xref: /openbmc/qemu/tcg/arm/tcg-target.c.inc (revision b2c514f9d5cab89814dc8a6b7c98c653ca8523d3)
1/*
2 * Tiny Code Generator for QEMU
3 *
4 * Copyright (c) 2008 Andrzej Zaborowski
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to deal
8 * in the Software without restriction, including without limitation the rights
9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10 * copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22 * THE SOFTWARE.
23 */
24
25#include "elf.h"
26
27int arm_arch = __ARM_ARCH;
28
29#ifndef use_idiv_instructions
30bool use_idiv_instructions;
31#endif
32#ifndef use_neon_instructions
33bool use_neon_instructions;
34#endif
35
36/* Used for function call generation. */
37#define TCG_TARGET_STACK_ALIGN          8
38#define TCG_TARGET_CALL_STACK_OFFSET    0
39#define TCG_TARGET_CALL_ARG_I32         TCG_CALL_ARG_NORMAL
40#define TCG_TARGET_CALL_ARG_I64         TCG_CALL_ARG_EVEN
41#define TCG_TARGET_CALL_ARG_I128        TCG_CALL_ARG_EVEN
42#define TCG_TARGET_CALL_RET_I128        TCG_CALL_RET_BY_REF
43
44#ifdef CONFIG_DEBUG_TCG
45static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
46    "%r0",  "%r1",  "%r2",  "%r3",  "%r4",  "%r5",  "%r6",  "%r7",
47    "%r8",  "%r9",  "%r10", "%r11", "%r12", "%sp",  "%r14", "%pc",
48    "%q0",  "%q1",  "%q2",  "%q3",  "%q4",  "%q5",  "%q6",  "%q7",
49    "%q8",  "%q9",  "%q10", "%q11", "%q12", "%q13", "%q14", "%q15",
50};
51#endif
52
53static const int tcg_target_reg_alloc_order[] = {
54    TCG_REG_R4,
55    TCG_REG_R5,
56    TCG_REG_R6,
57    TCG_REG_R7,
58    TCG_REG_R8,
59    TCG_REG_R9,
60    TCG_REG_R10,
61    TCG_REG_R11,
62    TCG_REG_R13,
63    TCG_REG_R0,
64    TCG_REG_R1,
65    TCG_REG_R2,
66    TCG_REG_R3,
67    TCG_REG_R12,
68    TCG_REG_R14,
69
70    TCG_REG_Q0,
71    TCG_REG_Q1,
72    TCG_REG_Q2,
73    TCG_REG_Q3,
74    /* Q4 - Q7 are call-saved, and skipped. */
75    TCG_REG_Q8,
76    TCG_REG_Q9,
77    TCG_REG_Q10,
78    TCG_REG_Q11,
79    TCG_REG_Q12,
80    TCG_REG_Q13,
81    TCG_REG_Q14,
82    TCG_REG_Q15,
83};
84
85static const int tcg_target_call_iarg_regs[4] = {
86    TCG_REG_R0, TCG_REG_R1, TCG_REG_R2, TCG_REG_R3
87};
88
89static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
90{
91    tcg_debug_assert(kind == TCG_CALL_RET_NORMAL);
92    tcg_debug_assert(slot >= 0 && slot <= 3);
93    return TCG_REG_R0 + slot;
94}
95
96#define TCG_REG_TMP  TCG_REG_R12
97#define TCG_VEC_TMP  TCG_REG_Q15
98#define TCG_REG_GUEST_BASE  TCG_REG_R11
99
100typedef enum {
101    COND_EQ = 0x0,
102    COND_NE = 0x1,
103    COND_CS = 0x2,	/* Unsigned greater or equal */
104    COND_CC = 0x3,	/* Unsigned less than */
105    COND_MI = 0x4,	/* Negative */
106    COND_PL = 0x5,	/* Zero or greater */
107    COND_VS = 0x6,	/* Overflow */
108    COND_VC = 0x7,	/* No overflow */
109    COND_HI = 0x8,	/* Unsigned greater than */
110    COND_LS = 0x9,	/* Unsigned less or equal */
111    COND_GE = 0xa,
112    COND_LT = 0xb,
113    COND_GT = 0xc,
114    COND_LE = 0xd,
115    COND_AL = 0xe,
116} ARMCond;
117
118#define TO_CPSR (1 << 20)
119
120#define SHIFT_IMM_LSL(im)	(((im) << 7) | 0x00)
121#define SHIFT_IMM_LSR(im)	(((im) << 7) | 0x20)
122#define SHIFT_IMM_ASR(im)	(((im) << 7) | 0x40)
123#define SHIFT_IMM_ROR(im)	(((im) << 7) | 0x60)
124#define SHIFT_REG_LSL(rs)	(((rs) << 8) | 0x10)
125#define SHIFT_REG_LSR(rs)	(((rs) << 8) | 0x30)
126#define SHIFT_REG_ASR(rs)	(((rs) << 8) | 0x50)
127#define SHIFT_REG_ROR(rs)	(((rs) << 8) | 0x70)
128
129typedef enum {
130    ARITH_AND = 0x0 << 21,
131    ARITH_EOR = 0x1 << 21,
132    ARITH_SUB = 0x2 << 21,
133    ARITH_RSB = 0x3 << 21,
134    ARITH_ADD = 0x4 << 21,
135    ARITH_ADC = 0x5 << 21,
136    ARITH_SBC = 0x6 << 21,
137    ARITH_RSC = 0x7 << 21,
138    ARITH_TST = 0x8 << 21 | TO_CPSR,
139    ARITH_CMP = 0xa << 21 | TO_CPSR,
140    ARITH_CMN = 0xb << 21 | TO_CPSR,
141    ARITH_ORR = 0xc << 21,
142    ARITH_MOV = 0xd << 21,
143    ARITH_BIC = 0xe << 21,
144    ARITH_MVN = 0xf << 21,
145
146    INSN_B         = 0x0a000000,
147
148    INSN_CLZ       = 0x016f0f10,
149    INSN_RBIT      = 0x06ff0f30,
150
151    INSN_LDMIA     = 0x08b00000,
152    INSN_STMDB     = 0x09200000,
153
154    INSN_LDR_IMM   = 0x04100000,
155    INSN_LDR_REG   = 0x06100000,
156    INSN_STR_IMM   = 0x04000000,
157    INSN_STR_REG   = 0x06000000,
158
159    INSN_LDRH_IMM  = 0x005000b0,
160    INSN_LDRH_REG  = 0x001000b0,
161    INSN_LDRSH_IMM = 0x005000f0,
162    INSN_LDRSH_REG = 0x001000f0,
163    INSN_STRH_IMM  = 0x004000b0,
164    INSN_STRH_REG  = 0x000000b0,
165
166    INSN_LDRB_IMM  = 0x04500000,
167    INSN_LDRB_REG  = 0x06500000,
168    INSN_LDRSB_IMM = 0x005000d0,
169    INSN_LDRSB_REG = 0x001000d0,
170    INSN_STRB_IMM  = 0x04400000,
171    INSN_STRB_REG  = 0x06400000,
172
173    INSN_LDRD_IMM  = 0x004000d0,
174    INSN_LDRD_REG  = 0x000000d0,
175    INSN_STRD_IMM  = 0x004000f0,
176    INSN_STRD_REG  = 0x000000f0,
177
178    INSN_DMB_ISH   = 0xf57ff05b,
179    INSN_DMB_MCR   = 0xee070fba,
180
181    /* Architected nop introduced in v6k.  */
182    /* ??? This is an MSR (imm) 0,0,0 insn.  Anyone know if this
183       also Just So Happened to do nothing on pre-v6k so that we
184       don't need to conditionalize it?  */
185    INSN_NOP_v6k   = 0xe320f000,
186    /* Otherwise the assembler uses mov r0,r0 */
187    INSN_NOP_v4    = (COND_AL << 28) | ARITH_MOV,
188
189    INSN_VADD      = 0xf2000800,
190    INSN_VAND      = 0xf2000110,
191    INSN_VBIC      = 0xf2100110,
192    INSN_VEOR      = 0xf3000110,
193    INSN_VORN      = 0xf2300110,
194    INSN_VORR      = 0xf2200110,
195    INSN_VSUB      = 0xf3000800,
196    INSN_VMUL      = 0xf2000910,
197    INSN_VQADD     = 0xf2000010,
198    INSN_VQADD_U   = 0xf3000010,
199    INSN_VQSUB     = 0xf2000210,
200    INSN_VQSUB_U   = 0xf3000210,
201    INSN_VMAX      = 0xf2000600,
202    INSN_VMAX_U    = 0xf3000600,
203    INSN_VMIN      = 0xf2000610,
204    INSN_VMIN_U    = 0xf3000610,
205
206    INSN_VABS      = 0xf3b10300,
207    INSN_VMVN      = 0xf3b00580,
208    INSN_VNEG      = 0xf3b10380,
209
210    INSN_VCEQ0     = 0xf3b10100,
211    INSN_VCGT0     = 0xf3b10000,
212    INSN_VCGE0     = 0xf3b10080,
213    INSN_VCLE0     = 0xf3b10180,
214    INSN_VCLT0     = 0xf3b10200,
215
216    INSN_VCEQ      = 0xf3000810,
217    INSN_VCGE      = 0xf2000310,
218    INSN_VCGT      = 0xf2000300,
219    INSN_VCGE_U    = 0xf3000310,
220    INSN_VCGT_U    = 0xf3000300,
221
222    INSN_VSHLI     = 0xf2800510,  /* VSHL (immediate) */
223    INSN_VSARI     = 0xf2800010,  /* VSHR.S */
224    INSN_VSHRI     = 0xf3800010,  /* VSHR.U */
225    INSN_VSLI      = 0xf3800510,
226    INSN_VSHL_S    = 0xf2000400,  /* VSHL.S (register) */
227    INSN_VSHL_U    = 0xf3000400,  /* VSHL.U (register) */
228
229    INSN_VBSL      = 0xf3100110,
230    INSN_VBIT      = 0xf3200110,
231    INSN_VBIF      = 0xf3300110,
232
233    INSN_VTST      = 0xf2000810,
234
235    INSN_VDUP_G    = 0xee800b10,  /* VDUP (ARM core register) */
236    INSN_VDUP_S    = 0xf3b00c00,  /* VDUP (scalar) */
237    INSN_VLDR_D    = 0xed100b00,  /* VLDR.64 */
238    INSN_VLD1      = 0xf4200000,  /* VLD1 (multiple single elements) */
239    INSN_VLD1R     = 0xf4a00c00,  /* VLD1 (single element to all lanes) */
240    INSN_VST1      = 0xf4000000,  /* VST1 (multiple single elements) */
241    INSN_VMOVI     = 0xf2800010,  /* VMOV (immediate) */
242} ARMInsn;
243
244#define INSN_NOP   (use_armv7_instructions ? INSN_NOP_v6k : INSN_NOP_v4)
245
246static const uint8_t tcg_cond_to_arm_cond[] = {
247    [TCG_COND_EQ] = COND_EQ,
248    [TCG_COND_NE] = COND_NE,
249    [TCG_COND_LT] = COND_LT,
250    [TCG_COND_GE] = COND_GE,
251    [TCG_COND_LE] = COND_LE,
252    [TCG_COND_GT] = COND_GT,
253    /* unsigned */
254    [TCG_COND_LTU] = COND_CC,
255    [TCG_COND_GEU] = COND_CS,
256    [TCG_COND_LEU] = COND_LS,
257    [TCG_COND_GTU] = COND_HI,
258};
259
260static int encode_imm(uint32_t imm);
261
262/* TCG private relocation type: add with pc+imm8 */
263#define R_ARM_PC8  11
264
265/* TCG private relocation type: vldr with imm8 << 2 */
266#define R_ARM_PC11 12
267
268static bool reloc_pc24(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
269{
270    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
271    ptrdiff_t offset = (tcg_ptr_byte_diff(target, src_rx) - 8) >> 2;
272
273    if (offset == sextract32(offset, 0, 24)) {
274        *src_rw = deposit32(*src_rw, 0, 24, offset);
275        return true;
276    }
277    return false;
278}
279
280static bool reloc_pc13(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
281{
282    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
283    ptrdiff_t offset = tcg_ptr_byte_diff(target, src_rx) - 8;
284
285    if (offset >= -0xfff && offset <= 0xfff) {
286        tcg_insn_unit insn = *src_rw;
287        bool u = (offset >= 0);
288        if (!u) {
289            offset = -offset;
290        }
291        insn = deposit32(insn, 23, 1, u);
292        insn = deposit32(insn, 0, 12, offset);
293        *src_rw = insn;
294        return true;
295    }
296    return false;
297}
298
299static bool reloc_pc11(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
300{
301    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
302    ptrdiff_t offset = (tcg_ptr_byte_diff(target, src_rx) - 8) / 4;
303
304    if (offset >= -0xff && offset <= 0xff) {
305        tcg_insn_unit insn = *src_rw;
306        bool u = (offset >= 0);
307        if (!u) {
308            offset = -offset;
309        }
310        insn = deposit32(insn, 23, 1, u);
311        insn = deposit32(insn, 0, 8, offset);
312        *src_rw = insn;
313        return true;
314    }
315    return false;
316}
317
318static bool reloc_pc8(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
319{
320    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
321    ptrdiff_t offset = tcg_ptr_byte_diff(target, src_rx) - 8;
322    int imm12 = encode_imm(offset);
323
324    if (imm12 >= 0) {
325        *src_rw = deposit32(*src_rw, 0, 12, imm12);
326        return true;
327    }
328    return false;
329}
330
331static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
332                        intptr_t value, intptr_t addend)
333{
334    tcg_debug_assert(addend == 0);
335    switch (type) {
336    case R_ARM_PC24:
337        return reloc_pc24(code_ptr, (const tcg_insn_unit *)value);
338    case R_ARM_PC13:
339        return reloc_pc13(code_ptr, (const tcg_insn_unit *)value);
340    case R_ARM_PC11:
341        return reloc_pc11(code_ptr, (const tcg_insn_unit *)value);
342    case R_ARM_PC8:
343        return reloc_pc8(code_ptr, (const tcg_insn_unit *)value);
344    default:
345        g_assert_not_reached();
346    }
347}
348
349#define TCG_CT_CONST_ARM  0x100
350#define TCG_CT_CONST_INV  0x200
351#define TCG_CT_CONST_NEG  0x400
352#define TCG_CT_CONST_ZERO 0x800
353#define TCG_CT_CONST_ORRI 0x1000
354#define TCG_CT_CONST_ANDI 0x2000
355
356#define ALL_GENERAL_REGS  0xffffu
357#define ALL_VECTOR_REGS   0xffff0000u
358
359/*
360 * r0-r3 will be overwritten when reading the tlb entry (system-mode only);
361 * r14 will be overwritten by the BLNE branching to the slow path.
362 */
363#define ALL_QLDST_REGS \
364    (ALL_GENERAL_REGS & ~((tcg_use_softmmu ? 0xf : 0) | (1 << TCG_REG_R14)))
365
366/*
367 * ARM immediates for ALU instructions are made of an unsigned 8-bit
368 * right-rotated by an even amount between 0 and 30.
369 *
370 * Return < 0 if @imm cannot be encoded, else the entire imm12 field.
371 */
372static int encode_imm(uint32_t imm)
373{
374    uint32_t rot, imm8;
375
376    /* Simple case, no rotation required. */
377    if ((imm & ~0xff) == 0) {
378        return imm;
379    }
380
381    /* Next, try a simple even shift.  */
382    rot = ctz32(imm) & ~1;
383    imm8 = imm >> rot;
384    rot = 32 - rot;
385    if ((imm8 & ~0xff) == 0) {
386        goto found;
387    }
388
389    /*
390     * Finally, try harder with rotations.
391     * The ctz test above will have taken care of rotates >= 8.
392     */
393    for (rot = 2; rot < 8; rot += 2) {
394        imm8 = rol32(imm, rot);
395        if ((imm8 & ~0xff) == 0) {
396            goto found;
397        }
398    }
399    /* Fail: imm cannot be encoded. */
400    return -1;
401
402 found:
403    /* Note that rot is even, and we discard bit 0 by shifting by 7. */
404    return rot << 7 | imm8;
405}
406
407static int encode_imm_nofail(uint32_t imm)
408{
409    int ret = encode_imm(imm);
410    tcg_debug_assert(ret >= 0);
411    return ret;
412}
413
414static bool check_fit_imm(uint32_t imm)
415{
416    return encode_imm(imm) >= 0;
417}
418
419/* Return true if v16 is a valid 16-bit shifted immediate.  */
420static bool is_shimm16(uint16_t v16, int *cmode, int *imm8)
421{
422    if (v16 == (v16 & 0xff)) {
423        *cmode = 0x8;
424        *imm8 = v16 & 0xff;
425        return true;
426    } else if (v16 == (v16 & 0xff00)) {
427        *cmode = 0xa;
428        *imm8 = v16 >> 8;
429        return true;
430    }
431    return false;
432}
433
434/* Return true if v32 is a valid 32-bit shifted immediate.  */
435static bool is_shimm32(uint32_t v32, int *cmode, int *imm8)
436{
437    if (v32 == (v32 & 0xff)) {
438        *cmode = 0x0;
439        *imm8 = v32 & 0xff;
440        return true;
441    } else if (v32 == (v32 & 0xff00)) {
442        *cmode = 0x2;
443        *imm8 = (v32 >> 8) & 0xff;
444        return true;
445    } else if (v32 == (v32 & 0xff0000)) {
446        *cmode = 0x4;
447        *imm8 = (v32 >> 16) & 0xff;
448        return true;
449    } else if (v32 == (v32 & 0xff000000)) {
450        *cmode = 0x6;
451        *imm8 = v32 >> 24;
452        return true;
453    }
454    return false;
455}
456
457/* Return true if v32 is a valid 32-bit shifting ones immediate.  */
458static bool is_soimm32(uint32_t v32, int *cmode, int *imm8)
459{
460    if ((v32 & 0xffff00ff) == 0xff) {
461        *cmode = 0xc;
462        *imm8 = (v32 >> 8) & 0xff;
463        return true;
464    } else if ((v32 & 0xff00ffff) == 0xffff) {
465        *cmode = 0xd;
466        *imm8 = (v32 >> 16) & 0xff;
467        return true;
468    }
469    return false;
470}
471
472/*
473 * Return non-zero if v32 can be formed by MOVI+ORR.
474 * Place the parameters for MOVI in (cmode, imm8).
475 * Return the cmode for ORR; the imm8 can be had via extraction from v32.
476 */
477static int is_shimm32_pair(uint32_t v32, int *cmode, int *imm8)
478{
479    int i;
480
481    for (i = 6; i > 0; i -= 2) {
482        /* Mask out one byte we can add with ORR.  */
483        uint32_t tmp = v32 & ~(0xffu << (i * 4));
484        if (is_shimm32(tmp, cmode, imm8) ||
485            is_soimm32(tmp, cmode, imm8)) {
486            break;
487        }
488    }
489    return i;
490}
491
492/* Return true if V is a valid 16-bit or 32-bit shifted immediate.  */
493static bool is_shimm1632(uint32_t v32, int *cmode, int *imm8)
494{
495    if (v32 == deposit32(v32, 16, 16, v32)) {
496        return is_shimm16(v32, cmode, imm8);
497    } else {
498        return is_shimm32(v32, cmode, imm8);
499    }
500}
501
502/* Test if a constant matches the constraint.
503 * TODO: define constraints for:
504 *
505 * ldr/str offset:   between -0xfff and 0xfff
506 * ldrh/strh offset: between -0xff and 0xff
507 * mov operand2:     values represented with x << (2 * y), x < 0x100
508 * add, sub, eor...: ditto
509 */
510static bool tcg_target_const_match(int64_t val, int ct,
511                                   TCGType type, TCGCond cond, int vece)
512{
513    if (ct & TCG_CT_CONST) {
514        return 1;
515    } else if ((ct & TCG_CT_CONST_ARM) && check_fit_imm(val)) {
516        return 1;
517    } else if ((ct & TCG_CT_CONST_INV) && check_fit_imm(~val)) {
518        return 1;
519    } else if ((ct & TCG_CT_CONST_NEG) && check_fit_imm(-val)) {
520        return 1;
521    } else if ((ct & TCG_CT_CONST_ZERO) && val == 0) {
522        return 1;
523    }
524
525    switch (ct & (TCG_CT_CONST_ORRI | TCG_CT_CONST_ANDI)) {
526    case 0:
527        break;
528    case TCG_CT_CONST_ANDI:
529        val = ~val;
530        /* fallthru */
531    case TCG_CT_CONST_ORRI:
532        if (val == deposit64(val, 32, 32, val)) {
533            int cmode, imm8;
534            return is_shimm1632(val, &cmode, &imm8);
535        }
536        break;
537    default:
538        /* Both bits should not be set for the same insn.  */
539        g_assert_not_reached();
540    }
541
542    return 0;
543}
544
545static void tcg_out_b_imm(TCGContext *s, ARMCond cond, int32_t offset)
546{
547    tcg_out32(s, (cond << 28) | INSN_B |
548                    (((offset - 8) >> 2) & 0x00ffffff));
549}
550
551static void tcg_out_bl_imm(TCGContext *s, ARMCond cond, int32_t offset)
552{
553    tcg_out32(s, (cond << 28) | 0x0b000000 |
554                    (((offset - 8) >> 2) & 0x00ffffff));
555}
556
557static void tcg_out_blx_reg(TCGContext *s, ARMCond cond, TCGReg rn)
558{
559    tcg_out32(s, (cond << 28) | 0x012fff30 | rn);
560}
561
562static void tcg_out_blx_imm(TCGContext *s, int32_t offset)
563{
564    tcg_out32(s, 0xfa000000 | ((offset & 2) << 23) |
565                (((offset - 8) >> 2) & 0x00ffffff));
566}
567
568static void tcg_out_dat_reg(TCGContext *s, ARMCond cond, ARMInsn opc,
569                            TCGReg rd, TCGReg rn, TCGReg rm, int shift)
570{
571    tcg_out32(s, (cond << 28) | (0 << 25) | opc |
572                    (rn << 16) | (rd << 12) | shift | rm);
573}
574
575static void tcg_out_mov_reg(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rm)
576{
577    /* Simple reg-reg move, optimising out the 'do nothing' case */
578    if (rd != rm) {
579        tcg_out_dat_reg(s, cond, ARITH_MOV, rd, 0, rm, SHIFT_IMM_LSL(0));
580    }
581}
582
583static void tcg_out_bx_reg(TCGContext *s, ARMCond cond, TCGReg rn)
584{
585    tcg_out32(s, (cond << 28) | 0x012fff10 | rn);
586}
587
588static void tcg_out_b_reg(TCGContext *s, ARMCond cond, TCGReg rn)
589{
590    /*
591     * Unless the C portion of QEMU is compiled as thumb, we don't need
592     * true BX semantics; merely a branch to an address held in a register.
593     */
594    tcg_out_bx_reg(s, cond, rn);
595}
596
597static void tcg_out_dat_imm(TCGContext *s, ARMCond cond, ARMInsn opc,
598                            TCGReg rd, TCGReg rn, int im)
599{
600    tcg_out32(s, (cond << 28) | (1 << 25) | opc |
601                    (rn << 16) | (rd << 12) | im);
602}
603
604static void tcg_out_ldstm(TCGContext *s, ARMCond cond, ARMInsn opc,
605                          TCGReg rn, uint16_t mask)
606{
607    tcg_out32(s, (cond << 28) | opc | (rn << 16) | mask);
608}
609
610/* Note that this routine is used for both LDR and LDRH formats, so we do
611   not wish to include an immediate shift at this point.  */
612static void tcg_out_memop_r(TCGContext *s, ARMCond cond, ARMInsn opc, TCGReg rt,
613                            TCGReg rn, TCGReg rm, bool u, bool p, bool w)
614{
615    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24)
616              | (w << 21) | (rn << 16) | (rt << 12) | rm);
617}
618
619static void tcg_out_memop_8(TCGContext *s, ARMCond cond, ARMInsn opc, TCGReg rt,
620                            TCGReg rn, int imm8, bool p, bool w)
621{
622    bool u = 1;
623    if (imm8 < 0) {
624        imm8 = -imm8;
625        u = 0;
626    }
627    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24) | (w << 21) |
628              (rn << 16) | (rt << 12) | ((imm8 & 0xf0) << 4) | (imm8 & 0xf));
629}
630
631static void tcg_out_memop_12(TCGContext *s, ARMCond cond, ARMInsn opc,
632                             TCGReg rt, TCGReg rn, int imm12, bool p, bool w)
633{
634    bool u = 1;
635    if (imm12 < 0) {
636        imm12 = -imm12;
637        u = 0;
638    }
639    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24) | (w << 21) |
640              (rn << 16) | (rt << 12) | imm12);
641}
642
643static void tcg_out_ld32_12(TCGContext *s, ARMCond cond, TCGReg rt,
644                            TCGReg rn, int imm12)
645{
646    tcg_out_memop_12(s, cond, INSN_LDR_IMM, rt, rn, imm12, 1, 0);
647}
648
649static void tcg_out_st32_12(TCGContext *s, ARMCond cond, TCGReg rt,
650                            TCGReg rn, int imm12)
651{
652    tcg_out_memop_12(s, cond, INSN_STR_IMM, rt, rn, imm12, 1, 0);
653}
654
655static void tcg_out_ld32_r(TCGContext *s, ARMCond cond, TCGReg rt,
656                           TCGReg rn, TCGReg rm)
657{
658    tcg_out_memop_r(s, cond, INSN_LDR_REG, rt, rn, rm, 1, 1, 0);
659}
660
661static void tcg_out_st32_r(TCGContext *s, ARMCond cond, TCGReg rt,
662                           TCGReg rn, TCGReg rm)
663{
664    tcg_out_memop_r(s, cond, INSN_STR_REG, rt, rn, rm, 1, 1, 0);
665}
666
667static void tcg_out_ldrd_8(TCGContext *s, ARMCond cond, TCGReg rt,
668                           TCGReg rn, int imm8)
669{
670    tcg_out_memop_8(s, cond, INSN_LDRD_IMM, rt, rn, imm8, 1, 0);
671}
672
673static void tcg_out_ldrd_r(TCGContext *s, ARMCond cond, TCGReg rt,
674                           TCGReg rn, TCGReg rm)
675{
676    tcg_out_memop_r(s, cond, INSN_LDRD_REG, rt, rn, rm, 1, 1, 0);
677}
678
679static void tcg_out_strd_8(TCGContext *s, ARMCond cond, TCGReg rt,
680                           TCGReg rn, int imm8)
681{
682    tcg_out_memop_8(s, cond, INSN_STRD_IMM, rt, rn, imm8, 1, 0);
683}
684
685static void tcg_out_strd_r(TCGContext *s, ARMCond cond, TCGReg rt,
686                           TCGReg rn, TCGReg rm)
687{
688    tcg_out_memop_r(s, cond, INSN_STRD_REG, rt, rn, rm, 1, 1, 0);
689}
690
691/* Register pre-increment with base writeback.  */
692static void tcg_out_ld32_rwb(TCGContext *s, ARMCond cond, TCGReg rt,
693                             TCGReg rn, TCGReg rm)
694{
695    tcg_out_memop_r(s, cond, INSN_LDR_REG, rt, rn, rm, 1, 1, 1);
696}
697
698static void tcg_out_st32_rwb(TCGContext *s, ARMCond cond, TCGReg rt,
699                             TCGReg rn, TCGReg rm)
700{
701    tcg_out_memop_r(s, cond, INSN_STR_REG, rt, rn, rm, 1, 1, 1);
702}
703
704static void tcg_out_ld16u_8(TCGContext *s, ARMCond cond, TCGReg rt,
705                            TCGReg rn, int imm8)
706{
707    tcg_out_memop_8(s, cond, INSN_LDRH_IMM, rt, rn, imm8, 1, 0);
708}
709
710static void tcg_out_st16_8(TCGContext *s, ARMCond cond, TCGReg rt,
711                           TCGReg rn, int imm8)
712{
713    tcg_out_memop_8(s, cond, INSN_STRH_IMM, rt, rn, imm8, 1, 0);
714}
715
716static void tcg_out_ld16u_r(TCGContext *s, ARMCond cond, TCGReg rt,
717                            TCGReg rn, TCGReg rm)
718{
719    tcg_out_memop_r(s, cond, INSN_LDRH_REG, rt, rn, rm, 1, 1, 0);
720}
721
722static void tcg_out_st16_r(TCGContext *s, ARMCond cond, TCGReg rt,
723                           TCGReg rn, TCGReg rm)
724{
725    tcg_out_memop_r(s, cond, INSN_STRH_REG, rt, rn, rm, 1, 1, 0);
726}
727
728static void tcg_out_ld16s_8(TCGContext *s, ARMCond cond, TCGReg rt,
729                            TCGReg rn, int imm8)
730{
731    tcg_out_memop_8(s, cond, INSN_LDRSH_IMM, rt, rn, imm8, 1, 0);
732}
733
734static void tcg_out_ld16s_r(TCGContext *s, ARMCond cond, TCGReg rt,
735                            TCGReg rn, TCGReg rm)
736{
737    tcg_out_memop_r(s, cond, INSN_LDRSH_REG, rt, rn, rm, 1, 1, 0);
738}
739
740static void tcg_out_ld8_12(TCGContext *s, ARMCond cond, TCGReg rt,
741                           TCGReg rn, int imm12)
742{
743    tcg_out_memop_12(s, cond, INSN_LDRB_IMM, rt, rn, imm12, 1, 0);
744}
745
746static void tcg_out_st8_12(TCGContext *s, ARMCond cond, TCGReg rt,
747                           TCGReg rn, int imm12)
748{
749    tcg_out_memop_12(s, cond, INSN_STRB_IMM, rt, rn, imm12, 1, 0);
750}
751
752static void tcg_out_ld8_r(TCGContext *s, ARMCond cond, TCGReg rt,
753                          TCGReg rn, TCGReg rm)
754{
755    tcg_out_memop_r(s, cond, INSN_LDRB_REG, rt, rn, rm, 1, 1, 0);
756}
757
758static void tcg_out_st8_r(TCGContext *s, ARMCond cond, TCGReg rt,
759                          TCGReg rn, TCGReg rm)
760{
761    tcg_out_memop_r(s, cond, INSN_STRB_REG, rt, rn, rm, 1, 1, 0);
762}
763
764static void tcg_out_ld8s_8(TCGContext *s, ARMCond cond, TCGReg rt,
765                           TCGReg rn, int imm8)
766{
767    tcg_out_memop_8(s, cond, INSN_LDRSB_IMM, rt, rn, imm8, 1, 0);
768}
769
770static void tcg_out_ld8s_r(TCGContext *s, ARMCond cond, TCGReg rt,
771                           TCGReg rn, TCGReg rm)
772{
773    tcg_out_memop_r(s, cond, INSN_LDRSB_REG, rt, rn, rm, 1, 1, 0);
774}
775
776static void tcg_out_movi_pool(TCGContext *s, ARMCond cond,
777                              TCGReg rd, uint32_t arg)
778{
779    new_pool_label(s, arg, R_ARM_PC13, s->code_ptr, 0);
780    tcg_out_ld32_12(s, cond, rd, TCG_REG_PC, 0);
781}
782
783static void tcg_out_movi32(TCGContext *s, ARMCond cond,
784                           TCGReg rd, uint32_t arg)
785{
786    int imm12, diff, opc, sh1, sh2;
787    uint32_t tt0, tt1, tt2;
788
789    /* Check a single MOV/MVN before anything else.  */
790    imm12 = encode_imm(arg);
791    if (imm12 >= 0) {
792        tcg_out_dat_imm(s, cond, ARITH_MOV, rd, 0, imm12);
793        return;
794    }
795    imm12 = encode_imm(~arg);
796    if (imm12 >= 0) {
797        tcg_out_dat_imm(s, cond, ARITH_MVN, rd, 0, imm12);
798        return;
799    }
800
801    /* Check for a pc-relative address.  This will usually be the TB,
802       or within the TB, which is immediately before the code block.  */
803    diff = tcg_pcrel_diff(s, (void *)arg) - 8;
804    if (diff >= 0) {
805        imm12 = encode_imm(diff);
806        if (imm12 >= 0) {
807            tcg_out_dat_imm(s, cond, ARITH_ADD, rd, TCG_REG_PC, imm12);
808            return;
809        }
810    } else {
811        imm12 = encode_imm(-diff);
812        if (imm12 >= 0) {
813            tcg_out_dat_imm(s, cond, ARITH_SUB, rd, TCG_REG_PC, imm12);
814            return;
815        }
816    }
817
818    /* Use movw + movt.  */
819    if (use_armv7_instructions) {
820        /* movw */
821        tcg_out32(s, (cond << 28) | 0x03000000 | (rd << 12)
822                  | ((arg << 4) & 0x000f0000) | (arg & 0xfff));
823        if (arg & 0xffff0000) {
824            /* movt */
825            tcg_out32(s, (cond << 28) | 0x03400000 | (rd << 12)
826                      | ((arg >> 12) & 0x000f0000) | ((arg >> 16) & 0xfff));
827        }
828        return;
829    }
830
831    /* Look for sequences of two insns.  If we have lots of 1's, we can
832       shorten the sequence by beginning with mvn and then clearing
833       higher bits with eor.  */
834    tt0 = arg;
835    opc = ARITH_MOV;
836    if (ctpop32(arg) > 16) {
837        tt0 = ~arg;
838        opc = ARITH_MVN;
839    }
840    sh1 = ctz32(tt0) & ~1;
841    tt1 = tt0 & ~(0xff << sh1);
842    sh2 = ctz32(tt1) & ~1;
843    tt2 = tt1 & ~(0xff << sh2);
844    if (tt2 == 0) {
845        int rot;
846
847        rot = ((32 - sh1) << 7) & 0xf00;
848        tcg_out_dat_imm(s, cond, opc, rd,  0, ((tt0 >> sh1) & 0xff) | rot);
849        rot = ((32 - sh2) << 7) & 0xf00;
850        tcg_out_dat_imm(s, cond, ARITH_EOR, rd, rd,
851                        ((tt0 >> sh2) & 0xff) | rot);
852        return;
853    }
854
855    /* Otherwise, drop it into the constant pool.  */
856    tcg_out_movi_pool(s, cond, rd, arg);
857}
858
859/*
860 * Emit either the reg,imm or reg,reg form of a data-processing insn.
861 * rhs must satisfy the "rI" constraint.
862 */
863static void tcg_out_dat_rI(TCGContext *s, ARMCond cond, ARMInsn opc,
864                           TCGReg dst, TCGReg lhs, TCGArg rhs, int rhs_is_const)
865{
866    if (rhs_is_const) {
867        tcg_out_dat_imm(s, cond, opc, dst, lhs, encode_imm_nofail(rhs));
868    } else {
869        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
870    }
871}
872
873/*
874 * Emit either the reg,imm or reg,reg form of a data-processing insn.
875 * rhs must satisfy the "rIK" constraint.
876 */
877static void tcg_out_dat_IK(TCGContext *s, ARMCond cond, ARMInsn opc,
878                            ARMInsn opinv, TCGReg dst, TCGReg lhs, TCGArg rhs)
879{
880    int imm12 = encode_imm(rhs);
881    if (imm12 < 0) {
882        imm12 = encode_imm_nofail(~rhs);
883        opc = opinv;
884    }
885    tcg_out_dat_imm(s, cond, opc, dst, lhs, imm12);
886}
887
888static void tcg_out_dat_rIK(TCGContext *s, ARMCond cond, ARMInsn opc,
889                            ARMInsn opinv, TCGReg dst, TCGReg lhs, TCGArg rhs,
890                            bool rhs_is_const)
891{
892    if (rhs_is_const) {
893        tcg_out_dat_IK(s, cond, opc, opinv, dst, lhs, rhs);
894    } else {
895        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
896    }
897}
898
899static void tcg_out_dat_IN(TCGContext *s, ARMCond cond, ARMInsn opc,
900                           ARMInsn opneg, TCGReg dst, TCGReg lhs, TCGArg rhs)
901{
902    int imm12 = encode_imm(rhs);
903    if (imm12 < 0) {
904        imm12 = encode_imm_nofail(-rhs);
905        opc = opneg;
906    }
907    tcg_out_dat_imm(s, cond, opc, dst, lhs, imm12);
908}
909
910static void tcg_out_dat_rIN(TCGContext *s, ARMCond cond, ARMInsn opc,
911                            ARMInsn opneg, TCGReg dst, TCGReg lhs, TCGArg rhs,
912                            bool rhs_is_const)
913{
914    /* Emit either the reg,imm or reg,reg form of a data-processing insn.
915     * rhs must satisfy the "rIN" constraint.
916     */
917    if (rhs_is_const) {
918        tcg_out_dat_IN(s, cond, opc, opneg, dst, lhs, rhs);
919    } else {
920        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
921    }
922}
923
924static void tcg_out_umull32(TCGContext *s, ARMCond cond, TCGReg rd0,
925                            TCGReg rd1, TCGReg rn, TCGReg rm)
926{
927    /* umull */
928    tcg_out32(s, (cond << 28) | 0x00800090 |
929              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
930}
931
932static void tcg_out_smull32(TCGContext *s, ARMCond cond, TCGReg rd0,
933                            TCGReg rd1, TCGReg rn, TCGReg rm)
934{
935    /* smull */
936    tcg_out32(s, (cond << 28) | 0x00c00090 |
937              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
938}
939
940static void tcg_out_udiv(TCGContext *s, ARMCond cond,
941                         TCGReg rd, TCGReg rn, TCGReg rm)
942{
943    tcg_out32(s, 0x0730f010 | (cond << 28) | (rd << 16) | rn | (rm << 8));
944}
945
946static void tcg_out_ext8s(TCGContext *s, TCGType t, TCGReg rd, TCGReg rn)
947{
948    /* sxtb */
949    tcg_out32(s, 0x06af0070 | (COND_AL << 28) | (rd << 12) | rn);
950}
951
952static void tcg_out_ext8u(TCGContext *s, TCGReg rd, TCGReg rn)
953{
954    tcg_out_dat_imm(s, COND_AL, ARITH_AND, rd, rn, 0xff);
955}
956
957static void tcg_out_ext16s(TCGContext *s, TCGType t, TCGReg rd, TCGReg rn)
958{
959    /* sxth */
960    tcg_out32(s, 0x06bf0070 | (COND_AL << 28) | (rd << 12) | rn);
961}
962
963static void tcg_out_ext16u(TCGContext *s, TCGReg rd, TCGReg rn)
964{
965    /* uxth */
966    tcg_out32(s, 0x06ff0070 | (COND_AL << 28) | (rd << 12) | rn);
967}
968
969static void tcg_out_ext32s(TCGContext *s, TCGReg rd, TCGReg rn)
970{
971    g_assert_not_reached();
972}
973
974static void tcg_out_ext32u(TCGContext *s, TCGReg rd, TCGReg rn)
975{
976    g_assert_not_reached();
977}
978
979static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg rd, TCGReg rn)
980{
981    g_assert_not_reached();
982}
983
984static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg rd, TCGReg rn)
985{
986    g_assert_not_reached();
987}
988
989static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg rd, TCGReg rn)
990{
991    g_assert_not_reached();
992}
993
994static void tcg_out_bswap16(TCGContext *s, ARMCond cond,
995                            TCGReg rd, TCGReg rn, int flags)
996{
997    if (flags & TCG_BSWAP_OS) {
998        /* revsh */
999        tcg_out32(s, 0x06ff0fb0 | (cond << 28) | (rd << 12) | rn);
1000        return;
1001    }
1002
1003    /* rev16 */
1004    tcg_out32(s, 0x06bf0fb0 | (cond << 28) | (rd << 12) | rn);
1005    if ((flags & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1006        /* uxth */
1007        tcg_out32(s, 0x06ff0070 | (cond << 28) | (rd << 12) | rd);
1008    }
1009}
1010
1011static void tcg_out_bswap32(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rn)
1012{
1013    /* rev */
1014    tcg_out32(s, 0x06bf0f30 | (cond << 28) | (rd << 12) | rn);
1015}
1016
1017static void tcg_out_deposit(TCGContext *s, ARMCond cond, TCGReg rd,
1018                            TCGArg a1, int ofs, int len, bool const_a1)
1019{
1020    if (const_a1) {
1021        /* bfi becomes bfc with rn == 15.  */
1022        a1 = 15;
1023    }
1024    /* bfi/bfc */
1025    tcg_out32(s, 0x07c00010 | (cond << 28) | (rd << 12) | a1
1026              | (ofs << 7) | ((ofs + len - 1) << 16));
1027}
1028
1029static void tcg_out_extract(TCGContext *s, ARMCond cond, TCGReg rd,
1030                            TCGReg rn, int ofs, int len)
1031{
1032    /* According to gcc, AND can be faster. */
1033    if (ofs == 0 && len <= 8) {
1034        tcg_out_dat_imm(s, cond, ARITH_AND, rd, rn,
1035                        encode_imm_nofail((1 << len) - 1));
1036        return;
1037    }
1038
1039    if (use_armv7_instructions) {
1040        /* ubfx */
1041        tcg_out32(s, 0x07e00050 | (cond << 28) | (rd << 12) | rn
1042                  | (ofs << 7) | ((len - 1) << 16));
1043        return;
1044    }
1045
1046    assert(ofs % 8 == 0);
1047    switch (len) {
1048    case 8:
1049        /* uxtb */
1050        tcg_out32(s, 0x06ef0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1051        break;
1052    case 16:
1053        /* uxth */
1054        tcg_out32(s, 0x06ff0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1055        break;
1056    default:
1057        g_assert_not_reached();
1058    }
1059}
1060
1061static void tcg_out_sextract(TCGContext *s, ARMCond cond, TCGReg rd,
1062                             TCGReg rn, int ofs, int len)
1063{
1064    if (use_armv7_instructions) {
1065        /* sbfx */
1066        tcg_out32(s, 0x07a00050 | (cond << 28) | (rd << 12) | rn
1067                  | (ofs << 7) | ((len - 1) << 16));
1068        return;
1069    }
1070
1071    assert(ofs % 8 == 0);
1072    switch (len) {
1073    case 8:
1074        /* sxtb */
1075        tcg_out32(s, 0x06af0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1076        break;
1077    case 16:
1078        /* sxth */
1079        tcg_out32(s, 0x06bf0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1080        break;
1081    default:
1082        g_assert_not_reached();
1083    }
1084}
1085
1086
1087static void tcg_out_ld32u(TCGContext *s, ARMCond cond,
1088                          TCGReg rd, TCGReg rn, int32_t offset)
1089{
1090    if (offset > 0xfff || offset < -0xfff) {
1091        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1092        tcg_out_ld32_r(s, cond, rd, rn, TCG_REG_TMP);
1093    } else
1094        tcg_out_ld32_12(s, cond, rd, rn, offset);
1095}
1096
1097static void tcg_out_st32(TCGContext *s, ARMCond cond,
1098                         TCGReg rd, TCGReg rn, int32_t offset)
1099{
1100    if (offset > 0xfff || offset < -0xfff) {
1101        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1102        tcg_out_st32_r(s, cond, rd, rn, TCG_REG_TMP);
1103    } else
1104        tcg_out_st32_12(s, cond, rd, rn, offset);
1105}
1106
1107static void tcg_out_ld16u(TCGContext *s, ARMCond cond,
1108                          TCGReg rd, TCGReg rn, int32_t offset)
1109{
1110    if (offset > 0xff || offset < -0xff) {
1111        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1112        tcg_out_ld16u_r(s, cond, rd, rn, TCG_REG_TMP);
1113    } else
1114        tcg_out_ld16u_8(s, cond, rd, rn, offset);
1115}
1116
1117static void tcg_out_ld16s(TCGContext *s, ARMCond cond,
1118                          TCGReg rd, TCGReg rn, int32_t offset)
1119{
1120    if (offset > 0xff || offset < -0xff) {
1121        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1122        tcg_out_ld16s_r(s, cond, rd, rn, TCG_REG_TMP);
1123    } else
1124        tcg_out_ld16s_8(s, cond, rd, rn, offset);
1125}
1126
1127static void tcg_out_st16(TCGContext *s, ARMCond cond,
1128                         TCGReg rd, TCGReg rn, int32_t offset)
1129{
1130    if (offset > 0xff || offset < -0xff) {
1131        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1132        tcg_out_st16_r(s, cond, rd, rn, TCG_REG_TMP);
1133    } else
1134        tcg_out_st16_8(s, cond, rd, rn, offset);
1135}
1136
1137static void tcg_out_ld8u(TCGContext *s, ARMCond cond,
1138                         TCGReg rd, TCGReg rn, int32_t offset)
1139{
1140    if (offset > 0xfff || offset < -0xfff) {
1141        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1142        tcg_out_ld8_r(s, cond, rd, rn, TCG_REG_TMP);
1143    } else
1144        tcg_out_ld8_12(s, cond, rd, rn, offset);
1145}
1146
1147static void tcg_out_ld8s(TCGContext *s, ARMCond cond,
1148                         TCGReg rd, TCGReg rn, int32_t offset)
1149{
1150    if (offset > 0xff || offset < -0xff) {
1151        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1152        tcg_out_ld8s_r(s, cond, rd, rn, TCG_REG_TMP);
1153    } else
1154        tcg_out_ld8s_8(s, cond, rd, rn, offset);
1155}
1156
1157static void tcg_out_st8(TCGContext *s, ARMCond cond,
1158                        TCGReg rd, TCGReg rn, int32_t offset)
1159{
1160    if (offset > 0xfff || offset < -0xfff) {
1161        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1162        tcg_out_st8_r(s, cond, rd, rn, TCG_REG_TMP);
1163    } else
1164        tcg_out_st8_12(s, cond, rd, rn, offset);
1165}
1166
1167/*
1168 * The _goto case is normally between TBs within the same code buffer, and
1169 * with the code buffer limited to 16MB we wouldn't need the long case.
1170 * But we also use it for the tail-call to the qemu_ld/st helpers, which does.
1171 */
1172static void tcg_out_goto(TCGContext *s, ARMCond cond, const tcg_insn_unit *addr)
1173{
1174    intptr_t addri = (intptr_t)addr;
1175    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1176    bool arm_mode = !(addri & 1);
1177
1178    if (arm_mode && disp - 8 < 0x01fffffd && disp - 8 > -0x01fffffd) {
1179        tcg_out_b_imm(s, cond, disp);
1180        return;
1181    }
1182
1183    /* LDR is interworking from v5t. */
1184    tcg_out_movi_pool(s, cond, TCG_REG_PC, addri);
1185}
1186
1187/*
1188 * The call case is mostly used for helpers - so it's not unreasonable
1189 * for them to be beyond branch range.
1190 */
1191static void tcg_out_call_int(TCGContext *s, const tcg_insn_unit *addr)
1192{
1193    intptr_t addri = (intptr_t)addr;
1194    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1195    bool arm_mode = !(addri & 1);
1196
1197    if (disp - 8 < 0x02000000 && disp - 8 >= -0x02000000) {
1198        if (arm_mode) {
1199            tcg_out_bl_imm(s, COND_AL, disp);
1200        } else {
1201            tcg_out_blx_imm(s, disp);
1202        }
1203        return;
1204    }
1205
1206    tcg_out_movi32(s, COND_AL, TCG_REG_TMP, addri);
1207    tcg_out_blx_reg(s, COND_AL, TCG_REG_TMP);
1208}
1209
1210static void tcg_out_call(TCGContext *s, const tcg_insn_unit *addr,
1211                         const TCGHelperInfo *info)
1212{
1213    tcg_out_call_int(s, addr);
1214}
1215
1216static void tcg_out_goto_label(TCGContext *s, ARMCond cond, TCGLabel *l)
1217{
1218    if (l->has_value) {
1219        tcg_out_goto(s, cond, l->u.value_ptr);
1220    } else {
1221        tcg_out_reloc(s, s->code_ptr, R_ARM_PC24, l, 0);
1222        tcg_out_b_imm(s, cond, 0);
1223    }
1224}
1225
1226static void tcg_out_mb(TCGContext *s, TCGArg a0)
1227{
1228    if (use_armv7_instructions) {
1229        tcg_out32(s, INSN_DMB_ISH);
1230    } else {
1231        tcg_out32(s, INSN_DMB_MCR);
1232    }
1233}
1234
1235static TCGCond tcg_out_cmp(TCGContext *s, TCGCond cond, TCGReg a,
1236                           TCGArg b, int b_const)
1237{
1238    if (!is_tst_cond(cond)) {
1239        tcg_out_dat_rIN(s, COND_AL, ARITH_CMP, ARITH_CMN, 0, a, b, b_const);
1240        return cond;
1241    }
1242
1243    cond = tcg_tst_eqne_cond(cond);
1244    if (b_const) {
1245        int imm12 = encode_imm(b);
1246
1247        /*
1248         * The compare constraints allow rIN, but TST does not support N.
1249         * Be prepared to load the constant into a scratch register.
1250         */
1251        if (imm12 >= 0) {
1252            tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, a, imm12);
1253            return cond;
1254        }
1255        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, b);
1256        b = TCG_REG_TMP;
1257    }
1258    tcg_out_dat_reg(s, COND_AL, ARITH_TST, 0, a, b, SHIFT_IMM_LSL(0));
1259    return cond;
1260}
1261
1262static TCGCond tcg_out_cmp2(TCGContext *s, const TCGArg *args,
1263                            const int *const_args)
1264{
1265    TCGReg al = args[0];
1266    TCGReg ah = args[1];
1267    TCGArg bl = args[2];
1268    TCGArg bh = args[3];
1269    TCGCond cond = args[4];
1270    int const_bl = const_args[2];
1271    int const_bh = const_args[3];
1272
1273    switch (cond) {
1274    case TCG_COND_EQ:
1275    case TCG_COND_NE:
1276    case TCG_COND_LTU:
1277    case TCG_COND_LEU:
1278    case TCG_COND_GTU:
1279    case TCG_COND_GEU:
1280        /*
1281         * We perform a conditional comparison.  If the high half is
1282         * equal, then overwrite the flags with the comparison of the
1283         * low half.  The resulting flags cover the whole.
1284         */
1285        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, ah, bh, const_bh);
1286        tcg_out_dat_rI(s, COND_EQ, ARITH_CMP, 0, al, bl, const_bl);
1287        return cond;
1288
1289    case TCG_COND_TSTEQ:
1290    case TCG_COND_TSTNE:
1291        /* Similar, but with TST instead of CMP. */
1292        tcg_out_dat_rI(s, COND_AL, ARITH_TST, 0, ah, bh, const_bh);
1293        tcg_out_dat_rI(s, COND_EQ, ARITH_TST, 0, al, bl, const_bl);
1294        return tcg_tst_eqne_cond(cond);
1295
1296    case TCG_COND_LT:
1297    case TCG_COND_GE:
1298        /* We perform a double-word subtraction and examine the result.
1299           We do not actually need the result of the subtract, so the
1300           low part "subtract" is a compare.  For the high half we have
1301           no choice but to compute into a temporary.  */
1302        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, al, bl, const_bl);
1303        tcg_out_dat_rI(s, COND_AL, ARITH_SBC | TO_CPSR,
1304                       TCG_REG_TMP, ah, bh, const_bh);
1305        return cond;
1306
1307    case TCG_COND_LE:
1308    case TCG_COND_GT:
1309        /* Similar, but with swapped arguments, via reversed subtract.  */
1310        tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR,
1311                       TCG_REG_TMP, al, bl, const_bl);
1312        tcg_out_dat_rI(s, COND_AL, ARITH_RSC | TO_CPSR,
1313                       TCG_REG_TMP, ah, bh, const_bh);
1314        return tcg_swap_cond(cond);
1315
1316    default:
1317        g_assert_not_reached();
1318    }
1319}
1320
1321/*
1322 * Note that TCGReg references Q-registers.
1323 * Q-regno = 2 * D-regno, so shift left by 1 while inserting.
1324 */
1325static uint32_t encode_vd(TCGReg rd)
1326{
1327    tcg_debug_assert(rd >= TCG_REG_Q0);
1328    return (extract32(rd, 3, 1) << 22) | (extract32(rd, 0, 3) << 13);
1329}
1330
1331static uint32_t encode_vn(TCGReg rn)
1332{
1333    tcg_debug_assert(rn >= TCG_REG_Q0);
1334    return (extract32(rn, 3, 1) << 7) | (extract32(rn, 0, 3) << 17);
1335}
1336
1337static uint32_t encode_vm(TCGReg rm)
1338{
1339    tcg_debug_assert(rm >= TCG_REG_Q0);
1340    return (extract32(rm, 3, 1) << 5) | (extract32(rm, 0, 3) << 1);
1341}
1342
1343static void tcg_out_vreg2(TCGContext *s, ARMInsn insn, int q, int vece,
1344                          TCGReg d, TCGReg m)
1345{
1346    tcg_out32(s, insn | (vece << 18) | (q << 6) |
1347              encode_vd(d) | encode_vm(m));
1348}
1349
1350static void tcg_out_vreg3(TCGContext *s, ARMInsn insn, int q, int vece,
1351                          TCGReg d, TCGReg n, TCGReg m)
1352{
1353    tcg_out32(s, insn | (vece << 20) | (q << 6) |
1354              encode_vd(d) | encode_vn(n) | encode_vm(m));
1355}
1356
1357static void tcg_out_vmovi(TCGContext *s, TCGReg rd,
1358                          int q, int op, int cmode, uint8_t imm8)
1359{
1360    tcg_out32(s, INSN_VMOVI | encode_vd(rd) | (q << 6) | (op << 5)
1361              | (cmode << 8) | extract32(imm8, 0, 4)
1362              | (extract32(imm8, 4, 3) << 16)
1363              | (extract32(imm8, 7, 1) << 24));
1364}
1365
1366static void tcg_out_vshifti(TCGContext *s, ARMInsn insn, int q,
1367                            TCGReg rd, TCGReg rm, int l_imm6)
1368{
1369    tcg_out32(s, insn | (q << 6) | encode_vd(rd) | encode_vm(rm) |
1370              (extract32(l_imm6, 6, 1) << 7) |
1371              (extract32(l_imm6, 0, 6) << 16));
1372}
1373
1374static void tcg_out_vldst(TCGContext *s, ARMInsn insn,
1375                          TCGReg rd, TCGReg rn, int offset)
1376{
1377    if (offset != 0) {
1378        if (check_fit_imm(offset) || check_fit_imm(-offset)) {
1379            tcg_out_dat_rIN(s, COND_AL, ARITH_ADD, ARITH_SUB,
1380                            TCG_REG_TMP, rn, offset, true);
1381        } else {
1382            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP, offset);
1383            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1384                            TCG_REG_TMP, TCG_REG_TMP, rn, 0);
1385        }
1386        rn = TCG_REG_TMP;
1387    }
1388    tcg_out32(s, insn | (rn << 16) | encode_vd(rd) | 0xf);
1389}
1390
1391typedef struct {
1392    ARMCond cond;
1393    TCGReg base;
1394    int index;
1395    bool index_scratch;
1396    TCGAtomAlign aa;
1397} HostAddress;
1398
1399bool tcg_target_has_memory_bswap(MemOp memop)
1400{
1401    return false;
1402}
1403
1404static TCGReg ldst_ra_gen(TCGContext *s, const TCGLabelQemuLdst *l, int arg)
1405{
1406    /* We arrive at the slow path via "BLNE", so R14 contains l->raddr. */
1407    return TCG_REG_R14;
1408}
1409
1410static const TCGLdstHelperParam ldst_helper_param = {
1411    .ra_gen = ldst_ra_gen,
1412    .ntmp = 1,
1413    .tmp = { TCG_REG_TMP },
1414};
1415
1416static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1417{
1418    MemOp opc = get_memop(lb->oi);
1419
1420    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1421        return false;
1422    }
1423
1424    tcg_out_ld_helper_args(s, lb, &ldst_helper_param);
1425    tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SIZE]);
1426    tcg_out_ld_helper_ret(s, lb, false, &ldst_helper_param);
1427
1428    tcg_out_goto(s, COND_AL, lb->raddr);
1429    return true;
1430}
1431
1432static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1433{
1434    MemOp opc = get_memop(lb->oi);
1435
1436    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1437        return false;
1438    }
1439
1440    tcg_out_st_helper_args(s, lb, &ldst_helper_param);
1441
1442    /* Tail-call to the helper, which will return to the fast path.  */
1443    tcg_out_goto(s, COND_AL, qemu_st_helpers[opc & MO_SIZE]);
1444    return true;
1445}
1446
1447/* We expect to use an 9-bit sign-magnitude negative offset from ENV.  */
1448#define MIN_TLB_MASK_TABLE_OFS  -256
1449
1450static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
1451                                           TCGReg addr, MemOpIdx oi, bool is_ld)
1452{
1453    TCGLabelQemuLdst *ldst = NULL;
1454    MemOp opc = get_memop(oi);
1455    unsigned a_mask;
1456
1457    if (tcg_use_softmmu) {
1458        *h = (HostAddress){
1459            .cond = COND_AL,
1460            .base = addr,
1461            .index = TCG_REG_R1,
1462            .index_scratch = true,
1463        };
1464    } else {
1465        *h = (HostAddress){
1466            .cond = COND_AL,
1467            .base = addr,
1468            .index = guest_base ? TCG_REG_GUEST_BASE : -1,
1469            .index_scratch = false,
1470        };
1471    }
1472
1473    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, false);
1474    a_mask = (1 << h->aa.align) - 1;
1475
1476    if (tcg_use_softmmu) {
1477        int mem_index = get_mmuidx(oi);
1478        int cmp_off = is_ld ? offsetof(CPUTLBEntry, addr_read)
1479                            : offsetof(CPUTLBEntry, addr_write);
1480        int fast_off = tlb_mask_table_ofs(s, mem_index);
1481        unsigned s_mask = (1 << (opc & MO_SIZE)) - 1;
1482        TCGReg t_addr;
1483
1484        ldst = new_ldst_label(s);
1485        ldst->is_ld = is_ld;
1486        ldst->oi = oi;
1487        ldst->addr_reg = addr;
1488
1489        /* Load cpu->neg.tlb.f[mmu_idx].{mask,table} into {r0,r1}.  */
1490        QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, mask) != 0);
1491        QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, table) != 4);
1492        tcg_out_ldrd_8(s, COND_AL, TCG_REG_R0, TCG_AREG0, fast_off);
1493
1494        /* Extract the tlb index from the address into R0.  */
1495        tcg_out_dat_reg(s, COND_AL, ARITH_AND, TCG_REG_R0, TCG_REG_R0, addr,
1496                        SHIFT_IMM_LSR(s->page_bits - CPU_TLB_ENTRY_BITS));
1497
1498        /*
1499         * Add the tlb_table pointer, creating the CPUTLBEntry address in R1.
1500         * Load the tlb comparator into R2 and the fast path addend into R1.
1501         */
1502        if (cmp_off == 0) {
1503            tcg_out_ld32_rwb(s, COND_AL, TCG_REG_R2, TCG_REG_R1, TCG_REG_R0);
1504        } else {
1505            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1506                            TCG_REG_R1, TCG_REG_R1, TCG_REG_R0, 0);
1507            tcg_out_ld32_12(s, COND_AL, TCG_REG_R2, TCG_REG_R1, cmp_off);
1508        }
1509
1510        /* Load the tlb addend.  */
1511        tcg_out_ld32_12(s, COND_AL, TCG_REG_R1, TCG_REG_R1,
1512                        offsetof(CPUTLBEntry, addend));
1513
1514        /*
1515         * Check alignment, check comparators.
1516         * Do this in 2-4 insns.  Use MOVW for v7, if possible,
1517         * to reduce the number of sequential conditional instructions.
1518         * Almost all guests have at least 4k pages, which means that we need
1519         * to clear at least 9 bits even for an 8-byte memory, which means it
1520         * isn't worth checking for an immediate operand for BIC.
1521         *
1522         * For unaligned accesses, test the page of the last unit of alignment.
1523         * This leaves the least significant alignment bits unchanged, and of
1524         * course must be zero.
1525         */
1526        t_addr = addr;
1527        if (a_mask < s_mask) {
1528            t_addr = TCG_REG_R0;
1529            tcg_out_dat_imm(s, COND_AL, ARITH_ADD, t_addr,
1530                            addr, s_mask - a_mask);
1531        }
1532        if (use_armv7_instructions && s->page_bits <= 16) {
1533            tcg_out_movi32(s, COND_AL, TCG_REG_TMP, ~(s->page_mask | a_mask));
1534            tcg_out_dat_reg(s, COND_AL, ARITH_BIC, TCG_REG_TMP,
1535                            t_addr, TCG_REG_TMP, 0);
1536            tcg_out_dat_reg(s, COND_AL, ARITH_CMP, 0,
1537                            TCG_REG_R2, TCG_REG_TMP, 0);
1538        } else {
1539            if (a_mask) {
1540                tcg_debug_assert(a_mask <= 0xff);
1541                tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addr, a_mask);
1542            }
1543            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0, t_addr,
1544                            SHIFT_IMM_LSR(s->page_bits));
1545            tcg_out_dat_reg(s, (a_mask ? COND_EQ : COND_AL), ARITH_CMP,
1546                            0, TCG_REG_R2, TCG_REG_TMP,
1547                            SHIFT_IMM_LSL(s->page_bits));
1548        }
1549    } else if (a_mask) {
1550        ldst = new_ldst_label(s);
1551        ldst->is_ld = is_ld;
1552        ldst->oi = oi;
1553        ldst->addr_reg = addr;
1554
1555        /* We are expecting alignment to max out at 7 */
1556        tcg_debug_assert(a_mask <= 0xff);
1557        /* tst addr, #mask */
1558        tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addr, a_mask);
1559    }
1560
1561    return ldst;
1562}
1563
1564static void tcg_out_qemu_ld_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1565                                   TCGReg datahi, HostAddress h)
1566{
1567    TCGReg base;
1568
1569    /* Byte swapping is left to middle-end expansion. */
1570    tcg_debug_assert((opc & MO_BSWAP) == 0);
1571
1572    switch (opc & MO_SSIZE) {
1573    case MO_UB:
1574        if (h.index < 0) {
1575            tcg_out_ld8_12(s, h.cond, datalo, h.base, 0);
1576        } else {
1577            tcg_out_ld8_r(s, h.cond, datalo, h.base, h.index);
1578        }
1579        break;
1580    case MO_SB:
1581        if (h.index < 0) {
1582            tcg_out_ld8s_8(s, h.cond, datalo, h.base, 0);
1583        } else {
1584            tcg_out_ld8s_r(s, h.cond, datalo, h.base, h.index);
1585        }
1586        break;
1587    case MO_UW:
1588        if (h.index < 0) {
1589            tcg_out_ld16u_8(s, h.cond, datalo, h.base, 0);
1590        } else {
1591            tcg_out_ld16u_r(s, h.cond, datalo, h.base, h.index);
1592        }
1593        break;
1594    case MO_SW:
1595        if (h.index < 0) {
1596            tcg_out_ld16s_8(s, h.cond, datalo, h.base, 0);
1597        } else {
1598            tcg_out_ld16s_r(s, h.cond, datalo, h.base, h.index);
1599        }
1600        break;
1601    case MO_UL:
1602        if (h.index < 0) {
1603            tcg_out_ld32_12(s, h.cond, datalo, h.base, 0);
1604        } else {
1605            tcg_out_ld32_r(s, h.cond, datalo, h.base, h.index);
1606        }
1607        break;
1608    case MO_UQ:
1609        /* We used pair allocation for datalo, so already should be aligned. */
1610        tcg_debug_assert((datalo & 1) == 0);
1611        tcg_debug_assert(datahi == datalo + 1);
1612        /* LDRD requires alignment; double-check that. */
1613        if (memop_alignment_bits(opc) >= MO_64) {
1614            if (h.index < 0) {
1615                tcg_out_ldrd_8(s, h.cond, datalo, h.base, 0);
1616                break;
1617            }
1618            /*
1619             * Rm (the second address op) must not overlap Rt or Rt + 1.
1620             * Since datalo is aligned, we can simplify the test via alignment.
1621             * Flip the two address arguments if that works.
1622             */
1623            if ((h.index & ~1) != datalo) {
1624                tcg_out_ldrd_r(s, h.cond, datalo, h.base, h.index);
1625                break;
1626            }
1627            if ((h.base & ~1) != datalo) {
1628                tcg_out_ldrd_r(s, h.cond, datalo, h.index, h.base);
1629                break;
1630            }
1631        }
1632        if (h.index < 0) {
1633            base = h.base;
1634            if (datalo == h.base) {
1635                tcg_out_mov_reg(s, h.cond, TCG_REG_TMP, base);
1636                base = TCG_REG_TMP;
1637            }
1638        } else if (h.index_scratch) {
1639            tcg_out_ld32_rwb(s, h.cond, datalo, h.index, h.base);
1640            tcg_out_ld32_12(s, h.cond, datahi, h.index, 4);
1641            break;
1642        } else {
1643            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1644                            h.base, h.index, SHIFT_IMM_LSL(0));
1645            base = TCG_REG_TMP;
1646        }
1647        tcg_out_ld32_12(s, h.cond, datalo, base, 0);
1648        tcg_out_ld32_12(s, h.cond, datahi, base, 4);
1649        break;
1650    default:
1651        g_assert_not_reached();
1652    }
1653}
1654
1655static void tcg_out_qemu_ld(TCGContext *s, TCGReg datalo, TCGReg datahi,
1656                            TCGReg addr, MemOpIdx oi, TCGType data_type)
1657{
1658    MemOp opc = get_memop(oi);
1659    TCGLabelQemuLdst *ldst;
1660    HostAddress h;
1661
1662    ldst = prepare_host_addr(s, &h, addr, oi, true);
1663    if (ldst) {
1664        ldst->type = data_type;
1665        ldst->datalo_reg = datalo;
1666        ldst->datahi_reg = datahi;
1667
1668        /*
1669         * This a conditional BL only to load a pointer within this
1670         * opcode into LR for the slow path.  We will not be using
1671         * the value for a tail call.
1672         */
1673        ldst->label_ptr[0] = s->code_ptr;
1674        tcg_out_bl_imm(s, COND_NE, 0);
1675
1676        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1677        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1678    } else {
1679        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1680    }
1681}
1682
1683static void tcg_out_qemu_st_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1684                                   TCGReg datahi, HostAddress h)
1685{
1686    /* Byte swapping is left to middle-end expansion. */
1687    tcg_debug_assert((opc & MO_BSWAP) == 0);
1688
1689    switch (opc & MO_SIZE) {
1690    case MO_8:
1691        if (h.index < 0) {
1692            tcg_out_st8_12(s, h.cond, datalo, h.base, 0);
1693        } else {
1694            tcg_out_st8_r(s, h.cond, datalo, h.base, h.index);
1695        }
1696        break;
1697    case MO_16:
1698        if (h.index < 0) {
1699            tcg_out_st16_8(s, h.cond, datalo, h.base, 0);
1700        } else {
1701            tcg_out_st16_r(s, h.cond, datalo, h.base, h.index);
1702        }
1703        break;
1704    case MO_32:
1705        if (h.index < 0) {
1706            tcg_out_st32_12(s, h.cond, datalo, h.base, 0);
1707        } else {
1708            tcg_out_st32_r(s, h.cond, datalo, h.base, h.index);
1709        }
1710        break;
1711    case MO_64:
1712        /* We used pair allocation for datalo, so already should be aligned. */
1713        tcg_debug_assert((datalo & 1) == 0);
1714        tcg_debug_assert(datahi == datalo + 1);
1715        /* STRD requires alignment; double-check that. */
1716        if (memop_alignment_bits(opc) >= MO_64) {
1717            if (h.index < 0) {
1718                tcg_out_strd_8(s, h.cond, datalo, h.base, 0);
1719            } else {
1720                tcg_out_strd_r(s, h.cond, datalo, h.base, h.index);
1721            }
1722        } else if (h.index < 0) {
1723            tcg_out_st32_12(s, h.cond, datalo, h.base, 0);
1724            tcg_out_st32_12(s, h.cond, datahi, h.base, 4);
1725        } else if (h.index_scratch) {
1726            tcg_out_st32_rwb(s, h.cond, datalo, h.index, h.base);
1727            tcg_out_st32_12(s, h.cond, datahi, h.index, 4);
1728        } else {
1729            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1730                            h.base, h.index, SHIFT_IMM_LSL(0));
1731            tcg_out_st32_12(s, h.cond, datalo, TCG_REG_TMP, 0);
1732            tcg_out_st32_12(s, h.cond, datahi, TCG_REG_TMP, 4);
1733        }
1734        break;
1735    default:
1736        g_assert_not_reached();
1737    }
1738}
1739
1740static void tcg_out_qemu_st(TCGContext *s, TCGReg datalo, TCGReg datahi,
1741                            TCGReg addr, MemOpIdx oi, TCGType data_type)
1742{
1743    MemOp opc = get_memop(oi);
1744    TCGLabelQemuLdst *ldst;
1745    HostAddress h;
1746
1747    ldst = prepare_host_addr(s, &h, addr, oi, false);
1748    if (ldst) {
1749        ldst->type = data_type;
1750        ldst->datalo_reg = datalo;
1751        ldst->datahi_reg = datahi;
1752
1753        h.cond = COND_EQ;
1754        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1755
1756        /* The conditional call is last, as we're going to return here. */
1757        ldst->label_ptr[0] = s->code_ptr;
1758        tcg_out_bl_imm(s, COND_NE, 0);
1759        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1760    } else {
1761        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1762    }
1763}
1764
1765static void tcg_out_epilogue(TCGContext *s);
1766
1767static void tcg_out_exit_tb(TCGContext *s, uintptr_t arg)
1768{
1769    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, arg);
1770    tcg_out_epilogue(s);
1771}
1772
1773static void tcg_out_goto_tb(TCGContext *s, int which)
1774{
1775    uintptr_t i_addr;
1776    intptr_t i_disp;
1777
1778    /* Direct branch will be patched by tb_target_set_jmp_target. */
1779    set_jmp_insn_offset(s, which);
1780    tcg_out32(s, INSN_NOP);
1781
1782    /* When branch is out of range, fall through to indirect. */
1783    i_addr = get_jmp_target_addr(s, which);
1784    i_disp = tcg_pcrel_diff(s, (void *)i_addr) - 8;
1785    tcg_debug_assert(i_disp < 0);
1786    if (i_disp >= -0xfff) {
1787        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_PC, i_disp);
1788    } else {
1789        /*
1790         * The TB is close, but outside the 12 bits addressable by
1791         * the load.  We can extend this to 20 bits with a sub of a
1792         * shifted immediate from pc.
1793         */
1794        int h = -i_disp;
1795        int l = -(h & 0xfff);
1796
1797        h = encode_imm_nofail(h + l);
1798        tcg_out_dat_imm(s, COND_AL, ARITH_SUB, TCG_REG_R0, TCG_REG_PC, h);
1799        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_R0, l);
1800    }
1801    set_jmp_reset_offset(s, which);
1802}
1803
1804void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
1805                              uintptr_t jmp_rx, uintptr_t jmp_rw)
1806{
1807    uintptr_t addr = tb->jmp_target_addr[n];
1808    ptrdiff_t offset = addr - (jmp_rx + 8);
1809    tcg_insn_unit insn;
1810
1811    /* Either directly branch, or fall through to indirect branch. */
1812    if (offset == sextract64(offset, 0, 26)) {
1813        /* B <addr> */
1814        insn = deposit32((COND_AL << 28) | INSN_B, 0, 24, offset >> 2);
1815    } else {
1816        insn = INSN_NOP;
1817    }
1818
1819    qatomic_set((uint32_t *)jmp_rw, insn);
1820    flush_idcache_range(jmp_rx, jmp_rw, 4);
1821}
1822
1823
1824static void tgen_add(TCGContext *s, TCGType type,
1825                     TCGReg a0, TCGReg a1, TCGReg a2)
1826{
1827    tcg_out_dat_reg(s, COND_AL, ARITH_ADD, a0, a1, a2, SHIFT_IMM_LSL(0));
1828}
1829
1830static void tgen_addi(TCGContext *s, TCGType type,
1831                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1832{
1833    tcg_out_dat_IN(s, COND_AL, ARITH_ADD, ARITH_SUB, a0, a1, a2);
1834}
1835
1836static const TCGOutOpBinary outop_add = {
1837    .base.static_constraint = C_O1_I2(r, r, rIN),
1838    .out_rrr = tgen_add,
1839    .out_rri = tgen_addi,
1840};
1841
1842static void tgen_and(TCGContext *s, TCGType type,
1843                     TCGReg a0, TCGReg a1, TCGReg a2)
1844{
1845    tcg_out_dat_reg(s, COND_AL, ARITH_AND, a0, a1, a2, SHIFT_IMM_LSL(0));
1846}
1847
1848static void tgen_andi(TCGContext *s, TCGType type,
1849                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1850{
1851    tcg_out_dat_IK(s, COND_AL, ARITH_AND, ARITH_BIC, a0, a1, a2);
1852}
1853
1854static const TCGOutOpBinary outop_and = {
1855    .base.static_constraint = C_O1_I2(r, r, rIK),
1856    .out_rrr = tgen_and,
1857    .out_rri = tgen_andi,
1858};
1859
1860static void tgen_andc(TCGContext *s, TCGType type,
1861                      TCGReg a0, TCGReg a1, TCGReg a2)
1862{
1863    tcg_out_dat_reg(s, COND_AL, ARITH_BIC, a0, a1, a2, SHIFT_IMM_LSL(0));
1864}
1865
1866static const TCGOutOpBinary outop_andc = {
1867    .base.static_constraint = C_O1_I2(r, r, r),
1868    .out_rrr = tgen_andc,
1869};
1870
1871static TCGConstraintSetIndex cset_idiv(TCGType type, unsigned flags)
1872{
1873    return use_idiv_instructions ? C_O1_I2(r, r, r) : C_NotImplemented;
1874}
1875
1876static void tgen_divs(TCGContext *s, TCGType type,
1877                      TCGReg a0, TCGReg a1, TCGReg a2)
1878{
1879    /* sdiv */
1880    tcg_out32(s, 0x0710f010 | (COND_AL << 28) | (a0 << 16) | a1 | (a2 << 8));
1881}
1882
1883static const TCGOutOpBinary outop_divs = {
1884    .base.static_constraint = C_Dynamic,
1885    .base.dynamic_constraint = cset_idiv,
1886    .out_rrr = tgen_divs,
1887};
1888
1889static const TCGOutOpBinary outop_eqv = {
1890    .base.static_constraint = C_NotImplemented,
1891};
1892
1893static void tgen_mul(TCGContext *s, TCGType type,
1894                     TCGReg a0, TCGReg a1, TCGReg a2)
1895{
1896    /* mul */
1897    tcg_out32(s, (COND_AL << 28) | 0x90 | (a0 << 16) | (a1 << 8) | a2);
1898}
1899
1900static const TCGOutOpBinary outop_mul = {
1901    .base.static_constraint = C_O1_I2(r, r, r),
1902    .out_rrr = tgen_mul,
1903};
1904
1905static const TCGOutOpBinary outop_mulsh = {
1906    .base.static_constraint = C_NotImplemented,
1907};
1908
1909static const TCGOutOpBinary outop_muluh = {
1910    .base.static_constraint = C_NotImplemented,
1911};
1912
1913static const TCGOutOpBinary outop_nand = {
1914    .base.static_constraint = C_NotImplemented,
1915};
1916
1917static const TCGOutOpBinary outop_nor = {
1918    .base.static_constraint = C_NotImplemented,
1919};
1920
1921static void tgen_or(TCGContext *s, TCGType type,
1922                     TCGReg a0, TCGReg a1, TCGReg a2)
1923{
1924    tcg_out_dat_reg(s, COND_AL, ARITH_ORR, a0, a1, a2, SHIFT_IMM_LSL(0));
1925}
1926
1927static void tgen_ori(TCGContext *s, TCGType type,
1928                     TCGReg a0, TCGReg a1, tcg_target_long a2)
1929{
1930    tcg_out_dat_imm(s, COND_AL, ARITH_ORR, a0, a1, encode_imm_nofail(a2));
1931}
1932
1933static const TCGOutOpBinary outop_or = {
1934    .base.static_constraint = C_O1_I2(r, r, rI),
1935    .out_rrr = tgen_or,
1936    .out_rri = tgen_ori,
1937};
1938
1939static const TCGOutOpBinary outop_orc = {
1940    .base.static_constraint = C_NotImplemented,
1941};
1942
1943static void tgen_sub(TCGContext *s, TCGType type,
1944                     TCGReg a0, TCGReg a1, TCGReg a2)
1945{
1946    tcg_out_dat_reg(s, COND_AL, ARITH_SUB, a0, a1, a2, SHIFT_IMM_LSL(0));
1947}
1948
1949static void tgen_subfi(TCGContext *s, TCGType type,
1950                       TCGReg a0, tcg_target_long a1, TCGReg a2)
1951{
1952    tcg_out_dat_imm(s, COND_AL, ARITH_RSB, a0, a2, encode_imm_nofail(a1));
1953}
1954
1955static const TCGOutOpSubtract outop_sub = {
1956    .base.static_constraint = C_O1_I2(r, rI, r),
1957    .out_rrr = tgen_sub,
1958    .out_rir = tgen_subfi,
1959};
1960
1961static void tgen_xor(TCGContext *s, TCGType type,
1962                     TCGReg a0, TCGReg a1, TCGReg a2)
1963{
1964    tcg_out_dat_reg(s, COND_AL, ARITH_EOR, a0, a1, a2, SHIFT_IMM_LSL(0));
1965}
1966
1967static void tgen_xori(TCGContext *s, TCGType type,
1968                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1969{
1970    tcg_out_dat_imm(s, COND_AL, ARITH_EOR, a0, a1, encode_imm_nofail(a2));
1971}
1972
1973static const TCGOutOpBinary outop_xor = {
1974    .base.static_constraint = C_O1_I2(r, r, rI),
1975    .out_rrr = tgen_xor,
1976    .out_rri = tgen_xori,
1977};
1978
1979static void tgen_neg(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
1980{
1981    tgen_subfi(s, type, a0, 0, a1);
1982}
1983
1984static const TCGOutOpUnary outop_neg = {
1985    .base.static_constraint = C_O1_I1(r, r),
1986    .out_rr = tgen_neg,
1987};
1988
1989static void tgen_not(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
1990{
1991    tcg_out_dat_reg(s, COND_AL, ARITH_MVN, a0, 0, a1, SHIFT_IMM_LSL(0));
1992}
1993
1994static const TCGOutOpUnary outop_not = {
1995    .base.static_constraint = C_O1_I1(r, r),
1996    .out_rr = tgen_not,
1997};
1998
1999
2000static void tcg_out_op(TCGContext *s, TCGOpcode opc, TCGType type,
2001                       const TCGArg args[TCG_MAX_OP_ARGS],
2002                       const int const_args[TCG_MAX_OP_ARGS])
2003{
2004    TCGArg a0, a1, a2, a3, a4, a5;
2005    int c;
2006
2007    switch (opc) {
2008    case INDEX_op_goto_ptr:
2009        tcg_out_b_reg(s, COND_AL, args[0]);
2010        break;
2011    case INDEX_op_br:
2012        tcg_out_goto_label(s, COND_AL, arg_label(args[0]));
2013        break;
2014
2015    case INDEX_op_ld8u_i32:
2016        tcg_out_ld8u(s, COND_AL, args[0], args[1], args[2]);
2017        break;
2018    case INDEX_op_ld8s_i32:
2019        tcg_out_ld8s(s, COND_AL, args[0], args[1], args[2]);
2020        break;
2021    case INDEX_op_ld16u_i32:
2022        tcg_out_ld16u(s, COND_AL, args[0], args[1], args[2]);
2023        break;
2024    case INDEX_op_ld16s_i32:
2025        tcg_out_ld16s(s, COND_AL, args[0], args[1], args[2]);
2026        break;
2027    case INDEX_op_ld_i32:
2028        tcg_out_ld32u(s, COND_AL, args[0], args[1], args[2]);
2029        break;
2030    case INDEX_op_st8_i32:
2031        tcg_out_st8(s, COND_AL, args[0], args[1], args[2]);
2032        break;
2033    case INDEX_op_st16_i32:
2034        tcg_out_st16(s, COND_AL, args[0], args[1], args[2]);
2035        break;
2036    case INDEX_op_st_i32:
2037        tcg_out_st32(s, COND_AL, args[0], args[1], args[2]);
2038        break;
2039
2040    case INDEX_op_movcond_i32:
2041        /* Constraints mean that v2 is always in the same register as dest,
2042         * so we only need to do "if condition passed, move v1 to dest".
2043         */
2044        c = tcg_out_cmp(s, args[5], args[1], args[2], const_args[2]);
2045        tcg_out_dat_rIK(s, tcg_cond_to_arm_cond[c], ARITH_MOV,
2046                        ARITH_MVN, args[0], 0, args[3], const_args[3]);
2047        break;
2048    case INDEX_op_add2_i32:
2049        a0 = args[0], a1 = args[1], a2 = args[2];
2050        a3 = args[3], a4 = args[4], a5 = args[5];
2051        if (a0 == a3 || (a0 == a5 && !const_args[5])) {
2052            a0 = TCG_REG_TMP;
2053        }
2054        tcg_out_dat_rIN(s, COND_AL, ARITH_ADD | TO_CPSR, ARITH_SUB | TO_CPSR,
2055                        a0, a2, a4, const_args[4]);
2056        tcg_out_dat_rIK(s, COND_AL, ARITH_ADC, ARITH_SBC,
2057                        a1, a3, a5, const_args[5]);
2058        tcg_out_mov_reg(s, COND_AL, args[0], a0);
2059        break;
2060    case INDEX_op_sub2_i32:
2061        a0 = args[0], a1 = args[1], a2 = args[2];
2062        a3 = args[3], a4 = args[4], a5 = args[5];
2063        if ((a0 == a3 && !const_args[3]) || (a0 == a5 && !const_args[5])) {
2064            a0 = TCG_REG_TMP;
2065        }
2066        if (const_args[2]) {
2067            if (const_args[4]) {
2068                tcg_out_movi32(s, COND_AL, a0, a4);
2069                a4 = a0;
2070            }
2071            tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR, a0, a4, a2, 1);
2072        } else {
2073            tcg_out_dat_rIN(s, COND_AL, ARITH_SUB | TO_CPSR,
2074                            ARITH_ADD | TO_CPSR, a0, a2, a4, const_args[4]);
2075        }
2076        if (const_args[3]) {
2077            if (const_args[5]) {
2078                tcg_out_movi32(s, COND_AL, a1, a5);
2079                a5 = a1;
2080            }
2081            tcg_out_dat_rI(s, COND_AL, ARITH_RSC, a1, a5, a3, 1);
2082        } else {
2083            tcg_out_dat_rIK(s, COND_AL, ARITH_SBC, ARITH_ADC,
2084                            a1, a3, a5, const_args[5]);
2085        }
2086        tcg_out_mov_reg(s, COND_AL, args[0], a0);
2087        break;
2088    case INDEX_op_mulu2_i32:
2089        tcg_out_umull32(s, COND_AL, args[0], args[1], args[2], args[3]);
2090        break;
2091    case INDEX_op_muls2_i32:
2092        tcg_out_smull32(s, COND_AL, args[0], args[1], args[2], args[3]);
2093        break;
2094    /* XXX: Perhaps args[2] & 0x1f is wrong */
2095    case INDEX_op_shl_i32:
2096        c = const_args[2] ?
2097                SHIFT_IMM_LSL(args[2] & 0x1f) : SHIFT_REG_LSL(args[2]);
2098        goto gen_shift32;
2099    case INDEX_op_shr_i32:
2100        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_LSR(args[2] & 0x1f) :
2101                SHIFT_IMM_LSL(0) : SHIFT_REG_LSR(args[2]);
2102        goto gen_shift32;
2103    case INDEX_op_sar_i32:
2104        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_ASR(args[2] & 0x1f) :
2105                SHIFT_IMM_LSL(0) : SHIFT_REG_ASR(args[2]);
2106        goto gen_shift32;
2107    case INDEX_op_rotr_i32:
2108        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_ROR(args[2] & 0x1f) :
2109                SHIFT_IMM_LSL(0) : SHIFT_REG_ROR(args[2]);
2110        /* Fall through.  */
2111    gen_shift32:
2112        tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1], c);
2113        break;
2114
2115    case INDEX_op_rotl_i32:
2116        if (const_args[2]) {
2117            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1],
2118                            ((0x20 - args[2]) & 0x1f) ?
2119                            SHIFT_IMM_ROR((0x20 - args[2]) & 0x1f) :
2120                            SHIFT_IMM_LSL(0));
2121        } else {
2122            tcg_out_dat_imm(s, COND_AL, ARITH_RSB, TCG_REG_TMP, args[2], 0x20);
2123            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1],
2124                            SHIFT_REG_ROR(TCG_REG_TMP));
2125        }
2126        break;
2127
2128    case INDEX_op_ctz_i32:
2129        tcg_out_dat_reg(s, COND_AL, INSN_RBIT, TCG_REG_TMP, 0, args[1], 0);
2130        a1 = TCG_REG_TMP;
2131        goto do_clz;
2132
2133    case INDEX_op_clz_i32:
2134        a1 = args[1];
2135    do_clz:
2136        a0 = args[0];
2137        a2 = args[2];
2138        c = const_args[2];
2139        if (c && a2 == 32) {
2140            tcg_out_dat_reg(s, COND_AL, INSN_CLZ, a0, 0, a1, 0);
2141            break;
2142        }
2143        tcg_out_dat_imm(s, COND_AL, ARITH_CMP, 0, a1, 0);
2144        tcg_out_dat_reg(s, COND_NE, INSN_CLZ, a0, 0, a1, 0);
2145        if (c || a0 != a2) {
2146            tcg_out_dat_rIK(s, COND_EQ, ARITH_MOV, ARITH_MVN, a0, 0, a2, c);
2147        }
2148        break;
2149
2150    case INDEX_op_brcond_i32:
2151        c = tcg_out_cmp(s, args[2], args[0], args[1], const_args[1]);
2152        tcg_out_goto_label(s, tcg_cond_to_arm_cond[c], arg_label(args[3]));
2153        break;
2154    case INDEX_op_setcond_i32:
2155        c = tcg_out_cmp(s, args[3], args[1], args[2], const_args[2]);
2156        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c],
2157                        ARITH_MOV, args[0], 0, 1);
2158        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
2159                        ARITH_MOV, args[0], 0, 0);
2160        break;
2161    case INDEX_op_negsetcond_i32:
2162        c = tcg_out_cmp(s, args[3], args[1], args[2], const_args[2]);
2163        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c],
2164                        ARITH_MVN, args[0], 0, 0);
2165        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
2166                        ARITH_MOV, args[0], 0, 0);
2167        break;
2168
2169    case INDEX_op_brcond2_i32:
2170        c = tcg_out_cmp2(s, args, const_args);
2171        tcg_out_goto_label(s, tcg_cond_to_arm_cond[c], arg_label(args[5]));
2172        break;
2173    case INDEX_op_setcond2_i32:
2174        c = tcg_out_cmp2(s, args + 1, const_args + 1);
2175        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c], ARITH_MOV, args[0], 0, 1);
2176        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
2177                        ARITH_MOV, args[0], 0, 0);
2178        break;
2179
2180    case INDEX_op_qemu_ld_i32:
2181        tcg_out_qemu_ld(s, args[0], -1, args[1], args[2], TCG_TYPE_I32);
2182        break;
2183    case INDEX_op_qemu_ld_i64:
2184        tcg_out_qemu_ld(s, args[0], args[1], args[2], args[3], TCG_TYPE_I64);
2185        break;
2186
2187    case INDEX_op_qemu_st_i32:
2188        tcg_out_qemu_st(s, args[0], -1, args[1], args[2], TCG_TYPE_I32);
2189        break;
2190    case INDEX_op_qemu_st_i64:
2191        tcg_out_qemu_st(s, args[0], args[1], args[2], args[3], TCG_TYPE_I64);
2192        break;
2193
2194    case INDEX_op_bswap16_i32:
2195        tcg_out_bswap16(s, COND_AL, args[0], args[1], args[2]);
2196        break;
2197    case INDEX_op_bswap32_i32:
2198        tcg_out_bswap32(s, COND_AL, args[0], args[1]);
2199        break;
2200
2201    case INDEX_op_deposit_i32:
2202        tcg_out_deposit(s, COND_AL, args[0], args[2],
2203                        args[3], args[4], const_args[2]);
2204        break;
2205    case INDEX_op_extract_i32:
2206        tcg_out_extract(s, COND_AL, args[0], args[1], args[2], args[3]);
2207        break;
2208    case INDEX_op_sextract_i32:
2209        tcg_out_sextract(s, COND_AL, args[0], args[1], args[2], args[3]);
2210        break;
2211    case INDEX_op_extract2_i32:
2212        /* ??? These optimization vs zero should be generic.  */
2213        /* ??? But we can't substitute 2 for 1 in the opcode stream yet.  */
2214        if (const_args[1]) {
2215            if (const_args[2]) {
2216                tcg_out_movi(s, TCG_TYPE_REG, args[0], 0);
2217            } else {
2218                tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2219                                args[2], SHIFT_IMM_LSL(32 - args[3]));
2220            }
2221        } else if (const_args[2]) {
2222            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2223                            args[1], SHIFT_IMM_LSR(args[3]));
2224        } else {
2225            /* We can do extract2 in 2 insns, vs the 3 required otherwise.  */
2226            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0,
2227                            args[2], SHIFT_IMM_LSL(32 - args[3]));
2228            tcg_out_dat_reg(s, COND_AL, ARITH_ORR, args[0], TCG_REG_TMP,
2229                            args[1], SHIFT_IMM_LSR(args[3]));
2230        }
2231        break;
2232
2233    case INDEX_op_divu_i32:
2234        tcg_out_udiv(s, COND_AL, args[0], args[1], args[2]);
2235        break;
2236
2237    case INDEX_op_mb:
2238        tcg_out_mb(s, args[0]);
2239        break;
2240
2241    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
2242    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
2243    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
2244    default:
2245        g_assert_not_reached();
2246    }
2247}
2248
2249static TCGConstraintSetIndex
2250tcg_target_op_def(TCGOpcode op, TCGType type, unsigned flags)
2251{
2252    switch (op) {
2253    case INDEX_op_goto_ptr:
2254        return C_O0_I1(r);
2255
2256    case INDEX_op_ld8u_i32:
2257    case INDEX_op_ld8s_i32:
2258    case INDEX_op_ld16u_i32:
2259    case INDEX_op_ld16s_i32:
2260    case INDEX_op_ld_i32:
2261    case INDEX_op_bswap16_i32:
2262    case INDEX_op_bswap32_i32:
2263    case INDEX_op_extract_i32:
2264    case INDEX_op_sextract_i32:
2265        return C_O1_I1(r, r);
2266
2267    case INDEX_op_st8_i32:
2268    case INDEX_op_st16_i32:
2269    case INDEX_op_st_i32:
2270        return C_O0_I2(r, r);
2271
2272    case INDEX_op_setcond_i32:
2273    case INDEX_op_negsetcond_i32:
2274        return C_O1_I2(r, r, rIN);
2275
2276    case INDEX_op_clz_i32:
2277    case INDEX_op_ctz_i32:
2278        return C_O1_I2(r, r, rIK);
2279
2280    case INDEX_op_divu_i32:
2281        return C_O1_I2(r, r, r);
2282
2283    case INDEX_op_mulu2_i32:
2284    case INDEX_op_muls2_i32:
2285        return C_O2_I2(r, r, r, r);
2286
2287    case INDEX_op_shl_i32:
2288    case INDEX_op_shr_i32:
2289    case INDEX_op_sar_i32:
2290    case INDEX_op_rotl_i32:
2291    case INDEX_op_rotr_i32:
2292        return C_O1_I2(r, r, ri);
2293
2294    case INDEX_op_brcond_i32:
2295        return C_O0_I2(r, rIN);
2296    case INDEX_op_deposit_i32:
2297        return C_O1_I2(r, 0, rZ);
2298    case INDEX_op_extract2_i32:
2299        return C_O1_I2(r, rZ, rZ);
2300    case INDEX_op_movcond_i32:
2301        return C_O1_I4(r, r, rIN, rIK, 0);
2302    case INDEX_op_add2_i32:
2303        return C_O2_I4(r, r, r, r, rIN, rIK);
2304    case INDEX_op_sub2_i32:
2305        return C_O2_I4(r, r, rI, rI, rIN, rIK);
2306    case INDEX_op_brcond2_i32:
2307        return C_O0_I4(r, r, rI, rI);
2308    case INDEX_op_setcond2_i32:
2309        return C_O1_I4(r, r, r, rI, rI);
2310
2311    case INDEX_op_qemu_ld_i32:
2312        return C_O1_I1(r, q);
2313    case INDEX_op_qemu_ld_i64:
2314        return C_O2_I1(e, p, q);
2315    case INDEX_op_qemu_st_i32:
2316        return C_O0_I2(q, q);
2317    case INDEX_op_qemu_st_i64:
2318        return C_O0_I3(Q, p, q);
2319
2320    case INDEX_op_st_vec:
2321        return C_O0_I2(w, r);
2322    case INDEX_op_ld_vec:
2323    case INDEX_op_dupm_vec:
2324        return C_O1_I1(w, r);
2325    case INDEX_op_dup_vec:
2326        return C_O1_I1(w, wr);
2327    case INDEX_op_abs_vec:
2328    case INDEX_op_neg_vec:
2329    case INDEX_op_not_vec:
2330    case INDEX_op_shli_vec:
2331    case INDEX_op_shri_vec:
2332    case INDEX_op_sari_vec:
2333        return C_O1_I1(w, w);
2334    case INDEX_op_dup2_vec:
2335    case INDEX_op_add_vec:
2336    case INDEX_op_mul_vec:
2337    case INDEX_op_smax_vec:
2338    case INDEX_op_smin_vec:
2339    case INDEX_op_ssadd_vec:
2340    case INDEX_op_sssub_vec:
2341    case INDEX_op_sub_vec:
2342    case INDEX_op_umax_vec:
2343    case INDEX_op_umin_vec:
2344    case INDEX_op_usadd_vec:
2345    case INDEX_op_ussub_vec:
2346    case INDEX_op_xor_vec:
2347    case INDEX_op_arm_sshl_vec:
2348    case INDEX_op_arm_ushl_vec:
2349        return C_O1_I2(w, w, w);
2350    case INDEX_op_arm_sli_vec:
2351        return C_O1_I2(w, 0, w);
2352    case INDEX_op_or_vec:
2353    case INDEX_op_andc_vec:
2354        return C_O1_I2(w, w, wO);
2355    case INDEX_op_and_vec:
2356    case INDEX_op_orc_vec:
2357        return C_O1_I2(w, w, wV);
2358    case INDEX_op_cmp_vec:
2359        return C_O1_I2(w, w, wZ);
2360    case INDEX_op_bitsel_vec:
2361        return C_O1_I3(w, w, w, w);
2362    default:
2363        return C_NotImplemented;
2364    }
2365}
2366
2367static void tcg_target_init(TCGContext *s)
2368{
2369    /*
2370     * Only probe for the platform and capabilities if we haven't already
2371     * determined maximum values at compile time.
2372     */
2373#if !defined(use_idiv_instructions) || !defined(use_neon_instructions)
2374    {
2375        unsigned long hwcap = qemu_getauxval(AT_HWCAP);
2376#ifndef use_idiv_instructions
2377        use_idiv_instructions = (hwcap & HWCAP_ARM_IDIVA) != 0;
2378#endif
2379#ifndef use_neon_instructions
2380        use_neon_instructions = (hwcap & HWCAP_ARM_NEON) != 0;
2381#endif
2382    }
2383#endif
2384
2385    if (__ARM_ARCH < 7) {
2386        const char *pl = (const char *)qemu_getauxval(AT_PLATFORM);
2387        if (pl != NULL && pl[0] == 'v' && pl[1] >= '4' && pl[1] <= '9') {
2388            arm_arch = pl[1] - '0';
2389        }
2390
2391        if (arm_arch < 6) {
2392            error_report("TCG: ARMv%d is unsupported; exiting", arm_arch);
2393            exit(EXIT_FAILURE);
2394        }
2395    }
2396
2397    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
2398
2399    tcg_target_call_clobber_regs = 0;
2400    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R0);
2401    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R1);
2402    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R2);
2403    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R3);
2404    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R12);
2405    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R14);
2406
2407    if (use_neon_instructions) {
2408        tcg_target_available_regs[TCG_TYPE_V64]  = ALL_VECTOR_REGS;
2409        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
2410
2411        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q0);
2412        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q1);
2413        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q2);
2414        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q3);
2415        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q8);
2416        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q9);
2417        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q10);
2418        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q11);
2419        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q12);
2420        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q13);
2421        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q14);
2422        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q15);
2423    }
2424
2425    s->reserved_regs = 0;
2426    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
2427    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP);
2428    tcg_regset_set_reg(s->reserved_regs, TCG_REG_PC);
2429    tcg_regset_set_reg(s->reserved_regs, TCG_VEC_TMP);
2430}
2431
2432static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg arg,
2433                       TCGReg arg1, intptr_t arg2)
2434{
2435    switch (type) {
2436    case TCG_TYPE_I32:
2437        tcg_out_ld32u(s, COND_AL, arg, arg1, arg2);
2438        return;
2439    case TCG_TYPE_V64:
2440        /* regs 1; size 8; align 8 */
2441        tcg_out_vldst(s, INSN_VLD1 | 0x7d0, arg, arg1, arg2);
2442        return;
2443    case TCG_TYPE_V128:
2444        /*
2445         * We have only 8-byte alignment for the stack per the ABI.
2446         * Rather than dynamically re-align the stack, it's easier
2447         * to simply not request alignment beyond that.  So:
2448         * regs 2; size 8; align 8
2449         */
2450        tcg_out_vldst(s, INSN_VLD1 | 0xad0, arg, arg1, arg2);
2451        return;
2452    default:
2453        g_assert_not_reached();
2454    }
2455}
2456
2457static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
2458                       TCGReg arg1, intptr_t arg2)
2459{
2460    switch (type) {
2461    case TCG_TYPE_I32:
2462        tcg_out_st32(s, COND_AL, arg, arg1, arg2);
2463        return;
2464    case TCG_TYPE_V64:
2465        /* regs 1; size 8; align 8 */
2466        tcg_out_vldst(s, INSN_VST1 | 0x7d0, arg, arg1, arg2);
2467        return;
2468    case TCG_TYPE_V128:
2469        /* See tcg_out_ld re alignment: regs 2; size 8; align 8 */
2470        tcg_out_vldst(s, INSN_VST1 | 0xad0, arg, arg1, arg2);
2471        return;
2472    default:
2473        g_assert_not_reached();
2474    }
2475}
2476
2477static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
2478                        TCGReg base, intptr_t ofs)
2479{
2480    return false;
2481}
2482
2483static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
2484{
2485    if (ret == arg) {
2486        return true;
2487    }
2488    switch (type) {
2489    case TCG_TYPE_I32:
2490        if (ret < TCG_REG_Q0 && arg < TCG_REG_Q0) {
2491            tcg_out_mov_reg(s, COND_AL, ret, arg);
2492            return true;
2493        }
2494        return false;
2495
2496    case TCG_TYPE_V64:
2497    case TCG_TYPE_V128:
2498        /* "VMOV D,N" is an alias for "VORR D,N,N". */
2499        tcg_out_vreg3(s, INSN_VORR, type - TCG_TYPE_V64, 0, ret, arg, arg);
2500        return true;
2501
2502    default:
2503        g_assert_not_reached();
2504    }
2505}
2506
2507static void tcg_out_movi(TCGContext *s, TCGType type,
2508                         TCGReg ret, tcg_target_long arg)
2509{
2510    tcg_debug_assert(type == TCG_TYPE_I32);
2511    tcg_debug_assert(ret < TCG_REG_Q0);
2512    tcg_out_movi32(s, COND_AL, ret, arg);
2513}
2514
2515static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
2516{
2517    return false;
2518}
2519
2520static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
2521                             tcg_target_long imm)
2522{
2523    int enc, opc = ARITH_ADD;
2524
2525    /* All of the easiest immediates to encode are positive. */
2526    if (imm < 0) {
2527        imm = -imm;
2528        opc = ARITH_SUB;
2529    }
2530    enc = encode_imm(imm);
2531    if (enc >= 0) {
2532        tcg_out_dat_imm(s, COND_AL, opc, rd, rs, enc);
2533    } else {
2534        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, imm);
2535        tcg_out_dat_reg(s, COND_AL, opc, rd, rs,
2536                        TCG_REG_TMP, SHIFT_IMM_LSL(0));
2537    }
2538}
2539
2540/* Type is always V128, with I64 elements.  */
2541static void tcg_out_dup2_vec(TCGContext *s, TCGReg rd, TCGReg rl, TCGReg rh)
2542{
2543    /* Move high element into place first. */
2544    /* VMOV Dd+1, Ds */
2545    tcg_out_vreg3(s, INSN_VORR | (1 << 12), 0, 0, rd, rh, rh);
2546    /* Move low element into place; tcg_out_mov will check for nop. */
2547    tcg_out_mov(s, TCG_TYPE_V64, rd, rl);
2548}
2549
2550static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
2551                            TCGReg rd, TCGReg rs)
2552{
2553    int q = type - TCG_TYPE_V64;
2554
2555    if (vece == MO_64) {
2556        if (type == TCG_TYPE_V128) {
2557            tcg_out_dup2_vec(s, rd, rs, rs);
2558        } else {
2559            tcg_out_mov(s, TCG_TYPE_V64, rd, rs);
2560        }
2561    } else if (rs < TCG_REG_Q0) {
2562        int b = (vece == MO_8);
2563        int e = (vece == MO_16);
2564        tcg_out32(s, INSN_VDUP_G | (b << 22) | (q << 21) | (e << 5) |
2565                  encode_vn(rd) | (rs << 12));
2566    } else {
2567        int imm4 = 1 << vece;
2568        tcg_out32(s, INSN_VDUP_S | (imm4 << 16) | (q << 6) |
2569                  encode_vd(rd) | encode_vm(rs));
2570    }
2571    return true;
2572}
2573
2574static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
2575                             TCGReg rd, TCGReg base, intptr_t offset)
2576{
2577    if (vece == MO_64) {
2578        tcg_out_ld(s, TCG_TYPE_V64, rd, base, offset);
2579        if (type == TCG_TYPE_V128) {
2580            tcg_out_dup2_vec(s, rd, rd, rd);
2581        }
2582    } else {
2583        int q = type - TCG_TYPE_V64;
2584        tcg_out_vldst(s, INSN_VLD1R | (vece << 6) | (q << 5),
2585                      rd, base, offset);
2586    }
2587    return true;
2588}
2589
2590static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
2591                             TCGReg rd, int64_t v64)
2592{
2593    int q = type - TCG_TYPE_V64;
2594    int cmode, imm8, i;
2595
2596    /* Test all bytes equal first.  */
2597    if (vece == MO_8) {
2598        tcg_out_vmovi(s, rd, q, 0, 0xe, v64);
2599        return;
2600    }
2601
2602    /*
2603     * Test all bytes 0x00 or 0xff second.  This can match cases that
2604     * might otherwise take 2 or 3 insns for MO_16 or MO_32 below.
2605     */
2606    for (i = imm8 = 0; i < 8; i++) {
2607        uint8_t byte = v64 >> (i * 8);
2608        if (byte == 0xff) {
2609            imm8 |= 1 << i;
2610        } else if (byte != 0) {
2611            goto fail_bytes;
2612        }
2613    }
2614    tcg_out_vmovi(s, rd, q, 1, 0xe, imm8);
2615    return;
2616 fail_bytes:
2617
2618    /*
2619     * Tests for various replications.  For each element width, if we
2620     * cannot find an expansion there's no point checking a larger
2621     * width because we already know by replication it cannot match.
2622     */
2623    if (vece == MO_16) {
2624        uint16_t v16 = v64;
2625
2626        if (is_shimm16(v16, &cmode, &imm8)) {
2627            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2628            return;
2629        }
2630        if (is_shimm16(~v16, &cmode, &imm8)) {
2631            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2632            return;
2633        }
2634
2635        /*
2636         * Otherwise, all remaining constants can be loaded in two insns:
2637         * rd = v16 & 0xff, rd |= v16 & 0xff00.
2638         */
2639        tcg_out_vmovi(s, rd, q, 0, 0x8, v16 & 0xff);
2640        tcg_out_vmovi(s, rd, q, 0, 0xb, v16 >> 8);   /* VORRI */
2641        return;
2642    }
2643
2644    if (vece == MO_32) {
2645        uint32_t v32 = v64;
2646
2647        if (is_shimm32(v32, &cmode, &imm8) ||
2648            is_soimm32(v32, &cmode, &imm8)) {
2649            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2650            return;
2651        }
2652        if (is_shimm32(~v32, &cmode, &imm8) ||
2653            is_soimm32(~v32, &cmode, &imm8)) {
2654            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2655            return;
2656        }
2657
2658        /*
2659         * Restrict the set of constants to those we can load with
2660         * two instructions.  Others we load from the pool.
2661         */
2662        i = is_shimm32_pair(v32, &cmode, &imm8);
2663        if (i) {
2664            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2665            tcg_out_vmovi(s, rd, q, 0, i | 1, extract32(v32, i * 4, 8));
2666            return;
2667        }
2668        i = is_shimm32_pair(~v32, &cmode, &imm8);
2669        if (i) {
2670            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2671            tcg_out_vmovi(s, rd, q, 1, i | 1, extract32(~v32, i * 4, 8));
2672            return;
2673        }
2674    }
2675
2676    /*
2677     * As a last resort, load from the constant pool.
2678     */
2679    if (!q || vece == MO_64) {
2680        new_pool_l2(s, R_ARM_PC11, s->code_ptr, 0, v64, v64 >> 32);
2681        /* VLDR Dd, [pc + offset] */
2682        tcg_out32(s, INSN_VLDR_D | encode_vd(rd) | (0xf << 16));
2683        if (q) {
2684            tcg_out_dup2_vec(s, rd, rd, rd);
2685        }
2686    } else {
2687        new_pool_label(s, (uint32_t)v64, R_ARM_PC8, s->code_ptr, 0);
2688        /* add tmp, pc, offset */
2689        tcg_out_dat_imm(s, COND_AL, ARITH_ADD, TCG_REG_TMP, TCG_REG_PC, 0);
2690        tcg_out_dupm_vec(s, type, MO_32, rd, TCG_REG_TMP, 0);
2691    }
2692}
2693
2694static const ARMInsn vec_cmp_insn[16] = {
2695    [TCG_COND_EQ] = INSN_VCEQ,
2696    [TCG_COND_GT] = INSN_VCGT,
2697    [TCG_COND_GE] = INSN_VCGE,
2698    [TCG_COND_GTU] = INSN_VCGT_U,
2699    [TCG_COND_GEU] = INSN_VCGE_U,
2700};
2701
2702static const ARMInsn vec_cmp0_insn[16] = {
2703    [TCG_COND_EQ] = INSN_VCEQ0,
2704    [TCG_COND_GT] = INSN_VCGT0,
2705    [TCG_COND_GE] = INSN_VCGE0,
2706    [TCG_COND_LT] = INSN_VCLT0,
2707    [TCG_COND_LE] = INSN_VCLE0,
2708};
2709
2710static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
2711                           unsigned vecl, unsigned vece,
2712                           const TCGArg args[TCG_MAX_OP_ARGS],
2713                           const int const_args[TCG_MAX_OP_ARGS])
2714{
2715    TCGType type = vecl + TCG_TYPE_V64;
2716    unsigned q = vecl;
2717    TCGArg a0, a1, a2, a3;
2718    int cmode, imm8;
2719
2720    a0 = args[0];
2721    a1 = args[1];
2722    a2 = args[2];
2723
2724    switch (opc) {
2725    case INDEX_op_ld_vec:
2726        tcg_out_ld(s, type, a0, a1, a2);
2727        return;
2728    case INDEX_op_st_vec:
2729        tcg_out_st(s, type, a0, a1, a2);
2730        return;
2731    case INDEX_op_dupm_vec:
2732        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
2733        return;
2734    case INDEX_op_dup2_vec:
2735        tcg_out_dup2_vec(s, a0, a1, a2);
2736        return;
2737    case INDEX_op_abs_vec:
2738        tcg_out_vreg2(s, INSN_VABS, q, vece, a0, a1);
2739        return;
2740    case INDEX_op_neg_vec:
2741        tcg_out_vreg2(s, INSN_VNEG, q, vece, a0, a1);
2742        return;
2743    case INDEX_op_not_vec:
2744        tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a1);
2745        return;
2746    case INDEX_op_add_vec:
2747        tcg_out_vreg3(s, INSN_VADD, q, vece, a0, a1, a2);
2748        return;
2749    case INDEX_op_mul_vec:
2750        tcg_out_vreg3(s, INSN_VMUL, q, vece, a0, a1, a2);
2751        return;
2752    case INDEX_op_smax_vec:
2753        tcg_out_vreg3(s, INSN_VMAX, q, vece, a0, a1, a2);
2754        return;
2755    case INDEX_op_smin_vec:
2756        tcg_out_vreg3(s, INSN_VMIN, q, vece, a0, a1, a2);
2757        return;
2758    case INDEX_op_sub_vec:
2759        tcg_out_vreg3(s, INSN_VSUB, q, vece, a0, a1, a2);
2760        return;
2761    case INDEX_op_ssadd_vec:
2762        tcg_out_vreg3(s, INSN_VQADD, q, vece, a0, a1, a2);
2763        return;
2764    case INDEX_op_sssub_vec:
2765        tcg_out_vreg3(s, INSN_VQSUB, q, vece, a0, a1, a2);
2766        return;
2767    case INDEX_op_umax_vec:
2768        tcg_out_vreg3(s, INSN_VMAX_U, q, vece, a0, a1, a2);
2769        return;
2770    case INDEX_op_umin_vec:
2771        tcg_out_vreg3(s, INSN_VMIN_U, q, vece, a0, a1, a2);
2772        return;
2773    case INDEX_op_usadd_vec:
2774        tcg_out_vreg3(s, INSN_VQADD_U, q, vece, a0, a1, a2);
2775        return;
2776    case INDEX_op_ussub_vec:
2777        tcg_out_vreg3(s, INSN_VQSUB_U, q, vece, a0, a1, a2);
2778        return;
2779    case INDEX_op_xor_vec:
2780        tcg_out_vreg3(s, INSN_VEOR, q, 0, a0, a1, a2);
2781        return;
2782    case INDEX_op_arm_sshl_vec:
2783        /*
2784         * Note that Vm is the data and Vn is the shift count,
2785         * therefore the arguments appear reversed.
2786         */
2787        tcg_out_vreg3(s, INSN_VSHL_S, q, vece, a0, a2, a1);
2788        return;
2789    case INDEX_op_arm_ushl_vec:
2790        /* See above. */
2791        tcg_out_vreg3(s, INSN_VSHL_U, q, vece, a0, a2, a1);
2792        return;
2793    case INDEX_op_shli_vec:
2794        tcg_out_vshifti(s, INSN_VSHLI, q, a0, a1, a2 + (8 << vece));
2795        return;
2796    case INDEX_op_shri_vec:
2797        tcg_out_vshifti(s, INSN_VSHRI, q, a0, a1, (16 << vece) - a2);
2798        return;
2799    case INDEX_op_sari_vec:
2800        tcg_out_vshifti(s, INSN_VSARI, q, a0, a1, (16 << vece) - a2);
2801        return;
2802    case INDEX_op_arm_sli_vec:
2803        tcg_out_vshifti(s, INSN_VSLI, q, a0, a2, args[3] + (8 << vece));
2804        return;
2805
2806    case INDEX_op_andc_vec:
2807        if (!const_args[2]) {
2808            tcg_out_vreg3(s, INSN_VBIC, q, 0, a0, a1, a2);
2809            return;
2810        }
2811        a2 = ~a2;
2812        /* fall through */
2813    case INDEX_op_and_vec:
2814        if (const_args[2]) {
2815            is_shimm1632(~a2, &cmode, &imm8);
2816            if (a0 == a1) {
2817                tcg_out_vmovi(s, a0, q, 1, cmode | 1, imm8); /* VBICI */
2818                return;
2819            }
2820            tcg_out_vmovi(s, a0, q, 1, cmode, imm8); /* VMVNI */
2821            a2 = a0;
2822        }
2823        tcg_out_vreg3(s, INSN_VAND, q, 0, a0, a1, a2);
2824        return;
2825
2826    case INDEX_op_orc_vec:
2827        if (!const_args[2]) {
2828            tcg_out_vreg3(s, INSN_VORN, q, 0, a0, a1, a2);
2829            return;
2830        }
2831        a2 = ~a2;
2832        /* fall through */
2833    case INDEX_op_or_vec:
2834        if (const_args[2]) {
2835            is_shimm1632(a2, &cmode, &imm8);
2836            if (a0 == a1) {
2837                tcg_out_vmovi(s, a0, q, 0, cmode | 1, imm8); /* VORRI */
2838                return;
2839            }
2840            tcg_out_vmovi(s, a0, q, 0, cmode, imm8); /* VMOVI */
2841            a2 = a0;
2842        }
2843        tcg_out_vreg3(s, INSN_VORR, q, 0, a0, a1, a2);
2844        return;
2845
2846    case INDEX_op_cmp_vec:
2847        {
2848            TCGCond cond = args[3];
2849            ARMInsn insn;
2850
2851            switch (cond) {
2852            case TCG_COND_NE:
2853                if (const_args[2]) {
2854                    tcg_out_vreg3(s, INSN_VTST, q, vece, a0, a1, a1);
2855                } else {
2856                    tcg_out_vreg3(s, INSN_VCEQ, q, vece, a0, a1, a2);
2857                    tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a0);
2858                }
2859                break;
2860
2861            case TCG_COND_TSTNE:
2862            case TCG_COND_TSTEQ:
2863                if (const_args[2]) {
2864                    /* (x & 0) == 0 */
2865                    tcg_out_dupi_vec(s, type, MO_8, a0,
2866                                     -(cond == TCG_COND_TSTEQ));
2867                    break;
2868                }
2869                tcg_out_vreg3(s, INSN_VTST, q, vece, a0, a1, a2);
2870                if (cond == TCG_COND_TSTEQ) {
2871                    tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a0);
2872                }
2873                break;
2874
2875            default:
2876                if (const_args[2]) {
2877                    insn = vec_cmp0_insn[cond];
2878                    if (insn) {
2879                        tcg_out_vreg2(s, insn, q, vece, a0, a1);
2880                        return;
2881                    }
2882                    tcg_out_dupi_vec(s, type, MO_8, TCG_VEC_TMP, 0);
2883                    a2 = TCG_VEC_TMP;
2884                }
2885                insn = vec_cmp_insn[cond];
2886                if (insn == 0) {
2887                    TCGArg t;
2888                    t = a1, a1 = a2, a2 = t;
2889                    cond = tcg_swap_cond(cond);
2890                    insn = vec_cmp_insn[cond];
2891                    tcg_debug_assert(insn != 0);
2892                }
2893                tcg_out_vreg3(s, insn, q, vece, a0, a1, a2);
2894                break;
2895            }
2896        }
2897        return;
2898
2899    case INDEX_op_bitsel_vec:
2900        a3 = args[3];
2901        if (a0 == a3) {
2902            tcg_out_vreg3(s, INSN_VBIT, q, 0, a0, a2, a1);
2903        } else if (a0 == a2) {
2904            tcg_out_vreg3(s, INSN_VBIF, q, 0, a0, a3, a1);
2905        } else {
2906            tcg_out_mov(s, type, a0, a1);
2907            tcg_out_vreg3(s, INSN_VBSL, q, 0, a0, a2, a3);
2908        }
2909        return;
2910
2911    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
2912    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
2913    default:
2914        g_assert_not_reached();
2915    }
2916}
2917
2918int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
2919{
2920    switch (opc) {
2921    case INDEX_op_add_vec:
2922    case INDEX_op_sub_vec:
2923    case INDEX_op_and_vec:
2924    case INDEX_op_andc_vec:
2925    case INDEX_op_or_vec:
2926    case INDEX_op_orc_vec:
2927    case INDEX_op_xor_vec:
2928    case INDEX_op_not_vec:
2929    case INDEX_op_shli_vec:
2930    case INDEX_op_shri_vec:
2931    case INDEX_op_sari_vec:
2932    case INDEX_op_ssadd_vec:
2933    case INDEX_op_sssub_vec:
2934    case INDEX_op_usadd_vec:
2935    case INDEX_op_ussub_vec:
2936    case INDEX_op_bitsel_vec:
2937        return 1;
2938    case INDEX_op_abs_vec:
2939    case INDEX_op_cmp_vec:
2940    case INDEX_op_mul_vec:
2941    case INDEX_op_neg_vec:
2942    case INDEX_op_smax_vec:
2943    case INDEX_op_smin_vec:
2944    case INDEX_op_umax_vec:
2945    case INDEX_op_umin_vec:
2946        return vece < MO_64;
2947    case INDEX_op_shlv_vec:
2948    case INDEX_op_shrv_vec:
2949    case INDEX_op_sarv_vec:
2950    case INDEX_op_rotli_vec:
2951    case INDEX_op_rotlv_vec:
2952    case INDEX_op_rotrv_vec:
2953        return -1;
2954    default:
2955        return 0;
2956    }
2957}
2958
2959void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
2960                       TCGArg a0, ...)
2961{
2962    va_list va;
2963    TCGv_vec v0, v1, v2, t1, t2, c1;
2964    TCGArg a2;
2965
2966    va_start(va, a0);
2967    v0 = temp_tcgv_vec(arg_temp(a0));
2968    v1 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
2969    a2 = va_arg(va, TCGArg);
2970    va_end(va);
2971
2972    switch (opc) {
2973    case INDEX_op_shlv_vec:
2974        /*
2975         * Merely propagate shlv_vec to arm_ushl_vec.
2976         * In this way we don't set TCG_TARGET_HAS_shv_vec
2977         * because everything is done via expansion.
2978         */
2979        v2 = temp_tcgv_vec(arg_temp(a2));
2980        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
2981                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
2982        break;
2983
2984    case INDEX_op_shrv_vec:
2985    case INDEX_op_sarv_vec:
2986        /* Right shifts are negative left shifts for NEON.  */
2987        v2 = temp_tcgv_vec(arg_temp(a2));
2988        t1 = tcg_temp_new_vec(type);
2989        tcg_gen_neg_vec(vece, t1, v2);
2990        if (opc == INDEX_op_shrv_vec) {
2991            opc = INDEX_op_arm_ushl_vec;
2992        } else {
2993            opc = INDEX_op_arm_sshl_vec;
2994        }
2995        vec_gen_3(opc, type, vece, tcgv_vec_arg(v0),
2996                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
2997        tcg_temp_free_vec(t1);
2998        break;
2999
3000    case INDEX_op_rotli_vec:
3001        t1 = tcg_temp_new_vec(type);
3002        tcg_gen_shri_vec(vece, t1, v1, -a2 & ((8 << vece) - 1));
3003        vec_gen_4(INDEX_op_arm_sli_vec, type, vece,
3004                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(v1), a2);
3005        tcg_temp_free_vec(t1);
3006        break;
3007
3008    case INDEX_op_rotlv_vec:
3009        v2 = temp_tcgv_vec(arg_temp(a2));
3010        t1 = tcg_temp_new_vec(type);
3011        c1 = tcg_constant_vec(type, vece, 8 << vece);
3012        tcg_gen_sub_vec(vece, t1, v2, c1);
3013        /* Right shifts are negative left shifts for NEON.  */
3014        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
3015                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
3016        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
3017                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
3018        tcg_gen_or_vec(vece, v0, v0, t1);
3019        tcg_temp_free_vec(t1);
3020        break;
3021
3022    case INDEX_op_rotrv_vec:
3023        v2 = temp_tcgv_vec(arg_temp(a2));
3024        t1 = tcg_temp_new_vec(type);
3025        t2 = tcg_temp_new_vec(type);
3026        c1 = tcg_constant_vec(type, vece, 8 << vece);
3027        tcg_gen_neg_vec(vece, t1, v2);
3028        tcg_gen_sub_vec(vece, t2, c1, v2);
3029        /* Right shifts are negative left shifts for NEON.  */
3030        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
3031                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
3032        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t2),
3033                  tcgv_vec_arg(v1), tcgv_vec_arg(t2));
3034        tcg_gen_or_vec(vece, v0, t1, t2);
3035        tcg_temp_free_vec(t1);
3036        tcg_temp_free_vec(t2);
3037        break;
3038
3039    default:
3040        g_assert_not_reached();
3041    }
3042}
3043
3044static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
3045{
3046    int i;
3047    for (i = 0; i < count; ++i) {
3048        p[i] = INSN_NOP;
3049    }
3050}
3051
3052/* Compute frame size via macros, to share between tcg_target_qemu_prologue
3053   and tcg_register_jit.  */
3054
3055#define PUSH_SIZE  ((11 - 4 + 1 + 1) * sizeof(tcg_target_long))
3056
3057#define FRAME_SIZE \
3058    ((PUSH_SIZE \
3059      + TCG_STATIC_CALL_ARGS_SIZE \
3060      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
3061      + TCG_TARGET_STACK_ALIGN - 1) \
3062     & -TCG_TARGET_STACK_ALIGN)
3063
3064#define STACK_ADDEND  (FRAME_SIZE - PUSH_SIZE)
3065
3066static void tcg_target_qemu_prologue(TCGContext *s)
3067{
3068    /* Calling convention requires us to save r4-r11 and lr.  */
3069    /* stmdb sp!, { r4 - r11, lr } */
3070    tcg_out_ldstm(s, COND_AL, INSN_STMDB, TCG_REG_CALL_STACK,
3071                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
3072                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
3073                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_R14));
3074
3075    /* Reserve callee argument and tcg temp space.  */
3076    tcg_out_dat_rI(s, COND_AL, ARITH_SUB, TCG_REG_CALL_STACK,
3077                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
3078    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
3079                  CPU_TEMP_BUF_NLONGS * sizeof(long));
3080
3081    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
3082
3083    if (!tcg_use_softmmu && guest_base) {
3084        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_GUEST_BASE, guest_base);
3085        tcg_regset_set_reg(s->reserved_regs, TCG_REG_GUEST_BASE);
3086    }
3087
3088    tcg_out_b_reg(s, COND_AL, tcg_target_call_iarg_regs[1]);
3089
3090    /*
3091     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
3092     * and fall through to the rest of the epilogue.
3093     */
3094    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
3095    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, 0);
3096    tcg_out_epilogue(s);
3097}
3098
3099static void tcg_out_epilogue(TCGContext *s)
3100{
3101    /* Release local stack frame.  */
3102    tcg_out_dat_rI(s, COND_AL, ARITH_ADD, TCG_REG_CALL_STACK,
3103                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
3104
3105    /* ldmia sp!, { r4 - r11, pc } */
3106    tcg_out_ldstm(s, COND_AL, INSN_LDMIA, TCG_REG_CALL_STACK,
3107                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
3108                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
3109                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_PC));
3110}
3111
3112static void tcg_out_tb_start(TCGContext *s)
3113{
3114    /* nothing to do */
3115}
3116
3117typedef struct {
3118    DebugFrameHeader h;
3119    uint8_t fde_def_cfa[4];
3120    uint8_t fde_reg_ofs[18];
3121} DebugFrame;
3122
3123#define ELF_HOST_MACHINE EM_ARM
3124
3125/* We're expecting a 2 byte uleb128 encoded value.  */
3126QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
3127
3128static const DebugFrame debug_frame = {
3129    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
3130    .h.cie.id = -1,
3131    .h.cie.version = 1,
3132    .h.cie.code_align = 1,
3133    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
3134    .h.cie.return_column = 14,
3135
3136    /* Total FDE size does not include the "len" member.  */
3137    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
3138
3139    .fde_def_cfa = {
3140        12, 13,                         /* DW_CFA_def_cfa sp, ... */
3141        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
3142        (FRAME_SIZE >> 7)
3143    },
3144    .fde_reg_ofs = {
3145        /* The following must match the stmdb in the prologue.  */
3146        0x8e, 1,                        /* DW_CFA_offset, lr, -4 */
3147        0x8b, 2,                        /* DW_CFA_offset, r11, -8 */
3148        0x8a, 3,                        /* DW_CFA_offset, r10, -12 */
3149        0x89, 4,                        /* DW_CFA_offset, r9, -16 */
3150        0x88, 5,                        /* DW_CFA_offset, r8, -20 */
3151        0x87, 6,                        /* DW_CFA_offset, r7, -24 */
3152        0x86, 7,                        /* DW_CFA_offset, r6, -28 */
3153        0x85, 8,                        /* DW_CFA_offset, r5, -32 */
3154        0x84, 9,                        /* DW_CFA_offset, r4, -36 */
3155    }
3156};
3157
3158void tcg_register_jit(const void *buf, size_t buf_size)
3159{
3160    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
3161}
3162