)
2363{
2364    switch (type) {
2365    case TCG_TYPE_I32:
2366        tcg_out_st32(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_VST1 | 0x7d0, arg, arg1, arg2);
2371        return;
2372    case TCG_TYPE_V128:
2373        /* See tcg_out_ld re alignment: regs 2; size 8; align 8 */
2374        tcg_out_vldst(s, INSN_VST1 | 0xad0, arg, arg1, arg2);
2375        return;
2376    default:
2377        g_assert_not_reached();
2378    }
2379}
2380
2381static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
2382                        TCGReg base, intptr_t ofs)
2383{
2384    return false;
2385}
2386
2387static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
2388{
2389    if (ret == arg) {
2390        return true;
2391    }
2392    switch (type) {
2393    case TCG_TYPE_I32:
2394        if (ret < TCG_REG_Q0 && arg < TCG_REG_Q0) {
2395            tcg_out_mov_reg(s, COND_AL, ret, arg);
2396            return true;
2397        }
2398        return false;
2399
2400    case TCG_TYPE_V64:
2401    case TCG_TYPE_V128:
2402        /* "VMOV D,N" is an alias for "VORR D,N,N". */
2403        tcg_out_vreg3(s, INSN_VORR, type - TCG_TYPE_V64, 0, ret, arg, arg);
2404        return true;
2405
2406    default:
2407        g_assert_not_reached();
2408    }
2409}
2410
2411static void tcg_out_movi(TCGContext *s, TCGType type,
2412                         TCGReg ret, tcg_target_long arg)
2413{
2414    tcg_debug_assert(type == TCG_TYPE_I32);
2415    tcg_debug_assert(ret < TCG_REG_Q0);
2416    tcg_out_movi32(s, COND_AL, ret, arg);
2417}
2418
2419static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
2420{
2421    return false;
2422}
2423
2424static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
2425                             tcg_target_long imm)
2426{
2427    int enc, opc = ARITH_ADD;
2428
2429    /* All of the easiest immediates to encode are positive. */
2430    if (imm < 0) {
2431        imm = -imm;
2432        opc = ARITH_SUB;
2433    }
2434    enc = encode_imm(imm);
2435    if (enc >= 0) {
2436        tcg_out_dat_imm(s, COND_AL, opc, rd, rs, enc);
2437    } else {
2438        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, imm);
2439        tcg_out_dat_reg(s, COND_AL, opc, rd, rs,
2440                        TCG_REG_TMP, SHIFT_IMM_LSL(0));
2441    }
2442}
2443
2444/* Type is always V128, with I64 elements.  */
2445static void tcg_out_dup2_vec(TCGContext *s, TCGReg rd, TCGReg rl, TCGReg rh)
2446{
2447    /* Move high element into place first. */
2448    /* VMOV Dd+1, Ds */
2449    tcg_out_vreg3(s, INSN_VORR | (1 << 12), 0, 0, rd, rh, rh);
2450    /* Move low element into place; tcg_out_mov will check for nop. */
2451    tcg_out_mov(s, TCG_TYPE_V64, rd, rl);
2452}
2453
2454static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
2455                            TCGReg rd, TCGReg rs)
2456{
2457    int q = type - TCG_TYPE_V64;
2458
2459    if (vece == MO_64) {
2460        if (type == TCG_TYPE_V128) {
2461            tcg_out_dup2_vec(s, rd, rs, rs);
2462        } else {
2463            tcg_out_mov(s, TCG_TYPE_V64, rd, rs);
2464        }
2465    } else if (rs < TCG_REG_Q0) {
2466        int b = (vece == MO_8);
2467        int e = (vece == MO_16);
2468        tcg_out32(s, INSN_VDUP_G | (b << 22) | (q << 21) | (e << 5) |
2469                  encode_vn(rd) | (rs << 12));
2470    } else {
2471        int imm4 = 1 << vece;
2472        tcg_out32(s, INSN_VDUP_S | (imm4 << 16) | (q << 6) |
2473                  encode_vd(rd) | encode_vm(rs));
2474    }
2475    return true;
2476}
2477
2478static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
2479                             TCGReg rd, TCGReg base, intptr_t offset)
2480{
2481    if (vece == MO_64) {
2482        tcg_out_ld(s, TCG_TYPE_V64, rd, base, offset);
2483        if (type == TCG_TYPE_V128) {
2484            tcg_out_dup2_vec(s, rd, rd, rd);
2485        }
2486    } else {
2487        int q = type - TCG_TYPE_V64;
2488        tcg_out_vldst(s, INSN_VLD1R | (vece << 6) | (q << 5),
2489                      rd, base, offset);
2490    }
2491    return true;
2492}
2493
2494static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
2495                             TCGReg rd, int64_t v64)
2496{
2497    int q = type - TCG_TYPE_V64;
2498    int cmode, imm8, i;
2499
2500    /* Test all bytes equal first.  */
2501    if (vece == MO_8) {
2502        tcg_out_vmovi(s, rd, q, 0, 0xe, v64);
2503        return;
2504    }
2505
2506    /*
2507     * Test all bytes 0x00 or 0xff second.  This can match cases that
2508     * might otherwise take 2 or 3 insns for MO_16 or MO_32 below.
2509     */
2510    for (i = imm8 = 0; i < 8; i++) {
2511        uint8_t byte = v64 >> (i * 8);
2512        if (byte == 0xff) {
2513            imm8 |= 1 << i;
2514        } else if (byte != 0) {
2515            goto fail_bytes;
2516        }
2517    }
2518    tcg_out_vmovi(s, rd, q, 1, 0xe, imm8);
2519    return;
2520 fail_bytes:
2521
2522    /*
2523     * Tests for various replications.  For each element width, if we
2524     * cannot find an expansion there's no point checking a larger
2525     * width because we already know by replication it cannot match.
2526     */
2527    if (vece == MO_16) {
2528        uint16_t v16 = v64;
2529
2530        if (is_shimm16(v16, &cmode, &imm8)) {
2531            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2532            return;
2533        }
2534        if (is_shimm16(~v16, &cmode, &imm8)) {
2535            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2536            return;
2537        }
2538
2539        /*
2540         * Otherwise, all remaining constants can be loaded in two insns:
2541         * rd = v16 & 0xff, rd |= v16 & 0xff00.
2542         */
2543        tcg_out_vmovi(s, rd, q, 0, 0x8, v16 & 0xff);
2544        tcg_out_vmovi(s, rd, q, 0, 0xb, v16 >> 8);   /* VORRI */
2545        return;
2546    }
2547
2548    if (vece == MO_32) {
2549        uint32_t v32 = v64;
2550
2551        if (is_shimm32(v32, &cmode, &imm8) ||
2552            is_soimm32(v32, &cmode, &imm8)) {
2553            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2554            return;
2555        }
2556        if (is_shimm32(~v32, &cmode, &imm8) ||
2557            is_soimm32(~v32, &cmode, &imm8)) {
2558            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2559            return;
2560        }
2561
2562        /*
2563         * Restrict the set of constants to those we can load with
2564         * two instructions.  Others we load from the pool.
2565         */
2566        i = is_shimm32_pair(v32, &cmode, &imm8);
2567        if (i) {
2568            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2569            tcg_out_vmovi(s, rd, q, 0, i | 1, extract32(v32, i * 4, 8));
2570            return;
2571        }
2572        i = is_shimm32_pair(~v32, &cmode, &imm8);
2573        if (i) {
2574            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2575            tcg_out_vmovi(s, rd, q, 1, i | 1, extract32(~v32, i * 4, 8));
2576            return;
2577        }
2578    }
2579
2580    /*
2581     * As a last resort, load from the constant pool.
2582     */
2583    if (!q || vece == MO_64) {
2584        new_pool_l2(s, R_ARM_PC11, s->code_ptr, 0, v64, v64 >> 32);
2585        /* VLDR Dd, [pc + offset] */
2586        tcg_out32(s, INSN_VLDR_D | encode_vd(rd) | (0xf << 16));
2587        if (q) {
2588            tcg_out_dup2_vec(s, rd, rd, rd);
2589        }
2590    } else {
2591        new_pool_label(s, (uint32_t)v64, R_ARM_PC8, s->code_ptr, 0);
2592        /* add tmp, pc, offset */
2593        tcg_out_dat_imm(s, COND_AL, ARITH_ADD, TCG_REG_TMP, TCG_REG_PC, 0);
2594        tcg_out_dupm_vec(s, type, MO_32, rd, TCG_REG_TMP, 0);
2595    }
2596}
2597
2598static const ARMInsn vec_cmp_insn[16] = {
2599    [TCG_COND_EQ] = INSN_VCEQ,
2600    [TCG_COND_GT] = INSN_VCGT,
2601    [TCG_COND_GE] = INSN_VCGE,
2602    [TCG_COND_GTU] = INSN_VCGT_U,
2603    [TCG_COND_GEU] = INSN_VCGE_U,
2604};
2605
2606static const ARMInsn vec_cmp0_insn[16] = {
2607    [TCG_COND_EQ] = INSN_VCEQ0,
2608    [TCG_COND_GT] = INSN_VCGT0,
2609    [TCG_COND_GE] = INSN_VCGE0,
2610    [TCG_COND_LT] = INSN_VCLT0,
2611    [TCG_COND_LE] = INSN_VCLE0,
2612};
2613
2614static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
2615                           unsigned vecl, unsigned vece,
2616                           const TCGArg args[TCG_MAX_OP_ARGS],
2617                           const int const_args[TCG_MAX_OP_ARGS])
2618{
2619    TCGType type = vecl + TCG_TYPE_V64;
2620    unsigned q = vecl;
2621    TCGArg a0, a1, a2, a3;
2622    int cmode, imm8;
2623
2624    a0 = args[0];
2625    a1 = args[1];
2626    a2 = args[2];
2627
2628    switch (opc) {
2629    case INDEX_op_ld_vec:
2630        tcg_out_ld(s, type, a0, a1, a2);
2631        return;
2632    case INDEX_op_st_vec:
2633        tcg_out_st(s, type, a0, a1, a2);
2634        return;
2635    case INDEX_op_dupm_vec:
2636        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
2637        return;
2638    case INDEX_op_dup2_vec:
2639        tcg_out_dup2_vec(s, a0, a1, a2);
2640        return;
2641    case INDEX_op_abs_vec:
2642        tcg_out_vreg2(s, INSN_VABS, q, vece, a0, a1);
2643        return;
2644    case INDEX_op_neg_vec:
2645        tcg_out_vreg2(s, INSN_VNEG, q, vece, a0, a1);
2646        return;
2647    case INDEX_op_not_vec:
2648        tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a1);
2649        return;
2650    case INDEX_op_add_vec:
2651        tcg_out_vreg3(s, INSN_VADD, q, vece, a0, a1, a2);
2652        return;
2653    case INDEX_op_mul_vec:
2654        tcg_out_vreg3(s, INSN_VMUL, q, vece, a0, a1, a2);
2655        return;
2656    case INDEX_op_smax_vec:
2657        tcg_out_vreg3(s, INSN_VMAX, q, vece, a0, a1, a2);
2658        return;
2659    case INDEX_op_smin_vec:
2660        tcg_out_vreg3(s, INSN_VMIN, q, vece, a0, a1, a2);
2661        return;
2662    case INDEX_op_sub_vec:
2663        tcg_out_vreg3(s, INSN_VSUB, q, vece, a0, a1, a2);
2664        return;
2665    case INDEX_op_ssadd_vec:
2666        tcg_out_vreg3(s, INSN_VQADD, q, vece, a0, a1, a2);
2667        return;
2668    case INDEX_op_sssub_vec:
2669        tcg_out_vreg3(s, INSN_VQSUB, q, vece, a0, a1, a2);
2670        return;
2671    case INDEX_op_umax_vec:
2672        tcg_out_vreg3(s, INSN_VMAX_U, q, vece, a0, a1, a2);
2673        return;
2674    case INDEX_op_umin_vec:
2675        tcg_out_vreg3(s, INSN_VMIN_U, q, vece, a0, a1, a2);
2676        return;
2677    case INDEX_op_usadd_vec:
2678        tcg_out_vreg3(s, INSN_VQADD_U, q, vece, a0, a1, a2);
2679        return;
2680    case INDEX_op_ussub_vec:
2681        tcg_out_vreg3(s, INSN_VQSUB_U, q, vece, a0, a1, a2);
2682        return;
2683    case INDEX_op_xor_vec:
2684        tcg_out_vreg3(s, INSN_VEOR, q, 0, a0, a1, a2);
2685        return;
2686    case INDEX_op_arm_sshl_vec:
2687        /*
2688         * Note that Vm is the data and Vn is the shift count,
2689         * therefore the arguments appear reversed.
2690         */
2691        tcg_out_vreg3(s, INSN_VSHL_S, q, vece, a0, a2, a1);
2692        return;
2693    case INDEX_op_arm_ushl_vec:
2694        /* See above. */
2695        tcg_out_vreg3(s, INSN_VSHL_U, q, vece, a0, a2, a1);
2696        return;
2697    case INDEX_op_shli_vec:
2698        tcg_out_vshifti(s, INSN_VSHLI, q, a0, a1, a2 + (8 << vece));
2699        return;
2700    case INDEX_op_shri_vec:
2701        tcg_out_vshifti(s, INSN_VSHRI, q, a0, a1, (16 << vece) - a2);
2702        return;
2703    case INDEX_op_sari_vec:
2704        tcg_out_vshifti(s, INSN_VSARI, q, a0, a1, (16 << vece) - a2);
2705        return;
2706    case INDEX_op_arm_sli_vec:
2707        tcg_out_vshifti(s, INSN_VSLI, q, a0, a2, args[3] + (8 << vece));
2708        return;
2709
2710    case INDEX_op_andc_vec:
2711        if (!const_args[2]) {
2712            tcg_out_vreg3(s, INSN_VBIC, q, 0, a0, a1, a2);
2713            return;
2714        }
2715        a2 = ~a2;
2716        /* fall through */
2717    case INDEX_op_and_vec:
2718        if (const_args[2]) {
2719            is_shimm1632(~a2, &cmode, &imm8);
2720            if (a0 == a1) {
2721                tcg_out_vmovi(s, a0, q, 1, cmode | 1, imm8); /* VBICI */
2722                return;
2723            }
2724            tcg_out_vmovi(s, a0, q, 1, cmode, imm8); /* VMVNI */
2725            a2 = a0;
2726        }
2727        tcg_out_vreg3(s, INSN_VAND, q, 0, a0, a1, a2);
2728        return;
2729
2730    case INDEX_op_orc_vec:
2731        if (!const_args[2]) {
2732            tcg_out_vreg3(s, INSN_VORN, q, 0, a0, a1, a2);
2733            return;
2734        }
2735        a2 = ~a2;
2736        /* fall through */
2737    case INDEX_op_or_vec:
2738        if (const_args[2]) {
2739            is_shimm1632(a2, &cmode, &imm8);
2740            if (a0 == a1) {
2741                tcg_out_vmovi(s, a0, q, 0, cmode | 1, imm8); /* VORRI */
2742                return;
2743            }
2744            tcg_out_vmovi(s, a0, q, 0, cmode, imm8); /* VMOVI */
2745            a2 = a0;
2746        }
2747        tcg_out_vreg3(s, INSN_VORR, q, 0, a0, a1, a2);
2748        return;
2749
2750    case INDEX_op_cmp_vec:
2751        {
2752            TCGCond cond = args[3];
2753            ARMInsn insn;
2754
2755            switch (cond) {
2756            case TCG_COND_NE:
2757                if (const_args[2]) {
2758                    tcg_out_vreg3(s, INSN_VTST, q, vece, a0, a1, a1);
2759                } else {
2760                    tcg_out_vreg3(s, INSN_VCEQ, q, vece, a0, a1, a2);
2761                    tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a0);
2762                }
2763                break;
2764
2765            case TCG_COND_TSTNE:
2766            case TCG_COND_TSTEQ:
2767                if (const_args[2]) {
2768                    /* (x & 0) == 0 */
2769                    tcg_out_dupi_vec(s, type, MO_8, a0,
2770                                     -(cond == TCG_COND_TSTEQ));
2771                    break;
2772                }
2773                tcg_out_vreg3(s, INSN_VTST, q, vece, a0, a1, a2);
2774                if (cond == TCG_COND_TSTEQ) {
2775                    tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a0);
2776                }
2777                break;
2778
2779            default:
2780                if (const_args[2]) {
2781                    insn = vec_cmp0_insn[cond];
2782                    if (insn) {
2783                        tcg_out_vreg2(s, insn, q, vece, a0, a1);
2784                        return;
2785                    }
2786                    tcg_out_dupi_vec(s, type, MO_8, TCG_VEC_TMP, 0);
2787                    a2 = TCG_VEC_TMP;
2788                }
2789                insn = vec_cmp_insn[cond];
2790                if (insn == 0) {
2791                    TCGArg t;
2792                    t = a1, a1 = a2, a2 = t;
2793                    cond = tcg_swap_cond(cond);
2794                    insn = vec_cmp_insn[cond];
2795                    tcg_debug_assert(insn != 0);
2796                }
2797                tcg_out_vreg3(s, insn, q, vece, a0, a1, a2);
2798                break;
2799            }
2800        }
2801        return;
2802
2803    case INDEX_op_bitsel_vec:
2804        a3 = args[3];
2805        if (a0 == a3) {
2806            tcg_out_vreg3(s, INSN_VBIT, q, 0, a0, a2, a1);
2807        } else if (a0 == a2) {
2808            tcg_out_vreg3(s, INSN_VBIF, q, 0, a0, a3, a1);
2809        } else {
2810            tcg_out_mov(s, type, a0, a1);
2811            tcg_out_vreg3(s, INSN_VBSL, q, 0, a0, a2, a3);
2812        }
2813        return;
2814
2815    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
2816    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
2817    default:
2818        g_assert_not_reached();
2819    }
2820}
2821
2822int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
2823{
2824    switch (opc) {
2825    case INDEX_op_add_vec:
2826    case INDEX_op_sub_vec:
2827    case INDEX_op_and_vec:
2828    case INDEX_op_andc_vec:
2829    case INDEX_op_or_vec:
2830    case INDEX_op_orc_vec:
2831    case INDEX_op_xor_vec:
2832    case INDEX_op_not_vec:
2833    case INDEX_op_shli_vec:
2834    case INDEX_op_shri_vec:
2835    case INDEX_op_sari_vec:
2836    case INDEX_op_ssadd_vec:
2837    case INDEX_op_sssub_vec:
2838    case INDEX_op_usadd_vec:
2839    case INDEX_op_ussub_vec:
2840    case INDEX_op_bitsel_vec:
2841        return 1;
2842    case INDEX_op_abs_vec:
2843    case INDEX_op_cmp_vec:
2844    case INDEX_op_mul_vec:
2845    case INDEX_op_neg_vec:
2846    case INDEX_op_smax_vec:
2847    case INDEX_op_smin_vec:
2848    case INDEX_op_umax_vec:
2849    case INDEX_op_umin_vec:
2850        return vece < MO_64;
2851    case INDEX_op_shlv_vec:
2852    case INDEX_op_shrv_vec:
2853    case INDEX_op_sarv_vec:
2854    case INDEX_op_rotli_vec:
2855    case INDEX_op_rotlv_vec:
2856    case INDEX_op_rotrv_vec:
2857        return -1;
2858    default:
2859        return 0;
2860    }
2861}
2862
2863void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
2864                       TCGArg a0, ...)
2865{
2866    va_list va;
2867    TCGv_vec v0, v1, v2, t1, t2, c1;
2868    TCGArg a2;
2869
2870    va_start(va, a0);
2871    v0 = temp_tcgv_vec(arg_temp(a0));
2872    v1 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
2873    a2 = va_arg(va, TCGArg);
2874    va_end(va);
2875
2876    switch (opc) {
2877    case INDEX_op_shlv_vec:
2878        /*
2879         * Merely propagate shlv_vec to arm_ushl_vec.
2880         * In this way we don't set TCG_TARGET_HAS_shv_vec
2881         * because everything is done via expansion.
2882         */
2883        v2 = temp_tcgv_vec(arg_temp(a2));
2884        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
2885                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
2886        break;
2887
2888    case INDEX_op_shrv_vec:
2889    case INDEX_op_sarv_vec:
2890        /* Right shifts are negative left shifts for NEON.  */
2891        v2 = temp_tcgv_vec(arg_temp(a2));
2892        t1 = tcg_temp_new_vec(type);
2893        tcg_gen_neg_vec(vece, t1, v2);
2894        if (opc == INDEX_op_shrv_vec) {
2895            opc = INDEX_op_arm_ushl_vec;
2896        } else {
2897            opc = INDEX_op_arm_sshl_vec;
2898        }
2899        vec_gen_3(opc, type, vece, tcgv_vec_arg(v0),
2900                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
2901        tcg_temp_free_vec(t1);
2902        break;
2903
2904    case INDEX_op_rotli_vec:
2905        t1 = tcg_temp_new_vec(type);
2906        tcg_gen_shri_vec(vece, t1, v1, -a2 & ((8 << vece) - 1));
2907        vec_gen_4(INDEX_op_arm_sli_vec, type, vece,
2908                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(v1), a2);
2909        tcg_temp_free_vec(t1);
2910        break;
2911
2912    case INDEX_op_rotlv_vec:
2913        v2 = temp_tcgv_vec(arg_temp(a2));
2914        t1 = tcg_temp_new_vec(type);
2915        c1 = tcg_constant_vec(type, vece, 8 << vece);
2916        tcg_gen_sub_vec(vece, t1, v2, c1);
2917        /* Right shifts are negative left shifts for NEON.  */
2918        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
2919                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
2920        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
2921                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
2922        tcg_gen_or_vec(vece, v0, v0, t1);
2923        tcg_temp_free_vec(t1);
2924        break;
2925
2926    case INDEX_op_rotrv_vec:
2927        v2 = temp_tcgv_vec(arg_temp(a2));
2928        t1 = tcg_temp_new_vec(type);
2929        t2 = tcg_temp_new_vec(type);
2930        c1 = tcg_constant_vec(type, vece, 8 << vece);
2931        tcg_gen_neg_vec(vece, t1, v2);
2932        tcg_gen_sub_vec(vece, t2, c1, v2);
2933        /* Right shifts are negative left shifts for NEON.  */
2934        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
2935                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
2936        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t2),
2937                  tcgv_vec_arg(v1), tcgv_vec_arg(t2));
2938        tcg_gen_or_vec(vece, v0, t1, t2);
2939        tcg_temp_free_vec(t1);
2940        tcg_temp_free_vec(t2);
2941        break;
2942
2943    default:
2944        g_assert_not_reached();
2945    }
2946}
2947
2948static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
2949{
2950    int i;
2951    for (i = 0; i < count; ++i) {
2952        p[i] = INSN_NOP;
2953    }
2954}
2955
2956/* Compute frame size via macros, to share between tcg_target_qemu_prologue
2957   and tcg_register_jit.  */
2958
2959#define PUSH_SIZE  ((11 - 4 + 1 + 1) * sizeof(tcg_target_long))
2960
2961#define FRAME_SIZE \
2962    ((PUSH_SIZE \
2963      + TCG_STATIC_CALL_ARGS_SIZE \
2964      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
2965      + TCG_TARGET_STACK_ALIGN - 1) \
2966     & -TCG_TARGET_STACK_ALIGN)
2967
2968#define STACK_ADDEND  (FRAME_SIZE - PUSH_SIZE)
2969
2970static void tcg_target_qemu_prologue(TCGContext *s)
2971{
2972    /* Calling convention requires us to save r4-r11 and lr.  */
2973    /* stmdb sp!, { r4 - r11, lr } */
2974    tcg_out_ldstm(s, COND_AL, INSN_STMDB, TCG_REG_CALL_STACK,
2975                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
2976                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
2977                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_R14));
2978
2979    /* Reserve callee argument and tcg temp space.  */
2980    tcg_out_dat_rI(s, COND_AL, ARITH_SUB, TCG_REG_CALL_STACK,
2981                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
2982    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
2983                  CPU_TEMP_BUF_NLONGS * sizeof(long));
2984
2985    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
2986
2987    if (!tcg_use_softmmu && guest_base) {
2988        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_GUEST_BASE, guest_base);
2989        tcg_regset_set_reg(s->reserved_regs, TCG_REG_GUEST_BASE);
2990    }
2991
2992    tcg_out_b_reg(s, COND_AL, tcg_target_call_iarg_regs[1]);
2993
2994    /*
2995     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
2996     * and fall through to the rest of the epilogue.
2997     */
2998    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
2999    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, 0);
3000    tcg_out_epilogue(s);
3001}
3002
3003static void tcg_out_epilogue(TCGContext *s)
3004{
3005    /* Release local stack frame.  */
3006    tcg_out_dat_rI(s, COND_AL, ARITH_ADD, TCG_REG_CALL_STACK,
3007                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
3008
3009    /* ldmia sp!, { r4 - r11, pc } */
3010    tcg_out_ldstm(s, COND_AL, INSN_LDMIA, TCG_REG_CALL_STACK,
3011                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
3012                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
3013                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_PC));
3014}
3015
3016static void tcg_out_tb_start(TCGContext *s)
3017{
3018    /* nothing to do */
3019}
3020
3021typedef struct {
3022    DebugFrameHeader h;
3023    uint8_t fde_def_cfa[4];
3024    uint8_t fde_reg_ofs[18];
3025} DebugFrame;
3026
3027#define ELF_HOST_MACHINE EM_ARM
3028
3029/* We're expecting a 2 byte uleb128 encoded value.  */
3030QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
3031
3032static const DebugFrame debug_frame = {
3033    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
3034    .h.cie.id = -1,
3035    .h.cie.version = 1,
3036    .h.cie.code_align = 1,
3037    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
3038    .h.cie.return_column = 14,
3039
3040    /* Total FDE size does not include the "len" member.  */
3041    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
3042
3043    .fde_def_cfa = {
3044        12, 13,                         /* DW_CFA_def_cfa sp, ... */
3045        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
3046        (FRAME_SIZE >> 7)
3047    },
3048    .fde_reg_ofs = {
3049        /* The following must match the stmdb in the prologue.  */
3050        0x8e, 1,                        /* DW_CFA_offset, lr, -4 */
3051        0x8b, 2,                        /* DW_CFA_offset, r11, -8 */
3052        0x8a, 3,                        /* DW_CFA_offset, r10, -12 */
3053        0x89, 4,                        /* DW_CFA_offset, r9, -16 */
3054        0x88, 5,                        /* DW_CFA_offset, r8, -20 */
3055        0x87, 6,                        /* DW_CFA_offset, r7, -24 */
3056        0x86, 7,                        /* DW_CFA_offset, r6, -28 */
3057        0x85, 8,                        /* DW_CFA_offset, r5, -32 */
3058        0x84, 9,                        /* DW_CFA_offset, r4, -36 */
3059    }
3060};
3061
3062void tcg_register_jit(const void *buf, size_t buf_size)
3063{
3064    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
3065}
3066